paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1908.05144 | 1 | 1908 | 2019-08-14T14:26:56 | Cardiac Mechano-Electrical Dynamical Instability | [
"physics.bio-ph",
"physics.med-ph"
] | In a computational study we reveal a novel dynamical instability of excitation waves in the heartmuscle. The instability manifests itself as gradual local increase in the duration of the actionpotential which causes formation and hypermeandering of spiral waves. The mechanism is causedby stretch-activated currents that cause wave front-tail collisions and beat to beat elongation of theaction potential duration due to biexcitability. We discuss the importance of the instability for theonset and dynamics of cardiac arrhythmias. | physics.bio-ph | physics |
Cardiac Mechano-Electrical Dynamical Instability
L. D. Weise1, 2 and A. V. Panfilov1
1Department of Physics and Astronomy, Ghent University, Krijgslaan 281, S9, Ghent, 9000, Belgium
2Theoretical Biology, Utrecht University, Padualaan 8, Utrecht, 3584 CH, The Netherlands
(Dated: August 14, 2019)
In a computational study we reveal a novel dynamical instability of excitation waves in the heart
muscle. The instability manifests itself as gradual local increase in the duration of the action
potential which causes formation and hypermeandering of spiral waves. The mechanism is caused
by stretch-activated currents that cause wave front-tail collisions and beat to beat elongation of the
action potential duration due to biexcitability. We discuss the importance of the instability for the
onset and dynamics of cardiac arrhythmias.
Spiral waves of excitation have been found in many
biological, physical, and chemical systems [1 -- 3]. Spi-
ral waves emerge in excitable media after wave break, a
temporary, local block of wave propagation, after which
the wave curls around it's back forming a spiral [4, 5].
The emergence of spiral waves in the heart muscle causes
life-threatening cardiac arrhythmias [6]. Therefore it is
of great interest to understand mechanisms that cause
wave break in the heart. Break formation can be a re-
sult of anatomical heterogeneity [7] or dynamical insta-
bility. The most studied dynamical instability in car-
diac tissue is "alternans" which can occur via various
mechanisms [8]. Alternans manifests in as a beat-to-beat
alternation in the duration of action potentials (short-
long-short), which grows in time and may result in wave
break formation.
The heart's contractions are governed by electrical
waves of excitation. Conversely, its deformation affects
the excitation processes of the cardiomyocytes, which is
called "mechano-electrical feedback" (MEF). MEF has
been shown to be able to cause, but also to abolish dan-
gerous cardiac arrhythmias [9]. However, so far no dy-
namical instability which is caused by MEF has been
identified yet. In this letter we report the finding of such
a mechano-electrical dynamical instability (MEDI) in a
model for human cardiac tissue.
Our method couples an ionic model for human epicar-
dial myocytes [10], with a discrete mechanical model for
cardiac tissue [11], and a model for excitation-contraction
coupling [12, 13] adjusted to human cardiac tissue [14].
The propagation of nonlinear waves of electrical excita-
tion in cardiac tissue is modeled via a reaction-diffusion
equation for the transmembrane potential V
∂V
∂t
= D∆V − Iion + Isac
Cm
,
(1)
with membrane capacitance density Cm = 2.0 µF/cm2
and diffusivity Dij = δij × 1.54 cm2/s. At bound-
aries of the medium no-flux boundary conditions are used
(∇V = 0). The transmembrane ion current Iion is mod-
eled by various time- and voltage-dependent ion chan-
nels [15]. The finite difference mesh for the explicit Euler
integration (space step 0.25 mm and time step 0.02 ms)
of Eq.(1) is coupled to a square lattice of mass points
connected with springs (see Figure 1 in [16]). Excitation
waves trigger a contraction of the tissue [16]. To solve
the mechanical model we assumed elastostatics, and used
Verlet integration [17]. To model MEF we use a linear,
time-independent model for stretch-activated currents
Isac = Gs
(λ − 1)
(λmax − 1)
(V − Es) , for λ > 1
(2)
where λ is a measure for local dilatation: strain in one-
dimensional (1D) simulations, and square root of the
relative area change of a quadrilateral formed by di-
rect neighboring mass points (see Figure 1 in [16]) in
two-dimensional (2D) simulations. Parameter λmax is
maximal normalized sarcomere length which we chose
as λmax = 1.1 as in [14]. Gs is the maximal conduc-
tance, Es the reversal potential of the stretch activated
channels. Es was measured in a range from −20 mV
to 0 mV [18, 19]. We set Es = 0 mV . We vary Gs in the
reported range from 0 to 100 S/F [9, 20]. Following sim-
ilar studies [21 -- 23] we fixed the boundaries of the model
to mimic isovolumic phases of the cardiac cycle. Our 2D
model relates to a thin slice of cardiac tissue with fixed
boundaries. For 1D simulations we assumed a constantly
stretched cable (λ = λmax) and vary Gs in Eq.(2).
Figure 1 and supplemental movie [35] shows develop-
ment of MEDI under periodic stimulation of cardiac tis-
sue. First, we stimulate the tissue with a constant period
of 340 ms and observe stable wave propagation. How-
ever, when we decrease the stimulation period to 300 ms,
we see development of MEDI. In Figure 1A, top we
show wave front-back collisions before wave break hap-
pens (compare supplemental movies [36] and figure [37]).
We see next, that at this location APD [38] gradually in-
creases, while DI decreases (Figure 1A), until wave break
occurs evolving to two counter-rotating spiral waves (Fig-
ure 1B). Note also, that in contrast to alternans insta-
bility MEDI occurs for longer stimulation periods than
classical APD alternans [10], and does not involve alter-
nations between long and short action potentials.
How robust is MEDI against the change of model pa-
rameters? To answer this question we used the setup
2
motes wave front-back collision. Here we have a similar
situation; however, the stretch is not caused by an ex-
ternal mechanical load, but the contraction is caused by
the excitation wave itself. The mechanism is the same:
stretch causes Isac which accelerates the wave front caus-
ing wave front-back collision.
However, why does the APD at the collision position
grow from wave to wave (see Figure 1A)? This is counter-
intuitive, because classical restitution theory predicts
that collision (short DI) should produce shorter APD.
However, classical restitution theory cannot be applied
here as the collision is a non-stationary spatio-temporal
process which cannot be reproduced by a periodic stim-
ulation of cardiac cell. Here one needs to consider the
interaction of the wavefront with the waveback of the
preceeding wave [39].
To systematically study wave front-back collisions we
developed a special electrophysiological setup.
In this
setup we use a moving obstacle in a cable to control the
velocity of a first wave (S1) (see also Figure 3 in [24]), and
initiate a wave train (see Figure 3). To study the effect
of MEF we assume the fiber to be constantly stretched
to λmax, thus Isac = Gs(V − Es). By changing the ve-
locity of the moving obstacle we can systematically vary
the degree of wave front-back interaction during collision.
We can see that we can reproduce the observed MEDI
FIG. 3: Stepwise APD increase and wave block in constantly
stretched cable. Top: Time-space plot. S1 is forced to con-
stant CV of 17.65 cm/s by obstacle (thick line). Middle panel:
APD of a wave train (S6 -- S10) vs position. Lower panel: DI
of a wave train (S6 -- S10) vs position. Waves were started
with period 290 ms in 5 cm long constantly stretched cable
(λ = λmax). Cable was prepared with twelve waves before S1
wave. Gs = 10.5 S/F .
FIG. 1: MEDI. Periodical wave initiation in left, upper cor-
ner: ten waves with period 340 ms, then seven waves with
period 300 ms. Five last waves ( 1 -- 5 ) are shown.
(A)
1 -- 4 : Wave front-back collisions and emergence of spatial
heterogeneity. Upper panel: Transmembrane voltage (V);
snapshots are taken, when wave front-back collision happens.
White cross indicates region where DI is minimal and MEDI
develops. Middle panel: Action potential duration (APD).
Lower panel: Diastolic interval (DI). (B) Wavebreak and spi-
ral formation. Time [ms] after first wave front-back colli-
sion is shown above a snapshot. Side length of model 15 cm.
Gs = 50 S/F .
shown in Figure 1; however, slowly decreased the period
of stimulation for different Gs. We show the results in
Figure 2. We see from Figure 2 that MEDI occurs in a
FIG. 2: Wave patterning as a function of period of stimula-
tion (T) and Gs. Robustness of MEDI. Above the line: stable
wave propagation. Thick line: onset of dynamical instability
resulting in wave break. Crosses: measurements. Dark grey
area: MEDI. Light grey area: alternans instability. Waves
were started with an initial period of 0.5 s. After ten stim-
ulations the period was decreased by 2.5 ms, and this was
repeated until wave break happened. Setup as in Figure 1
was used.
large parametric space of Gs and period of stimulation T .
We will now explain the mechanism of MEDI. First, we
need to explain what causes wave front-back collisions.
We reported in [24] that external dynamic stretching of
cardiac tissue causes acceleration of a wave front and pro-
02805708701704908101090320220APD[ms]20100DI[ms]35-87V[mV]1122334AB119013101380153016304550154575200400MEDIAlternansGs [S/F]T [ms]stable wave propagationmoving block (GNa=GCaL=0)2position [cm]11.52.56065DI [ms]220240260APD [ms]S9S8S7S6S10012345time [s]0123S1S2S3S4S5S6S7S8S9S10S11 in this setup by choosing a corresponding value of the
forced velocity (Figure 3). In particular, for forced ve-
locity of 17.65 cm/s we see collisions of successive waves,
gradual increase in APD (Figure 3B) and gradual de-
crease of the DI at the collision position (Figure 3C).
This process closely resembles the instability in the 2D
system (compare Figure 3 and Figure 1).
We performed additional simulations to systematically
study front-tail interactions by letting two waves (S1 and
S2) collide for different velocities of the moving obstacle.
Figure 4 shows wave characteristics at the collision point,
i.e. when SI between S1 and S2 is minimal. Note, that
this point has a different location for different forced CV.
We see from Figure 4A that, as expected, lower forced
CV results in closer front-tail interaction (DI decreases).
However, we also see that such decrease in DI results
in unexpected increase in APD which is counter to nor-
mal APD restitution, where shorter DI results in shorter
APD. Such abnormal dependency, can, in our view, ex-
plain the observed MEDI. Indeed, for periodic forcing
with a period T, DI=T-APD, thus increase in APD will
result in decrease in DI. However, if decrease in DI will
produce longer APD, as in the case of Fig. 4A this longer
APD will produce shorter DI and will further increase
APD and thus result in its gradual growth, what we see
in Figure 1 and Figure 3.
FIG. 4: APD elongation is caused by biexcitability. Electro-
physiological observables at position of S2-S1 collision (when
DI between S2 and S1 is minimal) vs forced CV. (A) Top:
APD vs forced CV. Bottom: DI vs forced CV. (B) Electro-
physiological observables of S2 wave vs time at S2-S1 collision
position as a function of forced CV. Time is shifted to start
of upstroke, when V = −60 mV . Top: transmembrane volt-
age V. Other panels: strongest polarizing and depolarizing
currents Waves were started with period 0.5 s in 20 cm long
cable. Cable was prepared with twelve waves before S1 wave.
Gs = 10.5 S/F .
3
How does
this "abnormal APD(DI) dependency
emerge? We find that it is related to the phenomenon
"biexcitability".
Under special conditions [25, 26]
bistable wave propagation can occur in the same tissue.
One type is fast propagation: characterized by a rapid
sodium-driven upstroke (INa) happening from the repo-
larized transmembrane potential, and the other is a slow
propagation where the upstroke is driven by L-type cal-
cium current (ICaL) from a depolarized potential when
sodium channels are mostly inactivated due to accom-
modation [27]. We found earlier in the moving obsta-
cle setup, that Isac causes biexcitability of the S2 wave
which manifests in its oscillation between sodium- (when
S2 is distant from the S1 wave back) and calcium-driven
upstroke (during wave front-back collision) [24]. Here we
find that such a transition of wave front propagation sub-
stantially affects APD. We can see (Figure 4B) that for
a forced CV of 17.5 cm/s, where APD is 219 ms the
upstroke of the action potential is steep, and driven by
sodium current (dotted lines). However, for slower forced
CV (straight and dashed lines) we can see that the slope
of the action potential becomes shallow, sodium current
is absent, and L-calcium current is the main depolariz-
ing current. The action potentials are also substantially
longer, 292 ms for forced CV of 13.75 cm/s, and 395 ms
for forced CV of 4.81 cm/s. We can explain the APD
elongation by a combination of the longer transient of
the calcium current, and a delay of the repolarizing cur-
rents IKs and IKr for slower forced CV happening as a
consequence of the change of the propagation type (Fig-
ure 4B lower panels). We studied the importance of ICaL
for the mechanism. In the setup of Figure 3 we found
that wavebreaks due to MEDI occur for forced velocities
[12.4; 15.26] cm/s. However, if we block ICaL we did not
observe MEDI; we could either see stable wave propaga-
tion (forced CV > 15.26 cm/s) or immediate block of S2
(forced CV < 15.26 cm/s), and no MEDI. Thus we can
conclude that ICaL and related to it biexcitability is a
key part of MEDI.
Overall we can explain the mechanism of MEDI as fol-
lows: MEF due to stretch-activated currents increases
the local velocity of the wavefront which causes wave
front-back collisions. A wave front-back collision results
in short DI, which however causes longer APDs at the col-
lision regions due to biexcitability. Elongation of APD
further decreases DI in the collision region which further
increases APD. This positive feedback results in wave by
wave increase of APD until the wave front dissipates.
How relevant is this mechanism for cardiac arrhyth-
mias? We showed in Figures 1, 2 that MEDI can lead
to wave break and creation of spiral waves in the heart.
Therefore MEDI may be relevant for the onset of cardiac
arrhythmia. Does MEDI also destabilize spiral waves?
We also studied what effect MEDI has on spiral wave
dynamics. We show the results in Figure 5. We see
that for Gs = 12.5 S/F the spiral wave has a circu-
-600-20-10-4-20100.40100200300400time [ms]voltage[mV]INa[A/F]ICaL[A/F]IKr[A/F]IKs[A/F]4.8113.7517.5forced CV [cm/s]APD[ms]DI[ms]030603004005101520forced CV [cm/s]ABno block 4
third of the wave velocity (see inset in subfigure 5A, wave
velocity is ca. 60 cm/s) which is too slow to cause ECG
patterns similar to ventricular fibrillation [31].
We found that the rapid spiral drift is caused by MEDI.
We illustrate it in Figure 5 for the spiral core trajectory
Gs = 37.5 S/F (green line), where we show collisions
I -- V. In subfigure 5B we show that MEDI occurs. Sim-
ilar to Figure 1A we see that APD gradually increases
(see collisions I -- III in subfigure 5B). We also see that it
causes wave block close to the spiral wave tip which does
not result in full spiral breakup. We observe that the spi-
ral core drifts along the "collision line" (region where DI
is minimal and APD maximal). We indicate the "drift
vector" in subfigure 5B,V as an arrow.
In this letter we reported on our finding of a novel dy-
namical instability "MEDI" for excitation waves in car-
diac tissue. MEDI emerges as a consequence of MEF
and causes the formation and hypermeandering of spiral
waves. These phenomena are relevant for the onset of
cardiac arrhythmias. MEDI causes wave break in a large
range of conductivities of stretch-activated channels and
stimulation periods. MEDI can occur for longer stimula-
tion periods than the alternans instability.
It is difficult to formulate an analytical theory for
MEDI. This is, because it emerges from the interplay
of the complex phenomena CV and biexcitability that
themself depend on the interplay of wave propagation
and MEF. To explain the mechanism we studied the wave
by wave increase of APD in a simplified 1D setup (Fig-
ure 4), in which we disabled the electromechanical cou-
pling. However, in the full model it is more complex, as
APD affects also the spatiotemporal strain distribution
in the medium, and thus affects CV. We see no easy way
to analytically reduce the complexity of this spatiotem-
poral problem.
We studied MEDI in a simplified 2D model for cardiac
tissue. As a next step it may be interesting to investigate
the novel mechanism in more detailed three-dimensional
electromechanical models for the human heart [14, 32].
For example, it is important to test if the novel insta-
bility can cause a breakup of vortices, as this is a key
mechanism for sudden cardiac death. It may be possi-
ble to experimentally study MEDI, for example by using
ultrasound-based strain imaging [33] in animal models.
It can also be interesting to design experiments similar
to our forced CV setup (compare Figure 4) using the
optogenetics approach [34].
L.D.W.
Acknowledgements:
thanks the Deutsche
Forschungsgemeinschaft for a research fellowship (Grant
No. WE 5519/1-1). We thank Dr. Vadim Biktashev, Dr.
Jan Kucera Dr. Hans Dierckx and Dr. Ivan Kazbanov
for valuable discussions. We thank Dr. Paul Baron for
critical comments on the manuscript.
FIG. 5: MEDI causes rapid spiral drift. (A) spiral tip trajec-
tories as function of Gs. Inset: drift velocity as function of Gs.
(B) APD and DI after wave back-front collisions show MEDI.
Lower, right quadrant of the medium is shown. Location of
spiral core is illustrated by a white dot. Snapshots are taken
for Gs = 37.5 S/F (compare subfigure A). In snapshot V we
illustrate position of spiral core of snapshot I as black dot,
and indicate drift direction with an arrow. (C) Illustration of
collision II (compare subfigure B) leading to temporary wave
block close to spiral tip. Time [ms] (starting at 2488 ms
simulation time) is shown above each snapshot. Length of
medium 12.5 cm. Spiral was initiated in the medium without
MEF (Gs = 0 S/F ), let rotate for 2 s, then system was saved
and used as starting point for simulations with MEF.
lar core which is the same as dynamics in absence of
MEF. However, for Gs = 25 S/F MEF causes mean-
dering of the spiral wave on a cycloidal trajectory. This
onset of meander can be explained by the resonant drift
theory [28] which predicts meandering under a period-
ical variation of the excitability of the medium, which
occurs here due to MEF [23, 29]. However, for larger val-
ues of Gs we observe a hyper-meandering trajectory (see
the orange line). Such hyper-meandering has substantial
consequences for the type of arrhythmia, as it induces
polymorphic ventricular tachycardia [30]. We found that
the maximal induced drift velocity is approximately one
12.52537.55062.575Gs [S/F]30456075Gs [S/F]0481216max. drift velocity [cm/s]06.256.25x [cm]y [cm]AIIIIVVIIIBC290190APD[ms]20100DI[ms]35-87V[mV]6.25012.56.25y[cm]052122152284IIIIIIIVVII[cm]x 5
[35] LINK TO SUPPLEMENTAL MOVIE
[36] LINK TO SUPPLEMENTAL MOVIE
[37] LINK TO SUPPLEMENTAL FIGURE
[38] APD and DI are recorded at −60 mV
[39] compare APD vs. DI plot for dynamic restitution proto-
col and collisions LINK TO SI
[1] M. A. Allessie, F. I. M. Bonke, and F. J. G. Schopman,
Circ. Res. 39, 168 (1976).
[2] G. Gerisch, Wilhelm Roux' Archiv fur Entwick-
lungsmechanik der Organismen 156, 127 (1965).
[3] K. Agladze and O. Steinbock, The Journal of Physical
Chemistry A 104, 9816 (2000).
[4] B. P. Belousov, Collection of short papers on radiation
medicine for 1958 (Med. Publ., Moscow, 1959), pp. 145 --
147, in Russian.
[5] A. Zaikin and A. Zhabotinsky, Nature 225, 535 (1970).
[6] A. Winfree and S. Strogatz, Nature 311, 611 (1984).
[7] D. P. Zipes and H. J. J. Wellens, Circulation 98, 2334
(1998), 01659.
[8] A. Garfinkel, Journal of electrocardiology 40, S70 (2007).
[9] P. Kohl, P. Hunter, and D. Noble, Prog. Biophys. Molec.
Biol. 71, 91 (1999).
[10] K. Ten Tusscher and A. Panfilov, Am. J. Physiol. Heart
Circ. Physiol. 291, H1088 (2006).
[11] L. D. Weise, M. P. Nash, and A. V. Panfilov, PLoS ONE
6(7), e21934 (2011).
[12] S. Niederer, P. Hunter, and N. Smith, Biophys. J. 90,
1697 (2006).
[13] S. Niederer and N. Smith, Prog. Biophys. Mol. Biol. 96,
90 (2008).
[14] R. H. Keldermann et al., Am J Physiol Heart Circ Physiol
299, H134 (2010).
[15] A. Hodgkin and A. Huxley, J. Physiol. 117, 500 (1952).
[16] L. D. Weise and A. V. Panfilov, PLoS ONE 8, e59317
(2013).
[17] L. Verlet, Phys. Rev. 159, 98 (1967).
[18] P. Kohl, P. Hunter, and D. Noble, Progress in Biophysics
and Molecular Biology 71, 91 (1999).
[19] K. Skouibine, N. Trayanova, and P. Moore, Math. Biosci.
166, 85 (2000).
[20] P. Kohl, K. Day, and D. Noble, Can. J. Cardiol. 14, 111
(1998).
[21] L. D. Weise and A. V. Panfilov, Phys. Rev. Lett. 108,
228104 (2012).
[22] A. Panfilov, R. Keldermann, and M. Nash, Phys. Rev.
Lett. 95, 258104 (2005).
[23] A. Panfilov, R. Keldermann, and M. Nash, Proc. Natl.
Acad. Sci. U.S.A. 104, 7922 (2007).
[24] L. D. Weise and A. V. Panfilov, Phys. Rev. Lett. 119,
108101 (2017).
[25] N. Vandersickel et al., PloS One 9, e84595 (2014).
[26] M. G. Chang et al., Heart Rhythm: The Official Journal
of the Heart Rhythm Society 9, 115 (2012).
[27] Z. Qu and J. N. Weiss, Annual Review of Physiology 77,
29 (2015).
[28] S. Grill, V. Zykov, and S. Muller, Phys. Rev. Lett. 75,
3368 (1995).
[29] H. Dierckx et al., New Journal of Physics 17, 043055
(2015).
[30] R. A. Gray et al., Circulation 91, 2454 (1995).
[31] R. Gray et al., Science 270, 1222 (1995).
[32] N. A. Trayanova, J. Constantino, and V. Gurev, Ameri-
can journal of physiology. Heart and circulatory physiol-
ogy 301, H279 (2011).
[33] J. Christoph et al., Nature 555, 667 (2018).
[34] N. Magome et al., Tissue Engineering Part A 17, 2703
(2011).
|
1509.05924 | 1 | 1509 | 2015-09-19T19:14:04 | Multiscale model of a freeze-thaw process for tree sap exudation | [
"physics.bio-ph"
] | Sap transport in trees has long fascinated scientists, and a vast literature exists on experimental and modelling studies of trees during the growing season when large negative stem pressures are generated by transpiration from leaves. Much less attention has been paid to winter months when trees are largely dormant but nonetheless continue to exhibit interesting flow behaviour. A prime example is sap exudation, which refers to the peculiar ability of sugar maple (Acer saccharum) and related species to generate positive stem pressure while in a leafless state. Experiments demonstrate that ambient temperatures must oscillate about the freezing point before significantly heightened stem pressures are observed, but the precise causes of exudation remain unresolved. The prevailing hypothesis attributes exudation to a physical process combining freeze-thaw and osmosis, which has some support from experimental studies but remains a subject of active debate. We address this knowledge gap by developing the first mathematical model for exudation, while also introducing several essential modifications to this hypothesis. We derive a multiscale model consisting of a nonlinear system of differential equations governing phase change and transport within wood cells, coupled to a suitably homogenized equation for temperature on the macroscale. Numerical simulations yield stem pressures that are consistent with experiments and provide convincing evidence that a purely physical mechanism is capable of capturing exudation. | physics.bio-ph | physics | Multiscale model of a freeze -- thaw process
for tree sap exudation
By Isabell Graf1, Maurizio Ceseri2, and John M. Stockie1,∗
1Department of Mathematics, Simon Fraser University, 8888 University Drive,
Burnaby, British Columbia, V5A 1S6, Canada
2Istituto per le Applicazioni del Calcolo 'Mauro Picone', via dei Taurini 19,
Consiglio Nazionale delle Ricerche, Rome, 00185, Italy
∗Author for correspondence: John M. Stockie ([email protected])
Sap transport in trees has long fascinated scientists, and a vast literature exists
on experimental and modelling studies of trees during the growing season when
large negative stem pressures are generated by transpiration from leaves. Much
less attention has been paid to winter months when trees are largely dormant but
nonetheless continue to exhibit interesting flow behaviour. A prime example is sap
exudation, which refers to the peculiar ability of sugar maple (Acer saccharum) and
related species to generate positive stem pressure while in a leafless state. Exper-
iments demonstrate that ambient temperatures must oscillate about the freezing
point before significantly heightened stem pressures are observed, but the precise
causes of exudation remain unresolved. The prevailing hypothesis attributes exu-
dation to a physical process combining freeze -- thaw and osmosis, which has some
support from experimental studies but remains a subject of active debate. We ad-
dress this knowledge gap by developing the first mathematical model for exudation,
while also introducing several essential modifications to this hypothesis. We derive a
multiscale model consisting of a nonlinear system of differential equations governing
phase change and transport within wood cells, coupled to a suitably homogenized
equation for temperature on the macroscale. Numerical simulations yield stem pres-
sures that are consistent with experiments and provide convincing evidence that a
purely physical mechanism is capable of capturing exudation.
Keywords: Tree sap exudation; sugar maple; multiphase flow and transport;
phase change; differential equations; periodic homogenization
1. Introduction
The study of tree sap flow has a long history that has given rise over time to the
concept of the hydraulic architecture of trees [54]. Despite the extensive literature
on this subject, several aspects of sap transport remain controversial, including the
cohesion-tension theory of sap ascent [4, 51, 58]; embolism formation and recov-
ery [35, 59], which is ubiquitous in species subject to drought- or freezing-induced
stresses; and sap exudation in maple and related species such as walnut, butternut
and birch [11]. Furthermore, there is a great deal of current interest in the possible
effects of recent changes in weather patterns on both individual trees and forest
ecosystems [5, 20], and their connections with sap hydraulics [46]. The problems
just described involve complex interactions between sap flow and other phenomena
Article submitted to Royal Society
LATEX Paper
5
1
0
2
p
e
S
9
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
4
2
9
5
0
.
9
0
5
1
:
v
i
X
r
a
2
I. Graf, M. Ceseri and J. M. Stockie
such as nutrient transport, photosynthesis, soil physics, atmospheric dynamics, cell
growth, etc. Despite the extensive work to date on mathematical and computational
modelling of trees and their interactions with the environment, many open questions
remain that can only be addressed by considering sap flow coupled with other pro-
cesses and building integrated models that connect flow and structure at different
spatial scales and levels of organization [29].
Sugar maple is a keystone species in the forests of central and eastern North
America [38] and so is worthy of special attention. Members of the maple family
are distinguished from other hardwoods by a number of unusual structural and
functional features that allow them to exude sap during winter [37, 49], to gen-
erate unusually high rates of nitrification [31], or to recover from freeze-induced
embolism [45, 57]. The potential impacts of climate change on maple have also at-
tracted recent attention [31, 38], motivated by the economic importance of the maple
syrup industry, not to mention maple's high timber value. In particular, maple sap
yields are sensitive to even small variations in temperature or snow cover during the
harvest season, so that recent unusual weather patterns underscore the importance
of developing a better understanding of the effects of local environmental conditions
on sap flow [26, 41].
Hundreds of scientific papers have addressed the phenomenon of sap exudation
during winter when maple trees are leafless and yet still exhibit pressure variations
that range over 150 -- 180 kPa [3, 11, 12, 49]. However, the precise mechanism driving
the generation of heightened exudation pressure is still not fully understood [54].
The first systematic study appeared in an 1860 article by Sachs [43], who attributed
exudation pressure to thermal expansion of gas within sapwood or xylem. The next
major advance in understanding followed from the exhaustive study of Wiegand [56],
who found to the contrary that thermal expansion of gas, water or wood has min-
imal impact on exudation. Instead, Wiegand proposed a vitalistic or 'living cell'
hypothesis wherein sugar is released into the sap by some cellular activity, which
gives rise to elevated pressure from osmotic gradients across selectively-permeable
membranes separating wood cells. Subsequently, this osmotic mechanism figured
prominently in the literature, although experimental studies have continued to yield
conflicting results that in turn stimulated development of new theories advocating
alternate (bio-)physical mechanisms. For example, some authors continued to sup-
port the thermal expansion hypothesis [36], while others advocated the various roles
of gas dissolution [24], cryostatic suction due to freezing [48], or temperature-induced
changes in bark thickness [32]. More recent studies have led to a new understanding
of exudation as a physical process deriving from a combination of freezing and thaw-
ing of sap [37] with osmosis [50]. Although some experimental evidence supports this
hypothesis [11] the precise mechanisms behind the exudation phenomenon is still
not fully understood.
We aim to resolve this long-standing open question by developing the first math-
ematical model for the freeze -- thaw process in maple. We uncover the essential role
played by two physical mechanisms whose significance has not yet been recognized
-- namely, root water uptake and freezing point depression due to sap sugar content.
Using numerical simulations of repeated freeze -- thaw cycles, we obtain computed
exudation pressures that are consistent with experimental results.
Although the focus of this paper is on developing a complete and physically
consistent model for sap exudation, our results also have more far-reaching conse-
Article submitted to Royal Society
Multiscale model for sap exudation
3
quences. This work affords new insights into the complex multi-physics processes
occurring in trees and also provides a framework for studying other practical ques-
tions of importance to tree physiologists and maple syrup producers. Our model also
provides a platform for studying related phenomena such as embolism that occur
in a much broader range of tree species, as well as evaluating the response of trees
to changes in environmental variables such as temperature and soil moisture arising
from various climate change scenarios.
2. Physical mechanism for sap exudation
(a) Milburn and O'Malley's hypothesis
Experimental work up to the 1980's demonstrated that no single physical mech-
anism is capable of capturing measured winter stem pressures [33], and sap exuda-
tion remained an unsolved puzzle until the ground-breaking study of Milburn and
O'Malley [37]. They proposed a physical mechanism based on freezing and thawing
of sap, motivated by the unique structural characteristic of xylem in maple (and
related trees) that a significant proportion of the libriform fibers (or simply fibers)
are primarily gas-filled rather than being liquid-filled as in most other hardwood
species [56]. This peculiar feature of the fibers should be contrasted with the two
other cell types that play an active role in sap transport -- vessels and tracheids --
which are mostly sap-filled and are connected hydraulically to each other via paired
pits (see figure 1a). Indeed, recent experiments [11, 44] suggest that fibers are essen-
tially non-conductive in comparison with other xylem elements because they lack
end-to-end cell connections and their lateral walls contain mostly unpaired (blind)
pits that are smaller and fewer in number than the more conductive vessels and
tracheids.
Milburn and O'Malley focused on the dynamics of a single fiber -- vessel pair as
pictured in figure 1b, and ignored other xylem elements such as parenchyma and
ray cells. Whereas fibers had previously been thought to play a purely structural
role, Milburn and O'Malley proposed that as temperature falls below freezing, liquid
from the vessel is drawn by cryostatic suction through the porous fiber -- vessel wall to
freeze on the inside of the fiber (figure 2, stages 1-2-3). As a result, any gas contained
within the fiber is compressed and acts as a pressure reservoir. When temperature
subsequently rises and ice thaws, the process reverses and the compressed gas bubble
forces melt-water into the vessel, thereby re-pressurizing the vessel sap (figure 2,
stages 3-4-1).
(b) Tyree's modified hypothesis, with gas dissolution and osmosis
This freeze -- thaw hypothesis was critically evaluated by Tyree [50], who proposed
a modified hypothesis featuring two important additions. First, he recognized that
gas under pressure will dissolve within an adjacent liquid [27] and that pressures
encountered in maple xylem are high enough that any bubbles should dissolve com-
pletely given sufficient time [52]. Therefore, some additional mechanism is required
to sustain gas bubbles in the fibers. Tyree's second observation was that measured
xylem pressures depend strongly on sugar concentration in the vessel sap, 80% of
which derives from sucrose [12], which led him to conclude that sucrose is required
for exudation. He recognized that although the axial conductivity of fibers is neg-
Article submitted to Royal Society
4
I. Graf, M. Ceseri and J. M. Stockie
(a)
tracheid
libriform fiber
vessel
parenchyma cell
pit
(b)
Rf
v
R
v
L
Vessel
Lf
100 µm
Fiber
Figure 1: Xylem microstructure. (a) Cross-sectional view of hardwood xylem,
showing tracheids connected hydraulically to vessels and other tracheids via paired
pits. Fibers appear similar to tracheids except that they have fewer pits, most of
which are blind or unpaired. The parenchyma are living cells whose main role is
carbohydrate storage and so they are ignored here. (b) A fiber -- vessel pair ap-
proximated as circular cylinders, showing typical dimensions of the fiber (length
Lf = 1.0 × 10−3 m and radius Rf = 3.5 × 10−6 ) and vessel (Lv = 5.0 × 10−4 m and
Rv = 2.0 × 10−5 m). The model domain corresponds to the horizontal cross-section
through the middle of the diagram.
ligible in comparison with vessels and tracheids, the lignified cellulose making up
secondary cell walls should admit a small radial conductivity. He then hypothesized
that the fiber -- vessel wall forms an osmotic barrier that allows water to penetrate
but not the larger sucrose molecules. Consequently, an additional osmotic pressure
difference exists between the sweet vessel sap and pure fiber water, which he argued
is responsible for preventing fiber gas bubbles from completely dissolving.
Tyree's modified freeze -- thaw hypothesis includes water phase transitions, gas
dissolution and osmosis, and is currently the prevailing hypothesis for sap exuda-
tion [50]. It depends strongly on the existence of a hydraulically isolated system of
fibers and a selectively permeable fiber -- vessel wall, both of which have since been
confirmed experimentally [11]. Although this evidence is compelling, there has been
no attempt yet to model this process mathematically (except for a related process
without phase change in the context of embolism recovery in maple [57]) and so it
remains unclear whether this physical description is capable of capturing exudation.
(c) Three essential physical mechanisms
Before proceeding further, we extend the freeze -- thaw hypothesis just described
by incorporating three additional mechanisms:
Gas bubbles in the vessel: Sap (like water) is an incompressible fluid so that in the
rigid, closed vessel network of a leafless tree there is no mechanism for fiber --
Article submitted to Royal Society
Multiscale model for sap exudation
5
1
Fiber
gas
Vessel
sap
gas
Thawing
Heat
influx
gas
4
sap
gas
Thawing
Freezing
Heat
outflux
ice
gas
Root
uptake
2
sap
gas
ice
gas
sap ice
gas
Freezing
3
Figure 2: Stages in the freeze -- thaw cycle. Various stages in the freeze -- thaw
cycle are depicted within an adjacent fiber -- vessel pair. Stages 1→2→3 depict the
freezing process: when temperature drops, an ice layer grows on the inner wall of the
gas-filled fiber as water is drawn via cryostatic suction through the porous cell wall.
Stages 3→4→1 depict the reverse process as temperature rises. Note the reversed
order of phase interfaces inside the fiber between stages 2 and 4. The blue arrows
denote water transport either through the fiber -- vessel wall or between roots and
vessels.
vessel mass transfer if vessels are completely saturated with sap. However, the
existence of gas bubbles within vessels is well-documented in maple [39, 56]
and other hardwood species [50]. Even if xylem pressures were high enough
to dissolve such bubbles, gas would eventually be forced out of solution upon
freezing and so at least a transient presence of gas bubbles is unavoidable.
Therefore, introducing a gas phase in the vessels provides a plausible mecha-
nism for fiber -- vessel pressure exchange.
Sap freezing point depression (FPD): Sap contains dissolved sugars and hence ex-
periences a reduced freezing point compared to pure water according to Blag-
den's Law [7], ∆Tf pd = Kb Cs/ρw, where Kb is the cryoscopic constant, Cs
is sugar concentration, and ρw is water density. For example, sap containing
3% sucrose by mass experiences a FPD of ∆Tf pd ≈ 0.162 ◦K. Although this
temperature difference may appear insignificant, we will see that it is actually
large when considered on the scale of individual cells, and indeed is sufficient to
account for the existence of ice in fibers while sap in adjacent vessels remains
in liquid form. This partitioning of ice and liquid in neighbouring fiber -- vessel
pairs induces cryostatic suction that draws liquid out of the vessel to form ice
on the inner fiber wall.
Root water uptake during freezing: No previous hypothesis for sap exudation ex-
plicitly considers the role of root water uptake. Furthermore, several studies
suggest that root pressure in maple has a negligible effect on exudation [3, 30].
Article submitted to Royal Society
6
I. Graf, M. Ceseri and J. M. Stockie
Nonetheless, it is well-known that during winter months trees can draw wa-
ter from the roots if soil temperatures are high enough [48, 54], which can
be caused by an insulating snow cover [42]. Recent experiments on maple
saplings [6, 40] have provided the first direct evidence that root water uptake
occurs in maple during winter while exudation is underway.
Our aim is now to incorporate these three modifications into a model of the freeze --
thaw process outlined previously, and then demonstrate that the resulting equations
are capable of reproducing observed behaviours.
3. Mathematical formulation
(a) Outline of the modelling approach
The freeze -- thaw mechanism outlined in the previous section involves processes
operating on two distinct spatial scales: the microscale corresponding to individ-
ual wood cells with dimensions ranging from 10 -- 100 microns; and the macroscale
corresponding to the tree stem with diameter tens of centimetres. The derivation
of our mathematical model for sap exudation therefore divides naturally over these
two scales. Firstly, we develop microscale equations that capture cell-level processes
within libriform fibers and vessels, combining the dynamics of freezing, thawing, gas
dissolution, osmotic pressure, heat transport, and porous flow through the fiber --
vessel wall. Secondly, we consider heat transport in the entire tree stem and apply
periodic homogenization to derive an equation for the macroscale temperature that
incorporates microscale cellular processes via appropriately defined transport coef-
ficients and source terms. We proceed as follows:
• Start from an existing microscale model for the thawing half of the freeze --
thaw process [8, 19] in which a 2D periodic microstructure is assembled from
copies of a reference cell Y containing a single fiber and vessel (see figure 3a).
The fiber is placed at the centre of Y (where the dashed line denotes the
fiber -- vessel wall) and the remainder of the reference cell corresponds to the
vessel. This choice of geometry is a mathematical idealization that captures the
volumes of the fiber and vessel compartments, but is not intended to accurately
represent the actual layout of wood cells. The governing equations consist of a
partial differential equation (PDE) for the microscale temperature along with
five ordinary differential equations (ODEs) for phase interface locations and
root water volume, coupled nonlinearly through source terms and algebraic
constitutive relations.
• Supplement the thawing model with analogous equations for the freezing pro-
cess, which have similar structure but differ slightly depending on the precise
state of freezing or thawing in the fibers and vessels.
• Apply periodic homogenization [2, 19] to derive a macroscopic equation for
temperature that is coupled to the microscale (reference cell) problem at each
point within the tree stem (see figure 3b). The macroscopic heat diffusion
equation contains an integral source term depending on the microscale tem-
perature and capturing all processes on the cellular level. A similar approach
has been applied to studying protein-mediated transport of water and solutes
in non-woody plant tissues [10].
Article submitted to Royal Society
Multiscale model for sap exudation
7
(a)
(b)
y
✻
r
Ω
x
✻
r
Figure 3: Multiscale problem geometry. (a) Idealized microscale fiber -- vessel
geometry, consisting of a square reference cell Y with side length ℓ. This diagram
depicts a thawing scenario (stage 4 in figure 2) wherein a fiber of radius Rf (dashed
line) contains a gas bubble surrounded by annular layers of ice and liquid water,
where the gas/ice and ice/water interfaces are concentric circles of radius sg and siw
respectively. The vessel contains a gas bubble (radius r) and liquid sap (water plus
sugar). The total liquid volume transferred from fiber to vessel is denoted by U . The
reference cell Y = Y 1 ∪ Y 2 is divided into two regions separated by a curve Γ, where
diffusion on Y 1 (light blue, outer vessel) is fast and on Y 2 (dark blue, fiber plus
fiber -- vessel overlap region) is slow. (b) A requirement for homogenization is that
the tree cross-section can be approximated by a periodic fine-structured domain,
tiled with copies of the reference cell. The macroscale problem is then solved on a
homogeneous domain Ω having the same size. Radial coordinates on the micro- and
macroscales are denoted y and x respectively.
• Exploit radial symmetry on the micro- and macroscales to reduce both PDEs
for temperature to a single spatial dimension. We will see later in section 4a
that the microscale equations need only be solved on the circular sub-region
Y 2 in figure 3a (consisting of the fiber and surrounding vessel overlap region)
which is clearly radially symmetric.
(b) Microscale equations for cell-level thawing process
The cell-level model is based on equations already developed for the thawing
half of the freeze -- thaw process by Ceseri and Stockie [8], which were subsequently
homogenized by Graf and Stockie [19]. We therefore begin by considering an inter-
mediate state in the thawing process corresponding to stage 4 in figure 2, during
which the vessel sap is completely thawed while the fiber contains both liquid and
ice. We extend the Ceseri -- Stockie model by incorporating additional physical ef-
fects that capture the influence of ice -- water surface tension, root water uptake, and
volume change due to ice/water phase transitions. We discuss some of the most
important assumptions and modifications next, leaving the reader to consult the
references [8, 19] for a complete derivation and discussion of assumptions.
Our model is based on the conceptual diagram in figure 1b that depicts a single
vessel -- fiber pair. Tracheids are not treated separately but instead 'lumped together'
Article submitted to Royal Society
8
I. Graf, M. Ceseri and J. M. Stockie
with vessels because, although they are connected hydraulically to vessels via paired
pits, they have a much smaller diameter and correspondingly lesser influence on sap
transport than vessels. Because multiple fibers adjoin and interact hydraulically
with each vessel, we introduce the parameter N f representing an average number of
fibers per vessel, which is estimated from SEM images [11] as N f ≈ 16. Our model
captures the dynamics of a single fiber and then scales all fiber -- vessel flux terms by
an appropriate factor of N f .
We assume that sapwood can be represented as a doubly-periodic array of ide-
alized reference cells Y as pictured in figure 3, where each reference cell contains a
circular fiber embedded within a surrounding square liquid region representing the
adjoining vessel. This choice of geometry is made for mathematical convenience in
the homogenization step, and can be justified because our aim is to derive a sys-
tem of equations that captures the net effect of sap flow and heat transport on the
microscale, keeping in mind that any specific geometric details will ultimately be
'averaged out' during the homogenization process anyways.
Our 2D geometry comes with the built-in assumption that axial (vertical) varia-
tions are neglected. In the absence of root water uptake, the model tree behaves as
a closed system that is essentially in equilibrium. Any pressure differences initiated
by phase change engender primarily horizontal flow between neighbouring cells, and
negligible axial flow. Furthermore, we have already shown [8] that phase change on
the microscale dominates the pressure exchange process and occurs very rapidly (on
the order of milliseconds). Root water uptake induces an axial flow but this is a
much slower process; therefore, over the time scales that dominate the microscale
problem, axial transients may be neglected.
The fiber is a circular cylinder with length Lf and cross-sectional radius Rf as
pictured in figure 3a. Situated at the centre of the fiber is a cylindrical gas bubble
with time-varying radius sg(t), outside of which lies an annular ice layer with outer
radius siw(t). The remaining volume extending to the fiber radius Rf contains melt-
water from thawed ice. We note that this configuration is specific to the thawing
process, and the ordering of ice and water layers would be reversed during freezing.
The vessel is represented by the portion of the reference cell lying outside the
fiber -- vessel wall (denoted by a dashed line) and the side length ℓ of the reference cell
is chosen so that the vessel cross-sectional area equals that of a cylinder of radius
Rv. Keeping in mind that there are actually N f fibers connected to each vessel, we
require that ℓ satisfy the area constraint
ℓ2 = πN f (Rf )2 + π(Rv)2.
(3.1)
Within the vessel is a gas bubble of radius r(t), which is surrounded by liquid sap
owing to the FPD effect that lowers freezing temperature below that in the fiber.
The cumulative volume of melt-water flowing through the porous fiber -- vessel wall
is denoted by U (t) and is measured positive from fiber to vessel. The final variable
that determines the local state of the fiber -- vessel system is the volume of root water
uptake, denoted Uroot(t).
We may now formulate a first-order system of five ODEs describing the time
evolution of siw, sg, r, U and Uroot. The fiber ice -- water interface is governed by the
Article submitted to Royal Society
Multiscale model for sap exudation
9
Stefan condition [1, 13]
∂tsiw = −
kw/ρw
(Ew − Ei)
∇T2 · ~n +
∂tU
2πsiwLf ,
(3.2)
where ∇T2 · ~n represents the normal temperature derivative on the interface (i.e.,
the curve along which temperature equals the melting (or freezing) point Tm) and
the final term accounts for the volume of water transferred between fiber and vessel.
This form of the Stefan condition assumes that liquid motion induced by phase
density differences is negligible [1]. The microscale temperature T2(y, t) is obtained
as the solution of a heat diffusion equation that will be stated in the next section,
where the microscale spatial coordinate is y. The parameters ρw and kw denote
density and thermal conductivity of liquid water, while (Ew − Ei) is the enthalpy
difference between water and ice (also called the latent heat or enthalpy of fusion)
at locations where T2 = Tm. The effects of thermal expansion are known to be
relatively small [56] and so have been neglected here.
Imposing mass conservation yields an equation for the fiber gas bubble radius
(which in this thawing scenario is a gas -- ice interface)
∂tsg = −
(ρw − ρi)siw∂tsiw
sgρi
+
ρw∂tU
2πsgρiLf ,
(3.3)
where ρi is the density of ice. An equation for the vessel gas bubble radius follows
from a similar mass conservation argument
∂tr = −
N f ∂tU + ∂tUroot
2πrLv
,
(3.4)
where Lv denotes the length of a vessel. This last equation expresses the balance
between water flux from neighbouring fibers and the slight volume change stemming
from the water/ice density difference. The effect of gas dissolution has been omitted
here but will be incorporated below in the gas density; this approximation was al-
ready justified in [9], which showed that incorporating dissolution in these equations
has negligible impact on the bubble radii sg and r.
Darcy's law governs liquid water flux through the porous fiber -- vessel wall
∂tU = −
w(t) − pf
w(t) − posm + pf
L A
N f hpv
i (t)i,
(3.5)
w) and fiber (pf
where the wall is characterized by hydraulic conductivity L and surface area A. The
pressure term in square parentheses derives from four contributions: liquid pressure
in the vessel (pv
w), osmotic pressure (posm), and capillary pressure
(pf
i ) due to ice -- water surface tension [18]. This latter contribution, also known as
cryostatic suction, follows hand-in-hand with FPD and arises whenever ice lies on
the inside surface of the wall and liquid sap is present on the vessel side, since then
the small capillary pores in the adjoining wall (with radius rcap) contain both ice
and liquid. For the thawing scenario under consideration here, water lies on both
sides of the fiber -- vessel wall and so pf
i = 0; however, other stages in the freeze -- thaw
process can give rise to non-zero pf
i as detailed in the next section (see also figure 4).
The final ODE comes from another application of Darcy's law to root flux
∂tUroot = maxn − LrAr(cid:0)pv
w(t) − psoil(cid:1), 0o,
(3.6)
Article submitted to Royal Society
10
I. Graf, M. Ceseri and J. M. Stockie
where Lr is the root hydraulic conductivity and Ar denotes the portion of root
surface area corresponding to a single vessel. The cut-off function 'max{ · , 0}' ensures
that water only flows inward from soil to roots and not outward, which is consistent
with experiments that demonstrate root outflow can be a factor of five smaller than
that for inflow [21]. Indeed, studies of root water transport in a variety of tree species
show that root conductivity can vary with factors such as temperature [17], root
age [14], and time of day [21] or season [34]. Many authors attribute this selective
control of water transport to membrane proteins known as aquaporins [23].
In the preceding discussion we introduced a number of constant parameters
whose values are listed in table 1. The remaining symbols correspond to interme-
diate variables whose definitions we provide next. First, the density of gas in the
fiber and vessel bubbles depends on initial values of density and volume, modified
to account for dissolved gas according to
ρf
g = V f
g (0) + HV f
V f
g + HV f
w (0)
w ! ρf
g (0),
ρv
g =(cid:18) V v
g (0) + HV v
g + HV v
V v
w
w (0)
(cid:19) ρv
g(0),
(3.7)
where H is the dimensionless Henry's constant for air in water. The various phase
volumes are determined from the cylindrical cell geometry as
V f
g = πLf s2
g,
V f
w = πLf(cid:16)(cid:0)Rf(cid:1)2
V v
g = πLvr2,
V v
w = πLv(cid:16)(Rv)2 − r2(cid:17) .
The corresponding gas pressures are given by the ideal gas law as
− s2
iw(cid:17) ,
pf
g =
ρf
RT
g
Mg
,
pv
g =
ρv
RT
g
Mg
,
(3.8)
(3.9)
(3.10)
where R is the universal gas constant and Mg is the molar mass of air. The water
and gas pressures in both fiber and vessel differ by an amount equal to the capillary
pressure, which is determined by the Young -- Laplace equation as
pf
w = pf
g −
2σgw
sg
,
pv
w = pv
g −
2σgw
r
,
(3.11)
where σgw is the air -- water surface tension. The osmotic pressure across the fiber --
vessel wall depends on sap sugar concentration according to
Finally, the sap sugar content induces a reduction in freezing temperature that obeys
posm = CsRT.
(3.12)
Tm,sap = Tm − ∆Tf pd = Tm −
KbCs
ρw
.
(3.13)
(c) Equations for other phase transitions
In the previous section we developed equations specific to the thawing process,
during which the vessel is completely thawed and the fiber contains a mix of gas,
water and ice (see stage 4 in figure 2). We describe next how these equations should
Article submitted to Royal Society
Multiscale model for sap exudation
11
Table 1: Model parameters for base case simulation
(Unless cited otherwise, all parameter values are taken from [8])
Symbol Description
Values
Units
Microscale variables (functions of time t and space x, y):
siw, sg
r
U
Uroot
V
p
ρ
interface locations in fiber
vessel bubble radius
water volume flowing from fiber to vessel
root water volume uptake
volume
pressure
density
m
m
m3
m3
m3
Pa
kg m−3
Subscripts: i, w, g for ice, water/sap, gas
Superscripts: f , v for fiber, vessel
Tree structural parameters:
A
Ar
ℓ
Lf
Lv
L
Lr
N f
Rf
Rv
rcap
W
area of fiber -- vessel wall
root area for a single vessel [15]
side length of reference cell, equation (3.1)
length of fiber
length of vessel element
conductivity of fiber -- vessel wall
conductivity of roots [47, 53]
number of fibers per vessel
inside radius of fiber
inside radius of vessel
radius of pores in fiber -- vessel wall [28]
thickness of fiber -- vessel wall
6.28 × 10−8
1.14 × 10−6
4.33 × 10−5
1.0 × 10−3
5.0 × 10−4
5.54 × 10−13
2.7 × 10−16
16
3.5 × 10−6
2.0 × 10−5
2.80 × 10−7
3.64 × 10−6
m2
m2
m
m
m
m s−1 Pa−1
m s−1 Pa−1
--
m
m
m
m
Water phase properties:
ci, cw
Ei, Ew
ki, kw
ρi, ρw
σiw, σgw
c∞
specific heat capacity
enthalpy at Tm
thermal conductivity
density
surface tension [18]
regularization parameter, equation (3.32)
Physical constants:
H
Kb
Mg
R
Tm
Henry's constant for air in water
cryoscopic (Blagden) constant
molar mass of gas (air)
universal gas constant
melting point for pure water
'Base case' simulation:
ice,
2100,
574,
2.22,
917,
0.033,
J ◦K−1 kg−1
kJ kg−1
liquid
4180
907
0.556 W m−1 ◦K−1
1000
0.076
kg m−3
N m−1
1.0 × 107
J ◦K−1 kg−1
0.0274
1.853
0.029
8.314
273.150
--
kg ◦K mol−1
kg mol−1
J ◦K−1 mol−1
◦K
Cs
psoil
R
Ta(t)
Tm,sap
sap sugar concentration (3% by mass)
soil pressure at roots = pv
tree cross-sectional radius
ambient temperature
melting point for sap = Tm − KbCs/ρw
w(0)
87.6
2.03 × 105
0.035
[−10, 20] + Tm
272.988
mol m−3
Pa
m
◦K
◦K
be modified to capture other freeze -- thaw states in the fiber and vessel. In particular,
we account for the fact that phase interfaces can appear or disappear whenever ice
completely thaws (or liquid completely freezes), as well as the reversal of the ice and
water layers in the fiber during freezing and thawing.
In fact, many equations remain unchanged throughout the entire freeze -- thaw
i . The required modifications
cycle, with the exception being those for sg, siw and pf
Article submitted to Royal Society
12
I. Graf, M. Ceseri and J. M. Stockie
1
Completely
thawed
gas
2
Vessel thawed
Fiber freezing
Heat
outflux
ice
gas
Vessel freezing
Fiber frozen
3
Completely
frozen
Vessel thawing
Fiber frozen
4
Vessel thawed
Fiber thawing
ice
gas
ice
gas
ice
gas
Heat
influx
gas
sap
gas
sap
gas
sap
gas
sap ice
gas
sap
gas
sap
gas
pf
i = 0
∂tsiw = 0
∂tsg =
∂tU
2πsgLf
Root uptake
pf
i =
∂tsiw =
∂tsg =
2σiw
rcap
V f
i
i + V f
V f
ki/ρi
w
∇T · ~n +
(Ew − Ei)
(ρw − ρi)siw∂tsiw
∂tU ρw
2πsiwLf ρi
+
∂tU
2πsgLf
sgρw
pf
i =
2σiw
rcap
∂tsiw = 0
∂tsg = min(cid:26) ρw∂tU
2πsgLf ρi
, 0(cid:27)
pf
i = 0
∂tsiw = 0
∂tsg = 0
∂tr = 0
∂tU = 0
∂tUroot = 0
pf
i =
2σiw
rcap
∂tsiw = 0
∂tsg = min(cid:26) ρw∂tU
2πsgLf ρi
, 0(cid:27)
pf
i = 0
∂tsiw = −
∂tsg = −
kw/ρw
∇T · ~n +
(Ew − Ei)
(ρw − ρi)siw∂tsiw
2πsiwLf
ρw∂tU
∂tU
sgρi
+
2πsgLf ρi
(3.14)
(3.15)
(3.16)
(3.17)
(3.18)
(3.19)
(3.20)
(3.21)
(3.22)
(3.23)
(3.24)
(3.25)
(3.26)
(3.27)
(3.28)
(3.29)
(3.30)
(3.31)
Figure 4: Microscale equations for all stages of the freeze -- thaw process.
for each case are listed in figure 4, referenced by the numbered stages in figure 2. We
emphasize that the ice -- water interface lies within pores in the fiber -- vessel wall and
forms a mushy layer wherein both solid and liquid phases coexist in the pore space.
Article submitted to Royal Society
Multiscale model for sap exudation
13
This type of phase interface (called a frozen fringe in the context of ice lensing in
soils [18]) differs from an idealized gas -- water interface in that the interfacial pressure
jump pf
i
increases with ice volume fraction according to
pf
i =
2σiw
rcap
V f
i
V f
i + V f
w
,
where σiw represents the ice -- water surface tension and V f
i,w are the corresponding
volume fractions. Finally, we note that when both fiber and vessel are completely
frozen (stage 3) the equations for r, U and Uroot also drop out of the system.
(d ) Homogenized equation for temperature
The equations derived in the preceding two sections govern microscale processes
at the cellular level whereas on the macroscale the temperature is of primary interest,
and it is transport of heat between the external (ambient) environment and the
interior of the tree stem that drives the freeze -- thaw process. Clearly, there exists a
two-way interaction between the global temperature and the local fiber -- vessel state,
wherein temperature governs phase change dynamics in fibers and vessels, while
cellular processes in turn influence heat transport through the Stefan condition and
local phase volume fractions. To simplify this complex multiscale problem, we exploit
a separation in spatial scales reflected in the fact that state variables describing the
fiber -- vessel configuration are essentially 'invisible' on the macroscale except through
their effect on heat transport properties of the sap- and gas-filled wood.
Because of the repeating microstructure of wood, this problem is ideally suited
to the application of periodic homogenization. The philosophy behind this approach
is to solve at each point in space a local problem on a reference cell Y that deter-
mines the solution state on the microscale. By using an appropriate homogenization
or averaging procedure, the effect of microscale variables on the macroscale may
then be incorporated into equations for the global solution variables. One technical
requirement is that the reference cell must divide into two sub-regions, Y = Y 1 ∪Y 2,
separated according to whether heat diffusion is fast (in Y 1, the outer portion of
the vessel) or slow (in Y 2, an overlap region covering the fiber and the remainder
of the vessel). The result is two heat equations: one governing temperature on the
macroscopic domain Ω and the second on Y 2 × Ω. When these two equations are
coupled together, we obtain a two-scale temperature solution on the domain Y × Ω.
Instead of fully coupling the micro- and macroscale equations, this homogenization
approach leads naturally to a simpler system of equations that captures the essential
aspects of coupling between scales. A similar homogenization approach has been ap-
plied by Chavarr´ıa-Krauser and Ptashnyk to a model of water and solute transport
in plants [10].
The dynamics of heat transport are best described using a mixed formulation
written in terms of both temperature and specific enthalpy, which are denoted re-
spectively by T2(x, y, t) and E2(x, y, t) on the reference cell region Y 2, and T1(x, t)
and E1(x, t) on the macroscale. The variables T1 and E1 depend on time t and the
global spatial coordinate x, whereas microscale quantities have an additional de-
pendence on the reference cell Y through a local spatial coordinate y. Temperature
and enthalpy are not independent variables but instead are related via the piecewise
Article submitted to Royal Society
14
I. Graf, M. Ceseri and J. M. Stockie
linear function
T (E) =
1
ci
E,
if E < Ei − δi,
Tm + 2E−Ei−Ew
2c∞
,
if Ei − δi ≤ E < Ew + δw,
(3.32)
Tm + 1
cw
(E − Ew),
if Ew + δw ≤ E.
We introduce the large parameter c∞ (taking c∞ = 107 in practice) to impose a
small but nonzero slope (1/c∞) on the central plateau region where T ≈ Tm. We
also make use of the fact that Ei = ciTm and choose
δi =
ci(Ew − Ei)
2(c∞ − ci)
and δw =
cw(Ew − Ei)
2(c∞ − cw)
,
(3.33)
so that the function T (E) is continuous. This form of T (E) is a regularization of the
exact temperature -- enthalpy relationship [55] that avoids numerical instabilities in
the calculation of temperature and also recovers the exact (piecewise linear) result
in the limit as c∞ → ∞ and δi, δw → 0.
During the homogenization procedure [19], we find that heat transport in the
reference cell must only be treated on the sub-region Y 2 where temperature obeys
cw∂tT2 − ∇y · (D(E2)∇yT2) = 0
in Y 2(x, t) × Ω,
(3.34)
and D(E2) is a thermal diffusion coefficient that is a piecewise linear and continuous
function of enthalpy [55]
D(E) =
ki
ρi
+ E−Ei
Ew−Ei(cid:16) kw
ρw
ki
ρi
,
kw
ρw
,
− ki
ρi(cid:17) ,
if E < Ei,
if Ei ≤ E < Ew,
if Ew ≤ E.
(3.35)
We employ this nonstandard definition of D (instead of the usual thermal diffusivity
having units m2 s−1) so that we can factor out the specific heat, thereby allowing
the same coefficient to be used in both this microscale heat equation and the mixed
temperature -- enthalpy form we develop below for the macroscale. We include an
explicit time- and global space-dependence in Y 2(x, t) to emphasize the fact that
the ice region within the fiber is bounded by a moving water -- ice interface, and that
the fiber configuration varies from point to point throughout the tree stem. On the
water-ice interface (corresponding to the inner boundary of Y 2) the temperature
equals the melting point value
T2 = Tm
on ∂Y 2(x, t) × Ω.
(3.36)
We thereby obtain the macroscale temperature equation
∂tE1 − ∇x · (ΠD(E1)∇xT1) =
1
Y 1ZΓ
D(E2)∇yT2 · ~n dS
in Ω,
(3.37)
where the coupling with microscale variables is embodied in a surface integral term.
The factor Π multiplying the diffusion coefficient is a 2 × 2 matrix whose entries
depend on the reference cell geometry according to
Πij =
1
Y 1ZY 1
(δij + ∇yµi) dy,
(3.38)
Article submitted to Royal Society
Multiscale model for sap exudation
15
for i, j = 1, 2. Here, δij is the Kronecker delta symbol and µi(y) are solutions of a
standard reference cell problem on Y 1 [2]. The temperature on the outer surface of
the tree is held at the ambient value
T1 = Ta(t)
on ∂Ω.
(3.39)
Finally, the micro- and macroscale solutions are coupled by matching temperature
on the interior boundary
T2 = T1
on Γ × Ω.
(3.40)
In summary, the governing equations consist of a system of differential -- algebraic
equations (3.2) -- (3.13) and (3.34) -- (3.36) for the microscale temperature and fiber --
vessel state variables within each local reference cell. These are supplemented by
equations (3.37) -- (3.40) for the macroscale temperature on Ω. Both problems are
solved at each spatial point x ∈ Ω and the two solutions are coupled by means of
the integral source term in (3.37) and the boundary condition (3.40). The geometry
of the local reference cell is also incorporated into the macroscale problem via the
(constant) pre-factors Π multiplying the diffusion coefficient in (3.37).
4. Simulating daily freeze -- thaw cycles
(a) Numerical solution algorithm
The radial symmetry of both micro- and macroscale domains implies that all
solution variables can be written as functions of a single radial coordinate and time.
We use a method of lines approach and discretize the temperature variables in
space using finite elements, yielding a large system of time-dependent ODEs. When
combined with the ODEs and algebraic equations governing microscale fiber -- vessel
dynamics, the resulting coupled system is integrated in time using a standard ODE
solver. The spatial discretization on the two scales proceeds as follows:
• Microscale (cell-level) equations: The fiber ice temperature is assumed to be
a uniform 0 ◦C, and gas temperature is also taken constant since the thermal
diffusivity of gas is so much larger than that for either ice or water. Therefore,
the PDE (3.34) for temperature on Y 2 must only be solved on the annular
region between Γ and the phase interface siw (see figure 3a). We find that
sufficient accuracy is obtained for T2 by using only 4 radial grid points within
the annulus. Because the phase interface evolves in time, we use a moving
mesh approach wherein the motion of grid points introduces an additional
'grid advection' term that is proportional to the mesh point velocity [22].
• Macroscale (tree-level) equation: The tree stem is similarly divided into equally-
spaced radial points, and here we find that taking 20 grid points yields suffi-
cient accuracy in T1. Owing to radial symmetry, the integral source term in
(3.37) reduces to multiplication by the curve length Γ. The factors Π de-
pend only on the reference cell geometry and so can be pre-computed at the
beginning of a simulation.
We employ an efficient split-step approach where in each time step the reference cell
problem is solved for the microscale variables, and then the macroscale temperature
equation is solved by holding the microscale variables constant.
Article submitted to Royal Society
16
I. Graf, M. Ceseri and J. M. Stockie
The algorithm described above has been implemented in Matlab using the built-
in stiff ODE solver ode15s to integrate the equations in time. The only algorithmic
detail remaining to be described is the switching between equations required as
phase interfaces appear or disappear. We can capture this switching simply and
robustly using the Events option provided in the ODE solver suite, which signals
an event based on zero-crossings of an 'indicator function'. During any portion of
the freeze -- thaw cycle, the indicator function is set equal to either the thickness of
a phase interface or the difference between the phase temperature and the melting
temperature. When the indicator crosses zero, the time integration halts, equations
are modified appropriately, and the ODE solver is restarted using the new set of
equations and taking the current solution as the new initial state. The time in-
tegration then proceeds until the next phase change event is signalled. A typical
simulation covering 4 daily temperature cycles requires between 30 -- 45 minutes of
clock time on an Apple MacBook Pro with 2.3 GHz quad-core Intel i7 processor.
(b) Choice of parameters
The algorithm just described is used to simulate freeze -- thaw dynamics in a
typical base case scenario for which all parameters are listed in table 1. We take
a 'sapling' of diameter 0.07 m consisting entirely of sapwood. The sugar content
of maple sap ranges from 1 -- 5% by mass [49] and so we choose a representative
value of 3% that induces a vessel FPD of ∆Tf pd = 0.162 ◦C. To mimic temperature
variations during late winter, we let ambient temperature vary sinusoidally between
−10 and +20 ◦C over a 24-hour period (this range is somewhat extreme but is
chosen to correspond with the experiments of Am´eglio et al. [3] that we will describe
shortly). We begin with a freezing event and initialize the tree in a thawed state
with uniform temperature 0.35 ◦C, just slightly above the freezing point. Each fiber
initially contains gas and water with 75% gas by volume, whereas the vessel has a
much smaller initial gas content of 8%.
There remain two parameters whose values we have not been able to obtain rea-
sonable estimates from the literature -- root hydraulic conductivity Lr and capillary
pore radius rcap -- and so we have had to adjust their values in order to match numer-
ical results with experimental data. First, we choose Lr = 2.7 × 10−16 m s−1 Pa−1
so that pressure and root uptake vary over time scales similar to those observed
in experiments [3, 49]. Then, we take rcap = 2.8 × 10−7 m so that the exudation
pressure build-up is within the observed range of 80 to 150 kPa [11, 12]. This pore
size is also consistent with that measured in other membranes that hinder transport
of sucrose molecules [28].
(c) Base case: Pressure build-up during temperature cycling
Using these base case parameters and initial conditions, we perform two nu-
merical simulations: one with root water uptake corresponding to a soil pressure of
psoil = 203 kPa, and a second with no root uptake (e.g., consistent with a completely
frozen soil). Vessel sap pressures are compared in figure 5a, and in both cases we
observe a periodic variation in pressure synchronized with daily temperature fluctu-
ations. Without root uptake, the vessel pressure simply oscillates between two fixed
values of 20 and 200 kPa and there is no pressure build-up over multiple freeze -- thaw
Article submitted to Royal Society
Multiscale model for sap exudation
17
(b)
Base case
No root uptake
∆p
1
∆p
2
1
2
time [d]
3
4
(a)
]
C
o
[
T
20
10
0
-10
]
a
P
k
[
vw
p
300
200
100
]
3
m
c
[
t
o
o
r
U
0.6
0.3
0
0
Figure 5: Comparison of base case simulation with Am´eglio's experi-
ments. (a) Simulated vessel sap pressure (middle, with and without root water)
and cumulative root water uptake (bottom) in response to an imposed periodic am-
bient temperature (top). The vertical dotted lines highlight times when temperature
crosses the freezing point. Two primary features of the vessel pressure curve are the
amplitude of pressure oscillations in each daily cycle (∆p1, arising from ice -- water
capillary effects) and the residual pressure increase at the end of a cycle (∆p2, due
to root water uptake). (b) Am´eglio et al.'s experiments on black walnut [3] (repro-
duced with permission of Oxford University Press). The measurements relevant to
our study are 'P control' (sap pressure) and 'T trunk' (temperature).
cycles. However, when root uptake is included there is a gradual pressure increase
superimposed on the background oscillations, with a total increase (measured from
the local maximum in each cycle) of roughly 80 kPa over the four days. The accom-
panying plot of total root uptake in figure 5a shows that the majority of root water
is absorbed during the first freeze -- thaw cycle, followed by a more gradual uptake
that is essentially complete after 3 days.
We next draw a direct comparison with the experiments of Am´eglio et al. [3]
who studied black walnut trees (Juglans nigra) in a controlled laboratory setting
where the living stump of an excised tree branch was connected via a sealed pipe
to a pressure transducer. We calculate vessel sap pressure in our simulations as an
average pressure across the stem cross-section to be as close as possible to such a
transducer measurement. We are unaware of any comparable data for sugar maple,
but we claim that a meaningful comparison may still be drawn with Am´eglio's
results since black walnut is closely related to maple and undergoes exudation under
similar conditions [3, 11, 16]. The curves to focus on in Am´eglio's figure 5b are the
air temperature (labelled 'T trunk') and stem pressure (labelled 'P control').
The qualitative agreement between simulated and experimental pressures is re-
markable considering the complexity of the processes involved and the minimal
parameter fitting required. The overall shape of pressure curves is similar, with each
freeze -- thaw cycle exhibiting a rapid increase whenever temperature exceeds the
freezing point. The pressure then attains a maximum, after which there is a slight
decrease over roughly 6 -- 8 hours, followed by a rapid drop as ambient temperature
crosses the freezing point again. We remark that there is also a rough quantita-
tive agreement between simulations and experiments in that pressure oscillations
Article submitted to Royal Society
18
I. Graf, M. Ceseri and J. M. Stockie
have an amplitude of 80 to 100 kPa, and the total pressure build-up over four days
also is similar. On the other hand, the maximum value of our simulated pressure is
290 kPa, which is almost double the 160 kPa observed in the black walnut experi-
ments; however, it is possible that more time is needed for the experiment to reach
steady state, not to mention that there are species -- specific differences that could
influence pressure.
A more quantitative comparison can be drawn based on two characteristic fea-
tures of the pressure in figure 5a labelled as ∆p1 and ∆p2. The first corresponds to
the amplitude of oscillations in the absence of root uptake, which derives mainly from
cryostatic suction and so can be estimated using the formula ∆p1 ≈ 2σiw/rcap ≈
236 kPa. This value is close to the computed amplitude of the 'no root uptake' curve,
as well as to the rise in vessel pressure during the initial thawing event for the 'base
case'. The second feature ∆p2 captures the exudation pressure build-up during the
first freeze -- thaw cycle which arises mainly from root water uptake. Because this
additional water acts to compress the gas in fiber and vessel, we apply the differ-
ential form of the ideal gas law at constant temperature, ∆p2 ≈ −p ∆V /V , during
the first freezing event. Substituting values of p ≈ 200 kPa for the initial vessel
pressure, ∆V ≈ 0.4 cm3 for the root water volume uptake (taken from figure 5a)
and V ≈ 1.15 cm3 for the initial gas volume in a slice through the tree cross-section
(with thickness equal to that of the reference cell, Lf ), we obtain ∆p2 ≈ 69 kPa.
The correspondence between this estimate and the computed value of 50 kPa is
reasonable, considering that it ignores effects such as gas dissolution.
Despite the abundance of experimental data available for sugar maple [11, 12, 25,
49], most experiments measure sap outflux from tapped trees [12] rather than the
'closed system' corresponding to an untapped tree that we consider here. Other mea-
surements have been taken of excised wood samples rather than living trees, while
yet others were taken in uncontrolled external conditions with irregular variations in
pressure and ambient temperature. Consequently, we hesitate to attempt a detailed
comparison between any of these experiments and our simulations; nonetheless, we
can still draw a few quantitative comparisons. For instance, Tyree [49] performed
experiments on excised maple branches that absorbed water at a maximum rate
of 12 cm3/h; for similar sized branches, our model yields a comparable maximum
absorption rate of roughly 13 cm3/h as well as qualitatively similar solution pro-
files. Another experiment by Johnson et al. [25] yielded total root uptake of 2.0 cm3
during freezing, followed by a much smaller uptake of 0.1 cm3 during a subsequent
freezing event. We see similar qualitative behaviour in our simulations, as well as
measuring 2.2 and 0.2 cm3 of water absorbed during the first and second freeze,
respectively.
(d ) Two crucial mechanisms: Root water uptake and FPD
To evaluate the relative importance of the various physical mechanisms, we
present in figure 6 the base case pressure and root water uptake alongside simula-
tions with each of the following mechanisms 'turned off': FPD, root uptake (repeated
from figure 5a), osmosis and gas dissolution. The first two mechanisms clearly have
the greatest impact on the build-up of exudation pressure. We already discussed
the crucial role of root uptake in facilitating pressure accumulation over multiple
freeze -- thaw cycles. This effect is underscored by the plots in figure 7a depicting the
Article submitted to Royal Society
Multiscale model for sap exudation
(a)
]
a
P
k
[
e
r
u
s
s
e
r
p
p
a
s
l
e
s
s
e
V
450
400
350
300
250
200
150
100
50
0
No osmosis
No gas dissolution
Base case
No FPD
No root uptake
1
2
time [d]
3
4
(b)
1
0.8
0.6
0.4
0.2
]
3
m
c
[
e
k
a
t
p
u
t
o
o
R
0
0
No osmosis
No gas dissolution
Base case
No FPD
No root uptake
19
1
2
time [d]
3
4
Figure 6: Comparison of various physical mechanisms. (a) An investigation
of the relative importance of various physical effects, depicting pressure with each
of the following mechanisms turned off: FPD, root water uptake, osmosis, gas dis-
solution. The 'base case' and 'no root uptake' curves are repeated from figure 5a for
easy comparison. (b) Corresponding curves for root uptake.
(a)
]
a
P
k
[
e
r
u
s
s
e
r
p
p
a
s
l
e
s
s
e
V
450
400
350
300
250
200
150
100
50
0
2e−15
5e−16
2.7e−16 (base case)
1e−16
0
1
2
time [d]
3
4
(b)
300
250
200
150
100
50
]
a
P
k
[
e
r
u
s
s
e
r
p
p
a
s
l
e
s
s
e
V
0
0
7%
5%
3% (base case)
1%
0%
3
4
1
2
time [d]
Figure 7: Sensitivity of exudation pressure to parameters. (a) Root hy-
draulic conductivity, Lr, in m s−1 Pa−1. (b) Sugar content in %.
pressure response when root conductivity Lr varies between zero (no uptake) and
nearly ten times the base value.
Without FPD, the vessel pressure remains nearly constant and there is minimal
root uptake, whereas without osmosis the vessel pressure increases. We therefore
conclude that the predominant impact of sugar on exudation is through FPD rather
than osmosis, and even though ∆Tf pd is small it nonetheless plays a critical role in
facilitating pressure transfer between fiber and vessel. This dependence is illustrated
further by figure 7b, where sugar content is varied between 0 and 7% and we ob-
serve that both net pressure build-up and oscillation amplitude increase with sugar
content.
One assumption requiring further investigation is that of zero conductivity to
root outflow in (3.6), which we motivated by citing experimental results that ex-
hibit a small but still nonzero root outflow [21]. To study this outflow effect, we
take four different outflow conductivities equal to the inflow value Lr scaled by a
Article submitted to Royal Society
20
I. Graf, M. Ceseri and J. M. Stockie
(a)
]
a
P
k
[
e
r
u
s
s
e
r
p
p
a
s
l
e
s
s
e
V
350
300
250
200
150
100
50
0
0.0 (base case)
0.1
0.2
0.5
1.0
1
2
time [d]
3
4
(b)
]
3
m
c
[
e
k
a
t
p
u
t
o
o
R
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.0 (base case)
0.1
0.2
0.5
1.0
1
2
time [d]
3
4
Figure 8: Sensitivity to root outflow. (a) Pressure curves with non-zero con-
ductivity to root outflow, where the ratio of outflow-to-inflow ratio varies between
0 and 1. (b) Corresponding plots of root uptake.
factor between 0 and 1 (where 0 corresponds to the base case). The results in fig-
ure 8 clearly show that allowing even a small outflow has a major influence on the
root water uptake by preventing accumulation of water over multiple freeze -- thaw
cycles and thereby reducing build-up of exudation pressure. Because of the obvious
sensitivity of these results to root outflow, a more extensive experimental study of
root conductivity in maple is warranted.
We end this section by addressing the seemingly counter-intuitive result in fig-
ure 6a that introducing osmosis decreases vessel sap pressure. This result can be most
easily explained by considering the water flux equation (3.5) over a long enough time
that the fiber and vessel have reached a quasi-steady state and ∂tU ≈ 0. Then (3.5)
reduces to the simple pressure balance
g −
(cid:18)pv
g −
2σgw
r (cid:19)
}
−(cid:18)pf
pv
w
{z
− posm + pf
i ≈ 0.
2σgw
sg (cid:19)
}
pf
w
{z
The ice -- water capillary pressure pf
is a constant, and our simulations show that
i
osmosis has relatively small impact on fiber bubble size and pressure (the latter
effect was discussed in [8]). Therefore, the primary influence of osmosis is within
the vessel: osmotically-driven flow from fiber to vessel compresses the vessel bubble
which not only increases the vessel gas pressure pv
g, but also increases the capillary
pressure term (via a reduction in bubble radius r). The contribution from surface
tension dominates and so the net effect is actually a decrease in vessel sap pressure
pv
w, which is consistent with figure 6a and the results reported in [8].
(e) Phase change dynamics on the microscale
When a completely frozen tree warms above 0 ◦C during the day, a thawing
front develops near the bark (wherein water and sap are frozen ahead of the front
and thawed behind) and advances into the stem; an analogous scenario occurs upon
freezing. Clearly, the 'interesting' solution dynamics will occur in the vicinity of
this front, and hence knowledge of phase change on the microscale is desirable for
Article submitted to Royal Society
(a)
]
C
◦
[
e
r
u
t
a
r
e
p
m
e
T
0.5
Tm
0
Tm,sap
−0.5
−1
−1.5
Multiscale model for sap exudation
21
(b)
freezing
∇T small
∇T large
Enthalpy [J/kg]
]
C
◦
[
e
r
u
t
a
r
e
p
m
e
T
0.5
0
−0.5
−1
∇T small
Tm
Tm,sap
∇T large
−1.5
0.005 0.01 0.015 0.02 0.025
Radius [m]
(c)
]
C
◦
[
e
r
u
t
a
r
e
p
m
e
T
1.5
1
0.5
0
−0.5
thawing
∇T large
Tm
Tm,sap
∇T small
Enthalpy [J/kg]
(d)
1.5
1
0.5
0
−0.5
]
C
◦
[
e
r
u
t
a
r
e
p
m
e
T
0.005
∇T large
Tm
∇T small
Tm,sap
0.01
0.015
Radius [m]
0.02
Figure 9: Local phase change dynamics. Plots of temperature versus radius
and enthalpy for a fixed time in the middle of a freezing event (a,b, top) and a
thawing event (c,d, bottom). Points correspond to discrete solution values on an
equally-spaced radial grid and are coloured according to the current state of fiber
and vessel: blue if both are frozen; red if both are thawed; purple if fiber is frozen
and vessel is thawed.
understanding solution behaviour. Over a century ago, Wiegand [56] recognized the
existence of freezing and thawing fronts that 'penetrate the wood in a wave-like
manner' and in which 'but few cells would actually take part in the production of
pressure at any one time'; however, there has so far been no attempt to develop a
mathematical model for this phenomenon. In particular, the role of FPD in govern-
ing the progress of these phase transitions throughout the sapwood has not been
investigated before.
Phase change dynamics are most easily studied by means of a temperature --
enthalpy diagram as depicted in figures 9a,c, which are each taken at a fixed time
during a freeze or thaw event. Both plots feature a plateau region at the melting
temperature, which has a horizontal extent equal to the enthalpy of fusion. Note
that there are two distinct melting temperatures in fiber and vessel equal to Tm and
Tm,sap = Tm − ∆Tf pd respectively. The corresponding plots of temperature versus
radius are shown in figures 9b,d which depict the local state of each point within the
tree stem. For example, in the freezing case (top) the three grid points closest to the
stem centre are completely thawed (red), the outermost point is frozen (blue), and
the intervening points are undergoing freezing (purple). Owing to FPD, water in the
fiber freezes before the vessel sap, thereby introducing a time delay in formation of
ice between the fiber and vessel.
For both freezing and thawing, the bulk of the stem is in a state located at the
leading edge of the enthalpy plateau (right edge for freezing, left edge for thawing).
This behaviour can be explained by considering the local rate of phase change:
Article submitted to Royal Society
22
I. Graf, M. Ceseri and J. M. Stockie
conservation of energy at a phase interface is expressed mathematically using the
well-known Stefan condition, which states that the rate of freezing (or thawing) is
proportional to the temperature gradient. Referring to figures 9b,d, the temperature
difference between adjacent points is smaller near the tree centre and larger near the
bark. Consequently, at any location in the tree a freezing event begins within the
fiber as a slow process, followed at a later time in the vessel which freezes relatively
quickly. In contrast, a thawing event begins with a slow thawing of the vessel sap,
followed by rapid thawing in the fiber.
5. Concluding remarks
We have developed the first complete mathematical model for the tree sap exuda-
tion process based on a prevailing freeze -- thaw hypothesis. We introduced a number
of additions to this hypothesis, and identified root water uptake and freezing point
depression (FPD) as the two main driving mechanisms for sap exudation. In par-
ticular, we showed that the primary mechanism whereby sugar induces exudation
pressure is via the FPD and not osmosis as was previously believed. Numerical sim-
ulations of the governing equations demonstrate qualitative and quantitative agree-
ment with experimental data on sugar maple and the related species black walnut.
The quality of agreement is striking considering that the model parameters were
determined using a minimum of parameter fitting. Our work clearly demonstrates
the need for further experiments on sugar maple that parallel the work of Am´eglio
et al. on walnut [3]. Our model results lead to the important conclusion that FPD
is a primary driver of sap exudation, which also requires experimental validation.
Furthermore, because we have only rough estimates at present for two of the model
inputs -- capillary pore size rcap and root conductivity Lr (especially differences
between conductivity to inflow and outflow) -- more accurate measurements of these
parameters are also required.
Our model provides an ideal platform from which to investigate other problems
related to sap flow in maple and related species. First of all, we aim to extend
our current model of a 2D stem cross-section to three dimensions. This will permit
us to incorporate variations in gravitational pressure head and sugar concentration
with height [56] and to study problems of practical importance to the maple syrup
industry such as optimizing tap-hole placement or determining sensitivity to changes
in soil or climatic conditions. Finally, there are a number of intriguing parallels
between exudation and the phenomenon of freeze-induced winter embolism [45, 57]
that are also worthy of future investigation.
Competing interests.
We have no competing interests.
Authors' contributions.
JMS designed the study. IG and MC designed the numerical algorithm. IG carried
out the computational studies. All authors derived the mathematical model, analyzed the
results, and wrote the manuscript. All authors gave final approval for publication.
Article submitted to Royal Society
Multiscale model for sap exudation
23
Acknowledgements.
We are indebted to Chris Budd (University of Bath) for his helpful comments on an
earlier version of this manuscript.
Funding statement.
This work was supported by research grants from the Natural Sciences and Engineering
Research Council of Canada and the North American Maple Syrup Council (to JMS), a
Postdoctoral Fellowship from Mitacs (to MC) and a Feodor Lynen Fellowship from the
Alexander von Humboldt Stiftung (to IG).
References
[1] Alexiades, V. & Solomon, A. D. 1993 Mathematical Modeling of Melting and
Freezing Processes. Washington, DC: Hemisphere Publishing Co.
[2] Allaire, G. 1992 Homogenization and two-scale convergence. SIAM J. Math.
Anal., 23(6), 1482 -- 1518. (doi:10.1137/0523084)
[3] Am´eglio, T., Ewers, F. W., Cochard, H., Martignac, M., Vandame, M., Bodet,
C. & Cruiziat, P. 2001 Winter stem xylem pressure in walnut trees: effects
of carbohydrates, cooling and freezing. Tree Physiol., 21(6), 387 -- 394. (doi:
10.1093/treephys/21.6.387)
[4] Angeles, G. et al. 2004 The cohesion-tension theory. New Phytol., 163, 451 -- 452.
(doi:10.1111/j.1469-8137.2004.01142.x)
[5] Beckage, B., Osborne, B., Gavin, D. G., Pucko, C., Siccama, T. & Perkins, T.
2008 A rapid upward shift of a forest ecotone during 40 years of warming in the
Green Mountains of Vermont. Proc. Natl. Acad. Sci. USA, 105(11), 4197 -- 4202.
(doi:10.1073/pnas.0708921105)
[6] Brown, J. E. 2013 Remaking maple: new method may revolutionize maple syrup
industry. University Communications, University of Vermont. Available online
at http://www.uvm.edu/$\sim$uvmpr/?Page=news&storyID=17209.
[7] Cavender-Bares, J. 2005 Impacts of freezing on long-distance transport in
woody plants. In Vascular Transport in Plants (eds N. M. Holbrook & M. Zwie-
niecki), chap. 19, pp. 401 -- 424. San Diego, CA: Academic Press.
[8] Ceseri, M. & Stockie, J. M. 2013 A mathematical model for sap exudation in
maple trees governed by ice melting, gas dissolution and osmosis. SIAM J.
Appl. Math., 73(2), 649 -- 676. (doi:10.1137/120880239)
[9] Ceseri, M. & Stockie, J. M. 2014 A three-phase free boundary problem involving
ice melting and gas dissolution. Euro. J. Appl. Math., 25(4), 449 -- 480. (doi:
10.1017/S0956792513000430)
[10] Chavarr´ıa-Krauser, A. & Ptashnyk, M. 2013 Homogenization approach to water
transport in plant tissues with periodic microstructures. Math. Model. Nat.
Phenom., 8(4), 80 -- 111. (doi:10.1051/mmnp/20138406)
Article submitted to Royal Society
24
I. Graf, M. Ceseri and J. M. Stockie
[11] Cirelli, D., Jagels, R. & Tyree, M. T. 2008 Toward an improved model of maple
sap exudation: the location and role of osmotic barriers in sugar maple, butter-
nut and white birch. Tree Physiol., 28, 1145 -- 1155. (doi:10.1093/treephys/28.
8.1145)
[12] Cortes, P. M. & Sinclair, T. R. 1985 The role of osmotic potential in spring
sap flow of mature sugar maple trees (Acer saccharum Marsh.). J. Exp. Bot.,
36(1), 12 -- 24. (doi:10.1093/jxb/36.1.12)
[13] Crank, J. 1984 Free and Moving Boundary Problems. Oxford: Clarendon Press.
[14] Dawson, T. E. 1997 Water loss from tree roots influences soil water and nutrient
status and plant performance. In Radical Biology: Advances and Perspectives
on the Function of Plant Roots (eds H. E. Flores, J. P. Lynch & D. Eissenstat).
Rockville, MD: American Society of Plant Physiologists.
[15] Day, S. D. & Harris, J. R. 2007 Fertilization of red maple (Acer rubrum) and
littleleaf linden (Tilia cordata) trees at recommended rates does not aid tree
establishment. Arbor. Urban For., 33(2), 113 -- 121.
[16] Ewers, F. W., Am´eglio, T., Cochard, H., Beaujard, F., Martignac, M., Vandame,
M., Bodet, C. & Cruiziat, P. 2001 Seasonal variation in xylem pressure of
walnut trees: root and stem pressures. Tree Physiol., 21, 1123 -- 1132.
(doi:
10.1093/treephys/21.15.1123)
[17] Fennell, A. & Markhart, A. H. 1998 Rapid acclimation of root hydraulic
(doi:
J. Exp. Bot., 49(322), 879 -- 884.
conductivity to low temperature.
10.1093/jxb/49.322.879)
[18] Fowler, A. C. & Krantz, W. B. 1994 A generalized secondary frost heave model.
SIAM J. Appl. Math., 54(6), 1650 -- 1675. (doi:10.1137/S0036139993252554)
[19] Graf, I. & Stockie, J. M. 2014 Homogenization of the Stefan problem, with
application to maple sap exudation. Submitted, arXiv:1411.3039 [math.AP].
[20] Groffman, P. et al. 2012 Long-term integrated studies show complex and sur-
prising effects of climate change in the northern hardwood forest. BioScience,
62(12), 1056 -- 1066. (doi:10.1525/bio.2012.62.12.7)
[21] Henzler, T., Waterhouse, R. N., Smyth, A. J., Carvajal, M., Cooke, D. T.,
Schaffner, A. R., Steudle, E. & Clarkson, D. T. 1999 Diurnal variations in
hydraulic conductivity and root pressure can be correlated with the expression
of putative aquaporins in the roots of Lotus japonicus. Planta, 210(1), 50 -- 60.
(doi:10.1007/s004250050653)
[22] Huang, W. & Russell, R. D. 2011 Adaptive Moving Mesh Methods, vol. 174 of
Applied Mathematical Sciences. New York: Springer.
[23] Javot, H. & Maurel, C. 2002 The role of aquaporins in root water uptake. Ann.
Bot., 90(3), 301 -- 313. (doi:10.1093/aob/mcf199)
Article submitted to Royal Society
Multiscale model for sap exudation
25
[24] Johnson, L. P. V. 1945 Physiological studies on sap flow in the sugar maple,
Acer saccharum Marsh. Can. J. Res. C: Bot. Sci., 23, 192 -- 197. (doi:10.1139/
cjr45c-016)
[25] Johnson, R. W., Tyree, M. T. & Dixon, M. A. 1987 A requirement for sucrose
in xylem sap flow from dormant maple trees. Plant Physiol., 84, 495 -- 500.
(doi:10.1104/pp.84.2.495)
[26] Karl, T. R., Melillo, J. M. & Peterson, T. C. (eds) 2009 Global Climate Change
Impacts in the United States. Cambridge University Press. US Global Change
Research Program, http://www.globalchange.gov/usimpacts.
[27] Keller, J. B. 1964 Growth and decay of gas bubbles in liquids. In Proceedings
of the Symposium on Cavitation in Real Liquids (ed. R. Davies), pp. 19 -- 29.
General Motors Research Laboratories, Warren, Michigan, Elsevier Publishing
Company.
[28] Khaddour, I. A., Bento, L. S. M., Ferreira, A. M. A. & Rocha, F. A. 2010
Kinetics and thermodynamics of sucrose crystallization from pure solution at
different initial supersaturations. Surface Sci., 604, 1208 -- 1214. (doi:10.1016/
j.susc.2010.04.005)
[29] Kim, H. K., Park, J. & Hwang, I. 2014 Investigating water transport through
the xylem network in vascular plants. J. Exp. Bot., 65(7), 1895 -- 1904. (doi:
10.1093/jxp/eru075)
[30] Kramer, P. J. & Boyer, J. S. 1995 The absorption of water and root and stem
pressures. In Water Relations of Plants and Soils, chap. 6, pp. 167 -- 200. London:
Academic Press.
[31] Lovett, G. M. & Mitchell, M. J. 2004 Sugar maple and nitrogen cycling in
the forests of eastern North America. Front. Ecol. Env., 2(2), 81 -- 88. (doi:
10.1890/1540-9295(2004)002[0081:SMANCI]2.0.CO;2)
[32] Marvin, J. W. 1949 Changes in bark thickness during sap flow in sugar maples.
Science, 109(2827), 231 -- 232. (doi:10.1126/science.109.2827.231)
[33] Marvin, J. W. 1968 Physiology of sap production. In Sugar Maple Conference,
pp. 12 -- 15. Houghton, MI: Michigan Technological University.
[34] McElrone, A. J., Bichler, J., Pockman, W. T., Addington, R. N., Linder, C. R.
& Jackson, R. B. 2007 Aquaporin-mediated changes in hydraulic conductivity
of deep tree roots accessed via caves. Plant Cell Env., 30(11), 1411 -- 1421.
(doi:10.1111/j.1365-3040.2007.01714.x)
[35] Meinzer, F. C., Clearwater, M. J. & Goldstein, G. 2001 Water transport in
trees: current perspectives, new insights and some controversies. Env. Exp.
Bot., 45, 239 -- 262. (doi:10.1016/S0098-8472(01)00074-0)
[36] Merwin, H. E. & Lyon, H. 1909 Sap pressure in the birch stem. Bot. Gazette,
48(6), 442 -- 458.
Article submitted to Royal Society
26
I. Graf, M. Ceseri and J. M. Stockie
[37] Milburn, J. A. & O'Malley, P. E. R. 1984 Freeze-induced sap absorption in
Acer pseudoplatanus: a possible mechanism. Can. J. Bot., 62(10), 2101 -- 2106.
(doi:10.1139/b84-285)
[38] Minorsky, P. V. 2003 The decline of sugar maples (Acer saccharum). Plant
Physiol., 133, 441 -- 442. (doi:10.1104/pp.900091)
[39] Perkins, T. D. & van den Berg, A. K. 2009 Maple syrup -- Production, com-
position, chemistry, and sensory characteristics. In Advances in Food and Nu-
trition Research (ed. S. L. Taylor), vol. 56, chap. 4, pp. 101 -- 143. Elsevier.
(doi:10.1016/S1043-4526(08)00604-9)
[40] Perkins, T. D. & van den Berg, A. K. 2015 Sap-collecting devices, systems and
methods for sap-producing saplings. U.S. Patent Application No. 20150040472.
[41] Reynolds, J. 2010 Will maple syrup disappear? The changing climate may alter
the syrup industry forever. Can. Geog., 130(5), 21 -- 22.
[42] Robitaille, G., Boutin, R. & Lachance, D. 1995 Effects of soil freezing stress on
sap flow and sugar content of mature sugar maples (Acer saccharum). Can. J.
For. Res., 25(4), 577 -- 587. (doi:10.1139/x95-065)
[43] Sachs, J. 1860 Quellungserscheinungen an Holzern. Bot. Zeit., 18(29), 253 -- 259.
[44] Sano, Y., Morris, H., Shimada, H., Ronse De Craene, L. P. & Jansen, S. 2011
Anatomical features associated with water transport in imperforate tracheary
elements of vessel-bearing angiosperms. Ann. Bot., 107, 953 -- 964. (doi:10.1093/
aob/mcr042)
[45] Sperry, J. S., Donnelly, J. R. & Tyree, M. T. 1988 Seasonal occurrence of xylem
embolism in sugar maple (Acer saccharum). Amer. J. Bot., 75(8), 1212 -- 1218.
[46] Sperry, J. S. & Love, D. M. 2015 What plant hydraulics can tell us about
responses to climate-change droughts. New Phytol. (doi:10.1111/nph.13354)
[47] Steudle, E. & Peterson, C. A. 1998 How does water get through roots? J. Exp.
Biol., 49(322), 775 -- 788. (doi:10.1093/jxb/49.322.775)
[48] Stevens, C. L. & Eggert, R. L. 1945 Observations on the causes of the flow of
sap in red maple. Plant Physiol., 20, 636 -- 648. (doi:10.1104/pp.20.4.636)
[49] Tyree, M. T. 1983 Maple sap uptake, exudation, and pressure changes corre-
lated with freezing exotherms and thawing endotherms. Plant Physiol., 73,
277 -- 285. (doi:10.1104/pp.73.2.277)
[50] Tyree, M. T. 1995 The mechanism of maple sap exudation. In Tree Sap: Pro-
ceedings of the First International Symposium on Sap Utilization (eds M. Ter-
azawa, C. A. McLeod & Y. Tamai), pp. 37 -- 45. Bifuka, Japan: Hokkaido Uni-
versity Press.
[51] Tyree, M. T. 2003 Plant hydraulics: the ascent of water. Nature, 423, 923.
(doi:10.1038/423923a)
Article submitted to Royal Society
Multiscale model for sap exudation
27
[52] Tyree, M. T. & Yang, S. 1992 Hydraulic conductivity recovery versus water
(doi:
pressure in xylem of Acer saccharum. Plant Physiol., 100, 669 -- 676.
10.1104/pp.100.2.669)
[53] Tyree, M. T., Yang, S., Cruiziat, P. & Sinclair, B. 1994 Novel methods of
measuring hydraulic conductivity of tree root systems and interpretation using
AMAIZED: a maize-root dynamic model for water and solute transport. Plant
Physiol., 104, 189 -- 199. (doi:10.1104/pp.104.1.189)
[54] Tyree, M. T. & Zimmermann, M. H. 2002 Xylem Structure and the Ascent of
Sap. Springer Series in Wood Science. Berlin: Springer-Verlag, 2nd edn.
[55] Visintin, A. 1996 Models of Phase Transitions, vol. 28 of Progress in Nonlinear
Differential Equations and Their Applications. Boston: Birkhauser.
[56] Wiegand, K. M. 1906 Pressure and flow of sap in wood. Amer. Nat., 40(474),
409 -- 453.
[57] Yang, S. & Tyree, M. T. 1992 A theoretical model of hydraulic conductivity
recovery from embolism with comparison to experimental data on Acer saccha-
rum. Plant Cell Env., 15, 633 -- 643. (doi:10.1111/j.1365-3040.1992.tb01005.x)
[58] Zimmermann, U., Schneider, H., Wegner, L. H. & Haase, A. 2004 Water ascent
in tall trees: does evolution of land plants rely on a highly metastable state?
New Phytol., 162(3), 575 -- 615. (doi:10.1111/j.1469-8137.2004.01083.x)
[59] Zwieniecki, M. A. & Holbrook, N. M. 2009 Confronting Maxwell's demon: bio-
physics of xylem embolism repair. Trends Plant Sci., 14(10), 530 -- 534. (doi:
10.1016/j.tplants.2009.07.002)
Article submitted to Royal Society
|
1801.08708 | 2 | 1801 | 2018-05-11T17:00:47 | Anti-margination of microparticles and platelets in the vicinity of branching vessels | [
"physics.bio-ph",
"cond-mat.soft",
"physics.comp-ph"
] | We investigate the margination of microparticles/platelets in blood flow through complex geometries typical for in vivo vessel networks: a vessel confluence and a bifurcation. Using 3D Lattice-Boltzmann simulations, we confirm that behind the confluence of two vessels a cell-free layer devoid of red blood cells develops in the channel center. Despite its small size of roughly one micrometer, this central cell-free layer persists for up to 100 $\mu$m after the confluence. Most importantly, we show from simulations that this layer also contains a significant amount of microparticles/platelets and validate this result by in vivo microscopy in mouce venules. At bifurcations, however, a similar effect does not appear and margination is largely unaffected by the geometry. This anti-margination towards the vessel center after a confluence may explain in vivo observations by Woldhuis et al. [Am. J. Physiol. 262, H1217 (1992)] where platelet concentrations near the vessel wall are seen to be much higher on the arteriolar side (containing bifurcations) than on the venular side (containing confluences) of the vascular system. | physics.bio-ph | physics |
Anti-margination of microparticles and platelets in
the vicinity of branching vessels
C. Bächer1, A. Kihm2, L. Schrack1,3, L. Kaestner4, M.W. Laschke5, C. Wagner2, and
S. Gekle1
1Biofluid Simulation and Modeling, University of Bayreuth, Bayreuth, Germany
2Experimental Physics, Saarland University, Saarbrücken, Germany
3Institute for Theoretical Physics, University of Innsbruck, Innsbruck, Austria
4Institute for Molecular Cell Biology, Research Centre for Molecular Imaging and
Screening, Center for Molecular Signaling (PZMS), Medical Faculty, Saarland
5Institute for Clinical & Experimental Surgery, Saarland University, Homburg/Saar,
University, Homburg/Saar, Germany
Germany
Abstract
We investigate the margination of microparticles/platelets in blood flow through complex geometries typical for in
vivo vessel networks: a vessel confluence and a bifurcation. Using 3D Lattice-Boltzmann simulations, we confirm
that behind the confluence of two vessels a cell-free layer devoid of red blood cells develops in the channel center.
Despite its small size of roughly one micrometer, this central cell-free layer persists for up to 100 µm after the
confluence. Most importantly, we show from simulations that this layer also contains a significant amount of
microparticles/platelets and validate this result by in vivo microscopy in mouce venules. At bifurcations, however,
a similar effect does not appear and margination is largely unaffected by the geometry. This anti-margination
towards the vessel center after a confluence may explain in vivo observations by Woldhuis et al. [Am. J. Physiol.
262, H1217 (1992)] where platelet concentrations near the vessel wall are seen to be much higher on the arteriolar
side (containing bifurcations) than on the venular side (containing confluences) of the vascular system.
Introduction
Red blood cells (RBCs) fill up to 45 volume percent of human blood [28, 35, 58, 67, 70] and thus represent by far
the major cellular blood constituent. Due to their high deformability, RBCs flowing through a cylindrical channel
or blood vessel prefer the low-shear rate region in the center of the channel. By hydrodynamic interactions with
the red blood cells stiffer particles such as platelets, white blood cells or artificial drug delivery agents are thus
expelled towards the wall. This separation of red blood cells and stiffer particles is known as margination and is
essential for the ability of blood platelets to quickly stop bleeding or for drug delivery agents to closely approach
the endothelial wall. One of the first observations of margination studied white blood cells in vivo as well as an in
vitro model system containing disks and spheres already in 1980 [75]. Interestingly, an in vivo study by Woldhuis
et al.
[94] demonstrated a striking difference between the platelet distribution in arterioles and venules, with
significantly more margination occurring on the arteriolar side of the vascular system. Since then, more detailed
insights were gained by experimental studies [9, 11, 12, 16, 17, 19, 21, 26, 42, 52, 63, 91], computer simulations
1
[22, 24, 27, 30, 33, 40, 44–46, 48, 56, 57, 61, 62, 73, 79, 83, 88, 89, 98, 99] as well as theoretical modeling
[13, 41, 71, 72, 85]. These studies all deal with margination in spatially constant geometries such as shear flow,
pipes with cylindrical or rectangular cross-section or plane-Couette systems.
In contrast, in living organisms blood vessels form a hierarchical structure where large arteries branch all the
way down to microcapillaries in a series of bifurcations followed by a reversed series of confluences leading
up to larger and larger vessels on the venular side. The typical distance between two bifurcations lies within
0.4 mm up to 1 mm in the microvascular system [35, 70]. Despite their importance, studies on RBC distribution
and margination using spatially varying geometries are surprisingly scarce. Platelets have been studied by 2D
simulations in the vicinity of an aneurysm [60, 95] and in the recirculation zone behind a sudden expansion
of a channel [100]. Near a vessel constriction the locally varying distribution of RBCs [20, 31, 90] and rigid
microparticles in suspension with RBCs have been investigated [8, 78, 93, 97]. In ref. [8] a local increase in
microparticle concentration in front of the constriction has been reported. On a technical side, microchannels
including bifurcations are investigated as a possible basis for microdevices separating blood plasma [53, 86]. In
asymmetric branches the Zweifach-Fung effect [32, 82] describes an asymmetric red blood cell distribution, i.e.
a larger hematocrit in the large flow rate branch [68, 69]. Under certain circumstances even an inversion of the
Zweifach-Fung effect may occur [77]. Combining experiments and simulations in a rectangular channel with
bifurcation [51] and for a diverging and converging bifurcation using 2D simulations [92] a cell-depleted zone
right after the confluence has been reported. Downstream a bifurcation an asymmetry of red blood cell distribution
has been seen [5, 51, 53, 65]. Ref. [64] reported margination of hardened RBCs while flowing through branching
vessels. Balogh and Bagchi [3] investigated transient behavior of red blood cell motion in more complex networks
[3, 4]. Besides the red blood cell behavior it is important to consider suspended particles like blood platelets or
synthetic particles, since possible influences of bifurcations and confluences may play a major role in medical
applications. Nevertheless, systematic studies covering particle blood suspensions in networks are scare. White
blood cell motion in asymmetric bifurcations has been studied experimentally in the context of a branched vessel
geometry [96] while Sun et al. [81] investigated the interaction of six red blood cells flowing behind a white blood
cell in the vicinity of vessel junctions.
In this report, we study margination of stiff spherical particles suspended among red blood cells in the vicinity
of a vessel confluence and bifurcation. This allows us first to confirm and investigate quantitatively the previously
observed cell-depleted layer of red blood cells after a confluence. Second, we provide results on the influence of
network geometries on stiff particle margination. Our generic stiff particles are a model for artificial drug delivery
agents, but also serve as a reasonable approximation for blood platelets. We investigate two cylindrical branches
either bifurcating from or forming a confluence into a larger vessel by means of 3D Lattice-Boltzmann simulations.
In order to realize simulations of these systems we implement inflow and outflow boundary conditions to the
Lattice-Boltzmann/Immersed-Boundary algorithm similar to a dissipative particle dynamics based approach by
Lykov et al. [55]. Behind a vessel confluence, we observe a RBC-free layer in the center of the channel persisting
for up to 100 µm after the confluence. Importantly, this central cell-free layer is not only devoid of red blood
cells, but in addition contains a significant amount of anti-marginated microparticles/platelets. Using fluorescent
microparticles in mouse microvessels we consistently observe this anti-margination also in vivo. At bifurcations,
no equivalent effect occurs. Our findings may explain in vivo observations of Woldhuis et al. [94], who found
that in the vascular system platelet margination is strongly present at the arteriolar side with bifurcations, but
less at the venular side with confluences. Similarly, recent observations show that thrombi formed in arterioles
contain significantly more platelets than thrombi formed in venules [10] which is another indication of increased
platelet margination on the arteriolar side. By considering the axial distribution of microparticles along the flow
direction we furthermore reveal the site of confluence as a spot with locally increased concentration. The paper is
organized as follows: we first introduce the simulation and experimental methods and then report two-dimensional
and one-dimensional concentration profiles first in the system with vessel confluence including the experimental
results then in the system with a bifurcation. Finally, we investigate the influence of larger hematocrit as well as
microparticle distribution in an asymmetric bifurcation.
2
Methods
Lattice-Boltzmann/Immersed Boundary Method
Fluid flow in the confluence/bifurcation geometry is modeled using a 3D Lattice-Boltzmann method (LBM) which
calculates fluid behavior by a mesoscopic description [1, 18, 80]. We use the implementation of LBM in the
framework of the simulation package ESPResSo [2, 54, 74]. Red blood cells and particles are modeled using the
immersed boundary method (IBM) [8, 33, 59, 66].
In order to mimic realistic conditions of blood flow we assign the blood plasma density ρplasma = 1000 kg/m3
and viscosity µplasma = 1.2 × 10−3 Pas. Due to the large size of red blood cells and microparticles we expect the
temperature to hardly affect collective flow behavior of cells and particles and thus neglect the effect of thermal
fluctuations. A typical fluid grid of the present simulations contains 170x110x58 nodes for a bifurcation and
288x110x58 nodes for a confluence. The time step is chosen as 0.09 µs with the time of a typical simulation being
about 2.5 s.
Red blood cells and microparticles are realized by an infinitely thin elastic membrane interacting with the fluid.
For the calculation of elastic forces imposed on the fluid, the membrane is discretized by nodes that are connected
by triangles. A red blood cell possesses 1280 triangles and 642 nodes and has a diameter of 7.82 µm. The
averaged distance between neighboring nodes is about one LBM grid cell. Nodes transfer forces to the fluid and
are themselves convected with the local fluid velocity. Interpolation between membrane nodes and fluid nodes
is done using an eight-point stencil. The viscosity contrast of the cells is λ = ηin/ηout = 1, i.e.
the fluid
inside and outside the cells has the same viscosity. The elastic properties of a red blood cell are achieved by
applying the Skalak model [6, 15, 28] with a shear modulus kS = 5 × 10−6 N/m and an area dilatation modulus
kA = 100kS. Additional bending forces are computed on the basis of the Helfrich model using a bending modulus
kB = 2 × 10−19 Nm [28, 37, 38]. For the calculation the algorithm denoted method A in [38] is used with
the bending energy being proportional to the angle of adjacent triangles and the actual forces being computed
by analytically differentiating the energy with respect to node position. This somewhat simplistic approach is
appropriate for the present work where we focus on collective rather than detailed single cell behavior and where
especially the behavior of the microparticles is of interest.
Microparticles are modeled in a similar fashion as the red blood cells with 320 triangles and 162 nodes. The
microparticles are chosen to have half the size of RBCs (a = 3.2 µm), which has been reported to show strong
margination [61]. In contrast to the RBCs the microparticles contain an additional inner grid to ensure the stiffness
and (approximate) non-deformability of the microparticles [33]. The inner grid is linked to the membrane nodes
by a harmonic potential. Elastic properties of microparticles are chosen 1000 times larger than for RBCs. For the
purpose of numerical stability we apply an empirical volume conservation potential [44] as well as a short ranged
soft-sphere repulsion, which decays with the inverse fourth power of the distance and with a cut-off radius equal
one grid cell. The latter potential acts between all particles and between the particles and the channel wall.
Stability and accuracy of our simulation method are extensively validated in [33, 38] and in the Supplemental
Information. The shapes of isolated red blood cells in cylindrical and rectangular channels have been validated to
agree with methodically very different Dissipative Particle Dynamics [25] and Boundary-Integral [39] simulations.
Channel geometry
The systems of interest, a confluence and a bifurcation, are shown in figure 1 a). Both geometries are constructed
in the same way: a main cylindrical channel of radius Rch branches into two symmetric daughter channels of
radius Rbr. In order to obtain a smooth boundary, i.e., the boundary itself and the first derivative is continuous, the
transition between main channel and the branches is modeled by third order polynomials: one polynomial, yc(x),
describes the bifurcating centerline, another, yup(x), describes the upper/lower boundary. By rotation around the
centerline with radius yup(x) − yc(x) we obtain a circular cross-section for each x forming the bifurcation along
the flow direction. Where the cross-sections of the two branches overlap the boundary is left out.
3
a)
b)
Fig. 1: a) Systems of interest: a suspension of red blood cells and microparticles flowing either through a confluence (left) or a bifurcation
(right). Rectangles with numbers refer to figures containing corresponding 2D radial/planar projections while dashed lines refer to
figures containing cross-sectional profiles. b) Inflow using straight cylinders as feeding systems: whenever a cell/particle crosses
the indicated plane, it is fed into one of the branches of the confluence system.
Inflow/outflow boundary condition for IBM-LBM
In order to investigate a confluence and a bifurcation as displayed in figure 1 a) periodic boundary conditions cannot
be employed. (Joining both geometries into one very large system would be computationally far too expensive due
to the very long-ranged influence of bifurcations/confluences as will be detailed in the course of this work.) At
the same time, at the entrance of the branches in figure 1 a) left as well as at the entrance of the main channel
in figure 1 a) right, we do not want the microparticles and red blood cells to enter in a randomly distributed
fashion, but instead obey a marginated configuration in order to match with the well-known behavior in a long
tube. The purpose behind this inflow condition is to bring out clearly how the behavior of the marginated fraction
of microparticles is influenced by the confluence/bifurcation.
To meet these requirements we implemented inflow and outflow boundary conditions to our IBM-LBM algo-
rithm similar to a recent work using Dissipative Particle Dynamics [55]. We start by simulating straight cylinders
with periodic boundaries and a body force driving the flow as depicted at the top of figure 1 b). These feeding
simulations yield a time dependent sequence which then serves for particle inflow in the complex system of inter-
est. During the simulation of the complex system we check a frame of the feeding sequence for cells and particles
crossing a certain (arbitrary) plane. Particles crossing this plane, are then inserted at the same radial position with
the same shape into the complex system as illustrated also in figure 1 b). For crossing of the plane the center-
of-mass serves as criterion similar to [55]. The inflow velocity is chosen to match the flow rate prescribed in the
straight cylinder. In order to prevent overlap of cells during inflow we sometimes increase the flow rate slightly
(about 10%). As a result we obtain a marginated pattern at the entrance of our complex systems as proven by the
cross-sectional concentration profiles shown in the Supplemental Information.
Since azimuthal motion of the dilute microparticles in the feeding channels is extremely slow, even a very long
feeding simulation would lead to a biased distribution of microparticles upon entering the complex channel. This is
prevented by applying a small angular random force to the microparticles in the feeding channel thus guaranteeing
an azimuthally homonegenous, yet well marginated distribution of microparticles. We furthermore show in the
Supplemental Information that after a first filling of the system the cell and particle number in the system reaches
a plateau and slightly fluctuates around a constant value.
For the fluid, in order to prescribe a distinct flow rate we assign a constant velocity to all fluid LBM nodes at
the beginning of the simulation box. The same is done at the end of the box taking into account the different
cross-sections of main channel and branches, thus matching fluid inflow and outflow. About 15 grid cells behind
the inflow the flow profile matches tube flow. This region with evolving flow profile is skipped for particle inflow
4
0150255075100125xinµm4a,e4d,h4c,g4b,f3a,c3b,d09015304560758c,f8b,e8a,dxinµm7a,c7b,dxyz and in data analysis.
Fig. 2: a) Color labeling and b) labeling criteria for red blood cells and microparticles at the entrance of the system with respect to the
in-plane position in the cross-section. Due to symmetry the particles left and right can be treated equally.
Analysis
In our work, we employ different concentration profiles at given positions along the channel. First, we compute
cross-sectionally averaged concentrations leading to 1D concentration profiles as a function of position along the
flow direction x. Some of these profiles consider only a certain fraction of cells or particles entering the channel in
specific regions which are labeled as a function of their lateral (r,φ) position in polar coordinates at the entrance of
the system. Corresponding concentrations are calculated taking into account only this particular fraction of cells
or particles. The labeling is illustrated by the color code in figure 2 a) and by the criteria for r and φ in figure 2 b).
Second, we use three types of two dimensional concentration profiles. For microparticle concentrations, 2D
radial projections in the r, x plane are calculated which reflect the radial symmetry of the main and the branch
channels. Such projections, however, are not appropriate to understand non-radially symmetric effects occuring
near confluences or bifurcations. We thus employ in addition, mainly for red blood cell concentrations, planar
projections of the 3D concentrations on the y, x plane by integrating the concentration over the z direction per-
pendicular to the plane of the paper. Finally, in order to get further insight into cell and particle distributions
perpendicular to the flow direction, we calculate cross-section profiles within the y, z plane.
All concentration profiles are averaged over the whole simulation time starting from the moment at which the
number of cells and particles does not vary significantly.
Preparation of dorsal skinfold chamber and in vivo imaging
Animals The in vivo experiments were performed in 10-12 week old male C57BL/6 mice (n = 3) with a body
weight of 23-26 g. The animals were bred and housed in open cages in the conventional animal husbandry of
the Institute for Clinical & Experimental Surgery (Saarland University, Germany) in a temperature-controlled
environment under a 12 h/12 h light-dark cycle and had free access to drinking water and standard pellet food
(Altromin, Lage, Germany). The experiment was approved by the local governmental animal care committee
(approval number 06/2015) and was conducted in accordance with the German legislation on protection of animals
and the NIH Guidelines for the Care and Use of Laboratory Animals (Institute of Laboratory Animal Resources,
National Research Council, Washington, USA).
Dorsal skinfold chamber model Microvessels were analyzed in the dorsal skinfold chamber model, which
provides continuous microscopic access to the microcirculation of the striated skin muscle and the underlying
subcutaneous tissue [49]. For the implantation of the chamber, the mice were anesthetized by i.p.
injection of
ketamine (75 mg/kg body weight; Ursotamin®; Serumwerke Bernburg, Bernburg, Germany) and xylazine (15
mg/kg body weight; Rompun®; Bayer, Leverkusen, Germany). Subsequently, two symmetrical titanium frames
(Irola Industriekomponenten GmbH & Co. KG, Schonach, Germany) were implanted on the extended dorsal
skinfold of the animals, as described previously in detail [50]. Within the area of the observation window, one layer
of skin was completely removed in a circular area of ∼15 mm in diameter. The remaining layers (striated skin
muscle, subcutaneous tissue and skin) were finally covered with a removable cover glass. To exclude alterations
of the microcirculation due to the surgical intervention, the mice were allowed to recover for 48 h.
In vivo microscopy In vivo microscopic analysis was performed, as previously described [7]. In detail, the mice
were anesthetized and a fine polyethylene catheter (PE10, 0.28 mm internal diameter) was inserted into the carotid
5
a)centerrightlefttopbottomb)rφcenterr<=Rch/2∀φtopr>Rch/245<=φ<135bottomr>Rch/2225<=φ<315leftr>Rch/2135<=φ<225rightr>Rch/2315<=φ&φ<45 RBC
RBC
a)
c)
micro
b)
d)
micro
Fig. 3: Concentration of red blood cells in a confluence in 2D planar projection a) along the upper branch and b) along the main channel.
Microparticle concentration in 2D radial projection c) along the upper branch and d) along the main channel. The cell-free layer
near the inner boundary decreases at the end of the branches while it increases near the outer boundary. Inside the main channel an
additional cell-free layer in the center develops.
artery for application of the plasma marker 5 % fluorescein isothiocyanate (FITC)-labeled dextran 150,000 (Sigma-
Aldrich, Taufkirchen, Germany) and microspheres (Fluoresbrite Plain YG 1.0 µm; Polysciences, Warrington, PA,
USA). Then, the animals were put in lateral decubital position on a plexiglas stage and the dorsal skinfold chamber
was attached to the microscopic stage of an upright microscope (Axiotech; Zeiss, Jena, Germany) equipped with a
LD EC Epiplan-Neofluar 50x/0.55 long-distance objective (Zeiss) and a 100 W mercury lamp attached to a filterset
(excitation 450-490 nm, emission > 520 nm). The microscopic images were recorded using a CMOS video camera
(Prime 95B, Photometrics, Tucson, AZ, USA) at an acquisition speed of 415 images per second controlled by a
PC-based acquisition software (NIS-Elements, Nikon, Tokyo, Japan).
Trajectory analysis The recorded video sequence was analyzed using a single particle tracking algorithm.
Hereby, the intensity profile of each frame was adjusted to have both the top and bottom 1% of all pixels saturated,
correcting for changes in illumination and exposure time. With the aid of a tailored MATLAB script, all spherical
(round) objects were detected and interconnected among all frames by cross-correlating consecutive images. To
only detect microspheres (and not red blood cells), we set a threshold of 0.9 as a lower limit in normalized intensity
values, since they are fluorescent. Further, we defined a minimal diameter for the detected particles (0.7 µm),
causing a trajectory to end if the measured value falls below this value. Combining the coordinates of all classified
microspheres in this way over the whole video sequence, we derived the respective trajectories.
Channel confluence
We first investigate the system with two branches of radius 16 µm merging into one main channel of radius 17.5 µm
as depicted in figure 1 a) left. The centerlines of the two branches are separated by 39 µm and the transition zone
from the end of the branches to the beginning of the main channel is about 13 µm. We choose the mean velocity at
the entrance of our system to be about v = 2.5 mm/s. Simulations are first performed for a physiologically realistic
hematocrit (red blood cell volume fraction) of Ht = 12%. Results for a higher Ht = 20% are qualitatively similar
and are presented at the end of this contribution. The Reynolds number calculated from the centerline velocity, the
red blood cell radius RRBC, and the kinematic viscosity of the fluid νplasma is Re = RRBC·v
= O(10−2).
ν
6
Fig. 4: 2D cross-sectional concentration in a confluence for red blood cells a)-d) and microparticles e)-h) at the end of the branches a),e), at
the beginning of the main channel b),f), at the middle of the main channel c),g), and at the end of the main channel d),h). Positions
are also indicated by the dashed black lines in figure 1 (a). In the main channel a clear cell-free layer in the center together with
microparticle anti-margination is present.
Cell and particle distribution
We start by considering 2D concentration profiles along flow direction in figure 3. In the two small branches
and far away from the confluence we observe a homogeneous distribution of red blood cells around the center
and the cell-free layer with vanishing concentration [23, 29, 43] near the wall as can be seen by figure 3 a). The
microparticle concentration in figure 3 c) exhibits the typical margination peak near the wall. The state of full
and azimuthally homogeneous margination is confirmed by the cross-sections at channel entrance shown in the
Supplemental Information. This behavior is the same as in a straight channel.
Approaching the confluence we observe an asymmetric cell-free layer: near the inner boundary of the branch
the cell-free layer decreases, near the outer boundary it increases. The asymmetry becomes more pronounced
towards the end of the branches (x ≈ 40 µm) stemming from cells flowing towards the main channel. However,
the motion of red blood cells towards the main channel already initiates at x ≈ 30 µm, i.e., about 10 µm before
the end of the branches. This also affects microparticle behavior as can be seen by the two separate peaks in the
radially projected concentration in figure 3 c). These two peaks stem from the particles near the inner and the outer
boundary, respectively: the microparticles near the inner boundary remain close to the wall, while the particles near
the outer boundary migrate away from the wall due the flow profile towards the main channel and the increased
cell-free layer.
Entering the main channel we observe a decreased cell-free layer near the upper and lower boundary in figure
3 b) right at the beginning. An interesting feature of the RBC concentration is the additional cell-free layer which
develops in the channel center after the confluence. This agrees well with the findings of ref. [51] that a cell-
depleted zone behind the apex of a confluence develops. We confirm that finding by considering the concentration
and in addition highlight the long-range stability of the central cell-free layer.
Most remarkably, this central cell-free layer contains, just as its classical near-wall counterpart, a significant
amount of microparticles as can clearly be seen in the radially projected microparticle concentration of figure 3 d).
Before investigating further this central cell-free layer, we consider cross-sectional concentration profiles in
figure 4. In figure 4 a), we observe how the circular pattern of red blood cells is shifted towards the inner boundary
at the end of the branches. This corresponds to the asymmetric cell-free layer in figure 3 a). Entering the main
channel the pattern of red blood cell concentration possesses two flattened and asymmetric spots (figure 4 b))
clearly separated by the central cell-free layer which shows vanishing concentration. At the left and right of the
main channel an additional large cell-free spot is obtained. This central-cell-free layer stems from the cells flowing
out of the upper and lower branch competing for the channel center.
7
a)RBCx=40µmb)RBCx=55µmc)RBCx=118µmd)RBCx=158µme)microx=40µmf)microx=55µmg)microx=118µmh)microx=158µm Microparticles in the two branches remain well marginated until the end of the branches as shown in 4 e). After
the confluence, however, figure 4 f) shows how a notable fraction of microparticles is now located very near the
channel center. This can be understood by the original location of the microparticles inside the branches: those
microparticles that are located near the inner boundary of the branch enter the main channel in the center. This
location is favorable due to the additional central cell-free layer observed in figure 4 b). In the Supplemental
Information we show the anti-margination also for platelet-shaped microparticles. Thus, the present geometry
leads to a re-distribution of microparticles from a near-wall marginated position before the confluence to a near-
center anti-marginated position after the confluence.
Lifetime of the central-cell-free layer, anti-margination, and physiological consequences
An interesting question is the stability of the RBC-depleted central cell-free layer and the corresponding anti-
marginated microparticles as the flow continues away from the confluence location along the main channel. Figure
3 b) shows that the central-cell-free layer is surprisingly stable, being visible all along the main channel and only
becoming slightly blurred towards the end. The same trend can be observed in the cross-sectional concentration at
three sites along the main channel in figure 4 b),c),d). Although it starts to become blurred after 60 µm in figure
4 c), the central cell-free layer and especially the cell-free spot left and right is visibly present until at least 100 µm
behind the confluence as shown in figure 4 d). Similarly, the corresponding microparticle concentration in figure
4 f),g),h) shows that microparticles are located in the center all along the main channel. Thus, a confluence of two
channels influences microparticle behavior over distances which are much longer than the channel diameter.
To gain a more mechanistic insight into this long-time stability of the central CFL, we calculate the (shear-
induced) diffusion coefficient [34, 36, 43, 47, 87, 98, 99] of red blood cells in the center. For this, we compute
the time-dependent mean squared displacement (MSD) which is shown in the Supplemental Information. By
modeling the increase in MSD with time by the theoretical expectation for normal diffusion h∆y(t)2i = 2Dt we
extract a diffusion coefficient for the red blood cells of DRBC ≈ 28 µm2/s in the case of Ht = 12%. This value
is of the same order as previous results in experiments with red blood cells [34, 36, 87] and simulations of spheres
and platelets [98, 99]. By assuming a thickness of 1.5 µm and a flow speed of 2.5 mm/s we calculate a distance
of 100 µm required to bridge the central cell-free layer. This length scale agrees well with the observation in the
concentration profiles that the central cell-free layer starts to become blurred after 100 µm. We calculate in the
same way the shear-induced diffusion coefficient of the microparticles Dmicro ≈ 25 µm2/s. Ref. [87] reports
for platelets in a perfusion chamber 34 µm2/s for a shear rate of 832 1/s and unknown hematocrit. Ref. [98] and
[99] obtain a diffusivity a factor 2 smaller in simulations of plane-Couette flow with Ht = 0.2. Considering not
only the gap of the central cell-free layer to be closed, but the larger spot left/right (assuming a distance of 5 µm
to be bridged) we can estimate a distance of 1.1 mm for red blood cell redistribution, which is comparable to the
estimation of 25D by Katanov et al. [43]. When we estimate the length scale required to migrate towards the
channel wall, i.e. to marginate, we get about 5 mm. Comparing this to the typical distance between successive
confluences of about 0.4-1 mm [35, 70] we conclude that full margination cannot be regained. This in turn may
explain in vivo observations that on the venular side of the vascular system margination is much less pronounced
than on the arterioral side [10, 94].
Furthermore, we want to address the question how strong the effect of anti-margination is. Therefore, we
calculate the fraction of particles that are not located directly next to the wall, i.e. we consider particles that are
more than one particle radius away from the vessel wall. For the concentration profiles in fig. 4 we obtain the
fractions 0 at x = 40 µm, 0.158 at x = 55 µm, 0.138 at x = 118 µm, and 0.135 at x = 158 µm. We find that the
fraction decreases very slowly with increasing distance from the confluence because of marginating microparticles
(most likely those located left/right).
Axial concentration
Further insight can be gained by considering 1D axial cell and particle concentration profiles in figure 5. The over-
all red blood cell concentration exhibits two plateaus inside the main channel and inside the branches, respectively.
Inside the branches the concentration is lower than in the main channel in agreement with the Fahraeus effect
[76]. Right at the confluence, we observe a zone where the RBCs become slightly depleted. The microparticle
8
Fig. 5: 1D axial profile of red blood cells and microparticles flowing through a vessel confluence. While the red blood cells are depleted at
the site of the confluence, the microparticles exhibit a concentration increase of about 50% compared to the branches and the main
channel.
concentration along the branches first increases slightly and right at the confluence a strong peak develops. In the
main channel the microparticles have a nearly constant concentration.
In order to elucidate red blood cell behavior further and, especially, in order to explain the peak in microparticle
concentration we label the cells/particles while entering the branches as explained in the Methods section. The
concentrations for the labeled cells and particles are shown in figure 6.
a)
b)
Fig. 6: Axial concentration of a) red blood cells and b) microparticles distinguished regarding their position inside the cross-section of the
branches as illustrated in figure 2. The microparticles entering at the top of the upper branch (or, equivalently, the bottom of the
lower branch) exhibit a pronounced peak.
The red blood cells in the center, those left/right and those at the bottom of the branches behave in a similar
fashion: at the end of the branches they are accelerated and thus have a decreased concentration. After a small
peak, they quickly reach a constant concentration in the main channel. Only the cells arriving at the top exhibit a
slightly increasing concentration at the end of the branches, but are depleted, as well, at the site of the confluence.
Microparticles arriving left/right or at the bottom show a similar concentration profile as the corresponding red
blood cells and thus do not cause the peak in overall concentration in figure 5. We note that these concentration
profiles can be understood by passive tracer particles similar as in constricted channels [8] and as detailed in the
Supplemental Information. From figure 6 b) we are thus able to conclude that the peak stems from the micropar-
ticles flowing at the top of the branches. The concentration of these microparticles increases more than two-fold
compared to the branches and the main channel. Due to the margination the microparticles are located right be-
sides the wall. Also the concentration profile of the microparticles at the top can be reproduced by passive tracer
particles as done in the Supplement Information. Thus, the local increase in microparticle concentration can be
understood by the underlying flow profile.
In vivo observation of microparticle anti-margination
To demonstrate the relevance of the anti-margination observed in our simulations, we inject fluorescent beads into
living mice and image their behavior when flowing through a microvessel confluence. In figure 7 we show a set
of trajectories obtained from the video microscopy images (a corresponding movie is included as Supplemental
9
0.08 0.1 0.12 0.14 0.16 0.180102030405060708090100110120130140150 0.6 0.8 1 1.2 1.4 1.6 1.8103 · ρmicro [µm-3]103 · ρRBC [µm-3]x [µm]microRBC 0.2 0.4 0.6 0.8 10102030405060708090100 0.1 0.15 0.2 0.25 0.3 0.35 0.4103 · ρRBC [µm-3]103 · ρRBC [µm-3]x [µm]centerleft/righttopbottom 0 0.02 0.04 0.06 0.08 0.1 0.120102030405060708090100103 · ρmicro [µm-3]x [µm]left/righttopbottom Information). In agreement with the predictions of our numerical simulations, beads which are initially marginated
at the outer walls (blue lines in figure 7) remain marginated whereas beads located initially at the inner walls (red
lines in figure 7) undergo anti-margination and end up near the channel center after passing through the confluence.
Fig. 7: In vivo measurement of tracked fluorescent beads in mouce microvessels. Blue lines show trajectories of beads which remain
marginated after the bifurcation while red lines show beads undergoing anti-margination. Yellow dashed lines denote the vessel
boundaries.
Channel bifurcation
Next, a bifurcation is investigated as depicted in figure 1 a) right. The suspension of red blood cells and mi-
croparticles flows through a straight channel of radius 16 µm branching into two daughter channels of radius
11.5 µm. Main channel and the combined branches have the same cross-sectional area and the centerlines of the
two branches are separated by 34 µm. The transition zone from the end of the main channel to the beginning of
the branches is about 13 µm.
RBC
RBC
a)
c)
micro
b)
d)
micro
Fig. 8: 2D planar projection within the bifurcation for red blood cells a) along the main channel and b) along the upper branch. 2D radial
projection of microparticle concentration c) along the main channel and d) along the branches. The cell-free layer decreases at the
end of the main channel and we observe an asymmetric cell distribution inside the branches. The margination peak of microparticles
is somewhat blurred after the bifurcation, but otherwise unaffected.
10
outflowinflow10 µminflow Fig. 9: 2D cross-sectional concentration of red blood cells a),b),c) and microparticles d),e),f) at the end of the main channel a),d), at the
beginning of the branches b),e), and at the end of the branches c),f). See figure 1 (a) for indications of the respective positions along
the channel. Inside the branches an asymmetric cell-free layer develops and microparticles suffer a loss of concentration directly
besides the inner wall.
Cell and particle distribution
In figure 8 we first investigate the two-dimensional concentration of red blood cells and microparticles along the
flow direction. At the very beginning around x ≈ 0 we again observe a homogeneous red blood cell distribution
(figure 8 a)) around the center and the cell-free layer with vanishing red blood cell concentration near the wall.
Approaching the bifurcation the cell-free layer decreases. The decrease in cell-free layer is of the same amount
at both locations near the upper boundary and near the lower boundary. It can be straightforwardly explained by
the bifurcating geometry which causes the red blood cells to flow upwards/downwards into the daughter channels.
This motion into the daughter channels starts already about 10 µm before the end of the main channel and makes
the cells migrate towards the outer wall.
An asymmetry in the cell-free layer occurs inside the daughter channels (figure 8 b)) as also observed in recent
work [53, 65]. The asymmetry is especially pronounced at the beginning of the branches stemming from cells
flowing in the center of the main channel, which enter either the upper or lower branch near the inner wall.
We observe a strongly decreased cell free layer near the inner boundary right at the beginning of the daughter
branches. The thickness of the cell-free layer near the outer boundary increases correspondingly. After about
10 µm the inner and outer cell-free layers both reach a constant value, which is similar to the length scale for
the re-establishment of the outer cell-free layer after the confluence. Interestingly, the inner and outer cell-free
layers do remain asymmetric and retain this asymmetry until the very end of our channel about 50 µm behind the
bifurcation.
Although the cell-free layer decreases at the end of the main channel hardly an effect is observed on micropar-
ticle behavior. Entering the branches the microparticle concentration peak only becomes more blurred because of
the microparticles located near the upper boundary entering the larger cell-free layer inside the branch.
The asymmetries in cell and particle distribution can be seen in more detail in the 2D cross-section profiles in
figure 9. At the end of the main channel in figure 9 a) the red blood cell distribution is still circular with only
small deviations corresponding to the decrease in cell-free layer seen in figure 8 a). At the beginning of the branch
the circular red blood cell concentration is strongly shifted towards the inner boundary in figure 9 b) and a less
pronounced, but still clearly visible asymmetry is still present at the end of the branch in figure 9 c). Furthermore,
a local spot with increased red blood cell concentration is observed near the inner boundary.
While the microparticles are hardly affected at the end of the main channel (figure 9 d)) a notable effect is
the vanishing concentration of microparticles near the inner boundary of the branches in figure 9 e) and f). Over
an angle-range of about 90 degrees at the bottom of figure 9 e) and f) the microparticle concentration vanishes
11
a)x=34µmRBCb)RBCx=57µmc)RBCx=74µmd)microx=34µme)microx=57µmf)microx=74µm completely. The vanishing microparticle concentration can be understood by the radial distribution in the main
channel due to margination: in order to reach positions near the lower boundary of the branch the microparticles
would have to be located near the center of the main channel, which is not the case because of margination. Thus,
we report a region within the branches that posses vanishing microparticle concentration in comparison to a simple
straight channel.
Fig. 10: Axial concentration of red blood cells and microparticles flowing through a bifurcating channel. Both red blood cells and mi-
croparticles exhibit a peak in front of the bifurcation apex. The microparticle concentration increases directly in front of the apex,
while the red blood cell concentration exhibits a second small peak due to a second cell flowing onto a cell already being stuck at
the apex.
Axial concentration
As for the confluence, we now investigate the behavior of cells and particles along the varying geometry by 1D
axial concentration profiles in figure 10. After a constant plateau inside the main channel both red blood cells and
microparticles show a clear peak ahead of the apex of the bifurcation. Inside the branches the red blood cells take
the same concentration as in the main channel while the microparticle concentration decreases. The latter can be
explained by the intrinsic velocity profile: flowing besides the boundary the stiff microparticles cover a certain
ring of tube diameter along the boundary. Due to the fixed particle size this ring has the same diameter in the main
channel and within the branches. Assuming a Poiseuille flow and averaging over such a ring around the boundary
leads to a higher flow rate inside the branches, since they cover a larger part of the steep velocity profile. Thus,
within the branch the microparticles experience a larger velocity, leading to lower residence time thereby causing
a decreasing concentration.
a)
b)
Fig. 11: Axial concentration of a) red blood cells and b) microparticles distinguished regarding their initial position inside the cross-section
of the main channel. The peak in red blood cell concentration stems from the cells trapped at the apex of the bifurcation. The
microparticles exhibit a similar peak when arriving left/right. However, for microparticles entering top/bottom no peak occurs.
We furthermore distinguish the cells and particles regarding their position inside the main channel and calculate
the axial concentration profiles for each cell/particle fraction in figure 11 a) and b). All fractions of red blood cells
behave in a similar manner with a small peak at the apex. This peak stems from cells being trapped at the apex
12
0.05 0.07 0.09 0.11 0.13 0.15 0.170102030405060708090 0.7 0.8 0.9 1 1.1 1.2 1.3103 · ρmicro [µm-3]103 · ρRBC [µm-3]x [µm]microRBC 0.2 0.3 0.4 0.5 0.60102030405060708090103 · ρcRBC [µm-3]x [µm]centertop/bottomleft/right 0 0.02 0.04 0.06 0.08 0.1 0.12 0.140102030405060708090103 · ρmicro [µm-3]x [µm]top/bottomleft/right Fig. 12: a) 2D planar projection within the confluence for red blood cells along the main channel behind a confluence as in figure 3, but with
larger hematocrit Ht = 20%. b)-d) 2D cross-section profiles for b) red blood cells and c),d) microparticles in the main channel as
in figure 4, but for larger hematocrit Ht = 20%.
of the bifurcation: arriving in the center of the main channel, red blood cells have to break symmetry and decide
for one branch. As visible in figure 1 a) right some cells flow directly onto the apex and are trapped there before
flowing in one of both branches, a phenomenon called "lingering" in [3]. We note that also the red blood cells
at the top/bottom show a peak, because they are still located close enough to the center to be influenced by more
central cells getting trapped at the apex.
In full analogy, a similar peak is observed for microparticles located left/right in figure 1 b). Also these mi-
croparticles flow onto the apex and become trapped for a short period of time. In contrast, the microparticles
located top/bottom are diluted at the bifurcation. After a subsequent little dip in concentration, microparticles
from both regions quickly reach a constant concentration inside the branches.
The concentration profile of both cells and particles can be understood again by considering tracer particles (see
Supplemental Information).
Influence of hematocrit
In the following we present simulations which have the same geometrical properties as the channels in figure 1 but
with a hematocrit of Ht = 20% for the inflow.
Figure 12 shows that behind the confluence of two branches the red blood cell distribution behaves qualitatively
similar as for the lower hematocrit. Although the cell-free layer near the vessel wall is reduced compared to
figure 3 b) the cell-free layer in the center of the main channel in figure 12 a) is of about the same size. Only
at the left and right of the cross-section the cell-free space clearly reduces compared to lower hematocrit (figure
12 b)). The central cell-free layer is very pronounced up to 40 µm behind the confluence, but becomes blurred
slightly faster towards the end of the channel when compared to the low hematocrit case. This faster decay can be
explained by the larger shear-induced diffusion coefficient of DRBC = 38 µm2/s. This agrees with the theoretical
expectation that the shear-induced diffusion coefficient depends on the number of cell-cell collision and thus on
the cell concentration [14, 36]. The microparticles are still located in the cell-free layer in the center all along
the main channel as can be seen in figure 12 c) and d). We can again calculate the fraction of anti-marginated
microparticles and obtain the fractions 0.145 at x = 55 µm and 0.137 at x = 158 µm. Also when we compare
the axial concentration of labeled red blood cells and microparticles in figure 13 a) and b) to the case of lower
hematocrit in figure 6 we see similar behavior in both cases.
When we investigate the influence of larger hematocrit on the system with bifurcation, we find that each cell-
free layer in the system decreases with increasing hematocrit (results shown in the Supplemental Information). At
the end of the main channel the cell-free layer still decreases and the pronounced asymmetry of cell-free layers
within the branches is present. The increase in concentration due to the apex of the bifurcation remains unaffected
by larger hematocrit as seen in figure 13 c) and d). Since a certain number of cells or microparticles stacks at the
apex of the bifurcation the effect is not modified when more cells are added to the system. Especially, the absolute
number of cells in the center region stays approximately the same for larger hematocrit.
13
a)RBCb)x=55µmRBCc)x=55µmmicrod)x=158µmmicro a)
c)
b)
d)
Fig. 13: 1D axial concentration of a),c) red blood cells and b),d) and microparticles flowing through a a),b) confluence or c),d) bifurcation
with larger hematocrit Ht = 20% labeled by their radial position at the entrance. The larger hematocrit hardly affects the behavior
of red blood cells and microparticles.
Asymmetric bifurcations
We finally touch briefly on the subject of asymmetric bifurcations. For this, we keep the main channel radius
Rch = 16 µm and the upper branch Rbr = 11.5 µm as in figure 1 right and only vary the diameter of the lower
branch. Two different simulations are done with radius Rlow = 8 µm and 5.5 µm. Here, we focus on the total
concentration of red blood cells and microparticles within the two branches as listed in table 1.
ρup
Rlow
RBC
8 µm
1.027
5.5 µm 1.005
ρlow
RBC
0.794
0.674
ρup
micro
0.0711
0.0722
ρlow
micro
0.0792
0.0806
Tab. 1: Concentration of red blood cells and microparticles in the upper (up) and lower (low) channel of an asymmetric bifurcation with
radius 11µm of the upper branch and Rlow of the lower branch. Concentrations are given in 103
µm−3.
The red blood cell concentration clearly differs between the upper and lower branch. The upper branch, being
the one with larger flow rate, receives clearly more cells than the lower branch. This effect is enhanced when the
diameter and correspondingly the flow rate further decreases as in the case of Rlow = 5.5 µm. The fraction of
concentration between lower and upper branch ρlow/ρup changes from 0.77 to 0.67 when changing Rlow from 8
to 5.5 µm. We note that the total flow rate at the outflow of the system is the same in both simulations to match the
flow rate at the entrance. As a consequence, the flow rate in the upper branch slightly differs in both simulations
(the fraction of flow rates is 0.5 and 0.26, respectively). The asymmetric distribution of red blood cells qualitatively
matches with the Zweifach-Fung effect observed earlier [4, 55, 68, 76] and can be attributed to the cell-free layer
[76] combined with RBC deformability.
In contrast to the asymmetric red blood cell distribution microparticles are nearly evenly distributed to the
daughter channels. Furthermore, the distribution is not affected by decreasing the diameter of the lower branch,
the fraction in both cases being about 1.11. Since the stiff microparticles are located within the cell-free layer the
different flow rate does not affect their distribution. Although the lower branch is significantly smaller the apex of
the bifurcation and thus the separation line between the two branches is located near the center of the main channel
14
0.2 0.4 0.60102030405060708090103 · ρRBC [µm-3]x [µm]centerleft/righttopbottom 0 0.02 0.04 0.06 0.08 0.1 0.120102030405060708090103 · ρmicro [µm-3]x [µm]left/righttopbottom 0.2 0.3 0.4 0.5 0.60102030405060708090103 · ρcRBC [µm-3]x [µm]centertop/bottomleft/right 0 0.02 0.04 0.06 0.08 0.1 0.12 0.140102030405060708090103 · ρmicro [µm-3]x [µm]top/bottomleft/right by construction of the geometry. Thus, arriving near the wall only the microparticles located around the equator
are drawn into the upper branch by the flow while all other microparticles may distribute equally into the branches.
All in all, the microparticles exhibit a very similar concentration in both branches.
Conclusion
We used a 3D Lattice-Boltzmann-Immersed-Boundary method with inflow/outflow boundary conditions to inves-
tigate a mixed suspension of red blood cells and stiff particles flowing through a vessel confluence as well as a
vessel bifurcation. The stiff particles can be regarded as models for synthetic drug delivery agents or naturally
occuring stiff cells such as platelets. In agreement with earlier studies, we observe and quantify the formation
of a pronounced central cell-free layer behind the confluence of two vessels. We find that the central cell-free
layer is very stable being still observable even 100 µm after the confluence. As a consequence, we show that stiff
particles at the confluence are strongly redistributed. Although all stiff particles arrive on a marginated position
inside the well-known near-wall cell free layer, while transversing the confluence a significant fraction of them
undergoes anti-margination ending up trapped in the central cell-free layer near the channel center. This position
is retained even longer than the 100 µm life time of the central cell-free layer itself. Calculating the fraction of
anti-marginated microparticles we found that more than 13% of the particles are located around the center 100 µm
behind a confluence. Under the assumption that at the succeeding confluence this fraction of microparticles is
still anti-marginated we estimate that after 5 confluences half of the initially completely marginated particles is
now evenly distributed across the cross-section of the channel. In contrast, a bifurcating geometry is found to not
significantly influence the margination propensity of stiff particles. For the confluence, we also conducted in vivo
measurements which proved the relevance of anti-margination of stiff microparticles in living mice.
In previous in vivo studies platelets have been observed to be mainly located near the wall in arterioles [84, 94]
but not in venules where the platelet concentration was rather continuous across vessel diameter. In a similar
direction, the recent work of Casa et al. [10] found that thrombi were platelet rich on the arterial but not on the
venous side of the blood vessel network. Our present findings may provide an explanation for these observations.
On the arterial side, the microvascular network consists mainly of bifurcations from larger into smaller and smaller
vessels which, according to our findings, do not significantly disturb the margination propensity of platelets. On
the venous side, however, small channels frequently merge into larger ones. At such confluences, our results
clearly demonstrated anti-margination, i.e., a tendency of platelets to be forced into the center of the vessel. In a
network with a cascade of confluences being only 400-1000 µm apart [35, 70] the platelet margination near the
channel wall will be further and further disturbed, ending up finally in a rather continuous concentration profile
and thus explaining the experimental observations of ref. [94] and [10].
Author Contributions
C.B. performed research, analyzed data, wrote the manuscript. L.S. contributed to simulation tools. S.G. designed
research, wrote the manuscript. A.K. performed image treatment and data analysis, C.W. designed and interpreted
research. L.K. and M.W.L. designed and performed the in vivo experiments.
Acknowledgments
The authors thank the Gauss Centre for Supercomputing e.V. for providing computing time on the GCS Super-
computer SuperMUC at the Leibniz Supercomputing Centre.
This work was supported by the Volkswagen Foundation.
We gratefully acknowledge the Elite Study Program Biological Physics and the Studienstiftung des deutschen
Volkes.
15
Supplemental Information
An online Supplemental Information to this article can be found by visiting BJ Online at http://www.biophysj.org.
References
[1] C. K. Aidun and J. R. Clausen. Lattice-boltzmann method for complex flows. Annual Review of Fluid
Mechanics, 42(1):439–472, 2010.
[2] A. Arnold, O. Lenz, S. Kesselheim, R. Weeber, F. Fahrenberger, D. Roehm, P. Košovan, and C. Holm.
Espresso 3.1: Molecular dynamics software for coarse-grained models. In Meshfree methods for partial
differential equations VI, pages 1–23. Springer Berlin Heidelberg, 2013.
[3] P. Balogh and P. Bagchi. Direct numerical simulation of cellular-scale blood flow in 3d microvascular
networks. Biophysical Journal, 113(12):2815–2826, 2017.
[4] P. Balogh and P. Bagchi. A computational approach to modeling cellular-scale blood flow in complex
geometry. Journal of Computational Physics, 334:280–307, 2017.
[5] J. O. Barber, J. M. Restrepo, and T. W. Secomb. Simulated Red Blood Cell Motion in Microvessel Bifur-
cations: Effects of Cell–Cell Interactions on Cell Partitioning. Cardiovasc Eng Tech, 2(4):349–360, Oct.
2011.
[6] D. Barthès-Biesel. Modeling the motion of capsules in flow. Current Opinion in Colloid & Interface
Science, 16(1):3 – 12, 2011.
[7] M. Brust, O. Aouane, M. Thiébaud, D. Flormann, C. Verdier, L. Kaestner, M. Laschke, H. Selmi, A. Beny-
oussef, T. Podgorski, et al. The plasma protein fibrinogen stabilizes clusters of red blood cells in microcap-
illary flows. Scientific reports, 4:4348, 2014.
[8] C. Bächer, L. Schrack, and S. Gekle. Clustering of microscopic particles in constricted blood flow. Physical
Review Fluids, 2(1):013102, 2017.
[9] E. J. Carboni, B. H. Bognet, G. M. Bouchillon, A. L. Kadilak, L. M. Shor, M. D. Ward, and A. W. Ma.
Direct Tracking of Particles and Quantification of Margination in Blood Flow. Biophysical Journal, 111(7):
1487–1495, 2016.
[10] L. D. C. Casa, D. H. Deaton, and D. N. Ku. Role of high shear rate in thrombosis. J. Vasc. Surg., 61(4):
1068–1080, 2015.
[11] P. Charoenphol, R. B. Huang, and O. Eniola-Adefeso. Potential role of size and hemodynamics in the
efficacy of vascular-targeted spherical drug carriers. Biomaterials, 31(6):1392 – 1402, 2010.
[12] H. Chen, J. I. Angerer, M. Napoleone, A. J. Reininger, S. W. Schneider, A. Wixforth, M. F. Schneider, and
A. Alexander-Katz. Hematocrit and flow rate regulate the adhesion of platelets to von willebrand factor.
Biomicrofluidics, 7(6), 2013.
[13] L. Crowl and A. L. Fogelson. Analysis of mechanisms for platelet near-wall excess under arterial blood
flow conditions. Journal of fluid mechanics, 676:348–375, 2011.
[14] F. R. Da Cunha and E. J. Hinch. Shear-induced dispersion in a dilute suspension of rough spheres. Journal
of Fluid Mechanics, 309:211–223, 1996.
[15] A. Daddi-Moussa-Ider, A. Guckenberger, and S. Gekle. Long-lived anomalous thermal diffusion induced
by elastic cell membranes on nearby particles. Phys. Rev. E, 93(1):012612, 2016.
16
[16] R. D'Apolito, F. Taraballi, S. Minardi, X. Liu, S. Caserta, A. Cevenini, E. Tasciotti, G. Tomaiuolo, and
S. Guido. Microfluidic interactions between red blood cells and drug carriers by image analysis techniques.
Medical Engineering and Physics, 38:17–23, 2015.
[17] R. D'Apolito, G. Tomaiuolo, F. Taraballi, S. Minardi, D. Kirui, X. Liu, A. Cevenini, R. Palomba, M. Ferrari,
F. Salvatore, E. Tasciotti, and S. Guido. Red blood cells affect the margination of microparticles in synthetic
microcapillaries and intravital microcirculation as a function of their size and shape. Journal of Controlled
Release, 217(C):263–272, 2015.
[18] B. Dünweg and A. J. Ladd. Advanced Computer Simulation Approaches for Soft Matter Sciences III, chapter
Lattice Boltzmann Simulations of Soft Matter Systems, pages 89–166. Springer, 2008.
[19] E. C. Eckstein, A. W. Tilles, and F. J. M. III. Conditions for the occurrence of large near-wall excesses of
small particles during blood flow. Microvascular Research, 36(1):31 – 39, 1988.
[20] M. Faivre, M. Abkarian, K. Bickraj, and H. A. Stone. Geometrical focusing of cells in a microfluidic device:
an approach to separate blood plasma. Biorheology, 43(2):147–159, 2006.
[21] M. E. Fay, D. R. Myers, A. Kumar, C. T. Turbyfield, R. Byler, K. Crawford, R. G. Mannino, A. Laohapant,
E. A. Tyburski, Y. Sakurai, M. J. Rosenbluth, N. A. Switz, T. A. Sulchek, M. D. Graham, and W. A.
Lam. Cellular softening mediates leukocyte demargination and trafficking, thereby increasing clinical blood
counts. Proceedings of the National Academy of Sciences, 113(8):1987–1992, 2016.
[22] D. A. Fedosov and G. Gompper. White blood cell margination in microcirculation. Soft Matter, 10:2961–
2970, 2014.
[23] D. A. Fedosov, B. Caswell, A. S. Popel, and G. E. Karniadakis. Blood flow and cell-free layer in microves-
sels. Microcirculation, 17:615–628, 2010.
[24] D. A. Fedosov, J. Fornleitner, and G. Gompper. Margination of white blood cells in microcapillary flow.
Phys. Rev. Lett., 108:028104, 2012.
[25] D. A. Fedosov, M. Peltomäki, and G. Gompper. Deformation and dynamics of red blood cells in flow
through cylindrical microchannels. Soft matter, 10(24):4258–4267, 2014.
[26] S. Fitzgibbon, A. P. Spann, Q. M. Qi, and E. S. G. Shaqfeh. In Vitro Measurement of Particle Margination
in the Microchannel Flow: Effect of Varying Hematocrit. Biophys J, 108(10):2601–2608, 2015.
[27] J. B. Freund. Leukocyte margination in a model microvessel. Physics of Fluids, 19(2), 2007.
[28] J. B. Freund. Numerical simulation of flowing blood cells. Annual Review of Fluid Mechanics, 46(1):67–95,
2014.
[29] J. B. Freund and M. M. Orescanin. Cellular flow in a small blood vessel. J Fluid Mech, 671:466–490, 2011.
[30] J. B. Freund and B. Shapiro. Transport of particles by magnetic forces and cellular blood flow in a model
microvessel. Phys. Fluids, 24(5):051904–12, 2012.
[31] H. Fujiwara, T. Ishikawa, R. Lima, N. Matsuki, Y. Imai, H. Kaji, M. Nishizawa, and T. Yamaguchi. Red
blood cell motions in high-hematocrit blood flowing through a stenosed microchannel. J Biomechanics, 42
(7):838–843, 2009.
[32] Y.-C. Fung. Stochastic flow in capillary blood vessels. Microvascular research, 5(1):34–48, 1973.
[33] S. Gekle. Strongly accelerated margination of active particles in blood flow. Biophysical Journal, 110(2):
514 – 520, 2016.
[34] H. Goldsmith and J. Marlow. Flow behavior of erythrocytes. ii. particle motions in concentrated suspensions
of ghost cells. Journal of Colloid and Interface Science, 71(2):383–407, 1979.
17
[35] G. Gompper and D. A. Fedosov. Modeling microcirculatory blood flow: current state and future perspec-
tives. WIREs Syst Biol Med, 8(2):157–168, 2015.
[36] X. Grandchamp, G. Coupier, A. Srivastav, C. Minetti, and T. Podgorski. Lift and Down-Gradient Shear-
Induced Diffusion in Red Blood Cell Suspensions. Phys. Rev. Lett., 110(10):108101, 2013.
[37] A. Guckenberger and S. Gekle. Theory and algorithms to compute Helfrich bending forces: A review.
Journal of Physics: Condensed Matter, 29(20):203001, 2017.
[38] A. Guckenberger, M. P. Schraml, P. G. Chen, M. Leonetti, and S. Gekle. On the bending algorithms for soft
objects in flows. Computer Physics Communications, 207:1–23, 2016.
[39] A. Guckenberger, A. Kihm, T. John, C. Wagner, and S. Gekle. Numerical-experimental observation of shape
bistability of red blood cells flowing in a microchannel. Soft Matter, 14:2032–2043, 2018.
[40] R. G. Henríquez Rivera, K. Sinha, and M. D. Graham. Margination regimes and drainage transition in
confined multicomponent suspensions. Phys. Rev. Lett., 114:188101, 2015.
[41] R. G. Henríquez Rivera, X. Zhang, and M. D. Graham. Mechanistic theory of margination and flow-induced
segregation in confined multicomponent suspensions: Simple shear and Poiseuille flows. Physical Review
Fluids, 1(6), 2016.
[42] A. Jain and L. L. Munn. Determinants of leukocyte margination in rectangular microchannels. PLoS One,
4(9):e7104, 2009.
[43] D. Katanov, G. Gompper, and D. A. Fedosov. Microvascular blood flow resistance: Role of red blood cell
migration and dispersion. Microvasc. Res., 99(C):57–66, 2015.
[44] T. Krüger. Effect of tube diameter and capillary number on platelet margination and near-wall dynamics.
Rheologica Acta, 55(6):511–526, 2016.
[45] A. Kumar and M. D. Graham. Segregation by membrane rigidity in flowing binary suspensions of elastic
capsules. Phys. Rev. E, 84:066316, 2011.
[46] A. Kumar and M. D. Graham. Mechanism of margination in confined flows of blood and other multicom-
ponent suspensions. Phys. Rev. Lett., 109:108102, 2012.
[47] A. Kumar and M. D. Graham. Margination and segregation in confined flows of blood and other multicom-
ponent suspensions. Soft Matter, 8:10536–10548, 2012.
[48] A. Kumar, R. G. H. Rivera, and M. D. Graham. Flow-induced segregation in confined multicomponent
suspensions: effects of particle size and rigidity. Journal of Fluid Mechanics, 738:423–462, 2014.
[49] M. Laschke and M. Menger. The dorsal skinfold chamber: a versatile tool for preclinical research in tissue
engineering and regenerative medicine. Eur Cell Mater, 32:202–215, 2016.
[50] M. Laschke, B. Vollmar, and M. Menger. The dorsal skinfold chamber: window into the dynamic interaction
of biomaterials with their surrounding host tissue. Eur Cell Mater, 22(147):e64, 2011.
[51] V. Leble, R. Lima, R. Dias, C. Fernandes, T. Ishikawa, Y. Imai, and T. Yamaguchi. Asymmetry of red blood
cell motions in a microchannel with a diverging and converging bifurcation. Biomicrofluidics, 5(4):044120,
2011.
[52] T.-R. Lee, M. Choi, A. M. Kopacz, S.-H. Yun, W. K. Liu, and P. Decuzzi. On the near-wall accumulation
of injectable particles in the microcirculation: smaller is not better. Sci. Rep., 3:2079, 2013.
[53] X. Li, A. S. Popel, and G. E. Karniadakis. Blood–plasma separation in y-shaped bifurcating microfluidic
channels: a dissipative particle dynamics simulation study. Physical biology, 9(2):026010, 2012.
18
[54] H. Limbach, A. Arnold, B. Mann, and C. Holm. Espresso-an extensible simulation package for research
on soft matter systems. Computer Physics Communications, 174(9):704 – 727, 2006.
[55] K. Lykov, X. Li, H. Lei, I. V. Pivkin, and G. E. Karniadakis.
Inflow/Outflow Boundary Conditions for
Particle-Based Blood Flow Simulations: Application to Arterial Bifurcations and Trees. PLOS Computa-
tional Biology, 11(8):e1004410, 2015.
[56] M. Mehrabadi, D. N. Ku, and C. K. Aidun. Effects of s hear rate, confinement, and particle parameters on
margination in blood flow. Phys. Rev. E, 93:023109, 2016.
[57] C. Migliorini, Y. Qian, H. Chen, E. B. Brown, R. K. Jain, and L. L. Munn. Red blood cells augment
leukocyte rolling in a virtual blood vessel. Biophysical Journal, 83(4):1834 – 1841, 2002.
[58] C. Misbah and C. Wagner. Living fluids. Comptes Rendus Physique, 14(6):447 – 450, 2013. Living fluids /
Fluides vivants.
[59] R. Mittal and G. Iaccarino. Immersed boundary methods. Annual Review of Fluid Mechanics, 37(1):239–
261, 2005.
[60] L. Mountrakis, E. Lorenz, and A. G. Hoekstra. Where do the platelets go? a simulation study of fully
resolved blood flow through aneurysmal vessels. Interface Focus, 3(2), 2013.
[61] K. Müller, D. A. Fedosov, and G. Gompper. Margination of micro-and nano-particles in blood flow and its
effect on drug delivery. Scientific reports, 4:4871, 2014.
[62] K. Müller, D. A. Fedosov, and G. Gompper. Understanding particle margination in blood flow: A step
toward optimized drug delivery systems. Medical Engineering & Physics, 38(1):2 – 10, 2016. Micro and
Nano Flows 2014 (MNF2014) - Biomedical Stream.
[63] K. Namdee, A. J. Thompson, P. Charoenphol, and O. Eniola-Adefeso. Margination Propensity of Vascular-
Targeted Spheres from Blood Flow in a Microfluidic Model of Human Microvessels. Langmuir, 29(8):
2530–2535, 2013.
[64] B. Namgung, Y. C. Ng, H. L. Leo, J. M. Rifkind, and S. Kim. Near-wall migration dynamics of erythrocytes
in vivo: Effects of cell deformability and arteriolar bifurcation. Frontiers in Physiology, 8:963, 2017.
[65] Y. C. Ng, B. Namgung, S. L. Tien, H. L. Leo, and S. Kim. Symmetry recovery of cell-free layer after
bifurcations of small arterioles in reduced flow conditions: effect of rbc aggregation. American Journal of
Physiology-Heart and Circulatory Physiology, 311(2):H487–H497, 2016.
[66] C. S. Peskin. The immersed boundary method. Acta numerica, 11:479–517, 2002.
[67] A. S. Popel and P. C. Johnson. Microcirculation and hemorheology. Annual review of fluid mechanics, 37:
43, 2005.
[68] A. Pries, K. Ley, M. Claassen, and P. Gaehtgens. Red cell distribution at microvascular bifurcations. Mi-
crovascular research, 38(1):81–101, 1989.
[69] A. Pries, T. W. Secomb, and P. Gaehtgens. Biophysical aspects of blood flow in the microvasculature.
Cardiovascular research, 32(4):654–667, 1996.
[70] A. R. Pries and T. W. Secomb. Chapter 1 Blood Flow in Microvascular Networks. In Microcirculation,
pages 3–36. Elsevier, Jan. 2008.
[71] Q. M. Qi and E. S. Shaqfeh. Time-dependent particle migration and margination in the pressure-driven
channel flow of blood. Physical Review Fluids, 3(3):034302, 2018.
[72] Q. M. Qi and E. S. G. Shaqfeh. Theory to predict particle migration and margination in the pressure-driven
channel flow of blood. Physical Review Fluids, 2(9), 2017.
19
[73] D. A. Reasor, M. Mehrabadi, D. N. Ku, and C. K. Aidun. Determination of critical parameters in platelet
margination. Annals of Biomedical Engineering, 41(2):238–249, 2013.
[74] D. Roehm and A. Arnold. Lattice boltzmann simulations on gpus with espresso. The European Physical
Journal Special Topics, 210(1):89–100, 2012.
[75] G. W. Schmid-Schönbein, S. Usami, R. Skalak, and S. Chien. The interaction of leukocytes and erythrocytes
in capillary and postcapillary vessels. Microvascular Research, 19(1):45 – 70, 1980.
[76] T. W. Secomb. Blood Flow in the Microcirculation. Annual Review of Fluid Mechanics, 49(1):443–461,
2017.
[77] Z. Shen, G. Coupier, B. Kaoui, B. Polack, J. Harting, C. Misbah, and T. Podgorski. Inversion of hematocrit
partition at microfluidic bifurcations. Microvascular research, 105:40–46, 2016.
[78] T. Skorczewski, L. C. Erickson, and A. L. Fogelson. Platelet motion near a vessel wall or thrombus surface
in two-dimensional whole blood simulations. Biophysical journal, 104(8):1764–1772, 2013.
[79] A. P. Spann, J. E. Campbell, S. R. Fitzgibbon, A. Rodriguez, A. P. Cap, L. H. Blackbourne, and E. S. G.
Shaqfeh. The Effect of Hematocrit on Platelet Adhesion: Experiments and Simulations. Biophys J, 111(3):
577–588, 2016.
[80] S. Succi. The lattice Boltzmann equation: for fluid dynamics and beyond. Oxford university press, 2001.
[81] C. Sun and L. L. Munn. Lattice-boltzmann simulation of blood flow in digitized vessel networks. Computers
& Mathematics with Applications, 55(7):1594–1600, 2008.
[82] K. Svanes and B. Zweifach. Variations in small blood vessel hematocrits produced in hypothermic rats by
micro-occlusion. Microvascular Research, 1(2):210–220, 1968.
[83] J. Tan, A. Thomas, and Y. Liu. Influence of red blood cells on nanoparticle targeted delivery in microcircu-
lation. Soft matter, 8(6):1934–1946, 2012.
[84] G. Tangelder, H. C. Teirlinck, D. W. Slaaf, and R. S. Reneman. Distribution of blood platelets flowing in
arterioles. American Journal of Physiology-Heart and Circulatory Physiology, 248(3):H318–H323, 1985.
[85] A. Tokarev, A. Butylin, E. Ermakova, E. Shnol, G. Panasenko, and F. Ataullakhanov. Finite platelet size
could be responsible for platelet margination effect. Biophysical Journal, 101(8):1835 – 1843, 2011.
[86] S. Tripathi, Y. B. V. Kumar, A. Prabhakar, S. S. Joshi, and A. Agrawal. Passive blood plasma separa-
tion at the microscale: a review of design principles and microdevices. Journal of Micromechanics and
Microengineering, 25(8):083001, 2015.
[87] V. Turitto and H. Baumgartner. Platelet deposition on subendothelium exposed to flowing blood: mathe-
matical analysis of physical parameters. ASAIO Journal, 21(1):593–601, 1975.
[88] K. Vahidkhah and P. Bagchi. Microparticle shape effects on margination, near-wall dynamics and adhesion
in a three-dimensional simulation of red blood cell suspension. Soft Matter, 11(11):2097–2109, 2015.
[89] K. Vahidkhah, S. L. Diamond, and P. Bagchi. Platelet Dynamics in Three-Dimensional Simulation of Whole
Blood. Biophys J, 106(11):2529–2540, 2014.
[90] K. Vahidkhah, P. Balogh, and P. Bagchi. Flow of red blood cells in stenosed microvessels. Scientific Reports,
6:28194, 2016.
[91] G. Wang, W. Mao, R. Byler, K. Patel, C. Henegar, A. Alexeev, and T. Sulchek. Stiffness dependent separa-
tion of cells in a microfluidic device. PLoS One, 8(10):e75901, 2013.
20
[92] T. Wang, U. Rongin, and Z. Xing. A micro-scale simulation of red blood cell passage through symmetric
and asymmetric bifurcated vessels. Scientific reports, 6, 2016.
[93] W. Wang, T. G. Diacovo, J. Chen, J. B. Freund, and M. R. King. Simulation of Platelet, Thrombus and
Erythrocyte Hydrodynamic Interactions in a 3D Arteriole with In Vivo Comparison. PLoS ONE, 8(10):
e76949–11, 2013.
[94] B. Woldhuis, G. Tangelder, D. W. Slaaf, and R. S. Reneman. Concentration profile of blood platelets
differs in arterioles and venules. American Journal of Physiology-Heart and Circulatory Physiology, 262
(4):H1217–H1223, 1992.
[95] W.-T. Wu, Y. Li, N. Aubry, M. Massoudi, and J. F. Antaki. Numerical simulation of red blood cell-induced
platelet transport in saccular aneurysms. Applied Sciences, 7(5):484, 2017.
[96] X. Yang, O. Forouzan, J. M. Burns, and S. S. Shevkoplyas. Traffic of leukocytes in microfluidic channels
with rectangular and rounded cross-sections. Lab on a Chip, 11(19):3231–3240, 2011.
[97] A. Yazdani and G. E. Karniadakis. Sub-cellular modeling of platelet transport in blood flow through mi-
crochannels with constriction. Soft Matter, 12:4339–4351, 2016.
[98] H. Zhao and E. S. G. Shaqfeh. Shear-induced platelet margination in a microchannel. Phys. Rev. E, 83:
061924, 2011.
[99] H. Zhao, E. S. G. Shaqfeh, and V. Narsimhan. Shear-induced particle migration and margination in a cellular
suspension. Physics of Fluids, 24(1), 2012.
[100] R. Zhao, J. Marhefka, F. Shu, S. Hund, M. Kameneva, and J. Antaki. Micro-flow visualization of red
blood cell-enhanced platelet concentration at sudden expansion. Annals of Biomedical Engineering, 36(7):
1130–41, 2008.
21
List of Figures
6
5
4
3
2
1
a) Systems of interest: a suspension of red blood cells and microparticles flowing either through
a confluence (left) or a bifurcation (right). Rectangles with numbers refer to figures containing
corresponding 2D radial/planar projections while dashed lines refer to figures containing cross-
sectional profiles. b) Inflow using straight cylinders as feeding systems: whenever a cell/particle
crosses the indicated plane, it is fed into one of the branches of the confluence system.
. . . . .
a) Color labeling and b) labeling criteria for red blood cells and microparticles at the entrance of
the system with respect to the in-plane position in the cross-section. Due to symmetry the particles
left and right can be treated equally.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Concentration of red blood cells in a confluence in 2D planar projection a) along the upper branch
and b) along the main channel. Microparticle concentration in 2D radial projection c) along the
upper branch and d) along the main channel. The cell-free layer near the inner boundary decreases
at the end of the branches while it increases near the outer boundary. Inside the main channel an
additional cell-free layer in the center develops.
. . . . . . . . . . . . . . . . . . . . . . . . . .
2D cross-sectional concentration in a confluence for red blood cells a)-d) and microparticles e)-h)
at the end of the branches a),e), at the beginning of the main channel b),f), at the middle of the
main channel c),g), and at the end of the main channel d),h). Positions are also indicated by the
dashed black lines in figure 1 (a). In the main channel a clear cell-free layer in the center together
with microparticle anti-margination is present.
. . . . . . . . . . . . . . . . . . . . . . . . . .
1D axial profile of red blood cells and microparticles flowing through a vessel confluence. While
the red blood cells are depleted at the site of the confluence, the microparticles exhibit a concen-
tration increase of about 50% compared to the branches and the main channel.
. . . . . . . . .
Axial concentration of a) red blood cells and b) microparticles distinguished regarding their posi-
tion inside the cross-section of the branches as illustrated in figure 2. The microparticles entering at
the top of the upper branch (or, equivalently, the bottom of the lower branch) exhibit a pronounced
peak.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
In vivo measurement of tracked fluorescent beads in mouce microvessels. Blue lines show trajecto-
ries of beads which remain marginated after the bifurcation while red lines show beads undergoing
anti-margination. Yellow dashed lines denote the vessel boundaries.
. . . . . . . . . . . . . . .
2D planar projection within the bifurcation for red blood cells a) along the main channel and b)
along the upper branch. 2D radial projection of microparticle concentration c) along the main
channel and d) along the branches. The cell-free layer decreases at the end of the main channel
and we observe an asymmetric cell distribution inside the branches. The margination peak of
microparticles is somewhat blurred after the bifurcation, but otherwise unaffected.
. . . . . . .
2D cross-sectional concentration of red blood cells a),b),c) and microparticles d),e),f) at the end of
the main channel a),d), at the beginning of the branches b),e), and at the end of the branches c),f).
See figure 1 (a) for indications of the respective positions along the channel. Inside the branches
an asymmetric cell-free layer develops and microparticles suffer a loss of concentration directly
besides the inner wall.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10 Axial concentration of red blood cells and microparticles flowing through a bifurcating channel.
Both red blood cells and microparticles exhibit a peak in front of the bifurcation apex. The mi-
croparticle concentration increases directly in front of the apex, while the red blood cell concen-
tration exhibits a second small peak due to a second cell flowing onto a cell already being stuck at
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
the apex. .
11 Axial concentration of a) red blood cells and b) microparticles distinguished regarding their initial
position inside the cross-section of the main channel. The peak in red blood cell concentration
stems from the cells trapped at the apex of the bifurcation. The microparticles exhibit a similar
peak when arriving left/right. However, for microparticles entering top/bottom no peak occurs.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
.
.
.
8
.
.
.
9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
5
6
7
9
9
10
10
11
12
12
22
12
13
a) 2D planar projection within the confluence for red blood cells along the main channel behind a
confluence as in figure 3, but with larger hematocrit Ht = 20%. b)-d) 2D cross-section profiles
for b) red blood cells and c),d) microparticles in the main channel as in figure 4, but for larger
hematocrit Ht = 20%.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1D axial concentration of a),c) red blood cells and b),d) and microparticles flowing through a
a),b) confluence or c),d) bifurcation with larger hematocrit Ht = 20% labeled by their radial
position at the entrance. The larger hematocrit hardly affects the behavior of red blood cells and
microparticles.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
14
23
List of Tables
1
Concentration of red blood cells and microparticles in the upper (up) and lower (low) channel of
an asymmetric bifurcation with radius 11µm of the upper branch and Rlow of the lower branch.
Concentrations are given in 103µm−3.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
14
24
SupplementalInformation-Anti-marginationofmicroparticlesandplateletsinthevicinityofbranchingvesselsC.Bächer,A.Kihm,L.Schrack,L.Kaestner,M.W.Laschke,C.Wagner,andS.GekleInflow:fullmarginationandconstantparticlenumberInfigureS1weshowthestateofcompletemarginationattheentranceofabifurcatingchannelandaconfluence,respectively.Thecross-sectionalconcentrationsshowtheredbloodcellsaccumulatedaroundthechannelcenterandthemicroparticlesclosetothewall.FigureS2showsthatafterashorttransienttimetheimplementedparticleinflow/outflowleadstoaconstantnumberofredbloodcellsandmicroparticlesinthesystem.a)x=9µm-15-10-5 0 5 10 15z [µm]-15-10-5 0 5 10 15y [µm] 0 1 2 3 4 5 6 7103 · ρRBC [µm-3]b)x=9µm-15-10-5 0 5 10 15z [µm]-15-10-5 0 5 10 15y [µm] 0 0.5 1 1.5 2 2.5 3103 · ρmicro [µm-3]c)x=9µm-15-10-5 0 5 10 15z [µm]-15-10-5 0 5 10 15y [µm] 0 1 2 3 4 5 6 7103 · ρRBC [µm-3]d)x=9µm-15-10-5 0 5 10 15z [µm]-15-10-5 0 5 10 15y [µm] 0 0.5 1 1.5 2 2.5 3103 · ρmicro [µm-3]Fig.S1:Attheentranceoftheconfluencesystema),b)andthebifurcationsystemc),d)wehaveastateoffullmargination:theredbloodcellsa),c)arelocatedinthechannelcenter,themicroparticlesb),d)nearthewall.ConcentrationprofilesoftracerparticlesInordertomodelthebehaviorofredbloodcellsandmicroparticlesweusepassivepointparticlesastracersflowingwiththeintrinsicvelocityprofile[1].Wefirststartwiththeconfluence.Homogeneouslydistributedparticlesexhibitasimilarconcentrationprofileastheredbloodcells.Asaconsequence,redbloodcellbehaviorcanbeexplainedbytheintrinsicvelocityprofile.Doingthesamecalculationsfortracersinthedifferentregionsresemblethelabeledredbloodcells.Italsofitstheconcentrationprofilesforthemicroparticles.Inthedivergingbifurcationstartingattoptheconcentrationprofileofthepointparticlesmatchesthatofthemicroparticlesquitewell.Alsothepointparticleslocatedrightreproducemicroparticlebehavior.Wenotethat1 0 2 4 6 8 10 12 14 16 18 20 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 0 20 40 60 80 100 120 140 160 180NtotmicroNtotRBCtime steps / 2500microRBCFig.S2:TotalnumberofRBCsandmicroparticleswithintheconfluencesystemdependingonthesimulationtime.Bothnumbersfluctuatearoundaconstantvalueafterinitialfillingofthesystem.a)1020304050607080ρpoint [a.u.]x [µm]homogeneousb)0102030405060708090ρpoint [a.u.]x [µm]centerrightbottomtopc)0102030405060708090ρpoint [a.u.]x [µm]centerrighttopFig.S3:Concentrationofa)homogeneouslydistributedtracerparticlesandb)tracerparticlesflowinginthedistinctregionsatsystemen-trancewithintheconfluence.c)Tracerparticlesflowinginthedistinctregionsatsystementrancewithinthedivergingbifurcation.Thesefiguresarecomparedtocellandparticleconcentrationinfigure6,11and13inthemaintext.startingpointparticlestoporbottomandleftorrightresultsinthesameconcentrationduetosymmetry.Redbloodcellbehaviorisalsosimilartothatofpointparticles,exceptthepeakatthebifurcationapex.Thedifferencesareeffectsduetothefinitesizeanddeformabilityofredbloodcells.2 Shear-induceddiffusiona) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035MSD in µm2t in sMSD2*D*t with D = 27.88b) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018MSD in µm2t in sMSD2*D*t with D = 25.03Fig.S4:Meansquaredisplacementovertimefora)redbloodcellsandb)microparticleslocatednearthecenterbehindaconfluence.Bymodelingthetheoreticalexpectationwecanextractashear-induceddiffusioncoefficient.Largerhematocrita) 0 5 10 15 20 25 30 35x [µm]-15-10-5 0 5 10 15y [µm] 0 1 2 3 4 5 6 7103 · ρRBC [µm-3]b) 55 60 65 70 75 80 85 90x [µm]-10-5 0 5 10y [µm] 0 1 2 3 4 5 6 7103 · ρRBC [µm-3]Fig.S5:2Dplanarprojectionforredbloodcellsalongthea)mainchannelandb)branchesofthebifurcationwithlargerhematocritHt=20%.Thecell-freelayerdecreases,butthebehaviorisqualitativelyunchanged.Anti-marginationofplatelet-shapedmicroparticlesInthemaintextwefocusonsphericalmicroparticles.Here,weshowadditionalresultsforoblatespheroids–ageometrythatmimicksmorecloselythatofrealplatelets.Theplateletshaveadiameterof3.9µmalongthetwolongaxesand2.3µmalongthesmallaxisandareillustratedinfigureS6.Similartothesphericalparticlesofthemaintext,about14%ofthespheroidalmicroparticlesareanti-marginateddirectlybehindtheconfluence.ValidationoftheusedIBM-LBMalgorithmInthefollowing,wesummarizeandextendthevalidationforourimmersedboundarymethod(IBM)andLattice-Boltzmannmethod(LBM).Inref.[3]thecalculationofshearandbendingforceshasbeenvalidatedforacapsuleinshearflow.Inref.[2]thehematocritprofilefortubeflowandplane-Poiseuilleflowhasbeenshowntoagreewithprevious,establishedstudies.Furthermore,thestabilityofthestiffsphericalparticlesusedhasbeendemonstratedandtheflowprofilepastaspherehasbeencomparedfavorablytotheanalyticalsolution.Inaddition,weherecalculateheretheStokesdrag1/(6πηa)thatrelatestheforceonasphereofradiusatoitsvelocityinasuspendingfluidofviscosityηforaspherewithtwodifferentparticleresolutionsinfigureS7.Weperformedsimulationswiththeresolutionusedinthemaintext(81nodesoftheinnerstiffgrid)andan3 a)b)x=55µm-15-10-5 0 5 10 15z [µm]-15-10-5 0 5 10 15y [µm] 0 0.5 1 1.5 2 2.5 3103 · ρmicro [µm-3]Fig.S6:(a)Simulationofmicroparticleswithaplatelet-likeoblateshapeflowingthroughaconfluence.(b)Similartothesphericalmi-croparticlesofthemaintext,theseparticlesalsoundergoanti-margination. 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 20 40 60 80 100 120 140 160velocity in mm/sforce in pN81 nodes485 nodestheoritcal predictionFig.S7:Velocityofasphericalparticlepulledthroughafluidwithagivenforce.SimulationswithtwodifferentparticleresolutionsarecomparedtothetheoreticalpredictiongivenbytheStokesdrag.increasedresolution(485nodesoftheinnergrid).Bothshowgoodagreementwiththetheoreticalpredictionandconvergencetothetheoryforincreasinggridresolution.Inordertoprovemeshinsensitivitywereducedtheresolutionofthemembranemeshoftheredbloodcellsto258nodesand512trianglesandtheresolutionofthestiffparticlesto66nodesand128triangles.Inthesamewaythefluidmeshchangesfrom288x110x58to200x82x42.Theredbloodcelldistributionbehindaconfluenceandthecross-sectionalmicroparticleconcentrationarecomparedtotheresultsofthemanuscriptinfigureS8.Theresultsareinverygoodagreement.Smalldiscrepanciesmaybecausedbyslightlydifferentinflowconcentrations.4 a)b)c)d)Fig.S8:Themainresultsfordifferentmeshresolutions:Thecellfree-layerina),b)andthemicroparticleanti-marginationinc),d).a),c)resolutionusedinthemaintextandb),d)decreasedresolution(RBC:512,microparticle:128,fluid:200x82x42).Bothresolutionsleadtosimilarresultsforcentralcellfreelayerstabilityandfractionofanti-marginatedmicroparticles(c)15.8%andd)16.2%).Figuresa)andc)arefromthemaintext.5 References[1]C.Bächer,L.Schrack,andS.Gekle.Clusteringofmicroscopicparticlesinconstrictedbloodflow.PhysicalReviewFluids,2(1):013102,2017.[2]S.Gekle.Stronglyacceleratedmarginationofactiveparticlesinbloodflow.BiophysicalJournal,110(2):514–520,2016.[3]A.Guckenberger,M.P.Schraml,P.G.Chen,M.Leonetti,andS.Gekle.Onthebendingalgorithmsforsoftobjectsinflows.ComputerPhysicsCommunications,207:1–23,2016.6 |
1802.04882 | 1 | 1802 | 2018-02-13T22:30:40 | Nanocommunication via FRET with DyLight Dyes using Multiple Donors and Acceptors | [
"physics.bio-ph",
"q-bio.MN"
] | The phenomenon of Forster Resonance Energy Transfer, commonly used to measure the distances between fluorophore molecules and to study interactions between fluorescent-tagged proteins in life sciences, can also be applied in nanocommunication networks to transfer information bits. The mechanism offers a relatively large throughput and very small delays, but at the same time the channel bit error rate is too high and the transmission ranges are too limited for communication purposes. In this paper, multiple donors at the transmitter side and multiple acceptors at the receiver side are considered to decrease the bit error rate. As nanoantennas, the DyLight fluorescent dyes, which are very well suited to long range nanocommunication due to their large Forster distances and high degrees of labeling, are proposed. The reported results of the recent laboratory experiments confirm efficient communication on distances over 10 nm. | physics.bio-ph | physics | Nanocommunication via FRET with DyLight
Dyes using Multiple Donors and Acceptors
Kamil Solarczyk, Krzysztof Wojcik, and Pawel Kulakowski
Abstract-The phenomenon of Förster Resonance Energy
Transfer, commonly used to measure the distances between
fluorophore molecules and to study
interactions between
fluorescent-tagged proteins in life sciences, can also be applied in
nanocommunication networks to transfer information bits. The
mechanism offers a relatively large throughput and very small
delays, but at the same time the channel bit error rate is too high
and the transmission ranges are too limited for communication
purposes. In this paper, multiple donors at the transmitter side
and multiple acceptors at the receiver side are considered to
decrease the bit error rate. As nanoantennas, the DyLight
fluorescent dyes, which are very well suited to long range
nanocommunication due to their large Förster distances and high
degrees of labeling, are proposed. The reported results of the
recent laboratory experiments confirm efficient communication
on distances over 10 nm.
Index Terms-Communication channels, FRET, MIMO,
DyLight dyes, molecular communication, nanocommunication.
I. INTRODUCTION
T
HE current development of nanomachines and nanorobots
and organizing them into systems is stimulated by
incredible applications promised by them in industrial
manufacturing, biology and, especially, medicine. The
progress in the latter field may be revolutionized with swarms
of tiny robots employed in targeted drug delivery, in vivo
imaging and diagnostics, tissues regeneration and engineering
[1, 2]. The future success of nanomachine systems depends,
however, on overcoming two challenges: the first one is
fabrication of nanodevices with proper precision, lifetime and
efficiency [3]. The second challenge, being the subject of this
paper, is the efficient communication between the nanorobots:
they obviously need to signal their actions and communicate
with each other.
The manuscript was submitted September 30, 2015. The work was
performed under Contract 11.11.230.018 and also funded by the National
Science Centre based on the decision number DEC-2013/11/N/NZ6/02003.
The confocal microscope was purchased through an EU structural funds grant
BMZ no. POIG.02.01.00-12-064/08.
K. Wojcik and K. Solarczyk are with Division of Cell Biophysics, Faculty
of Biochemistry, Biophysics and Biotechnology, Jagiellonian University, 7,
Gronostajowa
(e-mails:
[email protected], [email protected]). K. Wojcik is also with
2nd Department of Medicine, Jagiellonian University Medical College,
Kraków, Poland.
Kraków,
30-387
Poland
St.,
P. Kulakowski
is with Department of Telecommunications, AGH
University of Science and Technology, Al. Mickiewicza 30, 30-059 Kraków,
Poland (e-mail: [email protected]).
common
considered
mechanisms
The
in
nanocommunication literature are: calcium ion signaling [4-5],
flagellated bacteria carrying data in its DNA [6], the
movement of molecular motors [7-8], pheromones, pollen and
spores [9]. They are, however, based on mechanical
phenomena and thus their propagation delays and achievable
data throughputs are not satisfactory for communication
purposes. The delay issues were the reason to propose acoustic
communication techniques for nanoscale [10]. Here, we
consider even more rapid phenomena, namely Förster
Resonance Energy Transfer (FRET).
FRET has been already proposed as an efficient means for
nanoscale communication [9,11-12]. It is a non-radiative
process of transferring energy from an excited fluorophore
molecule called the donor to an adjacent molecule called the
acceptor, being in the ground state. Popular molecules that can
be donors and acceptors in a FRET process are: fluorescent
dyes, proteins and quantum dots. Fluorescent dyes, which are
quite small structures of about few nanometers, can serve as
nanoantennas being attached to larger nanomachines and
performing numerous tasks, e.g. transporting cargo (kynesins,
dyneins) or seeking other molecules (antibodies), see Fig. 1.
FRET is characterized by a very small delay, about 10-20
nanoseconds, and potentially a very high throughput of tens of
Mbit/s. The main drawbacks of the FRET mechanism are a
very high channel bit error rate and an effective transmission
range limited to a few nanometers. It has been also already
shown that the channel bit error rate can be decreased when
multiple donors at the transmitter side and multiple acceptors
at the receiver side are applied to create so called MIMO
(multiple-input multiple-output) FRET channel [13].
The limited transmission range, however, remains an
Achilles heel of the FRET-based communication technique.
The FRET efficiency decreases with sixth power of the donor-
acceptor separation (see Eq. 1 in the following section), what
means that increasing the transmission distance is much harder
than in wireless communication. In this paper, we report
experiments on DyLight dyes, which is quite a new family of
fluorophores
future
nanocommunication applications. We also consider multiple
donors and acceptors here, constructing 4 different
nanonetworks based on DyLight dyes. On the basis of a
laboratory measurements campaign, we report successful
communication over 12-13 nm, which is a 50% increase of
transmission range comparing with previous experiments. We
expect it will be crucial for the efficient cooperation of future
nanomachines. Many nanomachines, e.g. some antibodies,
kynesins, dyneins, are of size close to 10 nm (see Fig. 1), so it
suited
for
much
better
2
it can also be applied
not a rule, as it depends on the spectra shapes [14]. If the
excitation energy is not transferred via FRET, it can be also
emitted as a photon or dissipated. These two last phenomena
should be interpreted as channel losses increasing the error
rate.
The FRET phenomenon is well known in life sciences and
is commonly used to measure distances between fluorophores
(organic molecules whose emission spectra cover the range of
visible light) or to study interactions between fluorescent-
tagged proteins. However,
to
nanocommunications. Let's imagine a molecular structure
performing the function of a molecular transmitter. Let's have
a donor molecule attached to it, as a molecular transmit
antenna. On the other side, there could be another structure: a
molecular receiver with an acceptor attached (a molecular
receive antenna). Now, information can be sent from the Tx to
the Rx side via the FRET channel. This can be simply
achieved by exciting the donor when bit '1' is to be sent and
by keeping the donor in the ground state when transmitting bit
'0'. The FRET delay is relatively small, in the order of
nanoseconds. If FRET was efficient in 100% of cases, the
associated data throughput (calculated as the inverse of the
delay) could then be as high as tens or even hundreds of
Mbit/s. However, FRET efficiency is only 50% for the donor
and acceptor separated by R0 and it decreases with the sixth
power of that distance. This results in quite a high channel bit
error rate (BER): transmitting '0' is always successful, but
transmitting '1' is erroneous with the probability of 1-E. Thus
the FRET channel BER is equal to:
)
(
−⋅
15.0
BER
(2)
=
6
⋅
r
=
E
.
5.0
6
+
r
6
R
0
Consequently, in order to obtain BER=0.1%, which is a
common value in telecommunication standards, the donor-
acceptor separation should not exceed 0.355R0. This result
limits the efficiency of FRET-based nanocommunication to
distances below 3.5 nm.
FRET efficiency can be increased, if there is not one but
more acceptor molecules in the vicinity of the donor [15].
When an excited donor is surrounded by m equally distant
acceptors, the probability that the excitation energy is
transferred via FRET to one of the acceptors is given by [14]:
E m
,1
=
6
r
6
Rm
0
⋅
+
⋅
Rm
.
(3)
6
0
This means that having a molecular receiver with multiple
acceptor molecules (Rx antennas) can greatly increase FRET
efficiency and decrease the channel BER, which is then equal
to:
BER
m1,
=
6
⋅
5.0
r
⋅+
6
Rmr
0
.
(4)
6
Furthermore, we can also use molecular transmitters with
multiple donors. If there are n donor molecules, we can excite
them all when bit '1' is to be transmitted. Then, (again,
assuming they are equidistant from the acceptors) the
probability that at least one of them transfers its excitation
energy to an acceptor is much higher. Let us assume the FRET
events for all the donors are independent of each other, which
is reasonable, as the FRET efficiency does not depend on the
In particular, the contribution of this paper can be
Fig. 1. Communication is nanoscale between different protein-based
nanomachines: kynesins and dyneins walking over a microtubule and
antibodies. All proteins are labeled with fluorescent dyes (marked as small
circles with rays). The FRET process (marked as a violet zigzag) may
happen between the dyes on different proteins.
would be quite hard for them to approach each other and
communicate on distances of only few nanometers.
summarized as follows:
1. We propose DyLight dyes as efficient candidates for
nanoantennas in MIMO-FRET based nanocommunication.
2. We construct four different nanonetworks based on DyLight
dyes and perform the first nanocommunication experiments
with this fluorescent family.
3. We obtain
increase of
communication range comparing with previously reported
experiments [13].
4. Additionally, we calculate and present bit error rate curves
for the chosen MIMO-FRET channels based on DyLight dyes.
The rest of the paper contains the following sections. In
Section II, the principles of FRET theory are introduced
together with its extension to the case of multiple donors and
multiple acceptors. In Section III, a description is given of
nanonetworks built on antibodies and DyLight dyes. Section
IV contains the laboratory methodology. The results of the
experiments and the calculations are given in Section V.
Finally, in Section VI, conclusions and future directions of
study on the topic of FRET-based nanonetworks are outlined.
results showing a 50%
II. FRET WITH MULTIPLE DONORS AND ACCEPTORS
If we look at the FRET process from a communications
point of view we can think of donors and acceptors as transmit
and receive antennas, respectively. Consequently, the FRET
pair can be considered a wireless communication system. The
probability of the energy transfer is described as FRET
efficiency and is given by [14]:
E
=
R
0
+
6
r
6
6
R
0
.
(1)
In (1), r is the donor-acceptor separation, while R0, called
Förster distance, is the donor-acceptor separation where
E=50%. This separation is specific for each donor-acceptor
pair and can be measured experimentally (usually 3-8 nm). In
general, the Förster distance depends on the overlap between
the donor emission and acceptor absorption spectra as well as
on the quantum yield of the donor. The larger the R0, the better
the match of donor and acceptor spectra and the more efficient
the FRET process. Förster distances are usually larger when
the dyes' spectra are located in higher wavelengths, but this is
3
fact if there are other excited molecules nearby [14]. Now, the
probability that none of the donors transfers its energy to any
of the acceptors can be expressed as the n-th power of such a
probability for a single donor:
1
−
6
r
Thus, the probability that at least one donor succeeds in
transferring its excitation energy to an acceptor is equal to:
6
⋅
Rm
0
⋅
+
Rm
0
6
n
(5)
E
,
mn
−=
1
1
−
6
r
6
⋅
Rm
0
⋅+
Rm
0
n
6
.
(6)
the bit error rate of such a FRET
Consequently,
communication channel is much lower than in the basic case
(with a single donor and a single acceptor) and can be
calculated as:
BER
mn,
=
5.0
⋅
6
r
6
r
+
⋅
Rm
0
6
n
.
(7)
The number of donors/acceptors attached to a molecular
structure is a parameter called the degree of labeling (DoL)
and it is a parameter not easy to control, as will be further
discussed
the area of
communications, this reminds the known idea of MIMO
(multiple-input multiple-output) channels [16, 17]. Therefore,
we will henceforth use the term: MIMO-FRET channels.
following sections.
the
In
in
III. BUILDING NANONETWORKS WITH DYLIGHT DYES
The absorption and emission spectra of the DyLight
fluorescent dye family series cover much of the visible light
and extend into the infrared region (the spectra maxima range
from 350 nm to 794 nm), allowing the dyes detection using
most types of fluorescence and confocal microscopes, as well
as infrared imaging systems. In comparison with other
fluorescent dye families (Alexa Fluor, CyDye, LI-COR), the
main advantages of the DyLight group are as follows [18, 19]:
(a) absorption and emission spectra of a similar shape
(resulting in large Förster distances), (b) high photostability
and brightness, (c) low pH-sensitivity, (d) high dye-to-protein
(degree of labeling) ratio in water solutions to be achieved
without precipitation of conjugates. The most relevant
measurements on FRET between DyLight dyes reported in the
open literature are, until now, focused on describing spectral
properties of DyLight fluorophores [19] and using FRET, also
with multiple acceptors, in disease diagnostics [20].
For the purpose of studying the signal transfer via FRET
between DyLight dyes, we built special molecular structures
based on antibodies, i.e. proteins with the ability to recognize
and bind other molecules (Fig. 2). When thinking about future
applications in nanocommunication, antibodies can be treated
as nanomachines and DyLight dyes as nanoantennas attached
to them. They may diffuse in a cellular membrane and perform
some tasks, e.g. looking for antigens. When two antibodies are
close to each other, the communication process may occur. In
our case, in order to perform FRET measurements the chosen
antibodies were bound to a molecular skeleton consisting of
DNA and histone proteins in the nuclei of fixed HeLa cells
(Fig. 2). It should be emphasized that our choice of building
the network of antibodies in cells was solely technical, i.e.
Fig. 2. Four investigated scenarios: molecular structures prepared for the
purpose of the experiments. Each structure consists of a DNA molecule, a
histone, one or two linker (gray color here) antibodies and finally two more
antibodies with donor/acceptor dyes attached. The antibodies labeled with
DyLight dyes are marked as green/yellow/red for 549/594/649 dyes
respectively.
driven by fact that the nuclei contain a well-known and easily
accessible scaffold for binding of antibodies. We investigated
4 different scenarios: for each of them, we had a molecular
structure built of antibodies. In each scenario, two of the
antibodies were labeled with multiple DyLight dyes: donors
on one side and acceptors on the other side, creating a MIMO-
FRET communication channel. In order to obtain good
communication efficiency and low BER for long transmission
ranges the FRET dyes pairs should fulfill the following
criteria: (a)
the donor emission spectra and acceptor
absorption spectra should match each other well, resulting in
high values of Förster distances and (b) the degree of labeling
of the dyes used should be as high as possible. Thus, we
finally chose two pairs:
• DyLight 549 dyes as donors and DyLight 594 as acceptors,
donor-acceptor R0 = 5.62 nm,
• DyLight 594 as donors and DyLight 649 as acceptors,
donor-acceptor R0 = 8.16 nm.
All the dyes were supplied by Jackson ImmunoResearch
Laboratories, Inc. Their numbers, 549, 594 and 649, indicate
the main wavelengths of their absorption spectra given in
nanometers. Each pair was additionally tested in two spatial
configurations, with
slightly different donor-acceptor
distances and connections of antibodies, resulting in 4
scenarios in total, see Fig. 2. It allowed us to observe the
distance-dependent effects, the influence of the choice of
antibodies and the variability of possible lifetimes (especially
in the scenario C). All the experiments were performed on
fixed and permeabilized cells, in order to allow the antibodies
to easily enter the nucleus and guarantee that all chemical
reactions in the cells were stopped and did not affect the
measurements.
4
The details of the construction of the molecular structures
were as follows. The first element of the structure was always
a histone H1 bound to a DNA molecule. Next, one or two
linker antibodies were attached to the histone H1. Then, an
antibody labeled with donor DyLight dyes was bound to the
linker ones. Finally, the last antibody labeled with acceptor
dyes was attached directly to the one labeled with donors
(scenario B and D) or to the linker antibodies (scenarios A and
C), as shown in Fig. 2. Each antibody was a molecule of
Immunoglobulin G, being like a 3-element airscrew in shape
[21] with a radius of about 7 nm.
TABLE I
SEPARATIONS BETWEEN THE ANTIBODIES, DONOR AND ACCEPTOR DYES
scenario
A
B
C
D
Rabbit Anti-
antibody
on Tx side
Goat Anti-
Goat Anti-
Goat IgG, Fc
Goat Anti-
Mouse IgG
Mouse IgG
Fragment
Mouse IgG
115-505-008
115-505-008
Specific
115-515-062
305-515-046
Rabbit Anti-
Rabbit Anti-
Rabbit Anti-
Goat Anti-
Goat IgG, Fc
Goat IgG, Fc
Goat IgG, Fc
Mouse IgG
Fragment
Fragment
115-515-062
Specific 305-
Specific
Fragment
Specific
515-046
305-495-008
305-495-008
antibody
on Rx side
dyes on
Tx side
DyLight 549 DyLight 549
DyLight 594
DyLight 594
(donors)
dyes on
Rx side
DyLight 594 DyLight 594
DyLight 649
DyLight 649
(acceptors)
separation
between
antibodies
8-10 nm
11-13 nm
12-14 nm
11-13 nm
(centre to
centre)
In the labeling process, it is not possible to choose exact
positions of the dyes on the antibodies; the donors and
acceptors (DyLights 549, 594 and 649) can be attached to the
respective antibodies on their whole length1. Still, we have
constructed the antibodies chains in a way to have relatively
large distances between the donors and acceptors (on average
9 nm in scenario A and 12-13 nm in scenarios B-D). On the
basis of the known antibody geometries [21], we can assess
the separations between the antibodies where the dyes were
attached (see Table I). These separations are the average ones.
1 The positions of the dyes could be determined more precisely if smaller
molecules were used instead of Immunoglobulin G (IgG). For instance, the
donor and acceptor dyes could also be attached to fragment antigen-binding
(Fab) molecules, which are 3 times smaller than IgG [22]. The dyes on a Fab
particle would be located closer to each other, but having in mind smaller Fab
size comparing with IgG, it would be more difficult to obtain a high degree of
labeling.
Due to the distribution of the dyes on the surface of
antibodies, the real donor-acceptor distances may vary from
these estimated values by a few nanometers. Moreover, we
cannot exclude a slight rotation or bending of the antibodies,
what can change the final donor-acceptor distances by an
additional 1 nm. As the FRET probability decreases with the
sixth power of this distance, we expect that FRET happens
mainly between the donors and acceptors which are closest to
each other in the moment of the excitation of the donor. It is
another reason why the high DoL values are so important for
effective FRET-based nanocommunication.
According
ImmunoResearch Laboratories,
labeling for the dyes are:
• DyLight 549: DoL = 7
• DyLight 594 in scenarios A and C: DoL = 6.4
• DyLight 594 in scenarios B and D: DoL = 5.2
• DyLight 649: DoL = 5.52
These DoL values are not always integer numbers, which
means there are two possibilities of how many dyes are
attached to an antibody (e.g. 5 or 6, 6 or 7) and the resulting
DoL is the weighted mean of two numbers.
information provided by Jackson
the average degrees of
to
the
IV. LABORATORY METHODOLOGY
A. Cell culture
In order to allow their growth, HeLa cells were maintained
in Dulbecco's Modified Eagle Medium (containing all
necessary nutrients) supplemented with 10% fetal bovine
serum at 37ºC and 5% CO2. 24-48h before the experiment, the
cells were seeded on 18mm microscope slides. Before
immunolabeling, the cells were fixed in 4% formaldehyde,
permeabilized in 0.1% Triton X-100 and then blocked in 3%
bovine serum albumin (BSA). The fixed cells were incubated
with the primary antibody for 1h in 3% BSA at room
temperature, washed three times in phosphate-buffered saline
(PBS) and then incubated overnight at 4ºC with an appropriate
secondary (donor dye labeled) antibody. After performing
lifetime measurements of the structures with antibodies
labeled with the donor dye only, incubation with another
antibody (acceptor dye labeled) was performed at the
microscope stage in order to establish the required geometrical
configuration of the dyes.
B. FLIM measurements
FRET efficiency values cannot be measured directly.
Instead, two techniques are mainly used. The first one takes
advantage of the fact that fluorescence intensity of the donor
should be decreased in the presence of an acceptor. However,
accurate fluorescence
intensity measurements are often
impeded by spectral bleed-through and photobleaching of the
dyes. The second technique is based on the measurements of
fluorescence lifetimes of fluorophores. Fluorescence lifetime,
described as the average time a fluorophore spends in the
2 These DoL are quite high comparing with typical values reported in
fluorescence spectroscopy experiments. We have chosen these fluorophores
also because of their high possible DoL, as this factor determines the FRET
efficiency. The DyLight dyes with lower DoL could also be used here, but,
according to (5) and (6), the communication channel would have poorer
characteristics.
5
excited state before emitting a photon, is less influenced by
photobleaching and thus allows to measure FRET more
accurately. Because the lifetime of a donor should be
shortened in the presence of an acceptor, two measurements
are required: one before and one after the addition of the
acceptor. Then, the FRET efficiency can be calculated
according to the following formula [14]:
E
−=
1
with
−
acceptors
,
(8)
τ
τ
no
−
acceptors
where τno-acceptors and τwith-acceptors are the measured lifetimes of
the donor in the absence and presence of an acceptor,
respectively. In our experiments, to measure fluorescence
lifetimes of the DyLight dyes we used a Leica TCS SP5 II
SMD confocal microscope (Leica Microsystems GmbH)
integrated with FCS/FLIM TCSPC module from PicoQuant
GmbH (Fig. 3). The FLIM (Fluorescence Lifetime Imaging
Microscopy) module integrated with a confocal microscope
permits to record microscope images that contain information
about fluorescence intensity and lifetime in every pixel of the
image. In all measurements, the FLIM image was captured at
256x256 pixel resolution and at a speed of 200 lines/s. The
acquisition time for each image was set to 1 minute with a
laser repetition rate of 40 MHz. The laser power was set to
achieve a photon counting rate of 200-300 kCounts/s. In order
to measure FRET efficiency between DyLight dyes, two
lifetime images were collected for each scenario, one before
and one after the incubation with an acceptor-labeled
antibody. Both DyLight 549 (donors in scenarios A and B)
and DyLight 594 (donors in scenarios C and D) were excited
with 470 nm laser line and the emission was collected with
500-560 and 607-683 nm band-pass filters, respectively. Three
images of approximately 50x50 um were collected for each
pair, with 4-6 cells visible in this region. For the purposes of
the analysis, only the nuclei of the cells were selected in order
to exclude any signal resulting from non-specific binding of
the antibodies. The collected data was analyzed using
SymPhoTime II software (PicoQuant GmbH), excluding the
expected instrument response function time range (tail-fitting
method). The DyLight 549 and DyLight 594 lifetime decays
were fitted with a two-exponential function:
τ
/
2
where τ1 and τ2 are the lifetimes, while A1 and A2 represent the
amplitudes of the components. The final lifetime was
calculated as the amplitude-weighted average of τ1 and τ2:
⋅
eA
⋅
eA
)(
tF
(9)
+
=
,
τ
/
1
−
t
−
t
1
2
i
i
∑ ⋅
A τ
∑
A
i
i
τ
=
.
(10)
The goodness-of-fit was estimated based on the weighted
residuals and the chi squared value.
i
C. Image acquisition
Steady-state (containing information about fluorescence
intensity only) images of the cells were obtained using a Leica
TCS SP5 II SMD confocal microscope equipped with a 63x
NA 1.4 oil immersion lens. All images were captured at a
512x512 pixel resolution, with a scanning speed of 200 lines/s
and with 2 frames averaged. DyLight 549 was excited with a
488 nm laser line from an argon laser, while its emission was
Fig. 3. Schematic representation of the Laser Scanning Microscope
integrated with a FLIM module. The experiment begins when the sample
(HeLa cells) is excited with a pulsed laser (470 nm) controlled by a laser
driver. This unit allows us to control the laser power output and its repetition
rate. Upon excitation, the sample emits fluorescence photons which are
detected by SPAD detectors. In order to determine the exact time between
the excitation pulse and the arrival of the first photon at the detector, a
TCSPC module connected to a computer receives information from the
detectors, laser driver and laser scanner. In steady-state measurements
(image acquisition without information about fluorescence lifetimes), the
sample is excited with a continuous wave laser, the signal is detected in PMT
detectors and the resulting image is viewed on a LSM computer.
collected in a photo-multiplier tube (PMT) at 560-580 nm.
DyLight 594 was excited with a 594 laser line from a HeNe
laser and its emission collected at 600-620 nm. In analogy to
FLIM measurements, two images were collected for each
spatial configuration: one before and one after incubation with
the donor-labeled antibody. The equal laser power and
detector gain were used for each pair of the dyes.
V. EXPERIMENT RESULTS
Finally, for each of the scenarios A, B, C and D, we
performed the FLIM measurements for 13-17 nuclei of HeLa
cells. Every nucleus was analyzed separately. We estimated
that there were at least 3⋅106 constructed molecular structures
(a DNA molecule with a histone H1, antibodies labeled with
donor and acceptor dyes) with MIMO-FRET channels in
every nucleus. For each nucleus, we measured
the
fluorescence lifetimes of the acceptor molecules without and
in the presence of the donors (see Table II). The sets of data
together with fitting curves for all 4 scenarios are plotted in
Figure 4. These are the examples for the chosen nuclei. The
lifetimes may vary from one nucleus to another, as they are
influenced by the environment of the fluorophore (the density
of molecules not participating in the FRET process, like other
proteins and DNA) [23]. The lifetimes for the cases without
and with the donors are, however, usually proportional, what
means that the FRET efficiency, calculated with equation (8),
is stable. The average FRET efficiencies are given in Table
III.
Knowing the laser repetition rate (40 MHz) and the
achieved photon counting rate (200-300 kCounts/s), and
bearing in mind that the measured FRET efficiencies did not
exceed 50% (see the results below), we clearly see that we had
only isolated donor dyes excited (or no excitation at all) per
laser pulse. As each HeLa nucleus contained millions of donor
dyes, it was not possible to excite all the donor dyes
simultaneously
full MIMO-FRET
validate
and
the
6
transmission scheme. Instead, we measure the MIMO (1,m)
case, i.e. with a single donor excited and multiple acceptors
being able to receive the FRET energy.
TABLE II
MEASURED DONOR LIFETIMES FOR DIFFERENT HELA NUCLEI [NS]
A
no/with
acceptors
B
no/with
acceptors
Scenario
C
no/with acceptors
D
no/with
acceptors
0.81
0.8
0.82
0.83
0.75
0.82
0.83
0.84
0.79
0.8
0.75
0.8
0.73
0.78
0.59
0.58
0.59
0.59
0.6
0.62
0.61
0.62
0.6
0.61
0.59
0.61
0.58
0.6
0.80
0.77
0.84
0.80
0.78
0.80
0.81
0.83
0.83
0.75
0.75
0.75
0.79
0.73
0.77
0.76
0.75
0.80
0.77
0.75
0.72
0.74
0.74
0.75
0.70
0.68
0.71
0.73
0.68
0.71
2.10
2.12
2.11
2.15
2.12
2.17
2.32
2.25
2.77
2.79
1.95
1.87
2.07
1.83
2.05
2.06
1.95
1.51
1.48
1.52
1.50
1.57
1.51
1.51
1.49
1.65
1.75
1.15
1.13
1.24
1.09
1.24
1.14
1.17
2.03
2.00
2.04
2.07
2.03
2.20
2.22
2.22
2.19
2.18
2.21
2.22
2.20
Average lifetimes and their standard deviations for each scenario
0.8 ±
0.6 ±
0.79 ±
0.03
0.01
0.03
0.73
±
0.03
2.16 ±
1.39 ±
2.14 ±
0.26
0.21
0.09
1.35
1.37
1.41
1.43
1.39
1.25
1.27
1.24
1.21
1.26
1.27
1.26
1.23
1.3
±
0.08
)
1(5.0
E−⋅
The measured FRET efficiencies together with confidence
intervals for all 4 scenarios are shown in Table III. For each
FRET efficiency, we also give the respective BER values,
. At the same time, the FRET
calculated as:
phenomenon was observed in microscope images as a
decrease in donor fluorescence intensity after adding the
acceptor dyes, see Fig 5.
The measurement results can be verified by FRET theory.
Assuming the average donor-acceptor separation equal to the
distance between the respective antibodies (see Table I) and
putting it together with the known R0 into the equation (3), we
can calculate theoretical values of E(1,m). The calculations must
be performed twice, as we do not know the exact number of
acceptors: in scenario A we had 6 or 7 acceptors (DoL = 6.4),
in scenarios B-D we had 5 or 6 acceptors (DoL = 5.2 and 5.5).
Hence, we obtain a theoretical range for E(1,m): it is given in
the last row of Table III.
The measurement results match the theoretical ones quite
well, except in scenario C. In all scenarios, the real donor-
acceptor separations may vary, as we have no control
Fig. 4. Representative decay and fitting curves of the donor in the absence
and presence of the acceptor measured in HeLa nuclei for scenarios A, B, C
and D.
regarding exact positions of the DyLight dyes. Furthermore,
the antibodies chain shape depends on the axial (0-180 deg)
and segmental flexibility (0-180 deg); additionally the switch
peptide (elbow) also increases the flexibility of the antibody
Fab region (0-50 deg) [24], which may influence the donor-
acceptor separation. In scenario C, we assumed the labeled
antibodies bound the linker antibody in the best way from the
energetic point of view, which meant the distance between the
labeled antibodies was maximal. However, in practice, they
may bind closer to each other; thus the donor-acceptor
separation is smaller than predicted in Table III. As the FRET
efficiency is inversely proportional to the sixth power of this
separation, these effects cause higher than expected FRET
efficiency.
The experiments were performed for relatively
long
distances between the Tx and Rx sides, i.e. with the donor-
acceptor separations significantly exceeding their Förster
distances. Thus, as we can see from Table III, even with a
large number of acceptor molecules, it was not possible to
obtain FRET efficiency higher than 50%. Also the respective
bit error rates (30-50%) are highly non satisfactory for
telecommunication purposes. A solution for this (other than
exciting the donor multiple times per bit, what however
decreases the corresponding data throughput [25]) might lie in
exploiting full MIMO communication, i.e. to excite all the
donor dyes simultaneously. It might be realized with a very
strong external laser impulse or using the energy of a local
chemical
is called
Bioluminescence Resonance Energy Transfer. The receiver
decodes the bit as '1' if any of acceptors is excited, i.e. if at
least one donor transfers its excitation energy to at least one
acceptor. If no acceptors are excited, the bit is decoded as '0'.
For the proper data transmission, it is crucial to correctly
choose the rate (frequency) of sending bits. Obviously, from
the telecommunication point of view the higher rate the better,
but the FRET and fluorescence delays are going to put the
limits here. As documented in Table II and Fig. 4, with
average fluorescence lifetimes of about 1-2 ns, almost all
excited molecules will go back to the ground state after 15-20
ns. Since we do not know the exact FRET rate between
fluorophores in our experiments, we have to assume that the
energy transfer can occur in any time between a few to about
20 ns, i.e. as long as we observe photons after a single laser
pulse. Additionally, we should take into account the relaxation
time of the acceptor molecule. When the acceptor is excited, it
latter case
reaction,
this
in
it
7
cannot receive another signal from the donor3. Thus, the
acceptor should first pass the signal via another FRET to other
molecules or emit a photon. It adds another 20 ns to the total
delay of the signal transfer. Consequently, in this case the
transmission rate should not exceed 1 bit per 40 ns, i.e. 25
Mbit/s. Moreover, as discussed earlier, the fluorophore
excitation lifetimes may additionally vary because of the
influence of neighboring molecules, what should be also taken
into account when setting the transmission bit rate.
TABLE III
RESULTS OF THE FLIM MEASUREMENTS AND CALCULATIONS
scenario
n (DoL of donor molecules)
A
7
B
7
m (DoL of acceptor molecules)
6-7
5-6
C
6-7
5-6
D
5-6
5-6
measured E(1,m) [%]
25 ± 2
7 ± 2
36 ± 5
39 ± 6
respective BER(1,m)
R0 [nm]
0.375
0.465
5.62
5.62
av. donor-acceptor separation [nm]
9
12
0.32
8.15
13
0.305
8.15
12
theoretical range for E(1,m) [%]
26-29
5-6
23-27
33-37
The probability that at least one of many donors transfers its
excitation energy to at least one acceptor is much higher than
having a single donor only, as explained in Section II. We can
calculate FRET efficiency and BER for (n,m) case, comparing
equations (3) with (6) and (4) with (7):
)n
−−=
(11)
(
1
,
E
1
E
,1
m
mn
,
⋅
⋅
m1,
mn,
=
)n
.
BER
(
25.0
BER
and:
(12)
Consequently, on the basis of the measured values of E(1,m), we
can predict the performance of the full (n,m) MIMO-FRET
channels. The results of those are given in Table IV. For
scenarios C and D we have some ranges, as the number of
donor dyes varies. It is worth noting that bit error rates at the
level of 2-4% are easily achievable, even for a distance of 13
nm. The exception is the scenario B, where, however, the
distance between the donor and acceptor sides is more than
twice the length of the respective Förster distance.
In order to give the reader an even broader view as to what
bit error rates can be achieved in MIMO-FRET channels, we
also provide full BER curves for dye pairs being the subject of
our experiments (549-594, 594-649), as well as for a few other
DyLight dyes, well suited to nanocommunication. The curves
are presented in Figure 6; the respective Förster distances and
the numbers of donors/acceptors are also given. As we can
see, using multiple acceptors increases the range where FRET-
based communication can be used effectively. If we assume
the required BER level to be 0.1%, the transmission range is
about 35% longer in channels (1,6) compared to (1,1). Full
MIMO communication is, however, even more efficient: in
channels (6,6), the transmission range is 3.5 times longer than
3 In a scenario with multiple acceptors (MIMO-FRET), one might not wait
for the excited acceptor to release its energy, but instead, to transfer the signal
to other acceptor molecules. It, however, would complicate the decoding
process and decrease the FRET efficiency (as the number of available
acceptors is lower).
%1.0
BER <
Fig. 5. The decrease of fluorescence intensity being the result of FRET in
scenarios A, B, C and D. Two images for each scenario are shown: one
before (left column) and one after (right column) staining with acceptor-
bound secondary antibody. In all configurations, addition of the acceptor
resulted in the decrease of fluorescence intensity of the donor. Scale bar: 5
um.
in channels (1,1). As a rule of thumb, it is worth remembering
that for each donor-acceptor pair, the
can be
achieved at a distance equal to R0, if a FRET channel (4,4) or
larger is used (see Eq. (7)).
As FRET transmissions are so strictly limited in their range
(E decreases with the sixth power of the distance), the ability
to create multi-hop connections is critical when thinking about
future nanonetworks. Fluorophores have their emission spectra
shifted a little towards the higher wavelengths compared to
their respective absorption spectra [14], thus it is generally
quite difficult to send a signal via FRET among identical
fluorophores (there are some exceptions, see homo-FRET
[14]). Instead, chains of spectrally matched molecules can be
created. From Fig. 6, we see that a 6-element DyLight chain:
405→488→549→594→649→680 may effectively transfer a
signal over a distance of several dozen nanometers (assuming
each hop is equal to R0, see the rule of thumb above). Such
chains may be built not only with DyLight molecules, but also
with GFP and its derivatives, Alexa Fluor, CyDye, LI-COR
dyes, etc.
FRET EFFICIENCIES AND BER VALUES FOR (1,M) AND (N,M) MIMO-FRET
TABLE IV
CHANNELS
Scenario
n (DoL of donors)
m (DoL of acceptors)
A
7
6-7
measured E(1,m) [%]
25 ± 2
respective BER(1,m)
calculated E(n,m) [%]
calculated BER(n,m)
0.375
86.7
0.067
B
7
5-6
7 ± 2
0.465
40
0.3
C
6-7
5-6
D
5-6
5-6
36 ± 5
39 ± 6
0.32
93-96
0.305
91-95
0.02-0.04
0.025-0.04
VI. CONCLUSIONS AND FUTURE WORKS
than
other
techniques
considered
In this article, we have investigated the Förster Resonance
Energy Transfer as a means of nanocommunication. The
FRET phenomenon is characterized by a very short timescale,
in the order of nanoseconds, in which a signal is non-
radiatively
transferred between molecules. Consequently,
transmissions via FRET have much smaller propagation
delays
for
nanocommunications (even 4-5 orders of magnitude compared
with calcium signaling, flagellated bacteria, molecular or
catalytic motors) and a potential throughput of tens or even
hundreds of Mbit/s.
The main drawbacks of FRET-based nanocommunication
lie in its high bit error rate and short transmission ranges.
While the former issue can be solved using multiple donors
and acceptors, the latter requires a careful design of FRET
communication channels, especially by choosing
for
transmitter and receiver antennas the pairs of fluorescent dyes
characterized by large Förster distances and high degrees of
labeling. In this paper, we have proposed to use a fluorescent
family of DyLight dyes which
the
abovementioned criteria. We have constructed 4 different
nanonetworks based on Immunoglobulin G antibodies and
DyLight
dyes
the
communication
their nodes. The
experiments confirm successful communication over the
distances of 12-13 nm, which is a 50% increase comparing
with the results previously reported in the literature. Higher
ranges can be reached via multi-hop connections making use
of fluorophore chains.
While these research results are very promising, many
questions still remain. Importantly, we do not know how
future
their
nanoantennas (the fluorophore molecules) and send signals to
them. One option is Bioluminescence Resonance Energy
Transfer (BRET), where a nanomachine may induce a
chemical reaction generating energy that excites donors. In
order to fully exploit MIMO-FRET communication, all donors
should be excited simultaneously, which, however, may not be
straightforward to achieve.
The experimentally measured FRET efficiencies are in
reasonably good agreement with theoretical values, but more
exact assessments could be made if the distances between the
fluorophores were controlled with higher precision. These
nanomachines will
communicate with
efficiency between
and
experimentally
validated
fits very well
8
Fig. 6. Bit error rate curves for the chosen MIMO-FRET channels based on
DyLight dyes.
could be obtained with the aid of molecular biology and
genetic engineering techniques using polymers [26]. Polymers
are chains of biological molecules composed of many repeated
sub-units; the length of one polymer can easily reach 1000 nm,
whereas a single element can be 1 nm or less. In living
organisms we can find a few types of polymers: proteins (e.g.
actin filaments, microtubules), polysaccharides (e.g. starch)
and nucleic acids (DNA and RNA). If sub-units of a polymer
were tagged with fluorescent molecules, the distances between
these fluorophores could be determined with high precision:
about 1 nm or better. Additionally, a polymer molecule can be
composed of a main chain with one or more side chains or
branches, which may be an interesting scenario considering
possible
for FRET-based
nanocommunication.
research on signal
routing
REFERENCES
[1] B. Fadeel, Nanosafety: towards safer design of nanomachines, Journal of
[2]
Internal Medicine, 2013, 274, 578-580.
J. Gupta, Nanotechnology applications in medicine and dentistry,
Journal of Investigative and Clinical Dentistry, 2011, 2, 81-88.
[3] K. Kim, J. Guo, X. Xu, D. Fan, Recent Progress on Man-Made
Inorganic Nanomachines, Small, 2015, DOI: 10.1002/smll.201500407.
[4] T. Nakano, T. Suda, M. Moore, R. Egashira, A. Enomoto, K. Arima,
Molecular communication for nanomachines using intercellular calcium
signaling, Proceedings of
IEEE Conference on
Nanotechnology, June 2005, pp. 478–481.
the Fifth
[5] T. Nakano, A.W. Eckford, T. Haraguchi, Molecular Communications,
Cambridge University Press, 2013.
[6] M. Gregori, I.F. Akyildiz, A New NanoNetwork Architecture using
Flagellated Bacteria and Catalytic Nanomotors, IEEE Journal of
Selected Areas in Communications, vol. 28, pp. 612-619, May 2010.
[7] M. Moore, A. Enomoto, T. Nakano, R. Egashira, T. Suda, A. Kayasuga,
H. Kojima, H. Sakakibara, K. Oiwa, A design of a molecular
communication system for nanomachines using molecular motors,
Proceedings of the Fourth Annual IEEE International Conference on
Pervasive Computing and Communications, March 2006.
[8] V. Serreli, C. Lee, E.R. Kay, D.A. Leigh, A molecular information
ratchet, Nature 445: 523–527, 2007.
[9] L. Parcerisa, I.F. Akyildiz, Molecular Communication Options for Long
Range Nanonetworks, Computer Networks (Elsevier) Journal, vol. 53,
pp. 2753-2766, 2009.
[10] V. Loscri, A.M. Vegni, An Acoustic Communication Technique of
Nanorobot Swarms for Nanomedicine Applications, IEEE Transactions
on NanoBioscience, vol. 14, no. 6, pp. 598-607, September 2015.
9
[11] M. Kuscu, O.B. Akan, A Physical Channel Model and Analysis for
Nanoscale Communications with Förster Resonance Energy Transfer
(FRET), IEEE Transactions on Nanotechnology, vol. 11, no. 1, pp. 200-
207, January 2012.
[12] M. Kuscu, O.B. Akan, Multi-Step FRET-Based Long-Range Nanoscale
in
Communication Channel,
Communications, vol. 31, no. 12, pp. 715-725, December 2013.
IEEE Journal on Selected Areas
[13] K. Wojcik, K. Solarczyk, P. Kulakowski, Measurements on MIMO-
FRET nanonetworks based on Alexa Fluor dyes, IEEE Transactions on
Nanotechnology, vol. 14, no. 3, pp. 531-539, May 2015.
[14] Joseph R. Lakowicz, Principles of Fluorescence Spectroscopy 3 ed.,
Springer, 2006.
[15] Fábián ÁI, Rente T, Szöllosi J, Mátyus L, Jenei A., Strength in numbers:
on FRET efficiency.
effects
abundance
Chemphyschem. 2010 Dec 3;11(17):3713-21.
acceptor
of
[16] I.E. Telatar, Capacity of Multi-antenna Gaussian Channels, AT&T Bell
Laboratories, Technical Memorandum, June 1995.
[17] G.J. Foschini and M.J. Gans, On limits of wireless communications in a
fading environment when using multiple antennas, Wireless Personal
Communications, vol. 6, pp. 311–335, March 1998.
[18] DyLight
Fluors
–
Technology
and
Product
Guide,
http://www.piercenet.com/guide/dylight-fluors-technology-product-
guide
[19] P. Sarkar, S. Sridharan, R. Luchowski, S. Desai, B. Dworecki, M. Nlend,
Z. Gryczynski, I. Gryczynski, Photophysical properties of a new
DyLight 594 dye, Journal of Photochemistry and Photobiology B:
Biology, Volume 98, Issue 1, 21 January 2010, Pages 35-39.
[20] S. D'Auria, E. Apicella, M. Staiano, S. Di Giovanni, G. Ruggiero, M.
Rossi, P. Sarkar, R. Luchowski, I. Gryczynski, Z. Gryczynski,
Engineering resonance energy transfer for advanced immunoassays: The
case of celiac disease, Analytical Biochemistry, Volume 425, Issue 1, 1
June 2012, Pages 13-17.
[21] Yih Horng Tan, Maozi Liu, Birte Nolting, Joan G. Go, Jacquelyn
Gervay-Hague, Gang-yu Liu, A Nanoengineering Approach for
Investigation and Regulation of Protein Immobilization, ACS Nano 2
(11), pp. 2374–2384, 2008.
[22] Jeremy M. Berg, John L. Tymoczko, Lubert Stryer, Biochemistry, 5th
Edition. New York, W.H. Freeman and Company, 2002.
[23] Mikhail Y. Berezin, Samuel Achilefu, Fluorescence Lifetime
Measurements and Biological Imaging, Chem. Rev., 2010, 110(5), pp.
2641–2684.
[24] William E. Paul, Fundamental Immunology 7th Ed., Wolters Kluwer
Health/Lippincott Williams & Wilkins, 2013.
[25] M. Kuscu, O.B. Akan, FRET-Based Nanoscale Point-to-Point and
Broadcast Communications with Multi-Exciton Transmission and
Channel Routing, IEEE Transactions on NanoBioscience, vol. 13, no. 3,
pp. 315-326, September 2014.
[26] M. Rubinstein, R.H. Colby, Polymer physics. Oxford, New York,
Oxford University Press, 2003.
from
Kamil Solarczyk received the M.Sc. degree in
biophysics
the Jagiellonian University,
Kraków, Poland, in 2010. Since then, he has been
working toward the Ph.D. degree in the Faculty of
Biochemistry, Biophysics and Biotechnology. His
research
repair
processes, chromatin architecture and molecular
communication.
the DNA
interests
include
from
in biophysics
Krzysztof Wojcik received the M.Sc. and Ph.D.
degrees
the Jagiellonian
University in Krakow, Poland 2003 and 2015,
respectively; and M.D. from
the Jagiellonian
University Medical College in Kraków, Poland
2007. He was an Assistant at Division of Cell
Biophysics Faculty of Biochemistry, Biophysics
and Biotechnology Jagiellonian University (2007-
2014). He
is an Assistant at Allergy and
Immunology Clinic in II Chair of Internal Medicine
CMUJ. His research interests include the fields of confocal microscopy
techniques and its application in autoantibodies research and use of
fluorescent labeled antibodies in nanocommunications.
Pawel Kulakowski received the Ph.D. degree in
telecommunications from the AGH University of
Science and Technology in Krakow, Poland, in
2007. Currently he is working there as an Assistant
Professor. He was also working few years in Spain,
as a visiting post-doc or professor at Technical
University of Cartagena, University of Girona,
University of Castilla-La Mancha and University of
Seville. He co-authored about 30 scientific papers,
in journals, conferences and as technical reports.
He was also involved in numerous research projects, especially European
COST Actions: COST2100, IC1004 and CA15104 IRACON, focusing on
topics of wireless sensor networks, indoor localization and wireless
communications in general. His current research interests include molecular
communications and nano-networks. Dr. Kulakowski was recognized with
several scientific distinctions, including 3 awards for his conference papers
and a scholarship for young outstanding researchers.
|
1803.09071 | 1 | 1803 | 2018-03-24T07:38:08 | Chemical event chain model of coupled genetic oscillators | [
"physics.bio-ph",
"nlin.AO",
"q-bio.CB"
] | We introduce a stochastic model of coupled genetic oscillators in which chains of chemical events involved in gene regulation and expression are represented as sequences of Poisson processes. We characterize steady states by their frequency, their quality factor and their synchrony by the oscillator cross correlation. The steady state is determined by coupling and exhibits stochastic transitions between different modes. The interplay of stochasticity and nonlinearity leads to isolated regions in parameter space in which the coupled system works best as a biological pacemaker. Key features of the stochastic oscillations can be captured by an effective model for phase oscillators that are coupled by signals with distributed delays. | physics.bio-ph | physics | Chemical event chain model of coupled genetic oscillators
David J. Jorg,1, ∗ Luis G. Morelli,2, 3, 4 and Frank Julicher1, †
1Max Planck Institute for the Physics of Complex Systems, Nothnitzer Str. 38, 01187 Dresden, Germany
2Instituto de Investigaci´on en Biomedicina de Buenos Aires
(IBioBA)–CONICET–Partner Institute of the Max Planck Society,
Polo Cient´ıfico Tecnol´ogico, Godoy Cruz 2390, C1425FQD, Buenos Aires, Argentina
3Departamento de F´ısica, FCEyN UBA, Ciudad Universitaria, 1428 Buenos Aires, Argentina
4Max Planck Institute for Molecular Physiology, Department of
Systemic Cell Biology, Otto-Hahn-Str. 11, 44227 Dortmund, Germany
(Dated: March 28, 2018)
We introduce a stochastic model of coupled genetic oscillators in which chains of chemical events
involved in gene regulation and expression are represented as sequences of Poisson processes. We
characterize steady states by their frequency, their quality factor and their synchrony by the oscillator
cross correlation. The steady state is determined by coupling and exhibits stochastic transitions
between different modes. The interplay of stochasticity and nonlinearity leads to isolated regions in
parameter space in which the coupled system works best as a biological pacemaker. Key features
of the stochastic oscillations can be captured by an effective model for phase oscillators that are
coupled by signals with distributed delays.
generate an adequate stochastic description starting from
a chemical reaction scheme.
Genetic oscillators are a prime example of genetic feed-
back systems, in which stochastic properties are impor-
tant. A prominent genetic oscillator is the circadian
clock of humans, animals, and plants, where oscillations
are used to provide information about the daytime to
the organism [7–15]. Genetic oscillators also play an
important role during embryonic development, e.g., in
neuronal differentiation [16, 17] and the segmentation of
the body axis [17–19]. Genetic oscillators are charac-
terized by gene regulatory networks that autonomously
generate time-periodic changes in gene product num-
bers of so-called cyclic genes [5, 6, 20]. This is typ-
ically achieved by a negative transcriptional feedback
of the cyclic genes on themselves that involves a suffi-
ciently large time delay [5, 21, 22]. In recent years, ge-
netic oscillators have also been engineered in artificial
I.
INTRODUCTION
Biological cells are complex dynamic systems which use
specific proteins to activate and inhibit genes to ensure
robust control of cellular functions [1, 2]. The production
of such proteins themselves is mediated by gene activity,
giving rise to feedback systems and complex dynamics.
The production of a gene product from an active gene
comprises a series of chemical events such as transcrip-
tion of DNA to RNA, splicing, and translation of RNA
to a protein [3]. Gene regulation involves, e.g., the trans-
port and binding of regulatory proteins to DNA. Gene
products can also serve as chemical signals that are me-
diated across different cells through so-called signaling
pathways that involve production, transport, and bind-
ing of signaling molecules to receptor molecules or DNA
[4]. Such sequences of chemical events typically involve
the generation of intermediate products such as messen-
ger RNA (mRNA), transcription factors, and signaling
ligands. Because of their complexity, such systems are
often represented by simplified chemical rate equations
that bypass intermediate steps and often neglect fluc-
tuations [5]. Motivated by an earlier approach [6], we
propose to describe genetic feedbacks by chemical event
chains composed of a sequence of Poisson processes, see
Fig. 1. These chains represent the sequences of transi-
tions between intermediate chemical states, e.g., between
mRNAs of different lengths and from mRNA to spliced
mRNA. This approach captures generic stochastic prop-
erties of complex cellular processes and can be used to
∗ Present address: Cavendish Laboratory, Department of Physics,
University of Cambridge, JJ Thomson Avenue, Cambridge CB3
0HE, United Kingdom and The Wellcome Trust/Cancer Re-
search UK Gurdon Institute, University of Cambridge, Tennis
Court Road, Cambridge CB2 1QN, United Kingdom
† [email protected]
FIG. 1. Exemplary depiction of a simple genetic feedback
system with a gene x and a gene product y. (a) Simplified
feedback scheme that can be mapped onto a dynamic system
and (b) representation as a chemical chain involving interme-
diate chemical states xi and yi that can be mapped onto a
master equation.
8
1
0
2
r
a
M
4
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
1
7
0
9
0
.
3
0
8
1
:
v
i
X
r
a
xyxyabx1statedynamics...xnyny1 2
an effective phase oscillator model that captures key fea-
tures of this system. It is based on distributions of delay
times in the oscillator coupling that captures collective
frequency and stability properties of stochastic oscilla-
tions.
II. CHEMICAL EVENT CHAIN MODEL OF
COUPLED GENETIC OSCILLATORS
We first introduce a Markov model for chemical event
chains that captures the genetic interactions of coupled
zebrafish somitogenesis oscillators (Fig. 2), see Fig. 3a.
The state of the system is specified by the occupation
numbers of each step of the chain. We denote the number
of signaling molecules of oscillator µ = 1, 2 at step i =
0, . . . , n by xµi and the number of cyclic molecules at
step i = 0, . . . , n by xµi, see Fig. 3a. Synthesis of cyclic
molecules takes place at the initial step i = 0 of the
oscillators. Molecules undergo a transition from step i to
i + 1 at a constant transition rate and decay at the final
step i = n. The synthesis rate of both cyclic and signaling
molecules is regulated by the amount of molecules at the
final step i = n of the oscillator. Signaling molecules also
undergo transitions through a sequence of steps with the
last step of the signaling pathway regulating the synthesis
rate of cyclic molecules in the receiving oscillator.
A. Model formulation
We describe the dynamics of the system using a master
equation [50] that governs the time evolution of the prob-
ability P (x, t) to find the system in the state x at time
t, where
x = (x10, . . . , x1n, x10, . . . , x1n, x20, . . . , x2n, x20, . . . , x2n)
is the state vector of all occupation numbers. The master
equation is given by
FIG. 2. The zebrafish somitogenesis oscillator as an exam-
ple for coupled genetic oscillations. Shown are two cells that
act as autonomous oscillators and that are coupled through
the Delta–Notch signaling pathway, an example for juxtacrine
signaling [37, 49]. Coupling is bidirectional, that is, each cell
acts as both a sender and a receiver. During embryonic devel-
opment, a tissue comprising these cellular oscillators guides
the segmentation of the elongating body axis [19].
systems [20, 23–29]. Both natural and artificial genetic
oscillators exhibit pronounced amplitude and phase noise
[6, 12, 24, 30–36], which limits their precision when used
as a clock. To achieve temporal and spatial coherence
as well as high precision, cell-autonomous oscillators are
typically coupled [19, 37, 38]. Such coupling facilitates
synchronization and can affect the collective frequency
[39–47]. Moreover, coupling between cellular oscilla-
tors via paracrine or juxtacrine signaling (i.e., via dif-
fusible signals or contact-dependent signaling) typically
proceeds at time scales similar to the oscillation period,
implying the presence of coupling delays that can have
profound effects on the coupled dynamics [21, 40–42, 48].
In this paper, we present a framework to study the
stochastic properties of genetic oscillators that are cou-
pled by signaling pathways. As an example we consider
the zebrafish somitogenesis oscillator [17, 37], see Fig. 2.
We investigate the precision and stochastic properties of
collective oscillations and how they emerge from the ki-
netics of chains of chemical events. Finally, we present
(xµi + 1)E+
µi
µ,i+1 − xµi
E−
+ κ
(xµn + 1)E+
µn − xµn
+ ψ−
(cid:40)
2(cid:88)
µ=1
λ
(cid:124)
∂P
∂t
=
(cid:20)
n−1(cid:88)
n−1(cid:88)
i=0
i=0
+ λ
(cid:124)
(cid:20)
(cid:21)
(cid:125)
µ,i+1 − xµi
E−
(cid:20)
(cid:124)
(cid:21)
(cid:125)
(cid:20)
(cid:124)
(cid:123)(cid:122)
oscillator chain
(xµi + 1)E+
µi
(cid:123)(cid:122)
signaling chain
(E−
(cid:125)
µ0 − 1)
P .
(1)
(cid:123)(cid:122)
(cid:21)
(cid:125)
µn − xµn
(cid:124)
(cid:21)
(cid:125)
(cid:123)(cid:122)
(cid:18) xµn
p
+ αψ−
(cid:124)
α + βψ+
(cid:19)(cid:20)
(cid:18) xµn
(cid:19)
(cid:123)(cid:122)
q
(cid:19)(cid:21)
(cid:41)
q
(cid:18) x¯µn
(cid:123)(cid:122)
(cid:125)
µ0 − 1)
(E−
decay of the cyclic product
regulation of cyclic genes
+ κ
(xµn + 1)E+
decay of the signaling product
regulation of signaling genes
Here, ¯µ = 2δµ,1 + 1δµ,2 refers to the index of the respec-
tive other oscillator. The creation and annihilation op-
erators E±
µi increase or decrease the product levels xµi
Delta ligandNotch receptorNotch intracellular domain (NICD)cyclic genes
genetic oscillator 3
FIG. 3. (a) Schematics of the chemical event chain model of two coupled genetic oscillators as described by Eqs. (1,2). Boxes
mark the initial and final products of a multi-step process and broken lines the intermediate products. (b) Hill functions ψ−
and ψ+ as given by Eqs. (2) for different values of the exponent h.
µif (x10, . . . , xµi, . . . , x2n) = f (x10, . . . , xµi ±
by one, E±
1, . . . , x2n), and analogously for the operators Eµi and
product numbers xµi [51]. The kinetic parameters char-
acterizing the biochemical properties of gene expression
and interaction are listed and explained in Table I and
shown in Fig. 3a. Activation and repression of gene ex-
pression at the initial stages of the oscillators and the
signaling pathways are described by functions of the Hill
type [5],
ψ−(x) =
1
1 + xh ,
ψ+(x) =
xh
1 + xh ,
(2)
where ψ− describes inhibition and ψ+ describes activa-
tion and the exponent h determines the nonlinearity of
the feedback, see Fig. 3b. Here, we are interested in
steady-state solutions of the master equation (1), which
describe the long-term collective behavior of the system.
B. Characterization of oscillator coupling via
stochastic signaling
We first summarize features of the introduced coupling
process that are useful to parametrize the system and
understand its limiting cases. The coupling strength de-
pends on several model parameters: the production rate
α and inhibition threshold q of the signaling molecules
in the sending oscillator, as well as the activation rate β
and the activation threshold q in the receiving oscillator,
see Fig. 3 and Table I. The limiting case of uncoupled
oscillators can be realized through either (i) α = 0, (ii)
β = 0, (iii) q → ∞, (iv) q → 0, or any combination
thereof.
Moreover, the chain of Poisson processes of the signal-
ing pathway effectively generates a Gamma distribution
of arrival times for molecules starting at step i = 0 and
arriving at step i = n [6],
g(t) =
λn
(n − 1)!
tn−1e−λt .
(3)
Hence, the mean τ and variance σ2 of this distribution
characterize the mean signaling delay and the dispersion
of signaling delays,
τ = n/λ ,
σ2 = n/λ2 .
(4)
The arrival time distribution g can be interpreted as a
memory kernel for a probability that effectively summa-
rizes the effects of noise and delays introduced by stochas-
tic signaling.
C. Correlation functions and oscillator quality
Before studying the dynamics of the coupled genetic os-
cillator system, we introduce measures that characterize
their function: the quality factor measuring frequency
fluctuations and the cross correlation measuring syn-
chrony. To define these quantities, we introduce the tem-
poral correlation function Cµν between the final prod-
ucts xµn and xν n of the oscillators µ and ν,
Cµν(t) = (cid:104)xµn(t)xν n(0)(cid:105) − (cid:104)xµn(cid:105)(cid:104)xν n(cid:105) ,
(5)
where the brackets denote steady-state expectation val-
ues. The quality factor of an oscillator can be defined as
follows. The normalized temporal autocorrelation func-
tion of an oscillator is given by
G(t) =
Cµµ(t)
Cµµ(0)
.
(6)
Since both oscillators and signaling pathways are entirely
identical, the autocorrelation G does not depend on the
�������ψ-��������ψ+12abOscillator #1Signaling pathwaysOscillator #2x10~x1ñ~x2ñ~x20~x2nx20x1nx10activationdecaymulti-step processinhibitionh = 1h = 2h = 3h = 1h = 2h = 3 Param.
Unit
Value
Description
Oscillators
4
n
α
λ
κ
p
β
1
NT−1
T−1
T−1
N
NT−1
Signaling pathways
n
α
λ
κ
q
q
h
1
NT−1
T−1
T−1
N
N
1
18
60
1.5
0.5
20
20
10
60
0.5
0.5
100
20
2
number of steps
basal production rate
transition rate between steps
decay rate for the final product
threshold for cyclic auto-inhibition
activation strength due to signaling
number of steps
basal production rate
transition rate between steps
decay rate for the final product
threshold for activation by signaling
threshold for repression of signaling
Hill exponent (repression, activation)
TABLE I. Parameters used for numerical simulations.
refers to molecule numbers and 'T' is the unit of time.
'N'
see Fig. 4 for examples. The simulation method is de-
tailed in Appendix A.
A. Coupling delays determine the mode of
synchrony
First, we study how coupling via stochastic event chains
affects the mode of synchrony of the two oscillators. To
this end, we focus on the parameters β and q as a measure
of coupling strength and vary the mean signaling delay τ
by changing the transition rate λ (see Section II B). We
find that stochastic signaling delays determine whether
the oscillator system exhibits in-phase, anti-phase, or un-
correlated oscillations: Figs. 5a–c show density plots of
the cross correlation C as a function of the effective cou-
pling delay τ as well as the activation rate β, the acti-
vation threshold q, and the decay rate κ of the signaling
molecules. These plots reveal an alternation of in-phase
and anti-phase correlated regions as a function of the
signaling delay τ with in-phase regions located around
integer multiples of the uncoupled period T and anti-
phase regions located around odd multiples of T /2. This
behavior is generally known for coupled oscillators with
delayed coupling [53]. For increasing signaling delays τ ,
which imply increasing dispersions σ of delay times in our
parametrization, the correlations between the oscillators
decay until eventually, for large signaling delays, the os-
cillators become effectively uncoupled. Regions with a
high degree of correlation (large C) are separated by
regions of uncorrelated oscillations.
FIG. 4. Example time series of the cyclic final products x1n
and x2n for different synchronization scenarios: (a) in-phase,
(b) uncorrelated, and (c) anti-phase. For all plots, parameters
are given in Table I except for the transition rate λ, which
is chosen such that the effective coupling delay τ , given by
Eq. (4), takes values τ = 0.1T (a), τ = 0.2T (b), and τ = 0.5T
(c), where T = 28 is the period of the uncoupled oscillators.
oscillator index µ. For a noisy oscillator, this autocorre-
lation function typically exhibits a functional form of the
type G(t) (cid:39) cos(2πt/T )e−t/tc for large t, where T is the
period of oscillations and tc is the correlation time. We
define the quality factor Q as the dimensionless ratio of
the correlation time and the oscillation period [6, 52],
Q = tc/T .
(7)
The quality factor Q corresponds to the number of cycles
over which the oscillatory signal stays highly correlated,
thus quantifying the number of cycles over which the os-
cillators serve as a viable clock.
The synchrony of two stochastic trajectories is related
to the degree of correlation of their individual dynamics.
To quantify synchrony, we compute the normalized cross-
correlation of the final products of the oscillators,
(cid:112)C11(0)C22(0)
C12(0)
C =
,
(8)
with Cµν given by Eq. (5). The cross-correlation C de-
scribes the fraction of shared fluctuations between both
signals and its sign indicates the mode of synchrony: C
takes values in the interval [−1, 1], ranging from perfect
correlation (C = 1) to no correlation (C = 0) to perfect
anti-correlation (C = −1), which in the case of oscilla-
tions corresponds to in-phase oscillations, phase-drifting
oscillations and anti-phase oscillations, see Fig. 4.
III. FREQUENCY, QUALITY AND
SYNCHRONY OF COUPLED GENETIC
OSCILLATORS
We numerically investigate frequency, quality, and syn-
chrony of the coupled genetic oscillators by generating
multiple realizations of the stochastic process described
by the master equation (1) using numerical simulations,
As shown above, coupling tends to synchronize oscilla-
tors, implying that they attain a common collective fre-
quency. If coupling is delayed, as is the case here, this
B. Coupling affects the collective frequency
timetimetimeabcin-phaseuncorrelatedanti-phase 5
FIG. 5. Coupling delays determine the mode of synchrony and the collective frequency.
(a–c) Density plots of the cross
correlation C, Eq. (8). Blue colors indicate positive values of C corresponding to in-phase correlations, red colors indicate
negative values of C corresponding to anti-phase correlations. (d–f) Density plots of the collective frequency Ω. The parameters
that are not varied are given in Table I.
collective frequency can differ from the frequency of the
uncoupled oscillators [53, 54]. Figs. 5d–f show density
plots of the collective frequency Ω of both oscillators, ob-
tained from the autocorrelation, Eq. (6), as a function of
the same parameters as in Figs. 6 and 5. As a function
of the signaling delay τ , these plots reveal sharp changes
of the frequency at odd multiples of T /4. This indicates
that the collective frequency of in-phase correlated states
is distinct from those of anti-phase correlated states, com-
pare to Figs. 5a–c. For large signaling delays, the effect
of coupling on the collective frequency vanishes. We will
address the dependence of the collective frequency on the
signaling delay when studying a phase oscillator approx-
imation in Section IV.
C. Coupling enhances the quality of oscillations
Next we investigate how the precision of oscillations in
the coupled system is affected by stochastic coupling.
Fig. 6 shows the quality factor Q as a function of the
same parameters as in Fig. 5. As the delay τ increases,
the quality factor shows distinct maxima and minima of
decreasing magnitude. Maxima of the quality appear in
region where both oscillators show a high degree of in-
phase or anti-phase correlation, compare to Figs. 5a–c.
For large delays, the quality settles towards a low value
and eventually becomes independent of the delay. This
decay of the quality is due to the increase in the width σ
of the delay distribution which accompanies the increase
in τ , see Eq. (4). As the spread of the delay distribu-
tion increases, temporal information about the sending
oscillator's state is lost along the signal pathway due to
fluctuations and thus, information transmission becomes
unreliable.
The coupling strength affects the quality factor as well,
with stronger coupling leading to higher qualities: The
quality factor monotonically increases with the activation
strength β.
In the parameter range investigated here,
coupling can lead to an order of magnitude increase of
the quality compared to the uncoupled case, see Fig. 6a.
The activation threshold q has a qualitatively different
effect on Q. Quality is maximized for threshold levels q
close to the mean final product number (cid:104)xµn(cid:105) of the sig-
naling pathways and decays for smaller or larger values,
see Fig. 6b. Interestingly, since the quality depends on
the mean signaling delay τ in the same non-monotonic
way as above, islands of high quality containing local ex-
trema are observed in parameter space. The behavior of
the quality as a function of the decay rate κ as shown in
Fig. 6c suggests that in addition to the signaling delay τ ,
the decay time κ−1 of the signaling molecules effectively
contributes to the total coupling delay. This is indicated
by the leftward tilt of high quality islands for low decay
rates, see Fig. 6c. Again, for constant decay rate κ, a
non-monotonic behavior of the quality in the coupling
delay τ can be observed. Hence, we find that the non-
monotonic behavior of the quality in both the coupling
delay and the coupling strength results in multiple sep-
arated islands in parameter space that give rise to high
precision oscillations.
����������������τ/������������������τ/�β����������������τ/�β���������������τ/�κ���������������τ/�κ����������������τ/��abcdefCross correlationCollective frequencyanti-
phasein-
phaseincreasing coupling strengthdecreasing coupling strengthdecreasing coupling strength and delay 6
shows a density plot of a wavelet transform of one of
the time series, corresponding to a time-dependent pe-
riod spectrum (see Appendix B for technical details).
During the anti-phase state, the bright regions, indicat-
ing strong period components, are centered around the
white dashed line. After the transition to the in-phase
state, the bright regions fall almost entirely below the
white dashed line, indicating an increased period. This
frequency change is consistent with the frequency differ-
ences between in-phase and anti-phase correlated states,
compare to Figs. 5d–f.
Clearly, stochastic switching between different modes
of synchrony and collective frequencies including ex-
tended transient periods affects the long-time behavior
of the autocorrelation, effectively resulting in an impair-
ment of the precision. In addition, the presence of two
slightly detuned collective frequencies leads to beating
patterns in the autocorrelation function Eq. (6), so that
in these cases, the quality factor Q obtained from fits of
the autocorrelation captures the average period and an
effective correlation time (see Section II C) and can even
drop below the single-oscillator quality Q0. This impair-
ment contributes to the low quality regimes that separate
the regions of high quality observed in Fig. 6.
IV. EFFECTIVE PHASE MODEL
The chemical event chain model Eq. (1) describes how
stochastic coupling affects the collective modes and their
frequency, see Fig. 5. We aim to capture the key features
of these collective modes using a simpler theory of delay-
coupled phase oscillators. Phase oscillator models reduce
the complexity of limit cycle oscillators to the dynamics
of a phase variable φ ∈ [0, 2π) representing the state of
oscillator while neglecting the amplitude dynamics [55–
57].
A. Phase oscillators with distributed coupling
delay times
The dynamics for the phase φµ of oscillator µ = 1, 2 is
given by
(cid:90) ∞
0
FIG. 6. Coupling enhances precision. Density plots of the
quality factor Q, Eq. (7), as a function of the scaled mean
coupling delay τ /T and (a) the activation strength β, (b) the
activation threshold q, and (c) the decay rate κ of signaling
molecules. The graded bar on top of each panel indicates the
range of Q/Q0 values where Q0 = 20.8 is the quality of an
uncoupled oscillator. The parameters that are not varied are
given in Table I.
D. Stochastic switching between in-phase and
anti-phase synchrony
In the transition regions between in-phase and anti-phase
correlations (where τ ≈ (2n + 1)T /4 with integer n and
T being the period of the uncoupled oscillators), stochas-
tic switching between in-phase and anti-phase correlated
oscillations occurs in single realizations of the system.
Switching events can be observed by direct inspection of
the time series of final products, see Fig. 7a. The degree
of correlation within a single realization of the system
can be displayed by means of the windowed normalized
cross correlation,
(cid:112)c11(t)c22(t)
c12(t)
c(t) =
,
(9)
dφµ
dt
= ω + ε
g(s) sin(φ¯µ(t − s) − φµ(t)) ds ,
(11)
where
and (cid:104)(cid:104)f(cid:105)(cid:105)t = w−1(cid:82) w/2
cµν(t) = (cid:104)(cid:104)xµn xν n(cid:105)(cid:105)t − (cid:104)(cid:104)xµn(cid:105)(cid:105)t(cid:104)(cid:104)xν n(cid:105)(cid:105)t ,
(10)
−w/2 f (t + s) ds denotes a time average
over a time window with width w. Fig. 7b shows the
windowed cross correlation c for the realization shown
in A. Starting from anti-phase correlations (c ≈ −1),
the system goes through an extended phase of uncor-
related oscillations before attaining an in-phase synchro-
nized state (c ≈ 1). With the mode of synchrony, the col-
lective frequency of the system changes as well: Fig. 7c
where ω is the intrinsic frequency of the autonomous os-
cillators, ε is the coupling strength, g is the distribution of
delay times, and ¯µ denotes the respective other oscillator
as in Eq. (1). The coupling term in Eq. (11) dynamically
alters the instantaneous frequency of the oscillator de-
pending on the phase relationship to the other oscillator.
For the distribution g of delay times, we here choose the
Gamma distribution Eq. (3) that describes the distribu-
tion of arrival times of signaling molecules.
We now show that Eq. (11) can describe many qualita-
tive features of the in-phase and anti-phase synchronized
����������������τ/�β���������������τ/�κ����������������τ/��abcQuality factor 7
FIG. 7. Stochastic switching between anti-phase and in-phase synchronized oscillations. (a) Time series of products x1n and
x2n showing a switching event between anti-phase and in-phase oscillations. Parameters are given in Table I except for β = 120
and λ = 5/3. (b) Windowed cross correlation c of the time series in A, see Eq. (9), showing the three distinct regimes. The
width of the averaging window is w = 4T . (c) Wavelet scalogram of one of the time series in A. Bright regions indicate strong
period components. The white dashed line serves as a guide for the eye.
corresponds to the limit n → ∞ with λ = n/τ and fixed
τ and in this case, Eq. (12) becomes
Ω = ω − ε sin(Ωτ ) ,
(14)
states observed in the stochastic theory. The in-phase
synchronized state, given by φµ(t) = Ωt, is characterized
by both oscillators evolving with the same collective fre-
quency Ω and having no phase lag relative to each other.
Using this ansatz in Eq. (11) yields an implicit transcen-
dental equation for Ω,
(cid:18)
(cid:19)n/2
(cid:18)
(cid:19)
Ω
λ
Ω = ω − ε
1
1 + Ω2/λ2
sin
n arctan
.
(12)
a result well-known in the literature [42, 53, 54, 58].
The system can also exhibit an anti-phase synchronized
state, φ1(t) = ¯Ωt = φ2(t) + π, where the corresponding
collective frequency ¯Ω obeys Eq. (12) with ε replaced by
−ε. In both cases, the collective frequency satisfies the
bound
ω − ε ≤ Ω ≤ ω + ε ,
(13)
implying that Ω can only deviate from the intrinsic fre-
quency ω by the coupling strength ε. Moreover, we find
that Ω → ω for λ → 0, implying that for increasing delays
due to an increased jump rate, the collective frequency
becomes independent of coupling, a behavior that we also
found in the chemical event chain model, see Fig. 5d–f.
The special case of a single discrete delay, g(t) = δ(t− τ ),
Fig. 8 shows the collective frequencies for systems with
a distribution of delays, Eq. (12), and for systems with
a discrete coupling delay, Eq. (14). In both cases, for a
given set of parameters, this equation can exhibit multi-
ple solutions in Ω. However, compared to a discrete delay,
a distribution of delay times leads to a decaying depen-
dence of the collective frequency on the coupling delay if
the number of steps n is fixed. Using a standard linear
stability analysis, an analytical criterion for the stability
of the in-phase and anti-phase synchronized states can be
found, see Appendix C. In Fig. 8, stable states are indi-
cated by solid curves, unstable states by dashed curves.
This also illustrates that in-phase and anti-phase solu-
tions can be simultaneously stable in certain parameter
regions.
������������������μ�∼���������-������������� 8
of the event chain model is not obvious. For simplicity,
we here assume that ε scales linearly with the activation
strength β in the chemical event chain model (see also
Section II B) and fix the ratio r = ε/β by hand.
We assess whether the synchronized states of both
models agree by comparing the frequency and stabil-
ity solutions of the phase model to periodograms of the
chemical event chain system [59]. Figs. 9a–d show such
periodograms for different activation strengths β, with
bright regions indicating strong frequency components.
Figs. 9e–h show the same density plots, superimposed
with solutions for the collective frequency Ω, Eq. (12).
The dominant frequency components exhibit character-
istic jumps at delays being odd multiples of T /4, where T
is the uncoupled period, as already observed earlier, com-
pare to Figs. 5d–f. This is expected because of stochas-
tic switching between in-phase and anti-phase synchrony
with different frequencies, see Section III D. Moreover, as
the signaling delay τ increases, the dominant frequency
components approach the intrinsic frequency of the un-
coupled oscillator, a behavior that the phase model cap-
tures as described in the previous section and Fig. 8a.
Interestingly, the phase model exhibits regions in which
the in-phase and anti-phase solution are simultaneously
stable. This implies that sufficiently strong fluctuations
can drive the system out of one synchronized state into
the basin of attraction of the other, consistent with
the stochastic switching between in-phase and anti-phase
synchrony found in the chemical event chain model. To
obtain a proxy for the size of the basins of attraction of
the two states in the phase oscillator model, we numer-
ically solve the deterministic Eq. (11) with a constant
phase difference ∆φ between the two oscillators as an
initial history, (φ1 − φ2)t≤0 = ∆φ, and monitor their
long-time phase difference to determine their final state
[60]. The region plots in Fig. 8 display these the final
states depending on ∆φ and illustrate how the relative
size of such basins change as the mean signaling delay τ
is varied.
V. DISCUSSION
Considering two coupled genetic oscillators, we have
shown how stochastic coupling by signaling chains affects
their frequency and quality and promotes synchroniza-
tion. An important feature of the chemical event chain
framework is that it naturally accounts for distributed
signaling and transcriptional delays that are a conse-
quence of the sequences of chemical steps. These dis-
tributed signaling delays have profound consequences for
oscillator dynamics and fluctuations that cannot be cap-
tured by simplified descriptions such as rate equations or
coupled oscillators with a discrete time delay. In particu-
lar, we found that synchrony and quality are maximized
in isolated islands in parameter space characterizing cou-
pling delay and coupling strength. Moreover, noisy cou-
pling can lead to stochastic switching between in-phase
FIG. 8. Collective frequency Ω of the phase oscillator system
Eq. (11) (curves) and its asymptotic state for given constant
initial phase differences ∆φ (region plots) as a function of
the scaled coupling delay τ /T (see Section IV B). Blue curves
show the in-phase synchronized state, red curves show the
anti-phase synchronized state. Solid curves show stable solu-
tions, dashed curves show unstable solutions. (a) System with
a distribution of delay times for n = 10 and different mean
delays τ obtained by varying λ. The collective frequency is
determined by Eq. (12). (b) Collective frequency for a dis-
crete delay time τ , determined by Eq. (14).
In both plots,
ω = 0.224 and ε = ω/4.
B. Comparison of phase oscillator model and
chemical event chain model
We now show that the phase oscillator model Eq. (11) can
capture the key features of the collective modes described
by the chemical event chain model Eq. (1). We compare
the collective frequency Ω, Eq. (12), obtained from the
phase model to the frequency spectrum of oscillations
from the chemical event chain model. For the distribu-
tion g of delay times in the phase model, we adopt the
parameter values of λ and n used in the chemical event
chain model. For the intrinsic frequency ω in the phase
model, we use an estimate provided in Ref. [6] for a sin-
gle uncoupled oscillator of the same type as investigated
here,
ω (cid:39)
π
nλ−1 + κ−1
.
(15)
The coupling strength ε is the only parameter in the
phase model whose relationship to the kinetic parameters
��/���/���πτ/�Δϕ�πΔϕω-ϵωω+ϵΩω-ϵωω+ϵΩabin-phaseanti-phase 9
FIG. 9. Collective frequency as a function of the scaled coupling delay τ /T for different activation strengths β: (a, e) β = 30, (b,
f) β = 60, (c, g) β = 90, (d, h) β = 120. (a–d) The density plots show logarithmic periodograms of oscillations in the chemical
event chain model, where bright regions correspond to strong frequency components. (e–h) The curves (superimposed on the
same density plots as in a–d) show the collective frequencies of the in-phase (black) and anti-phase (white) solutions of the
phase model, Eq. (12). Solid lines show stable solutions, dashed lines show unstable solutions. The delay is given in multiples
of the uncoupled period T , the collective frequency Ω is given in multiples of the uncoupled frequency ω = 2π/T . Solid lines
indicate stable solutions, dashed lines indicate unstable solutions. The parameters for the chemical event chain model are given
in Table I. The parameters for the phase model are adopted from the chemical event chain model with ω ≈ 0.224 estimated
via Eq. (15) and ε = rβ with r = 1.33 × 10−4.
and anti-phase states, a behavior also found in other cou-
pled noisy oscillators, e.g., Hodgkin–Huxley neurons [61]
and delay-coupled phase oscillators [62]. Our findings
may shed light on the operating regime of cellular ge-
netic oscillator systems in which precise timing is vital,
such as the circadian clock [10] and the vertebrate seg-
mentation clock [19]. The Delta–Notch signaling path-
way may provide an experimental system where in-phase
and anti-phase oscillations have been observed in the
context of the segmentation clock and neurogenesis re-
spectively [17, 44]. Stochastic mode switching could be
explored in this system using synthetic approaches [63]
and optogenetic perturbations [47]. Here we have cho-
sen intercellular coupling as an example, but coupled
genetic oscillators also occur within cells [64], or, on a
coarse-grained level, as coupled subpopulations of oscil-
lators, such as different regions of the mammalian cir-
cadian clock [65, 66], for which our modeling framework
can be adapted as well. Moreover, using a unidirectional
signaling mechanism, effects of stochastic signaling on en-
trainment to external signals can be studied, an aspect
relevant for circadian clock research [14, 48, 67].
Key effects of distributed signaling delays that result
from chemical chains can be captured by an effective
phase oscillator model. This phase model can also be
extended to include noise which enables to study preci-
sion of collective oscillations in a simplified picture.
Our results demonstrate the interplay of stochastic-
ity and nonlinear effects in genetic regulatory networks
containing chemical event chains. It extends existing ap-
proaches to represent fluctuations in biochemical systems
and captures the statistics of non-equilibrium noise that
arises in such chemical processes. This approach is not
limited to oscillatory systems investigated here but can
also be applied to other genetic feedback systems such
as homeostatic systems, switches, and feed-forward cas-
cades.
ACKNOWLEDGMENTS
We thank Andy Oates for inspiring discussions. LGM
acknowledges support from ANPCyT PICT 2012 1954,
PICT 2013 1301 and FOCEM Mercorsur (COF 03/11).
abcdefgh 10
where ¯µ refers to the respective other oscillator as in
Eq. (1) and where we have defined the delayed variable
µ (t) = ξµ(t − s). We decouple the dynamics by defin-
ξ(s)
ing the collective modes ϕk = ξ1 + kξ2 with k = +1,−1.
Inverting this definition yields ξ1 = (ϕ+ + ϕ−)/2 and
ξ2 = (ϕ+ − ϕ−)/2, which shows that exciting the collec-
tive mode ϕ+ shifts both oscillators by the same amount
and thus corresponds to a global phase shift, whereas ϕ−
is the phase difference between both oscillators. The dy-
namics of these collective modes are given by
g(s) cos(Ωs)(kϕ(s)
k − ϕk) ds .
(C2)
The characteristic equation for these modes is obtained
using the exponential ansatz ϕk(t) = eηkt with ηk be-
ing complex. The sign of Re ηk then determines whether
perturbations decay (Re ηk < 0) or grow (Re ηk > 0) and
thus whether the synchronized state is stable or unsta-
ble [71]. Using this ansatz in Eq. (C2), we obtain
(cid:90) ∞
0
1
ε
dϕk
dt
=
(cid:90) ∞
0
Appendix A: Stochastic simulations
Direct numerical solutions of the master equation,
Eq. (1), is impracticable due to the high dimensionality
of the state space.
Instead, a stochastic simulation al-
gorithm of the Gillespie-type has been used to compute
exact realizations of trajectories of the model [68]. Ex-
pectation values were obtained by computing averages of
the respective observable over multiple realizations. The
data shown in Figs. 5 and 6 were obtained by averaging
over 50 realizations of duration 80 000 units of time for
each data point.
Appendix B: Wavelet transform
The continuous wavelet transform of a discrete time series
(x1, . . . , xm) sampled with time interval δt is defined by
[69]
xjΨ∗(cid:18) j − k
(cid:19)
s
m(cid:88)
j=1
W (s, k) =
1√
s
where s is the wavelet scale. We here choose the Gabor
wavelet function, given by Ψ(t) = π−1/4e6it−t2/2. For
the Gabor wavelet, the scale s corresponds to a period
of T (s) ≈ 1.033s · δt. The wavelet scalogram in Fig. 7c
displays the squared magnitude W (s, k)2 as a density
plot, where the abscissa shows time t = k · δt and the
ordinate shows the period T = T (s).
Appendix C: Stability analysis of the synchronized
state in the phase model
To assess the stability of the in-phase synchronized state
φµ(t) = Ωt, we linearize the dynamics around this
state [70]. We use the standard ansatz φµ(t) = Ωt + ξµ(t)
in Eq. (11), where ξµ is a small perturbation. We ob-
tain the time evolution of the perturbation by expanding
Eq. (11) to first order in ξ, which yields
g(s) cos(Ωs)(ξ(s)
¯µ − ξµ) ds ,
(C1)
(cid:90) ∞
0
1
ε
dξµ
dt
=
,
(B1)
ηk
ε
=
g(s) cos(Ωs)(ke−ηks − 1) ds .
(C3)
Since the Gamma distribution g, Eq. (3), decays as e−λs,
the integral on the rhs of Eq. (C3) is only well-defined
if Re ηk > −λ. In this case, the integral can be solved
analytically and the resulting characteristic equation is
kΓ(ηk) − ηk/ε = Γ(0) ,
(C4)
(cid:20)
where
Γ(η) =
λn
2
1
(λ + η + iΩ)n +
1
(λ + η − iΩ)n
(cid:21)
.
(C5)
In general, Eq. (C4) can have multiple solutions in ηk.
The synchronized state is linearly stable if and only if
Re ηk < 0 holds for all solutions ηk to Eq. (C3) for both
k = +1 and k = −1. The stability of the anti-phase syn-
chronized state is determined by Eq. (C4) with ε replaced
by −ε. For Figs. 8 and 9, we determine the stability of a
given synchronized state by solving Eq. (C4) numerically
and determining the sign of the solution with largest real
part.
[1] U. Alon, An Introduction to Systems Biology: Design
Principles of Biological Circuits (Chapman & Hall/CRC
Press, Boca Raton, Florida, 2006).
[2] A. Goldbeter, Biochemical Oscillations and Cellular
Rhythms: The Molecular Bases of Periodic and Chaotic
Behaviour (Cambridge University Press, 1997).
[3] B. Alberts, A. Johnson, J. Lewis, M. Raff, K. Roberts,
and P. Walter, Molecular Biology of the Cell, 4th ed.
(Garland Science, New York, 2002).
[4] M. X. G. Ilagan and R. Kopan, "SnapShot: Notch sig-
naling pathway," Cell 128, 1246 (2007).
[5] B. Nov´ak and J. J. Tyson, "Design principles of biochem-
ical oscillators," Nat. Rev. Mol. Cell Biol. 9, 981–991
(2008).
[6] L. G. Morelli and F. Julicher, "Precision of genetic oscil-
lators and clocks," Phys. Rev. Lett. 98, 228101 (2007).
[7] P. E. Hardin, J. C. Hall, and M. Rosbash, "Feedback
of the Drosophila period gene product on circadian cy-
cling of its messenger RNA levels." Nature 343, 536–540
(1990).
11
[8] J. C. Dunlap, "Molecular bases for circadian clocks," Cell
[27] J. Kim and E. Winfree, "Synthetic in vitro transcrip-
96, 271–290 (1999).
[9] P. Smolen, D. A. Baxter, and J. H. Byrne, "A Reduced
Model Clarifies the Role of Feedback Loops and Time
Delays in the Drosophila Circadian Oscillator," Biophys.
J. 83, 2349–2359 (2002).
[10] U. Schibler and F. Naef, "Cellular oscillators: Rhythmic
gene expression and metabolism," Curr. Opin. Cell Biol.
17, 223–229 (2005).
[11] M. Hastings, J. S. O'Neill, and E. S. Maywood, "Cir-
regulators of endocrine and metabolic
cadian clocks:
rhythms," J. Endocrinol. 195, 187–198 (2007).
[12] D. Zwicker, D. K. Lubensky, and P. R. ten Wolde, "Ro-
bust circadian clocks from coupled protein-modification
and transcription–translation cycles," Proc. Natl. Acad.
Sci. USA 107 (2010).
[13] E. E. Zhang and S. A. Kay, "Clocks not winding down:
unravelling circadian networks," Nat. Rev. Mol. Cell Biol.
11, 764–776 (2010).
[14] U. Abraham, A. E. Granada, P. O. Westermark,
M. Heine, H. Herzel, and A. Kramer, "Coupling gov-
erns entrainment range of circadian clocks," Mol. Sys.
Biol. 6, 438 (2010).
[15] A. E. Granada, G. Bordyugov, A. Kramer,
and
H. Herzel, "Human Chronotypes from a Theoretical Per-
spective," PLOS ONE 8, e59464 (2013).
[16] I. Imayoshi and R. Kageyama, "bHLH Factors in Self-
Renewal, Multipotency, and Fate Choice of Neural Pro-
genitor Cells," Neuron 82, 9–23 (2014).
[17] H. Shimojo and R. Kageyama, "Oscillatory control of
Delta-like1 in somitogenesis and neurogenesis: A uni-
fied model for different oscillatory dynamics," Semin. Cell
Dev. Biol. 49, 76–82 (2016).
[18] D. J. Jorg, L. G. Morelli, D. Soroldoni, A. C. Oates,
and F. Julicher, "Continuum theory of gene expression
waves during vertebrate segmentation," New J. Phys. 17,
093042 (2015).
[19] A. C. Oates, L. G. Morelli, and S. Ares, "Patterning em-
bryos with oscillations: structure, function and dynamics
of the vertebrate segmentation clock," Development 139,
625–639 (2012).
[20] H. Niederholtmeyer, Z. Z. Sun, Y. Hori, E. Yeung, A. Ver-
poorte, R. M. Murray, and S. J. Maerkl, "Rapid cell-
free forward engineering of novel genetic ring oscillators,"
eLife 4, e09771 (2015).
[21] I. T. Tokuda, D. Ono, B. Ananthasubramaniam,
S. Honma, K.-I. Honma, and H. Herzel, "Coupling Con-
trols the Synchrony of Clock Cells in Development and
Knockouts," Biophysical Journal 109, 2159 (2015).
[22] D. J. Jorg, "Amplitude bounds for biochemical oscilla-
tors," EPL (Europhys. Lett.) 119, 58004 (2017).
[23] M. B. Elowitz and S. Leibler, "A synthetic oscillatory
network of transcriptional regulators," Nature 403, 335–
338 (2000).
[24] J. Garcia-Ojalvo, M. B. Elowitz, and S. H. Strogatz,
"Modeling a synthetic multicellular clock: repressilators
coupled by quorum sensing." Proc. Natl. Acad. Sci. USA
101, 10955 (2004).
[25] J. Stricker, S. Cookson, M. R. Bennett, W. H. Mather,
L. S. Tsimring, and J. Hasty, "A fast, robust and tunable
synthetic gene oscillator," Nature 456, 516–519 (2008).
[26] T. Danino, O. Mondrag´on-Palomino, L. Tsimring, and
J. Hasty, "A synchronized quorum of genetic clocks," Na-
ture 463, 326–330 (2010).
tional oscillators," Mol. Sys. Biol. 7, 465–465 (2011).
[28] L. Potvin-Trottier, N. D. Lord, G. Vinnicombe,
and
J. Paulsson, "Synchronous long-term oscillations in a
synthetic gene circuit." Nature 538, 514 (2016).
[29] A. M. Tayar, E. Karzbrun, V. Noireaux, and R. H. Bar-
Ziv, "Synchrony and pattern formation of coupled genetic
oscillators on a chip of artificial cells." Proc. Natl. Acad.
Sci. USA 114, 11609–11614 (2017).
[30] N. Barkai and S. Leibler, "Circadian clocks limited by
noise," Nature 403, 267–268 (2000).
[31] J. M. G. Vilar, H. Y. Kueh, N. Barkai, and S. Leibler,
"Mechanisms of noise-resistance in genetic oscillators,"
Proc. Natl. Acad. Sci. USA 99, 5988–5992 (2002).
[32] D. Gonze, J. Halloy,
and P. Gaspard, "Biochemical
clocks and molecular noise: Theoretical study of robust-
ness factors," J. Chem. Phys. 116, 10997–11010 (2002).
[33] I. Mihalcescu, W. Hsing, and S. Leibler, "Resilient cir-
cadian oscillator revealed in individual cyanobacteria,"
Nature 430, 81–85 (2004).
[34] D. A. Potoyan and P. G. Wolynes, "On the dephasing
of genetic oscillators," Proc. Natl. Acad. Sci. USA 111,
2391–2396 (2014).
[35] A. B. Webb, I. M. Lengyel, D. J. Jorg, G. Valentin,
F. Julicher, L. G. Morelli, and A. C. Oates, "Persis-
tence, period and precision of autonomous cellular oscil-
lators from the zebrafish segmentation clock," eLife 5,
e08438 (2016).
[36] I. M. Lengyel and L. G. Morelli, "Multiple binding sites
for transcriptional repressors can produce regular burst-
ing and enhance noise suppression," Phys. Rev. E 95,
042412 (2017).
[37] J. Lewis, "Autoinhibition with transcriptional delay: A
simple mechanism for the zebrafish somitogenesis oscil-
lator." Curr. Biol. 13, 1398–1408 (2003).
[38] A. C. Liu, D. K. Welsh, C. H. Ko, H. G. Tran, E. E.
Zhang, A. A. Priest, E. D. Buhr, O. Singer, K. Meeker,
I. M. Verma, F. J. Doyle III, J. S. Takahashi, and S. A.
Kay, "Intercellular Coupling Confers Robustness against
Mutations in the SCN Circadian Clock Network," Cell
129, 605–616 (2007).
[39] D. J. Needleman, P. H. E. Tiesinga, and T. J. Sejnowski,
"Collective enhancement of precision in networks of cou-
pled oscillators," Physica D: Nonlinear Phenomena 155,
324–336 (2001).
[40] L. G. Morelli, S. Ares, L. Herrgen, C. Schroter,
F. Julicher, and A. C. Oates, "Delayed coupling theory
of vertebrate segmentation," HFSP J. 3, 55–66 (2009).
[41] L. Herrgen, S. Ares, L. G. Morelli, C. Schroter,
F. Julicher, and A. C. Oates, "Intercellular coupling reg-
ulates the period of the segmentation clock," Curr. Biol.
20, 1244–1253 (2010).
[42] S. Ares, L. G. Morelli, D. J. Jorg, A. C. Oates, and
F. Julicher, "Collective modes of coupled phase oscil-
lators with delayed coupling," Phys. Rev. Lett. 108,
204101 (2012).
[43] M. C. Cross, "Improving the frequency precision of os-
cillators by synchronization," Phys. Rev. E 85, 046214
(2012).
[44] H. Shimojo, A. Isomura, T. Ohtsuka, H. Kori, H. Miy-
achi, and R. Kageyama, "Oscillatory control of Delta-
like1 in cell interactions regulates dynamic gene expres-
sion and tissue morphogenesis," Genes Dev. 30, 102–116
(2016).
[45] B.-K. Liao, D. J. Jorg, and A. C. Oates, "Faster embry-
onic segmentation through elevated Delta-Notch signal-
ing," Nat. Commun. 7, 11861 (2016).
[46] D. J. Jorg, "Stochastic Kuramoto oscillators with discrete
phase states," Phys. Rev. E 96, 032201 (2017).
[47] A. Isomura, F. Ogushi, H. Kori, and R. Kageyama, "Op-
togenetic perturbation and bioluminescence imaging to
analyze cell-to-cell transfer of oscillatory information."
Genes Dev. 31, 524–535 (2017).
[48] B. Ananthasubramaniam, E. D. Herzog, and H. Herzel,
"Timing of neuropeptide coupling determines synchrony
and entrainment in the mammalian circadian clock."
PLoS Comp. Biol. 10, e1003565 (2014).
[49] J. Lewis, A. Hanisch, and M. Holder, "Notch signaling,
the segmentation clock, and the patterning of vertebrate
somites," J. Biol. 8, 44 (2009).
[50] C. Gardiner, Stochastic Methods: A Handbook for the
Natural and Social Sciences, Springer Series in Synerget-
ics (Springer, 2009).
[51] N. G. van Kampen, Stochastic Processes in Physics and
Chemistry, 3rd ed. (Elsevier, 2007).
[52] R. L. Stratonovich, Topics in the Theory of Random
Noise, Mathematics and its applications No. Bd. 1 (Gor-
don and Breach, 1963).
[53] H. G. Schuster and P. Wagner, "Mutual entrainment of
two limit cycle oscillators with time delayed coupling."
Prog. Theor. Phys. 81, 939–945 (1989).
[54] M. K. S. Yeung and S. H. Strogatz, "Time delay in the
Kuramoto model of coupled oscillators." Phys. Rev. Lett.
82, 648–651 (1999).
[55] Y. Kuramoto, "Cooperative dynamics of oscillator com-
munity: a study based on lattice of rings," Prog. Theor.
Phys. 79, 223–240 (1984).
[56] J. A. Acebr´on, L. L. Bonilla, and C. J. P. Vicente, "The
Kuramoto model: A simple paradigm for synchronization
phenomena," Rev. Mod. Phys. 77, 137 (2005).
[57] F. A. Rodrigues, T. K. DM. Peron, P. Ji, and J. Kurths,
"The Kuramoto model in complex networks," Phys. Rep.
610, 1–98 (2016).
[58] M. G. Earl and S. H. Strogatz, "Synchronization in os-
cillator networks with delayed coupling: A stability cri-
terion." Phys. Rev. E 67, 036204 (2003).
[59] The logarithmic periodogram P of a time series
(x1, . . . , xm) is given by Pω = 2 lnxω, where xω is the
discrete Fourier transform of xk.
12
[60] Note that the phase model Eq. (11) is an infinite-
dimensional system due to the presence of delays and
therefore requires an entire phase history φi(t)t≤0 as an
initial condition [53]. Therefore, the final state of the sys-
tem may depend on the entire time dependence of the
initial history. For simplicity, we here only focus on con-
stant initial conditions.
[61] X. Ao, P. Hanggi,
and G. Schmid, "In-phase and
anti-phase synchronization in noisy Hodgkin-Huxley neu-
rons," Math. Biosci. 245, 49–55 (2013).
[62] O. D'Huys, Th. Jungling, and W. Kinzel, "Stochastic
switching in delay-coupled oscillators," Phys. Rev. E 90,
032918–9 (2014).
[63] M. Matsuda, M. Koga, K. Woltjen, E. Nishida,
and
M. Ebisuya, "Synthetic lateral inhibition governs cell-
type bifurcation with robust ratios," Nat. Commun. 6,
6195 (2015).
[64] C. Schroter, S. Ares, L. G. Morelli, A. Isakova, K. Hens,
D. Soroldoni, M. Gajewski, F. Julicher, S. J. Maerkl,
B. Deplancke, and A. C. Oates, "Topology and dynamics
of the zebrafish segmentation clock core circuit," PLoS
Biology 10, e1001364 (2012).
[65] S. Bernard, D. Gonze, B. Cajavec, H. Herzel,
and
A. Kramer, "Synchronization-Induced Rhythmicity of
Circadian Oscillators in the Suprachiasmatic Nucleus,"
PLoS Comp. Biol. 3, e68 (2007).
[66] A. Azzi, J. A. Evans, T. Leise, J. Myung, T. Takumi,
A. J. Davidson, and S. A. Brown, "Network Dynam-
ics Mediate Circadian Clock Plasticity," Neuron 93, 441
(2017).
[67] A. Erzberger, G. Hampp, A. E. Granada, U. Albrecht,
and H. Herzel, "Genetic redundancy strengthens the cir-
cadian clock leading to a narrow entrainment range," J.
Royal Soc. Interface 10, 20130221 (2013).
[68] D. Gillespie, "Exact stochastic simulation of coupled
chemical-reactions," J. Phys. Chem. 81, 2340–2361
(1977).
[69] C. Torrence and G. P. Compo, "A practical guide to
wavelet analysis," B. Am. Meteorol. Soc. 79, 61–78
(1998).
[70] S. H. Strogatz, Nonlinear Dynamics and Chaos
(Addison-Wesley, Reading, MA, 1994).
[71] A. Amann, E. Scholl, and W. Just, "Some basic remarks
on eigenmode expansions of time-delay dynamics," Phys-
ica A 373, 191–202 (2007).
|
1706.02260 | 1 | 1706 | 2017-05-24T19:20:25 | Apex predator and the cyclic competition in a rock-paper-scissors game of three species | [
"physics.bio-ph",
"cond-mat.stat-mech"
] | This work deals with the effects of an apex predator on the cyclic competition among three distinct species that follow the rules of the rock-paper-scissors game. The investigation develops standard stochastic simulations but is motivated by a novel procedure which is explained in the work. We add the apex predator as the fourth species in the system that contains three species that evolve following the standard rules of migration, reproduction and predation, and study how the system evolves in this new environment, in comparison with the case in the absence of the apex predator. The results show that the apex predator engenders the tendency to spread uniformly in the lattice, contributing to destroy the spiral patterns, keeping biodiversity but diminishing the average size of the clusters of the species that compete cyclically. | physics.bio-ph | physics |
Apex predator and the cyclic competition in a
rock-paper-scissors game of three species
C. A. Souza-Filho,1, 2 D. Bazeia,1 and J. G. G. S. Ramos1
1Departamento de Física, Universidade Federal da Paraíba, 58051-970, João Pessoa, PB, Brazil
2Instituto Federal de Educação, Ciência e Tecnologia da Paraíba,
Campus Princesa Isabel, 58755-000, Princesa Isabel, PB, Brazil
This work deals with the effects of an apex predator on the cyclic competition among three distinct
species that follow the rules of the rock-paper-scissors game. The investigation develops standard
stochastic simulations but is motivated by a novel procedure which is explained in the work. We
add the apex predator as the fourth species in the system that contains three species that evolve
following the standard rules of migration, reproduction and predation, and study how the system
evolves in this new environment, in comparison with the case in the absence of the apex predator.
The results show that the apex predator engenders the tendency to spread uniformly in the lattice,
contributing to destroy the spiral patterns, keeping biodiversity but diminishing the average size of
the clusters of the species that compete cyclically.
PACS numbers: 87.10.Mn, 87.23.Cc
I.
INTRODUCTION
An intriguing problem in Biology concerns the un-
derstanding of how distinct species interact to maintain
the mechanisms underlying biodiversity in nature. Sev-
eral models focusing on the competing relations among
species have been proposed and studied in the last few
decades [1–3]. In the simple case where the species com-
pete for a single and restricted resource, their abilities
appear to be hierarchical, involving a transitive relation-
ship. In this case, one expects a winner species, leading
all the other species to extinction [4, 5]. However, there
are other possibilities and when resource is abundant one
may observe intransitive competing relationship, as it
happens if one uses the rules of the rock-paper-scissors
(rps) game, for instance. In this case, when one considers
the system with three species, individuals of the species
A predate those of the species B, B predates C and C
predates A in a cyclic competition environment. In this
rps dynamics, all the species are treated equally and the
system is known to lead to biodiversity [6–11].
On the other hand, apex predators are being described
as highly interactive from the biological point of view and
their importance in the ecological environment has been
the focus of several investigations [11–15]. The presence
of an apex predator in a given ecosystem may favor co-
existence of species, since it can diminish the process of
competitive exclusion, imposing its own order to the set
of species. This is known as predator-mediated coexis-
tence [16] and has been identified in several distinct en-
vironments, as in coral reef communities [17–19], in com-
munities of birds [20], in vegetational diversity [21–23],
and in other scenarios.
Recent works also emphasize the use of the apex preda-
tor to restore ecosystems [24] that have been weakened
due to distinct causes and motivations.
In particular,
one has identified secondary extinctions that could be
maintained or restored in order to minimize damage or
improve performance, as in the case of invasion of non
native species [25, 26], the transmission of diseases [27]
or the effects of climate changes on the dynamics of the
species [28]. Such effects, which are triggered by the
presence or action of the apex predator and may influ-
ence other levels of the chain, contributes to the so-called
trophic cascade [29]. Distinct factors such as changes in
prey cascade behavior as a survival strategy and as pre-
dation that decreases the abundance of specific prey and
interfere at distinc levels of the chain, may induce the
occurrence of the trophic cascade.
In this work we investigate systems with three and four
species and select three systems, one with three species
that compete cyclically and two with four species, one
with the four species competing cyclically and the other
with the fourth species representing the apex predator,
which predates all the other three species and is not pre-
dated by any of them. We focus mainly on the behavior
of the apex predator and how it changes the behavior
of the other species in the stochastic evolutions in the
square lattice which we consider below. As we have com-
mented above, the apex predador has been studied in
several distinct scenarios, but the idea to be pursued in
this work is original and of current interest. It develops
standard stochastic simulations and was motivated by
the recent work [30], which makes use of the density of
maxima related to the abundance of the species, and by
the algorithm developed in Ref. [31], which allows that
we explore the clustering of species in the square lattice.
The work is organized as follows. In the next section
we introduce the four systems and study some specific
time evolutions to show how they evolve under the rules
there discussed. We move on and in Sec. III one selects
and investigates three distinct systems, with the results of
the stochastic simulations adding novelties to the current
understanding of the behavior of the apex predator and
how it changes the behavior of the other three species
that compete cyclically. We end the work in Sec. IV,
where we include some comments and conclusions.
II. THE MODELS
In this work we investigate systems described by three
and four species.
In Fig. 1 one illustrates the systems
and the interactions among the species, with the arrows
going from the predator to the prey. In Fig. 1 (a) one
shows three distinct species that compete cyclically, as
in the rock-paper-scissors game. This is the system X3
and several studies were already implemented; see, e.g.,
Refs. [32–36]. Among several interesting characteristics,
one notes the presence of spiral patterns. In Fig. 1 (b)
one shows another system with four distinct species that
also interact cyclically, but in this case the fourth species
adds the next-to-next neighbor which does not compete
and so form partnerships. This and other similar cases
were studied in [37–41] and may generate patterns as the
one shown in the figure, with two subsets of two partner
species. In Fig. 1 (c) and (d) we show the systems X4 and
SX3, respectively. They have all the four species inter-
acting, but in (c) they are all equivalent, and in (d) one
selects the yellow species to represent the apex predator.
In order to implement the stochastic simulations, we
consider a square lattice with N = L2 sites and use pe-
riodic boundary conditions. We take L = 512, so we
deal with a square lattice of 512 × 512 sites. Each site in
the lattice is occupied by one of the species A (red), B
(green), C (blue) or D (yellow), or is empty E (black).
The interactions follows the rules [8, 9]:
BC
σ
µ
AB
σ
µ
CA
σ
µ
CE
C(cid:3) ε
AE
A(cid:3) ε
BE
B(cid:3) ε
−→ AE,
−→ AA,
−→ (cid:3)A,
−→ CE,
−→ CC,
−→ (cid:3)C,
−→ BE,
−→ BB,
−→ (cid:3)B,
(1)
(2)
(3)
where (cid:3) represents a site that can be empty or occu-
pied by any individual. The relations in (1) describe pre-
dation, which is characterized by the σ parameter that
shows the cyclic interactions. The relations in (2) and (3)
show reproduction and migration, which occur controlled
by the µ and ε parameters, respectively.
In the last system in Fig. 1 (d), the D or yellow species
interacts obeying the following rules
γ
DX
D (cid:3) β
D (cid:3)
−→ DD,
−→ E (cid:3),
−→ (cid:3) D,
(4)
(5)
(6)
where X represents one among the three species A, B,
or C that compete cyclically. These rules show that the
D or yellow species describes the apex predator. The
relation (4) ensures that it can reproduces after predating
any of the three species under the same ratio γ, and the
relations (5) and (6) describe dead and migration, which
are controlled by the parameter β and , respectively.
Before starting the simulation one prepares the initial
state, in which all the species and empty sites are evenly
distributed in the square lattice with the same proba-
bility. Each time step randomly selects a site and one
2
(a)
(b)
(c)
(d)
FIG. 1: In the left panel one displays the diagrams to show
how predation works for the three and four species. The black
arrows indicate unidirectional predation and the gray dashed
arrows indicate bidirectional predation. In the right panel one
displays typical snapshots of stochastic simulations that run
for 104 generations. In (d) one shows the system where the
fourth species constitutes the apex predator.
of its four nearest neighbors, and for each selected pair,
the random process continues using the normalized ra-
tios controlled by σ, µ, ε, γ, β, and , as described by the
above processes (1)-(6). To describe the time evolution
of the system, we use generation, which is the time spent
to account for N time steps.
In the right panels in Fig. 1 (a) and (c), one sees the ap-
3
pearance of spiral patterns, which is typical of the cyclic
evolutions that follow the rules of the rock-paper-scissors
game. However, in Fig. 1 (d) one notices the absence of
spirals and the diminishing of the size of the clusters of
the species that evolve in cyclic competition. This behav-
ior is due to the presence of the apex predator, and we
study it below, showing that although the apex predator
does not destroy biodiversity, it diminishes the average
size of the clusters of species that compete cyclically. We
have added in https://goo.gl/QexPxf three videos to
illustrate how the three systems X3, X4, and SX3 evolve
in time under the stochastic simulations.
III. RESULTS
Let us now implement the stochastic simulations, tak-
ing the parameters for predation, reproduction and mi-
gration in Fig. 1 (a) and (c) as σ = 0.25, µ = 0.25 and
ε = 0.50, respectively. For the system with the apex
predator which appears in Fig. 1 (d), one uses σ = 0.30,
µ = 0.30, ε = 0.40, γ = 0.25, β = 0.15 and = 0.60.
We have checked that the results in this work are robust
against changes in the values of the parameters, if they
are chosen in a way that maintain coexistence among the
species. The main focus of the current work is to investi-
gate the systems shown in Fig. 1 (a), (c), and (d), which
we refer to as X3, X4, and SX3, respectively.
In Fig. 2 one displays the density of species, ρx(t),
with x ∈ {A, B, C, D}, as a function of the generation
time. The colored curves represent the corresponding
species, and the black curve identify the empty sites. The
cases displayed in Fig. 2 (a), (b) and (c) represent the
systems X3, X4 and SX3 respectively, and one notices
that after a transient time interval, all the systems evolve
maintaining coexistence of the species.
One sees that in the absence of the apex predator, the
density of empty sites diminishes rapidly. This happens
because the initial density of species and empty sites fa-
vor the reproduction, since the species can frequently
meet empty sites. In the presence of the apex predator,
the density of empty sites increases before diminishing
rapidly, and now the reason is that the presence of empty
sites contributes to the death of the apex predators, be-
fore they starts to increase to reach an equilibrium state
that oscillates around a given average.
In all the above cases, the density of species oscillates
around an average value. This behavior is expected,
since the systems are similar to the famous predator-
prey model of Lotka-Volterra [42, 43]; see, e.g., Fig. 6
of Ref. [44], which shows similar mean field and stochas-
tic network simulations in a 5 species model in the ab-
sence of the apex predator. However, in the presence
of the apex predator which we investigate in the cur-
rent work, the frequency of oscillation of the density of
species increases, with the decreasing of the amplitude.
This occurs because the apex predator acts uniformly in
the square lattice, and is expected to equally suppress all
FIG. 2: Density of species as a function of time. The colored
curves represent the respective species, and the black curve
stands for the empty sites. The (a) and (b) panels represent
the X3 and X4 systems, and the (c) panel stands for the
last system, SX3, which contains the apex predator. The
long time evolution shows that all the systems evolve ensuring
coexistence of species.
the species. The average value of the density of species
depends on the values used for the parameters that we
consider in each case; for instance, the value obtained for
the X3 system was ρx = 0.301, with σx = 0.012, where
σx is the standard deviation, and for the system X4,
ρx = 0.225, with σx = 0.018. For the system SX3, we
obtained ρx = 0.226 with σx = 0.006 for the three species
A, B, C that evolve in cyclic competition, and ρd = 0.142
with σd = 0.001, for the apex predator.
We provided 3 × 103 realizations for the systems X3
and X4, and 3 × 104 for the system SX3, in order to
obtain the average density of maxima (cid:104)ρt(cid:105) in the inter-
val of 103 generations. The Fig. 3 shows the histogram
normalized for the number of maxima in the interval of
103 generations (black dots) and in Table I one shows the
values for the average density of maxima and its standard
deviations. We use the average and standard deviations
values in Table I to fit a Gaussian curve, as can be seen
in red dashed line in Fig. 3. Note that the peak distri-
butions have a Gaussian behavior.
In order to get further information concerning the dy-
namical evolution of the systems, we computed the au-
tocorrelation function, which is defined by
(cid:48)
Cxx(t, t
1
σ2(cid:104)x(t)x(t
(cid:48)
)(cid:105) −
(cid:48)
1
σ2(cid:104)x(t)(cid:105)(cid:104)x(t
)(cid:105)
) ≡
(7)
where x ∈ A, B, C, D, σ2 = (cid:104)x(t)2(cid:105) − (cid:104)x(t)(cid:105)2, and (cid:104)···(cid:105)
is to be understood as an average in the ensamble. In
0,10,20,30,10,2020040060080010000,10,2(a)(b)(c)t(generations)density 4
FIG. 4: The autocorrelation functions for the systems X3 (a),
X4 (b) and SX3 (c). The colors identify the species, and the
black dots curve stand for the empty sites. In (c) the three
species compete cyclically in the presence of an apex predator,
in yellow.
possible to map the data of the autocorrelation with
C(t) = cos(ωt + φ),
(9)
where ω and φ are real parameters. Thus, one uses Eq.
(8) to write the average density of maxima for species
that evolve under cyclic competition in the form
(cid:104)ρt(cid:105) =
ω
2π
.
(10)
FIG. 3: The black dots represent the histograms of the num-
ber of maxima in the interval of 103 generations and the red
dashed lines show the fit as Gaussian curves. The cases (a)
and (b) represent the systems X3 e X4, in which all the
species compete cyclically. The figure (c) is for the three
species A, B e C of the system SX3, and the case (d) stands
for the the species D, the apex predator.
TABLE I: Average number of maxima in the interval of 103
generations. In the systems X3 and X4 all the species evolve
in cyclic competition, while in the system SX3 the species D
represents the apex predator.
System
Species
(cid:104)ρt(cid:105)
σt
X4
A, B, C, D A, B, C
5.544
1.052
X3
A, B, C
6.858
0.862
SX3
D
16.496
1.807
6.625
1.222
SX3
However, for the apex predator the initial form of its
autocorrelation function can be described by
C(t) = exp(a0t)
antn,
(11)
n(cid:88)
n=1
where a0, a1, a2, ..., are real parameters. Since the data
are strongly correlated for n ≥ 3, we used n = 4 to get
the density of maxima for the apex predator. It gives
(cid:115)
.
(8)
(cid:104)ρt(cid:105) =
1
2π
−
a 0
4 +4a0
3a1 +12a0
a0
2 + 2a0a1 + 2a2
2a2 +24a0a3 +24a4
.
(12)
(cid:114)
(cid:12)(cid:12)(cid:12)(cid:12)δt=0
Ref. [45], the authors have shown (in a different context)
that the density of maxima of an observable that fluctu-
ates can be obtained from its correlation function. This
result was recently used [30] to provide a way to connect
the density of maxima with the correlation length of the
density of species. The result is obtained with the use
of the maximum entropy principle and can be written in
the form [30, 45]
(cid:104)ρt(cid:105) =
1
2π
T4
T2
;
Tj ≡
−
dj
d(δt)j C(δt)
We then develop stochastic simulations in the square
lattice for the three systems X3, X4 and SX3 and show
in Fig. 4 the autocorrelation function for each system,
there represented by the curves shown in the panels (a),
(b) and (c), respectively. The black-dot curves stand
for the empty sites, and the colored curves represent the
species with their respective colors. One notices that
there is strong accord among the results for the species
that evolve in cyclic competition, with an oscillating be-
havior similar to the one found in Ref. [30].
Let us now focus on the species that compete cyclically.
One notices from Fig. 4 that for t near the origin, it is
In Fig. 5 one depicts the autocorrelation functions for
the three distinct systems, X3, X4 and SX3, with the
red curve representing the X3 system, the green curve
the X4 system, and the blue and yellow curves for one
of the three competing species and for the apex predator
of the SX3 system, respectively. Also, in the inset one
displays the same curves together with the corresponding
fitting functions as the empty-ball curves that show ex-
cellent accord with the colored curves. One recalls that
the colored curves result from the stochastic simulations,
0481200,20,40481200,20,405101500,10,20,30,4714212800,10,20,3number of maximamaxima distribution(a)(c)(b)(d)-0,500,5100,5102004006008001000-0,500,51(a)(b)(c)(generations)tautocorrelation 5
apex predator to spread uniformly in the lattice is kept;
as we see from Fig. 7, the average size of the group of
apex predators is very small, being 3 individuals in the
simulations shown in the figure.
FIG. 6: Average number of clusters (cid:104)N (t)(cid:105) as a function of
time. The red, green and blue lines represent the species in
cyclic competition in the systems X3, X4 and SX3, respec-
tively. The yellow curve stands for the apex predator.
FIG. 7: The mono-log behavior of the average size of the
clusters (cid:104)S(t)(cid:105) as a function of time. The colors follow as in
Fig. 6.
IV. COMMENTS AND CONCLUSIONS
In this work we studied the presence of an apex preda-
tor in a system composed of three distinct species that
compete cyclically following the standard rules of repro-
duction, migration, and predation, which predation con-
trolled as in the paper-rock-scissors game. The apex
predator is a superpredator, since it predates all the other
species and is not predated by any of them. The popu-
lation of apex predator does not increase indefinitely be-
cause it dies with the ratio controlled by β, as suggested
by the rule (5).
We studied the system X3, which is composed of three
distinct species that evolve in cyclic competition, and
FIG. 5: The autocorrelation function for the systems X3 (red
curve), X4 (green curve) and SX3 (blue and yellow curves).
The yellow curve represents the apex predator. The inset
shows perfect agreement between the numerical simulation
and the approximations used to fit the curves, shown as the
empty-ball curves.
TABLE II: Average number of maxima in a interval of 103
generations, obtained from the relation (8).
System
Species
(cid:104)ρt(cid:105)
X4
A, B, C, D A, B, C
4.402
X3
A, B, C
6.081
SX3
D
16.537
SX3
5.802
and the empty-ball curves come from the approximations
(9) and (11) used above to fit the colored curves. The
numerical values are shown in Table II.
In the systems X3, X4, and SX3 that we have studied,
one noticed that species in cyclic competition appear to
have the tendency to clusters or agglomerate, as a mecha-
nism to survive in the competing environment. However,
the apex predator behaves differently, tending to spread
uniformly in the lattice, as it is shown in the right panel
of Fig. 1 (d).
In order to quantify this behavior, one
should investigate the number and size of the clusters of
the species. To implement this possibility, we used the
Hoshen-Kopelman algorithm [31] to compute the average
number of clusters ns(t) of size s as a function of time.
This allows that we introduce the quantities
(cid:88)
(cid:80)
(cid:80)
s
N (t) =
S(t) =
ns(t)
s s2 ns(t)
s s ns(t)
(13)
(14)
which represent the number of clusters and the average
size of the clusters, respectively.
We display the results in Figs. 6 and 7, which show
the average of the number of clusters and the average
size of such clusters as a function of time in the case
of 128 realizations, respectively. The results show that
the presence of the apex predator contributes to diminish
the size of the clusters of the species that compete cycli-
cally, increasing its number. Also, the tendency of the
02004006008001000-0,500,51015300,500,75t(generations)autocorrelation05010015020025030001×1042×1043×104t(generations)hN(t)i050100150200250300100101102103t(generations)hS(t)i then two systems of four species, one, the system X4,
where all the species compete among themselves and have
similar behavior, and the other, SX3, where three species
compete cyclically and the other one represents the apex
predator. We followed ideas developed before in the
works [30, 31] to show that when the species evolve under
cyclic competition, the abundance or density of species
oscillates around an average value, producing maxima
and minima. Also, we computed the average number
of maxima for the three models and related it with the
correlation length of the abundance of species. We also
noticed that the presence of the apex predator decreases
the amplitude of oscillation of the three species.
Other results indicated that the autocorrelation func-
tion for the abundance of species that evolve under cyclic
competition presents a senoidal behavior, but for the
apex predator the behavior is exponential. It was also
noticed that while the species that compete cyclically
tend to cluster and form spiral patterns to survive, the
presence of the apex predator diminishes the average size
of these clusters, increasing its number in a way that con-
tributes to destroy the spiral patterns without jeopardiz-
ing biodiversity. We also found that the apex predator
does not form large clusters, preferring to spread out uni-
formly into the lattice.
The model for the apex predator studied in this work
6
does not account for several aspects such as the interfer-
ence among individuals of the same species in the search
for the corresponding prey, the conversion of successful
predator-prey hunts into new predators, the age of the
predators, which may interfere in the predation ratio,
and so on. The addition of new rules with focus on spe-
cific scenarios is a challenging task that may appear as
natural extensions of the current work. Despite the sim-
plicity of the systems studied, the results of this work are
of current interest and suggest the need of new, more de-
tailed investigations on the behavior of the apex predator
and how it may change basic aspects of the environment
that favors biodiversity. Other issues concern investi-
gations related to ordering of heterogeneous systems and
complex networks, with the predator-prey Lotka-Volterra
mean field theory in the presence of an apex predator and
also, with Monte Carlo simulations of n-state Potts mod-
els and other complex fluid models in Statistical Physics
[46]. We hope to further report on this in the near future.
Acknowledgments
This work is partially supported by The Brazilian
agencies CAPES and CNPq.
[8] T. Reichenbach, M. Mobilia, and E. Frey, Nature (Lon-
[26] A. D. Wallach, W. J. Ripple, and S. P. Carroll, Trends
[1] M. Frean and E. Abraham, Proc. R. Soc. London, Ser.
B, 268, 1323 (2001).
[2] N. Bacaër, A short history of mathematical population
dynamics. (Springer-Verlag, London, 2011).
[3] C. Adami and A. Hintze, Nat. Commun. 4, 2193 (2013)
[4] G. Hardin, Science 131, 1292 (1960).
[5] G. Hutchinson, American Naturalist 95, 137 (1961).
[6] B. Kerr, M. A. Riley, M. W. Feldman, and B. J. M.
Bohannan, Nature (London) 418, 171 (2002).
[7] G. Károlyi, Z. Neufeld, and I. Scheuring, J. Theor. Biol.
236, 12 (2005).
don) 448, 1046 (2007).
Lett. 99, 238105 (2007).
[9] T. Reichenbach, M. Mobilia, and E. Frey, Phys. Rev.
[10] J. Nahum, B. Harding, and B. Kerr, Proc. Natl. Acad.
Sci. USA 108, 10831 (2011).
[11] A. Sih, P. Crowley, M. McPeek, J. Petranka, and K.
Strohmeier, Annu. Rev. Ecol. Syst. 16, 269 (1985).
[12] F. Palomares, P. Gaona, P. Ferrerasa, and M. Delibes,
Conserv. Biol. 9, 295 (1995).
[13] J. Terborgh, J. A. Estes, P. Paquet, K. Ralls, D. Boyd-
Herger, B. J. Miller, and R. F. Noss, in Continental con-
servation, edited by M. E. Soulé and J. Terborgh (Island
Press, California, 1999).
[14] J. B. Shurin and E. G. Allen, American Naturalist 158,
624 (2001).
[18] R. T. Paine, American Naturalist 103, 91 (1969).
[19] J. W. Porter, American Naturalist 106, 487 (1972).
[20] C. Kullberg and J. Ekman, Oikos 89, 41 (2000).
[21] M. G. Jones, Emp. Jour. Exp. Agri. 1, 43 (1933).
[22] J. L. Harper, Brookhaven Symp. Bio. 22, 48 (1969).
[23] R. T. Paine, Science, 296, 736 (2002).
[24] E. G. Ritchie, B. Elmhagen, A. S. Glen, M. Letnic, G.
Ludwig, and R. A. McDonald, Trends Ecol. Evol. 27, 265
(2012).
[25] A. D. Wallach, C. N. Johnson, E. G. Ritchie, and A. J.
O'Neill, Ecol. Lett. 13, 1008 (2010).
Ecol. Evol. 30, 146 (2015).
[27] M. J. Pongsiri, J. Roman, V. O. Ezenwa, T. L. Goldberg,
H. S. Koren, S. C. Newbold, R. S. Ostfeld, S. K. Pat-
tanayak, and D.J. Salkeld, BioScience, 59, 945 (2009).
[28] C. C. Wilmers, E. Post, R. O. Peterson, and J. A.
Vucetich, Ecol. Lett. 9, 383 (2006).
[29] B. R. Silliman and C. Angelini, Nat. Educ. Knowl. 3, 44
[30] D. Bazeia, M. B. P. N. Pereira, A. V. Brito, B. F.
Oliveira, and J. G. G. S. Ramos, Scientific Reports 7,
44900 (2017).
[31] J. Hoshen and R. Kopelman, Phys. Rev. B 14, 3438
[32] M. Peltomäki and M. Alava, Phys. Rev. E 78, 031906
(2012).
(1976).
(2008).
[15] A. D. Wallach, I. Izhaki, J. D. Toms, W. J. Ripple, and
[33] W. Wang, Y. Lai, and C. Grebogi, Phys. Rev. E 81,
U. Shanas, Oikos 124, 1453 (2015).
[16] H. Caswell, American Naturalist 112, 127(1978).
[17] R. T. Paine, American Naturalist 100, 65 (1966).
046113 (2010).
030901 (2010).
[34] H. Shi, W. Wang, R. Yang, and Y. Lai, Phys. Rev. E 81,
7
[35] L. Jiang, T. Zhou, M. Perc, and B. Wang, Phys. Rev. E
[42] A. J. Lotka, Journal of the American Chemical Society
[36] D. Grošelj, F. Jenko, and E. Frey, Phys. Rev. E 91,
[43] V. Volterra, Leçons sur la Theorie Mathematique de la
42, 1595 (1920).
[37] G. Szabó and G. A. Sznaider, Phys. Rev. E 69, 031911
[44] P. P. Avelino, D. Bazeia, J. Menezes, and B. F. de
84, 021912 (2011).
033009 (2015).
(2004).
[38] G. Szabó, J. Phys. A: Math. Gen. 38, 6689 (2005).
[39] P. P. Avelino, D. Bazeia, L. Losano, J. Menezes, and B.
F. Oliveira, Phys. Rev. E 86, 036112 (2012).
[40] B. Intoy and M. Pleimling, J. Stat. Mech. P08011 (2013).
[41] A. F. Lütz, S. Risau-Gusman, and J. J. Arenzon, J.
Theor. Biol. 317, 286 (2013).
Lutte pour la Vie. (Gauthier-Villars, Paris, 1931).
Oliveira, Phys. Lett. A 378, 393 (2014).
[45] J. G. G. S. Ramos, D. Bazeia, M. S. Hussein, and C. H.
Lewenkopf, Phys. Rev. Lett. 107, 176807 (2011).
[46] D. P. Landau and K. Binder, A Guide to Monte Carlo
Simulations in Statistical Physics. (Cambridge Univer-
sity Press, Cambridge, UK, 2009).
|
1804.08231 | 1 | 1804 | 2018-04-23T02:51:33 | Non-neurotoxic Nanodiamond Probes for Intraneuronal Temperature Mapping | [
"physics.bio-ph"
] | Optical bio-markers have been used extensively for intracellular imaging with high spatial and temporal resolution. Extending the modality of these probes is a key driver in cell biology. In recent years, the NV centre in nanodiamond has emerged as a promising candidate for bio-imaging and bio-sensing with low cytotoxicity and stable photoluminescence. Here we study the electrophysiological effects of this quantum probe in primary cortical neurons. Multi-electrode array (MEA) recordings across five replicate studies showed no statistically significant difference in 25 network parameters when nanodiamonds are added at varying concentrations over various time periods 12-36 hr. The physiological validation motivates the second part of the study which demonstrates how the quantum properties of these biomarkers can be used to report intracellular information beyond their location and movement. Using the optically detected magnetic resonance from the nitrogen vacancy defects within the nanodiamonds we demonstrate enhanced signal-to-noise imaging and temperature mapping from thousands of nanodiamond probes simultaneously. This work establishes nanodiamonds as viable multi-functional intraneuronal sensors with nanoscale resolution, that may ultimately be used to detect magnetic and electrical activity at the membrane level in excitable cellular systems. | physics.bio-ph | physics | Non-neurotoxic Nanodiamond Probes for
Intraneuronal Temperature Mapping
David A. Simpson1,2*, Emma Morrisroe3, Julia M. McCoey1, Alain H. Lombard4, Dulini C.
Mendis5, François Treussart4, Liam T. Hall1, Steven Petrou2,3,6,7, Lloyd C. L. Hollenberg1,2,8
1School of Physics, University of Melbourne, Parkville, 3010, Australia.
2Centre for Neural Engineering, University of Melbourne, Parkville, 3010, Australia.
3Florey Neuroscience Institute, University of Melbourne, Parkville, 3010, Australia.
4Laboratoire Aimé Cotton, CNRS, Univ. Paris-Sud, ENS Paris-Saclay, Université Paris-Saclay,
91405 Orsay, France.
5Department of Mechanical Engineering, University of Melbourne, Parkville, VIC 3010,
Australia.
6Centre for Integrated Brain Function, University of Melbourne, Parkville, 3010, Australia.
7Department of Medicine, Royal Melbourne Hospital, University of Melbourne, Parkville, 3010,
Australia.
8Centre for Quantum Computation and Communication Technology, University of Melbourne,
Parkville, 3052, Australia.
KEYWORDS: nanodiamonds, nitrogen-vacancy color center, primary neuron culture, multi-
electrode arrays, wide field optical microscopy, quantum sensing, temperature sensing.
1
ABSTRACT
Optical bio-markers have been used extensively for intracellular imaging with high spatial and
temporal resolution. Extending the modality of these probes is a key driver in cell biology. In
recent years, the NV centre in nanodiamond has emerged as a promising candidate for bio-imaging
and bio-sensing with low cytotoxicity and stable photoluminescence. Here we study the
electrophysiological effects of this quantum probe in primary cortical neurons. Multi-electrode
array (MEA) recordings across five replicate studies showed no statistically significant difference
in 25 network parameters when nanodiamonds are added at varying concentrations over various
time periods 12-36 hr. The physiological validation motivates the second part of the study which
demonstrates how the quantum properties of these biomarkers can be used to report intracellular
information beyond their location and movement. Using the optically detected magnetic resonance
from the nitrogen vacancy defects within the nanodiamonds we demonstrate enhanced signal-to-
noise imaging and temperature mapping from thousands of nanodiamond probes simultaneously.
This work establishes nanodiamonds as viable multi-functional intraneuronal sensors with
nanoscale resolution, that may ultimately be used to detect magnetic and electrical activity at the
membrane level in excitable cellular systems.
Imaging neuronal action potentials (APs) at the single cell level is a challenging and active field
of research. A variety of approaches have been developed with success from patch clamping1 and
multi-electrode arrays (MEAs)2 to optical recording via genetically encoded voltage indicators.3, 4
However, many of these techniques suffer from at least one drawback of scalability, low
sensitivity, slow temporal response, toxicity, low specificity and limited spatial resolution. Optical
2
probes have attracted significant interest in recent years due to the relatively high signal to noise
ratios, ∼20 % increase in the fluorescence signal at the peak of the AP,5 and the possibility of high
resolution sub-micron imaging of APs in real time. Unfortunately, these approaches are often
accompanied by toxicity effects due to the overexpression of the genetically encoded probe, and
can in some cases modify the network activity of the system under investigation. Therefore, the
development of alternative sensing probes that exhibit low neurotoxicity and can report cellular
physico-chemical events such as changes in the electric potential, pH, ionic currents, magnetic
field and potentially temperature is essential.
Diamond nanocrystals containing colour centres have been identified as promising candidates
for bio-imaging with long term photo-stability and inherently low cytotoxicity.6 These properties
have led to a myriad of fluorescence based applications including neuronal differentiation and
tracking. The presence of nanodiamonds (NDs) in such systems has been shown not to alter the
cell viability,7 morphology8 or production of neuron-specific markers such as β-III-tubulin during
cell differentiation.9 The photo-stability of the probes has also been used to identify cellular
internalization pathways,10 and monitor and track the intraneuronal endosomal transport in model
and transgenic systems.11 The prominent optical defect centre used in these biological applications
is the negatively charged nitrogen-vacancy (NV) centre. In addition to the photo-stability of this
defect, the quantum properties of the NV provide opportunities to sense a range of physio-chemical
parameters including magnetic fields,12, 13 temperature,14 electric potentials15 and pH16 which are
of particular interest in excitable cellular systems. In order to utilise these nanoscale quantum
sensors in functioning neuronal networks, we must first establish whether the presence of the probe
itself has an adverse effect on the connectivity and performance of the neuronal network.
Neurotoxicity can often proceed in the absence of other biochemical or morphological changes;
3
therefore, traditional cell toxicity assays are not optimized for detecting this type of toxicity.17
MEAs have demonstrated the capacity to screen for neurotoxic effects such as synaptic function,
action potential generation and propagation, plasticity, and network formation and function.17, 18
Here we use MEAs to assess whether the presence of NDs in primary cortical neurons elicits a
neurotoxic response. We conduct a detailed network analysis across multiple control and ND
groups at varying ND concentrations. Comparison of 25 separate network performance indicators
over time demonstrate no neurotoxicity from the uptake of NDs at concentrations up to 20 µg/ml.
In addition to the neurotoxicity study, optical imaging and quantum sensing protocols were
performed on the primary cultures. We demonstrate improved signal to background fluorescence
imaging by taking advantage of the optically detected magnetic resonance (ODMR) associated
with the NV centre and use wide-field optical techniques to simultaneously report the ODMR
spectra from thousands of NDs within seconds. As a proof-of-principle demonstration we apply
quantum control techniques to map the intracellular temperature of a neuronal network.
Action potentials which govern interactions and connectivity in excitable cells and neuronal
networks generate a range of measurable parameters. To date, optical sensors have been developed
to respond to changes in electrical potential and/or ionic currents. Nanodiamonds represent a
promising approach for the detection of APs with demonstrated sensitivity to magnetic and electric
fields, pH and temperature. At present, ND magnetic19 and electric field15 achieved sensitivities
are far from that required to detect single AP events.20 However, recent reports by Barry et al. 21
have shown that magnetic detection of single APs in squid and worm axons is possible by
integrating the signal from ensembles of NV centres in single crystal diamond over large volumes
(0.013×0.2×2 mm3). Beyond magnetic sensing, opportunities exist based on photoluminescence
(PL) changes in response to electrostatic potential modulation,15 but it remains to be seen if the
4
electric signaling from ion exchange across axonal membranes, as opposed to charge injection,
can modulate NV fluorescence. These potential applications motivate further investigation of NDs
in biological systems, in particular, excitable cellular systems where these types of signals are
generated. However, before deploying such sensors into such complicated systems, it is imperative
to evaluate the neurotoxicity of these probes. The motivation of this work is to assess whether the
interaction of NDs with neurons impacts the functional dynamics and performance of neuronal
networks and to show that the imaging protocols required for quantum sensing can be achieved on
a timescale relevant to cellular processes.
RESULTS
Multi-electrode array recordings of neuron primary cultures containing fluorescent NDs.
Here we use MEAs to systematically study the effect of NDs on the electrophysiological activity
of cultured mouse cortical neurons between DIV 11 and DIV 13. Extracellular electrical recordings
were carried out using the Multiwell Screen from Multi Channel Systems in 5 minute recordings
at 20 kHz. During data acquisition, cultures were kept at 37 °C and enclosed in chambers with
perfused carbogen (95% O2, 5% CO2). The network electrical activity was benchmarked with
control cultures injected with commensurate amounts of sterile water. Figures 1(a-d) show typical
AP spikes from a single MEA channel over a 10 s period, for the control and ND-containing
networks. The network parameters across the 12 channels per well can be visually represented in
a raster plot as shown in Fig. 1(e-h), where each vertical line represents a single AP and each
horizontal row represents one channel of the 12 in each MEA well. From this plot, it is possible to
visualise spikes, bursts, and network bursts across each channel. The raster plots for the baseline
5
control and ND groups show a high level of network bursting, consistent with a relatively mature
neuronal network. Five sets of MEA recordings were performed on 48 individual wells at varying
ND concentrations. The first set of MEA baseline recordings were undertaken at Day 11, prior to
the introduction of the NDs. The control recordings were performed 24 hours post media change
to avoid stress-related responses. The ND suspension was dispersed in cell media, sonicated for
5 min and then applied to the primary cultures during a routine change of cell media. Network
recordings were then carried out at 12, 24 and 36 hours intervals.
Figure 1: Electrical recordings obtained from the Multiwell Screen for two conditions at baseline (a) and
(c) and 36 hours after the addition of water vehicle (b) and 20 µg/ml ND (d). The data signals have been
high-pass filtered at 300 Hz to detect AP spikes. (e-h). Raster plots showing AP spikes from each of the 12
electrodes per well for the baseline, control and highest concentration ND group. The plots are shown for
the last minute of recording.
6
To quantify the level of network activity, key parameters of the network spiking such as the burst
rate, firing rate, spike amplitude and percentage of spikes in network bursts are compared from all
channels across the 48 individual wells for the control and ND groups. The results for the baseline
and highest ND concentration (Fig. 1) are typical for the results observed across all measured ND
concentrations. The raw comparisons of four specific network parameters are shown in Figure 2(a-
d). The plots show no statistically significant difference in any of the key network comparisons
both in terms of absolute values and trend across the data set. In all cases, the rate of network
bursting increased with time as did the mean firing rate and average spike amplitude. This provides
strong evidence that the network is continuing to mature and strengthen its connectivity across
computational units, even after the addition of NDs. The highly synchronous nature of the activity
and the fact that greater than 80 % of spikes occur in a network burst indicates highly
interconnected neurons within the network. A more detailed network comparison is shown in
Figs. 2(e) and (f). For each parameter, the percentage change from the baseline was calculated at
each time point. Any differences between a ND group and the corresponding control that are
statistically significant would be coloured while those that are not significant will be shown in
grey.
7
Figure 2: Neuronal network parameter comparison for varying ND concentrations. (a-d) The network burst
rate, mean firing rate, average spike amplitude and percentage of spikes in the network bursts for the control
and highest ND concentration (20 µg/ml) at 0 hours and 36 hours across the 5 replica sets. (e-f) Iris plots
comparing 25 network parameters between control and ND groups. Grids in each heat map represent
statistical comparison between a ND condition (e) 12, 24, 36 hours after the application of 20 µg/ml NDs,
8
(f) 36 hr after the application of 1, 5, 10, 20 µg/ml concentrations of NDs and its control. Sections in the
circles correspond to network parameters. p-values resulting from statistical comparisons (adjusted for
multiple comparisons) are represented by the colour scale, with grey showing no difference between the
pair of conditions being compared. Red and blue represent p-values for increase/decrease in the parameter
values in the ND group compared to control as indicated within the iris plots (* 0.01 ≤ p < 0.05, ** 0.001
≤ p < 0.01, *** 0.0001 ≤ p < 0.001, **** 0.00001 ≤ p < 0.0001).
As can be seen from the iris plot, there is no statistically significant difference in 25 separate
network comparison parameters investigated over time or as a function of ND concentration. This
detailed comparison demonstrates that NDs present at concentration ranges up to 20 µg/mL
(1.6 µM) have a negligible impact on the neuronal activity, growth or development. The next
section demonstrates how the quantum properties of NV-containing NDs can be optically
addressed, manipulated and read out simultaneously over wide fields of view. As a proof of
principle application, we show how to report the local intracellular temperature using the quantum
properties of the probes.
Wide-field imaging and ODMR spectroscopy of NDs.
The optical and quantum properties of the NV centre in diamond set it apart from other
fluorescence based probes. These properties can be controlled and manipulated at room
temperature in biological environments22 to report changes in the local pH, electric, magnetic and
temperature fields. Over the past decade there has been a renewed interest in probing the
intracellular temperature with several fluorescence based methods proposed to image the
temperature distribution including: fluorescence proteins23 and polymers,24 quantum dots,25 rare-
9
earth doped nanocrystals,26 metal nanoparticles27 and NDs.14 Many of these approaches rely on
fluorescence changes in the excited state lifetime or the fluorescence spectrum of the probe. The
temperature response from NDs differs from most fluorescence probes, whereby the ground state
spin levels are shifted in response to a change in temperature.28 This provides a robust readout of
temperature with reduced cross sensitivity to other physiological parameters such as pH and redox
potential. Therefore, in this section we demonstrate how these nanoscale quantum probes can be
used to report intracellular temperature distributions over wide-fields of view within a temporal
resolution of order a few seconds. To demonstrate optical and quantum sensing protocols in
primary cortical neurons we culture neuronal networks onto a tailored glass coverslip with a gold
microwave resonator coated with polydimethylsiloxane (PDMS) using the same culturing
procedure outlined for the MEA investigation (see Methods). The ND suspension was dispersed
in cell media (6 µg/ml), sonicated for 5 min and then applied to the primary cultures during a
routine change of cell media. Imaging was carried out 12 hours after the incubation. Using ODMR
spectroscopy29 in combination with standard wide-field microscopy techniques, see Fig. 3(a), we
are able to simultaneously determine the ground state energy level structure from thousands of
individual NDs over a wide field of view (80 × 80 µm) within seconds. The ground state spin
levels of the NV centre respond to local magnetic, electric and temperature fields in a variety of
ways. For example, in the presence of a magnetic field, the spin states Zeeman split with a
gyromagnetic ratio of 2.8 MHz/G.12 An electric field Stark splits the energy levels by 1.7 kHz
m/V,30 whilst changes in temperature shift the spin levels by -74 kHz/K.28 In addition to these
effects, the ODMR provides a mechanism to modulate the photoluminescence from NV centres
within the NDs.31 At zero magnetic field, the majority of NDs have a crystal field splitting of
D = 2.87 GHz. Under 532 nm laser excitation, the NV spin can be conveniently polarised into the
10
mS = 0 bright state. By applying a microwave frequency of 2.87 GHz across the field of view, the
NV spins can be driven into the mS = ± 1 dark state with a photoluminescence (PL) ratio ΔPL/PL
of ∼10 %. By taking the difference of two photoluminescence images with and without the
microwave field as shown in Fig. 3(b) unwanted cell/media auto-fluorescence can be removed31.
This is particularly important for neuronal cell imaging which often requires tailored cell culture
media along with well controlled temperature environments to maintain functional networks. PL
images of primary cortical neurons incubated with NDs at a concentration of 6 µg/mL of cell media
are shown in Fig. 3(c), along with alternative ODMR contrast images of the same area obtained
with microwaves "on" and "off" resonance at 2.87 GHz and 2.84 GHz, respectively. By comparing
a single diffraction limited spot from an individual ND in the PL and ODMR contrast image, we
find an 50% increase in the signal to background ratios for the ODMR contrast image due to the
removal of cell and media auto-fluorescence. The ODMR contrast imaging for the "off" resonance
case shows no fluorescence contrast from the same imaging area demonstrating the ODMR
contrast enhancement acts over the entire field of view and is highly selective.
11
Figure 3: Wide-field imaging of NDs containing NV centres in primary cortical neurons. (a) Structure and
energy levels of the NV centre in diamond and schematic of the wide-field fluorescence microscope. A
532 nm laser is used to excite the NDs with the resulting fluorescence collected from 650-750 nm. (b)
Acquisition sequences for conventional PL and ODMR contrast imaging. ODMR contrast imaging is
performed "on" (2.87 GHz) and "off" (2.84 GHz) resonance with a microwave pulse applied over the field
of view. (c) Series of images using the acquisition sequences described in (b). The PL image (left) shows
NDs distributed throughout the cellular components. The ODMR contrast image (middle) "on" resonance
shows background free imaging with improved signal to noise ratios, while the "off" resonance image
(right) demonstrates the microwave pulse acts efficiently over the entire field of view. (d) 3D fluorescence
intensity maps for a single diffraction limited ND spot highlighted in (c). Analysis shows an 50% increase
in the signal to background ratio for the ODMR contrast imaging technique.
12
The entire ODMR spectrum from each individual ND can be acquired through an image series at
various microwave frequencies. The wide-field NV fluorescence image in Fig. 4(a) taken 6 µm
above the coverslip identifies thousands of NDs within the neuronal structures. Figure 4(b) shows
typical ODMR spectra from two individual NDs from Fig. 4(a), with a typical ODMR peak
splitting observed due to intrinsic strain within the ND. The ODMR spectrum is obtained from a
set of 100 images integrated for 30 ms per image with a total acquisition time of 6 seconds.
Although each individual ND may experience slightly modified crystal field splitting D from local
strain,22 the 170 nm NDs used in this work exhibit a spread in D smaller than 170 kHz measured
over 9078 different NDs which ensures the ODMR contrast imaging technique is effective in
addressing all fluorescent NDs at zero magnetic field.
Figure 4: Wide-field ODMR imaging of NDs. (a) NV fluorescence image of NDs measured 6 µm above
the coverslip, recorded at 37 °C. (b) ODMR spectra from two individual NDs corresponding to the points
#1 and #2 in (a). The ODMR spectra is fit with a sum of two Lorentzians (red line). The total acquisition
13
time for the simultaneous recording of ODMR spectra of all NDs was 6 seconds. (c) Histogram of the
distribution of D from 9078 NDs in Fig. 4(a). The mean crystal field splitting was 2868.59 MHz with a
standard deviation of 170 kHz.
The ND concentration used for imaging (6 µg/mL) provides good coverage of the network and
may be used to spatially map intracellular components. To demonstrate the sensing capabilities of
the ND probes, the temperature of the environmental chamber was reduced by 1.9 °C. A
temperature map from the intracellular NDs, shown in Fig. 5a, was obtained by identifying
individual ND spots in the fluorescence images at 37.3 and 35.2 °C (see Methods for procedure).
The spots were then co-localized within a maximum pixel distance of N=9, corresponding to 2 µm.
This analysis resulted in 255 co-localised regions from which the ODMR spectrum was obtained
at each temperature. The crystal field splitting, D, from each ND was determined by fitting a single
Lorentzian peak to the ODMR spectrum, in place of the two Lorentzian peak fit shown in Fig. 4b.
The noise around the strain split ODMR lines causes significant uncertainly in the crystal field
splitting parameter. We instead remove 4 central points and fit a single Lorentzian to the data to
achieve a more precise estimate of D, as shown in Fig. 5b. The D values from the co-localised
NDs were retained for ODMR fits with errors in the peak position less than 0.01%. A temperature
map was then obtained by subtracting the D at each location at 37.3 and 35.2 °C and converting
the shift in D into temperature, using dD/dt = -74 kHz/K. The acquisition time per point was 120
ms with 50 points in total used for the fitting analysis. With pixel binning of 1.2×1.2 µm and a
total acquisition time of 12 s the signal to noise ratio for the temperature recording just above one.
Figure 5(b) shows an example of the ODMR fit from a single co-localised spot with a shift in D
of 110 ± 90 kHz and a change in local temperature of -1.5 ± 1.2 °C. The uncertainty in D is
dominated by the broad ODMR line associated with the noise from the nitrogen and 13C spin bath
14
intrinsic to the ND. The precision can be improved using the statistical power of many NDs as
shown below or by utilizing tailored ultra-pure NDs currently under development.32 Temporal
improvements can be made by moving to a simplified two33 or four point34 frequency measurement
technique and increasing the pixel binning at the expense of spatial resolution. A histogram of the
temperature recordings from Fig. 5(a) is shown in Fig. 5(c). The distribution reports a mean
temperature change of -1.36 °C with a standard error on the mean (s.e.m.) of 0.08 °C. This is
consistent with the reduction in environmental temperature measured externally from the network
culture. The measurement technique was verified by performing the same fitting procedure on two
sets of fluorescence images at the same environmental temperature i.e. 37.0 ± 0.1 °C. The
histogram from 209 reported NDs found a mean temperature change of 0.2 ± 0.1 (s.e.m.) °C, see
Supplementary Information.
Figure 5: Intracellular temperature map from NDs in primary cortical neurons. (a) Temperature map from
co-localised NDs overlaid with the ND fluorescence image at 37.3 °C. (b) ODMR spectra from a typical
15
ND in (a) at 37.3 °C (top right) and 35.2 °C (bottom right). The shift in D of 110 ± 90 kHz represents a
local temperature change of -1.5 ± 1.2°C. The temperature error is obtained from the square root of the sum
of the square errors on the peak position from the respective fits at each temperature. (c) Histogram of
temperature changes measured from 255 co-localised NDs. The mean temperature change reported across
the neuronal network was -1.36 ± 0.08 (s.e.m) °C consistent with the reduction in environmental
temperature.
This proof of principle demonstration showed no functional or morphological neuronal
degradation from the weak optical and microwave fields used in the temperature sensing protocol,
indicating that NDs can act as effective multi-modal sensors in functioning neuronal networks.
Recent studies have identified temperature heterogeneity in neuronal cells,24 however debate
remains about whether biochemical reactions at the single cell level are sufficient to generate such
heterogeneity.35, 36 Other forms of temperature variations have been proposed which may
accompany current flow in open ionic channels 37. NDs therefore represent a realistic pathway
towards characterising such changes at these boundaries.14 The sensitivity of the ND probes in
vitro is 1K/√Hz; this is favorable in comparison to other nanoscale probe techniques.14 The
accuracy of the temperature recordings can be improved to less than 1 K with modest integration
times of tens of seconds. In terms of the potential for magnetic and electric field sensing with these
probes, the sensitivity remains limited by the quality of the diamond material. At present, the bulk
of ND material arises from low quality high pressure high temperature synthesised diamond. The
relatively high nitrogen (100 ppm) and 13C (1 ppt) content in this material severely impacts the
linewidths of the ODMR spectrum to of order MHz. This is three orders of magnitude away from
that needed to detect single action potential events. With material improvements, such as tailored
CVD growth32 and isotopically enriched 12C diamond38, combined with targeted functionalisation
16
strategies, these functional quantum probes may provide sufficient sensitivity to evaluate
intracellular processes of excitable cells, with nanoscale resolution20.
DISCUSSION
In this work, we have shown that NDs containing NV centres present a promising pathway for
multi-modal imaging in biology. We have shown that the presence of NDs in neuronal networks
up to a concentration of 20 µg/mL has a negligible impact on the network dynamics and
connectivity. Network electrophysiological parameters such as the mean firing rate, network burst
rate and spike amplitude show no statistically significant deviation from the measured controls for
the five replicates at each ND concentration studied in this work. Furthermore, we demonstrate
that ODMR spectroscopy can be performed in the intracellular environment using wide-field
microscopy, with quantitative ODMR spectra obtained in seconds. The measurement protocol is
the basis of the majority of quantum sensing protocols of the NV centre in diamond. We use the
ODMR signature from individual NDs to demonstrate intracellular temperature mapping with high
spatial resolution. This approach may be of interest for measuring temperature changes in response
to external stimulus such as light. Optogenetics is a well-established and widely used technique
for neuronal stimulation. The heat generated from the optical stimulation has been found to impact
the firing dynamics of the system under investigation39, however, quantifying these effects is
challenging and has so far relied on the use of thermistors39. The ND-based thermometry
demonstrated here appears ideally suited to this task. The temperature sensitivity of the NDs is
sufficient for this type of application and the optimal NV excitation wavelength (560 nm) is distinct
from the one of the most common light-gated ion channel proteins (e.g. channel rhodopsin ChR2,
17
absorption maximum at 460 nm wavelength). Moreover, we have shown that NDs do not elicit a
neurotoxic response in vitro. Other future applications may include the study of signaling cascade
resulting from locally induced temperature changes. For example, in neurological disorders such
as epilepsy, where the firing rate of neurons is found to increase; this may result from a local
increase in temperature related to increased metabolism. Our method provides the opportunity to
correlate temperature and neuronal activity changes. ODMR is certainly not the only spectroscopy
and sensing methodology applicable to NDs; the coherence of the NV spin can also be exploited
for detection of electronic40-43 and nuclear spin44-47 species. Therefore, applying these functional
quantum probes to the intracellular environment of excitable cells may provide insight into the
mechanisms and processes governing intracellular dynamics and communication.
METHODS
Culture Preparation.
Multiwell MEAs with gold electrodes and an epoxy (FR-4) substrate from Multi Channel Systems
(Reutlingen, Germany) with a 4×6 wells format were used. MEAs were plasma cleaned 4-7 days
prior to plating and were treated with PolyEthyleneImine 24 hours before plating. Wells were
coated with Laminin 2-3 hours before plating and kept in the incubator. Laminin was removed
immediately prior to plating. The cortex was dissected from C57BL/6 mice 0-2 days post-natal.
Cell media was prepared with 89.3 % Minimum Essential Medium (MEM), 0.9 % 1M Hepes,
glucose (6 mg per 1 mL of MEM), 8.9 % Fetal Bovine Serum (FBS) and 0.9 % P/S. Cortical pieces
were subjected to dissociation using Trypsin (0.25 %), DNase (0.032 %) and trituration with a
glass pipette. 375,000 cells were plated in 120 µL of media to cover the base of each well and after
the initial media was removed, a 500 µL of culture media containing Neurobasal-A medium
18
supplemented with 1.9 % B27, 0.95 % GlutaMax, 0.95 % HEPES and 0.95 % PenStrep was added.
MEAs were covered with lids and kept in an incubator (37 °C, 65 % humidity, 9 % O2, 5 % CO2).
5 µM Ara-C was added at DIV 3 and was removed on DIV 5 with a 100 % medium change to
control the proliferation of glial cells. 100 % culture medium changes were carried out three times
a week. All animal experiments were approved by the Howard Florey Institute Animal Ethics
Committee, and performed in accordance with the Prevention of Cruelty to Animals Act and the
NHMRC Australian Code of Practice for the Care and Use of Animals for Scientific Purposes.
MEA Analysis.
Voltage data recorded with Multiwell Screen system (Multi Channel Systems) were imported into
Matlab (The MathWorks, Inc., Natick, Massachusetts, USA) where subsequent analysis was
carried out. In order to extract high frequency components of the signals, voltage signals of each
electrode were high-pass filtered at 300 Hz. Spikes were detected with custom Matlab scripts based
on precision timing spike detection,48 with spikes being detected if the highest phase of the spike
was greater than 6 times the standard deviation of noise. Bursts on single channels were detected
depending on average firing rates as described in Mendis et al.49 with the minimum number of
spikes for detecting a network burst set at three. Network bursts (NBs) were detected when more
than 25 % of channel-wise bursts overlapped in time.
Network characteristics were extracted in terms of firing and bursting parameters. Average
amplitude of spikes (avgSpkAmp) and mean firing rates (MFRAll) were calculated for each
channel over the entire time period and were then averaged across channels. Mean firing rates
inside bursting periods (MFRIn), in non-bursting periods (MFROut) as well as the ratio
MFROut/MFRIn (MFRRatio) were calculated. Burst features were calculated for both single-
19
channel bursts as well as network bursts. Single-channel burst parameters include the number of
spikes inside bursts (SCBSize) and burst durations (SCBDur). Network burst parameters features
consists of NB rate, NB durations (NBDur), inter NB intervals (INBI-time interval between the
end of a NB and the beginning of the next NB), the jitter between channels participating in the NB
(time interval between the starts of the first and last single-channel bursts inside a NB), the average
number of spikes in a NB (avgSpkInNB), the amplitude of spikes in a NB averaged over the NB
duration (avgNBAmp), the percentage of active channels participating in NBs (%ChInNBs) and
the percentage of spikes in NBs (%SpkInNBs). For the parameters SCBSize, SCBDur, NBDur,
INBI and jitter, the mean, coefficient of variation (cv) and the range were calculated.
Statistical analysis.
The statistical differences for each network parameter was characterised by looking at the
percentage change from the measurement baseline between the control and ND groups. If the
values in each group followed a Gaussian distribution (as determined by Jarque-Bera tests), t-tests
were carried out to obtain the statistical difference between them. Otherwise, Wilcoxon Mann
Whitney-U tests were carried out. The resulting p-values were corrected for multiple comparisons
using the Bonferroni method and were used to generate iris plots, which are radial heat maps that
show color-coded p-values. Lower p-values (ex: p < 0.0001) represent a larger difference between
the two groups while higher p-values (ex: 0.01 < p < 0.05) represent smaller differences.
ODMR fitting procedure.
The majority of NDs exhibit strain splitting in their ODMR spectrum as shown in Fig. 4(b). To
improve the peak fitting precision, the four central data points were removed and a single
Lorentzian fit applied. For the temperature mapping, NDs were located using a spot detection
20
algorithm using the open source software package Icy. Bright spots were detected over a dark
background with a lower scale of: 100 and an upper scale of: 700. Nanodiamond locations were
then co-located using the "Co-localisation studio" package for the respective temperatures of 37.3
and 35.2 °C with a maximum pixel distance of N=9. The ODMR peak position was then obtained
from a 6×6 bin area around each co-localised spot. The data was retained for Lorentzian fits with
less than 0.01% uncertainty in the peak position. The change in peak position at each location was
then converted into temperature using the relationship dD/dT = -74 kHz/K.
Materials.
The NDs (brFND-100) used in this work were sourced from FND Biotech Inc. (Taiwan). The NDs
were dispersed in water at a concentration of 1 mg/ml. The Z-average particle size from dynamic
light scattering measurements was 170 nm (see supplementary information) and the average
number of NVs per particle was ∼500. No specific ligands/moieties were used to facilitate cellular
uptake. The uptake was passive and nonspecific endocytosis.
Fluorescence Imaging.
The widefield imaging and quantum sensing experiments were undertaken by preparing neuronal
cultures onto a glass coverslip with an evaporated gold microwave resonator. The glass coverslip
was coated with a thin layer of PDMS (10:1 ratio) to prevent contact between the active and ground
plane of the resonator when the cultures were immersed in cell media. The networks were cultured
under the same conditions as described above. The ND suspension (brFND) was dispersed in cell
media (6 µg/ml), sonicated for 5 min and then applied to the primary cultures during a routine
change of cell media. Imaging was carried out 12 hours after the incubation. The wide-field
imaging was performed on a modified inverted microscope (Eclipse Ti-U, Nikon, Japan). Optical
21
excitation from a diode-pumped solid-state laser emitting at 532 nm wavelength (Sprout G,
Lighthouse Photonics, Inc, San Jose, USA) was focused onto an acousto-optic modulator (AOM)
(Model 3520-220 from Crystal Technologies/ Gooch & Housego, Palo Alto, USA) and then
expanded and collimated (GBE05-A beam expander, from Thorlabs Inc, Newton, USA) to a beam
diameter of 10 mm. The collimated beam was focused using a wide-field lens (focal
length: 300 mm) to the back aperture of the Nikon x40 (1.2 NA) oil immersion objective via a
dichroic mirror (Ref. Di02-R561-25×36, Semrock Inc, Rochester, USA). The NV fluorescence
was filtered using two bandpass filters before being imaged using a tube lens (focal length:
300 mm) onto a sCMOS camera (Neo, Andor Technology Ltd, Belfast, UK). The optical
configuration relies on a standard widefield optical microscope using laser light excitation source,
with the addition of the AOM for optical modulation. Microwave excitation was provided by a
microwave generator (Ref. N5182A, Agilent Technologies, Santa Clara, USA) and switched using
a RF switch (Ref. ZASWA-2-50DR+, Miniciruits Inc, Brooklyn, USA). The microwaves were
amplified (Ref. 20S1G4, Amplifier Research Corp., Bothell, USA) before being sent to the
microwave resonator. A computer board (Ref. PulseBlasterESR-PRO 500 MHz, SpinCore
Technologies, Inc, Gainesville, USA) was used to control the timing sequences of the excitation
laser, microwaves and sCMOS camera and the images where obtained and analysed using custom
LabVIEW code. Optical imaging was performed at 37 °C unless stated otherwise with an
excitation power density of 30 W/mm2. The temperature studies were undertaken by reducing the
temperature of the environment and monitoring these changes with a k-type thermocouple and
reader (Fluke Corp., Everett, USA).
ASSOCIATED CONTENT
22
The Supporting Information is available free of charge on the ACS Publications website at DOI:
XXXXX. More information on the temperature measurement protocol are provided in the
Supplementary Information, along with dynamic light scatting and zeta potential characterization
of the NDs used in this work (PDF).
AUTHOR INFORMATION
Corresponding Author
*[email protected].
Author Contributions
D.A.S., E.M., S.P., L.C.L.H conceived the study. D.A.S. and E.M. performed the electrical
recordings. E.M cultured the cortical neurons for the multi electrode array recordings and imaging
experiments. D.A.S., J.M.M. and A.L. performed the optical imaging and quantum sensing
experiments. D.A.S. constructed the quantum imaging microscope and acquisition software.
D.C.M. performed the statistical analysis of the electrical recordings. D.A.S., L.T.H. and F.T.
analyzed the data from the quantum measurements. D.A.S., S.P. and L.C.L.H. supervised the
project and all authors contributed to writing the manuscript.
ACKNOWLEDGMENT
The authors acknowledge Prof Michel Simonneau for helpful discussions regarding applications
of the technology. Dr Philipp Reineck for performing the DLS and Zeta potential measurements
on the ND solutions and Dr Babak Nasr for the helium ion microscope images. This research was
supported in part by the Australian Research Council Centre of Excellence for Quantum
Computation and Communication Technology (Project number CE110001027). This work was
23
also supported by the University of Melbourne through the Centre for Neural Engineering and the
Centre for Neuroscience. L.C.L.H acknowledges the support of the Australian Research Council
Laureate Fellowship Scheme (FL130100119). S.P acknowledges the support from the NHMRC
fellowship scheme (1005050). D.A.S acknowledges support from the Melbourne Neuroscience
Institute Fellowship Scheme.
REFERENCES
1. Mathias, R. T.; Cohen, I. S.; Oliva, C., Limitations of the Whole Cell Patch Clamp Technique in the
Control of Intracellular Concentrations. Biophys. J. 1990, 58, 759-770.
2. Spira, M. E.; Hai, A., Multi-Electrode Array Technologies for Neuroscience and Cardiology. Nat.
Nanotechnol. 2013, 8, 83-94.
Neurosci. 2012, 13, 687-700.
3. Knöpfel, T., Genetically Encoded Optical Indicators for the Analysis of Neuronal Circuits. Nat. Rev.
4. Gong, Y.; Huang, C.; Li, J. Z.; Grewe, B. F.; Zhang, Y.; Eismann, S.; Schnitzer, M. J., High-Speed
Recording of Neural Spikes in Awake Mice and Flies with a Fluorescent Voltage Sensor. Science
2015, 350, 1361-1366.
5. Hochbaum, D. R.; Zhao, Y.; Farhi, S. L.; Klapoetke, N.; Werley, C. A.; Kapoor, V.; Zou, P.; Kralj, J.
M.; Maclaurin, D.; Smedemark-Margulies, N.; Saulnier, J. L.; Boulting, G. L.; Straub, C.; Cho, Y. K.;
Melkonian, M.; Wong, G. K.; Harrison, D. J.; Murthy, V. N.; Sabatini, B. L.; Boyden, E. S.; Campbell,
R. E.; Cohen, A. E., All-Optical Electrophysiology in Mammalian Neurons Using Engineered
Microbial Rhodopsins. Nat. Methods 2014, 11, 825-33.
6. Paget, V.; Sergent, J. A.; Grall, R.; Altmeyer-Morel, S.; Girard, H. A.; Petit, T.; Gesset, C.; Mermoux,
M.; Bergonzo, P.; Arnault, J. C.; Chevillard, S., Carboxylated Nanodiamonds Are Neither Cytotoxic
nor Genotoxic on Liver, Kidney, Intestine and Lung Human Cell Lines. Nanotoxicology 2014, 8 Suppl
1, 46-56.
7. Thalhammer, A.; Edgington, R. J.; Cingolani, L. A.; Schoepfer, R.; Jackman, R. B., The Use of
Nanodiamond Monolayer Coatings to Promote the Formation of Functional Neuronal Networks.
Biomaterials 2010, 31, 2097-2104.
8. Huang, Y.-A.; Kao, C.-W.; Liu, K.-K.; Huang, H.-S.; Chiang, M.-H.; Soo, C.-R.; Chang, H.-C.; Chiu,
T.-W.; Chao, J.-I.; Hwang, E., The Effect of Fluorescent Nanodiamonds on Neuronal Survival and
Morphogenesis. Sci. Rep. 2014, 4.
9. Hsu, T.-C.; Liu, K.-K.; Chang, H.-C.; Hwang, E.; Chao, J.-I., Labeling of Neuronal Differentiation
and Neuron Cells with Biocompatible Fluorescent Nanodiamonds. Sci. Rep. 2014, 4.
10. Faklaris, O.; Joshi, V.; Irinopoulou, T.; Tauc, P.; Sennour, M.; Girard, H.; Gesset, C.; Arnault, J. C.;
Thorel, A.; Boudou, J. P.; Curmi, P. A.; Treussart, F., Photoluminescent Diamond Nanoparticles for
Cell Labeling: Study of the Uptake Mechanism in Mammalian Cells. ACS Nano 2009, 3, 3955.
11. Haziza, S.; Mohan, N.; Loe-Mie, Y.; Lepagnol-Bestel, A.-M.; Massou, S.; Adam, M.-P.; Le, X. L.;
Viard, J.; Plancon, C.; Daudin, R.; Koebel, P.; Dorard, E.; Rose, C.; Hsieh, F.-J.; Wu, C.-C.; Potier,
B.; Herault, Y.; Sala, C.; Corvin, A.; Allinquant, B.; Chang, H.-C.; Treussart, F.; Simonneau, M.,
Fluorescent Nanodiamond Tracking Reveals Intraneuronal Transport Abnormalities Induced by Brain-
Disease-Related Genetic Risk Factors. Nat. Nanotechnol. 2017, 12, 322-328.
12. Balasubramanian, G.; Chan, I. Y.; Kolesov, R.; Al-Hmoud, M.; Tisler, J.; Shin, C.; Kim, C.; Wojcik,
A.; Hemmer, P. R.; Krueger, A.; Hanke, T.; Leitenstorfer, A.; Bratschitsch, R.; Jelezko, F.; Wrachtrup,
24
J., Nanoscale Imaging Magnetometry with Diamond Spins under Ambient Conditions. Nature 2008,
455, 648-651.
13. Maze, J. R.; Stanwix, P. L.; Hodges, J. S.; Hong, S.; Taylor, J. M.; Cappellaro, P.; Jiang, L.; Dutt, M.
V. G.; Togan, E.; Zibrov, A. S.; Yacoby, A.; Walsworth, R. L.; Lukin, M. D., Nanoscale Magnetic
Sensing with an Individual Electronic Spin in Diamond. Nature 2008, 455, 644-648.
14. Kucsko, G.; Maurer, P.; Yao, N.; Kubo, M.; Noh, H.; Lo, P.; Park, H.; Lukin, M., Nanometre-Scale
Thermometry in a Living Cell. Nature 2013, 500, 54-58.
15. Karaveli, S.; Gaathon, O.; Wolcott, A.; Sakakibara, R.; Shemesh, O. A.; Peterka, D. S.; Boyden, E. S.;
Owen, J. S.; Yuste, R.; Englund, D., Modulation of Nitrogen Vacancy Charge State and Fluorescence
in Nanodiamonds Using Electrochemical Potential. Proc. Natl. Acad. Sci. U. S. A. 2016, 113, 3938-
3943.
16. Rendler, T.; Neburkova, J.; Zemek, O.; Kotek, J.; Zappe, A.; Chu, Z.; Cigler, P.; Wrachtrup, J., Optical
Imaging of Localized Chemical Events Using Programmable Diamond Quantum Nanosensors. Nat.
Commun. 2017, 8, 14701.
17. Johnstone, A. F.; Gross, G. W.; Weiss, D. G.; Schroeder, O. H.; Gramowski, A.; Shafer, T. J.,
Microelectrode Arrays: A Physiologically Based Neurotoxicity Testing Platform for the 21st Century.
Neurotoxicology 2010, 31, 331-50.
18. Hogberg, H. T.; Sobanski, T.; Novellino, A.; Whelan, M.; Weiss, D. G.; Bal-Price, A. K., Application
of Micro-Electrode Arrays (Meas) as an Emerging Technology for Developmental Neurotoxicity:
Evaluation of Domoic Acid-Induced Effects in Primary Cultures of Rat Cortical Neurons.
Neurotoxicology 2011, 32, 158-68.
19. Knowles, H. S.; Kara, D. M.; Atatüre, M., Observing Bulk Diamond Spin Coherence in High-Purity
Nanodiamonds. Nat. Mater. 2014, 13, 21-25.
20. Hall, L. T.; Beart, G. C. G.; Thomas, E. A.; Simpson, D. A.; McGuinness, L. P.; Cole, J. H.; Manton,
J. H.; Scholten, R. E.; Jelezko, F.; Wrachtrup, J.; Petrou, S.; Hollenberg, L. C. L., High Spatial and
Temporal Resolution Wide-Field Imaging of Neuron Activity Using Quantum Nv-Diamond. Sci. Rep.
2012, 2, 1.
21. Barry, J. F.; Turner, M. J.; Schloss, J. M.; Glenn, D. R.; Song, Y.; Lukin, M. D.; Park, H.; Walsworth,
R. L., Optical Magnetic Detection of Single-Neuron Action Potentials Using Quantum Defects in
Diamond. Proc. Natl. Acad. Sci. U. S. A. 2016, 113, 14133-14138.
22. McGuinness, L. P.; Yan, Y.; Stacey, A.; Simpson, D. A.; Hall, L. T.; Maclaurin, D.; Prawer, S.;
Mulvaney, P.; Wrachtrup, J.; Caruso, F.; Scholten, R. E.; Hollenberg, L. C. L., Quantum Measurement
and Orientation Tracking of Fluorescent Nanodiamonds inside Living Cells. Nat. Nanotechnol. 2011,
6, 358-363.
23. Donner, J. S.; Thompson, S. A.; Kreuzer, M. P.; Baffou, G.; Quidant, R., Mapping Intracellular
Temperature Using Green Fluorescent Protein. Nano Lett. 2012, 12, 2107-2111.
24. Tanimoto, R.; Hiraiwa, T.; Nakai, Y.; Shindo, Y.; Oka, K.; Hiroi, N.; Funahashi, A., Detection of
Temperature Difference in Neuronal Cells. Sci. Rep. 2016, 6, 22071.
25. Liu, H.; Fan, Y.; Wang, J.; Song, Z.; Shi, H.; Han, R.; Sha, Y.; Jiang, Y., Intracellular Temperature
Sensing: An Ultra-Bright Luminescent Nanothermometer with Non-Sensitivity to Ph and Ionic
Strength. Sci. Rep. 2015, 5, 14879.
26. Takei, Y.; Arai, S.; Murata, A.; Takabayashi, M.; Oyama, K.; Ishiwata, S. i.; Takeoka, S.; Suzuki, M.,
A Nanoparticle-Based Ratiometric and Self-Calibrated Fluorescent Thermometer for Single Living
Cells. ACS Nano 2014, 8, 198-206.
27. Shang, L.; Stockmar, F.; Azadfar, N.; Nienhaus, G. U., Intracellular Thermometry by Using
Fluorescent Gold Nanoclusters. Angewandte Chemie 2013, 52, 11154-7.
28. Acosta, V. M.; Bauch, E.; Ledbetter, M. P.; Waxman, A.; Bouchard, L. S.; Budker, D., Temperature
Dependence of the Nitrogen-Vacancy Magnetic Resonance in Diamond. Phys. Rev. Lett. 2010, 104,
070801.
25
29. Gruber, A.; Drabenstedt, A.; Tietz, C.; Fleury, L.; Wrachtrup, J.; vonBorczyskowski, C., Scanning
Confocal Optical Microscopy and Magnetic Resonance on Single Defect Centers. Science 1997, 276,
2012-2014.
30. Van Oort, E.; Glasbeek, M., Electric-Field-Induced Modulation of Spin Echoes of Nv Centers in
Diamond. Chem. Phys. Lett. 1990, 168, 529-532.
31. Igarashi, R.; Yoshinari, Y.; Yokota, H.; Sugi, T.; Sugihara, F.; Ikeda, K.; Sumiya, H.; Tsuji, S.; Mori,
I.; Tochio, H.; Harada, Y.; Shirakawa, M., Real-Time Background-Free Selective Imaging of
Fluorescent Nanodiamonds in Vivo. Nano Lett. 2012, 12, 5726-5732.
32. Trusheim, M. E.; Li, L.; Laraoui, A.; Chen, E. H.; Gaathon, O.; Bakhru, H.; Schroder, T.; Meriles, C.
A.; Englund, D., Scalable Fabrication of High Purity Diamond Nanocrystals with Long-Spin-
Coherence Nitrogen Vacancy Centers. Nano Lett. 2013.
33. Maletinsky, P.; Hong, S.; Grinolds, M. S.; Hausmann, B.; Lukin, M. D.; Walsworth, R. L.; Loncar,
M.; Yacoby, A., A Robust Scanning Diamond Sensor for Nanoscale Imaging with Single Nitrogen-
Vacancy Centres. Nat. Nanotechnol. 2012, 7, 320.
34. Tzeng, Y.-K.; Tsai, P.-C.; Liu, H.-Y.; Chen, O. Y.; Hsu, H.; Yee, F.-G.; Chang, M.-S.; Chang, H.-C.,
Time-Resolved Luminescence Nanothermometry with Nitrogen-Vacancy Centers in Nanodiamonds.
Nano Lett. 2015, 15, 3945-3952.
35. Baffou, G.; Rigneault, H.; Marguet, D.; Jullien, L., A Critique of Methods for Temperature Imaging
in Single Cells. Nat. Methods 2014, 11, 899-901.
36. Suzuki, M.; Zeeb, V.; Arai, S.; Oyama, K.; Ishiwata, S. i., The 105 Gap Issue between Calculation and
Measurement in Single-Cell Thermometry. Nat. Methods 2015, 12, 802-803.
37. Chen, D. P.; Eisenberg, R. S.; Jerome, J. W.; Shu, C. W., Hydrodynamic Model of Temperature
Change in Open Ionic Channels. Biophys. J. 1995, 69, 2304-2322.
38. Balasubramanian, G.; Neumann, P.; Twitchen, D.; Markham, M.; Kolesov, R.; Mizuochi, N.; Isoya,
J.; Achard, J.; Beck, J.; Tissler, J.; Jacques, V.; Hemmer, P. R.; Jelezko, F.; Wrachtrup, J., Ultralong
Spin Coherence Time in Isotopically Engineered Diamond. Nat. Mater. 2009, 8, 383-387.
39. Stujenske, J. M.; Spellman, T.; Gordon, J. A., Modeling the Spatiotemporal Dynamics of Light and
Heat Propagation for in Vivo Optogenetics. Cell reports 2015, 12, 525-534.
40. Hall, L. T.; Kehayias, P.; Simpson, D. A.; Jarmola, A.; Stacey, A.; Budker, D.; Hollenberg, L. C. L.,
Detection of Nanoscale Electron Spin Resonance Spectra Demonstrated Using Nitrogen-Vacancy
Centre Probes in Diamond. Nat. Commun. 2016, 7, 10211.
41. Ermakova, A.; Pramanik, G.; Cai, J. M.; Algara-Siller, G.; Kaiser, U.; Weil, T.; Tzeng, Y. K.; Chang,
H. C.; McGuinness, L. P.; Plenio, M. B.; Naydenov, B.; Jelezko, F., Detection of a Few Metallo-
Protein Molecules Using Color Centers in Nanodiamonds. Nano Lett. 2013, 13, 3305-3309.
42. Kaufmann, S.; Simpson, D. A.; Hall, L. T.; Perunicic, V.; Senn, P.; Steinert, S.; McGuinness, L. P.;
Johnson, B. C.; Ohshima, T.; Caruso, F.; Wrachtrup, J.; Scholten, R. E.; Mulvaney, P.; Hollenberg, L.,
Detection of Atomic Spin Labels in a Lipid Bilayer Using a Single-Spin Nanodiamond Probe. Proc.
Natl. Acad. Sci. U. S. A. 2013, 110, 10894-10898.
43. Simpson, D. A.; Ryan, R. G.; Hall, L. T.; Panchenko, E.; Drew, S. C.; Petrou, S.; Donnelly, P. S.;
Mulvaney, P.; Hollenberg, L. C. L., Electron Paramagnetic Resonance Microscopy Using Spins in
Diamond under Ambient Conditions. Nat. Commun. 2017, 8, 458.
44. Shi, F.; Zhang, Q.; Wang, P.; Sun, H.; Wang, J.; Rong, X.; Chen, M.; Ju, C.; Reinhard, F.; Chen, H.;
Wrachtrup, J.; Wang, J.; Du, J., Single-Protein Spin Resonance Spectroscopy under Ambient
Conditions. Science 2015, 347, 1135-1138.
45. Mamin, H. J.; Kim, M.; Sherwood, M. H.; Rettner, C. T.; Ohno, K.; Awschalom, D. D.; Rugar, D.,
Nanoscale Nuclear Magnetic Resonance with a Nitrogen-Vacancy Spin Sensor. Science 2013, 339,
557-560.
46. Staudacher, T.; Shi, F.; Pezzagna, S.; Meijer, J.; Du, J.; Meriles, C. A.; Reinhard, F.; Wrachtrup, J.,
Nuclear Magnetic Resonance Spectroscopy on a (5-Nanometer)3 Sample Volume. Science 2013, 339,
561-563.
26
47. Wood, J. D. A.; Tetienne, J.-P.; Broadway, D. A.; Hall, L. T.; Simpson, D. A.; Stacey, A.; Hollenberg,
L. C. L., Microwave-Free Nuclear Magnetic Resonance at Molecular Scales. Nat. Commun. 2017, 8,
15950.
48. Maccione, A.; Gandolfo, M.; Massobrio, P.; Novellino, A.; Martinoia, S.; Chiappalone, M., A Novel
Algorithm for Precise Identification of Spikes in Extracellularly Recorded Neuronal Signals. J.
Neurosci. Methods 2009, 177, 241-9.
49. Mendis, G.; Morrisroe, E.; Petrou, S.; Halgamuge, S., Use of Adaptive Network Burst Detection
Methods for Multielectrode Array Data and the Generation of Artificial Spike Patterns for Method
Evaluation. Journal of neural engineering 2016, 13, 026009.
27
|
1109.6730 | 1 | 1109 | 2011-09-30T07:26:14 | Structural Rigidity of Paranemic (PX) and Juxtapose (JX) DNA Nanostructures | [
"physics.bio-ph",
"cond-mat.soft"
] | Crossover motifs are integral components for designing DNA based nanostructures and nanomechanical devices due to their enhanced rigidity compared to the normal B-DNA. Although the structural rigidity of the double helix B-DNA has been investigated extensively using both experimental and theoretical tools, to date there is no quantitative information about structural rigidity and the mechanical strength of parallel crossover DNA motifs. We have used fully atomistic molecular dynamics simulations in explicit solvent to get the force-extension curve of parallel DNA nanostructures to characterize their mechanical rigidity. In the presence of mono-valent Na+ ions, we find that the stretch modulus (\gamma_1) of the paranemic crossover (PX) and its topo-isomer JX DNA structure is significantly higher (~ 30%) compared to normal B-DNA of the same sequence and length. However, this is in contrast to the original expectation that these motifs are almost twice rigid compared to the double-stranded B-DNA. When the DNA motif is surrounded by a solvent with Mg2+ counterions, we find an enhanced rigidity compared to Na+ environment due to the electrostatic screening effects arising from the divalent nature of Mg2+ ions. This is the first direct determination of the mechanical strength of these crossover motifs which can be useful for the design of suitable DNA for DNA based nanostructures and nanomechanical devices with improved structural rigidity. | physics.bio-ph | physics |
Structural Rigidity of Paranemic (PX) and
Juxtapose (JX) DNA Nanostructures
Mogurampelly Santosh
Center for Condensed Matter Theory, Department of Physics,
Indian Institute of Science, Bangalore, India 560012.
E-mail: [email protected]
Prabal K. Maiti∗
Center for Condensed Matter Theory, Department of Physics,
Indian Institute of Science, Bangalore, India 560012.
E-mail: [email protected]
∗Corresponding author. Address: Center for Condensed Matter Theory, De-
partment of Physics, Indian Institute of Science, Bangalore, India, 560012. E-mail:
[email protected], Tel.: (0091)80-22932865
Abstract
Crossover motifs are integral components for designing DNA based nanos-
tructures and nanomechanical devices due to their enhanced rigidity com-
pared to the normal B-DNA. Although the structural rigidity of the double
helix B-DNA has been investigated extensively using both experimental and
theoretical tools, to date there is no quantitative information about struc-
tural rigidity and the mechanical strength of parallel crossover DNA motifs.
We have used fully atomistic molecular dynamics simulations in explicit
solvent to get the force-extension curve of parallel DNA nanostructures to
characterize their mechanical rigidity. In the presence of mono-valent Na+
ions, we find that the stretch modulus (γ1) of the paranemic crossover (PX)
and its topo-isomer JX DNA structure is significantly higher (∼ 30 %) com-
pared to normal B-DNA of the same sequence and length. However, this
is in contrast to the original expectation that these motifs are almost twice
rigid compared to the double-stranded B-DNA. When the DNA motif is
surrounded by a solvent with Mg2+ counterions, we find an enhanced rigid-
ity compared to Na+ environment due to the electrostatic screening effects
arising from the divalent nature of Mg2+ ions. This is the first direct de-
termination of the mechanical strength of these crossover motifs which can
be useful for the design of suitable DNA for DNA based nanostructures and
nanomechanical devices with improved structural rigidity.
Key words: PX/JX DNA motif; Atomistic Simulations; Force-extension
Curves; Stretch Modulus; Torsion Angles
Structural Rigidity of PX and JX DNA
2
1
Introduction
The structural properties that enable DNA to serve so effectively as genetic
material can also be used for DNA based computation and for DNA nan-
otechnology (1). The success of these applications to large extent, depends
on the complementarity that leads to the pairing of the strands of the DNA
double helix. This pairing can be exploited to assemble more complex mo-
tifs, based on branched structures with sticky ends that are promising for
building macromolecular structures (1, 2). The combination of sticky ended
ligation (3) and stable branched DNA species (4) has permitted investiga-
tors to construct DNA molecules whose edges have connectivity of a cube (5)
and of a truncated octahedron (6). Similarly, branched DNA molecules have
provided the basis for the syntheses of variety of structures such as knots
(7, 8), Borromean rings (9) and DNA tweezers (10). One of the early goals
of DNA nanotechnology was to construct precisely configured materials on
a much larger scale (1, 11). However the key feature that was lacking in
such constructions were rigid DNA molecules, as flexible components failed
to maintain the same spatial relationships between each member of a set
(12). DNA anti-parallel double crossover molecules (DX) (13), analogous
to intermediates in genetic recombination, were found to be about twice as
rigid as a normal DNA molecules (14) and were subsequently used for cre-
ating two dimensional DNA lattices (15).
The concept of double crossover molecules stems from the concept of
formation of Holliday junction (16), which is the most prominent interme-
diate in genetic recombination. These crossover molecules usually contain
two Holliday junctions connected by two double helical arms. Depending
on whether the crossovers are between strands of the same (called parallel)
or opposite (called anti-parallel) polarity, as well as the number of dou-
ble helical half-turns (even or odd) between the two crossovers, different
double crossover molecules are generated.
It is known that anti-parallel
double helical molecules are better behaved than the parallel ones and as
mentioned above these anti-parallel molecules have found numerous appli-
cations in nano construction (17). The concept of crossover molecules also
led to the discovery of paranemic crossover molecules (PX) and its topoi-
somers, the various JX molecules. The PX motif arises from the fusion of
two parallel double helices by reciprocal exchange at every possible point
where the two helices come in contact, whereas its topoisomer JXi contains
i adjacent sites where backbones juxtapose without crossing over (18). The
rigidity of these crossover motifs has led to wide applicability in DNA based
Structural Rigidity of PX and JX DNA
3
nanotechnology (19).
Nevertheless, there is lack of quantitative knowledge about the rigidity of
the paranemic PX DNA and its topoisomers JX DNA, even though they find
applications in nano constructions, owing to their inherent rigidity. Using
atomistic simulation studies on PX and JX structures we have shown that
among all synthesized PX and JX DNA structures, PX-6:5 motif is the most
stable and in comparison with JX motifs the rigidity order runs as PX-6:5
> JX1 > JX2 > JX3 > JX4 (20, 21). Our predictions on thermodynamic
calculations are in consistence with the recent experimental studies using
calorimetric and denaturing gradient gel electrophoretic methods (22). In
this report we make an attempt to get a quantitative measure of the rigidity
of these structures from their force-extension curves. The concept that has
been adapted is to pull these structures with an external force in constant-
force ensemble (23) and from the internal energy as a function of extension
of the DNA motif, we obtain the stretch modulus using Hooke's law of elas-
ticity. The stretch modulus in turn gives an estimate of the rigidity of the
motifs.
Experimental and theoretical studies have established that when sub-
jected to an external force, double-stranded DNA (dsDNA) exhibits different
force-extension regimes, stemming from its unique double-helix structures
(24 -- 27). For example, in the low force regime, the elasticity of dsDNA is
entropy dominated and the experimental force-extension data obtained with
the magnetic bead method can be described very well by the standard en-
tropic worm-like chain model (24, 28, 29); in the high force regime (starting
from 10 pN to 70 pN), which is accessible with optical tweezers or atomic
force microscopy, when the external force is comparable with the basepair
stacking interaction in dsDNA the polymers can be suddenly stretched to
about 1.7 times its B-form length in a very narrow force range (30, 31). An
explanation of this regime is attributed to the short-range nature of base-
pair stacking interactions. Typically in this regime, the increase in length of
the DNA molecule is due to large distortion in the double helical structure
resulting in a ladder-like structure which is termed as S-DNA (31 -- 34) where
most of the H-bonds involving basepairing are intact. It has been shown that
depending on the experimental conditions such as the force attachment and
salt concentration, dsDNA can undergo either strand-unpeeling transition
to ssDNA (35, 36), or transition to S-DNA where basepairs are intact (or
B-to-S transition) (31, 37). In the elastic regime of force-extension curve,
DNA behaves similar to that of a spring and follows the simple Hooke's Law:
Structural Rigidity of PX and JX DNA
4
F = Y A∆L/L where Y is the Young Modulus and A is the cross-sectional
area of the DNA. The product of Young modulus and area, Y × A gives
stretch modulus, γ1. At large forces, the stacking potential can no longer
stabilize the B-form conformation of dsDNA and the (optimally) stacked
helical pattern is severely distorted (38), and therefore a structural transi-
tion from canonical B-form to a new overstretched conformation. While all
these features are well established for dsDNA and ssDNA, no theoretical
or experimental studies exist for the force-extension behavior of cross-over
DNA molecules such as DX, PX and various JX structures. This paper is
the first attempt to address such lacunas. The paper is organized as follows:
in the next section we describe building of PX/JX nanostructures and the
simulation methodologies. In section 3 we describe the results from our fully
atomistic DNA pulling simulation and in section 4 we give a summary of
the force-extension results and conclude.
2 Computational Details and Methodology
2.1 Building PX/JX structures
In order to characterize the stiffness of various cross-over DNA motifs, the-
oretically we have tried to mimic nano-manipulation techniques like AFM,
magnetic tweezers, or laser tweezers in our fully atomistic simulations. The
basic protocol in our simulation technique was to put the DNA structure in
counterion solution, keep one end of the structure anchored, and pull the
other end along the PX molecule dyad axis (parallel to the helix axes) with
an external force and determine the force-extension curves. The various
DNA structures studied are PX-6:5, JX1, JX2, AT-6:5, GC-6:5 and normal
B-DNA which are four-strand complexes of DNA paired with one another.
Initial structures of PX-6:5, JX1 and JX2 were shown in Fig. 1. We have
also studied the force-extension behavior of a double strand DNA having
the same length (38 bps) and the sequence of one of the helices of the PX
structure. This will allow us to compare the force-extension behavior of the
normal double strand B-DNA with that of the cross-over DNA nanostruc-
tures. The procedure for constructing these structures is as follows:
a. Building the DNA double helices: Regular B-DNA has between 10 and
10.5 base pairs per helical turn. Hence by varying the twist angle of
a selected number of base pairs, we can create B-DNA structures with
between 8 and 18 base pairs per helical turn. Table 1 of (20) shows the
Structural Rigidity of PX and JX DNA
5
twist angles used for building the various PX structures. We assigned
the same twist angle for all the base pairs in the helical half turn. The
helical rise value of 3.4 A was used to build the PX structures.
b. Building the crossover points: When a double helix is built in Namot2
(39), the molecules are oriented so that the 5′ and 3′
(version 2.2.)
ends of the double helices are parallel to the y-axis. To create realistic
crossover structures, it is necessary to rotate the individual helices so
that the desired crossover points are closest to each other (rotation an-
gles shown in Table 2 of (20)). To find this point we wrote a program
that starts with the first crossover point and rotates the first helix in
1◦ increment to find the rotation leading to the shortest distance be-
tween these crossover points. Once found, the first helix is rotated by
the prescribed value and held steady while the second helix is rotated
and the shortest distance between the crossover points determined. The
second helix is rotated 180◦ more than the 1st helix so that the helices
are arranged as shown in Fig 1. The crossovers were then created using
the "nick" and "link" commands in Namot2. These structures are saved
in the Protein Database (PDB) file format. Fig. 1 shows the snapshot
of the built PX, JX1 and JX2 structures after minimizations.
2.2 Simulation Details for the PX/JX structures
All molecular dynamics simulations were carried out using AMBER8 suite
of programs (40) using the Amber 2003 (ff03) force fields (41, 42) and the
TIP3P model (43) for water. For Mg2+ we have used the Aqvist (44) inter-
action parameters. This initial crossover DNA motif is then solvated with
TIP3P model (43) water box using the LEaP module in AMBER8 (40).
In addition, some water molecules were replaced by Na+ (Mg2+) counte-
rions to neutralize the negative charge on the phosphate backbone groups
of the PX/JX DNA structure. We have used Aqvist parameter (44) set
to describe the ion-water as well as ion-DNA interactions. Aqvist (44) pa-
rameters reproduce the location of the first peak of the radial distribution
function for the corresponding ion-water and the solvation free energies of
several ionic species. After neutralizing the system with counterions, the
concentration of Na+ (Mg2+) in the PX-6:5, JX1, JX2, AT-6:5, GC:6-5 and
normal B-DNA crossover systems is 246 mM (160 mM) where as for a double
helix B-DNA, the concentration of Na+ (Mg2+) is 254 mM (140 mM). The
ion specificity and electrostatic interactions play crucial in the DNA func-
tioning and protein-DNA binding mechanism (45, 46). The LEaP module
Structural Rigidity of PX and JX DNA
6
works by constructing Coulombic potential on a grid of 1 A resolution and
then placing ions one at a time at the highest electrostatic potential. Once
the placements of all the ions are done using the above method long MD
simulations ensure that they sample all the available space around DNA
and preferentially visit electronegative sites. In fact the initial placement
of ions should not influence the final results provided long simulation was
performed before the pulling runs were done. The radial distribution func-
tion provided in the Fig. S1 and S2 of supplementary material also reveal
this fact that the counterions remain associated in the close proximity (∼
10 A) of the DNA. We have also made sure that no counterion is "stuck"
to the DNA molecule anywhere and observed diffusive behavior. Hence-
forth we believe that the initial ion distribution is important to study DNA
properties but ultimately equilibration is the key for avoiding such initial
ion position dependency. The box dimensions were chosen in order to en-
sure a 10 A solvation shell around the DNA structure in its fully extended
form when the DNA is in overstretched regime. This procedure resulted
in solvated structures, containing 99734 atoms for PX-6:5, JX1 and JX2;
97326 atoms for 38-mer B-DNA; 99781 atoms for PX-6:5 structure having
only AT sequence and 99705 atoms for PX-6:5 structure having only GC
3
sequence in a box of dimensions 54 × 77 × 299 A
when Na+ ions are present
as counterions. Translational center of mass motions were removed every
1000 MD steps. The trajectory was saved at a frequency of 2 ps. We have
used periodic boundary conditions in all three directions for the water box
during the simulation. Bond lengths involving bonds to hydrogen atoms
were constrained using SHAKE algorithm (47). During the minimization,
the PX/JX structure was fixed in their starting conformations using har-
monic potential with a force constant of 500 kcal/mol-A2. This allowed the
water molecules to reorganize to eliminate bad contacts within the system.
The minimized structures were then subjected to 40 ps of MD, using 1 fs
time step for integration. During the MD, the system was gradually heated
from 0 to 300 K using weak 20 kcal/mol-A2 harmonic constraints on the
solute to its starting structure. NPT-MD was used to get the correct solvent
density corresponding to experimental value of the density. Lastly, we have
carried out pulling the motif in NVT MD with 2 fs integration time step
using a heat bath coupling time constant of 1 ps.
The external pulling force was applied at one end on O3′ and O5′ atoms
on each strand and the other O5′ and O3′ atoms of the strands on the other
end were fixed with large force constant of 5000 kcal/mol-A2. The force
Structural Rigidity of PX and JX DNA
7
was applied along the direction of the end-to-end vector joining O3′ and O5′
atoms. The external force started at 0 pN and increased linearly with time
up to 1000 pN with a rate of force of 10−4 pN/fs. For comparison, we have
also done the simulation with faster pulling rate at 10−3 pN/fs. It should
be mentioned that the typical pulling rate in an AFM experiment is of the
order of 104 pN/s whereas the slowest pulling rate achieved in our simula-
tion is 1011 pN/s (or 10−4 pN/fs) due to computing limitations. Theoretical
model suggests that with the increased rate of force, the dynamic strength
of the molecular adhesion bonds increases (48). Therefore, since our sim-
ulation employs higher rates of force compared to AFM pulling rates, the
calculated stretch modulus will be higher than that of the stretch modulus
calculated in an AFM pulling experiment. Our pulling protocol was vali-
dated for B-DNA (23) and is expected to be applicable for PX/JX motifs
also. In Fig. S3(a) we have shown the force-extension curve that is consis-
tent with experimental curve. Note that the plateau region was observed at
95 pN instead of 65 pN which has its origin in fast rate of pulling which is
inherent to computer simulations. This MD simulation has been done with
a pulling rate of 10−5 pN/fs, the slowest pulling rate we could achieve as
of now. We have also done the pulling simulation at various rates. In Fig.
S3(b) we plot the plateau force as a function of the pulling rate. From this
plot we can see that by extrapolating to slower pulling rates, one can get
a plateau region like that of AFM pulling experiments. Hence we expect
that we can exactly get the experimental force-extension curve when pulled
much slower than 10−5 pN/fs such as in AFM or optical tweezers. Apart
from the rate dependency, the stretch modulus calculation is in accordance
with the experimental result.
3 Results and Discussion
3.1 Force-extension behavior
Fig. 2 gives the force-extension curve for PX-6:5 DNA motif in the pres-
ence of Na+ counterions. The force-extension curve consists of an entropic
regime where, the extension of DNA beyond its contour length is negligi-
ble and this regime continues until 150 pN. Beyond this forces and up to
200 pN, DNA extends to about 30 % with most of the H-bonds are still
intact. It has been shown that depending on the experimental conditions
such as the force attachment and salt concentration, dsDNA can undergo ei-
ther strand-unpeeling transition to ssDNA (35, 36), or transition to S-DNA
Structural Rigidity of PX and JX DNA
8
where basepairs are intact (or B-to-S transition) (31, 37). This is followed
by an overstretching plateau region resulting in an extension of 1.7 times
the initial contour length where the DNA helical structure starts to deform.
Beyond this plateau regime is the overstretched structure of DNA which
is followed by strand separation. It is worth mentioning that at low rates
of force, the above mentioned regimes in the force-extension curves shifts
to much lower force values. For comparison we have also shown the force-
extension behavior of the normal double strand B-DNA which has same
length and sequence as one of the helices of the PX-6:5 structure. From the
force-extension curve we can calculate the stretch modulus of DNA in the
elastic regime using Hooke's Law. However, it is known from previous theo-
retical and experimental studies on the DNA force-extension behavior that
at high force regimes a conventional WLC model and its variant does not
reproduce the force-extension behavior very well. Previous theoretical stud-
ies have demonstrated that at high force regime, non-linear elasticity plays
an important role governing the force-extension behavior of the DNA (49)
and the WLC model is inadequate to describe the force-extension behavior
(50, 51). Therefore, we have also used the following polynomial function to
fit the simulated force-extension curve for whole force regime
F =
∞
Xn=1
γn(cid:18) L
L0
− 1(cid:19)n
=
∞
Xn=1
γnεn
Integrating the above equation, we get energy
E = E0 + L0
∞
Xn=2
γn−1
n (cid:18) L
L0
− 1(cid:19)n
= E0 + L0
∞
Xn=2
γn−1
n
εn
(1)
(2)
Where E0 is the energy of the DNA when the extension is zero, L is the
length of the DNA and L0 is the initial contour length of the DNA. The coef-
ficients γn for various values of n gives various elastic modulii. For example,
γ1 gives the linear stretch modulus. Stretch modulus γ1 can be calculated
from both the above equations by fitting the force and energy as a function
of the applied strain. From the above two equations, it can be seen that
the linear term coefficient and the quadratic term coefficient will give the
stretch modulus, respectively. We have done a sequential fit to the energy
vs strain curve similar to ref (49) where the fitting is done for the leading
quadratic form of energy vs strain plot for restricted data that are taken
around the energy minima and obtain the stretch modulus as a coefficient
of the quadratic term. By giving more weightage to the quadratic term, we
Structural Rigidity of PX and JX DNA
9
have done the higher order fit due to the difficulty in identifying the exact
Hooke's law region. The term 'sequential' essentially means that the leading
terms near the energy minima are fit to quadratic term and the rest are fit
to higher order terms. This procedure is repeated self consistently so that
the value of γ1 is independent of the γn for large n. However γ1 depends
very weakly on γ2 and γ3 (that means γ1 has lesser dependence on choice
of γn for n > 1). The energy vs strain curve has a minimum around which
a quadratic fitting was done and the obtained γ1 is the stretch modulus as
shown in equation 2. The stretch modulus obtained from a polynomial fit
to the force-extension curve and also from a sequential fit to the energy-
strain curve are listed in table 1. The values of stretch modulus from the
above two fitting methods are very similar. From the sequential fit to the
energy-strain curve, we find that PX-6:5 has a stretch modulus of 1636 pN
at 246 mM of Na+ concentration when pulled with 10−4 pN/fs rate, which
is 30 % more rigid compared to the stretch modulus 1269 pN of B-DNA at
same concentration and of the same sequence. From the polynomial fit to
the force-extension curve, the same trend was observed (Table 1). During
the pulling the bond lengths and bond angles (Fig. 3) are changing very
less whereas the torsion angles (Fig. 3) are changing almost 100 % with
respect to the initial values at zero force (See next para for more discus-
sion). This implies that the backbone helix has large internal dynamics and
offer less resistance to the applied force. Note that the backbone helix gives
the structural stability to the DNA molecule. On the other hand, bond
lengths and bond angles offer more resistance to the applied force. Hence
the crossover links between helical strands contribute more to the structural
rigidity of these molecules. Although 30 % of rigidity increase seems small,
yet this enables one to construct promising DNA nanotechnology devices
with enhanced rigidity. For the JX1, JX2 topoisomers at the same coun-
terion concentration we obtain 1515 pN, 1349 pN from sequential fitting
and 1373 pN, 1521 pN from polynomial fitting respectively. Force-extension
plot for PX-6:5, JX1 and JX2 is shown in Fig. 2(b) and can see a slight
change in the slope of linear region. Among all structures, PX-6:5 has large
stretch modulus both in Na+ and Mg2+ medium due to the most number
of crossover points. It was also shown that the optimal design of repeated
stacks and bundles of nanostructures provide great strength to the molecule
(52). Experimental reports of stretch modulus for lambda phase DNA is of
the order of 1000 pN depending on the environmental conditions like ionic
strength (24, 53). Though we get the stretch modulus of the same order,
the helix to ladder transition occurs at a much larger force regime than that
observed experimentally (54, 55) due to the higher rates of force employed
Structural Rigidity of PX and JX DNA
10
in our simulation. Experimental results have shown that the ladder trans-
formation occurs at a much lower (∼ 60-70 pN) force than that observed in
our simulation data. It has been observed that a higher force rate leads to
higher stiffness of the short DNA (Fig. 4). This shift of curve and increase
of slopes in force-extension curves is expected since higher rates means that
our simulation is far from being reversible, in which case it dissipates energy
to the environment (since our simulations were constant T = 300 K) and
higher the irreversibility the higher is the work of dissipation which is given
by the area under the force-extension curve. A possible reason may also be
the length of DNA that has been used. Since a 38-mer DNA is too short, its
contour length being significantly smaller than the persistence length of a
lambda DNA, its behavior is likely to deviate from that observed for lambda
DNA (24, 53).
To gain further microscopic understanding of the structural changes with
the applied force we have looked at the energetics of the PX/JX DNA as
a function of the applied force. In Fig. 5 we plot the total internal energy
as a function of force for PX-6:5, JX1 and JX2 DNA structures. The con-
formational entropy of the structure is dominating for small forces during
which the change in length is almost negligible. For forces up to 150 pN,
the energy is decreasing and seems to attain minimum value which corre-
sponds to a most stable configuration. With further increase in force above
150 pN, the energy of the DNA molecule increases. This implies that the
PX/JX DNA structures become thermodynamically unstable under elon-
gation. Various backbone parameters were calculated when PX-6:5 struc-
ture is pulled at one end when the structure is neutralized with Na+ and
−P(cid:1) and average angles
Mg2+ counterions. The average bond length (cid:0)rO3′
(θP-O5′-C5′ and θC5′-C4′-C3′) were changing less than 2 % with applied force
with respect to the zero force equilibrium values (Fig. 3). Correspondingly,
the increase in the energy contribution from bond stretching and angle bend-
ing is about to 5-10 % compared to the zero force case. Interestingly, all
the torsion angles (α, β, γ, δ, ǫ, ζ) (56) were changing over 100 % (Fig. 3)
with respect to the values at zero force, which is causing the structural de-
formation to great extent. The change in torsion angles is not considerable
for forces up to 200 pN during which the change in DNA contour length is
also very less. Dramatic change is observed in the torsion angles when the
DNA motif elongates to almost twice of its initial contour length at critical
force. At this stage, except the breaking of few H-bonds (23), we see no
considerable change in average bond length or average angle. We have used
geometry measurement based criteria for the H-bond calculation in simula-
Structural Rigidity of PX and JX DNA
11
tion. Generally the H-bond is represented as D−H ...A where D is the donor
and A is the acceptor which is bonded to D through the H atom. In the
case of DNA, D is a N atom and A is either a N or O atoms depending on
AT and GC base pairing. When the distance between D and A atoms is less
than 2.7 A and the angle DHA is greater than 130◦, we say that the atom A
is H-bonded to atom D otherwise the H-bond is broken. It is justifiable to
argue that the backbone atoms in DNA motif were drastically re-oriented
to give the elastic rod a sudden elongation, a clear signature indicating very
large change in torsion angles. The torsion energy is also increasing greatly
compared to all other contributions to the total internal energy of the motif.
So a closer look at the various energy contributions reveals that, when the
DNA motif is pulled along the helix axis, there is an increase of about 5-10
% in the bond stretching and angle bending energy, 15 % increase in van
der Waals energy but dramatic increase in the torsion energy. Apart from
all the energy components like bond, angle, van der Waals, Coulomb energy,
this increase in the torsion energy contributes to rapid increase in the total
internal energy at a critical force. Instantaneous snapshots of PX-6:5 struc-
ture at different force in the presence of Na+ counterions is shown in Fig. 6
(see Fig. S4-S13 for instantaneous snapshots of PX-6:5, JX1 and JX2 in the
presence of Na+ and Mg2+ at various forces, respectively). As the applied
force is increased beyond 200 pN, H-bonds in the basepair were broken and
the DNA overstretches to 1.7 times its initial contour length.
The original experiments on the PX/JX DNA molecules were performed
in Mg2+ buffer (18). To understand the pulling response of these DNA
nanostructures with divalent cations we have also done the pulling simula-
tion of PX-6:5, JX1 and JX2 structure in the presence of divalent Mg2+ ions.
The counterion concentration in our simulation is close to 160 mM and the
rate of pulling is 10−4 and 10−3 pN/fs. We have analyzed the force-extension
spectrum of all DNA structures in Na+ and Mg2+ ions at 10−4 pN/fs (Fig.
7). Also, Fig. S14(a) shows the force-extension behavior of PX-6:5 molecule
in presence of Mg2+ ions. In the presence of Mg2+, PX-6:5 have a stretch
modulus of 1840 pN with a pulling rate of 10−4 pN. In contrast, the stretch
modulus of B-DNA of same length and sequence as one of the helical do-
mains of PX-6:5 is 1590 pN. Again we see the enhanced rigidity of the PX-6:5
motif as was in the presence of Na+ ions. However, this increase is not as
dramatic as in the presence of Na+ ions. Possibly the rigidity can be further
enhanced by increasing the ionic strength. For all the structures we find the
stretch modulus is more in the presence of Mg2+ than in the presence of
Na+ counterions. This is due to the strong phosphate-Mg2+ coordination
Structural Rigidity of PX and JX DNA
12
that resist the external applied force. This result is in accordance with the
experimental results on DNA stretch modulus (53) where it is shown that
the presence of divalent cations strongly reduce the persistence length of the
DNA and hence increase the stretch modulus of the DNA. To understand
the effect of crossover on the rigidity of the DNA motifs in the presence of
Mg2+ ions we have also done the pulling simulation for JX1 and JX2 motifs.
Fig. S14(b) gives the force-extension behavior of the JX1 and JX2 motifs
and compared with the PX-6:5 in presence of Mg2+ ions. We get the stretch
modulus from the sequential fitting of the energy vs strain plot: JX1 and
JX2 have stretch modulus of 1465 pN and 1654 pN respectively. We see
the same trend with both 10−4 and 10−3 pN/fs pulling rates as calculated
from polynomial fitting. To get more of a molecular level picture we have
plotted the internal energy of the PX/JX structure as a function of pulling
force as shown in Fig. 5 and we see a similar behavior compared to the en-
ergy variation observed in the presence of Na+ ions. Energy is decreasing in
the small force regime thus making the DNA structures thermodynamically
more stable. The increase in the energy with further load is rapid compared
to the case of pulling in Na+ medium. Increase in the total energy with the
applied force implies that the PX/JX DNA structure is thermodynamically
unstable, with some of H-bonds are broken. We give all the stretch modulus
results in both Na+ and Mg2+ counterion solution in Fig. S16. For compar-
ison in Fig. S1 and S2 we show the radial distribution functions of the Na+
and Mg2+ ions with the O1P and O2P Oxygen of the phosphate backbone.
3.2 Effect of DNA basepair sequence and rate of pulling
To investigate the effect of sequence on the structural rigidity of the PX
crossover DNA motifs we have calculated the force-extension profile for PX-
6:5 made of only AT or GC basepairs in presence of Na+ counterions. The
stretch modulus obtained from the sequential fits comes out to be 1592 pN
for AT sequence and 1780 pN for GC sequence. As expected, a PX struc-
ture made of only GC sequence is much stiffer than one made of only AT
sequence due to extra H-bonding in GC basepair. The stretch modulus of
crossover PX structures made of only AT or only GC is still larger than the
stretch modulus of 1269 pN of B-DNA double helix of same length with a
combination of AT and GC sequence. Stretch modulus γ1 values for various
PX-6:5 and JX structures are tabulated in table 1 for comparison.
It is known from previous theoretical study that the stiffness of the poly-
mer increase when pulled at faster rate (48). So when pulling with faster rate
Structural Rigidity of PX and JX DNA
13
the bond strength would increase dynamically and hence stretch modulus is
expected to increase. Typical AFM pulling rates are of 104 pN/s whereas the
slowest pulling rate achieved in our simulation is 1011 pN/s (i.e., 10−4 pN/fs).
Therefore due to the higher pulling rates employed in our simulation, the
obtained stretch modulus is expected to be higher in magnitude compared to
the results obtained from single molecule experiments. Similarly the magni-
tude of force where the PX/JX structure elongates roughly twice its initial
contour length would also be more than that obtained from experiments.
To see the effect of rate of force on force-extension behavior and stretch
modulus, we have done simulations at two different pulling rates viz., 10−4
pN/fs and 10−3 pN/fs. Fig. 4(a) shows the force-extension curve for PX
structure with two different pulling rates. It is clear from the plot that the
force at which the PX structure extends double its initial contour length is
very high when pulled with 10−3 pN/fs pulling rate compared to 10−4 pN/fs
pulling rate. This is due to the dynamic stiffening of the H-bonds in PX
structure when pulled with 10−3 pN/fs compared to 10−4 pN/fs. Breaking
of H-bonding is the major signature of mechanical deformation of the DNA
molecule (57, 58) We have calculated the fraction of surviving H-bonds as
a function of pulling force at two different pulling rates (Fig. 4(b)). As the
applied force is increased, the fraction of survived H-bonds decreases and
at large enough force the fraction of survived H-bonds goes to zero. From
Fig. 4(b) it is clear that the fraction of surviving H-bonds goes to zero at
smaller forces when pulled with 10−4 pN/fs pulling rate compared to 10−3
pN/fs pulling rate. The stretch modulus obtained from energy-strain and
force-extension curve are listed in table 1 in the presence of Na+ (Mg2+)
counterion medium.
4 Conclusion
We have calculated the rigidity of PX DNA molecules in the presence of Na+
and Mg2+ counterions by directly calculating their force-extension behavior
under axial stretching. Earlier we demonstrated (20, 21, 59) the rigidity of
these crossover DNA motifs from the vibrational density of states analysis
but a quantitative estimate of their structural rigidity was lacking. Now we
give a quantitative estimate of the stretch modulus of these DNA strictures.
In the presence of Na+ ions at a counterion concentration 246 mM the stretch
modulus of the PX-6:5 structure is almost 30 % more than that of normal
B-DNA double helix of same length and having sequence of one of the double
helical domain of PX-6:5. The computational cost of these calculations is
Structural Rigidity of PX and JX DNA
14
enormous, thus restricting us to use intermediate pulling rates of 10−4 and
10−3 pN/fs. To understand the effect of crossovers on the stretch modulus
of these DNA motifs we have also calculated the force-extension profile of
the JX1/JX2 motifs in the presence of Na+ counterions. We find that JX1
has stretch modulus of 1515 pN which is slightly smaller than the stretch
modulus of PX-6:5 (1635 pN). JX2 has a stretch modulus of 1349 pN which
is 286 pN smaller than the stretch modulus of PX-6:5. When the DNA is
pulled, among all contributions to the total energy of the DNA, there is a
dramatic increase in torsion energy. This increase in torsion energy is due to
a very large change in different torsion angles (Fig. 3) which cause the DNA
to destabilize with increased force. Interestingly there is almost no change in
various bond distances or angles as a function of force (Fig. 3). The similar
behavior was observed in the presence of divalent Mg2+ ions. In the presence
of Mg2+ we find the stretch modulus of PX-6:5, JX1 and JX2 to be 1875 pN,
1465 pN and 1654 pN, respectively. PX-6:5 has the highest stretch modulus
than any other DNA motif as in the case of Na+ medium. Interestingly all
structures in Mg2+ medium have more stretch modulus compared to Na+
medium due to large electrostatic screening arising from the divalency of
Mg2+ ions. This could be due to the fact that presence of Mg2+ ions gives
extra stability to the structure, owing to the strong coordination of Mg2+
with the phosphate atoms of the two double helical domains. We have also
studied the effect of the rate of force on the rigidity of DNA motif and found
that the increased rate of force enhances the rigidity of the structure.
5 Acknowledgements
We acknowledge Department of Science and Technology (DST), Government
of India for financial support. PKM also thanks Alexander von Humboldt
foundation for sponsoring his visit to Technical University Munich where
part of the work was done and Roland Netz for help with the sequential
energy fitting and valuable comments. We are also grateful to Prof. Ned
Seeman for a critical reading of the manuscript and valuable suggestions.
MS thanks University Grants Commission (UGC), India for senior research
fellowship.
Structural Rigidity of PX and JX DNA
15
References
1. Seeman, N., 2003. DNA in a material world. Nature 421:427 -- 431.
2. Seeman, N., 2001. DNA nicks and nodes and nanotechnology. Nano
Lett. 1:22 -- 26.
3. Cohen, S., A. Chang, H. Boyer, and R. Helling, 1973. Construction of
biologically functional bacterial plasmid in-vitro. Proc. Natl. Acad. Sci.
USA. 70:3240 -- 3244.
4. Seeman, N., 1982. Nucleic-Acid junctions and lattices. J. Theor. Biol.
99:237 -- 247.
5. Chen, J., and N. Seeman, 1991. Synthesis from DNA of molecule with
the connectivity of a cube. Nature 350:631 -- 633.
6. Zhang, Y., and N. Seeman, 1994. Construction of a DNA-truncated
octahedron. J. Am. Chem. Soc. 116:1661 -- 1669.
7. Du, S., B. Stollar, and N. Seeman, 1995. A synthetic DNA molecule in
3 knotted topologies. J. Am. Chem. Soc. 117:1194 -- 1200.
8. Du, S., and N. Seeman, 1994. The construction of a trefoil knot from a
DNA branched junction motif. Biopolymers 34:31 -- 37.
9. Mao, C., W. Sun, and N. Seeman, 1997. Assembly of Borromean rings
from DNA. Nature 386:137 -- 138.
10. Yurke, B., A. Turberfield, A. Mills, F. Simmel, and J. Neumann, 2000.
A DNA-fuelled molecular machine made of DNA. Nature 406:605 -- 608.
11. Buehler, M. J., and T. Ackbarow, 2007. Fracture mechanics of protein
materials. Materials Today 10:46 -- 58.
12. Seeman, N., H. Wang, X. Yang, F. Liu, C. Mao, W. Sun, L. Wenzler,
Z. Shen, R. Sha, H. Yan, M. Wong, P. Sa-Ardyen, B. Liu, H. Qiu, X. Li,
J. Qi, S. Du, Y. Zhang, J. Mueller, T. Fu, Y. Wang, and J. Chen, 1998.
New motifs in DNA nanotechnology. Nanotechnol. 9:257 -- 273.
13. Fu, T., and N. Seeman, 1993. DNA double-crossover molecules. Bio-
chemistry 32:3211 -- 3220.
14. Sa-Ardyen, P., A. Vologodskii, and N. Seeman, 2003. The flexibility of
DNA double crossover molecules. Biophys. J. 84:3829 -- 3837.
Structural Rigidity of PX and JX DNA
16
15. Winfree, E., F. Liu, L. Wenzler, and N. Seeman, 1998. Design and
self-assembly of two-dimensional DNA crystals. Nature 394:539 -- 544.
16. Holliday, R., 1964. Mechanism for gene conversion in fungi. Genet. Res.
5:282 -- &.
17. Li, X., X. Yang, J. Qi, and N. Seeman, 1996. Antiparallel DNA double
crossover molecules as components for nanoconstruction. J. Am. Chem.
Soc. 118:6131 -- 6140.
18. Shen, Z., H. Yan, T. Wang, and N. Seeman, 2004. Paranemic crossover
DNA: A generalized Holliday structure with applications in nanotech-
nology. J. Am. Chem. Soc. 126:1666 -- 1674.
19. Yan, H., X. Zhang, Z. Shen, and N. Seeman, 2002. A robust DNA
mechanical device controlled by hybridization topology. Nature 415:62 --
65.
20. Maiti, P., T. Pascal, N. Vaidehi, and W. Goddard, 2004. The stability
of Seeman JX DNA topoisomers of paranemic crossover (PX) molecules
as a function of crossover number. Nucleic Acids Res. 32:6047 -- 6056.
21. Maiti, P., T. Pascal, N. Vaidehi, J. Heo, and W. Goddard, 2006.
Atomic-level simulations of Seeman DNA nanostructures: The parane-
mic crossover in salt solution. Biophys. J. 90:1463 -- 1479.
22. Spink, C. H., L. Ding, Q. Yang, R. D. Sheardy, and N. C. Seeman,
2009. Thermodynamics of Forming a Parallel DNA Crossover. Biophys.
J. 97:528 -- 538.
23. Santosh, M., and P. K. Maiti, 2009. Force induced DNA melting. J.
Phys.: Condens. Matter 21:034113.
24. Bustamante, C., J. Marko, E. Siggia, and S. Smith, 1994. Entropic
elasticity of lambda-phase DNA. Science 265:1599 -- 1600.
25. Smith, P., and B. Pettitt, 1996. Ewald artifacts in liquid state molecular
dynamics simulations. J. Chem. Phys. 105:4289 -- 4293.
26. Allemand, J., D. Bensimon, R. Lavery, and V. Croquette, 1998.
Stretched and overwound DNA forms a Pauling-like structure with ex-
posed bases. Proc. Natl. Acad. Sci. USA. 95:14152 -- 14157.
Structural Rigidity of PX and JX DNA
17
27. Rief, M., H. Clausen-Schaumann, and H. Gaub, 1999.
Sequence-
dependent mechanics of single DNA molecules. Nat. Struct. Biol. 6:346 --
349.
28. Vologodskii, A., 1994. DNA extension under the action of an external
force. Macromolecules 27:5623 -- 5625.
29. Marko, J., and E. Siggia, 1995. Stretching DNA. Macromolecules
28:8759 -- 8770.
30. Smith, S., Y. Cui, and C. Bustamante, 1996. Overstretching B-DNA:
The elastic response of individual double-stranded and single-stranded
DNA molecules. Science 271:795 -- 799.
31. Cluzel, P., A. Lebrun, C. Heller, R. Lavery, J. Viovy, D. Chatenay, and
F. Caron, 1996. DNA: An extensible molecule. Science 271:792 -- 794.
32. Fu, H., H. Chen, J. F. Marko, and J. Yan, 2010. Two distinct over-
stretched DNA states. Nucleic Acids Res. 38:5594 -- 5600.
33. Fu, H., H. Chen, X. Zhang, Y. Qu, J. F. Marko, and J. Yan, 2011. Transi-
tion dynamics and selection of the distinct S-DNA and strand unpeeling
modes of double helix overstretching. Nucleic Acids Res. 39:3473 -- 3481.
34. Chen, Hu and Yan, Jie , 2008. Effects of kink and flexible hinge defects
on mechanical responses of short double-stranded DNA molecules. Phys.
Rev. E 77:041907.
35. Rouzina, I., and V. Bloomfield, 2001. Force-induced melting of the DNA
double helix. 2. Effect of solution conditions. Biophys. J. 80:894 -- 900.
36. van Mameren, J., P. Gross, G. Farge, P. Hooijman, M. Modesti,
M. Falkenberg, G. J. L. Wuite, and E. J. G. Peterman, 2009. Unraveling
the structure of DNA during overstretching by using multicolor, single-
molecule fluorescence imaging. Proc. Natl. Acad. Sci. USA. 106:18231 --
18236.
37. Cocco, S., J. Yan, J. Leger, D. Chatenay, and J. Marko, 2004. Over-
stretching and force-driven strand separation of double-helix DNA.
Phys. Rev. E 70:011910.
38. Lai, P., and Z. Zhou, 2003. B- to S-form transition in double-stranded
DNA with basepair interactions. Physica A 321:170 -- 180.
Structural Rigidity of PX and JX DNA
18
39. Tung, C., and E. Carter, 1994. Nucleic-acid modeling tool (NAMOT) -
an interactive graphic tool for modeling nucleic-acid structures. Comput.
Appl. Biosci. 10:427 -- 433.
40. Case, D. A., T. A. Darden, T. E. Cheatham, I., C. L. Simmerling,
J. Wang, R. E. Duke, R. Luo, K. M. Merz, B. Wang, D. A. Pearlman,
M. Crowley, S. Brozell, V. Tsui, H. Gohlke, J. Mongan, V. Hornak,
G. Cui, P. Beroza, C. Schafmeister, J. W. Caldwell, W. S. Ross, and
P. A. Kollman, 2004. AMBER8.
41. Duan, Y., C. Wu, S. Chowdhury, M. Lee, G. Xiong, W. Zhang, R. Yang,
P. Cieplak, R. Luo, T. Lee, J. Caldwell, J. Wang, and P. Kollman,
2003. A point-charge force field for molecular mechanics simulations of
proteins based on condensed-phase quantum mechanical calculations. J.
Comput. Chem. 24:1999 -- 2012.
42. Wang, J., P. Cieplak, and P. Kollman, 2000. How well does a restrained
electrostatic potential (RESP) model perform in calculating conforma-
tional energies of organic and biological molecules? J. Comput. Chem.
21:1049 -- 1074.
43. Jorgensen, W., J. Changrasekhar, J. Madura, R. Impey, and M. Klein,
1983. Comparison of simple potential function for simulating liquid
water. J. Chem. Phys. 79:926 -- 935.
44. Aqvist, J., 1990.
Ion water interaction potentials derived from free-
energy perturbation simulations. J. Phys. Chem. 94:8021 -- 8024.
45. Baker, N., D. Sept, S. Joseph, M. Holst, and J. McCammon, 2001.
Electrostatics of nanosystems: Application to microtubules and the ri-
bosome. Proc. Natl. Acad. Sci. USA. 98:10037 -- 10041.
46. Baker, N., 2005.
Improving implicit solvent simulations: a Poisson-
centric view. Current Opinion in Structural Biology 15:137 -- 143.
47. Ryckaert, J., G. Ciccotti, and H. Berendsen, 1977.
Numerical-
integration of cartesian equations of motion of a system with constraints
- molecular-dynamics of N-alkanes. J. Comput. Phys. 23:327 -- 341.
48. Evans, E., and K. Ritchie, 1997. Dynamic strength of molecular adhe-
sion bonds. Biophys. J. 72:1541 -- 1555.
Structural Rigidity of PX and JX DNA
19
49. Hugel, T., M. Rief, M. Seitz, H. Gaub, and R. Netz, 2005. Highly
stretched single polymers: Atomic-force-microscope experiments versus
ab-initio theory. Phys. Rev. Lett. 94:048301.
50. Seol, Y., J. Li, P. C. Nelson, T. T. Perkins, and M. D. Betterton, 2007.
Elasticity of short DNA molecules: Theory and experiment for contour
lengths of 0.6-7 mu m. Biophys. J. 93:4360 -- 4373.
51. Wiggins, P. A., T. Van der Heijden, F. Moreno-Herrero, A. Spakowitz,
R. Phillips, J. Widom, C. Dekker, and P. C. Nelson, 2006. High flexibil-
ity of DNA on short length scales probed by atomic force microscopy.
Nature Nanotechnol. 1:137 -- 141.
52. Ackbarow, T., and M. J. Buehler, 2009. Alpha-helical protein domains
unify strength and robustness through hierarchical nanostructures. Nan-
otechnology 20:075103.
53. Baumann, C., S. Smith, V. Bloomfield, and C. Bustamante, 1997. Ionic
effects on the elasticity of single DNA molecules. Proc. Natl. Acad. Sci.
USA. 94:6185 -- 6190.
54. Konrad, M., and J. Bolonick, 1996. Molecular dynamics simulation of
DNA stretching is consistent with the tension observed for extension
and strand separation and predicts a novel ladder structure. J. Am.
Chem. Soc. 118:10989 -- 10994.
55. Lebrun, A., and R. Lavery, 1996. Modelling extreme stretching of DNA.
Nucleic Acids Res. 24:2260 -- 2267.
56. Saenger, W., 1984. Principles of Nucleic Acid Structure. Springer Ver-
lag, New York.
57. Keten, S., and M. J. Buehler, 2008. Asymptotic Strength Limit of
Hydrogen-Bond Assemblies in Proteins at Vanishing Pulling Rates.
Phys. Rev. Lett. 100:198301.
58. Keten, S., and M. J. Buehler, 2008. Geometric confinement governs the
rupture strength of H-bond assemblies at a critical length scale. Nano
Lett. 8:743 -- 748.
59. Maiti, P. K., T. A. Pascal, N. Vaidehi, and W. A. Goddard, III, 2007.
Understanding DNA based nanostructures. J. Nanosci. Nanotechnol.
7:1712 -- 1720.
Structural Rigidity of PX and JX DNA
20
Table 1: Stretch modulus γ1 (pN) from sequential fit to the energy-strain
curve and polynomial fit to the force-extension curve in the presence of Na+
(Mg2+) counterions with different rates of force of 10−4 and 10−3 pN/fs (Fig.
S15). Note that the values in brackets corresponds to Mg2+ case.
DNA
c (mM )
B-DNA 254(140)
246(160)
PX-6:5
246(160)
JX1
246(160)
JX2
AT65
246
246
GC65
γ1 for Na+ (Mg2+) (pN)
Sequential Fit
10−3
(pN/fs)
1358
1737
1558
1480
1731
1862
10−4
(pN/fs)
1269
1636(1875)
1515(1465)
1349(1654)
1592
1780
Polynomial Fit
10−3
(pN/fs)
1644
2322
2176
2051
1981
2138
10−4
(pN/fs)
1578(1591)
1772(1841)
1374(1404)
1521
1378
1546
Structural Rigidity of PX and JX DNA
21
Figure Legends
Fig. 1.
Atomic level structure of PX/JX DNA molecules:
(a) The basepair se-
quences used in the generations of PX-6:5, JX1 and JX2 crossover molecules.
Initial structure of (b) PX-6:5, (c) JX1 and (d) JX2 used in our pulling sim-
ulation.
Fig. 2.
Force-extension curves: (a) Force-extension behavior of PX-6:5 and B-DNA
molecule in the presence of Na+ counterions. (b) Force-extension behavior
of PX-6:5, JX1 and JX2 in presence of Na+ ions. The rate of force is 10−4
pN/fs.
Fig. 3.
Variation of backbone parameters with force: O3′-P bond distance, P-O5′-
C5′ angle and C5′-C4′-C3′ angle as a function of force applied on PX-6:5
in Na+ (black) and Mg2+ (blue dashed line) medium. There is very little
change in the bond distance and the angles with the increase in force. Vari-
ation of torsion angles alpha (α), beta (β), gamma (γ), delta (δ), epsilon
(ǫ) and zeta (ζ) as a function of force applied at the end of PX-6:5 in Na+
(black) and Mg2+ (blue dashed line) medium. There is a very large change
over 100 % of the equilibrium zero force limits (dotted lines) in all the torsion
angles with the force applied.
Fig. 4.
Effect of rate of pulling on PX-6:5 structure: (a) Force-extension curve at
two different pulling rates. Higher the pulling rate steeper is the response
and hence large stretch modulus. (b) Fraction of survived H-bonds as a
function of pulling force at different pulling rate. At high pulling rate melt-
ing/breaking of H-bonds occurs at larger magnitude of pulling force.
Fig. 5.
Energy of various PX/JX structures as a function of pulling force in Na+
and Mg2+ counterions. The rate of pulling is 10−4 pN/fs. The minimum in
the energy curve corresponds to the most stable structure during pulling.
Structural Rigidity of PX and JX DNA
22
Fig. 6.
Instantaneous snapshots of the PX-6:5 structure at various forces (a) 200
pN, (b) 400 pN, (c) 600 pN, (d) 800 pN and (e) 1000 pN in the presence
of Na+ counterions. As the applied force is increased beyond 200 pN, H-
bonds in the basepair were broken where the extension in the DNA is almost
double its initial contour length.
Fig. 7.
Force-extension spectrum analysis. Upper color bar indicates the strain in-
crease percentage on applying the external force on the various DNA struc-
tures. The extension in Mg2+ medium is less than the extension in Na+
medium which implies the enhanced rigidity of the PX/JX DNA molecules
in Mg2+ medium.
Structural Rigidity of PX and JX DNA
23
(a)
(b)
(c)
(d)
Figure 1:
Structural Rigidity of PX and JX DNA
24
200
100
1000
800
600
400
200
)
N
p
(
F
0.2
0.4
ւ
↑
←
PX-6:5
B-DNA
1000
800
600
400
200
)
N
p
(
F
PX-6:5
JX1
JX2
0
0
0.2
0.4
0.6
ε
0.8
1
1.2
0
0
0.2
0.4
0.6
ε
0.8
1
1.2
(a)
(b)
Figure 2:
Structural Rigidity of PX and JX DNA
25
rO3'-P (Å)
-65o
α
118o
1.62
1.60
1.58
100o
0o
-100o
140o
120o
124o
122o
120o
100o
0o
-100o
0o
-75o
δ
ε
θ
P-O5'-C5'
θ
C5'-C4'-C3'
117o
114o
111o
100o
77o
-65o
39o
-40o
β
0o
γ
-100o
0o
-50o
-100o
ζ
100o
0 pN
-150o
1000 pN
0 pN
1000 pN
0 pN
1000 pN
Figure 3:
Structural Rigidity of PX and JX DNA
26
(a)
(b)
Figure 4:
Structural Rigidity of PX and JX DNA
27
-210
-220
-230
-240
-250
-260
)
p
b
-
l
o
m
/
l
a
c
k
(
E
Na+
Mg2+
PX-6:5
JX1
JX2
B-DNA
-270
0
200
400
600
800
1000
F (pN)
Figure 5:
Structural Rigidity of PX and JX DNA
28
200
pN
400
pN
600
pN
800
pN
1000
pN
(a)
(b)
(c)
(d)
(e)
Figure 6:
Structural Rigidity of PX and JX DNA
29
Figure 7:
|
1305.5929 | 1 | 1305 | 2013-05-25T14:45:17 | Two-Input Enzymatic Logic Gates Made Sigmoid by Modifications of the Biocatalytic Reaction Cascades | [
"physics.bio-ph",
"cond-mat.soft",
"q-bio.MN"
] | Computing based on biochemical processes is a newest rapidly developing field of unconventional information and signal processing. In this paper we present results of our research in the field of biochemical computing and summarize the obtained numerical and experimental data for implementations of the standard two-input OR and AND gates with double-sigmoid shape of the output signal. This form of response was obtained as a function of the two inputs in each of the realized biochemical systems. The enzymatic gate processes in the first system were activated with two chemical inputs and resulted in optically detected chromogen oxidation, which happens when either one or both of the inputs are present. In this case, the biochemical system is functioning as the OR gate. We demonstrate that the addition of a "filtering" biocatalytic process leads to a considerable reduction of the noise transmission factor and the resulting gate response has sigmoid shape in both inputs. The second system was developed for functioning as an AND gate, where the output signal was activated only by a simultaneous action of two enzymatic biomarkers. This response can be used as an indicator of liver damage, but only if both of these of the inputs are present at their elevated, pathophysiological values of concentrations. A kinetic numerical model was developed and used to estimate the range of parameters for which the experimentally realized logic gate is close to optimal. We also analyzed the system to evaluate its noise-handling properties. | physics.bio-ph | physics | Two-Input Enzymatic Logic Gates Made Sigmoid by
Modifications of the Biocatalytic Reaction Cascades
Oleksandr Zavalov,1 Vera Bocharova,2 Jan Halámek,3 Lenka Halámková,3 Sevim Korkmaz,4
Mary A. Arugula,5 Soujanya Chinnapareddy,3 Evgeny Katz,3 Vladimir Privman1
1 Department of Physics, Clarkson University, Potsdam, NY 13699
2 Chemical Sciences Division, Oak Ridge National Laboratory, Oak Ridge, TN
3 Department of Chemistry and Biomolecular Science, Clarkson University, Potsdam, NY
4 Department of Chemical and Biomolecular Engineering, Clarkson University, Potsdam, NY
5 Department of Chemistry, University of Utah, Salt Lake City, UT 84112
International Journal of Unconventional Computing 8, 347-365 (2013)
Abstract: Computing based on biochemical processes is a newest rapidly developing field of
unconventional information and signal processing. In this paper we present results of our
research in the field of biochemical computing and summarize the obtained numerical and
experimental data for implementations of the standard two-input OR and AND gates with
double-sigmoid shape of the output signal. This form of response was obtained as a function of
the two inputs in each of the realized biochemical systems. The enzymatic gate processes in the
first system were activated with two chemical inputs and resulted in optically detected
chromogen oxidation, which happens when either one or both of the inputs are present. In this
case, the biochemical system is functioning as the OR gate. We demonstrate that the addition of
a “filtering” biocatalytic process leads to a considerable reduction of the noise transmission
factor and the resulting gate response has sigmoid shape in both inputs. The second system was
developed for functioning as an AND gate, where the output signal was activated only by a
simultaneous action of two enzymatic biomarkers. This response can be used as an indicator of
liver damage, but only if both of these of the inputs are present at their elevated,
pathophysiological values of concentrations. A kinetic numerical model was developed and used
to estimate the range of parameters for which the experimentally realized logic gate is close to
optimal. We also analyzed the system to evaluate its noise-handling properties.
Keywords:
Biocomputing, enzyme logic, logic gate, binary logic, sigmoid response
– 1 –
Introduction
1.
Recently, there has been considerable interest in the design and development of chemical
[1-6] and biochemical [7-12] systems which implement binary logic gates and networks. These
systems were investigated for novel computational processing of chemical and biochemical
signals. Special attention has been paid to biomolecular systems [13,14], including those that are
based on enzymatic reactions [15-18], DNA [9,19,20], or RNA [21] and even entire cells [22,23]
as a new research direction in unconventional computing [24,25]. Enzymatic reactions can be
used to simulate digital logic gates [26-31] and perform elementary arithmetic [32], as well as
“networked” in binary logic circuits [33-36]. Processes in such systems can emulate Boolean
logic and thus used as “digital” biochemical sensors with multiple inputs [37-39], e.g., in
biomedical applications [29-31].
In biotechnological environments, biomolecular computing systems promise [40] design
of original biosensors with multiplexing capabilities, processing several biological signals in the
digital format (0 or 1) [39,41]. In particular, devices based on implementations of biomolecular
logic have been extensively studied for possible applications in analysis of biomarkers
characteristic of various pathophysiological conditions and for diagnosis of disease or injury. For
example, the output chemical concentration reaching a certain logic-1 value could signal that an
“action” is needed, however the output concentration close to logic-0 would indicate “no action”
is needed. Most developed biochemical systems consider logic-0 as zero concentrations of the
signaling substances, whereas logic-1 is defined as experimentally convenient nonzero
concentrations. However, for real biomedical conditions the logic values 0 and 1 of
concentrations (or possibly ranges of concentrations) should correspond to normal physiological
and pathophysiological states of the organism, respectively. In most cases the difference between
the inputs at the logic 0 and 1 is comparable with the level of natural noise, making the
discrimination of the 0 and 1 output signals difficult unless careful optimization of the signal
processing systems is performed [42].
Noise amplification tendency of the simplest biochemical processes has to be taken into
consideration in designing of biochemical process as “devices” and networks. There are several
– 2 –
sources of noise in biochemical systems, in addition to the noise in the actual input data.
However, the noise amplification is controlled by the transmission of random fluctuations of the
input concentrations to the output response [43]. The amplification of the noise can build up
from one step to another, preventing connection of these systems in a network for information
processing [44]. Thus, suppression of the noise amplification from the input(s) to output is
crucial.
A typical response in biochemical systems has a convex shape with an initial linear
region, followed by saturation caused by the limitation of reaction rates and the amount of input
catalysts and reagents. Generally, in order to avoid the build-up of noise, it is useful to pass the
signals through biochemical filters, which give the sigmoid response profile with zero or small
slopes close to the “logic point” values. Such response is quite difficult to achieve in standard
biochemical reactions catalyzed by enzymes [45-47], despite the fact that the response of
sigmoid shape is often observed in nature [48-50]. Figure 1 (the figures are appended after the
Nomenclature summary, as Pages 21-25) illustrates the response function of an OR gate, as an
example of a system with high noise amplification levels. This system has its maximum noise
transmission factor about 690% at the logic-00 point, i.e., sevenfold noise amplification. This
transmission factor can be estimated for smooth gate-response surfaces by calculating the
absolute value of the gradient in terms of the logic variables, to be defined below. Therefore,
despite the fact that the system response function visually may look like a reasonable OR gate
(see Figure 1), without an additional filtering step the system is not appropriate for network
applications because of such a large noise amplification.
Novel biochemical reactions with an added “filter” have recently been designed [45,51-
54,57,58] and optimized as separate elements for inclusion as “logic gates” in biochemical logic
networks. Integration of a filtering step into a digital gate process allows to significantly improve
the parameters and performance of the latter. This applies particularly to the behavior near the
binary values 0 and 1 in the output response, corresponding, for instance, to the normal
physiological and pathophysiological concentration of biomarkers [51] for liver dysfunction,
abdominal trauma, and soft tissue injury.
– 3 –
Quite often it is necessary to analyze the information simultaneously from two
biomarkers [55,56]. In such cases, logic gates can be applied, such as OR, AND, etc., to
facilitate making the appropriate decision, depending on the presence or absence of each
biomarker or various biomarker combinations. In recent papers [57,58], we have demonstrated
implementations and analyses of such biochemical systems based on enzymatic reactions with
two inputs. The common feature of these works has been the incorporation of the filtering
processes, which have enabled to achieve a double-sigmoid shape of the output response
(sigmoid in both inputs). It is important to develop and test various biochemical filtering
approaches within a “toolbox” of binary gates, including OR and AND.
In this paper we summarize the investigation of the first reported realizations of an AND
gate and an OR gate in enzymatic filtered systems with two inputs. When discussing the AND
gate system, we will focus on the most studied case in this context [51], that of a pair of liver-
injury biomarkers [55,56], the enzymes alanine transaminase (ALT) and lactate dehydrogenase
(LDH). Elevated levels of both enzymes simultaneously have been used as an indication of liver
injury [59,60]. We consider a model system [31] in aqueous solution rather than in serum [61].
The aqueous solution system has not only been realized (as the AND gate) to optically detect the
presence of both enzymes, but has also been coupled to a signal-responsive polymer-brush thin-
film deposited on an electrode [62]. The practical implementation of such a system requires large
gate times, for which the gate realization precision by the original enzymatic cascade was low.
The added filtering process not only reduced noise transmission, but also—for the larger gate
times—improved the precision of the AND binary logic realization [51].
In the analysis and consideration of the OR gate implementation, we describe
biocatalytic reactions which produce changes in pH [63-67], coupled with a pH-buffering
mechanism [53]. The latter is performed by the addition of an appropriate amount of buffer for
the “filtering” effect. Experimental results [58] demonstrated that the addition of the
appropriately designed filtering step changes the gate response to the double-sigmoid shape.
We have developed and applied a kinetic modeling approach specially designed for the
study of the binary-logic gate-response system properties. This numerical model, applied to the
– 4 –
OR gate function realization with double-sigmoid response
experimental data, was used for a quantitative evaluation of the sources of noise, input-to-output
noise transmission factors [43,52], and tolerance properties of the realized AND [57] and OR
gates [58] as components for biochemical logic and networks. Section 2 outlines the
experimental procedure of an OR gate implementation with the double-sigmoid response. In
Section 3, we present a recent realization of a biochemical AND gate logic function with the
double-sigmoid response shape. In Section 4, we detail the kinetic modeling of logic gate
functions. Finally, in Section 5 we present results of the experimental data fitting using the
numerical model, and discuss the gate optimization and noise reduction approach.
2.
In this section we describe the OR gate realization designed in a biochemical system
biocatalysing the first process is esterase, of concentration to be denoted E(cid:4666)(cid:1872)(cid:4667), where (cid:1872) is the
outlined in Scheme 1 (the schemes are appended at the end of this preprint as Pages 26-27). This
system includes two enzymatic reactions, and also the buffering process. The enzyme
A(cid:4666)(cid:1872)(cid:4667)), Input 2 is methyl butyrate (concentration B(cid:4666)(cid:1872)(cid:4667)). Here esterase reacts with either one of
process time. This first enzymatic part has two inputs: Input 1 is ethyl butyrate (concentration
acid, U(cid:4666)(cid:1872)(cid:4667), as a byproduct of these processes. The increase in the butyric acid concentration in
the solution leads to a decrease of pH from its initial value of pH(cid:4666)(cid:1872) (cid:3404) 0(cid:4667) (cid:3404) 9.0. The pH of the
them, biocatalyzing the production of ethanol and methanol, respectively, and also of butyric
system lowers to the value which depends on the initial concentrations of the inputs, A(cid:4666)0(cid:4667) and
B(cid:4666)0(cid:4667), of Tris buffer (concentration T(cid:4666)0(cid:4667)), and on other process parameters, but no less than to
4.2. The Tris buffer, T(cid:4666)(cid:1872)(cid:4667), which is added at the initial moment of time, as T(cid:4666)0(cid:4667) moles per unit
The experiments were carried out in several stages, first without buffering (T(cid:4666)0(cid:4667) (cid:3404) 0
volume, is the main component for the filtering process, as will be described below.
buffered system yielding a high-quality OR gate, with T(cid:4666)0(cid:4667) (cid:3404) 4 mM, Figure 2, and finally a
mM), see Figure 1. As expected, the output values are equal to 0 at the logic 00, and are close to
system with excessive buffering which deteriorates the OR gate realization: T(cid:4666)0(cid:4667) (cid:3404) 8 mM,
each other at the logic points 01, 10, 11, defining the binary 1 for the output. Next, we present a
– 5 –
Figure 3. Adding the Tris buffer, T(cid:4666)0(cid:4667), into the solution instantly leads to the production of its
protonated form TH(cid:2878) . The process with the Tris buffer involves the absorption of hydrogen ions
from the solution and, as a consequence, prevention of the decrease in pH. This mild alkaline
buffering effect persists only in the range of pH from 9 to value about 7.2, and only as long as
initial buffer concentration, T(cid:4666)0(cid:4667), we can delay the effect of the pH decrease, which is due to the
process of continuous production of butyric acid with (pK a(cid:4667)(cid:2912)(cid:2931)(cid:2930)(cid:2935)(cid:2928)(cid:2919)(cid:2913) (cid:2911)(cid:2913)(cid:2919)(cid:2914) (cid:3404) 4.82 [68].
the Tris buffer is not completely turned into the protonated form. Therefore, controlling the
At first, when the pH is high, the second enzyme laccase (of concentration L(cid:4666)(cid:1872)(cid:4667), see
approximately 8.1, biocatalytic activity of laccase increases. Using K(cid:2872)Fe(cid:4666)CN(cid:4667)(cid:2874) as a co-substrate,
this enzyme then produces K(cid:2871)Fe(cid:4666)CN(cid:4667)(cid:2874) with increasing rate. The concentrations of K(cid:2872)Fe(cid:4666)CN(cid:4667)(cid:2874)
Scheme 1) is practically not active. However, as the pH of the system decreases past
and K(cid:2871)Fe(cid:4666)CN(cid:4667)(cid:2874) will be denoted F(cid:4666)(cid:1872)(cid:4667) and P(cid:4666)(cid:1872)(cid:4667), respectively. The output, which is the product of
the entire process, P(cid:4666)(cid:1872)(cid:3034) (cid:4667), is measured at the gate time tg = 800 s by absorbance (denoted Abs) at
(cid:2019) (cid:3404) 420 nm. For the present system, Abs(cid:4666)(cid:1872)(cid:4667) and P(cid:4666)(cid:1872)(cid:4667) are numerically almost identical (but have
physical zeros, A0(cid:4666)0(cid:4667) (cid:3404) B0 (cid:4666)0(cid:4667) (cid:3404) 0, as the signals’ binary 0s, and experimentally convenient
input values A1(cid:4666)0(cid:4667) (cid:3404) 10 mM and B1(cid:4666)0(cid:4667) (cid:3404) 10 mM as logic 1s. Logic 1 of the output, P1(cid:4666)(cid:1872)(cid:3034) (cid:4667), is
different units) [58]. For our modeling of the biochemical logic-gate system, here we take the
set by the gate function itself.
For the considered OR gate system implementation of [58] to be accurate, the appropriate
output values close to 0, 1, 1, 1, at the four logic-input combinations, 00, 01, 10, 11, respectively,
should be realized. However, for the study of the noise control, the behavior of the system
response should be investigated [16,17,43] not only at, but also around the logic input values.
The latter study is carried out in terms of the scaled, binary-range (zero to one) dimensionless
(cid:1876)(cid:3404)A(cid:4666)0(cid:4667)/A1(cid:4666)0(cid:4667); (cid:1877)(cid:3404)B(cid:4666)0(cid:4667)/B1(cid:4666)0(cid:4667); (cid:1878)(cid:4666)(cid:1876) , (cid:1877)(cid:4667)(cid:3404)P(cid:4666)(cid:1872)(cid:3034) (cid:4667)/P1(cid:4666)(cid:1872)(cid:3034) (cid:4667),
variables,
Note that in practical applications the logic one values of the inputs will be determined [43] by
the environment in which the gate is used, but the logic zero values of the inputs (and output)
– 6 –
(1)
AND logic gate with double-sigmoid response
need not be exactly at the physical zeros [42,43,57].
3.
Scheme 2 shows the sequence of biocatalytic processes involved in the AND gate
function, catalyzed by the two input enzymes ALT and LDH, and the added filter process
catalyzed by an additional enzyme, glucose-6-phosphate dehydrogenase (G6PDH). Additional
details of the system functioning [51,57] are presented below. We study a biocatalytic cascade to
detect the high levels of the two enzyme inputs, ALT and LDH. Their simultaneous increase in
concentrations, from normal to pathophysiological levels, provides [59,60] an evidence of liver
α-KTG (cid:3397) Ala → Glu (cid:3397) Pyr
injury. The sequence of the biochemical processes is shown in the “gate” section of Scheme 2:
Pyr (cid:3397) NADH → Lac (cid:3397) NAD+
Oxidation of NADH is followed by measuring the change in the absorbance, yielding the
concentration of the output, β-nicotinamide adenine dinucleotide (NAD+), which is produced
only in the presence of the two input enzymes (and the “gate machinery” reactants α-KTG, Ala,
NADH). Note that without the added filter process, we would obtain a response function that
would not have a sigmoid shape in any direction. Here Glu abbreviates glutamate, Pyr stands for
pyruvate, and Lac for lactate. Figure 4 illustrates the result for the “gate time” of tg = 600 s,
relevant for applications involving signal-responsive membranes [57,62].
As mentioned above, adding the filter will cause a delay in the signal increase and the
appearance of the sigmoid shape. The “filter” section of Scheme 2 shows the added filtering
process. Our filtering component of the AND logic gate function is based on enzyme glucose-6-
Glc6P (cid:3397) NAD(cid:3397) → NADH (cid:3397) 6‐PGluc
phosphate dehydrogenase (G6PDH). The biocatalytic reaction is shown in the following:
– 7 –
(2)
(3)
(4)
This reaction has two inputs: glucose-6-phosphate (Glc6P) and cofactor (NAD+), and one output:
the reduced cofactor (NADH); the other reaction product in aqueous solution is 6-phospho-
gluconic acid (6-PGluc). The system now displays double-sigmoid behavior in both inputs ALT
and LDH (Figure 5).
The output signal, [NAD+], was quantified by detecting the decrease in the concentration
function (Figures 4, 5), it is necessary to obtain the output signal values, (cid:4670)NAD(cid:2878) (cid:4671)(cid:4666)(cid:1872)(cid:3034) (cid:4667), not only
of NADH, measured optically at = 340 nm. It was calculated using the extinction coefficient,
for the inputs at the logical points 0 and 1, which were (cid:4670)ALT(cid:4671)0(cid:4666)0(cid:4667) (cid:3404) 0.02 U/mL, (cid:4670)ALT(cid:4671)1 (cid:4666)0(cid:4667) (cid:3404)
6.22 mM–1cm–1, for NADH [69]. In order to study the 3D response surface of the AND gate
2 U/mL for Input 1 and (cid:4670)LDH(cid:4671)0(cid:4666)0(cid:4667) (cid:3404) 0.15 U/mL, (cid:4670)LDH(cid:4671)1(cid:4666)0(cid:4667) (cid:3404) 1 U/mL for Input 2, but also
for varying input concentrations. Here and below, the subscripts 0 and 1 refer to the logic-point
values. However, the actual noise-property analysis is best carried out in terms of the rescaled
(cid:1876) (cid:3404) (cid:3435)(cid:4670)ALT(cid:4671)(cid:4666)0(cid:4667)‐(cid:4670)ALT(cid:4671)0(cid:4666)0(cid:4667)(cid:3439) (cid:3435)(cid:4670)ALT(cid:4671)1(cid:4666)0(cid:4667)‐(cid:4670)ALT(cid:4671)0(cid:4666)0(cid:4667)(cid:3439)
(cid:3415)
variables, defined as
(cid:1877) (cid:3404) (cid:3435)(cid:4670)LDH(cid:4671)(cid:4666)0(cid:4667)‐(cid:4670)LDH(cid:4671)0(cid:4666)0(cid:4667)(cid:3439) (cid:3435)(cid:4670)LDH(cid:4671)1(cid:4666)0(cid:4667)‐(cid:4670)LDH(cid:4671)0(cid:4666)0(cid:4667)(cid:3439)
(cid:3415)
(cid:1878) (cid:3404) (cid:4672)(cid:4670)NAD(cid:3397) (cid:4671)(cid:3435)tg(cid:3439)‐(cid:4670)NAD(cid:3397) (cid:4671)0(cid:3435)tg(cid:3439)(cid:4673) (cid:4672)(cid:4670)NAD(cid:3397) (cid:4671)1(cid:3435)tg(cid:3439)‐(cid:4670)NAD(cid:3397) (cid:4671)0(cid:3435)tg(cid:3439)(cid:4673)
(cid:3415)
and output (z). Experimentally, we can map out the gate response surface, (cid:1878)(cid:4666)(cid:1876) , (cid:1877)(cid:4667).
These equations (5-7) determine the normalized variables for the first input (x), second input (y),
As shown in our recent papers [43,44,57,58], the design of a large network, which
includes several biochemical logic systems, requires careful control of the built-in noise level.
kept under control. It is desirable to have a response function, (cid:1878)(cid:4666)(cid:1876) , (cid:1877)(cid:4667), which has small absolute
The noise level depends generally on the gate’s environment. In order to ensure stable operation
of complex networks with scalable processes, the degree of the noise amplification should be
– 8 –
(5)
(6)
(7)
Kinetic modeling of the logic gate functions
estimated from the absolute values of gradients, (cid:3627)(cid:1487)(cid:4652)(cid:4652)(cid:1318)(cid:1878)(cid:4666)(cid:1876) , (cid:1877)(cid:4667)(cid:3627)(cid:2868)(cid:2868),(cid:2868)(cid:2869),(cid:2869)(cid:2868),(cid:2869)(cid:2869) , of the function (cid:1878)(cid:4666)(cid:1876) , (cid:1877)(cid:4667) at
values of gradients at all the logic points. The noise scaling properties of the logic gates can be
the logic points in the (x,y) plane. If the absolute value of the gradient is less than one, then the
spread of noise in the logic gate system is suppressed. However, as mentioned above, without the
inclusion of a properly designed filtering process it is difficult to obtain such a response function.
Typically, without filtering the response function always has a gradient greater than 1 (see Figure
4), and, consequently, amplifies the noise level. Our results [57] with filtering have yielded a
realization of an AND gate function with double-sigmoid filter response in two inputs; see
Figure 5.
4.
In this section we outline how kinetic modeling can be applied to gate optimization. As
mentioned above, the addition of the filtering process in the biocatalytic system allows an
elimination of the built-in noise amplification in the response function. Therefore the logic gate
can be used for designing large biochemical networks. We focus on the study of the properties
and the analysis of such processes by a kinetic modeling approach. Consider as an example the
OR gate function introduced earlier. The biocatalytic function of esterase can be described using
(cid:1831) (cid:3397) (cid:1827) (cid:1863)(cid:2869)⇄(cid:1863)(cid:2879)(cid:2869)(cid:1829)(cid:3002) (cid:1863)(cid:2870)→ (cid:1847) (cid:3397) (cid:1831) (cid:3397) ⋯
the standard schemes
(cid:1831) (cid:3397) (cid:1828) (cid:1863)(cid:2871)⇄(cid:1863)(cid:2879)(cid:2871)(cid:1829)(cid:3003) (cid:1863)(cid:2872)→ (cid:1847) (cid:3397) (cid:1831) (cid:3397) ⋯
(9)
where (cid:1829)(cid:3002) and (cid:1829)(cid:3003) are (time-dependent) concentrations of complexes. As the pH of the system
changes over a wide range, the effective rate constants in eq. (8) and (9) will also change. As in
the earlier work [53], we described this pH dependence by assuming fast acidification
(8)
– 9 –
equilibrium, e.g., (cid:1863)(cid:2869) (cid:3404) (cid:1863)(cid:3364)(cid:2869)(cid:1837)(cid:3006) /(cid:4666)(cid:4670)H(cid:2878) (cid:4671)(cid:4666)(cid:1872)(cid:4667) (cid:3397) (cid:1837)(cid:3006) (cid:4667), with (cid:1863)(cid:3364)(cid:2869) (cid:3404) 1.018 mM–1sec–1 [70]. Similar
expressions were used for the other rate constants [70], with the values (cid:1863)(cid:3364)(cid:2870) (cid:3404) 1.603 sec–1, (cid:1863)(cid:3364)(cid:2871) (cid:3404)
0.639 mM–1sec–1, and (cid:1863)(cid:3364)(cid:2872) (cid:3404) 3.990 sec–1. The acidification equilibrium constant,
(cid:1837)(cid:3006) = [E][H+]/[EH+] = 0.0068 mM
and also the values for (cid:1837)(cid:3004)(cid:3250) = 0.026 mM and (cid:1837)(cid:3004)(cid:3251) = 0.039 mM were taken from the literature
(10)
[53,70,71]. Thus, the effective rate constants become time-dependent via pH(t). Next, we set up
rate equations for the chemicals involved in the “esterase” part of the cascade. Here we only
(cid:1856)(cid:1829)(cid:3003) (cid:4666)(cid:1872)(cid:4667)(cid:1856)(cid:1872) (cid:3404) (cid:3397)(cid:1863)(cid:2871)(cid:1828)(cid:4666)(cid:1872)(cid:4667)(cid:1831)(cid:4666)(cid:1872)(cid:4667) (cid:3398) (cid:1863)(cid:2872)(cid:1829)(cid:3003) (cid:4666)(cid:1872)(cid:4667)
show one of these equations for illustration,
The buffering step of the kinetics is actually practically instantaneous. The conventional
estimates for buffer functioning cannot be used because of the large range of pH variation (from
numerically. Specifically, for Tris buffer [72] we have (cid:4666)p(cid:1837)(cid:3028) (cid:4667)(cid:2904)(cid:2928)(cid:2919)(cid:2929) (cid:3404) 8.06. The processes
9 to about 4.2). The charge-balance equations, as well as dissociation equilibrium equations for
each of the relevant chemicals, were therefore included in the overall set of equations solved
4H(cid:2878) (cid:3397) (cid:1838) (cid:3397) 4(cid:1832) (cid:3397) O(cid:2870) (cid:1863)(cid:2873)⇄(cid:1863)(cid:2879)(cid:2873)(cid:1829)(cid:3013) (cid:1863)(cid:2874)→ (cid:1838) (cid:3397) 4(cid:1842) (cid:3397) 2H(cid:2870)O
biocatalyzed by laccase can be schematically described as follows [73]:
The actual mechanism of action of laccase is more complicated than that schematically shown in
eq. (12): A more detailed description of the kinetic model is explained in our earlier paper [58].
We can now write the set of rate equations for the concentrations of the various chemicals
(cid:1856)(cid:1838)(cid:1856)(cid:1872) (cid:3404) (cid:3398)(cid:1863)(cid:3560) (cid:2873)(cid:4666)pH(cid:4666)(cid:1872)(cid:4667)(cid:4667)(cid:1838)(cid:4666)(cid:1872)(cid:4667)(cid:1832)(cid:4666)(cid:1872)(cid:4667) (cid:3397) (cid:1863)(cid:3560) (cid:2874)(cid:4666)pH(cid:4666)(cid:1872)(cid:4667)(cid:4667)(cid:1829)(cid:3013) (cid:4666)(cid:1872)(cid:4667)
involved, for example:
– 10 –
(13)
(11)
(12)
Results and Discussion
A similar approach was used for the simulation of the biochemical AND gate logic
system. A set of kinetic differential equations was based on chemical reactions described by the
eq. (2-4), and was numerically solved to study the kinetic processes in the system.
5.
The kinetic numerical model can be used for investigations of the biochemical system not
only with the filter (OR gate, Figures 2, 3 and AND gate, Figure 5), but also without the filter
(with the same rate parameter values: OR gate, Figure 1 and AND gate, Figure 4). In the figures,
the experimental data are shown as compared to a model calculation. We focus on the results at
the selected gate time. Using this model we have received a good-quality semi-quantitative fit for
the measured experimental data, with the exception, perhaps, of those cases when the
concentrations of inputs are close to the points of logical 0, which is explained generally by the
gates” for information processing needs a description of the system response function (cid:1878)(cid:4666)(cid:1876) , (cid:1877)(cid:4667)
high level of relative noise for small concentrations. As described in earlier works [42-46,57,58],
estimation of noise-handling properties and quality of the realized biochemical systems as “logic
using the logic variable ranges. The positions of the logical points are usually chosen on the basis
the output, (cid:3427)NAD(cid:2878) (cid:3431)0(cid:3435)(cid:1872)(cid:3034) (cid:3439), as the value at the 00 input. Therefore, for the parameters used for our
of the relevant physiological conditions. For example, the non-filtered AND function (Figure 4)
has insufficient quality to determine with satisfactory accuracy the position of a logical zero of
model, (cid:3427)NAD(cid:2878) (cid:3431)0(cid:3435)(cid:1872)(cid:3034) (cid:3439) ≃ 0.2·10–2 mM and (cid:3427)NAD(cid:2878) (cid:3431)1(cid:3435)(cid:1872)(cid:3034) (cid:3439) ≃ 29.8·10–2 mM. The actual definition
"filtered" experiment, see Figure 5, we used the calculated values [57] from our numerical
of the expected values or ranges of the system outputs corresponding to logic 1 may vary
according to the logic gate application. As mentioned earlier, the systems without the filter
(Figures 1, 4) are characterized by an appreciable increase of the input signal noise upon
processing by the gate to yield the output.
The correctly filtered OR gate (Figure 2) and AND gate (Figure 5) systems do not have
this problem, because the added filtering process makes the noise transmission factor at all four
– 11 –
approximately 14% for the T(cid:4666)0(cid:4667) (cid:3404) 4 mM buffer concentration (OR gate realization, Figure 2)
logic points into actual noise suppression. For example, the largest transmission factor is
100%. The third filtered OR gate system (Figure 3), T(cid:4666)0(cid:4667) (cid:3404) 8 mM, has minor noise
and it is about 8% for the realized filtering AND gate system (Figure 5), which is safely below
amplification, about 113%, and the gate realization is inaccurate: this system is an example of
undesirable “overfiltering.”
logic-point values the slopes (cid:3627)(cid:1487)(cid:4652)(cid:4652)(cid:1318)(cid:1878)(cid:4666)(cid:1876) , (cid:1877)(cid:4667)(cid:3627), in terms of the rescaled “logic-range” variables, are well
As expected for a good-quality filter of AND gate and OR gate logic systems, at the
below 1 (below 100%), resulting in noise suppression. Note that we can estimate the quality of
the filter using the “noise tolerance” parameter specifying the extent of the possible input noise
near the logical points. For the non-filtered system, the noise tolerance is zero, because it is clear
from Figure 1 and Figure 4 that the largest slopes of these gate-response functions are
the value of (cid:3627)(cid:1487)(cid:4652)(cid:4652)(cid:1318)(cid:1878)(cid:4666)(cid:1876) , (cid:1877)(cid:4667)(cid:3627) (cid:3406) 4.5 (indicating
significantly larger than 1. This is a common property of most biocatalytic reactions [15,43,52].
The non-filtered AND and OR gates will significantly amplify any input noise. Numerical
estimates have shown
input-to-output noise
amplification by a factor of approximately 450%) for the AND gate and 6.94 for the OR gate,
respectively. The main reason to adding the filter processes has been to reduce the slopes at all
the logic points to well below 1 and to maximize the noise tolerance.
To conclude, we reviewed the experimental realization and numerical performance
analysis of enzymatic OR [58] and AND [57] gate systems with a noise-reducing double-
sigmoid response in two of the inputs. For the investigated enzymatic reactions, we identified the
regimes of the experimental conditions necessary to have a relatively small degree of analog
noise amplification.
The authors gratefully acknowledge funding of this research by the U. S. National
Science Foundation (via awards CCF-1015983 and CBET-1066397 at Clarkson University), and
sponsorship by the Laboratory Directed Research and Development Program of Oak Ridge
National Laboratory (managed by UT-Battelle, LLC, for the U. S. Department of Energy).
– 12 –
References
[1] Credi, A. (2007). Molecules that make decisions. Angew. Chem. Int. Ed. 46, 5472-5475.
[2] De Silva, A.P., Uchiyama, S., Vance, T.P., Wannalerse, B. (2007). A supramolecular
chemistry basis for molecular logic and computation. Coord. Chem. Rev. 251, 1623-
1632.
[3] Pischel, U. (2007). Chemical approaches to molecular logic elements for addition and
subtraction. Angew. Chem. Int. Ed. 46, 4026-4040.
[4] Szacilowski, K. (2008). Digital information processing in molecular systems. Chem. Rev.
108, 3481-3548.
[5] Andreasson, J., Pischel, U. (2010). Smart molecules at work-mimicking advanced logic
operations. Chem. Soc. Rev. 39, 174-188.
[6] Pischel, U. (2010). Digital operations with molecules - Advances, challenges, and
perspectives. Austral. J. Chem. 63, 148-164.
[7] Ashkenasy, G., Ghadiri, M.R. (2004). Boolean logic functions of a synthetic peptide
network. J. Am. Chem. Soc. 126, 11140-11141.
[8] Benenson, Y., Gil, B., Ben-Dor, U., Adar, R., Shapiro, E. (2004). An autonomous
molecular computer for logical control of gene expression. Nature 429, 423-429.
[9] Stojanovic, M.N., Stefanovic, D., LaBean, T., Yan, H. (2005). Computing with nucleic
acids. In: Willner, I., Katz, E. (Eds.), Bioelectronics: From Theory to Applications,
Wiley-VCH, Weinheim, pp. 427-455.
[10] Shapiro, E., Gil, B. (2007). Biotechnology - Logic goes in vitro. Nature Nanotechnol. 2,
84-85.
[11] Win, M.N., Smolke, C.D. (2008). Higher-order cellular information processing with
synthetic RNA devices. Science 322, 456-460.
[12] Benenson, Y. (2009). Biocomputers: from test tubes to live cells. Molecular Biosystems
5, 675-685.
[13] Benenson, Y. (2012). Biomolecular computing systems: principles, progress and
potential. Nature Rev. Genetics 13, 455-468
[14] Adar, R., Benenson, Y., Linshiz, G., Rosner, A., Tishby, N., Shapiro, E. (2004).
Stochastic computing with biomolecular automata. Proc. Natl. Acad. USA 101, 9960-
– 13 –
9965.
[15] Katz, E., Privman, V. (2010). Enzyme-based logic systems for information processing.
Chem. Soc. Rev. 39, 1835-1857.
[16] Katz, E. (2012). Molecular and Supramolecular Information Processing: From Molecular
Switches to Logic Systems. (Ed.), Willey-VCH, Weinheim, (ISBN-10: 3-527-33195-6).
[17] Katz, E. (2012). Biomolecular Computing: From Logic Systems to Smart Sensors and
Actuators. (Ed.), Willey-VCH, Weinheim, (ISBN-10: 3-527-33228-6).
[18] Unger, R., Moult, J. (2006). Towards computing with proteins. Proteins 63, 53-64.
[19] Ezziane, Z. (2006). DNA computing: applications and challenges. Nanotechnology 17,
R27-R39.
[20] Margulies, D., Hamilton, A.D. (2009). Digital analysis of protein properties by an
ensemble of DNA quadruplexes. J. Am. Chem. Soc. 131, 9142-9143.
[21] Rinaudo, K., Bleris, L., Maddamsetti, R., Subramanian, S., Weiss, R., Benenson, Y.
(2007). A universal RNAi-based logic evaluator that operates in mammalian cells. Nat.
Biotechnol. 25, 795-801.
[22] Tamsir, A., Tabor, J.J., Voigt, C.A. (2011). Robust multicellular computing using
genetically encoded NOR gates and chemical ‘wires’. Nature 469, 212-215.
[23] Li, Z., Rosenbaum, M.A., Venkataraman, A., Tam, T.K., Katz, E., Angenent, L.T.
(2011). Bacteria-based AND logic gate: A decision-making and self-powered biosensor.
Chem. Commun. 47, 3060-3062.
[24] Calude, C.S., Costa, J.F., Dershowitz, N., Freire, E., Rozenberg, G. (2009).
Unconventional Computation. Lecture Notes in Computer Science, (Eds.), Vol. 5715,
Springer, Berlin.
[25] Adamatzky, A., De Lacy Costello, B., Bull, L., Stepney, S., Teuscher, C. (2007).
Unconventional Computing, (Eds.), Luniver Press, UK.
[26] Baron, R., Lioubashevski, O., Katz, E., Niazov, T., Willner, I. (2006). Logic gates and
elementary computing by enzymes. J. Phys. Chem. A 110, 8548-8553.
[27] Strack, G., Pita, M., Ornatska, M., Katz, E. (2008). Boolean logic gates using enzymes as
input signals. ChemBioChem 9, 1260-1266.
[28] Zhou, J., Arugula, M.A., Halámek, J., Pita, M., Katz, E. (2009). Enzyme-based universal
NAND and NOR logic gates with modular design. J. Phys. Chem. B 113, 16065-16070.
– 14 –
[29] Manesh, K.M., Halámek, J., Pita, M., Zhou, J., Tam, T.K., Santhosh, P., Chuang, M.-C.,
Windmiller, J.R., Abidin, D., Katz, E., Wang, J. (2009). Enzyme logic gates for the
digital analysis of physiological level upon injury. Biosens. Bioelectron. 24, 3569-3574.
[30] Pita, M., Zhou, J., Manesh, K.M., Halámek, J., Katz, E., Wang, J. (2009). Enzyme logic
gates for assessing physiological conditions during an injury: Towards digital sensors and
actuators. Sens. Actuat. B 139, 631-636.
[31] Halámek, J., Windmiller, J.R., Zhou, J., Chuang, M.-C., Santhosh, P., Strack, G.,
Arugula, M.A., Chinnapareddy, S., Bocharova, V., Wang, J., Katz, E. (2010).
Multiplexing of injury codes for the parallel operation of enzyme logic gates. Analyst
135, 2249-2259.
[32] Baron, R., Lioubashevski, O., Katz, E., Niazov, T., Willner, I. (2006). Elementary
arithmetic operations by enzymes: A model for metabolic pathway based computing.
Angew. Chem. Int. Ed. 45, 1572-1576.
[33] Niazov, T., Baron, R., Katz, E., Lioubashevski, O., Willner, I. (2006). Concatenated logic
gates using four coupled biocatalysts operating in series. Proc. Natl. Acad. Sci. USA 103,
17160-17163.
[34] Strack, G., Ornatska, M., Pita, M., Katz, E. (2008). Biocomputing security system:
Concatenated enzyme-based logic gates operating as a biomolecular keypad lock. J. Am.
Chem. Soc. 130, 4234-4235.
[35] Privman, M., Tam, T.K., Pita, M., Katz, E. (2009). Switchable electrode controlled by
enzyme logic network system: Approaching physiologically regulated bioelectronics, J.
Am. Chem. Soc. 131, 1314-1321.
[36] Tam, T.K., Pita, M., Katz, E. (2009). Enzyme logic network analyzing combinations of
biochemical inputs and producing fluorescent output signals: Towards multi-signal
digital biosensors. Sens. Actuat. B 140, 1-4.
[37] Windmiller, J.R., Strack, G., Chuan, M.-C., Halámek, J., Santhosh, P., Bocharova, V.,
Zhou, J., Katz, E., Wang, J. (2010). Boolean-format biocatalytic processing of enzyme
biomarkers for the diagnosis of soft tissue injury. Sens. Actuat. B 150, 285-290.
[38] May, E.E., Dolan, P.L., Crozier, P.S., Brozik, S., Manginell, M. (2008). Towards de novo
design of deoxyribozyme biosensors for GMO detection. IEEE Sens. J. 8, 1011-1019.
[39] Wang, J., Katz, E. (2010). Digital biosensors with built-in logic for biomedical
– 15 –
applications - biosensors based on biocomputing concept. Anal. Bioanal. Chem. 398,
1591-1603.
[40] Kahan, M., Gil, B., Adar, R., Shapiro, E. (2008). Towards molecular computers that
operate in a biological environment. Physica D 237, 1165-1172.
[41] Wang, J., Katz, E. (2011). Digital biosensors with built-in logic for biomedical
applications. Isr. J. Chem. 51, 141-150.
[42] Melnikov, D., Strack, G., Zhou, J., Windmiller, J.R., Halámek, J., Bocharova, V.,
Chuang, M.-C., Santhosh, P., Privman, V., Wang, J., et al. (2010). Enzymatic AND logic
gates operated under conditions characteristic of biomedical applications. J. Phys. Chem.
B 114, 12166-12174.
[43] Privman, V., Strack, G., Solenov, D., Pita, M., Katz, E. (2008). Optimization of
enzymatic biochemical
logic for noise reduction and scalability: How many
biocomputing gates can be interconnected in a circuit? J. Phys. Chem. B 112, 11777-
11784.
[44] Privman, V., Arugula, M.A., Halámek, J., Pita, M., Katz, E. (2009). Network analysis of
biochemical logic for noise reduction and stability: A system of three coupled enzymatic
AND gates. J. Phys. Chem. B 113, 5301-5310.
[45] Privman, V., Halámek, J., Arugula, M. A., Melnikov, D., Bocharova, V., Katz, E. (2010).
Biochemical filter with sigmoidal response: Increasing the complexity of biomolecular
logic. J. Phys. Chem. B 114, 14103-14109.
[46] Privman, V., Pedrosa, V., Melnikov, D., Pita, M., Simonian, A., Katz, E. (2009).
Enzymatic AND-gate
based
on
electrode-immobilized
glucose-6-phosphate
dehydrogenase: Towards digital biosensors and biochemical logic systems with low
noise. Biosens. Bioelect. 25, 695-701.
[47] Researchers from Clarkson University report details of new studies and findings in the
area of biosensors and bioelectronics. Electronics Newsweekly (Atlanta, GA), Issue: Jan.
13, 2010, Page: 161,
(online version at http://www.verticalnews.com/article.php?articleID=3010969).
[48] Buchler, N.E., Gerland, U., Hwa, T. (2005). Nonlinear protein degradation and the
function of genetic circuits. Proc. Natl. Acad. Sci. USA 102, 9559-9564.
[49] Setty, Y., Mayo, A.E., Surette, M.G., Alon, U. (2003). Detailed map of a cis-regulatory
– 16 –
input function. Proc. Natl. Acad. Sci. USA 100, 7702-7707.
[50] Alon, U. (2007). An Introduction to Systems Biology. Design Principles of Biological
Circuits, Boca Raton, Florida: Chapman & Hall/CRC Press.
[51] Halámek, J., Zhou, J., Halámková, L., Bocharova, V., Privman, V., Wang, J., Katz, E.
(2011). Biomolecular Filters for Improved Separation of Output Signals in Enzyme Logic
Systems Applied to Biomedical Analysis. Anal. Chem. 83, 8383-8386.
[52] Privman, V. (2011). Control of Noise in Chemical and Biochemical Information
Processing. Isr. J. Chem. 51, 118-131.
[53] Pita, M., Privman, V., Arugula, M.A., Melnikov, D., Bocharova, V., Katz, E. (2011).
Towards Biochemical Filter with Sigmoidal Response to pH Changes: Buffered
Biocatalytic Signal Transduction. Phys. Chem. Chem. Phys. 13, 4507-4513.
[54] Rafael, S.P., Vallée-Bélisle, A., Fabregas, E., Plaxco, K., Palleschi, G., Ricci, F. (2012).
Employing the Metabolic “Branch Point Effect” to Generate an All-or-None, Digital-like
Response in Enzymatic Outputs and Enzyme-Based Sensors. Anal. Chem. 84, 1076-1082.
[55] Moser, I., Jobst, G., Svasek, P., Varahram, M., Urban, G. (1997). Rapid liver enzyme
assay with miniaturized liquid handling system comprising thin film biosensor array.
Sens. Actuators B 44. 377-380.
[56] Kato, G.J., McGowan, V., Machado, R.F., Little, J.A., Taylor, VI, J., Morris, C.R.,
Nichols, J.S., Wang, X., Poljakovic, M., Morris, Jr., S.M., Gladwin, M.T. (2006). Lactate
dehydrogenase as a biomarker of hemolysis-associated nitric oxide resistance, priapism,
leg ulceration, pulmonary hypertension, and death in patients with sickle cell disease.
Blood 107, 2279-2285.
[57] Halámek, J., Zavalov, O., Halámková, L., Korkmaz, S., Privman, V., Katz, E. (2012).
Enzyme-Based Logic Analysis of Biomarkers at Physiological Concentrations: AND
Gate with Double-Sigmoid “Filter” Response. The Journal of Physical Chemistry B 116,
4457-4464.
[58] Zavalov, O., Bocharova, V., Privman, V., Katz, E. (2012). Enzyme-Based Logic: OR
Gate with Double-Sigmoid Filter Response. The Journal of Physical Chemistry B 116,
9683–9689.
[59] Kotoh, K., Enjoji, M., Kato, M., Kohjima, M., Nakamuta, M., Takayanagi, R. (2008). A
new parameter using serum lactate dehydrogenase and alanine aminotransferase level is
– 17 –
useful for predicting the prognosis of patients at an early stage of acute liver injury: A
retrospective study. Compar. Hepatol. 7, 6-14.
[60] Khalili, H., Dayyeh, B.A., Friedman, L.S. (2010). In: Clinical Gastroenterology: Chronic
Liver Failure, Ginès, P., Kamath, P.S., Arroyo, V. (Eds.) Humana Press, New York, 47-
76.
[61] Zhou, J., Halámek, J., Bocharova, V., Wang, J., Katz, E. (2011). Talanta 83, 955-959.
[62] Privman, M., Tam, T.K., Bocharova, V., Halámek, J., Wang, J., Katz, E. (2011).
Responsive interface switchable by logically processed physiological signals: toward
"smart" actuators for signal amplification and drug delivery. ACS Appl. Mater. Interfaces
3, 1620-1623.
[63] Tokarev, I., Gopishetty, V., Zhou, J., Pita, M., Motornov, M., Katz, E., Minko, S. (2009).
Stimuli-responsive hydrogel membranes coupled with biocatalytic processes. ACS Appl.
Mater. Interfaces 1, 532-536.
[64] Motornov, M., Zhou, J., Pita, M., Tokarev, I., Gopishetty, V., Katz, E., Minko, S. (2009).
Integrated multifunctional nanosystem from command nanoparticles and enzymes. Small
5, 817-820.
[65] Motornov, M., Zhou, J., Pita, M., Gopishetty, V., Tokarev, I., Katz, E., Minko, S. (2008).
“Chemical transformers” from nanoparticle ensembles operated with logic. Nano Lett. 8,
2993-2997.
[66] Bychkova, V., Shvarev, A., Zhou, J., Pita, M., Katz, E. (2010). Enzyme logic gate
associated with a single responsive microparticle: Scaling biocomputing to microsize
systems. Chem. Commun. 46, 94-96.
[67] Bocharova, V., Tam, T.K., Halámek, J., Pita, M., Katz, E. (2010). Reversible gating
controlled by enzymes at nanostructured interface. Chem. Commun. 46, 2088-2090.
[68] Adler, A.J., Kristiakowsky, G.B. (1962). J. Am. Chem. Soc. 84, 695-703.
[69] Bergmeyer, H.U. (1974). Methods of Enzymatic Analysis, 2nd ed., (Ed.) Academic Press,
New York, Vol. 4, 2066-2072.
[70] Craig, N.C., Kistiakowsky, G.B. (1958). Kinetics of Ester Hydrolysis by Horse Liver
Esterase. J. Am. Chem. Soc. 80, 1574-1579.
[71] Kyger, E.M., Riley, D.J., Spilburg, C.A., Lange, L.G. (1990). Pancreatic cholesterol
esterases. 3. Kinetic characterization of cholesterol ester resynthesis by the pancreatic
– 18 –
cholesterol esterases. Biochemistry 29, 3853-3858.
[72] Millero, F.J. (2009). Use of the Pitzer Equations to Examine the Dissociation of TRIS in
NaCl Solutions. J. Chem. Eng. Data. 54, 342-344.
[73] Koudelka, G.B., Hansen, F.B., Ettinger, M.J. (1985). Solvent isotope effects and the pH
dependence of laccase activity under steady-state conditions. J. Biol. Chem. 260, 15561-
15565.
– 19 –
Nomenclature
Alanine transaminase
Lactate dehydrogenase
Esterase concentration
Ethyl butyrate concentration
Methyl butyrate concentration
Butyric acid concentration
Tris buffer concentration
Protonated form of Tris buffer
Laccase concentration
K4Fe(CN)6 concentration
K3Fe(CN)6 concentration
Logic input
Logic input
Logic gate response surface
Glucose-6-phosphate dehydrogenase
D-glucose-6-phosphate
β-Nicotinamide adenine dinucleotide cofactor
β-Nicotinamide adenine dinucleotide reduced cofactor
6-phospho-gluconic acid
α-Ketoglutaric acid
Alanine
Glutamate
Pyruvate
Lactate
0.5 U = 4.85×10–5 mM
0.5 U = 3.78×10–4 mM
1 mM
10 mM
200 mM
– 20 –
ALT
LDH
[E]
[A]
[B]
[U]
[T]
TH+
[L]
[F]
[P]
x
y
z(x,y)
G6PDH
Glc6P
NAD+
NADH
6-PGluc
α-KTG
Ala
Glu
Pyr
Lac
Figure 1: Realization of an OR gate for the biochemical system without filtering, by
biochemical processes detailed in the text. Spherical symbols show the 7×7 grid of 49
experimental data points. These were measured at the gate time 800 s, for various values of the
two input concentrations (ethyl butyrate [A] and methyl butyrate [B]), the logic point values for
both of which were set at 0 and 10 mM concentrations (for 0 and 1, respectively). The logic point
values of the measured output, measured by the absorbance of the product (see text), are set by
the gate function itself. The shown surface was calculated from the model described in the text.
– 21 –
Figure 2: Experimental data and model-calculated surface for various values of the two input
concentrations of ethyl butyrate and methyl butyrate (as in Figure 1) for the OR gate system with
4 mM buffer added initially (the optimal filtering process), gate time 800 s.
– 22 –
Figure 3: Experimental data and model-calculated surface (as in Figures 1 and 2) for the OR
gate system with 8 mM buffer added initially, gate time 800 s. The OR function is not well
realized in this over-filtered system.
– 23 –
Figure 4: Realization of an AND gate for the biochemical system without the filter. Spherical
symbols show the 108 (12×9) experimental data points. These were measured at the gate time t =
600 s, for various values of the two input enzyme concentrations as detailed in the text. The lines
under the surface are drawn at the selected logic-0 and 1 values of the inputs and thus delineate
the "logic" range for mapping out the gate-response function. The surface was calculated from
the model described in the text, with the parameters fitted based on the full time-dependent data
set for the “filtered” system.
– 24 –
Figure 5: The AND function realization for the biochemical system with the filter process.
Spherical symbols show the 108 (12×9) experimental data points, measured at the gate time t =
600 s. The lines under the surface mark the selected logic-0 and 1 values of the inputs. The
surface was calculated from the model described in the text, with the parameters fitted based on
the full time-dependent data set for the “filtered” system.
– 25 –
Scheme 1. Chemical and biochemical processes in the OR gate system. All the notations are
defined in the text. The two inputs are marked by the dashed-line boxes, whereas the output is
marked by the solid box.
– 26 –
Scheme 2. Biocatalytic cascade realizing the AND gate function with the filter process added.
All the notations are defined in the text. The two inputs are marked by the dashed-line boxes,
whereas the output is marked by the solid box.
– 27 –
|
1512.06024 | 2 | 1512 | 2017-03-08T14:21:45 | Persistence-driven durotaxis: Generic, directed motility in rigidity gradients | [
"physics.bio-ph",
"cond-mat.soft",
"q-bio.CB"
] | Cells move differently on substrates with different elasticities. In particular, the persistence time of their motion is higher on stiffer substrates. We show that this behavior will result in a net transport of cells directed up a soft-to-stiff gradient. Using simple random walk models with controlled persistence and stochastic simulations, we characterize this propensity to move in terms of the durotactic index measured in experiments. A one-dimensional model captures the essential features of this motion and highlights the competition between diffusive spreading and linear, wavelike propagation. Since the directed motion is rooted in a non-directional change in the behavior of individual cells, the motility is a kinesis rather than a taxis. Persistence-driven durokinesis is generic and may be of use in the design of instructive environments for cells and other motile, mechanosensitive objects. | physics.bio-ph | physics | Persistence-driven durotaxis: Generic, directed motility in rigidity gradients
Elizaveta A. Novikova1,4, Matthew Raab2, Dennis E. Discher3 and Cornelis Storm4,5
1 Institute for Integrative Biology of the Cell(I2BC),
Institut de Biologie et de Technologies de Saclay(iBiTec-S), CEA,
CNRS, Universite Paris Sud, F-91191 Gif-sur-Yvette cedex, France,
2 CNRS UMR144, Institut Curie, 12 rue Lhomond, 75005 Paris,
France 3Molecular & Cell Biophysics and Graduate Group in Physics,
University of Pennsylvania, Philadelphia, Pennsylvania 19104, USA,
4Department of Applied Physics, 5Institute for Complex Molecular Systems,
Eindhoven University of Technology, P. O. Box 513, NL-5600 MB Eindhoven, The Netherlands
(Dated: September 24, 2018)
7
1
0
2
r
a
M
8
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
4
2
0
6
0
.
2
1
5
1
:
v
i
X
r
a
Cells move differently on substrates with different rigidities: The persistence time of their motion is
higher on stiffer substrates. We show that this behavior -- in and of itself -- results in a net flux of cells
directed up a soft-to-stiff gradient. Using simple random walk models with varying persistence and
stochastic simulations, we characterize the propensity to move in terms of the durotactic index also
measured in experiments. A one-dimensional model captures the essential features and highlights
the competition between diffusive spreading and linear, wavelike propagation. Persistence-driven
durokinesis is generic and may be of use in the design of instructive environments for cells and other
motile, mechanosensitive objects.
PACS numbers: 87.17.Aa, 87.17.Jj, 87.10.-e
Cells are acutely aware of the mechanical properties
of their surroundings. The rigidity, or lack thereof, of
the substrate to which a cell is adhering determines a
number of crucial processes: Differentiation, gene expres-
sion, proliferation, and other cellular decisions have all
been shown to be affected by the stiffness of the sur-
rounding matrix [1 -- 7]. Cells also move differently de-
pending on the rigidity of the substrate. One of the
more striking manifestations of this is the near-universal
tendency of motile cells to travel up rigidity gradients
in a process generally referred to as durotaxis [8 -- 15], a
term that emphasizes the similarity to chemotaxis, the
ability of cells to move directedly in chemical gradients.
Chemotaxis -- generally believed to offer significant evo-
lutionary advantage -- allows cells, for instance, to move
towards sources of nutrients. For durotaxis, such advan-
tage is less obvious. Motion in stiffness gradients could
allow neutrophils and cancer cells to seek out optimal
locations for extravasation [16 -- 18], or stem cells to con-
tribute to mitigation or regeneration of stiff scars and
injured tissues [19]. Durotactic motion is universal: with-
out exception it is away from softer, towards stiffer. In
addition to an overall motion in a gradient region, the na-
ture of cellular motion itself was shown to change quan-
titatively depending directly on the local rigidity of the
substrate, with cells moving more persistently on more
rigid substrates. In this Letter, we examine how locally
different, persistent motility affects the global transport
of cells. We find that soft-to-stiff durotaxis is a necessary
consequence of stiffness-dependent persistence, with or
without a rigidity-dependent crawling speed. The mech-
anism we uncover is fundamentally different from those
reported in earlier theoretical works on durotaxis [20, 21]:
the cells take no directional cues from the gradient re-
FIG. 1: Persistence-dependent motility. Simulated trajecto-
ries (2D model) of 25 cells, departing from the origin at t = 0
with a linear velocity of 50 µm/hr. Total time is 12 hrs, cel-
lular positions are recorded at 6-minute intervals. A black
dot marks the end of each cell trajectory. (a) Cells on a soft
substrate, with a low persistence time τp = 0.2 hrs. (b) stiff
substrate; persistence time τp =2 hrs. (c) Gradient substrate,
with persistence time increasing linearly from 0.2 to 2 hrs over
the x-range [−0.1, 0.1] mm (i.e., ∆τp/∆x = 9 hrs/mm). (d)
averaged x-displacement in the gradient region, for different
widths of the gradient regions, and hence the gradient steep-
nesses (top to bottom: ∆τp/∆x = 90 hrs/mm, 18 hrs/mm, 9
hrs/mm, 4.5 hrs/mm, 1.8 hrs/mm).
gion, but their persistence -- a nondirectional property --
is stiffness dependent. This experimentally established
2
PRW alike: (cid:104)(cid:126)r(cid:105)(t) = (cid:126)0 -- this is no longer the case for
durotactic processes. A meaningful question, now, is to
ask how the parameters that quantify persistence and
directed displacement change with the properties of the
substrate. While the tendency to move from soft to stiff
substrates has been broadly noted and characterized [24 --
28], the persistence of cells as they do so has only recently
begun to be quantitatively addressed. A potential rela-
tion between the two has been hinted at in passing [12],
but not further substantiated. In experiments recording
the motility of fibroblasts on uniformly rigid PEG hy-
drogels, Missirlis and Spatz [13] demonstrate that the
persistence time, quantified by a Directionality Index
∆(t) =(cid:112)(cid:104)(cid:126)r2(cid:105)(t)/(vct) ∝ (τp/t)1/2 recorded at the same
time on substrates coated with different ligands, rises by
about a factor of 3 when the substrate stiffness is in-
creased from 5.5 to 65.7 kPa. Over the same range of
stiffnesses, a decrease of vc by about 33% (from 60 µm/hr
to 40 µm/hr) is reported. House et al [29] place fibrob-
lasts on uniformly rigid PAM hydrogels, and report that
their persistence time increases by a factor of 3 when the
gel stiffness is varied from 10 kPa to 150 kPa. Interest-
ingly, and in contrast to Missirlis and Spatz, House et
al. report an increase of vc with substrate stiffness by a
factor of about 2 from 21.6 µm/hr to 42.7 µm/hr over
the same stiffness range. A preliminary test, reported in
[29], suggests the cells move in the direction of increased
persistence.
In earlier work [12], Raab et al. quantify
the motility of mesenchymal stem cells on uniform PAM
substrates - likewise showing an increase in persistence
time of about a factor of 3 from 0.7 hrs to 2.1 hrs when
the substrate stiffness is varied from 1 kPa to 34 kPa.
Raab et al. report no significant change in the cell ve-
locity vc over the entire range of stiffnesses they study.
Importantly, however, Raab et al. also show that the
same cells, on the same substrates that are now gradi-
ented in stiffness from 1 kPa to 34 kPa, move towards
the stiff side with a durotaxis index that over the course
of about 2 hrs rises from 0 to 0.2. In summary, exper-
iments unanimously suggest that cells move more per-
sistently on stiffer substrates, and that they move from
soft to rigid. This behavior is independent of the relation
between velocity and stiffness, which appears to be more
cell-type dependent although a recent work suggests that
speed and persistence may be correlated [15]. The empir-
ical fact that two behaviors -- increasing persistence and
soft-to-stiff motion -- coincide suggests they might not be
independent. We now examine whether there is indeed a
causation underlying the correlation.
Simulation setup and results. We consider a 2D sub-
strate, endowed with a gradient in stiffness that mani-
fests itself as a position-dependent persistence time τp(x)
and a position-dependent velocity vc(x). To simulate the
variable-persistence, variable cell speed PRW in the gra-
dient region, we generate trajectories as follows: Starting
in the origin at t = 0, a random initial direction θ0 is
FIG. 2: Evolution of probability with time. Simulated trajec-
tories (2D model) of 50 cells, departing from the origin at
t = 0 with a linear velocity of 50 µm/hr on a persistence gra-
dient, increasing linearly from 0.2 to 2 hrs over the x-range
[−0.1, 0.1] mm (i.e., ∆τp/∆x = 9 hrs/mm). The cells were
tracked for 12 hrs, their positions recorded at 6-minute in-
tervals. A black dot marks the end of each cell trajectory.
(a)-(d) As time progresses, the asymmetry becomes increas-
ingly clear. (e) The probability distribution P(x, y) (rescaled
such that its maximal value is 1) at t = 4 hrs clearly shows
a double-peaked structure: a diffusive peak on the soft side,
and a wavefront further out on the rigid side.
fact, alone, suffices to generate durotactic motion.
Definitions and experimental observations. For cells
moving on uniformly rigid substrates, most experiments
record the paths of motile cells by tabulating, at fixed
time intervals ∆t = ti+1 − ti, their position (cid:126)r(ti) =
{x(ti), y(ti)}. The resulting time series constitutes a
discrete-time Random Walk (RW). These cellular RW
paths display a certain amount of persistence; the ten-
dency to keep moving along the same direction (or,
equivalently, the cell's inability to turn on very short
timescales). This persistence is quantified by the per-
sistence time τp. For cells moving at a constant linear
velocity vc, this persistence time may be obtained by an-
alyzing the displacement statistics of the path, either as
the decay time of the tangent autocorrelation, or by fit-
ting to the mean squared displacement for a persistent
random walk (PRW) [22]
(cid:104)(cid:126)r2(cid:105)(t) = 2v2
c τ 2
p
(cid:18) t
τp
(cid:19)
+ e−t/τp − 1
.
(1)
We note, that while the PRW correctly describes cellular
motility in 2D, it fails in 3D [23] -- one of many important
differences between 2D and 3D processes of cellular ad-
hesion and migration. We restrict ourselves to the case
of 2D motility here, to make our general point. The
limiting behavior of Eq.1 is instructive: for short times
t (cid:28) τp it describes ballistic motion (cid:104)(cid:126)r2(cid:105)(t) ≈ (vct)2,
whereas for long times t (cid:29) τp the motion is a pure ran-
dom walk; (cid:104)(cid:126)r2(cid:105)(t) ≈ 2v2
c τpt. Thus, the persistence time
is the characteristic timescale for the crossover between
ballistic and diffusive motion. A trivial point, which
nonetheless bears repeating here, is that the first mo-
ment of the vectorial displacement vanishes, for RW and
3
stiffness). Our main finding is summarized in Fig. 1:
a gradient in persistence produces a soft-to-stiff flux of
cells, and confers upon them, for typical values, an av-
erage velocity up the stiffness gradient of 2-10 µm/hr.
The origin of the effect is readily read off from Fig. 1
(a)-(c); PRW trajectories become asymmetric in the gra-
dient region, and those trajectories that either depart
up the rigidity gradient, or at some point in time first
turn towards the stiff direction, travel further in the
stiff direction, on average. As Fig. 1(d) illustrates, this
leads to a nonzero (cid:104)x(cid:105)(t), and the average velocity - over
the ∼ 12 hr course of a typical experiment, increases
with increasing gradient steepness. We note, that in the
limit of sufficiently small ∆t, the dimensionless number
V = vc × (∂τp/∂x) combines both parameters into a sin-
gle quantity, and allows for a universal characterization
of the durotactic motion. We choose to retain dimen-
sional quantities to provide a sense of the magnitudes
of velocities that may be expected in experimental set-
tings. As Fig. 2 (a)-(d) shows, the asymmetry of a set of
PRW trajectories on a substrate with gradient stiffness
increases with time. Fig. 2(e) plots the probability dis-
tribution P(x, y) of finding a cell at position (x, y) after
t = 4 hrs and shows the crucial statistical feature that
gives rise to the nonzero center-of-mass motion. On the
left, less persistent, side of the substrate the distribution
resembles that of a diffusive process. On the right side,
where motion is more persistent, a narrower peak moves
outward at constant velocity.
The net motion that results from differentially persis-
tent PRW's executed in a stiffness gradient is reminiscent
of the motion that chemotactic bacteria execute in, for
instance, a gradient in nutrient concentration [30]. To be
sure, in both cases an environmental gradient sets up a
flux, but to what extent are these processes truly simi-
lar? Following [27, 28], it is instructive to scrutinize the
motility using a durotactic (vector-)index
(cid:126)DI(t) = {DIx(t), DIy(t)} ≡ (cid:104)(cid:126)r(cid:105)(t)
(2)
.
vct
by dividing (cid:104)(cid:126)r(cid:105)(t) by rpath =(cid:82)vc((cid:126)r(t(cid:48)))dt(cid:48), the length of
In the case of variable cell-speed vc, we compute (cid:126)DI(t)
the path traveled up to time t. For all - persistent and
non-persistent - non-directional processes (cid:126)DI(t) = (cid:126)0. For
the gradients studied here DIy(t) = 0; we report only
the x-component. In the main panel of Fig. 3, we plot
DIx(t) for a representative set of parameters (listed in
the caption). The general behavior is, that DIx(t) ini-
tially rises, peaks at a few times the persistence time,
and then slowly drops back down, proportional to t−1/2
(cf., inset Fig. 4). Fig 3 also shows, that this behavior
remains qualitatively the same regardless of whether vc
increases, decreases or stays the same through the gradi-
ent region. Since the DI is directly proportional to the
average velocity in the direction of the gradient, this is
also the expected behavior for the average velocity (see
FIG. 3: Durotactic index as function of time. Main figure:
x-component of the durotactic index vs time for cells mov-
ing in a rigidity gradient, with τp increasing linearly from
0.2 to 2 hrs over the x-range [−0.1, 0.1] mm. Averages com-
puted over 5·104 trajectories (2D model). Black line, black
dots:
stiffness-independent velocity vc = 50µm/hr every-
where. Red-dashed line: the same system, but with a ve-
locity that rises with persistence; vc = 20 − 80µm/hr accross
the gradient region. Blue-dashed line: velocity decreases with
persistence; vc = 80 − 20µm/hr accross the gradient region.
Inset: The average velocity over the 12 hr window as a func-
tion of the gradient strength. All gradients had τp varying
from 0.2 to 2 hrs, but over different spatial ranges.
chosen, along which the cell is displaced by a distance
∆r1 = vc(0)∆t. For all subsequent steps, a deviation
angle −π < δθ < π is picked randomly from a Gaus-
sian distribution centered around δθ = 0 with variance
σ2 = 2∆t/τp(x) using the Box-Muller transform, x being
the instantaneous x-position. The next point is placed a
distance ∆r2 = vc(x)∆t in the θ0 + δθ direction, this last
step is repeated N = ttot/∆t times to complete a trajec-
tory representing a total time ttot. The time interval ∆t
is chosen such that ∆t < minx(τp(x)); smaller than the
smallest persistence time in the system. In all simulations
shown here we chose ∆t = 0.1 hrs (corresponding to 6-
minute intervals between measurements). The substrate
has a finite gradient region x ∈ [−W, W ], with persis-
tence time and velocity τp,min and vc,left at x ≤ −W , and
τp,max and vc,right for x ≥ W . Both τp and vc transi-
tion linearly (but with variable steepness) controlled by
W between their max and min values. Much like most
experimental settings, the gradient region thus occupies
only part of the system, and is flanked by uniformly rigid
regions to either side (i.e., the rigidity gradient changes
discontinuously at the boundaries of the gradient region).
We will always choose left to right to be the direction of
increasing persistence but will, for demonstrational pur-
poses, allow the velocity to decrease or increase from left
to right. For each realization of the gradient region, on
the order of 105 trajectories are generated to obtain ac-
curate averages.
We assume, for now, that vc(x) ≡ vc; a constant (later
on, we will briefly demonstrate that our findings are
largely insensitive to increases or decreases in vc with
inset, Fig. 3) which is thus a time dependent quantity
for this process. The large-time drop in DIx(t) is the re-
sult of walkers leaving the finite gradient region, and is
generic; it also features when walkers exhibit a true direc-
tional bias in the same region, such as would occur for the
'run-and-tumble' behavior that underlies chemotaxis in,
for instance, E. coli where it leads to a constant cellular
drift velocity vd while in a chemical gradient region [30].
The short-time behavior, however, is completely different
for these two processes: the biased run-and-tumble walk
displays a DIx(t) rapidly saturating to its plateau value
vd/vc, in contrast to a very gradual increase from zero
for the differentially persistent walk. Thus, the presence
or absence of a short-time regime of increasing DIx(t) is
a reliable way to discriminate the motion we discuss here
from a 'regular' taxis.
1D model and an inhomogeneous telegraph equation.
We map the process to one dimension by studying the
dispersal of walkers on a line. The equivalent of a spa-
tially dependent persistence, here, is a spatially depen-
dent turning frequency λ(x). Typical behavior in the
presence of a gradient region is collected in Fig. 4 and
confirms the dual behavior also seen in two dimensions:
the softer side is diffusion-dominated while the more rigid
side displays a wavelike propagation. To derive the ap-
propriate continuum equation, we apply a similar ap-
proach to the one presented for uniform turning rates
in [27], and consider separately the two densities of left-
and right movers; ρ−(x, t) and ρ+(x, t), normalized such
that P(x, t) = ρ+ + ρ−, is the total probability density.
After a time step ∆t, each walker reverses direction with
a probability π = λ(x)∆t, or continues (with probability
1−π(x)) along its prior direction. During each time step,
it travels a distance ∆x = vc∆t. The densities ρ+ and
ρ− then obey
ρ+(x, t+∆t) = [1 − λ(x−∆x)∆t] ρ+(x−∆x, t)
+ [λ(x−∆x)∆t]ρ−(x−∆x, t) ,
ρ−(x, t+∆t) = [λ(x+∆x)∆t]ρ+(x+∆x, t)
(3)
+ [1 − λ(x+∆x)∆t] ρ−(x+∆x, t) . (4)
Expanding these two equations to first order in ∆x and
∆t and combining them using P = ρ+ + ρ− yields the
following governing PDE
∂2
t P + 2λ(x)∂tP = v2
c ∂2
xP .
(5)
A spatially varying velocity may be included by replac-
ing vc → vc(x). This inhomogeneous telegraph equa-
tion is also the appropriate model to use for effectively
one-dimension migration experiments. To connect with
the two-dimensional case, we may identify 2λ(x) (cid:39) τ−1
p .
The two competing behaviors are readily recognized in
4
FIG. 4: Evolution of 1D inhomogeneous telegraph probability.
Probability distributions P(x, t) determined by direct inte-
gration of Eq. 5 (1D model). The turning frequency λ(x)
decreased linearly from 0.4 to 0.02 over the x-interval [−5, 5].
From left to right, we plot distributions for t = 10 . . . 100 with
10 unit time intervals. Clearly visible is the diffusive spread-
ing on the left, vs. the wave-like propagation to the right.
The inset shows the long-time t−1/2 behavior of DI(t).
the PDE; for large turning frequencies (i.e., short persis-
tence times) the second order time derivative is dom-
inated by the first order term, and diffusive behavior
emerges. For low turning frequencies -- highly persistent
motion -- a wave equation is recovered. This equation,
supplemented with a specific form for the persistence
gradient λ(x), and the appropriate boundary conditions
(generally, P(x, 0) = δ(x) and ∂tP(x, 0) = 0), allows one
to compute averaged displacements as moments in this
distribution.
Conclusions and Outlook. In this Letter, we demon-
strate how a broadly reported feature of cellular
motility -- a dependence of the persistence of movement
on the rigidity of the substrate -- leads, without further
assumptions, to universal soft-to-stiff motion on gradi-
ented substrates. The motion is faster, on experimental
timescales, for steeper gradients. For the type of motion
we report here, the term durotaxis may be a bit of a
misnomer. Following the suggestions laid out in [31], the
flux set up by gradients in the local, substrate-informed
persistence is perhaps more accurately described as a
(positional) kinesis -- an "almost instantaneous response
induced by a purely positional signal". That is, a non-
directional change in behavior as opposed to the direc-
tional changes typical for chemotaxis. This distinction
goes beyond semantics:
it suggests that durotaxis in a
stiffness gradient is not to be interpreted as the existence
of a preferred stiffness for the cell, which it is purpose-
fully migrating towards. Without dismissing the possi-
bility that other mechanisms not considered here could
lead to such properly durotactic motion, we show here
that -- at the very least to an extent that is worth deter-
mining in much greater detail -- soft-to-stiff migration is
an unavoidable consequence of stiffness-dependent per-
sistence. The short-time behavior of DIx(t) may help
distinguish this kinesis from properly tactic motion. The
generic nature of durokinesis suggests it as a potentially
worthwhile mechanism to pursue in the development of
instructive environments (for an early demonstration see,
for instance, [32]); our results show that any stochastic,
particulate system whose persistence is informed, locally,
by some external parameter has the potential to harness
this kinetic transport mechanism.
This work was supported by funds from the Nether-
lands Organization for Scientific Research (NWO-FOM)
within the program on Mechanosensing and Mechan-
otransduction by Cells (FOM-E1009M). We thank all
FOM staff for 70 years of passionate commitment to
Dutch physics, and Prof. James P. Butler for valuable
discussions.
[1] N. D. Evans, C. Minelli, E. Gentleman, V. LaPointe,
S. N. Patankar, M. Kallivretaki, X. Chen, C. J. Roberts,
and M. M. Stevens, Eur. Cell. Mater. 18, 1 (2009).
[2] T. Yeung, P. C. Georges, L. A. Flanagan, B. Marg,
M. Ortiz, M. Funaki, N. Zahir, W. Ming, V. Weaver,
and P. A. Janmey, Cell Motility and the Cytoskeleton
60, 24 (2005).
[3] D. Discher, P. Janmey, and Y.-L. Wang, Science 310,
[4] A. J. Engler, S. Sen, H. L. Sweeney, and D. E. Discher,
1139 (2005).
Cell 126, 677 (2006).
[5] B. Trappmann, J. Gautrot, J. Connelly, D. Strange,
Y. Li, M. Oyen, M. Cohen Stuart, H. Boehm, B. Li,
V. Vogel, et al., Nature Mat. 11, 642 (2012).
[6] R. T. Justin and A. J. Engler, PloS one 6 (2011).
[7] J. T. Pham, L. Xue, A. del Campo, and M. Salierno,
Acta Biomaterialia (2016), ISSN 1742-7061.
[8] C.-M. Lo, H.-B. Wang, M. Dembo, and Y.-l. Wang, Bio-
physical Journal 79, 144 (2000).
[9] G. Charras and E. Sahai, Nat. Rev. Mol. Cell Biol. 15,
5
Álamo, and A. J. Engler, Biotech. J. 8, 472 (2013).
[12] M. Raab, J. Swift, D. P. C. Dingal, P. Shah, J.-W. Shin,
and D. E. Discher, J. Cell Biol. 199, 669 (2012).
[13] D. Missirlis and J. P. Spatz, Biomacromolecules 15, 195
(2013).
[14] B. C. Isenberg, P. A. DiMilla, M. Walker, S. Kim, and
J. Y. Wong, Biophys. J. 97, 1313 (2009).
[15] P. Maiuri, J.-F. Rupprecht, S. Wieser, V. Ruprecht,
O. Bénichou, N. Carpi, M. Coppey, S. D. Beco, N. Gov,
C.-P. Heisenberg, et al., Cell 161, 374 (2015).
[16] J. Brábek, C. T. Mierke, D. Rösel, P. Vesel`y, and
B. Fabry, Cell Commun. Signal. 8, 22 (2010).
[17] R. A. Jannat, G. P. Robbins, B. G. Ricart, M. Dembo,
and D. A. Hammer, J. Phys.: Cond. Matt. 22, 194117
(2010).
[18] P. W. Oakes, D. C. Patel, N. A. Morin, D. P. Zitterbart,
B. Fabry, J. S. Reichner, and J. X. Tang, Blood 114,
1387 (2009).
[19] P. D. P. Dingal, A. M. Bradshaw, S. Cho, M. Raab,
A. Buxboim, J. Swift, and D. E. Discher, Nature Mat.
14, 951 (2015).
[20] B. Harland, S. Walcott, and S. X. Sun, Physical biology
8, 015011 (2011).
[21] F. Stefanoni, M. Ventre, F. Mollica, and P. A. Netti,
Journal of theoretical biology 280, 150 (2011).
[22] M. Rubinstein and R. Colby, Polymers Physics (Oxford,
[23] P.-H. Wu, A. Giri, S. X. Sun, and D. Wirtz, Proc. Natl.
Acad. Sci. USA 111, 3949 (2014).
[24] E. F. Keller and L. A. Segel, J.Theor. Biol. 30, 225
2003).
(1971).
[25] D. Horstmann et al., Jahresbericht der Deutschen
Mathematiker-Vereinigung 105 (2003).
[26] E. A. Codling, M. J. Plank, and S. Benhamou, J. Roy.
Soc. Interface 5, 813 (2008).
[27] H. G. Othmer, S. R. Dunbar, and W. Alt, J. Math. Biol.
26, 263 (1988).
[28] M. McCutcheon, Physiol. Rev. 26 (1946).
[29] D. House, M. L. Walker, Z. Wu, J. Y. Wong, and
M. Betke, in Computer Vision and Pattern Recognition
Workshops, 2009. CVPR Workshops 2009. IEEE Com-
puter Society Conference on (IEEE, 2009), pp. 186 -- 193.
[30] H. C. Berg, D. A. Brown, et al., Nature 239, 500 (1972).
[31] R. Tranquillo and W. Alt, Biological Motion 89, 510
[32] D. S. Gray, J. Tien, and C. S. Chen, J. Biomed. Mater.
Res. A 66, 605 (2003).
[10] R. J. Pelham and Y.-l. Wang, Proc. Natl. Acad. Sci. USA
(1990).
813 (2014).
94, 13661 (1997).
[11] L. G. Vincent, Y. S. Choi, B. Alonso-Latorre, J. C. del
|
1610.05745 | 1 | 1610 | 2016-10-18T19:16:43 | Protease-sensitive atelocollagen hydrogels promote healing in a diabetic wound model | [
"physics.bio-ph",
"q-bio.TO"
] | The design of exudate-managing wound dressings is an established route to accelerated healing, although such design remains a challenge from material and manufacturing standpoints. Aiming towards the clinical translation of knowledge gained in vitro with highly swollen rat tail collagen hydrogels, this study investigated the healing capability in a diabetic mouse wound model of telopeptide-free, protease-inhibiting collagen networks. 4 vinylbenzylation and UV irradiation of type I atelocollagen (AC) led to hydrogel networks with chemical and macroscopic properties comparable to previous collagen analogues, attributable to similar lysine content and dichroic properties. After 4 days in vitro, hydrogels induced nearly 50 RFU% reduction in matrix metalloproteinase (MMP)-9 activity, whilst showing less than 20 wt.-% weight loss. After 20 days in vivo, dry networks promoted 99% closure of 10x10 mm full thickness wounds and accelerated neodermal tissue formation compared to Mepilex. This collagen system can be equipped with multiple, customisable properties and functions key to personalised chronic wound care. | physics.bio-ph | physics | Protease-sensitive atelocollagen hydrogels promote healing in a diabetic wound model
Giuseppe Tronci,1,2* Jie Yin,1,2 Roisin A. Holmes,2 He Liang,1,2 Stephen J. Russell,1 David J.
Wood2
1 Nonwovens Research Group, School of Design, University of Leeds, Leeds, United
Kingdom
2 Biomaterials and Tissue Engineering Research Group, School of Dentistry, University of
Leeds, Leeds, United Kingdom
Table of contents entry
Protease-sensitive atelocollagen hydrogels were chemically designed to promote accelerated
wound healing in vivo compared to a dressing gold standard.
* Corresponding author: Level 7 Wellcome Trust Brenner Building, St. James's University Hospital, University
of Leeds, Leeds LS9 7TF, UK. E-mail address: [email protected] (G. Tronci)
Abstract
The design of exudate-managing wound dressings is an established route to accelerated
healing, although such design remains a challenge from material and manufacturing
standpoints. Aiming towards the clinical translation of knowledge gained in vitro with
highly-swollen rat tail collagen hydrogels, this study investigated the healing capability in a
diabetic mouse wound model of telopeptide-free, protease-inhibiting collagen networks.
4-vinylbenzylation and UV irradiation of type I atelocollagen (AC) led to hydrogel networks
with chemical and macroscopic properties comparable to previous collagen analogues,
attributable to similar lysine content and dichroic properties. After 4 days in vitro, hydrogels
induced nearly 50 RFU% reduction in matrix metalloproteinase (MMP)-9 activity, whilst
showing less than 20 wt.-% weight loss. After 20 days in vivo, dry networks promoted 99%
closure of 10×10 mm full thickness wounds and accelerated neo-dermal tissue formation
compared to Mepilex®. This collagen system can be equipped with multiple, customisable
properties and functions key to personalised chronic wound care.
Keywords: atelocollagen, hydrogel wound dressings, chronic wounds, MMPs
1. Introduction
Chronic wounds in the form of venous leg ulcers (VLUs), diabetic foot ulcers (DFUs)
and pressure ulcers (PUs) fail to repair in an orderly and timely self-healing process.1,2 With
increasing life expectancy and the associated occurrence of vascular diseases and type II
diabetes, costs associated with chronic wound care represent a significant burden to healthcare
systems worldwide and are expected to continue to rise. In the UK alone, ~ 650,000 patients
are affected by such pathological conditions, resulting in a £3 billion annual cost to the
National Health Service (NHS).3 Additionally, chronic wounds are responsible for prolonged
pain and morbidity in patients.
Dressing materials have been widely employed in the clinic for the treatment of chronic
wounds.4,5 In contrast to skin substitutes,6 wound dressings are temporarily applied to the
wound bed, in order to ensure a defined environment in terms of moisture (to minimise the
risk of tissue maceration) and exudate management (to retain growth factors, MMPs and
specific cells key to healing).7,8 Furthermore, an ideal wound dressing should (i) provide
thermal insulation and oxygen exchange; (ii) protect damaged tissue from secondary
infections and bacterial contamination; (iii) display low adherence in situ to enable complete
dressing removal without debris formation and integration with the host tissue; (iv) control
activity of up-regulated MMPs, such as MMP-9,9,10,11,12,13,14,15,16,17 in order to promote wound
healing; (v) not induce any toxic response to tissue microenvironment. Although these
requirements can be individually provided by many existing commercial dressings, such
controlled multi-functionality is still challenging to accomplish in a single, soluble factor-free
material system. Here, we report a synthetically processed, triple helix preserved collagen
system that fulfils the above requirements and successfully leads to complete wound closure
in diabetic mice. Moreover, the system can be customised to provide bespoke material
architectures, i.e. hydrogels, fibres and fabrics.
As they are based on hydrophilic building blocks, hydrogels have been widely employed
for the design of wound dressings;18 their water content can be tuned in order to ensure
defined levels of exudate in situ,19,20,21 whilst the moist interface with the skin prevents
dressing adherence and allows for easy removal. Additionally, hydrogels can be customised
into fibres and fabrics, whereby the creation of both internal pores and fabric architecture
offers advantages with regard to wound exudate management and material dressability.22,23
Polyurethane24 and methylcellulose25 have been successfully employed for the development
of wound dressing products, i.e. Mepilex® and Aquacel®, respectively. Particularly the
methylcellulose-based materials can become significantly weaker
in
the wet-state,
highlighting the narrow trade-off between water absorbency and hydrated mechanical
properties.26 Also, such materials, based on either synthetic or polysaccharide-based
backbones, do not contain MMP cleavage sites, which are key for the dressing to act as a
substrate for up-regulated proteases in order to stimulate wound healing.
In an effort to introduce MMP sensitivity, polymer networks have recently been
synthesised as active systems to accelerate diabetic wound healing in vivo.27 Protease-
cleavable hyaluronic acid macromers have been electrospun to ensure controlled enzymatic
degradability.28 Polyethylene glycol (PEG) hydrogels have been synthesised to either provide
sustained, enzymatically-responsive release of peptide drugs,29 or provide real-time
quantification of MMP activity30 aiming to establish a relationship to cell viability following
drug treatment.31 Ultimately, Francesko et al. investigated polyphenolic compounds in
multi-component collagen-based sponges as enzymatic inhibitors.32
Rather than synthetically introducing MMP sensitivity in covalent networks, we have
recently reported the creation of inherently enzymatically degradable rat tail collagen
hydrogels with retained triple helix organisation that have higher compressive modulus and
comparable water uptake with respect to commercial dressings, e.g. Aquacel®.33,34,35 Wet
stable fibres have also been produced by sequential wet-spinning of collagen triple helices and
post-spinning covalent crosslinking with an aromatic diacid.36,37,38 In light of these promising
reports, we investigated in this study whether functional systems could be realised by building
molecular architectures defined by the biochemical attributes of the chronic wound
microenvironment in order to achieve accelerated healing in vivo. We hypothesised that type I
telopeptide-free, non-hydrolysed, medical grade AC could be used as highly-purified building
block for the formation of largely hydrated systems with minimal antigenicity.39,40 AC triple
helices were chemically functionalised with 4-vinylbenzyl chloride (4VBC) and integrated in
a covalent network via UV irradiation. 4VBC was employed due to the remarkable swelling
ratio and hydrated compressive modulus previously observed in rat tail collagen materials.41
Introduced 4VBC adducts were also hypothesised to mediate complexation with upregulated
MMPs found in chronic wound exudates. Experiments were carried out in vitro to investigate
the impact of either AC hydrogels or two commercial dressings, i.e. Aquacel® (Convatec) and
Mepilex® (Mölnlycke Health Care), on MMP-9 activity. MMP-9 was selected since
significantly elevated levels in MMP-9 have been reported in chronic compared with acute
wound fluids, suggesting a direct correlation with the clinical wound state.42 Further to the
study in vitro, full thickness wounds were created in db/db diabetic mice as well-accepted
animal models in wound healing research,43,44 and treated with either AC hydrogel or above-
mentioned wound dressing gold standards.
2. Materials and methods
2.1 Materials
An aqueous solution of medical grade type I AC from bovine corium (3 mg∙ml-1 in 10
mM HCl) was kindly provided by Collagen Solutions PLC. Rat tails were provided by the
School of Dentistry, University of Leeds (UK) from which type I collagen was isolated in-
house via acidic treatment of rat tail tendons. Phosphate-buffered saline (PBS, w/o Ca2+ and
Mg2+ ions) was purchased from Lonza. All the other chemicals were purchased from Sigma-
Aldrich.
2.2 Synthesis of atelo- and rat tail collagen hydrogels
Medical grade AC hydrogels were prepared by adopting previously-published
protocols.33,34,41 A solution of type I AC in 10 mM HCl (3 mg∙ml-1) was neutralized to pH 7.4.
1 wt.-% of Tween-20 (with respect to the initial solution volume) and 30 molar excess of both
4-vinylbenzyl chloride (4VBC) and triethylamine (TEA) (with respect to the molar content of
free amino groups of collagen, i.e. ~ 3×10-4 moles∙g-1) were added. After 24 h reaction, the
reaction mixture was precipitated in 10 – 15 volume excess of pure ethanol and stirred
overnight before centrifugation (10000 rpm, 30 min) and air-drying. The retrieved dry product
was characterised via 2,4,6-trinitrobenzenesulfonic acid (TNBS) assay45 and circular
dichroism.33,34 Hydrogels were prepared by dissolving reacted collagen in 10 mM HCl
solution containing 1 wt.-% 2-hydroxy-1-[4-(2-hydroxyethoxy) phenyl]-2-methyl-1-
propanone (I2959) photoinitiator. The suspension was cast in a 12-well plate (1 ml/well) and
UV irradiated (346 nm, Spectroline) for 30 min from above and below. Formed hydrogels
were thoroughly washed in distilled water and used for compression tests. Sample
dehydration was carried out in ethanol-distilled water mixtures with increasing ethanol
content. Rat tail collagen hydrogel controls were prepared in the same manner, except that the
reaction with 4VBC was carried out with a 0.25 wt.-% collagen solution in 10 mM
hydrochloric acid, as previously reported.33,34
2.3 Compression tests
PBS-equilibrated hydrogel discs (Ø: 18 mm; h: 7 mm) were compressed at room
temperature with a compression rate of 3 mm∙min-1 (Instron ElectroPuls E3000). A 250 N
load cell was operated up to complete sample compression.34 Stress-strain curves were
recorded and the compression modulus quantified as the slope of the plot linear region at
25-30% strain. Six replicates were employed for each composition. Data are presented as
mean ± standard deviation (SD).
2.4 Swelling ratio and gel content
Dry atelo- and rat tail collagen networks of known mass (md) were individually incubated
in PBS-containing vials at 25 °C for 24 hours. The swelling ratio (SR) was calculated
according to Equation 1:
(Equation 1)
where ms is the mass of PBS-equilibrated samples.
In addition to the swelling ratio, the gel content was determined as the overall portion of the
covalent hydrogel network insoluble in 10 mM HCl solution.46 Dry, freshly synthesised,
atelo- and rat tail collagen networks of known weight (md) were equilibrated in 10 mM HCl
solution for 24 hour. Resulting hydrogels were air dried and weighed. The gel content (G) was
calculated according to equation 2:
(Equation 2)
where m1 is the dry sample mass following incubation in 10 mM HCl. Five replicates were
employed for both swelling and gel content tests. Data are presented as mean ± SD.
2.5 MMP-9 activity and enzymatic degradation study
Full length human proenzyme MMP-9 (PF038, Merck Millipore,UK) was activated in
fluorescence buffer (50mM Tris-HCL, 10mM CaCl2, 150mM NaCl and 0.05% Brij-35, pH
7.5, Sigma-Aldrich) in the presence of p-aminophenylmercuric acetate (Sigma-Aldrich) at
37°C for 2 hours, according to manufacturer's specifications. Samples of known mass (md:
3-5 mg, n=4) of dry AC networks and gold standard dressings, i.e. Mepilex® (Mölnlycke
Health Care) and Aquacel® (Convatec), were individually placed in a 24-well plate and
incubated with the activated MMP-9 solution (100 ng∙mL-1, 1.5 mL) in an orbital shaker (160
100ddsmmmSR1001dmmG rpm, MaxQ mini 4450) at 37°C. The concentration of MMP-9 was selected according to
MMP-9 levels found in chronic wound fluids.42 Sample-free solution controls of either
activated or non-activated MMP-9 were also included. After 4 days, the MMP-9 activity of
the supernatant was determined via a standard commercial fluorometric assay (ab112146,
Abcam, UK); whilst, 4-day incubated samples were collected, washed in an increasing
distilled water/ethanol series (0, 25, 50, 80, 100 vol.-% ethanol) and air-dried. The mass of
retrieved, air-dried samples (m4) was measured and the relative mass (µrel) calculated
according to Equation 3, as described previously:38
(Equation 3)
Statistical analysis was carried out using OriginPro 8.5.1. Data normality was confirmed via
the Shapiro-Wilk test. Significance of difference was analysed using one-way ANOVA with
Bonferroni test. A p value of less than 0.05 was considered significant. Data are presented as
mean ± SD.
2.6 In vivo study
2.6.1 Animals
Forty-one diabetic, male, 10-week old mice (BKS.Cg-m Dock7m +/+ Leprdb /J, The
Jackson Laboratory, USA) were housed in groups of two animals according to Home Office
regulations and specific requirements for diabetic animals. Prior to experimentation, they were
housed for a period of more than 7 days without disturbance, other than to refresh their
bedding and to replenish their food and water provisions. After experimental wounding,
animals were housed in individual cages (cage size 500 cm2 with sawdust bedding, changed
three times per week) at 23C with 12-hour light/dark cycles. Animals were provided with
food (Standard Rodent Diet) and water ad libitum. All in vivo procedures were carried out in a
1004drelmm Home Office licensed establishment under Home Office Licences (PCD: 50/2505; PPL:
40/3614; PIL: IBCEFDF55; PIL: I34817249).
2.6.2 Creation of full-thickness experimental wounds, treatment and monitoring
Following randomisation into four experimental groups, animals were anaesthetised on
day 0 (isofluorane & air) and the dorsa of animal groups 1-3 shaved and cleaned with saline-
soaked gauze. A single standardised full-thickness wound (10 × 10 mm) was created in the
left dorsal flank skin of each experimental animal. The wound site was hydrated with 25 µl of
sterile saline prior to application of either a commercial dressing or a dry collagen network.
Wounds of animal group 1 (n=11) received gamma-sterilised AC networks placed in such a
manner as to overlay the wound margins. The upper surface of the collagen network was
hydrated with 25 µl of sterile saline following application in situ. Wounds in groups 2 and 3
(n=10) received either a commercial polyurethane (Mepilex®, Mölnlycke Health Care) or
carboxymethyl cellulose dressing (Aquacel®, Convatec), respectively, so that the dressings
overlaid the wound margins by ~5 mm in all directions (Figure S1). All wounds of groups 1-3
also received a transparent polyurethane adhesive film (Bioclusive, Systagenix Wound
Management) in order to secure previously-applied dressings. Control wounds (n=10) were
treated with Bioclusive (Systagenix Wound Management), only. On post-wounding days 4,
8, 12 and 16, all animals were re-anaesthetised and the outer film dressing and any free debris
removed. Wounds were gently cleaned using saline-soaked sterile gauze, blotted dry with
sterile gauze and digitally photographed. Each wound was scored by two independent
observers, as to whether it was displaying initiation of neo-dermal tissue generation activity,44
and the percentage of responsive wounds compared between experimental groups. Following
this, fresh, sterile samples of collagen networks, commercial dressings and secondary
adhesive dressings were then applied. Following all anaesthetic events, animals were placed in
a warm environment and were monitored until they had fully recovered from the surgical
procedure. All animals received appropriate analgesia (buprenorphine) after surgery and
additional analgesics as required.
At day 20, wounds of all animal groups were cleaned, assessed and digitally photographed.
Animals received an intraperitoneal injection (30 µg∙g-1 body weight) of 5-bromo-2′-
deoxyuridine (BrdU, Sigma) one hour prior to termination, in order to facilitate future
detection of proliferating cells by immuno-histochemical analysis of tissue sections. Wounds
and marginal tissues were harvested and processed for histological investigation.
2.6.3 Image analysis of wound closure
Image Pro image analysis software (version 4.1.0.0, Media Cybernetics, USA) was used
to calculate wound closure from wound images over time in each of the experimental groups.
At selected time points, the open area was measured for each wound and wound closure
expressed as percentage ratio between post-wounding open area and initial area.
2.6.4 Harvesting and processing of wound tissues
On post-wounding day 20, all animals were humanely sacrificed by CO2 asphyxiation
(confirmed by cervical dislocation). Wounds and surrounding normal tissue were excised and
fixed in 10% (v/v) neutral buffered formalin (Sigma, UK) for histological assessment.
Excised tissue was sandwiched between two pieces of foam, prior to being placed in fixative,
to reduce the extent of tissue curling. Fixed specimens were trimmed and bisected, generating
two half wounds per site. Both halves were processed and embedded in paraffin wax.
Specimens were orientated in such a fashion as to ensure that appropriate transverse sections
of the wound could be taken.
2.6.5 Histological evaluations
Wax-embedded tissues from post-wounding day 20 were sectioned at 6 μm thickness and
stained with haematoxylin and eosin (H&E). The histological wound width and the
granulation tissue depth were quantitatively assessed using Image Pro image analysis
software. Wound width measurements were taken through the centre of each wound.44
Cranio-caudal contraction was expressed as the percentage of the histological wound width
measured at day 20 with respect to the central wound width calculated digitally at day 0 as an
average of three separate measurements.
2.6.6 Statistical analysis
Data were tested for normality using the Shapiro-Wilk test. Fisher's exact test (for
proportionate data) was carried out on a two-sample basis to analyse the impact of treatment
on initiation of neo-dermal tissue repair activity. Non-parametric analysis (Kruskal-Wallis
followed by ad hoc two-sample Mann Whitney U-test) was used to test the significance of any
inter-group differences in wound closure, contraction and re-epithelialisation as well as
granulation tissue depth and cranio-caudal contraction.
3. Results and discussion
The creation of a single-material AC system is presented, whereby macroscopic properties
and functions can be controlled by structural parameters introduced at the molecular level in
order to fulfil the complex requirements of the chronic wound microenvironment. The system
is built from a defined molecular architecture of AC triple helices that are covalently
functionalised and integrated in a covalent network following photo-activation (Scheme 1).
The AC network swells upon contact with the wound exudate, forming hydrogels with
superior exudate uptake and compression modulus with respect to two wound dressings
routinely employed in the clinic, whilst providing additional regulation of upregulated MMP-
9 activity. In the following, results on AC structural organisation and respective 4VBC-
mediated functionalisation are presented and discussed in relation to the ones of type I rat tail
collagen, as typical source employed for the formation of collagen-based materials.
Subsequently, the attention moves to the characterisation of AC hydrogels in vitro and in vivo,
aiming to explore their potential applicability in chronic wound care. Nomenclature used to
present the results is as follows: type I AC networks are coded as '4VBC*', where '4VBC'
refers to the functionalised precursor synthesised via reaction of AC with 30 molar excess of
4VBC with respect to the primary amino groups and amino termini of collagen, whilst '*'
indicates the resulting covalent network obtained via UV irradiation.
Scheme 1. Research strategy undertaken to realise AC hydrogels with superior wound exudate and MMP-9
management capabilities compared to current leading wound dressings. AC triple helices were functionalised
with 4VBC adducts via lysine-initiated substitution reaction. UV irradiation of AC precursor solutions resulted
in the creation of 4VBC-covalently crosslinked networks of retained AC triple helices. In contact with wound
exudate (WE), networks swell resulting in dressing-forming hydrogels at the macroscopic scale. When applied to
hard-to-heal wounds, hydrogels display high exudate uptake and compression modulus, and control the activity
of exudate-carried upregulated MMPs via complexation with 4VBC adducts.
3.1 Synthesis of type I AC networks
A molar content of primary amino groups of about 3×10-4 mol∙g-1 was measured via TNBS on
pristine type I AC, in agreement with previously-reported values.33-35,37,41 Likewise, far-UV
CD analysis indicated typical dichroic spectral patterns of type I collagen, i.e. a positive triple
helix-related maximum absorption band at 220-225 nm, and a negative minimum absorption
band around 195-200 nm, associated with the presence of polyproline II helices (Figure S2).
Quantitative analysis of the AC CD spectrum resulted in a magnitude ratio between positive
and negative peak intensities (RPN) of 0.13, in line with previous findings.33,34 Reaction of
type I AC with 4VBC was carried out prior to UV irradiation in the presence of I2959
photoinitiator and resulted in the formation of covalent hydrogel networks (G: 97 ± 4 wt.-%,
Table 1). 4VBC-functionalised precursors displayed 45 mol.-% averaged functionalisation of
primary amino groups, whilst Far-UV CD confirmed retention of triple helices (RPN: 0.14,
Table 1). PBS-equilibrated hydrogels exhibited a swelling ratio of about 2000 wt.-% and a
compressive modulus of about 80 kPa, confirming comparable macroscopic properties with
respect to rat tail collagen-based analogues.
Table 1. Chemical, structural and hydrogel properties of 4VBC-functionalised AC networks. (F): Degree of
functionalisation, (RPN): CD magnitude ratio between positive and negative peak intensities, (SR): swelling ratio
following equilibration with PBS, (G): gel content, (Ec): compressive modulus. Data is presented as mean ± SD.
Sample ID
F /mol.-%
RPN
G /wt.-%
SR /wt.-%
4VBC*
45 ± 4
44 ± 1(a)
0.14
0.12(a)
97 ± 4
100 ± 0(a)
2065 ± 191
1900 ± 200(a)
Ec /kPa
84 ± 17
81 ± 9(a)
(a) Rat tail collagen control.
These values proved to be higher than the ones displayed by two gold wound dressing
standards, i.e. Aquacel® (SR: 1759 ± 107 wt.-%; Ec: 34 ± 18 kPa) and Mepilex® (SR: 447 ± 53
wt.-%).
3.2 Protease sensitivity and degradability in vitro
Samples of 4VBC*, Mepilex® and Aquacel® were incubated in a buffer solution containing
activated MMP-9. Figure 1 displays the MMP-9 activity of sample-treated supernatants
relative to the one of sample-free supernatants (containing activated MMP-9), as well as the
relative mass (µrel) of collected samples, after 4 days (as the typical application time of a
chronic wound dressing). The MMP-9 activity was found to be decreased to 51±5 RFU%
when the respective solution was in contact with samples of 4VBC*. In contrast, a
non-statistically significant variation in MMP-9 activity was observed following application
of either Mepilex® or Aquacel®.
Figure 1. (A) MMP-9 activity measured in the supernatant following 4-day incubation with either no sample
(MMP-9) or sample of 4VBC*, Mepilex® or Aquacel®. Data are expressed as relative percentages with respect to
the activity of untreated MMP-9 solution. (B): Relative sample mass (rel) of either 4VBC*, Mepilex® or
Aquacel® following 4-day incubation in an aqueous solution with or without activated MMP-9. *: p <0.05; **:
p < 0.01; ***: p < 0.001.
In the presence of MMP-9, samples of 4VBC* displayed more than 80 wt.-% averaged
relative mass (µrel: 83±2 wt.-%), a value found to be between the one measured in samples of
Mepilex® (µrel: 97±2 wt.-%) and Aquacel® (µrel: 66±13 wt.-%). On the other hand, control
experiments in MMP-9–free media revealed minimal weight loss in the case of 4VBC* (µrel:
97±1 wt.-%) and Mepilex® (µrel: 99±1 wt.-%, data not shown), whilst similar trends were
observed in the case of Aquacel® (µrel: 70±4 wt.-%, data not shown).
3.3 Macroscopic wound assessment
A schematic overview of the overall strategy pursued to monitor the healing process is shown
in Figure 2. Macroscopic closure was expressed as the ratio between the post-wounding open
and original areas, whereby wound dimensions were quantified digitally (Figure 2, A and B).
The extent of granulation tissue formation, re-epithelialisation and contraction to wound
closure was determined either macroscopically or histologically. Wound contracted and
re-epithelialised areas were identified from time-specific digital macrographs (Figure 2, B),
whilst granulation tissue depth and wound width were quantified via histological H&E tissue
sections at day 20 post-wounding (Figure 2, C). All experimental groups displayed nearly-
complete closure of open wound areas at the end of the study in vivo (Figure 3, I-K), in
contrast to the control group showing minimal dimensional changes (Figure 3, L).
Among the three dressing groups, a major reduction of open wound area was observed after 8
days in animals treated with both 4VBC* and Aquacel®. All mice were found to tolerate the
application of the AC hydrogel and the two commercial dressings, although some erythema
was noted in the skin surrounding the wounds (4VBC*: n=3, Mepilex®: n=4, Aquacel®: n=2).
This may be attributed to the repeated application and removal of the adherent film
BioclusiveTM. Other than that, hydrogel 4VBC* remained intact on the wound surface during
the 4-d application time window and could be easily removed. In contrast, samples of
Aquacel® were found in different locations on day 4 with respect to the underlying wound
position.
Figure 2. Strategy adopted to monitor the wound
healing process over a 20-d in vivo study. (A-B): the
wound area was quantified via equally-scaled digital
macrographs taken at specific time points. Wounds
were optically assessed in terms of neo-dermal tissue
generation, contraction and re-epithelialisation. (C):
Histology sections were obtained at the end of the
study so that wound width and granulation tissue
depth were quantified.
3.4 Temporal profiles of wound closure, re-epithelialisation and contraction
Post-wounding digital macrographs were analysed to determine (i) the percentage of
wounds showing initiation of neo-tissue formation and (ii) the temporal profiles of wound
closure, contraction and re-epithelialisation (Figure 4). In comparison to the adhesive-treated
control group, significant improvements (p≤0.011) in both neo-dermal tissue generation
activity and closure were observed in dressing-treated wounds at all assessment time points
(Figure 4, A and B) irrespective of the dressing type. More than 80% response was identified
at day 4 in wounds in receipt of AC hydrogel, a percentage found to be between the one
recorded in Aquacel®- (100%) and Mepilex®- (60%) treated experimental groups (Figure 4,
A); whilst no responding wound was found at this assessment point in the adhesive-treated
control group. By day 8, neo-dermal tissue formation occurred in all wounds in receipt of both
4VBC* and Aquacel® dressings, whilst application of Mepilex® dressing promoted neo-tissue
stimulation in 80% wounds, in stark contrast with the control group (10%). Likewise,
temporal profiles of wound closure (Figure 4, B) indicated a significantly reduced open area
in either 4VBC*- and Aquacel®-, compared to Mepilex®-, treated wounds (p=0.024). From
day 12 onwards, all experimental groups confirmed initiation of tissue formation, whilst this
was still minimal in the control group. Significantly increased wound closure was observed
following application of Aquacel®, compared to Mepilex®, dressing on day 12 (p = 0.043),
whilst comparable results were observed when sample 4VBC* was employed (Figure 4, B).
Other than neo-tissue initiation and wound closure, re-epithelialisation and contraction
temporal profiles (Figure 4, C and D) showed that Mepilex® dressing induced the highest
level of re epithelialisation (day 20: 34 ± 15 %) at all selected post wounding time points,
followed by 4VBC* hydrogel (day 20: 27 ± 10 %) and Aquacel® dressing (day 20: 22 ± 5 %),
whilst the opposite trend became apparent with regard to the contraction profile (day 20:
4VBC*: 70 ± 9 %; Mepilex®: 66 ± 14 %; Aquacel®: 78 ± 5 %).
Figure 3. Digital macrographs of the wound site (10 × 10 mm) at day 0 (A-D), 8 (E-H) and 20 (I-L), following
application of dry AC hydrogel (4VBC* (A, E, I)), commercial polyurethane foam dressings (Mepilex® (B, F,
J)), commercial cellulose-based dressings (Aquacel® (C, G, K)), and commercial adhesive controls
(BioclusiveTM (D, H, L)). The scale bar is 5 mm.
Levels of re-epithelisation were found to increase up to day 12 post-wounding, after which
they tended to plateau or reduce due to the on-going contraction of the wounds. From day 16
onwards, wounds in receipt of AC hydrogel showed significantly greater re-epithelialisation
(p ≤ 0.043) with respect to wounds in receipt of Aquacel® dressing; and a similar situation
was apparent when comparing groups treated with Mepilex® and Aquacel® dressings from
day 12 onwards (p ≤ 0.029). The latter dressing promoted significantly higher wound
contraction with respect to 4VBC* hydrogel from day 12 onwards (p ≤ 0.024), and similar
results were obtained when comparing the former with the Mepilex® dressing group in the
same time window (p ≤ 0.011).
Figure 4. (A): Percentage of responding wounds showing initiation of neo-dermal tissue formation at selected
post-wounding time points following application of commercial adhesive controls (Bioclusive®, white columns),
AC hydrogels (4VBC*, grey columns), commercial polyurethane foam dressings (Mepilex®, black columns), and
commercial cellulose-based dressings (Aquacel®, patterned columns). **: p < 0.01. At day 4, no responding
wound was found in the control group, so that no column is reported at this assessment point. (B-D): Temporal
profiles of remaining wound area (B), wound re-epithelialisation (C) and wound contraction (D) determined
from post-wounding digital macrographs upon receipt of AC hydrogels (4VBC*, –■–), commercial polyurethane
foam dressings (Mepilex®, ∙∙─●─∙∙), commercial cellulose-based dressings (Aquacel®, ∙∙─○∙∙─) and commercial
adhesive controls (BioclusiveTM, ∙∙─▼∙∙─). Curves are guidelines to the eye. Data are presented as the mean ±
SD.
3.5 Histological analysis of cranio-caudal contraction and granulation tissue depth
Typical low power wound histology images for all treatment groups at day 20 post-
wounding are shown in Figure 5. Granulation tissue depth (GTD) was significantly increased
in wounds with either 4VBC* (Figure 5, A), Mepilex® (Figure 5, B) or Aquacel® (Figure 5, C)
dressing, compared to control wounds (Figure 5, D), confirming previous trends obtained at
the macro-scale. Comparable GTD levels were found upon application of AC and cellulose-
based materials, with marginally reduced levels in wounds treated with Mepilex® dressing
(Figure 6).
Figure 5. Low power images of typical histological
sections related to 20-d wounds treated with either AC
hydrogels (4VBC*, A), commercial polyurethane
foam dressings (Mepilex®, B), commercial cellulose-
(Aquacel®, C) or commercial
based dressings
adhesive controls (BioclusiveTM, D). The scale bar is
500 µm.
Further to GTD, cranio-caudal contraction (CCC) was quantified in order to express the
reduction in histology-based central wound width at the time of tissue harvesting (Figure 2, C)
relative to that at the time of injury. CCC values (Figure 6) proved to be comparable to
macroscopic values of wound contraction measured at day 20 (Figure 4, C). Wounds in
receipt of either 4VBC*, Mepilex® or Aquacel® dressings exhibited significantly increased
CCC with respect to control wounds (p = 0.000). Samples of Aquacel® promoted the greatest
levels of CCC, followed by the AC hydrogel and ultimately by samples of Mepilex®, although
no significant differences were observed.
Figure 6. Histology-based cranio-caudal contraction (CCC, black columns, left-handed y axis) and granulation
tissue depth (GTD, grey columns, right-handed y axis). CCC was calculated as the ratio between the wound
width measured on 20-d H&E histology sections and the wound width measured on day 0 digital macrographs.
'*' and '**' (p = 0.000) indicate means significantly different from the means of the other groups (non-parametric
Mann Whitney U-test).
3.6 Discussion
Several natural compounds have been proposed for the treatment of hard-to-heal wounds,
such as Manuka honey,23 bee venom,47 propolis48 and whey protein,49,50 in light of their
antibacterial and antioxidant activity, although their potential effect in wound healing is still
matter of current research. Aiming to develop intelligent wound dressings capable to
accelerate the healing of either delayed or impaired wounds, collagen has recently gained a
great deal of attention. Several collagen-containing systems have been approved for clinical
use, including Promogran® (Systagenix),51 a lyophilized collagen-oxidised regenerated
cellulose matrix that gels on contact with wound exudate, and BiostepTM (Smith & Nephew,
Inc),52 a blend consisting of collagen, sodium alginate, carboxyl methylcellulose, and
ethylenediaminetetraacetic acid (EDTA). In hydrated conditions, these dressings present
documented form-instability in biological fluids,53 in comparison with non-collagen-based
standards of care, e.g. Aquacel® and Mepilex® (both used in this study), resulting in handling
difficulties, debris formation and bacterial contamination in vivo. Such poor behaviour is
partially explained by the fact that collagen has been reported to lose its native triple helix
architecture when blended with other polymers (potentially requiring the use of organic
solvents) and processed with state of the art manufacturing methods,54 resulting in
hydrolysed, denatured derivatives.37,55 In light of these challenges, alternative biomaterials
have therefore been proposed and explored either in vitro or pre-clinically for the treatment of
burn injuries and impaired wounds, including copper-doped glass microfibres inducing
angiogenesis and wound closure,56 nanofibrous peptidic hydrogels,57 protease-cleavable
polymer systems,28-31 as well as hydrogels containing protease-inhibiting soluble factors13,32
or cells,58 often requiring considerations with regard to activity and release of payload as well
as customisation of material format.
The use of individual, ECM-derived biopolymers is a promising strategy for chronic
wound care aiming to create wound-customised systems capable to inherently regulate MMP
activity and manage wound exudate via structural parameters introduced at the molecular
level. Here, the selection of purified, epitope-free sources59 and the development of multiscale
approaches60 enabling bespoke structure-property-function relationships and material formats
are key tasks to enable successful translation to clinical use. The results of this study provide a
systematic and comprehensive investigation on the design of a polymer-, soluble factor- and
cell-free AC system with controlled MMP-9 sensitivity and enzymatic degradability,
displaying retained triple helices, and capable of promoting nearly 99% closure in diabetic
wound models, comparable to established gold standards of care.
To investigate wound healing capability, medical grade non-hydrolysed AC was selected
as a pristine backbone and compared to rat tail collagen. The influence of the collagen source
on the physical and structural properties of resulting materials is known59 and was therefore
addressed by characterising both native as well as functionalised and crosslinked states.
Especially for applications in vivo, the selection of antigen-free sources is a key to the design
of medical collagen products, whilst
the presence of non-collagenous, potentially
immunogenic impurities, e.g. deriving from the material synthesis, should also be considered.
Antigenic determinants of collagen can be found in the (triple) helical regions, with variations
in the amino acid sequences not exceeding more than a few percent between mammalian
species.39,40 A far greater degree of variability is found in the non-helical terminal regions, i.e.
telo-peptides, with up to half of the amino acid residues in these regions exhibiting
interspecies variation.61 Although a systematic immune-toxicity evaluation was not the focus
of this study, two strategies were adapted to ensure the formation of collagen materials with
minimal antigenicity and immunogenicity: (1) selection of a telopeptide-free collagen source
compatible with the synthesis of 4VBC-functionalised collagen networks, and (2) intensive
material washing following each synthetic step aiming to accomplish complete removal of
non-collagenous impurities. Selected type I AC was obtained via pepsin-mediated extraction
of bovine corium, resulting in a highly purified telopeptide-free backbone with similar lysine
content and dichroic properties to our previously-used in-house extracted collagen (Table 1).
Telopeptide-free hydrogels were successfully synthesised with enhanced swelling ratio,
compressive modulus (Table 1) and enzymatic degradability (Figure 1), compared to both
Aquacel® and Mepilex®. Samples proved to display less than 20 wt.-% mass loss in both
enzymatic and hydrolytic conditions, whilst still inducing nearly 50 RFU% reduction in
MMP-9 activity in vitro, in contrast to both commercial dressings. The synthesis of the
covalent network was a key to provide AC hydrogels with statistically decreased
degradability, with respect to the case of Aquacel® (as based on non-crosslinked,
water-soluble carboxymethyl cellulose). Covalently-bound electron-rich aromatic adducts in
hydrogels 4VBC* were expected to be responsible for the observed inhibition of MMP-9
activity via chelation with zinc sites of active MMPs, minimising the need of soluble,
protease-chelating factors, e.g. EDTA, in the collagen system. The combination of these
swelling, degradation and compression properties together with the preserved triple helix
organisation at the molecular level proved to be crucial in order to promote accelerated wound
healing in the hydrogel- with respect to Mepilex®-treated wounds; after 20 days
post-wounding and in contrast to control wounds, similar levels of wound contraction, re-
epithelialisation (Figure 4), cranio-caudal contraction and granulation tissue depth (Figure 6)
were observed in wounds in receipt of 4VBC* with respect to wounds treated with both
dressings. The high hydrogel swellability in physiological conditions promoted sequestration
and deactivation of up-regulated proteases, so that proteases could be diverted from the neo-
tissue to the dressing and healing accelerated. In line with gravimetric enzymatic degradation
data, AC dressings exhibited minimal macroscopic change following a 4-day application in
vivo, so that they could be easily located at and removed from the wound bed with no sample
break. After 20 days post-wounding, telopeptide-free hydrogels were found to induce both
wound contraction and re-epithelialisation. These events are essential for successful skin
wound healing, in order to quickly re-establish barrier function by migration of keratinocytes
in the direction of the injury and minimise risks of environment-triggered insults.62 Wound
and histological analyses (Figure 2 and 4) confirmed that the regeneration of the epidermis
was almost complete for hydrogel-treated wounds, suggesting the presence of healthy
epithelial cells on the top of the injured tissue, whilst new epidermis appeared as thick as that
of healthy skin wounds.
4. Conclusions
The present study provided new information on the capacity of a functionalised collagen
network to stimulate healing of full-thickness wounds in diabetic mice, avoiding the use of
soluble factors, encapsulated cells or multi-compartment systems. A medical grade, AC
source was identified as telopeptide-free analogue of in-house extracted rat tail collagen,
enabling the formation of purified hydrogels with accelerated neo-dermal tissue initiation
capability in comparison to a clinical polyurethane dressing, i.e. Mepilex®. The retention of
native triple helices in and high swellability of collagen networks enabled the development of
a hydrated wound bed in situ, whereby controlled enzymatic degradability and inherent
control in MMP activity could be accomplished. When used to treat full thickness wounds in
diabetic mice, hydrogels showed a significantly better capacity to stimulate wound closure
and re-epithelialisation than the untreated wounds (control), whilst similar healing levels were
observed when either Aquacel® or Mepilex® gold-standard dressings were applied. The
versatility of this collagen system is expected to promote the design of advanced, protease-
degradable wound dressings with systematically adjusted molecular, microscopic and
macroscopic architecture to enable superior wound exudate management capability and
healing performance in vivo.
Acknowledgments
The authors thank the Clothworkers Centre for Textile Materials Innovation for Healthcare,
EPSRC Centre for Innovative Manufacturing in Medical Devices (MeDe Innovation), EPSRC
MeDe Innovation Fresh Ideas Fund and the University of Leeds Medical Technologies IKC
for financial support. Dr. Jeff Hart and Andrea Bell (Cica Biomedical Ltd., UK) are gratefully
acknowledged for carrying out the in vivo work. The authors thank Dr Graeme Howling and
Martin Fuller for their support with the study in vivo and TEM, respectively.
Supporting information
Figure S1. (A-D): digital macrographs taken during in vivo surgery. (A): The mouse dorsum was shaved and
incised to create a 10×10 mm wound (A). (B-D): AC hydrogel (B) and two commercial dressings, i.e. Aquacel®
(C) and Mepilex® (D), were applied on the top of the wound. Samples were replaced every four days over a 20-d
time window.
Figure S2. Far-UV CD spectrum of AC solution in 10 mM HCl. 195-200 and 220-225 nm peaks are clearly
detected, confirming the presence of collagen polyproline II and triple helices, respectively.
References
1 W.A Sarhan, H.M.E. Azzazy, I.M. El-Sherbiny, ACS Appl. Mater. Interfaces, 2016, 8,
6379-6390.
2 F. Werdinemail, M. Tenenhaus, H.-O. Rennekampff, The Lancet, 2008, 372, 1860-1862.
3 A. McLister, J. McHugh, J. Cundell, J. Davis, Adv Mater 2016, 28, 5732-5737.
4 K. Vowden, P. Vowden, Surgery, 2014, 32, 462-467.
5 J.L. Schiefer, R. Rath, M. Held, W. Petersen, J.O. Werner, H.E. Schaller, A. Rahmanian-
Schwarz, Adv Skin Wound Care, 2016, 29, 73-78.
6 H. Lagus, M. Sarlomo-Rikala, T. Böhling, J. Vuola, Burns, 2013, 39, 1577-1587.
7 J.S. Boateng, K.H. Matthews, H.N.E. Stevens, G.M. Eccleston, J Pharm Sci, 2008, 97,
2892-2923.
8 J.-F. Jhong, A. Venault, L. Liu, J. Zheng, S.-H. Chen, A. Higuchi, J. Huang, Y. Chang, ACS
Appl. Mater. Interfaces, 2014, 6, 9858-9870.
9 C.G. Decker, Y. Wang, S.J. Paluck, L. Shen, J.A. Loo, A.J. Levine, L.S. Miller, H.D.
Maynard, Biomaterials, 2016, 81, 157-168.
10 M. Kulkarni, A. O'Loughlin, R. Vazquez, K. Mashayekhi, P. Rooney, U. Greiser, E.
O'Toole, T. O'Brien, M.M. Malagon, A. Pandit, Biomaterials, 2014, 35, 2001-2010.
11 L.I.F. Moura, A.M.A. Dias, E. Carvalho, H.C. De Sousa, Acta Biomaterialia, 2013, 9,
7093-7114.
12 S.A. Castleberry, B.D. Almquist, W. Li, T. Reis, J. Chow, S. Mayner, P.T. Hammond, Adv
Mater, 2016, 28, 1809-1817.
13 M. Gao, T.T. Nguyen, M.A. Suckow, W.R. Wolter, M. Gooyit, S. Mobashery, M. Chang,
Proc Natl Acad Sci USA, 2015, 112, 15226-15231.
14 N.J. Trengove, M.C. Stacey, S. Macauley, N. Bennett, J. Gibson, F. Burslem, G. Murphy,
G. Schultz, Wound Repair Regen, 1999, 7, 442-52.
15 S. Eming, H. Smola, B. Hartmann, G. Malchau, R. Wegner, T. Krieg, S. Smola-Hess,
Biomaterials, 2008, 29, 2932-2940.
16 B.P. Purcell, D. Lobb, M.B. Charati, S.M. Dorsey, R.J. Wade, K.N. Zellars, H. Doviak, S.
Pettaway, C.B. Logdon, J.A. Shuman, P.D. Freels, J.H. Gorman, R.C. Gorman, F.G. Spinale,
J.A. Burdick, Nature Materials, 2014, 13, 653-661.
17 S.S. Anumolu, A.R. Menjoge, M. Deshmukh, D. Gerecke, S. Stein, J. Laskin, P.J. Sinko,
Biomaterials, 2011, 32, 1204-1217.
18 S. Sakai, M. Tsumura, M. Inoue, Y. Koga, K. Fukano, M. Taya, J. Mater. Chem. B, 2013,
1, 5067-5075.
19 G. Tronci, H. Ajiro, S.J. Russell, D.J. Wood, M. Akashi, Acta Biomater 2014, 10, 821-830.
20 Z. Fan, B. Liu, J. Wang, S. Zhang, Q. Lin, P. Gong, L. Ma, S. Yang, Adv. Funct. Mater.,
2014, 24, 3933-3943.
21 M. Panca, K. Cutting, J.F. Guest, J Wound Care, 2013, 22, 109-118.
22 A. Yaari, Y. Schilt, C. Tamburu, U. Raviv, O. Shoseyov, ACS Biomater Sci Eng, 2016, 2,
349-360.
23 S.E.L. Bulman, P. Goswami, G. Tronci, S.J. Russell, C. Carr, J Biomater Appl, 2015, 29,
1193-1200.
24 H.J. Yoo, H.D. Kim, J Biomed Mater Res Part B: Appl Biomater, 2008, 85B, 326-333.
25 M.J. Waring, D. Parsons, Biomaterials, 2001, 22, 903-912.
26 S.M. Bishop, M. Walker, A.A. Rogers, W.Y.J. Chen, J Wound Care, 2003, 12, 125-128.
27 E.A. Rayment, T.R. Dargaville, G.K. Shooter, G.A. George, Z. Upton, Biomaterials, 2008,
29, 1785-1795.
28 R.J. Wade, E.J. Bassin, C.B. Rodell, J.A. Burdick, Nature Communications, 2015, 6, 6639.
29 A.H. Van Hove, M.-J.G. Beltejar, D.S.W. Benoit, Biomaterials, 2014, 35, 9719-9730.
30 J.L. Leight, D.L. Alge, A.J. Maier, K.S. Anseth, Biomaterials, 2013, 34, 7344-7352.
31 J.L. Leight, E.J. Tokuda, C.E. Jones, A.J. Lina, K.S. Anseth, Proc Natl Acad Sci USA,
2015, 112, 5366-5371.
32 A. Francesko, D.S. Da Costa, R.L. Reis, I. Pashkuleva, T. Tzanov, Acta Biomaterialia,
2013, 9, 5216-5225.
33 G. Tronci, S.J. Russell, D.J. Wood, J Mater Chem B, 2013, 1, 3705-3715.
34 G. Tronci, C.A. Grant, N.H. Thomson, S.J. Russell, D.J. Wood, J. R. Soc. Interface, 2015,
12, 20141079.
35 G. Tronci, A. Doyle, S.J. Russell, D.J. Wood, D.J. Mater. Res. Soc. Symp. Proc., 2013,
1498, 145-150.
36 M.T. Arafat, G. Tronci, J. Yin, D.J. Wood, S.J. Russell, Polymer, 2015, 77, 102-112.
37 G. Tronci, R.S. Kanuparti, M.T. Arafat, J. Yin, D.J. Wood, S.J. Russell, Int J Biol
Macromol., 2015, 81, 112-120.
38 G. Tronci, A. Doyle, S.J. Russell, D.J. Wood, J. Mater. Chem. B, 2013, 1, 5478-5488.
39 A.K. Lynn, I.V. Yannas, W. Bonfield, J Biomed Mater Res Part B: Appl Biomater, 2004,
71B, 343-354.
40 J. Glowacki, S. Mizuno, Biopolymers, 2008, 89, 338-344.
41 G. Tronci, C.A. Grant, N.H. Thomson, S.J. Russell, D.J. Wood, MRS Advances, 2016, 1,
533-538.
42 E.A. Rayment, Z. Upton, G.K. Shooter, British Journal of Dermatology, 2008, 158, 951-
961.
43 U. Freudenberg, A. Zieris, K. Chwalek, M.V. Tsurkan, M.F. Maitz, P. Atallah, K.R.
Levental, S.A. Eming, C. Werner, J Control Rel, 2015, 220(A), 79-88.
44 J.T. Hardwicke, J. Hart, A. Bell, R. Duncan, D.W. Thomas, R. Moseley, J Control Rel,
2011, 152, 411-417.
45 W.A. Bubnis, C.M. Ofner, Anal Biochem., 1992, 207, 129-133.
46 R. Jin, L.S.M. Teixeira, P.J. Dijkstra, M. Karperien, C.A. van Blitterswijk, Z.Y. Zhong, J.
Feijen, Biomaterials, 2009, 30, 2544-2551.
47 G. Badr, W.N. Hozzein, B.M. Badr, A.A. Ghamdi, H.M.S. Eldien, O. Garraud, J Cell
Physiol, 2016, 231, 2159-2171.
48 W.N. Hozzein, G. Badr, A.A.A. Ghamdi, A. Sayed, N.S. Al-Waili, O. Garraud, Cell
Physiol Biochem, 2015, 37, 940-954.
49 G. Badr, Cell Physiol Biochem, 2012, 29, 571-582.
50 G. Badr, Lipids in Health and Disease, 2013, 12, 46.
51 F. Vin, L. Teot, S. Meaume, J Wound Care, 2002, 11, 335-341.
52 A. Landsman, D. Taft, K. Riemer, Clinics in Podiatric Medicine and Surgery, 2009, 26,
525-533.
53 G. Tronci, A.T. Neffe, B.F. Pierce, A. Lendlein, J Mater Chem, 2010, 20, 8875-8884.
54 E.R. Durham, G. Tronci, X.B. Yang, D.J. Wood, S.J. Russell, in Biomedical Textiles for
Orthopaedic and Surgical Applications, ed. T. Blair, Elsevier; 1st Edition, 2015, 3, 45-65.
55 X. Qiao, S.J. Russell, X. Yang, G. Tronci, D.J. Wood, J Funct Biomater, 2015, 6, 667-686.
56 S. Zhao, H. Wang, Y. Zhang, X. Cheng, N. Zhou, M.N. Rahaman, Z. Liu, W. Huang, C.
Zhang, Biomaterials, 2015, 53, 379-391.
57 Y. Loo, Y.-C. Wong, E.Z. Cai, C.-H. Ang, A. Raju, A. Lakshmanan, A.G. Koh, H.J. Zhou,
T.C. Lim, S.M. Moochhala, C.A. Hauser, C.A. Biomaterials, 2014, 35, 4805-4814.
58 Y. Dong, W.U. Hassan, R. Kennedy, U. Greiser, A. Pandit, Y. Garcia, W. Wang, Acta
Biomaterialia, 2014, 10, 2076-2085.
59 S. Majumdar, Q. Guo, M. Garza-Madrid, X. Calderon-Colon, D. Duan, P. Carbajal, O.
Schein, M. Trexler, J. Elisseeff, J Biomed Mater Res Part B: Appl Biomater, 2016, 104B,
300-307.
60 P.R. Stoessel, R.N. Grass, A. Sánchez-Ferrer, R. Fuhrer, T. Schweizer, R. Mezzenga, W.J.
Stark, Adv Funct Mater, 2014, 24, 1831-1839.
61 C. Soo, C. Rahbar, R.L. Moy, J Dermatol Surg Oncol., 1993, 19, 431-434.
62 K. Safferling, T. Sütterlin, K. Westphal, C. Ernst, K. Breuhahn, M. James, D. Jäger, N.
Halama, N. Grabe, The Journal of Cell Biology, 2013, 203, 691-709.
|
1706.06446 | 2 | 1706 | 2017-09-17T06:43:51 | The Concepts and Applications of Fractional Order Differential Calculus in Modelling of Viscoelastic Systems: A primer | [
"physics.bio-ph",
"cond-mat.soft"
] | Viscoelasticity and related phenomena are of great importance in the study of mechanical properties of material especially, biological materials. Certain materials show some complex effects in mechanical tests, which cannot be described by standard linear equation (SLE) mostly owing to shape memory effect during deformation. Recently, researchers have been applying fractional calculus in order for probing viscoelasticity of such materials with a high precision. Fractional calculus is a powerful tool for modeling complex phenomenon. In this tutorial based paper, we try present clear descriptions of the fractional calculus, its techniques and its implementation. The intention is to keep the details to a minimum while still conveying a good idea of what and how can be done with this powerful tool. We try to expose the reader to the basic techniques that are used to solve the fractional equations analytically and/or numerically. More specifically, modeling the shape memory phenomena with this powerful tool are studied from different perspectives, as well as presented some physical interpretation in this case. Moreover, in order to show the relationship between fractional models and standard linear equations, a fractal system comprising spring and damper elements is considered, and the constitutive equation is approximated with a fractional element. Finally, after a brief literature review, two fractional models are utilized to investigate the viscoelasticity of the cell, and the comparison is made among them, experimental data, and previous models. Verification results indicate that not only does the fractional model match the experimental data well, but it also can be a good substitute for previously used models. | physics.bio-ph | physics | The Concepts and Applications of Fractional Order Differential Calculus
in Modelling of Viscoelastic Systems: A primer
2 Computational physical Sciences Research Laboratory, School of Nano-Science, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran
1 Biomathematics Laboratory, Department of Applied Mathematics, Tarbiat Modares University, Iran
Mohammad Amirian Matlob
1
, Yousef Jamali
*1,2
Viscoelasticity and related phenomena are of great importance in the study of mechanical properties of material especially, biological
materials. Certain materials show some complex effects in mechanical tests, which cannot be described by standard linear equation
(SLE) mostly owing to shape memory effect during deformation. Recently, researchers have been applying fractional calculus in order
for probing viscoelasticity of such materials with a high precision. Fractional calculus is a powerful tool for modeling complex
phenomenon. In this tutorial based paper, we try present clear descriptions of the fractional calculus, its techniques and its
implementation. The intention is to keep the details to a minimum while still conveying a good idea of what and how can be done with
this powerful tool. We try to expose the reader to the basic techniques that are used to solve the fractional equations analytically
and/or numerically. More specifically, modeling the shape memory phenomena with this powerful tool are studied from different
perspectives, as well as presented some physical interpretation in this case. Moreover, in order to show the relationship between
fractional models and standard linear equations, a fractal system comprising spring and damper elements is considered, and
the constitutive equation is approximated with a fractional element. Finally, after a brief literature review, two fractional models are
utilized to investigate the viscoelasticity of the cell, and the comparison is made among them, experimental data, and previous models.
Verification results indicate that not only does the fractional model match the experimental data well, but it also can be a good
substitute for previously used models.
Keywords: Viscoelasticity, Fractional calculus, Mechanical properties, Cell biomechanics, Fractional modeling, Fractal system.
Author Summary
Fractional Calculus is a new powerful tool which has been recently employed to model complex biological systems with non-linear
behavior and long-term memory .In spite of its complicated mathematical background, fractional calculus came into being of some
simple questions which were related to the derivation concept; such questions as while the first order derivative represents the slope
of a function, what a half order derivative of a function reveal about it? Finding answers to such questions, scientists managed to open
a new window of opportunity to mathematical and real world, which has arisen many new questions and intriguing results. For
example, the fractional order derivative of a constant function, unlike the ordinary derivative, is not always zero. In this tutorial-
based paper it is sought to answer the aforementioned questions and to construct a comprehensive picture of what fractional calculus
is, and how it can be utilized for modelisation purpose. The focus of the research has been on viscoelastic materials. After an extensive
literature review of the concepts and application of this potent tool, a novel application of this tool is developed for simulating a
dynamic system in order to investigate the mechanical behavior of a cell.
*Corresponding author. Tel: +98-21-8288-4762
E-mail addresses: [email protected] (Y.Jamali), [email protected] (M.Amirian Matlob).
1
Contents
Introduction ............................................................................................................................................ 3
Methods .................................................................................................................................................. 3
Fractional calculus ................................................................................................................................... 3
Definition of fractional calculus ......................................................................................................... 6
The interpretation of fractional calculus ................................................................................................ 9
Laplacean interpretation ...................................................................................................................... 9
Analytical and numerical methods ....................................................................................................... 10
Analytical methods ........................................................................................................................... 10
Numerical solution ............................................................................................................................ 13
Viscoelastic systems .............................................................................................................................. 14
Creep test .......................................................................................................................................... 15
Periodic test ....................................................................................................................................... 16
Fractional viscoelastic model ................................................................................................................ 17
Periodic strain ................................................................................................................................... 21
Relation between fractional equations and SLE ................................................................................... 21
Application ............................................................................................................................................ 23
Two-dimension fractional network model ........................................................................................... 24
Discussion ............................................................................................................................................. 27
Appendix ............................................................................................................................................... 28
Acknowledgment .................................................................................................................................. 33
References ............................................................................................................................................ 34
2
Introduction
The concept of derivative is the main idea of calculus. It shows the sensitivity to change of a function
i.e. the rate or slope of a quantity. The current definition of derivative was suggested by Newton in
1666. Newton with a physical viewpoint of derivative interpreted the instantaneous velocity [1]. The
intuition of researchers from derivative and integral is based on their geometrical or physical meaning,
e.g. the first and second order derivative of displacement is called velocity and acceleration respectively
(also jerk and jounce for 3th and 4th derivatives). As well, integral of a curve function means the area
under the curve. This form of classical calculus was developed extensively over four centuries. Today,
scientists are able to describe and model many physical phenomena with an ordinary differential
equation [2]. In many cases, however, the classical calculus is not able to describe exactly these complex
phenomena. Under a small deformation, for example, the relationship between the force and
displacement in an ideal spring (small deformations in elastic materials) is linear, i.e. force is related to
the zero derivative of displacement, while in an ideal damper, the force is proportional to the velocity
of extension or compression. In other words, force is related to the first derivative of deformation.
Which law, however, governs the materials with intermediate mechanical properties (i.e. between ideal
spring and ideal damper)? With this aim, many researchers try to answer the question and to model as
well as to analyse the mechanical behavior of these non-linear systems by means of fractional calculus
[3].
Fractional calculus is a generalization of ordinary differentiation and integration to arbitrary non-integer
order, but with this definition, many interesting questions will arise; for example, if the first derivative
of a function gives you the slope of the function, what is the geometrical meaning of half derivative? In
half order, which operator must be used twice to obtain the first derivative? The early history of this
questions back to the birth of fractional calculus in 1695 when Gottfried Wilhelm Leibniz suggested
the possibility of fractional derivatives for the first time
[2]
.
In this article, we aim to introduce fractional calculus as a new tool for modeling the complex systems,
especially viscoelastic material. First, we briefly discuss the basic concepts of fractional calculus and
explain the essential steps of the fractionalization algorithm. Next, we present an interpretation of
fractional derivative and elaborate upon how fractional equations could be solved analytically. Then,
we briefly look at the modeling of viscoelastic systems by the help of this approach. Ultimately, after
overviewing some recent works, we present an application of the approach in modeling biomechanical
properties of a cell and indicate that the proposed model predicts the cell behavior much better than the
previous spring-dashpot models, as well as the model outputs are in good agreement with experimental
data. To sum, we are going to give the minimum need to get reader "feet wet", so that a reader can
quickly get into building a fractional calculus model for a complex system.
Methods
Fractional calculus
In mathematics, many complex concepts developed from simple concepts. For example, we can refer
to the extension of natural number to the real one in some mathematical formulae. Let's give an example
to clarify: the factorial of a non-negative integer n, denoted by n!, is the product of all positive integers
less than or equal to n. On the other hand, there is a concept which named Gama function and defined
as follows:
Γ(𝑥) = ∫ 𝑡𝑧−1𝑒−𝑡𝑑𝑥
∞
0
(1)
3
One property of the function for 𝑛 ∈ ℝ+ is
Γ(𝑛 + 1) = nΓ(𝑛)
(2)
Hence, this function is equal to factorial for the integer numbers. As a result, the gamma function could
be considered as an extension of factorial function to real numbers. For instance, according to the above
formalism, a factorial of 1/2 can be obtained as follows:
1
(
2
) ! = Γ (
3
2
) =
3
2
Γ (
1
2
) =
3√𝜋
2
(3)
Indeed, according to Wikipedia, the gamma function can be seen as the solution to find a smooth curve
that connects the points (x, y) given by y = (x − 1)! at the positive integer values for x. Let's use this
approach to extend the concept of derivative to non-integer order; consider 𝑛𝑡ℎ derivation of power
function 𝑔(𝑥):
𝑔(𝑥) = 𝑥𝑘 𝑠. 𝑡 𝑥 ≥ 0
𝑑𝑛
𝑑𝑥𝑛 𝑔(𝑥) =
𝑘!
(𝑘 − 𝑛)!
𝑥𝑘−𝑛 =
Γ(1 + 𝑘)
Γ(1 + 𝑘 − 𝑛)
𝑥𝑘−𝑛
(4)
(5)
which 𝑘 and 𝑛 are real integer number respectively, and 𝑘 ≥ 𝑛. To generalize the above equation, it
could be possible to extend the integer number 𝑛 to a real value which named 𝛼:
𝑑𝛼
𝑑𝑥𝛼 𝑔(𝑥) =
Γ(1 + 𝑘)
Γ(1 + 𝑘 − 𝛼)
𝑥𝑘−𝛼
(6)
Then for fractional derivative of an arbitrary function, expand the function in a power series of 𝑥 first,
and then by using 𝑒𝑞. 5, derivate the expansion. For example, for derivative 𝑓(𝑥) = 𝑒𝑘𝑥 to 𝛼 order, we
rewrite 𝑓(𝑥) function as follows:
Hence [4]:
That Γ(−𝛼, 𝑘𝑥) is incomplete gamma function (discussed in Box 3).
(7)
(8)
Of course, this is an arbitrary way to define fractional derivative and not the only way; for example, it
is possible to use an exponential function 𝑓(𝑥) = 𝑒𝑘𝑥 instead of a power function, i.e. we can
define
𝐷𝛼𝑓(𝑥) = 𝑘𝛼𝑒𝑘𝑥
(9)
The fractional derivative of the exponential function obtained by Liouville in 1832, and the fractional
derivative of power function got by Riemann in 1847 [4]. By comparing Eq. 8 with 9, it is noticed that
when the order is an integer, the results are same, but at the non-integer order they are different! In fact,
contrary to the integer-order derivative in which the definition and the output under these operators are
same and unique in the fractional derivatives, under different operators, the result is not the same and
not necessarily unique. In other words, there are multiple definitions for fractional derivative and all of
them are mathematically correct. From the physical perspective, each definition has its own application
and interpretation.
4
23()12!3!xxfxx12,1111123kxkxdfxxxxsignxsignxkedx Fractional derivative has many interesting and counter-intuitive properties; for example, the Riemann
derivative
(Eq. 5) of a constant
is not zero!
𝑑𝛼
𝑑𝑥𝛼 𝑥0 =
𝑑𝛼
𝑑𝑥𝛼 𝑐𝑜𝑛𝑠 =
that
1
is,
𝑥−𝛼
Γ(1−𝛼)
(10)
Putting it all together, there are many ways to define a fractional derivative, provided that each
definition approach to ordinary derivative in the integer order limit-this method known as
the fractionalization algorithm-i.e.:
Where Δ represents an arbitrary operator. As well, two more examples provided in Box 1 which have
also been used in the following sections.
Box 1
In this box, a basic definition of both the ordinary integral and the ordinary derivative is presented by
two examples, as well as its extension to fractional operators is proposed by the help of fractionalization
algorithm (ref section 2); resulting we obtain two important fractional functions.
Example 1. Assuming 𝑓(𝑥) ∈ [𝑎, 𝑏]. The first order derivative of function 𝑓(𝑥) is defined as follows:
by applying derivative operator on the above equation, we get:
by the help of induction and above procedure, we gain:
(B1.1)
(B1.2)
(B1.3)
𝑡−𝑎
𝑛(𝑛−1)(𝑛−2)…(𝑛−𝑟+1)
, and (
Where 𝑎 ≤ 𝑡 ≤ 𝑏, ℎ =
Since the equation (B1.3) hold for all 𝑛 ∈ 𝑁, we expand 𝑛 order to 𝛼 ∈ 𝑅 by the fractionalization
algorithm; thereby achieving a really important equation known as fractional Grunwald-Letnikov
differintegral mentioned in section 2.
) =
.
𝑟!
𝑛
𝑛
𝑟
Example 2. Consider the following equation which is known as the integral operator of order n:
We can rewrite it as follow according to the Cauchy equation [6]:
(B1.4)
Since the above equation holds for all 𝑛 ∈ 𝑁, we apply the fractionalization algorithm to expand the
obtained equation to fractional equation. In other words, if n → α, we will able to gain an integral known
as Riemann-Liouville. Hence, it seems that fractional calculus is an expansion of integer calculations.
(B1.5)
(B1.6)
5
lim()()(cid:160)(cid:160),(cid:160)(cid:160)0,1,....nnfxfxn(1)0()()()limhdfftfthftdth2(2)220()2()(2)()limhdfftfthfthftdth()2001()lim(1)()nnnrnhrndfftftrhrdth()001()lim(1)()rnhrftftrhrh11001()dxnnxxxnnDffxdx11()()(1)!xnnDfxfdn Definition of fractional calculus
As mentioned in the previous section, the fractional order derivative is not necessarily unique; in this
regard, there are some accepted and common definitions in the literature. In the following, we
mentioned a number of important ones.
Grünwald–Letnikov derivative
Grünwald–Letnikov derivative is a basic extension of the natural derivative to fractional one, which
derived in Box 1 (Eq. B1.4). It was introduced by Anton Karl Grünwald in 1867, and then by Aleksey
Vasilievich Letnikov in 1868. Hence, it is written as [5]:
(11)
where 𝑛 ∈ 𝑁, and the binomial coefficient is calculated by the help of the Gamma function.
)12(
Riemann-Liouville fractional derivative
Riemann-Liouville fractional derivative acquiring by Riemann in 1847 is defined as follows.
(13)
where 𝛼 > 0; this operator is an extension of Cauchy's integral (Box 1) from the natural number to real
one. Based on the fractionalization algorithm, it seems logical to reach relation (13) by n order
derivative of the Eq. B1.6
In addition, according to the above relation, if 0 < 𝛼 < 1 then the Riemann-Liouville operator reduced
to
(14)
It is worth noting that this relation is the same of Eq. 5. [5], and also RL, 𝛼, 𝑎, and 𝑡 are the abbreviation
of Riemann-Liouville, fractional order, the lower and upper bound of the above integral respectively.
Caputo derivative
Since Riemann-Liouville fractional derivatives failed in the description and modeling of some complex
phenomena, Caputo derivative was introduced in 1967 [6]. The Caputo derivative of fractional order α
(𝑛 − 1 ≤ 𝛼 < 𝑛 ) of function 𝑓(𝑡) defined as
(15)
6
001()lim(1)()mhmDftftmhmh(1)(2)...(1)!mmm()1()()()()1()(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(cid:160)(),([]1,)()()RLnnatattnnadDftDftdtdfxdxntandttx1()()(1)()tRLatadfxDftdxdttx11()()()()ntCatnaDfDftdnt where 𝐷𝑛 is 𝑛𝑡ℎ derivative operation, and C represents the Caputo word. In other words, based on box
2, it can be found that the Caputo derivative is equal to the Riemann–Liouville integral of a
𝑛𝑡ℎ derivative of a function.
It is worth noting again that "the behavior of all of the fractional derivatives, when the order is integer
are same; it, however, could be different in the non-integer order; for example, in the non-integer order,
the Caputo derivative of a constant, unlike the Riemann–Liouville derivative is zero.
There are more fractional derivatives for more detailed information, the interested reader is referred to
[4, 5, 7]. Since we try to analyse and to investigate viscoelastic systems via the Riemann-Liouville and
Caputo fractional derivative, the main focus of the article is on these two types of derivative.
As previously mentioned, different definitions for fractional derivative with the different properties can
be proposed, which all of them are valid and mathematically acceptable. However, the main question
is "which relation should be applied in modeling of a specific phenomenon? In other words, which
definition would be more appropriate for a specific problem?" As a rule of thumb, since they tend to
interpret natural phenomena, the definition which is more consistent with the experimental results have
more privilege than the other fractional definitions.
BOX 2
In this box, some properties of fractional Riemann-Liouville differintegral, such as commutative
property, distributive laws, and so forth are investigated2.
Riemann–Liouville integral and derivative
To find a profound understanding of Riemann–Liouville integral and derivative, some of the most
crucial properties of this operator are mentioned in the following.
Lemma 1. Assuming arbitrary function 𝑓(𝑥) and 𝑚, 𝑛 ≥ 0 the following equations hold3.
1. Semi-group property:
2.
commutative property:
Lemma 2. Let 𝑓1 and 𝑓2 are two functions on [a,b] as well as 𝑐1, 𝑐2 ∈ ℝ, 𝑛 > 0, and 𝑚 > 𝑛. Regarding
these, the following equations hold4:
1. Linearity rules:
2. Zero rule:
3. Product rule:
𝐷0𝑓 = 𝑓
4. In the general, semi-group property does not hold for Riemann-Liouville fractional derivative.
Indeed, the following equation is not always true.
2 All the mentioned properties in this box hold almost everywhere (a.e) on [a, b] and also, if (X, Σ, μ) is a measure space, a quality 𝑃 is said to hold almost
𝑏.
everywhere in 𝑋 if 𝜇({𝑥 ∈ 𝑋: 𝑃(𝑥)}) = 0. Also, integral operator on interval [𝑎, 𝑏] is defined as 𝐼𝑎
3 𝑓 ∈ 𝐿1[𝑎, 𝑏] where 𝐿1 ≔ {𝑓: [𝑎, 𝑏] → ℝ; 𝑓 𝑖𝑠 𝑚𝑒𝑎𝑠𝑢𝑟𝑎𝑏𝑙𝑒 𝑜𝑛 [𝑎, 𝑏] 𝑎𝑛𝑑 ∫ 𝑓(𝑥)1𝑑𝑥 < ∞
}.
4 Provided that 𝐷𝑎
𝑛𝑓2 almost everywhere define.
𝑛𝑓1 and 𝐷𝑎
𝑎
𝑏
7
mnmnaaaIIfIf()()mnnmaaaaIIfxIIfx12121111(),()()RLnRLnRLnRLnRLnaaaaaDffDfDfDcfcDf0()()()RLqRLjRLqjtttjqDfgDfDgjRLaRLbRLabDDfDf NB: To prove the above lemmas see [8].
Caputo derivative
In this part, some fundamental properties related to Caputo operator are represented based on ref [5, 8].
Let 𝑓 is a enough differentiable function, 𝑐1, 𝑐2 ∈ ℝ, and 𝑚 > 𝑛 ≥ 0. Considering these,
1. Caputo derivative is the left inverse of Riemann-Liouville integral.
2.
3. Distributive law in Caputo derivative.
4. Leibniz equation5.
5. The semi-group property, the following equation, holds under some special condition6.
NB: The part 5 of the above Lemma) does not hold for Riemann–Liouville derivative generally. To
prove this claim, assuming 𝒇 function with identity function 𝒇(𝒙) = 𝟏, as well as 𝑎 = 0, 𝑛 = 1, and
1
2𝑓 )(𝑥) =
𝜀 = 1/2. Thus, if Riemann–Liouville operator is applied to the left side, we will obtain (𝐷0
1
20 = 0. On the other hand, by applying this operator to the right side, we will gain
𝐷0
The relationship between fractional integral and fractional derivative
In this part, the relationship between fractional integral and fractional derivative is investigated. There
are always several properties in the classical integral and derivative area accepted as constant roles;
these properties, however, might not hold always in the fractional sense. The following equations, for
instance, are almost same in the classical calculation, but it has a major difference in the fractional
calculation (𝑚 > 𝛼).
and
Indeed, a constant and a variable term in calculation represent this difference, which the constant and
variable term are related to non-fractional and fractional sense respectively. In other words, assuming
function f: [a, b] → ℝ where −∞ < a < b < ∞; if the function 𝑓(𝑥) on closed interval [𝑎, 𝑏] has been
𝑛 order differentiable, then for all 𝑥 ∈ (𝑎, 𝑏), we have:
5 Assuming 0 < 𝑛 < 1, 𝑔 and 𝑓 are analytical on (𝑎 − ℎ, 𝑎 + ℎ) are essential to hold the equation.
6 Special condition for existing this property is 𝑓 function must belong to 𝐶𝑘[𝑎, 𝑏] where 𝐶𝑘[𝑎, 𝑏] ≔ {𝑓: [𝑎, 𝑏] → ℝ; 𝑓 ℎ𝑎𝑠 𝑎 𝑐𝑜𝑛𝑡𝑖𝑛𝑢𝑜𝑢𝑠 𝐾𝑡ℎ 𝑑𝑒𝑟𝑖𝑣𝑎𝑡𝑖𝑣𝑒} for
some 𝑎 < 𝑏 and some 𝑘 ∈ ℕ. Existing ℓ ∈ ℕ with ℓ ≤ 𝑘 and 𝑛, 𝑛 + 𝜀 ∈ [ℓ − 1, ℓ] is essential as well (𝑛, 𝜀 > 0).
NB: The existence of ℓ with the mentioned properties is essential. Otherwise, it is enough to assume that 𝑛 = 𝜀 = 7/10(𝑖. 𝑒 7/10 = 𝑛 < 1 < 𝑛 + 𝜀 = 7/5),
𝑎 = 0, and 𝑓(𝑥) = 𝑥, causing the right side of lemma become zero owing to
𝑥13/10. As a result,
7/10𝑓(𝑥)=
; however, 𝑐𝐷0
1
Γ(13/10)
the left side of lemma will obtain as follow:
8
CnnaaDIff10()()()()!kmnCnkaakDfaIDfxfxxak11221122CnCnCnaaaDcfcfcDfcDf1()()()()()()()()()(1)nCnCnknCkaaaaknxaDfgxgafxfaDgxfxIgxDfxknCCCaaaDDfDf3/221/23/2001()()(1/2)DfxDIfxxmmmmamdDDDIdxmmmmamdDDDIdx7/53/53/5000()()()00CDfxIfxI7/107/102/5001()()(3/5)CCDDfxx
In a simple word, if the fractional derivative with order 𝛼 ∈ ℝ is applied on the fractional integral with
the same order, the output of that will be 𝑓(𝑥). Thus, we can say, "Caputo derivative is the left inverse
of Riemann-Liouville integral." Also
NB: To prove the mentioned theorems see [9].
With regarding the Riemann-Liouville as a derivative operator, the following equation hold (𝒏 > 𝟎,
𝒎 > 𝒏):
and particularly, for 0 < 𝑛 < 1, we have:
The interpretation of fractional calculus
There is a dozen of suggestions for interpretation of fractional calculus [4, 7]. However, most of them
are a little bit abstract and give no physical intuition. Hence, we just present a most useful interpretation
of fractional calculus. The core of these interpretations is memory concept. In general, when the output
of a system at each time 𝑡 depends only on the input at time 𝑡, such systems are said to be memoryless
systems. On the other side, when the system has to remember previous values of the input in order to
determine the current value of the output, such systems are said non-memoryless systems, or memory
systems. For example, all of the Markov chain phenomena are memoryless [3, 4], and human decision
making or shape memory alloy are non-memoryless.
Laplacean interpretation
Suppose that 𝑌(𝑡) is a quantity whose value in terms of 𝑓(𝑡) can be achieved as follows:
(16)
i.e. the output 𝑌(𝑡) can be viewed as a power-weighted sum which stores the previous input of function
𝑓(𝑡). Based on the above definition, such system is a non-memoryless system and in such systems,
memory decays at the rate of 𝑤(𝑡) = 𝑡𝛼−1/Γ(α).
Applying the Caputo derivative of order 𝛼 to the both sides of the last relation leads to
(17)
As a result, the differential equation governing the system memory 𝑌(𝑡) is described by a fractional
derivative. Therefore, the fractional derivative is a good candidate to explain the system with memory.
The nature of weighted function determines the type of fractional derivative which describes a system
memory. For example, If the weight function of a system is defined by 𝑡1−𝛼 ∕ 𝛤(𝛼) , the Riemann-
9
()()()(cid:160)CaaDIfxfx1()0()()()(())!jnCjaaxajxaIDfxfxfxj1110()()()lim()()nkmnRLnmkmnaaazakxaIDfxfxDIfznk11()()()lim()()nnRLnnaaazaxaIDfxfxIfzn(1)0()()())(ttYtfd0()()CtDtftY Liouville elements, and by 𝑡1−𝛼𝜃(𝑥 − 𝑡)/Γ(𝛼) the Caputo elements are used--which θ is the Heaviside
function [4].
Since the convolution operation representing the integral of the product of the two functions after one
is reversed and shifted, so if one of the functions considered as a weighted function, it may be claimed
that the task of this function is collecting the system memory over time. For example, in the Riemann-
Liouville, "The relationship indicates that the information function 𝐹(𝑡) is memorized (storage) with a
power low rate function". In Ref [4] some other weight function is reviewed.
From the physical point of view, what the memory is and how it is defined in a system depends on a
deep understanding of the phenomena. Also, there is no certain rules and methods to select the
fractional type in a modeling. In other words, any definition that its result is more consistent with the
experimental data is the suitable definition to us.
Analytical and numerical methods
Analytical methods
There are different methods to solve fractional differential equations analytically. One of the most
common and widely used methods is the Laplace transform. In the following, by an example, this
method is illustrated (for other methods see reference [7]).
Before proceeding, it is worth noting that in general, the number of initial conditions that are required
for a given ordinary differential equation will depend upon the order of the differential equation.
However, in the fractional differential equation, number of initial condition equals to the integer lower
bound of order value (𝛼) [10, 11].
Consider the following differential equation
(18)
Which 𝑥(𝑡) is displacement, 𝑘, and 𝜉 are constants, as well as the fractional derivative is also Caputo
and 0 < 𝛼 < 1--In what follows, it has been shown that this differential equation model is the dynamics
of a purely elastic spring and a viscoelastic element connecting in parallel with a body of mass m, which
a force f is applied on a body. In order to solve, the first step is to take the Laplace transform of both
sides of the original differential equation (the Laplace transformation is concisely explained in Box 3).
Based on the equation B3.2 in Box 3, we have
(19)
where 𝛼 and 𝑠 are fractional order and Laplace domain variable respectively. Also, it is supposed that
𝑥(0) = 0. To find the solution, all we need to do is to take the inverse transform (Box 3.1):
Which 𝐸𝛼,𝛼(𝑡) is Mittag-Leffler function [for more information see the Box 3 footnote]
(20)
10
()()Dxtkxtf()(/)fXssk1,()()fkxttEt In the last example, if the spring is ignored, the Eq. 18 will be reduced to
𝑓 = 𝜉𝐷𝛼𝑥(𝑡)
(21)
Then as above, by a) taking the Laplace transforms of both sides of the equation, b) simplifying
algebraically the result to solve the obtained equation in terms of 𝑠, and c) finally finding the inverse
transform, we have:
𝑥(𝑡) = 𝐾𝑡𝛼
(22)
Which 𝐾 = 𝑓/𝜉Γ(1 + 𝛼). Although the Laplace transformation method is one of the simple and
practical methods for solving the fractional equations--same as the ordinary differential equations, most
of the fractional equation could not be solved analytically. In what follows, we present a numerical
technique to solve Caputo fractional differential equation.
BOX 3
In this box, the concept of Laplace Transform is presented for several fractional definitions. To begin
with, let us bring up some basic facts in this regard [7].
The function 𝐹(𝑠) of the complex variable 𝑠 is defined as the following equation-which is known as
the Laplace transform of the function 𝑓(𝑡):
To exist integral (B3.1), the function 𝑓(𝑡) must have been an exponential order 𝛼. In other words, the
existence of two positive constants, such as 𝑀 and 𝑇 which satisfy the following condition is essential.
Indeed, when 𝑡 → ∞, the function 𝑓(𝑡) cannot grow faster than a certain exponential function.
With the help of the inverse Laplace transform, the original 𝑓(𝑡) can be gained from the Laplace
transform 𝐹(𝑠)
where the integration is done along the vertical line 𝑅𝑒(𝑠) = 𝑎 in the complex plane such that 𝑎 is
greater than the real part of all singularities of 𝐹(𝑠), which guarantees that the contour path is in the
convergent region. If either all singularities are in the left half-plane or 𝐹(𝑠) is a smooth function on
−∞ < 𝑅𝑒(𝑠) < ∞ (i.e., no singularities), then 𝑎 can be set to zero and the above inverse integral
formula becomes identical to the inverse Fourier transform.
The direct calculation of the inverse Laplace transform (B3.2) is often sophisticated. Sometimes,
however, it gives useful information on the unknown original 𝑓(𝑡) behavior which we look for.
The following formula seems to be another useful property for the Laplace transform of the derivative
of an integer order 𝑛 of the function 𝑓(𝑡):
which can be obtained from the definition (B3.1). To facilitate applying the Laplace transform, we
gather some formulae of vital functions in the table (B3.1) for both the integer mode and the non-integer
mode.
11
0(){()}()(3.1)stFsLfteftdtB().teftMforalltT11(){()}lim()(3.2)2aibstaibbftLFseFsdsBi1()12(1)1()0{()}()(0)(0)(0)()(0)(3.3)nnnnnnnnkkkLftsFssfsffsFssfB
Table B3.1. Laplace transform table of some basic fractional calculus.
Fractional order
Integer order
𝑓(𝑡) = 𝐿−1{𝐹(𝑠)}
𝐹(𝑠) = 𝐿{𝑓(𝑡)}
𝑓(𝑡) = 𝐿−1{𝐹(𝑠)}
7
8
9
𝑭(𝒔) = 𝑳{𝒇(𝒕)}
The Laplace transform of fractional operators
In this part, the Laplace transform of fractional operators is represented, and some formulae of which
is summarized in table (B3.2).
We introduced Riemann- Liouville integral in Box 1 as follow:
With applying the Laplace transform on the Eq. B3.4, we obtain (based on Table B3.1):
7 incomplete gamma function
8 𝐸𝛼,𝛽 is denoted to Mittage-Leffler function such that
Also, if 𝛼 = 𝛽 = 1, then this function can be equaled with exponential function.
9 Hypergeometric functions
12
()t11()t1s1,1,2,3,...()ktkk1ks1()tte1()sa()();,aftbgtab()()aFsbGs1(,)()at()assa()(),0nftn12(1){()}(0)(0)(0)nnnnsLftsfsff()Eat()sssasin()at22asa1,()tEat1sacos()at22ssa1()Eat()assa()cteft()Fsc11(;1;)Fat()sssa(),1,2,3,...ntftn()(1)()nnFs1,1()tEat1()ssa0()()tftgd()()FsGs1,1(),01tEat()ssa1()ftt()sFudu1,()tEatssa0()tfd()Fss111(;;)()tFat()ssa001111()()[*()](3.()4)()ttftfIdtftB101{}{}{()}()()LIfLtLftsFs1(,)atzaztedt,0(),,0()kkxExk11011()()({};{};)!()()qpnjjiipqijpqnijijbanzFabznabn
Similarly, other Laplace transform of fractional operators will be gained by utilizing the proposed
formulae in Table B3.1 [7].
Table B3.2. The Laplace transform of some fractional operators with order 𝛼.
Operator
Fractional formula
Riemann-
Liouville integral
Riemann-
Liouville
derivative
Caputo derivative
Griinwald-
Leitnikov
fractional
derivative
10
Numerical solution
Laplace Transform of the fractional order
operators
Until now, various numerical methods have been proposed for solving fractional differential equations
which are based on discretizing the relationships [12-17]. In this section, a new algorithm which was
recently published in [18] is proposed discretizing the Caputo derivative.
Considering the Caputo derivative of function 𝑢(𝑡) with order 0 < 𝛼 < 1, the following equations hold
for discretizing 𝑢(𝑡) on [0,T] in a uniform way.
(23)
Where 𝑑𝛼,𝑗 = (𝑗 + 1)1−𝛼 − 𝑗1−𝛼 and 𝑗 = 0,1,2, … , 𝑛, as well as 𝑟𝛼,𝜏
𝑛+1is local truncation error.
To utilize the above algorithm, equation (21), which used in the last section simulation, is considered.
The following equation will be obtained by applying the equation (23) on equation (21).
(24)
Comparison between the numerical method and the analytical answer is shown in both the figure (1)
and the table (1)
10 The Laplace transform of the Griinwald- Letnikov fractional derivative of order 𝛼 > 1 does not exist in the classical sense.
13
()sFs1001()()()()ttfttfId11000()(),1nkRLktktsFssDftnn011()()()()()tRLntnadfxDftdxndttx11()0()(0),1nkkksFssfnn011()()()()ntCtnaDfDftdnt(),01sFs000()lim(1)()nmthmnhtaDfthftmhm11110111,011,,01()()()(1)()()1()(1)()()1(2)jjjjnttnntjntjjnntjnnjnjnjjusuttsdssututtsdsrututdr1,11()(2)()()(()())nnnjnjnjjxtfxtdxtxt
Fig 1. A schematic comparison of analytical and numerical answer with the represented mesh in Table 1.
Table 1. Comparison between the numerical results, equation 24, and the analytical answers of equation (25) with 𝑓 = 1100𝑝𝑁 stress (𝜉 and
𝛼 are chosen from table 4).
Mesh point (𝝉)
Numerical result in 𝒕 = 𝟑
Analytical result in 𝒕 = 𝟑
Difference
Percentage error
0.1
0.01
0.001
848.3334
849.8002
849.9445
849.9605
849.9605
849.9605
1.6271
0.1603
0.0160
0.1914
0.0188
0.0018
Viscoelastic systems
Elasticity is the ability of a material to resist a distorting or deforming force and return to its original
form when the force is removed. According to the classical theory in the infinitesimal deformation,
most elastic materials, based on Hooke's Law (Eq. 25), can be described by a linear relation between
the stress 𝜎 and strain 𝜀 [19].
𝜎(𝑡) = 𝐸𝜀(𝑡)
(25)
where 𝐸 is a constant known as the elastic or Young's modulus.
The viscosity of a fluid is a quantity which describes its resistance to deformation under shear stress or
tensile stress. In the ideal viscose fluid, according to the Newtonian fluid (Eq. 26) the stress is
proportional to the local shear velocity (i.e. the rate of deformation over time):
𝜎(𝑡) = 𝜂
𝑑𝜀(𝑡)
𝑑𝑡
(26)
where 𝜂 is the shear modulus.
However, many liquids briefly respond like elastic solids when subjected to abrupt stress. On the
contrary, many solids will flow like liquids, albeit very slowly, under small stress. Such materials
possessing both elasticity (reaction to deformation) and viscosity (reaction to the rate of deformation)
are known as viscoelastic materials. When these materials subjected to sinusoidal stress, the
corresponding strain is neither in the same phase as the applied stress (like an ideal elastic material) nor
in the 𝜋/2 out of phase (like an ideal viscose material). In these materials, some part of input energy is
stored and recover in each cycle and the other part of the energy is dissipated as heat. Materials whose
behavior exhibits these characterized are called viscoelastic. If both strain and stress be not only
infinitesimal but dependent upon the time as well, strain-stress relation (constitutive equation) can be
described by a linear differential equation with constant coefficient i.e. the material which exhibits
linear viscoelastic behavior. One of the main consequent of this assumption is the stress (strain)
responses to successive strain (stress) stimuli are additive. In other words, in the creep experiment
consist of applying a step stress 𝜎0, (a stress increment at time t = 0, which is kept constant for t>0),
and measuring the corresponding stain respond 𝜀(𝑡), the constitutive equation is 𝜀(𝑡) = 𝜎0𝐽(𝑡), where
𝐽(𝑡)-which termed creep compliance-is the strain at time t owing to a unit stress increment at time
14
0. Suppose that various stress increments Δ𝜎 occur at the successive time interval Δ𝑡 based on the
superposition of effects, we get:
After N steps, we have
(27)
(28)
This relation recognized as Boltzmann superposition principle which states that the response of a
material to a given load is independent of the response of the material to any load, which is already on
the material [19].
Fig 2. A graph versus time of (a) a step-plus increment stress applied to linear viscoelastic material and (b) the resulting strain.
in the limit Δ𝑡 → 0
(29)
This relation is termed hereditary or superposition integral. Similarly, with applying a step strain, we
get
(30)
Where G(t) is relaxation modulus.
In the mechanical science, to evaluate mechanical properties of materials, after sudden stress and
dynamic (periodic) experiments, researchers implement certain tests, such as creep tests. In the
following, we will briefly discuss these two tests.
Creep test
One of the most common tests on the viscoelastic system is creep test. In this test, a system is subjected
to constant tension and the amount of deformation is recorded over time. For viscoelastic materials
creep vs. time diagram shows power-law behavior (figure 3).
(31)
15
0()()()tJtJtt01()()()NntnttJtJtnttt00()()()()tdttJtJttdtdt00()()()()tdttGtGttdtdt()()tJt Where 𝐽(𝑡) = 1/𝐺(𝑡) is termed creep compliance.
Fig 3. The profile of a shear creep experimental.
Periodic test
In this common experiment, a system subjected to a cyclic applied stress. Therefore, as mentioned
above, for a viscoelastic system, strain in response to an imposed periodic stress of angular frequency
𝜔 is also periodic with the same frequency but in the different phase (figure 4). Based on relation (30),
if the periodic stress 𝜀 = 𝜀0sin (𝜔𝑡) applied, we expect that the stress follows a periodic response with
the same frequency and different phase:
And by extension of relaxation modulus, we define two shear storage modulus G', and shear loss
modulus G" as follows
(32)
(33)
Fig 4. Geometry and time profile of a simple shear experiment with sinusoidally varying shear [19].
16
000cos()cossinsincosttt0000(/)cos(/)sin/tanGGGG Fractional viscoelastic model
There are different models to evaluate and to predict the constitutive equation in the viscoelastic system,
these models commonly, composing of a different combination of springs and dampers elements [20,
21]. For example, two well-known models to name are Maxwell and Kelvin models (Table 2).
In this section, a model which is known as fractional order element (FOE) with equation (39) is
presented in order to investigate the mechanical response of viscoelastic systems from different points
of view. It is indicated that this fractional model can be a good substitute for modeling the viscoelastic
materials than SLE models. With this aim, let us take a closer look at Maxwell model; based on this
model, the constitutive equation will be as follows:
1
𝐸
𝑑𝜎
𝑑𝑡
+
𝜎
𝜂
=
𝑑𝜀
𝑑𝑡
(34)
where E(0+) = η(0+). Also, 𝜎, 𝜀, 𝜂, and 𝐸 are stress, strain, shear modulus, and Young modulus
respectively. The following equation will be obtained by taking Laplace transformation and considering
a step strain function.
𝑡
𝐺(𝑡) = 𝐸𝑒−(
𝜏
)
(35)
Where 𝑡 > 0 and 𝐺(𝑡) is relaxation modulus.
In nature, the 𝐺(𝑡) value in viscoelastic materials does not usually follow a simple exponential behavior,
and a more realistic expression for its behavior is a power-low response [3]:
𝐺(𝑡) =
𝐸
Γ(1 − 𝛼)
−𝛼
)
(
𝑡
𝜏
(36)
Table 2. Viscoelastic models, which 𝜎, 𝜀 𝑎𝑛𝑑 𝜏 are stress, strain and 𝜂/𝐸 respectively; also, 𝑋𝛼 = 𝐸1−𝛼𝜂𝛼 and 0 < 𝛼 < 1.
Name
Model
Constitutive equation
Maxwell fluid
Voigt fluid
FOE
Relaxation modulus
(𝑮(𝒕))
As it mentioned in the previous section, the value of stress in a pure elastic system, an ideal spring, is
proportional to the derivative zero displacements. On the other hand, it is proportional to the first-order
derivative of displacement in the pure viscous system. It is expected that the viscoelastic systems have
a behavior between these two systems, an elastic and viscous material, resulting to conclude that stress
is proportional to a fractional derivative whose order is between zero and one (Fig 5). Hence, from the
mathematical point of view, this physical interpretation can be modelled as the following equation.
𝜎(𝑡) = 𝑋𝛼𝐷𝛼𝜀(𝑡)
(37)
Where 0 ≤ 𝛼 ≤ 1 and 𝑋𝛼 are the constant coefficients to make an equal relation between stress and
strain-the 𝛼 index represents the dependence of the constant 𝑋 to expanded order.
17
()()()tDtEDt/tEe()()()tEtDt()Et()()tXDt(1)Et Fig 5. A schematic representation of viscoelastic regime and stress-strain relation from a physical point of view.
To obtain the coefficients, the essential condition of fractionalization algorithm (section 2) is applied
first. Thus, the equation (41) must contain both Hook's law and Newton fluid in boundary points, i.e.
𝛼 = 0 𝑎𝑛𝑑 𝛼 = 1. In other words
(38)
One of the specific cases which can satisfy the mentioned boundary condition is 𝑋𝛼 = 𝐸1−𝛼𝜂𝛼. In what
follows, it is shown that the proposed 𝑋𝛼 is a good suggestion. To make a sufficient condition in
fractionalization algorithm, the physical point of view must be considered and to this end, the equation
(37) with the assumed coefficients is rewritten.
𝜎(𝑡) = 𝐸1−𝛼𝜂𝛼 𝑑𝛼𝜀(𝑡)
𝑑𝑡𝛼
(39)
By applying the Laplace transform to the above FOE the following equation (40) will be obtained-
derivative is Caputo.
(40)
As a result, the relaxation modulus will be achieved by considering a unit step strain function.
𝐺(𝑡) =
𝐸
Γ(1 − 𝛼)
−𝛼
𝑡
(
𝜏
)
Where 𝜏 = 𝐸/𝜂. Thus, we gained the same equation (40) by the help of physical interpretation and
mathematical methods which have better match with the experimental results.
In equation (43), 𝛼 and 𝑋𝛼 are quantities which represent the system property. In other words, it could
be possible to consider the 𝐸1−𝛼𝜂𝛼 coefficient as an independent quantity which is introduced by
variable 𝑋𝛼-it means that whatever determine the quantity and the structure of the system are just the
value of 𝛼 and 𝑋𝛼; indeed, it is possible that a system with 𝐸′ and 𝜂′ modulus has equal dynamical
behavior as another system which has different modulus as 𝐸 and 𝜂 i.e. 𝜂 ≠ 𝜂′ and 𝐸 ≠ 𝐸′, but
𝐸(1−𝛼)𝜂𝛼 = 𝐸′(1−𝛼)𝜂′𝛼. Hence, the product of two moduli which expressed in equation (40) will be
just one independent quantity.
If 𝛼 → 0, the equation (40) will lead to the Hook's law and if 𝛼 → 1, then the equation turn to Newton
fluid model. As a result, based on the physical interpretations, it looks logical to conclude that the
proposed model can be a proper tool for investigating the mechanical properties of viscoelastic
materials.
To take a more precisely look at the FOE equation in the viscoelastic systems, in what follows, the
formalism of inheritance integral for the fractional state is investigated. Equation (41) is obtained by
taking partial integral from equation (29).
(41)
18
01lim,limXEX1()()sEss0()()()(0)()()tdJttttJtdtdtt Also, the creeping state with unit step stress function will be as follow based on FOE equation.
(42)
By taking derivative from equation (42) and putting the obtained result in equation (41), equation (43)
will be achieved.
(43)
Which the equation inside the bracket is Riemann-Liouville integral; thus, by rewriting the above
equation, we have
(44)
According to the abovementioned subjects, in equation (44), the previous events are embedded in the
fractional kernel function 𝑡𝛼−1/Γ(𝛼); which the obtained result has match with the proposed
interpretation of section (3). Regarding those interpretation, the information of a system, which is
related to deformation, are storing with a power-law function which has fractional order 𝛼 from the
initial step (𝑡+ = 0) until the present moment (𝑡)-which the fractional order is the indication of saving
order information. According to the kernel function 𝑡𝛼−1/Γ(𝛼) in equation (44), 𝛼 = 0 and 𝛼 = 1 are
equal with a memory-less system and full-memory system in creeping state respectively owing to
Γ(0) = ∞. Indeed, in a pure viscose system, system's position completely depend on this fact that how
stress applies on the system over time. In other words, the nature of stress and the amount of which has
a direct impact on the present position of the system. However, owing to this fact that the effect of
memory at different times has a unit weight function, the system is able to save all the prior history of
itself precisely and equally. In contrast, in the pure elastic system, the present stress is an effective
element on the amount of displacement, not previous stresses, because the weight function is zero.
Similarly, the above procedure can be followed for relaxation modulus. The following equation will be
obtained by putting equation (36) in (30).
(45)
As a result, by the help of equation (15) and rewriting the above equation, constitutive equation is gained
as follow.
Based on the equation (45), in the relaxation state, 𝛼 = 0 and 𝛼 = 1 represent a full-memory system
and a memory-less system respectively-Γ(0) = ∞. In other words, in the pure elastic systems, the
displacements of an object in prior times completely keep in the memory of object owing to Because
of storing displacement changes over time by an unit function.-known as weight function-over time,
and there is no differences in very far and very near time. As a result, all the previous displacements
will have been an effect on the value of stress. In the pure viscose system, however, such a memory as
the mentioned one does not exist, as well as the previous displacement will not have impacted on the
present stress of system; thus, the system is absolutely devoid of any memory in such condition. This
is a really intriguing, fascinating interpretation of FOE model with equation (39) because on the one,
19
1(/)()(1)tJtE1011()()(0)()()()tttJtttdtE01()()(0)()tttJItE00()()()()(1)tEdttGtttdtdt00()()()CttGtEDt hand the fractional order 𝛼can represent a system which is devoid of any memory, i.e. a memory-less
system, when the order is zero, but on the other hand, it can be the representation of a full-memory
system.
It is indicated that the fractional models are a valuable tool for describing the dynamic properties of real
materials, specifically, in polymers which use for controlling the sounds and vibrations [22-33]. Also,
the combination of fractional models with other linear elements can make intriguing structures; for
example, it can be referred to fractional Maxwell and Kelvin models which are known as spring-spot
models (Table 2). Since SLEs are not always precisely successful in investigating many viscoelastic
systems, researchers applied spring-pot models to address the problem, as well as by virtue of these
fractional models and the combination of them, they achieved the astonishing results [33-42]-to more
details see the "Application" section.
Table 3. Fractional order Voigt Models.
Model
Constitutive equation
Table 4. Fractional order Maxwell Models
Model
Constitutive equation
Relaxation modulus (𝑮(𝒕))
Relaxation modulus (𝑮(𝒕))
Assuming the fractional Maxwell model which made of two FOE models with (𝛼, 𝐸1, 𝜏1) and (𝛽, 𝐸2, 𝜏2)
elements, where 0 < 𝛼, 𝛽 < 1. Regarding this, the constitutive equation, in relaxation state, will be as
follow [to prove, see Appendix]:
(46)
20
11()()()tEDtEt11(1)EtE11()()()tEDtDt11()(1)Ett1122()()()tEDtEDt1112(1)(1)EEtt()()()tDtEDt1tEE112()()()tDtDt11,1ttEE()()()tDtEDt,1ttEE,1()ttGtEE Where 𝐸𝛼,𝛽 is Mittag-Lefler function [to see the definition of this function, see Appendix]. Above
equation will be reduced to the relaxation modulus of classical Maxwell by assuming 𝛼 = 0 and 𝛽 = 1.
In other words, this equation includes not only the classical relaxation state but also further expansion
as compared to the SLEs (more fractional models is presented in the table 3 and 4).
Periodic strain
As previously detailed, by imposing a periodic strain upon a viscoelastic object, stress is neither exactly
in the same phase with strain nor 90o out of phase. In this section, the trueness of this fact is investigated
for a FOE model. Assuming a FOE is subjected to a sinusoidal strain, i.e, 𝜀 = 𝜀0sin (𝜔𝑡), then the stress
of which will be obtained as follow by the help of equation (39) and Caputo derivative (15).
As a result, the following equation hold [to prove the equation, see Appendix].
(74)
)
(74
In equation (74), if 𝛼 = 1, stress will be 90o out of phase with strain as a pure viscous object, and if
𝛼 = 0, it will be exactly in the same phase with strain as a pure elastic object. Therefore, when 0 <
𝛼 < 1, then stress will neither exactly in the same phase with strain nor 90o out of phase. Hence, the
result of equation (74) thoroughly matches with physical interpretations.
Relation between fractional equations and SLE
In this section, with the help of a new fractal system which made of quite a few springs and dampers
(Fig 6), the relation between fractional equations and SLEs is investigated.
Increasing the number of elements in the SLEs is one approach to improving the accuracy of a model
in the viscoelastic systems. Calculation of and work with such sophisticated systems are not usually
easy. Thus, in order to investigate the constitutive equation in infinite systems, which made of many
both springs and dampers, different models were derived so far, and their outputs were reduced to
fractional equations [24,48,49]. Regarding this, a new fractal system is considered and after massive
mathematical calculation, it reveals that the constitutive equation is equal to a spring-pot11.
Consider the spring-damper circuit in Fig 6, so equation (74) will be obtained with the help of governing
equations on the stress/strain elements.
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
= 𝜂 0(1 −
1+
𝐸0
𝜂 1
(1+
𝜂 1
𝐸1
(1−
1
1
1+
𝐸1
𝜂 2
(1+
𝜂 2
𝐸2
(1−
1
⋮
1+
𝐸𝑛−2
𝜂 𝑛−1
(1+
𝜂 𝑛−1
𝐸𝑛−1+𝜂 𝑛
)
)
) (74)
)
Where 𝜀(𝑠) = 𝐿{𝜀(𝑡)} , 𝜎(𝑠) = 𝐿{𝜎(𝑡)}.
11 From the mathematical point of view, by imposing a local force to the proposed system, an equivalent topology for investigating the relationship between stress
and strain over time is found based on the mathematical formulae. In other words, the govern equation in figure 9 is in R space with the Euclidean norm, while
the output of the assumed system is located in 𝐿1 space.
𝐿1 ≔ {𝑓: [𝑎, 𝑏] → ℝ; 𝑓 𝑖𝑠 𝑚𝑒𝑎𝑠𝑢𝑟𝑎𝑏𝑙𝑒 𝑜𝑛 [𝑎, 𝑏] 𝑎𝑛𝑑 ∫ 𝑓(𝑥)1𝑑𝑥 < ∞
𝑏
𝑎
}
21
10cos()()()(1)tattdE10()sin()exp(21)22titiE Based on equation (05), after rewriting equation (74) to (05), the following equation is obtained12.
𝑥(𝑥 + 1)𝛼−1 =
𝑥
(1−𝛼)𝑥
1+
1+
𝑥
1.(0+𝛼)
1.2
1+
1.(2−𝛼)
2.3
1+
𝑥
2.(1+𝛼)
3.4
1+
𝑥
2.(3−𝛼)
4.5
1+
𝑥
… (05)
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
= 𝜂0 (1 −
𝑐0/𝑠
(1−𝛾)𝑐0/𝑠
1+
1+
⋯
(𝑛−1)(𝑛−𝛾)
(2𝑛−1)(2𝑛−2)
𝑐0/𝑠
1
) (02)
Where 0 < 𝑥, 𝛼 < 1 and 𝑐0 = 𝐸0/𝜂0. The equation (05) will be reduced to the following equation by
assuming 𝑐0 ≪ 𝑠 when 𝑛 → ∞.
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
≈ 𝑠𝜂0(1 − (𝑐0/𝑠)𝛾 )
Fig. 6 Diagram of the finite mechanical arrangement applied to simulate the generalized damper.
By assuming 𝐸0 = 0 and rewriting the above equation, we have [to follow the proof in detail, see
Appendix]:
𝜎𝑑(𝑡) = 2𝜂0
𝑑𝜀𝑑(𝑡)
𝑑𝑡
− 𝜂0
𝑑1−2𝛾𝜀𝑑(𝑡)
𝑑𝑡1−2𝛾
(07)
In other words, the output of the considered system is a combination of a FOE and damper connecting
in a series way when 0 < 𝛾 <
1
. Also, based on the achieved result, derivative in the above equation is
2
Caputo.
12 Notations
𝑏1 +
𝑏2 +
𝑎1
𝑎2
𝑎3
𝑏3+ ⋱
+
=
𝑎1
𝑏1 +
𝑎2
𝑏2 +
… … …
𝑎𝑛
𝑏𝑛
𝑎𝑛
𝑏𝑛
22
Application
As the abovementioned sections, fractional calculus is a powerful tool for modeling complex systems,
specifically for viscoelastic materials. In this section, at first, a brief review of the application of
fractional calculus in the modelisation of viscoelastic materials in the variety fields are presented. In the
next step, a one-dimensional fractional model for investigating a cell deformation in creeping state is
proposed, and then a 2D fractional model for simulating the network of actin filaments on the cellular
surface is proposed. Owing to the perfect match with experimental data and simulated model, we
logically conclude that fractional models can be a great replacement for previous models, which are
made of spring and damper, in this regard.
Fractional calculus has been employed for modelling viscoelastic systems which cover many fields and
subjects. In ref [43], Djordjević, V.D., et al. utilized fractional calculation in four parametric model
with equation (05) to investigate a cell's viscoelasticity in the range of 10−2 − 102𝐻𝑧 for a periodic
test. The output of this model was compared with experimental data obtained from the magnetic
oscillatory cytometry method, and it revealed that the proposed model highly matched with
experimental data.
)
(05
It is noteworthy Riemann-Liouville (fractional) derivative was considered in equation (05).
Investigation of asphalt mixtures' property during their service life is a real challenge owing to its
complexity and sensitivity to environmental and loading conditions. It has been shown that asphalt
mixtures behave as linear viscoelastic materials when it is subjected to loading conditions.
Traditionally, the viscoelastic material is modeled via creep/recovery functions by the help of spring
and damper models. The output of these models have shown an exponential behavior, in nature,
however, the mechanical response of these materials have a power-law behavior rather than an
exponential. Considering this fact, spring-pot models were proposed for predicting creep/recovery
behavior of asphalt mixtures. Finally, the result of proposed models were compared with experimental
data, showing that they have in good agreement with the data [39].
Scholars believe that some pulmonary diseases are related to lung viscoelasticity. Thus, profound
understanding of viscoelastic models in this field can bring about a new progress in lung pathology and
trauma. As a result, stress relaxation test was applied on a pig's lung, and the output of this test showed
a power-law behavior than an exponential. Regarding this, two integer standard linear solid-
generalized Maxwell models-and a Fractional standard linear solid model were proposed for
predicting the mechanical behavior of a system. The result of this investigation indicated that the
fractional model by far better matches with experimental data than integer models [41].
To investigate the viscoelastic response of human breast tissue cells, the fractional Zener model was
recently proposed [44]. The authors demonstrated that the proposed fractional model has a better result
as compared to that of the non-fractional models showed for probing the mechanical response in
relaxation state.
According to the recent studies [42], fractional Maxwell and Kelvin model were proposed for analyzing
and interpreting the mechanical response of polymer materials, and the result of which have shown that
the fractional Maxwell model was not able to investigate the polymer behavior. On the other hand, the
fractional Kelvin model is a good proper model with this aim [42].
In 2008, a fractional model was utilized to investigate the arterial viscoelasticity. In that research, the
output of the fractional model and SEL model was pinpointed by the help of least-squire. The result
23
012()()()()tataDtat that validation showed that the fractional model is able to analyze the system mechanical response to
tensile perfectly [35].
Two-dimension fractional network model
In the last section of the paper, to investigate the mechanical response of a cell, two (new) fractional
models are suggested. In the first step, a one-dimension fractional model is proposed, and in the second,
a more realistic model is put forward in two dimensions by extending the first model.
We define a cell as the smallest vital unit of the configuration of any live existence. The cell is made of
different parts, such as cytoplasm, nucleus, cytoskeleton (CSK), etc. The CSK plays a remarkably
essential role in the intracellular force transmission, cellular contractility, as well as the structural
integrity of cells not only in the static but in dynamic states as well. The biological functions of cells,
such as growth, differentiation, and apoptosis are associated with changes in the cell shape, are related
to the mechanical behavior of the CSK [45-47]. The CSK structures are composed of three main
elements: actin filaments, intermediate filaments, and microtubules. The structural changes in the CSK,
including deformation and rearrangement, are related to changes in the tensile forces that are applied to
the actin filaments [45, 46, 48-50]; as a result, the actin filaments have mainly tension-bearing roles.
The viscoelastic properties of cells have been found to play an important role in distinguishing
diagnostic, stem cells and etc. Regarding this, several two-dimensional (2D) and three-dimensional
(3D) CSK network models have been proposed to calculate the mechanical response of cells [56,58-
61].
The viscoelasticity of a cell under some specific external forces was measured with the help of Magnetic
Bead Microrheometry methods in creeping state with 𝐹 = 1100𝑝𝑁 [51]. To investigate the gained
results, the author utilized a model which made of two springs and two dampers (Fig 7). The result of
this model is compared with the experimental data (Fig 8).
Fig. 7. Bausch model has been proposed to investigate a cell viscoelasticity under deformation [51].
The mechanical response of a living cell is quite intricate. The complex, heterogeneous characteristics
of cellular structure cause that simple linear models of viscoelasticity cannot predict the mechanical
properties of cells under the deformation quite well, especially under minuscule deformation. Regarding
this, fractional models are applied to address the point mentioned. In the first step, a FOE model with
Eq. 39 is proposed for modeling a cell deformation in the creeping state. To validate the proposed
model, the model result is compared with the experimental data which expressed in the ref [51]. In order
to make the same condition as experimental data, after solving equation (39) by regarding the creep
test, equation (22) is obtained. According to Fig 8, it is clear that the FOE model is a great agreement
with the experimental data. By the help of simulated annealing optimization methods, the best
coefficients-𝛼, 𝐸, 𝑎𝑛𝑑 𝜂-in equation (22) is gained as well (Table 5). Hence, it looks logical to
conclude that the fractional model can be a suitable replacement for modeling the mechanical properties
of a cell than the Bausch model.
24
Fig 8. Comparison between the FOE model and the Bausch model. The red points and the black line represent the experimental and the
response of the FOE models respectively. Which the right side related to the FOE model and the left side represents the Bausch model.
Table 5. The optimized coefficients of the FOE model which obtained by annealing optimization methods.
Fractional order (𝛼)
Young's modulus (Pa)
Shear modulus (𝑃𝑎. 𝑠−𝛼)
MSE
0.112889249504105
0.003456900326295
1.207297449819304e-6
69.1723
To investigate the models' error with experimental data, normalized root-mean-square deviation
(NRMSD) method is applied.
Also, NRMSE was calculated 10.21% and 3.16% for Bausch model and FOE model by the help of
above formula respectively, showing the fractional model is highly accurate for probing and predicting
the mechanical properties of a subjected cell.
Fig 9. A schematic representation of the preposed network model. Which the black points are fixed and, the others are dynamic. Also, the
lines between two points represent the actin filaments connecting by cross-linking protein (in cell cortex).
In the next step, a two-dimension fractional model, so-called 2D fractional network model, is proposed
to investigate the cell viscoelasticity. The cell can deal with the external and intercellular forces imposed
on it by virtue of CSK. Regarding this, these protein filaments, CSK, is considered as a started point of
making a 2D model.
In most cells, actin filaments are intensely concentrated in a layer just beneath the plasma membrane,
even though they are found throughout the cytoplasm of a eukaryotic cell. In this region, which is called
25
maxmax100MSENRMSDyy the cell cortex, these filaments are linked by actin-binding proteins into a meshwork that gives
mechanical strength to cells and supports the outer surface of them as well. As a simplification, without
loss of generality, imaging the whole mechanical properties of a cell is related to actin filaments [52].
Considering this, we can propose a dynamic network model, shown in Fig. 8, for probing the cell
mechanical behavior.
In 2003, a model as the proposed network model in this paper was suggested predicting the mechanical
response of a cell [53]. The mechanical properties of actin filaments were approximated with pure
elastic materials based on an experimental paper [54]; however, according to the experimental data,
these filaments have shown an elastic response in millisecond scale under deformation. Neglecting this
scale as a simplification may cause a remarkable error in long time simulation. In other words, the
deformation of any material under minuscular stress is divided into three phases which the first one is
related to the elastic limit. Also, the mechanical response of an actin filament was reported in this phase,
and by overgeneralizing this point, a pure elastic network model-which every actin was approximated
with a spring-was proposed in the scale of sec for probing the whole mechanical behavior of a cell
[53]. Consequently, the elastic network model, so-called pre-stressed cable network model, was failed
to predict the experimental data from bead micrometry method [52]. Regarding this, a new model with
FOE elements is suggested in this paper. Two significant merits of which, connected to the system
memory and the viscosity term in arbitrary scale time, are embedded in proposed model; resulting the
fractional model is able to highly match with the experimental data. The fractional network model is
based on the principle that actin filaments, according to ref [52, 55], have shown viscoelastic behavior.
Therefore, the FOE model can be a proper model to investigate the mechanical response of each actin
filament. By expanding the FOE model into a 2D model with the same order to each element, the
fractional network model, which depicts in the Fig. 9, will be obtained.
Fig 10. A schematic representation of 2D dynamic fractional simulated network.
Since the direct observation of a 𝑙0 (the length of an actin filament) is not exist from a living cell, 𝑙0 is
approximated from CSK pore (𝐷𝑝) in an adhere cell based on equation 𝐷𝑝 = √3𝑙0/3 as paper [53]. In
other words, based on the observations, the value of 𝐷𝑝 is considered in the middle of experimental
data-that is Dp = 100nm; resulting l0 = 173.2nm. The space of network is approximated 4.5464 ×
106nm2 based on 𝑙0, which is almost same with the reported value in the ref [53].
To validate the simulation result and to compare it with the previous model, we simulate the FOE and
spring network, Fig 11 show the behavior of these two model under the same stress. Considering the
26
results, the fractional model is better matched with the experimental data. Hence, it seems that we could
conclude that 2D fractional network model is a good candidate for modeling and predicting CSK in
cells.
Fig 11. A comparison between the 2D dynamic fractional simulated network, the left side figure, and the elastic model which represented in
ref [53], the right side figure. Experimental data is obtained from ref [51].
Discussion
The fractional calculus is a powerful tool for describing the complex physical systems which have long-
term memory and long-term spatial interactions. In this paper, different aspects of this powerful tool
are introduced through a concise literature review to provide the reader with a picture of what fractional
calculus is. Then, with the help of developed concepts, it is sought to find a relation between fractional
and viscoelastic systems and to employ the obtained relation for investigating the mechanical property
of a cell.
In the first step, to investigate the mechanical response of elastic, viscous, and viscoelastic materials,
scholars typically apply spring, damper models and some combination of them respectively. In nature,
pure elastic or viscous systems can rarely found and moreover, the mechanical behavior of them are
modelled by the help of the Hook's law and Newtonian fluid. Regarding a minuscule deformation, the
response of a material is modelled by these laws, thereby making some errors in both the simulation
and the calculation of theoretical models-specifically, bio-material such as cells and tissues. In spite
of neglecting the arisen errors in some cases, scholars are usually making a highly complex model for
improving the accuracy and reducing the error in minute systems. Considering this fact, proposing a
new method to cope well with the sophisticated problem can be a worthwhile tool. Based on the
investigation which is done in this paper, it is indicated that fractional calculus is a suitable candidate
to this end. To prove this claim, a fractal system, made of the combination of quite a few springs and
dampers on a large scale, was considered, as well as the output of the proposed system was
approximated only with the simple fractional model.
In the next step, a version of fractional models, known as FOE, was employed to model the mechanical
response of a cell in one-dimension. By virtue of the achieved result, a 2D model so-called 2D fractional
network model was proposed to investigate the mechanical response of actin filaments, exist in CKS
and cover a cellular outer layer. The creeping state was considered in both models, and the result of
which was validated by not only the experimental date but previous models as well. The results revealed
that in the one-dimensional model, data variance relative to a previous model presented in ref [51] was
about 11%, while in the FOE model, it was estimated 3%. In other words, the FOE model precision has
evaluated 97% in comparison with experimental data. Furthermore, the simulation of two-dimension
elastic lattice, represented in ref [53], for actin filaments on the cell surface, is totally invalid in
27
Magnetic Bead Microrheometry method. Meanwhile, the 2D fractional network model has able to
analyze same experimental date precisely and to consider the viscosity term of actin filament as well.
The fractional order 𝛼 in both models is about 0.11. Hence, not only does the investigated model shows
more elasticity behavior, but it also has high memory. As a result of done research the paper and many
other studies, applying the fractional tool to investigate and to model the mechanical response of
phenomena which shows power-law behavior is more efficient than the SLE models.
Appendix
To facilitate the context for readers, the proof of some equations presenting in the text is presented in
this part.
The proof of Equation 46:
According to equation (39)
where the initial value of 𝜀1(𝑡) and 𝜀2(𝑡) is assumed zero. Considering 𝜀 = 𝜀1 + 𝜀2 based on Maxwell
model, we have
In addition, 𝜎 = 𝜎1 = 𝜎2 and 0 < 𝛼, 𝛽 < 1 then
Now let 𝜏𝛼−𝛽 = 𝐸1𝜏1
𝛽
𝛼/𝐸2𝜏2
and
, so
Taking the Laplace transform and applying the unit step strain function on the above equation, we
finally gain the following equation
Hence, the following equation by utilizing the inverse Laplace on the above equation is achieved.
The proof of Equation 48:
By taking inverse Laplace from equation (47) and by the help of convolution operation (see table
B3.1), the following equation is obtained
according to the table (B3.1)
By taking inverse Laplace from the above equation, we have
28
12121122ddanddtEdtE12112211()ddtEdtEdt11211122()()EdtdtEEdtdt11(/)EE()()()dtdttEdtdt1()(1/)Esgss,1()ttGtEE10()().(){},(){cos(1)(())}FsGsstFssELtGsLt1202()ssEs
In order to
solve the above equation, it is enough
to
calculate
. Thus, according to the
(B3.2) equation
By assuming ℎ(𝑧) = 𝑧𝛼𝑒 𝑧𝑡/(𝑧2 + 𝜔2) and Cauchy theorem [56]
Where 𝑧1 = 𝑧2 = ±𝑖𝜔 are non-analytical points of function ℎ(𝑧), resulting
And also, we have
According to (50), (56), and equations (57)
29
11220(){}(54)stELs122{}sLs122221{}()2cistcissLedssis11222{}()()sLhzhzs121(1)21((1))((31))22()()()()22()((1))2()2()(55)2ititititiitiitititiihzhzeeiiieeieeeeee((1))((31))22()cos()cos((1))sin()sin((1))cos()cos((31))sin()sin((31))2222sin()cos((1))cos()sin((1))(cos()sin((31))sin()cos((31)))2222ititeettttittttcos()cos((1))cos((31))sin()sin((1))sin((31))2222sin()cos((1))cos((31))cos()sin((1))sin((31))22222cos()sin((21))sin22ttittt2sin()sin((21))cos222sin()cos((21))cos2cos()sincos((21))22222sin((21))cos()sinsin()cos2222cos((21))sin()coscos(22tittttitt)sin22sin((21))sin()2cos((21))sin()22222sin()sin((21))cos((21))2222sin()exp(21)(56)22tittiiti10()sin()exp(21)22titiE The proof of Equation 52:
According to figure (6) the following equations hold.
𝑑𝑡
(65)
(62)
𝑠 ∶ 𝑘 = 0,1, … . , 𝑛 − 1 (57)
∶ 𝑘 = 0,1, … . , 𝑛 − 2 (58)
(54)
(65)
(65)
𝑑 + 𝜎𝑘
𝑑 = 𝜎𝑘+1
𝜎𝑘
𝑑
𝑠 + 𝜀𝑘+1
𝑠 = 𝜀𝑘+1
𝜀𝑘
𝑠 + 𝜀0
𝑑
𝜀 = 𝜀0
𝑠 = 𝜀𝑛
𝑑
𝜀𝑛−1
𝑑
𝜎 = 𝜎0
Regarding Hook's law and Newtonian fluid, we have
𝑑𝜀𝑑(𝑡)
𝜎𝑑(𝑡) = 𝜂
𝜎 𝑠(𝑡) = 𝐸𝜀 𝑠(𝑡)
where 𝜎𝑑،𝜀𝑑 ،𝜎 𝑠 and 𝜀 𝑠related to stress and strain in damper as well as stress and strain in spring
respectively. Accordingly, the following equations is obtained by taking Laplace transform from
equations (62) and (65).
𝜎𝑑(𝑠) = 𝜂𝑠𝜀𝑑(𝑠)
𝜎 𝑠(𝑠) = 𝐸𝜀𝑠(𝑠)
Based on equation (54)
𝑑(𝑡) = 𝜎𝑘+1
𝜎𝑘
𝑑(𝑡)
𝑑𝜀𝑘
(6)
→ 𝜂𝑘
𝑑𝑡
(9)
𝑑 (𝑠) + 𝜎𝑘
𝑑(𝑠) = 𝜂𝑘+1𝜀𝑘+1
→ 𝜂𝑘𝑠𝜀𝑘
(9)
𝑑 (𝑠)
𝑠(𝑠)
𝜀𝑘+1
𝜎𝑘
=⏞
𝑑(𝑠)
𝑑(𝑠)
𝜀𝑘
𝜀𝑘
(67)
(60)
𝑠(𝑡)
𝑑 (𝑡)
𝑑𝜀𝑘+1
𝑑𝑡
= 𝜂𝑘𝑠 − 𝜂𝑘+1𝑠
𝑑 (𝑡) + 𝜎𝑘
= 𝜂𝑘+1
𝑑 (𝑠)
𝜂𝑘𝑠 −
𝑠(𝑠)
𝑠(𝑡)
+ 𝜎𝑘
𝜎𝑘+1
1
𝜂𝑘𝑠
(
) 𝜎𝑘
𝑑(𝑠)
= 𝜂𝑘𝑠 −
Thus, we have
1
𝑑(𝑠)
𝜎𝑘
𝑑 (𝑠)
𝜎𝑘+1
(
1
𝜂𝑘𝑠
)
(1)
) =⏞
𝜂𝑘𝑠(1 −
1
𝑠 (𝑠)
𝜎𝑘
𝑑 (𝑠)
𝜎𝑘+1
1+
)
(66)
𝑠(𝑠)
𝜎𝑘
𝑑(𝑠)
𝜀𝑘
= 𝜂𝑘𝑠(1 −
1
𝑑(𝑠)
𝜎𝑘
𝑑 (𝑠)
𝜎𝑘+1
In addition, we have
(2)
=⏞
(10)
=⏞
𝑠 (𝑠)
𝑑 (𝑠)
𝐸𝑘𝜀𝑘
𝜎𝑘+1
𝑠(𝑠)
𝜎𝑘
𝑑 (𝑠)
𝜎𝑘+1
𝐸𝑘 (
𝑠
𝜀𝑘+1
𝑑 (𝑠)
(𝑠)+𝜀𝑘+1
𝑑 (𝑠)
𝜎𝑘+1
)
(9)
=⏞
𝐸𝑘
𝜂𝑘+1𝑠
𝑠
𝜀𝑘+1
(
𝑑 (𝑠)
(𝑠)+𝜀𝑘+1
𝑑 (𝑠)
𝜀𝑘+1
)
Therefore, the following equation holds
(10)
=⏞
𝑠
1
𝐸𝑘
𝜂𝑘+1𝑠
(1 +
𝑠(𝑠)
𝜎𝑘
𝑑 (𝑠)
𝜎𝑘+1
If we put equation (64) into equation (66), we will obtain
𝑠(𝑠)
𝜎𝑘
𝑑(𝑠)
𝜀𝑘
(𝑠)
𝜎𝑘+1
𝑑 (𝑠)
𝜀𝑘+1
= 𝜂𝑘𝑠(1 −
𝐸𝑘+1
𝐸𝑘
)
1
1
)
)
(1+
1+
𝜂𝑘+1𝑠
𝐸𝑘+1
𝑠
(𝑠)
𝜎𝑘+1
𝑑 (𝑠)
𝜀𝑘+1
(64)
(64)
In the step 𝑛𝑡ℎ, the following equation will be got
30
𝑠
𝜎𝑛−1
𝑑
𝜀𝑛−1
(𝑠)
(𝑠)
(9)
=⏞
𝜂𝑛−1𝑠
1
𝜎𝑛−1
𝜎𝑛−1
𝑑 (𝑠)
𝑠
(𝑠)
(
(1)
=⏞
𝜂𝑛−1𝑠
(
1 +
)
1
𝑑(𝑠)
𝜎𝑛
𝑠
𝜎𝑛−1
(𝑠))
(9,10)
=⏞
𝜂𝑛−1𝑠
(
1 +
1
𝜂𝑛𝑠
𝐸𝑛−1
(4)
=⏞
𝑑(𝑠)
𝜀𝑛
𝑠
𝜀𝑛−1
(𝑠))
𝜂𝑛−1𝑠
𝜂𝑛𝑠
𝐸𝑛−1
(1 +
)
𝑠
=
)
(1+
𝜂𝑛−1𝑠
𝜂𝑛𝑠
𝐸𝑛−1
Thus
(𝑠)
𝜎𝑛−1
𝑑
(𝑠)
𝜀𝑛−1
Assuming 𝜂𝑘𝑠 = 𝜂𝑘 to simplification, so based on equation (64) and (64), in step (𝑛 − 1)𝑡ℎ we will
have
𝜎𝑛−2
𝑑
𝜀𝑛−2
(45)
= 𝜂 𝑛−2(1 −
(41)
(𝑠)
(𝑠)
1
𝑠
)
)
1+
𝐸𝑛−2
𝜂 𝑛−1
(1+
𝜂 𝑛−1
𝜂 𝑛
𝐸𝑛−1
𝐸𝑛−1(
+1)
By rewriting the above equation
𝜎𝑛−2
𝑑
𝜀𝑛−2
= 𝜂 𝑛−2(1 −
(𝑠)
(𝑠)
1
𝑠
(1+
1+
𝐸𝑛−2
𝜂 𝑛−1
𝜂 𝑛−1
𝐸𝑛−1+𝜂 𝑛
)
)
(42)
In similar manner, we obtain
𝜎𝑛−3
𝑑
𝜀𝑛−3
= 𝜂 𝑛−3(1 −
(𝑠)
(𝑠)
𝑠
1+
𝐸𝑛−3
𝜂 𝑛−2
(1+
1
𝜂 𝑛−2
𝐸𝑛−2
(1−
1
1+
𝐸𝑛−2
𝜂 𝑛−1
(1+
𝜂 𝑛−1
𝐸𝑛−1+𝜂 𝑛
)
)
)
(43)
The following equation is achieved in step (𝑛 − 4)𝑡ℎ by the help of above approach
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
= 𝜂 0(1 −
1+
𝐸0
𝜂 1
(1+
𝜂 1
𝐸1
(1−
1
1
1+
𝐸1
𝜂 2
1+
𝜂 2
𝐸2
(1−
(
1
⋮
1+
𝐸𝑛−2
𝜂 𝑛−1
(1+
𝜂 𝑛−1
𝐸𝑛−1+𝜂 𝑛
)
)
)
)
) (44)
NB: Donating the following symbols to simplification are considered
𝑏1+
𝑏2+
𝐸𝑖−1
𝜂 𝑖
𝐴 =
𝑎1
𝑏1+
𝑎2
𝑏2+
… … …
𝑎𝑛
𝑏𝑛
𝑎1
𝑎2
𝑎3
𝑏3+ ⋱
+
𝑎𝑛
𝑏𝑛
=
= 𝑎𝑖 ,
𝜂 𝑖
𝐸𝑖
= 𝑏𝑖
1
1+𝑎1(1+𝑏1(1−
1+𝑎2(1+𝑏2(1−
1
1
⋱
(45)
(46)
)
)
1+𝑎𝑛−1 (1+
𝜂 𝑛−1
𝐸𝑛+𝜂 𝑛
)
Thus
A =
1
a1(1+b1)
[1+a1(1+b1)]−
[1+a2(1+b2)]−
…
an−2(1+bn−2)
1+
an−1 (1+
η n−1
En+η n
1
)
(47)
Let us to consider 𝑎𝑖(1 + 𝑏𝑖) = 𝑑𝑖, by rewriting equation (46), we obtain
𝐴 =
1
𝑑1
[1+𝑑1]−
𝑎𝑛−1η n−1 (𝐸𝑛+η n)−1
[1+𝑑2]−
To take factor form term 1/𝑠, assuming
1
…
𝑑𝑛−2
1+
𝑎𝑛−1 (1+
1
η n−1
𝐸𝑛+η n
)
=
[1+𝑑1]−1
𝑑1(−[1+𝑑2])−1
1+
1+
…
𝑑𝑛−2(1+𝑎𝑛−1)−1
1+
=
(78)
31
𝐸𝑖−1
𝜂𝑖
𝐸𝑖−1
𝐸𝑖
, ℎ𝑖 =
𝑎 𝑖 =
Thus
1
𝑑𝑖 =
Based on equation (76), the equation (44) is rewritten as follow
𝑎 𝑖 + ℎ𝑖 , 𝑎𝑖 =
𝑎 𝑖
1
𝑠
𝑠
(79)
(80)
1
[1 + ℎ1 +
𝑠
1 +
1
𝑠
(ℎ𝑛−2 +
𝐴 =
−1
𝑎 1]
(ℎ1 +
1
𝑠
𝑎 1) (− [1 + ℎ1 +
−1
1
𝑠
𝑎 2])
…
1 +
−1
1
𝑠
𝑎𝑛−1)
𝑎 𝑛−2) (1 +
𝑎 𝑛−1𝜂𝑛−1 (𝐸𝑛 + 𝑠𝜂𝑛)−1
1 +
[𝑠(1 + ℎ1) + 𝑎 1]−1
1
1
𝑠
(𝑠ℎ1 + 𝑎 1) (− [1 + ℎ1 +
1
𝑠
=
−1
1
𝑠
𝑎 2])
…
1
𝑠
(𝑠ℎ𝑛−2 + 𝑎 𝑛−2) (1 +
1
𝑎 𝑛−1𝜂𝑛−1 (
𝑠
1
The following expansion holds according to ref [57].
𝑎𝑛−1)
1 +
1
𝑠
−1
1 +
1
𝑠
1 +
−1
𝐸𝑛 + 𝜂𝑛)
(81)
𝑥
(1−𝛾)𝑥
1.(0+𝛾)𝑥
1+
1+
1.2
1+
…
𝑥(𝑥 + 1)𝛾−1 =
The coefficient of equation (81) is chosen in a way that the equation (82) holds.
[𝑠(1 + ℎ1) + 𝑎 1]−1 = 𝑐0 ,
(𝑠ℎ1 + 𝑎 1) (− [1 + ℎ1 +
𝑎 2])
Based on equation (63)
= (1 − 𝛾)𝑐0 , 𝑒𝑡𝑐
−1
1
𝑠
(82)
(83)
= 𝜂0(1 −
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
Now let 𝑛 → ∞ then
𝑐0/𝑠
1 +
(1 − 𝛾)𝑐0/𝑠
1 +
…
(𝑛 − 1)(𝑛 − 𝛾)
(2𝑛 − 1)(2𝑛 − 2)
𝑐0/𝑠
1
)
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
= 𝜂0(1 − (
𝑐0
𝑠
) (
𝑐0
𝑠
𝛾−1
+ 1)
)
Therefore
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
≈ 𝑠𝜂0(1 − (
𝛾
𝑐0
𝑠
)
)
(84)
Before applying Laplace transform, the left side of the equation (84) must depend on only a damper.
Thus, the following calculation is done to this end.
(1)
=⏞
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
𝑑(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
−
𝑑(𝑠)
𝜎1
𝑑(𝑠)
𝜀0
(85)
Also
(9)
=⏞
𝑑(𝑠)
𝜎1
𝑑(𝑠)
𝜀0
𝑠𝜂0
(1)
=⏞
𝑠𝜂0
1
𝑑(𝑠)
𝜎0
𝑑(𝑠) )
𝜎1
(
(
1 +
1
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜎1
)
(10)
=⏞
𝑠𝜂0(
1 +
Considering
𝑑(𝑠)
𝜎1
𝑑(𝑠)
𝜀0
= 𝑦
As a result
𝑦 = 𝑠𝜂0(
1
1 +
)
𝐸0
𝑦
)
1
𝐸0
𝑑(𝑠)
𝜎1
𝑑(𝑠)
𝜀0
32
By putting equation (86) into equation (85), we will obtain
𝑦 = 𝑠𝜂0 − 𝐸0
𝑑(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
Hence, based on equation (84) we will have
=
𝑠(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
+ 𝑠𝜂0 − 𝐸0
(86)
By rewriting the above equation
𝑑(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
= 𝑠𝜂0(1 − (
𝛾
)
) + 𝑠𝜂0 − 𝐸0
𝑐0
𝑠
𝑑(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
= 2𝑠𝜂0 − 𝑠𝜂0(𝑐0/𝑠)𝛾 − 𝐸0
ز
(87)
By the help of equation (83), we obtain
𝑑(𝑠)
𝜎0
𝑑(𝑠)
𝜀0
= 2𝑠𝜂0 − 𝑠𝜂0(𝑠(𝑠(1 + ℎ1) + 𝑎 1))−𝛾 − 𝐸0
In other words
𝜎𝑑(𝑠)
𝜀𝑑(𝑠)
= 2𝑠𝜂0 − 𝐸0 − 𝜂0𝑠1−𝛾(𝑠(1 + ℎ1) + 𝑎 1)−𝛾
(88)
By the use of binomial distribution function for every arbitrary 𝛽, the following equation holds.
1
∞
(1 − 𝑧)𝛽+1 = ∑ (
𝑘=0
𝑘 + 𝛽
𝑘
) 𝑍𝑘
Thus
(𝑠(1 + ℎ1) + 𝑎 1)−𝛾 =
1
(𝑠(1+ℎ1))
𝛾
+(1−(−𝑎 1/𝑠(1+ℎ1))
𝛾 =
1
(𝑠(1+ℎ1))
𝛾 ∑
∞
𝑘=0
(
𝑘 + 𝛾
𝑘
) (−1)𝑘 (
𝑎 1
𝑠(1+ℎ1)
𝑘
)
(89)
The following equation will have achieved by putting equation (89) into equation (88), Therefor
𝜎𝑑(𝑠)
𝜀𝑑(𝑠)
= 2𝑠𝜂0 − 𝐸0 − 𝜂0𝑠1−2𝛾
1
(1+ℎ1)𝛾 ∑
∞
𝑘=0
(
𝑘 + 𝛾
𝑘
) (−1)𝑘 (
𝑎 1
)
𝑠(1+ℎ1)
𝑘
(90)
By the use of Laplace transform from the above equation, we obtain
𝜎𝑑(𝑡) = 2𝜂0 (
) − 𝐸0𝜀𝑑(𝑡) − 𝜂0
𝜂0
1
(1+ℎ1)𝛾 ∑
∞
𝑘=0
(
) (−1)𝑘 (
𝑘
)
𝑎 1
1+ℎ1
𝑑𝜀𝑑(𝑡)
𝑑𝑡
𝑘 + 𝛾
𝑘
1
(1+ℎ1)𝛾
𝑑1−2𝛾𝜀𝑑(𝑡)
𝑑𝑡1−2𝛾 +
𝐿−1{𝑠1−2𝛾−𝑘𝜀𝑑(𝑠)}
where 0 < 𝛾 < 1.
Now let us assume that 𝐸0 = 0 in equation (91), resulting in we gain a combination of a damper and
FOE connected to each other in a series way.
(91)
𝜎𝑑(𝑡) = 2𝜂0 (
𝑑𝜀𝑑(𝑡)
𝑑𝑡
) − 𝜂0
𝑑1−2𝛾𝜀𝑑(𝑡)
𝑑𝑡1−2𝛾
Acknowledgment
We would also like to show our gratitude to the Dr. Mahdieh Tahmasebi for sharing their pearls of
wisdom with us during the course of this research, and we thank Miss Mina Moeini for comments that
greatly improved the mathematical proofs in the paper.
33
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
Boyer, C.B. and U.C. Merzbach, A history of mathematics. 2011: John Wiley & Sons.
Dalir, M. and M. Bashour, Applications of fractional calculus. Applied Mathematical Sciences,
2010. 4(21): p. 1021-1032.
Magin, R.L., Fractional calculus in bioengineering. 2006: Begell House Redding.
Herrmann, R., Fractional calculus: An introduction for physicists. 2011: World Scientific.
Diethelm, K., The analysis of fractional differential equations: An application-oriented
exposition using differential operators of Caputo type. 2010: Springer.
Caputo, M., Linear models of dissipation whose Q is almost frequency independent-II.
Geophysical Journal International, 1967. 13(5): p. 529-539.
Podlubny, I., Fractional differential equations: an introduction to fractional derivatives,
fractional differential equations, to methods of their solution and some of their applications.
Vol. 198. 1998: Academic press.
Kilbas, A.A., H.M. Srivastava, and J.J. Trujillo, Preface. 2006, Elsevier.
Grigoletto, E.C. and E.C. de Oliveira, Fractional versions of the fundamental theorem of
calculus. Applied Mathematics, 2013. 4(07): p. 23.
Diethelm, K., Efficient solution of multi-term fractional differential equations using P (EC) m E
methods. Computing, 2003. 71(4): p. 305-319.
Diethelm, K., N.J. Ford, and A.D. Freed, Detailed error analysis for a fractional Adams
method. Numerical algorithms, 2004. 36(1): p. 31-52.
Meerschaert, M.M. and C. Tadjeran, Finite difference approximations for two-sided space-
fractional partial differential equations. Applied numerical mathematics, 2006. 56(1): p. 80-
90.
Jafari, H., C.M. Khalique, and M. Nazari, An algorithm for the numerical solution of nonlinear
fractional-order Van der Pol oscillator equation. Mathematical and Computer Modelling,
2012. 55(5): p. 1782-1786.
Jiang, H., et al., Analytical solutions for the multi-term time–space Caputo–Riesz fractional
advection–diffusion equations on a finite domain. Journal of Mathematical Analysis and
Applications, 2012. 389(2): p. 1117-1127.
Gülsu, M., Y. Öztürk, and A. Anapalı, Numerical approach for solving fractional relaxation–
oscillation equation. Applied Mathematical Modelling, 2013. 37(8): p. 5927-5937.
Bu, W., Y. Tang, and J. Yang, Galerkin finite element method for two-dimensional Riesz space
fractional diffusion equations. Journal of Computational Physics, 2014. 276: p. 26-38.
Bu, W., et al., Finite difference/finite element method for two-dimensional space and time
fractional Bloch–Torrey equations. Journal of Computational Physics, 2015. 293: p. 264-279.
Jin, B., et al., The Galerkin finite element method for a multi-term time-fractional diffusion
equation. Journal of Computational Physics, 2015. 281: p. 825-843.
Markovitz, H., Viscoelastic properties of polymers: By John D. Ferry, Wiley, New York, 1980,
xxiv+ 641 pp., $49.50. 1981, Academic Press.
McCrum, N.G., C. Buckley, and C.B. Bucknall, Principles of polymer engineering. 1997: Oxford
University Press, USA.
Tanner, R.I., Engineering rheology. Vol. 52. 2000: OUP Oxford.
Caputo, M. and F. Mainardi, A new dissipation model based on memory mechanism. Pure
and Applied Geophysics, 1971. 91(1): p. 134-147.
Bagley, R.L. and J. TORVIK, Fractional calculus-a different approach to the analysis of
viscoelastically damped structures. AIAA journal, 1983. 21(5): p. 741-748.
34
Rogers, L., Operators and fractional derivatives for viscoelastic constitutive equations.
Journal of Rheology, 1983. 27(4): p. 351-372.
Bagley, R.L. and P.J. Torvik, On the fractional calculus model of viscoelastic behavior. Journal
of Rheology, 1986. 30(1): p. 133-155.
Koh, C.G. and J.M. Kelly, Application of fractional derivatives to seismic analysis of base‐
isolated models. Earthquake engineering & structural dynamics, 1990. 19(2): p. 229-241.
Makris, N. and M. Constantinou, Fractional-derivative Maxwell model for viscous dampers.
Journal of Structural Engineering, 1991. 117(9): p. 2708-2724.
Friedrich, C. and H. Braun, Generalized Cole-Cole behavior and its rheological relevance.
Rheologica Acta, 1992. 31(4): p. 309-322.
Fenander, A., Modal synthesis when modeling damping by use of fractional derivatives. AIAA
journal, 1996. 34(5): p. 1051-1058.
Pritz, T., Analysis of four-parameter fractional derivative model of real solid materials.
Journal of Sound and Vibration, 1996. 195(1): p. 103-115.
Shimizu, N. and W. Zhang, Fractional calculus approach to dynamic problems of viscoelastic
materials. JSME International Journal Series C Mechanical Systems, Machine Elements and
Manufacturing, 1999. 42(4): p. 825-837.
Rossikhin, Y.A. and M. Shitikova, Analysis of dynamic behaviour of viscoelastic rods whose
rheological models contain fractional derivatives of two different orders. ZAMM‐Journal of
Applied Mathematics and Mechanics/Zeitschrift für Angewandte Mathematik und Mechanik,
2001. 81(6): p. 363-376.
Mainardi, F. and G. Spada, Creep, relaxation and viscosity properties for basic fractional
models in rheology. The European Physical Journal-Special Topics, 2011. 193(1): p. 133-160.
Barpi, F. and S. Valente, Creep and fracture in concrete: a fractional order rate approach.
Engineering Fracture Mechanics, 2002. 70(5): p. 611-623.
Craiem, D., et al., Fractional calculus applied to model arterial viscoelasticity. Latin American
applied research, 2008. 38(2): p. 141-145.
Lewandowski, R. and B. Chorążyczewski, Identification of the parameters of the Kelvin–Voigt
and the Maxwell fractional models, used to modeling of viscoelastic dampers. Computers &
structures, 2010. 88(1): p. 1-17.
Meral, F., T. Royston, and R. Magin, Fractional calculus in viscoelasticity: an experimental
study. Communications in Nonlinear Science and Numerical Simulation, 2010. 15(4): p. 939-
945.
Lewandowski, R. and Z. Pawlak, Dynamic analysis of frames with viscoelastic dampers
modelled by rheological models with fractionalderivatives. Journal of sound and Vibration,
2011. 330(5): p. 923-936.
Celauro, C., et al., Experimental validation of a fractional model for creep/recovery testing of
asphalt mixtures. Construction and Building Materials, 2012. 36: p. 458-466.
Grzesikiewicz, W., A. Wakulicz, and A. Zbiciak, Non-linear problems of fractional calculus in
modeling of mechanical systems. International Journal of Mechanical Sciences, 2013. 70: p.
90-98.
Dai, Z., et al., A model of lung parenchyma stress relaxation using fractional viscoelasticity.
Medical engineering & physics, 2015. 37(8): p. 752-758.
Jóźwiak, B., M. Orczykowska, and M. Dziubiński, Fractional generalizations of maxwell and
Kelvin-Voigt models for biopolymer characterization. PloS one, 2015. 10(11): p. e0143090.
Djordjević, V.D., et al., Fractional derivatives embody essential features of cell rheological
behavior. Annals of biomedical engineering, 2003. 31(6): p. 692-699.
Carmichael, B., et al., The fractional viscoelastic response of human breast tissue cells.
Physical biology, 2015. 12(4): p. 046001.
35
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
Ingber, D.E., et al., Cellular tensegrity: exploring how mechanical changes in the cytoskeleton
regulate cell growth, migration, and tissue pattern during morphogenesis. International
review of cytology, 1994. 150: p. 173-224.
Chicurel, M.E., C.S. Chen, and D.E. Ingber, Cellular control lies in the balance of forces.
Current opinion in cell biology, 1998. 10(2): p. 232-239.
Zemel, A., R. De, and S.A. Safran, Mechanical consequences of cellular force generation.
Current Opinion in Solid State and Materials Science, 2011. 15(5): p. 169-176.
Alberts, B., et al., Essential cell biology. 2013: Garland Science.
STAMENOVIĆ, D. and M.F. COUGHLIN, The role of prestress and architecture of the
cytoskeleton and deformability of cytoskeletal filaments in mechanics of adherent cells: a
quantitative analysis. Journal of Theoretical Biology, 1999. 201(1): p. 63-74.
Maurin, B., et al., Mechanical model of cytoskeleton structuration during cell adhesion and
spreading. Journal of biomechanics, 2008. 41(9): p. 2036-2041.
Bausch, A.R., et al., Local measurements of viscoelastic parameters of adherent cell surfaces
by magnetic bead microrheometry. Biophysical journal, 1998. 75(4): p. 2038-2049.
Chen, T.-J., C.-C. Wu, and F.-C. Su, Mechanical models of the cellular cytoskeletal network for
the analysis of intracellular mechanical properties and force distributions: A review. Medical
engineering & physics, 2012. 34(10): p. 1375-1386.
Coughlin, M.F. and D. Stamenović, A prestressed cable network model of the adherent cell
cytoskeleton. Biophysical journal, 2003. 84(2): p. 1328-1336.
Kojima, H., A. Ishijima, and T. Yanagida, Direct measurement of stiffness of single actin
filaments with and without tropomyosin by in vitro nanomanipulation. Proceedings of the
National Academy of Sciences, 1994. 91(26): p. 12962-12966.
Käs, J., et al., F-actin, a model polymer for semiflexible chains in dilute, semidilute, and liquid
crystalline solutions. Biophysical journal, 1996. 70(2): p. 609-625.
Brown, J.W., Complex variables and applications. 2009.
Jones, W.B. and W.T.C. Fractions, Analytic Theory and Applications. Encyclopedia of
Mathematics and its Applications, 1980.
36
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
|
1806.08001 | 1 | 1806 | 2018-06-20T21:50:54 | Analog control with two Artificial Axons | [
"physics.bio-ph",
"q-bio.NC"
] | The artificial axon is a recently introduced synthetic assembly of supported lipid bilayers and voltage gated ion channels, displaying the basic electrophysiology of nerve cells. Here we demonstrate the use of two artificial axons as control elements to achieve a simple task. Namely, we steer a remote control car towards a light source, using the sensory input dependent firing rate of the axons as the control signal for turning left or right. We present the result in the form of the analysis of a movie of the car approaching the light source. In general terms, with this work we pursue a constructivist approach to exploring the nexus between machine language at the nerve cell level and behavior. | physics.bio-ph | physics | Analog control with two Artificial Axons
Hector G. Vasquez and Giovanni Zocchi∗
Department of Physics and Astronomy, University of California - Los Angeles
The artificial axon is a recently introduced synthetic assembly of supported lipid bilayers and voltage
gated ion channels, displaying the basic electrophysiology of nerve cells. Here we demonstrate the
use of two artificial axons as control elements to achieve a simple task. Namely, we steer a remote
control car towards a light source, using the sensory input dependent firing rate of the axons as the
control signal for turning left or right. We present the result in the form of the analysis of a movie
of the car approaching the light source. In general terms, with this work we pursue a constructivist
approach to exploring the nexus between machine language at the nerve cell level and behavior.
8
1
0
2
n
u
J
0
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
1
0
0
8
0
.
6
0
8
1
:
v
i
X
r
a
Introduction. Action potentials form the machine lan-
guage of the nervous system. While the mechanisms and
characteristics of action potentials have been known for
over half a century, these most interesting excitations
have not been produced in vitro, in a synthetic biology
setting, until very recently [1]. The artificial axon (AA)
[1, 2] is an excitable node which supports action poten-
tials, that is, voltage spikes produced by the same ionic
mechanism as action potentials in neurons. Our aim is
to develop it into a platform for synthetic biology con-
structions which echo neuronal systems. Since the AA
works, as far as electrical excitability, through the same
physical mechanism as the real neuron, it can be endowed
with some of the same capabilities, and also suffers from
similar constraints. Learning to build useful networks
within these constraints is a viable engineering approach
to understanding brain function. Existing constructive
approaches are computational [3], electronic [4, 5], or
based on real neurons, directing their pattern of connec-
tion [6–11], and re-programming stem cells in 3D cultures
[12]. Especially the first two are much more advanced
than what we show here; they also have a longer history,
and, crucially, they are based on electronics. Specifi-
cally, neuromorphic chips (NMCs) are currently compet-
itive with traditional von Neumann architecture neural
networks in solving complex problems in pattern recog-
nition [13]. However, NMCs are simulators of spiking
nerves, based on a fundamentally different microscopic
process, namely electronics. Although the higher level
architecture may be designed similar to a real network
of neurons, the microscopics is different. Cultured neu-
rons have been used by the Moses group to construct
logic functions such as AND, also demonstrating a re-
markable reliability achieved through a redundancy of
connections [10]. Our approach is an attempt to sim-
plify the patterned neuron paradigm, by introducing a
simpler, synthetic "neuron". In contrast to NMCs, AAs
are to be based on the same microscopic process as real
neurons, namely ionics. This is, we believe, an incisive
experimental approach to study how the microscopics of
neurons may generate complex macroscopic responses,
patterns, and behaviors. The latter program partakes of
the underlying thread of condensed matter physics, and
FIG. 1: Opening probability for the potassium channel
used in this study (KvAP); reproduced from [15].
our approach is informed by that discipline.
Coming back to NMCs, interest in the field is both sci-
entific and practical, the latter because NMCs based AI
may enjoy possibly orders of magnitude better energy
consumption characteristics compared to von Neumann
architecture AI. AA chips could in principle enjoy simi-
lar architectural advantages, plus additional power sav-
ing benefits due to the fact that ionics circuitry works
at ∼ 100 mV vs the ∼ 1 V of electronic circuitry [1].
However, it should be mentioned that where power den-
sity is the sole criterion, notwithstanding, for example,
portability, there are low temperature devices which can
fare much better [14]. Our objective with the AA in the
near future is to develop a breadboard or "tool kit" with
which to construct task performing networks which are
based on the same microscopics as neurons, and there-
fore are "possible brains". At the moment, even one AA
is sufficiently delicate, and complicated to make, that we
thought it would be useful to give a demonstration of an
actual task performing device based on AAs, even though
we can only make very simple ones at present. Namely,
we use a system of two AAs to steer a remote control toy
car towards a light source.
First, we briefly describe the AA, which was introduced
previously [1, 2], and give an overview of the present sys-
tem. In the next section, we describe the results: they
consist of a demo in the form of a movie, with accompa-
nying analysis. In Materials and Methods we describe in
detail the EE aspects, in order to report exactly what we
do with the AAs and the electronic circuitry. We con-
clude with a brief discussion of relevance and scope.
The present AA is an excitable 100 µm size "node" con-
sisting of a supported phospholipid bilayer with ∼ 100
voltage gated potassium channels (KvAP) embedded and
oriented. A concentration gradient of potassium ions is
maintained across the bilayer by means of reservoires;
typically [K +]in ≈ 30 mM for the "inside" of the axon
and [K +]out ≈ 150 mM for the "outside". In the follow-
ing, all potentials represent the potantial difference be-
tween the inside and outside of the axon: V = Vin−Vout.
The aforementioned ionic gradient results in an equilib-
rium (Nernst) potential
kT
e ln
[K +]out
[K +]in
≈ + 40 mV
VN =
(1)
Here, kT ≈ 25 meV is the thermal energy at room tem-
perature and e the electronic charge. In the real neuron,
there is in addition an opposite gradient of sodium ions
across the membrane, which results in a "resting poten-
tial" Vr intermediate between the Nernst potentials of
the K + and N a+ ions, and corresponding voltage gated
sodium channels. In the AA, that role is played by the
current limited voltage clamp (CLVC) [2], which keeps
the AA voltage out of equilibrium, typically at a "rest-
ing potential" Vr ≈ −120 mV . The response curve of the
KvAP channels (Fig. 1) shows that they are closed at
V = Vr = −120 mV , but if a stimulus brings the volt-
age above V ∼ −20 mV the channels open, the chemical
potential driven channel current overwhelms the CLVC
current, and the AA "fires". The subsequent inactiva-
tion of the KvAP channels, which stochastically enter a
third, closed, state, allows the CLVC to pull the mem-
brane potential back to the resting value. In Fig. 2 we
show a train of two action potentials in an artificial axon.
The input stimulus is a constant current (∼ 128 pA), de-
livered with a current clamp. The membrane potential
Vm is measured through a separate electrode, and corre-
sponds to the ordinate scale on the left in the figure. The
dotted trace represents the command voltage Vc to the
CLVC, and corresponds to the ordinate scale on the right.
Instead of simply keeping Vc constant at ∼ −364 mV ,
which would maintain, in this case, a resting potential
Vm ≈ −120 mV , the protocol for Vc is that when the axon
"fires", Vc is lowered to −636 mV for a fixed time (= 1 s),
then returns to −364 mV . This maneuver is necessary
but incidental to the particular inactivation dynamics of
the ion channel we are using (the KvAP). Namely, after
a firing we must pull the membrane potential Vm down
to about −150 mV in order that the channels re-activate
fast enough to be ready for the next firing.
2
FIG. 2: Action potentials in the AA, elicited by a
constant current input. Plotted is the measured
membrane potential (Vm: left scale). The dotted trace
shows the protocol of the CLVC (Vc: right scale). The
downward step in Vc is triggered by Vm crossing 10 mV ;
see text for explanations.
We have previously documented some basic electrophys-
iology with this system, such as integrate-and-fire dy-
namics [1]. For our present purpose, we use two AAs
as a control module to steer a toy car towards a light
source. Fig. 3 shows a diagram of the system. The
AAs sit on an optical table in the lab; communication
with the car is by radio waves. The remote control car is
modified with two sets of photodiodes ("eyes") accepting
light from the right (R) and left (L) side of the car, re-
spectively, and corresponding voltage-to-frequency con-
verters (VFCs) and transmitters. Following for example
the signal from the R photodiode, its voltage output is
converted to frequency (in the kHz range) by the VFC
and transmitted; this signal is received by a receiver on
the optical table, converted to voltage by another VFC,
and used as the input to a current clamp, or "synapse",
which injects a proportional current (in the tens of pA
range) into the right (R) axon. Action potentials in the
AA trigger a threshold detector which inputs into the
remote control module of the car the signal to turn the
wheels to the right. We use the actual remote control
of the toy car, and so the same receiver and right/left
control built into the car. A similar but independent
pathway conditions the signal from the left photodiode
set. In summary, this system realizes a very simple ana-
logue control protocol: each time the R axon spikes, the
wheels of the car are turned to the right and stay there
until the next signal comes in, and similarly for the L
axon, which turns the wheels to the left. Action poten-
tials in the R/L axon are induced by the light intensity
falling onto the R/L photodiode set. So while all the
peripheral systems are, at the moment, electronic, the
"decision making algorithm" is implemented by ionics.
There is no interaction between the two AAs in this
realization, and the only property of the AA which
we really exploit is the "integrate-and-fire" dynamics.
The system has a degree of stochasticity (due to the
3
FIG. 3: System block diagram. The complete system
consists of two such circuits, one for the right (R)
photodiodes and AA and one for the left (L)
components. VFC: voltage to frequency converter;
FVC: frequency to voltage converter.
FIG. 4: Still of the room of the room of the demo, seen
from the ceiling. The car and light source are at
diametrically opposite corners. Visible on the right is
the optical table with the artificial axons and the
electronics, as well as H.G.V.
relatively small number of ion channels in the axons),
lots of noise which is not only thermal in origin, makes
many mistakes, has many defects, and ends up looking
"biological" (see movie).
Results. The main result we present is a demonstration
of the system in the form of the accompanying movie (see
Ancillary files), which we now describe. The car moves
in a (previously decluttered) laboratory room of about
5×5 m2; in the movie, the light source is in the SW corner
of the screen, and the car starts at the NE corner, facing
W. The other bright spots on the screen are reflections of
the light source from objects at the perifery of the room.
Fig. 4 is a picture of the room seen from the ceiling.
In the movie, watch the front wheels of the car
repeatedly switch between L and R, and the overall
progress. We had to slow down the speed of the car
to match it to our rather slow axons, so we adopted
short, regular spurts of forward motion. The actual
average speed of the car is however not constant, as
you see in the movie, because at times the tires slip
FIG. 5: Car trajectory corresponding to the run shown
in the movie. The light source is at the origin; the
scales on the axes are in m, and the car starts at
(x, y) = (3.4, 3.0).
on the polished floor. From an engineering standpoint,
we may view this circumstance as simply one of many
"defects" or sources of noise in the system. There are
many such sources of randomness, from the microscopic
scale of the individual ion channels in the AA to the
macroscopic scale of the tires. As a result, the trajectory
of the car, the "behavior", is not deterministic (starting
from identical initial conditions, different realizations of
the car's trajectory will be different); however, the car
does find the light source in the end. Fig. 5 shows the
trajectory corresponding to the movie.
Let us now describe the very simple "machine lan-
guage" with which the system operates. Fig. 6a shows
action potentials in the two AAs over a time of 10 s;
the blue trace is the membrane potential of the left AA,
the yellow trace is the right AA. The response of the car
is that when the blue trace crosses 4.5 mV from below,
the wheels turn left and stay there until further notice;
similarly, when the yellow trace crosses 0 mV , the wheels
turn right and stay there. What decides, then, whether
overall the car is turning L or R is the relative phase of
the spikes in the two AAs.
In the example shown, for
45 < t < 50.5 s the car is, overall, turning R because the
time interval between a yellow and the next blue zero
crossing is larger than the time between a blue and the
next yellow. On the other hand, for 50.6 < t < 56 s the
car is overall turning L; this is a consequence of the R
photodiodes seeing less light for 50 < t < 52 s (Fig. 6b),
which causes a delay in the yellow spikes, changing the
phase relation between blue and yellow spikes. Even with
identical stimuli (input currents from the "synapses"),
the firing rates of the two AAs are not the same (due to
4
FIG. 7: Whole time series of action potentials
corresponding to the run of the movie. Time t = 0 s
corresponds to the start of the video. The data
recording begins 1.75 seconds later.
(a) Left axon; (b) right axon.
AAs corresponding to the run of the movie, and in Fig.
8 two different representations, among the many possible
ones, of the "behavior" of the car.
Materials and Methods. In this section, we describe
in detail the EE aspetcs of the demo, and summarize
the construction and operation of the AA; the latter has
been described before [1, 2].
Molecular biology. The KvAP gene in vector pQE60
is expressed in E. coli strain XL1-Blue competent cells
(Aligent) and reconstituted in DPhPC vesicles. Vesicles
are fused to the supported bilayer to introduce the
channels in the AA. Protein expression, purification,
and reconstitution protocols are described at length in
previous publications [1, 2, 15].
The "eyes" of the car. For each eye, three Burr-Brown
OPT 301 photodiodes are positioned at 90 degrees rela-
tive to each other. This arrangement is shown in Figure
9 for right eye. One photodiode faces the direction of for-
ward motion, another is oriented 180 degrees relative to
the first in the backward direction, and the third is per-
pendicular to the other two, facing outward. The sum
of their outputs is the turn input to one artificial axon.
FIG. 6: (a) Membrane potential in the Left artificial
axon (blue trace) and the Right AA (yellow trace) for
part of the run shown in the movie.
(b) The signal at the output of the voltage to frequency
converter (VFC), for the R and L circuits, and the same
time interval as in (a). The currents injected by the
synapses into the respective AAs are proportional to
these signals.
physical differences between the AAs, for instance, differ-
ent number of ion channels, different leak currents, etc.).
This circumstance introduces "phase noise", yet another
source of (non-thermal) stochasticity which however does
not prevent the overall working of the system. That is,
the two AAs do not need to be perfectly tuned as far as
firing rates.
Fig.
6b shows,
for the same run as in part a),
the frequency coming out of the voltage to frequency
converter, for the L (blue) and R (yellow) circuit. The
current injected by the corresponding "synapse" into the
L / R axon is proportional to this frequency. While the
firing rates of the two AAs do not need to be perfectly
tuned, if, for equal light, one firing rate is larger than
the other, this introduces a bias in the approach to the
light source.
In the realization shown in Fig. 5, the
right AA had a faster firing rate, and a right turn bias is
visible in the trajectory. We come back to the question
of how much difference in firing rates can be tolerated in
the Discussion.
In Fig. 7 we report the whole time series for the two
5
to zero at θ ∼ 60 degrees. As an example, and referring
to Fig. 9, with the light source in the NE direction,
the forward and outward-facing photodiodes contribute
input to the AA for right turns. The backward-facing
photodiode is in the dark and therefore does not con-
tribute to the AA input. For each eye, the ground pin
of one photodiode is connected to the amplifier output
of the next. The result is one output voltage from all
three photodiodes, equal to the sum of their individual
outputs. The summed output is converted to a frequency
and sent wirelessly to the corresponding Artificial Axon.
Although it is not necessary, for these navigation exper-
iments we chose the photodiode circuitry such that even
far away from the light source but at normal incidence,
the photodiode output saturates. Specifically, we chose
the transimpedance resistance so that the photodiode
saturates at 5 meters from the 625 nm, 100 W LED light
source.
Source and terrain. Heating and background in the high-
wattage LED are minimized with an active heat sink
and brightness shields, respectively. Silicone thermal
grease is applied to the interface between the LED and a
heat sink for improved thermal conduction. Background
refers to reflection of light from the floor and walls that
adds a constant to the photodiode voltage output and
is roughly independent of car orientation. The CLVC
competes with this background and membrane leaks to
keep the axon at the resting potential, and struggles
when leaks alone are large. A brightness shield above
the LED lowers background from reflection off the
ceiling and walls, and a shield below the LED reduces
background from floor reflection. Walls and reflective
surfaces are covered with black tarps for background
reduction. In total, shielding reduces the background to
less than 5% of the photodiode saturation voltage.
Wireless communication. The Burr-Brown VFC32 con-
verts the photodiode voltage output to a digital pulse
input for the wireless transmitter. An external resistor
value is chosen to make the VFC frequency linearly pro-
portional to the photodiode output, and an external ca-
pacitor value is chosen to cap the frequency maximum at
4 kHz. Two separate radio wave transmitters are used
for left/right turning, 433 and 315 MHz, one transmitter
for each direction. Encoder/decoder components typi-
cally paired to transmitters/receivers are excluded here
because there is no interference between transmitters,
and there is no interference from other devices in the lab
where the navigation demonstration is performed. Fur-
thermore, these simple receivers struggle with transmit-
ted data frequencies above 17 kHz, but encoder oscillator
frequencies must be larger than this to transmit VFC in-
put in the kHz range.
Unlike voltage-to-frequency,
the frequency-to-voltage
conversion is done by software. While the VFC32
FIG. 8: (a) Distance from the target vs time, calculated
for the trajectory of Fig. 5;
(b) time series showing intervals when the car is turning
left (+1) and right (−1). This plot corresponds to the
spike trains of Fig. 6.
FIG. 9: Top-view schematic of the photodiode
arrangement for right-turns.
With this arrangement, if the light source in Figure 9 was
to the left of the car, the right photodiodes would output
no turning voltage and the car would turn left toward
the light source.
The voltage output of one photodiode depends on the an-
gle of incidence of the light falling on it. For small angles
θ (counted from normal incidence), the voltage drops as
cos(θ), and for larger angles the voltage drops sigmoidally
can convert frequency to voltage, the wireless receiver
outputs a frequency duty cycle that is different from
the VFC. To correct for the mismatch by hardware is
unnecessarily complicated. Therefore, the frequency to
voltage conversion is instead done by computer using
LabVIEW. The program converts frequency to voltage,
which is then fed to the input of the current clamp
which forms the "synapse" injecting into the AA. The
command voltage to the clamp is thus Vcc = α f where
f is the receiver frequency and α is a proportionality
constant. There are two independent circuits for the
right and left axons. The constant α is chosen so
that it matches the electrophysiology characteristics
of the corresponding axon (mainly dependent on leak
current, Nernst potential, number of channels) and is
therefore different between axons. For the run of the
movie, we had αL = 2.1 mV /kHz for the left axon and
αR = 1.8 mV /kHz for the right axon. The receiver
output becomes noisy at frequencies near zero, so an
added filter is written into LabVIEW to filter out this
frequency noise.
LabVIEW interprets this noise as
large frequencies as high as 15kHz. The filter removes
frequencies higher than 3900Hz, before the voltage
conversion.
Current clamp. A schematic of the current clamp is
shown in Figure 10. All op-amps used are the low-noise
FET precision op-amp AD795. The summation amplifier
adds the command voltage VCC to the membrane voltage
Vm as −(VCC + Vm), and the inverter flips the voltage
reference so that the sum is positive. The potential
difference across the 100 M Ω current clamp resistor
RCC is VCC, and the current injected is VCC/RCC.
The high-impedance voltage follower measures Vm for
feedback to the summation amplifier.
CLVC protocol.
The CLVC protocol during firing
(VC, F ig.2) is incidental to the particular inactivation
dynamics of the KvAP. Channels will completely inac-
tivate if the membrane potential is not pulled down to
a large negative value after firing. The channel recovery
rate from inactivation has a sigmoidal dependence on
the membrane voltage, with the turning point at about
−100 mV .
Pulling the membrane voltage down to
∼ −100 mV is typically sufficient to maintain firing.
For the run of the movie, the following settings were
used. For the left axon: when the membrane voltage
reaches the trigger value VT = 4.5 mV , the command
voltage to the CLVC changes from VC(1) = −127 mV to
VC(2) = −455 mV for tT = 1.3 s, pulling the membrane
voltage to a large negative value. For the right axon:
VT = 0 mV , VC(1) = −145 mV , VC(2) = −364 mV ,
tT = 1.0 s.
The clamp value VC(2) is chosen with a big safety margin
to address the fact that sometimes the leak conductance
of the AA changes in the course of a run. For example,
6
FIG. 10: Schematic of the current clamp. The green
dots represent the positions of the stated voltage values,
also in green. The current clamp has three components:
a summation amplifer circuit, a voltage inverter, and a
high-impedance voltage follower. The potential
difference between the voltage inverter and follower
determines the current injected into the AA as
ICC = VCC/RCC.
you can see in Fig. 7a, looking at the negative swings
of the spikes, that the envelope of the spikes is roughly
constant (at ∼ −280 mV ) for 0 < t < 80 s, then increases
for 80 < t < 100 s, then stabilizes again (at ∼ −120 mV )
for t > 100 s. This increase is caused by an increase
in leak conductance of the axon, from ∼ (83 GΩ)−1 to
∼ (2.4 GΩ)−1, approximately. However, even with the
increased leak, the same CLVC protocol is able to pull
the resting voltage down below −100 mV , allowing the
channels to recover from inactivation and so be able
to fire repeatedly. Similarly, you see that in the right
axon (Fig. 7b) the leak conductance increases and then
decreases again for 20 < t < 50 s. The origin of these
slow fluctuations in leak conductance is presumably that
the interface between lipid bilayer and solid support is
not as stable as one would wish, in the present system.
Similar fluctuations in leak conductance are observed
even in the absence of channels, so this is a membrane
phenomenon. In our present system, a membrane with
channels lasts typically ∼ 10 min before it breaks;
exceptionally we have lifetimes of ∼ 1 hour. Without
channels, a membrane lasts typically 1 hr. Thus we will
need to significantly improve the stability of the system
if we want to scale it up even modestly.
Electrophysiology parameters. For the run of the movie,
the axon parameters were set / measured as follows.
Left axon: at maximum photodiode output, the current
clamp injects I max
CC = 74 pA into the axon; the number
of open channels at peak voltage is approximately
N = 380; the membrane capacitance is C = 190 pF .
Right axon: I max
CC = 64 pA, N = 720, and C = 185 pF .
The vehicle.
From the manufacturer (GPTOYS),
left/right movement on the model S911 car's remote
is controlled by a 5 kΩ potentiometer configured as
a voltage divider.
In our system, we removed the
potentiometer and connected the car remote to a
National Instruments NI USB-6008 data acquisition
device (DAQ). Analog turn signals are given to the car
remote by LabVIEW through the DAQ, in place of the
potentiometer. The negative terminal of the remote's
battery is connected to the DAQ's ground channel, and
turn voltages are supplied by the DAQ to the remote's
"signal" pin in the voltage divider circuit.
The smallest-radius left turn corresponds to a 0 V
signal with respect to ground. The smallest-radius
right turn corresponds to 3 V, and forward directed
wheel orientation corresponds to 1.5 V. Signal values
between 0 V and 3 V correspond to larger turn radii
that decrease linearly as the signal moves away from
1.5 V in either direction. When an axon's membrane
potential exceeds the set trigger voltage, LabVIEW
sends an analog voltage signal to the car remote to make
a smallest-radius turn in the direction of the axon that
fired. The analog turn signal persists until LabVIEW
detects a voltage signal (from the other axon) to turn in
the opposite direction.
From the manufacturer, forward motion of the car is
also controlled by a 5 kΩ potentiometer configured as
a voltage divider in the car remote. The stop position
corresponds to a 3 V signal, and maximum speed
corresponds to a 0 V signal.
In the remote's circuit
protocols for forward motion, the applied signal must
be at 3 V when the remote is turned on. The car
begins to move at 150 mV below 3V, i.e. 2.85 V. For
compatibility with the slow (∼ 1 Hz) firing rate of our
axons, we had to slow down the car. We introduced two
modifications. First, a LabVIEW function generator
supplies square pulses with amplitude ∼ 150 mV and
period 500 ms for forward motion, with an offset chosen
so that the maximum voltage is 3 V . Second, four 50W
resistors are connected in parallel to the car's motor
to reduce the motor's current. This is a high current
RC motor, so the power resistors are necessary. These
modifications bring the car's speed down to 20−30 cm/s.
Discussion. Our goal with this demo is to instigate
the development of "ionic networks" [1]. We submit
that a large network of artificial axons connected by
tunable synapses would form an interesting neuroscience
breadboard. One use would be to analyze principles of
how the "microscopics" of action potentials may give
rise to macroscopic behavior. We note in passing that
such a program is within the traditional focus of con-
densed matter physics, which seeks to understand "emer-
gent" macroscopic properties starting from the micro-
scopic components and interactions. At a higher level
of description, the relation between information flow and
behavior need not be based on complicated rules in or-
der to produce complex behavior. In his delightful book
7
FIG. 11: Car trajectory obtained from a simulation
where the right AA has a firing rate 1.9 times higher
than the left AA, for the same light seen. The car still
"finds" the light source, which is at the origin.
"Vehicles", Valentino Braitenberg explains how simple
control mechanisms can lead to surprisingly complex be-
havior [16]. His very first example in the book is the
car with left/right control. However, the specific micro-
scopics of action potentials puts constraints on the flow
of information and also provides specific mechanisms for
the interaction of different bits of information. If we be-
lieve that the latter process is essential for "thought", we
want our test network to be based on nodes which sup-
port action potentials. Even our simple, non-interacting
system is not trivial to analyze, if one gets into a little
detail, though it is easy enough to simulate. Let us come
back to the issue of different firing rates for equal light
intensity. Fig. 11 shows the trajectory from a simulation
with similar initial conditions as the movie (see Materials
and Methods for details). In the simulation, the right AA
had a firing rate 1.9 times higher than the left AA, for
the same light received at the photodiode. The right turn
bias is visible in the car's trajectory, but overall the car
still finds the light source. The end state is a limit cycle
which is a circle containing the light source. How much
difference in the firing rates can be tolerated depends
on the other parameters of the system. For example,
with the firing rate ν, the car speed u, the turn radius
of the car r, the initial distance to the light source L, we
can form the two dimensionless numbers χ = u/(νr) and
ρ = r/L. Then we can discuss, in this parameter space,
the basin of attraction of the set of limit cycles which
form the desirable end states. However, this is already a
complicated question to explore analytically, for such a
simple dynamical system !
Looking to the future, we are far from being able
to construct a self-contained ionic network. Some of
the difficulties seem surmountable with present day
engineering, others would require new inventions.
In
the context of this demo, for example, we can see a
path for substituting some of the electronic components
with ionics. The CLVC could be dispensed with by
of N a+, and
adding a second ionic gradient, e.g.
corresponding voltage gated ion channels. Photodiodes
could in principle be replaced by AAs with embedded
channel rhodopsin. On the other hand,
ionics based
"synapses" compatible with the 100 mV scale of ionic
action potentials require new inventions. Finally, 3D
printing technology currently being developed to pro-
duce scaffolds for directed neuronal growth [17] could
probably form the basis for scaling up our AA network.
ACKNOWLEDGEMENTS
This work was supported by funds from the Dean of
Physical Sciences at UCLA. We thank Albert Libchaber
for pointing us to Valentino Breitenberg's book.
∗ [email protected]
[1] H.G. Vasquez and G. Zocchi, "Coincidences with the ar-
tificial axon," EPL 119, 48003 (2017).
[2] Amila Ariyaratne and Giovanni Zocchi, "Toward a mini-
mal artificial axon," J. Phys. Chem. B 120, 6255 (2016).
[3] Rajagopal Ananthanarayanan, Steven K. Esser, Horst D.
Simon,
and Dharmendra S. Modha, "The cat is out
of the bag: Cortical simulations with 109 neurons, 1013
synapses," Proceedings of 2009 IEEE/ACM Conference
on High Performance Computing Networking, Storage
and Analysis (2009).
[4] Giacomo Indiveri et al, "Neuromorphic silicon neuron cir-
cuits," Frontiers in Neuroscience 5, 73 (2011).
8
[5] Thomas Pfeil et al, "Six networks on a universal neuro-
morphic computing substrate," Frontiers in Neuroscience
7, 11 (2013).
[6] Claire Wyart et al, "Constrained synaptic connectivity in
functional mammalian neuronal networks grown on pat-
terned surfaces," J. Neurosci. Methods 117, 123 (2002).
and Elisha Moses,
"Signal propagation along unidimensional neuronal net-
works," J. Neurophysiol. 94, 3406 (2005).
[7] Ofer Feinerman, Menahem Segal,
[8] Ofer Feinerman and Elisha Moses, "Transport of infor-
mation along unidimensional layered networks of dissoci-
ated hippocampal neurons and implications for rate cod-
ing," J. Neurosci. 26, 4526 (2006).
[9] John C. Chang, Gregory J. Brewer,
and Bruce C.
Wheeler, "Neuronal network structuring induces greater
neuronal activity through enhanced astroglial develop-
ment," J. Neural Eng. 3, 217 (2006).
[10] Ofer Feinerman, Assaf Rotem, and Elisha Moses, "Reli-
able neuronal logic devices from patterned hippocampal
cultures," Nature Physics 4, 967–973 (2008).
[11] Emanuele Marconi et al, "Emergent functional proper-
ties of neuronal networks with controlled topology," PLoS
ONE 7, e34648 (2012).
[12] Anca M Pasca et al, "Functional cortical neurons and
astrocytes from human pluripotent stem cells in 3d cul-
ture," Nature Methods 12, 671 (2015).
[13] Catherine D. Schuman et al, "Survey of neuromor-
phic computing and neural networks in hardware,"
arXiv:1705.06963 (2017).
[14] Michael L. Schneider et al, "Ultralow power artificial
synapses using nanotextured magnetic josephson junc-
tions," Science Advances 26, 4526 (2006).
[15] A. Ariyaratne and G. Zocchi, "Nonlinearity of a voltage
gated potassium channel revealed by the mechanical sus-
ceptibility," PRX 3, 011010 (2013).
[16] Valentino Braitenberg, Vehicles (MIT Press, Cambridge,
MA, 1984).
[17] Daniela Espinosa-Hoyos et al, "Engineered 3D-printed
artificial axons," Scientific Reports 8, 478 (2018).
|
1712.06817 | 1 | 1712 | 2017-12-19T08:28:26 | Optical Antenna-based Fluorescence Correlation Spectroscopy to Probe the Nanoscale Dynamics of Biological Membranes | [
"physics.bio-ph",
"physics.optics"
] | The plasma membrane of living cells is compartmentalized at multiple spatial scales ranging from the nano- to the meso-scale. This non-random organization is crucial for a large number of cellular functions. At the nanoscale, cell membranes organize into dynamic nano-assemblies enriched by cholesterol, sphingolipids and certain types of proteins. Investigating these nano-assemblies known as lipid rafts is of paramount interest in fundamental cell biology. However, this goal requires simultaneous nanometer spatial precision and microsecond temporal resolution which is beyond the reach of common microscopes. Optical antennas based on metallic nanostructures efficiently enhance and confine light into nanometer dimensions, breaching the diffraction limit of light. In this Perspective, we discuss recent progress combining optical antennas with fluorescence correlation spectroscopy (FCS) to monitor microsecond dynamics at nanoscale spatial dimensions. These new developments offer numerous opportunities to investigate lipid and protein dynamics in both mimetic and native biological membranes. | physics.bio-ph | physics | Optical Antenna-based Fluorescence
Correlation Spectroscopy to Probe the
Nanoscale Dynamics of Biological Membranes
Pamina M. Winkler†, Raju Regmi‡,†, Valentin Flauraud§, Jürgen Brugger§, Hervé Rigneault‡,
Jérôme Wenger*‡, María F. García-Parajo*†⊥
† ICFO-Institut de Ciencies Fotoniques, The Barcelona Institute of Science and Technology,
Barcelona, Spain; ‡ Aix Marseille Univ, CNRS, Centrale Marseille, Institut Fresnel, Marseille,
France; § Microsystems Laboratory, Institute of Microengineering, Ecole Polytechnique
Fédérale de Lausanne, 1015 Lausanne, Switzerland; ⊥ ICREA, Pg. Lluís Companys 23, 08010
Barcelona, Spain.
*E-mail: [email protected]; [email protected]
The plasma membrane of living cells is compartmentalized at multiple spatial scales ranging
from the nano- to the meso-scale. This non-random organization is crucial for a large number
of cellular functions. At the nanoscale, cell membranes organize into dynamic nano-
assemblies enriched by cholesterol, sphingolipids and certain types of proteins. Investigating
these nano-assemblies known as lipid rafts is of paramount interest in fundamental cell
biology. However, this goal requires simultaneous nanometer spatial precision and
microsecond temporal resolution which is beyond the reach of common microscopes. Optical
antennas based on metallic nanostructures efficiently enhance and confine light into
nanometer dimensions, breaching the diffraction limit of light. In this Perspective, we discuss
recent progress combining optical antennas with fluorescence correlation spectroscopy (FCS)
to monitor microsecond dynamics at nanoscale spatial dimensions. These new developments
offer numerous opportunities to investigate lipid and protein dynamics in both mimetic and
native biological membranes.
Keywords: optical antenna, fluorescence correlation spectroscopy (FCS), living cell
membrane, lipid rafts, plasmonics
1
The plasma membrane is a complex, versatile and essential signaling interface that separates
the cell cytoplasm from the extracellular space.1,2 Its spatiotemporal organization and
biological function are intricately interlaced at the nanoscale.3–5 The heterogeneous landscape
of the cell membrane is shaped by a variety of lipids and proteins that differ in their physico-
chemical properties. Sphingolipids, cholesterol and certain types of proteins such as
glycosylphospatidylinositol-anchored proteins (GPI-APs) can assemble
into dynamic
nanoscale clusters or nanodomains, also known as lipid rafts.6,7. Although lipid rafts have
been implicated in a large number of cellular functions5,8–15, their existence in living cells
have been heavily debated given the enormous challenge associated with observing these
transient nanodomains.7,16–23 This Perspective aims to briefly review and to place into a
broader context very recent works using advanced nano-optical techniques to investigate lipid
rafts and their dynamics in biological membranes.
The basis for understanding cell membrane structure was proposed nearly 50 years ago by
Singer and Nicolson.24 This fluid mosaic model captures the general characteristics of the cell
membrane as a lipid bilayer dressed with embedded proteins. However, intensive research in
the last twenty years has revealed that biological membranes are highly heterogeneous and
with a much higher complex architecture that goes well beyond what it was initially proposed
by the fluid mosaic model. Within the plane of the membrane, certain types of proteins,
sphingolipids and cholesterol arrange in transient nanoscopic domains, also denoted as lipid
rafts.1,2,6,17,21 These highly dynamic and fluctuating nanoscale assemblies can be stabilized in
the presence of lipid- or protein-mediated activation events to compartmentalize cellular
processes.2,18 By means of physically segregating specific molecular components within the
membrane, lipid rafts are believed to modulate the activity of raft-associated proteins, and
influence signaling and function of a broad range of membrane receptors.5,8,10,15 Moreover,
recent research indicates that the biophysical properties of lipid rafts (size, composition and
dynamics) can be modulated by the proximal actin cytoskeleton7,25,26 and components of the
extracellular matrix27–31, adding an extra complexity to the sub-compartmentalization of the
plasma membrane. While the overwhelming diversity of membrane nanodomains makes their
study particularly challenging, understanding the fundamental mechanisms that lead to raft
formation as the first organizing principle of the cell membrane, is of paramount importance.
Artificial lipid bilayers have been extensively used as model systems since they recapitulate
some of the most important features of biological membranes.32–35 On the microscopic scale,
2
ternary lipid membranes composed of unsaturated phospholipids, saturated sphingolipids and
cholesterol separate into two distinct liquid phases which can be resolved by diffraction-
limited optics: a liquid disordered (Ld) phase comprised mainly of unsaturated phospholipids
and a liquid ordered (Lo) phase mostly composed of saturated lipids and cholesterol.2,5,36 This
Lo phase has been considered to represent the potential physical model for living rafts in
cellular membranes.2,5,32–35 Microscopic and stable liquid-liquid phase separation has been
observed on both supported lipid bilayers (SLB) and giant unilamellar vesicles prepared from
cell membrane lipid extracts.37,38 However, such phase coexistence has remained so far
largely unresolved on biological membranes. Interestingly, some studies have shown that the
cell membrane in all its complexity is fully capable to phase segregate into a micrometer-
sized two-phase fluid-fluid system, upon a temperature decrease39, or through ganglioside
GM1 (a raft lipid) tightening by its ligand cholera toxin-β (CTxB) at physiological
temperatures40, provided that the membrane is separated from the influence of the cortical
cytoskeleton. Based on these results, it has been proposed that an underlying selective
connectivity mediated by cholesterol must exist among membrane rafts even at the resting
state.2,40 This connectivity will thus be responsible for the large-scale phase segregation
induced far beyond the valency of initial GM1 tightening through CTxB.40,41 Yet, most of the
experimental proof for such raft connectivity has been based on the visualization of the end
stage of an activated condition and in the absence of the cytoskeleton and/or membrane
traffic, where the transient rafts are amplified to coalesce into larger, stable micrometer-sized
raft domains. It is only at this stage that standard fluorescence microscopy is able to observe
this segregation.
In the context of fully intact living cells, early investigations on membrane organization
yielded conflicting results regarding the sizes, distribution and dynamics of lipid rafts,
including experimental results that refuted their existence.2,5,16,18,23 Most of the earliest work
was performed using fluorescence recovery after photobleaching (FRAP)42,43 and more
recently, using single particle
tracking
(SPT)3,23,43,44 and
fluorescence correlation
spectroscopy (FCS).19,23,43 FCS has been widely adopted for studying structural dynamics and
biomolecular interactions on cell membranes as it features several key advantages.19,45,46 The
working principle of FCS is to analyze the temporal correlation of fluorescence intensity
fluctuations.47 This allows to determine the mean transit time averaging over thousands of
single molecule diffusion events. The local molecular mobility can thus be investigated with a
3
high temporal resolution in the sub-microsecond regime together with a broad dynamic range
of timescales from microseconds to seconds.
While the FCS correlation function contains rich information on the molecular mobility, it is
however hard to extract a complete description of the diffusion process (free, anomalous,
constrained, directed …) out of a single FCS measurement. Alternatively, a more powerful
method consists in performing diffusion measurements over a range of observation areas, as
first introduced by Yechiel and Edidin in the context of FRAP.48 This concept has been
further generalized by Lenne and coworkers to establish the so-called "FCS diffusion
law"20,49, which is a graph representing the average FCS diffusion time as a function of
increasing observation areas (Fig.1). Based on a series of FCS measurements for different
observation areas, the shape of the FCS diffusion law allows to determine the nature of the
diffusion process and the underlying membrane organization at scales smaller than the
accessible experimental observation area.20,50 Free diffusion is characterized by a strict linear
proportionality between the diffusion time and the area, hence the curve crosses the origin
(Fig.1 c). The presence of impermeable obstacles constrains the diffusion and increases the
apparent time to cross a given observation area, thus the slope of the FCS diffusion law is
higher, but the origin (0,0) is still crossed. Notably, the presence of confinement affecting the
lateral diffusion is revealed by a deviation of the intercept on the time axis t0 from the origin
(Fig.1 c). Extrapolating the experimental curve to the intercept with the time axis, hindered
diffusion due to nanodomains is regarded as a positive intercept on the time axis, while the
meshwork model is related to a negative intercept. This approach was established on
diffraction-limited confocal microscopes, where the FCS diffusion law for the transferrin
receptor TfR-GFP (known to interact with the cytoskeleton meshwork) yielded a negative t0
value, while that of the fluorescent ganglioside GM1 exhibited a positive t0 value.20,51 By
extrapolating to the origin, the FCS diffusion laws can predict the occurrence of membrane
heterogeneities affecting the lateral diffusion at spatial scales well beyond the optical
resolution. However, the size of lipid rafts is expected to be around 10-100 nm18,19,52, so their
areas are 5 to 500 x smaller than the smallest diffraction-limited observation area on confocal
microscopes. Reducing this gap between optical resolution and the size of lipid rafts to gain
better insights on membrane organization at the nanoscale is currently a field of active
research.
4
Figure 1. Principles of FCS diffusion laws to reveal biomembrane organization at the
nanoscale. (a) FCS diffusion laws are constructed by measuring the diffusion times of
molecules traversing illumination areas of increasing sizes. (b) Different diffusion models
depending on the membrane organization can be distinguished by varying the illumination
areas. Molecules can freely diffuse on the membrane or show hindered diffusion due to their
dynamic partitioning into nanodomains or due to the cortical actin meshwork. (c) FCS
diffusion laws showing diffusion times versus observation area. The type of diffusion is
retrieved by extrapolation of the curves through the y-axis intercept t0. Free diffusion and
impermeable obstacles are characterized by t0 = 0, while a positive t0 intercept indicates the
presence of nanodomains transiently trapping the molecular probe. A negative t0 intercept
relates to a meshwork of barriers separating adjacent domains. The observation areas
accessible with various super-resolution techniques are indicated as grey lines.
With the advent of super-resolution optical microscopy approaches such as single molecule
localization methods53–55, stimulated emission depletion (STED) microscopy56–58 and near-
field scanning optical microscopy (NSOM)8,59–62, it is now becoming clearer that lipids and
proteins can indeed organize in nanometric compartments on the cell membrane, albeit a
5
consensus in terms of their sizes and dynamics has not yet been reached. In terms of dynamic
measurements at the nanoscale, NSOM has been combined with FCS to show anomalous
diffusion of ganglioside GM1 on living cell membranes at sizes smaller than 120nm.63 STED
has been also combined with FCS to explore the nanoscale dynamics occurring in lipid
membranes.22,64–68 Notably, STED-FCS experiments on living cell membranes revealed that
unlike phosphoglycerolipids, sphingolipids and GPI-APs are
transiently
trapped
in
cholesterol-mediated molecular complexes of sub-20 nm dimensions.22 These landmark
results rely on
the key benefits of STED-FCS
to provide sub-diffraction spatial
(super)resolution, tunable observation areas, and microsecond temporal resolution. STED-
FCS was also applied to study ternary lipid-cholesterol model membranes featuring
microscopic liquid-liquid phase separation into Ld and Lo phases, without observing any
direct evidence of the presence of nanoscopic domains at the spatial scales down to 40nm.65
However, nanoscale assemblies smaller than the 40 nm minimum STED resolution may still
exist without being observed. Further recent technical developments extended the FCS
functionalities to scanning STED-FCS67 being specially adapted to membrane studies of slow
diffusion, and to fluorescence lifetime filtering STED-FLCS.68 Recently, Basu and coworkers
detected dynamic heterogeneities at length scales of ~ 100-150 nm in binary phospholipid-
cholesterol bilayers of high cholesterol content (50 %) by applying STED-FCS (with a
resolution of ~ 80nm).69 The occurrence of these heterogeneities in binary model membranes
showing no macroscopic phase separation indicates that the domain formation is driven by
cholesterol packing and influenced by the phospholipid type. However, the high cholesterol
content (50 %) used in the binary mixtures complicates a direct comparison to cellular
membranes.
Advanced SPT3,26,70 and the recently introduced high-speed SPT interferometric scattering
microscopy (iSCAT) technique71 enable direct visualization of single particle trajectories.
iSCAT microscopy allows nowadays nanometer localization precision together with
microsecond time resolution by means of using 20-40nm gold nanoparticles as labeling
probes.72,73 Recent high-speed SPT experiments using 20 nm gold beads attached to
individual lipids in multicomponent model membranes showed anomalous diffusion in the Lo
phase consistent with the occurrence of nanoscale heterogeneities, while homogeneous lipid
diffusion was observed in the Ld phase.74 The estimated sizes of the nanodomains in the Lo
phase varied between 20 to 40 nm with lipid trapping times inside the domains below 1 ms.
iSCAT thus constitutes an attractive tool to investigate dynamic biophysical processes in
mimetic systems at the nanoscale, yet additional investigations are still required to rule out
6
potential artifacts related to the large size of the gold nanoparticle label with respect to the
lipids under study.
Beside these enormous progresses in super-resolution microscopy and single-molecule
dynamic approaches, advances from the nanophotonics field have led to the concept of
photonic nanostructures to confine light on a subwavelength scale and reach sub-diffraction
observation areas in FCS.75–77 A conceptually simple yet powerful approach uses single
nanometric apertures milled in a metallic film also known as zero-mode waveguides (ZMW)
to confine the illumination spot directly in the sample plane.78 Typically, the apertures have
radii between 50 to 250 nm and are milled in an opaque aluminum film covering a glass
coverslip.79,80 Their combination with FCS has been used to probe model lipid membranes81,82
and living cell membranes83–86, revealing for instance that fluorescent chimeric ganglioside
proteins partition into 30 nm structures within the cell membrane.86 While the ZMW approach
is very efficient at confining light within nanospots of diameters between 100 to 200 nm, this
technique has difficulties reaching spot sizes below 80 nm. Indeed, the FCS signal-to-noise
ratio rapidly deteriorates for ZMW diameters below 100 nm as a consequence of fluorescence
quenching induced by the metallic aperture edges.87 An additional issue affecting the use of
ZMWs for living cell membrane studies is the lack of control on the membrane invagination
into the aperture. This problem has been addressed by introducing a planarization procedure
in the nanofabrication process filling the aperture volume with fused silica.88,89 Thanks to the
absence of a height difference between the ZMW and the surrounding metal layer, the cells
can lie on a nearly perfectly flat surface. The best results achieved so far reach a nanospot
diameter of 60 nm and microsecond resolution.89
The concept of resonant optical nanoantennas has been introduced to further confine the
excitation light down to sub-20 nm scales.90–92 Optical nanoantennas are metallic (plasmonic)
nanostructures that localize and enhance the incident optical radiation into highly confined
nanometric regions (plasmonic hotspots),
leading
to greatly enhanced
light-matter
interactions.93,94 Over thousand-fold enhancement of the single molecule fluorescence signal
was reported with lithographically fabricated gold nanoantennas in the shape of bowties95,
with dimers of gold nanoparticles assembled with DNA origami96,97 and at the apex of single
gold nanorods.98,99. Recent advances in nanofabrication using colloid nanosphere lithography
combined with plasma processing100 and nanostencil lithography101 enable nowadays large
scale production of reproducible nanoantennas with narrow gaps as required for the study of
the plasma membrane of living cells. However, the applications of plasmonic antennas to
living cells remain scarce. Plasmon-enhanced fluorescence was recently observed inside
7
living bacterial cell membranes102, highlighting the need to develop well-tuned substrates to
maximize fluorescence enhancement and signal-to-noise ratio. In the highly active field of
biosensing in the context of nanomedicine103 plasmonic antennas have enabled to perform
Raman spectroscopy in a microfluidic device on the single cell level104 or to detect single
amino acid mutations in breast cancer cells.105
A major issue limiting the use of optical nanoantennas for living cell membrane studies is the
efficient rejection of the background fluorescence light originating from the molecules that
are sufficiently away (tens of nanometers) from the antenna hotspot but still within the
diffraction limited confocal volume.106 For antennas made of individual nanoparticles107–111 or
dimers of nanoparticles112–114 deposited on a glass substrate, the fluorescence background can
be significantly larger than the antenna-enhanced fluorescence signal from the plasmonic
hotspot, challenging single molecule detection and FCS using nanoantennas. The initial
approach to deal with this challenge employed low quantum yield emitters (quantum yield
below 8 %) leading to maximizing the apparent fluorescence enhancement while minimizing
the background110–112,114 Another solution relies on time gating and lifetime filtering, taking
advantage of the reduced lifetime of the emitters in the vicinity of the plasmonic hotspot.115 A
third approach, and the one that we will detail further in this Perspective, uses a dedicated
antenna design termed "antenna-in-box".116,117
The "antenna-in-box" platform features a metal dimer nanogap antenna centered inside a
nanoaperture and is specifically designed for FCS and single molecule analysis at
physiologically relevant (micromolar) concentrations (Fig. 2). The central nanogap antenna
provides the nanoscale plasmonic hotspot, while the surrounding metal cladding screens the
fluorescence background by preventing the excitation of the molecules diffusing away from
the nanogap.116 A challenge associated with classical nanofabrication techniques such as
focused ion beam milling or electron beam lithography is that the region of maximum field
localization is buried into the nanostructure and not directly accessible for fluorescent emitters
embedded in a membrane. We recently overcame this issue by combining electron beam
lithography with planarization, etch back and template stripping.118 The planarization strategy
fills the aperture volume with a transparent polymer, yielding a flat top surface (of a planarity
better than 3 nm, see Fig. 2 c), compatible with membrane studies on living cells. Possible
curvature induced effects on the cell membrane are thus avoided.88,89 The etch back approach
produces reproducible arrays of nanoantennas with controlled gap sizes and sharp edges. With
a gap size of 10 nm, the antenna gap area can be as small as 300 nm2 (Fig. 2 d,e), realizing a
reduction of 200 x as compared to the diffraction-limited confocal area. Lastly, the template
8
stripping flips the plasmonic hotspot to the top surface and place it in the immediate vicinity
of the cell membrane. Owing to these nanofabrication advances, planar plasmonic
nanoantennas drastically improve optical performance leading to fluorescence enhancement
factors above 10,000 x (for crystal violet dyes of 2 % quantum yield) and detection volumes
in the zeptoliter range.118
Figure 2. Planar optical nanoantennas to investigate lipid biological membranes. (a)
Schematics of the experimental arrangement. Arrays of planar optical antennas are fabricated
by electron beam lithography. Each antenna consists of a dimer of gold nanoparticles
separated by a nanometric gap embedded in a polymer filling a rectangular aperture. The gold
dimer confines the excitation light into a nanometric hotspot while the metal cladding
prevents direct excitation of the surrounding fluorescently labeled membrane. The lipid
membrane is directly prepared on top of the planar nanoantennas. (b) Transmission electron
microscope (TEM) image of a representative antenna with 10 nm gap size. (c) Atomic force
microscope (AFM) image of the antenna sample top surface. The topography profile (red
curve) along the antenna axis shows variations in height below 3 nm across the antenna. (d)
Finite-difference time-domain (FDTD) simulations of the electric field intensity profile within
a 10 nm gap antenna for an illumination wavelength of 633 nm. The color scale indicates the
enhancement of the local excitation intensity. (e) Simulations of the observation areas as a
function of the gap size. Each observation area is computed as the product of the gap size
times the full width at half maximum of the intensity profile along the direction perpendicular
to the antenna main axis.
9
We have recently used these planar plasmonic nanoantennas in combination with FCS to
assess the dynamic nanoscale organization of mimetic biological membranes.119 As already
mentioned, tertiary model lipid membranes composed of phospholipids, sphingolipids and
cholesterol separate into stable coexisting Lo and Ld phases which are microscopic in size
and easily observable by diffraction-limited optical microscopy.32–35,120,121 However, there are
intriguing indications that the microscopically homogeneous Lo and Ld phases in model lipid
membranes might in fact be also heterogeneously organized at the nanometer scale,
resembling the scenario occurring in living cells.18,74,122 Indeed, atomistic and coarse-grained
simulations have predicted the existence of highly transient lipid clusters around 10 nm in size
and with microsecond lifetimes within both phases, namely raft and non-raft domains of
multicomponent membranes.123 In support of these simulations, recent NMR experiments
demonstrated that a significant amount of saturated lipids and cholesterol is present in the Ld
phase, and that unsaturated lipids are also found in the Lo phase, strongly pointing towards
the existence of nanoscopic assemblies in both phases.122 However, other workers have
observed the presence of transient nanoscopic domains only in the Lo phase74,124,125, while
others have shown the occurrence of nanoscopic heterogeneities in the Ld phase.126,127
Finally, a STED-FCS study showed no nanodomain formation down to 40nm (set by the
STED resolution) in any of the phases, indicating that both phases are homogeneously
distributed.65
To investigate the potential existence of transient nanoscopic heterogeneities within the
microscopically homogeneous Lo and Ld phases in model lipid membranes we thus took
advantage of the planar plasmonic nanogap antenna platform combined with FCS at various
nanoscale illumination areas.119 We analyzed the diffusion of individual DiD fluorescent
molecules inserted in lipid bilayers composed of the unsaturated phospholipid 1,2-dioleoyl-
sn-glycerol-3-phosphocholine (DOPC) alone, DOPC in combination with sphingomyelin
(SM) (1:1 molar proportions) and of the two ternary mixtures of DOPC, SM (1:1) with
addition of 10 or 20 mol % cholesterol (Chol) (Fig. 3a). Using nanoantennas of gap sizes from
10 to 45 nm, the FCS diffusion laws could be extended down to areas of a few hundreds of
nm2 (Fig. 3b). The results for the diffusion of the dye in pure DOPC membranes indicated
free Brownian diffusion down to the nanoscale, as expected for a homogeneous lipid
distribution. Importantly, these results confirm that the planar nanoantenna platform does not
introduce any artifacts hindering the diffusion of fluorescent dyes. In the presence of
cholesterol, microscopic phase separation occurred into the Lo and Ld phases, enabling the
investigation of putative nanoscopic heterogeneities in both phases. Interestingly, the FCS
10
diffusion laws for both the Lo and Ld phase displayed positive y-axis intercepts t0
significantly deviating from free Brownian diffusion (Fig. 3b). These results indicate the
presence of transient nanoscopic domains in both the Ld and Lo phases of ternary lipid
mixtures with sizes about 10 nm and short characteristic times around 30 μs for the Ld, and
100 μs for the Lo phases (Fig. 3c).119 The extremely short-lived average residence time for the
heterogeneities in the Ld phase is most likely the reason why they have not been detected
before even with high-resolution SPT.74 Although the plasma membrane of living cells bears
a much higher complexity, these nanoscale assemblies in lipid model membranes might
illustrate a general underlying principle setting the basis for lipid raft formation in living cells.
Figure 3. Transient nanodomains in biological model membranes of ternary lipid
mixtures resolved by optical nanoantennas. (a) FCS correlation curves for DOPC:SM(1:1)
+ Cholesterol 20 mol % lipid mixtures recorded for the liquid ordered (Lo) and liquid
disordered (Ld) phases in the confocal setup and with a 12 nm gap antenna. (b) FCS diffusion
laws at the nanoscale for the ternary lipid mixture DOPC:SM(1:1) + Cholesterol 20 mol % in
the Lo and Ld phases. The results for pure DOPC bilayers are shown for comparison. (c) The
positive y-axis intercept t0 (highlighted by red arrows) in (b) reveals the existence of transient
nanoscopic domains in both the Lo and Ld phases for lipid mixtures containing cholesterol.
To extend the applications of planar nanoantennas to membrane studies, we recently used
these platforms in combination with FCS to measure the nanoscale dynamics of different
lipids in fully intact living cells.128 For this, living Chinese hamster ovary (CHO) cells were
incubated in a cell culture well on the antenna platform at 37 °C so that the cells could nicely
11
adhere on the nanoantenna substrate (Fig. 4 a).128 We discuss here the results obtained by
measuring the sphingomyelin (SM) lipid analog labeled with the fluorescent dye Atto647N
before and after methyl-β-cyclodextrin (MCD) treatment. MCD depletes cholesterol from the
cell membrane, which is expected to play a significant role in the formation of lipid rafts.
With the diffraction-limited resolution of the confocal microscope, the typical fluorescence
bursts (Fig.4 b) and FCS traces (Fig.4 c) for both SM before and after MCD treatment (MCD-
SM) showed barely distinguishable features. In stark contrast, clear differences between the
SM and MCD-SM signals were observed with a 12 nm hotspot from the nanoantenna,
indicating the influence of cholesterol in hindering the diffusion of SM.
Figure 4. Lipid membrane organization in living cells probed by optical nanoantennas.
(a) Experimental scheme: live CHO cells are seeded and grown directly on the planar antenna
substrate. (b) Example of fluorescence bursts and FCS correlation traces (c) for
sphingomyelin (SM) before and after cholesterol depletion using MCD treatment, recorded in
the confocal setup and with a 12 nm gap antenna. (d,e) FCS diffusion laws for SM before and
12
after MCD treatment. Linear fits through the median values (continuous lines) are
extrapolated through the y-axis intercept t0 (dashed lines, red arrow). The box plots in (e)
represent the 25th, 50th and 75th percentiles while the bars indicate the 10th and 90th
percentiles (for 48 different antennas).
We further plotted the FCS diffusion laws recorded on nanoantennas for SM and SM after
MCD treatment (Fig. 4 d,e). The slope of the fitted curves allows to determine the diffusion
coefficients to DSM = 0.38 ± 0.19 μm2/s and DMCD-SM = 0.46 ± 0.07 μm2/s, which are
consistent with the confocal measurements. Extrapolating the fits to estimate the time-axis
intercepts, we found for SM a positive t0,SM ~ 110 ± 80 μs, while after MCD treatment, the
intercept comes close to zero with t0, MCD-SM ~ 20 ± 15 μs. It should be pointed out that a
positive time-axis intercept t0 does not exactly correspond to the trapping time of the
nanodomains. As shown by Ruprecht and coworkers50, in the case of immobile nanodomains
and an exponential distribution of trapping and diffusion times, the time offset t0 is the
product of the trapping time τtrap times the fraction β of trapped fluorophores: t0 = β τtrap.
Likewise, the effective diffusion coefficient Deff measured from the FCS diffusion laws can be
expressed Deff = (1-β) Dfree where Dfree is the diffusion coefficient for the free dye. Using the
experimental values measured for SM before and after MCD treatment (and substituting in the
previous equations DSM = Deff and DMCD-SM = Dfree), we obtain β=0.17 and τtrap =0.6 ms. A
slightly modified set of equations allows to take into account also the mobility of the
nanodomains.50 Assuming that the diffusion coefficient for the nanodomains is ten times
slower than for free diffusion Dtrap = Dfree /10, we obtain slightly modified values for the
trapped fraction and trapping time, i.e., β =0.19 and τtrap. =0.9 ms. These results stand in good
agreement with the 1-2~ms trapping time inferred from STED-FCS using an anomalous
diffusion fitting (Fig. S3 of Ref. 22). Altogether, these results indicate the occurrence of
cholesterol-dependent nanodomains hindering SM diffusion in living cell membranes with
sub-millisecond characteristic times and typical sizes below 10 nm (as inferred from the
smallest gap size of our antennas). These experimental observations extend the previous
works20,51 on FCS diffusion laws to the nanoscale dimension far below the diffraction limit.
Taking advantage of the narrow gap sizes down to 10 nm, this approach also allows to explore
membrane organization on areas below the 10-3-10-4 nm2 spatial scale probed by STED-
FCS.22 The nanodomain characteristics stand in good agreement with the predictions from
13
stochastic models18,123,129, and with the current understanding of lipid rafts as highly transient
and fluctuating nanoscale assemblies of sterol and sphingolipids.2,7,52
In conclusion, planar plasmonic nanoantennas with accessible surface nanogaps offer a
promising new approach to investigate the dynamic nanoscale organization of living cell
membranes. The proof-of-principle demonstrations on model lipid membranes119 and CHO
cell membranes128 constitute a significant step forward in our ability to address native
biological membranes with ultrahigh spatiotemporal resolution at the nanometer and
microsecond scales. The nanoantennas provide an encouraging outlook to investigate the
dynamics and interactions of lipids and raft-associated proteins and their recruitment into
molecular complexes. These studies will ultimately improve our understanding of the cell
membrane organization and its link to the cell's function.
Working on multicomponent mimetic biological membranes permits to investigate the
nanoscale dynamic organization of biological membranes and its impact in biological function
in a controllable manner. Future directions involve the addition of more complex components
into the mimetic system such as membrane signaling proteins, components of the glycan
network or a cortical actin mesh. The envisioned next steps in living cells will explore the
native influence of the adjacent inner and outer environment (the cortical actin cytoskeleton
and the glycan network, respectively) on templating the dynamic nanoscale organization of
the plasma membrane. Reaching these goals will also require pushing the nanoantenna
technology even further, to narrow the antenna gap, sharpen the metal edges, improve the
overall reproducibility over the full antenna arrays and enabling multiplexed, parallel
detection from hundreds of antennas simultaneously. Additional challenges comprise the
development of antennas with broadband resonance enabling multi-color fluorescence
detection. Altogether, this outlook preludes a new class of biomolecular studies with ultrahigh
spatial and temporal resolutions, reaching the long-awaited goal of nanometer spatial
precision combined with microsecond temporal resolution and with full biocompatibility.
Additional Information
The authors declare no competing financial interests.
14
Acknowledgements
The research leading to these results has received funding from the European Commission's
Seventh Framework Programme (FP7-ICT-2011-7) under grant agreements ERC StG 278242
(ExtendFRET), 288263 (NanoVista), Spanish Ministry of Economy and Competitiveness
("Severo Ochoa" Programme for Centres of Excellence in (SE - 0 -0 ) and
F S 0 - 0 - ), Fundaci CE E (Barcelona) and CERCA Programme/Generalitat de
Catalunya. P.M.W is supported by the ICFOstepstone Fellowship, a COFUND Doctoral
Programme of the Marie-Sklodowska-Curie-Action of the European Commission. R.R. is
supported by the Erasmus Mundus Doctorate Program Europhotonics (Grant 159224-1-2009-
1-FR-ERA MUNDUS-EMJD).
References
(1)
Brown, D. A.; London, E. Functions of Lipid Rafts in Biological Membranes. Annu.
Rev. Cell Dev. Biol. 1998, 14, 111–136.
(2)
Lingwood, D.; Simons, K. Lipid Rafts as a Membrane-Organizing Principle. Science
2010, 327, 46–50.
(3) Kusumi, A.; Nakada, C.; Ritchie, K.; Murase, K.; Suzuki, K.; Murakoshi, H.; Kasai,
R. S.; Kondo, J.; Fujiwara, T. Paradigm Shift of the Plasma Membrane Concept from
the Two-Dimensional Continuum Fluid to the Partitioned Fluid: High-Speed Single-
Molecule Tracking of Membrane Molecules. Annu. Rev. Biophys. Biomol. Struct.
2005, 34, 351–378.
(4) Gowrishankar, K.; Ghosh, S.; Saha, S.; C., R.; Mayor, S.; Rao, M. Active Remodeling
of Cortical Actin Regulates Spatiotemporal Organization of Cell Surface Molecules.
Cell 2012, 149, 1353–1367.
(5)
Sezgin, E.; Levental, I.; Mayor, S.; Eggeling, C. The Mystery of Membrane
Organization: Composition, Regulation and Roles of Lipid Rafts. Nat. Rev. Mol. Cell
Biol. 2017, 18, 361–374.
(6)
Simons, K.; Ikonen, E. Functional Rafts in Cell Membranes. Nature 1997, 387, 569.
(7) Mayor, S.; Rao, M. Rafts: Scale-Dependent, Active Lipid Organization at the Cell
Surface: Raft Hypothesis. Traffic 2004, 5, 231–240.
(8)
van Zanten, T. S.; Cambi, A.; Koopman, M.; Joosten, B.; Figdor, C. G.; Garcia-
Parajo, M. F. Hotspots of GPI-Anchored Proteins and Integrin Nanoclusters Function
15
as Nucleation Sites for Cell Adhesion. Proc. Natl. Acad. Sci. 2009, 106, 18557–
18562.
(9)
Raghu, H.; Sodadasu, P. K.; Malla, R. R.; Gondi, C. S.; Estes, N.; Rao, J. S.
Localization of uPAR and MMP-9 in Lipid Rafts Is Critical for Migration, Invasion
and Angiogenesis in Human Breast Cancer Cells. BMC Cancer 2010, 10, 647.
(10) Lingwood, D.; Binnington, B.; Róg, T.; Vattulainen, I.; Grzybek, M.; Coskun, Ü.;
Lingwood, C. A.; Simons, K. Cholesterol Modulates Glycolipid Conformation and
Receptor Activity. 2011, 7, 260.
(11) Rios, F. J. O.; Ferracini, M.; Pecenin, M.; Koga, M. M.; Wang, Y.; Ketelhuth, D. F. J.;
Jancar, S. Uptake of oxLDL and IL-10 Production by Macrophages Requires PAFR
and CD36 Recruitment into the Same Lipid Rafts. PLOS ONE 2013, 8, e76893.
(12) Laganowsky, A.; Reading, E.; Allison, T. M.; Ulmschneider, M. B.; Degiacomi, M.
T.; Baldwin, A. J.; Robinson, C. V. Membrane Proteins Bind Lipids Selectively to
Modulate Their Structure and Function. Nature 2014, 510, 172–175.
(13) Farnoud, A. M.; Toledo, A. M.; Konopka, J. B.; Del Poeta, M.; London, E. Raft-Like
Membrane Domains in Pathogenic Microorganisms. Curr. Top. Membr. 2015, 75,
233–268.
(14) Larsen, J. B.; Jensen, M. B.; Bhatia, V. K.; Pedersen, S. L.; Bjørnholm, T.; Iversen,
L.; Uline, M.; Szleifer, I.; Jensen, K. J.; Hatzakis, N. S.; et al. Membrane Curvature
Enables N-Ras Lipid Anchor Sorting to Liquid-Ordered Membrane Phases. 2015, 11,
192.
(15) Varshney, P.; Yadav, V.; Saini, N. Lipid Rafts in Immune Signalling: Current
Progress and Future Perspective. Immunology 2016, 149, 13–24.
(16) Munro, S. Lipid Rafts: Elusive or Illusive? Cell 2003, 115, 377–388.
(17) Pike, L. J. Rafts Defined: A Report on the Keystone Symposium on Lipid Rafts and
Cell Function. J. Lipid Res. 2006, 47, 1597–1598.
(18) Hancock, J. F. Lipid Rafts: Contentious Only from Simplistic Standpoints. Nat Rev
Mol Cell Biol 2006, 7, 456–462.
(19) Marguet, D.; Lenne, P.; Rigneault, H.; He, H. Dynamics in the Plasma Membrane:
How to Combine Fluidity and Order. EMBO J. 2006, 25, 3446.
(20) Lenne, P.-F.; Wawrezinieck, L.; Conchonaud, F.; Wurtz, O.; Boned, A.; Guo, X.-J.;
Rigneault, H.; He, H.-T.; Marguet, D. Dynamic Molecular Confinement in the Plasma
Membrane by Microdomains and the Cytoskeleton Meshwork. EMBO J. 2006, 25,
3245–3256.
16
(21)
Jacobson, K.; Mouritsen, O. G.; Anderson, R. G. W. Lipid Rafts: At a Crossroad
between Cell Biology and Physics. Nat Cell Biol 2007, 9, 7–14.
(22) Eggeling, C.; Ringemann, C.; Medda, R.; Schwarzmann, G.; Sandhoff, K.; Polyakova,
S.; Belov, V. N.; Hein, B.; von Middendorff, C.; Schönle, A.; et al. Direct
Observation of the Nanoscale Dynamics of Membrane Lipids in a Living Cell. Nature
2009, 457, 1159–1162.
(23) Klotzsch, E.; Schütz, G. J. A Critical Survey of Methods to Detect Plasma Membrane
Rafts. Philos. Trans. R. Soc. B Biol. Sci. 2013, 368, 20120033.
(24) Singer, S. J.; Nicolson, G. L. The Fluid Mosaic Model of the Structure of Cell
Membranes. Science 1972, 175, 720.
(25) Goswami, D.; Gowrishankar, K.; Bilgrami, S.; Ghosh, S.; Raghupathy, R.; Chadda,
R.; Vishwakarma, R.; Rao, M.; Mayor, S. Nanoclusters of GPI-Anchored Proteins Are
Formed by Cortical Actin-Driven Activity. Cell 2008, 135, 1085–1097.
(26) Fujiwara, T.; Ritchie, K.; Murakoshi, H.; Jacobson, K.; Kusumi, A. Phospholipids
Undergo Hop Diffusion in Compartmentalized Cell Membrane. J. Cell Biol. 2002,
157, 1071–1082.
(27) Lajoie, P.; Goetz, J. G.; Dennis, J. W.; Nabi, I. R. Lattices, Rafts, and Scaffolds:
Domain Regulation of Receptor Signaling at the Plasma Membrane. J. Cell Biol.
2009, 185, 381–385.
(28) Subramaniam, A. B.; Guidotti, G.; Manoharan, V. N.; Stone, H. A. Glycans Pattern
the Phase Behaviour of Lipid Membranes. Nat. Mater. 2012, 12, 128–133.
(29) Groves, J. T. Cell Membranes: Glycans' mprints. Nat. Mater. 2013, 12, 96–97.
(30) Garcia-Parajo, M. F.; Cambi, A.; Torreno-Pina, J. A.; Thompson, N.; Jacobson, K.
Nanoclustering as a Dominant Feature of Plasma Membrane Organization. J. Cell Sci.
2014, 127, 4995–5005.
(31) Blouin, C. M.; Hamon, Y.; Gonnord, P.; Boularan, C.; Kagan, J.; Viaris de Lesegno,
C.; Ruez, R.; Mailfert, S.; Bertaux, N.; Loew, D.; et al. Glycosylation-Dependent
IFN-γR Partitioning in Lipid and Actin Nanodomains Is Critical for JAK Activation.
Cell 166, 920–934.
(32) Dietrich, C.; Bagatolli, L. A.; Volovyk, Z. N.; Thompson, N. L.; Levi, M.; Jacobson,
K.; Gratton, E. Lipid Rafts Reconstituted in Model Membranes. Biophys. J. 2001, 80,
1417–1428.
(33) Veatch, S. L.; Keller, S. L. Organization in Lipid Membranes Containing Cholesterol.
Phys. Rev. Lett. 2002, 89, 268101.
17
(34) Kahya, N.; Scherfeld, D.; Bacia, K.; Poolman, B.; Schwille, P. Probing Lipid Mobility
of Raft-Exhibiting Model Membranes by Fluorescence Correlation Spectroscopy. J.
Biol. Chem. 2003, 278, 28109–28115.
(35) Chiantia, S.; Ries, J.; Kahya, N.; Schwille, P. Combined AFM and Two-Focus SFCS
Study of Raft-Exhibiting Model Membranes. ChemPhysChem 2006, 7, 2409–2418.
(36) Simons, K.; Vaz, W. L. C. Model Systems, Lipid Rafts, and Cell Membranes. Annu.
Rev. Biophys. Biomol. Struct. 2004, 33, 269–295.
(37) Tamm, L. K.; McConnell, H. M. Supported Phospholipid Bilayers. Biophys. J. 1985,
47, 105–113.
(38) Sezgin, E.; Kaiser, H.-J.; Baumgart, T.; Schwille, P.; Simons, K.; Levental, I.
Elucidating Membrane Structure and Protein Behavior Using Giant Plasma Membrane
Vesicles. Nat Protoc. 2012, 7, 1042–1051.
(39) Baumgart, T.; Hammond, A. T.; Sengupta, P.; Hess, S. T.; Holowka, D. A.; Baird, B.
A.; Webb, W. W. Large-Scale Fluid/Fluid Phase Separation of Proteins and Lipids in
Giant Plasma Membrane Vesicles. Proc. Natl. Acad. Sci. 2007, 104, 3165–3170.
(40) Lingwood, D.; Ries, J.; Schwille, P.; Simons, K. Plasma Membranes Are Poised for
Activation of Raft Phase Coalescence at Physiological Temperature. Proc. Natl. Acad.
Sci. 2008, 105, 10005–10010.
(41) Hammond, A. T.; Heberle, F. A.; Baumgart, T.; Holowka, D.; Baird, B.; Feigenson,
G. W. Crosslinking a Lipid Raft Component Triggers Liquid Ordered-Liquid
Disordered Phase Separation in Model Plasma Membranes. Proc. Natl. Acad. Sci. U.
S. A. 2005, 102, 6320–6325.
(42) Kenworthy, A. K. Fluorescence Recovery After Photobleaching Studies of Lipid
Rafts. In Lipid Rafts; McIntosh, T. J., Ed.; Humana Press: Totowa, NJ, 2007; pp. 179–
192.
(43) Chen, Y.; Lagerholm, B. C.; Yang, B.; Jacobson, K. Methods to Measure the Lateral
Diffusion of Membrane Lipids and Proteins. Anal. Methods Sci. Lipidomics Membr.
Organ. Protein-Lipid Interact. 2006, 39, 147–153.
(44) Dietrich, C.; Yang, B.; Fujiwara, T.; Kusumi, A.; Jacobson, K. Relationship of Lipid
Rafts to Transient Confinement Zones Detected by Single Particle Tracking. Biophys.
J. 82, 274–284.
(45) Bacia, K.; Kim, S. A.; Schwille, P. Fluorescence Cross-Correlation Spectroscopy in
Living Cells. Nat Meth 2006, 3, 83–89.
18
(46) He, H.-T.; Marguet, D. Detecting Nanodomains in Living Cell Membrane by
Fluorescence Correlation Spectroscopy. Annu. Rev. Phys. Chem. 2011, 62, 417–436.
(47) Maiti, S.; Haupts, U.; Webb, W. W. Fluorescence Correlation Spectroscopy:
Diagnostics for Sparse Molecules. Proc. Natl. Acad. Sci. 1997, 94, 11753–11757.
(48) Yechiel, E.; Edidin, M. Micrometer-Scale Domains in Fibroblast Plasma Membranes.
J. Cell Biol. 1987, 105, 755.
(49) Wawrezinieck, L.; Rigneault, H.; Marguet, D.; Lenne, P.-F. Fluorescence Correlation
Spectroscopy Diffusion Laws to Probe the Submicron Cell Membrane Organization.
Biophys. J. 2005, 89, 4029–4042.
(50) Ruprecht, V.; Wieser, S.; Marguet, D.; Schütz, G. J. Spot Variation Fluorescence
Correlation Spectroscopy Allows for Superresolution Chronoscopy of Confinement
Times in Membranes. Biophys. J. 2011, 100, 2839–2845.
(51) Lasserre, R.; Guo, X.-J.; Conchonaud, F.; Hamon, Y.; Hawchar, O.; Bernard, A.-M.;
Soudja, S. M.; Lenne, P.-F.; Rigneault, H.; Olive, D.; et al. Raft Nanodomains
Contribute to Akt/PKB Plasma Membrane Recruitment and Activation. 2008, 4, 538.
(52) Simons, K.; Gerl, M. J. Revitalizing Membrane Rafts: New Tools and Insights. Nat.
Rev. Mol. Cell Biol. 2010, 11, 688–699.
(53) Betzig, E.; Patterson, G. H.; Sougrat, R.; Lindwasser, O. W.; Olenych, S.; Bonifacino,
J. S.; Davidson, M. W.; Lippincott-Schwartz, J.; Hess, H. F. Imaging Intracellular
Fluorescent Proteins at Nanometer Resolution. Science 2006, 313, 1642.
(54) Hess, S. T.; Girirajan, T. P. K.; Mason, M. D. Ultra-High Resolution Imaging by
Fluorescence Photoactivation Localization Microscopy. Biophys. J. 2006, 91, 4258–
4272.
(55) Rust, M. J.; Bates, M.; Zhuang, X. Sub-Diffraction-Limit Imaging by Stochastic
Optical Reconstruction Microscopy (STORM). Nat Meth 2006, 3, 793–796.
(56) Hell, S. W.; Wichmann, J. Breaking the Diffraction Resolution Limit by Stimulated
Emission: Stimulated-Emission-Depletion Fluorescence Microscopy. Opt. Lett. 1994,
19, 780–782.
(57) Klar, T. A.; Jakobs, S.; Dyba, M.; Egner, A.; Hell, S. W. Fluorescence Microscopy
with Diffraction Resolution Barrier Broken by Stimulated Emission. Proc. Natl. Acad.
Sci. 2000, 97, 8206–8210.
(58) Hell, S. W. Far-Field Optical Nanoscopy. Science 2007, 316, 1153–1158.
19
(59) Hwang, J.; Gheber, L. A.; Margolis, L.; Edidin, M. Domains in Cell Plasma
Membranes Investigated by near-Field Scanning Optical Microscopy. Biophys. J.
1998, 74, 2184–2190.
(60) De Lange, F.; Cambi, A.; Huijbens, R.; de Bakker, B.; Rensen, W.; Garcia-Parajo, M.;
van Hulst, N.; Figdor, C. G. Cell Biology beyond the Diffraction Limit: Near-Field
Scanning Optical Microscopy. J. Cell Sci. 2001, 114, 4153–4160.
(61) van Zanten, T. S.; Gómez, J.; Manzo, C.; Cambi, A.; Buceta, J.; Reigada, R.; Garcia-
Parajo, M. F. Direct Mapping of Nanoscale Compositional Connectivity on Intact Cell
Membranes. Proc. Natl. Acad. Sci. 2010, 107, 15437–15442.
(62) van Zanten, T. S.; Cambi, A.; Garcia-Parajo, M. F. A Nanometer Scale Optical View
on the Compartmentalization of Cell Membranes. Biochim. Biophys. Acta BBA -
Biomembr. 2010, 1798, 777–787.
(63) Manzo, C.; van Zanten, T. S.; Garcia-Parajo, M. F. Nanoscale Fluorescence
Correlation Spectroscopy on Intact Living Cell Membranes with NSOM Probes.
Biophys. J. 2011, 100, L8–L10.
(64) Mueller, V.; Ringemann, C.; Honigmann, A.; Schwarzmann, G.; Medda, R.;
Leutenegger, M.; Polyakova, S.; Belov, V. N.; Hell, S. W.; Eggeling, C. STED
Nanoscopy Reveals Molecular Details of Cholesterol- and Cytoskeleton-Modulated
Lipid Interactions in Living Cells. Biophys. J. 2011, 101, 1651–1660.
(65) Honigmann, A.; Mueller, V.; Hell, S. W.; Eggeling, C. STED Microscopy Detects and
Quantifies Liquid Phase Separation in Lipid Membranes Using a New Far-Red
Emitting Fluorescent Phosphoglycerolipid Analogue. Faraday Discuss. 2013, 161,
77–89.
(66) Honigmann, A.; Sadeghi, S.; Keller, J.; Hell, S. W.; Eggeling, C.; Vink, R. A Lipid
Bound Actin Meshwork Organizes Liquid Phase Separation in Model Membranes.
Elife 2014, 3, e01671.
(67) Honigmann, A.; Mueller, V.; Ta, H.; Schoenle, A.; Sezgin, E.; Hell, S. W.; Eggeling,
C. Scanning STED-FCS Reveals Spatiotemporal Heterogeneity of Lipid Interaction in
the Plasma Membrane of Living Cells. Nat. Commun. 2014, 5, 5412.
(68) Vicidomini, G.; Ta, H.; Honigmann, A.; Mueller, V.; Clausen, M. P.; Waithe, D.;
Galiani, S.; Sezgin, E.; Diaspro, A.; Hell, S. W.; et al. STED-FLCS: An Advanced
Tool to Reveal Spatiotemporal Heterogeneity of Molecular Membrane Dynamics.
Nano Lett. 2015, 15, 5912–5918.
20
(69) Sarangi, N. K.; Ayappa, K. G.; Basu, J. K. Complex Dynamics at the Nanoscale in
Simple Biomembranes. Sci. Rep. 2017, 7, 11173.
(70) Manzo, C.; Garcia-Parajo, M. F. A Review of Progress in Single Particle Tracking:
From Methods to Biophysical Insights. Rep. Prog. Phys. 2015, 78, 124601.
(71) Ortega-Arroyo, J.; Kukura, P. Interferometric Scattering Microscopy (iSCAT): New
Frontiers in Ultrafast and Ultrasensitive Optical Microscopy. Phys. Chem. Chem.
Phys. 2012, 14, 15625.
(72) Spillane, K. M.; Ortega-Arroyo, J.; de Wit, G.; Eggeling, C.; Ewers, H.; Wallace, M.
I.; Kukura, P. High-Speed Single-Particle Tracking of GM1 in Model Membranes
Reveals Anomalous Diffusion due to Interleaflet Coupling and Molecular Pinning.
Nano Lett. 2014, 14, 5390–5397.
(73) Spindler, S.; Ehrig, J.; König, K.; Nowak, T.; Piliarik, M.; Stein, H. E.; Taylor, R. W.;
Garanger, E.; Lecommandoux, S.; Alves, I. D.; et al. Visualization of Lipids and
Proteins at High Spatial and Temporal Resolution via Interferometric Scattering
(iSCAT) Microscopy. J. Phys. Appl. Phys. 2016, 49, 274002.
(74) Wu, H.-M.; Lin, Y.-H.; Yen, T.-C.; Hsieh, C.-L. Nanoscopic Substructures of Raft-
Mimetic Liquid-Ordered Membrane Domains Revealed by High-Speed Single-
Particle Tracking. Sci. Rep. 2016, 6, 20542.
(75) Holzmeister, P.; Acuna, G. P.; Grohmann, D.; Tinnefeld, P. Breaking the
Concentration Limit of Optical Single-Molecule Detection. Chem Soc Rev 2014, 43,
1014–1028.
(76) Punj, D.; Ghenuche, P.; Moparthi, S. B.; de Torres, J.; Grigoriev, V.; Rigneault, H.;
Wenger, J. Plasmonic Antennas and Zero-Mode Waveguides to Enhance Single
Molecule Fluorescence Detection and Fluorescence Correlation Spectroscopy toward
Physiological Concentrations. Wiley Interdiscip. Rev. Nanomed. Nanobiotechnol.
2014, 6, 268–282.
(77) Wenger, J.; Rigneault, H. Photonic Methods to Enhance Fluorescence Correlation
Spectroscopy and Single Molecule Fluorescence Detection. Int. J. Mol. Sci. 2010, 11,
206–221.
(78) Levene, M. J.; Korlach, J.; Turner, S. W.; Foquet, M.; Craighead, H. G.; Webb, W. W.
Zero-Mode Waveguides for Single-Molecule Analysis at High Concentrations.
Science 2003, 299, 682.
(79) Genet, C.; Ebbesen, T. W. Light in Tiny Holes. Nature 2007, 445, 39–46.
21
(80) Moran-Mirabal, J. M.; Craighead, H. G. Zero-Mode Waveguides: Sub-Wavelength
Nanostructures for Single Molecule Studies at High Concentrations. Single Mol.
Detect. Ther. Technol. 2008, 46, 11–17.
(81) Samiee, K. T.; Moran-Mirabal, J. M.; Cheung, Y. K.; Craighead, H. G. Zero Mode
Waveguides for Single-Molecule Spectroscopy on Lipid Membranes. Biophys. J.
2006, 90, 3288–3299.
(82) Wenger, J.; Rigneault, H.; Dintinger, J.; Marguet, D.; Lenne, P.-F. Single-Fluorophore
Diffusion in a Lipid Membrane over a Subwavelength Aperture. J. Biol. Phys. 2006,
32, SN1-SN4.
(83) Edel, J. B.; Wu, M.; Baird, B.; Craighead, H. G. High Spatial Resolution Observation
of Single-Molecule Dynamics in Living Cell Membranes. Biophys. J. 2005, 88, L43–
L45.
(84)
Jose M Moran-Mirabal and Alexis J Torres and Kevan T Samiee and Barbara A Baird
and Harold G Craighead. Cell Investigation of Nanostructures: Zero-Mode
Waveguides for Plasma Membrane Studies with Single Molecule Resolution.
Nanotechnology 2007, 18, 195101.
(85) Richards, C. I.; Luong, K.; Srinivasan, R.; Turner, S. W.; Dougherty, D. A.; Korlach,
J.; Lester, H. A. Live-Cell Imaging of Single Receptor Composition Using Zero-Mode
Waveguide Nanostructures. Nano Lett. 2012, 12, 3690–3694.
(86) Wenger, J.; Conchonaud, F.; Dintinger, J.; Wawrezinieck, L.; Ebbesen, T. W.;
Rigneault, H.; Marguet, D.; Lenne, P.-F. Diffusion Analysis within Single Nanometric
Apertures Reveals the Ultrafine Cell Membrane Organization. Biophys. J. 2007, 92,
913–919.
(87) Gérard, D.; Wenger, J.; Bonod, N.; Popov, E.; Rigneault, H.; Mahdavi, F.; Blair, S.;
Dintinger, J.; Ebbesen, T. W. Nanoaperture-Enhanced Fluorescence: Towards Higher
Detection Rates with Plasmonic Metals. Phys. Rev. B 2008, 77, 045413.
(88) Kelly, C. V.; Baird, B. A.; Craighead, H. G. An Array of Planar Apertures for Near-
Field Fluorescence Correlation Spectroscopy. Biophys. J. 2011, 100, L34–L36.
(89) Kelly, C. V.; Wakefield, D. L.; Holowka, D. A.; Craighead, H. G.; Baird, B. A. Near-
Field Fluorescence Cross-Correlation Spectroscopy on Planar Membranes. ACS Nano
2014, 8, 7392–7404.
(90) Schuller, J. A.; Barnard, E. S.; Cai, W.; Jun, Y. C.; White, J. S.; Brongersma, M. L.
Plasmonics for Extreme Light Concentration and Manipulation. Nat Mater 2010, 9,
193–204.
22
(91) Novotny, L.; van Hulst, N. Antennas for Light. Nat. Photonics 2011, 5, 83–90.
(92) Biagioni, P.; Huang, J.-S.; Hecht, B. Nanoantennas for Visible and Infrared Radiation.
Rep. Prog. Phys. 2012, 75, 024402.
(93) Halas, N. J.; Lal, S.; Chang, W.-S.; Link, S.; Nordlander, P. Plasmons in Strongly
Coupled Metallic Nanostructures. Chem. Rev. 2011, 111, 3913–3961.
(94) Koenderink, A. F. Single-Photon Nanoantennas. ACS Photonics 2017, 4, 710–722.
(95) Kinkhabwala, A.; Yu, Z.; Fan, S.; Avlasevich, Y.; Müllen, K.; Moerner, W. E. Large
Single-Molecule Fluorescence Enhancements Produced by a Bowtie Nanoantenna.
Nat. Photonics 2009, 3, 654–657.
(96) Acuna, G. P.; Moller, F. M.; Holzmeister, P.; Beater, S.; Lalkens, B.; Tinnefeld, P.
Fluorescence Enhancement at Docking Sites of DNA-Directed Self-Assembled
Nanoantennas. Science 2012, 338, 506–510.
(97) Puchkova, A.; Vietz, C.; Pibiri, E.; Wünsch, B.; Sanz Paz, M.; Acuna, G. P.;
Tinnefeld, P. DNA Origami Nanoantennas with over 5000-Fold Fluorescence
Enhancement and Single-Molecule Detection at 25 μM. Nano Lett. 2015, 15, 8354–
8359.
(98) Yuan, H.; Khatua, S.; Zijlstra, P.; Yorulmaz, M.; Orrit, M. Thousand-Fold
Enhancement of Single-Molecule Fluorescence Near a Single Gold Nanorod. Angew.
Chem. Int. Ed. 2013, 52, 1217–1221.
(99) Khatua, S.; Paulo, P. M. R.; Yuan, H.; Gupta, A.; Zijlstra, P.; Orrit, M. Resonant
Plasmonic Enhancement of Single-Molecule Fluorescence by Individual Gold
Nanorods. ACS Nano 2014, 8, 4440–4449.
(100) Lohmüller, T.; Iversen, L.; Schmidt, M.; Rhodes, C.; Tu, H.-L.; Lin, W.-C.; Groves, J.
T. Single Molecule Tracking on Supported Membranes with Arrays of Optical
Nanoantennas. Nano Lett. 2012, 12, 1717–1721.
(101) Flauraud, V.; van Zanten, T. S.; Mivelle, M.; Manzo, C.; Garcia Parajo, M. F.;
Brugger, J. Large-Scale Arrays of Bowtie Nanoaperture Antennas for Nanoscale
Dynamics in Living Cell Membranes. Nano Lett. 2015, 15, 4176–4182.
(102) Flynn, J. D.; Haas, B. L.; Biteen, J. S. Plasmon-Enhanced Fluorescence from Single
Proteins in Living Bacteria. J. Phys. Chem. C 2016, 120, 20512–20517.
(103) Fabrizio, E. D.; Schlücker, S.; Wenger, J.; Regmi, R.; Rigneault, H.; Calafiore, G.;
West, M.; Cabrini, S.; Fleischer, M.; van Hulst, N. F.; et al. Roadmap on Biosensing
and Photonics with Advanced Nano-Optical Methods. J. Opt. 2016, 18, 063003.
23
(104) Perozziello, G.; Candeloro, P.; De Grazia, A.; Esposito, F.; Allione, M.; Coluccio, M.
L.; Tallerico, R.; Valpapuram, I.; Tirinato, L.; Das, G.; et al. Microfluidic Device for
Continuous Single Cells Analysis via Raman Spectroscopy Enhanced by Integrated
Plasmonic Nanodimers. Opt. Express 2016, 24, A180–A190.
(105) Coluccio, M. L.; Gentile, F.; Das, G.; Nicastri, A.; Perri, A. M.; Candeloro, P.;
Perozziello, G.; Proietti Zaccaria, R.; Gongora, J. S. T.; Alrasheed, S.; et al. Detection
of Single Amino Acid Mutation in Human Breast Cancer by Disordered Plasmonic
Self-Similar Chain. Sci. Adv. 2015, 1, e1500487.
(106) Langguth, L.; Femius Koenderink, A. Simple Model for Plasmon Enhanced
Fluorescence Correlation Spectroscopy. Opt. Express 2014, 22, 15397–15409.
(107) Estrada, L. C.; Aramendía, P. F.; Martínez, O. E. 10000 Times Volume Reduction for
Fluorescence Correlation Spectroscopy Using Nano-Antennas. Opt. Express 2008, 16,
20597–20602.
(108) Wang, Q.; Lu, G.; Hou, L.; Zhang, T.; Luo, C.; Yang, H.; Barbillon, G.; Lei, F. H.;
Marquette, C. A.; Perriat, P.; et al. Fluorescence Correlation Spectroscopy near
Individual Gold Nanoparticle. Chem. Phys. Lett. 2011, 503, 256–261.
(109) Lu, G.; Liu, J.; Zhang, T.; Li, W.; Hou, L.; Luo, C.; Lei, F.; Manfait, M.; Gong, Q.
Plasmonic near-Field in the Vicinity of a Single Gold Nanoparticle Investigated with
Fluorescence Correlation Spectroscopy. Nanoscale 2012, 4, 3359–3364.
(110) Punj, D.; de Torres, J.; Rigneault, H.; Wenger, J. Gold Nanoparticles for Enhanced
Single Molecule Fluorescence Analysis at Micromolar Concentration. Opt. Express
2013, 21, 27338.
(111) Khatua, S.; Yuan, H.; Orrit, M. Enhanced-Fluorescence Correlation Spectroscopy at
Micro-Molar Dye Concentration around a Single Gold Nanorod. Phys Chem Chem
Phys 2015, 17, 21127–21132.
(112) Kinkhabwala, A. A.; Yu, Z.; Fan, S.; Moerner, W. E. Fluorescence Correlation
Spectroscopy at High Concentrations Using Gold Bowtie Nanoantennas. Single Mol.
Spectrosc. Curr. Status Perspect. 2012, 406, 3–8.
(113) Dutta Choudhury, S.; Ray, K.; Lakowicz, J. R. Silver Nanostructures for Fluorescence
Correlation Spectroscopy: Reduced Volumes and Increased Signal Intensities. J. Phys.
Chem. Lett. 2012, 3, 2915–2919.
(114) Punj, D.; Regmi, R.; Devilez, A.; Plauchu, R.; Moparthi, S. B.; Stout, B.; Bonod, N.;
Rigneault, H.; Wenger, J. Self-Assembled Nanoparticle Dimer Antennas for
24
Plasmonic-Enhanced Single-Molecule Fluorescence Detection at Micromolar
Concentrations. ACS Photonics 2015, 2, 1099–1107.
(115) Pradhan, B.; Khatua, S.; Gupta, A.; Aartsma, T.; Canters, G.; Orrit, M. Gold-
Nanorod-Enhanced Fluorescence Correlation Spectroscopy of Fluorophores with High
Quantum Yield in Lipid Bilayers. J. Phys. Chem. C 2016, 120, 25996–26003.
(116) Punj, D.; Mivelle, M.; Moparthi, S. B.; van Zanten, T. S.; Rigneault, H.; van Hulst, N.
F.; García-Paraj , M. F.; Wenger, J. A Plasmonic "antenna-in-Box" Platform for
Enhanced Single-Molecule Analysis at Micromolar Concentrations. Nat.
Nanotechnol. 2013, 8, 512–516.
(117) Ghenuche, P.; de Torres, J.; Moparthi, S. B.; Grigoriev, V.; Wenger, J. Nanophotonic
Enhancement of the Förster Resonance Energy-Transfer Rate with Single
Nanoapertures. Nano Lett. 2014, 14, 4707–4714.
(118) Flauraud, V.; Regmi, R.; Winkler, P. M.; Alexander, D. T. L.; Rigneault, H.; van
Hulst, N. F.; García-Parajo, M. F.; Wenger, J.; Brugger, J. In-Plane Plasmonic
Antenna Arrays with Surface Nanogaps for Giant Fluorescence Enhancement. Nano
Lett. 2017, 1703–1710.
(119) Winkler, P. M.; Regmi, R.; Flauraud, V.; Brugger, J.; Rigneault, H.; Wenger, J.;
García-Parajo, M. F. Transient Nanoscopic Phase Separation in Biological Lipid
Membranes Resolved by Planar Plasmonic Antennas. ACS Nano 2017, 11, 7241–
7250.
(120) Brown, D. A. Seeing Is Believing: Visualization of Rafts in Model Membranes. Proc.
Natl. Acad. Sci. 2001, 98, 10517–10518.
(121) Bagatolli, L. A.; Gratton, E. Two Photon Fluorescence Microscopy of Coexisting
Lipid Domains in Giant Unilamellar Vesicles of Binary Phospholipid Mixtures.
Biophys. J. 2000, 78, 290–305.
(122) Yasuda, T.; Tsuchikawa, H.; Murata, M.; Matsumori, N. Deuterium NMR of Raft
Model Membranes Reveals Domain-Specific Order Profiles and Compositional
Distribution. Biophys. J. 2015, 108, 2502–2506.
(123) Apajalahti, T.; Niemela, P.; Govindan, P. N.; Miettinen, M. S.; Salonen, E.; Marrink,
S.-J.; Vattulainen, I. Concerted Diffusion of Lipids in Raft-like Membranes. Faraday
Discuss. 2010, 144, 411–430.
(124) Sodt, A. J.; Sandar, M. L.; Gawrisch, K.; Pastor, R. W.; Lyman, E. The Molecular
Structure of the Liquid-Ordered Phase of Lipid Bilayers. J. Am. Chem. Soc. 2014,
136, 725–732.
25
(125) Sodt, A. J.; Pastor, R. W.; Lyman, E. Hexagonal Substructure and Hydrogen Bonding
in Liquid-Ordered Phases Containing Palmitoyl Sphingomyelin. Biophys. J. 2015,
109, 948–955.
(126) Silvius, J. R. Fluorescence Energy Transfer Reveals Microdomain Formation at
Physiological Temperatures in Lipid Mixtures Modeling the Outer Leaflet of the
Plasma Membrane. Biophys. J. 2003, 85, 1034–1045.
(127) de Almeida, R. F. M.; Loura, L. M. S.; Fedorov, A.; Prieto, M. Lipid Rafts Have
Different Sizes Depending on Membrane Composition: A Time-Resolved
Fluorescence Resonance Energy Transfer Study. J. Mol. Biol. 2005, 346, 1109–1120.
(128) Regmi, R.; Winkler, P. M.; Flauraud, V.; Borgman, K. J. E.; Manzo, C.; Brugger, J.;
Rigneault, H.; Wenger, J.; García-Parajo, M. F. Planar Optical Nanoantennas Resolve
Cholesterol-Dependent Nanoscale Heterogeneities in the Plasma Membrane of Living
Cells. Nano Lett. 2017, 17, 6295–6302.
(129) Nicolau, D. V.; Burrage, K.; Parton, R. G.; Hancock, J. F. Identifying Optimal Lipid
Raft Characteristics Required To Promote Nanoscale Protein-Protein Interactions on
the Plasma Membrane. Mol. Cell. Biol. 2006, 26, 313–323.
26
|
1302.2665 | 1 | 1302 | 2013-02-11T23:03:38 | Travelling waves in hybrid chemotaxis models | [
"physics.bio-ph",
"math.AP",
"q-bio.PE"
] | Hybrid models of chemotaxis combine agent-based models of cells with partial differential equation models of extracellular chemical signals. In this paper, travelling wave properties of hybrid models of bacterial chemotaxis are investigated. Bacteria are modelled using an agent-based (individual-based) approach with internal dynamics describing signal transduction. In addition to the chemotactic behaviour of the bacteria, the individual-based model also includes cell proliferation and death. Cells consume the extracellular nutrient field (chemoattractant) which is modelled using a partial differential equation. Mesoscopic and macroscopic equations representing the behaviour of the hybrid model are derived and the existence of travelling wave solutions for these models is established. It is shown that cell proliferation is necessary for the existence of non-transient (stationary) travelling waves in hybrid models. Additionally, a numerical comparison between the wave speeds of the continuum models and the hybrid models shows good agreement in the case of weak chemotaxis and qualitative agreement for the strong chemotaxis case. In the case of slow cell adaptation, we detect oscillating behaviour of the wave, which cannot be explained by mean-field approximations. | physics.bio-ph | physics | Submitted to Bulletin of Mathematical Biology
Travelling waves in hybrid chemotaxis models
Benjamin Franz · Chuan Xue · Kevin J.
Painter · Radek Erban
Preprint version: April 3, 2018
Abstract Hybrid models of chemotaxis combine agent-based models of cells
with partial differential equation models of extracellular chemical signals. In
this paper, travelling wave properties of hybrid models of bacterial chemo-
taxis are investigated. Bacteria are modelled using an agent-based (individual-
based) approach with internal dynamics describing signal transduction. In ad-
dition to the chemotactic behaviour of the bacteria, the individual-based model
also includes cell proliferation and death. Cells consume the extracellular nu-
trient field (chemoattractant) which is modelled using a partial differential
equation. Mesoscopic and macroscopic equations representing the behaviour
of the hybrid model are derived and the existence of travelling wave solutions
for these models is established. It is shown that cell proliferation is neces-
sary for the existence of non-transient (stationary) travelling waves in hybrid
models. Additionally, a numerical comparison between the wave speeds of the
continuum models and the hybrid models shows good agreement in the case of
weak chemotaxis and qualitative agreement for the strong chemotaxis case. In
the case of slow cell adaptation, we detect oscillating behaviour of the wave,
which cannot be explained by mean-field approximations.
Keywords hybrid model · travelling wave · bacterial chemotaxis
Benjamin Franz · Radek Erban
Mathematical Institute, University of Oxford
24-29 St. Giles', Oxford, OX1 3LB, United Kingdom
E-mail: [email protected], [email protected]
Chuan Xue
Department of Mathematics, Ohio State University
231 West 18th Avenue, Columbus, OH 43210, USA
E-mail: [email protected]
Kevin J. Painter
Department of Mathematics, Heriot-Watt University
Edinburgh, EH14 4AS, United Kingdom
E-mail: [email protected]
3
1
0
2
b
e
F
1
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
5
6
6
2
.
2
0
3
1
:
v
i
X
r
a
2
1 Introduction
Franz, Xue, Painter, Erban
The wavelike spread of cell populations plays a fundamental role in many bio-
logical processes, including development [24], wound healing [38] and tumour
invasion [16]. Bacterial populations show similar phenomena, with the pio-
neering studies of Adler [1] confirming the capacity of an E. coli population to
form travelling bands via chemotaxis to extracellular signals. Mathematically,
the extent to which chemotaxis can generate and sustain stationary travelling
bands has motivated a number of studies, including the Keller-Segel model of
Adler's experiments which is written in the form of coupled partial differential
equations (PDEs) [20]. This early model necessitated a biologically unrealis-
tic singularity in the chemotactic sensitivity to generate stationary travelling
waves: a requirement that allows bacteria behind the wave to acquire infinite
speeds and to avoid "dropping-out", an effect that leads to gradual dispersal
of the band [40,15].
This singularity requirement can be circumvented by incorporating other
processes. The well known Fisher's equation [14] demonstrates travelling waves
in systems coupling diffusion with logistic growth terms [14]. Parabolic chemo-
taxis models with non-singular sensitivities but incorporating either logistic
[22,23,30] or non-logistic [21,36] growth terms also admit travelling wave so-
lutions. Other studies have shown that introduction of more complex nutrient
terms can give rise to travelling waves, even when growth is absent [34,35]. An
experimental system which also included two chemicals – a chemoattractant
and a nutrient source – was presented in [6,7], with stationary or transient
travelling waves obtained according to t he formulation of the model [5,40].
Travelling waves in chemotaxis models have also been recently studied in [26,
25]; we also note the articles [19] and [37] for a review and analysis of travelling
waves in PDE-based models. A comparison between mesoscopic (hyperbolic)
and macroscopic (parabolic) PDEs has been presented in [27].
Relatively little exploration has been conducted into travelling wave for-
mation for chemotactic models extending beyond PDE systems, in particular
those introducing terms to account for the inherent noise of biological systems.
One exception is the study of [9], in which a multiplicative noise term was in-
troduced into the Keller-Segel model and the existence of travelling waves has
been demonstrated within this setting. Hybrid models, in which an individual-
based model for bacterial behaviour is coupled to a continuum description of
extracellular signals, naturally introduce stochastic effects and will be the focus
of the present paper. Such a hybrid model was formulated in [15] where it was
shown that under finite cell speeds only transient travelling waves formed, even
with singular chemotactic sensitivity. The individual-based model was formu-
lated in terms of the velocity-jump model with internal dynamics [12,13,41]
and, in this paper, we extend the model in [15] to incorporate proliferation
and death of bacteria. We analyse this system numerically and analytically
with respect to its travelling wave properties, employing the biologically in-
spired chemotactic sensitivity presented in [40] and a linear growth term. We
Travelling waves in hybrid chemotaxis models
3
show that stationary travelling waves can be observed even in the absence of
chemotaxis, although wave speeds are substantially increased in its presence.
The organisation of the paper is as follows: the full hybrid model is pre-
sented in Section 2 along with illustrative simulation results, while the corre-
sponding continuum equations are derived under certain assumptions in Sec-
tion 3; in Section 4 these continuum equations are analysed with respect to
travelling wave properties; in Section 5 where a computational analysis and
comparison of the models is presented; finally, we discuss our observations in
Section 6.
2 Hybrid model of bacterial chemotaxis
In this section we formulate the hybrid model of bacterial chemotaxis which
will be investigated in this paper. The model is motivated by the behaviour
of the bacterium E.coli and, in its most general form, includes cell movement,
sensing and response to a chemical signal, consumption of the chemoattrac-
tant, cell proliferation and death. However, for analytical tractability, we will
also explore simplified hybrid models which exclude some of these processes.
Bacteria are modelled as agents with internal dynamics that represent the sig-
nal processing and response of each individual while the extracellular chemical
is modelled using a PDE to describe its spatio-temporal concentration. The
mathematical framework and simulation techniques are reviewed in [15]. We
consider the model in an effectively one-dimensional domain representing a
long but narrow tube, similar to the experimental set up considered in [1].
The motion of E. coli bacteria is controlled through the coordinated rota-
tion of flagella distributed over the cell surface [2]. Counterclockwise rotation
generates a propulsive bundle that results in straight line motion of the bac-
terium – a so-called "run" [3]. Alternatively, clockwise rotation results in the
outward flaying of flagella and a "tumble" – rotation with insignificant dis-
placement. At the end of each tumble the bacterium chooses a new direction
of movement, seemingly at random, and returns to the run phase. The lengths
of the individual phases are independent from each other and distributed ex-
ponentially, yet they can be influenced by internal dynamics [2].
Internal dynamics of the E. coli bacteria possess two principal features [4]:
a quick excitation phase followed by slower adaptation. Specifically, changes in
the extracellular signal concentration lead to quick excitation of the internal
metabolism, signified through altered chemical concentrations inside the cell.
Following excitation the internal concentrations revert slowly to normal in an
adaptation process, even when the external signal remains at the raised level.
2.1 Velocity jump model with internal dynamics
Run-and-tumble dynamics are aptly modelled as a velocity-jump process [31,
12]. We denote by Na(t) the number of bacteria (agents) in the system at time
4
Franz, Xue, Painter, Erban
t. The current state of the i-th individual, i = 1, 2, . . . , Na(t), will be described
using its position xi ∈ R, its velocity vi = ±s ∈ R and a set of internal state
variables yi ∈ Rm that represent the states of components in the intracellular
signal transduction network.
Here we concentrate on a cartoon version of the internal dynamics of bac-
teria written in terms of two internal variables [32,12], i.e m = 2. Internal
variables y(1) and y(2) are governed by the equations
dy(1)
dt
dy(2)
dt
=
=
S(x(t), t) − y(1) − y(2)
te
,
S(x(t), t) − y(2)
ta
,
(2.1)
where te is the excitation time, ta is the adaptation time, te ≪ ta and S(x(t), t)
is the concentration of chemoattractant at the position of the bacterium x(t)
at time t. Furthermore, bacteria move with the velocity vi = ±s governed
through a velocity jump process with a turning frequency λ = λ(y) that
depends on the internal dynamics. In this paper, we will use the biologically
motivated nonlinear turning kernel developed in [40]. Hence, the full model of
one individual over (a small) time step ∆t can be written as:
x(t + ∆t) = x(t) + v(t) ∆t,
v(t + ∆t) = (cid:26)−v(t), with probability λ(y(t)) ∆t,
λ(y(t)) = λ0(cid:18)1 −
v(t), otherwise ,
y(1)(t)
κ + y(1)(t)(cid:19) ,
y(1)(t + ∆t) = y(1)(t) +
S(x(t), t) − y(1)(t) − y(2)(t)
te
y(2)(t + ∆t) = y(2)(t) +
S(x(t), t) − y(2)(t)
ta
∆t,
(2.2)
(2.3)
(2.4)
∆t,
(2.5)
(2.6)
where λ0 and κ are positive constants.
In addition to the behaviour of an individual bacterium we define a signal-
dependent proliferation function h(S) : R+ 7→ R. We thereby interpret a
positive value of h(S) as a proliferation rate, meaning that in the infinitesimal
interval [t, t + ∆t) a bacterium at position x generates an exact copy of itself
with probability h(S(x(t), t)) ∆t. Similarly, a negative value of h(S) means
that the bacterium disappears (dies) with the probability −h(S(x(t), t)) ∆t.
In this paper, we will use the following form for the proliferation rate h(S):
h(S) = α(S − Sc) ,
(2.7)
where α and Sc are positive constants.
Travelling waves in hybrid chemotaxis models
5
2.2 Evolution of the extracellular chemoattractant
For the extracellular signal S(x, t) we formulate a PDE that incorporates dif-
fusion (with diffusion constant DS ≥ 0) and signal consumption by bacteria,
the latter with signal dependent rate k(S) : R+ → R+. The equation for S
therefore takes the form
∂S
∂t
= DS
∂2S
∂x2 − k(S)
Na(t)
Xi=1
δ(x − xi(t)) .
(2.8)
For the remainder of the paper we employ a linear form for the consumption
function k(S):
k(S) = βS ,
(2.9)
where β is a positive constant.
2.3 Illustrative example
The hybrid model framework presented in Sections 2.1 and 2.2 includes es-
sential features of the more complicated hybrid chemotaxis models formulated
in [10,39]. In this section we numerically show that these processes can give
rise to travelling waves. For the numerical simulation we employ techniques
described in [15]. In particular, for the extracellular signal S(x, t), this means
that the simulation is performed on the one-dimensional domain [0, L] with
initial condition S(x, 0) = S∞ > 0 and zero-flux boundary conditions. We
consider M + 1 regularly spaced grid points rj = j ∆x, j = 0, . . . , M, where
∆x = L/M and the values of S(xi, t) are advanced by a small time step ∆t
and a forward Euler update rule:
S(rj, t + ∆t) = S(rj, t) + DS ∆t
S(rj−1, t) + S(rj+1, t) − 2S(rj, t)
(∆x)2
− k(S(rj, t)) ∆t
Na(t)
Xi=1
K(rj − xi(t)) .
(2.10)
In the above K : R → R+ is the symmetric, normalised and non-negative
kernel
K(ξ) =
1
√2πσ2
exp(cid:20)−
ξ2
2σ2(cid:21) ,
where the kernel width σ is a positive real number. Here, K(rj − xi) represents
the influence a bacterium at position xi has on grid point j.
The simulation of the individual bacterium is given in the full system (2.2)–
(2.6) and complemented by the birth and death processes described in Sec-
tion 2.1, where we use the same time step ∆t as in (2.10). To calculate the
necessary off-grid values of extracellular signal, we linearly interpolate from the
6
Franz, Xue, Painter, Erban
two nearest grid points. We further simplify the system (2.2)–(2.6) by exploit-
ing the separate time scales for excitation and adaptation (i.e. te ≪ ta): specif-
ically, we assume the update equation (2.5) for y(1) is in a quasi-equilibrium,
which is identical to the assumption te = 0. The value for y(1) can therefore
be calculated by
y(1)(t) = S(x(t), t) − y(2)(t) .
(2.11)
Illustrative results are presented in Figure 1. For this simulation, Na(0) = 104
bacteria were initialised at positions xi(0), randomly generated as the absolute
value of a Gaussian random variable with variance much smaller than the
domain length L. The initial velocity (direction of movement) is generated
uniformly at random and initial values of the extracellular signal and internal
variables are taken as
y(1)
i
(0) = S∞, y(2)
i
(0) = 0,
S(x, 0) ≡ S∞
for i = 1, 2, . . . , Na(0),
for x ∈ [0, L],
where S∞ = 1. We simulate the system until time Tfinal = 100 and plot both
the distribution of bacteria and concentration of chemoattractant S in Figure
1(a). We also estimate the wave speed as a function of time in Figure 1(b).
We clearly see formation of a travelling band of bacteria, moving rightwards
with average speed v = 0.51 (plotted as the dashed line in Figure 1(b)).
Influence of the growth term
To investigate the influence of the growth term on the existence of travel-
ling waves, we simulate the full hybrid model (2.2)–(2.6) and (2.10) including
(α = 1) and excluding (α = 0) growth and death processes. We use identical
parameters to those described above and present the results in Figure 2. In
Figure 2(a) the position of the wave front (defined as the right-most position
for which S(x) < 0.9) is compared. The full hybrid system (dashed line) gen-
erates a straight line, indicating a wave moving with constant speed. While
the system excluding growth and death (solid line) moves with a similar initial
speed, speed is gradually lost over time: the shape of n(x, t) at different times
for this case is shown in Figure 2(b). We clearly see that no true travelling
wave forms, with many agents being left far behind the wave front, leading to
its slowing down. Thus, we can interpret growth and death terms in terms of
a stabilising role on the wave profile: although not all agents can keep up with
the wave, new agents are constantly created at the front and the agents that
drop out eventually die, resulting in a travelling band of agents.
3 From hybrid models to macroscopic PDEs
In this section we derive macroscopic PDEs for the spatio-temporal density of
bacteria n(x, t) at given position x ∈ R and time t ≥ 0. An implicit assumption
of the derivation is spatial independence of bacteria, which allows formulation
Travelling waves in hybrid chemotaxis models
7
(a)
S
,
n
(c)
S
,
n
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
20
40
x
60
80
100
(b)
)
t
(
x
1
0.8
0.6
0.4
0.2
0
0
1
(d)
l
s
s
a
m
d
e
a
c
s
e
r
,
)
t
(
v
0.8
0.6
0.4
0.2
20
40
20
40
x
60
80
100
0
0
20
40
60
80
100
60
80
100
t
t
Fig. 1 Numerical solutions of the hybrid chemotaxis model (2.2)–(2.6) and (2.10) and PDE
System A (3.1)–(3.3).
(a) Wave form for the hybrid model after time t = 100. Solid line: estimated density of
bacteria, dashed line: extracellular chemical signal S.
(b) Measured speed of travelling wave (solid line). Dashed line denotes the average speed.
(c) Wave form for PDE system A after time t = 100. Solid line: estimated density of
bacteria, dashed line: extracellular chemical signal S.
(d) Measured speed of travelling wave (solid line) for PDE System A. Note that the spike
near t = 0 is a product of the wave speed calculation method.
The dimensionless parameters are: α = β = s = 1, Sc = 0.5, S∞ = 1, ∆t = 10−3,
∆x = 0.25, L = 100, λ0 = 10, κ = 0.01, DS = 0, ta = 0.1, σ = 0.5.
of a continuous mesoscopic system. We then use results from [12] to obtain
the macroscopic equations. To illustrate the successive formulation of models
we construct two systems of PDEs – denoted System (A) and System (B) –
to be referred to in the remainder of the paper.
3.1 System (A)
We define the mesoscopic densities p±(x, y(2), t) for left and right-moving bac-
teria, depending on their position x ∈ R, their internal variable y(2) ∈ R and
t ≥ 0. If the signal profile S ≡ S(x, t) was uninfluenced by bacteria, densities
8
(a)
60
50
40
)
t
(
x
30
20
10
0
0
(b)
0.2
0.15
)
x
(
n
0.1
0.05
0
0
Franz, Xue, Painter, Erban
t1 = 20
✁
✁
t2 = 40
t3 = 60
✟✟ t4 = 80
10
20
x
30
40
20
40
t
60
80
100
Fig. 2 Numerical solutions of the hybrid chemotaxis model (2.2)–(2.6) and (2.10) without
growth and death terms.
(a) Comparison of position of wave front over time. Solid line: without growth/death (α =
0), dashed line: with growth/death (α = 1).
(b) Wave form at different times during simulation with α = 0. From left to right: t =
20, 40, 60, 80.
Remaining parameters as in Figure 1.
p± would satisfy the following system of hyperbolic PDEs:
+ s
∂p+
∂t
∂p−
∂t − s
∂p+
∂x
∂p−
∂x
+
+
∂
∂
∂y(2) (cid:18) S(x, t) − y(2)
∂y(2) (cid:18) S(x, t) − y(2)
ta
ta
p+(cid:19) = − λp+ + λp− + h(S(x, t))p+ ,
p−(cid:19) =+λp+ − λp− + h(S(x, t))p− ,
where λ is defined in (2.4) which, under (2.11), can be simplified to
λ = λ0(cid:18)1 −
S(x) − y(2)
κ + S(x) − y(2)(cid:19) .
(3.1)
(3.2)
The signal dynamics is described by (2.8) which can be rewritten in terms of
p± as
∂S
∂t
= DS
∂2S
∂x2 − k(S)ZR
(p+ + p−)dy(2) .
(3.3)
We denote the system of equations (3.1)–(3.3) as System (A).
The system (3.1) (for the one-particle distribution) can be derived by inte-
grating the probability distribution function p(x1, v1, y1; x2, v2, y2; . . . S(x, t))
for the many particle system, utilizing the fact that the movement of individ-
uals are biased by the signal function S(x, t), but independent to each other.
However, for the hybrid chemotaxis models described in Sections 2.1 and 2.2,
individual bacteria interact via the extracellular signal S which complicates
the derivation of (3.1). In [11], a kinetic description has been derived for a
model of interacting locusts, using a modified version of the BBGKY hier-
archy from the classical kinetic theory of gases [8]. The system we consider
here is much more complicated to analyse than the locust model studied in
[11], due to the variable number of bacteria and internal variables. Thus the
Travelling waves in hybrid chemotaxis models
9
kinetic description (3.1) can only be considered as an approximation to the
one particle distributions of the interacting system.
The capacity of the above mesoscopic system to generate travelling bands
analogous to those observed in the hybrid model is illustrated in Figure 1(c)-
(d). For details of the numerical method employed for this and other simula-
tions of the continuous model, we refer to [40]. The qualitatively and quan-
titatively close correspondence in solutions under equivalent parameters and
initial conditions corroborates the use of the above approximation.
3.2 System (B)
We consider a macroscopic model in this section. Define the macroscopic den-
sities
p±(x, t) = ZR
p±(x, y(2), t)dy(2),
(3.4)
and let them satisfy the following system
+ s
∂p+
∂t
∂p−
∂t − s
∂p+
∂x
∂p−
∂x
= −λ+(cid:18) ∂S
= +λ+(cid:18) ∂S
∂x(cid:19) p+ + λ−(cid:18) ∂S
∂x(cid:19) p+ − λ−(cid:18) ∂S
∂x(cid:19) p− + h(S)p+ ,
∂x(cid:19) p− + h(S)p− ,
(3.5)
where the turning rates λ± are given by
λ± = λ0(cid:18)1 ∓ χ
∂S
∂x(cid:19) with χ =
sta
κλ0(1 + 2λ0ta)
.
(3.6)
Using (3.4), equation (3.3) can be written as
∂S
∂t
= DS
∂2S
∂x2 − k(S)(p+ + p−) .
(3.7)
We will denote (3.5) and (3.7) along with the definition of λ± in (3.6) as
System (B). According to the analysis in [12,41], System (B) is quantita-
tively consistent with System (A) when the external signal S(x) changes slow
enough such that cells are close to their fully adapted state, in which case cell
movement is only moderately modified by the signal.
In the rest of the paper, we assume diffusion of extracellular signal to
occur on a much slower time scale than the active motion of the bacteria,
hence DS = 0. The number of parameters of the above models can be reduced
by setting s, S∞, α, β to one through rescaling. We show this in detail for
System (B) as follows. Rescaling the variables S = SS∞, p± = p±αS∞/β,
t = t/(αS∞), x = xs/(αS∞) and the parameters Sc = ScS∞, λ0 = λ0αS∞,
10
Franz, Xue, Painter, Erban
taking (2.7) and substituting into System (B) we obtain, after dropping hats
for notational simplicity,
+
∂p+
∂t
∂p−
∂t −
∂p+
∂x
∂p−
∂x
∂S
∂t
∂x(cid:19) p+ + λ−(cid:18) ∂S
∂x(cid:19) p+ − λ−(cid:18) ∂S
∂x(cid:19) p− + (S − Sc)p+ ,
∂x(cid:19) p− + (S − Sc)p− ,
= −λ+(cid:18) ∂S
= +λ+(cid:18) ∂S
= −S(p+ + p−) .
(3.8)
We are interested in travelling wave solutions that develop from a pointwise
inoculation of cells into a domain containing uniformly distributed nutrient S.
In this scenario, p± (defined as in each system) should form travelling pulses
while S forms a travelling front and relevant boundary conditions will be
,
∂S
∂x → 0
p±,
∂p±
∂x
S → 1
S → S−
as x → ±∞ ,
as x → +∞ ,
as x → −∞ .
(3.9)
Note that S− is currently unknown; we determine its value in the travelling
wave analysis of Section 4. Since p± and S are physical quantities, we search
for nonnegative travelling wave solutions, i.e.
p± ≥ 0, S ≥ 0.
It is clear that a travelling wave of this form cannot exist for Sc ≥ 1 (extinction
of bacteria) or for Sc ≤ 0 (infinite growth) and we will therefore only consider
systems that satisfy Sc ∈ (0, 1). In the next section we analyse System (B)
with respect to travelling wave solutions in order to obtain further insight. To
do that, we use the rescaled system (3.8).
4 Travelling wave analysis
In this section we first apply the standard travelling wave ansatz to system
(3.8) and derive a necessary condition for the existence of non-negative trav-
elling wave solutions. We then reduce the resulting ODE system to two com-
ponents through a change of variables and utilizing an invariant manifold
identified for the problem. Finally we use phase plane methods to analyse the
existence and properties of travelling wave solutions.
4.1 A necessary condition for the existence of travelling wave solutions
Let us apply the travelling wave ansatz p±(x, t) = p±(ξ) = p±(x − ct) and
S(x, t) = S(ξ) = S(x − ct), where c is the unknown wave speed [29]. System
Travelling waves in hybrid chemotaxis models
11
(3.8) becomes
(1 − c)(p+)′ = −λ0 (1 − χ S ′) p+ + λ0 (1 + χ S ′) p− + (S − Sc)p+ ,
−(1 + c)(p−)′ = +λ0 (1 − χ S ′) p+ − λ0 (1 + χ S ′) p− + (S − Sc)p− ,
−cS ′ = −S(p+ + p−) ,
(4.1)
where the primes denote derivatives with respect to the travelling wave variable
ξ. Note that any point on the S-axis is a steady state of the system (4.1) and
that linear stability of such a steady state, (p+, p−, S) = (0, 0, S∗), is governed
by the eigenvalues of the matrix A−1B, where
0
A =
1 − c
0
0 −1 − c 0
,
0 −c
0
B =
−λ0 + S∗ − Sc
λ0
−S∗
The eigenvalues of A−1B are
λ0
0
−λ0 + S∗ − Sc 0
.
0
−S∗
µ1 = 0,
µ2,3 =
c(−λ0 + S∗ − Sc) ±p∆1(S∗)
1 − c2
,
(4.2)
where
∆1(S∗) = c2λ2
0 + (S∗ − Sc − 2λ0)(S∗ − Sc).
(4.3)
Under the boundary conditions (3.9) we look for nonnegative solutions to (4.1)
connecting steady states (p+, p−, S) = (0, 0, S−) and (p+, p−, S) = (0, 0, 1). To
admit such a solution the latter must be a stable node, since a stable spiral
would imply negative values for p±. Hence, a necessary condition is ∆1(1) ≥ 0,
which is equivalent to
1
c ≥ c∗ =
λ0p(2λ0 − 1 + Sc)(1 − Sc) .
Given 2λ0 > (1 − Sc) it is easy to show that c∗ ∈ [0, 1].
(4.4)
Theorem 1 A necessary condition for the existence of nonnegative travelling
wave solutions of the system (3.8) is
2λ0 > (1 − Sc).
(4.5)
The above condition is reasonable, as we expect the run duration to occur on
a much faster time scale than proliferation processes.
12
Franz, Xue, Painter, Erban
4.2 Dimension reduction
Let us now perform a change of variables by introducing the cell density n =
p+ + p− and the cell flux j = p+ − p−. The travelling wave system (4.1) can
then be written as
− cn′ + j ′ = (S − Sc)n ,
−cj ′ + n′ = 2λ0χ S ′n + (S − Sc − 2λ0)j ,
−cS ′ = −Sn ,
(4.6)
(4.7)
(4.8)
where the boundary conditions for this system are
,
n, j,
∂n
∂x
S → 1
S → S−
∂j
∂x
,
∂S
∂x → 0
as
as
as
ξ → ±∞ ,
ξ → +∞ ,
ξ → −∞ .
From (4.8), we have Sn = cS ′ and, hence, n = c(ln S)′. Substituting into (4.6)
we obtain
−cn′ + j ′ = cS ′ − cSc(ln S)′ .
Integrating and applying the boundary conditions at ξ → +∞, an invariant
manifold of the problem is given by
−cn + j = c(S − 1) − cSc ln S .
With the definition f (S) ≡ S − 1 − Sc ln S, we obtain j = cn + cf (S), which
can be used to eliminate j from the system (4.6)–(4.8). For c 6= 1 we can solve
for n′ and obtain the reduced system
n′ =
S ′ =
c
1 − c2 (cid:20) 2λ0χ Sn2
c2
1
c
Sn .
+ 2 n(S − Sc − λ0) + (S − Sc − 2λ0)f (S)(cid:21) ,
(4.9)
(4.10)
For c = 1, we obtain
n =
λ0 − S + Sc −p(λ0 − S + Sc)2 − 2λ0χ S(S − Sc − 2λ0)f (S)
2λ0χ S
S ′ = Sn ,
, (4.11)
(4.12)
where we chose the solution to the quadratic equation for n that satisfies the
boundary conditions n → 0 as ξ → ±∞.
It can be easily shown that f (S) = 0 has two solutions in the region (0, 1]
for all Sc ∈ (0, 1) as follows. Since f ′(S) = 1 − Sc/S, f (S) is monotonically
decreasing for S ∈ (0, Sc) and monotonically increasing for S ∈ (Sc, 1]. With
f (1) = 0, this implies f (Sc) < 0 and, using f (S) → ∞ for S → 0, we obtain the
existence and uniqueness of the second root of f (S) = 0: we call it S1 ∈ (0, Sc).
The existence of S1 and the negativity of f (S) for S ∈ (S1, 1), together with the
condition 2λ0 > 1−Sc, implies that n as given in (4.11) is positive everywhere,
and that the given solution therefore satisfies the nonnegativity condition.
Travelling waves in hybrid chemotaxis models
13
4.3 Steady states and their linear stability
Using the two roots of f (S) = 0 and under the condition (4.5), it is clear
that there are two steady states of the system (4.9)-(4.10): (n, S) = (0, 1) and
(n, S) = (0, S1). Linearising the system (4.9)-(4.10) about its steady states
generates a system of the form
S(cid:19)′
(cid:18)n
= A(cid:18)n
S(cid:19) ,
where, for the general steady state S∗ ∈ {S1, 1}, we have
A =
2c
1 − c2 (S∗ − Sc − λ0)
S∗
c
c
1 − c2 (S∗ − Sc − 2λ0)
S∗ − Sc
S∗
0
with
trace A =
2c
1 − c2 (S∗ − Sc − λ0) ,
det A = −
1
1 − c2 (S∗ − Sc − 2λ0)(S∗ − Sc) .
The eigenvalues of A are identical to µ2,3 as given in (4.2). The steady state
(0, 1) is therefore a stable node for all c ∈ (c∗, 1) with c∗ as defined in (4.4).
Similarly, it can be seen that the steady steady (0, S1) is a saddle point. The
eigenvectors corresponding to the eigenvalues µ2,3 take the form
v1,2 = (cid:18)µ2,3 ,
S∗
c (cid:19)T
.
In the n − S plane, the slopes of the eigenvectors are given by
k1,2(c) =
µ2,3c
S∗
.
(4.13)
For the steady state (n, S) = (0, 1) this slope can be written in the form
k1,2(c) =
c2λ2
0
1 − Sc − λ0 ∓
,
√∆
(4.14)
where we define ∆ = c2λ2
0 + (1 − Sc − 2λ0)(1 − Sc) similarly to (4.3).
4.4 Case I: No chemotaxis (κ = ∞)
We first consider the case where the chemotactic sensitivity χ (given by (3.6))
vanishes, i.e cells do not respond chemotactically to changes in S. Here, trav-
elling waves are generated solely through proliferation of bacteria at the wave
14
(a)
0.8
0.6
n
0.4
0.2
0
0.2
(b)
S
,
n
1
0.8
0.6
0.4
0.2
0
0
Ω
0.4
0.6
S
0.8
1
Franz, Xue, Painter, Erban
5
10
ξ
15
Fig. 3 Illustration of the travelling wave solution calculated using the ODE system
(4.9)–(4.10) for χ = 0, λ0 = 10, c = c∗ = 0.3122 and Sc = 0.5.
(a) Trajectory of travelling wave solution. Solid line: trajectory, dashed line: n-nullcline,
dotted line: circumference of invariant region Ω introduced in the proof of Theorem 2.
(b) Travelling wave solution in ξ. Solid line: n, dashed line: S.
front. To understand the wave behaviour we perform a phase plane analysis
for the ODE system (4.9)–(4.10). Using κ = ∞ (i.e. χ = 0), it reduces to
n′ =
S ′ =
c
1 − c2h2n(S − Sc − λ0) + (S − Sc − 2λ0)f (S)i ,
1
c
Sn .
(4.15)
Thus, the slope of a trajectory in the n − S plane can be written as
dn
dS
=
c2
1 − c2
2n(S − Sc − λ0) + (S − Sc − 2λ0)f (S)
Sn
.
Additionally, an expression for the n−nullcline Γn is given by
n = −
S − Sc − 2λ0
2(S − Sc − λ0)
f (S) ,
and the S-nullcline is simply
n = 0,
or S = 0 .
Let us now show that travelling waves exist for the reduced system (4.15).
Theorem 2 For the case χ = 0 (which is equivalent to κ = ∞), a unique
travelling wave solution for the system (3.8) exists for all c ∈ (c∗, 1).
Proof For any c ∈ (c∗, 1) we can define a region Ω (see Figure 3(a)), enclosed
by the line n = k2(S − 1) (with k2 defined in (4.14)), the S-nullcline n = 0
and the line S = S1. We will first show that Ω is an invariant region of the
system (3.8). Since S is non-decreasing everywhere in Ω and n′ is non-negative
for n = 0 and S ∈ [S1, 1], we need only to show that the direction field on
Travelling waves in hybrid chemotaxis models
15
the segment Γ1 = {(n, S) : n = k2(S − 1), S ∈ [S1, 1)} points from the top
half of the plane above this segment towards the bottom. Since S is strictly
increasing we require
Indeed,
1 − c2
c2
dn
dS(cid:12)(cid:12)(cid:12)(cid:12)Γ1 ≤ k2 (≤ 0) .
dn
dS(cid:12)(cid:12)(cid:12)(cid:12)Γ1
= 2
S − Sc − λ0
S
= 2
S − Sc − λ0
S
S − Sc − λ0
S
≤ 2
+
+
+
(S − Sc − 2λ0)f (S)
S(S − 1)k2
(S − Sc − 2λ0)f (S)
S(S − 1)c2λ2
0
(S − Sc − 2λ0)(1 − Sc)
Sc2λ2
0
,
(1 − Sc − λ0 + √∆) ,
(1 − Sc − λ0 + √∆) ,
where we used (4.14) in the first step and the relation f (S)/(S − 1) ≤ 1 − Sc
for all S ∈ [S1, 1]. Using the fact that k2 and (S − Sc − 2λ0) are negative, we
can use the definition of c∗ and the fact that S ≤ 1 to obtain
1 − c2
c2
dn
dS(cid:12)(cid:12)(cid:12)(cid:12)Γ1 ≤ 2
S − Sc − λ0
S
−
2λ0 + Sc − S
S(2λ0 + Sc − 1)
(1 − Sc − λ0 + √∆) ,
= −2λ2
0 + 3λ0(S − Sc) − (S − Sc)(1 − Sc)
S(2λ0 + Sc − 1)
2λ0 + Sc − S
S(2λ0 + Sc − 1)
√∆ ,
0 + 2λ0(S − Sc) + λ0(1 − Sc) − (S − Sc)(1 − Sc)
−
≤ −2λ2
−
= −
S(2λ0 + Sc − 1)
√∆ ,
2λ0 + Sc − S
2λ0 + Sc − 1
(2λ0 + Sc − 1)(λ0 + Sc − S)
S(2λ0 + Sc − 1)
√∆ =
1 − c2
c2 k2 ,
≤ −λ0 + 1 − Sc −
−
2λ0 + Sc − S
2λ0 + Sc − 1
√∆ ,
where we used S ≤ 1 throughout the derivation. We can therefore conclude
that Ω is an invariant region of the system (3.8). Noting that at the steady state
(n, S) = (0, S1) the unstable manifold has a positive slope (k1,2 = µ2,3c/S∗),
i.e. it points into the region Ω, and using the fact that S is strictly increasing
inside Ω for n > 0 we can conclude that, for each c ≥ c∗, there is a heteroclinic
orbit starting from (0, S1) and finishing at (0, 1), corresponding to a travelling
wave solution of the PDE system (3.8).
16
Franz, Xue, Painter, Erban
4.5 Case II: Increasing chemotaxis (0 < κ < ∞)
Decreasing κ corresponds to an increase in the chemotactic sensitivity χ in
the ODE system (4.9)–(4.10) and the slope of trajectories in the n − S plane
is determined by
dn
dS
=
c2
1 − c2
2n(S − Sc − λ0) + (S − Sc − 2λ0)f (S)
Sn
+
2λ0χ
1 − c2 n .
It is noted that the above slope is larger than that for the non-chemotaxis case
within the region of interest n > 0. Due to this increase the region Ω for the
proof of Theorem 1 is no longer invariant for this system and a travelling wave
solution to (3.8) does not necessarily exist for all c ∈ (c∗, 1). The n-nullcline
for the full ODE system (4.9)–(4.10) is given as the solution of the quadratic
equation
2λ0χS
c
n2 + 2c(S − Sc − λ0)n + c(S − Sc − 2λ0)f (S) = 0.
For a given wave speed c, the n-nullcline can therefore be calculated as
n =
c
2λ0χS hc(λ0 + Sc − S) ±p∆2(S)i ,
with
∆2(S) = c2(λ0 + Sc − S)2 − 2λ0χ S(S − Sc − 2λ0)f (S) .
We can see that ∆2(S) → −∞ as S → ∞ due to its leading order term
−2λ0χ S3. Therefore, as S becomes large, no n-nullcline exists and n′ is posi-
tive everywhere. Additionally, ∆2(S) might have further roots and, in partic-
ular, ∆2(S) might be negative in parts (or the whole) of region S ∈ [S1, 1].
This again means that n is strictly growing in these parts of the domain.
We detect three different types of behaviours of trajectories starting close
to (n, S) = (0, S1), plotted in Figure 4. In particular, we can see each of these
behavioural types for different values of χ and despite different configurations
of the nullclines. In the top two plots of Figure 4 we present the case of a
diverging solution. Examining ODE (4.9), we observe that for large n, n grows
quicker than O(n2) and the divergence can be identified as a finite-time blow-
up. In the second case, depicted in the two plots in the middle of Figure 4,
the trajectory converges to the steady state (0, 1), but does so after entering
the region S > 1 and thereafter the region n < 0. Note that the steady state
(0, 1) is still a stable node in this case and that this overshoot is therefore not
a spiralling effect. Since these trajectories do not correspond to a non-negative
solution of the ODE system (4.9)–(4.10), they do not represent travelling wave
solutions to the original problem. The last case, presented in the plots on the
bottom of Figure 4, corresponds to an acceptable solution and is characterised
by the convergence to (0, 1) without crossing the line S = 1.
Travelling waves in hybrid chemotaxis models
17
χ = 1, c = 0.5884
χ = 0.3, c = 0.3
n
0.4
0.3
0.2
0.1
0
n
0.4
0.3
0.2
0.1
0
−0.1
0.2
0.4
0.6
S
0.8
1
−0.1
0.2
0.4
0.6
S
0.8
1
χ = 1, c = 0.5885
χ = 0.3, c = 0.328
n
0.4
0.3
0.2
0.1
0
n
0.4
0.3
0.2
0.1
0
−0.1
0.2
0.4
0.6
S
0.8
1
−0.1
0.2
0.4
0.6
S
0.8
1
χ = 1, c = 0.59
χ = 0.3, c = 0.35
n
0.4
0.3
0.2
0.1
0
n
0.4
0.3
0.2
0.1
0
−0.1
0.2
0.4
0.6
S
0.8
1
−0.1
0.2
0.4
0.6
S
0.8
1
Fig. 4 Trajectories of the ODE system (4.9)–(4.10) that highlight the three different cases.
Parameters in all plots are λ0 = 10, Sc = 0.5. Solid line: trajectory, dashed line: n-nullcline,
dotted lines: n = 0 and S = 1.
4.6 Case III: Infinite chemotactic sensitivity (κ = 0)
As κ decreases further we observe that the minimal wave speed necessary to
allow a non-negative travelling wave solution of (3.8) increases. In the limit
κ → 0, the ODE system (4.9)–(4.10) no longer has convergent solutions. How-
ever, in this limit the linearisation assumption leading to these ODEs and the
18
Franz, Xue, Painter, Erban
system (3.8) is no longer valid and we must consider the original turning kernel
as defined in (3.2). In the limit κ → 0 the turning rate in the hybrid model
therefore becomes
λ = (cid:26) 0,
2λ0,
for y(1) > 0,
for y(1) < 0 .
(4.16)
Hence, bacteria moving in a favourable direction do not turn, indicating that
the wave speed achieved in this limit should evolve to c = s = 1. In [40] it was
shown, for a slightly different turning kernel, that travelling waves can exist
even without growth terms and that their wave speed satisfies c = s.
5 Computational analysis of the wave speed
In this section we computationally compare wave speeds from the hybrid
model with those of the fully continuous models. Specifically, we investigate
the regimes in which the latter provide an acceptable insight into the travel-
ling wave behaviour of the hybrid model, and where they differ. We begin by
investigating the non-chemotaxis case, where the minimum wave speed c∗ for
the continuum systems was determined in (4.4). In Section 5.2 we show how
the wave speed depends on the value of κ, and correspondingly the chemotac-
tic sensitivity χ in the macroscopic model. A comparison with hybrid models
without cell proliferation is given in Section 2.3. We conclude this section with
a discussion into the effect and origin of oscillations observed under increasing
the adaptation time ta.
5.1 Case I: No chemotaxis (κ = ∞)
In Section 4.4 we analysed the macroscopic PDEs in the absence of chemotaxis.
Travelling wave solutions were shown to exist for all wave speeds c ∈ (c∗, 1),
with c∗ determined by (4.4). In Figure 5(a), variation of (4.4) as a function of
λ0 is illustrated; we note that wave speeds determined through simulation of
the PDE systems correspond exactly (to accuracy of the numerical approxi-
mation) with the analytical wave speeds. We now numerically investigate the
wave speed for the case χ = 0 in the hybrid model.
For our simulations we consider the same parameters and methods as de-
scribed in Section 2.3: specifically, we set the system parameters Sc = 0.5,
s = 1 and DS = 0. For the computations we consider a time step ∆t = 10−3,
a spatial resolution of ∆x = 0.25 on a domain with length L = 100, and sim-
ulate the system until the value of S at x = 60 falls below 0.5. The profiles at
this time, together with the time when S at x = 20 falls below 0.5, are used
to estimate the wave speed.
The measured wave speed for varying λ0 is illustrated in Figure 5(a), along
with c∗ as predicted from the travelling wave analysis. While the relationship
is similar in shape, we note that at all values of λ0 tested the measured wave
speed lies below the analytical value c∗. In the literature it has been observed
Travelling waves in hybrid chemotaxis models
19
(a)
1
0.8
0.6
0.4
d
e
e
p
s
e
v
a
W
0.2
0
2
4
λ
0
6
8
10
(b)
d
e
e
p
s
e
v
a
W
0.32
0.31
0.3
0.29
0.28
0.27
0.26
103
104
N
0
105
Fig. 5 Measured wave speed in the hybrid model. Crosses: individual simulations, dots:
ensemble averages. Parameters are as described in the text.
(a) Wave speed in dependence of λ0 for N0 = 10, 000. Dashed line: c∗ given by (4.4).
(b) Wave speed as a function of N0 with λ0 = 10. Dashed line: c∗ computed by (4.4).
that inaccuracies in numerical schemes can lead to an increase in wave speeds
[33], therefore rendering the lower wave speed seen in Figure 5(a) as counter
intuitive.
Nevertheless, we can provide the following explanation for the distinct val-
ues in the continuum and hybrid models. For the zero-chemotaxis case, wave
generation and movement is solely determined by growth ahead and death
behind the wave. In the continuum model an outermost "fractional bacteria
population" can extend significantly beyond the wave front, since some pro-
portion of the initial population never turns left, and hence far into the region
where S is very close to its initial value of 1. Yet this fractional population still
grows exponentially (∂p±/∂t ≈ (1− Sc)p±), seeding the growth and expansion
of the population. The finite/discrete nature of the hybrid model precludes any
fractional bacterium: the forward "tail" is necessarily finite and growth will
not occur beyond the outermost individual.
For the above explanation to hold we would expect a dependence of the
measured wave speed on the initial number of bacteria N0: continuous densi-
ties provide a closer approximation under larger numbers of bacteria and we
would expect convergence in the wave speed to c∗. Simulations in Figure 5(b)
demonstrate this property, corroborating our interpretation.
5.2 Case II: Increasing chemotaxis (0 < κ < ∞)
In the second set of numerical experiments we measure the dependency of
the wave speed on the critical parameter κ, i.e. we determine the effect of
increasing chemotaxis as κ decreases. We compare the results measured for
the hybrid system with the continuous Systems (A) and (B).
We use the same parameters as in Section 5.1 and results are shown in
Figure 6. The results demonstrate the regimes where correspondence across
the varying modelling levels occurs: while the hybrid model (dotted line) cor-
20
Franz, Xue, Painter, Erban
d
e
e
p
s
e
v
a
W
1
0.8
0.6
0.4
0.2
0
10−2
10−1
κ
100
Fig. 6 Comparison between wave speeds of the various models in dependence of κ. Dot-
ted line: hybrid model, red solid line: mesoscopic System (A), dashed line: linearised Sys-
tem (B). Parameters are as described in the text.
responds well with its closest continuous version (mesoscopic System (A), red
solid line) over a wide range of κ, it only corresponds with System (B) (black
dashed line) for larger κ, diverging as κ decreases. Note that the turning rate
(3.6) used for System (B) becomes negative at small values of κ and we limit
the range of κ studied accordingly.
At larger κ all three models converge to a value close to c∗ as κ grows: in
this regime the main assumption proposed for the linearisation (S(x)−y(2) ≪
κ) holds and we obtain good quantitative agreement. While this assumption
becomes less acceptable as we decrease κ, leading to the divergent behaviour
described above, we note that all models show the same qualitative agreement:
increasing chemotactic responses leads to an increase in the wave speed. Note
that the results for System (B) can be identically replicated using the ODE
system (4.9)–(4.10) and a search algorithm for the smallest value of c that
admits a nonnegative solution to the system.
These numerical experiments demonstrate that chemotaxis has a signifi-
cant effect on the speed of movement and that the waves cannot solely be
explained by growth and death terms. Rather, we interpret birth and death
processes as stabilisers to what would otherwise be transient waves [15,40].
This interpretation is in agreement with the results presented in Figure 2, as
the initial wave speed for the system without growth seems to be similar to
the wave speed of the system including growth and death terms.
5.3 Oscillations in the wave speed
An additional observation we made during the numerical experiments of the
hybrid model is that for increasing values of the adaptation time ta, the wave
Travelling waves in hybrid chemotaxis models
21
speed starts to differ strongly from the mesoscopic System (A), an effect that
we identified to be due to oscillations in the behaviour of the wave. In Fig-
ure 7(a) we present an example of strongly oscillating wave speeds (where the
wave speed is measured as rate of change of the average position of bacteria).
This example occurred for the parameters Sc = 0.5, λ0 = 10, κ = 0.001 and
ta = 4. We can also clearly see that the wave speed is correlated to the current
number of agents in the system. In the literature similar effects of oscillating
waves in stochastic models have been observed [28,32].
In Figure 7(b), we present the form of the wave at different times through-
out the simulation. It is clearly visible that the shape differs significantly at
different times. One reason these oscillations occur when ta is very high is
that a bacterium that happens to be in front of the wave experiences a very
high value of S, whilst its internal dynamics only adapt very slowly. This,
in combination with the low value of κ, leads to a bacterium that does not
switch direction for a long time and will proliferate at a high rate. This implies
that a spike of bacteria forms in front of the wave that moves faster than the
rest of the wave. We can clearly see such a spike in the left-most waveform
in Figure 7(b). Once the frontrunning bacterium and its copies have turned,
the wave goes into a reordering phase (second and third waveform), until,
eventually, a new spike emerges (4th waveform).
In Figure 7(c) we plot the wave speed over time for a smaller value of
ta. We can see that the oscillations are less severe and more frequent than in
Figure 7(a), which is in agreement with the explanation above. As we decrease
ta the frontrunning bacteria will adapt quicker to their surroundings and are
thereby more likely to turn. We show the influence of changing N0 on the
oscillating behaviour in Figure 7(d). The oscillations seem to occur with a
similar frequency but more regular to those before, which can be explained
by the increased likelihood of frontrunning bacteria with a higher number of
agents and reduced noise in the system.
6 Discussion
In this paper we presented a hybrid model of chemotaxis, incorporating a bio-
logically realistic turning kernel introduced in [40]. We analysed the travelling
wave behaviour of this hybrid system using mesoscopic and macroscopic equa-
tions, deriving an analytical value for the expected wave speed in the case
of no chemotaxis. As chemotaxis increases we demonstrated (analytically and
numerically) that the expected wave speed increases, indicating that the wave
that forms is not solely driven by growth and death processes. In contrast to
the transient waves observed for the hybrid model in the absence of growth
and death terms [15], the (numerical) waves observed here in their presence
are stable, indicating the stabilising effect of birth and death. The numerical
analysis reveals that the macroscopic equations derived through linearisation
of the turning kernel can qualitatively describe the change in wave speed as
chemotaxis increases, but that there are significant quantitative differences
22
(a)
1.5
1
0.5
l
s
s
a
m
d
e
a
c
s
e
r
,
)
t
(
v
0
0
100
200
(c)
1.5
l
s
s
a
m
d
e
a
c
s
e
r
,
)
t
(
v
1
0.5
0
0
100
200
t
t
Franz, Xue, Painter, Erban
(b)
0.2
0.15
)
x
(
n
0.1
0.05
300
400
500
0
150
160
170
x
180
190
200
(d)
1.5
l
s
s
a
m
d
e
a
c
s
e
r
,
)
t
(
v
1
0.5
300
400
500
0
0
100
200
t
300
400
500
Fig. 7 Oscillations in the wave speed of the hybrid model (2.2)–(2.6) and (2.10).
(a) Wave speed in comparison to current number of particles for N0 = 10, 000, ta = 4. Solid
line: wave speed, dashed line: number of particles, dotted lines: times of wave forms shown
in panel (b).
(b) Waveform at 4 distinct times marked in panel (a) from left to right.
(c) As in (a) with N0 = 10, 000, ta = 2.
(d) As in (a) with N0 = 50, 000, ta = 4.
Other parameters are given in Section 5.3.
between the two systems. Additionally, we observed oscillations in the wave
movement, an effect that had been seen in similar systems in the literature
[32] and that cannot be explained using mean-field approximations.
To date, travelling waves in chemotaxis models have mainly been analysed
from the perspective of macroscopic PDE models of chemotaxis [19,18]. The
existence of travelling waves for continuum models with growth terms is well
established [36,30,22]. While hybrid models have been used to study pattern
formation in bacterial chemotaxis [17,39], these studies have not analysed the
travelling wave patterns observed in bacterial cell populations.
Recently, experimental studies using microfluidic techniques tracked cell
trajectories within a traveling pulse, and revealed that persistence of direction
in cell movement accounts for 30% of the macroscopic speed of the travel-
ing pulse [35]. The hybrid model framework studied here provides a natural
method for direct comparison of model predictions with experimental mea-
surements of cell trajectory, and this is left as future work.
Travelling waves in hybrid chemotaxis models
23
Acknowledgements
The research leading to these results has received funding from the European
Research Council under the European Community's Seventh Framework Pro-
gramme (FP7/2007-2013) / ERC grant agreement No. 239870. This publica-
tion was based on work supported in part by Award No KUK-C1-013-04, made
by King Abdullah University of Science and Technology (KAUST). Radek Er-
ban would also like to thank the Royal Society for a University Research Fel-
lowship; Brasenose College, University of Oxford, for a Nicholas Kurti Junior
Fellowship; and the Leverhulme Trust for a Philip Leverhulme Prize. This prize
money was used to support research visits of Chuan Xue and Kevin Painter in
Oxford. Kevin Painter acknowledges a Leverhulme Trust Research Fellowship
award (RF-2011-045).
References
1. J. Adler. Chemotaxis in bacteria. Science, 153:708–716, 1966.
2. H. Berg. How bacteria swim. Scientific American, 233:36–44, 1975.
3. H. Berg and D. Brown. Chemotaxis in Esterichia coli analysed by three-dimensional
tracking. Nature, 239:500–504, 1972.
4. R. Bourret, K. Borkovich, and M. Simon. Signal transduction pathways involving pro-
tein phosphorylation in prokaryotes. Annual Review of Biochemistry, 60:401–441, 1991.
5. M. Brenner, L. Levitov, and E. Budrene. Physical mechanisms for chemotactic pattern
formation by bacteria. Biophysical Journal, 74(4):1677–1693, 1998.
6. E. Budrene and H. Berg. Complex patterns formed by motile cells of Esterichia coli.
Nature, 349:630–633, February 1991.
7. E. Budrene and H. Berg. Dynamics of formation of symmetrical patterns by chemotactic
bacteria. Nature, 376:49–53, July 1995.
8. C. Cercignani, R. Illner, and M. Pulvirenti. The Mathematical Theory of Dilute Gases.
Applied Mathematical Sciences, 106, Springer-Verlag, 1994.
9. P. Chavanis. A stochastic Keller-Segel model of chemotaxis. Communications in non-
linear science and numerical simulations, 15:60–70, 2010.
10. R. Erban. From individual to collective behaviour in biological systems. PhD thesis,
University of Minnesota, 2005.
11. R. Erban and J. Haskovec. From individual to collective behaviour of coupled velocity
jump processes: A locust example. Kinetic and Related Models, 5(4):817–842, 2012.
12. R. Erban and H. Othmer. From individual to collective behaviour in bacterial chemo-
taxis. SIAM Journal on Applied Mathematics, 65(2):361–391, 2004.
13. R. Erban and H. Othmer. From signal transduction to spatial pattern formation in E.
coli: A paradigm for multi-scale modeling in biology. Multiscale Modeling and Simula-
tion, 3(2):362–394, 2005.
14. R. Fisher. The wave of advance of advantageous genes. Annals of Eugenics, 7:355–369,
1937.
15. B. Franz and R. Erban. Hybrid modelling of individual movement and collective be-
haviour. In M. Lewis, P. Maini, and S. Petrovskii, editors, Dispersal, individual move-
ment and spatial ecology: A mathematical perspective. Springer, 2013.
16. A. Gerisch and K. Painter. Mathematical modelling of cell adhesion and its applications
to developmental biology and cancer invasion. In A. Chauviere and L. Preziosi, editors,
Cell Mechanics: From Single Scale-Based Models to Multiscale Modeling, Chapter 12,
pages 319–350. CRC Press, 2010.
17. Z. Guo, P. Sloot, and J. Tay. A hybrid agent-based approach for modeling microbiolog-
ical systems. Journal of Theoretical Biology, 255:163–175, 2008.
18. T. Hillen and K. Painter. A user's guide to pde models for chemotaxis. Journal of
Mathematical Biology, 58:183–217, 2009.
24
Franz, Xue, Painter, Erban
19. D. Horstmann. From 1970 until present: the Keller-Segel model in chemotaxis and its
consequences II. Jahresbericht des Deutschen Mathematiker Vereins, 106:51–69, 2004.
20. E. Keller and L. Segel. Traveling bands of chemotactic bacteria: A theoretical analysis.
Journal of Theoretical Biology, 30:235–248, 1971.
21. C. Kennedy and R. Aris. Traveling waves in a simple population model involving growth
and death. Bulletin of Mathematical Biology, 42:397–429, 1980.
22. K. Landman, G. Petter, and D. Newgreen. Chemotactic cellular migration: smooth
and discontinuous travelling wave solutions. SIAM Journal on Applied Mathematics,
63(5):1666–1681, 2003.
23. K. Landman, M. Simpson, J. Slater, and D. Newgreen. Diffusive and chemotactic
cellular migration: Smooth and discontinuous travelling wave solutions. SIAM Journal
of Applied Mathematics, 65:1420–1442, 2005.
24. K.A. Landman, M.J. Simpson, and D.F. Newgreen. Mathematical and experimental
insights into the development of the enteric nervous system and hirschsprung's disease.
Development, growth and differentiation, 49:277–286, 2007.
25. T. Li and Z. Wang. Asymptotic nonlinear stability of traveling waves to conservation
laws arising from chemotaxis. Journal of Differential Equations, 250:1310–1333, 2011.
26. T. Li and Z. Wang. Steadily propagating waves of a chemotaxis model. Mathematical
Biosciences, 240:161–168, 2012.
27. R. Lui and Z. Wang. Traveling wave solutions from microscopic to macroscopic chemo-
taxis models. Journal of Mathematical Biology, 61:739–761, 2010.
28. M. Metcalf, J. Merkin, and S. Scott. Oscillating wave fronts in isothermal chemical
systems with arbitrary powers of autocatalysis. Proceedings of the Royal Society London
A, 447:155–174, 1994.
29. J. Murray. Mathematical Biology. Springer Verlag, 2002.
30. G. Nadin, B. Perthame, and L. Ryzhik. Traveling waves for the Keller-Segel system
with Fisher birth term. Interfaces and Free Boundaries, 10:517–538, 2008.
31. H. Othmer, S. Dunbar, and W Alt. Models of dispersal in biological systems. Journal
of Mathematical Biology, 26:263–298, 1988.
32. H. Othmer and P. Schaap. Oscillatory cAMP signaling in the development of Dic-
tyostelium discoideum. Comments on Theoretical Biology, 5:175–282, 1998.
33. R.D. Reitz. A study of numerical methods for reaction-diffusion equations. SIAM
Journal on Scientific and Statistical Computing, 2:95–106, 1981.
34. J. Saragosti, V. Calvez, N. Bournaveas, A. Buguin, P. Silberzan, and B. Perthame.
Mathematical description of bacterial traveling pulses. PLoS Computational Biology,
6:e1000890, 2010.
35. J. Saragosti, V. Calvez, N. Bournaveas, B. Perthame, A. Buguin, and P. Silberzan.
Directional persistence of chemotactic bacteria in a traveling concentration wave. Pro-
ceedings of the National Academy of Sciences, 108(39):16235–16240, 2011.
36. R. Satnoianu, P. Maini, F. Garduno, and J. Armitage. Travelling waves in a nonlinear
degenerate diffusion model for bacterial pattern formation. Discrete and Continuous
Dynamical Systems B, 1:339–362, 2001.
37. Z.A. Wang. Mathematics of traveling waves in chemotaxis – review paper. Discrete and
Continuous Dynamical Systems Series B, 13:601–641, 2013.
38. M. B. Witte and A. Barbul. General principles of wound healing. Surgical Clinics of
North America, 77:509–528, 1997.
39. C. Xue, E. Budrene, and H. Othmer. Radial and spiral streams in proteus mirabilis
colonies. PLoS Computational Biology, 7(12):e1002332, 2011.
40. C. Xue, H. Hwang, K. Painter, and R. Erban. Travelling waves in hyperbolic chemotaxis
equations. Bulletin of Mathematical Biology, 73(8):1695–1733, 2011.
41. C. Xue and H. Othmer. Multiscale models of taxis-driven patterning in bacterial pop-
ulations. SIAM Journal on Applied Mathematics, 70(1):133–167, 2009.
|
1911.13197 | 1 | 1911 | 2019-11-29T17:07:20 | Robust increase in supply by vessel dilation in globally coupled microvasculature | [
"physics.bio-ph",
"q-bio.TO"
] | Neuronal activity induces changes in blood flow by locally dilating vessels in the brain microvasculature. How can the local dilation of a single vessel increase flow-based metabolite supply, given that flows are globally coupled within microvasculature? Solving the supply dynamics for rat brain microvasculature, we find one parameter regime to dominate physiologically. This regime allows for robust increase in supply independent of the position in the network, which we explain analytically. We show that local coupling of vessels promotes spatially correlated increased supply by dilation. | physics.bio-ph | physics |
Robust increase in supply by vessel dilation in globally coupled microvasculature
Felix J. Meigel,1 Peter Cha,2 Michael P. Brenner,2 and Karen Alim1, 3
1Max Planck Institute for Dynamics and Self-Organization, 37077 Gottingen, Germany
2John A. Paulson School of Engineering and Applied Sciences and Kavli Institute for Bionano Science and Technology,
3Physik-Department, Technische Universitat Munchen, 85748 Garching, Germany∗
Harvard University, Cambridge, MA 02138, USA
Neuronal activity induces changes in blood flow by locally dilating vessels in the brain microvas-
culature. How can the local dilation of a single vessel increase flow-based metabolite supply, given
that flows are globally coupled within microvasculature? Solving the supply dynamics for rat brain
microvasculature, we find one parameter regime to dominate physiologically. This regime allows for
robust increase in supply independent of the position in the network, which we explain analytically.
We show that local coupling of vessels promotes spatially correlated increased supply by dilation.
Vascular networks pervade all organs of animals
and are the paradigm of adaptive transport networks.
Their self-organized architecture continuously inspires
the search for their underlying physical principles [1 -- 4]
and at the same time serves as a template for design-
ing efficient networks in engineering [5]. The blood flow-
ing through vessels transports nutrients, hormones, and
metabolites to adjacent tissues. Metabolite exchange pri-
marily occurs within the fine vessel meshwork formed by
microvasculature. In the brain, local metabolite demand
can abruptly rise due to an increase in neural activity [6],
altering blood flow [7, 8] in the same brain region, observ-
able in fMRI [9]. During the process of increased neu-
ronal activity, neurons signal their increased demand to
adjacent astrocyte cells, which in turn trigger small ring
muscles surrounding blood vessels to relax [10]. Thus,
neural activity drives local dilation of a vessel [11, 12],
and hence regulates metabolite supply [7, 13]. However,
from a fluid dynamics perspective there is a mystery:
blood vessels form a highly interconnected network in
the microvasculature [8], resulting in a global coupling of
blood flow. A single dilating vessel can potentially change
the metabolite supply in a broad region of the network -
and thus the local increase due to dilation is a function
of specific network topology. Quantitatively, how much
control over changes in blood-based supply resides in a
single dilating vessel?
Models considering metabolite spread in tissue date
back more than a hundred years to A. Krogh [14].
Krogh's model estimates the supply pattern in a tissue
enclosed by vessels assuming that supply is constant on
all vessel walls. Yet, on a larger tissue scale, supply spa-
tially varies along the vasculature since resources sup-
plied upstream are not available downstream. Alterna-
tive models consider vessel-based transport [15], yet only
diffusive transport is taken into account. The combined
importance of advection and diffusion for transporting
solutes in a single tube was discovered by G.I. Taylor
[16, 17], with subsequent work outlining modifications
due to solute absorption at the tube boundary [18 -- 20].
Yet, there has been much less work capturing the cou-
pling of advection and diffusion in tubular network struc-
tures [21, 22], including solute absorption [23]. The im-
pact of a dilating vessel is hard to estimate since not only
the absorption dynamics on the level of single vessels is
changed, but also solute flux throughout the network is
rerouted since fluid flow and thus solute flux are glob-
ally coupled. However, to connect fMRI, which relies
on a fluid dynamic signal [9, 24, 25], and the change in
blood flow with neuronal activity [7, 11, 26 -- 28], we need
to understand how vessel dilations affect the supply with
metabolites.
In this letter, we present a theoretical model to deter-
mine the change in supply resulting from the dilation of a
single vessel. On the level of an individual vessel, we an-
alytically identify three regimes, each yielding a different
functional dependence of the overall supply by absorp-
tion along the vessel wall on vessel geometry, blood flow,
and blood flow based solute flux. Numerically analyz-
ing supply dynamics in a microvasculature excerpt of a
rat brain supplied from the Kleinfeld laboratory [8], we
find that a single regime dominates. This regime has the
important property that dilating a single vessel robustly
increases the supply along the dilated vessel independent
of the exact location of the vessel in the network. We ex-
plain analytically how a single vessel can buffer the global
coupling of solute fluxes within the network and yield a
robust local increase independent of network topology.
We further discuss how a single dilating vessel impacts
the solute flux downstream and thereby induces spatial
correlations in supply increase.
To understand how a change in flow induces changes in
solute flux and supply dynamics, we first focus on a single
vessel. We assume that the flow is laminar with longi-
tudinal velocity profile U (r) = 2 ¯U (1 − (r/R)2) [29, 30],
where ¯U denotes the cross-sectional averaged longitudi-
nal flow velocity. The dispersion of soluble molecules of
concentration C by the fluid flow within a tubular vessel
of radius R and length L is then given by
∂C
∂t
+ U (r)
∂C
∂z
= κ∇2C,
(1)
where κ denotes the molecular diffusivity of the solute,
and r and z parameterize the radial and longitudinal
2
where Pe = ¯U L/κ is the P´eclet number, α = γL, and
S = κγL/R ¯U measures the ratio of absorption rate to ad-
vection rate. Note, that the concentration decays along
the vessel starting from an initial concentration C0 that
itself is determined by the solute flux entering a vessel
J0. Also for the solute influx into a vessel advective and
diffusive transport contribute,
J0 = πR2C0(cid:18) ¯U +
κβ
L (cid:19) = πR2C0 ¯U(cid:18)1 +
β
Pe(cid:19) .
(5)
We define as supply of a vessel φ the integrated diffusive
flux through the entire vessel surface S of the cylindrical
vessel,
κ
φ = −(cid:90)S
·(cid:32) α2
∂C
2πRdz.
∂r(cid:12)(cid:12)(cid:12)(cid:12)r=R
4SPe (cid:33) · (1 − exp (−β)) .
12SPe + 2 S
1 + α2
β
resulting in,
φ =J0
1
1 + β
Pe
(6)
(7)
For physical intuition on how flow and vessel proper-
ties affect supply, we partition the phase space of sup-
ply dynamics spanned by Pe and S into three regimes,
keeping α fixed, see Fig. 1. At large values of S (cid:29) 1
and S (cid:29) 1/Pe the solute decays very quickly along the
vessel. Here, all solute that flows into the vessel of cross-
sectional area πR2 is absorbed at the wall, here denoted
all-absorbing regime
φall ≈ J0 = πR2C0 ¯U(cid:18)1 +
β
Pe(cid:19) .
(8)
For a network this implies that after a vessel in this
regime, no solute for further absorption downstream of
this vessel is available, which indeed is physiologically
rare, 1.0% in the rat brain microvasculature considered
here. A second regime occurs at Pe (cid:28) 1/S, Pe (cid:28) S
where diffusive transport dominates, here denoted diffu-
sive regime. We distinguish a third regime, which we
denote advective regime where advective transport dom-
inates, defined by S (cid:28) 1 and S (cid:28) Pe.
In both cases
the solute decay is very shallow, β (cid:28) 1 in Eqs. (3), (6),
resulting in supply independent of flow velocity, except
for the dependence on the initial concentration C0
φadvective ≈ φdiffusive ≈ 2πRL · κγ · C0.
(9)
Yet, note that the reason for the solute decay, i.e. β being
small, arises from entirely different transport dynamics,
see Fig. 1. This is reflected in the very different relation
between initial solute concentration at the start of the
vessel C0 and the solute influx J0 for the two regimes
(see the Supplemental Material S1 for derivation)
J0,advective ≈ C0 · πR2 · ¯U ,
2 · κ(cid:112)2γ.
J0,diffusive ≈ C0 · πR
3
(10)
(11)
FIG. 1. Supply φ by a single vessel can be partitioned into
three distinctive regimes as a function of dimensionless pa-
rameters characterizing flow and absorption, Pe = ¯U L
κ and
S = κγL
R ¯U . Dotted lines indicate separation of regimes. Re-
maining non-dimensional parameter fixed at α = 0.001. Error
ellipsoids contain the annotated percentage of vessels of the
here considered rat brain microvasculature [8] with physiolog-
ical parameters for γ and κ, see main text.
component of the vessel. The soluble molecule is ab-
sorbed at the vessel boundary, following
κ
+ κγC(R) = 0,
(2)
∂C
∂r(cid:12)(cid:12)(cid:12)(cid:12)r=R
with absorption parameter γ. In analogy to the deriva-
tion of Taylor Dispersion [16, 17, 23], we simplify the
multidimensional diffusion-advection for C = ¯C + C to an
equation for the cross-sectionally averaged concentration
¯C if the cross-sectional variations of the concentration C
are much smaller than the averaged concentration itself.
This is true if the time scale to diffuse radially within the
vessel is much shorter than the time scale of advection
along the vessel, R2/κ (cid:28) L/ ¯U , if the vessel itself can be
characterized as a long, slender vessel, R (cid:28) L, and if the
absorption parameter is small enough to keep a shallow
gradient in concentration across the vessel's cross-section
γR (cid:28) 1, which states that the length scale of absorption
is much bigger than the vessel radius. All these approx-
imations are valid for the rat brain microvasculature ex-
ample considered here [8]. With these assumptions, the
concentration profile along the vessel approaches a steady
state over a timescale L/ ¯U given by (see the Supplemen-
tal Material S1 for derivation)
¯C(z) = C0 exp(cid:16)−β(Pe, S, α)
24 · Pe
48 + α2
S2 (cid:32)(cid:114)1 +
8S
Pe
z
L(cid:17) ,
6PeS − 1(cid:33) ,
α2
+
β(Pe, S, α) =
(3)
(4)
10-2100102S (absorption rate / advection rate)10-310-210-1100101102103Pe (advection rate / diffusion rate)advectivealldiffusiveVessels R >150 %90 %99 %-3-2.5-2-1.5-1-0.50log10(/J0) 3
Hence, under constant solute influx J0 the diffusive and
the advective regime show a fundamentally different, yet
both non-linear dependence on the vessel radius,
φadvective ≈ J0
φdiffusive ≈ J0
2γκL
R ¯U
√2γL
√R
,
.
(12)
(13)
Based on these results for a single vessel we expect largely
varying increase in supply in response to vessel dilation.
The coupling of flows and solute flux in a network is likely
to make supply changes even more complex.
in,jUin,j = (cid:80)k πR2
Within a network not only fluid flows are cou-
pled with every network node obeying Kirchhoff's law
out,kUout,k but also solute flux
(cid:80)j πR2
J is conserved at every node(cid:80)j Jin,j =(cid:80)k J0,k. Here,
the solute influx Jin,j is determined by the inlet's vessel
inflow J0,j upstream reduced by the amount of supply,
φj, via that vessel, see Eq. (7). The influxes J0,k down-
stream a node, defined by Eq. (5), follow from the solute
concentration at the network node C0, given by
(cid:80)j Jin,j
C0 =
(cid:80)k πR2
out,k( ¯Uout,k + κβout,k/Lout,k)
.
(14)
Thus, solute fluxes are subsequently propagated from
network inlets throughout the network.
To now investigate the impact of single vessel dilation
on supply within a network, we turn to an experimentally
mapped rat brain microvasculature [8]. The data speci-
fies R, U , and L for all vessels as well as the pressures at
network inlets and outlets. Focussing on glucose as pri-
mary demand, we account for glucose's diffusion constant
κ = 6 × 10−10 m2 s−1 [31] and estimate glucose's perme-
ability rate and include γ = 200 m−1, see Supplemental
Material S2. Interestingly, we find 98% of all vessels to
be in the advective regime. Is there a functional property
that makes the advective regime stand out?
We next quantify the change in supply due to vessel
radius dilation in a capillary bed excerpt of the mapped
rat brain microvasculature excluding pial and penetrat-
ing vessels. To this end, we use the pressures given in the
data set [8] and impose the pressure values at inlet and
outlet vessels of a network excerpt. To be consistent with
the flows determined within the data set we use a modi-
fied hydraulic vessel resistance to account for additional
blood hematocrit resistance [32, 33] in accordance with
Blinder et al. [8]. Note, that a vessel's hydraulic resis-
tance is only important to calculate fluid flow velocities
within vessels but does not modify the supply dynam-
ics derived above. Pressures and hydraulic resistances
then fully determine the flow velocities throughout the
network due to Kirchhoff's law.
To identify differences in the behaviour of the three
supply regimes that may justify the physiological abun-
dance of the advective regime, we sample the effect of
FIG. 2. The advective regime is robust in increasing supply
by dilation. Histogram of change in supply ∆φ due to a single
vessel dilating by 10%. Lines indicate a range covering 69%
with both a percentage of 15.5% showing a lower or higher
supply outside the indicated range. Big dots indicate the me-
dian, with values of 0.17, 0.10, and 0.11 for the all-absorbing,
advective, and diffusive regime, respectively. For each his-
togram 120 vessels of the respective regime were randomly
chosen and dilated.
vessel dilation for all three regimes, drawing randomly
120 vessels in each regime out of the total number of
21793 vessels. The sheer total number of vessels allows
us to sample underrepresented diffusive and all-absorbing
regime without introducing a statistical bias due to sam-
ple size. Each vessels radius is dilated by 10%, and
the flow and solute flux is recalculated throughout the
network keeping the networks inlet and outlet pressures
fixed. The relative change in supply in the dilated ves-
sel itself is evaluated in a histogram, see Fig. 2. Ves-
sels in the all-absorbing regime show a broad response
to vessel dilation. Vessels in the advective regime, in
contrast, peak sharply at a robust 10% increase in sup-
ply, ∆φ = 0.1. The diffusive regime is also somewhat
peaked around ∆φ = 0.1, but in addition shows a sig-
nificant amount of vessels with smaller supply increase
of ∆φ < 0.1. Particularly the advective regime shows a
robust increase in supply matching the increase in ves-
sel diameter independent of the vessels' exact position
within the network topology. This observation is robust
against changes in the choice of the diffusion constant
and permeability rate, see Supplemental Material S6.
Despite our expectations of a non-linear change in sup-
ply from single vessel dynamics, Eqs. (12), (13), we find
a robust increase of 10% for 10% vessel dilation, which
would be reconciled within Eq. (9), if the initial concen-
tration at the inlet of a dilating vessel C0, Eq. (14), stays
constant despite changes in flow and solute flux through-
out the network. Which network properties allow C0 to
stay constant? What makes the advective regime more
robust than the diffusive?
Let us consider a network node, where all vessels are in
the advective regime with one inlet vessel and two outlet
vessels, out of the latter one is being dilated. Following
Eq. (14) and the simplification of the solute fluxes from
Eq. (10) for the advective regime the initial concentration
at the node is
C0 ≈ Cin
= Cin,
(15)
πR2
inUin
out,kUout,k
(cid:80)k πR2
where Kirchhoff's law was used for further simplifica-
tion. Hence, even though vessel radius dilation induces
changes in the flow, C0 ≈ Cin remains unchanged, though
Cin might be affected by upstream changes in the sup-
ply. However, we find that upstream effects on Cin are
small if the upstream vessels are in the advective or diffu-
sive regime, see Supplemental Material S3 and S5, which
leaves Cin and thus C0 approximately constant during
vessel dilation. This result generalizes to good approxi-
mation to the case where the non-dilating outlet vessel
is in the diffusive rather than in the advective regime,
see Supplemental Material S3. Note, that the case where
two inlet vessels merge into one outlet vessels is funda-
mentally different, as then the initial concentration at
the node is a mixture from the two inlet vessels. Dila-
tion of the outlet vessel changes flow in inlets differently
and thereby changes the mixing ratio non-linearly. Phys-
iologically, we find this pattern especially closer toward
venules. Taken together, these analytical results are in
agreement with the statistics of Fig. 2 and explain in
particular the robust increase in supply by dilation if the
vessel is in the advective regime.
We next probe why the diffusive regime is less robust
and revisit the setting of one inlet and one outlet in the
advective regime, and the second outlet in the diffusive
regime. But now we compute the initial concentration at
the node given that we dilate the vessel in the diffusive
regime,
C0 ≈ Cin,adv
πR2
in,advUin,adv
πR2
out,advUout,adv + πR
3
2
out,dif κ√γ
.
(16)
Now the dilation of the vessel in the diffusive regime in-
creases the denominator and thus leads to a decrease in
resulting C0, rendering the diffusive vessel's response less
robust compared to the advective. The same effect hap-
pens if all vessels at a node are in the diffusive regime,
even more so as no vessel in the advective regime can
buffer the dilation and diffusion dominated solute flux in-
dependent of flow velocity, see Eq. (11). Together, these
analytical arguments explain why the diffusive regime
yields a less robust increase in supply upon vessel di-
lation.
We found in Fig. 2, that the supply in upstream ves-
sel remains approximately constant during a single vessel
4
FIG. 3. Advective and diffusive regime robustly increase sup-
ply downstream of a dilating vessel at the cost of decreasing
supply in parallel vessels. (a) Enlargement of microvascula-
ture excerpt exemplifying the neighborhood change in supply
due to a single vessel dilation of 10% (advective regime, black
arrow). Inlet marked by yellow star. Blue denotes a decrease,
red an increase in supply in the individual vessels. The total
change in supply is ∆φtot = 6.4% in the downstream vessels
and φtot = 0.8% in the parallel vessels. Change in C0 for the
dilating vessel is below ∆C0 < 3 × 10−4. (b) Neighborhood
statistics of supply increase '+' or decrease '-' due to a dilat-
ing vessel in the respective regime. Evaluated is the overall
change in supply in up to four vessels downstream or parallel
to the dilated vessel chosen at the main inlet of a loop, respec-
tively. The dilated vessel itself is excluded from the statistics
here.
dilation. What is the effect on vessels downsteam the di-
lated vessel? For this, we focus on the dilating vessel's
immediate neighborhood and find that change in sup-
ply is spatially correlated, Fig. 3. We distinguish the
vessels in the direct neighbourhood of the dilated vessel
in two categories: downstream vessels are vessels that
are located directly downstream of the dilated vessel and
parallel vessels are vessels that are downstream the node
the dilated vessels branches off, but not downstream the
dilated vessel itself. The microvasculature data set is
known to show predominantly loop topologies, with a
median size of eight vessels within a loop [8]. We thus
considered only vessels with a topological distance of four
vessels to the dilated vessel for the analysis of the imme-
diate neighbourhood. We find that the typical response
of a dilating vessel in both advective and diffusive regime
is to increase supply downstream at the cost of reducing
supply in the parallel vessels, Fig. 3 (b). More solute is
drawn along the branch of the loop containing a dilating
vessel than the dilating vessel itself is taking up, which
increases the supply in downstream vessels. This is at
the expense of the vessels in the parallel branch, reduc-
ing the supply there. See also Supplemental Material S4.
While this applies qualitatively, the strength of this effect
depends on the exact network topology.
We here provided a theoretical framework to investi-
gate supply dynamics in a dynamically adapting tubular
network, where flows are globally coupled by topology.
We find that individual vessels can be classified in three
regimes by vessel geometry and flow rate. Among those
particularly the regime governed by advective transport
inletparalleldownstreamall-absorbing advectivediffusive Tot,down<0 Tot,down>0 Tot,par<0 Tot,par>06.3%23%14%20%47%52%47%20%19%26%10%16%(a)(b) 5
[18] E. M. Lungu and H. K. Moffatt, J. Eng. Math. 16, 121
(1982).
[19] M. Shapiro and H. Brenner, Chem. Eng. Sci. 41, 1417
(1986).
[20] G. N. Mercer and A. J. Roberts, Japan J. Indust. Appl.
Math. 11, 499 (1994).
[21] S. Marbach, K. Alim, N. Andrew, A. Pringle, and M. P.
Brenner, Phys. Rev. Lett. 117, 178103 (2016).
[22] Q. Fang, S. Sakadzi, L. Ruvinskaya, A. Devor, A. M.
Dale, and D. A. Boas, Opt. Express 16, 17530 (2008).
[23] F. J. Meigel and K. Alim, Roy. Soc. Interface 15,
20180075 (2018).
[24] Y. He, M. Wang, X. Chen, R. Pohmann, J. R. Polimeni,
K. Scheffler, B. R. Rosen, D. Kleinfeld, and X. Yu, Neu-
ron 97, 925 (2018).
[25] P. Tian, I. C. Teng, L. D. May, R. Kurz, K. Lu, M. Sca-
deng, E. M. C. Hillman, A. J. D. Crespigny, H. E.
DArceuil, J. B. Mandeville, J. J. A. Marota, B. R. Rosen,
T. T. Liu, D. A. Boas, R. B. Buxton, A. M. Dale,
and A. Devor, Proc. Natl. Acad. Sci. U.S.A. 107, 15246
(2010).
[26] O. B. Paulson, S. G. Hasselbalch, E. Rostrup, G. M.
Knudsen, and D. Pelligrino, Journal of Cerebral Blood
Flow & Metabolism 30, 2 (2010).
[27] P. O'Herron, P. Y. Chhatbar, M. Levy, Z. Shen, A. E.
Schramm, Z. Lu, and P. Kara, Nature 534, 378 (2016).
[28] C. M. Peppiatt, C. Howarth, P. Mobbs, and D. Attwell,
Nature 443, 700 (2006).
[29] S.-S. Chang, S. Tu, K. I. Baek, A. Pietersen, Y.-H.
Liu, V. M. Savage, S.-P. L. Hwang, T. K. Hsiai, and
M. Roper, PLOS Comp. Biol. 13, e1005892 (2017).
[30] D. Obrist, B. Weber, A. Buck, and P. Jenny, Philos.
Trans. Roy. Soc. A 368, 2897 (2010).
[31] W. D. Stein, Channels, Carriers, and Pumps: An Intro-
duction to Membrane Transport (Academic Press, 2012).
[32] A. Pries, Cardiovasc. Res. 32, 654 (1996).
[33] D. Rubenstein, W. Yin, and M. D. Frame, Biofluid Me-
chanics, An Introduction to Fluid Mechanics, Macrocir-
culation, and Microcirculation (Academic Press, 2015).
- and to lesser extend also the regime governed by dif-
fusive transport - yield a robust increase in supply upon
vessel dilation within the dilating vessel, notably leaving
the supply pattern upstream unchanged and increasing
supply immediately downstream. Interestingly, the most
robust advective regime is found to dominate in brain
microvasculature. Our findings therefore promote that
vessel dilation results in a robust increase in supply inde-
pendent of the exact position of the vessel in the network.
Our results are important for understanding the link be-
tween neural activity and patterns of change in supply
invoked by vessel dilations and changes in blood flow un-
derlying fMRI. Moreover, our framework is instrumental
to predict drug delivery, design blood vessel architecture
in synthetic organs but may also open entire new avenues
for the programming of soft robotics and smart materials.
We thank David Kleinfeld and collaborators for shar-
ing their data on the rat brain microvasculature with us.
This work was supported by the Max Planck Society and
the National Science Foundation Division of Mathemati-
cal Sciences DMS 1411694 and DMS 1715477. K.A. fur-
ther acknowledges the stimulating environment of Amer-
ican Institute of Mathematics' Square Meetings. M.P.B.
is an investigator of the Simons Foundation.
∗ [email protected]
[1] C. D. Murray, Proc. Natl. Acad. Sci. U.S.A. 12, 207
(1926).
[2] G. B. West, J. H. Brown, and B. J. Enquist, Science
276, 122 (1997).
[3] D. Hu and D. Cai, Phys. Rev. Lett. 111, 138701 (2013).
[4] H. Ronellenfitsch and E. Katifori, Phys. Rev. Lett. 117,
138301 (2016).
[5] X. Zheng, G. Shen, C. Wang, D. Dunphy, T. Hasan, C. J.
Brinker, Y. Li, and B.-L. Su, Nat. Commun. 8, 1 (2017).
[6] G. M. Boynton, S. A. Engel, G. H. Glover, and D. J.
Heeger, J. Neurosci. 16, 4207 (1996).
[7] M. E. Raichle and M. A. Mintun, Annu. Rev. Neurosci.
29, 449 (2006).
[8] P. Blinder, P. S. Tsai, J. P. Kaufhold, P. M. Knutsen,
H. Suhl, and D. Kleinfeld, Nat. Neurosci. 16, 889 (2013).
[9] N. K. Logothetis, Nature 453, 869 (2008).
[10] B. A. MacVicar and E. A. Newman, Cold Spring Harb.
Perspect. Biol. 7, a020388 (2015).
[11] C. Cai, J. C. Fordsmann, S. H. Jensen, B. Gesslein,
M. Lnstrup, B. O. Hald, S. A. Zambach, B. Brodin,
and M. J. Lauritzen, Proc. Natl. Acad. Sci. U.S.A. 115,
E5796 (2018).
[12] R. A. Hill, L. Tong, P. Yuan, S. Murikinati, S. Gupta,
and J. Grutzendler, Neuron 87, 95 (2015).
[13] I. Vanzetta, R. Hildesheim, and A. Grinvald, J. Neurosci.
25, 2233 (2005).
[14] A. Krogh, J. Physiol. 52, 409 (1919).
[15] M. Schneider, J. Reichold, B. Weber, G. Szkely, and
S. Hirsch, Med. Image Anal. 16, 1397 (2012).
[16] G. I. Taylor, Proc. R. Soc. Lond. A 219, 186 (1953).
[17] R. Aris, Proc. R. Soc. Lond. A 235, 67 (1956).
SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.S1.DERIVATIONOFSTEADY-STATESUPPLYACROSSACYLINDRICALVESSELWALLA.Formulationofthedynamicsintermsofthecross-sectionalaveragedconcentrationWebeginwiththeadvectiondiffusionequationinasinglestraightcylindricalvessel,whereweassumethatthefluidflowobeysaPoiseuilleprofileU(r),∂C∂t+U(r)∂C∂z=κ∇2C.(S1)Theabsorptionthroughavesselwallsetstheboundaryconditionto∂C∂r(cid:12)(cid:12)(cid:12)(cid:12)r=R+γC(R)=0,(S2)wherewedefinedtheabsorptionparameterγ.InanalogytoTaylor'sderivationofsheardispersion,wedefinethecross-sectionalaveragedvariables¯C=1AZCdA.Withthisdefinition,wecanwritetheconcentrationCassumofthecross-sectionalaveragedconcentration¯CandadeviationtermC,C=¯C+C.Motivatedbythisseparation,wedistinguishbetweenderivativesalongandperpendiculartotheflowdirection.Theadvectiondiffusionequationandtheboundaryconditionnowread∂¯C∂t+∂C∂t+(¯U+U)∂¯C∂z+(¯U+U)∂C∂z=κ∇2k¯C+κ∇2kC+κ∇2⊥Cand∂C∂r(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)r=R+γ(C(R)+¯C)=0.Wetakeintoaccountthattheaverageanddifferentiationcommute.TogetherwithU=C=0,wetakethecross-sectionalaverageoftheadvectiondiffusionequationandfind∂¯C∂t+¯U∂¯C∂z+U∂C∂z=κ∇2k¯C+κ∇2⊥C.(S3)Wecancalculatethelasttermofaboveequationusingtheboundaryconditionκ∇2⊥C=κ2R2ZR0∇2⊥Crdr=−2κγR(¯C+C(R)).InEq.(S3),weonlyneedtoestimatethetermU∂zC.Forthis,wesubtractthecross-sectionalaveragedadvectiondiffusionequationfromEq.(S1)andfind∂C∂t+U∂¯C∂z+(¯U+U)∂C∂z−U∂zC=κ∇2C+2κγR(¯C+C(R)).Wenextemploythreeapproximations,whichareinlinewithreducingthedynamicstocentermanifolddynamics.ThefirstassumptionisthatthetimescaleofdiffusionismuchsmallerthanthetimescaleofadvectionthroughthevesselR2/κ(cid:28)L/U.ThisreducestheequationtoU∂¯C∂z+(¯U+U)∂C∂z−U∂zC=κ∇2C+2κγR(¯C+C(R)).1 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.Next,weassumethatthecross-sectionalvariationaresmall¯C(cid:29)C.Asahighabsorptionparameterimpliesalargeconcentrationgradientwithinthecross-sectionwedemandthatthelength-scaleofadvectionismuchbiggerthanthecross-sectionγR(cid:28)1.ThisimpliesU∂¯C∂z=κ∇2C+2κγR(¯C+C(R)).Inalaststep,weassumethatthevariationofCismuchbiggerinradialdirectionthaninflowdirection∂2rC(cid:29)∂2zC.Hence,thelength-scaleofdiffusionalongthevesselismuchlargerthanthelength-scaleofdiffusionoverthevessel'scross-section,whatlongslendervesselsimply.U∂¯C∂z=κ∇2rC+2κγR(¯C+C(R)).(S4)Foraknownflowprofile-hereweassumedaPoiseuilleflowprofile-wecansolvetheaboveequationasdifferentialequationforCindependenceofrand¯C.¯U(cid:18)1−2r2R2(cid:19)∂¯C∂z=κ∇2rC+2κγR(¯C+C(R)).1κ(cid:18)¯U∂z¯C−2κγR(¯C+C(R))−¯U2R2∂z¯Cr2(cid:19)=∇2rCα1−α2r2=1r∂r(r∂rC).Wefindthroughsimpleintegration12α1r2−14α2r4+α3=r∂rC.Wecanimmediatelydeterminethatα3yieldsalogarithmicterminthenextintegration.Sincewedemandthattheconcentrationisfiniteoverthewholecross-section,wefindthattheintegrationconstantα3=0.Inthenextintegrationstep,wefind14α1r2−116α2r4+α4=C.Wefixthevalueofα4byaccountingforC=0.ZR0(cid:18)14α1r3−116α2r5+α4r(cid:19)dr=α1R416−α2R616·6+α4R22,whichfixesthevalueα4=α2R448−α1R28.WefindthatC(R)isgivenbyC(R)=¯UR224κ(cid:18)1+γR4(cid:19)−1∂z¯C−γR/41+γR/4¯C.Equippedwiththisresult,wecannoweasilydeterminetheexpressionU∂zC.U∂zC=2¯UR2ZR0(cid:18)14∂zα1r3−116∂zα2r5+∂zα4r(cid:19)−2r2R2(cid:18)14∂zα1r3−116∂zα2r5+∂zα4r(cid:19)dr,=2¯UR2(cid:20)R4∂zα116−R6∂zα296+R2∂zα42−2R6∂zα124R2+2R8∂zα216·8·R2−2R4∂zα44R2(cid:21)=−R2¯U∂zα124+R4¯U∂zα296Insertingalldefinitionsandsortingforordersin∂nz¯C,wefind∂¯C∂t=−2κR24γR4+γR¯C−12+5γR12+3γR¯U∂¯C∂z+(cid:18)κ+12+γR12+3γR¯U2R248κ(cid:19)∂2¯C∂z2.(S5)2 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.B.Solutionofthecross-sectionalaverageddynamicsWenowproceedbysolvingtheabovepartialdifferentialequationforsteadystate,andwefind0=a¯C+b∂z¯C+c∂2z¯C.Makinguseofanexponentialansatz¯C=C0exp(−β0z),wefindthatβ0isgivenbyβ0=b±√b2−4ac2c=b2c(cid:18)1−r1−4acb2(cid:19),whereweusedinthelaststepthatthesolutionhastoconverge.Wedefinenowtheparameterβ=β0·L,suchthatthesolutionisgivenby¯C=C0exp(−βz/L).WethenapproximatethetwocoefficientsinβmakinguseoftheaboveapproximationγR(cid:28)1andfindb2cL≈¯U2(cid:16)κ+¯U2R248κ(cid:17)=24¯ULκ48+γ2L2κ2γ2L2R2¯U2=24·Pe48+α2S2.Wewillfindthatit'susefultosortthetermsinthethreenon-dimensionalvariablesPe,Sandα.HerePeisthePecletnumberwithPe=¯UL/κ.SisinthestyleofaDamkohlernumberS=κγL/R¯U.αisgivingtheratioofthelength-scaleofabsorptionandthelength-scaleofthevesselα=γL.LikewisewefindforthesecondcoefficientinβagainusingγR(cid:28)11−4acb2=1+42κR2γR(cid:16)κ+¯U2R248κ(cid:17)¯U2=1+8γRκ2R2¯U2+γR6=1+8SPe+α26PeS.Inthecontextofbloodflow,itbecomesusefulnottoexpress¯CintermsofaninitialconcentrationatavesselentranceC0,butwithaninitialfluxintothevessel.Thefluxisdefinedwithcontributionfromtheadvectiveandthediffusiveflux.WethuswriteJ=¯U¯C−κ∂z¯C.Atthebeginningofthevessel,wherez=0,wefindtherelationC0=J0¯U+κβL.Finally,wedefinethesupplyφasconcentrationfluxthroughthesurfaceofthevesselwallandthusφ=−ZSκ∇Cda.MakinguseoftheboundaryconditioninEq.(S2),wefindφ=−ZSκ∇Cda=−κZS∂C∂r(cid:12)(cid:12)(cid:12)(cid:12)r=Rda=2πκγRZL0(¯C+C(R))dz.(S6)Insertingalldefinitions,wefindforthesupplyφ=J011+βPe· α212SPe+2Sβ1+α24SPe!·1−exp−24·Pe·q1+8SPe+α26PeS−148+α2S2.(S7)Thislengthyexpressionisunintuitiveatfirstsight.InthemanuscriptwemappedoutthedynamicsofEq.(S7).Thoughthisexpressionislengthy,onecaneasilyapplytheapproximationsweuseinthemanuscripttodefinethethreedifferentregimesintermsofPeandS.3 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.C.ApproximationofthesupplyexpressionWenowproceedtoestimatehowbigdifferenttermsareforthedifferentregimes.WestartbyfocusingontheTaylorapproximationsconditionsthatneedtobefulfilledforallregimesR(cid:28)L=⇒αSPe(cid:28)1,(S8)γR(cid:28)1=⇒α2SPe(cid:28)1,(S9)R2κ(cid:28)LU=⇒α2S2Pe(cid:28)1=⇒α2S2(cid:28)Pe.(S10)1.AdvectiveregimeWenextapproximatethesupplyinthedifferenttransportregimes.Westartwiththeadvectiveregime,whereS(cid:28)1andS(cid:28)Pe→S/Pe(cid:28)1.WefirstapproximatethebracketexpressioninβwherewefindusingS/Pe(cid:28)1 r1+8SPe+α26PeS−1!−→12α26PeS,(S11)asweseriesexpandedthesquarerootexpressionforsmallvaluesofα2SPe(cid:28)1.Fortheprefactorofthebracket,wefindsinceS(cid:28)124·Pe48+α2S2−→24Peα2S.(S12)Asaresult,βisgivenbyβ(Pe,S,α)−→2S.(S13)Notethat2S(cid:28)1,whichimpliesthatβissmallfortheadvectiveregime.2.DiffusiveregimeWenextfocusonthediffusiveregime,wherePeS(cid:28)1andPe(cid:28)S→S/Pe(cid:29)1.ThisimpliesPe(cid:28)1.Note,thatthisfurtherimplieswiththeTaylorapproximationsα2S2(cid:28)Pe(cid:28)1.(S14)Focusingnowontheprefactorofthebracketexpression,wefind24·Pe48+α2S2−→Pe2.(S15)Forthebracketexpression,wefindusingS/Pe(cid:29)1 r1+8SPe+α26PeS−1!−→r8SPe.(S16)Asaresult,βisgivenbyβ(Pe,S,α)−→√2PeS.(S17)NotethatPeS(cid:28)1,whichimpliesthatβissmallforthediffusiveregime.4 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.3.AllabsorbingregimeWenextfocusontheallabsorbingregime,whereS(cid:28)1andPeS(cid:29)1.Withthiswefindthat24·Pe48+α2S2 r1+8SPe+α26PeS−1!−→Pe2 r1+8SPe−1!.(S18)WenextdistinguishthecasesS/Pe(cid:29)1andS/Pe(cid:28)1.StartingwithS/Pe(cid:29)1,wefindβ(Pe,S,α)−→√2PeS,(S19)whereβisbigduetoPeS(cid:29)1.ForS/Pe(cid:28)1,wefindβ(Pe,S,α)−→2S(S20)whereβisbigduetoS(cid:29)1.β(Pe,S,α)=24·Pe48+α2S2 r1+8SPe+α26PeS−1!(S21)S2.DERIVATIONOFTHEABSORPTIONPARAMETERγToourknowledgenomeasurementsoftheabsorptionparameterγhaveyetbeenconducted.Toestimatevaluesfortheabsorptionparameterseitheramicroscopicapproachoramacroscopicapproachcanbeused.Forthemicroscopicapproachweestimateanabsorptionparameterusingpermeabilitiesthroughlipidbilayers,channelsconductance,channelabundanceandotherphysiologicalfeaturesallowingtheextractionofsolutefromthebloodvesselthroughthevesselwall.Thoughthereareexperimentsassessingsomeoftheseparameters,seee.g.[1 -- 3],toconcludeontheabsorptionparameterusingthemicroscopicapproachdemandsalargenumberofadditionalassumptionswhicharenotvalidatedbyinvivoexperiments.Insteadweusehereamacroscopicapproach.ForthemacroscopicapproachweestimatetheabsorptionparameterusingthemodeldevelopedinSupplementalMaterialS1andthemacroscopicobservableoftheoverallmeasuredabsorptioninthebrain.Thisapproachallowsustoderiveamorephysiologicalestimateoftheabsorptionparameters,aswerelyoninvivoexperimentsandmakelessassumptionscomparedtothemicroscopicapproach.Leybaertetal.[4]measuredthetotalabsorptionofglucoseintheratbraintobe10%.Weassumethatthemainabsorptionhappensincapillarybedsandthusbetweenpenetratingateriolandpenetratingvenuleandthatthenumbernofvesselsbetweenateriolandvenuleisoftheordern=O(10).Inthedatasetconsideredbyus[5]themedianvesselradiusisR=2µm,vessellengthL=50µm,andflowvelocityU=400µms−1.Fortheestimatehere,weconsiderthetotalsupplyandthusassumeaneffectivelengthofLtot=nL.WeestimatetheabsorptionparametermakinguseoftherelationCabsCtot=1−e−β.(S22)Wesolvethisequationforγ,wherewetakeintoaccountthatγR(cid:28)1.Wefindwiththeaboveequationandthemedianparametersofthevascularnetworkavaluefortheabsorptionparameteroftheorderγ=O(100m−1).Here,wechoseγ=200m−1.Wechoseavalueontheupperboundaryoftheestimatetoincreasethestatisticsinthediffusiveandallabsorbingregime.Notethatwiththischoiceofγ,westillfulfilltheassumptionγR(cid:28)1.Furthermorenote,thattherobustresultsareindependentoftheexactvalueofγ,comparewiththefiguresinSupplementalMaterialS6.S3.ESTIMATIONOFUPSTREAMCHANGESONC0Wefindthatfortheadvectiveandthediffusiveregime,wecanestimatethesupplybyφadvective≈φdiffusive≈2πRL·κγ·C0,5 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.whichrenderssupplyindependentoftheflowvelocity.Forupstreamvesselsintheadvectiveregime,thesolutefluxexitingavesselisgivenbyJup,out=J0,up−φup.(S23)ThisisthesolutefluxthatentersinEq.(14)asinfluxintoanetworknode.WerewritetheequationaboveintermsoftheinitialconcentrationC0,up,whichistheinitialconcentrationatthebeginningoftheupstreamvessel,tofindJup,out=C0,up·πR2U−C0,up·2πRL·κγ,(S24)whereflowvelocity,radius,andlengthoftheupstreamvesselwereused.WiththisandthesupplythroughtheupstreamvesselwecannowcomputethesoluteoutfluxoftheupstreamvesselasJup,out=C0,up·πR2U(cid:18)1−2κγLRU(cid:19)=C0,up·πR2U(1−2S).(S25)Thus,wefindthatC0,upandhenceCininEq.(14)changebythefactor(1−2Snew)/(1−2S).Thisfactorisfunctionof∆U.TakingtheTaylorseriestofirstorderin∆U,wefindthatCin,dilischangedbyafactorCin,dil=Cin(cid:18)1+2S1−2S∆U+O(cid:0)(∆U)2(cid:1)(cid:19).(S26)WepreviouslyidentifiedtheadvectiveregimetobecharacterizedbyS(cid:28)1.Andevenaswecannotestimatethepreciseexpressionof∆Uwithoutknowingtheexactnetworktopology,itisreasonabletoassumethatthechangeforavesseldilationof10%isbelow100%fortheflowvelocity,whatimplies∆U<1.Asaconsequence,wefindthatthecontributionofofupstreameffectsonCinandhenceC0inEq.(15)isnegligible,whichisalsoinlinewiththesimulationresults,seeforthisthefigureintheSupplementalMaterialS5.Foroneoftheoutflowtubesbeinginthediffusiveregime,whiletheinlettubeandatleastoneotherinflowtubeintheadvectiveregime,weneedtofocusonthedenominatorofEq.(14).Note,thatinthedatasetconsideredhere,vesselbranchingofhaveasimilarradius.WefindthatavesselinthediffusiveregimehavealowflowQ.Thesizeofthediffusivefluxisapproximatelythesameforvesselsinthediffusiveregimeandvesselsintheadvectiveregime.ForthedenominatorofEq.(14)thisimpliesthattheadvectivefluxtermsdominatethedenominatoraslongasatleastonoftheoutflowvesselsisintheadvectiveregime.Withthesamereasoningasbefore,wecanthusneglectalldiffusivetermsintheequation.Iftheupstreaminletvesselisinthediffusiveregime,allvesselsbranchingoffmustalsobeinthediffusiveregime.Forthiscase,theapproximationabovefailsandthedenominatorincreasesasoneofthevesselsisdilated.FocusingonthenominatorofEq.(14)andhowJup,outchangesforanupstreaminletvesselinthediffusiveregime,wefindtheanalogousexpressionJup,out=C0,up·πR32p2γκ(cid:16)1−√2SPe(cid:17).(S27)Thus,foravesselinthediffusiveregimeCinistrulyindependentofupstreameffectsasthisexpressionisfullyindependentofU.Thusavesseldilationwillcausenoupstreameffects.ThenominatorofEq.(14)willnotchangeduetoavesseldilation,leavingCin,dilwithasmalldecrease.S4.SPATIALCORRELATIONOFSUPPLYINDOWNSTREAMVESSELSWefindthatthetypicalresponseofadilatingvesselinboththeadvectiveandthediffusiveregimeistoincreasesupplydownstreamatthecostofreducingsupplyintheparallelbranch.Canweunderstandhowspatialcorrelationsindownstreamvesselsarisewithoutknowingtheexactnetworktopology?Forthiswefocusagainfirstatadilatingvesselintheadvectiveregime.FollowingthesamelineofreasoningasintheSupplementalMaterialS3,weunderstandthedilatedtubeasinlettubeforvesselsthataredirectlydownstreamthedilatedvessel.IncontrasttothecalculationsintheSupplementalMaterialS3,herethetubeitselfisdilatedbyafactor∆R.Tofirstorderin∆U,wefindthatCin,dilisalteredbythefactor.Cin,dil=Cin(cid:18)(1+∆R)3−2S(1+∆R)(1−2S)+2S(1−2S)(1+∆R)∆U+O(cid:0)(∆U)2(cid:1)(cid:19).(S28)6 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.With(1+∆R)2−2S(1+∆R)(1−2S)>1forS(cid:28)1,whichistrueforvesselsintheadvectiveregime.Thesecondtermcanbepositiveornegativedependingonthesignof∆U.Topredictthevalueof∆U,thefullnetworktopologymustbeknown.Furthermore,non-linearcorrectionsforthehematocritaffect∆U.Nevertheless,wecanestimate∆Uwitharuleofthumb.Forthis,weassumeaconstantpressuredropoverthedilatingvesselandneglectcorrectionforthehematocrit.Inthiscase,thechangeinthevelocityscaleslike∆U∝(∆R)2andthesecondtermisalsopositive.Moresoluteisdrawnintothedilatedbranchthanisconsumedmorebyincreasedabsorptionalongthedilatedvessel.FurtherdownstreameffectsfollowthesameformasdescribedintheSupplementalMaterialS5,wherewefindCin,dil=Cin(cid:18)1+2S1−2S∆U+O(cid:0)(∆U)2(cid:1)(cid:19).Asforthedilatedtube,whetherthecorrectiontermincreasesordecreasesthefurthersupplydownstreamdependsonthesignof∆U.InlinewiththeargumentinSupplementalMaterialS5,thefurthercorrectiontermsdownstreamthedilatedvesselaresmall,as∆Uissmall.Note,howeverthatthereisatopologicaldifferencebetweenvesselsupstreamandvesselsdownstreamadilatedvessel.Theconsequenceofthetopologicaldifferencecanbeunderstooddrawingtheanalogytocircuitnetworks.Recall,thatthemicrovasculaturehasaloopytopology.Forthefurtherargument,wefocusontheloopinwhichthedilatedvesselislocated.Recallnow,thattheflowthroughthebranchofaloopisdeterminedbythetotalresistancealongthebranch.Thechangeinresistanceisstrongestinthebranchinwhichthedilatedvesselislocated.However,fortheupstreamvessels,thetotalpaththroughthevasculaturemustbeconsidered.Here,thetotalchangeinresistanceisbufferedbytheparallelbranchesleavingasmallerchangeintheflow.Asaresult∆Uisbiggerinthedownstreamvesselsthanintheupstreamvessel.Areductionofsupplyinthevesselsoftheparallelbranchescanbeunderstoodalongthesamelines.Decreasingtheresistancealongthebranchwiththedilatedvessel,moreflow-andhencemoreadvectiveflux-isdrawnintothebranchwiththedilatedtube.Asthisisonexpenseofthethefluxintheparallelbranches,thesupplyintheparallelbranchesisdecreased.Theargumentiseasierforvesselsinthediffusiveregime.Here,thesupplyindownstreamvesselsisincreasedastheinfluxscaleswith(∆R)3/2,whiletheincreasedsupplyalongthedilatedvesselscaleswith∆R.Thusthenetoutfluxexitingthedilatedtubeisalsoincreased.Notethatontheonehand,theeffectofalteredsupplydownstreamisweakonthelevelofindividualvessels,butcorrelationssumuptheeffectinaspatialregionofthemicrovsasculatureandbythisincreasetheeffect.Noteontheotherhand,thatchangesindownstreamandparallelvesselsstronglydependonthenetworktopologyandthat∆Uissubjecttonon-linearcorrectionsduetothehematocrit.Anincreaseinsupplydownstreamthedilatedvesselandadecreaseinsupplyinvesselsparalleltothedilatedvesselcanthusonlybeconsideredasaruleofthumb,inlinewithFig.3(b).Toestimatethefulleffecttheexactnetworktopologyofthemicrovasculaturemustbeknown.[1]JoaoM.N.Duarte,FlorenceD.Morgenthaler,HongxiaLei,CarolPoitry-Yamate,andRolfGruetter,"Steady-StateBrainGlucoseTransportKineticsRe-EvaluatedwithaFour-StateConformationalModel,"Front.Neuroenerg.1,1 -- 10(2009).[2]H.Fischer,R.Gottschlich,andA.Seelig,"Blood-BrainBarrierPermeation:MolecularParametersGoverningPassiveDiffusion,"J.MembraneBiol.165,201 -- 211(1998).[3]WarrenL.LeeandAmiraKlip,"Shuttlingglucoseacrossbrainmicrovessels,withalittlehelpfromGLUT1andAMPkinase.FocusonAMPkinaseregulationofsugartransportinbraincapillaryendothelialcellsduringacutemetabolicstress,"Am.J.Physiol.CellPhysiol.303,C803 -- C805(2012).[4]LucLeybaert,"NeurobarrierCouplingintheBrain:APartnerofNeurovascularandNeurometabolicCoupling?"J.Cereb.BloodFlowMetab.25,2 -- 16(2005).[5]PabloBlinder,PhilbertS.Tsai,JohnP.Kaufhold,PerM.Knutsen,HarrySuhl,andDavidKleinfeld,"Thecorticalangiome:aninterconnectedvascularnetworkwithnoncolumnarpatternsofbloodflow,"Nat.Neurosci.16,889 -- 897(2013).7 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.S5.CHANGEINC0INSIMULATIONSFIG.S1.HistogramofchangeininitialconcentrationC0duetoasinglevesseldilatingby10%.TheadvectiveregimeshowsnochangeintheinitialconcentrationC0atthebeginningofavesselthatisdilated.Also,vesselsinthediffusiveregimeshowlittleresponseforsinglevesseldilation,withmorevesselsshowingadecreaseinC0comparedtovesselsintheadvectiveregime.ForthishistogramthesamesimulationsasshowninFig.2areevaluated.Linesindicatearangecovering69%withbothapercentageof15.5%showingalowerorhigherchangeinC0outsidetheindicatedrange.Foreachhistogram120randomvesseloftherespectiveregimewererandomlychosenanddilated.8 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.S6.CHANGEINSUPPLYFORALTERNATIVECHOICESOFγANDκANDDIFFERENTNETWORKBOUNDARYCONDITIONS-0.100.10.20.30.4Change in supply 05101520253035Probability density functiondiffusiveadvectiveall absorbingFIG.S2.TheadvectiveregimeisrobustinincreasingsupplybydilationalsoforanalternativeparametersetthanshowninFig.2.Here,andiffusivityofκ=1×10−8m2s−1andanabsorptionparameterofγ=200m−1arechosen.Histogramofchangeinsupply∆φduetoasinglevesseldilatingby10%.Linesspanfromthefirstquartileofthedatatothethirdquartile.Bigdotsindicatethemedian.Foreachhistogram450randomvesseloftherespectiveregimewererandomlychosenanddilated.-0.100.10.20.30.4Change in supply 05101520Probability density functiondiffusiveadvectiveall absorbingFixed pressure Fixed in(cid:31)owFIG.S3.TheadvectiveregimeisrobustinincreasingsupplybydilationalsoforanalternativeparametersetthanshowninFig.2andforafixedinflowboundarycondition.Here,andiffusivityofκ=1×10−8m2s−1andanabsorptionparameterofγ=800m−1arechosen.ThischoiceofparameterscorrespondstocenteringthedataaroundPe=1andS=1yieldingthemostequaldistributionofdatainthethreeidentifiedtransportregimes.Ontheleft,nodesattheboundaryofthenetworkwereheldatfixedpressureassinglevesselsweredilated.Ontheright,tubesstartingfromtheboundarywereheldatafixedinflowQassinglevesselswerdedilated.Histogramofchangeinsupply∆φduetoasinglevesseldilatingby10%.Linesspanfromthefirstquartileofthedatatothethirdquartile.Bigdotsindicatethemedian.Foreachhistogram450randomvesseloftherespectiveregimewererandomlychosenanddilated.AsmallernetworkexcerptcomparedtoFig.2waschosen,whichcontainsonly3340vessels.9 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.S7.IMPACTOFHEMATOCRITONROBUSTNESS-0.100.10.20.30.4Change in supply 010203040Probability density functiondiffusiveadvectiveall absorbingFIG.S4.Histogramofchangeinsupply∆φduetoasinglevesseldilatingby10%forthesameparametersasFig.2butneglectingtheeffectofhematocritwhencalculatingtheflowswithinthenetworksshowsnodifferencetooriginalFig.2.Linesspanfromthefirstquartileofthedatatothethirdquartile.Bigdotsindicatethemedian.Foreachhistogram450randomvesseloftherespectiveregimewererandomlychosenanddilated10 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.S8.DEVIATIONFROMROBUSTNESS00.511.52log10(qmin)-0.100.10.20.30.4Change in supply 012345log10(qmin)-0.100.10.20.30.4Change in supply 00.511.522.5log10(qmin)-0.100.10.20.30.4Change in supply all absorbingadvectivediffusiveFIG.S5.Thequalityofthetheoreticalpredictionisdecreasedclosetotheborderoftheregimes.Mappedoutonthey-axisisthechangeinsupplydisplayedinthehistograminFig.2.Mappedoutonthex-axisisthedistancetotheborderoftheregimeidentifiedinFig.1.qminisafoldchangeneededtoletavesselchangetheregime.Intuitively,log10(qmin)canbeunderstoodastheminimaldistanceofeachdatapointtoaregimeborderinthelogarithmicspace,i.e.distanceofpointfromtheborderinFig.1.Notablyvesselsfarintheadvectiveregimedisplaya10%increaseinsupplywhilevesselsclosetotheintersectionwiththeotherregimesarenotasrobustinagreementwithanalyticalcalculations.11 SupplementaryMaterial:RobustincreaseinsupplybyvesseldilationMeigeletal.S9.FLOWCHARTFORCALCULATIONOFABSORPTIONPROFILEKleinfeld data set Load in network topology:vessel length and radiusvessel connectivity matrixposition of verticesDetermine vessels at the edge of network excerptLoad in boundary conditionsOptional: modify vessel radiusCompute flow velocitiesCompute absorption profileCompare absorption profile with unpertubated absorption profile Kleinfeld data setFlow chart of full programCompute flow velocitiesComput conductance for each vesselCorrect conductance to account for hematocritCompute flows in the network using the matrix formalism of Kirchho's circuit lawsCompute flow velocities from flows through networkChoose boundary condition for network excerptKleinfeld data setCompute absorption profileInitialize empty lists:CalculatedVesselsCalculatedVerticesIs absorption calculated in all vessels?Find all calculatebale vertices:Vertices on boundary of network excerptVertices with all inflowing vessel calculatedSet dierence to list with all calculated verticesCalculate the absorption in all vessels outflowing from caluclateable verticesAdd newly calculated vessels to CalculatedVessels listAdd vertices to calculatedVertices listCalculated absorption profile in network excerptNoYesFIG.S6.Flowchartgivinganoverviewoverthenumericalproceduretocomputeabsorptionprofiles.Ontheleft,thefulloutlineofthenumericalprocedureisshown.Inthemiddle,thecalculationsofflowvelocitiesisvisualised,showinghowtheeffectforthehematocritisaccountedfor.Ontheright,theprocedureofthecalculationoftheabsorptionprofilesisshown.Wearehappytoprovideelementsorthefullcodeuponrequest.12 |
1911.00863 | 1 | 1911 | 2019-11-03T11:02:31 | Synchrony and symmetry-breaking in active flagellar coordination | [
"physics.bio-ph",
"cond-mat.soft",
"q-bio.CB"
] | Living creatures exhibit a remarkable diversity of locomotion mechanisms, evolving structures specialised for interacting with their environment. In the vast majority of cases, locomotor behaviours such as flying, crawling, and running, are orchestrated by nervous systems. Surprisingly, microorganisms can enact analogous movement gaits for swimming using multiple, fast-moving cellular protrusions called cilia and flagella. Here, I demonstrate intermittency, reversible rhythmogenesis, and gait mechanosensitivity in algal flagella, to reveal the active nature of locomotor patterning. In addition to maintaining free-swimming gaits, I show that the algal flagellar apparatus functions as a central pattern generator which encodes the beating of each flagellum in a network in a distinguishable manner. The latter provides a novel symmetry-breaking mechanism for cell reorientation. These findings imply that the capacity to generate and coordinate complex locomotor patterns does not require neural circuitry but rather the minimal ingredients are present in simple unicellular organisms. | physics.bio-ph | physics | Synchrony and symmetry-breaking in
active flagellar coordination
Kirsty Y. Wan1,2
2. College of Engineering, Mathematics, and Physical Sciences, University of Exeter, Exeter, UK
1. Living Systems Institute, University of Exeter, Exeter, UK
Keywords: Cilia coordination, basal coupling, locomotion, CPG oscillations, mechanosensitivity
Summary
Living creatures exhibit a remarkable diversity of locomotion mechanisms, evolving structures specialised for
interacting with their environment. In the vast majority of cases, locomotor behaviours such as flying,
crawling, and running, are orchestrated by nervous systems. Surprisingly, microorganisms can enact
analogous movement gaits for swimming using multiple, fast-moving cellular protrusions called cilia and
flagella. Here, I demonstrate intermittency, reversible rhythmogenesis, and gait mechanosensitivity in algal
flagella, to reveal the active nature of locomotor patterning. In addition to maintaining free-swimming gaits, I
show that the algal flagellar apparatus functions as a central pattern generator which encodes the beating of
each flagellum in a network in a distinguishable manner. The latter provides a novel symmetry-breaking
mechanism for cell reorientation. These findings imply that the capacity to generate and coordinate complex
locomotor patterns does not require neural circuitry but rather the minimal ingredients are present in simple
unicellular organisms.
1. The locomotor gaits of organisms
Since the invention of photography, the natural habits of organisms have come under increasing scrutiny.
Modern optical technologies have enabled resolution of ever-finer detail, so that we can visualise and track
behaviour across scales [1, 2]. Consider for example the gallop of a horse, how are its four feet coordinated?
This fast action cannot be resolved by the human eye, so the question remained unanswered until 1878, when
(reputedly to settle a bet) the photographer Eadweard Muybridge successfully imaged the gait sequences of a
horse [3]. Control of rhythmic limb movements in both vertebrates and invertebrates is determined by
neuronal networks termed central pattern generators (CPG), where motor feedback is not required [4]. Indeed,
CPGs in in vitro preparations can produce activity patterns that are identifiable with the in vivo behaviour. For
instance, the pteropod Clione limacine ("sea butterfly") swims with two dorsal wings, the neuronal pattern for
basic synchronous movement ("fictive swimming") can be reproduced in isolated pedal ganglia [5]. A striking
example of the genetic basis of gait-control and CPGs pertains to the Icelandic horse, where the ability to pace
(legs on the same side of the body moving synchronously) was shown to be associated with a point mutation
[6].
At the microscale, organisms interact with their world following very different physical principles.
Here inertia is negligible and many species developed motile appendages that are slender and suited to drag-
based propulsion through fluids. From these simplest designs, myriad locomotion strategies have evolved.
The model microswimmer E. coli is peritrichously flagellated -- bundles of rigid (prokaryotic) flagella rotate
synchronously in the same sense to elicit forward swimming (a "run"), but lose synchrony and unbundle (a
"tumble") when one or more of the motors reverse direction (see [7] and references therein). The latter results
in rapid changes in reorientation. In this way, a bacterium controls its propensity for reorientation to bias its
swimming towards or away from chemo- attractants or repellents [8]. It has been shown that bundling is
largely due to hydrodynamic interactions between the rotating filaments [9], but unbundling is stochastic. In
contrast, eukaryotes exhibit more deterministic responses to vectorial cues [10]. Phototactic reorientation by
the biflagellate alga Chlamydomonas reinhardtii requires bilateral symmetry-breaking in a pair of apparently
identical flagella [11], though differences in signal transduction must be present [12, 13, 14]. Eukaryotic flagella
and cilia possess a flexible and distributed molecular architecture allowing for many more degrees of freedom
than prokaryotic flagella. This versatility, particularly when multiple appendages are attached to the same
cell, allows algal flagellates to orchestrate diverse swimming gaits such as the breaststroke, trot, or gallop [15].
*Author for correspondence ([email protected])
Extensive research has already shown that distinct waveforms and beat frequencies can be produced by the
same ciliary structure [16, 17] (also Man et al, this Issue), depending on intrinsic and extrinsic forcing [18, 19,
20]. This ability to modulate dynein motor cooperativity to produce distinct beating modes on a single cilium
does not however explain the higher-level coordination over a network of such oscillating structures. To
achieve the latter, intracellular control mechanisms are implicated (reviewed in [21]). In this article, I present
new findings on gait patterning in microalgae, which reveal that single-celled multiflagellates can actively
dictate the dynamics and activity of each individual flagellum. I further propose that symmetry-breaking
processes in the flagellar apparatus are causal to this distinguishability between flagella. The attainment of
control specificity of locomotor appendages may be a key innovation in the evolution toward increasingly
deterministic movement in eukaryotes.
NUMBER OF LOCOMOTOR APPENDAGES
two
four
eight
sixteen
EXAMPLE SWIMMING GAITS
breaststroke
trot
π
0
0
π
π
0
0
π
π
0
COMPARISON
rotary breaststroke
0
π
π
0
0
π
0
π
π
0
wave
0
π
π
0
Figure 1. Locomotor patterning in unicellular algae. Algae with 2, 4, 8, and 16 flagella are shown with representative swimming
gaits. These are characterised by specific phase relationships between the flagella (0 to 2π for a complete beat cycle). (Different
species exhibit different gaits and swimming behaviours even for an identical arrangement of flagella.) Animal models with an
equivalent number of locomotor appendages are shown for comparison. These are respectively, a human swimmer, a horse, a jellyfish
ephyra (juvenile jellyfish), and the predatory sunflower sea star Pycnopodia helianthoides which walks with 16-24 limbs.
2. Gait species-specificity
We focus on unicellular algae bearing 2, 4, 8, or 16 flagella, and study the patterns of flagellar actuation and
movement. Representative gaits are compared to animals with analogous limb positioning (Figure 1), for
which the neuronal basis of movement control has been clearly demonstrated (except in the sea star which is
yet to be subject to systematic investigation). In all cases, movement operates in a very different physical and
size regime [22, 23] from that encountered by the microeukaryotes from this study. Species-specificity of
locomotor behaviours also applies to flagellates, organisms with an identical number and arrangement of
flagella can still assume different locomotor gaits [15]. Concurrently, a single species can be capable of multiple
gaits associated with significant changes in the flagellar beat pattern. For instance, sudden environmental
perturbations elicit so-called shock responses in multiple species, in which flagella reverse their beating
direction (but not the direction of wave propagation) to produce transient backward movement or even
cessation of swimming [24, 25]. Some species have a stop state [26], in which all flagella become reversibly
quiescent. We organise this section by characterising first the gaits used for forward swimming, and then
separately, introducing the phenomenon of selective beat activation in free-swimming cells.
2.1. Run gaits for forward swimming
The biflagellate C. reinhardtii is favoured for biophysical models of swimmers [27]. Cells execute a forward
breaststroke gait in which the two flagella are synchronized, interrupted by phase slips. Stochastic switching
between this characteristic in-phase breaststroke and a perfectly antiphase (or freestyle) gait was also
discovered in a phototaxis mutant [28]. Experimental and theoretical evidence overwhelmingly implicate
intracellular coupling as the dominant mechanism for flagellar coupling [29, 15, 30, 31]. Phase-synchrony is
not a given however, heterotrophic biflagellates such as Polytoma uvella have two flagella of slightly unequal
lengths and different beat frequencies, which remain largely asynchronous. With more than two flagella, new
gaits become accessible. Depending on flagella configuration and ultrastructure, different interflagellar phase
patterns are sustained during steady, forward swimming. Comparison with quadrupedal locomotion is
Figure 2: Active gait reconfiguration and heterogeneity in single-cell swimming trajectories. Persistent directional swimming (a)
is contrasted with asymmetric and chiral trajectories (b-f). Free-swimming cells of the quadriflagellate Tetraselmis subcordiformis are
shown in (b,c). In (b), a cell transitions from a spinning gait of two flagella (i), to a galloping gait of all four flagella (ii). In (c), beating
is restricted to only one of four flagella, resulting in a strongly helical trajectory. A different quadriflagellate Pyramimonas parkeae
exhibits intermittency: in trajectory (d) a cell alternates between a trot gait of four flagella (active state), and a stop state with no flagella
activity (inactive state). In (e), the octoflagellate Pyramimonas octopus undergoes reorientation following a shock response (see also
[26]). (f) A P. parkeae cell presenting a single actively beating flagellum. (g) The swimming speed of P. parkeae was measured in a
population of individuals, which shows a characteristic activity timescale for "bursting". [SM Videos 1-3]
instructive here [32]. Among species with four flagella, a robust trotting gait comprising two alternating pairs
of synchronous breaststrokes is assumed by two marine Pyramimonas species: P. parkeae and P. tychotreta,
whereas galloping gaits are consistently displayed by Tetraselmis sp. and Carteria crucifera (see SM Videos 1-5).
Only three extant species are known to have eight flagella [33]. Of these, Pyramimonas octopus coordinates its
eight flagella into a unique forward propulsion gait which we term the rotary breaststroke (see section 4).
Identifying equivalence between this and octopedal animal movement is more challenging. The spider
possesses obvious bilateral symmetry not present in the algae, though alternating patterns of leg movements
are observed [34]. Jellyfish ephyrae use cycles of contraction and relaxation involving simultaneous movement
of their eight soft appendages, but move in an inertial regime [35]. Finally, the Arctic species Pyramimonas
cyrtoptera has 16 flagella [36] which can be actuated metachronously rather than in discrete patterns. This may
be the critical number at which the flagella begin to interact hydrodynamically [15]. In this limiting scenario,
the wave-like coordination of many closely-separated appendages may help to avoid collisions, and mirrors
the locomotion mechanisms adopted by arthropods [37, 38].
2.2. Gait-switching and partial-activation
In animals, the capacity to produce multiple locomotor modes is critical for survival. Gait often depends on
the speed of desired movement. For example, the ghost crab becomes bipedal at the highest speeds [39].
Likewise, we find that unicellular flagellates dynamically reconfigure their motor apparatus to produce
different gaits that are coupled to photo- or mechano- sensory pathways. Next, we compare the motion
repertoire of different multiflagellate species. Freely-swimming cells were imaged at kilohertz and sub-
micrometer resolution, so that flagella motion could be discerned at the same time as the cell body movement.
Strong behavioural stereotypy was observed within each species, but heterogeneity across species (Figure 2).
Consistently, tracks presenting high curvature loops were found to result from a previously uncharacterised
motility mechanism -- namely, the partial activation of flagella. For the same cell, beating can be restricted to
a subset of flagella with the remaining flagella quiescent. Asymmetries arising from the actively beating
Figure 3: Quadriflagellate gaits resolved at the single-flagellum level in free-swimming cells. Displacement from tip to centre of
the cell body was used as proxy for flagellar phase for each flagellum, labelled (f1-4). Frames corresponding to specific timepoints are
highlighted. Scalebars for cell size reference. (a) Carteria crucifera has a galloping gait during forward runs, but exhibits noise-induced
shocks involving reorientation of the flagella (time 3). Recovery of the original rhythm (phase-resetting), completed within 4 beat
cycles. (b) Pyramimonas tychotreta has a canonical trotting gait during forward runs, but exhibits partial activation of flagella. Beating
is restricted to the pair f1,3, but suppressed in f2,4. Synchrony of f1,3 drifts from antiphase (times 3,4) to in-phase (time 5) over the
course of this recording. [SM Videos 4, 5]
flagellar subgroup produces large-angle turns in the free-swimming trajectory. A previously quiescent
flagellum can subsequently become active again, and vice versa (Figure 2b, and SM Video 1). Four different
quadriflagellate species showed selective activation of flagella (P. parkeae, T. suecica, T. subcordiforms, and P.
tychotreta). In cells exhibiting motion intermittency such as P. parkeae (see Figure 2d), flagella alternate between
periods of bursting activity and total quiescence (Figure 2g, SM), analogous to the physiological behaviour of
motor neurons.
Next, we resolve gait transients and interflagellar phase dynamics while keeping track of flagellum
identity. Automated methods could not follow the flagella movement reliably. Due to the 3D movement and
continuous cell axial rotation, features of interest do not always remain in focus. Instead, flagella tip positions
were manually annotated in Trackmate [40]. The distance between flagellum tip and cell centroid undergoes
oscillations for an actively beating flagellum, but remains constant for a quiescent flagellum. We use this
distance as proxy for flagellar phase (Figure 3). The quadriflagellate C. crucifera displays stochastic gait-
switching between forward swimming and transient shocks, as does P. octopus [26]. Shocks involve
simultaneous conversion of all four flagella to a hyperactivated undulatory state (Figure 3a). After ~0.3 s, the
canonical ciliary beat pattern is recovered. A rhythmic galloping gait resumes after only a few asynchronous
beat cycles. The rapidity of phase-resetting responses in these organisms further implies gait control is active
[41]. In Figure 3b, we demonstrate partial activation in P. tychotreta [42]. Here, full beat cycles are sustained in
flagella f1,3, while beating in the diametrically opposite pair f2,4 is suppressed throughout. Small fluctuations
detected in the inactive flagella are due to passive movement induced by the beating of nearby flagella. Further
examples of quadriflagellates in which beating is dynamically constrained to one, two, or three out of a total
possible four flagella, are provided in the SM.
3. Intermittency and temporal ordering in a flagellar network
Free-swimming individuals are not amenable to long-time imaging. In order to investigate the long-time
statistics of locomotor patterning, experiments were also conducted on micropipette-held (body-fixed)
organisms. Single or double micropipette configurations were used to position, manipulate single cells, and
to apply localised mechanical perturbations to single flagella. Not all species were suitable for the micropipette
technique however - in some cases captured cells prematurely shed their flagella. This is likely related to a
stress-induced, calcium-dependent deflagellation response [43, 44]. We identified two quadriflagellate species
that were minimally affected by pipette-capture. These are P. parkeae and T. suecica, which are representative
of the two known quadriflagellar arrangements: type I and II in [15] (see also Figure 4). In P. parkeae the flagella
form a cruciate arrangement, while in T. suecica they align into two antiparallel pairs. Both flagellates show
partial activation and beat intermittency.
For pipette-fixed individuals, flagellar phases could be extracted automatically from imaging data to
characterise the emergence and decay of coordination in the flagellar network. Phase patterns are reminiscent
of classical studies of footfall patterns and limbed locomotion in animals [45, 46]. In long-time, high-speed
recordings, reversible transitions in locomotor patterning were found in both species. These occurred
spontaneously, but could also be induced by mechanical forcing. This touch-dependent, gait-sensitivity is a
novel manifestation of a mechanoresponse in motile cilia and flagella. At the moment of capture, contact with
the pipette can induce rhythmogenesis in a cell with initially quiescent flagella (Figure 4a, SM Video 6).
Similarly, changes in locomotor patterning were induced when fluid was injected manually (~1s pulses) via a
second pipette in the vicinity of a quiescent cell (Figure 4b). The first pulse administered at time t1 induced
beating in f1,3 only. This biflagellate state continued until the second pulse (time t2), which then resulted in
beating in all four flagella (SM Video 7). In both cases, even after cessation of external perturbations, the cells
continued to exhibit intermittent activity over tens of seconds. We conclude that spontaneous transitions in
behaviour measured during free-swimming (Figure 2) are replicated in pipette-fixed individuals.
Next, we emphasize two aspects of gait reconfigurability in multiflagellates. First, neither activation
(e.g. Figure 4c) nor deactivation of beating (e.g. Figure 4d) necessarily occurs simultaneously in all four
flagella. However, the shock response (which is an all-or-none response depending on threshold perturbation
magnitude [26, 47] is associated with a strong process asymmetry, the forward reaction (activation) is
simultaneous but the reverse (deactivation) is sequential - different flagella can cease beating at different times
(Figure 4b). Second, after any gait transition, the reestablishment of an expected run gait (whether the trot, or
gallop) is not instantaneous, but rapid. The canonical temporal ordering of the flagella is recovered after only
a few beat cycles.
Figure 4: Reversible rhythmogenesis and gait-mechanosensitivity in quadriflagellates. (a) The four flagella of P. parkeae have a
cruciate arrangement (type I, top view). The captured cell has two flagella in the focal plane (f1,3) and two others moving transversely
-- these are labelled as f2 and are tracked as one (see SM Video 6). (b) The four flagella of T. suecica have an anti-parallel arrangement
(type II, top view). Mechanical perturbations are introduced at times t1,2, and spontaneous shocks are detected at times indicated by
red pointers. (c-f) are zoomed-in plots of (a-b), at time points of interest. At the moment of capture the flagella begin to beat, but not
simultaneously; a regular trot gait emerges within 3 beat cycles (c). Spontaneous termination of beating does not occur at the same
time in all flagella (d). Induced changes in locomotor patterning (e) are compared with the regular phase patterning (f) observed during
the trotting gait. (See SM Video 7). Inset: flagellar phase is defined according to progression through the beat cycle.
4. Symmetry-breaking in the flagellar apparatus
In this section we elaborate on the non-identity of flagella in multiflagellate algae. Adapting terminology from
gait analyses of limbed animal locomotion, we distinguish between symmetric gaits (e.g. the trot, pronk), and
asymmetric gaits (e.g. the gallop, bound). A quadrupedal gait is referred to as being symmetric if footfall
patterns of both pairs of feet are evenly-spaced in time, and as asymmetric if the activity of at least one pair is
unevenly spaced in time. Different gaits can therefore arise from the same underlying locomotor circuit. The
presence of more than two flagella also allows for diverse possibilities for symmetry-breaking even during
regular gait patterning (not transitional gaits).
We focus now on the rotary breaststroke of P. octopus, which has not been described in any other
organism. Gait ordering in multiflagellates is particularly difficult to visualise, due to the highly three-
dimensional swimming. Additionally, P. octopus rotates slowly clockwise about its own axis when viewed
from the anterior end (SM Video 8). All eight flagella are distinguishable simultaneously only when the cell is
swimming into or out of the focal plane (Figure 5, SM Video 9). Recordings of such transients were made, from
which the extremal positions of each flagellum could be followed in time and tracked manually in ImageJ. As
before, a normalised flagellar phase was computed for each flagellum. We find that phase ordering propagates
directionally, in the same sense as the about-axis rotation (Figure 5c). (Gaits are similarly directional in the
related P. cyrtoptera, which swims with 16 flagella.) A pairwise Pearson-correlation matrix was computed for
each tracked cell, to quantify the degree of correspondence between the eight flagella (Figure 5d). Strong
positive correlation was measured between diametrically opposite flagella (dashed white lines are guides),
which tend to move synchronously to produce a noisy sequence of intercalated breaststrokes (first 1:5, then
2:6, and so on).
triplets
Individual basal bodies
Given the directional nature of this swimming gait, asymmetries in developmental patterning must
be present. This is corroborated by the only ultrastructural study available on this species [48]. The flagella
emerge from separate basal bodies (centrioles during cell division), which are localised to the anterior of the
organism (Figure 5a). Electron microscopy sections reveal a complex basal architecture lacking in rotational
are
symmetry.
positioned uniquely according to generational
age [48]. A contractile network of thick and thin
fibres (some are striated) connect specific
numbered microtubule
to provide
intracellular coupling between flagella [15]. No
obvious structural or morphological differences
have been reported in the flagellar axonemes.
Consequently, it is not possible from light
microscopy data to prescribe the identity of the
individual flagella
in accordance with the
numbering system devised from TEM images.
Flagella numbering is thus distinguishable only
up to circular permutation. Nonetheless, robust in-
free-swimming
phase
pertains
flagella
(positive phase correlation, Figure 5d). This is in
good agreement with the presence of direct
physical coupling between distinguished pairs
of flagella. Particularly, the principal pair b1 &
b5 is connected by a thick striated structure
known as the synistosome, which is functionally
related to the distal fibre in C. reinhardtii [49].
The lack of correlation between certain basal
body pairs (Figure 4d) will be investigated
further in future studies to determine if this is
related to the detailed topology of the fibre
network (e.g. b2 and 8 appear unconnected).
synchrony during
to diametrically opposite
the canonical
Finally, we find that in ~5% of cases,
free-swimming cells exhibit a distinguished
flagellum which is held extended in front of the
cell body (SM Video 10). Figure 5 compares this
"search gait" with
forward
swimming gait of this organism. The beat
pattern in this single flagellum is undulatory
(sperm-like), in contrast to the lateral (ciliary)
Figure 5: P. octopus exhibits symmetry-breaking in both structure
stroke used by the other seven flagella. These
and movement. a) Schematic (redrawn from [48]) of the basal
observations further highlight the capacity for
apparatus comprising eight basal bodies (b1-b8) whence emanate the
producing heterodynamic behaviour
in a
flagella (f1-f8). Thick and thin filaments couple the flagella at specific
network of outwardly-identical flagella.
locations to produce the octoflagellate's unique rotary breaststroke (b).
Stereotypical phase dynamics for the numbered flagella are displayed
in (c). Cross-correlation matrices for one cell, and group average (n =
5. Discussions and Outlook
8), both show strong synchrony between diametrically-opposite flagella
(d). (e) A small number of cells adopt a "search" gait, in which a single
The present article is devoted to the gait
flagellum (marked by the red arrow) is held extended in front of the cell.
dynamics of
free-living unicellular algal
(SM Video 10). Canonical swimming is also shown for comparison.
flagellates, advancing the state-the-art from
single-cell to single-flagellum resolution. Performing a comparative study, I showed that gait coordination in
algal flagellates is not only intracellular but also active. The same network of flagella can undergo dynamic
reconfiguration over fast timescales to generate different locomotor patterns. Multiple species share an ability
to reversibly activate beating in a subset of flagella, while suppressing the activity of the remainder. This is an
extreme instantiation of symmetry-breaking in the algal flagella apparatus. Individually, a eukaryotic
flagellum or cilium can be in an oscillatory or non-oscillatory state, beating can be modulated by extrinsic
factors including mechanosensitivity [50, 26] (and Bezares Calderon et al, this Issue), but how can drastically
distinct beat patterns be produced in different flagella of the same cell? We suggest two sources of this apparent
control specificity in coupled flagella: i) physical asymmetries in the basal apparatus, and ii) differences in
signal transduction. The two contributions are not mutually exclusive.
In some species, physical asymmetries are obvious. Heterokont biflagellates can have one long
flagellum and one short, or one smooth and one hairy (bearing mastigonemes). In our case, control specificity
is all the more surprising as we only considered species whose flagella exhibit no exterior morphological
differences. During flagellar assembly and lengthening, tubulin subunits are shuttled from an intracellular
pool by molecular motors, and added to the tips of growing flagella by a process known as intraflagellar
transport [51]. This process is unlikely to result in structural heterogeneities between the individual flagella
axonemes. Instead, asymmetries arising from physiological differences in basal body age [52] and associated
contractile structures may be more significant. In flagellate green algae, basal body duplication is semi-
conservative. The two flagella of C. reinhardtii are termed cis or trans according to proximity to a single eyespot:
the trans basal body is inherited from the parental cell but the cis is formed anew. This distinction likely
underlies their differential calcium response, important during phototaxis [11]. The octoflagellate exhibits an
extreme form of this asymmetric ultrastructural patterning, where the eight flagella are again distinguishable
by the age of the attached basal body and organisation of accessory structures (Figure 5a).
Asymmetries in the control architecture itself are sufficient but not necessary to template asymmetric
dynamics. A radially-symmetric nerve ring innervates the 5-legged starfish [53], yet the organism routinely
performs a breaststroke gait differentiating between two pairs of side arms and a single leading arm. It would
therefore be interesting to explore whether the "special" flagellum in P. octopus is always the same flagellum
(as uniquely identified by basal body numbering), or whether different flagella can take turns to assume the
"exploratory" position (SM Video 10). Given the dual sensory and motor capabilities of cilia and flagella, this
could signify an early evolution of division of labour. We suggest that symmetry-breaking in excitatory or
inhibitory signalling in the algal flagella apparatus specifies the local state of contractility of intracellular fibres
to control the activation state of individual flagella. This may be likened to the case of vertebrate limbs, where
antagonistic control of flexor and extensor muscles is provided by motor neurons, or to marine invertebrate
larvae, where specialised ciliomotor neurons induce coordinated ciliary arrest [54]. We showed that in
multiflagellates the ability to activate subsets of flagella provides a novel mechanism for trajectory
reorientations which is distinct from steering in sperm (Alvarez et al, this Issue) and in other uniflagellates
which rely upon a change in the beat pattern of a single flagellum. A quadriflagellate can turn left, right, up
or down, depending on which of its four flagella is active. However, the extent to which organisms make use
of these capabilities warrants further study.
The diverse gaits of multiflagellates may have arisen from different evolutionary pressures associated
with a need to occupy different ecological niches. Even small genetic changes could cause sufficient genetic
rewiring of the basal apparatus and differences in locomotor pattern. With increasing flagella number, we
suggest that new gaits arose from superpositions of gaits from flagellates with fewer flagella. In this
perspective of gait modularity, the octoflagellate rotary breaststroke may have originated from intercalation
of two quadriflagellate trotting gaits, and the trot or gallop from intercalcation of two pairs of biflagellate gaits.
Indeed, it is thought that P. octopus branched from a related quadriflagellate Pyramimonas species via
incomplete cell division [48].
In conclusion, the capacity to generate and coordinate complex behaviour is by no means exclusive to
organisms possessing neural circuitry. In diverse unicellular flagellates, phase-resetting processes and
transitions in locomotor patterning occurred much faster than is possible by any passive means, implicating
excitable signalling. In future work, we will seek to identify structures capable of functioning as pacemakers
for locomotor patterning in flagellates. By analogy with CPGs, we suggest that parts of the algal cytoskeleton
may be independently capable of generating coupled oscillations. This has enabled these organisms to
dynamically reconfigure their gaits in response to environmental changes and uncertainty, without the need
for a central controller. The latter feat harkens to the embodiment perspective applicable to designing
bioinspired robots, in which "control of the whole", is "outsourced to the parts" [56]. As modern technology
strives towards greater automation, engineering adaptability into artificial systems has remained a formidable
challenge. In this sense, cell motility is in fact a form of physical embodiment, wherein the compliant
cytoskeleton and appendages enact morphological computation. In light of these findings, we may wish to
extend the scope of gait research beyond model vertebrates and invertebrates [23, 55] to include aneural
organisms. Much can be gained from exploring how simple unicellular organisms sculpt motor output and
achieve sensorimotor integration -- for herein lies the evolutionary origins of decentralised motion control.
Acknowledgements
I thank Ray Goldstein and Gáspár Jékely for discussions, and an anonymous referee for suggesting the
relevance of embodiment concepts from control theory. Financial support is gratefully acknowledged from
the University of Exeter.
References
[1] K. M. Taute, S. Gude, S. J. Tans and T. S. Shimizu, "High-throughput 3D tracking of bacteria on a
standard phase contrast microscope," Nature Communications, vol. 6, p. 8776, 11 2015.
[2] G. Catavitello, Y. Ivanenko, F. Lacquaniti, R. L. Calabrese and R. Chowdhury, "A kinematic synergy for
terrestrial locomotion shared by mammals and birds," eLife, vol. 7, p. e38190, 10 2018.
[3] E. Muybridge, Muybridge's complete human and animal locomotion. vol. 1, Dover Publications, 1979.
[4] A. J. Ijspeert, "Central pattern generators for locomotion control in animals and robots: A review,"
Neural Networks, vol. 21, pp. 642-653, 5 2008.
[5] Y. I. Arshavsky, T. G. Deliagina, G. N. Orlovsky, Y. V. Panchin, L. B. Popova and R. I. Sadreyev,
"Analysis of the central pattern generator for swimming in the mollusk Clione.," Annals of the New York
Academy of Sciences, vol. 860, pp. 51-69, 11 1998.
[6] L. S. Andersson, M. Larhammar, F. Memic, H. Wootz, D. Schwochow, C.-J. Rubin, K. Patra, T. Arnason,
L. Wellbring, G. Hjälm, F. Imsland, J. L. Petersen, M. E. McCue, J. R. Mickelson, G. Cothran, N. Ahituv,
L. Roepstorff, S. Mikko, A. Vallstedt, G. Lindgren, L. Andersson and K. Kullander, "Mutations in
DMRT3 affect locomotion in horses and spinal circuit function in mice," Nature, vol. 488, pp. 642-646, 8
2012.
[7] E. Lauga, "Bacterial hydrodynamics," Annual Review of Fluid Mechanics, vol. 48, pp. 105-130, 7 2019.
[8] D. F. Blair, "How bacteria sense and swim," Annual Review of Microbiology, vol. 49, pp. 489-522, 1995.
[9] Y. Man, W. Page, R. J. Poole and E. Lauga, "Bundling of elastic filaments induced by hydrodynamic
interactions," Physical Review Fluids, vol. 2, p. 123101, 12 2017.
[10] J. F. Jikeli, L. Alvarez, B. M. Friedrich, L. G. Wilson, R. Pascal, R. Colin, M. Pichlo, A. Rennhack, C.
Brenker and U. B. Kaupp, "Sperm navigation along helical paths in 3D chemoattractant landscapes,"
Nature Communications, vol. 6, p. 7985, 8 2015.
[11] U. Ruffer and W. Nultsch, "Comparison of the beating of cis- and trans-flagella of Chlamydomonas cells
held on micropipettes," Cell Motility and the Cytoskeleton, vol. 7, pp. 87-93, 1987.
[12] P. Hegemann, "Vision in microalgae," Planta, vol. 203, pp. 265-274, 11 1997.
[13] U. Ruffer and W. Nultsch, "Flagellar photoresponses of Chlamydomonas cells held on micropipettes II.
Change in flagellar beat pattern," Cell Motility and the Cytoskeleton, vol. 18, pp. 269-278, 1991.
[14] K. Y. Wan, K. C. Leptos and R. E. Goldstein, "Lag, lock, sync, slip: the many 'phases' of coupled
flagella," Journal of the Royal Society Interface, vol. 11, p. 20131160, 5 2014.
[15] K. Y. Wan and R. E. Goldstein, "Coordinated beating of algal flagella is mediated by basal coupling,"
Proceedings of the National Academy of Sciences of the United States of America, vol. 113, pp. E2784--E2793, 5
2016.
[16] D. M. Woolley and G. G. Vernon, "A study of helical and planar waves on sea urchin sperm flagella,
with a theory of how they are generated.," The Journal of experimental biology, vol. 204, pp. 1333-45, 4
2001.
[17] P. Sartori, V. F. Geyer, A. Scholich, F. Julicher and J. Howard, "Dynamic curvature regulation accounts
for the symmetric and asymmetric beats of Chlamydomonas flagella," Elife, vol. 5, p. e13258, 5 2016.
[18] A. C. H. Tsang, A. T. Lam and I. H. Riedel-Kruse, "Polygonal motion and adaptable phototaxis via
flagellar beat switching in the microswimmer Euglena gracilis," Nature Physics, vol. 14, pp. 1216-1222, 12
2018.
[19] B. Qin, A. Gopinath, J. Yang, J. P. Gollub and P. E. Arratia, "Flagellar kinematics and swimming of
algal cells in viscoelastic fluids," Scientific Reports, vol. 5, p. 9190, 3 2015.
[20] S. Y. Sun, J. T. Kaelber, M. Chen, X. Dong, Y. Nematbakhsh, J. Shi, M. Dougherty, C. T. Lim, M. F.
Schmid, W. Chiu and C. Y. He, "Flagellum couples cell shape to motility in Trypanosoma brucei,"
Proceedings of the National Academy of Sciences, vol. 115, pp. E5916-5925, 2018.
[21] K. Y. Wan, "Coordination of eukaryotic cilia and flagella," Essays in Biochemistry, 2018.
[22] J. C. Nawroth, K. E. Feitl, S. P. Colin, J. H. Costello and J. O. Dabiri, "Phenotypic plasticity in juvenile
jellyfish medusae facilitates effective animal-fluid interaction," Biology Letters, vol. 6, pp. 389-393, 6
2010.
[23] P. S. Katz, "Evolution of central pattern generators and rhythmic behaviours," Philosophical Transactions
of the Royal Society B: Biological Sciences, vol. 371, p. 20150057, 11 2015.
[24] J. A. Schmidt and R. Eckert, "Calcium couples flagellar reversal to photostimulation in Chlamydomonas
reinhardtii," Nature, vol. 262, pp. 713-715, 1976.
[25] I. Kunita, S. Kuroda, K. Ohki and T. Nakagaki, "Attempts to retreat from a dead-ended long capillary
by backward swimming in Paramecium," Frontiers in Microbiology, vol. 5, p. 270, 6 2014.
[26] K. Y. Wan and R. E. Goldstein, "Time irreversibility and criticality in the motility of a flagellate
microorganism," Physical Review Letters, vol. 121, p. 058103, 8 2018.
[27] R. Jeanneret, M. Contino and M. Polin, "A brief introduction to the model microswimmer
Chlamydomonas reinhardtii," European Physical Journal-Special Topics, vol. 225, pp. 2141-2156, 11 2016.
[28] K. C. Leptos, K. Y. Wan, M. Polin, I. Tuval, A. I. Pesci and R. E. Goldstein, "Antiphase synchronization
in a flagellar-dominance mutant of Chlamydomonas," Physical Review Letters, vol. 111, p. 158101, 10 2013.
[29] G. S. Klindt, C. Ruloff, C. Wagner and B. M. Friedrich, "In-phase and anti-phase flagellar
synchronization by waveform compliance and basal coupling," New Journal of Physics, vol. 19, p.
113052, 11 2017.
[30] G. Quaranta, M. E. Aubin-Tam and D. Tam, "Hydrodynamics versus intracellular coupling in the
synchronization of eukaryotic flagella," Physical Review Letters, vol. 115, p. 238101, 11 2015.
[31] Y. J. Liu, R. Claydon, M. Polin and D. R. Brumley, "Transitions in synchronization states of model cilia
through basal-connection coupling," Journal of the Royal Society Interface, vol. 15, p. 20180450, 10 2018.
[32] M. Golubitsky, I. Stewart, P. L. Buono and J. J. Collins, "A modular network for legged locomotion,"
Physica D, vol. 115, pp. 56-72, 4 1998.
[33] O. Moestrup and D. R. A. Hill, "Studies on the Genus Pyramimonas (Prasinophyceae) from Australian
and European Waters - P. propulsa sp. nov and P. mitra sp. nov," Phycologia, vol. 30, pp. 534-546, 11
1991.
[34] D. Wilson, "Stepping patterns in tarantula spiders," Journal of Experimental Biology, vol. 47, p. 133, 8
1967.
[35] Z. Ren, W. Hu, X. Dong and M. Sitti, "Multi-functional soft-bodied jellyfish-like swimming," Nature
Communications, vol. 10, 7 2019.
[36] N. Daugbjerg and O. Moestrup, "Ultrastructure of Pyramimonas cyrtoptera Sp nov (prasinophyceae), a
Species with 16 Flagella from Northern Foxe Basin, Arctic Canada, Including Observations on Growth-
rates," Canadian Journal of Botany-revue Canadienne De Botanique, vol. 70, pp. 1259-1273, 6 1992.
[37] B. Anderson, J. Shultz and B. Jayne, "Axial kinematics and muscle activity during terrestrial
locomotion of the centipede Scolopendra heros," Journal of Experimental Biology, vol. 198, p. 1185, 5 1995.
[38] T. Kano, K. Sakai, K. Yasui, D. Owaki and A. Ishiguro, "Decentralized control mechanism underlying
interlimb coordination of millipedes," Bioinspiration & Biomimetics, vol. 12, p. 036007, 4 2017.
[39] M. Burrows and G. Hoyle, "The Mechanism of rapid running in the ghost crab, Ocypode
ceratophthalma," Journal of Experimental Biology, vol. 58, pp. 327-349, 4 1973.
[40] J. Y. Tinevez, N. Perry, J. Schindelin, G. M. Hoopes, G. D. Reynolds, E. Laplantine, S. Y. Bednarek, S. L.
Shorte and K. W. Eliceiri, "TrackMate: An open and extensible platform for single-particle tracking,"
Methods, vol. 115, pp. 80-90, 2 2017.
[41] H. Guo, K. Y. Wan., J. Nawroth and E. Kanso, "Transitions in the synchronization modes of elastic
filaments through basal-coupling," in Bulletin of the American Physical Society, 2018.
[42] N. Daugbjerg, "Pyramimonas tychotreta, sp nov (Prasinophyceae), a new marine species from Antarctica:
Light and electron microscopy of the motile stage and notes on growth rates," Journal of Phycology, vol.
36, pp. 160-171, 2 2000.
[43] L. M. Quarmby and H. C. Hartzell, "Two distinct, calcium-mediated, signal-transduction pathways can
trigger deflagellation in Chlamydomonas reinhardtii," Journal of Cell Biology, vol. 124, pp. 807-815, 3 1994.
[44] G. L. Wheeler, I. Joint and C. Brownlee, "Rapid spatiotemporal patterning of cytosolic Ca2+ underlies
flagellar excision in Chlamydomonas reinhardtii," Plant Journal, vol. 53, pp. 401-413, 2 2008.
[45] J. Aguilar, T. N. Zhang, F. F. Qian, M. Kingsbury, B. McInroe, N. Mazouchova, C. Li, R. Maladen, C. H.
Gong, M. Travers, R. L. Hatton, H. Choset, P. B. Umbanhowar and D. I. Goldman, "A review on
locomotion robophysics: the study of movement at the intersection of robotics, soft matter and
dynamical systems," Reports on Progress in Physics, vol. 79, p. 110001, 11 2016.
[46] M. Grabowska, E. Godlewska, J. Schmidt and S. Daun-Gruhn, "Quadrupedal gaits in hexapod animals
- inter-leg coordination in free-walking adult stick insects," Journal of Experimental Biology, vol. 215, pp.
4255-4266, 12 2012.
[47] H. Harz., C. Nonnengasser and P. Hegemann, "The photoreceptor current of the green-alga
Chlamydomonas," Philosophical Transactions of the Royal Society B-biological Sciences, vol. 338, pp. 39-52, 10
1992.
[48] O. Moestrup and T. Hori, "Ultrastructure of the flagellar apparatus in Pyramimona octopus
(prasinophyceae) II: Flagellar roots, connecting fibers, and numbering of individual flagella in green-
algae," Protoplasma, vol. 148, pp. 41-56, 1989.
[49] D. L. Ringo, "Flagellar motion and fine structure of flagellar apparatus in Chlamydomonas," Journal of
Cell Biology, vol. 33, pp. 543-571, 1967.
[50] A. Mathijssen, J. Culver, M. S. Bhamla and M. Prakash, "Collective intercellular communication
through ultra-fast hydrodynamic trigger waves," Nature, vol. 10, pp. 560-564, 2019.
[51] J. L. Rosenbaum and G. B. Witman, "Intraflagellar transport," Nature Reviews Molecular Cell Biology, vol.
3, pp. 813-825, 11 2002.
[52] M. Bornens, "Cell polarity: having and making sense of direction-on the evolutionary significance of
the primary cilium/centrosome organ in Metazoa," Open Biology, vol. 8, p. 180052, 8 2018.
[53] W. Watanabe, T. Kano, S. Suzuki and A. Ishiguro, "A decentralized control scheme for orchestrating
versatile arm movements in ophiuroid omnidirectional locomotion," Journal of The Royal Society
Interface, vol. 9, pp. 102-109, 7 2011.
[54] C. Veraszto, N. Ueda, L. A. Bezares-Calderon, A. Panzera, E. A. Williams, R. Shahidi and G. Jekely,
"Ciliomotor circuitry underlying whole-body coordination of ciliary activity in the Platynereis larva,"
Elife, vol. 6, p. e26000, 5 2017.
[55] K. G. Pearson, "Common principles of motor control in vertebrates and invertebrates," Annual Reviews
of Neuroscience, vol. 16, pp. 265-297, 1993.
[56] R. Pfeifer, M. Lungarella and F. Iida, "Self-organization, embodiment, and biologically inspired
robotics," Science, vol. 318, pp. 1088-1093, 11 2007.
Legends and captions for electronic supplementary materials.
The following multimedia files (V1-10) accompany this study -- all scalebars are 10 µm.
1. V1 - A quadriflagellate gait transition from a spinning gait to a trotting gait.
This movie shows a free-swimming Tetraselmis subcordiformis cell undergoing a gait transition (track
corresponds to Figure 2b) from a spinning gait of two flagella, to a trotting gait of four flagella.
2. V2 - A quadriflagellate symmetry breaking gait.
This movie shows a free-swimming Tetraselmis subcordiformis cell exhibiting a symmetry-breaking gait
(track corresponds to Figure 2c) in which beating is restricted to only one of its four flagella.
3. V3 - Another quadriflagellate symmetry breaking gait.
This movie shows a free-swimming Pyramimonas parkeae cell exhibiting a symmetry-breaking gait (track
corresponds to Figure 2f) in which beating is restricted to only one of its four flagella.
4. V4 - A quadriflagellate resetting its forward-swimming gait after a shock response.
This movie shows a free-swimming Carteria crucifera cell before, during, and after a spontaneous shock
response (Figure 3a). Note the rapid phase-resetting dynamics of the flagella.
5. V5 - A quadriflagellate gait with two of four flagella active.
This movie shows a free-swimming Pyramimonas tychotreta cell in which beating is restricted to two of its
four flagella (Figure 3b).
6. V6 -- A quadriflagellate being caught by micropipette aspiration.
This movie shows a Pyramimonas parkeae cell with initially quiescent flagella being captured by a
micropipette (Figure 4a). Flagella beating is initiated at the moment of capture.
7. V7 -- Demonstrating gait-mechanosensitivity in a micropipette-fixed quadriflagellate.
This movie shows a micropipette-fixed Tetraselmis suecica cell exhibiting spontaneous and induced gait
transitions (Figure 4b).
8. V8 -- The axial rotation of an octoflagellate during swimming.
This movie shows the clockwise axial rotation of Pyramimonas octopus during free-swimming, and the
synchrony between pairs of diametrically opposite flagella.
9. V9 -- The rotary breaststroke of an octoflagellate during swimming.
This movie shows the rotary breaststroke of Pyramimonas octopus during free-swimming (Figure 5b). The
cell is viewed from posterior end.
10. V10 -- The octoflagellate search gait in which one flagellum is extended.
This movie shows the search gait of Pyramimonas octopus in which one flagellum is held extended in
front of the cell (Figure 5e).
|
1505.03313 | 2 | 1505 | 2015-10-02T10:14:53 | A semi-flexible model prediction for the polymerization force exerted by a living F-actin filament on a fixed wall | [
"physics.bio-ph",
"cond-mat.soft",
"cond-mat.stat-mech"
] | We consider a single living semi-flexible filament with persistence length l_p in chemical equilibrium with a solution of free monomers at fixed monomer chemical potential mu_1 and fixed temperature T. While one end of the filament is chemically active with single monomer (de)polymerization steps, the other end is grafted normally to a rigid wall to mimick a rigid network from which the filament under consideration emerges. A second rigid wall, parallel to the grafting wall, is fixed at distance L<<l_p from the filament seed. In supercritical conditions the filament tends to grow and impinges onto the second surface which, in suitable conditions (non-escaping filament regime) stops the filament growth. We first establish the grand-potential and derive some general properties, in particular the filament size distribution and the force exerted by the living filament on the obstacle wall. We apply this formalism to the semi-flexible, living, discrete Wormlike chain (d-WLC) model with step size d and persistence length l_p, hitting a hard wall. By original Monte-Carlo calculations we justify the use of the weak bending universal expressions of Gholami et al. (Phys.Rev.E. 74,(2006), 041803) over the whole non escaping filament regime. Employing this universal form for living filaments, we find that the average force exerted by a living filament on a wall at distance L is in practice L independent and very close to the value predicted by Hill, his expression being strictly valid in the rigid filament limit. The average filament force results from the product of the cumulative size fraction x, where the filament is in contact with the wall, times the buckling force on a filament of size L_c ~ L. We discuss several consequences of the L independence of the stalling force for our specific filament model. | physics.bio-ph | physics | A semi-flexible model prediction for the polymerization force exerted by a living
F-actin filament on a fixed wall.
Carlo Pierleoni,1, ∗ Giovanni Ciccotti,2, 3, † and Jean-Paul Ryckaert4, ‡
1Department of Physical and Chemical Sciences, University of L'Aquila,
and CNISM UdR L'Aquila, Via Vetoio 10, 67100 L'Aquila, Italy
2Physics Department, Sapienza University of Rome, P. A. Moro 5, 00185 Rome, Italy
3School of Physics, University College Dublin (UCD), Belfield, Dublin 4, Ireland
4Physics Department, Université Libre de Bruxelles (ULB),
Campus Plaine, CP 223, B-1050 Brussels, Belgium
(Dated: Monday 9th August, 2021)
We consider a single living semi-flexible filament with persistence length 'p in chemical equilibrium
with a solution of free monomers at fixed monomer chemical potential µ1 and fixed temperature T .
While one end of the filament is chemically active with single monomer (de)polymerization steps,
the other end is grafted normally to a rigid wall to mimick a rigid network from which the filament
under consideration emerges. A second rigid wall, parallel to the grafting wall, is fixed at distance
L << 'p from the filament seed. In supercritical conditions where monomer density ρ1 is higher
than the critical density ρ1c, the filament tends to polymerize and impinges onto the second surface
which, in suitable conditions (non-escaping filament regime) stops the filament growth. We first
establish the grand-potential Ω(µ1, T, L) of this system treated as an ideal reactive mixture and
derive some general properties, in particular the filament size distribution and the force exerted
by the living filament on the obstacle wall. We apply this formalism to the semi-flexible, living,
discrete Wormlike chain (d-WLC) model with step size d and persistence length 'p, hitting a hard
wall. Explicit properties require the computation of the mean force ¯fi(L) exerted by the wall at
L and associated potential ¯fi(L) = −dWi(L)/dL on a filament of fixed size i. By original Monte-
Carlo calculations for few filament lengths in a wide range of compression, we justify the use of
the weak bending universal expressions of Gholami et al.(Phys.Rev.E. 74,(2006), 041803) over the
whole non escaping filament regime. For a filament of size i with contour length Lc = (i − 1)d,
this universal form is rapidly growing from zero (non compression state) to the buckling value
over a compression range much narrower than the size d of a monomer.
fb(Lc, 'p) = π
Employing this universal form for living filaments, we find that the average force exerted by a living
filament on a wall at distance L is in practice L independent and very close to the value of the
stalling force F H
s = (kBT /d) ln(ρ1) predicted by Hill, this expression being strictly valid in the rigid
filament limit. The average filament force results from the product of the cumulative size fraction
x = x(L, 'p, ρ1), where the filament is in contact with the wall, times the buckling force on a filament
of size Lc ≈ L, namely F H
−2
for given ('p, ρ1) and x ∝ ln ρ1 for given ('p, L). At fixed (L, ρ1), one also has x ∝ '
−1
p which
indicates that the rigid filament limit 'p → ∞ is a singular limit in which an infinite force has zero
weight. Finally we derive the physically relevant threshold for filament escaping in the case of actin
filaments.
s = xfb(L; 'p). The observed L independence of F H
implies that x ∝ L
kB T 'p
4L2
c
2
s
5
1
0
2
t
c
O
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
3
1
3
3
0
.
5
0
5
1
:
v
i
X
r
a
∗ [email protected]
† [email protected]
‡ [email protected]
I.
INTRODUCTION.
2
Cytoskeleton actin filaments, with the help of a wide variety of auxiliary proteins, are at the root of dynamical
processes involved in cell motility[1, 2]. The growth of lamellipodium and filopedia is directly related to actin filaments
pushing or sometimes pulling (with the help of trans membrane proteins) with their barbed end pointing against the
cellular membrane. A subtle interplay of polymerizing or depolymerizing steps, involving single G-actin monomers at
the barbed end, provides the essential mechanism allowing the cytoskeletal network to keep contact while maintaining
a permanent pressure force on a load resisting membrane.
In vitro experiments on biofilaments, like actin and tubulin, to measure in supercritical conditions either the force-
velocity relationship in detectable non-zero velocity conditions[3, 4], or the approach to stalling where the applied load
effectively stops the net polymerization of living filaments and the stalling force is effectively measured[5], have been
deviced. To simplify the analysis and to concentrate on the fundamental process of force generation by polymerizing
filaments, the experiments deal with bundles of parallel filaments hitting an orthogonal moving wall, a network having
strong analogy with the structure of actin filopedia[1]. Data analysis requires models for bundle dynamics and stalling
force predictions, and in general, most models treat living filaments as perfectly rigid.
In a series of pioneering papers on this topics in the eighties, Hill was the first to propose the expression [6]
bun = Nf
F H
kBT
d
ln ρ1
(1)
for the force generated by a bundle of Nf growing (proto)filaments stopped by a normal wall. In Eq.(1), T is the
absolute temperature, d is the effective monomer size along the filament contour (equal to half the G-actin diameter
as there are two interwined protofilaments in F-actin) and ρ1 = U0/W0 > 1 is the reduced free monomer density
equal to the ratio of bulk polymerizing and depolymerizing rates U0 and W0. Using a combination of thermodynamic
and mean-field arguments [6], this expression has been established as the equilibrium state (zero growth velocity) of
a more general expression linking the wall velocity and the load force for a bundle of filaments in a generally non
equilibrium framework. As stressed by Hill, eq. (1) is derived from a one-dimensional longitudinal and incompressible
model which implies a proportionality between the average polymerization rate and the average wall velocity.
The 1D brownian ratchet model for an individual rigid filament hitting a moving wall was later proposed to offer
a physically justified stochastic model [7]. The living filament is subject to random polymerizing and depolymerizing
steps with respective rates U0 and W0 with U0 > W0 to treat supercritical conditions. While the depolymerizing step
is possible even in presence of the wall, the polymerizing step is only accepted if it does not lead to an overlap with the
moving wall. In addition the wall undergoes a 1D brownian motion characterized by a diffusion coefficient and by a
load which biases the wall dynamics towards the filament's end. The coupling between the filament (de)polymerization
dynamics and the wall random motion leads to a stalling force in agreement with Eq.(1) and to a stationary drift
velocity of the wall, which for large wall diffusion coefficient, agrees with Hill's prediction of the load-velocity law.
When many parallel filaments act together as a bundle, the brownian rachet model can be generalized to a multi-rigid
filament system while remaining essentially 1D [4, 8 -- 10]. The dynamical coupling among filaments via the common
wall evolution, which is very sensitive to the relative longitudinal disposition of the filaments, has strong implication on
the velocity-load relationship [3, 4, 9]. Let us note that all the above models consider non interacting filaments and a
single kind of actin-Adenosine triphosphate (actin-ATP) complex for the monomers whether free or incorporated into
filaments. The above dynamical models can be generalized to take into account lateral interactions among filaments
[11, 14] mainly to treat a many-protofilaments model and/or the hydrolysis of the ATP(Guanosine triphosphate-GTP)
in the filament actin-complexes (tubulin-complexes) by considering additional types of complexes, requiring in turn
additional information on specific rates [12 -- 14].
If rigid filament models are certainly satisfactory as long as the elementary working filaments (being isolated and
uncrosslinked) remain sufficiently short, the flexibility of F-actin should be properly considered for longer filaments.
Flexibility was found to be relevant in some important experimental cases. The bending shape of single F-actin
filaments observed by fluorescence spectroscopy, was precisely exploited to measure for the first time the typical
polymerization force generated by single living actin filaments [15, 16]. In the optical trap experiment of Footer et
al.[5], a bundle of about ten filaments, with their seeds glued to a trapped colloidal particle, push with their active
side (barbed end) on a fixed rigid wall. The polymerization force they progressively develop to reach equilibrium is
inferred by measuring the colloidal displacement in the trap. The interpretation of the experiments was possible only
by assuming the presence of escaping filaments in the bundle (filaments growing parallel to the obstacle wall after a
large angle bending fluctuation), a phenomenon interpreted by the authors as rod buckling related the beam elastic
instability [1]. Despite care in eliminating data potentially polluted by escaping filaments, the measured stalling force
for a eight filament bundle was (repeatedly) found to be close to Eq.(1) with Nf ≈ 1 instead of the expected Nf = 8
filaments number, a result still presently not understood.
3
That flexibility leads, in some extreme cases, to escaping filaments was reported and analyzed in a non equilibrium
simulation of a model of single living filament hitting a moving wall in which filament flexibility was explicitly taken
into account[17, 18]. Quite generally, in these pseudo-stationary simulations with constant load, a wall velocity
enhancement was found with respect to the predictions of the "rigid filament-hard wall" ratchet model, in agreement
with theoretical considerations which have generally predicted an enhancement of the efficiency of the conversion of
chemical free energy into useful work when realistic filament flexibility is included[19, 20]. For large loads (still below
the stalling force) and for large seed-wall distances, some escaping filaments were detected during the drift of the
wall [17, 18]. It was argued that this phenomenon is related but distinct from rod buckling and hence was denoted
as the "pushing catastrophe". The consensus seems to be that to efficiently grow against membrane resistance, actin
filaments should be neither too short (short filaments are too rigid to intercalate easily a polymerizing monomer
between the tip of the filament and the wall) nor too long as the load would simply buckle them, the optimal range
70nm − 500nm being cited in a recent review article[21].
In this paper we concentrate on the equilibrium Statistical Mechanical treatment of a semi-flexible filament in a
slab.
In section II, extending previous work [22], we establish within the reactive grand canonical ensemble, the
grand potential for a living filament in contact with an obstacle wall at fixed temperature and fixed free monomer
chemical potential. Formal expressions for the size distribution and the equilibrium force on the obstacle wall are
established. Section III deals specifically with F-actin modeled as a living discrete Wormlike Chain (d-WLC). We
first define the model and the related range of physical parameters to probe the non-escaping regime of the filament.
We then compute, by Monte Carlo simulation, the compression-force law for a dead (non-reacting) d-WLC in the
slab and validate, in the non-escaping regime, the weak bending expression of Gholami et al [25]. Subsequently we
define the filament force averaged over a distance equal to a monomer size d, crucial for the comparison with Hill's
prediction, In section IV we introduce the stalling force and compare the predictions for flexible (finite 'p) against
rigid ('p → ∞) models, proving for the latter Hill's expression for the stalling force. In the entire range of filament
lengths corresponding to the non-escaping regime, the flexible filament has a stalling force only few percents larger
than a rigid filament (Hill's law). Nonetheless, the specific L-dependence of the force (∼ L−2) resulting from buckled
filaments hitting the obstacle wall, induces a spectacular, previously undetected, effect of flexibility. Since the stalling
force is nearly independent of the slab's width L in the non-escaping regime, this requires a systematic evolution with
L2 of the fraction of sizes of the filament touching the wall. This is discussed in section V. Finally, section VI provides
some general conclusions and perspectives on the flexibility issue for many filaments bundles, including both static
properties and dynamic aspects linked directly to the exploration of the force-load relationship.
II. THE SINGLE GRAFTED LIVING FILAMENT IN A SLAB SYSTEM
A. The single grafted living filament concept
We consider a reacting ideal mixture in a slab at temperature T consisting of Nt monomers which can either be free
(G actin-ATP complex) or integrated within a single self-assembled filament (F-actin) with fixed persistence length
'p. In the F-actin case, 'p = 5370d and d = 2.7nm is the effective monomer size in the filament. The filament, with a
variable size i and associated contour length Lc = (i−1)d, is grafted normally to one of the walls of the slab considered
as an orthorhombic volume of transverse area A and width (wall to wall distance) L << 'p. The filament undergoes
single monomer (de)polymerization events with a polymerization rate U0 = konρ1, proportional to the free monomers
density ρ1, and a depolymerization rate W0 = kof f, independent on the free monomer density, where kon/of f are
the kinetic constants for the (de)polymerization reactions. Supercritical conditions are realized whenever the bulk
polymerization rate is larger than the depolymerization rate, which happens for ρ1 > kof f /kon = 1/K0, where K0 is
the bulk reaction equilibrium constant[23]. We define ρ1 ≡ ρ1K0 as the reduced free monomers density; supercritical
conditions correspond to ρ1 > 1. In super-critical conditions where polymerization dominates, the filament will grow
and hit the opposite wall as soon as Lc > L.
The series of possible chemical reaction will be denoted as
Ai−1 + A1 *) Ai
(2)
where Ai and A1 represent respectively the grafted filament of size i and a free monomer. At global equilibrium,
the chemical potentials µi of the different species involved in any reaction must satisfy the chemical equilibrium
requirement
(2 < i ≤ z∗)
(3)
This series of reactions is considered as limited to a size window going from a minimum filament size of two (to be
considered as an effective permanent seed of the filament) up to a maximum size z∗. Fixing a maximum filament
µi = µi−1 + µ1
4
FIG. 1. Two configurations of the same living filament grafted normally to the left wall through its first two monomers drawn
in dark grey, in chemical equilibrium with a free monomers solution. According to the whole set of (de)polymerizing reactions
Eqs.(2), the filament can polymerize or depolymerize at its free end by addition or removal of one single monomer as illustrated
by the arrows. In our illustrations (a,b), the contour length Lc of the polymerizing filament in (a) appears to be longer than
the distance L between the walls while in (b), the contour length Lc of the depolymerizing filament is shorter than the distance
L.
length is only necessary when considering flexible filaments. In fact for rigid 1D filaments, the obstacle hard wall will
necessarily limit the filament growth. Instead for flexible filaments, equilibrium Statistical Mechanics based on the
concept of a steady equilibrium state, can only be applied if we consider a mechanism limiting the filament growth in
particular if supercritical conditions are considered. Again if the slab is narrow enough with respect to the filament
persistence length, the obstacle wall will effectively limit the filament growth but for wider slabs we need to introduce
an artificial limit. In our present geometry (see fig. 1) we impose a maximum filament size
z∗ = IN T(cid:18) πL
2d(cid:19) .
(4)
In absence of this size limit, corresponding to filaments
where IN T(x) means the integer part of the argument.
adopting a planar configuration which covers a quarter of a cycle of radius L, filaments longer than z* could grow
unhindered parallel to obstacle wall, preventing the establishment of an equilibrium state. The imposed upper limit
will bias the statistical mechanics averages except for suitable external conditions (choice of control variables Nt/AL
and L in particular) for which the statistical weight of filaments of size z∗ or longer is negligibly small. This defines
what we call the "non-escaping" regime at stalling conditions.
In terms of the temperature T, volume V = AL, total number of monomers Nt and total number of grafted filaments
Nf, the reversible change of the relevant Helmholtz thermodynamic potential F R for this reactive system (hence the
R superscript) is
(5)
where S is the system entropy and where pN A and pT A are the total normal and tangential forces exerted by the
obstacle wall on the system. The last two terms involve the chemical potential µ1 of free monomers and the chemical
dF R = −SdT − pN AdL + pT LdA + µ1dNt + (µ2 − 2µ1)dNf
4LFigure ALFigure BFIG.1:Twoconfigurationsofthesamelivingfilamentgraftednormallytotheleftwallthroughitsfirsttwomonomersdrawnindarkgrey,inchemicalequilibriumwithafreemonomerssolution.Accordingtothewholesetof(de)polymerizingreactionsEqs.(2),thefilamentcanpolymerizeordepolymerizeatitsfreeendbyadditionorremovalofonesinglemonomerasillustratedbythearrows.Inourillustrations(a,b),thecontourlengthLcofthepolymerizingfilamentin(a)appearstobelongerthanthedistanceLbetweenthewallswhilein(b),thecontourlengthLcofthedepolymerizingfilamentisshorterthanthedistanceL.whereINT(x)meanstheintegerpartoftheargument.Inabsenceofthissizelimit,correspondingtofilamentsadoptingaplanarconfigurationwhichcoversaquarterofacycleofradiusL,filamentslongerthanz*couldgrowunhinderedparalleltoobstaclewall,preventingtheestablishmentofanequilibriumstate.Theimposedupperlimitwillbiasthestatisticalmechanicsaveragesexceptforsuitableexternalconditions(choiceofcontrolvariablesNt/ALandLinparticular)forwhichthestatisticalweightoffilamentsofsizezúorlongerisnegligiblysmall.Thisdefineswhatwecallthe"non-escaping"regimeatstallingconditions.IntermsofthetemperatureT,volumeV=AL,totalnumberofmonomersNtandtotalnumberofgraftedfilamentsNf,thereversiblechangeoftherelevantHelmholtzthermodynamicpotentialFRforthisreactivesystem(hencetheRsuperscript)isdFR=≠SdT≠pNAdL+pTLdA+µ1dNt+(µ2≠2µ1)dNf(5)whereSisthesystementropyandwherepNAandpTAarethetotalnormalandtangentialforcesexertedbytheobstaclewallonthesystem.Thelasttwotermsinvolvethechemicalpotentialµ1offreemonomersandthechemicalpotentialµ2ofgraftedfilaments(seeds)ofminimumsize2.ThesetwolasttermsresultfromimposingchemicalequilibriumEqs.(3)forallreactions(2)totheoriginalseriesoftermsqzúi=1µidNiinvolvingallspeciesofthemixture.Inthelattersum,thechemicalpotentialofanyfilamentwithsizei>2issubstitutedbyµi=µ2+(i≠2)µ1whichfollowsrecursivelyfromEqs.(3)andafinalregroupingoftermsyieldsEq.(5),takingintoaccounttheexpressionsofthetotalnumberofmonomersNt=N1+qzúi=2iNiandofthetotalnumberoffilamentsNf=qzúi=2Ni. potential µ2 of grafted filaments (seeds) of minimum size 2. These two last terms result from imposing chemical
equilibrium Eqs.(3) for all reactions (2) to the original series of termsPz∗
i=1 µidNi involving all species of the mixture.
In the latter sum, the chemical potential of any filament with size i > 2 is substituted by µi = µ2 + (i − 2)µ1 which
follows recursively from Eqs.(3) and a final regrouping of terms yields Eq.(5), taking into account the expressions of
the total number of monomers Nt = N1 +Pz∗
i=2 i Ni and of the total number of filaments Nf =Pz∗
i=2 Ni.
B. Free energy of a grafted living filament under confinement
5
Applying equilibrium statistical mechanics to a closed reacting ideal system[23], the canonical partition function
QR = exp [−βF R] for a single grafted filament in a solution of free monomers is given by
(Nt−z∗)
1
(Nt − z∗)! qz∗
(Nt−2)
QR(A, L, T, Nt, Nf = 1) = q
1
(Nt−3)
1
(Nt − 3)! q3 + ..
The sum over all distinct microscopic states compatible with the macroscopic variables is expressed in Eq.(6) as a
sum over (z∗ − 1) similar terms, each of them corresponding to one particular size of the single grafted filament and
the remaining free monomers. Each term of this ideal system involves the canonical partition function qi(L, T) of the
filament of size 2 ≤ i ≤ z∗ grafted in the slab and the corresponding contribution from the free monomers
(Nt − 2)! q2 + q
(6)
q
q1(L, A, T) = AL
Λ3
where Λ is the free monomer thermal de Broglie wavelength.
To each term qi(L, T) corresponds a canonical partition functions q0
i in the absence of the opposite wall. Keeping the temperature dependence implicit, we define the ratio's
As long as the intra-filament interactions have a local and homogeneous character, the ratio between successive parti-
tion functions q0
i is independent of i. Hence, as further detailed in section II D, we introduce a temperature
dependent equilibrium constant K0 [23]
i−1 and q0
α(i, L) = qi(L)
q0
i
(8)
i (T) relative to the same grafted filament of size
Using Eqs.(8,9), the partition functions of the filaments of any size i can be written as
q0
i
K0 ≡
q0
i−1q1/V
i
= q0
q0
i−1
Λ3
(7)
(9)
(10)
(13)
(14)
qi(L) = α(i, L)q0
Λ3(cid:19)i−2
2(cid:18) K0
(2 ≤ i ≤ z∗).
where q0
2 is the partition function of the grafted seed.
Eq.(6) can now be combined with expressions (7,10), to give
QR(A, L, T, Nt, Nf = 1) = q0
1
+ ..
2qNt−2
q−(z∗−2)
1
× α(2, L)
1 α(3, L)(cid:0) K0
Λ3(cid:1)
α(z∗, L)(cid:0) K0
(Nt − 2)! + q−1
(Nt − 3)!
(Nt − z∗)!
× α(2, L)
(Nt − 2)! + α(3, L)(cid:0) K0
V (cid:1)
V (cid:1)(z∗−2)
(Nt − 3)! + ..
(Nt − 2)! ×"α(2, L) + α(3, L)(cid:18) K0
(Nt − z∗)!#
V (cid:19)(z∗−2) (Nt − 2)!
V (cid:19) (Nt − 2) + ..α(z∗, L)(cid:18) K0
= q0
α(z∗, L)(cid:0) K0
(Nt − z∗)!
Λ3(cid:1)(z∗−2)
2qNt−2
2qNt−2
= q0
1
1
(11)
(12)
qNt−2
1
and the partition function can further be transformed as
2V 2
QR = q0
K2
0
2V 2
= q0
K2
0
(Nt − 2)! ×"α(2, L)(cid:18) K0
V (cid:19)2
+ α(3, L)(cid:18) K0
V (cid:19)3
(Nt − 2) + ..α(z∗, L)(cid:18) K0
Nt! ×"α(2, L)(cid:18) K0
(Nt − 2)! + α(3, L)(cid:18) K0
V (cid:19)2
V (cid:19)3
(Nt − z∗)!#
V (cid:19)z∗ (Nt − 2)!
(Nt − 3)! + ..α(z∗, L)(cid:18) K0
V (cid:19)z∗
qNt−2
1
Nt!
Nt!
Nt!
(Nt − z∗)!# (15)
Note that in the T.L. we may replace Nt by N1 = Nt − lf il (the average length of the filament) and ρt by ρ1, so
that the first term in the r.h.s. of Eq.(19) is the Helmholtz free energy of the bath of free monomers. While the third
term of Eq.(19) is the relevant free energy of the living filament, the middle term, function of T only, must be linked
to the free energy required to graft the filament seed (fixed dimer).
The probability for the living filament to have a size j, defined as P(j) ≡ P(j; L, ρt), is the term of index j in the
global partition function Eq.(6), properly normalized. Using the equivalent version of Eq.(16), leads to
P(j) = α(j, L)ρj
t
D
(20)
for j ∈ [2, z∗], where D is given by Eq.(17).
Performing a Legendre transform of the reactive Helmholtz free energy F R to the reactive grand potential ΩR =
F R − µ1Nt, Eq.(5) becomes
dΩR = −SdT − pN AdL + pT LdA − Ntdµ1 + (µ2 − 2µ1)dNf
(21)
To obtain ΩR, one needs to express µ1 in terms of the old variables according to Eq.(5) using Eq.(19) for F R and
associated D and ρt Eqs. (17,18). We get
βµ1 = ∂βF R
∂D(ρt)/∂ ρt
(22)
D(ρt)
where we have introduced the average length of the filament
= ln(cid:0)Λ3ρt(cid:1) −
ρt∂D(ρt)/∂ ρt
∂Nt
= ln(cid:0)Λ3ρt(cid:1) −
D(ρt) K0/V
1
Nt
= ln(cid:0)Λ3ρt(cid:1) −
lf il
Nt
6
(16)
(17)
(18)
(19)
(23)
(24)
(25)
(26)
In the thermodynamic limit (T.L.), here Nt → ∞, A → ∞ with fixed ratio Nt/A = ρtL, one gets
QR(A, L, T, Nt, Nf = 1) = q0
2Λ6
K2
0
qNt1
Nt! × D(ρt, L)
where we have defined
D(ρt, L) = α(2, L)ρ2
t + α(3, L)ρ3
t ... + α(z + 1, L)ρ
(z+1)
t
+ .. + α(z∗, L)ρ
(z∗)
t
ρt = ρtK0 = NtK0
V
and where we have replaced
We have thus
Nt!
t (Nt−i)! ≈ 1 in all terms of D(ρt, L).
N i
βF R(A, L, T, Nt, Nf = 1) = Nt [ln(cid:0)Λ3ρt(cid:1) − 1] − ln(cid:18) q0
2Λ6
K2
0 (cid:19) − ln D(ρt, L)
lf il(L, ρt) = Pz∗
j=2 j α(j, L) ρj
t
D
Formally the Legendre transform requires the inversion of Eq.(23) as Nt = Nt(µ1) to estimate
βΩR(A, L, T, µ1, Nf = 1) =(cid:2)βF R(A, L, T, Nt, Nf = 1) − Ntβµ1(cid:3)Nt=Nt(µ1)
Using Eqs.(19,23), one gets successively
βΩR(A, L, T, µ1, Nf = 1) =(cid:20)−Nt + lf il − ln(cid:18) q0
2Λ6
K2
0 (cid:19) − ln D(ρt)(cid:21)Nt=Nt(µ1)
="−Nt + lf il −
l2
f il
Nt − ln(cid:18) q0
2Λ6
K2
0 (cid:19) − ln D(ρ1)#Nt=Nt(µ1)
where we have developed D(ρt) around ρ1 up to first order. We also note that Eq.(23) can be rewritten as
βµ1 = ln(cid:0)Λ3ρ1(cid:1) + O(cid:18) lf il
Nt(cid:19)2
≈ ln(cid:0)Λ3 ρ1/K0(cid:1)
where the central expression is the expected relationship for the chemical potential of one species in an ideal mixture,
the negligible correction coming from the approximations made earlier to simplify the QR Eq. (15).
Neglecting the term l
in Eq.(26), the grand potential can finally be reformulated as
2
f il
Nt
βΩR(A, L, T, µ1, Nf = 1) =(cid:20)−N1 − ln(cid:18) q0
2Λ6
K2
0 (cid:19) − ln D(ρ1)(cid:21)ρ1= K0
Λ3 exp (βµ1)
In the biophysics literature it is customary to use the reduced free monomer density ρ1 as the independent variable
instead of the more appropriate chemical potential µ1. Therefore we will re-express the grand potential in Eq.(28) as
(29)
(30)
βΩR(A, L, T, µ1, Nf = 1) = βΩfree(A, L, T, ρ1) + βΩfil(L, T, ρ1)
βΩfree(A, L, T, ρ1) = −
AL
K0
ρ1
Ωfil(L, T, ρ1) = −kBT(cid:20)ln(cid:18) q0
2Λ6
K2
0 (cid:19) + ln D(ρ1)(cid:21)
where we indentify the grand canonical contributions Ωfree(A, L, T, ρ1) and Ωfil(L, T, ρ1) for the free monomers and
the grafted living filament respectively.
The filament size distribution, the normalization factor D and the average size of the filament, given respectively
by Eqs.(20,17,24), take the final form
P(j) ≡ P(j; L, µ1) ≡
Nj
Nf
= α(j, L)ρj
1
D
α(j, L)ρj
1
D(µ1) =
z∗Xj=2
¯lf il(L, µ1) = Pz∗
j=2 j α(j, L) ρj
1
D
= ρ1
∂ ln D
∂ ρ1
= ∂ ln D
∂βµ1
where ρ1 in the r.h.s. is again used instead of µ1 and where Nj is the average number of filaments with size j within
the microscopic states of the reactive grand canonical ensemble.
C. Single filament force exerted on the opposite wall
Combining Eqs. (21) and (29), and noting that pN A is the sum of a free monomer contribution and the single
filament average force f⊥(L, ρ1), one gets
7
(27)
(28)
(31)
(32)
(33)
(34)
(35)
(36)
(38)
i=2
∂α(i,L)
∂L (ρ1)i
D
β ¯fi(L)P(i; L, ρ1)
βf⊥(L, ρ1) = −(cid:18) ∂(βΩf il)
=
z∗Xi=2
∂L
∂ ln α(i, L)
P(i; L, ρ1) =
∂L (cid:19) =(cid:18) ∂ ln D
∂L (cid:19) = Pz∗
z∗Xi=2
Wi(L) = −kBT ln α(i, L).
¯fi(L) = −
∂Wi(L)
∂L
where we have used Eq.(32) and introduced a filament mean force potential and associated mean force at fixed length
(37)
Eq. (36) gives the equilibrium force exerted on a living grafted filament by a fixed planar wall located at a distance L
from the grafting wall. As expected, it is the average of the force exerted by the wall on a fixed length "dead" grafted
filament (this latter force is an average over its internal degrees of freedom), weighted by the absolute probability
P(i; L, ρ1) of having a filament of length i. Of course, only the filaments sufficiently long to interact with the wall
(α(i, L) 6= 1) contribute to the average.
D. Equilibrium constants and rates
8
Here, we discuss a few properties of the equilibrium constants valid for the arbitrary grafted flexible filament.
For the considered ideal mixture of N1 free monomers and the series of Ni grafted filaments of size i, one has[23]
βµ1 = − ln(cid:18) q1
βµi = − ln(cid:18) qi
N1(cid:19) = ln Λ3 + ln ρ1
Ni(cid:19) = − ln(cid:18) α(i, L)q0
Nf
i
(cid:19) + ln P(i)
(39)
(40)
(41)
using Eqs.(7,8,32).
Substituting Eqs.(39,40) in Eq.(3) for any reaction given by Eq.(2), one gets
= α(i, L)
α(i − 1, L) K0(T)
= α(i, L)
α(i − 1, L)
Ki(L, T) ≡
P(i − 1)ρ1
i Λ3
q0
q0
i−1
P(i)
which defines the equilibrium constants Ki and its link with the equilibrium constant K0, already defined in Eq.(9),
which would apply in absence of wall. Eq.(41) expresses the evolution of the equilibrium constant Ki(L, T) with
increasing i, as a result of interferences between filaments of sizes i and (i − 1) and the wall.
Some considerations on the related wall influence on the (de)polymerisation reaction rates are provided in appendix
A, given their close connection to the equilibrium constants Ki. These rates become essential ingredients of the present
approach when extended to the study of the coupling of a mobile wall dynamics and the filament (de)polymerization
steps.
III. THE WORMLIKE CHAIN MODEL AND THE F-ACTIN CASE.
A. The discrete model.
To model the living grafted filament with fluctuating size in the range 2 ≤ i ≤ z∗, we adopt the d-WLC model with
discrete contour length step d and persistence length 'p. Using a cartesian reference frame where the grafting wall is
at x = 0 and the obstacle wall at x = L, the filament normally grafted at the wall at x = 0 has its two first monomers
located at ¯r1 = (0, 0, 0), ¯r2 = (d, 0, 0). The filament with i monomers, having a contour length Lc,i = (i − 1)d has a
configuration fully specified by the set of coordinates [ ¯rj]j=1,i including the grafted dimer. The instantaneous internal
potential energy of the filament of size i is
i−1Xk=2
iYj=2
E([¯r]i) = −(i − 1)00 + κ
d
[1 − cos θk]
(42)
where κ = kBT 'p is the bending modulus of the filament, 00 the bonding energy associated to the chemical step
Eq.(2)[22] and θk the angle between successive bonds implying monomers (k − 1, k, k + 1). The configuration having
the minimum energy corresponds to the straight filament with all bending angles at zero. The monomer-wall potential
is zero or infinite depending whether the articulation point (monomer) j is in the slab space (0 ≤ xj ≤ L) or lies
inside the obstacle wall (xj > L). If we represent by U w the global filament-wall interaction potential, being the sum
of all monomer-wall potentials, according to Eq.(8) the factors α(i, L) become
α(i, L) =< exp−[βU w] >i,0=<
Θ(L − xj) >i,0
(43)
where < ... >i,0 denotes a canonical average with weight exp (−βE([¯r]i)) (Eq.(42)) of a grafted filament of size i in
absence of the obstacle wall, and Θ is the Heaviside function. For this model, of contour length Lci = (i − 1)d, we
have qi(L) = qi0 and hence α(i, L) = 1, as long as i ≤ z, where z is the integer given by
(44)
In the case of a living filament undergoing (de)polymerizing reactions and for short enough filaments (i ≤ z), this
WLC model leads to the following expression for the equilibrium constant, as defined in Eq.(9), [22]
z = IN T(L/d) + 1
K0 = 2π exp (β00) d4
'p
[1 − exp (−2'p/d)] ≈ 2π exp (β00) d4
'p
.
(45)
in terms of the fundamental parameters d, 'p, 00 of the filament model. The equilibrium constant for filaments hitting
the wall is given by Eq.(41) where the α factors are given by Eq.(43).
9
B. Explicit calculations for the F-actin case.
1. The relevant L and ρ1 regime to probe single F-actin polymerization force.
The essential ingredients to get static properties of grafted actin filaments are the wall factors α(i, L) for a grafted
d-WLC hitting a hard wall (z∗ > i > z), with 'p = 5370d and d = 2.7nm. On this basis, all properties can be derived
for any supercritical value of the reduced monomer density ρ1 > 1.
We first consider the relevant range of wall position L and the range of reduced free monomer concentrations ρ1 for
which the polymerization force is operative and of interest for a quantitative comparison with in vitro experiments[4,
5, 15]. According to Mogilner[21], to produce a working force, individual filaments should be longer than Lc = 70nm
(about 25 monomers) to avoid being too rigid but should remain below Lc = 500nm (about 185 monomers) to avoid
what he refers to as buckling. In Footer et al. experiments[5], the polymerization force was measured for non buckled
filaments of length 200nm (about 70 monomers). In filopedia bundles[24], parallel filaments are cross linked by fascin
but free portions of filaments at the leading edge are supposed to be of the order of 20− 200nm. Finally in the recent
experiment of Démoulin et al.[4] the bundle length studied to get the velocity load relationship is of the order of
100 − 400nm (about 40 − 150 monomers per filament) (see supplemental information of ref. [4]). Further, it has to
be noted that in vitro experiments probe the polymerization force in moderate supercritical conditions (1 < ρ1 < 3)
to avoid too rapid buckling and interferences with spontaneous nucleation of new filaments [4, 5].
In our illustrative section of the F-actin case, we will concentrate on the supercriticality regime by considering two
values of the reduced density ρ1 = 1.7 and ρ1 = 2.5. We will be interested to the wall position regime 20d < L < 100d
where actin filaments are sufficiently long to avoid unphysical influence of minimum size filaments (j = 2) but still
sufficiently short to avoid escaping filaments, as it will be made more precise later.
2. The compression law of a grafted (fixed size) filament.
The basic input of our theory are the α functions (see Eq.(8)) of "dead" filaments of contour length Lc = (n − 1)d
(z < n ≤ z∗). The force ¯f(L, T; Lc) exerted by a wall located at L on a d-WLC filament with 'p = 5370d and of
contour length Lc ≥ L (see Eq.(38)) (using here a notation without size index as we now deal with a unique dead
filament hitting the wall) has been computed by Monte-Carlo simulation. The resulting force-compression laws for
three filament sizes (n = 41, 77, 158) are shown in Figure 2. The MC sampling was realized by a mixture of two
types of attempted moves, i) local crankshaft moves, where a sequence of three, four or five articulation points are
rotated as a rigid body around an axis joining the two surrounding articulation points, and ii) pivot moves implying a
global rigid rotation around a bond of the end chain fragment starting from that bond (the size of the fragment being
sampled between 1 and (n − 3) articulation points). The force exerted by the filament on the wall was estimated as
(46)
1
¯f(L, T; Lc)/kBT = lim
∆→0
∆ ln(cid:20)
q(L, T; Lc)
q(L − ∆, T; Lc)(cid:21)
where q(L, T; Lc) is the partition function of a single grafted filament of contour length Lc. This force is easily
estimated during the MC sampling by measuring the probability that the filament configuration has an articulation
point located in the region of thickness ∆ adjacent to the wall.
In Figure 2, we observe that as L decreases down from L = Lc (where the force vanishes) the force quickly increases
up to a pseudo-plateau before undergoing a final steep rise as L approaches a value of ≈ 10d − 15d on its way down
2 L/d(cid:1) corresponding to the non escaping
to zero. As we are interested to filament lengths limited to z∗ = IN T(cid:0) π
regime, it implies that only the elasticity law for the regime Lc < L < 2
π Lc needs to be exploited for each Lc. The
relevant regime for our present study terminates within the pseudo plateau at Lc − L = (1 − 2/π)Lc indicated by a
vertical arrow for each of the lengths reported in the Figure. As it will be discussed elsewhere[26], the fast increase
of the elasticity force at short L corresponds to a Lc independent behavior ¯f = AkBT lp/L2, with A ≈ 2/3, valid for
escaped filament lengths Lc/d >> z∗.
Gholami et al.[25] derived a weak bending approximation for the elasticity law of a grafted continuous wormlike
chain. Their prediction for the L << 'p regime, leads to the universal law summarized here. Using notations of
reference[25], the identification with our formalism for a dead filament of size n and contour length Lc = (n − 1)d
10
FIG. 2. Compressional force ¯f(L, T ; Lc) in units kBT /d exerted by a grafted dead d-WLC filament of contour length Lc
(Lc = (n − 1)d with n = 41, 77, 158) and persistence length 'p = 5370d on an obstacle hard wall oriented normally to the
filament grafting direction and located at distance L from the filament's seed. For each filament size considered, we observe
three successive regimes as L decreases from Lc. First, a rapid rise corresponding to the weak bending regime, followed by
2 divergent
a pseudo-plateau regime which terminates with the onset of the escaping filament regime characterized by a 1/L
behavior as L approaches 0 (as (Lc − L) approaches Lc on the Figure). The L threshold value where the filament enters the
escaping filament regime is indicated by a vertical arrow for each filament length (see text). In inset, all force data ¯f(L, T ; Lc)
are reduced by the theoretical weak bending plateau value fb(L, Lc) for the continuous WLC model (see Eq.(52)) and plotted
versus the renormalized compression distance η (see Eq.(48)) in order to test the weak bending expression Eq.(53) shown by a
continuous green line.
leads to
in terms of a new reduced compression distance
involving the characteristic length
α(n, L) ≡ Zk(η)
η = Lc − L
Lk
Lk = L2
c
'p
.
(47)
(48)
(49)
(50)
The central quantity Zk is (see eqs. (36) or (38) in reference [25])
∞Xk=1(cid:2)(−1)k+1λ−1
Zk(η) = 2
kη](cid:3)
k exp [−λ2
2 . The (microscopically) averaged force ¯f(L, T; Lc), defined by eq.(38), which is the force exerted
where λk = (2k−1) π
by the wall on a (dead) grafted WLC filament of contour length Lc hitting the normal hard wall at seed-wall distance
L, is[25]
¯f(L; Lc) = fb fk(η)
(51)
0.011100 Lc-L0.0010.010.1110 f_(L,T;Lc)0.1110100 η∼01f~ Lc=157 Lc=76 Lc=40 where fb turns out to be equivalent to the buckling force for a clamped rod of contour length Lc [1], namely
fb = π2
4
kBT
Lk
and where fk(η) is a universal function defined by
fk(η) = −
4
π2
∂ ln Zk(η)
∂η
11
(52)
(53)
This function, shown in the inset of Figure 2, starts from 0 at η = 0 and increases monotonically to a unity plateau
which is reached around η = 0.25. We argue that the Gholami et al. elasticity function is worth exploiting not only
for the weak bending regime (limited to η ≈ 0.6) where it is rather precise, but also for the intermediate pseudo
plateau regime up to Lc/L = π/2 ≈ z∗/z where the force appears to be underestimated by 10 − 15 percent only.
When this approximation is made for our purpose, the gain is enormous as we do not have to run a large number of
single filament MC simulations to get the force for each specific filament size n (Lc = (n − 1)d) as a function of the
continuous L variable. We get all the needed expressions as functions of a single universal variable η (Eq.(48)) under
the form of an explicit convergent series easy to compute (see Eqs.(50,53)).
3. Living filament force in the grand canonical ensemble and the L average force concept.
Given the properties of the WLC discussed in the the previous subsection, the general expression of the polymer-
ization force for the d-WLC model, Eq.(36), takes the explicit form
βf⊥(L, ρ1) =
z∗Xi=z+1
β ¯fi(L)P(i; L, ρ1)
(54)
where z is defined in Eq.(44) and where ¯fi(L) and P(i; L, ρ1) are given by Eqs.(38,32,33), computed with model
functions Eqs.(51,47).
In Figure 3, we report the polymerization force f⊥ for the d-WLC model adapted to F-actin persistence length, as
given by Eq.(54), for two values of ρ1 and highlighted within specific ranges of L. We observe a pseudo periodic signal
of period λ ≈ d. In the lower L regime (upper panel), only the first term in the rhs of Eq.(54) (i = z + 1) contributes
to the force. To discuss the behavior of the polymerization force, let us focus on the interval where L changes from
L = 28d up to L = 29d (where z = 29), noting that the force f⊥(L) is essentially zero at both boundaries. In this
interval, f⊥(L) ∝ ¯f30(L)α(30; L)/D(L). Given the moderate variation of D with L and the fact that the force ¯f30
is essentially constant except when L approaches the filament contour length Lc = 29d by less than 0.05d (η ≈ 0.3)
where it drops quickly to zero when L reaches Lc, the rise of f⊥(L) reflects the L dependence of P(30; L, ρ1) and thus
the dependence of α(30, L) from practically zero (for L < 28.5d) towards unity. The fast drop of f⊥(L) towards zero
at L = 29d comes from the drop of ¯f30. As the L domain increases (lower panel), the variation becomes more complex
as it involves increasingly more terms in Eq.(54). We note in Figure 3 that Hill's prediction for the polymerisation
force lies very close to the average value f⊥(L) within any interval.
In his seminal paper, Hill[6] introduced the stalling force through the work needed to add reversibly a new monomer
to a rigid filament pressing normally against a wall, as the wall moves by a distance d (See Eq.2 in ref[6]). This
implicitly defines the average of the living filament force over an interval [L, L+ d]. On the basis of a reversible change
of the grand potential Eq.(21), the reversible work at constant T and constant µ1 performed by the filament pressing
against a wall moving from L up to L + d satisfies
WL,L+d(L, µ1, T) = Ωf il(L, µ1, T) − Ωf il(L + d, µ1, T) = kBT ln(cid:20) D(L + d, ρ1)
D(L, ρ1) (cid:21) ≡ F av
f ild.
(55)
where we have defined the average force F av
f il over the d interval.
The average force concept appearing in Eq.(55) can be used for rigid filaments but also for flexible filaments as
modeled by the d-WLC model, if one adapts specifically the D(L, ρ1) terms (in particular the α functions) in Eq.(33).
In order to do so, we first reexpress this average force in an equivalent but more appropriate way to discuss the
specificities of filament flexibility. Given the definition of z in Eq.(44), we have z(L + d) = z(L) + 1. For the upper
limit z∗ given by Eq.(4), one has z∗(L + d) = z∗(L) + k where k = 1 or k = 2 depending upon the precise L value
12
FIG. 3. L dependence of the polymerization force Eq.(36) for a F-actin single living filament, modeled by the living d-WLC.
The force is shown for two values of the reduced free monomer concentration, ρ1 = 1.7 (black curve) and ρ1 = 2.5 (red curve),
as a function of L in the interval [27d, 30d] (upper panel) and in the interval [67d, 70d] (lower panel). The dashed horizontal
lines show Hill's predicted value for the reduced force ln ρ1 at each considered free monomer reduced concentration.
which is truncated by the integer value operator in eq. (4). We can then rewrite D(L + d) in alternative equivalent
ways:
ρi−1
1 +
α(j, L + d) ρj−1
D(L + d) = ρ1
z(L+d)Xi=2
= ρ1ρ1 +
z(L)Xm=2
= ρ1ρ1 +
z(L)Xm=2
z∗(L+d)Xj=z(L+d)+1
z∗(L+d)−1Xn=z(L)+1
z∗(L)Xn=z(L)+1
ρm
1 +
ρm
1 +
1
1
α(n + 1, L + d)ρn
(56)
(57)
(58)
α(n + 1, L + d)ρn
(z∗(L+d)−1)
1 + α(z∗(L + d), L + d)ρ
1
Dropping the last term, present only when z∗(L+ d)− z∗(L) = 2 but which is anyway negligibly small in non escaping
regime conditions, it leads to
D(L + d) ' ρ1ρ1 +
z∗(L)Xn=z(L)+1
f il(L, µ1, T) = ln ρ1 + ln [ ρ1 + Dshif t(L)]
z(L)Xm=2
ρm
1 +
α(n + 1, L + d)ρn
1
(59)
(60)
βdF av
D(L)
2727.52828.52929.530 L/d0246810 f⊥ d /kBT ρ^1=1.7 ρ^1=2.56767.56868.56969.570 L/d00.511.52 f⊥ d/kBT where use of Eq.(55) has been made and where Dshif t is obtained by substitution of all terms α(j, L) in D by
αshif t(j, L) = α(j +1, L+ d) for all j > z(L). Note that the first term in eq. (60) is the single filament polymerization
force of Hill.
A similar calculation for the variation of the average size of the filament as the wall moves reversibly from L to
L + d gives, according to Eqs.(34,55,60),
13
∆¯lf il(L, ρ1) ≡ ¯lf il(L + d, ρ1) − ¯lf il(L, ρ1) = ∂
∂βµ1
= 1 + ρ1
ln(cid:20) D(L + d, ρ1)
D(L, ρ1) (cid:21) = ∂
ln(cid:20)[ρ1 + Dshif t(L, ρ1)]
D(L, ρ1)
∂
∂ ρ1
∂βµ1
(cid:21)
[βWL,L+d]
(61)
(62)
where the first unity term is also the Hill's result for rigid filaments hitting normally the obstacle wall[6].
In fact, in both Eqs.(60) and (62), the second term on the r.h.s. gives the correction arising from two different
effects. The first effect is the imposed minimal size of filaments which manifests itself at low L by the additive term
ρ1 in the numerator of the argument of the logarithm. The second effect linked to the ratio Dshif t(L,ρ1)
is the effect
of flexibility by opposition to the purely rigid case ('p = ∞) where Dshif t = D. Before embarking on this analysis,
presented in Section IV, we derive the precise criteria to be satisfied, in supercritical conditions, to remain in the non
escaping regime for the flexible filament case.
D(L,ρ1)
4. Non-escaping regime criteria.
To avoid the presence of escaping filaments, one needs to have at equilibrium a negligible probability for filament
size of the order of z∗. This can best stated by comparing this probability to the (near) maximum value at i = z in
supercritical conditions, giving
P(z)(cid:21) = α(z∗, L)ρ
(cid:20) P(z∗)
(z∗−z)
1
<< 1
(63)
where we have used Eq.(32). Taking the logarithm of both sides and using Eq.(37), one gets
(64)
The mean force potential Wz∗(L) is the reversible work to compress a grafted filament of size z∗ until it fits within
the space limited by a hard wall at L. Using the approximate universal expression of the force in the weak bending
limit Eq.(51) and treating the plateau value fbz∗ as constant over the whole compression interval, one gets
−βWz∗(L) + (z∗ − z) ln ρ1 < 0
−βfbz∗( π
2 − 1)L + ( π
2 − 1) L
d
ln ρ1 < 0,
Substituting Eq.(52) for fbz∗ in Eq.(65), the final expression of the non escaping regime condition reads
ρ1 < exp(cid:18) 'pd
L2(cid:19)
(65)
(66)
which can be used either to limit ρ1 at given L or to limit L at given ρ1.
A comment about the evolution of size populations in the intermediate size window z < i ≤ z∗ for any situation
where condition (63) or equivalently (66) is met is in order. According to Eqs.(63), (64), one has
P(i)/P(z) ≈ exp(cid:18)(i − z)(cid:20)ln ρ1 −
π2'p
4d
1
(i − 1)2(cid:21)(cid:19)
z < i
(67)
where we have again assumed the compressive force to be constant over the whole compression interval (Wi(L) =
(Lc − L)fbi = [(i − 1)d − L]π2'p/[4(i − 1)2d]) and we have assumed L ' (z − 1)d according to Eq.(44). The argument
of the exponential is the product of a positive term (i − z) and the factor in square brackets where the first constant
and positive term is dominated by the negative second term at the lowest i = z + 1 values as condition (66) is met.
According to Eq.(67), P(i) must diverge as P(i) ∝ ρi1 when i grows to infinity. Therefore, the ratio in Eq.(67) must
pass through a minimum (lower than unity) at some size imin. So if z∗ < imin (low ρ1 value), the ratio P(i)/P(z)
decreases monotonously over the relevant physical regime limited to z∗, down to a small value required by Eq.(66).
14
∗)/P (z) (in logarithmic scale) of the probability
FIG. 4. Single F-actin filament with 'p = 5370d. L dependence of the ratio P (z
∗ monomers over the probability to have z monomers. We report the ratio for two values of the free monomer
to have z
reduced density ρ1 = 1.7 (blue, continuous line) and ρ1 = 2.5 (red, dashed line). The horizontal green line marks the value
∗)/P (z) = 0.001 for which we consider that the bias induced by the constraint is negligibly small. The figure suggests
P (z
maximal values of L to be Lmax/d ' 89 for ρ1 = 1.7 and Lmax/d ' 70 for ρ1 = 2.5 (see also Eq.(68)). The red horizontal
corresponding to P (z
∗)/P (z) = 1 level is reached for L
2
l = 'pd/ ln ρ1 which directly follows from Eq.(66).
Otherwise, if imin is located in the relevant z < i < z∗ regime (higher ρ1 value), the criteria (63) implies that the
ratio P(imin)/Pz must be even lower than P(z∗)/P(z) so that kinetically, small filaments growing against the wall
will see their size limited at values below imin.
Eq.(66) predicts that the range of L values where the wall can effectively stop the bundle polymerization in
supercritical conditions is limited by Ll =p'pd/ ln ρ1, namely Ll ' 100 and Ll ' 76 for ρ1 = 1.7 and 2.5 respectively.
In practice, for fixed ρ1 > 1, we will limit the non escaping regime at the lower value Lmax which corresponds to
P(z∗)/P(z) = 0.001, as illustrated in Figure 4. The dependence of Lmax upon ρ1, empirically established, is
Lmax(ρ1) = Ll(ρ1) − ∆(ρ1) =q'pd/ ln ρ1 − ∆(ρ1)
∆(ρ1) = 24.538 − 10.695 ρ1 + 1.3965 ρ2
(68)
(69)
where the ∆ term has been fitted in the ρ1 range between 1.7 and 4. This relation provides Lmax ' 89 and Lmax ' 70
for ρ1 = 1.7 and 2.5 respectively, as seen in Fig.4. As a point of comparison, the measurement of the polymerisation
force in an optical trap set up for what appears to be a single actin filament in supercritical conditions at ρ1 = 1.7
[5] (see Fig 4b of that reference), involves an elongation of L ≈ 200nm which corresponds to about L = 74d and is
compatible with the present definition of the non-escaping regime.
Condition (66) indicates that the concept of "non-escaping regime" is valid for flexible filaments only, since when
'p → ∞ the inequality is satisfied for any finite ρ1 value.
1
IV. THE STALLING FORCE AND ENERGY CONVERSION FOR F-ACTIN.
A. The rigid living filament case.
The living filament polymerization force F av
f il(L, µ1, T), shown for various cases in Figure 5, is derived in Eq. (60)
as the average of the L dependent force over an interval equal to the single monomer size d. In the limiting case of a
rigid filament, any α(i, L) is a step function being unity as long as L ≥ Lci (implying i ≤ z(L)) and zero otherwise.
Hence one has, using the Eq.(33) of D and Eq.(60) of Dshif t,
Dshif t(L) = D(L) = ρ2
1
1(1 − ρz−1
1 − ρ1
)
.
(70)
406080100120L/d1e-181e-151e-121e-091e-060.00111e+031e+06 P(z*)/P(z) 15
FIG. 5. The polymerization force, averaged over an interval from L = nd to L = (n + 1)d and reduced by kBT /d, is shown
as a function of n for a single living filament at supercritical conditions ρ1 = 1.7 (black curve) and ρ1 = 2.5 (red curve). At
each density, the flexible d-WLC case ('p = 5370d, data points) is compared to the rigid limit case ('p = ∞, dashed lines)
on the basis of Eq.(60), which simplifies to Eq.(71) for the rigid case. The peculiar short L behavior, common to all curves,
essentially reflects the boundary effect related to the imposition of a filament minimum size imin = 2. Beyond L > ¯L (see text),
the rigid filament average force goes to the asymptotic value ln ρ1 in agreement with Hill's expression. Curves are deliberately
interrupted in the Figure at L = Lmax which is the upper range of the non escaping regime according to Eq.(68).
Using this relation, the average force in Eq. (60) can be recast for the rigid filament case as
dF av
f il
kBT
= ln ρ1 + ln(cid:20)1 +
ρ1z(L) − ρ1(cid:21).
ρ1 − 1
(71)
The correction (second) term in Eq.(71) is numerically important at small L only. For it to be of order , one has to
go beyond ¯L given by
¯L/d ≈ IN T(cid:18)ln(ρ1 − 1) − ln()
ln(ρ1)
(cid:19)
(72)
which follows from Eq.(71) and from the link between z and L in Eq.(44). This boundary problem for rigid filaments
is illustrated in Figure 5 where it can be observed that the Hill's result, ln ρ1, is indeed valid asymptotically beyond
a value of L = ¯L ≈ 8d computed from Eq.(72) at ρ1 = 2.5 for = 0.001.
Similarly in the rigid filament case, the size increment as the wall position L is displaced by d is given by combination
of Eq.(62) and Eq.(70),
∆lf il(L, ρ1)) = 1 + ρ1
ln(cid:20)1 +
ρ1z(L) − ρ1(cid:21)
ρ1 − 1
ρ1z−1 − 1(cid:21)
ρ1z − 1 + (1 − z)
Again, the correction to unity vanishes for large L/d (L > ¯L where ¯L is provided by Eq.(72)). Figure 6 shows the
increment becoming asymptotically unity for the rigid filament case as L increases. As commented by Hill[6] and
shown in Figure 7, the ratio of the reversible work performed by the polymerization force to displace the wall by
a distance d over the corresponding chemical free energy (µ1 − µ1c)∆'f il used to increase the average length of the
filament, goes to unity asymptotically for the rigid filament case.
∂
∂ ρ1
z
= 1 +(cid:20)
(73)
020406080 L/d0.50.60.70.80.911.1 d Ffilav/kBT ρ^1=2.5 ρ^1=1.7 16
FIG. 6. The increase ∆lf il(L, ρ1)) of the average length of a living filament pressing against a wall as the latter is moved from
L = nd to L = (n + 1)d under supercritical conditions specified by ρ1 = 1.7 (black curve) or ρ1 = 2.5 (red curve) is reported
according to Eq.(61). The rigid model case ('p = ∞, dashed lines) (also given by Eq.(73)) and the flexible case ('p = 5370d,
data points) are compared. The behavior at short L (L < ¯L) results from the imposition of a lower end boundary condition on
filament length, namely imin = 2.
FIG. 7. L dependence of the ratio of the reversible work of the polymerization force WL,L+d over the interval L ∈ [nd, (n + 1)d]
and the corresponding chemical energy estimated as ∆lf il(L, ρ1)kBT ln ρ1. The results are shown for flexible filaments ('p =
5370d) at two free monomers reduced densities, ρ1 = 1.7 (black points) and ρ1 = 2.5 (red points). For L > ¯L, the observed
central plateau value of unity reflects a perfect energy conversion. For the flexible filament case, the ratio starts decreasing
progressively as L approaches the upper limit of the non-escaping regime.
020406080 L/d0.70.80.911.1∆lfilρ^=1.7ρ^=2.5020406080 L/d0.960.9811.021.04βWL,L+d /(∆lfil log(ρ^1))ρ^1=1.7ρ^1=2.5 17
B. The flexible living filament case adapted to F-actin.
The effect of flexibility for a living grafted filament on the average polymerization force and on its size increment
as the wall is displaced by the monomer size d needs to be investigated for supercritical conditions in the regime
¯L < L < Lmax. The higher limit, Eq.(68), was justified in section III B 4 while the lower limit turns out to be in
practice identical for flexible and rigid cases as illustrated in Figure 5 for two ρ1 values. Adopting = 0.001, one has
for ρ1 = 1.7, ¯L = 12d and Lmax = 89d and for ρ1 = 2.5, one has ¯L = 7.9 and Lmax = 70.
and for the model of a WLC hitting a hard wall, the correction term is necessarily positive given the inequality
Considering the general expression for the average force Eq.(60), we first establish that in the relevant L regime
Dshif t(L) > D(L)
(f lexible f ilaments)
(74)
This inequality basically follows from the property that α(i, L), as given by Eqs.
(47), (50), is a monotonously
decreasing function when L decreases, or equivalently when η increases (see Eq.(48)). This property is intuitively
obvious and verified by visual inspection illustrative figures in ref.
[25]. To justify this on the basis of Eq.(50), we
note that it is a sum of decaying exponentials in terms of η but with alternating sign. Grouping terms in pairs, the
even k− th term and the odd (k +1)− th term, we obtain an absolutely converging series. The justification of Eq.(74)
follows from the property that, provided the sums of terms in D(L) and Dshif t(L) converge sufficiently fast (before
reaching i = z∗, thus well within the non-escaping regime), the two expressions can be compared term by term.
The strict inequality α(j + 1, L + d) > α(j, L) for each pair of corresponding terms follows from the fact that, while
Lc − L is identical, the corresponding reduced compressions η (Eq.(48)) is smaller for the Dshif t(L) term, implying
a larger value for α. In Figure 5, the average force F av
f il for a flexible actin living filament is shown as a function of
L. The resulting curve in the regime ¯L < L < Lmax is never very different from the rigid case, and thus from Hill's
prediction. The averaged force is slightly above the Hill's plateau value by a marginal 1− 2 % in the upper domain of
the non escaping regime. This can be interpreted by noting that the α's are related to the fraction of possible chain
conformations for a chain of given number of monomer in presence of a rigid obstacle at distance L < Lc. Intuitively
this number should increase with the chain flexibility which is equivalent of taking longer chains Lc + d at larger
distance L + d for given 'p.
The effect of flexibility on the filament size increment in the regime L > ¯L gives, starting with Eq. (62),
∆¯lf il(L, ρ1) ≈ 1 + ρ1
∂
∂ ρ1
= 1 + [¯lshif t
(cid:21)
ln(cid:20) Dshif t(L, ρ1)
(L, ρ1) − ¯lf il(L, ρ1)]
D(L, ρ1)
f il
(75)
f il
where ¯lf il(L, ρ1) is given by Eq.(34) and ¯lshif t
(L, ρ1) is given by the same expression with probabilities P(i) =
α(i, L)ρi1/D(L, ρ1) for i > z substituted by P(i)shif t = αshif t(i, L)ρi1/Dshif t(L, ρ1) (see also Eq.(60)).
On Figure 6, ∆¯lf il(L, ρ1) computed with Eq.(62) for the flexible case is again found to be close to the rigid limit.
The value is however a few percents higher than unity in the upper part of the non escaping filaments domain, a
logical result arising from the bending fluctuations of the filaments. Note that the approximate expression Eq.(75)
(data not shown) gives identical results, except in the L < ¯L domain.
In Figure 7, the ratio of the reversible work of the polymerization force over the corresponding chemical free energy
used to polymerize the living filament, goes also to unity for the flexible case at least in the central domain of the
non escaping filament regime. At larger L, the ratio becomes lower than unity by a few percents, indicating that the
conversion of chemical energy into work becomes affected by the flexible character of the filaments. Obviously, the
situation quickly worsens if the filaments start to escape, which would happen with large probability if L gets larger
than Lmax by 5 − 10 monomer units (Eqs.(68,69)).
V. DISTRIBUTION OF FILAMENT SIZES PRESSING AGAINST A FIXED WALL FOR F-ACTIN.
In this last section, we analyze the influence of flexibility on the equilibrium distribution of filament sizes when a
living filament in supercritical conditions, is stopped by a normal hard wall. The size distribution, given by Eqs.(32),
(33), takes in the rigid limit the form of a truncated growing exponential (P(i; L, ρ1) ∝ exp (i(ln ρ1)) for (i = 2, z(L))
and P(i; L, ρ1) = 0 for (i > z(L)) and for (i = 1)). In order to avoid large fluctuations of the probabilities of hitting
filaments as L varies over a monomer size distance d, we discuss results for the size distribution of flexible filaments in
terms of an average over wall positions, as discussed earlier for the equilibrium polymerization force. Let Qk(L, ρ1) be
the probability to have a filament of relative size k = i − z(L) with respect to the fixed wall position L. The average
< Qk >n, computed as the average of Qk over the interval L/d ∈ [n, n + 1], is shown in Fig. 8 for several values of L
covering the entire non-escaping regime. In Fig. 9 we show the average fraction hx0in of filament sizes touching the
wall, obtained as the cumulative sum of < Qk >n over the positive values of k.
18
FIG. 8. Average normalized distribution < Qk >n of filament relative sizes k = i − z(L) at ρ1 = 1.7, for wall position averaged
over the interval [nd, (n + 1)d] shown for various n = L/d values within the non escaping regime. The exponential rise is
observed for filaments sizes avoiding the wall, while the decay for filament sizes hitting the wall becomes increasingly sharper
as L decreases.
FIG. 9. Fraction hx0i of filaments hitting the wall divided by ln ρ1 plotted versus L for two supercritical densities ρ1 = 1.7
(black circles) and at ρ1 = 2.5 (red squares). A clear trend hx0i ∝ ln ρ1L
2 for all data in the non escaping regime is indicated
by the green continuous line.
These results show the most spectacular features of semi-flexible filaments with respect to rigid case. The (average)
f il for filaments like F-actin in supercritical conditions and in the non escaping
s =
ln ρ1. However, the way this force is produced by the living filament is highly L dependent and it is essentially
equilibrium polymerization force F av
regime, is observed to remain essentially L independent and equal to the standard rigid filament result of Hill F H
kB T
d
-25-20-15-10-50510 k10-610-510-410-310-210-1100 <Qk>n L=25d L=35d L=45d L=55d L=65d L=75d L=85d100 L0.11 <x0> / ln( ρ^1) ρ^1=1.7 ρ^1=2.5 obtained as the product of two factors with inverse L dependencies
F av
f il ≈ hx0i
π2
4
kBT 'p
L2
19
(76)
as seen in Fig. 9. This first order expression means that the required force is produced at wall position L by recruiting
a buckled filament of length Lc ≈ L with a weight hx0i ∝ ln ρ1L2, while with weight (1 − hx0i), the filament does not
contribute to the force on the wall, being it shorter than L. Indeed, filaments in contact with the wall are mostly in
a compressed state corresponding to a force into the plateau region (see Fig. 2) therefore providing a force fb given
by Eq.(52) in the force expression Eq.(51). This means for a wall located at L = 50d that the plateau is reached
as soon as (Lc − L) > 0.1d which is most often the case (see Fig. 8). In eq. (76) we also replaced 1/L2
c ≈ 1/L2.
Finally, it should be noticed that Eq. (76) also predicts that hx0i ∼ '−1
p hence, in the rigid limit 'p → ∞, hx0i → 0
which demonstrates the impulsive character of the force when the rigid filament hits the wall during its Brownian
fluctuations.
VI. DISCUSSION AND CONCLUSIONS.
The ability of actin filaments to sustain in supercritical conditions a compressive force has predominantly been
justified with the aid of the rigid living filament model, effectively a one dimensional model, both for the single
filament case and for bundles of parallel filaments[4, 6 -- 13]. The 1D filaments which are fluctuating in length as a
result of (de)polymerizing steps are hitting a fluctuating obstacle usually subject to load by producing instantaneous
kicks which result into a time averaged force biasing the obstacle brownian motion. These brownian ratchet dynamical
models do lead to an effective polymerizing force, compatible with Hill's force expression at stalling, which satisfies
a specific velocity-load relationship either for the single filament case [7] or for few bundle models differing mainly
by the longitudinal disposition of filaments [4, 8 -- 10]. When Hill's prediction for a bundle of Nf > 1 actin filaments
has been found to fail in interpreting experimental data [5], the role of flexibility has been invoked by introducing an
ad-hoc maximum value for the force that a bunble can exert and beyond which the bundle buckles. However even
this ad-hoc extension of the 1D model did not provide a satisfactory interpretation of the experiments.
In the present paper and in its future extensions, we develop a Statistical Mechanics theory for a flexible filament
model and we show that, quite generally, it leads to a systematic and continuous evolution of the filament behavior
from a rigid rod character at short contour lengths to a pronounced flexible character at longer contour lengths,
ending ultimately with the so called pushing catastrophe limit[17, 18], when the filament(s) has(ve) acquired a finite
probability to grow unimpeded by the wall and escape laterally. While our approach will involve ultimately multi-
filament bundles, moving obstacles under various loads thus implying non equilibrium situations, in this work we have
focused on the already very rich phenomenology offered by the basic equilibrium properties of a single grafted filament
in supercritical conditions, as it hits a fixed wall oriented normally to its grafting direction. We have incorporated
the living character and the flexibility of the filament explicitly into a statistical mechanics approach based on the
reactive grand canonical ensemble, dealing with the model of a discrete WLC hitting a hard wall. This formalism
has been illustrated by the F-actin/free G-actin reacting mixture restricted to a single kind of actin monomer-ATP
complexes.
The results and the new phenomenology which emerges from the present work can be summarized as follows
• We provide a statistical mechanics justification of the popular Hill's expression for the single filament stalling
force[6]. Rigorously, this expression corresponds to the average force exerted by the filament (defined as the
ratio of the mechanical work over a finite displacement and the displacement itself), as the wall moves reversibly,
under mechanical and chemical equilibrium, over a distance corresponding to one monomer size d. The Hill's
expression is found to be strictly valid only for the rigid filament case and we derive explicitly the correction
terms for the semi-flexible case. The correction is positive and L-dependent for the model of a hard wall hit by
a discrete WLC (the force is larger than for the rigid case) but these flexibility effects are only of the order of
the percent when the experimental value of the actin persistence length is used. So we conclude that the L-
independent Hill's expression remains a very good approximation for the stalling force of semi-flexible filaments
like actin. It should be stressed however that the force exerted by the filament on a fixed wall in the reactive
grand canonical ensemble, that is the force which was integrated to get the work over a finite displacement of
size d, shows large fluctuations around the mean. These fluctuations which decrease progressively in amplitude
as L increases, find their origin in commensurability effects related to the degree of matching of the filament
contour length, necessarily an integer number of monomer sizes d, with respect to the gap width L. These effects
become less pronounced at large L as the amplitude of tip transverse fluctuations due to bending become more
important.
20
• Like for the rigid case[10], the equilibrium size distribution of the living flexible filament whose net polymerization
is stopped in supercritical conditions by a wall at position L, starts as a growing exponential, as long as the
filament size is too short for its set of fluctuating configurations to interact directly with the hard wall. Filament
configurations larger than the slab gap have zero probability for the 1D rigid model, while for flexible filaments
the size distribution generally presents a fast decay which involves some finite but rapidly decreasing probability
to get filament contour lengths larger than L. This rapid decay results from a filament bending work penalty
which systematically exceeds the gain in chemical free energy, as a result of polymerization steps beyond the
largest size z(L) of non touching filaments. Given the mentioned large oscillations in equilibrium properties as
the gap width L is varied over sub monomer length scales, the filament size distribution properties are better
discussed in terms of their d-averaged (average over a d window around the wall position L). The knowledge
of these distributions (function of slab gap L) allows to adopt a quantitative definition for the limit of the
non-escaping regime. In particular we require that the probability of a planar filament configuration of length
πL/2, the minimum length to laterally escaping, be three orders of magnitude smaller than the probability
of having filaments of length just below L. Our work provides the opportunity to establish more precisely the
characteristics of the crossover towards the escaping regime, a point of high relevance in in-vitro experiments[4, 5].
2
4 kBT 'p
2
4 kBT 'p
c ≈ π
L2
• At stalling, in the non escaping regime, the quasi L independence of the d-average force is produced by the
fraction of filament configurations hitting the wall. The cumulative probability hx0i(L, ρ1) of the size distribution
involving hitting filaments has been shown to increase like hx0i ∝ ln ρ1 L2/'p. This observation is compatible
with the buckled filament state of the large majority of hitting filaments of the ensemble, each of them exerting
adiabatically the classical plateau force expression fb = π
L2 (here, adiabatic refers to the
assumption that the life time of a given filament size is long with respect to the microscopic relaxation time of a
fixed contour length filament). In this way the product hx0ifb is compatible with the L independent stalling force
expression of Hill, which allows us to pinpoint a major distinction between rigid and flexible living filaments,
a distinction established here for the case of a single filament at equilibrium but which will be relevant for
multi-filament bundles at and outside equilibrium. For finite 'p, the L → 0 limit of very short semi-flexible
filaments leads to a contact probability hx0i → 0 and a buckling force fb → ∞, just like in the case of rigid
filaments ('p = ∞) at arbitrary L . Hitting the obstacle takes the form of instantaneous kicks both for single
filaments and multi-filament bundles. For flexible filaments of given 'p, the fraction of hitting filaments grows
quadratically with L while the force of each (buckled) filament decreases quadratically with L. If we consider a
dynamical trajectory of a single living flexible filament at equilibrium against a wall at distance L, the fraction
of time the filament is in contact with the wall is finite together with the associated (buckling) exerted force.
For bundles of Nf > 1 filaments at equilibrium, supposed to act independently, the force is produced by the
permanent recruitment of a subset of hx0iNf filaments pressing each with the buckling force fb(L), the subset
of hitting filaments permuting continuously among the Nf equivalent filaments as the result of continuous
(de)polymerization steps. Finally, as L approachesp'pd/ ln ρ1 under stalling conditions, which coincides with
the upper limit of the non-escaping regime, the fraction hx0i should approach unity as the polymerizing force
exerted by a filament cannot exceed fb. A more quantitative analysis of the limit is provided by eq. (69) which
shows that the probability to get escaping filaments starts to be non negligible when hx0i approaches 0.5.
A conjecture is possible when extending the criteria to observe the pushing catastrophe to the stationary situation
of a wall moving at constant velocity, pushed by the polymerizing bundle of Nf ≥ 1 filaments and subject to a load
FL = γFstal smaller than the stalling value (γ < 1) [17, 18]. To keep a constant velocity of the wall, the bundle must
exert a force equal to the load FL (or sightly larger if solvent friction is considered), therefore the number of filaments
N0 = hx0iNf needed to press on the wall should go as N0(L) ≈ γFstal/fb(L) ∼ L2 in order to compensate for the
variation of the single filament force fb ∼ 1/L2. If we assume that the non-escaping limit in stationary conditions
(v > 0) would correspond to the recruitment of all Nf filaments (or a permanent contact with the wall for the single
filament case Nf = 1), the maximum gap tolerated should be at least a factor γ−1/2 larger than the limiting value
for the non-escaping regime at stalling.
It is illuminating to compare the above considerations with two reported experimental measurements of the actin
polymerizing force. The experiments of Footer et al.
[5] use an optical trap set up to measure the stalling force
of a few actin filaments in supercritical conditions. In particular, they report in Figure 4b data corresponding to a
polymerizing force of F ≈ 1pN at reduced concentration ρ1 = 1.7 which, as they observe, corresponds to the stalling
force of a single actin filament. The average filament length is ≈ 180nm for the chosen optical trap. Eq.(76) applied
to this case would imply a contact time fraction of hx0i ≈ 0.25 with a force intensity of fb ≈ 4pN, the probed
filament length being indeed lower than the limit Lmax = 240nm predicted by Eq.(69). This experiment was in fact
dealing with a bundle of Nf = 8 filaments but surprisingly enough the stalling force of a single filament was effectively
recorded. The issue here is still under debate but, according to our present work, to detect a force eight times larger
21
at the same reduced concentration in free monomers, a trap force constant 5 − 10 larger would be required in order
to avoid laterally escaping filaments.
The Demoulin et al. experiment [4] probes the force-velocity relationship for a set of actin bundles[4], implying a
total of Nf ≈ 130 filaments at ρ1 ≈ 3, pressing together against a bead. While the stalling force in this case is around
200pN, the bead is subject to load forces ranging from a few pN (largest velocity probed) up to 100pN covering a
range 0.02 < γ < 0.5 for the load over stalling forces ratio. Looking at figure 2 in ref.[4], if we take a typical length of
200nm for the actin filaments beyond their lateral connection by fascin bridges, the number of filaments at contact
able to press on the obstacle bead with a buckling force of ≈ 4pN should lie between 1 and 25 over the explored force
range. Further, considering the results for the longest filaments (≈ 400nm) at FL = 3.9pN, our criterium above for
stationary non-escaping conditions is still justified since Lmax(ρ1) = 65d = 175nm at ρ1 = 3, and γ−1/2 ≈ 7 in these
conditions leading to a maximum length of the stationary non-escaping regime of Lmax/γ1/2 = 1225nm, still larger
than the probed bundle length.
Further consequences of filament semi-flexible character on actin bundles at and outside equilibrium will be discussed
in future publications[26, 28].
VII. ACKNOWLEDGEMENTS
We thank M. Baus, M. J. Footer and B. Mognetti for useful discussions and G. Destree and P. Pirotte for technical
help. This work has been supported by the Italian Institute of Technology (IIT) under the SEED project Grant 259
SIMBEDD and by the Italian Ministery of Research under project PRIN2012 -- 2012NNRKAF.
Appendix A: Some considerations on (de)polymerization rates.
of f are usually associated to the reactions described by Eqs.(2).
Phenomenological kinetic rate constants k
In terms of such kinetic constants the equilibrium micro-reversibility conditions, i.e. the equality between the number
of polymerizations of filaments of size (i − 1) to size i and the number of de-polymerizations of filaments of size i to
size (i − 1) per unit of time, are written as
and ki
on
(i−1)
ρ1P(i − 1) = ki
which implies, according to Eq.(41) and thermodynamics, the link
k(i−1)
on
of f P(i)
Eq.(A1) is often written equivalently as
Ki = k(i−1)
on
/ki
of f
U(i−1)P(i − 1) = WiP(i)
(A1)
(A2)
(A3)
where U(i−1) ≡ k
(i−1)
on
ρ1 and Wi ≡ ki
of f are (de-)polymerisation rates, respectively. Using again Eq.(41), their ratio is
U(i−1)
Wi
= P(i)
P(i − 1) = ρ1
α(i, L)
α(i − 1, L)
(A4)
The (de)polymerization rates for filament ends in bulk, denoted by U0 and W0 with U0/W0 = ρ1, are valid for our
grafted filaments, as long as they do not interact with the wall. When ρ1 = 1 (that is the free monomer density is
critical ρ1 = ρ1c = 1/K0), one has U0 = W0 so that the rate of polymerization and rate of depolymerization are equal
and the distribution should be uniform in the short filaments region (α = 1). We are interested to the supercritical
regime ρ1 > 1 and U0 > W0 for which the distribution in the same short filament region is a growing exponential.
For (de)polymerization reactions implying filaments hitting the wall, the rates satisfying Eq.(A4) are often chosen in
applications assuming that the rates of depolymerisation are not affected by the presence of the wall [7, 10, 18, 27],
namely
Wi = W0
Ui−1 = α(i, L)
α(i − 1, L) U0
(A5)
(A6)
22
[1] J. Howard, Mechanics of Motor Proteins and the Cytoskeleton (publisher Sinauer, address Sunderland, MA, year 2001).
[2] K. Sneppen and G. Zocchi, Physics in Molecular Biology (publisher Cambridge University Press, address Cambridge, UK,
year 2005).
[3] M. Dogterom and B. Yurke, Science, 278, (1997), 856.
[4] D. Démoulin, M-F. Carlier, J. Bibette, and J. Baudry,Proc.Natl.Acad.Sci. 111 (2014), 17845.
[5] M. J. Footer, J.W.J. Kerssemakers, J.A. Theriot and M. Dogterom, Proc. Natl. Acad. Sci. USA, 104 (2007), 2181.
[6] T.L. Hill,Proc. Natl. Acad. Sci. USA, 78 (1981), 5613.
[7] C.S. Peskin, G.M. Oster and G.S. Odell, Biophys. J, 65, (1993), 316.
[8] A. Mogilner and G. Oster, Eur. Biophys J., 28, (1999), 235.
[9] G. Sander van Doorn, C. Tanase, B.M. Mulder and M. Dogterom, Eur. Biophys., 29, (2000), 2.
[10] K. Tsekouras, D. Lacoste, K. Mallick, and J.-F. Joanny, New J. Phys., 13, (2011), 103032.
[11] J. Krawczyck and J. Kierfeld EurphysicsLetters, 93, (2011), 28006.
[12] A.E. Carlsson, Biophys. J., 81, (2001), 1907.
[13] D. Das, D. Das and R. Padinhateeri, New J. of Physics, 16 (2014), 063032.
[14] X. Li and A.B. Kolomeisky, J. Phys. Chem. B, 119, 4653-4661 (2015).
[15] D.R. Kovar and T.D. Pollard, Proc. Natl. Acad. Sci. USA, 101 (2004), 14725.
[16] J. Berrot, A. Michelot, L. Blanchoin, D. Kovar and J. Martiel, Biophys. J. 92, 2546 (2007).
[17] N.J. Burroughs and D. Marenduzzo, J. Chem. Phys. 123 (2005) 174908.
[18] N.J. Burroughs and D. Marenduzzo, J. Phys.: Condens. Matter 18 (2006) S357
[19] A. Mogilner and G. Oster, Biophys J., 71, (1996), 3030.
[20] T.E. Schaus and G.G. Borisy, Biophys J., 95, (2008), 1393.
[21] A. Mogilner J. Math. Biol., 58, (2009), 105.
[22] J.-P. Ryckaert and S. Ramachandran, Mol. Phys., 111, (2013), 3515.
[23] T.L. Hill, An Introduction to Statistical Thermodynamics, ( Dover, New York, 1986).
[24] A. Mogilner and B. Rubinstein, Biophys. J, 89, (2005), 782.
[25] A. Gholami, J. Wilhelm and E. Frey, Phys.Rev.E, 74, (2006), 041803.
[26] A. Perilli, C. Pierleoni, G. Ciccotti and J.P. Ryckaert, "Polymerization Force of a Bundle of Living Actin Filaments in an
Optical Trap Set-up" in preparation, (2015).
[27] S. Ramachandran and J.-P. Ryckaert, J. Chem. Phys. 139, (2013), 064902.
[28] T. Hunt, S. Mogurampelly, C. Pierleoni, G. Ciccotti and J.P. Ryckaert, "Dynamical relaxation of a bundle of living actin
filaments in an optical trap apparatus", in preparation (2015).
|
1612.07184 | 3 | 1612 | 2018-01-16T06:32:14 | Performance limits and trade-offs in entropy-driven biochemical computers | [
"physics.bio-ph",
"cond-mat.stat-mech",
"q-bio.MN"
] | The properties and fundamental limits of chemical computers have recently attracted significant interest as a model of computation, an unifying principle of cellular organisation and in the context of bio-engineering. As of yet, research in this topic is based on case-studies. There exists no generally accepted criterion to distinguish between chemical processes that compute and those that do not. Here, the concept of entropy driven computer (EDC) is proposed as a general model of chemical computation. It is found that entropy driven computation is subject to a trade-off between accuracy and entropy production, but unlike many biological systems, there are no trade-offs involving time. The latter only arise when it is taken into account that the observation of the state of the EDC is not energy neutral, but comes at a cost. The significance of this conclusion in relation to biological systems is discussed. Three examples of biological computers, including an implementation of a neural network as an EDC are given. | physics.bio-ph | physics |
Performance limits and trade-offs in entropy-driven biochemical computers
Dominique Chu
School of Computing
University of Kent
CT2 7NF, Canterbury
Tel.: +123-45-678910
Fax: +123-45-678910
Abstract
It is now widely accepted that biochemical reaction networks can perform computations. Examples are kinetic
proof reading, gene regulation, or signalling networks. For many of these systems it was found that their
computational performance is limited by a trade-off between the metabolic cost, the speed and the accuracy
of the computation. In order to gain insight into the origins of these trade-offs, we consider entropy-driven
computers as a model of biochemical computation. Using tools from stochastic thermodynamics, we show
that entropy-driven computation is subject to a trade-off between accuracy and metabolic cost, but does
not involve time-trade-offs. Time trade-offs appear when it is taken into account that the result of the
computation needs to be measured in order to be known. We argue that this measurement process, although
usually ignored, is a major contributor to the cost of biochemical computation.
Keywords: biological computing, information thermodynamics, cost of computation, linear noise
approximation
1. Introduction
Computing architectures based on biochemistry, rather than semi-conductor technologies, are attracting
increasing interest as alternative models of computation [1]. Biochemistry can be used to engineer novel
types of computers based on biological components. Examples include, DNA based computers [2, 3], robots
controlled by slimemolds [4], or logic gates implemented in living cells [5, 6, 7, 8]. Beside this technolog-
ical importance of biochemical computers, there is now also an increasing appreciation that information
processing may be an important fitness contributing function for natural organisms [9, 10]. There are a
number of biosystems that have been studied as in vivo special purpose computations. For example, kinetic
proofreading [11, 12] greatly enhances the copying fidelity during translation and is often interpreted as an
in vivo computation. A classical example of biochemical computation is bacterial sensing [13, 14, 15, 16],
whereby cells measure molecular concentrations in their environment and modify internal pathways and
gene expression levels in response. Chemotaxis [17], for instance, depends on organisms sensing a molecular
concentration gradient by computing the difference between several measurements, either in time or across
the cell volume. Most recently even bacterial growth dynamics has been interpreted as a computational
process [18, 19].
Detailed case studies of biological computers often find performance limits to biochemical computations.
For a simple gene-switch, Zabet and coworker found a trade-off between the cost, the accuracy and the
speed of the computation [20, 21]. Similar trade-offs were established for other biological systems, including
chemotaxis [22], regulation of nutrient uptake [18], and translation [23]; lower limits on the cost of sensing
(not involving trade-offs) have also been found recently [24, 14].
Intuitively such trade-offs are to be expected. Bio-chemical networks are stochastic systems and as such
subject to noise. Overcoming this noise requires energy input and time. Energy-time-accuracy trade-offs are
also implied by the classical results on the physics of computation [25, 26, 27]. While there does not seem to
Email address: [email protected] (Dominique Chu)
Preprint submitted to Journal of Theoretical Biology
May 17, 2018
be a lower limit for the energy used during a computation, Bennett [26] pointed out that in the zero energy
limit the speed of computation goes to zero. Computations that complete within a finite time, therefore
require finite energy resources.
The question is now whether one can go beyond both the individual case studies of biochemical computers
and the intuitive arguments and establish a model which provides insights into the origins of the performance
limits to biochemical computations. The task is a difficult one. For one, there is a wide variety of approaches
to biochemical computation (only some of which are mentioned above). At the same time, there is no general
definition of biological computation, i.e. it is not clear how to distinguish a reaction network that computes
from one that does not.
For the purpose of this article, we will take a pragmatic approach with respect to the latter question and
simply identify (in section 2.1) computation with out-of-equilibrium biochemical processes. According to
this, every biochemical process that is not in equilibrium performs a computation. As far as the wide variety
of biochemical computations are concerned, we will abstract away from specific models and define the concept
of entropy driven computers (EDC) in section 2.2. This will capture many properties of in vivo computers
as they appear in biological systems. EDCs are in many aspects different from real biological networks, but
we will argue that they share important characteristics with a wide range of biochemical computers. It is
perhaps best to think of EDCs as test-tube biochemistry, as it is frequently used in biological research to
study reaction networks in vitro.
We model EDCs as continuous time Markov chain models of biochemical systems. We assume that each
model is initialised in some state and then left to relax to equilibrium. We will then interpret this relaxation
process as a computation. Throughout this article we will assume that the EDC is of mesoscopic scale.
By this we mean that it is still affected by stochastic fluctuations, but that it is also within the range of
validity of the linear noise approximation [28]. Simply put, this assumption states that the stochastic system
behaves like the deterministic equivalent, plus some noise. The linear noise approximation is a very good
approximation for mesoscopic systems and holds true for a wide range of biochemical systems, and hence
for a wide range of biological "computers" such as gene regulatory networks, protein-protein interactions or
intra and inter-cellular signalling systems, although clearly there are systems that will not be captured by
this assumption.
For the purpose of the present contribution, we will identify the cost in energy of a computation with
the entropy produced during the computation. While this does not quantify the actual metabolic cost
of this computation, it is directly related to it. We will first show that the linear noise approximation
implies that the entropy production scales linearly with the system size, while the time-scale to approach
equilibrium (which we interpret as the computing time) remains invariant. This means that there is a trade-
off between the cost of the computation and its accuracy, but there is no trade-off involving the speed of the
computation. Contrary to previous work (or apparently so), this suggests that speed-energy trade-offs are
not a fundamental property of biochemical computation per se
A trade-off involving time emerges only when it is taken into account that the result of the computation
must be measured in order for the computation to have any impact in the world. Any measurement of
the outcome of the computation in turn requires a measurement device. This device needs to be brought
into contact with the computer to determine its state. Device and computer then form a joint system,
which initially will be out of equilibrium but relaxes to an equilibrium. This relaxation constitutes the
measurement process. Formally a measurement is thus also an entropy driven computation. As we will show
below, restoring the computer to its original state, while at the same time leaving the measurement device
in a state that indicates the result of the computation, requires both energy input and time. It also leads
to a trade-off between the energy used and the speed with which the restoration can be completed with a
given confidence. A second trade-off involving time arises from the stochastic nature of the computer. A
single measurement only indicates the correct result with a certain probability. Repeated measurements are
necessary in order to sample the state of the computer reliably, thus leading to a trade-off between accuracy
and time.
2
2. Results
2.1. Computation by biochemical systems
The current modus operandi in the field of biochemical computing is to identify a biological system (such
as sensing or proof-reading) as a computation when it implements a function that is naturally interpreted
as a computation. This approach enables deep insights into specific examples, but is likely to miss most
instantiations of biochemical computation. It would be much more useful to have a concept of biochemical
computation that is independent of its function, just as in computer science computation is defined with
respect to a number of specific mathematical models, not by reference to what is computed.
The best known model of computation is the Turing machine. This is a mathematical construct consisting
of a "reading head" that is reading and writing a tape, while changing its internal states in the process, until
it reaches a "halting state," at which point the computation stops. It is believed that for every computable
function there is a corresponding Turing machine that computes it. Based on this, one could be tempted to
define a biochemical process as a computation if there is a Turing machine that simulates this process. This
does not work however: The natural equivalent of a halting state in biochemical systems is the equilibrium
state, i.e. the state of the biochemical system where reactions are in detailed balance. Unlike the halting
state of a Turing machine, the equilibrium state is of a statistical nature. This means that on average there
are no net-fluxes across the network of reactions [29, 30]. This does not mean, however, that reactions
stop. Even in equilibrium there is an ongoing chemical activity. Crucially though, the sequence of reaction
events is symmetric in time [28], such that an observer would not be able to tell apart an actual sequence
of reactions from a (hypothetical) reversed sequence. Equilibrium is not time-directed. Computation, on
the other hand, is necessarily time directed, mapping a particular input to a particular output. Equilibrium
systems are therefore not able to compute. Sample paths of equilibrium biochemical systems can still be
simulated and are thus computable by Turing machines, whether or not the system is in equilibrium. This
demonstrates that not all processes that can be simulated by Turing machines are also themselves processing
information.
For the purpose of this paper, we will adopt the simplest working hypothesis and postulate that the
equilibrium state is the only halting state of biochemical computers. This implies that every biochemical
system that is not in equilibrium is in the process of computing. By adopting this definition, we also accept
that most biochemical computers will not do any useful calculations, just as almost all Turing machines do
not compute anything of interest.
2.2. Entropy driven computation
In this section we define an EDC as a closed, stochastic, biochemical system, denoted by a fraktur S,
S. The system does not exchange particles with the environment. We conceptualise S as consisting of a
(typically very large) number of discrete microstates s0, s1, . . . , sm (see SI section 1 for more details). An EDC
is initialised in a macrostate M S
0 characterised by a specified abundance for each of its constituent biochemical
species at time t = 0; see SI section 1 for a detailed explanation of what we mean by "macrostate." After
a transient period, the biochemical system approaches an equilibrium state M S
∞ characterised by detailed
balance. The approach to equilibrium is the computation. Strictly speaking, the transition to equilibrium
takes an infinite amount of time. In practice, EDCs will be very close to equilibrium after a finite, possibly
very short, time. We will model EDCs here as continuous time Markov chains. Then the time scale for the
approach to equilibrium depends on the kinetic parameters appearing in the master-equation that defines
S. We will take this time scale as the speed of the computation.
In contrast to EDCs, in vivo computations, i.e. cells, never approach equilibrium, but rather operate
around non-equilibrium steady states. We acknowledge this, but still argue that for the present purpose
equilibrium models are more convenient. They are also more revealing of underlying principles, for the
following reasons: Firstly, steady state processes have an additional "maintenance" entropy production.
This may be substantial, but is not related to the cost of the computation per se. Mathematically, it is
possible to separate the contribution from the relaxation from the contribution to the entropy production
that arises from the maintenance of the non-equilibrium state. By focussing on the simpler equilibrium case,
we circumvent this mathematical complication that does not add anything to the present purpose. Secondly,
as will become clear below, considering equilibrium processes naturally forces a conceptual separation between
the computation per se and the reading of the result. Both are separate processes that limit computational
3
performance in qualitatively different ways. Finally, the equilibrium state is conceptually reminiscent of the
halting state in Turing machines, which in turn is of fundamental importance for the theory of computation.
Indeed, there is an interesting analogy in the relationship between on the one hand a non-equilibrium steady
state, sustained by an organism and test-tube biochemical systems approaching equilibrium and on the
other hand an operating system - which is not supposed to halt - and an algorithm - which must halt.
Models of entropy driven computation therefore naturally link biology, physics and theoretical computer
science. With this being said, the basic conclusions that we will reach depend primarily on the linear noise
approximation, which in turn depends on system size, not on the distance from equilibrium.
2.3. Cost of the EDC proper
2.3.1. The model
We assume that the stochastic system S is defined by the master equation, which is a differential equation
for the probability to observe a particular combination of molecular abundances n at time t [31].
P (n, t) =Xi
(wi(n − σi)P (n − σi, t) − wi(n)P (n, t))
(1)
where wi(n) := kihi(n) is the total rate of reaction i and h is the multiplier indicating how many combinations
of molecules can realise this reaction; n and σi are the particle vector and the stoichiometric vector of reaction
i. Altogether, the right hand side of the equation contains two terms. The first one expresses the possibility
that the current state characterised by a particle composition n was obtained from reaction i and the state
before the reaction happened was n − σi. The second term formulates the possibility that before the last
reaction the state of the system was n, but reaction i took it away from this state (to a new state n + σi).
Given such a stochastic model, we can scale the total number of particles in the initial conditions by
a factor c and the reaction rate constants of bimolecular reactions by 1/c. This amounts to scaling the
volume V of the system while keeping the concentration fixed.
In this way we construct an equivalence
In general, the members in this class will behave differently. Most importantly,
class S of systems S.
they show different amounts of fluctuations and approach equilibrium at different speeds. For mesoscopic
volumes, however, the behaviour of the master equation is increasingly well approximated by the first order
linear noise approximation [28] whereby the mean behaviour of the system is invariant to scaling. Scaling the
system size in the regime of the linear noise approximation only affects the noise around the mean behaviour,
which scales with V −1/2. Importantly, scaling the volume does not affect the time-scale to reach equilibrium,
which is determined by the mean behaviour. In the limiting case of an infinite volume the noise goes to zero
and the system is described by a differential equation whose trajectory corresponds (for monostable systems)
to the mean of the linear noise approximation.
2.3.2. Entropy production
In this sub-section, we will show that the operation of an EDC is subject to a trade-off between the noise
and the cost. An EDC starts with the stochastic system S initialised in some macrostate M0. The system
then relaxes into an equilibrium state M∞, performing a computation in the process. We will identify the cost
of the computation with the amount of entropy generated during the relaxation. There are two components
to the entropy. Firstly, the exported entropy or heat dissipated to the environment which accrues whenever
the system makes the transition from state si to state sj. This is given by the ratio of the forward rate of
the transition wji and the corresponding backwards rate wji [32].
∆SQ = kB ln
wji
wij
(2)
Using this definition the heat dissipated to the environment is positive, whereas heat extracted from the
reservoir is negative. No heat is generated by a system that transitions from a state si to a state sj and then
back again. More generally the heat generated by a system that transitions from an initial state s0 to some
final state sj via a number of intermediate states is independent of the particular path, and depends only
on the initial and final state. Taking into account that the heat generated in each transition is additive, and
applying the detailed balance condition that relates the steady state probabilities πi, πj of states si and sj
of a stochastic system with the transition probabilities, i.e.
πiwij = πj wji
(detailed balance condition)
4
y
p
o
r
t
n
e
m
e
t
s
y
S
14000
12000
10000
8000
6000
4000
2000
0
0
500 1000 1500 2000 2500 3000 3500 4000 4500
7
6
5
4
3
2
1
y
p
o
r
t
n
e
n
o
n
n
a
h
S
0
0
50
100
150
200
250
300
Number of particles in system
Number of particles in system
Figure 1: Numerical calculations of the exported entropy (left) and the Shannon entropy (right) for the system A + B ⇋
C + D ⇋ E. The system size is the number of A and B at time t = 0. The concentration was kept constant as the particle size
was increased. The entropy is given in units where kB = 1. The forward rate constant was set to 0.3 and the backwards rate
constants from E and C + D were 0.05 and 0.1 respectively. Prism [33] was used to solve the numerics; for the model file see
supplementary information.
we obtain:
∆SQ(s0 → sj) = kB ln
πj
π0
(3)
In addition to the exported entropy, the computation also produces Shannon entropy. This Shannon
entropy of a microstate si can be written as ∆Sint = −kB ln pi, where pi is the probability to observe the
system S in microstate si. On average this gives a contribution hSinti = −kBPi pi ln(pi). The Shannon
entropy produced during the computation is then given by hSint(∞)i − hSint(0)i.
For realistically sized systems it is usually not possible to solve the master eq. 1 exactly, even in steady
state. Analytic expressions for the entropy production are therefore difficult to obtain. However, general
scaling arguments can be given in the limit of the linear noise approximation. In this regime, one expects the
total entropy during the approach to equilibrium to scale linearly with the system size. Consider, for example,
a biochemical system with a mono-modal steady state distribution and discrete, steady state probabilities
{π0, π1, . . . , πk, . . . , πm}, where m + 1 ≫ 1 is the total number of states of the system. Assume now that the
initial state is s0, then the heat dissipation is given by averaging eq. 3 over all possible end states:
hSQi = kB
m
Xi=0
πi ln(cid:18) πi
π0(cid:19) = kB
m
Xi=0
πi ln (πi) − kB ln π0.
(4)
The first term on the right hand side equals the Shannon entropy (up to the sign). The total entropy
produced hSi = hSQi + hSinti is therefore given by
hSi = −kB ln π0.
(5)
In order to understand the scaling of the entropy production, we approximate the discrete steady state
distribution by a continuous Gaussian distribution with mean µ and variance σ2. According to the linear
noise approximation, the variance and the mean scale linearly with V , such that we obtain:
hSi = −kB ln π0 ∼
x2
0 − 2x0µ + µ2
2σ2
∼ V.
(6)
We stress that this scaling argument is only valid for mesoscopic systems when the linear noise approx-
imation is good. In particular, for very small systems there can be a non-linear relationship between the
system size and the heat dissipation. In those cases the master equation needs to be solved to determine
the entropy production. This is usually problematic. See SI section 5 for an explicit calculation of the heat
dissipation for a minimal example and figure 1 for a graphical representation.
In summary: The ability to determine the result of the EDC is limited by the noise of the system in
2 in the linear noise approximation. The cost of the computation scales
equilibrium, which scales like V −
1
5
linearly with the size of the system. Hence there is a trade-off between entropy production (and thus energy
cost) and the accuracy of the EDC. No time-trade-offs arise here. In the linear noise approximation, the
time evolution of the mean does not depend on the system size.
2.4. Trade-offs arising from the measurement
The equilibration of the EDC is only one part of the computation.
In order, for the outcome of the
computation to be known, a measurement must be performed to determine the macrostate M S
∞ and the
result needs to be recorded. This comes at a cost [34, 35]. We assume that the measurement is done by
means of the measurement device ¯S, which is itself an EDC; see SI section 1 for details on the concept of
measurement.
We will first discuss how ¯S can be used for binary measurements on observables of S, i.e. in order to
determine whether a particular biochemical species L of S is above or below a threshold abundance. We
will assume that ¯S is a bistable system, i.e. in equilibrium, it can be in one of two (transient) macrostates,
indicating the result of the computation.
Measurement is only possible if S and ¯S are temporarily brought into contact. We will assume that
the contact remains limited to specific interfaces. The biochemical systems S and ¯S should not be allowed
to mix, as it would be difficult to separate the systems again after the measurement. Instead, we consider
here a protocol whereby S and ¯S remain separated by a wall. We model the interface between system and
measuring device as a single "trans-membrane" receptor placed in the surface of ¯S. The external part of the
sensor contains a number of binding sites for molecules of type L of S. The inside of the receptor can interact
with ¯S, as will be described in the next section. We will assume a Monod-Wyman-Changeux (MWC) [36]
receptor (but other models are possible too). When its external sites are bound then the receptor is heavily
biased towards the "active" state. Otherwise it is heavily biased towards being deactivated. When there are
several binding sites, then the receptor can support ultra-sensitivity. This means that for low abundances of
L the binding sites are almost never bound; above a threshold abundance the sites are almost always bound.
In this way, MWC receptors can be used for binary measurements of external concentrations.
Upon contact, the system and the measurement device form a new system S ¯S. The joint system will
therefore evolve from an initial macrostate M S ¯S
t+Tmeas,
where Tmeas is the time required for the measurement. Upon separation of S and ¯S the new macrostate of
the measurement device should be a (transient) record of the state of S. Once the measurement is completed
it is necessary to restore S as an independent system, i.e. to separate S and ¯S.
at time t of the measurement, to a macrostate M S ¯S
t
Conceptually, there are thus two aspects to the measurement process that are relevant for the question
we consider here. Firstly, the process of measurement itself and secondly, the separation of the measurement
device from the system S. We find that the binary measurement - while it cannot be done for free - does
not imply any trade-offs. In contrast, the separation step leads to a trade-off between energy, cost and time.
2.4.1. The measurement proper
We choose ¯S to be bistable, i.e. there are two macrostates M m and M g characterised by a high amount
of m or g respectively. Bistability can occur in equilibrium biochemical systems. Stochastic bistable systems
will switch spontaneously, but possibly rarely, between the two macrostates. The expected time between
switching events will depend on the system size and the "well depth." The latter is essentially determined
by the kinetic parameters of the system and indicates the difficulty of escaping a metastable state. Larger
systems are more stable but even relatively small systems can be sufficient to virtually guarantee stability
over any practically relevant time-scale.
In order to switch a bistable system at a particular time, it is necessary to take it out of equilibrium by
introducing a net probability flux to the desired state of the system. This implies breaking detailed balance
and hence dissipates heat. A controlled switch of the bistable system can therefore only be achieved at a
certain expense of work. Formally, it would be sufficient to connect a source of free energy (a "chemical
battery") to ¯S during the switch only and disconnect it afterwards. In practice, due to the nature of the
biochemical systems, it is difficult to remove such a battery without extra work input. It is therefore much
better to operate the device permanently from a battery away from equilibrium.
6
l
a
n
g
i
S
f
o
y
t
i
l
i
b
a
t
S
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
N=4
N=8
N=16
0
2
4
6
8
10
12
Dissipation
Figure 2: The trade-off between dissipation and stability. For networks of different size N , defined as the total number of
particles in the system, the energy dissipation after 600 time units was recorded against the stability of the memory. Small
systems are performing better than large systems. Stability is defined as the number of m molecules divided by the number of m
and g molecules combined after 600 time units. Each point in the graph represents the average over 5000 individual simulations.
For each value of N we sampled initial amounts of ATP from 5 to 300 in steps of 5. The dissipation is the difference between
the number of ATP at time 0 and at time 600 averaged over 5000 simulations.
2.4.2. A possible design of a measurement device
We will now construct a measurement device ¯S consisting of the molecular species {c, d, e, g, m}. It is
bistable in m and g, i.e. at any one time only one of those two species is present in a high concentration.
We will identify the two metastable states by the macrostate M g and M m respectively. These states will
indicate whether the concentration of species L of S is above or below some threshold. There are many
ways to achieve bi-stability in biochemical systems. We choose a mechanisms based on two competing auto-
catalytic reactions: Molecules of type c are converted into d. This conversion is catalysed by g. Molecules
of type d are reversibly converted into g. Altogether, g is thus autocatalytic. To achieve bi-stability, we
add symmetrically a second autocatalytic circuit consisting of e which is produced from c catalysed by m.
Molecules of e are also reversibly converted into m. The reactions can be summarised as follows:
g
c
−−⇀↽−− d
m−−⇀↽−− e
c
d + d −−⇀↽−− g
e + e −−⇀↽−− m
To create bi-stability, this system needs to be extended by an antagonistic interaction between g and m.
d + m −−⇀↽−− m + c
e + g −−⇀↽−− g + c
Finally, we also posit that g and m inter-convert reversibly, albeit at a very low rate. Forward and backwards
rates are equal.
g −−⇀↽−− m
If initialised with a sufficient number of c and one g and m each, then the system will, after a transient
period, evolve to a macrostate M g
∞ characterised by m > g. The probabilities of either
of these states will be equal if the parameters are symmetric between the autocatalytic pathways of m and
g.
∞ where g > m or M m
2.4.3. Switching
In order to switch the state of ¯S, we couple the chemistry of the device to the internal part of the MWC
receptor, which can be in an active state r∗ or an inactive state r. Here we assume that the active receptor
7
catalyses the inter-conversion of m and g, adding the following reactions to the measurement device:
r −−⇀↽−− r∗
m
r∗
−−⇀↽−− g
In the absence of a driving force, the active receptor accelerates the approach to m = g. Given the right
range of parameters it will thus temporarily disrupt bi-stability and bring ¯S into a mono-stable regime.
Upon removal of the catalysing reaction, i.e. when the internal receptor reverts to the inactive state, ¯S will
relax quickly to M g or M m, both with equal probability. No directed switch is possible.
A directed switch is only possible if the catalytic reaction is driven preferentially into one direction. One
possibility is to couple it to a biochemical battery, i.e. a reaction that is not in equilibrium. A modified
reaction scheme could then be the following:
m + ATP
r∗
−−⇀↽−− g + ADP
If there is a large excess of ATP over ADP then this would result in a net drive towards g whenever the
receptor is activated. Assuming the excess of g is sufficient, the activation of the receptor leads to a switch
of the macrostate to M g.
2.4.4. Measurement and recording
The measurement procedure is as follows: (i) Reset ¯S. (ii) Initiate the measurement proper on S by
bringing S and ¯S into contact mediated by the MWC receptor. (iii) Wait for a fixed amount of time Tmeas.
(iv) Separate S and ¯S.
The reset step is necessary so as to ensure that the measurement device is in a known state prior to
starting the measurement. Without this step, ¯S is in state ¯M m with probability 1/2. This is problematic,
because the measurement of S will only effect a switch in ¯S if the concentration of L is indeed above the
threshold. The reset can be implemented as a measurement of some species I of a reference volume that is
in a known state. The measurement should be mediated by an auxiliary receptor different from the one that
measures S. The active state of this auxiliary receptor should effect a switch to M g when the concentration
of I in the reference system is high (which should always be the case).
Once the reset is completed, the measurement of S can be performed. Following this step the macrostate
of ¯S is a record of the state of S. The metastable states of ¯S are only stable over a finite (although possibly
very long) time, because the system may transition spontaneously to a different state.
In principle it is possible to remove the battery from the bistable system post-measurement. In practice
this will be difficult because it entails extracting the ATP and ADP molecules from ¯S and requires bio-
chemical work. Therefore, it is better not to remove the battery. In this case, however, the macrostate of
¯S remains only stable for as long as the battery is sufficiently charged, i.e. as long as there is an excess of
ATP over ADP. Once the battery has run out ¯S will no longer be bistable (see SI fig. 1).
The biochemical battery is discharged by two processes. Firstly, during switching the receptor is active
with high probability and the battery will drive the conversion from g to m while using ATP. When the
external binding sites of the MWC receptor are occupied the receptor will be active most of the time, resulting
in a high rate of discharge. Secondly, there is spontaneous activation of the receptor even if no ligands L are
binding to the outside receptor. The rate of spontaneous activation of the receptor determines the base-rate
of ATP usage/battery discharge.
Note that the ability of ¯S to measure and record the state of S is not subject to a trade-off between
system size and accuracy. In the particular model we present here, small measurement devices ¯S perform
better than those with a large number of molecules, while also dissipating substantially less energy (see SI
fig. 1). Similarly, there are no trade-offs between the speed of the measurement and its accuracy. The speed
is related to the scale of the reaction rates inside of ¯S which does not affect the heat dissipation of the
system.
There are, however, other sources of trade-offs that are closely connected to the measurement process.
The accuracy of the binary measurement corresponds to the ability of ¯S to indicate whether or not the
concentration of the system is below or above a threshold. The parameter that determines this accuracy is
the number of binding sites for the ligand L. The receptor number does not influence the cost of switching,
but it does increase the cost and time of separating S and ¯S, as will be discussed in the next section.
8
(i)
(ii)
S
V1
V2
¯S
S
V1
V2
¯S
PSfrag replacements
x0
x1
x2
x3
x0
x1
x2
x3
(iv)
(iii)
S
V1
V2
¯S
S
V1
V2
¯S
x0
x1
x2
x3
x0
x1
x2
x3
Figure 3: A schematic representation of the resetting procedure. The steps (i) to (iv) are illustrated from top left to bottom
right. In this particular case, the particle was successfully captured in the first step and is thus successively transported into
the volume V1.
2.4.5. The separation step
Once the measurement of the system has been completed the coupled system S ¯S needs to be separated,
while restoring S to a state that is statistically equivalent to the state before the measurement. This simply
reflects the condition that the measurement should not alter the system. In our specific case the restoration
of the system requires the separation of S and ¯S and that the ligand L is returned to the system. We will
find that this separation step gives rise to a trade-off involving time. To do this, we use an abstract model
of separation which, using ideas from thermodynamics, estimates minimal work requirements. While the
model is not biological realistic, we will nonetheless find that it leads to biologically relevant trade-offs in
terms of binding rates that also relate to known limitations of biochemical sensors.
We stipulate that contact between S and ¯S is mediated by a straight wall. The system S is fixed in
space, but ¯S can be moved to the right, which opens up a volume V0 between the two systems (see fig. 3).
We also assume a number of removable walls that can be inserted at arbitrary (but fixed) points to further
sub-partition V0. We now describe a protocol to restore S. This model involves 3 walls. A simpler - but
energetically less favourable - model is possible and described in the SI section 4.
The idea is to insert a wall (Wall 1) close to the membrane of ¯S at x0; see fig. 3. It separates S from ¯S.
The measurement device is then moved to the right, opening up the volume V0. Wall 2 is then inserted at
some distance to the right of wall 1 at position x1 so as to form the volume V1 between it and wall 1. The
separation procedure is as follows. (i) After waiting T time units insert wall 3 at x3, which is immediately
to the left of the membrane of ¯S. Wall 3 separates the receptor (and possibly ligands bound to it) from the
volume V0.(ii) Slide wall 3 to the point x2. Wall 3 now forms a volume V2 between itself and wall 2 and
a volume V0 − V1 − V2 between itself and the membrane of ¯S. (iii) Remove wall 2; this opens the volume
V1 + V2 between S and wall 3. (iv) Slide wall 3 from position x2 to position x1. At this point the cycle is
started again at (i) by inserting wall 2 at x3.
The minimal work required to separate the system is given by (see SI section 3 for the details of the
9
calculation):
Wtot = ln(cid:18) x−1
+ ln
κ + x−1(cid:19) +
(V0 − x−1 − κ) (x−1 + κ + V2)! − 1
V2(cid:0)x−1 + κ(cid:1)
(7)
Here 0 ≤ κ ≤ V0 − V2 is a constant expressing the size of V1; see SI section 3 for details. This shows that if
the concentration x ≫ 1 of particles in ¯S is high, then the work scales as the logarithm of the concentration.
More importantly though, within any particular equivalence class S of systems, where the concentration
does not change, the separation work is independent of the system size and not affected by scaling of the
volume V .
2.5. Trade-offs involving time
The separation step of the two systems gives rise to two different trade-offs involving time. Firstly, the
work estimates calculated above are valid for quasi-static processes. Finite time processes will always require
more work [37]. A second trade-off emerges from the waiting time for the ligand-receptor bond to equilibrate
during step (i) of the separation procedure. The time scale for equilibration is τ ∼ ku + kb/V0, where kb and
ku are the binding and binding/dissociation rates to/from the receptor. The equilibrium probability to find
the particle unbound is pu = V0/(V0 + η) where η = kb/ku. From this, we obtain the number of iterations
hnǫi of the resetting procedure necessary to ensure that the average number of bound ligands is ǫ:
hnǫi =&
ln(η) + ln V0 + η' ,
ln(cid:0) ǫ
l(cid:1)
(8)
where ⌈y⌉ denotes the smallest integer > y and l is the number of ligand binding sites at the external
receptor. Multiplying this by the time scale τ gives a time-scale indicator for the restoration of S (see fig.
4b). A trade-off between the measurement time and the work required arises here, because both quantities
depend on the volume V0. A larger volume makes it more likely for the unbinding receptor to be captured,
but equally increases the amount of work that needs to be done (see fig. 4a).
Further trade-offs arise when a concentration needs to be determined with an accuracy higher than
binary. To see this, assume that dmax is the maximal abundance of L in S. Given a set of perfect binary
measurement devices ¯Si one could use half-interval search to determine the abundance of the particles with
an accuracy of dmax/N by using ⌈log(N )⌉ sequential binary measurements. Each binary measurement comes
at a fixed cost and takes a finite amount of time. Altogether, this means that a higher accuracy can only be
achieved at the expense of a longer waiting time and a higher energy cost. Instead of performing half-interval
search, one could alternatively perform a number of concurrent measurements. In this case, there would
be no additional time penalty, but N measurements would be required in order to achieve an accuracy of
dmax/N , which is much worse than the logarithmic scaling of the half-interval search.
3. Discussion
Computational processes in biological systems typically show a trade-off between cost, accuracy and
speed. Intuitively such trade-offs are expected, but their precise origin remained unclear. Our model shows
that the origins of time trade-offs and trade-offs involving accuracy are quite distinct. The key-parameter
controlling the accuracy of the computation of an EDC is the system size, which also controls the cost. Yet,
in mesoscopic systems the speed of the computation is independent of the system size, but depends only on
the kinetic parameters. Trade-offs involving time arise during the process of measuring the outcome of the
computation.
One could take the point of view that the measurement process is not an integral part of the computation,
and that it is therefore unreasonable to charge this cost to the computation account.
Indeed, it is true
that there are many stochastic processes that happen in the world where the state of the system is never
determined. However, information processing is only useful to the cell when the result is somehow utilised
10
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
τ
h
n
ǫ
i
3
2.8
2.6
2.4
2.2
2
τ
1.8
1.6
1.4
1.2
1
Probability of success
τ
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
τ
y
t
i
l
i
b
a
b
o
r
P
i
ǫ
n
h
τ
14
12
10
8
6
4
2
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
1 receptor
2 receptors
0
1
2
3
4
5
6
7
8
4
6
8
Work
(a)
12
14
16
10
Work
(b)
Figure 4: (a) Trade-off between the time and the work spent on the restoration. Here we chose k = ku = 1 and V1 = V2 = 0.25.
The reaction volume V0 was varied from 0.5 to 200. The line labelled "Probability" denotes the probability that the particle
bound to the receptor has been "captured," i.e. pushed to V1. (b) Trade-off between the time spent on computation and the
work, for 1 and 2 receptors. Here we chose ǫ = 0.01, k/ku = 1, N = 200, VS = 10 and V2 = 1.
by the cell. For instance, during chemotaxis a bacterium determines the external concentration of some
molecular species and then alters the motion of its flagella. More generally, if any computation is to have
an effect on the world at all, then its result must be known. A computation that just happens without the
result ever being communicated is thus not a meaningful computation at all. Measurement, therefore, is
crucial to computation and cannot be separated.
Within our model, there are two parts to the measurement process. (i) The cost of recording the result of
the computation and (ii) the separation/restoration cost. As expected, the cost of recording the measurement
does not scale with the size of the measurement system. There are no fundamental reasons to assume that
large systems are necessary in order to store information. Indeed, every electronic computer uses microscopic
magnetic systems in order to store bits on disc drives. At a first glance more surprising is that the time-cost
stems from the time required to separate the system from the measurement device.
The model of separation we used is, of course, not biologically realistic. Nonetheless, it reflects a real
biochemical effect. The separation cost is directly related to the work required to break the bond between
the external protein and the receptor, which features prominently in various biological incarnations of what
is essentially the same effect. A well known example is the classical insight that the unbinding rate of the
ligand from the receptors at the cell surface fundamentally limits the ability of cells to measure molecular
concentrations in their environment. This is the ultimate origin of the Berg-Purcell limit [15, 38, 24]. Also
relevant in this context is [39], where it was shown how in a biological control system trade-offs arise from
the dynamics of binding and unbinding of external ligands.
Further trade-offs, beyond the ones described here, arise because the system S is subject to noise. This
means that a single measurement of the system will only result in a probabilistic answer. In order to increase
the confidence that the correct answer was obtained, repeated measurements need to be performed (see for
example [20]). In terms of our model, each of these sampling events comes at a measurement cost. Altogether,
the accuracy of the biochemical computation can be increased in two ways: Either the EDC is made bigger,
and thus noise reduced, while the cost is increased; or, alternatively, the EDC remains constant but its result
is sampled more frequently. This increases the time and the cost.
References
References
[1] M. Amos, Cellular Computing, Oxford University Press, 2004.
11
[2] G. Seelig, D. Soloveichik, D. Y. Zhang, E. Winfree, Enzyme-free nucleic acid logic circuits, Science
314 (5805) (2006) 1585–1588. doi:10.1126/science.1132493.
URL http://dx.doi.org/10.1126/science.1132493
[3] M. Lakin, A. Phillips, Modelling, simulating and verifying Turing-powerful strand displacement systems,
in: Proceedings of the 17th International Conference on DNA Computing and Molecular Programming,
DNA'11, Springer-Verlag, Berlin, Heidelberg, 2011, pp. 130–144.
URL http://dl.acm.org/citation.cfm?id=2042033.2042047
[4] A. S. Tsuda S., Zauner K.P., The Phi-Bot: A Robot Controlled by a Slime Mould, Springer-Verlag
London, 2009, Ch. 10, pp. 213–232.
[5] A. Friedland, T. Lu, X. Wang, D. Shi, G. Church, J. Collins, Synthetic gene networks that count.,
Science 324 (5931) (2009) 1199–1202. doi:10.1126/science.1172005.
URL http://dx.doi.org/10.1126/science.1172005
[6] R. Sole, J. Macia, Expanding the landscape of biological computation with synthetic multicellular consortia,
Natural Computing (2013) 1–13doi:10.1007/s11047-013-9380-y.
URL http://dx.doi.org/10.1007/s11047-013-9380-y
[7] R. Silva-Rocha, V. de Lorenzo, Implementing an OR-NOT (ORN) logic gate with components of the SOS regulatory network of Escherichia coli.,
Molecular Biosystems 7 (8) (2011) 2389–2396. doi:10.1039/c1mb05094j.
URL http://dx.doi.org/10.1039/c1mb05094j
[8] M. Amos,
I. Axmann, N. Bluthgen, F. de la Cruz, A. Jaramillo, A. Rodriguez-Paton,
F. Simmel, Bacterial computing with engineered populations, Philosophical Transactions of
the
Royal Society A: Mathematical, Physical and Engineering Sciences 373 (2046) (2015) 20140218.
doi:10.1098/rsta.2014.0218.
URL https://doi.org/10.1098/rsta.2014.0218
[9] S. Walker, H. Kim, P. Davies, The informational architecture of the cell, Phil. Trans. R. Soc. A
374 (2063) (2016) 20150057. doi:10.1098/rsta.2015.0057.
URL http://dx.doi.org/10.1098/rsta.2015.0057
[10] P. Davies, S. Walker, The hidden simplicity of biology, Reports on Progress in Physics 79 (10) (2016)
102601. doi:10.1088/0034-4885/79/10/102601.
URL http://dx.doi.org/10.1088/0034-4885/79/10/102601
[11] J. Ninio, Kinetic amplification of enzyme discrimination., Biochimie 57 (5) (1975) 587–595.
[12] A. Fluitt, E. Pienaar, H. Viljoen, Ribosome kinetics and aa-trna competition determine rate and fidelity of peptide synthesis.,
Computional
doi:10.1016/j.compbiolchem.2007.07.003.
URL http://dx.doi.org/10.1016/j.compbiolchem.2007.07.003
Chemistry
31
Biology
and
(5-6)
(2007)
335–346.
[13] T. Gregor, D. Tank, E. Wieschaus, W. Bialek, Probing the limits to positional information., Cell 130 (1)
(2007) 153–164. doi:10.1016/j.cell.2007.05.025.
URL http://dx.doi.org/10.1016/j.cell.2007.05.025
[14] C. Govern, P. ten Wolde, Optimal resource allocation in cellular sensing systems., Proceedings of the
National Academy of Science USA 111 (49) (2014) 17486–17491. doi:10.1073/pnas.1411524111.
URL http://dx.doi.org/10.1073/pnas.1411524111
[15] H. C. Berg, E. M. Purcell, Physics of chemoreception., Biophys J 20 (2)
(1977) 193–219.
doi:10.1016/S0006-3495(77)85544-6.
URL http://dx.doi.org/10.1016/S0006-3495(77)85544-6
[16] P. Mehta, D. Schwab, Energetic costs of cellular computation., Proceedings of the National Academy
of Science USA 109 (44) (2012) 17978–17982. doi:10.1073/pnas.1207814109.
URL http://dx.doi.org/10.1073/pnas.1207814109
12
[17] U. Alon, M. Surette, N. Barkai, S. Leibler, Robustness in Bacterial Chemotaxis, Nature 397 (1999)
168–171.
[18] D. Chu, D. Barnes, The lag-phase during diauxic growth is a trade-off between fast adaptation and high growth rate.,
Scientific Reports 6 (2016) 25191. doi:10.1038/srep25191.
URL http://dx.doi.org/10.1038/srep25191
[19] D. Chu,
In silico evolution of diauxic growth., BMC Evolutionary Biology 15
(2015)
211.
doi:10.1186/s12862-015-0492-0.
URL http://dx.doi.org/10.1186/s12862-015-0492-0
[20] N. Zabet, D. Chu, Computational limits to binary genes., Journal of the Royal Society Interface 7 (47)
(2010) 945–954. doi:10.1098/rsif.2009.0474.
URL http://dx.doi.org/10.1098/rsif.2009.0474
[21] D. Chu, N. Zabet, A. Hone, Optimal parameter settings for information processing in gene regulatory networks,
BioSystems 104 (2011) 99–108. doi:10.1016/j.biosystems.2011.01.006.
URL http://www.cs.kent.ac.uk/pubs/2011/3081
[22] G. Lan, P. Sartori, S. Neumann, V. Sourjik, Y. Tu, The energy-speed-accuracy trade-off in sensory
adaptation, Nature Physics 8 (5) (2012) 422–428.
[23] M. Johansson, J. Zhang, M. Ehrenberg, Genetic code translation displays a linear trade-off between efficiency and accuracy of trna selection.,
of
Proceedings
doi:10.1073/pnas.1116480109.
URL http://dx.doi.org/10.1073/pnas.1116480109
the National Academy
of
Science USA 109
[24] K. Kaizu, W.
de Ronde,
J.
Paijmans,
K. Takahashi,
The Berg-Purcell limit revisited.,
Wolde,
doi:10.1016/j.bpj.2013.12.030.
URL http://dx.doi.org/10.1016/j.bpj.2013.12.030
Biophysical
Journal
106
(1)
(2012)
131–136.
F. Tostevin,
(2014)
(4)
P.
ten
976–985.
[25] R. Landauer, C. Bennett, R. Laing, W. Zurek, Maxwells Demon, Information Erasure, and Computing,
Princeton University Press, 1990, pp. 187–288.
URL http://www.jstor.org/stable/j.ctt7zts1p.8
[26] C. Bennett, The thermodynamics of computation. A review,
International Journal of Theoretical
Physics 21 (12) (1982) 905–940. doi:10.1007/bf02084158.
URL http://dx.doi.org/10.1007/bf02084158
[27] R. P. Feynman, A. Hey, Feynman Lectures On Computation, Westview Press, 2000.
[28] N. van Kampen, Stochastic Processes in Physics and Chemistry, Elsevier, Amsterdam, 2007, third
edition.
[29] H. Qian, D. Beard, Thermodynamics of stoichiometric biochemical networks in living systems far from equilibrium,
Biophysical Chemistry 114 (2-3) (2005) 213220. doi:10.1016/j.bpc.2004.12.001.
URL http://dx.doi.org/10.1016/j.bpc.2004.12.001
[30] D. Beard, S. Liang, H. Qian, Energy balance for analysis of complex metabolic networks., Biophys J
83 (1) (2002) 79–86. doi:10.1016/S0006-3495(02)75150-3.
URL http://dx.doi.org/10.1016/S0006-3495(02)75150-3
[31] D. Gillespie, A rigorous derivation of the chemical master equation, Physica A: Statistical Mechanics
and its Applications 188 (1-3) (1992) 404–425. doi:10.1016/0378-4371(92)90283-v.
URL http://dx.doi.org/10.1016/0378-4371(92)90283-V
[32] U. Seifert, Entropy Production along a Stochastic Trajectory and an Integral Fluctuation Theorem,
Physical Review Letters 95 (4) (2005) 040602. doi:10.1103/PhysRevLett.95.040602.
13
[33] M. Kwiatkowska, G. Norman, D. Parker, PRISM: Probabilistic symbolic model checker, in: P. Kemper
(Ed.), Proc. Tools Session of Aachen 2001 International Multiconference on Measurement, Modelling
and Evaluation of Computer-Communication Systems, 2001, pp. 7–12, available as Technical Report
760/2001, University of Dortmund.
[34] T. Ouldridge, C. Govern, P. ten Wolde, Thermodynamics of computational copying in biochemical systems,
Physical Review X 7 (2). doi:10.1103/physrevx.7.021004.
URL https://doi.org/10.1103/physrevx.7.021004
[35] T. Ouldridge, P. ten Wolde, Fundamental costs in the production and destruction of persistent polymer copies,
Physical Review Letters 118 (15). doi:10.1103/physrevlett.118.158103.
URL https://doi.org/10.1103/physrevlett.118.158103
[36] S. Marzen, H. Garcia, R. Phillips, Statistical mechanics of Monod–Wyman–Changeux (MWC) models,
Journal of Molecular Biology 425 (9) (2013) 1433–1460. doi:10.1016/j.jmb.2013.03.013.
URL http://dx.doi.org/10.1016/j.jmb.2013.03.013
[37] I. Bena, C. V. den Broeck, R. Kawai, Jarzynski equality for the Jepsen gas, Europhysics Letters (EPL)
71 (6) (2005) 879–885. doi:10.1209/epl/i2005-10177-0.
URL http://dx.doi.org/10.1209/epl/i2005-10177-0
[38] W. Bialek, S. Setayeshgar, Cooperativity, sensitivity, and noise in biochemical signaling., Physical Re-
view Letters 100 (25) (2008) 258101.
[39] D. Chu, Limited by sensing - a minimal stochastic model of the lag-phase during diauxic growth, Jour-
nal of Theoretical Biology 414 (2017) 137–146. doi:10.1016/j.jtbi.2016.10.019.
URL https://doi.org/10.1016%2Fj.jtbi.2016.10.019
14
|
1006.5497 | 1 | 1006 | 2010-06-29T03:23:13 | Non-specific cellular uptake of surface-functionalized quantum dots | [
"physics.bio-ph",
"cond-mat.mes-hall"
] | We report a systematic empirical study of nanoparticle internalization into cells via non-specific pathways. The nanoparticles were comprised of commercial quantum dots (QDs) that were highly visible under a fluorescence confocal microscope. Surface-modified QDs with basic biologically-significant moieties, e.g. carboxyl, amino, streptavidin were used, in combination with the surface derivatization with polyethylene glycol (PEG) in a range of immortalized cell lines. Internalization rates were derived from image analysis and a detailed discussion about the effect of nanoparticle size, charge and surface groups is presented. We find that PEG-derivatization dramatically suppresses the non-specific uptake while PEG-free carboxyl and amine functional groups promote QD internalization. These uptake variations displayed a remarkable consistency across different cell types. The reported results are important for experiments concerned with cellular uptake of surface-functionalized nanomaterials, both when non-specific internalization is undesirable and also when it is intended for material to be internalized as efficiently as possible.
Published article at: http://iopscience.iop.org/0957-4484/21/28/285105/ | physics.bio-ph | physics | Non-specific cellular uptake of surface-functionalized
quantum dots
T A Kelf1, V K A Sreenivasan1, J Sun1, E J Kim2, E M Goldys1 and A V
Zvyagin1
1 MQ Photonics Centre, Faculty of Science, Macquarie University, Sydney,
Australia
2Department of Science Education-Chemical Education Major, Daegu
University, Gyeonbuk, South Korea
E-mail: [email protected]
Abstract. We report a systematic empirical study of nanoparticle internalization
into cells via non-specific pathways. The nanoparticles were comprised of
commercial quantum dots (QDs) that were highly visible under a fluorescence
confocal microscope. Surface-modified QDs with basic biologically-significant
moieties, e.g. carboxyl, amino, streptavidin were used, in combination with the
surface derivatization with polyethylene glycol (PEG) in a range of immortalized
cell lines. Internalization rates were derived from image analysis and a detailed
discussion about the effect of nanoparticle size, charge and surface groups is
presented. We find that PEG-derivatization dramatically suppresses the non-
specific uptake while PEG-free carboxyl and amine functional groups promote
QD internalization. These uptake variations displayed a remarkable consistency
across different cell types. The reported results are important for experiments
concerned with cellular uptake of surface-functionalized nanomaterials, both
when non-specific internalization is undesirable and also when it is intended for
material to be internalized as efficiently as possible.
PACS: 42.62.Be, 81.07.Ta, 87.16.dp, 87.85.Rs
Introduction
To sustain the lifecycle of living cells material must be absorbed from the extracellular medium
via certain mechanisms. These mechanisms fall into two categories: specific internalization, that
requires the cell to actively recruit molecules into the cytoplasm, and non-specific internalization,
which constitutes random processes in which the cell has no active control. Receptor-mediated
endocytosis is an important example of the specific uptake mechanism that facilitates import of
selected extracellular macromolecules and allows inter-cellular signalling. This process is
regulated by plasma membrane receptors that are only activated by receptor-specific ligands [1-
4]. As a result, only biomolecular complexes grafted with these ligands can gain entry into the
cell. In the process of endocytosis the plasma membrane is engulfed inwards from specialized
membrane micro-domains forming either clathrin- or caveolin-coated pits. In recent years this
specific cellular uptake mechanism has been a subject of intense research driven by the
motivation to better understand cellular molecular trafficking and potential applications in
targeted drug delivery [5]. Non-specific cellular uptake refers to a process of extraneous material
internalization; it can vary depending on the cell type and has poor material selectivity. For
virtually all eukaryotic cells, non-specific uptake is dominated by pinocytosis [6, 7]. This
process involves continual budding of vesicles from the cell surface, allowing fluid and extra-
1
cellular material to enter the cell. After either specific or non-specific internalization the material
trapped in the vesicles is sequestered into endosomes, subsequently it is transported to locations
within the cell [4], moved back to the extra-cellular medium, or taken to the lysosomes for acidic
degradation. Nanoparticles incubated with cells are eventually concentrated in lysosomes or in
the perinuclear recycling compartment [8] as they are of no use to the cell. Often, due to the
similarity of pinocytosed and receptor-mediated internalization patterns, the discrimination of the
non-specific versus specific uptake processes is not straightforward. The literature reports
addressing non-specific cellular uptake and the key factors determining its rate are dispersed over
various research fields from gene transfection [9, 10] and drug delivery [11, 12] to nanoparticle
synthesis [13, 14]. They provide evidence that the key factors determining internalization are
size [15, 16, 17], charge [18, 19] and surface functional groups [12], where polyethylene glycol
(PEG) is of particular importance [20-22].
Nanoparticles serving as biomolecular cargo vehicles hold considerable promise for targeted
delivery, with their large surface areas hosting various functional moieties that can dock different
biomolecules. Quantum dots (QDs) have been intensely investigated due to their exceptional
luminescent properties: unprecedented efficiency (high action cross-section, which is a product of
quantum efficiency and absorption cross-section [23]). Spectral emission tunability by size
variations allowing spectral multiplexing and their convenient excitation band in the UV spectral
range make QDs particularly attractive as fluorescent labels and they have already been widely
used in a number of biological studies, as molecular reporters [24]. Their large two-photon
action cross-sections compared to the best organic fluorophores add further to their appeal [23].
The QD technology is maturing rapidly with a range of QD products available commercially,
including QDs with amino, carboxyl and streptavidin surface groups. The toxicity of QDs
remains a hotly debated issue, which may limit medical applications [25-28], however their
brightness and photostability make them ideal for in-vitro biomedical applications.
In this paper, we investigate the cellular internalization rates of a range of different commercially
available QDs
including streptavidin-coated QDs and carboxyl- and amino- surface
functionalized QDs with and without PEG linker chains. Carboxy- and amino- functionalised
QDs have been chosen because they are readily amenable to further conjugation with
biomolecules following various chemical treatments. Streptavidin-coated QD are important
because, by virtue of the exceptional affinity of streptavidin for biotin (dissociation constant ~ 10-
15 M),
streptavidin QDs
can be
easily
attached
to biomolecules of
interest
(ligands/proteins/antibodies), which have been coupled to biotin. Some of the QDs investigated
here were also modified chemically to further understand the effect of chemical groups on
cellular internalisation dynamics. This modification revealed significant modulation of the
cellular uptake rates.
The aim of this paper is to provide an initial guide for the use of nanoparticles in cellular systems.
While this work focuses specifically on QDs, the results and discussion can be applied to various
other types of nanoparticles as the effects of size, surface charge and molecular coatings are
generally more important than the chemical identity of nanoparticles. This general discussion
aims to shed light in issues arising in early stages of the research on application of nanoparticles
in cellular systems. Our experience suggests that without proper controls and time-consuming
investigations the reasons for particle internalization can be hard to elucidate.
2
Experimental Procedure
1.
Quantum Dots
All QDs used in this study were obtained from two major suppliers: Invitrogen (Oregon, USA),
(referred to as QDi) and eBiosciences (San Diego, USA), (referred to as QDe). Both have the
same basic nanostructure consisting of a cadmium selenide core passivated with a layer of zinc
sulphide (CdSe/ZnS). For this study, QDs with a fluorescence emission maximum at 605 nm
were chosen; the QD excitation band is broad, mostly in the ultraviolet spectral range, and has an
additional excitation peak at 545 nm allowing convenient excitation in visible. All QDs are
capped with an amphiphilic polymer coating by the manufacturer. This coating serves three
purposes: to reduce cytotoxicity, improve colloidal stability, and enable the grafting of functional
moieties. Hence, the overall size of these QD complexes ranged from 15 nm to 25 nm depending
on the surface functional groups. This work used seven different types of QDs whose full list is
shown in figure 1. They were, as follows: Carboxyl-QDs (referred to as C-QDi), Amino-PEG
QDs (N-pQDi, where p refers to the presence of polyethylene-glycol, PEG) and Streptavidin-QDs
(SA-QDi), all purchased from Invitrogen. Carboxyl-PEG QDs (C-pQDe) and Amino-PEG QDs
(N-pQDe), were purchased from eBiosciences. To further understand the effect of the surface
terminal groups, C-QDi and C-pQDe were conjugated to ethylene-diamine to obtain amine-
terminated quantum dots, referred to as N-CC-QDi and N-CC-pQDe. The QDs can be divided
into PEG-derivatized and PEG-free, these are drawn in the left and the right column in figure 1,
respectively. It is also possible to classify the QDs with respect to the chemical terminal group
(amine, carboxyl or streptavidin), some of which have a reacted variant.
For the coupling reaction 400 fmole (1 molar equivalent) of QDs were diluted in 400 µL (MES)
buffer with 2 mM (EDTA, pH=6) in a vial. Freshly dissolved room temperature equilibrated 1-
ethyl-3-(3-dimethylaminopropyl)
carbodiimide
(EDC)
(10 mg/mL)
and
sulfo- N-
Hydroxysuccinimide (sulfo-NHS) (20 mg/mL) solutions in deionized water were added to the
QD vial for activation and the solution was incubated at room temperature with gentle shaking
for 15 minutes. An excess of ethylene-diamine was added along with 50 µL of 5x borate buffer
(pH = 8.4). After 1-hour room temperature incubation with gentle shaking free EDC, sulfo-NHS
and ethylene-diamine were removed by using a desalting column (Prepacked Sephadex G-25
medium column, GE Healthcare) equilibrated with phosphate buffer saline PBS (pH = 7.2).
3
Figure 1. List of commercially available QDs from Invitrogen and eBiosciences. Chemical
structure of the surface groups is shown along with the short hand notation. Additional chemical
reactions have been performed where noted.
the concentration and quantum efficiency of the reacted QDs were measured using an absorption
spectrometer and fluorimeter, by comparing with untreated QDs. We found that Invitrogen QDs
have approximately 2.5 times greater quantum efficiency than that of the eBiosciences QDs.
After the EDC reaction the Invitrogen QDs (N-CC-QDi) and eBiosciences QDs (N-CC-pQDe)
suffered a 2.5- and 10- fold drop in quantum efficiency, respectively.
2.
Cell Experiments
Two rat tumour cell lines were obtained from ATCC (USA), these were GH4C1 (pituitary) and
AR42J (pancreas) cells. A pituitary tumour cell line, AtT20 was a kind gift from Dr M Connor
(The University of Sydney, Australia). Ar42J cells were cultured in F-12K medium (DKSH)
with 20% fetal bovine serum. Both GH4C1 and AtT20 cells were cultured in F-10 Ham’s
(DKSH) with 2.5% fetal bovine serum and 15% horse serum. All cells were grown in culture
flasks (BD Falcon) inside a humidified incubator at 37oC with 5% CO2.
The day before experiments cells were transferred to 8 chambered culture slides (BD Falcon) and
incubated in medium, as above. The cells were then incubated with a 20-nM QD solution in 300-
μL phosphate buffer saline (PBS, Invitrogen, pH = 7.2) with additional CaCl2 (0.9 mM), MgCl2
(0.5 mM), BSA (0.1%) and D-Glucose (20 mM) at 37oC, 5% CO2 for 60 min. After QD
incubation, cells were washed twice with PBS then fixed using 3.7% (w/v) paraformaldehyde
solution for 20 min followed by a final PBS wash.
Imaging and image processing
3.
A confocal laser-scanning microscope (Leica TCM SP2) with a 488-nm Ar-ion laser was used for
acquisition of fluorescence images of the QDs. An external Nikon DS-Qi1Mc camera was
attached to acquire the respective bright field, transmission images. The internalization was
quantified by fluorescence intensity of the QDs inside the cells. For each experiment,
approximately 20 cells were imaged using a spectral filter of the bandwidth 585 nm – 625 nm to
accommodate for the QD emission and reject illumination light. All imaging parameters were
kept constant for all experiments to allow direct comparisons. After background correction and
intensity scaling to account for different QD quantum efficiencies, the total brightness inside each
cell was calculated (software Igor Pro, Wavemeterics). These values were normalized to the cell
size and averaged over all cells to provide a single number that corresponded to the level of
internalization for each QD type into each cell line. These values were subsequently normalized
to the internalization rate of N-CC-QDi into AtT20 cells to facilitate comparison between
different cell lines and QD types.
Results and Discussion
Results
Firstly, we examined the internalization mechanism of QDs. Figure 2a shows representative
images of 20 nM of C-QDi incubated with Ar42J cells for one hour at 4oC and 37oC. No
internalization was observed at 4oC, with the QDs localized on the cell membrane. The cellular
machinery was switched on at 37ºC manifested by the QD uptake in the cellular cytoplasm at the
expense of the reduced concentration of QDs localized on the membrane. This internalization
4
mechanism was active, as opposed to, for example, the passive diffusion cellular uptake of small-
size molecules [29]. This observation is consistent with receptor-mediated endocytosis [30] that
is not effective at 4ºC. The receptor-mediated endocytosis mechanism of the QD uptake was
additionally confirmed by pre-treatment of the Ar42J cells with sucrose before the incubation of
20 nM of C-QDi, with the result shown in figure 2b. In this case, no internalization was
observed. Sucrose is known to disrupt the formation of clathrin vesicles [16, 31], hence this
shows that C-QDi were predominantly internalized via endocytosis that involved clathrin-coated
vesicles. Figure 2c shows the detected level of C-QDi in Ar42J cells as a function of incubation
time. The internalization rates were calculated by analysing the number of QDs inside the cells
(ignoring those localized on the membrane), and averaging over 20 cells for each time point. The
internalization saturated at approximately 10 minutes, a period limited by the number and activity
of the receptors recruited for the internalization process. The variation in the internalization level
at longer incubation times is explained by statistical variations in the cell uptake. We verified that
there was no observable variation in the internalization rate up to 90 minutes. The incubation
time of 60 minutes was used in further experiments to ensure that the internalization of the
nanoparticles was as complete as possible.
Figure 2. (a) C-QDi incubated at 4oC and 37oC in Ar42J cells, QDs are pseudo-colour-coded in
red, nuclei were stained with blue fluorescent dye Hoerst and shown in blue colour. Image size
37.5 µm × 37.5 µm. (b) C-QDi incubated (i) with sucrose and (ii) without sucrose in Ar42J cells.
Image size and colours are the same as in (a). (c) plot of the C-QDi internalization into Ar42J
cells versus incubation time.
Figure 3 shows representative images of different QDs in the tested cell lines. Figure 4 shows
their calculated internalization levels, with the values summarized in table 1. QD sizes and zeta
potentials, measured using a dynamic light scattering (DLS) method (Zetasizer, Malvern, UK)
are also presented in table 1. QD hydrodynamic diameters (which overestimate the geometric
average diameter by roughly 10%) were measured to be approximately 20 nm with the exception
of N-CC-QDi (100 nm, table 1). This larger measured size reflects the process of QD
aggregation in the buffer solution. We have tested all QD solutions in water and PBS a number
of times and have sometimes found aggregates present in the data. The use of PEG and the
amphiphilic polymer coating act to reduce aggregation significantly over that of bare particles but
5
never completely remove the effect. Hence, sizes of the nanoparticle complexes formed in buffer
medium, rather than their primary sizes, should be taken into account in the context of cellular
internalization. The effect of nanoparticle size will be discussed later; however we note that QD
aggregates were under 200 nm in size, with smaller aggregate/primary particle often present in
buffers. Such aggregation levels still permitted the efficient QD uptake into the cells, as shown
in figure 3 for N-CC-QDi since the vesicles can accommodate up particles up to ~ 200 nm.
We now discuss the general trends that can be seen in the data. On average, Ar42J and GH4C1
cells had a similar level of internalization, which was, roughly, half that of AtT20 cells. QDs
surface-derivatized with PEG showed comparatively low levels of cellular internalization. We
note internalization of N-pQDi was virtually negligible, at the noise level of the detection. The
level of the SA-QDi internalization was slightly greater than that of most of the PEG-derivatized
QDs, except for AtT20 where low internalization rate is observed.
Figure 3. Representative bright-field images of the tumour cells superimposed with the
epifluorescence images of quantum dot fluorescence (red). The image acquisition parameters
were constant for all experiments to facilitate direct comparison. Each image panel size 37.5 µm
× 37.5 µm.
6
Figure. 4. Bar plot of the QD uptake levels for the three tumour cell lines: (a) normalized to that
of N-CC-QDi quantum dots in AtT20 cells; (b) normalized across each cell line data to that of N-
CC-QDi.
Table1. Normalized QD uptake rates into the 3 tumour cell lines, QD size and zeta potential.
Internalization
Ar42J GH4C1 AtT20
0.225
0.419
1
0.082
0.12
-
0.027
0.03
0.078
Zeta Potential
(mV)
100
28
23
-32
-31
-2
Size
(nm)
QD Type
N-CC-QDi
C-QDi
N-pQDe
N-CC-
pQDe
C-pQDe
SA-QDi
N-pQDi
0.024
0.029
0.039
0.005
0.035
0.002
0.043
0.008
0.022
0.018
0.012
0.002
24
22
20
21
-12
-12
-15
-2
7
Discussion
It is well known that a number of factors affect the rate of non-specific particle internalization
into cells. These include charge [18, 19], size [15, 16, 17], and surface functional groups [12].
Additionally, hydrophobic nanomaterials which have an affinity to the lipid bilayer of the cell
facilitate uptake [32]. Pristine amphiphilic-polymer-coated QDs and PEG-derivatized QDs have
similar hydrophilic properties, and yet as shown, their cellular uptakes are dramatically different.
To understand this difference we will discuss the roles of the receptor-specific mediated pathway,
size effects, PEG-derivatization, and charge, in the internalization process.
Firstly, we discuss whether direct receptor mediated pathways plays a significant role in the
internalization of the streptavidin-coated QDs in comparison with non-specific mechanisms.
Streptavidin has a high binding affinity to biotin, which belongs to the class of vitamins. Since
biotin is present in the culture medium and can be adsorbed onto the cell membrane this could
enhance the internalization of Sav-QDi through streptavidin biotin binding [33]. Also, it has been
reported that the RYD peptide chain in streptavidin is sufficiently similar to the sequence RGD
that is used by many cell adhesion-related molecules, and, thus, unmodified streptavidin can be
directly internalized [34]. Thirdly, there can be non-specific internalization through pinocytosis,
a common pathway for a broad class of extraneous material internalisation, including QDs. We
tested the first pathway by pre-binding biotin to the streptavidin-coated QDs, and found no
discernable change in the level of internalization. This ruled out the possibility of internalization
via the biotin-mediated pathway. The second pathway was more difficult to test, however the
total observed internalization levels of the streptavidin-coated QDs were broadly comparable to
the lowest levels of internalization of the other (PEG-containing) QDs, that were limited to the
pinocytosis pathway. This suggests that pinocytosis was the dominant mechanism. The added
complexity of the potential multiple internalization pathways makes streptavidin-coated QDs less
attractive, as a reference nanoparticle unless preliminary work is conducted to ensure the
receptor-mediated internalization is significantly greater compared to alternative processes.
With respect to the carboxyl group mediated internalization, it has been shown that carboxylic
QDs were endocytosed via lipid rafts [35]. However, since the sucrose-induced inhibition of the
C-QDi internalization strongly suggested the clathrin-vesicle pathway (figure 2b), the lipid raft
internalization mechanism is ruled out as a dominant factor for these cell lines. At the same time,
its active role in the reported internalization scenarios [35] indicates that the process can be more
complex and makes it difficult to explain C-QDi internalization within a simple framework.
Size has been shown to have a strong influence over the rate of nanoparticle internalization.
Whilst our primary QDs were of a similar size, it was noted that there can be aggregation in the
culture medium, leading to distribution of actual sizes. The effect of size on internalisation has
been addressed in several papers [16, 17, 37-39], and although consensus has not been reached, a
general trend is emerging, as schematically shown in Figure 5. The size effects fall
approximately into three classes: Clathrin, Caveolae, Macro pinocytosis.
8
Figure 5. Qualitative plot of the internalization level versus the particle size. This plot was
compiled using published data [16, 17, 37-39] obtained for a variety of cell models and particle
types.
Clathrin-mediated endocytosis pathways are predominant for nanoparticles smaller than 200 nm,
and it has been shown both experimentally [16] and theoretically [38] that 50 nm particles are
most efficiently endocytosed. The non-specific internalization in this size range is hypothesized
to occur via adsorption of the proteins from the culture medium onto the particles surface,
followed by their internalization via receptor-mediated pathways. Thus the rate of internalization
is dependent on the available receptors on the cell surface and the number of receptors a single
nanoparticle triggers when it adheres to the membrane. It has been shown experimentally that for
particles <50 nm, several closely spaced particles are required to trigger the formation of a
clathrin-coated vesicle [17]. As a result, an internalization rate per particle is reduced, as
manifested by a descending slope at sizes below 50 nm. A theoretical model based on the
balance between free energy required to drive nanoparticles into the cell and diffusion kinetics of
the recruitment of receptors to the binding site [38] helps explain this decreasing internalization
level. For particle sizes in excess of 50 nm, more receptors are engaged by the particle
internalization process followed by their translocation into the cytoplasm. To sustain the
internalization, these receptors must be recycled back to the membrane, which takes time and
limits the internalization rate. When nanoparticles reach sizes greater than 200 nm they can no
longer be internalized through clathrin vesicles, and cells recruit alternative internalization
pathways. It has been shown [17] that caveolae-mediated pathways were of the greatest
importance for particles between 200 nm and 1000 nm. This process is slower than that of the
clathrin internalization, hence the efficiency is lower. For particles greater than a micron in size
(up to 5 µm in diameter), the effect of macro-pinocytosis becomes important. This requires
ruffling of the cell membrane to form large troughs, which can subsequently bud in the form of
large size vesicles into the cell. This is a genuinely non-specific process, and highly dependent of
cell type and internalization conditions. The internalization rate drops dramatically above around
one micron due to difficulty of forming these large vesicles.
Particle shape has also been shown to play a role in the internalization [37], which adds another
layer of complexity. We note briefly that a cell perceives the particle in terms of its surface area
in contact with the cell membrane, and thus this size governs internalization, as discussed above.
9
We now turn to the role of PEG in the non-specific nanoparticle uptake in cells. While precise
reasons behind the significantly reduced internalization of PEG-derivatized nanoparticles are still
debated, a number of consistent observations have been presented in the literature [12, 32, 35,
36]. PEG is an uncharged, hydrophilic polymer, which is non-immunogenic and characterized by
a high steric stabilization. Figure 6 shows the zeta potential of the investigated QDs in PBS (pH
= 7.2), along with the average internalization rate over the investigated cell lines.
Figure 6. Plot of the QD zeta potentials in PBS, pH 7.2, and average internalization levels.
As can be seen, PEG-derivatized QDs had zeta potentials closer to zero than their pristine
counterparts. This surface charge neutralization occurs due as the PEG molecules stopping the
liquid flow in their vicinity. The fluid slippage plane was, therefore, shifted, and, as a result,
modified the zeta potential (the zeta potential is defined the electrostatic potential measured at the
slippage plane – the further the slippage plane from the particle surface, the smaller the potential,
and hence, the zeta-potential ) [14, 18]. This effect is influenced by both PEG chain length, and
the density of PEG molecules on the nanoparticle surface. To test this dependency, we
performed EDC reactions to conjugate two different length PEG molecules (molecular weight
3400 and 5000 Da) in different surface concentrations to C-QDi, as shown in figure 7.
10
Figure 7. Zeta-potential of QDs derivatized by 3400-Da and 5000-Da PEG molecules at
different surface densities.
We calculated that up to 400 PEG molecules can covalently bind to an individual QD. At this
PEG surface density the measured zeta potential was close to zero, showing almost complete
screening of the QD surface charge. A PEG layer thickness of 2 nm was inferred from the
hydrodynamic diameter measured by DLS. As the surface density was reduced, the surface
became more highly charged (the zeta potential shifted to the negative), approaching the value for
the uncoated C-QDi. The PEG surface density was correspondingly reduced, as confirmed by the
DLS measurements. Note that 3400-Da and 5000-Da PEG-derivatization showed similar result
within the experimental accuracy, in agreement with the literature reporting the effect of PEG-
derivatization on the internalization at the 660-Da molecular weight [12].
A balance between the PEG length and surface density is important for optimal design of
nanoparticle PEG-coating. When densely packed, PEG covers the majority of surface binding
sites, and the molecules align perpendicular to the particles surface. At a lower density, due to the
flexibility of PEG, it conforms to a ‘mushroom’ type state, where the chains are bent over the
surface, again covering a large number of surface sites [18]. The main difference between the
two conformational states is that at the lower density, the PEG chains tangle and cause the
terminal groups to become covered with the PEG layer. As a result, non-specific internalisation
[32, 41] is reduced. By the same token, a number of bioconjugation-active terminals are also
reduced. While only limited information about the precise nature of the different QDs has been
provided by the manufacturers, a different PEG length and density ratio is likely to account for
the difference in internalization between N-pQDi and N-pQDe.
It has been shown that charge has a profound effect on the rate of non-specific internalization [9,
40], through electrostatic interactions. The PEG molecules reduce the ability of the QD to
undergo these interactions by physically separating the surface from that of the cell (steric
hindrance), hence reducing QDs adherence to the cell membrane [13, 22, 41], and leading to
reduced internalization. PEG also acts efficiently to reduce protein adsorption onto the
nanoparticle surface [12]. This property has been shown prevent phagocytosis [11, 12], as well
as reducing the binding of nanoparticles to the cell membrane, and hence lowering
internalization [16].
We now discuss the surface charge effect. The importance of surface charge on the non-specific
internalization rate has been discussed in the past [12, 18, 19], and it has been shown that both
11
positively and negatively charged molecules can be efficiently internalized into cell, utilizing
interactions with various charged proteins on the cell membrane [9, 10, 40]. It has also been
shown that nearly neutral QDs with hydroxyl (-OH) surface functionalization, also exhibit greatly
reduced non-specific internalization [13], when compared with the carboxyl or amino
counterparts, so charge is of a definitive importance to understand the internalization
characteristics.
As noted above, presence of the PEG-layer reduces an effective number of the terminal groups
due to the entanglement of the chains, and hence it reduces the apparent charge. The zeta-
potential measurements showed that PEG-derivatized QDs with amino terminal groups, while
still negative, were charged more positively in comparison with those of the carboxyl groups, as
expected. This negative charge is attributed to the intrinsic zeta-potential of the pristine QDs.
This was confirmed by the measurements of the PEG-free QDs, which both showed large
negative zeta-potentials regardless of the terminal groups.
Positively charged amino-functionalized particles exhibit the greatest internalization rates [10,
42], that can be ten-fold higher than negatively-charged carboxylated particles in neutral pH [40].
This is in agreement with our results, where both N-CC-QDi and C-QDi were internalized to a
greater extent than SA-QDi and amino-functionalized particle internalization rate was, on
average, ten-fold that of their carboxyl equivalents. It should be noted that the overall charge of
the QD will influence electrostatic interactions with the cell membrane, although molecular
interactions with individual proteins become important in close proximity. We propose that that
this is the reason that N-CC-QDi internalization exceeds that of the C-QDi, while the zeta
potential is similar.
Summary
The level of cellular internalization of various commercially available QDs by three different
tumour cell lines has been investigated. Overall, we showed that the presence of PEG on the
particle surface significantly reduced the internalization. This was due to the PEG-layer-induced
steric hindrance: the direct docking of QDs to the cellular membrane was obstructed, as well as
protein adsorption on the QD surface, which otherwise would promote internalization. Charge
has the greatest effect on internalization level, so should be most carefully considered when
conducting culture experiments in live cells. While the uptake mechanisms of nanoparticles into
cells is defined as non-specific, these pathways were initiated by surface receptors, either through
direct interactions between the charged particle and the receptor or via proteins adsorbed on the
nanoparticles surface.
Commercially available, PEG-derivatized QDs are appropriate as control particles due to their
low non-specific binding/internalization into the investigated cell types, of these we find N-pQD i
to be best, showing almost zero levels of internalization. Amino or carboxyl, PEG-free QDs can
be absorbed by the cells in relatively large quantities due to their large surface charges, and while
both positive and negatively charged particles are effectively internalized, positive amino capped
particles are the most efficient. Such particles can serve as useful staining models, or for toxicity
evaluation, where pathway-specific trafficking into the cell is of little importance. The
importance of particle size should also be carefully considered, particles of the order of 50 nm
will be most efficiently internalized into the cell, but this must be balanced with need of the
particle once internalized, for protein pathway tracking for example, this size will affect the
proteins movements in the cell.
12
This paper highlights that non-specific uptake is an important factor, when considering the use of
nanoparticles in cellular systems. We believe that these observations will be of use to researchers
moving from the fields of physics and nanotechnology into cell biology in providing the basis
needed to perform efficient studies into nanoparticle internalization and interactions with cellular
systems.
Acknowledgment
We appreciate help and equipment loans from David Inglis. Mark Connor, Medical Foundation of
The University of Sydney, provided the AtT20 cells. The work was supported by the Macquarie
University Research Innovation Fund #1136900.
Reference
1. Goldstein J L, Anderson R G W and Brown M S 1979 Nature. 279 679-85
2. Chen B, Liu Q L, Zhang Y L, Xu L and Fang X H 2008 Langmuir. 24 11866
3. Dahan M, Levi S, Luccardini C, Rostaing P, Riveau B and Triller A 2003 Science. 302 442-45
4. Delehanty J B, Mattoussi H and Medintz I L 2009 Ana.l Bioanal. Chem. 393 1091-05
5. Rajendran L, Knolker H -J and Simons K 2010 Nat. Rev. Drug. Discov. 9 29-42
6. Alberts B, Johnson A, Lewis J, Raff M, Roberts K and Walter P, Molecular Biology of the
Cell. 5th ed. 2008: Garland Science
7. Brodsky F 1988 Science. 242 1396-1402
8. Barua S and Rege K 2009 Small. 5 370-76
9. Belting M 2003 Trends. Biochem. Sci. 28 145-51
10. Schwartz B, Benoist C, Abdallah B, Scherman D, Behr J P and Demeneix B A 1995 Hum.
Gene. Ther. 6 1515-24
11. Shan X Q, Yuan Y, Liu C S, Tao X Y, Sheng Y and Xu, F 2009 Biomed. Microdevices. 11
1187-94
12. Vonarbourg A, Passirani C, Saulnier P and Benoit J-P 2006 Biomaterials. 27 4356-73
13. Kairdolf B A, Mancini M C, Smith A M and Nie S 2008 Anal. Chem. 80 3029-34
14. Liufu S, Xiao H and Li Y 2004 Powder. Technol. 145 20-24
15. Howarth M, Liu W H, Puthenveetil S, Zheng Y, Marshall L F, Schmidt M M, Wittrup K D,
Bawendi M G and Ting A Y 2008 Nat. Methods. 5 397
16. Chithrani B D and Chen W C W 2007 Nano. Lett. 7 1542-50
17. Rejman J, Oberle V, Zuhorn I S and Hoekstra D 2004 Biochem. J. 377 159-169
18. Mosqueira V C F, Legrand P, Gulik A, Bourdon O, Gref R, Labarre D and Barratt G 2001
Biomaterials. 22 2967-79
19. Ogris M, Steinlein P, Carotta S, Brunner S and Wagner E 2001 Aaps. Pharmsci. 3 43-53
20. Liu W, Howarth M, Greytak A B, Zheng Y, Nocera D G, Ting A Y and Bawendi M G 2008
J. Am. Chem. Soc. 130 1274-84
21. Warnement M R, Tomlinson I D, Chang J C, Schreuder M A, Luckabaugh C M and
Rosenthal S J 2008 Bioconjugate. Chem. 19 1404-13
13
22. Bentzen E L, Tomlinson I D, Mason J, Gresch P, Warnement M R, Wright D, Sanders-Bush
E, Blakely R and Rosenthal S. J. 2005 Bioconjugate. Chem. 16 1488-94
23. Larson D R, Zipfel W R, Williams R M, Clark S W, Bruchez M P, Wise F W and Webb W
W 2003 Science. 300 1434
24. Medintz I L, Uyeda H T, Goldman E R and Mattoussi H 2005 Nat. Mater. 4 435
25. Hardman R A 2006 Environ. Health. Persp. 114 165
26. Hoshino A, Manabe N, Fujioka K, Suzuki K, Yasuhara M and Yamamoto K 2007 J. Artif.
Organs. 10 149-57
27. Li H, Zhou Q, Liu W, Yan B, Zhao Y and Jiang G B 2008 Sci. China. Ser. B. 51 393-400
28. Maysinger D, Lovric J, Eisenberg A and Savic R 2007 Eur. J. Pharm. Biopharm. 65 270-81
29. Ross J A, Zvyagin A V, Heckenberg N R, Upcroft J, Upcroft P and Rubinsztein-Dunlop H
2006 J. Biomed. Opt. 11 014008
30. Silverstein S C, Steinman R M and Cohn Z A 1977 Annu. Rev. Biochem. 46 669-722
31. Daukas G and Zigmond S H 1985 J. Cell. Biol. 101 1673-79
32. Hu Y, Xie J, Tong Y W and Wang C-H 2007 J. Control. Release. 118 7-17
33. Howarth M, Takao K, Hayashi Y and Ting A Y 2005 P. Natl. Acad. Sci. USA. 102 7583-88
34. Alon R, Bayer E A and Wilchek M 1990 Biochem. Bioph. Res. Co. 170 1236-41
35. Zhang LW and Monteiro-Riviere N A 2009. Toxicol. Sci. 110 138-55
36. Deshpande M C, Davies M C, Garnett M C, Williams P M, Armitage D, Bailey L,
Vamvakaki M, Armes S P and Stolnik S 2004 J. Control. Release. 97 143-56
37. Gratton S E A, Ropp P A, Pohlhaus P D, Luft J C, Madden V J, Napier M E and DeSimone J
M 2008 Proc. Natl. Acad. Sci. U.S.A. 105 11613-18
38. Zhang S, Li J, Lykotrafitis G, Bao G and Suresh S 2009 Adv. Mater. 21 419-24
39. Chithrani B D Ghazani A A and Chen W C W 2006 Nano. Lett. 6 662-68
40. Foster K A, Yazdanian M and Audus K L 2001 J. Pharm. Pharmacol. 53 57-66
41. McNamee C E, Yamamoto S and Higashitani K 2007 Biophys. J. 93 324-34
42. Duan H W and Nie S M 2007 J. Am. Chem. Soc. 129 3333
14
|
1708.03138 | 5 | 1708 | 2018-11-30T11:00:24 | Stability and roughness of interfaces in mechanically-regulated tissues | [
"physics.bio-ph",
"cond-mat.soft"
] | Cell division and death can be regulated by the mechanical forces within a tissue. We study the consequences for the stability and roughness of a propagating interface, by analysing a model of mechanically-regulated tissue growth in the regime of small driving forces. For an interface driven by homeostatic pressure imbalance or leader-cell motility, long and intermediate-wavelength instabilities arise, depending respectively on an effective viscosity of cell number change, and on substrate friction. A further mechanism depends on the strength of directed motility forces acting in the bulk. We analyse the fluctuations of a stable interface subjected to cell-level stochasticity, and find that mechanical feedback can help preserve reproducibility at the tissue scale. Our results elucidate mechanisms that could be important for orderly interface motion in developing tissues. | physics.bio-ph | physics |
Stability and roughness of interfaces in mechanically-regulated tissues
John. J. Williamson and Guillaume Salbreux
The Francis Crick Institute, 1 Midland Road, London NW1 1AT, UK
(Dated: December 3, 2018)
Cell division and death can be regulated by the mechanical forces within a tissue. We study
the consequences for the stability and roughness of a propagating interface, by analysing a model
of mechanically-regulated tissue growth in the regime of small driving forces. For an interface
driven by homeostatic pressure imbalance or leader-cell motility, long and intermediate-wavelength
instabilities arise, depending respectively on an effective viscosity of cell number change, and on
substrate friction. A further mechanism depends on the strength of directed motility forces acting
in the bulk. We analyse the fluctuations of a stable interface subjected to cell-level stochasticity,
and find that mechanical feedback can help preserve reproducibility at the tissue scale. Our results
elucidate mechanisms that could be important for orderly interface motion in developing tissues.
Interfaces are ubiquitous in tissue biology, between a
tissue and its environment [1 -- 3] or between cell popu-
lations [4 -- 8]. There is great interest in how interfaces
propagate smoothly or maintain their shape in the face
of cell proliferation and renewal [1, 5, 9 -- 12], for example
by line tension acting at tissue boundaries [10, 13 -- 15].
Theoretical efforts have focused on contour instabili-
ties in cancer [16 -- 20], branching [21, 22] or folding [23],
and wound healing [3, 24, 25].
In models that include
nutrient diffusion, protruding regions access more nutri-
ent, triggering further growth [17, 18, 26], reminiscent of
the Mullins-Sekerka instability in non-living systems [27].
An epithelium-stroma interface could form undulations
due to mechanical stresses from cell turnover [28, 29],
while a Saffman-Taylor-like instability based on viscosity
contrast has been proposed to underlie branching in the
developing lung [21]. A recent cell-based simulation of
imbalanced mechanically-regulated growth between two
epithelia observed a stable interface, and quantified its
roughness [8]. A related simulation of cells in an inert
medium found finger-like protrusions, arising for higher
friction in the medium relative to the cells [30, 31]. Ref.
[32] calculated the steady-state surface fluctuations of a
non-growing tissue maintained in its homeostatic state.
such as in the developing
Drosophila abdominal epidermis [33, 34], interface prop-
agation occurs. This may be driven by imbalances in
pressure associated with cell division, and/or directed cell
motility, which cause the expansion of one tissue at the
other's expense.
In tissue replacement,
In this Letter, we ask whether factors that drive an
interface's propagation can also affect its stability and
roughness. We are particularly interested in the conse-
quences of mechanically-regulated cell division and death
for the behaviour of interfaces. If cell number change is
sensitive to mechanical forces [11, 35 -- 44] it leads to a
"homeostatic pressure" [8, 45 -- 47] which can drive inter-
face propagation without coherently-directed cell motil-
ity forces. Alternatively, active, directed migration is
proposed to drive interface motion in wound healing
[2, 48, 49] or tissue replacement [34]. We study a model
FIG. 1. A) Side-on schematic of competing epithelial tissues.
Each tissue is described with coarse-grained fields of cell den-
sity ρ, stress σij and velocity vi. Each has an elastic modulus
χ (Eq. 3), substrate friction ξ (Eq. 4), and division/death re-
sponsive to strain on a timescale τ (Eq. 2). B) Top-down view
of the tissues illustrating an interface contour fluctuation. We
use a 2D description, with z a coordinate parallel to x in the
comoving frame of the interface.
of competing epithelial tissues, with an interface driven
by homeostatic pressure or by directed motility, acting
either at the interface (a "leader-cell" limit) or in bulk
[50] [51 -- 53]. Our results encompass also the cases of sta-
tionary interfaces maintained under constant cell renewal
[9], and of a single growing tissue [1].
We find a Saffman-Taylor-like instability involving sub-
strate friction, and a long-wavelength instability depen-
dent on an effective viscosity of cell number change. Bulk
motile forces induce an instability depending on their
strength and direction in each tissue. The free bound-
ary of a growing single tissue is generally stabilised by
the mechanisms studied here. Adding a driving noise
to represent, e.g., stochastic cell division, we calculate
the roughness and centre-of-mass diffusion of a stably-
propagating interface, and find that mechanical feedback
can help preserve reproducibility at the tissue scale [1].
We use a 2D hydrodynamic description in terms of
cell density and velocity fields. Tissues A and B cover
an infinite domain, meeting at a flat interface (Fig. 1A).
Since we assume a sharp interface, we state equations for
a general tissue, unless decorated with A or B.
We begin with continuity of the areal cell density, ρ,
∂tρ + ∂i(ρvi) = kdρ ,
where vi is the velocity field and
kd =
1
τ
ρd − ρ
ρd
(1)
(2)
is an expansion of the net division/death rate about the
homeostatic density ρd, with τ a characteristic timescale
[45]. The ρ units of each tissue are independent, so we set
ρd ≡ 1. We consider a linearised, isotropic elastic stress,
σij = σδij , σ ≡ σh − χ∆ρ ,
(3)
where σh is a tissue's homeostatic stress, χ its elastic
modulus and ∆ρ ≡ ρ − 1. The homeostatic pressure
imbalance is −∆σh ≡ −(σhA − σhB). The quantity χτ is
an effective bulk viscosity for cell number change [21]; on
a timescale τ , a tissue loses its elastic character as cells
are lost or created in response to elastic stress [32, 54].
Based on parameter estimates (see supplement [50]), we
neglect viscous stresses ∼ ¯η∂ivj, anticipating χτ ≫ ¯η
[47, 50] . Force balance expresses the tissue velocity as
vi = ξ−1(∂iσ + fi) ,
(4)
for substrate friction ξ and a density of active motil-
ity forces fi = δixf directed normal to the interface.
"Leader-cell" motility at the interface gives an effective
contribution to ∆σh [50]. We thus take f in Eq. 4 as
uniform in a given tissue, to account for "bulk" directed
motility forces, as may arise from cryptic lamellipodia
away from tissue edges [51 -- 53].
Moving steady state. We first solve for the steady
propagation of a flat interface (cf. Refs. [8, 47]). The
comoving coordinate is z ≡ x − V0t, with V0 the velocity
of the interface at z0 = 0, propagating in z. Assuming
driving forces small enough that nonlinear terms in ∆ρ,
v can be neglected in Eq. 1, we write
∂t∆ρ + ∂zvz + ∂yvy = −
1
τ
∆ρ .
(5)
where vz ≡ vx − V0. The resulting propagating steady
state is derived in the supplement [50] by matching the
tissues' stress and velocity at the interface. Density per-
turbations ∝ e±z/ℓ decay from the interface (Fig. 2) gov-
erned by each tissue's hydrodynamic length ℓ = pχτ /ξ.
Their sign (see Eq. S2 [50]) depends on −∆σh, and on
∆vf ≡ fA/ξA − fB/ξB, a difference in "bare" velocities
f /ξ associated to the bulk directed motilities.
In Fig.
2A, the growing tissue has decreased density so, by Eq.
2
FIG. 2. A) Steady state density perturbation profile ∆ρ
(solid line) and velocity vx (dashed) for tissues A (z < 0)
and B (z > 0). Parameters: homeostatic pressure differ-
ence −∆σh = 0.5χA,
frictions ξB = ξA, elastic moduli
χB = χA, division/death timescales τB = τA, bulk motil-
ity forces fA = fB = 0. B) As A, but with ∆σh = 0,
fA = 0.1χA/ℓA. C) As A, but with ∆σh = 0, fB = 0.1χA/ℓA.
D) As A, but with ∆σh = 0, fA = fB = 0.1χA/ℓA. In this
particular case the tissue moves uniformly and the density
perturbation cancels to zero.
2, is proliferative near the interface, while the shrinking
tissue has increased density so undergoes net apoptosis
near the interface. In Fig. 2B, both tissues are apoptotic
near the interface. The steady interface velocity,
V0 ≡ vxz=0 =
−∆σh + ℓAfA + ℓBfB
ξAℓA + ξBℓB
,
(6)
is, for fA = fB = 0 and ξB = ξA, that found in Ref.
[8]. To justify ignoring nonlinear terms, we require that
stresses from homeostatic pressure, motility and inter-
face line tension γ are small relative to the tissues' elas-
tic moduli: ∆σh ≪ χ, f /ξ ≪ ℓ/τ , γ/ℓ ≪ χ. Much
stronger stresses would lead to nonlinear responses and,
eventually, tissue rupture [47, 50, 55].
Interface stability.
In the supplement [50], starting
from Fourier and Laplace transforms y → q and t → s of
Eq. 5, we perturb the propagating steady state calculated
above, to find the fate of an interface fluctuation (Fig.
1B) δz0 = ǫ(t) cos(qy), where ǫ(0) = ǫ0. For line tension
γ ≥ 0 (e.g., increased myosin at heterotypic junctions
[10] or a supracellular actin cable [56]),
(σ + δσ)δz0,A − (σ + δσ)δz0,B = −γq2δz0 ,
(7)
where δσ is the deviation from stress σ of the propagat-
ing steady state. The dominant growth rate is denoted
s∗(q), with s∗ < 0, > 0 indicating stability or instability
(in the applicable parameter regime we do not find com-
plex poles, so treat s∗ as real [50]). Dispersion relations
3
FIG. 4. A) Root-mean-squared interface deviation for a
single tissue using example physical parameters (Eq. 12).
The homeostatic pressure/leader-cell motility parameter σh
is varied, and the arrow shows increasing friction ξ =
103, 104, 105 Pa s/µm. Other parameters: χ = 104 Pa µm,
τ = 3.6 × 104 s, γ = 103 pN, L⊥ = 1000 µm, D = 0.4 µm2/h.
B) Interface centre-of-mass diffusion coefficient for a single
tissue (Eq. 14) using example physical parameters. The fric-
tion ξ is varied, and the arrow shows increasing elastic mod-
ulus χ = 103, 104, 105 Pa µm. Other parameters: σh = 0,
τ = 3.6 × 104 s, γ = 103 pN, L⊥ = 1000 µm, D = 0.4 µm2/h.
and
− ∆σh >
2γ(ξAℓA + ξBℓB)
χBτB − χAτA
, χBτB > χAτA .
(9)
where Eq. 8 is approximate [50]. Two types of transition
arise. Fig. 3A and Eq. 8 show a "type Is" transition in the
Cross-Hohenberg classification [57], where an intermedi-
ate band qmin < q < qmax , qmin 6= 0 becomes unstable
("s" indicates that the instabilities found are stationary,
not oscillatory). Fig. S2A [50] and Eq. 9 show a "type
IIs" transition, with unstable band 0 < q < qmax and on-
set at q → 0. Then, one expects near threshold that the
characteristic wavelength scales with system size. Eqs. 8,
9 are combined with phase diagrams of q∗ in Fig. 3B,C.
Interface driven by bulk directed motility. We now
consider the case ∆σh = 0, with bulk directed motil-
ity forces fA 6= 0, fB 6= 0.
Instability occurs when
∆vf ≡ fA/ξA − fB/ξB [50] satisfies
∆vf >
2γ(ℓAξA + ℓBξB)
ξAξBℓAℓB(ℓA + ℓB)
.
(10)
Fig. S2B [50] shows dispersion relations crossing the type
IIs transition of Eq. 10. The phase diagram in Fig. 3D
shows that a static interface (V0 = 0), marginally stable
for f = 0 [50], can be stable or unstable depending on
the direction of fA, fB.
Single tissue. The free boundary of a growing single
tissue (e.g., epithelium invading empty substrate [2]), is
stabilised by the mechanisms studied here (Eqs. S20 -- S22
[50]). Protrusion formation is often observed in wound-
healing, via a number of proposed mechanisms we have
not included [3, 24, 25, 58].
A stable interface subject to noise.
Interface propa-
gation and maintenance takes place in the presence of
FIG. 3. A) Example numerically-determined dispersion re-
lations. Parameters: frictions ξB = 1.5ξA, moduli χB =
0.5χA, division/death timescales τB = τA, bulk motility
forces fA = fB = 0,
line tension γ = 0.001ℓAχA. The
homeostatic pressure imbalance/leader-cell motility parame-
ter varies, ∆σh = −0.1χA, −0.2χA, . . . , −0.5χA, in the direc-
tion of the arrow, crossing a "type Is" instability transition
[57]. B) Phase diagram in (∆σh, ξB) of the most unstable
wavenumber q∗ (white if no instability), using χB = 0.5χA,
τB = τA, fA = fB = 0, γ = 0.001ℓAχA. The dashed line
indicates the approximation of the type Is transition line by
Eq. 8. C) Phase diagram in (χB, ξB) using ∆σh = −0.5χA,
τB = τA, fA = fB = 0, γ = 0.001ℓAχA. The meaning of
the dashed line is as in C, whereas the solid line indicates the
type IIs transition (Eq. 9). D) Phase diagram in bulk directed
motilities (fA, fB), with other parameters ξB = ξA, χB = χA,
τB = τA, γ = 0.001ℓAχA. The black line is the transition ap-
proximated by Eq. 10. The dotted line is V0 = 0, with the
upper half-space V0 > 0 (tissue A growing) and the lower
V0 < 0. In each quadrant, cartoons illustrate the direction of
the bulk motilities.
(e.g., Fig. 3A) maximised over q yield phase diagrams
(Fig. 3B,C,D) of the most-unstable wavenumber q∗. We
approximate the dispersion relation in limits of q [50] to
find the analytic criteria discussed below.
Interface driven by homeostatic pressure or leader-cell
motility. We first discuss growth of tissue A driven by
∆σh < 0 (Fig. 3A,B,C), without bulk directed motility
(fA = fB = 0). Analytic dispersion relations [50] show
that, for strong enough ∆σh, the interface is unstable if
tissue B has greater friction ξB > ξA or effective viscosity
χBτB > χAτA. Instability criteria (given ∆σh < 0) are
−∆σh &
27
4
γ
ξB > ξA ,
(ξAℓB − ξBℓA)2(ξAℓA + ξBℓB)
Aℓ2
ℓ2
B(ξB − ξA)3
,
(8)
stochasticity in cell divisions, motilities, material param-
eters, etc. In the supplement [50] we model this with i) a
driving noise in Eq. 5, ∂t∆ρ+∂zvz +∂yvy = −(1/τ )∆ρ+k
where hk(z, y, t)k(z ′, y′, t′)i = DA(B)δ(z−z ′)δ(y−y′)δ(t−
t′), corresponding to a contribution of random cell divi-
sion; or (ii) a noisy motile force contribution to Eq. 4,
hfi(z, y, t)fj(z, y, t)i = Df δ(y − y′)δ(z − z ′)δ(t − t′)δij.
Noisy motility could arise, e.g., from 'swirling' patterns
[3], provided that the correlation length of these patterns
is small compared to other length scales discussed. We
focus here on cell division noise, but find qualitatively
similar results for noise on the motile force [50].
Excluding the q = 0 mode (discussed below), we cal-
culate the correlation function hδz0(y + y′, t + t′)δz0(y, t)i
of an interface in the stable parameter regime. The satu-
ration (late-time) roughness as the system size in y, L⊥,
becomes large, is
hδz0(y, t)2i ∼
L⊥(ξ2
Aℓ3
ADA + ξ2
Bℓ3
BDB)
2π2N
,
N ≡ 2(ξAℓA + ξBℓB)γ − (χAτA − χBτB)∆σh
− ξAξBℓAℓB(ℓA + ℓB)∆vf .
(11)
The dependence on L⊥ is as in 1-dimensional Edwards-
Wilkinson deposition [59].
The positive denomina-
tor N is expanded in the applicable regime of small
∆vf , γ, ∆σh. Roughness can be reduced by three now-
familiar mechanisms: line tension γ > 0; stabilising ef-
fective viscosity contrast χAτA > χBτB, ∆σh < 0; sta-
bilising bulk motilities ∆vf < 0. For identical tissues
without line tension, the "interface" is an arbitrary line
in the tissue: Eq. 11 then diverges, i.e., the roughness
grows indefinitely.
Identical tissues with line tension
yield hδz2
0i ∼ L⊥χτ D/(4π2γ), so that mechanical regu-
lation (i.e., smaller τ ) reduces boundary roughness. This
is true also for a single tissue,
hδz2
0i ∼ L⊥ξ2ℓ3D/ (cid:0)2π2(2ξℓγ − χτ σh)(cid:1) ,
(12)
where if the tissue is growing (σh < 0) the roughness is
decreased. Fig. 4A shows this behaviour quantitatively
for estimated physical values of the parameters [50].
The q = 0 mode leads to an effective diffusion coeffi-
cient for the interface centre-of-mass,
¯D ∼
1
L⊥
BDB + ξ2
Bℓ3
ξ2
4(ξAℓA + ξBℓB)2
Aℓ3
ADA
,
(13)
and for a single tissue,
¯D ∼
ℓD
4L⊥
.
(14)
The behaviour of Eq. 14 for varying friction ξ and elastic
modulus χ is shown in Fig. 4B. These equations control
the accumulating uncertainty in tissue size as the inter-
face centre-of-mass progressively diffuses away from its
noise-free trajectory. A larger friction coefficient leads to
more precise growth (Fig. 4B) but decreases the velocity
(Eq. 6), which suggests trade-offs might be necessary to
optimise the speed and precision of growth.
4
evidence
experimental
Discussion. Given
of
mechanically-regulated cell number change [11, 35 --
43], models of the type used here are widely studied
[8, 31, 45 -- 47]. There is much interest in mechanisms
of boundary maintenance between cell populations
[10, 56, 60]. Recent simulations showed how the topol-
ogy of cellular interactions can stabilise anomalously
smooth interfaces [13], while experiments suggest that
interface maintenance is not only local but is connected
to mechanical waves and jamming processes deep within
neighbouring tissues [61]. Our results add to this pic-
ture, showing that mechanically-regulated cell number
change within in the tissue bulk can exert an important
influence on the properties of interfaces. We have shown
how the forces driving overall interface propagation can
also generate instabilities, and affect the response of
interfaces to cell-level stochasticity.
A Saffman-Taylor-like instability based on substrate
friction [30, 31] (Eqs. 8, S17 [50]) accords with the tumour
literature, where tissues with weaker cell-matrix adhe-
sions tend to be more invasive [62]. A longer-wavelength
instability (Eq. 9) occurs if the effective bulk viscosity
for cell number change is smallest in the growing tissue.
Cell-based simulations [8] could explore our predictions,
which could in turn be extended to include, e.g., cell
growth anisotropy, proposed to play a role in the stable
interfaces found in Ref. [8].
The effects of bulk directed motility depend on ∆vf
(Fig. 3D). Repulsive migration, known to occur due to
Eph and ephrin signalling [14, 60], should yield ∆vf < 0,
favouring stability. In Drosophila abdominal epidermis,
larval epithelial cells being replaced by histoblasts are
proposed to actively migrate away from the propagating
interface [34]. This motility force would promote sta-
bility, presumably desirable to ensure reproducible, well-
controlled tissue replacement. This Drosophila system,
or model experiments [61], could be used to test our the-
ory by perturbing, e.g., motility, substrate friction, or
cell division, and observing the effect on interfaces.
We found that mechanical feedback can help to smooth
a stable interface in the presence of noise, as well as deter-
mining how quickly the interface centre-of-mass diffuses
away from its noise-free position. These findings are rel-
evant to the question of how tissue-level reproducibility
is achieved despite cell-level stochasticity [1].
We thank Anna Ainslie, John Robert Davis, Feder-
ica Mangione and Nic Tapon for discussions. J. J. W.
acknowledges support by a Wellcome Trust Investiga-
tor award to Dr Nic Tapon (107885/Z/15/Z), and ac-
knowledges discussions with Ruth Curtain, Claire McIl-
roy, Pasha Tabatabai and Ansgar Trachtler. G. S. and
J. J. W. acknowledge support by the Francis Crick Insti-
tute which receives its core funding from Cancer Research
UK (FC001317, FC001175), the UK Medical Research
Council (FC001317, FC001175), and the Wellcome Trust
(FC001317, FC001175).
[1] L. Hong, M. Dumond, S. Tsugawa, A. Sapala, A.-L.
Routier-Kierzkowska, Y. Zhou, C. Chen, A. Kiss, M. Zhu,
O. Hamant, R. S. Smith, T. Komatsuzaki, C.-B. Li,
A. Boudaoud, and A. H. K. Roeder, Developmental Cell
38, 15 (2016).
[2] A. Ravasio, I. Cheddadi, T. Chen, T. Pereira, H. T.
Ong, C. Bertocchi, A. Brugues, A. Jacinto, A. J. Kabla,
Y. Toyama, X. Trepat, N. Gov, L. Neves de Almeida, and
B. Ladoux, Nature Communications 6, 7683 EP (2015).
[3] M. Basan, J. Elgeti, E. Hannezo, W. J. Rappel, and
H. Levine, Proceedings of the National Academy of Sci-
ences 110, 2452 (2013).
[4] N. E. Baker, Curr. Biol. 21, R11 (2011).
[5] N. Ninov and E. Martin-Blanco, Nature Protocols 2, 3074
(2007).
[6] J.-P. Vincent, A. G. Fletcher, and L. A. Baena-Lopez,
Nature Reviews Molecular Cell Biology 14, 581 (2013).
[7] R. Gogna, K. Shee, and E. Moreno, Annual review of
genetics 49, 697 (2015).
[8] N. Podewitz, F. Julicher, G. Gompper, and J. Elgeti,
New Journal of Physics 18, 083020 (2016).
[9] A. Marianes and A. C. Spradling, eLife 2, e00886 (2013).
[10] K. P. Landsberg, R. Farhadifar, J. Ranft, D. Umetsu,
and
T. J. Widmann, T. Bittig, A. Said, F. Julicher,
C. Dahmann, Current Biology 19, 1950 (2009).
[11] M.
D.
de
la
Loza
and
B.
Thompson,
Mechanisms of Development 144, Part A, 23 (2017).
[12] K. Curtius, N. A. Wright, and T. A. Graham, Nature
Reviews Cancer 18, 19 (2017).
[13] D. M. Sussman, J. M. Schwarz, M. C. Marchetti, and
M. L. Manning, Phys. Rev. Lett. 120, 058001 (2018).
[14] J. Cayuso, Q. Xu, and D. G. Wilkinson, Developmental
Biology 401, 122 (2015).
[15] C. Bielmeier, S. Alt, V. Weichselberger, M. La Fortezza,
H. Harz, F. Julicher, G. Salbreux, and A.-K. Classen,
Current Biology 26, 563 (2016).
[16] P. Tracqui, Reports on Progress in Physics 72, 056701
(2009).
[17] V. Cristini, H. B. Frieboes, R. Gatenby, S. Caserta,
M. Ferrari, and J. Sinek, Clinical Cancer Research 11,
6772 (2005).
[18] N. J. Poplawski, A. Shirinifard, U. Agero, J. S. Gens,
M. Swat, and J. A. Glazier, PLOS ONE 5, 1 (2010).
[19] P. Ciarletta, L. Foret, and M. Ben Amar, Journal of
Royal Society Interface 8, 345 (2011).
[20] M. Ben Amar, C. Chatelain, and P. Ciarletta, Phys. Rev.
Lett. 106, 148101 (2011).
[21] S.
R.
Lubkin
and
J.
D.
Murray,
Journal of Mathematical Biology 34, 77 (1995).
[22] T. Miura, The Journal of Biochemistry 157, 121 (2015).
[23] P. V. Bayly, R. J. Okamoto, G. Xu, Y. Shi, and L. A.
Taber, Physical Biology 10, 016005 (2013).
[24] J. Zimmermann, M. Basan, and H. Levine, The Euro-
pean Physical Journal Special Topics 223, 1259 (2014).
[25] V. Tarle, A. Ravasio, V. Hakim, and N. S. Gov, Integr.
Biol. 7, 1218 (2015).
[26] M. Castro, C. Molina-Par´ıs, and T. S. Deisboeck, Phys.
Rev. E 72, 041907 (2005).
5
[27] W. W.
F.
Journal of Applied Physics 34, 323 (1963).
Mullins
and
R.
Sekerka,
[28] M. Basan, J.-F. Joanny, J. Prost, and T. Risler, Phys.
Rev. Lett. 106, 158101 (2011).
[29] T. Risler and M. Basan, New Journal of Physics 15,
065011 (2013).
[30] D. Drasdo and S. Hoehme, New Journal of Physics
(2012).
[31] T. Lorenzi, A. Lorz, and B. Perthame, Kinetic and Re-
lated Models 10, 299 (2017).
[32] T. Risler, A. Peilloux, and J. Prost, Phys. Rev. Lett.
115, 258104 (2015).
[33] N. Ninov, S. Menezes-Cabral, C. Prat-Rojo, C. Manj´on,
A. Weiss, G. Pyrowolakis, M. Affolter, and E. MartIn-
Blanco, Current Biology 20, 513 (2010).
[34] M. Bischoff, Developmental Biology 363, 179 (2012).
[35] F. Montel, M. Delarue, J. Elgeti, L. Malaquin, M. Basan,
T. Risler, B. Cabane, D. Vignjevic, J. Prost, G. Cappello,
and J.-F. Joanny, Phys. Rev. Lett. 107, 188102 (2011).
Hoerner,
and L. Hufnagel,
[36] S.
T.
Proceedings of the National Academy of Sciences 111, 5586 (2014).
Schneidt, D. Holzer,
Streichan,
R.
C.
J.
[37] A. Puliafito, L. Hufnagel, P. Neveu, S. Streichan, A. Si-
gal, D. K. Fygenson, and B. I. Shraiman, Proceedings
of the National Academy of Sciences 109, 739 (2012).
[38] B. W. Benham-Pyle, B. L. Pruitt, and W. J. Nelson,
Science 348, 1024 (2015).
[39] T. Aegerter-Wilmsen, C. M. Aegerter, E. Hafen, and
K. Basler, Mechanisms of Development 124, 318 (2007).
[40] Y. Pan, I. Heemskerk, C. Ibar, B. I. Shraiman, and K. D.
Irvine, Proceedings of the National Academy of Sciences
113, E6974 (2016).
[41] R. Levayer, C. Dupont, and E. Moreno, Current Biology
26, 670 (2016).
[42] S. A. Gudipaty, J. Lindblom, P. D. Loftus, M. J. Redd,
K. Edes, C. F. Davey, V. Krishnegowda, and J. Rosen-
blatt, Nature (2017).
[43] G. C. Fletcher, M.-d.-C. Diaz-de-la Loza, N. Borreguero-
and B. J.
Munoz, M. Holder, M. Aguilar-Aragon,
Thompson, Development (2018), 10.1242/dev.159467.
[44] L. LeGoff and T. Lecuit, Cold Spring Harbor perspectives
in biology 8, a019232 (2016).
[45] M. Basan, T. Risler, J.-F. Joanny, X. Sastre-Garau, and
J. Prost, HFSP Journal 3, 265 (2009).
[46] J. Ranft, M. Aliee, J. Prost, F. Julicher,
and J.-F.
Joanny, New Journal of Physics , 1 (2014).
[47] P. Recho, J. Ranft, and P. Marcq, Soft Matter 12, 2381
(2016).
[48] O. Cochet-Escartin, J. Ranft, P. Silberzan,
and
P. Marcq, Biophys. J. 106, 65 (2014).
Poujade,
[49] M.
A.
B. Ladoux, A. Buguin,
Proceedings of the National Academy of Sciences 104, 15988 (2007).
Grasland-Mongrain,
Chavrier,
Silberzan,
P.
and P.
Jouanneau,
Hertzog,
E.
J.
[50] See Supplemental Material for further theoretical calcu-
lations, estimates of parameters, and Refs. [63 -- 71].
[51] R. Farooqui and G. Fenteany, J. Cell Sci. 118, 51 (2004).
[52] X. Trepat, M. R. Wasserman, T. E. Angelini, E. Millet,
D. A. Weitz, J. P. Butler, and J. J. Fredberg, Nature
Physics 5, 426 (2009).
[53] E. Scarpa and R. Mayor, The Journal of Cell Biology
212, 143 (2016).
[54] J. Ranft, M. Basan, J. Elgeti, J.-F. Joanny, J. Prost, and
[63] P.
G.
Saffman
and
G.
Taylor,
F. Julicher, Proc. Natl. Acad. Sci. 107, 20863 (2010).
Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences 245, 312 (1958).
[55] A. R. Harris, L. Peter, J. Bellis, B. Baum, A. J. Kabla,
[64] S.
F.
Edwards
and
D.
R. Wilkinson,
and G. T. Charras, PNAS 109, 16449 (2012).
Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences 381, 17 (1982).
6
[56] P. Hayes and J. Solon, Mechanisms of Development 144,
2 (2017).
C.
[57] M.
P.
Rev. Mod. Phys. 65, 851 (1993).
Cross
and
C.
Hohenberg,
[58] S. Mark, R. Shlomovitz, N. S. Gov, M. Poujade,
and P. Silberzan, Biophysical
E. Grasland-Mongrain,
Journal 98, 361 (2010).
[59] T. Antal and Z. R´acz, Physical Review E (Statistical
Physics 54, 2256 (1996).
[60] H. B. Taylor, A. Khuong, Z. Wu, Q. Xu, R. Morley,
L. Gregory, A. Poliakov, W. R. Taylor, and D. G. Wilkin-
son, Journal of The Royal Society Interface 14, 20170338
(2017).
[61] P. Rodr´ıguez-Franco, A. Brugues, A. Mar´ın-Llaurad´o,
V. Conte, G. Solanas, E. Batlle, J. J. Fredberg, P. Roca-
Cusachs, R. Sunyer, and X. Trepat, Nature Materials
16, 1029 (2017).
[62] H. B. Frieboes, Cancer Research 66, 1597 (2006).
[65] C. Blanch-Mercader, R. Vincent, E. Bazelli`eres, X. Serra-
Picamal, X. Trepat, and J. Casademunt, Soft Matter 13,
1235 (2017).
[66] I. Bonnet, P. Marcq, F. Bosveld, L. Fetler, Y. Bellaiche,
and F. Graner, Journal of The Royal Society Interface 9,
2614 (2012).
Bambardekar,
[67] K.
C.
Proceedings of the National Academy of Sciences 112, 1416 (2015).
Blanc,
Lenne,
Cl´ement,
Chards,
P.-F.
and
R.
O.
[68] N. Sep´ulveda, L. Petitjean, O. Cochet, E. Grasland-
Mongrain, P. Silberzan, and V. Hakim, PLoS Computa-
tional Biology 9, e1002944 (2013).
[69] R. Curtain and K. Morris, Automatica 45, 1101 (2009).
(2016),
[70] A.
e-prints
Trachtler,
ArXiv
arXiv:1603.01059 [math.DS].
[71] A. Trachtler, private communication.
|
1709.02183 | 1 | 1709 | 2017-09-07T11:29:36 | Catch Bonding in the Forced Dissociation of a Polymer Endpoint | [
"physics.bio-ph",
"cond-mat.soft"
] | Applying a force to certain supramolecular bonds may initially stabilize them, manifested by a lower dissociation rate. We show that this behavior, known as catch bonding and by now broadly reported in numerous biophysics bonds, is generically expected when either or both the trapping potential and the force applied to the bond possess some degree of nonlinearity. We enumerate possible scenarios, and for each identify the possibility and, if applicable, the criterion for catch bonding to occur. The effect is robustly predicted by Kramers theory, Mean First Passage Time theory, and finally confirmed in direct MD simulation. Among the catch scenarios, one plays out essentially any time the force on the bond originates in a polymeric object, implying that some degree of catch bond behavior is to be expected in {\em any} protein-protein bond, as well as in more general settings relevant to polymer network mechanics or optical tweezer experiments. | physics.bio-ph | physics | Catch Bonding in the Forced Dissociation of a Polymer Endpoint
Department of Applied Physics and Institute for Complex Molecular Systems, Eindhoven University of Technology,
Cyril Vrusch and Cornelis Storm
P.O. Box 513, NL-5600 MB Eindhoven, The Netherlands
(Dated: August 6, 2018)
Applying a force to certain supramolecular bonds may initially stabilize them, manifested by a
lower dissociation rate. We show that this behavior, known as catch bonding and by now broadly
reported in numerous biophysics bonds, is generically expected when either or both the trapping
potential and the force applied to the bond possess some degree of nonlinearity. We enumerate
possible scenarios, and for each identify the possibility and, if applicable, the criterion for catch
bonding to occur. The effect is robustly predicted by Kramers theory, Mean First Passage Time
theory, and finally confirmed in direct MD simulation. Among the catch scenarios, one plays out
essentially any time the force on the bond originates in a polymeric object, implying that some
degree of catch bond behavior is to be expected in any protein-protein bond, as well as in more
general settings relevant to polymer network mechanics or optical tweezer experiments.
PACS numbers: 87.15.A-,V87.15.Fh,05.40.-a
How long does it take a fluctuating particle to escape
the trap of a confining potential well? The question is one
of the staples of statistical mechanics and, in its simplest
incarnation, gives rise to Kramers' well-known expressions
for the rate at which a particle crosses a potential energy
barrier -- the rate exponentially decaying with increasing
barrier height [1]. This escape problem features in a wide
range of problems in statistical mechanics, and has impor-
tant applications and consequences in materials science,
soft matter- and biological physics for its capacity to pre-
dict, under general external conditions, the dissociation
rate of bound states and, thereby, the mean lifetime (cid:104)τ(cid:105) of
bonds between, for instance, receptor-ligand pairs [2 -- 4].
Escape kinetics change when forces are taken into con-
sideration. Generally, an applied force aligning with the
escape path will hasten dissociation; the naive Kramers
prediction is, again, an exponential decrease in lifetime
with increasing force [1]. As [5] showed, forced escape
scenarios become considerably richer when the energy
landscape is multidimensional -- in particular, they note
the curious possibility of the lifetime of the bound state
initially increasing with the applied force, a phenomenon
they termed rollover. This counter-intuitive behavior --
a bond is strengthened by applying a force to it -- is no
longer a theoretical curiosity but has, in recent years, been
experimentally demonstrated in a range of non-covalent
biophysical bonds where it has become widely known
as a catch bond [6, 7]. For protein-protein bonds, such
behavior is generally ascribed to specific conformational
properties of the molecules involved [8 -- 12]; we will show
that it is generic. In this paper, we will define a catch
bond to mean a bound state whose average lifetime (cid:104)τ(cid:105)(f )
possesses an initial regime of increase with force:
(cid:12)(cid:12)(cid:12)(cid:12)f =0
(cid:104)τ(cid:105)(f )
d
df
> 0 .
(1)
What we will show, is that this behavior is generic when
FIG. 1:
(a) Illustration of the bond, the bead (red) is trapped
in an potential well while an external potential acts on it. (b)
Scenarios for the energy landscapes with corresponding catch
bond criteria.
the applied force is non-linear (i.e., depends asymmetri-
cally on the position of the particle in the trap), even in
one-dimensional energy landscapes, and moreover even
in symmetric one-dimensional potentials. Our findings
demonstrate the broad generality, without the need for
7
1
0
2
p
e
S
7
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
3
8
1
2
0
.
9
0
7
1
:
v
i
X
r
a
further assumptions, of the implication of entropic elastic-
ity in catch bonding [13]. We also show, that this scenario
plays out essentially any time the force is transmitted to
the escaping particle by a polymeric object, and thus that
some regime of catch bond behavior is to be expected in
any protein-protein bond, as well as more general settings
relevant to polymer network mechanics or optical tweezer
experiments.
Our paper is organized as follows: First, we outline the
general framework of a forced one-dimensional escape pro-
cess in the Kramers sense (with rates depending only on
barrier height) and summarize the various trapping and
forcing scenarios that give rise to catch bonding. Next,
we specify to the case of a particle that is pulled out of
a confining trap by a polymer under tension, and study
the full differential equation determining the mean first
passage time (MFPT) for a general asymmetric potential.
Our explicit analytical solution proves, that a nonlinear
force-extension relation for the polymer is a necessary
condition for catch bonding to occur. We conclude our
paper with a Molecular Dynamics (MD) simulation of
the escape of a tensed semiflexible polymer with one end
trapped in a symmetric, harmonic potential well show-
ing significant (approximately 10% increase in unbinding
time) catch bonding, while also highlighting some of the
more subtle features of polymer unbinding.
Kramers Catch Bonding Scenarios An intuitive model
to rationalize equilibrium catch bond behavior is the
so-called Two-Pathway Model (TPM), developed by
Pereverzev et al.[14]. A particle is confined to a local
minimum of the trap's potential energy Ut(x), where x
is the position of the particle in the trap. This local
minimum Ut = 0 (the bound state) is flanked by two
different energy barriers. In any trap configuration, there
are two escape paths; one left (L) and one right (R). By
applying a force f that is orthogonal to the pathway with
the lowest barrier, however, the effective barrier height
in the secondary direction is lowered and escape along
the second unbinding pathway becomes increasingly likely
and frequent. Provided the dissociation lengths of the two
pathways are suitably chosen, the same applied force may
simultaneously increase the barrier height of the original
pathway, making it less likely. The combined escape prob-
lem over the two barriers determines the overall escape
time, and the differential response to the applied force
may result in an increase in the average unbinding time as
a force is applied, that is, catch bond behavior. As such,
catch bonding in the TPM is a result of an asymmetric
trap, in combination with a linear forcing.
We generalize the one-dimensional TPM by making no
assumptions on the energy barriers, but only assuming
that there are two pathways by which the bond can unbind.
We will continue to call those pathways left (L) and right
(R), and shall call the dissociation lengths to the left
and right bL and bR, respectively. The bare escape rates
corresponding to the pathways are called k0
R, and
L and k0
in the Kramers picture are determined by the barrier
heights presented by the trap potential only
2
i = ν e−βUt(bi),
k0
(i = L, R),
(2)
with ν the attempt frequency and β = (kBT )−1. For
two otherwise independent processes, we thus expect the
combined rate of escape to be k0 = k0
R. These rates
define the survival probabilities SL(t) and SR(t) -- the
probabilities that unbinding did not occur along the L or
R pathway prior to time t -- as
L + k0
SL(t) = e−k0
Lt ; SR(t) = e−k0
Rt ,
(3)
and permit to compute πL(t) dt (πR(t) dt), the likelihood
that the first unbinding event occurs along L (R) between
t and t + dt, as the joint probability of unbinding along L
(R) between those times, and having survived unbinding
along R (L) up until t:
πL(t) = − dSL(t)
dt
· SR(t) ; πR(t) = − dSR(t)
dt
· SL(t) . (4)
The asymptotic probabilities PL and PR of unbinding
along either direction are then given by
(cid:90) ∞
(cid:90) ∞
0
0
PL =
PR =
πL(t) dt =
πR(t) dt =
k0
L
k0
L + k0
R
k0
R
k0
L + k0
R
,
.
(5)
(6)
This makes intuitive sense; if one of the two rates is very
high compared to the other, fast unbinding preempts dis-
sociation along the slow pathway. It is, however, also
instructive to consider the average unbinding times, sepa-
rately for each of the two directions. The mean unbinding
times along L and R are computed as
(cid:90) ∞
(cid:90) ∞
0
0
t
t
L(cid:105) =
(cid:104)τ 0
(cid:104)τ 0
R(cid:105) =
(cid:18) πL(t)
(cid:18) πR(t)
PL
(cid:19)
(cid:19)
PR
dt =
dt =
1
k0
L + k0
R
1
k0
L + k0
R
,
,
(7)
(8)
where the factors PL and PR are included to normalize the
L and R unbinding distributions; (cid:104)τL(cid:105) and (cid:104)τR(cid:105) are thus
the mean unbinding times of each path, averaged over only
the unbinding events in that particular direction. This
reveals two interesting properties: the presence of a second
unbinding path always decreases the average unbinding
lifetime along a primary path (the unbinding is, of course,
as least as fast as it was originally) but the average time
it takes to escape along either direction becomes the same.
Counter-intuitively, perhaps, the escape along a very slow
direction is greatly sped up by the presence of a secondary,
fast unbinding route; this is because only the rare, very
quick event along the slow path is not precluded by the
abundant, fast rebinding. The overall number of such
events may become very small, though, as borne out by
PL and PR. Averaging the escape time over all events,
both L and R, gives the expected result mean lifetime
(cid:104)τ(cid:105) of the bond
(cid:104)τ 0(cid:105) = PL(cid:104)τ 0
L(cid:105) + PR(cid:104)τ 0
R(cid:105) =
1
k0
L + k0
R
=
1
k0 .
(9)
Thus, we find that due to the presence of the other path-
way, each path acquires the same escape time. This
escape time, moreover, is equal to the overall lifetime of
the bound state
(cid:104)τ 0(cid:105) = (cid:104)τ 0
L(cid:105) = (cid:104)τ 0
R(cid:105) .
(10)
How do external manipulations, such as the application
of a force, affect (cid:104)τ(cid:105)? Let us consider now the case in which
the unbinding process takes place in the presence of an ad-
ditional, external potential Uext(x, ε) which defines a (pos-
sibly position-dependent) force fext(x, ε) = −dUext/dx.
The parameter ε quantifies some manner of tuning this
external potential; in the simplest case it is the force
itself (Uext(x, ε) = −εx), but it may also represent some
more general way of altering the externally applied poten-
tial. Assuming that the rate at which the particle crosses
the boundaries bR or bL of the trap still depends on the
barrier height only (we will examine the validity of this
assumption in the latter part of this article), Uext(x, ε)
shifts the transition rates according to
ki = k0
i e−β Uext(bi,ε),
(i = L, R).
(11)
To assess the possibility of catch bonding, we compute
the change in the overall lifetime (cid:104)τ(cid:105)
(cid:104)τ(cid:105)(ε) = (cid:104)τ(cid:105)(ε = 0) + (cid:104)∆τ(cid:105)(ε) ,
(12)
due to a small change in ε. This is done by substituting
Eq. (11) into Eq. (9) and expanding to lowest (first)
order in ε, to yield
(cid:19)
×
(cid:18)
(cid:104)∆τ(cid:105)(ε) = βε
(cid:32)
×
eβ(Uext(bL,0)+Uext(bR,0))
(eβUext(bR,0)k0
L + eβUext(bL,0)k0
∂Uext
R)2
+
(cid:34)
(cid:34)
∂ε
∂Uext
∂ε
(cid:35)
(cid:12)(cid:12)(cid:12)(cid:12)(bL,0)
(cid:35)(cid:33)
(cid:12)(cid:12)(cid:12)(cid:12)(bR,0)
+ O(ε2) .
eβUext(bR,0)k0
L
+ eβUext(bL,0)k0
R
3
Note, first, that the left hand side is entirely set by the trap
potential Ut(x), while the right hand side is determined
Uext is shorthand for the ε-derivative of
only by Uext(x).
Uext, which is to be taken before ε is set to zero. Either
Ut or Uext, or indeed both, may be used to create a catch
bond effect. This equation thus encodes a number of
scenarios as illustrated in Fig.1:
(i): Symmetric trap, constant external force. Ut(x) is
symmetric (that is, Ut(x) = Ut(−x)), and Uext(x, ε) =
−εx (that is, the external force f = ε is constant for given
ε, and points in the R-direction throughout the trap region
[bL, bR]). In this case, because of the symmetry of Ut it
must be true that bL = −bR, as well as k0
R. Since
Uext(x, 0) = 0, and the ε-derivatives at the boundaries
are equal but opposite, the catch criterion cannot be met;
both sides of Eq. (14) are 1.
(ii): Asymmetric trap, constant external force. Ut(x) (cid:54)=
Ut(−x), and Uext(x) = −εx in [bL, bR]. An asymmetry
in Ut may be relevant in two different ways; either (iia)
the dissociation lengths bL and bR to either side are equal
in absolute magnitude, but Ut(bL) (cid:54)= Ut(bR), that is,
the barrier heights are different; or (iib) the dissociation
lengths differ (bL (cid:54)= bR) while the trap barrier heights
In case (iia), calling
are the same Ut(bL) = Ut(bR).
the dissociation length −bL = bR = b, the criterion for
increased lifetime reduces to
L = k0
(iia) :
k0
L
k0
R
> 1 .
(15)
This is the TPM catch bonding scenario: if an external
force opposing escape along the initially favored (most
frequent) pathway is applied, the overall escape rate in-
creases. However, also in case (iib) catch bonding may
occur; in this case, the criterion reduces to
(iib) :
1 > − bR
bL
.
(16)
Thus, even for identical barrier heights applying a pulling
force in the direction where the dissociation length is
shortest will increase the overall lifetime of the bond. Ob-
viously, suitably chosen combinations of trap asymmetries
(iia) and (iib) also produce catch bonding.
(iii): Symmetric trap, harmonic external potential. For
symmetric Ut(x), and thus again −bL = bR = b and
k0
L = k0
R, the left hand side of Eq. (14) is equal to one.
Suppose now, that the external potential is supplied by
a harmonic spring of rest length (cid:96)0 and spring constant
κ(2), whose left end point is attached to the particle in
the trap, and whose right end is at x = (cid:96)0(1 + ε). ε thus
measures the extension of the spring and relates to the
tension in the spring. In that case, the external potential
(with x = 0 the center of the trap) is given by
(13)
The essential information in this cumbersome expression
may be condensed somewhat. The first factor, between
brackets, is strictly positive and does not affect the sign
of (cid:104)∆τ(cid:105)(ε). For (cid:104)∆τ(cid:105) to increase with increasing ε, the
following 'catch criterion' must hold
(cid:18) eβUext(bL,0)
(cid:19)(cid:32) Uext(bR, 0)
(cid:33)
> −
k0
L
k0
R
eβUext(bR,0)
Uext(bL, 0)
.
(14)
Uext(x, ε) =
κ(2)
2
(ε(cid:96)0 − x)2 .
(17)
4
Straightforward substitution into Eq. (14) yields that in
this case, too, the catch criterion cannot be met, as both
the left- and right hand sides of Eq. (14) are 1. Moreover,
expanding the change in lifetime to second order in ε
reveals that (cid:104)∆τ(cid:105)(ε) ∼ −ε2, and thus suggests that the
lifetime at zero displacement (i.e., zero force applied to
the spring) is a maximum.
(iv): Symmetric trap, anharmonic external potential.
We may also consider the same trap conditions as case
(iii), but with an anharmonic, nonlinear external potential
(we will use a third power here, but the argument is valid
with the inclusion of arbitrary odd powers)
Uext(x, ε) =
κ(2)
2
(ε(cid:96)0 − x)2 +
κ(3)
3
(ε(cid:96)0 − x)3 .
(18)
In this case, the catch criterion may be cast as
(cid:18) κ(2) − bκ(3)
(cid:19)
κ(2) + bκ(3)
2
3 βb3κ(3)
e
< 1 ,
(19)
demonstrating that for suitably chosen nonlinearity κ(3)
catch bonding is possible. In particular, one recognizes
the limits κ(3) → 0 (recovering case (iii)) to preclude
catch bonding, whereas κ(2) → 0 will always give catch
bonding, for all positive values of κ(3).
Summarizing, a range of configurations of external forc-
ing and trap properties give rise to an increase in lifetime
with applied force (or with the extension of a springlike
tether to which the escaping particle is attached).
In
particular, we show here that such a regime is generic
for nonlinearly elastic tethers. Examples of such are all
polymers, flexible and semiflexible. Most proteins also
display highly nonlinear force-extension relationships, and
our result suggests a route towards catch binding that
is not due to any specific allosteric or conformational
mechanism, but rather is encoded within the universal
departures from nonlinearity in protein mechanics. To
explore this mechanism in more detail, we focus on cases
(iii) and (iv) in the following.
The Kramers analysis assumes L and R transition rates
to depend only on barrier height, and ignores the shape
of the potential along the different unbinding pathways.
In order to account for those, and check the robustness of
the Kramers predictions, we now compute the lifetimes
via the full mean first passage time formalism.
Mean first passage time The average time for the par-
ticle to escape the trap, and thus for dissociation of the
bond, is the mean first passage time (MFPT) for the parti-
cle to pass the boundaries of the trap. The MFPT (cid:104)τ (x)(cid:105),
with x the point of departure inside the trap, obeys the
differential equation [15]
ftot(x)
kBT
d
dx
(cid:104)τ (x)(cid:105) +
d2
dx2(cid:104)τ (x)(cid:105) = − 1
D
,
(20)
supplemented with the boundary condition that (cid:104)τ (x)(cid:105) =
(a) The average unbinding time (cid:104)τ (x0)(cid:105) obtained by
FIG. 2:
Eq. 23 for a crosslink which released at x0 at t = 0. The sum
of κc and κs is 2 (black line), 4 (blue line), and 6 (red line).
(b) The first three terms of terms of the perturbed system for
κ = 2.
0 at the boundaries bL and bR of the trap. Here is D the
diffusion constant for the fluctuating particle, and ftot(x)
is the total force acting on the bead, defined as − dUtot
with Utot(x) = Ut(x) + Uext(s).
dx
We will focus on potential catch bonding for symmet-
ric traps, in combination with harmonic or anharmonic
external forcing to mimick the situation where the escap-
ing particle is actually one end of a polymeric object or
protein. Thus, we choose as our trapping potential
(cid:26) 1
2 κ(t)(x2 − b2) x ≤ b
Ut(x) =
0
,
(21)
x > b
i.e., harmonic with trap spring constant κ(t) in the re-
gion [−b, b], and zero elsewhere (this choice, as well as
the discontinuity at the boundaries, is immaterial; the
potential outside of the trap region is irrelevant to the
escape problem). Escape is defined as first passage across
the left or right boundary.
(iii): Symmetric trap, harmonic external potential
When the particle is attached to a harmonic spring with
an energy of the of the form of Eq. (17), the total force
(23)
fWLC(x, ε) = 90kBT
So, with the identification
(cid:96)p. The force-extension relation for such a polymer is
strongly nonlinear, and given by [16, 17]
(cid:115)
(cid:115)
(cid:96)(f ) = (cid:96)c − kBT
2f
2
f (cid:96)c
kBT (cid:96)p
coth
2
f (cid:96)c
kBT (cid:96)p
5
− 1
.
(26)
2/(cid:96)p.
The equilibrium length for such a polymer is (cid:96)0 = (cid:96)c
Expanding (cid:96)(f ) − (cid:96)0 to second order in the force f , and
inverting the relation yields the force-extension relation
to first anharmonic order for the semiflexible WLC:
(ε(cid:96)0 − x) +
(ε(cid:96)0 − x)2 .
2
(cid:96)c
(cid:18) (cid:96)p
(cid:19)2
(cid:18) (cid:96)p
(cid:19)3
(cid:18) (cid:96)p
(cid:19)2
(cid:18) (cid:96)p
(cid:96)c
(cid:96)c
2
2
kBT
2
(cid:96)c
(cid:19)3
,
+
5400
7
kBT
κ(2) = 90kBT
κ(3) =
5400
7
(27)
(28)
(29)
(30)
on the particle is given by
ftot(x, ε) = −κ(t)x + κ(2)(ε(cid:96)0 − x)
(22)
When we set ε = 0, the end point of the spring is fixed a
distance (cid:96)0 away from the center of the trap and the total
force becomes simply ftot(x, 0) = −(κ(t) + κ(2))x. In this
case, Eq. (20), with boundary conditions (cid:104)τ0(±b)(cid:105) = 0
may be solved analytically:
(cid:104)τ0(x0)(cid:105) =
b2
2F2({1, 1},{ 3
2D
0 2F2({1, 1},{ 3
2D
− x2
2 , 2}, b2(κ(t)+κ(2))
2kBT
)
2 , 2}, x2
0(κ(t)+κ(2))
2kBT
)
,
where pFq(a; b; z) is the generalized hypergeometric func-
tion. x0 is the position of the bead at t = 0. (cid:104)τ (x0)(cid:105) is
plotted in Fig. 2(a) for various values of κ = κ(t) + κ(2).
As expected, the escape time is maximal when the particle
departs from the center of the trap; we shall take this
value (cid:104)τ (0)(cid:105) at ε = 0 as our reference time.
To study the effect of putting the spring under tension,
we now increase ε away from zero. This results in an
extended tether in the center of the trap, and ftot is now
given by Eq. (22). The effect of this small change may be
studied perturbatively. When we expand the perturbed
solution to second order in ε as
(cid:104)τ(x0)(cid:105) = (cid:104)τ0(x0)(cid:105) + (cid:104)τ1(x0)(cid:105) + 2 (cid:104)τ2(x0)(cid:105) + O(ε3),
(24)
with (cid:104)τ0(x0)(cid:105) is the exact solution for = 0, (cid:104)τ1(x0)(cid:105)
and (cid:104)τ2(x0)(cid:105) (and, indeed, the higher order corrections)
may be obtained from an order-by-order set of recursive
differential equations
d(cid:104)τn(cid:105)
dx0
−x0(κ(t) + κ(2))
d(cid:104)τn−1(cid:105)
dx0
κ(2)(cid:96)0
kBT
d2(cid:104)τn(cid:105)
dx2
0
= 0 ,
kBT
(25)
with boundary conditions (cid:104)τn(±b)(cid:105) = 0. These equations,
too, may be analytically solved for (cid:104)τ1(x0)(cid:105) and (cid:104)τ2(x0)(cid:105).
Fig. 2(b) graphs these first two corrections to the ε = 0
result, and confirms what was already suggested by the
Kramers analysis: (cid:104)τ1(0)(cid:105) = 0, meaning that there is no
effect, to first order in ε, on the lifetime of the bond (i.e.,
no catch bond). Moreover, we also see that indeed the
second order correction (cid:104)τ2(0)(cid:105) = 0 is negative, proving
that the zero-ε lifetime is indeed a maximum. Attaching
the escaping particle to a linear spring cannot increase
the lifetime in a symmetric potential.
+
+
(iv): Symmetric trap, anharmonic external potential
Kramers theory predicts catch bonding in the case of
an anharmonic external potential. To connect this to
a more realistic setting, we will specify to the case of a
Worm-Like Chain tether; instead of a Hookean spring
the escaping particle is now connected to a semiflexible
polymer with a contour length (cid:96)c and a persistence length
FIG. 3:
(a) Potentials Uext(x, 0) and (b) average unbinding
times relative to their ε = 0 values for a particle attached to a
semi-flexible fiber with (cid:96)p = (cid:96)c = 20 (dashed black line), the
first harmonic approximation to the WLC (dotted blue line)
and the first anharmonic approximation (red line).
we may use the external potential of Eq. (18) to capture
the lowest relevant nonlinear order of tethering by a semi-
flexible WLC. Obviously, for larger extensions (and thus
forces) we may need to go to higher orders in (ε(cid:96)0 − x).
To verify the catch bonding effect predicted by Kramers
for this anharmonic force extension relation on the mean
unbinding time, We numerically solve Eq. (20) for a semi-
flexible polymer with (cid:96)c = (cid:96)p = 20, confined to a trap
with a spring constant strength of κ(2) = 10 and a disso-
ciation length b = 1. While we can no longer treat this
case analytically, we do have access to the full range of ε
allowing to compute, also, the total extent of the catch
effect. Comparing the full WLC force-extension relation
with its harmonic and first anharmonic approximations
(potentials graphed in Fig. 3(a)) confirms the Kramers
predictions; to harmonic order; there is no catch bond ef-
fect but the inclusion of the first anharmonic term creates
a regime of increasing lifetime with rising ε. Thus, indeed,
a nonlinear anharmonicity (third order terms or higher,
odd powers, in Uext(x, ε)) are a prerequisite for this type
of catch bond effect. The full WLC curve shows that the
rise in lifetime continues over an extended range of ε, and
that the total induced lifetime increase can attain values
of over 10% (see Fig. 3(b)).
So, both the Kramers analysis and the MFPT compu-
tations show a catch bonding effect for a particle in a
symmetric, one-dimensional trap, attached to a polymer
or, in fact, any tether possessing some anharmonic re-
sponse. In the final part of this manuscript, we assess the
real-life validity of the effect by a Molecular Dynamics
(MD) simulation in three dimensions.
Molecular Dynamics of the escape of a particle attached
to a WLC in three dimensions. We use LAMMPS MD [18]
to simulate a fluctuating semiflexible bead-spring chain
consisting of N beads, connected by identical Hookean
springs with a spring constant κ(MD) and rest length (cid:96)0
and include a bending contribution to the chain energy,
quantified by a bend stiffness K. The resulting chain
energy is
N(cid:88)
i=1
N−1(cid:88)
i=1
UMD =
κ(MD)
2
((cid:96)i − (cid:96)0)2 +
K
2
θ2
i
(31)
where (cid:96)i the length is of spring i, and θi the angle between
spring i and spring i+1. One of the end beads is trapped in
a harmonic potential with spherical symmetry, the other
end is fixed at a distance (cid:96) = (cid:96)0(1 + ε) from the center of
the trap. In the simulation a polymer consisting of N = 40
springs that each have a rest length (cid:96)0 = 0.5 (that is, the
contour length of the entire polymer is 20) and spring
constant κ(MD) = 2500 is used (all in Lennard-Jones
units). This is very high, to supress backbone extension
and approximate the inextensible WLC. The bending
stiffness K is set to 20, corresponding to a persistence
length (cid:96)p of 20 when we set kBT = 1.
6
FIG. 4:
(a) Normalized lifetime of particle attached to a bead-
spring polymer obtained by MD simulation. (b) Mean square
displacement of the end-bead showing subdiffusive motion.
The three dimensional MD simulations again confirm
the existence of the effect. As Fig. 4(a) shows, the life-
time of the particle inside the trap rises with rising ε,
qualitatively in the manner seen in the MFPT apprach.
There is, however, no satisfactory quantitative agreement
between the numerical solution of the Eq. (20) and the
MD simulations, despite the identical parameters. We
believe this to be due to the anomalous diffusion of the
polymer end point within the trap. As we show in Fig.
4(b), the Mean Squared Displacement (cid:104)δx2(cid:105) of the end
bead scales as (cid:104)δx2(cid:105) ∼ t 3
4 , that is -- slower than the ex-
pected power of 1. This subdiffusive motion of a polymer
end point is well-known [19], but invalidates the deriva-
tion that produces Eq. (20). To fully capture the escape
problem of a polymer end point, we suggest a version of
Eq. (20) with fractional derivatives might be required,
but this is beyond the scope of our current presentation.
Summary and conclusion Our results prove, that catch
bonding -- bound states whose lifetime increases with ap-
plied force -- is a generic feature of bonds where either
the trap itself, or the external structures that a ligand is
connected to, or both, display some degree of nonlinear
response. These effects are robustly predicted by Kramers
theory, MFPT theory, and are confirmed in direct MD
simulation.
The mechanism we describe suggest catch bonding may
be ubiquitous, and does not require finely tuned struc-
tural or conformational properties of bond participants.
The same effects are predicted to occur in reversible (non-
covalent) bonding in, for instance, polymeric materials;
these reversible links should, under very general condi-
tions, also show a regime of some strengthening when
loaded. It will be instructive, in future analysis, to assess
the theoretical limits of this effect to inspire the design of
novel, responsive catch bond materials.
[1] H. A. Kramers, Physica 7, 284 (1940).
[2] R. Merkel, P. Nassoy, A. Leung, K. Ritchie, and E. Evans,
Nature 397, 50 (1999).
[3] M. Escud´e, M. K. Rigozzi, and E. M. Terentjev, Biophys-
ical journal 106, 124 (2014).
[4] B. T. Marshall, M. Long, J. W. Piper, T. Yago, et al.,
Nature 423, 190 (2003).
[5] Y. Suzuki and O. K. Dudko, The Journal of chemical
physics 134, 02B616 (2011).
[6] B. T. Marshall, M. Long, J. W. Piper, T. Yago, et al.,
Nature 423, 190 (2003).
7
[7] F. Kong, A. J. Garc´ıa, A. P. Mould, M. J. Humphries,
and C. Zhu, The Journal of cell biology 185, 1275 (2009).
[8] W. E. Thomas, V. Vogel, and E. Sokurenko, Annual
Review of Biophysics 37, 399 (2008).
[9] W. Thomas, Annu. Rev. Biomed. Eng. 10, 39 (2008).
[10] J. Lou and C. Zhu, Biophysical journal 92, 1471 (2007).
[11] K. K. Sarangapani, J. Qian, W. Chen, V. I. Zarnitsyna,
P. Mehta, T. Yago, R. P. McEver, and C. Zhu, Journal
of Biological Chemistry 286, 32749 (2011).
[12] H. Chen and A. Alexander-Katz, Biophysical journal 100,
174 (2011).
[13] Y. Wei, Physical Review E 77, 031910 (2008).
[14] Y. V. Pereverzev, O. V. Prezhdo, M. Forero, E. V.
Sokurenko, and W. E. Thomas, Biophys. J. 89, 1446
(2005).
[15] S. Redner, A Guide to First-Passage Processes (Cam-
bridge University Press, 2001).
[16] M. Dennison, M. Jaspers, P. Kouwer, C. Storm, A. Rowan,
and F. MacKintosh, Soft Matter 12, 6995 (2016).
[17] E. Huisman, C. Storm, and G. Barkema, Physical Review
E 78, 051801 (2008).
[18] S. Plimpton, Journal of computational physics 117, 1
(1995).
[19] M. Hinczewski, X. Schlagberger, M. Rubinstein,
O. Krichevsky, and R. R. Netz, Macromolecules 42, 860
(2009).
|
1712.04832 | 1 | 1712 | 2017-12-13T16:04:38 | Seeding Dispersal Modeling For Systems of Planar Microbial Biofilms | [
"physics.bio-ph"
] | We propose a modeling approach to study how mature biofilms spread and colonize new surfaces by predicting the formation and growth of satellite colonies generated by dispersing biofilms. This model provides the basis for better understanding the fate and behavior of dispersal cells, phenomenon that cannot, as yet, be predicted from knowledge of the genome. The model results were promising as supported by the experimental results. The proposed approach allows for further improvements through more detailed sub-models for front propagation, seeding, availability and depletion of resources. The present study was a successful proof-of-concept in answering the following questions: Can we predict the colonization of new sites following biofilm dispersal? Can we generate patterns in space and time to shed light on seeding dispersal? That are fundamental issues for developing novel approaches to manipulate biofilm formation in industrial, environmental and medical applications. | physics.bio-ph | physics | Seeding Dispersal Modeling For Systems of Planar Microbial Biofilms
Andrea Trucchia1,2, Federica Villa3, Luigi Frunzo4, Gianni Pagnini1,5∗
1BCAM - Basque Center for Applied Mathematics,
Alameda de Mazarredo 14, E-48009 Bilbao, Basque Country -- Spain
2Universidad del Pa´ıs Vasco/Euskal Herriko Unibertsitatea UPV/EHU,
Campus de Leioa, E-48949 Leioa, Basque Country -- Spain
3Department of Food, Environmental and Nutritional Sciences,
University of Milan, via Celoria 2, I-20133 Milan,Italy
4Department of Mathematics and Applications "Renato Caccioppoli",
5Ikerbasque - Basque Foundation for Science, Calle de Mar´ıa D´ıaz de Haro 3, E-48013 Bilbao, Basque Country -- Spain
University of Naples "Federico II", via Cintia, Monte S. Angelo, I-80126 Naples, Italy
We propose a modeling approach to study how mature biofilms spread and colonize new sur-
faces by predicting the formation and growth of satellite colonies generated by dispersing biofilms.
This model provides the basis for better understanding the fate and behavior of dispersal cells,
phenomenon that cannot, as yet, be predicted from knowledge of the genome. The model results
were promising as supported by the experimental results. The proposed approach allows for fur-
ther improvements through more detailed sub-models for front propagation, seeding, availability
and depletion of resources. The present study was a successful proof-of-concept in answering the
following questions: Can we predict the colonization of new sites following biofilm dispersal? Can
we generate patterns in space and time to shed light on seeding dispersal? That are fundamental
issues for developing novel approaches to manipulate biofilm formation in industrial, environmental
and medical applications.
It is now well accepted that microorganisms lead social
lives and engage in complex behavior in response to other
organisms and the extracellular environment. By adopt-
ing coordinated chemical and physical interactions, mi-
croorganisms establish complex communities attached to
a surface and embedded in a self-produced extracellular
polymeric matrix, enabling cells to develop efficient sur-
vival strategies [1]. This sessile lifestyle is called biofilm,
and it represents the dominant mode of microbial life
in many natural, medical and engineered systems [2, 3].
Cells in biofilms undergo developmental programs result-
ing in an ordered and predictable transition through a se-
ries of stages, each based on stage-specific expression of
genes [4]. The biofilm developmental program culminates
with the release of free-living cells that can colonize new
habitats, possibly richer in resources [5], as seen Fig. 1.
While detachment is a passive process of cell loss result-
ing from sloughing of cells and erosion from the biofilm,
active or seeding dispersal is coordinated via regulatory
systems in response to a number of cues (e.g., alteration
in the availability of nutrients, oxygen depletion, levels
of iron) and signals (e.g., acyl-homoserine lactones, dif-
fusible fatty acids, cell-cell autoinducing peptides) [6].
Thus, seeding dispersal can occur in the complete ab-
sence of flowing conditions, and does not depend upon
shear forces that removes cells from the biofilm. Another
interesting feature of seeding dispersal is that cells ap-
pear to have a distinct phenotypes different from those
of biofilm and planktonic cells, increasing cell ability to
colonize a greater range of habitats important for niche
expansion [7, 8]. Thus, dispersal represents an important
adaptive strategy with profound impacts on the survival
and fitness of microorganisms. It allows biofilm popula-
tions to spread and colonize new surfaces, avoiding over-
crowding, depletion of resources and competition among
cells in the local environment, and promoting the reju-
venation of biofilms [9]. Furthermore, dispersal is linked
to the generation or maintenance of genetic variation,
with significant outcomes for the success of those bac-
teria in the environment [7, 8, 10]. Although dispersal
is advantageous from the microbial standpoint, it may
negatively affect some industrial and medical processes.
For instance, through dispersed cells, biofilm can spark
new infections within the host and result in the trans-
mission of bacteria between different hosts [11]. Further-
more, dispersal may promote, for example, the spread of
parasitism phenomena in animals and plants [12], biode-
terioration of historical and artistic objects [13, 14] and
fouling in food-processing equipment [15]. The existence
of a programmed generation of dispersed cells appears in-
creasingly clear, but the challenge now is to provide the
mechanistic understanding of biofilm dispersal. Thus,
the principal questions that motivate this work are: Can
we predict the colonization of new sites following biofilm
dispersal? Can we generate patterns in space and time
to shed light on seeding dispersal? We propose a model-
ing approach to study the growth of mono-layer micro-
bial biofilm on inert surfaces by focusing on the biofilm
spread induced by dispersal, predicting the formation
and growth of satellite colonies generated by dispers-
ing biofilms. The importance of this work relies on the
fact that the fate and behavior of dispersal cells can-
not, as yet, be predicted from knowledge of the genome.
Thus, a mathematical modelling of biofilm dispersion is
urgently needed. The planar geometry we focus on is
proper of biofilm growth in oligotrophic environments
7
1
0
2
c
e
D
3
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
2
3
8
4
0
.
2
1
7
1
:
v
i
X
r
a
2
of the trajectory of a dispersed cell with an average po-
sition x = x(t) and initially located in x(0) = x0, such
that Xω(0, x) = x0. Cell trajectories are described by the
one-particle density function pω(x; t) = δ (x − Xω (t, x)),
where δ (x) is the Dirac δ-function. Moreover, let the re-
gions Ω occupied by the colonies be conveniently marked
by an indicator function IΩ(x, t). Then, an effective in-
dicator ϕe, ϕe(x, t) : S × [0, +∞[→ [0, 1], of the region
surrounded by a random front is obtained by using the
sifting property of the δ-function and by averaging the
indicator function:
ϕe(x, t) = (cid:104)
IΩ(x, t)δ(x − Xω(t, x)) dx(cid:105)
S
IΩ(x, t)(cid:104)δ(x − Xω(t, x))(cid:105) dx
IΩ(x, t)p(x; t x) dx
(cid:90)
(cid:90)
(cid:90)
(cid:90)
S
S
=
=
=
(cid:90) t
1
τ (x, )
0
p(x; t x) dx ,
(2)
Ω(t)
where p (x; t x) = (cid:104)δ (x − Xω (t, x))(cid:105) is the PDF of the
seeding bacteria. In this work, p (x; t x) is assumed to
be Gaussian.
Function ϕe(x, t) provides the probability that dis-
persed bacteria cells arrive in a point x from different
sources Ωi(t). However, to relate this probability of ar-
rival to a successful formation of a new biofilm colony
spot, a criterion associated with a reversible/irreversible
attachment due to environmental conditions and biolog-
ical time scales is needed. With this aim, we introduce
the integral field
ψ(x, t) =
ϕe(x, ) d ,
(3)
that stores the signals received from the active biofilm
domain Ω in the temporal interval [0, t]. We denote by
τ (x, t) the timescale of signal storing and it is determined
by
τ (x, t) = τe(x, t) + V (x, t) ,
(4)
where τe(x, t) represents the environmental distribution
of resources in absence of biofilm and V (x, t) accounts
for the resource depletion performed by the biofilm.
The feedback mechanism between ψ and ϕ is given by
the procedure
ψ(x, t) (cid:62) 1 → IΩ(x, t) = 1 ,
(5)
that is: when into a considered spot a certain amount
of dispersed cells have established and endured a cer-
tain amount of time (that accounts for the environmental
availability of resources) then a new colony is generated.
FIG. 1. Graphic picture of biofilm seeding mechanism, with
motile bacteria abandoning the main colony in order to attach
into favorable spots where to start new colonies.
(e.g., reverse osmosis membranes, stone monuments, sur-
gical gauze, contact lenses, water supply pipes), where
nutrient constraints limit microbial growth to thin mono-
layered biofilms. The growth of this biofilm is character-
ized by two main phenomena: the biomass expansion due
to the growth of primary existing colonies, and the forma-
tion of new colonies due to the attachment of dispersed
cells released by the primary ones, i.e., seeding dispersal.
In analogy with an approach originally introduced for
turbulent premixed combustion [16] and wild-land fire
propagation [17, 18], the proposed mechanistic model is
built up as follows. Biofilm colony growth is modeled by
using the Level Set Method [19], while seeding dispersal
is simulated through the Probability Density Function
(PDF) corresponding to the diffusive process that gov-
erns the bacteria dispersal behavior. The seeds attach-
ment depends on their concentration and environmental
resources availability, with the latter characterized by its
initial spatial distribution, and by the depletion effect
due to the presence of mature biofilm colonies. The ini-
tial configuration of environmental resource availability
can be modeled by setting a specific scenario or by using
a random distribution.
The surface of mature biofilm colony Ω is generally
composed by an ensemble of biofilms spots Ωi with
i = 1, ..., n(t) where the total number n depends on
time t because of merging and birth of colonies. Let
ϕ : S × [0, +∞[→ IR be a function defined on the domain
of interest S ⊆ IR2 such that the iso-line ϕ (x, t) = c de-
scribes the evolution the boundaries of Ωi, i.e., the evo-
lution of the colonies fronts. Then the motion of the
fronts of biofilm colonies is determined by the Level Set
Equation:
∂ϕ(x, t)
∂t
= u(x, t)(cid:107)∇ϕ(x, t)(cid:107) .
(1)
In the following, the outward normal velocity u(x, t) is
assumed constant, i.e., u(x, t) = u.
Let the mature colonies be able to release a sufficient
large number of cells whose dispersion is characterised
by a random motion. Let Xω(t, x) be the ω-realization
Hence, the indicator function IΩ(x, t) results to be
1 ,
0 ,
IΩ(x, t) =
if ϕ (x, t) (cid:54) c or ψ(x, t) (cid:62) 1 , x ∈ Ω ,
elsewhere , x (cid:54)∈ Ω .
(6)
Equation (4) shows an interplay between the avail-
ability of resources offered from the surrounding envi-
ronment and the resource depletion performed by the
growth of the biofilm colonies. This simple formulation
of the timescale for the waiting times of free cells seeding
is able to generate a plethora of patterns of biological rel-
evance. In the following, the term τe is assumed constant
in time, because it represents the availability of resources
before the action of biofilm, and this changes slower than
the biofilm evolution. The term V (x, t) is modeled by
the following Poisson problem
α∆V (x, t) = ρb ,
V (x, t)x∈∂S = 0 ,
(7a)
(7b)
where ρb is the bacterial density inside the colonies and
α an absorption kinetic coefficient. In our case, the bac-
terial density inside the colony is constant, and the latter
equation becomes
α∗∆V (x, t) = IΩ(x, t) ,
(8)
where α∗ corresponds to α in the rescaled setting and dif-
fers for the physical dimensions. The dynamic governed
by (8) depends only on α∗ and, in spite of its simplicity,
it manages to represent availability of biofilm resources,
determining the temporal dynamics of seeding dispersal.
In order to prove the potentiality of the proposed
approach, an experimental test case has been designed
and realized. Pseudomonas aeruginosa strain PAO1
(MH873) was used in this study as a model system of
bacterial biofilms. In fact, the metabolically versatile P.
aeruginosa PAO1 is an opportunistic pathogen of plants,
animals, and humans and is ubiquitously distributed in
soil and aquatic habitats. Furthermore, the bacterium is
genetically characterized and amenable to mutagenesis
and "omics" based approaches [20, 21]. The microorgan-
ism was maintained at -80◦C in suspensions containing
20% glycerol and 2% peptone, and was grown aerobically
in Tryptic Soy Broth (TSB medium) for 15h at 30◦C. Dis-
persion experiments were conducted by using the colony-
biofilm culturing system. Briefly, 2 sterile black polycar-
bonate filter membranes (0.22 µm pore size and 25mm
diameter) were placed in each Petri dish containing Tryp-
tic Soy Agar (TSA medium), at a distance of 2 mm from
each other. Bacterial cells are trapped completely by the
membrane filters having a pore size smaller than the bac-
terial size, while nutrients and metabolites diffuse across
membranes easily. Fifty µl of cell suspension contain-
ing 1 × 108 cells were used to inoculate the central filter
3
membrane. The plates were incubated at 30◦C for 72h.
Every 24h the Petri dishes were observed, and the disper-
sal phenomenon was documented by capturing imagines
with both a camera and a stereomicroscope (magnifica-
tion 12X).
Numerical solution of model (1) -- (8) has been com-
puted by setting the physical parameters as follows: α∗ =
0.05 ms−2, u = 10.0 ms−1 inside the membranes and zero
outside, and the diffusion coefficient of the Gaussian PDF
equal to 103 ms−2. The numerical set-up is based on a
2D mesh [0, 220]x[0, 370] with grid step δx = δy = 1.0.
The numerical test concerns the two membranes: the in-
oculated one and seeding target (the two external dashed
lines in Fig. 2). These circular membranes have radius
R = 60 in grid step units and center in (110, 118) (the
inoculated membrane) and (110, 252) (the target mem-
brane). At the initial instant, a mature biofilm colony is
assumed to be present in the inoculated membrane with
circular profile centered in the center of the membrane
and radius r = 35. Furthermore, the availability of the
environmental food needs to be set and it is represented
by τe in (4). In particular, τe is assumed to be constant
in time and ranging through a linear interpolation pro-
cedure from 0.01s, when x is inside the inner disk with
radius < 0.70R, to 600.00s, when x is outside the mem-
branes (see the dashed circles in the right side of Fig. 2).
This assumption corresponds to a very large timescale for
generating a new colony outside the membranes, which
corresponds to unfavorable conditions.
in
Fortran2008/OpenMP,
The computation was done by using the facilities of
BCAM by running an OpenMp-parallel finite difference
C/Fortran code. Its routines rely on a general-purpose
library written
LSMLib
(http://ktchu.serendipityresearch.org/software/
lsmlib/). The latter provides robust and efficient tools
for studying the evolution of co-dimensional fronts mov-
ing in one-, two- and three-dimensional domains. ENO
algorithms are used for the sake of computing accurate
space derivatives, while for the advancement in time a
second order Runge -- Kutta scheme was implemented.
Figure 2 shows the growth of a primary colony in the
inoculated membrane and its colonization of the target
membrane by a seeding dispersal mechanism both for
the experimental data (the pair of membranes on the
left) and the proposed modeling approach (the pair on
the right). The Level Set Method describes the growth
of the colony: first the primary one that is living in the
inoculated membrane (left side membrane) and later the
secondary one in the target membrane. The seeding and
the attachment mechanism, which are responsible for the
colonization of the target membrane, are well reproduced
by the model. In spite of the fact that the present com-
parison is qualitative, it shows that the present approach
is able through its modular structure to model the growth
of the biomass colony and to take into account the differ-
ent processes that simultaneously occur. In particular,
the present approach provides a method to link a sharp
interface model for the growth of biofilm colonies and a
statistical treatment for biofilm seeding. The modular
structure allows for a detailed front propagation through
a more detailed expression for the normal velocity of the
colony front u(x, t) and a more detailed bacterial mi-
gration through a new statical characterization. The
comparison between the experimental pictures and some
frame of evolution of the model is promising and, thanks
to the modular structure, the present approach emerges
as a novel and useful method for understanding the com-
plex dynamics displayed by microbial biofilm.
FIG. 2. Left column: pictures from the inoculated membrane
and the host membrane, taken at t = 24h, 48h and 72h. The
biofilm is contoured by a red dashed line. Right column: three
stages of the numerical simulation of the experiment, where
the biofilm is marked by the purple bold surface.
To conclude, we remark that one of the main motiva-
tions for studying biofilm dispersal is to provide a mecha-
nistic model to predict how cells attach and proliferate to
seed new biofilms. An increased understanding of the fate
of dispersed cells will offer a broad conceptual framework
for developing novel approaches to manipulate biofilm
formation (either discouraging or promoting biofilm de-
velopment) in industrial, environmental and medical ap-
plications. Thus, the ability to unravel the mechanisms
of dispersal would have a great socio-economical signif-
icance, with profound implications for global health, as
well as for the management of environmental microorgan-
isms in biogeochemical cycling processes and biotechno-
logical applications of biofilms. Another argument sup-
porting the significance of this model is that could be
potentially applied to eukaryotes that show "biphasic"
life cycles characterized by a dispersive phase and a ses-
sile phase (e.g., corals, bryozoans, cancer cells).
GP acknowledges the support by the Basque Gov-
ernment through the BERC 2014-2017 program and
by Spanish Ministry of Economy and Competitive-
4
ness MINECO through BCAM Severo Ochoa excel-
lence accreditation SEV-2013-0323 and through project
MTM2016-76016-R "MIP". AT is supported by the PhD
Grant "La Caixa" 2014. LF acknowledges Progetto Gio-
vani GNFM 2016 "Comportamenti emergenti ed auto-
organizzazione in sistemi iperbolici di reazione-diffusione
in ambito biologico ed ecologico." FV has received fund-
ing from the European Union Seventh Framework Pro-
gramme (FP7-PEOPLE-2012-IOF) under grant agree-
ment No. 328215.
∗ [email protected]
[1] C. D. Nadell, K. Drescher, and K. R. Foster, Nat. Rev.
Microbiol. 14, 589 (2016).
[2] L. Hall-Stoodley, J. W. Costerton, and P. Stoodley, Nat.
Rev. Microbiol. 2, 95 (2004).
[3] M. R. Mattei, L. Frunzo, B. D'Acunto, Y. Pechaud,
F. Pirozzi, and G. Esposito, J. Math. Biol. (2017), doi
10.1007/s00285-017-1165-y.
[4] H. C. Flemming, J. Wingender, U. Szewzyk, P. Steinberg,
S. A. Rice, and S. Kjelleberg, Nat. Rev. Microbiol. 14,
563 (2016).
[5] D. McDougald, S. A. Rice, N. Barraud, P. D. Steinberg,
and S. Kjelleberg, Nat. Rev. Microbiol. 10, 39 (2011).
[6] C. Guilhen, C. Forestier, and D. Balestrino, Mol. Micro-
biol. 105, 188 (2017).
[7] S. L. Chua, Y. Liu, J. K. Yam, Y. Chen, R. M. Vejborg,
B. G. Tan, S. Kjelleberg, T. Tolker-Nielsen, M. Givskov,
and L. Yang, Nat. Commun. 5, 4462 (2014).
[8] B. D'Acunto, L. Frunzo, I. Klapper, and M. R. Mattei,
Math. Biosci. 259, 20 (2015).
[9] N. Barraud, S. Kjelleberg,
and S. A. Rice, Micro-
biol. Spectr. 3 (2015), doi:10.1128/microbiolspec.MB-
0015-2014.
[10] B. Purevdorj-Gage, W. J. Costerton, and P. Stoodley,
Microbiology 151, 1569 (2005).
[11] H. Koo, R. N. Allan, R. P. Howlin, P. Stoodley, and
L. Hall-Stoodley, Nat. Rev. Microbiol. 15, 740 (2017).
[12] F. Villa, F. Cappitelli, P. Cortesi, and A. Kunova, Front.
Microbiol. 8, 654 (2017).
[13] F. Cappitelli, O. Salvadori, D. Albanese, F. Villa, and
C. Sorlini, Biofouling 28, 257 (2011).
[14] F. Villa, P. Stewart, I. Klapper, J. Jacob, and F. Cap-
pitelli, BioScience 66, 285 (2016).
[15] F. Cappitelli, A. Polo, and F. Villa, Food Eng. Rev. 6,
29 (2014).
[16] G. Pagnini and E. Bonomi, Phys. Rev. Lett. 107, 044503
(2011).
[17] G. Pagnini and A. Mentrelli, Nat. Hazards Earth Syst.
Sci. 14, 2249 (2014).
[18] I. Kaur, A. Mentrelli, F. Bosseur, J.-B. Filippi,
and
G. Pagnini, Commun. Nonlinear Sci. Numer. Simul. 39,
300 (2016).
[19] J. A. Sethian and P. Smereka, Ann. Rev. Fluid Mech. 35,
341 (2003).
[20] C. K. Stover, X. Q. Pham, A. L. Erwin, S. D. Mizoguchi,
P. Warrener, M. J. Hickey, F. S. Brinkman, W. O. Huf-
nagle, D. J. Kowalik, M. Lagrou, R. L. Garber, L. Goltry,
E. Tolentino, S. Westbrock-Wadman, Y. Yuan, L. L.
Time Brody, S. N. Coulter, K. R. Folger, A. Kas, K. Larbig,
R. Lim, K. Smith, D. Spencer, G. K. Wong, Z. Wu, I. T.
Paulsen, J. Reizer, M. H. Saier, R. E. Hancock, S. Lory,
and M. V. Olson, Nature 406, 959 (2000).
[21] T. S. Walker, H. P. Bais, E. Deziel, H. P. Schweizer, L. G.
Rahme, R. Fall, and J. M. Vivanco, Plant Physiol. 134,
320 (2004).
5
|
1907.10533 | 3 | 1907 | 2019-12-11T02:25:31 | Chromatin state switching in a polymer model with mark-conformation coupling | [
"physics.bio-ph",
"cond-mat.soft",
"cond-mat.stat-mech"
] | We investigate the phase transition properties of the polymer-Potts model, a chain composed of monomers with magnetic degrees of freedom, with the motivation to study the conformation and mark switching dynamics of chromatin. By the mean-field approximation, we find that the phase transition between the swollen-disordered state and the compact-ordered state is discrete; it is first-order as in the long-range Potts model, but with a significantly larger jump in magnetization (i.e., mark coherence) upon the ordering transition. The results imply how small changes in epigenetic writer concentrations can lead to a macroscopic switching of the chromatin state, suggesting a simple mechanism of discrete switching observed, for instance, in cell differentiation. | physics.bio-ph | physics | Chromatin state switching in a polymer model with mark-conformation coupling
Kyosuke Adachi1, 2, ∗ and Kyogo Kawaguchi1, 3, †
1Nonequilibrium Physics of Living Matter RIKEN Hakubi Research Team,
RIKEN Center for Biosystems Dynamics Research, 2-2-3 Minatojima-minamimachi, Chuo-ku, Kobe 650-0047, Japan
2RIKEN Interdisciplinary Theoretical and Mathematical Sciences Program, 2-1 Hirosawa, Wako 351-0198, Japan
3RIKEN Cluster for Pioneering Research, 2-2-3 Minatojima-minamimachi, Chuo-ku, Kobe 650-0047, Japan
(Dated: December 12, 2019)
9
1
0
2
c
e
D
1
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
3
v
3
3
5
0
1
.
7
0
9
1
:
v
i
X
r
a
We investigate the phase transition properties of the polymer-Potts model, a chain composed of monomers
with magnetic degrees of freedom, with the motivation to study the conformation and mark switching dynamics
of chromatin. By the mean-field approximation, we find that the phase transition between the swollen-disordered
state and the compact-ordered state is discrete; it is first-order as in the long-range Potts model, but with a
significantly larger jump in magnetization (i.e., mark coherence) upon the ordering transition. The results imply
how small changes in epigenetic writer concentrations can lead to a macroscopic switching of the chromatin
state, suggesting a simple mechanism of discrete switching observed, for instance, in cell differentiation.
Introduction. -- Chromatin is a large polymer composed
of monomers called nucleosomes, which are histone protein
complexes wrapped with DNA [1, 2]. The switching of cell
states is encoded in the changes in epigenetics [3] such as in
the molecular and structrual changes in the chromatin. The
chromatin states have been traditionally categorized into two:
euchromatin and heterochromatin, which correspond to active
(open) and inactive (closed) parts of the chromatin in terms of
gene expression and accessibility. Consistent with this, recent
chromosome conformation capture (Hi-C) experiments [4, 5]
have identified the existence of two major compartments in
the genome. The regions within the same compartment share
similar marks (i.e., chemical modifications) in the nucleo-
somes and tend to interact with each other more frequently
than across [6]. Experiments have shown that different marks
on the nucleosome induce distinct interactions due to the nat-
ural attraction and repulsion between nuclesomes [7, 8], or
by mediating proteins such as HP1 [9, 10], and the polycomb
repressive complexes [11, 12].
It has also been established that cell differentiation is ac-
companied by a large (megabase) scale transition in the com-
partments as well as changes in the states of epigenetic
marks [5]. The mechanism behind this switching, however,
remains elusive. Previous modeling studies have assumed
mark-dependent interactions between nucleosomes in order to
explain the observed contact maps and 3D structures of the
chromatin [13, 14]. Other models have considered how the
interaction between the marked histones lead to bistability in
the coherent epigenetic marks [15 -- 17]. A natural question is
then how the interplay of chromatin chain dynamics and the
kinetics of nucleosome modifications can lead to the drastic
switching of compartments observed in differentiation.
To model chromatin polymer dynamics under stochastic
modifications of nucleosomes, the polymer-Potts model has
been considered [18 -- 20]. In this model, the random motion
of the polymer chain is accompanied by monomer-monomer
interactions that depend on the histone marks, and the hi-
∗ [email protected]
† [email protected]
stone marks can stochastically switch due to enzymatic re-
actions and histone turnover [21, 22].
It has been numer-
ically shown [19, 23] that even for the Ising-type model,
where there is essentially only two distinct states of the his-
tone marks, there is a first-order-like transition between the
swollen-disordered state, which corresponds to a loose poly-
mer with spatially random marks, and the compact-ordered
state, where the conformation is globular and the marks are
coherent. This abrupt transition is likely due to the coupling
between the conformation change and the epigenetic switch-
ings, although a concrete theory is still lacking.
In this paper, we investigate the phase transition properties
of the polymer-Potts model by considering a polymer chain
in continuum space with stochastic histone mark exchange.
Employing the Flory-type mean-field approximation for the
dynamics of the chain, we write the pseudo free energy of
the generic polymer-Potts model as a function of the order
parameters representing the magnetization (i.e., mark coher-
ence) and the polymer conformation. For the Ising-type in-
teraction, the transition between the swollen-disordered state
and a compact-ordered state is first-order, consistent with sim-
ulations and theories investigating mark dynamics on self-
avoiding random walks [18, 20]. In the general case with mul-
tiple types of marks, we find that the jump in magnetization
at the transition point is always larger in the polymer-Potts
model compared with the Potts-model counterpart, and also
obtain a criteria for the absense of a continuous transition. We
further study the switching transition upon stretching of the
chain, which serves as a simple model of force-induced epi-
genetic modification.
Model. -- We consider a polymer model with Potts-like in-
teractions between monomers. The interactions are medi-
ated in a histone-mark-dependent way by proteins that we
call readers, and the marks can change stochastically due to
enzymatic reactions caused by the writers (Fig. 1). We as-
sume that there are q (≥ 2) types of marks. Extending the
Flory-type mean-field approximation [24] to the present situ-
ation, the pseudo free energy (per monomer) at a temperature
2
FIG. 1.
(a) Chromatin consists of a DNA (gray line) wrapping
histones (green cylinders). The readers (colored rectangles) con-
nect histones with the same marks (colored circles), while the writ-
ers (colored stars) edit the histone marks.
(b) The polymer-Potts
model. A monomer constituting the polymer has a changeable his-
tone mark (colored circle), and there are mark-specific interactions
between monomers (colored dashed lines, Ji j), in addition to the
mark-independent interactions (v and w).
FIG. 2. Color plot of the pseudo free energy f (φ, m) [Eq. (1)] as a
function of φ, which represents the square root of the globule den-
sity, and the magnetization m for the case of the Ising-type interac-
tion with v = 0.1. For (a) J < J(c), the minimum of f is realized at
φ = 0 and m = 0 (swollen-disordered state), while for (c) J > J(c),
it is realized at φ > 0 and m > 0 (compact-ordered state). There
is bistability at (b) J = J(c), meaning that the transition between the
swollen-disordered state and the compact-ordered state is discontin-
uous.
T reads
(cid:88)
f (ρ,{xi}) = vρ + wρ2 − ρ
Ji jxix j
(kBT xi ln xi − hixi) .
1≤i, j≤q
+
(cid:88)
1≤i≤q
(1)
the i-th type mark occupation ((cid:80)q
Here ρ is the average monomer concentration given as ρ ∼
N/R3 with the polymer end-to-end distance R and the total
number of monomers N. The other variables, {xi}, represent
i=1 xi = 1). The first two
terms in Eq. (1) correspond to the volume exclusion effect
(v, w > 0). Note that the parameters v and w represent the
second and third virial coefficients, respectively. The third
term originates from the Potts-like two-body interactions be-
tween monomers mediated by the readers. The detail of the
interactions between the different types of marks is coded in
Ji j, which is a real symmetric matrix. The fourth term is the
entropy associated with the mark degrees of freedom. The
last term represents the effect of external fields {hi}, which
describes how much a specific epigenetic mark is favored, re-
flecting, for example, the concentration of the histone mod-
ification enzymes, i.e., the writers. The equilibrium state is
determined by minimizing Eq. (1) with respect to ρ and {xi}.
A few remarks are to be made for Eq. (1). Firstly, fluc-
tuation effects are neglected compared with the microscopic
model, although its inclusion will likely not change the key re-
sults [25, 26]. Secondly, while higher-order interaction terms
are irrelevant near the conventional second-order coil-globule
transition point [24], inclusion of these terms may shift the
transition point if the coil-globule transition becomes first-
order, as in the situations explained below. Nevertheless, we
expect that Eq. (1) captures the key characters observed in
simulations of similar systems [19, 20] and is useful to gen-
erally analyse models with multiple kinds of marks. Lastly,
the chromatin state transitions of our interest is at the level
of sub-regions of compartments or several topologically asso-
ciated domains [27], which is megabase scale corresponding
to N = 103−5. Although we have omitted all the terms that
vanish in the limit of N → ∞ in Eq. (1), it is straightforward
to include higher-order terms and discuss their effects on the
properties of transition [24, 28].
For the sake of understanding, let us first fix the mark de-
grees of freedom, {xi}. Then, Eq. (1) is equivalent to the free
energy of a classic polymer [24] in the large N limit. As in-
vestigated in [24], there exists a transition between the coiled
state, a swollen polymer with the average length scaling as
R ∼ N3/5 [29, 30], and the globule state, a densely packed
polymer with R ∼ N1/3 [24], upon changing the overall two-
i, j Ji jxix j) from re-
pulsive to attractive. Such coil-globule transitions have been
observed in experiments using DNA [31] and chromatin [32].
The coil-globule transition in this case is continuous in the
limit of N → ∞ [24], even beyond the mean-field approxima-
tion [25, 26].
body interaction (in the present case, v −(cid:80)
In another direction of simplification, we can consider the
order-disorder transition of the marks under a fixed polymer
density, ρ. Assuming a globular configuration (R ∼ N1/3) and
Ji j = Jδi j, Eq. (1) represents the mean-field free energy of the
Potts model [33]. In the limit of N → ∞, the order-disorder
transition is continuous for q = 2 while discontinuous for q ≥
3, which has been believed to be correct in three dimensions
even beyond the mean-field approximation [33].
ing the dimensionless globular order parameter φ
Phase transition by interactions and fields. -- Introduc-
:=
ρ, we can express the equilibrium free energy as
.
+
1≤i≤q
xi ln xi − hixi
kBT
√
Here, the dimensionless two-body interaction strengths are
defined as Ji j
kBTw. As we
have seen, if {xi} or φ is fixed to some value, the system de-
scribed by Eq. (2) will show the conventional coil-globule or
magnetic transition, respectively.
√
kBTw and v := v/
:= Ji j/
(2)
f = [minφ,{xi} f (φ,{xi}) s.t.(cid:80)q
(w/kBT)1/4 √
v − (cid:88)
(cid:32)
(cid:88)
f (φ,{xi})
kBT
=
1≤i, j≤q
i=1 xi = 1], where
Ji jxix j
φ2 + φ4
(cid:33)
(a)(b)Figures for ver. 0.7(a) !"=1.5<!"((b) !"=1.77=!"((c) !"=2>!"( 3
(Fig. 2), we can approximate minφ,m f (φ, m) (cid:39) minφ f (φ, m =
±1) = −( J − v)2kBT/4. We then obtain
(cid:16)
J(c) (cid:39) v + 2
(cid:113)
ln 2
( J − v)/2
for J > J(c)(cid:17)
∗ (cid:39)
(3)
(4)
√
φ
shown as the black dotted line in Fig. 3(a), giving a good ap-
proximation. We have confirmed that the features such as the
jump of m∗ from 0 to (cid:39)1 at the transition point are observed
for a broad range of the values of v [28].
We further consider the effect of external mark-specific
fields by setting h1 = h and h2 = −h in the Ising-type model.
The phase diagram in the case of v = 0.1 is shown in Fig. 3(b).
We find that the field-induced transition from the swollen-
disordered state to the compact-ordered state is of first-order
around the zero-field transition point, meaning that increasing
or decreasing specific writers can also induce the switching
behavior. Interestingly, within a certain range of J, sequential
second- and first-order transitions occur as the field becomes
stronger [Fig. 3(c)]. For a smaller J, a single continuous tran-
sition is induced by the field. Note that such a field-induced
transition has been discussed in the context of the magnetic
polymer within mean-field approaches [18] as well as in sim-
ulations on self-avoiding walk models [18, 34].
exp(hi/kBT)/(cid:80)q
the continuous transition occurs at { J(c)
Propeties of transitions under general settings. -- Here
we consider the condition for the transition between the
swollen and compact states to be countinuous under gen-
eral q, {Ji j}, and {hi}. Minimizing Eq. (2) on the assump-
i j } for a
tion that
given set of {hi}, we obtain φ∗ = 0 and x∗
= S i( h) :=
j=1 exp(h j/kBT) [28]. The order parameters
will grow in response to the deviation of the interaction
strengths from their critical values: φ∗ = ∆φ and x∗
=
S i( h) + ∆xi for
+ ∆ Ji j. Minimizing Eq. (2) at
Ji j = J(c)
i j
+ ∆ Ji j will give the following relation:
Ji j = J(c)
i j
i
i
(cid:17)(cid:104)
(cid:16) h
(cid:17)
(cid:105)(cid:104)
(cid:16) h
(cid:17)
(cid:105)− v.
2 (∆φ)2 =
+ ∆ Ji j
S i
+ ∆xi
S j
+ ∆x j
(5)
For the transition to be continuous,
the order parameters
should smoothly change at the transition point: ∆φ → 0 and
∆xi → 0 for ∆ Ji j → 0, meaning that J(c)
(cid:16) h
(cid:17)
i j should obey
(cid:16) h
(cid:17)
(cid:88)
= v.
(6)
S j
J(c)
i j S i
(cid:88)
(cid:16) J(c)
i j
1≤i, j≤q
Therefore, if (cid:80)
1≤i, j≤q
i, j Ji jS i( h)S j( h) < v is satisfied, which is
when the mean effective two-body interaction is repulsive,
any continuous swollen-compact transition is prohibited and
only switch-like transitions can occur. This condition can in
principle be checked in experiment by measuring the effective
interactions between nucleosomes.
out external fields.
A simple example is again the Ising-type model with-
i, j Ji jS i( h)S j( h) =
i, j Ji j/4 = 0 < v, the continuous conformation transition is
always prohibited, consistent with our numerical results that
In this case, since (cid:80)
(cid:80)
FIG. 3. (a) The globular order parameter φ∗ (green solid line) and
magnetization m∗ (orange dashed line) as a function of the mark-
specific interaction strength J. The approximate functional form of
φ∗ [Eq. (4)] is plotted as black dotted line. (b) Phase diagram in the
J-h plane. The analytical expressions of the second-order transition
line [28] (brown dotted line) and the approximate first-order transi-
tion line [28] (black solid line) are also plotted. (c) The order pa-
rameters, φ∗ and m∗, as a function of the external field strength h for
the case of J = 1.25. (d) Magnetization jump (∆m) at the magnetic
transition point as a function of the number of mark types (q) for the
mean-field polymer-Potts model [red (dark gray) points] and for the
mean-field Potts model [blue (light gray) points]. For the polymer-
Potts model, ∆m is plotted for several values of v (v = 0, 0.1, 0.5, 1, 5
for q = 2; v = 0, 5 for q ≥ 3). (e) The stretched-state order parameter
η∗ (purple dotted line), in addition to φ∗ (green solid line) and m∗ (or-
ange dashed line), as a function of the external force strength F for
the case of J = 3. (f) Phase diagram in the J-F plane. The small-J
and large-J regions represent the stretched-disordered and compact-
ordered phases, respectively. The phase boundary estimated with
Eqs. (2), (7), and (8) is shown with red circles, along with the ap-
proximate F(c)( J) curve [28] (black line). The transition is always of
first-order.
To see the effect of the coupling between {xi} and φ, we
first study the Ising-model case, q = 2 and Ji j = J(2δi j − 1)
with hi = 0. Introducing the magnetization, m := x1 − x2,
f (φ, m) for the case of v = 0.1 is shown in Fig. 2. Let us
denote the equilibrium values of φ and m as φ∗ and m∗, re-
spectively. In this model, there is a critical value J(c) such that
for J < J(c), the swollen-disordered phase is the equilibrium
[φ∗ = 0 and m∗ = 0, Fig. 2(a)], whereas for J > J(c), this
switches to the compact-ordered state [φ∗ > 0 and m∗ > 0,
Fig. 2(c)]. At the transition point [ J = J(c), see Fig. 2(b)],
both the swollen-disordered state and the compact-ordered
state are stable, meaning that there is a first-order transition.
Thus, a switching transition can occur by simply changing the
strength of the reader-mediated interaction, as has been seen
numerically in a similar model [19].
In Fig. 3(a) the J dependence of φ∗ and m∗ is plotted for
v = 0.1, showing a clear jump of the order parameters at the
transition point. Since m∗ (cid:39) 1 in the compact-ordered state
(a)R~ N3/5R~ N1/3(b)(c)(e)FR~ N−F(f)F−FR~ N1/3FR ~ N−FF−FR~ N1/3(d)"#=5"#=0 the first-order transition occurs irrespective of the value of
v [28].
i } − min{x∗
To investigate the order-parameter jump at the first-order
transition more specifically, we consider the Potts-type inter-
action: Ji j = Jδi j with hi = 0. Figure 3(d) shows the mag-
i } for q ≥ 2 in the
netization jump ∆m = max{x∗
mean-field polymer-Potts model [Eq. (2)] compared with the
conventional mean-field Potts model. In the Potts model [33],
the transition is of second-order for q = 2 while first-order for
q ≥ 3, and ∆m is given as (q− 2)/(q− 1). In the polymer-Potts
model, on the other hand, the magnetic transition is always
first-order, and ∆m is a monotonically increasing function of
v and q, while the dependency on v is almost negligible for
q ≥ 4. Notice that ∆m in the polymer-Potts model is always
larger than that in the corresponding Potts model, and ∆m is
practically unity for q ≥ 4. This suggests that the polymer
conformation change that accompanies the magnetic transi-
tion reinforces the all-or-none switching property.
Mechanical discontinuous transition. -- We here consider
what happens when a stretching force is applied to the edges
of a polymer chain with mark degrees of freedom. For sim-
plicity, let us investigate the effects of an external force term
added to the pseudo free energy [Eq. (2)] with the Ising-type
interaction. The force term can be written as fF = −F · R/N
with an external force F and the polymer end-to-end vector
R.
Within the mean-field level, the free energy including the
effects of an external force is given as
f = min {min
φ, m
f (φ, m), min
η, m
f (cid:48)(η, m)},
(7)
where f (cid:48)(η, m) is another pseudo free energy including the ef-
fect of the external force:
= − Fbη
kBT
1 + m
f (cid:48)(η, m)
kBT
+
3
2 η2
1 + m
ln
(8)
+
1 − m
2
1 − m
2
.
ln
+
2
2
Here, the dimensionless polymer length, η := R/Nb, can be
interpreted as an order parameter characterizing a stretched
state with the scaling R ∼ N [35].
In Eq. (8), the second
term represents the entropic elasticity [30], which is essential
under stretched conditions. We denote the equilibrium point
as (φ∗, η∗, m∗). Since the globule state (φ∗ > 0, R ∼ N1/3) and
the stretched state (η∗ > 0, R ∼ N) are incompatible, only one
of (φ∗, η∗) can be finite and the other should be zero.
Figure 3(e) shows the changes in order parameters upon
varying of the external force F for the case of v = 0.1 and
J = 3, in which the compact-ordered state (φ∗ > 0, η∗ = 0,
and m∗ > 0) is stabilized when F = 0. We can see that
a first-order transition occurs at a certain critical value F(c),
above which a stretched-disordered state (φ∗ = 0, η∗ > 0, and
m∗ = 0) emerges. The numerically obtained phase diagram
in the J-F plane is shown in Fig. 3(f). Note that the force-
induced coil-globule transitions are believed to be discontin-
uous also in classical polymer models at N → ∞ [35 -- 38].
In the polymer-Potts model, we find that the mark degrees
of freedom become immediately disordered accompanying
4
(a) Lennard-Jones-type interaction U(i j)
LJ (r) between two
FIG. 4.
monomers (1 ≤ i, j ≤ 2). σ is the interaction length, and is the at-
tractive interaction strength between monomers with the same mark.
(b) Optimized order parameters φ∗ and m∗ as a function of the inter-
action strength /kBT.
this stretching transition. Similar discontinuous transitions
between a compact-ordered state and a stretched-disordered
state have recently been seen in molecular dynamics simula-
tions with short-range interactions [39].
Relation to molecular dynamics simulations. -- To compare
our results with the molecular dynamics simulations [19] us-
ing Lennard-Jones-type interactions [Fig. 4(a)], we consider
the virial expansion. By neglecting O(φ6) terms and the ex-
istence of the neutral mark, we obtain the pseudo free energy
[Eq. (2)] with additional terms proportional to m2φ4 [28].
By minimizing the pseudo free energy, we obtain the op-
timized φ∗ and m∗ as a function of the interaction strength
/kBT. Figure 4(b) shows that a discontinuous transition
with a large jump of magnetization occurs at the interaction
strength /kBT (cid:39) 0.85, which is close to the simulation re-
sult [19] (/kBT (cid:39) 0.9 for N = 2000). Although the virial
expansion is not generally justfied for cases with a first-order
transition, this result suggests that a simplified framework can
connect the molecular level measurement of histone interac-
tions [7] to the compartment level chromatin state transition.
Discussion and conclusion. -- Here we have studied the
polymer-Potts model at the mean-field level and found that
switch-like transitions are largely enhanced, compared with
the transitions in a polymer model with unchangeable marks
or the conventional Potts model, due to the coordination of the
coil-globule and magnetic transition. The bistable property
leading to the first-order transition fits with the phenomenol-
ogy of chromatin state transition and cell differentiation. For
instance, it has been shown that elimination of small kilobase-
scale genome regions can induce compartment switching of
a whole megabase-scale region [40]. This can be explained
by the bistability of the chromatin state, which allows local-
ized histone mark biases induced by transcription factors to
spread macroscopically. The hysteresis effect, which is ex-
pected to accompany the chromatin discontinuous transition,
may also improve the stability of the epigenetic regulation
against chemical and mechanical perturbations and cell divi-
sion.
Additional to the equilibrium phase transition scenario pro-
posed in this paper, nonequilibrium features of the chemical
dynamics [19, 20] and the phase separation properties of the
(a)(b) key components in chromatin dynamics [8, 41 -- 44] may play
roles in enhancing or diminishing the switch-like behavior.
Nevertheless, the fact that a simple mark-conformation cou-
pling can lead to a discrete switch indicates that nonlinear
dynamics and well-designed chemical networks may not be
essential in explaining cell fate dynamics. In real differenti-
ation, state switching occurs in sub-regions and does not ex-
pand to the whole chromosome [5]. It is interesting to explore
how specific regions in the genome set boundaries to prevent
the phase transition dynamics from spreading into undesired
regions [45].
Acknowledgements. -- We are grateful to Yohsuke T. Fukai
and Soya Shinkai for fruitful discussions and the reading of
the manuscript. This work was supported by JSPS KAKENHI
Grant Numbers JP18H04760, JP18K13515, JP19H05275,
JP19H05795.
5
[1] R. Cortini, M. Barbi, B. R. Car´e, C. Lavelle, A. Lesne, J. Mozzi-
conacci, and J.-M. Victor, Rev. Mod. Phys. 88, 025002 (2016).
[2] B. Fierz and M. G. Poirier, Annu. Rev. Biophys. 48, 321 (2019).
[3] C. H. Waddington, The strategy of the genes (1957).
[4] E. Lieberman-Aiden, N. L. van Berkum, L. Williams,
M. Imakaev, T. Ragoczy, A. Telling, I. Amit, B. R. Lajoie,
P. J. Sabo, M. O. Dorschner, R. Sandstrom, B. Bernstein, M. A.
Bender, M. Groudine, A. Gnirke, J. Stamatoyannopoulos, L. A.
Mirny, E. S. Lander, and J. Dekker, Science 326, 289 (2009).
[5] J. R. Dixon, I. Jung, S. Selvaraj, Y. Shen, J. E. Antosiewicz-
Bourget, A. Y. Lee, Z. Ye, A. Kim, N. Rajagopal, W. Xie,
Y. Diao, J. Liang, H. Zhao, V. V. Lobanenkov, J. R. Ecker, J. A.
Thomson, and B. Ren, Nature 518, 331 (2015).
[6] M. Di Pierro, R. R. Cheng, E. L. Aiden, P. G. Wolynes, and
J. N. Onuchic, Proc. Natl. Acad. Sci. U.S.A. 114, 12126 (2017).
[7] J. J. Funke, P. Ketterer, C. Lieleg, S. Schunter, P. Korber, and
H. Dietz, Sci. Adv. 2, e1600974 (2016).
[8] B. A. Gibson, L. K. Doolittle, L. E. Jensen, N. Gamarra, S. Red-
ding, and M. K. Rosen, bioRxiv 523662 (2019).
[9] A. J. Bannister, P. Zegerman, J. F. Partridge, E. A. Miska, J. O.
Thomas, R. C. Allshire, and T. Kouzarides, Nature 410, 120
(2001).
[10] S. Machida, Y. Takizawa, M. Ishimaru, Y. Sugita, S. Sekine, J.-
i. Nakayama, M. Wolf, and H. Kurumizaka, Mol. Cell 69, 385
(2018).
[11] A. Angel, J. Song, C. Dean, and M. Howard, Nature 476, 105
(2011).
[12] A. N. Boettiger, B. Bintu, J. R. Moffitt, S. Wang, B. J. Beliv-
eau, G. Fudenberg, M. Imakaev, L. A. Mirny, C.-t. Wu, and
X. Zhuang, Nature 529, 418 (2016).
[13] D. Jost, P. Carrivain, G. Cavalli, and C. Vaillant, Nucleic Acids
Res. 42, 9553 (2014).
[14] M. Barbieri, M. Chotalia, J. Fraser, L.-M. Lavitas, J. Dostie,
A. Pombo, and M. Nicodemi, Proc. Natl. Acad. Sci. U.S.A.
109, 16173 (2012).
[15] I. B. Dodd, M. A. Micheelsen, K. Sneppen, and G. Thon, Cell
P. R. Cook, and D. Marenduzzo, Nucleic Acids Res. 46, 83
(2017).
[24] P. G. de Gennes, J. Phys. Lett. 36, 55 (1975).
[25] M. A. Moore, J. Phys. A 10, 305 (1977).
[26] I. M. Lifshitz, A. Y. Grosberg, and A. R. Khokhlov, Rev. Mod.
Phys. 50, 683 (1978).
[27] J. R. Dixon, S. Selvaraj, F. Yue, A. Kim, Y. Li, Y. Shen, M. Hu,
J. S. Liu, and B. Ren, Nature 485, 376 (2012).
[28] See Supplemental Material for derivations and detailed analy-
sis.
[29] M. Doi and S. F. Edwards, The Theory of Polymer Dynamics
(Clarendon Press, Oxford, 1986).
[30] P. G. de Gennes, Scaling Concepts in Polymer Physics (Cornell
University Press, Ithaca, NY, 1979).
[31] K. Yoshikawa, M. Takahashi, V. V. Vasilevskaya, and A. R.
Khokhlov, Phys. Rev. Lett. 76, 3029 (1996).
[32] A. Zinchenko, N. V. Berezhnoy, S. Wang, W. M. Rosencrans,
N. Korolev, J. R. C. van der Maarel, and L. Nordenskild, Nu-
cleic Acids Res. 46, 635 (2017).
[33] F. Y. Wu, Rev. Mod. Phys. 54, 235 (1982).
[34] J.-H. Huang and M.-B. Luo, Polymer 45, 2863 (2004).
[35] P.-Y. Lai, Physica A 221, 233 (1995).
[36] A. Halperin and E. B. Zhulina, Europhys. Lett. 15, 417 (1991).
[37] P. Grassberger and H.-P. Hsu, Phys. Rev. E 65, 031807 (2002).
[38] P. L. Geissler and E. I. Shakhnovich, Phys. Rev. E 65, 056110
[39] D. Michieletto, E. Orlandini, and D. Marenduzzo, Sci. Rep. 7,
(2002).
14642 (2017).
[40] J. Sima, A. Chakraborty, V. Dileep, M. Michalski, K. N. Klein,
N. P. Holcomb, J. L. Turner, M. T. Paulsen, J. C. Rivera-Mulia,
C. Trevilla-Garcia, D. A. Bartlett, P. A. Zhao, B. K. Washburn,
E. P. Nora, K. Kraft, S. Mundlos, B. G. Bruneau, M. Ljungman,
P. Fraser, F. Ay, and D. M. Gilbert, Cell 176, 816 (2019).
[41] A. R. Strom, A. V. Emelyanov, M. Mir, D. V. Fyodorov,
X. Darzacq, and G. H. Karpen, Nature 547, 241 (2017).
[42] A. G. Larson, D. Elnatan, M. M. Keenen, M. J. Trnka, J. B.
Johnston, A. L. Burlingame, D. A. Agard, S. Redding, and
G. J. Narlikar, Nature 547, 236 (2017).
[43] A. J. Plys, C. P. Davis, J. Kim, G. Rizki, M. M. Keenen, S. K.
Marr, and R. E. Kingston, bioRxiv 467316 (2018).
[44] R. Tatavosian, S. Kent, K. Brown, T. Yao, H. N. Duc, T. N.
Huynh, C. Y. Zhen, B. Ma, H. Wang, and X. Ren, J. Biol.
Chem. 294, 1451 (2019).
[45] M. J. Obersriebnig, E. M. H. Pallesen, K. Sneppen, A. Trusina,
and G. Thon, Nat. Commun. 7, 11518 (2016).
[46] J. E. Mayer and M. G. Mayer, Statistical Mechanics (Wiley,
[16] K. Sneppen, M. A. Micheelsen, and I. B. Dodd, Mol. Syst.
[17] D. Jost, Phys. Rev. E 89, 010701 (2014).
[18] T. Garel, H. Orland, and E. Orlandini, Eur. Phys. J. B 12, 261
[19] D. Michieletto, E. Orlandini, and D. Marenduzzo, Phys. Rev.
[20] D. Coli, D. Michieletto, D. Marenduzzo, and E. Orlandini,
129, 813 (2007).
Biol. 4, 182 (2008).
(1999).
X 6, 041047 (2016).
arXiv:1807.11101 .
(2010).
[21] R. B. Deal, J. G. Henikoff, and S. Henikoff, Science 328, 1161
1977).
[22] M. F. Dion, T. Kaplan, M. Kim, S. Buratowski, N. Friedman,
and O. J. Rando, Science 315, 1405 (2007).
[23] D. Michieletto, M. Chiang, D. Col, A. Papantonis, E. Orlandini,
Chromatin state switching in a model with mark-conformation coupling
Supplemental Material for
S1
Kyosuke Adachi and Kyogo Kawaguchi,
(Dated: December 12, 2019)
I. POLYMER-POTTS MODEL
In this section, we describe how to obtain the pseudo free energy of the polymer-Potts model starting a generic form of
microscopic interactions. The Hamiltonian of the polymer-Potts model with external field and force terms reads
H = HGC({Rn}) + HP({Rn},{sn}) + HM({sn}) + HVE({Rn}) + HF(R0, RN).
(S1)
Here, the degrees of freedom are the marks of each monomer, sn (1 ≤ n ≤ N), which represent the types of chemical modifi-
cations attached to the n-th histone along the chain, and the positions of the monomers, Rn. We assume that there are q (≥ 2)
types of histone marks that are distinguishable in the sense of interactions: sn ∈ {1, 2, ..., q}. Also, we take {Rn}N
n=1 as a set of
variables. The first term of Eq. (S1),
(Rn − Rn−1)2 ,
(S2)
N(cid:88)
n=1
HGC({Rn}) =
3kBT
2b2
HP({Rn},{sn}) = −(cid:88)
J(sn, sm)δ (Rn − Rm) ,
corresponds to the Gaussian-chain interaction that restricts the positions of monomers adjacent to each other in the polymer,
where b corresponds to the Kuhn length (the size of the monomer), and T is the temperature. The second term,
detail of interactions between the different types of marks can be coded in J(sn, sm) = (cid:80)
describes the attractive or repulsive interaction between the monomers when they become close in contact with each other. The
1≤i, j≤q Ji jδi,sn δ j,sm, where Ji j is a real
symmetric matrix. For example, the Ising-type interaction can be realized if we set q = 2 and Ji j = J(2δi j − 1), and the q-state
Potts-type interaction can be realized if we choose Ji j = Jδi j. The third term,
n(cid:44)m
(cid:88)
HM({sn}) = −(cid:88)
(cid:88)
hi
n
i
δ (Rn − Rm) + w
n(cid:44)m(cid:44)k
(cid:88)
n(cid:44)m
δsn,i,
describes the effect of an external bias field, which favors a certain type of histone marks over other types. The fourth term,
HVE({Rn}) = v
δ (Rn − Rm) δ (Rm − Rk) ,
describes the two- and three-body volume exclusion interaction terms that we assume have no histone-mark dependence (v, w >
0). The last term,
HF(R0, RN) := −F · (RN − R0),
(S6)
represents a mechanical stretching force that is applied to the edges of the polymer.
A. Modified polymer-Potts model with long-range interactions
rewrite the interaction terms (HP and HVE) with the polymer density, ρ(r,{Rn}) :=(cid:80)
i-th state, σi(r,{Rn},{sn}) :=(cid:80)
To perform a mean-field analysis of Eq. (S1), let us replace the local quantities with the averaged counterparts. First, we
n δ(r − Rn), and the mark density of the
(cid:90)
n δi,sn δ(r − Rn). We then obtain
HP({Rn},{sn}) = −(cid:88)
(cid:90)
Ji j
i, j
HVE({Rn}) = v
drσi(r,{Rn},{sn})σ j(r,{Rn},{sn}),
(cid:90)
drρ(r,{Rn})3,
drρ(r,{Rn})2 + w
and
(S3)
(S4)
(S5)
(S7)
(S8)
by neglecting the terms proportional to δ(0). Next, we define the average density and mark density as
and
ρav(r, RN0) :=
3N
4πRN0
3 θ(RN0 − r),
(cid:88)
1
N
δi,sn ,
n
σi,av(r, RN0,{sn}) := ρav(r, RN0)
(cid:82)
where RN0 := RN − R0 is the end-to-end distance vector and θ(x) is the Heaviside step function. Note that
drρ(r,{Rn}) = N and
As a mean-field approximation, we replace ρ and σi with ρav and σi,av, respectively. The obtained Hamiltonian is a modified
drσi,av(r, RN0,{sn}) =(cid:82)
drσi(r,{Rn},{sn}) =(cid:80)
drρav(r, RN0) =
n δi,sn.
(cid:82)
(cid:82)
S2
(S9)
(S10)
(S11)
(S12)
(S13)
P (RN0,{sn}) + HM({sn}) + HLR
VE(RN0) + HF(R0, RN).
polymer-Potts model with long-range interactions:
HLR = HGC({Rn}) + HLR
The long-range interaction terms (HLR
VE) are given as
P and HLR
P (RN0,{sn}) := −(cid:88)
HLR
(cid:90)
(cid:88)
J(cid:48)
i j
3
i, j
1
N
(cid:88)
n
where J(cid:48)
i j := 3Ji j/4π, and
i, j
= − N2
RN0
(cid:90)
HLR
VE(RN0) := v
δi,sn
δ j,sm
,
(cid:88)
m
1
(cid:90)
N
drρav(r, RN0)2 + w
drρav(r, RN0)3
Ji j
drσi,av(r, RN0,{sn})σ j,av(r, RN0,{sn})
where v(cid:48) := 3v/4π and w(cid:48) := (3/4π)2w.
= v(cid:48) N2
RN0
3
+ w(cid:48) N3
RN0
6 ,
B. Derivation of pseudo free energy
In the following, we derive the functional form of the pseudo free energy (per monomer) through analyzing the equilibrium
free energy of the modified polymer-Potts model, f := −(kBT/N) ln Z, where the canonical partition function is defined as
(cid:90)
N(cid:89)
n=1
(cid:82)
− 3
+
2b2
n=1
(cid:33)
− HLR
kBT
sn=1
(cid:32)
exp
q(cid:88)
exp
1
.
Z :=
dRn
.
(S14)
Using the explicit form of HLR and an identity,
Z =
dR exp
(cid:32)
(cid:90)
×
(cid:90)
N(cid:89)
n=1
− v(cid:48)
kBT
N2
N3
R6
R3 − w(cid:48)
δ(R − RN0) exp
kBT
dRn
q(cid:88)
(cid:88)
dR δ(R − RN0) = 1, we can obtain
J(cid:48)
F · R
i j
kBT
kBT
(cid:33) N(cid:89)
N(cid:88)
(Rn − Rn−1)2
sn=1
R3
n=1
i, j
(cid:88)
n,m
δi,sn δ j,sm
+
q(cid:88)
i=1
hi
kBT
N(cid:88)
n=1
δi,sn
(S15)
Note that R represents the end-to-end distance vector, which is clear from its definition.
To perform the integration with respect to {Rn}N
n=1, we change the integration variables from {Rn}N
n=1 to {Rn := Rn−Rn−1}N
n=1.
n=1 Rn)], we can proceed the calculation as
=
3
n=1
n=1
n=1
n=1
=
=
Rn
2b2
exp
N3/2
dRn
dRn
3
1
We then obtain
− 3
− Nb2k2
(cid:33)
2b2
+ ik · R
(Rn − Rn−1)2
N(cid:88)
n=1 Rn) = (2π)−3(cid:82)
dk exp[ik · (R −(cid:80)N
Using δ(R − RN0) = δ(R −(cid:80)N
δ(R − RN0) exp
N(cid:89)
(cid:90)
N(cid:88)
− 3
exp
N(cid:89)
(cid:90) dk
(cid:90)
(2π)3 exp (ik · R)
(cid:33)3N/2(cid:90) dk
(cid:33)
(cid:32)
(cid:32)2πb2
(cid:32)
(cid:32)2πb2
(cid:33)3(N−1)/2
(2π)3 exp
6
− 3R2
2Nb2
(cid:32)
(cid:33)3(N−1)/2(cid:90)
(cid:32)2πb2
R3 − w(cid:48)
2Nb2 − v(cid:48)
1
exp
N(cid:89)
− 3R2
q(cid:88)
q(cid:88)
N(cid:88)
(cid:88)
(cid:88)
Ni=0 δ(Ni,(cid:80)N
n=1, we use the identity,(cid:80)N
exp
1
N(cid:89)
q(cid:88)
(cid:88)
N(cid:88)
(cid:88)
exp
1
q(cid:89)
N(cid:89)
q(cid:89)
N(cid:88)
(cid:88)
q(cid:88)
q(cid:89)
exp
1
q(cid:88)
(cid:88)
N(cid:88)
N!(cid:81)q
R3
To perform the summation with respect to {sn}N
J(cid:48)
i j
kBT
J(cid:48)
i j
kBT
J(cid:48)
i j
kBT
Kronecker delta. We then obtain
dR exp
J(cid:48)
i j
kBT
q(cid:88)
q(cid:88)
N3/2
×
i=1
hi
kBT
q(cid:88)
hi
kBT
hi
kBT
hi
kBT
NiN j +
NiN j +
δi,sn δ j,sm
δi,sn δ j,sm
kBT
kBT
Z =
δi,sn
n=1
sn=1
n=1
sn=1
n=1
sn=1
i=1
Ni=0
N2
R3
R3
R3
Ni
Ni
n=1
=
=
n=1
n,m
n,m
i=1
i=1
i=1
1
3
+
+
i, j
i, j
.
i, j
i, j
i=1
i=1
Ni=0
2 − ik · N(cid:88)
n=1
Rn
(cid:33)
F · R
kBT
+
N3
R6
.
δi,sn
n=1 δi,sn) = 1, where δ(a, b) := δa,b is the
Ni,
δ
δi,sn
(S18)
N(cid:88)
.
n=1
δ
i=1 Ni!
i=1
Ni, N
Note that Ni represents the occupation number of the i-th mark state. Applying Stirling's formula, we can estimate the terms
originated from the entropy as
S3
(S16)
(S17)
(S19)
(S20)
Replacing the summation with respect to Ni with an integration over the occupation ratio of the i-th mark state, xi := Ni/N, we
obtain
.
i=1
− q(cid:88)
(cid:88)
Ni
i, j
J(cid:48)
i j
kBT
Ni ln
Ni
N
− 1
2
(q − 1) ln N + O(N0)
Ni, N
N!(cid:81)q
i=1 Ni!
J(cid:48)
i j
kBT
xix j + N
δ
i=1
q(cid:88)
q(cid:88)
q(cid:88)
i=1
hi
kBT
i=1
xix j + N
xi − N
q(cid:88)
hi
kBT
xi − N
i=1
q(cid:88)
i=1
xi ln xi − q − 1
2
ln N + O(N0)
xi ln xi +
q − 1
2
ln N + O(N0)
.
R3
i=1
q(cid:89)
q(cid:89)
exp
1
N(cid:88)
q(cid:89)
(cid:90) 1
δ
(cid:90) 1
Ni=0
= Nq−1
dxi
i=1
=
0
i=1
0
N!(cid:81)q
i=1 Ni!
= exp
dxi
i, j
J(cid:48)
i j
kBT
q(cid:88)
(cid:88)
δ
q(cid:88)
(cid:90) ∞
i=1
i=1
xi − 1
NiN j +
i=1
R3
xi − 1
hi
kBT
q(cid:88)
N2
exp
N2
exp
(cid:88)
q(cid:89)
(cid:90) 1
dxi
R3
i, j
i=1
0
dR
0
Z =
Based on the above calculation, after performing the solid-angle integration of R, we finally obtain the following reduced
form of the partition function:
δ
q(cid:88)
i=1
(cid:32)
exp
xi − 1
− N f (R,{xi})
kBT
− N f0
kBT
(cid:33)
+ O(N0)
,
(S21)
where the constant parts of the free energy, f0, is defined as
f0
kBT
and the pseudo free energy, f (R,{xi}), is given as
:=
3
2
ln
(cid:32)2πb2
(cid:33)
3
− q − 2
2
ln N
N ,
S4
(S22)
f (R,{xi})
kBT
3R2
2N2b2 − FR
kBT N
:=
(S23)
In the limit of N → ∞, the last term in Eq. (S23) is negligible, and we can apply Laplace's method to estimate Eq. (S21). Thus,
the free energy, f = −(kBT/N) ln Z, is obtained as
ln
+
+
xi ln xi − 2
N
R√
Nb
.
i jxix j
i, j J(cid:48)
kBT
N
R3
w(cid:48)
kBT
v(cid:48) −(cid:80)
min
N→∞ f = lim
lim
N→∞
N2
R6 − q(cid:88)
q(cid:88)
i=1
hi
kBT
i=1
xi +
q(cid:88)
+ lim
N→∞ f0.
xi = 1
i=1
Now we introduce two kinds of dimensionless order parameters, φ and η, regarding the conformation of the polymer. The first
parameter, φ, is defined as
R,{xi} f (R,{xi}) subject to
(cid:114) w(cid:48)
(cid:115)
φ :=
N
R3
kBT ,
(S24)
(S25)
(S27)
(S28)
(S29)
(S30)
(S31)
which corresponds to the square root of the globule density (N/R3) and becomes finite in the compact (or globule) state satisfying
R ∼ N1/3. The second parameter, η, is defined as
R
Nb ,
η :=
(S26)
which represents how the polymer is stretched and becomes finite in the stretched state satisfying R ∼ N. Note that either of φ
or η can be finite in the limit of N → ∞, which can be seen from the definitions of these order parameters. In the following, the
order parameters that minimize the pseudo free energy (i.e., equilibrium state) are denoted as φ∗ and η∗. Considering the scaling
R ∼ Nν, the swollen state (ν (cid:39) 3/5 [29, 30]) is included in the phase characterized by φ∗ = η∗ = 0. Although our framework
cannot determine ν beyond the inequality 1/3 < ν < 1 within the swollen state in the N → ∞ limit, ν = 3/5 is obtained in the
parameter regime far from the transition point if we consider the finite size effect using Eq. (S23) [24].
equilibrium mark order parameter as x∗
Regarding the macroscopic mark, the occupation ratios of each mark, {xi = Ni/N}q
i , and call that the i-th mark is ordered when x∗
Using φ, η, and {xi}, we can rewrite the formula of the free energy [Eq. (S24)] as
i=1, are order parameters. We denote the
i > 1/q.
(cid:110)
(cid:111)
The compact-state free energy, f C, is defined as
f C :=
where the corresponding pseudo free energy is given as
fC(φ,{xi})
kBT
kBTw and Ji j := J(cid:48)
i j/
fmark({xi})
,
xi = 1
N→∞ f = min
lim
f C, f S
N→∞ f0.
+ lim
φ,{xi} fC(φ,{xi}) subject to
min
v −(cid:88)
:= − q(cid:88)
:=
√
kBTw(cid:48) = Ji j/
i, j
Ji jxix j
√
i=1
q(cid:88)
φ2 + φ4 +
q(cid:88)
hixi +
xi ln xi,
i=1
i=1
fmark({xi})
kBT
.
kBT
min
f S :=
η,{xi} fS(η,{xi}) subject to
xi = 1
q(cid:88)
i=1
,
Here, v := v(cid:48)/
√
kBTw(cid:48) = v/
√
kBTw are dimensionless interaction strengths, and
is the non-interacting pseudo free energy of the marks with hi := hi/kBT being the dimensionless external field. On the other
hand, the stretched-state free energy, f S, is defined as
where the corresponding pseudo free energy is given as
fS(η,{xi})
kBT
:=
3
2 η2 − Fη +
fmark({xi})
kBT
.
S5
(S32)
Here, F := Fb/kBT is a dimensionless stretching force strength.
To sum up, in the case of f C < f S with φ∗ > 0 and η∗ = 0, the compact state (R ∼ N1/3) is stable and the realized order
parameters minimize fC(φ,{xi}); in the case of f C > f S with φ∗ = 0 and η∗ > 0, the stretched state (R ∼ N) is stable and the
realized order parameters minimize fS(η,{xi}); in the case of f C = f S = min{xi} fmark({xi}) with φ∗ = η∗ = 0, the swollen state
(R ∼ N3/5) is stable. The stable polymer state is also classified into ordered and disordered states according to the occupancy of
each mark: the i-th mark is ordered (disordered) in the case of x∗
i > 1/q (x∗
i ≤ 1/q).
C. Derivation of equations for the order parameters
Based on (S27), we derive the equations for the order parameters.
to the constraint(cid:80)q
To obtain the order parameters in the compact (or globule) state (φ∗ > 0 and η∗ = 0), we need to minimize fC(φ,{xi}) subject
i=1 xi = 1. Instead of directly treating the constraint, we introduce the Lagrange multiplier µ. Then, as a
necessary condition, the optimized order parameters should correspond to a stationary point of the following function:
1. Compact state
hixi +
xi ln xi − µ
xi − 1
q(cid:88)
i=1
.
q(cid:88)
i=1
We can find the stationary points of f (cid:48)
C(φ,{xi}, µ) by solving ∂ f (cid:48)
C/∂φ = 0, {∂ f (cid:48)
C/∂xi = 0}q
i=1, and ∂ f (cid:48)
C/∂µ = 0, i.e.,
where φ∗, x∗
i , and η∗ are the solution of these equations.
Since we here consider the compact state, we look for a solution with φ∗ > 0. Then, Eq. (S34) reduces to
(cid:16)
(cid:17)
∗ − v
,
∗T J x
x
φ
∗2 =
1
2
:= xi and [J]i j
e µ∗−1 exp (2φ∗2(cid:80)q
where we use matrix representations, [x]i
Ji jx∗
j
j=1
+ hi). Substituting this expression for x∗
hi)]−1. Therefore, we obtain the following expression of x∗
i :
∗2 J x
2φ
x∗
i = S i
(cid:16)
:= Ji j. On the other hand, Eq. (S35) can be rewritten as x∗
Ji jx∗
i in Eq. (S36), we obtain e µ∗−1 = [(cid:80)q
(cid:17)
i=1 exp (2φ∗2(cid:80)q
j=1
i
j
∗ + h
(1 ≤ i ≤ q),
=
+
where [ h]i := hi, and the softmax function, S i (y), is defined as
exp (zi)
i=1 exp (zi)
i }q
As a consequence, the order parameters in the compact state, φ∗ and {x∗
i=1, satisfy Eqs. (S37) and (S38).
S i (z) :=
.
(cid:80)q
C(φ,{xi}, µ)
f (cid:48)
kBT
φ2 + φ4 − q(cid:88)
φ
∗ = 0
∗2
i=1
:=
i, j
Ji jxix j
v −(cid:88)
v −(cid:88)
Ji jx∗
i x∗
j + 2φ
q(cid:88)
i, j
− 2φ
∗2
q(cid:88)
Ji jx∗
j=1
i − 1 = 0,
x∗
i=1
j − hi + ln x∗
i + 1 − µ
∗ = 0
(1 ≤ i ≤ q)
(S33)
(S34)
(S35)
(S36)
(S37)
(S38)
(S39)
S6
In the stretched state (φ∗ = 0 and η∗ > 0), we need to minimize fS(η,{xi}) subject to(cid:80)q
2. Stretched state
uncoupled, we can easily obtain
i=1 xi = 1. Noticing that η and {xi} are
and
x∗
i = S i
where S i (z) is the softmax function defined in Eq. (S39).
∗ =
η
(cid:17)
(cid:16) h
F
3 .
(1 ≤ i ≤ q),
(S40)
(S41)
3. Swollen state
In the swollen state (φ∗ = η∗ = 0), we only have to minimize fmark({xi}) subject to(cid:80)q
i }q
parameters, {x∗
i=1, is already obtained in Eq. (S41).
We can derive a sufficient condition for the swollen state to be stabilized as explained in the following. Assuming F = 0, we
to be zero if(cid:80)
can see η∗ = 0 since from Eq. (S32), fS(η,{xi}) is a monotonously increasing function of η. We thus show below a sufficient
condition for φ∗ = 0 on the assumption that F = 0. Focusing on the stationary condition of φ∗ [Eq. (S34)], we can see that φ∗ has
i, j Ji jxix j = xT J x is smaller than v for arbitrary values of {xi}q
i=1 satisfying 0 ≤ xi ≤ 1. Introducing an orthogonal
matrix M which diagonalizes the symmetric matrix J as MT J M = Λ := diag(λ1,··· , λq), writing the maximum eigenvalue of Λ
q(cid:88)
as λmax, and defining y := MTx, we obtain the following inequality,
i=1 xi = 1. The solution of the order
q(cid:88)
q(cid:88)
q(cid:88)
2 ≤ λmax
λiyi
yi
2 = λmax
2 ≤ λmax
xi
xi = λmax.
xT J x = yT Λ y =
(S42)
The second inequality is due to 0 ≤ xi ≤ 1. Therefore, if the maximum eigenvalue of J, λmax, is smaller than v, the only solution
of Eq. (S34) is φ∗ = 0. In short, a sufficient condition for the swollen state to be realized is that the maximum eigenvalue of J is
smaller than v (i.e., interaction is effectively repulsive) and also F = 0 (i.e., no stretching force).
i=1
i=1
i=1
i=1
D. Criteria for second-order transition
the second-order transition between a swollen state (φ∗ = 0) and a compact state (φ∗ > 0) can occur.
Let us consider the case with some external field (h (cid:44) 0) and no external force (F = 0), and explore under what conditions
Recall first that in the swollen state, φ∗ = 0 and x∗
= S i( h) (Sec. I C 3). Then, if we assume that a continuous transition occurs
i j , the system will smoothly change from the swollen state to the compact state in response to a certain small change
+ ∆ Ji j is obtained from
+ ∆ Ji j. The pseudo free energy for Ji j = J(c)
i j
= S i( h) + ∆xi for Ji j = J(c)
i j
i
at Ji j = J(c)
in Ji j: φ∗ = 0 + ∆φ and x∗
Eq. (S29) as
i
Minimizing this function with respect to ∆φ, we obtain
The assumption of a continuous transition leads to the continuity of the order parameters at Ji j = J(c)
∆ Ji j → 0; therefore, we obtain the following expression of the possible second-order transition point:
i j : ∆φ → 0 and ∆xi → 0 for
= v.
(S45)
(cid:16) h
(cid:17)
(cid:16) h
(cid:17)
J(c)
i j S i
S j
(cid:88)
i, j
fC(φ,{xi})
kBT
=
+ ∆ Ji j
(cid:105)
+ ∆xi
+
i j
i, j
v −(cid:88)
(cid:16) J(c)
(cid:104)
(cid:16) h
(cid:17)
(cid:118)(cid:117)(cid:117)(cid:116)1
(cid:88)
(cid:16) J(c)
−hi
S i
2
i j
i, j
+ ∆xi
S j
+ ∆x j
(cid:16) h
(cid:17)
+ ∆xi
ln
S i
+ ∆xi
(cid:17)(cid:104)
q(cid:88)
S i
i=1
(cid:17)(cid:104)
(cid:17)
(cid:16) h
(cid:16) h
(cid:104)
(cid:17)
(cid:16) h
(cid:17)
S i
(cid:105)(cid:104)
(cid:105)
(cid:105)(cid:104)
S j
(cid:16) h
(cid:17)
(cid:104)
(cid:16) h
(cid:17)
(cid:105) ∆φ2 + ∆φ4
(cid:105)
.
(cid:105) − v
.
∆φ =
+ ∆ Ji j
S i
+ ∆xi
+ ∆x j
(S43)
(S44)
S7
FIG. S1. Compact-state order parameter (φ∗) as a function of the interaction strength ( J) in the case of the mark-independent interaction and
v = 0.1.
This means that within the modified polymer-Potts model [Eq. (S11)], any second-order conformation transition accompanying
the changes in the order parameters should satisfy Eq. (S45). Conversely, if Eq. (S45) is not satisfied, the continuous swollen-
compact phase transition cannot occur. We will see in the following subsection I E that a discontinuous transition can occur even
if Eq. (S45) is not satisfied.
E. Analysis of simple cases
In this subsection, we provide some examples to explain how the equilibrium states and transition properties of the polymer-
Potts model can change according to the histone-mark specificity of the interactions, {Ji j}. For simplicity, we assume F = 0 and
hi = 0.
1. Mark-independent interaction
Let us consider interactions which are not specific to histone marks (Ji j = J). In this case, the formula of the free energy
[Eq. (S27)] is reduced to limN→∞ f = f C + limN→∞ f0 with the following pseudo free energy:
φ2 + φ4 +
xi ln xi.
(S46)
(cid:16)
(cid:17)
v − J
fC(φ,{xi})
kBT
=
q(cid:88)
i=1
(cid:16)
(cid:17)
v − J
The third term originated from the entropy of the marks can be minimized with xi = 1/q, so that we obtain the further reduced
expression, limN→∞ f = minφ fC,U(φ) + const., where the pseudo free energy for a uniform interaction, fC,U(φ), is defined as
Minimizing this function, we can see that there is a continuous transition as the interaction strength J is varied through a critical
point, J(c) = v, and obtain the stable equilibrium states as follows:
fC,U(φ) :=
φ2 + φ4.
(S47)
Continuous transition at J = J(c) = v
J < J(c) ⇒ Swollen-disordered state: φ
J > J(c) ⇒ Compact-disordered state: φ
∗ =
(1 ≤ i ≤ q)
∗ = 0, x∗
1
q
i =
(cid:114) J − J(c)
, x∗
i =
1
q
(1 ≤ i ≤ q).
2
(S48)
(S49)
(S50)
The obtained value of J(c) is consistent with the discussion in the subsection I D [take h = 0 and J(c)
i j
Figure S1 shows the J dependence of the order parameter φ∗ for the case of v = 0.1.
Note that the mean-field model of the coil-globule transition introduced in [24] is reduced to the same functional form as
fC,U(φ) in the limit of N → ∞ (see also Sec. II). Thus, with a uniform Potts-like interaction, the marks are always disordered
(x∗
= 1/q), and the properties of the polymer-Potts model are reduced to those of a usual polymer model with two-body and
three-body monomer-monomer interactions.
= J(c) in Eq. (S45)].
i
S8
FIG. S2. Two kinds of order parameters (φ∗ and m∗ = x∗
interaction.
1 − x∗
2) as a function of the interaction strength ( J) in the case of the Ising-type
2. Mark-specific interaction: Ising-type
√
kBTw.
Here and in the main text, we consider one of the simplest histone mark-specific interaction, the Ising-type interaction:
Ji j = J(2δi j − 1) with q = 2. According to the general discussion in the subsection I D [see Eq. (S45)], there is no second-
order transition in this system since the possible second-order transition point cannot be achieved irrespective of the value of
J = J/
It is natural to redefine the mark order parameter with the magnetization, m := x1 − x2. The numerically obtained order
parameters are shown in Fig. S2. In stark contrast to the non-mark-specific interaction case discussed in the subsection I E 1,
there is a first-order transition between the swollen-disordered state (φ∗ = 0, m∗ = 0) and the compact-ordered state (φ∗ > 0 and
m∗ > 0) at a v-dependent transition point, J(c).
Since the magnetization amplitude m∗ shows a jump from 0 to (cid:39) 1, we can obtain an approximate formula of the first-
order transition point J(c) and that of the J dependence of φ∗ in the compact state, as explained in the following. According to
Eqs. (S27), (S28), and (S29), the free energy of the swollen-disordered state is given as f C = fC(φ = 0, x1 = x2 = 1/2) = − ln 2.
On the other hand, the free energy of the compact-ordered state is given as f C (cid:39) minφ fC(φ, x1 = 1, x2 = 0) [= minφ fC(φ, x1 =
( J − v)/2, and finally we obtain in the
0, x2 = 1)] = minφ[(v − J)φ2 + φ4]. The last expression can be minimized with φ =
compact-ordered state f C (cid:39) −( J − v)2/4. At the transition point ( J = J(c)), the free energy of the swollen-disordered state and
that of the compact-ordered state coincide with each other: − ln 2 = −( J(c) − v)2/4, and we thus obtain J(c) (cid:39) v + 2
(cid:112)
ln 2.
√
To summarize, the stable equilibrium state for the Ising-type interaction is given as follows:
Discontinuous transition at J = J(c) (cid:39) v + 2
J < J(c) ⇒ Swollen-disordered state: φ
1
2
(i = 1, 2)
(S51)
(S52)
J > J(c) ⇒ Compact-ordered state: φ
∗ (cid:39)
(S53)
Let us shortly consider the case with finite fields (h1 = h and h2 = −h). Under sufficiently high fields, the possible second-order
2 , max
i
i
{x∗
i } (cid:39) 1, min
{x∗
i } (cid:39) 0.
transition point is given from Eq. (S45) as
√
ln 2
∗ = 0, x∗
(cid:114) J − v
i =
(cid:114)
J(c)(h) =
v
[tanh(h/kBT)]2 .
(S54)
As h gets larger and the mark type becomes virtually fixed, J(c)(h) gets close to v [18], which is the second-order transition point
of a homogeneous polymer with only mark-independent interactions [see Sec. I E 1 and Eq. (S48)]. In addition, for low fields
where the magnetization jump is expected around unity as in the zero-field case, we can estimate the first-order transition point
as in Eq. (S51):
These expressions [Eqs. (S54) and (S55)] are plotted in Fig. 3(b) in the main text and compared with the numerical results.
J(c)(h) (cid:39) v + 2
ln 2 − h
kBT
(S55)
(a) !"=0.01(b) !"=0.1(c) !"=1 S9
FIG. S3. Order parameters (φ∗ and m∗ = x∗
interaction.
1 − x∗
2) as a function of the interaction strength ( J) in the case of the 2-state (q = 2) Potts-type
FIG. S4. Order parameters (φ∗ and ∆x∗ = maxi{x∗
Potts-type interaction.
i } − mini{x∗
i }) as a function of the interaction strength ( J) in the case of the 3-state (q = 3)
FIG. S5. Order parameters (φ∗ and ∆x∗ = maxi{x∗
Potts-type interaction.
i } − mini{x∗
i }) as a function of the interaction strength ( J) in the case of the 10-state (q = 10)
Lastly, we consider external-force effects. In this case, the first-order transition occurs as a function of F and J. Since the
magnetization jump is almost unity, we can estimate the first-order transition point in a similar way to Eq. (S51):
(cid:114)
F(c)( J)b/kBT (cid:39)
3( J − v)2
2
− 6 ln 2,
(S56)
which is also compared with the numerical results [Fig. 3(f) in the main text].
(a) !"=0.1(b) !"=0.5(c) !"=2(a) !"=0.1(b) !"=0.5(c) !"=1.5(a) !"=0.1(b) !"=0.5(c) !"=1 S10
FIG. S6. Phase diagram in the J-v plane for the case of 2-state (q = 2) Potts-type interaction. The compact-state order parameter (φ∗) is
plotted as the color depth. The continuous transition line [Eq. (S57)] between the swollen-disordered and the compact-disordered states is
shown with a brown dotted line. Also, the approximate discontinuous transition line between the swollen-disordered and compact-ordered
states [Eq. (S58)] as well as that between the compact-disordered and compact-ordered states [Eq. (S59)] is shown with a black solid line. The
approximate cross point [Eqs. (S60) and (S61)] is given as J(cross) (cid:39) 3.33 and v(cross) (cid:39) 1.67.
3. Mark-specific interaction: q-state Potts-type interaction
As another example of a mark-specific interaction, let us think of the q-state Potts-type interaction: Ji j = Jδi j with q (≥ 2)
kinds of marks. In contrast to the Ising-type interaction case, according to Eq. (S45), there is a possible second-order transition
at
J(c)
2nd
= qv
(from swollen-disordered to compact-disordered)
In Fig. S3, the J dependence of the order parameters, φ∗ and m∗ := x∗
(S57)
2, is shown for the case of q = 2. For small values
of v [Figs. S3(a) and (b)], two-step transitions occur: the compact-disordered state (φ∗ > 0, m∗ = 0) starts to appear at J = J(c)
through a second-order transition, and then the compact-ordered state (φ∗ > 0 and m∗ > 0) emerges at J = J(c) through a
2nd
first-order transition. For large values of v [Fig. S3(c)], on the other hand, a single-step transition occurs: the compact-ordered
state emerges at J = J(c) through a first-order transition. Defining the magnetization as ∆x∗ := maxi{x∗
i }, we can see
that such qualitative features of the order parameters seem to be invariant even when the number of different marks (q) is larger
than 2 (see Figs. S4 and S5 for the case of q = 3 and q = 10, respectively).
For the case of large v, noticing that the magnetization size (∆x∗) shows a jump from 0 to (cid:39)1 we can estimate the first-order
transition point ( J(c)), just as in the case of the Ising-type interaction (see the subsection I E 2). Comparing the free energy in the
swollen-disordered state and that of the compact-ordered state, we can obtain
i } − mini{x∗
1 − x∗
(S58)
On the other hand, for the case of small v, a similar comparison of the free-energy difference between the compact-disordered
state and the compact-ordered state leads to −( J/q − v)2/4 − ln q = −( J − v)2/4, and we then obtain
(from swollen-disordered to compact-ordered).
ln q
(cid:112)
J(c) (cid:39) v + 2
(cid:115)
J(c) (cid:39) qv
q + 1
+
q
q + 1
v2 + 4
q + 1
q − 1
ln q
(from compact-disordered to compact-ordered).
(S59)
The approximate cross point (v(cross), J(cross)) of the first-order and second-order transition lines can be estimated with Eqs. (S57)
and (S58) [or Eqs. (S57) and (S59)] as
(cid:112)
ln q
q − 1
v(cross) (cid:39) 2
J(cross) = qv(cross) (cid:39) 2q
(cid:112)
ln q
q − 1 .
(S60)
(S61)
The numerically acquired phase diagram in the J-v plane for the case of q = 2 is shown in Fig. S6 with the analytic expressions
of the transition lines [Eqs. (S57), (S58), and (S59)].
Swollen-disordered(!∗=0, %∗=0)Compact-disordered(!∗>0, %∗=0)Compact-ordered(!∗>0, %∗>0) II. POLYMER MODEL WITH LONG-RANGE INTERACTIONS
In this section, we consider a polymer model with long-range interactions and show that this model is essentially equivalent
to the mean-field model of the coil-globule transition introduced in [24].
Let us consider a polymer model with long-range interactions:
S11
(S62)
where the Gaussian-chain term, HGC({Rn}), is defined in Eq. (S2), and the long-range volume exclusion interaction term,
HLR
VE(RN0), is defined in Eq. (S13). In this section, we assume that v can be positive or negative, while w is positive.
The equilibrium partition function, Zpolymer, is defined as
VE(RN0),
polymer
HLR
= HGC({Rn}) + HLR
Zpolymer :=
n=1
(cid:32)
dRn
(cid:90)
(S63)
.
polymer
kBT
exp
− HLR
(cid:33)
− 3
N(cid:89)
(cid:82)
dR δ(R− RN0) = 1, we can rewrite Zpolymer in a similar way to Eq. (S15)
R3 − w(cid:48)
δ(R − RN0) exp
(cid:32)
(cid:33)3(N−1)/2(cid:90)
2Nb2 − v(cid:48)
− 3R2
R3 − w(cid:48)
(Rn − Rn−1)2
− v(cid:48)
kBT
N(cid:88)
,
dR exp
N3
R6
N3
R6
(S65)
(S64)
dRn
kBT
2b2
(cid:33)
N2
N2
n=1
kBT
kBT
.
Introducing the end-to-end vector R with an identity,
as
Zpolymer =
dR exp
(cid:90)
n=1
×
(cid:90)
N(cid:89)
(cid:32)2πb2
(cid:34)3(N − 1)
N3/2
3
1
Zpolymer =
(cid:90) ∞
0
where v(cid:48) = 3v/4π and w(cid:48) = (3/4π)2w as in Sec. I. Through the same calculation process as Eq. (S16), we obtain
Performing the solid-angle integration of R and exponentiating all the factors lead to the following form:
Zpolymer =
dR exp
ln
2πb2
3
2
Following [24], we define the expansion factor, α := R/
Zpolymer =
where Feff(α) is an effective free energy defined as
2Nb2 − v(cid:48)
kBT
N2
R3 − w(cid:48)
kBT
ln N + ln(4π) + 2 ln R − 3R2
− 3
2
√
(cid:35)
Nb, which leads to
(cid:34)
(cid:90) ∞
dα exp
+ const.
,
− Feff(α)
kBT
0
:=
Feff(α)
kBT
3
2 α2 − 2 ln α +
x
α3
+
y
2α6
(cid:35)
.
N3
R6
(S66)
(S67)
(S68)
with x := v(cid:48) √
N/kBTb3 and y := 2w(cid:48)/kBTb6. The effective free energy, Feff(α), has the same functional form as the mean-
field free energy, Eq. (1) in [24], apart from the numerical coefficient of the second term [−3 ln α in [24] instead of −2 ln α
in Eq. (S68)]. This numerical difference of the coefficient only produces quantitative difference in the coil-globule transition
properties for finite N, and moreover, the term proportional to ln α becomes irrelevant in the limit of N → ∞ since ln α =
o(N) (see also the discussion in the subsection I E 1). Therefore, the polymer model with long-range two-body and three-body
interactions is essentially equivalent to the mean-field free energy introduced in [24].
III. RELATION TO MICROSCOPIC MODELS
In this section, we employ the virial expansion method to obtain a relation between the pseudo free energy of the polymer-
Potts model and a microscopic Hamiltonian describing a polymer with mark-dependent interactions.
Let us consider the microscopic two-body interaction term in the Hamiltonian of a virtual gas of q types of monomers:
A. Gas model
(cid:88)
Hint =
1≤n<m≤N
U(rn − rm; sn, sm)
S12
(S69)
Here, rn and sn are the position and mark of the n-th monomer, respectively, and U(r− r(cid:48); s, s(cid:48)) is the interaction energy between
one monomer with the mark s placed at r and another monomer with the mark s(cid:48) placed at r(cid:48). A basic inversion symmetry,
U(rn − rm; sn, sm) = U(rm − rn; sn, sm) = U(rm − rn; sm, sn) is assumed. In the following, U(rn − rm; sn, sm) is also expressed
as Unm for simplicity.
Extending the standard procedure of the virial expansion [46] to the multiple-mark case, we can formally expand the interac-
tion part of the partition function Q := V−N(cid:82) (cid:81)N
− 1
exp
(cid:90)
1
V N
d3rn
Q =
(cid:90) N(cid:89)
1
(cid:88)
= exp
n=1
V2
1≤n<m≤N
B. Virial expansion
kBT
Unm
(cid:88)
(cid:88)
n=1 d3rn exp(−Hint/kBT) as
(cid:90)
(cid:32)
1≤n<m≤N
1
V3
1≤n<m<k≤N
d3rnd3rm fnm +
where
fnm = f (rn − rm; sn, sm) = exp
− Unm
kBT
− 1
d3rnd3rmd3rk fnm fnk fmk + ···
(cid:33)
,
(S70)
(S71)
is the Mayer f-function.
Using an abbreviation, U(i j)
12
virial coefficient as
(cid:88)
(cid:90)
1≤n<m≤N
:= U(r1 − r2; i, j) and correspondingly f (i j)
:= f (r1 − r2; i, j), we can obtain the second-order
(cid:88)
(cid:88)
(cid:90)
(cid:90)
12
2
d3r1d3r2 f (i j)
12
+ O(N0).
(S72)
d3rnd3rm fnm =
d3r1d3r2 f (ii)
12
+
NiN j
2V2
1≤i(cid:44) j≤q
Using similar expressions, we can also obtain the third-order virial coefficient as
d3rnd3rmd3rk fnm fnk fmk =
d3r1d3r2d3r3 f (ii)
12 f (ii)
13 f (ii)
23
+
(cid:88)
(cid:90)
2N j
Ni
2V3
d3r1d3r2d3r3 f (ii)
12 f (i j)
13 f (i j)
23
d3r1d3r2d3r3 f (i j)
+ O(N0).
(S73)
1≤i(cid:44) j≤q
12 f (ik)
13 f ( jk)
23
1
V2
(cid:90)
(cid:88)
1
V3
1≤n<m<k≤N
Ni
2V2
1≤i≤q
(cid:90)
(cid:88)
1≤i≤q
+
3
Ni
6V3
(cid:88)
1≤i(cid:44) j(cid:44)k(cid:44)i≤q
(cid:90)
NiN jNk
6V3
Higher-order virial coefficients can be obtained in a similar manner.
C. Relation to polymer-Potts model with Ising-type interaction
Here, let us consider the Ising-type case where q = 2 and U(r; 1, 1) = U(r; 2, 2) (and thus f (11)
12
and
v =
kBT
V
J =
kBT
V
(cid:90)
(cid:90)
−1
4
(cid:32)
(cid:32)1
4
(cid:90)
12 − 1
4
(cid:90)
(cid:33)
d3r1d3r2 f (12)
12
(cid:33)
,
d3r1d3r2 f (12)
12
d3r1d3r2 f (11)
d3r1d3r2 f (11)
12 − 1
4
= f (22)
12 ). Taking v and J as
(S74)
(S75)
we can readily obtain from Eq. (S72)
(cid:88)
(cid:90)
1
V2
1≤n<m≤N
(cid:16)
vρ2 − Jm2ρ2(cid:17)
+ O(N0),
d3rnd3rm fnm = − V
kBT
S13
(S76)
where the density ρ = N/V and the magnetization m = (N1 − N2)/N are introduced. The form of Eq. (S76) corresponds to the
two-body interaction terms in the polymer-Potts pseudo free energy [Eq. (1) in the main text] if we take V ∼ R3 with the polymer
length R.
(cid:33)
(cid:33)
,
(S77)
(S78)
(S79)
(cid:90)
23 − 1
f (11)
8
(cid:90)
1
8
d3r1d3r2d3r3 f (11)
12
f (11)
13
d3r1d3r2d3r3 f (11)
12
f (12)
13
f (12)
23
d3r1d3r2d3r3 f (11)
12
f (11)
13
f (11)
23
+
d3rnd3rmd3rk fnm fnk fmk = − V
kBT
d3r1d3r2d3r3 f (11)
12
f (12)
13
f (12)
23
(cid:16)
wρ3 + w(cid:48)m2ρ3(cid:17)
+ O(N0).
(cid:32)
(cid:32)
In a similar way, taking w and w(cid:48) as
− 1
24
kBT
V
w =
and
w(cid:48) =
kBT
V
−1
8
we can obtain from Eq. (S73)
1
V3
(cid:88)
1≤n<m<k≤N
(cid:90)
(cid:90)
(cid:90)
(cid:34)
(cid:16)
(cid:17)
(cid:35)
The first term of Eq. (S79) corresponds to the three-body interaction term in the polymer-Potts pseudo free energy [Eq. (1) in
the main text].
To sum up, the interaction part of the partition function can be represented as
− V
kBT
vρ2 − Jm2ρ2 + wρ3 + w(cid:48)m2ρ3 + O(ρ4)
Q = exp
(S80)
Thus, replacing V = const.×R3 and adding the entropic contribution, we obtain the mean-field pseudo free energy (per monomer)
of the polymer-Potts model in N → ∞ limit:
(cid:16)
1 + w(cid:48)m2(cid:17)
[(1 + m) ln (1 + m) + (1 − m) ln (1 − m)] + O
v − Jm2(cid:17)
(cid:16)
φ6(cid:17)
(S81)
φ4 +
φ2 +
(cid:16)
=
.
.
+ O(N0)
√
Here, the dimensionless globular order parameter φ = (w/kBT)1/4 √
kBTw,
kBTw, and w(cid:48) = w(cid:48)/w are used as in the main text. Note that the term proportional to m2φ4 does not appear in the
J = J/
minimal model considered in the main text. Neglecting the higher-order terms O(φ6) (although, strictly speacking, this operation
is not justified when a discontinuous coil-globule transition occurs), we can discuss based on Eq. (S81) the phase transition
nature of the polymer-Potts model with general interactions satisfying the Ising symmetry [U(r; 1, 1) = U(r; 2, 2)].
ρ and the dimensionless coupling constants v = v/
√
f (φ, m)
kBT
1
2
D. Virial expansion of specific model
In the simulation of [19], three kinds of marks are assumed (1 ≤ i ≤ 3): two of them (i = 1, 2) with Ising-type interactions
explained below and one of them (i = 3) with no mark-specific interactions. For simplicity, we here neglect this third mark
although its existence may shift the transition point slightly, as can the O(φ6) terms. The polymer model with two kinds of
marks, where the interaction between a monomer with the mark i and another monomer with the mark j (1 ≤ i, j ≤ 2) is given
as the truncated Lennard-Jones type [19]:
(cid:18) σ
r
(cid:19)12 −(cid:18) σ
(cid:19)6 −
r
σ
r(i j)
c
12
σ
r(i j)
c
+
6 (for r ≤ r(i j)
c
U(i j)
LJ (r) =
4(i j)
N
),
(S82)
LJ (r) = 0 (for r > r(i j)
c
and U(i j)
interaction strength, r(i j)
so that minr U(ii)
LJ (r) = − [see Fig. S7(a)].
). Here, σ corresponds to the interaction range, i j = δi j + kBT(1− δi j) represents the mark-specific
c = 1.8σδi j +21/6σ(1− δi j) is the mark-specific interaction cutoff, and N = 1 +4(1.8−12−1.8−6) is chosen
Noticing the Ising symmetry of U(i j)
LJ (r) as U(r; i, j) in Sec. III C and calculate
the dimensionless coupling constants v, J, and w(cid:48) based on Eqs. (S74), (S75), (S77), and (S78). Since the Ising-type interaction
LJ (r), we identify U(i j)
LJ (r), i.e., U(11)
LJ (r) = U(22)
S14
FIG. S7. (a) Truncated Lennard-Jones-type interaction potential [Eq. (S82)]. (b) Dimensionless coupling constants [v (blue solid line), J (red
dashed line), and w(cid:48) (purple dotted line)] and (c) equilibrium order parameters [φ∗ (green solid line) and m∗ (orange dashed line)] as a function
of the interaction strength /kBT for the case of the Lennard-Jones-type interaction [Eq. (S82)].
strength can be quantified by a dimensionless parameter /kBT, we plot v, J, and w(cid:48) as a function of /kBT in Fig. S7(b). The
mark-specific interaction strength J is an increasing function of /kBT as expected, while v and w(cid:48) also depend on the value of
/kBT since also affects the mark-nonspecific interactions.
By minimizing Eq. (S81), we can obtain the equilibrium globular order parameter φ∗ and magnetization m∗, as shown in
Fig. S7(c) as a function of /kBT. There is a first-order transition between the swollen-disordered state (φ∗ = 0, m∗ = 0) and the
compact-ordered state (φ∗ > 0, m∗ > 0) at /kBT (cid:39) 0.85, which is of the same order as the critical value (/kBT (cid:39) 0.9) obtained
in the simulation with N = 2000 [19].
Lastly, we stress that the necessary condition for a second-order coil-globule transition derived in the main text [see the
discussions around Eq. (6)] is applicable even when m2φ4 terms or O(φ6) terms are present in the pseudo free energy as Eq. (S81).
Indeed, v is positive around the transition point [see Figs. S7(b) and (c)], and thus the second-order coil-globule transition is
prohibited, consistent with the numerical result.
(a)(b)(c) |
1412.3981 | 7 | 1412 | 2017-02-03T13:24:03 | Do biological molecular machines act as Maxwell's demons? | [
"physics.bio-ph",
"cond-mat.stat-mech"
] | The nanoscopic isothermal machines are not only energy but also information transducers. We show that the generalized fluctuation theorem with information creation and entropy reduction can be fulfilled for the enzymatic molecular machines with the stochastic dynamics, which offers a choice of the work performance in a variety of ways. A model of such dynamics, specified by a critical complex network, is investigated. The main conclusion of the study is that the processing of free energy has to be distinguished from the processing of organization, which we identify with an adequately defined thermodynamic variable. Maxwell's demon utilizes entropy reduction for creation of information, which, from the former point of view, may be used for a reduction of energy losses, hence ultimately, for the performance of work. From the latter point of view, however, it can be used for other purposes, for example molecular recognition. This can be the case of biological molecular machines. From the biological perspective, the ascertainment is important, that the information creation and storage take place in the long lasting transient stages before completing the free energy transduction cycles. From a broader physical perspective, a supposition could be of special importance, that information is a change of organization, the thermodynamic function of state of the system. | physics.bio-ph | physics | Do biological molecular machines act as Maxwell's demons?
Michal Kurzynski∗, Przemyslaw Chelminiak
Faculty of Physics, A. Mickiewicz University, Umultowska 85, 61-614 Poznan, Poland
7
1
0
2
b
e
F
3
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
7
v
1
8
9
3
.
2
1
4
1
:
v
i
X
r
a
Abstract
The nanoscopic isothermal machines are not only energy but also information transducers. We show that the
generalized fluctuation theorem with information creation and entropy reduction can be fulfilled for the enzymatic
molecular machines with the stochastic dynamics, which offers a choice of the work performance in a variety of
ways. A model of such dynamics, specified by a critical complex network, is investigated. The main conclusion of
the study is that the processing of free energy has to be distinguished from the processing of organization, which we
identify with an adequately defined thermodynamic variable. Maxwell's demon utilizes entropy reduction for creation
of information, which, from the former point of view, may be used for a reduction of energy losses, hence ultimately,
for the performance of work. From the latter point of view, however, it can be used for other purposes, for example
molecular recognition. This can be the case of biological molecular machines. From the biological perspective, the
ascertainment is important, that the information creation and storage take place in the long lasting transient stages
before completing the free energy transduction cycles. From a broader physical perspective, a supposition could be of
special importance, that information is a change of organization, the thermodynamic function of state of the system.
Keywords: Thermodynamics far from equilibrium, Fluctuation theorem, Native protein dynamics, Critical complex
networks
PACS: 05.70.Ln, 87.15.H-, 87.15.Ya, 89.75.Hc
1. Introduction
In the intention of its creator [1], Maxwell's demon was thought to be an intelligent being, able to perform work at
the expense of the entropy reduction of a closed operating system. The perplexing notion of the demon's intelligence
was formalized in terms of memory and information processing by Szilard [2], Landauer [3] and subsequent followers
[4–6], who pointed out that, in order for the total system to obey the second law of thermodynamics, the entropy
reduction should be compensated for by, at least, the same entropy increase, related to the demons information gain
on the operating system's state.
The present, almost universal consensus on this issue is expressed in terms of the feedback control [7–12]. First,
information is transferred from the operating system to memory in the process of measurement (observation) and
next, when the system is externally loaded, the transfer of information from memory to the operating system con-
trols the work performance process. The both processes can occur simultaneously [13–18]. Following Landauer's
principle, the memory content must be erased at the expense of some entropy production. It should be stressed that
the information transfer may [19] but in general is not related to energy transfer between the operating system and
memory.
A non-informational formulation of the problem was proposed by Smoluchowski [20] and popularized by Feyn-
man [21] as the ratchet and pawl machine. It can operate only in agreement with the second law, at the expense of
an external energy source [22]. A. F. Huxley [23] and consequent followers [24, 25] adopted this way of thinking
to suggest numerous ratchet mechanisms for the protein molecular machines action, but no entropy reduction takes
place for such models [22], thus, they do not act as Maxwells demons.
∗Corresponding author
Email address: [email protected] (Michal Kurzynski)
Preprint submitted to Elsevier
July 11, 2018
More general models of protein dynamics have been put forward [26–34] with a number of intramolecular states
organized in a network of stochastic transitions. Here we show that if such models offer work performance in a variety
of ways [34], the generalized fluctuation theorem [7–18, 35–37] is fulfilled with possible entropy reduction. A ques-
tion appears as to whether they can be considered to act as Maxwell's demons like the artificial nanoscopic machines
recenly constructed [38–41]. The hypothetical computer model of the network with the Markovian stochastic dynam-
ics is studied, displaying, like networks of the systems biology [42, 43], a transition from the fractal organization on
a small length-scale to the small-world organization on the large length-scale.
We start from the general theory of free energy transduction in mesoscopic isothermal machines and the relation-
ship between the entropy and information production. Then, we present the protein molecular machines as isothermal
chemo-chemical machines and state a specific model of the proteins stochastic dynamics. A study of this model,
using computer simulations, leads us to the result that the free energy transduction in the fluctuating systems must
be distinguished from the arrangement transduction. We identify arrangement with an adequately defined thermody-
namic variable and show that its increase, related to an entropy reduction and an information creation, takes place in
the transient stage before completing the free energy transduction cycle. The latter, in the case of protein machines,
can last quite long. Some biological as well as general physical implications of this statement are the subject of the
concluding section.
2. Theory: Formulation of the problem
2.1. Stationary isothermal machines
A long story has been made up for it that the word machine has several different meanings. In our context, we
understand a machine to be any physical system that enables two other systems to perform work on one another
[44]. The demon considered by Maxwell changed the temperature to perform work at the expense of absorbed
heat. However, the biological molecular machines, like the machine considered by Szilard, operate at a constant
temperature. Under isothermal conditions, the internal energy is uniquely divided into the free energy, the component
that can be turned into work, and the bound energy (the entropy multiplied by the temperature), the component that
can be turned into heat [45]. Both the thermodynamic quantities can make sense in the non-equilibrium steady state,
if the latter is treated as a partial equilibrium state [44]. The free energy can be turned irreversibly into the bound
energy in the process of internal entropy production, having a meaning of the energy dissipation [44]. In accordance
to such internal energy division, the protein molecular machines are referred to as free energy transducers [46], thus,
when supposed to act as Maxwells demon, they should perform work at the expense of a part of the dissipation.
During the stationary isothermal processes, both the free energy and the bound energy remain constant. The
energy processing pathways in any stationary isothermal machine are shown in Fig. 1(a), where the role of all the
physical quantities being in use is also indicated. Xi denotes the input (i = 1) and the output (i = 2) thermodynamic
variable, Ai is the conjugate thermodynamic force and the time derivative, Ji = dXi/dt, is the corresponding flux. T is
the temperature and S is the entropy. The thermodynamic variables X1 and X2 may be mechanical – displacements,
electrical – charges, or chemical – numbers of distinguished molecules, hence the fluxes J1 and J2 may be velocities,
or electrical or chemical current intensities, respectively. The conjugate forces are then the mechanical forces, or the
differences of electrical or chemical potentials (voltages or affinities).
Machines often work as a gear. To clearly specify the transmission ratio n between the fluxes J1 and J2, determin-
ing their tight coupling J2 = nJ1, it is important that the thermodynamic variables X1 and X2 be dimensionless. Then,
the fluxes are counted in the turnover numbers of the machine per unit time and the forces are of energy dimension.
By convention, the fluxes J1 and J2 are assumed to be of the same sign. Then, one system performs work on the other
when the forces A1 and A2 are of the opposite sign. We assume J1, J2, A1 > 0 and A2 < 0 throughout this paper.
According to the second law of thermodynamics, the net dissipation flux (the internal entropy production rate,
multiplied by the temperature) A1J1 + A2J2 must be nonnegative. However, it consists of two components. The first
component, (A1 + nA2)J1, achieved when the input and output fluxes are tightly coupled, J2 = nJ1, and the output
energy flux A2J2 = nA2J1 is completely transmitted between the thermodynamic variables X1 and X2 (see Fig. 1(a)),
must also be, according to the same law, nonnegative. Open to discussion is the sign of the complement A2(J2 − nJ1)
of (A1 + nA2)J1 to the net dissipation flux A1J1 + A2J2.
In the macroscopic systems, the entropy S is additive and can always be divided into the following two parts S 1
and S 2, relating to the input and output thermodynamic variables X1 and X2, respectively. As a consequence, the flux
2
(a)
(b)
Constraints
Constraints
A J1
1
A J2
2
Work flux
A J1
1
nA J2
1
A J nJ
-
(
2
2
)
1
-nA J2
1
-A X1
1
-A X2
2
Free energy
(
)
A nA J
+
1
2
1
A J nJ
(
-
2
2
1
)
Dissipation flux
-(A nA X
+
)
1
2
1
-
A X nX
-
(
2
2
)
1
( +
)
A nA J
2
1
1
A J nJ
-
(
2
2
)
1
Work flux
Free energy
Dissipation flux
TS
Bound energy
Heat flux
Entropy
Internal dynamics
A J A J
+
1
1
2
2
t
n
e
m
n
o
r
i
v
n
E
Heat bath
Figure 1: Energy processing in the stationary (cyclic) isothermal machine. See the text for the notation explanation. The constraints keep fixed
stationary values of the thermodynamic variables Xi. We assume these variables to be dimensionless, hence the conjugate forces Ai are of energy
dimension and the fluxes Ji = dXi/dt are counted in the turnover numbers of the machine per unit time. (a) The division of the machine's internal
energy into the free energy F = −A1X1 − A2X2 and the bound energy T S . In the steady state, the work flux (the resultant power) equals the
dissipation flux, and that equals the heat flux. The transmission ratio n determines the free energy flux from −A1X1 to −A2X2. By convention, the
fluxes J1 and J2 are assumed to be of the same sign. Then, one system performs work on the other when the forces A1 and A2 are of the opposite
sign. The directions of the energy fluxes shown are for J1, J2, A1 > 0 and A2 < 0, what is assumed throughout this paper. The actual direction of
the flux A2(J2 − nJ1), denoted by the forward-reverse arrow, is the subject of this research. (b) The alternative view of free energy processing in
the stationary isothermal machine, determined in the text. Here, both the thermodynamic variables X1 and X2 − nX1 are energetically independent.
Only the free energy is specified. The bound energy and the environment are considered to be determined jointly by the internal dynamics of the
machine, modified by an interaction with the environment.
A2(J2 − nJ1), which corresponds to S 2, must also be, under isothermal conditions, nonnegative. This means that, for
the assumed negative A2, the output flux J2 should not surpass more than n times the input flux J1. Macroscopically,
the second component of the dissipation flux has the obvious interpretation of a slippage in the case of the mechanical
machines, a short-circuit in the case of the electrical machines, or a leakage in the case of pumps. However, because
of non-vanishing correlations within the bound energy subsystem [7–18, 35–37], in the mesoscopic systems like the
protein molecular machines, entropy S is not additive and cannot be divided into two parts like the free energy. This
allows the transfer of information within the bound energy subsystem, which could result in the partial reduction of
energy dissipation.
In fact, for the biological molecular machines, the output flux J2 can surpass the input flux J1 [34]. Such a
surprising case was observed by Yanagida and his co-workers [47, 48], who found that the single myosin II head
can take several steps along the actin filament per ATP molecule hydrolyzed. Whether it changes the sign of the flux
A2(J2 − nJ1) to the negative depends on establishing the value of the transmission ratio n, which, as opposed to the
macroscopic machines, is not a simple task in the case of molecular machines, and is one of the main topics of the
present paper.
From the point of view of the output force A2, subsystem 1 carries out work on subsystem 2 while subsystem
2 carries out work on the environment. Jointly, the flux of the resultant work (the resultant power) A2(J2 − nJ1)
is driven by the force A2. The complement to A1J1 + A2J2 is the flux (A1 + nA2)J1 driven by the force A1 + nA2.
Consequently, the free energy processing from Fig. 1(a) can be alternatively presented as in Fig. 1(b), with the free
energy transduction path absent. Here, the subsystems described by two variables X1 and X2 − nX1, respectively,
are energetically independent. However, like the subsystems described by the two variables X1 and X2, they are still
statistically correlated. Note that in Fig. 1(b), only the free energy is specified. The bound energy and the environment
are considered as determined jointly by the internal dynamics of the machine, modified by an interaction with the
environment.
2.2. Generalized fluctuation theorem
In the mesoscopic machines, the work, dissipation and heat are fluctuating random variables and their variations,
proceeding forward and backward in time, are related to each other by the fluctuation theorem [35–37].
In fact,
the feedback control description of the Maxwell's demon action [7–12], mentioned in the Introduction, was based
3
on the fluctuation theorem. For the stationary process, the probability distribution function for the input and output
fluxes, in general, depending on the time period t of determination, satisfies the stationary fluctuation theorem in the
Andrieux-Gaspard form [49, 50]:
p( j1(t), j2(t))
p(− j1(t), − j2(t))
= exp β[A1 j1(t) + A2 j2(t)]t .
(1)
Here, β = 1/kBT , where kB is the Boltzmann constant, and p is the joint probability distribution function for the
statistical ensemble of the fluxes j1(t) and j2(t) over the time period t and their inverses. Eq. (1) can be equivalently
rewritten as the Jarzynski equality [51]
with the stochastic dimensionless entropy production (the energy dissipation divided by kBT )
hexp(−σ)i = 1
σ = Xi
βAiJi(t)t .
(2)
(3)
Ji(t) in (3) denotes the random variable of the mean net flux over the time period t, (i = 1, 2), whereas ji(t) in (1)
denotes its particular value. h. . .i is the average over the ensemble of the fluxes j1(t) and j2(t). Time t must be long
enough for the considered ensemble to comprise only stationary fluxes. The convexity of the exponential function
provides the second law of thermodynamics:
hσi ≥ 0
(4)
to be a consequence of (2). Only the averages of the random fluxes can be identified with the stationary fluxes Ji.
They are time-independent, hJi(t)i = Ji for arbitrary t.
In further discussion, for brevity, we will omit the argument t specifying all the fluxes. In the context of the
transition from Fig. 1(a) to (b), the two-dimensional probability distribution function p( j1, j2) can be treated as a
two-dimensional probability distribution function of two variables j1 and j2 − n j1, with j2 − n j1 as a whole treated as
a single variable. If we calculate the marginal probability distributions, then, from the fluctuation theorem (1) for the
total entropy production in both the stationary fluxes J1 and J2, the generalized fluctuation theorems for J1 and the
difference J2 − nJ1 follow, respectively, in the logarithmic form:
ln
p( j1)
p(− j1)
= β(A1 + nA2) j1t + βA2h(J2 − nJ1)ti −*ln
p(J2 − nJ1 j1)
p(−J2 + nJ1 − j1)+
(here, the averages h. . .i are taken over the ensemble of the flux differences j2 − n j1) and
ln
p( j2 − n j1)
p(− j2 + n j1)
= βA2( j2 − n j1)t + β(A1 + nA2)hJ1ti −*ln
p(J1 j2 − n j1)
p(−J1 − j2 + n j1)+
(5)
(6)
(here, the averages h. . .i are taken over the ensemble of the fluxes j1). Above, we introduced conditional probabilities.
The first components on the right of Eqs. (5) and (6) describe the entropy production, now only in the separate fluxes
J1 and J2 − nJ1, respectively, but the interpretation of the remaining components is not so easy. The problem is that for
the stationary processes, as opposed to the transient nonequilibrium protocols considered in Refs. [7–11], the notion
of the mutual information is not well defined, as it should be exchanged continuously, without any delay [12].
This disadvantage may, however, be used to determine the value of the transmission ratio n. If it is chosen such
that the remaining terms in Eq. (5) cancel each other out and only the entropy production term remains:
ln
p( j1)
p(− j1)
= β(A1 + nA2) j1t ,
then, Eq. (6) for the flux J2 − nJ1 can be rewritten in terms of the partly averaged mutual information differences:
ln
p( j2 − n j1)
p(− j2 + n j1)
= βA2( j2 − n j1)t −*ln
p(J1, j2 − n j1)
p(J1)p( j2 − n j1)+ +*ln
p(−J1, − j2 + n j1)
p(−J1)p(− j2 + n j1)+ .
4
(7)
(8)
Like the antecedent entropic term, both the informative components in Eq. (8) depend only on the single variable
j2 − n j1.
The replacement of (5) by (7) and (6) by (8) corresponds to treating the thermodynamics of the system as bipartite
[12]. Although the fully averaged mutual information does not distinguish between the sender from the recipient,
the unilaterally averaged informations may differ substantially [6].
In our case, we are dealing with completely
asymmetrical coarse graining of the fluxes J1 and J2 − nJ1. This means that J1, when averaged over the ensemble
of the flux differences j2 − n j1, is unable to make any choice related to the creation of information, whereas the
flux difference J2(t) − nJ1(t), when averaged over the ensemble of the fluxes j1, is able to choose to create some
information. In other words, Eq. (7) can be considered as the condition for J1 to be a hidden thermodynamic variable,
from which and to which the information does not flow [16, 17, 52, 53]. The specific value of n results from the internal
dynamics of the system; an illustrative example for the biological molecular machine will be presented further on in
Fig. 6, which is a confirmation of the method of the reduction ratio designation that we have chosen.
To conclude, we write the fluctuation theorem for the flux J1, Eq. (7), in the form analogous to the Jarzynski
equality (2), from which the second law inequality (4) follows. However, the Jarzynski equality for the flux J2 − nJ1,
Eq. (8), should be written in the generalized form [7–18, 35–37]:
hexp(−σ − ι)i = 1 ,
from which the generalized second law inequality follows:
hσi + hιi ≥ 0 .
As in Eq. (2), σ has the meaning of the random dimensionless entropy production, while the additional quantity
ι = − ln
p(J1, J2 − nJ1)
p(J1)p(J2 − nJ1)
+ ln
p(−J1, −J2 + nJ1)
p(−J1)p(−J2 + nJ1)
(9)
(10)
(11)
represents the difference of the stochastic information, which the fluctuating flux J2(t) − nJ1(t) sends outside to the
flux J1(t), when proceeding, respectively, in the forward and backward directions. Accordingly, hιi has the direct
interpretation of the information production by the flux J2 − nJ1. Because the variable J1 is hidden, this information
production is, from the point of view of this variable, the information loss, hence positive. This statement remains
true for any nanoscopic isothermal machine. Whether the complementing dissipation hσi in Inequality (10) may be
negative, it depends on the specific dynamics of the system. Further on, we study this problem for the biological
molecular machines.
2.3. Proteins as chemo-chemical machines
From a theoretical point of view, it is convenient to treat all biological molecular machines as chemo-chemical
machines [44]. The protein chemo-chemical machines are enzymes, that simultaneously catalyze two effectively
unimolecular reactions: the free energy-donating input reaction R1 ↔ P1 and the free energy-accepting output reaction
R2 ↔ P2 (Fig. 2(a)). Also, pumps and molecular motors can be treated in the same manner. Indeed, the molecules
present on either side of a biological membrane can be considered to occupy different chemical states (Fig. 2(b)),
whereas the external load influences the free energy involved in binding the motor to its track (Fig. 2(c)), which can
be expressed as a change in the effective concentration of this track [27, 28, 33, 44].
The system considered consists of a single enzyme macromolecule, surrounded by a solution of its substrates,
possibly involving the track (Fig. 2). It is an open system with constraints controlling the mean numbers of incoming
and outgoing molecules and, in particular, the number of steps performed by the motor along the track. Under
specified relations between the concentration of the substrates [34], two independent stationary (nonequilibrium)
molar concentrations of the products [P1] and [P2], related to the enzyme total concentration [E], are to be treated as
the input and output dimensionless thermodynamic variables X1 and X2, respectively, presented in Fig. 1. The fluxes
Ji with the conjugate thermodynamic forces Ai are determined as [44, 46]
Ji ≡
d
dt
Xi =
d
dt
[Pi]
[E]
,
βAi = ln
[Pi]eq
[Ri]eq
[Ri]
[Pi]
,
5
(12)
(a)
R1
P1
P1
R1
R2
R1
P2
R1
R2
R1
R2
P2
P2
(b)
R1
P1
R1
P1
R2
R1
R1
R2
R2
(c)
P
R
P
R
P
R R
R
R1
P2
P2
P2
Figure 2: A simplified representation of the three types of the biological molecular machines: (a) enzymes that simultaneously catalyze two
reactions, (b) pumps placed in a biological membrane, and (c) motors moving along a track. Constraints are symbolized by the frame of the broken
line. It shoud be stressed that the entire system inside the frame is the machine: the number of substrate molecules determines its free energy (the
thermodynamic state) whereas the enzyme's internal dynamics determines its bound energy (the entropy).
where the superscript eq denotes the corresponding equilibrium concentrations. The fluxes can be calculated if a
model of the systems dynamics is at disposal [33, 34].
It is now well established that most if not all enzymatic proteins display slow stochastic dynamics of transitions
between a variety of conformational substates composing their native state [26, 54–58]. Two types of experiments
imply this statement. The first includes observations of the non-exponential initial stages of a reaction after the special
preparation of an initial microscopic state in a statistical ensemble of biomolecules by, e.g., the laser pulse [59, 60].
The second type of experiments is imaging and the manipulation of single biomolecules in various processes [61, 62].
The even more convincing proof of the conformational transition dynamics of native proteins has been provided by
molecular dynamics simulations [63, 64]. Here, as a symbol of progress made recently, the study of conformational
transitions in intrinsically disordered native kinases could be quoted, which a few years ago resulted in a network of
25 substates [65] and now offers a network of several hundred items [66].
It follows that on the mesoscopic level, the dynamics of a specific biological chemo-chemical machine is the
Markov process described by a system of master equations, determining a network of conformational transitions that
satisfy the detailed balance condition (we mentioned more representative examples in the Introduction [26–34]), and
a system of pairs of distinguished nodes (the gates) between which the input and output chemical reactions force
transitions, that break the detailed balance [33, 34] (Fig. 3(a)). Recently, we proved analytically [34] that, for a single
output gate, when the enzyme has no opportunity for any choice, the ratio ǫ = J2/J1 cannot exceed one. This case
also includes the various ratchet models [24, 25], which assume the output and input gates to coincide. The output
flux J2 can only exceed the input flux J1 in the case of many output gates, that seems to be the rule rather than the
exception (the output 'fluctuating reaction rate') [26, 67–71].
At this point, it is worth explaining demonstratively the relationship between the degree of coupling ǫ and the
transmission ratio n, which is essential for determining the direction of the flux A2(J2 − nJ1) in Figs. 1(a) and (b). For
the assumed negative force A2, the sign of A2(J2 − nJ1) = A2(ǫ − n)J1 can be negative only if ǫ > n. The value of
the degree of coupling ǫ depends, in general, on the values of the forces Ai [33, 34, 44], whereas the specific value
of n results from the topology of the network. Two examples of extremely simple networks with two output gates
are presented in Fig. 3(b). In the absence of internal transitions indicated by dashed lines, for the gates connected in
series, both n = 2 and ǫ = 2, independently on the forces Ai, whereas for the gates connected in parallel, both n = 1
and ǫ = 1. Consequently, the flux A2(J2 − nJ1) = A2(ǫ − n)J1 vanishes in both cases, which is obvious, since there is
no short circuit of the input gate that omits the output gates. The only difference between both examples is that in the
case of the gates connected in series, the system passes along the output gates successively, whereas in the case of the
gates connected in parallel, it has a choice which output gate to pass along first. However, by assumption, the network
of internal conformational transitions should be one-component, not broken, so that transitions shown by dashed lines
must be taken into account. It allows the choice of the output gate in both examples and results in decreasing the value
of ǫ [34]. Adding yet other possible transitions changes the value of the transmission ratio n to be established in the
whole range from 1 to 2.
Determination of the values of ǫ and n in actual, more complex networks is not a simple task. Our goal in the
following is to consider the biological molecular machines with a dynamics with many output gates allowing a choice.
Since very poor experimental support is still available for actual conformational transition networks in native proteins,
we restrict our attention only to a model network.
6
(a)
R1
P1
(b)
(c)
1''
2'
M
1'
2''
P2
R2
1''
1'
1''
1'
2'b
2'a
2''b
2''a
2'b
2'a
2''b
2''a
2'a
2'c
2'b
2''b
2''d
2''c
1''
2'd
2''a
1'
Figure 3: The kinetics of the enzymatic chemo-chemical machine. (a) General scheme. The grey box represents an arbitrary one-component
network of transitions between conformational substates, composing either the enzyme or the enzyme-substrates native state [34]. All these
transitions satisfy the detailed balance condition. A single pair (the gate) of conformational states 1′′ and 1′ is distinguished, between which the
input reaction R1 ↔ P1 brakes the detailed balance. Also, a single or a variety (ovals) of pairs of conformational states 2′′ and 2′ is distinguished,
between which the output reaction R2 ↔ P2 does the same. All the reactions are reversible; the arrows indicate the directions assumed to be
forward. (b) The simplest examples of the network of transitions with two output gates connected in series (up) or in parallel (down). In the
absence of internal transitions indicated by the dashed lines, both the transmission ratio n = 2 and the degree of coupling ǫ = 2 in the serial case,
whereas both n = 1 and ǫ = 1 in the parallel case. In the presence of internal transitions, the value of n is lower in the serial case and higher in the
parallel case, whereas the value of ǫ is lower in the both cases. (c) Exemplifying implementation of the 100-node network, constructed following
the algorithm described in Methods. The stochastic dynamics on this network is also described there. Note the two hubs, the states of the lowest
free energy, that can be identified with the two main conformations of the protein machine, e.g., open and closed, or bent and straight, usually the
only ones occupied sufficiently high to be notable under equilibrium conditions. The single pair of the output transition states (the gate) chosen for
the simulations is (2′′a, 2′d). The alternative four output pairs (2′′a, 2′a), (2′′b, 2′b), (2′′c, 2′c) and (2′′d, 2′d) are chosen tendentiously to lie one
after another.
3. Methods: Specification of the computer model
Various models of the networks with several output gates have been considered [34], but one class of models seems
to be the most realistic, based on a hypothesis that the protein conformational transition networks, like the higher level
biological networks, the protein interaction network and the metabolic network, have evolved in the process of a self-
organized criticality [72, 73]. Most networks of the systems biology are scale-free and display a transition from the
fractal organization on the small length-scale to the small-world organization on the large length-scale [42, 43].
In Ref. [34] we have shown that the case of many different output gates was reached in a natural way on scale-free
fractal trees, extended by long-range shortcuts. A network of 100 nodes with such properties is depicted in Fig. 3(c).
The algorithm of constructing the stochastic scale-free fractal trees was adopted after Goh et al. [74]. Shortcuts,
though more widely distributed, were considered by Rozenfeld, Song and Makse [42]. Here, we randomly chose
three shortcuts from the set of all the pairs of nodes distanced by six links. The network of 100 nodes in Fig. 3(c) is
too small to determine its scaling properties, but a similar procedure of construction applied to 105 nodes results in a
scale-free network, which is fractal on a small length-scale and a small world on a large length-scale. Very recently,
an actual network, which seems to possess similar properties and comprising some 250 nodes, was obtained in a long,
17 µs molecular dynamics simulation [66].
To provide the network with stochastic dynamics, we assumed the probability of changing a node to any of its
neighbors to be the same in each random walking step [34, 75]. Then, following the detailed balance condition, the
free energy of a given node is proportional to the number of its links (the node degree). The most stable nodes are
the hubs, which are the only practically observed conformational substates under equilibrium conditions. For a given
7
node l, the transition probability to one of the neighboring nodes is inversely proportional to the number of links, thus
to the equilibrium occupation probability peq
l τint)−1, where τint is the mean time
of internal transition counted in the random walking steps. This time is determined by the doubled number of links
minus one [75], τint = 2(100 + 3 − 1) = 204 random walking steps for the 100 node tree network with 3 shortcuts
assumed. To compare, we found the mean first passage time between the most distant nodes to equal 710 random
walking steps.
l of the output node, and equals (peq
i τext)−1 per random walking step, peq
The forward external transition probability, related to the stationary concentration [Pi], is determined by the mean
time of external transition τext, and equals (peq
i denoting the equilibrium occu-
pation probability of the initial input or output node (i = 1, 2, respectively). The corresponding backward external
transition probabilitiy is modified by the detailed balance breaking factors exp(−βAi). The assumed mean time of
forward external transition τext = 20 was one order of the magnitude shorter than the mean time of internal transition
τint = 204, so that the whole process is controlled by the internal dynamics of the system, not the external. The as-
sumption that most biochemical reactions are controlled by the slow dynamics of the proteins has a very rich literature
confirmations [26, 44]. It is this, in fact, that led to the now widely accepted change of the fundamental paradigm of
molecular biology, that not only structure but also dynamics determine the function of the proteins [54–58, 76].
4. Results
4.1. Determination of the mean cycle duration
All the averages in the equations from (2) to (10) are to be performed over a statistical ensemble of the stationary
fluxes determined for the time period t. We can obtain such an ensemble by dividing a long stochastic trajectory of
random walk on the studied network into the segments of the length t, and, next, by determining the net numbers of
external transitions, hence, the values of the fluxes j1(t) and j2(t) for each segment. However, we have to be sure that
the time period t is long enough for the considered ensemble to comprise only stationary, but not transient fluxes.
Because the initial state is random in the successive divisions of the trajectory into the segments of equal length,
the averaged value of the flux hJi(t)i (but not the higher moments!) coincides practically with its stationary value Ji,
even for the very short time period t. To evaluate the actual time, after which the fluxes become stationary, we have
first to divide the whole trajectory into the cycles or 'protocols' [35–37] of transient trajectories, starting and ending
in the same state, say 1′. For the ensemble of the cycles of the length t, or t within a certain small range, we can
determine the time-dependent averages J1(t) and J2(t), and in such a way find the mean cycle duration τcycle, long
after which the dynamics of the studied system passes from the transient to the stationary stage.
Figure 4: The time dependence of the averages over cycles J2(t) (the white circles) and J2(t) − J1(t) (the black circles) found in the model random
walk simulations on the network shown in Fig. 3(c) with the four output gates and the stationary flux ratio J2/J1 = 1.59. The solid line corresponds
to exponential fitting, with τcycle = 100 random walking steps, of the flux difference J2(t) − J1(t) downfall to zero.
8
We performed Monte Carlo simulations of random walk in 1010 computer steps on the network shown in Fig. 3(c)
with the dynamics specified above, with both the single and the fourfold output gate. The gates were chosen tenden-
tiously to maximize the value of the degree of coupling ǫ = J2/J1. We assumed βA1 = 1 and a few smaller, negative
values of βA2 determining the ratio J2/J1. The presence of external transitions, breaking the detailed balance, makes
some computational complications, which are discussed and explained in details elsewhere [75].
In Fig. 4, the time dependence of J2(t) and the difference J2(t) − J1(t) for the fourfold output gate and a chosen
value of the ratio J2/J1 is depicted. We do not quote the time dependence of the separate means over cycles J1(t)
and J2(t), because they both do not tend to zero in the cycle duration possible to designate. This means that they
are not determined completely in the transient stages of dynamics (they have a longer memory). Only the difference
J2(t) − J1(t) tends to zero, which is understandable, because the transition through the input gate each time erases
the memory of this difference. Exponentially fitting the downfall of the flux difference J2(t) − J1(t) to zero allows us
to evaluate the mean cycle duration in the case of the fourfold output gate to be approximately τcycle = 100 random
walking steps.
4.2. Determination of the transmission ratio
Many trials with divisions of the whole trajectory into segments of different lengths result in the conclusion that
(i) The obtained two-dimensional probability distribution functions p( j1(t), j2(t)) actually satisfy the
Andrieux-Gaspard fluctuation theorem (1) for t longer than the mean cycle duration τcycle.
To generate the statistical samples of the stationary fluxes numerously enough, we have chosen the time of the sta-
tionary averaging t = 104 random walking steps for the single gate and t = 103 random walking steps for the fourfold
gate. Exemplary two-dimensional probability distribution functions p( j1, j2) are shown in Fig. 5.
Figure 5: Exemplifying two-dimensional probability distribution functions p( j1(t), j2(t)) found in the model random walk simulations on the
network shown in Fig. 3(c) with the single output gate (left) and the fourfold output gate (right). For the single output gate, we assumed t = 104
random walking steps and J2/J1 = 0.95. For the fourfold output gate, we assumed t = 103 random walking steps and J2/J1 = 1.59. The averaged
values of the individual fluxes, multiplied by the time t of the determination, are marked by the dashed lines.
Next, we calculated the marginals p( j1) of the two-dimensional probability distribution functions p( j1, j2). Upon
analyzing the results, we found that
(ii) The logarithm of the ratio of the marginals p( j1)/p(− j1) can be described by the formula (5) with both
the entropic and additional components linearly depending on j1. The correction to entropy production is
negative.
The dependences of ln p( j1)/p(− j1) on j1t, when divided by βA1, are plotted in Fig. 6 and fitting these dependences
to Eq. (7) allowed us to determine the value of the transmission ratio to be n = 1.00 for the single output gate and
9
Figure 6: The generalized fluctuation theorem dependence (5) found in the model random walk simulations. The examined network was the one
shown in Fig. 3(c) with the single output gate (the squares) and the fourfold output gate (the circles). We assumed βA1 = 1 and a few smaller,
negative values of βA2 determining the ratio ǫ = J2/J1 of the average output and input fluxes noted in the figure. At the beginning, the dependence
was related to the force βA1 (the lowest diagram). Relating it to βA1 + nA2, Eq. (7), allowed us to fit the value of the reduction ratio n to be
1.00 for the single output gate (the higher diagram), and 2.27 for the fourfold output gate (the highest diagram). In the latter two diagrams, many
simulation points practically cover each other, which means that the value of n depends only on the network topology and confirms our method of
its designation.
n = 2.27 for the fourfold output gate. Let us note, that the determined values of n do not depend on the values of βA2,
hence are the property of only the assumed topology of the network.
Knowing the values of n, we determined the non-diagonal marginal distributions p( j2−n j1) of the two-dimensional
probability distributions p( j1, j2), and found that
(iii) The logarithm of the ratio of the marginals p( j2 − n j1)/p(− j2 + n j1) can be described by the formula (8)
with both the entropic and informative components linearly depending on j2 − n j1. The informative correction
is, as expected, positive.
Exemplary verifications of Eq. (8) are depicted in Fig. 7 both for the single (left) and the fourfold (right) output gate.
Three markedly different values of the ratio ǫ = J2/J1 were chosen for each case. The thick, solid line corresponds
to the lack of information production in Eq. (8). Such a case takes place for the force βA2 = −0.670, that stalls the
flux through the single output gate (J2/J1 = 0). The same effect of stalling the flux through the fourfold output gate is
reached for βA2 = −0.173, but this is accompanied by a non-zero information production (the line with a larger slope)
resulting from the possibility of the choice of the output gate even if the resultant net flux is zero. The greatest value
of the degree of coupling ǫ = J2/J1 is reached for βA2 = 0, which is represented by the vertical line in both diagrams.
It corresponds to ǫ = 0.98 for the single output gate, and ǫ = 2.01 for the fourfold output gate. As a consequence,
because of ǫ < n in both cases, the entropy production rate βA2(J2 − nJ1) = βA2(ǫ − n)J1 is always positive. The
important conclusion is that the direction of the flux considered in Figs. 1(a) and (b) to be forward or reverse, should
always be forward.
10
A2
β
/
]
)
j1
n
+
j2
-
(
p
/
)
-
j1
n
j2
p
[
n
l
(
n = 1.0
0
-5
0
-10
0.95
n = 2.27
0
-5
0
-10
1.59
0.98
-10
-8
-6
-4
(j2-nj1)t
-2
0
-10
-8
-6
-4
(j2-nj1)t
2.01
-2
0
Figure 7: The generalized fluctuation theorem dependence (8) for the marginal j2 − n j1, found in the model random walk simulation for the single
(n = 1, the squares, left) and the fourfold (n = 2.27, the circles, right) output gate. The thick, solid line corresponds to the lack of information
production in Eq. (8). Three markedly different values of the ratio ǫ = J2/J1 were chosen for each case (the vertical line, βA2 = 0, corresponds
to the maximum value of ǫ both for the single and the fourfold output gate). Because the corresponding values of βA2 are negative, the simulation
points lie in the lower half-plane of the graph. See the text for more detailed discussion.
4.3. Contribution from transient fluxes
The nanoscopic machines are not only energy but also information transducers. In other words, 'information can
serve as a thermodynamic resource similar to free energy' there [37]. Up to then, we treated information as an appro-
priately averaged mutual information exchanged between two subsystems [7–12], but such creation of information at
the expense of the entropy reduction appeared not possible in the biological molecular machines. Information may,
however, be also considered as the content of a sequence of bits [13–17], and the latter approach is equivalent to the
bipartite measurement-control approach [18]. Two types of information in the second sense are produced during the
time course of the simulated process with many output gates. The first is in the form of a string with a more or less
random sequence of letters, e.g.
. . . b c d d a c b b a b c c . . . ,
labelling successive gates, which the system passed through. And the second is in the form of a string with a more or
less random sequence of signs, e.g.
. . . + + − + − − + + + + . . . ,
describing the directions (forward or backward) of the successive transitions through the output gates.
Nowadays, it is impossible to determine the information of the first type experimentally, and for its theoretical
determination, detailed studies of individual, successive transient protocols related to individual trajectories are needed
[35–37]. However, the information of the first type is strongly related to the information of the second type, directly
registered in the single biomolecule experiments [47, 48, 61, 62]. The information of the second type flows out of
the machine and defines the number of the product molecules P2, hence, the fluctuating variable X2. Both types
of information are erased each time the system passes forward through the input gate 1′′1′ and the molecule P1 is
created, hence, the variable X1 increases by one. In other words, the information created by the molecular machine
in the successive free energy transduction cycles is written in the fluctuating number of product molecules P2, created
per product molecule P1, i.e., in the fluctuating variable X2 − X1 which differs from the variable X2 − nX1 in the case
of the multiple output gate. We had already found that the corresponding stationary flux J2 − J1 is determined entirely
by the transient stage dynamics (Fig. 4).
In accordance with the dual interpretation of information [18], also for the flux J2 − J1, we consider Eq. (6) to result
in the generalized second law of thermodynamics (10). Formally, upon substituting n = 1, the average dimensionless
entropy production (the dissipation D divided by kBT ) is determined as
βD ≡ hσi = βA2(J2 − J1)t
11
(13)
A2
β
/
]
)
j1
+
-j2
p
(
/
)
-
(
j1
j2
p
[
n
l
1.00
0.50
0
0
-5
-10
-10
-8
-6
-4
-2
1.00
1.21
1.59
1.77
2.01
0
(j2-j1)t
2
4
6
8
10
Figure 8: The generalized fluctuation theorem dependence (6) with n = 1, found in the model random walk simulations for the fourfold output
gate. The examined network was that shown in Fig. 3(c). We assumed βA1 = 1 and a few negative values of βA2 determining the difference J2 − J1
of the stationary output and input fluxes (or the degree of coupling ǫ = J2/J1 noted in the figure). The results of each simulation were obtained
symmetrically on the whole axis of the fluctuating difference ( j2 − j1)t but in the figure, they are presented, divided by the (negative!) coefficient
βA2, only on the left from zero for the negative average J2 − J1, and on the right from zero for the positive average J2 − J1. The thick, solid line
corresponds to the first, entropic component to the right-hand side of Eq. (6). For the multiple output gate, the information gain, resulting from the
possibility of a choice, surpasses the information loss in favor of the flux J1, and reduces the effects of the entropy production in part, until the limit
of J2 − J1 = 0 (J2/J1 = 1), above which information production prevails.
whereas the average information production I, as
I ≡ hιi = β(A1 + A2)J1t −*ln
p(J1(t) J2(t) − J1(t))
p(−J1(t) −J2(t) + J1(t))+ .
(14)
The averages are taken over the joint probability function of the fluxes p( j1(t), j2(t)).
For the macroscopic machines, there is no correlation between the fluxes, so that only the first term in (14)
contributes to the information production I. It indirectly represents the entropy production in the hidden variable X1
(in the case of n = 1, compare Eq. (7)). For the macroscopic machines as well as the mesoscopic enzymatic machines
with the single output gate, the information production I and the entropy production βD must always be positive, of
the same sign (see Fig. 7, left). Accordingly, I is to be interpreted as the information loss. It is the general case of the
free energy transduction for any value of n (see Fig. 7, right).
However, the case of the direct difference J2 − J1 for values of n greater than one is different. In order to determine
the temporal fluctuations of the flux J2 − J1 for the fourfold output gate, we calculated the diagonal marginal p( j1 − j2)
of the two-dimensional distributions p( j1, j2), and found that
(iv) The logarithm of the ratio of the diagonal marginals p( j1 − j2)/p(− j2 + j1) can be described by the formula
(6) with n = 1 and all the components linearly depending on the difference j2 − j1. The informative correction
is of the opposite sign to dissipation and large.
The dependences of ln p( j1 − j2)/p(− j2 + j1) on ( j2 − j1)t are depicted in Fig. 8 for a few chosen negative values of
βA2 that determine the difference of the averages of the output and input fluxes J2 − J1, thus, the ratio of those averages
ǫ = J2/J1. We discarded the rare results for the values of ( j2 − j1)t higher than 10, as being burdened with a statistical
error too large. The simulation data are presented in such a way that the transition from the entropy production to the
entropy reduction might be clearly seen. The thick, solid line corresponds to the first, entropic term to the right-hand
side of Eq. (6). For many output gates, as opposed to the case of a single output gate, the information gain, resulting
from the possibility of a choice, surpasses the information loss and reduces the effects of the entropy production in
part, until the limit of J2 − J1 = 0 (J2/J1 = 1), above which information production prevails.
The stationary values of J1 and J2 are unambiguously related to the forces A1 and A2, so we can directly determine
the average entropy production βD, Eq. (13), for the given time period t. Fig. 8 clearly shows the linear relation
βD + I = αβD
12
(15)
Figure 9: The dependence of the information production I (the squares) and the entropy production βD (the circles) per t = 103 random walking
steps of determination, on the flux difference J2 − J1. The points represent the values obtained from the data given in Fig. 8. The sum of βD and I,
we refer to as the cost C of information processing (the triangles), is also shown.
where α is the tangent of the adequate straight line slope. From (15), knowing βD and α, we can determine the average
information production I without referring to the much more complex formula (14). The values of I and βD, obtained
from the values of α found from Fig. 8, are presented in Fig. 9 as the functions of J2 − J1. Note that for J2 − J1 > 0,
the negative entropy production βD is associated with the positive information production I, which should now be
referred to as actual information creation.
5. Discussion: Procesing of free energy versus processing of organization
The fundamental task of statistical thermodynamics is to link the internal dynamics of 'microstates' of a studied
system with the dynamics of 'macrostates' – the thermodynamic variables. In the mesoscopic systems, the thermody-
namic variables are fluctuating quantities. In any case, the transition from the 'microstates' to 'macrostates' consists
in averaging over time. The change of the availble knowlege of the 'microstate' is refered to as entropy and the change
of the available knowledge of the 'macrostate' as information.
In the present paper, we realized such program for the internal dynamics determined by the complex network
presented in Fig. 3 and the macroscopic dynamics in the form of a network connecting two chains of natural numbers,
which represent the fluctuating concentrations of reagents X1 and X2 related to a single molecule of protein enzyme.
We considered the open system at steady state, so the concentrations X1 and X2 had to be replaced by the corresponding
fluxes J1 and J2. Study of the fluctuating flux difference J2 − J1 turned out the most important. It is the time derivative
of the thermodynamic variable X2 − X1 that characterizes organization of the system. In the case of the macroscopic
mechanical machines, it could be, for example, the possibly slipping angle of a component wheel to the axle. In
the case of the macroscopic battery (the electrochemical machine), it is the difference between the available electric
charge and the unused amount of reductor, both quantities being determined in molecular units. The organization
X = X2 − X1 is controlled by force A2 and can be expressed in the energy units as A2X. The organization of the
macroscopic machines always decreases: the wheels slide with respect to the axles, the batteries wear out. This need
not be the case of the fluctuating mesoscopic machines.
The main result of our study is that the free energy processing has to be distinguished from the organization
processing, which is schematically illustrated in Fig. 10, where we distinguish between the action of the perfect
machine and the losses. Note that both free energy of the perfect machine F, energy losses F0 and organization X are
functions of state of the system. As in Fig. 1(b), the bound energy and the environment in Fig. 10 are considered to
be determined jointly by the internal dynamics of the machine modified by an interaction with the environment. In
the case of the purely stochastic dynamics, the presence of the thermal bath influences the probabilities of the internal
transitions, whereas the constraints influence the probabilities of the external transitions.
For the systems operating under isothermal conditions, the first and second laws of thermodynamics:
W1 + W2 − ∆F = D ≥ 0
13
(16)
(a)
Perfect machine
(b)
Losses / Profits
(c)
Information creator
s
c
i
m
a
n
y
d
l
a
n
r
e
t
n
I
Constraints
W1
W2
Work
F X~ 1
Free energy
D
Dissipation
Entropy
s
c
i
m
a
n
y
d
l
a
n
r
e
t
n
I
Constraints
W0
Work
F X nX
~
-
0
2
1
Free energy
D0
b-1I0
Dissipation
Entropy
s
c
i
m
a
n
y
d
l
a
n
r
e
t
n
I
Constraints
C
Arrangement
X X X
=
-2
1
Organization
bD
I
Information
Entropy
Figure 10: Free energy processing versus arrangement processing in the machine participating in a stationary isothermal process. (a) In the case
of the perfect machine, the output flux is tightly coupled to the input flux, hence the free energy of the system is uniquely determined by its input
thermodynamic variable X1. (b) In the actual macroscopic as well as mesoscopic machine, the remaining part of the free energy, proportional to
the difference X2 − nX1, corresponds to energy losses both for the transmission ratio n = 1 and greater (see the discussion just after Eq. (14)). (c) In
the mesoscopic enzymatic machine with many gates, the transmission ratio n is greater than unity and it is the additional variable X2 − X1 that has
the meaning of the system's organization. See the text for more detailed discussion. The wishful directions of the input and output work fluxes as
well as the input and output information fluxes are indicated. The actual directions of the fluxes denoted by the forward-reverse arrows depend on
the specific internal dynamics of the system. In the biological molecular machines with the internal dynamics studied here, these are forward.
and
W0 − ∆F0 = D0 + β−1I0 ≥ 0
(17)
are fulfilled for the perfect free energy processing and losses, respectively. For the stationary processes, the free
energies F and F0 remain constant, ∆F = ∆F0 = 0. All the components of work W1, W2, and W0, dissipation D
and D0, and information I0 are functions of the process. The individual components of Eqs. (16) and (17), which
have been considered in our study, are specified in Fig. 1 and in the text. The free energy F is proportional to X1,
and the free energy losses F0 to X2 − nX1. In the perfect machine, the output flux is tightly coupled to the input flux,
J2 = nJ1, and the output work is maximum. Dissipation D tends to zero when the perfect stationary machine works
infinitely slowly, J1 → 0, but, then, its input and output powers also drop to zero. Loss in the free energy processing
means that no work on the environment is performed despite the forced energy dissipation. We have proven that loss
of information I0 occurs in any case of the free energy transduction. In the biological molecular machines also the
dissipation D0 is positive. However, this must not be the case of other molecular machines and then, the loss W0
changes into the profit.
In the mesoscopic enzymatic machines with many gates, hence the possibility of a choise, the loss is determined
by the thermodynamic variable X2 − nX1 with the transmission ratio n greater than unity. However, this variable
jointly with X1 does not characterize fully the system. The variable X2 − X1 remains, which now has the meaning of
the system's organization. For the processing of organization, the generalized first and second law can be written as
βD + I − ∆X = C ≥ 0 .
(18)
Under stationary conditions, the mean organization X = X2 − X1 is constant, ∆X = 0. Here, the two components
of the generalized second law (10), given by Eqs. (13) and (14), are more generally treated as the two components
of information. Both βD and I as well as C, the quantity which balances two former quantities and which could be
referred to as the organization cost or arrangement of the system, are functions of the process.
All three changes of organization in Eq. (18) are in fact the three types of information and are associated with the
three strings of bits. When presenting results of our simulations, we described two signals. The first, in the form of a
string of letters, corresponds to the arrangement of organization, whereas the second, in the form of a string of signs,
corresponds to the actual information. We did not mention the third signal from the random number generator, which
simulates a noise corresponding to entropy production or reduction.
Only the free energy transducer, for which the works W1 and W2 are of the opposite sign, can be referred to as
the machine. Similarly, only the organization transducer, for which the information production I and the entropy
production βD are of the opposite sign, can be referred to as the information creator (see Fig. 10(c)). For I being
positive, βD must be negative (the entropy reduction instead of production). The biological molecular machines, for
which this is the case, may be said to act as Maxwell's demons, although they do not utilize the information creation
for the reduction of energy losses, but only for other purposes, for example molecular recognition.
14
When considering the coupling of the free energy donating process with the free energy accepting one, the case
of J2 = J1 is exceptional. Our simulation clearly points out that for n > 1, the case of J2 − J1 = 0 occurs for the total
compensation of entropy reduction βD by information creation I, albeit, on average, both then tend to zero (compare
Fig. 9). This case is the optimum in the sense that the arrangement C of the total process is then zero. We may refer
such a situation to as the system's selforganization.
6. Concluding remarks
The two powerful theoretical physics tools created at the turn of the century, the fluctuation theorem [35–37] and
the idea of self-organized criticality [72, 73], force a significant change in our views on the nature of the biological
molecular machines action. Progress in the theory coincides with the intensive experimental and numerical studies of
protein dynamics [54–66].
Under the physiological conditions, the protein molecular machines fluctuate constantly between lots of conforma-
tional substates composing their native state. The probabilities of visiting individual substates are far from equilibrium
and determined by the concentration of the substrate molecules surrounding the machine. The machine can be con-
sidered as an enzyme that simultaneously catalyzes two reactions: the free energy-donating and free energy-accepting
ones. During the transient stages before completing the free energy transduction cycles, the subsequent realizations
of the free energy-accepting reaction force transitions between separate regions of the conformational substates of the
protein machine, which brakes ergodicity [77] of its dynamics until the next free energy-donating reaction begins. The
machine transduces information about the types of successive transitions on the information about their successive
directions. In this transduction the third signal can be utilized. It is the purely random noise of the environment, which
decides the realization of actual transitions between the conformational substates. Information about the direction of
the successive transitions during the transient stage of the free energy transduction cycle determines the fluctating
difference of the concentrations of the free-energy-accepting and free-energy-donating molecules, surrounding the
machine. This difference is a thermodynamic measure of the dynamical organization of the molecular machine and
can be the source of information passed to a further use.
A possible proposal, not to be underestimated, is that the biological structural memory, traditionally associated
with DNA and RNA, can in this way be complemented by a dynamical memory of proteins. The transient dynam-
ics stages during the succeeding free energy transduction cycles can last quite long [66, 78], especially when more
shortcuts and external transitions in the conformational transition network are taken into account [76]. Projection
of the trajectory on the fluctuating thermodynamic variable of organization has a form of time series representing
continuous time random walk [66, 78, 79]. It is very likely, that the natively disordered transcription factors, in their
one-dimensional search for its target on DNA, intermittent by three-dimensional flights [80, 81], perform not the
passive diffusion but the active continuous time random walk.
As mentioned earlier [34], our approach is able to explain the observation, that the single myosin II head can take
several steps along the actin filament per ATP molecule hydrolyzed [47, 48]. It is likely that the mechanism of the
action of the small G-proteins such as Ras, which have a common ancestor with the myosin [82] and an alike partly
disordered structure after binding the nucleoside triphosphate [83–85], is, after a malignant transformation, similar. Of
course, we cannot claim that the model considered here has something to do with the dynamics of the myosin had or
the small G-proteins, but in future, a more adequate model is worth considering. The greatest weakness of the current
model is the temperature independence of conformational transitions and this is also considered to be improved in the
near future.
In the healthy cells, the G-proteins play the role of signal transducers that activate various kinases, which further
transmit the signal to the target. The chemotactic response of Escherichia coli to methylation of a receptor, leading to
activation of the flagellar motor to move, is the relatively simple experimental model investigated in more detail [86].
Theory of the biochemical signal transduction in this system is the subject of intensive research [87–93]. Here, the
information is also stored in the form of the concentration of specific molecules like in the system considered in the
present paper. It would be interesting to combine both approaches.
There are some arguments, for instance, for the kinesin motor [94–98], the myosin V motor [99], the cytoplasmic
dynein motor [100, 101], the quinol:cytochrome c oxidoreductase [102, 103], and the very Ras molecules [104,
105] that an organization transduction with ǫ = 1 is achieved in dimeric protein complexes, which are composed
15
of two identical monomers. We have ascertained that the case of the total compensation of the entropy production
by information creation is the optimum in the sense that the cost of the total process is then zero. This may occur
for a binary system, whose components create and collect information alternately in such a way that the resultant
information and entropy productions are zero. The cyclic, alternate behavior is important for the corresponding cost
in the stationary process could not be assigned to individual components. In this respect, the corresponding models
should differ essentially from those of the bipartite systems [11, 12], and are worth a separate study. Research of
this kind could help to answer the certainly interesting question: why do most protein machines operate as dimers or
higher organized structures?
The main conclusion of the paper is that the free energy processing has to be distinguished from the organization
processing. From the former point of view, Maxwell's demon utilizes entropy reduction for the performance of work
and more precisely, for a reduction of energy losses. From the latter point of view, it can be used for other purposes,
for example molecular recognition. Our answer to the question posed in the title is that the biological molecular
machines can, under certain conditions, act as Maxwell's demons, but only creating information, and not performing
work. This statement is based on the relationship ǫ < n between the degree of coupling ǫ and the transmission ratio n
shown to hold for the studied model of dynamics. There is still the need to prove the generality of this relationship.
We know that work, heat and dissipation (the entropy production) are changes of energy. But there is still contro-
versy, the change of which physical quantity is information [37, 106]? Here, we suggest the answer that information
is the change of organization, a quantity being the difference of two physical quantities X2 and X1 taking part in
the free energy-transduction process and describing the free energy reception and delivery, respectively. In such an
approach, the dual nature of entropy becomes clear as the physical quantity that connects the processes of the free
energy transduction and the organization transduction.
At the end, let us take the liberty for a couple of speculation. The first and second laws of thermodynamics (16) and
(17) are valid for any isothermal processes, whereas the suggested first and second laws of arrangement transduction
(18) were justified only for the stationary isothermal processes. The open problem remains the generality of these
laws. One thing is certain: the necessary conditions for the information and entropy productions to be of the opposite
signs, is the presence of fluctuations and the possibility of a choice. Besides the mesoscopic machines, we know three
macroscopic systems sharing such properties and intriguing long. The first and the best known are the systems with
critical thermodynamic fluctuations, whose organization is determined by new thermodynamic variables that survive
stochastization [107], referred to as the order parameters [45] or, in various contexts, the emergent [108] or structural
[44] variables. Here, the long-living transient stage can be identified, e.g., with the nonergodic condensation of gas
through the state of fogg or the nonergodic solidification (the glass transition) of liquid [45]. The second example are
the systems displaying quantum fluctuations entangled with the environment, which organization is determined by
classical variables that survive decoherence [109, 110]. And the third, most controviersial example are living systems
that display a non-gradual stochastic Darwinian evolution, and their organization is determined by the survival degree
of a species, long-resistant against mutations [111, 112].
Acknowledgements. MK thanks Yasar Demirel and Herve Cailleau for discussing the problem in the early stages of
the investigation.
Author Contributions. The general concept and the theory is mainly due to MK, who also wrote the manuscript.
The specification of the critical branching tree model and the numerical simulations are mainly due to PC.
Conflict of Interest. None declared.
References
[1] J. C. Maxwell, Theory of Heat, Logmans, London, 1871.
[2] L. Z. Szilard, Z. Phys. 53 (1929) 840-857.
[3] R. Landauer, IBM J. Res. Dev. 5 (1961) 183-191.
[4] H. S. Leef, E. Rex eds., Maxwell's Demon 2: Entropy, Classical and Quantum Information, Computing, Institute of Physics Publish-
ing, Philadelphia, 2003.
[5] K. Maruyama, F. Nori, V. Verdal, Revs. Mod. Phys. 81 (2009) 1-23.
[6] T. Sagawa, Thermodynamics of Information Processing in Small Systems, Springer, Berlin, 2013.
[7] T. Sagawa, M. Ueda, Phys. Rev. Lett. 104 (2010) 090602.
16
[8] M. Ponmurugan, Phys. Rev. E 82 (2010) 031129.
[9] T. Sagawa, M. Ueda, Phys. Rev. E 85 (2012) 021104.
[10] T. Sagawa, M. Ueda, New J. Phys. 15 (2013) 125012.
[11] B. Hartich, A. C. Barato, U. Seifert, J. Stat. Mech. 14 (2014) P02016.
[12] J. M. Horowitz, M. Esposito, Phys. Rev. X 4 (2014) 031015.
[13] D. Mandal, C. Jarzynski, Proc. Natl. Acad. Sci. USA 109 (2012) 11641-11645.
[14] A. C. Barato, U. Seifert, Phys. Rev. Lett. 112 (2014) 090601.
[15] J. M. Horowitz, T. Sagawa, J. M. R. Parrondo, Phys. Rev. Lett. 111 (2013) 010602.
[16] N. Shiraishi, T. Sagawa, Phys. Rev. E 91 (2015) 012130.
[17] N. Shiraishi, T. Ito, K. Kawaguchi, T. Sagawa, New J. Phys. 17 (2015) 045012.
[18] N. Shiraishi, T. Matsumoto, T. Sagawa, New J. Phys. 18 (2016) 013044.
[19] S. Deffner, C. Jarzynski, Phys. Rev. X 3 (2013) 041003.
[20] M. Smoluchowski, Phys. Z. 13 (1912) 1069-1080.
[21] R. P. Feynman, R. B. Leighton, M. Sands, The Feynman Lectures on Physics, vol. 1, chap. 46, Addison-Wesley, Reading, 1963.
[22] C. Jarzynski, O. Mazonka, Phys. Rev. E 59 (1999) 6448-6459.
[23] A. F. Haxley, Prog. Biophys. Biophys. Chem. 7 (1957) 255-318.
[24] J. Howard, Mechanics of Motor Proteins and the Cytoskeleton, Sinauer, Sunderland, 2001.
[25] J. Howard, Motor proteins as nanomachines: the roles of thermal fluctuations in generating force and motion, in Biological Physics - Poincare
Seminar 2009, B. Duplainer, V. Rivasseau, eds., 47-60, Springer, Basel, 2003.
´E. Rold´an, I. A. Martinez, J. M. R Parrondo, D. Petrov, Nature Phys. 10 (2014) 457-461.
[26] M. Kurzynski, Prog. Biophys. Molec. Biol. 69 (1998) 23-82.
[27] M. A. Fisher, A. B. Kolomeisky, Proc. Natl. Acad. Sci. USA 96 (1999) 6597-6602.
[28] A. B. Kolomeisky, M. A. Fisher, Annu. Rev. Phys. Chem. 58 (2007) 675-695.
[29] R. D. Astumian, I. Derenyi, Biophys. J. 77 (1999) 993-1002.
[30] C. Bustamante, D. Keller, G. Oster, Acc. Chem. Res. 34 (2001) 412-420.
[31] R. Lipowsky, Phys. Rev. Lett. 85 (2000) 4401-4406.
[32] R. Lipowsky, S. Liepelt, A. Valleriani, J. Stat. Phys. 135 (2009) 951-975.
[33] M. Kurzynski, P. Chelminiak, J. Stat. Phys. 110 (2003) 137-181.
[34] M. Kurzynski, M. Torchala, P. Chelminiak, Phys. Rev. E 89 (2014) 012722.
[35] C. Jarzynski, Annu. Rev. Condens. Matter Phys. 2 (2011) 329-351.
[36] U. Seifert, Rep. Prog. Phys. 75 (2012) 126001(58pp).
[37] J. M. R. Parrondo, J. M. Horowitz, T. Sagawa, Nature Phys. 11 (2015) 131-139.
[38] S. Toyabe, T. Sagawa, M. Ueda, E. Muneyuki, M. Sano, Nature Phys. 6 (2010) 988-992.
[39] A. B´erut, A. Arakelyan, A. Petrosyan, S. Ciliberto, R. Dillenschneider, E. Lutz, Nature 483 (2012) 187-190.
[40] J. V. Koski, Y. F. Maisi, T. Sagawa, E. Pekola, Phys. Rev. Lett. 113 (2014) 030601.
[41]
[42] H. D. Rozenfeld, C. Song, H. A. Makse, Phys. Rev. Lett. 104 (2010) 025701.
[43] F. Escolano, E. R. Handcock, M. A. Lozano, Phys. Rev. E 85 (2012) 036206.
[44] M. Kurzynski, The Thermodynamic Machinery of Life, Springer, Berlin, 2006.
[45] H. B. Callen, Thermodynamics and an Introduction to Thermostatistics 2nd edn., Wiley, New York, 1989.
[46] T. L. Hill, Free Energy Transduction and Biochemical Cycle Kinetics, Springer, New York, 1989.
[47] K. Kitamura, M. Tokunaga, A. H. Iwane, T. Yanagida, Nature 397 (1997) 129-134.
[48] K. Kitamura, M. Tokunaga, S. Esaki, A. H. Iwane, T. Yanagida, Biophysics 1 (2005) 1-19.
[49] D. Andrieux, P. Gaspard, J. Stat. Phys. 127 (2007) 107.
[50] P. Gaspard, New J. Phys. 15 (2013) 115014.
[51] C. Jarzynski, Phys. Rev. Lett. 78 (1997) 2690-2693.
[52] J. Mehl, B. Lander, C. Bechinger, V. Blickle, U. Seifert, Phys. Rev. Lett. 108 (2012) 220601.
[53] M. Borrelli, J. V. Koski, S. Maniscalco, J. P. Pekola, Phys. Rev. E 91 (2015) 012145.
[54] K. Henzler-Wildman, D. Kern, Nature 450 (2007) 964-972.
[55] V. N. Uversky, A. K. Dunker, Biochim. Biophys. Acta 1804 (2010) 1231-1263.
[56] D. Chouard, Nature 471 (2011) 151-153.
[57] J. D. Chodera, F. Noe, Curr. Opin. Struct. Biol. 25 (2014) 135-144.
[58] D. Shukla, C. X. Hernandez, J. K. Weber, V. S. Pande, Acc. Chem. Res. 48 (2015) 414-422.
[59] R. H. Austin, K. W. Beeson, L. Eisenstein, H. Frauenfelder, I. C. Gunsalus, Biochemistry 14 (1975) 5355-5373.
[60] H. Frauenfelder, The Physics of Proteins: An Introduction to Biological Physics and Molecular Biophysics, Springer, Berlin, 2010.
[61] A. Ishijima, T. Yanagida, Trends Biochem. Sci. 26 (2001) 438-444.
[62] T. Yanagida, E. Ishii, eds., Single Molecule Dynamics in Life Science, Wiley-VCH, Weinheim, 2008.
[63] A. Kitao, S. Hayward, N. Go, Proteins 33 (1998) 496-517.
[64] M. Karplus, J. A. McCammon, Nature Struct. Biol. 9 (2002) 646-652.
[65] S. Yang, N. K. Banavali, B. Roux, Proc. Natl. Acad. Sci. USA. 106 (2009) 3776-3781.
[66] X. Hu, L. Hong, M. D. Smith, T. Neusius, X. Cheng, J. C. Smith, Nature Phys. 12 (2016) 171-174.
[67] R. Zwanzig, Acc. Chem. Res. 23 (1990) 148-152.
[68] H. P. Lu, L. Xun, X. S. Xie, Science 282 (1998) 1877-1882.
[69] B. P. English, W. Min, A. M. van Oijen, K. T. Lee, G. Luo, H. Sun, Nature Chem. Biol. 2 (2006) 87-94.
[70] M. Kurzynski, Cell. Mol. Biol. Lett. 13 (2008) 502-513.
[71] X. S. Xie, Science 342 (2013) 1457-1459.
17
[72] P. Bak, How Nature Works, Springer, New York, 1996.
[73] K. Sneppen, G. Zocchi, Physics in Molecular Biology, Cambridge University Press, New York, 2005.
[74] K. I. Goh, G. Salvi, B. Kahng, D. Kim, Phys. Rev. Lett. 96 (2006) 018701.
[75] P. Chelminiak, M. Kurzynski, Physica A 468 (2017) 540-551.
[76] M. Kurzynski, P. Chelminiak, Entropy 16 (2014) 1969-1982.
[77] R. G. Palmer, Adv. Phys. 31 (1982) 669-783.
[78] R. Metzler, Nature Phys. 12 (2016) 113-114.
[79] R. Metzler, J.-H. Jeon, A. G. Cherstvy, E. Barkai, Phys. Chem. Chem. Phys. 16 (2014) 24128-24164.
[80] A. Tafvizi, F. Huang, A. R. Ferhst, L. A. Mirny, A. M. A. van Oijen, Proc. Natl. Acad. Sci. USA 108 (2011) 263-268.
[81] G.-W Li, X. S. Xie, Nature 475 (2011) 308-315.
[82] F. J. Kull, R. D. Vale, R. J. Fletterick, J. Muscle Res. Cell Motil. 19 (1998) 877-886.
[83] A. Houdusse, H. L. Sweeney, Curr. Res. Struct. Biology. 11 (2001) 182-194.
[84] I. Kosztin, R. Bruinsma, P. O'Lague, K. Schulten, Proc. Natl. Acad. Sci. USA 99 (2002) 3575-3580.
[85] Y. Arai, A. H. Iwane, T. Wazawa, T. Yokota, Y. Ishii, T. Kataoka, T. Yanagida, Biochem. Biophys. Res. Commun. 343 (2006) 809-815.
[86] Y. Tu, T. S. Shimizu, H. C. Berg, Proc. Natl. Acad. Sci. USA 105 (2008) 14855-14860.
[87] S. Ito, T. Sagawa, Phys. Rev. Lett. 111 (2013) 180603.
[88] A. C. Barato, D. Hartrich, U. Seifert, Phys. Rev. E 87 (2013) 042104.
[89] A. C. Barato, D. Hartrich, U. Seifert, New J. Phys. 16 (2014) 103024.
[90] P. Sartori, L. Granger, C. F. Lee, J. M. Horowitz, PLOS Comput. Biol. 10 (2014) e1003974.
[91] S. Ito, T. Sagawa, Nature Commun. 6 (2015) 2-6.
[92] S. Bo, M. Del Giudice, A. Celani, J. Stat. Mech. (2015) P01014.
[93] D. Hartrich, A. C. Barato, U. Seifert, Phys. Rev. E 93 (2016) 022116.
[94] Y. Taniguchi, P. Karagiannis, M. Nishiyama, T. Yanagida, BioSystems 88 (2007) 283-292.
[95] M. Bier, BioSystems 88 (2007) 301-307.
[96] R. D. Astumian, BioSystems 93 (2007) 8-15.
[97] S. Liepelt, R. Lipowsky, Phys. Rev. E 79 (2009) 011917.
[98] R. D. Astumian, Biophys. J. 98 (2010) 2401-2409.
[99] M. Nishikawa, H. Takai, T. Shibata, A. F. Ivane, T. Yanagida, Phys. Rev. Lett. 101 (2008) 12103.
[100] D. Tsygankov, A. W. R. Serohijos, N. V. Dokholyan, T. C. Elston, J. Chem. Phys. 130 (2009) 025101.
[101] D. Tsygankov, A. W. R. Serohijos, N. V. Dokholyan, T. C. Elston, Biophys. J. 101 (2011) 144-150.
[102] M. Swierczek, E. Cieluch, M. Sarewicz, A. Borek, C. C. Moser, P. L. Dutton, A. Osyczka, Science 329 (2010) 451-454.
[103] M. Sarewicz, A. Osyczka, Physiol. Rev. 95 (2015) 219-243.
[104] L. Iversen, H.-L. Tu, W.-C. Lin, S. M. Christensen, S. M. Abel, J. Iwig et al., Science 345 (2014) 50-54.
[105] Y. Nakamura, K. Hibino, T. Yanagida, Y. Sako, Biophys. Physbiol. 13 (2016) 1-11.
[106] J. P. Crutchfield, Nature Phys. 8 (2012) 17-24.
[107] O. Penrose, Foundations of Statistical Mechanics. A Deductive Treatment., Pergamon Press, Oxford, 1970.
[108] P. W. Anderson, Science 177 (1972) 393-396.
[109] W. H. Zurek, Revs. Mod. Phys. 75 (2003) 715-765.
[110] W. H. Zurek, Nature Phys. 5 (2009) 181-188.
[111] N. Eldredge, S. J. Gould, Punctuated equilibria: an alternative to phyletic gradualism., in T. J. M. Schopf ed., Models in Paleobiology,
82-115, Freeman, San Francisco 1972.
[112] P. Bak, P. Sneppen, Phys. Rev. Lett. 71 (1993) 4083-4086.
18
|
1311.6362 | 2 | 1311 | 2015-02-04T13:52:16 | Calculating the contribution of different binding modes to Quinacrine - DNA complex formation from polarized fluorescence data | [
"physics.bio-ph",
"q-bio.BM"
] | Binding of acridine derivative quinacrine (QA) to chicken erythrocyte DNA was studied by methods of absorption and polarized fluorescent spectroscopy. Measurements were carried out in aqueous buffered solutions (pH 6.9) of different dye concentrations (QA concentration range from $10^{-6}$ till $10^{-4}$ M) and ionic strengths ($Na^{+}$ concentration rang from $10^{-3}$ till 0.15 M) in a wide range of phosphate-to-dye molar ratios ($P/D$). It is established that the minimum of fluorescent titration curve plotted as relative fluorescence intensity $vs$ $P/D$ is conditioned by the competition between the two types of QA binding to DNA which posses by different emission parameters: (i) intercalative one dominating under high $P/D$ values, and (ii) outside electrostatic binding dominating under low $P/D$ values, which is accompanied by the formation of non-fluorescent dye associates on the DNA backbone. Absorption and fluorescent characteristics of complexes formed were determined. The method of calculation of different binding modes contribution to the complex formation depending on $P/D$ value is presented. It was shown that the size of binding site measured as the number of DNA base pairs per one QA molecule bound in the case of the electrostatic interaction is 8 times less than that for the intercalative one that determines the competitive ability of the outside binding against the stronger intercalative binding mode. | physics.bio-ph | physics | Calculating the contribution of different binding modes to
Quinacrine–DNA complex formation from polarized fluorescence data
I.M. Voloshin, О.A. Ryazanova, V.A. Karachevtsev and V.N. Zozulya
Department of Molecular Biophysics, B. Verkin Institute for Low
Temperature Physics and Engineering of NAS of Ukraine, 47 Lenin
ave., 61103, Kharkiv, e-mail: [email protected]
Binding of acridine derivative quinacrine (QA) to chicken erythrocyte DNA was
studied by methods of absorption and polarized
fluorescent spectroscopy.
Measurements were carried out in aqueous buffered solutions (pH 6.9) of different dye
concentrations ([QA] = 10-6 ÷ 10-4 M) and ionic strengths ([Na+] = 10-3 ÷ 0.15 M) in a
wide range of phosphate-to-dye molar ratios (P/D). It is established that the minimum
of fluorescent titration curve plotted as relative fluorescence intensity vs P/D is
conditioned by the competition between the two types of QA binding to DNA which
posses by different emission parameters: (i) intercalative one dominating under high
P/D values, and (ii) outside electrostatic binding dominating under low P/D values,
which is accompanied by the formation of non-fluorescent dye associates on the DNA
backbone. Absorption and fluorescent characteristics of complexes formed were
determined. The method of calculation of different binding modes contribution to the
complex formation depending on P/D value is presented. It was shown that the size of
binding site measured as the number of DNA base pairs per one QA molecule bound in
the case of the electrostatic interaction is 8 times less than that for the intercalative one
that determines the competitive ability of the outside binding against the stronger
intercalative binding mode.
Key words: quinacrine, DNA, outside binding, intercalation, cooperative binding,
fluorescence, absorption.
I. INTRODUCTION
Study of interaction between different organic dyes and intracellular biopolymers is one of
major scientific approaches which give the opportunity to obtain information about molecular
mechanisms of biological system activities. In particular, the dyes give the possibility to
visualize the intracellular structures. So, after the works of Caspesson et al. [1-3] devoted to
staining of fixed chromosomes, acridine derivative called quinacrine (QA – quinacrine,
mepacrine, atabrine – Fig.1) become widely applied in cytology. This dye fluoresces
intensively in yellow-green range of spectra, that gives the possibility to obtain the specific
staining pattern (characteristic transverse bands that appear on chromosomes), that was
2
successfully used under the chromosome analysis. High biological activity provides wide
range of its application in biology and medicine as antiprotozoal, antirheumatic, antihelmintic
agent. Quinacrine was first synthesized by Bayer (Germany) in 1931. During the II World
War it has been widely used for prevention and therapy of malaria. In the present time QA is
widely used for therapy of lambliasis, diphyllobothriasis, leishmaniasis, taeniasis, lupus
erythematosus et al. [4, 5]. It was shown that QA bound to prion peptides blocks their
transformation to pathogenic form that makes it promise candidate for design of drugs against
prion encephalopathy [6]. Such a wide range of QA action causes a great scientific interest
and the need to clarify the molecular mechanisms of complex formation between them and
biopolymers, as well as to account their structural features. The dye represents dication
compound with planar aromatic tricyclic structure, therefore its interaction with charged
nucleic acids (NA) and peptides is of special interest. Being the key participants of biological
processes, these biopolymers undergo the structural transformations as a result of the complex
formation, which can affect their functioning. For example, interaction of DNA with acridine
derivatives leads to the frameshift mutation, photocleavage and another effect [7-10].
It is well known that under the interaction of acridine derivatives with DNA, together
with strong intercalation binding, weaker external binding takes place which is conditioned by
electrostatic coupling of positively charged chromophores with negatively charged phosphate
groups of DNA backbone [11, 12]. Its contribution to the complex formation can be
significant even at physiological ionic strength of solution and high values of molar
phosphate-to-dye ratios (P/D) [13, 14]. However, in spite of great number of works
devoted to study of the dye–NA interaction, up
to now no any reliable data about quantitative
ratio between contributions of these binding
modes that can be conditioned by small number
of experimental techniques suitable to separate
them efficiently. Analysis of literature available
shows that main attention was paid to the
FIG. 1. Molecular structure of quinacrine
intercalation binding of acridines to NA at high
dye (QA).
P/D ratios [15]. The contribution of the external electrostatic binding was neglected,
considering it as unsubstantial at physiological ionic conditions. This could leads to some
error in the evaluation of binding cooperativity. External binding have also to be taken into
account when using fluorescent probes for investigation of intracellular processes.
3
Polarized fluorescent spectroscopy is powerful experimental technique which can be used
to estimate quantitatively contribution of different binding modes to dye–NA complex
formation since it is very sensitive to any changes in environment properties and
chromophore mobility. However the data about changes in fluorescence polarization degree
of organic dyes during its binding to NA are very limited. In particular, no data available on
the change in fluorescence polarization degree when QA binds to DNA in the wide range of
concentrations, different solution ionic strength and P/D ratios.
In the present work binding of quinacrine to DNA have been studied at different
experimental conditions (in the wide range of phosphate-to-dye molar ratios, P/D; at different
values of dye concentrations (10-6÷10-4 М) and solution ionic strength (0.0012, 0.01 and
0.15 М Na+)) using techniques of absorption and polarized fluorescent spectroscopy to
estimate quantitatively the contributions of intercalative and external electrostatic binding to
complex formation.
II. EXPERIMENTAL
The quinacrine dihydrochloride (QA, Fig. 1) was purchased from Serva (Heidelberg,
Germany). High-polymer DNA sodium salt from chicken erythrocyte (containing 41% of GC
base pairs) was obtained from Reanal (Hungary).
For all experiments 1 mM sodium cacodylate buffer, pH 6.9, containing 0.1 mM
Na2EDTA and prepared in fresh deionized distilled water to which NaCl (Sigma Chemical
Co.) was added to give concentrations of 0.01 and 0.15 M Na+. Buffers were ultra filtered
through nitrocellulose filters (Millipore-Q system, USA) with a pore diameter of 0.22
microns.
The concentrations of the dye and DNA were determined spectrophotometrically using
the molar extinction coefficient of 260 = 6600М-1сm-1 for DNA and 424 = 9750 М-1см-1 for
QA [16]. Samples for spectral measurements were prepared by mixing the stock solutions of
known concentration in predetermined volume proportions.
Electronic absorption spectra were measured on a SPECORD UV-VIS spectrophotometer
(VEB Carl Zeiss, Jena). Measurements of steady-state fluorescence intensity were carried out
by the method of photon counting with a laboratory spectrofluorimeter based on double
monochromator DFS-12 (LOMO, Russia). Fluorescence excitation was performed by
linearly-polarized beam of He-Cd laser LPM-11 (λ = 441.6 nm), which power was stabilized
during the experiment using hand-made set-up described in [17]. The fluorescence intensity
4
was registered at the right angle to the incident beam. Ahrens prisms were used to polarize
linearly the exciting beam as well as to analyze the fluorescence polarization. The
spectrofluorimeter was equipped with a quartz depolarizing optical wedge to exclude the
monochromator polarization-dependent response. When measuring the fluorescence intensity,
the pulses from photomultiplier tube were accumulated during 10 s for each data point and
measurements were repeated five times, at that the measurements error was about 0.5%. Also
the correction was made to the absorption of the laser beam in the solution layer of the front
wall to the cell center. Fluorescence spectra were corrected on the spectral sensitivity of the
spectrofluorimeter. Experimental set-up and the measurement procedure were described
earlier [18]. The total fluorescence intensity, I, its polarization degree, p, and anisotropy, μ,
were calculated using formulas [19]:
I
p
II
I
II
I
I
II
I
I
I II
I
2
I
I
p
2
p
3
(1)
(2)
(3)
III
and I - are measured intensities of the emitted light, which are polarized parallel
where
and perpendicular to the polarization direction of the exciting light beam, respectively.
The binding of quinacrine to DNA was followed by changes in parameters of the dye
fluorescence under titration experiments at several fixed values of QA concentration being in
the range of 10-6 ÷ 10-4 М. The QA solution was added with QA–DNA complex containing
the same concentration of the dye that gives the possibility to obtain the sample of required
P/D ratio without changing the concentration of the dye.
Fluorescence intensity and polarization degree of complexes were registered at the
wavelength corresponding to the maximum of free QA, λobs = 510 nm.
All measurements were carried out in 1 cm quartz cell at room temperature from 22 to
24 C.
III. RESULTS AND DISCUSSION
Visible electronic absorption and fluorescence spectra of quinacrine in a free state and bound to
DNA are depicted in Fig. 2. Absorption spectrum of the free dye represents the superposition of two
intense bands with maxima at 424 and 445 nm and less intensive shoulder. The fluorescence
spectrum is a broad intense unstructured band centered at 510 nm. Fluorescence polarization degree,
p, amounts to 0.035.
5
The results of the titration of QA solution with QA–DNA complex are presented in Figs. 3(a,b)
as P/D dependent changes in the relative fluorescence intensity, I/I0 (Fig. 3a), and polarization
degree, p (Fig. 3b). Here I0 and I are fluorescence intensities of the free and bound dye
correspondingly, measured at λ = 510 nm. The curves illustrate dependence of QA binding to DNA
on the solution ionic strength at the low dye concentrations (СQA = 10-6 М). It is clearly seen that the
binding results in the strong changes in the QA absorption and fluorescent characteristics, especially
at P/D 30 where two parts of the titration curves can be identified. So in the solutions of low ionic
strength, [Na+] = 1.2 mM, the initial part of titration curve (P/D = 0 ÷ 4) is linear, that is typical for
outside electrostatic binding of cationic dyes to polyanions [20-22].
ions,
its
Increase in solution ionic strength
reduces the dye binding to DNA due
the competition between QA
to
molecules and sodium
that
results in decrease of the depth of
fluorescence titration curve minimum
for the sample with [Na+] = 0.01 M
(I/I0 = 30 %)
and
total
disappearance
in sample of near
physiological
strength,
[Na+] = 0.15 M. In the last case the
P/D dependent rise of fluorescent
polarization degree is substantially
FIG. 2. Normalized absorption (left) and fluorescence
spectra (right) of free quinacrine (1) and its complex with
slighter in comparison with that in
DNA registered at P/D = 100 (2) in aqueous buffered
solution of low ionic strength (Fig.
solution at room temperature.
3b). This evidences that the sodium ions weaken not only outside electrostatic binding of QA to
DNA, but the intercalative one also. However fluorescence polarization degree also reaches the
constant level of p = 0.3 at P/D ≈ 200 for the sample containing 0.01 M of Na+, and at P/D ≈ 1000
for that with [Na+] = 0.15 M (not shown).
ionic
In Fig. 4 we can see the fluorescent titration curves registered in solutions of low and high ionic
strengths at higher concentration of the dye, СQA = 10-4 М. For these samples stronger fluorescent
quenching were observed as compared with previous system (СQA = 10-6 М, Fig 3a): in the curve
6
minimum at P/D = 4 the emission is only 9 % from initial, since the fraction of the dye externally
bound to DNA is proportional to its concentration.
More visually the dependence of QA to DNA binding on dye concentration are demonstrated
in Fig. 5, where QA–DNA fluorescent titration curves are represented for the three samples at the
following dye concentrations, СQA = 10-6; 2·10-5; 10-4 М, and low Na+ content (1.2 mM). From the
figure it is seen, that that growth of QA concentration strengthens appreciably quenching of its
emission that results in a deepening of a minimum on the titration curve. However, practically total
fluorescence quenching have not observed, as it was registered earlier for QA complex with
inorganic polyphosphate (see dashed line in Fig. 5 [27]), where only pure electrostatic binding
occurred, since even at P/D = 4 nonzero contribution of intercalation binding mode to QA–DNA
complex formation takes place.
Simultaneous measurements of fluorescence intensity and anisotropy of complexes vs
P/D ratio allow to separate the contributions of outside and intercalative binding to QA –
DNA complex formation, as well as to determine the fraction of free and bound dye
molecules in the samples, since these parameters are substantially different.
FIG. 3. Dependence of QA normalized fluorescence intensity (left) and polarization degree (right) on
molar DNA-to-dye ratio (P/D) obtained in the solutions of different ionic strength: () – 1.2 mM
Na+, () – 0.01 M Na+, () – 0.15 M Na+, CQA = 10-6 M; λexc = 441.6 nm; λobs = 510 nm.
7
FIG. 4. Titration curves plotted as normalized
fluorescence intensity vs P/D for QA–DNA
complexes, СQA = 10-4 М, in solutions of different
() –
ionic
0.1 М Na+; λexc = 441.6 nm; λobs = 510 nm.
() – 1.2 mМ Na+,
strength:
FIG. 5. Titration curves plotted as normalized
intensity vs P/D for QA–DNA
fluorescence
complexes at
following QA concentrations:
() – 10-4 М, () – 2·10-5 М, () – 10-6 М, in
1.2 mМ Na+;
solutions
λexc = 441.6 nm; λobs = 510 nm.
line
corresponds
quinacrine–polyphosphate
complex, СQA = 5·10-5 М.
containing
Dashed
to
For this purpose, the system of equations were used [28,29]:
2
0i
i
fI
i
i
I
2
0i
fI
i
i
I
if
1
,
2
i
0
(4)
(5)
(6)
where
if - fraction,
iI – fluorescence intensity and
i – fluorescence anisotropy for the dye
in i-state; here i = 0 corresponds to the free QA, i = 1 – to the dye intercalated into DNA and
if vs P/D were
i = 2 – to externally bound molecules. From the data obtained the plots of
constructed (Fig. 6). The curves presented in this figure relates to QA–DNA samples containing
2·10-5 М of quinacrine and 1.2 mМ Na+. They illustrate the contributions of two QA–DNA binding
modes at different ratios of the dye and polymer in solution. So, intercalative binding mode
dominates at high P/D values, whereas at low ones an external electrostatic binding prevails. With
decreasing P/D, the population density of intercalation binding sites is reduced, that leads to the
redistribution of QA molecules between two bound states in favor of external binding. For example,
8
at P/D = 50 approximately 91.2 % of quinacrine molecules are intercalated to double helix, 8,3 %
involved to external electrostatic binding, and only 0.5 % does not participate in the complex
formation. In the minimum of fluorescent titration curve at P/D = 4 the major part of dye molecules
(78.5 %) are included in the complexes of second type (external binding), and only 14 % of the
molecules are intercalated into double helix; 7.5 % of the dye remains unbound.
to DNA: () – intercalation,
FIG. 6. Fractions of QA in a free state () and
bound
() –
external binding. Data were calculated for QA-
DNA complexes (CQA = 2·10-5 M) in the solution
containing 0.0012 M Na+.
FIG. 7. Scatchard plot constructed for QA bound
to DNA in the solution containing 0.15 M Na+,
САХ = 10-6 M. Solid line corresponds to the curve
calculated from McGhee and von Hippel equation
(7) at К = 1.1·105 M-1 and m = 8.
Parameters of QA intercalative binding to DNA were determined from Scatchard plot (Fig.
7) constructed for the sample with low dye concentration, [QA] = 10-6 M, and high content of
sodium ions, 0.15 M Na+, where external electrostatic binding can be neglected. This binding
type is characterized by non-linear adsorption isoterm, which is well described by theoretical
dependence, obtained for non-cooperative binding of the ligands to homopolymer [30]:
r
C
K
(1
m
[1 (
mr
)
m
r
1) ]
m
1
(7),
where К is apparent binding constant, m is a size of binding site, equal to the number of base pairs
per one dye molecule, r – number of bound dye molecules per one base pair, С – concentration
of free dye. The best fit of experimental points was obtained using the next parameter in the
equation (7): m = 8, K = 1.1·105 M-1 (Fig. 7). The result means that the saturation of intercalation
binding occurs at one dye molecule per 8 base pairs of DNA. The obtained m value of is in good
9
agreement with the data [31]. Under reduction of solution ionic strength to [Na+] = 1.2 mM the
binding constant K characterizing the intercalation increases to approximately 2·109 M-1 [16].
The external electrostatic binding is characterized by substantially smaller binding sites, m = 1,
and apparent binding constant which is three orders of magnitude less than that for the intercalation.
It was estimated using model system of quinacrine – inorganic polyphosphate where only the pure
electrostatic binding of QA to DNA occurs. In solution with low Na+ content, 1.2 mM, the apparent
binding constant was found to be equal K = 1,7·106 М-1 [27]. The described above transition of the
dye from the intercalated to externally bound state can be explained using McGhee and von Hippel
theory [30, 32], videlicet, under the binding saturation the total increment of free energy can be
greater in the case of less strong interaction at the cost of higher density of the lattice binding sites.
IV. CONCLUSIONS
Quinacrine binding to DNA is realized via two competing mechanisms: (i) intercalation of the
dye between the biopolymer base pairs and (ii) external electrostatic binding of the dye to DNA
polyanionic backbone with - stacking of its chromophores. Simultaneous measurements of
fluorescence intensity and polarisation degree give the possibilty not only determine the fraction of
free and bound dye molecules, but quantitatively separate the contributions of these two binding
modes. It was shown that in the case of outside electrostatical binding the size of binding site
(measured in the DNA base pairs per one quinacrine molecule) is 8 times less than that for
intercalation, that allows him to compete successfully. Upon the saturation of the QA binding an
external type of complex formation can be dominant.
REFERENCES
1. T. Caspersson, L. Zech, E.J. Modest, G. E. Foley, U. Wagh, E. Simonsson, Exp.Cell Res.
58 (1), 128 (1969).
2. T. Caspersson, L. Zech, E.J. Modest, G. E. Foley, U. Wagh, E. Simonsson, Exp. Cell
Res. 58 (1), 141 (1969).
3. T. Caspersson, L. Zech, C. Johansson, E. J. Modest, Chromosoma 30 (2), 215 (1970).
4. L.S. Goodman, in The Pharmacological Basis of Therapeutics (Macmillan, New York,
1975) 1030.
5. D. J. Wallace, Semin. Arthritis Rheum. 18, 282 (1989).
6. K. Doh-ura, K. Ishikawa, I. Murakami-Kubo, K. Sasaki, S. Mohri, R. Race, T. Iwaki,
J. Virol. 78, 4999 (2004).
10
7. L. R. Ferguson, W. A. Denny, Mutat. Res. 623, 14 (2007).
8. K. Krzyminski, A. D. Roshal, A. Niziolek, Spectrochim. Acta A Mol. Biomol. Spectrosc.
70, 394 (2008).
9. C. Valencia, E. Lemp, A.L. Zanocco, J. Chil. Chem. Soc. 48 (4), 17 (2003).
10. C. Di Giorgio, A. Nikoyan, L. Decome, C. Botta, M. Robin, J.-P. Reboul, A.-S. Sabatier,
A. Matta, M. De Méo, Mutat Res. 650 (2) , 104 (2008).
11. H. Porumb, Progr. Biophys. Mol. Biol. 37, 175 (1978).
12. H. M. Berman, P. R. Young, Annu. Rev. Biophys. Bioeng. 10, 87 (1981).
13. H. J. Li, D. M. Crothers, J. Mol. Biol. 39, 461 (1969).
14. D. E. V. Schmechle, D. M. Crothers, Biopolymers 10, 465 (1971).
15. K. Nakamoto, M. Tsuboi, G. D. Strahan, Drug-DNA interactions: structures and spectra
(John Wiley & Sons, New York, 2008).
16. W. D. Wilson, J. G. Lopp, Biopolymers, 18, 3025 (1979).
17. I. N. Govor, V. M. Nesterenko, Pribory i tekhnika experimenta (Rus) 3 168 (1974).
18. V. Zozulya, Yu. Blagoi, G. Lober, I. Voloshin, S. Winter, V. Makitruk, A. Shalamay,
Biophys. Chem. 65, 55 (1997).
19. J. R. Lakowicz, Principles of Fluorescent Spectroscopy (Kluwer Academic/Plenum Press,
New York, 1999).
20. G. Schwarz, S. Klose, Eur. J. Biochem. 29, 249 (1972).
21. Z. Balcarova, V. Kleinwachter; J. Koudelka, Biophys. Chem. 8, 17 (1978).
22. V. N. Zozulya, V. F. Fyodorov, Yu. P. Blagoi, Stud. Biophys. 137(1-2), 17 (1990).
23. U. Pachman, R. Rigler, Exper. Cell. Res. 72, 602 (1972).
24. O. F. Borisova, A. P. Razjivin, V. I. Zaregorodzev, FEBS Letters. 46, 239 (1974).
25. G. Baldini, S. Doglia, G. Sassi, G. Lucchini, Int. J. Biol. Macromol. 3, 248 (1981).
26. G. G. Sheina, A. P. Limansky, T. F. Stepanova, Yu. P. Blagoi, Stud. biophys. 100(3), 187
(1984) .
27. V. N. Zozulya, I. M. Voloshin, Biophys. Chem. 48, 353 (1994).
28. D. A. Deranleau, Th. Binkert, P. Bally, J. Theor. Biol. 86, 477 (1980).
29. G. Weber, Biochem. J. 51, 145 (1952).
30. J. D. McGhee, P. H. Von Hippel, J. Mol. Biol. 86, 469 (1974).
31. M. Aslanoglu, G. Ayne, Anal. Bioanal. Chem. 380, 658 (2004).
32. B. Gaugain, J. Barbert, N. Capelle, B. P Roques, J. B. Le Pecq, Biochemistry 17, 5078
(1978).
|
1612.08995 | 1 | 1612 | 2016-12-28T21:55:08 | Controlling uncertainty in aptamer selection | [
"physics.bio-ph",
"q-bio.PE"
] | The search for high-affinity aptamers for targets such as proteins, small molecules, or cancer cells remains a formidable endeavor. Systematic Evolution of Ligands by EXponential Enrichment (SELEX) offers an iterative process to discover these aptamers through evolutionary selection of high-affinity candidates from a highly diverse random pool. This randomness dictates an unknown population distribution of fitness parameters, encoded by the binding affinities, toward SELEX targets. Adding to this uncertainty, repeating SELEX under identical conditions may lead to variable outcomes. These uncertainties pose a challenge when tuning selection pressures to isolate high-affinity ligands. Here, we present a novel stochastic hybrid model that describes the evolutionary selection of aptamers in order to explore the impact of these unknowns. To our surprise, we find that even single copies of high-affinity ligands in a pool of billions can strongly influence population dynamics, yet their survival is highly dependent on chance. We perform Monte Carlo simulations to explore the impact of environmental parameters, such as the target concentration, on selection efficiency in SELEX and identify new strategies to control these uncertainties to ultimately improve the outcome and speed of this time- and resource-intensive process. | physics.bio-ph | physics |
Controlling Uncertainty in Aptamer Selection
Fabian Spilla,b, Zohar B. Weinsteinc, Nga Hob, Atena Irani Shemiranib, Darash Desaib,1, and Muhammad H. Zamanb,d,1
aDepartment of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA; bDepartment of Biomedical Engineering, Boston University,
Boston MA 02215, USA; cDepartment of Pharmacology and Experimental Therapeutics, Boston University School of Medicine, Boston, MA 02118; dHoward Hughes Medical
Institute, Boston University, Boston, MA 02215, USA
This manuscript was compiled on June 30, 2021
The search for high-affinity aptamers for targets such as proteins,
small molecules, or cancer cells remains a formidable endeavor. Sys-
tematic Evolution of Ligands by EXponential Enrichment (SELEX)
offers an iterative process to discover these aptamers through evo-
lutionary selection of high-affinity candidates from a highly diverse
random pool. This randomness dictates an unknown population dis-
tribution of fitness parameters, encoded by the binding affinities, to-
ward SELEX targets. Adding to this uncertainty, repeating SELEX
under identical conditions may lead to variable outcomes. These
uncertainties pose a challenge when tuning selection pressures to
isolate high-affinity ligands. Here, we present a novel stochastic hy-
brid model that describes the evolutionary selection of aptamers in
order to explore the impact of these unknowns. To our surprise, we
find that even single copies of high-affinity ligands in a pool of bil-
lions can strongly influence population dynamics, yet their survival
is highly dependent on chance. We perform Monte Carlo simulations
to explore the impact of environmental parameters, such as the tar-
get concentration, on selection efficiency in SELEX and identify new
strategies to control these uncertainties to ultimately improve the
outcome and speed of this time- and resource-intensive process.
Aptamer SELEX Evolutionary Dynamics Stochastic Process Hybrid
Model
Understanding and exploiting target-ligand binding are
bedrocks of the biomedical sciences and support a host of
applications ranging from diagnostics, therapeutics, and drug
discovery to biosensing, imaging, and gene regulation. Anti-
bodies and rational design provide a constructive playground
to develop these applications, yet there generally remains a
paucity of strong and specific binders for the innumerable
viral, protein, and small molecule targets under investigation.
Aptamers offer an alternative to antibodies, yet in spite
of their growth [1 -- 4], the discovery of high-affinity aptamers
remains a challenge, especially for small molecule targets [5, 6].
Systematic Evolution of Ligands by EXponential Enrichment
(SELEX) [7, 8]
is the premier framework for aptamer de-
velopment and isolates high-affinity ligands from an initial
library similar to how advantageous traits are enriched in a
biological population through Darwinian selection. In a cyclic
process, ligands are incubated with the target, and those that
exhibit preferential binding are amplified and survive to the
next round. Target molecules are typically immobilized on a
substrate material to facilitate easy separation of target-bound
and unbound ligands. Through numerous rounds of selection,
an initial library can be reduced to a handful of high-affinity
aptamers. Nucleic acids comprise the vast majority of libraries
used in SELEX, where sequence regions are randomized to
generate tremendous structural diversity. While this diversity
underpins the evolutionary nature of SELEX, numerous works
suggest that initial library design is a significant contributor
to its overall success [9].
LEX is plagued by uncertainty. Despite the impact of library
design, the initial affinity distribution for any library toward a
specific target remains a priori unknown. Target immobiliza-
tion further complicates the procedure, particularly for small
molecules. In comparison to large molecular weight targets
such as proteins [10], viruses [11], and whole cells [12, 13]; the
immobilization of small molecules eliminates ligand binding
sites and is thus impractical. Newer approaches instead bind
the library itself to a substrate material using non-covalent
equilibrium binding, but this introduces the opportunity for
competitive losses of high-affinity ligands that are initially
present in extremely low numbers. Wash steps and other
experimental procedures may lead to further random losses,
while non-specific selection of ligands can counter environmen-
tal pressures and stall selection. In short, these uncertainties
may quickly compound to apply tremendous risk toward the
guarantee of successful selection.
Mathematical modeling therefore has great potential to
help understand the uncertainties of aptamer selection and
devise strategies to optimize environmental parameters and
improve selection outcomes. Previous models have explored
SELEX for protein targets, considering parameters such as
target concentration [14 -- 16], separation efficiency of target-
bound and unbound ligand [17], nonspecific binding of DNA
to target [18], and negative selection steps [19]. These studies
predict that, in spite of its experimental complexity, the evo-
lutionary nature of SELEX guarantees selection of the highest
affinity ligand from the initial library. However, these works
focus primarily on the use of deterministic equilibrium equa-
Significance Statement
Oligonucleotide aptamers have increasing applications as a
class of molecules that bind with high affinity and specificity to
a target. Aptamers are typically selected from a large pool of
random candidate nucleic acid libraries through competition for
the target. Using a novel stochastic hybrid model, we are able
to study the combined impact of important evolutionary success
factors such as competition, randomness, and changes in the
environment. While the environment may be tuned with experi-
mental parameters such as target concentration, competition
varies with differences in the initial distribution of aptamer-target
binding affinities, and random events can eliminate even the
ligands with the highest affinity.
Author contributions: F.S., Z.B.W., D.D. and M.H.Z. designed research, F.S. developed the model,
F.S. and D.D. implemented the model, F.S. and D.D. analyzed the data, Z.B.W, N.H., A.I.S. and D.D.
contributed new reagents/analytic tools, F.S., D.D. and M.H.Z. wrote the paper.
The authors declare no conflict of interest.
1To whom correspondence should be addressed. E-mail: [email protected] (DD), [email protected]
(MHZ)
While conceptually simple, the practical application of SE-
www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX
PNAS June 30, 2021
vol. XXX no. XX 1 -- 6
tions [14], whereas the presence of ligands in low copy numbers
and the role of other experimental uncertainties suggest the
use of more fundamental stochastic models rather than de-
terministic approximations. Mathematically, the chemical
master equation provides a framework to test this hypothesis
and generalize the above-mentioned deterministic models to
include intrinsic stochasticity [20]. While this approach could
be applied toward a purely stochastic model for SELEX, the
result cannot currently be solved analytically or simulated by
conventional techniques such as the Gillespie algorithm [21],
due to the large number of molecules present. These limita-
tions are common for many stochastic multiscale problems in
biology, chemistry and physics; and the development of novel
analytic approximations or numerical techniques to address
this problem is an important ongoing research topic [22].
Using these ideas as our foundation, we introduce a new
hybrid model for aptamer selection that builds on the chemical
master equation to introduce stochastic uncertainty in SELEX
modeling. Here, ligands are separated into two categories of
high and low copy number. In the former case, the master
equation is simplified toward a deterministic equilibrium sys-
tem, whereas in the latter it can be approximately solved
analytically. Unlike previous efforts to incorporate stochastic-
ity into aptamer modeling [23, 24], our framework allows us
to simultaneously investigate the impact of low copy number
ligands and their competitive binding to target molecules and
immobilization substrates among the presence of high copy
number ligands. Most importantly, this approach can capture
total loss of individual ligands, which can strongly contribute
to protocol outcome. Such events have not previously been
investigated and cannot be captured by other approximations
of the master equation such as the Langevin approximation,
which rely on the presence of sufficiently high numbers of
molecules and thereby diminish the possibility of extinction
events [25].
Using this framework, we investigate unexplored sources
of uncertainty in SELEX, beginning with a systematic anal-
ysis of the role the initial library affinity distribution plays
in selection. We further challenge the assumption that this
distribution is continuous at its tails and evaluate the impact
of adding noise at these extremes. We find that introducing
as few as 20 additional ligands outside the bulk distribution
of 1015 molecules can strongly affect the outcome of selection.
In light of these results, we revisit the topic of optimizing
target concentration as discussed in previous works [14 -- 16],
and show that the assumed initial KD distribution strongly
influences protocol optimizations. We also provide additional
insights regarding non-covalent ligand immobilization to sup-
port more recent efforts to develop robust protocols for small
molecule SELEX [26 -- 28].
Integrating these ideas, we show
that simultaneously lowering the target concentration and the
substrate binding dissociation constant over the SELEX cycles
can lead to improved selection outcomes for a wide range of
initial conditions.
Computational Model of Selection Dynamics
The original SELEX protocol [7, 8] serves as the basis for our
model, with additional modifications to accommodate small
molecule targets as described in [26]. While this marks the
first model that specifically considers small molecule targets,
the main ideas and conclusions derived from this work remain
Fig. 1. A sample candidate library of ligands Ai is prepared by letting the ligands
bind to a substrate S. Then, the target is added, leading to competitive binding
between the different apatamers for substrate and target molecules T . The ligands
still bound to the substrate are then separated from those which are either bound to a
target, or have randomly unbound from the substrate. The latter two are subsequently
amplified and taken into the next cycle.
i
applicable for other targets and selection schemes. The main
steps of our approach are summarized in Fig. 1. We begin
ligands of type i, where i = {1, . . . , M A}
with a library of Atot
and M A is the total number of unique ligands. The ligands are
then non-covalently immobilized using S substrate molecules,
where KS is the ligand-substrate dissociation constant. These
complexes are then subjected to wash steps to remove unbound
ligands, from which AI
i ligands of type i survive. Surviving
ligands are then incubated with T target molecules, where a
ligand of type i binds to the target with a dissociation constant
KD,i. Ligands that are bound to a target or have unbound
from the substrate are partitioned from those that remain
bound to the substrate. Finally, the partitioned ligands are
amplified via PCR, modeled as a constant factor increase of
αP CR, and used to begin the next cycle. The proceeding
sections highlight the notable details of our hybrid approach,
while a more thorough description and derivation of the model
can be found in the supporting information. Throughout
these sections, quantities that refer to an absolute number
of molecules are denoted with a tilde, while those without
represent concentrations.
Deterministic Model of Ligand Binding. Earlier works use
equilibrium conditions to characterize ligand-target interac-
tions during selection [14 -- 17] , focusing on changes in bulk
properties, such as the mean dissociation constant, to study
the enrichment of a single best candidate. We instead monitor
the full ligand affinity distribution in an effort to better under-
stand how parameters such as the initial standard deviation
also impact selection dynamics. Since modeling each of the
M A ≈ 1015 unique ligands is computationally intractable, we
discretize the initial distribution of M A unique ligands into
M B bins, each containing Ai ligands of dissociation constant
KD,i, where i = {1, . . . , M B}. We choose M B to be large
enough that the results do not depend on the binning, and
small enough to optimize simulation performance. We further
build on this analysis by introducing additional equilibrium
2 www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX
Spill et al.
conditions for non-specific ligand-substrate interactions repre-
sented by a dissociation constant KS.
In [26], substrate-ligand
binding is accomplished through DNA base pairing using a
fixed sequence, and is thus constant. Altering the length of
this fixed sequence is a means to tune KS. Moreover, different
immobilization techniques, such as the use of graphene ox-
ide [27, 28], will lead to variations of KS within a given pool,
but we do not consider such cases here and consider KS to
be constant throughout a single cycle of SELEX. Combining
ligand-target and ligand-substrate binding, the full system of
steady-state equilibrium binding conditions can be described
by the set of equations:
(AI
i − [SAi] − [T Ai])Sf ree,
i − [SAi] − [T Ai])T f ree,
(AI
i = 1, . . . , M B,
i = 1, . . . , M B,
KS
[SAi] = 1
[T Ai] = 1
M BX
Stot =
KD,i
i=1
M BX
i=1
[SAi] + Sf ree, T tot =
[T Ai] + T f ree
[1]
Here, [SAi] and [T Ai] denote the concentration of ligand-
substrate and ligand-target complexes, representing 2M B in-
dependent variables that are solved; the quantities T tot, T f ree
and Stot, Sf ree denote the concentrations of total and free
target and substrate, respectively. From these results, we
determine the concentration of ligands which survive selection,
denoted by AS,D
, and are amplified by PCR for the next cycle.
The superscripts denote that this number is obtained after
selection and using the deterministic model defined by Eq. (1).
This concentration is simply the sum of free and target-bound
ligands, and is hence given by
i
i = [T Ai] + Af ree
AS,D
i
= AI
i − [SAi].
[2]
Stochastic Model of Ligand Selection. Chemical reactions are
fundamentally stochastic in nature, with forward and backward
reactions occurring constantly. While powerful and simple,
Eq. (1) is based on real-valued concentrations which require
sufficiently high molecular copy numbers to make discreteness
and random fluctuations negligible. This is challenged at the
tails of the KD distribution, where appropriate binning results
in few ligands per bin. To address this, a hybrid approach
is used where additional stochastic analysis is applied when
Eq. (1) predicts AS,D
to be below a threshold Θ. To distinguish
these quantities for stochastic analysis, we denote them as
AS,D
ψ , where ψ represents the subset of indices i that satisfy
the condition AS,D
i < Θ. Results exploring the choice for Θ
are provided in SI Fig. S6. We then calculate the probability
for selecting AS,S
that the number is obtained after selection and using the
stochastic model. As described in the supporting information,
we find that by starting with the chemical master equation,
(cid:1); the superscripts denotes
(cid:1) is well-approximated by a binomial distribution:
p(cid:0) AS,S
ligands, p(cid:0) AS,S
p(cid:0) AS,S
(cid:18) Atot
(1 − pψ) Atot
(cid:1) =
ψ − AS,S
ψ ,
(cid:19)
ψ
ψ
ψ
ψ
i
AS,S
ψ
ψ
ψ
[3]
AS,S
p
ψ = 0, . . . , Atot
ψ
ψ
for AS,S
Here, the quantity pψ represents the probability that a single
ligand is selected out of Atot
ψ ligands of type ψ. To provide the
most accurate description, we account for stochastic contribu-
tions from both the immobilization and incubation steps. The
contribution from immobilization is approximately the same for
all candidates, and is given by AI
Atot , the fraction of remaining
immobilized ligands after wash steps over those present before
i=1 AI
i=1 Atot
immobilization, where AI = ΣM B
.
The contribution from incubation is calculated as the fraction
of predicted ligands, AS,D
ψ , out of an initial number of AI
ψ.
Using these contributions, the total probability that a ligand
in bin ψ survives is given by:
i and Atot = ΣM B
i
pψ =
AI AS,D
AtotAI
ψ
ψ
[4]
ψ = Atot
ψ,N + Atot
Finally, Eq. (3) requires Atot
ψ to be integer-valued, as it denotes
a number of molecules. However, the deterministic equations
yield real-valued concentrations that must be renormalized to
an integer. We separate Atot
ψ into its integer and fractional
parts, Atot
ψ,f < 1
as the probability to have an extra molecule present. We then
draw a uniformly distributed random number 0 ≤ r ≤ 1, and
set Atot
ψ,N otherwise.
Following this renormalization, we finally draw a random
variate distributed according to Eq. (3) to simulate the set
of ligands AS,S
that remain after both immobilization and
selection.
ψ,f, and then interpret 0 ≤ Atot
ψ,N + 1 if r < Atot
ψ,f, and Atot
ψ = Atot
ψ = Atot
ψ
Results and Discussion
Utilizing a hybrid computational approach, our model pro-
vides a generalized framework that can be used to analyze
both deterministic and stochastic effects in SELEX. We use
the model to deconstruct two main forms of uncertainties
in aptamer selection. The first is parameter uncertainty, in-
cluding the unknown initial KD distribution as well as the
experimentally tunable quantities KS and T tot. These are
analyzed using a parameter study that observes the impact of
these factors on SELEX dynamics. The second is stochastic
uncertainty associated with low copy number binding phe-
nomena. As this form of uncertainty is random in nature, we
employ Monte Carlo simulations to observe the variability in
outcomes between repeated SELEX procedures and extract
conclusions which are robust with respect to stochastic fluc-
tuations. Unless mentioned otherwise, the parameters from
Table S1 are used in all simulations.
Effect of KD Distribution on Selection Efficiency. Gaussian
distributions describing the initial ligand pool dominate SE-
LEX models in literature [16], yet we are not aware of any
prior systematic approach to study the impact of various distri-
butions on the outcome of SELEX. While strong justifications
have been made for the assumption of a log-normal Gaussian
description [29], we explore various Gaussian as well as non-
Gaussian distributions and their impact on selection. Our
convention for log-normal KD distributions is such that a
Gaussian N(µ, σ) with mean µ and standard deviation σ in
log-space translates to a mean of 10µ in KD space; we do not
2 σ2 as is customary in Ito calculus. Fig. 2
shift the mean by 1
highlights the dramatic difference observed for just two differ-
ent assumed distributions , and demonstrates the significant
role the initial KD distribution plays in SELEX. This point
Spill et al.
PNAS June 30, 2021
vol. XXX no. XX 3
Fig. 2. Initial Distribution affects SELEX dynamics. We plot the distribution of ligand
binding affinities with increasing SELEX cycles for the same experimental parameters
and two different assumed Gaussian distributions at cycle 1, N(−3, 0.4) (blue
triangles) and N(−5, 0.8) (red dots). The dynamics of the two cases are totally
different. For N(−5, 0.8), the distribution shifts to the left and becomes considerably
narrower, and for N(−3, 0.4), the distribution additionally skews to the left, such
that from cycle 12 on the highest-affinity binders have outcompeted the rest of the
distribution.
Fig. 4.
Impact of target concentration on SELEX dynamics. Evolution of KD
distribution for three different values of the target concentrations is shown. Under
a high target concentration of T tot = 10−2M, the distribution shifts to the left
and narrows, but does not skew towards high-affinity ligands. Additional skewing is
achieved by reducing to T tot = 10−4M, which increases selection pressure by
intensifying ligand competition. However, further reduction to T tot = 10−8M has
the opposite affect and actually halts selection. In this case, the target concentration
is so low that non-specific ligand-substrate equilibria dominate selection dynamics
and nullifies the selection pressure.
Fig. 3. Noise affects SELEX dynamics. We fix the experimental parameters, and
the same initial Gaussian distribution N(−4, 0.4) with the same added noise of only
20 additional ligands initially present between KD = 10−10M and 5 × 10−8M.
Two different Monte Carlo simulations show dynamics of selection under random
loss of the 20 strongest binders, (blue triangles), versus dynamics when only two of
those strong binders with affinities between 10−10 and 10−9M are selected, (red
dots). In the latter cases, these two high-affinity binders completely dominate the
distribution from Cycle 12 on and outcompete the remaining ligands with low affinities
(KD > 10−7M).
Fig. 5. Optimal target concentrations strongly depend on assumed initial KD
distribution. Mean KD as a measure of pool binding strength for SELEX pool at cycle
20 using different constant target concentration. Depending on the initial distribution
of ligands, we find vastly different optimal target concentrations, i.e. concentrations
with lower mean KD.
is further accentuated by the fact that different selection tar-
gets may significantly alter the initial KD distribution for any
given library. SI Fig. S1 confirms that for a variety of other
distributions, including non-Gaussians, distribution shape has
a dramatic impact on selection dynamics.
In addition to shape, we also explore the assumption that
the KD distribution is continuous everywhere. While this
assumption is credible near the distribution mean where the
frequency of molecules is sufficiently high, we expect it to fail at
the extreme tail where stochastic effects dominate and highly
specific sequences can create gaps in the affinity distribution.
Indeed, it is well-known that even single base-pair changes in
DNA can dramatically impact binding [30]. Ligands in this
regime are highly prized, but may also be at highest risk to
be lost to stochastic effects due to low copy numbers.
We investigate this risk by using an initial N(−4, 0.4) dis-
tribution and adding a fixed noise component that is randomly
sampled from a uniform distribution in log-space. Fig. 3
and Movie S1 show a comparison of two Monte Carlo simu-
lations where there are only 20 ligands present in the range
of KD < 10−7M, i.e. where the continuous Gaussian dis-
tribution is effectively zero. We find that random binding
effects can lead to total loss of those 20 ligands, resulting in a
very different evolution of the KD distribution from cycle 12
onward in comparison to the case where only 2 of those ligands
survive. SI Fig. S2 shows a distribution of the mean ligand
KD at cycle 20 obtained from 250 Monte Carlo simulations,
confirming this enormous variability in outcomes, where the
mean KD value spans three orders of magnitude.
These results demonstrate the tremendous sensitivity of
selection dynamics to both distribution shape and noise. They
illustrate that selection pressures are parameterized not only by
extrinsic environmental conditions given by the experimental
setup, such as the tunable quantities KS and T tot, but just
as importantly by inherently uncertain intrinsic population
parameters that govern relative competition between ligands
of varying affinities.
Revisiting Target Concentration. Optimization of the target
concentration, T tot, has long stood as a critical step in adjust-
ing selection pressure based on experimental parameters [14 --
16]. However, the results from the previous section now suggest
that in addition to these experimental factors, the intrinsic
affinity distribution of the initial ligand pool may have a sig-
nificant influence on the impact T tot exerts on the overall
selection pressure. In light of this, we revisit the topic to study
this impact by varying both T tot and the initial distribution.
Fig. 4 and Movie S2 first show the dramatic impact of
drug concentration on selection dynamics. The results indicate
that T tot = 10−4M (blue) provides optimal selection out of
4 www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX
Spill et al.
KD value of ligands present at cycle 20 for nine different initial
KD distributions (Figs. 7, S7(d)-(f)). Similar to target concen-
tration, we find an optimum in the intermediate ranges of KS
and a clear dependence on the initial distribution. However,
contrary to target concentration, the mean KD for smaller
KS is relatively insensitive. Thus, these results suggest that
a lower value of KS = 10−16M would provide similar results
across a multitude of initial distributions.
As it pertains to small molecule selection schemes, these
results provide useful insights into the impact that substrate
binding affinity has on selection efficiency, and may offer some
guidance in the appropriate selection of a substrate material.
The results also provide general insights into the impact of
partitioning efficiency and non-specific binding on selection
across various initial distributions and suggest that a given
partitioning efficiency or fraction of non-specific selection can
impact different initial distributions in vastly different ways.
Improving Selection Efficiency. We have shown that the initial
KD distribution has a tremendous impact on selection effi-
ciency and plays a significant role in modulating the impact of
experimental parameters such as T tot and KS. These results
highlight that while established protocols are expected to per-
form well for some distributions, they may perform moderately
for others. To address this variability in outcomes, we finally
explore strategies to mitigate these impacts using only the
experimental parameters T tot and KS. As a metric for our
analysis, we introduce the quantity φ(c), which describes the
fraction of ligands with KD < 10−10M at cycle c = {1, . . . , C}.
Using this quantity, we further introduce two measures of ef-
ficiency: success probability Φ = φ(C) and success speed SC
defined as the cycle c at which φ(c) = 0.5φ(C).
We have seen that KS and T tot play distinct roles in the
evolutionary dynamics of the KD distribution. However, both
parameters exhibit regimes of optimal selection that depend
heavily on the initial distribution mean and width. Figs. 5
and 7 show that high values for T tot and KS have a similar
impact across all distributions, and suggest a conservative
approach of beginning at these high values for the initial
cycles. This reduces the risk of eliminating high-affinity, low
copy number ligands early on. As these high-affinity ligands are
amplified in subsequent rounds, T tot and KS can be lowered
to rapidly eliminate the remaining low-affinity ligands (see SI
Figs. S3, S4). While ideas to lower the target concentrations
have been discussed previously [26], our results indicate that
other parameters such as KS can be tuned simultaneously to
improve outcome across a multitude of initial distributions and
stochastic conditions. Fig. 8 shows Φ and SC obtained from
50 Monte Carlo simulations of an improved protocol where
both T tot and KS are decreased over the cycles as described in
Table S2. These results are compared to the original protocol
with constant values T tot = 10−4M and KS = 10−12M [26];
SI Fig. S5 shows φ(c) including the standard deviations. Using
six different initial Gaussian distributions with noise added
similar to Fig. 3, we observe that the improved protocol
with decreasing T tot and KS is faster and leads to a higher
fraction of high affinity binders than the original protocol. As
an alternative metric of protocol performance, SI Fig. S8 shows
the evolution of mean KD across the cycles, and also introduces
two alternative protocols where T tot or KS are decreased
faster than in the improved protocol. The results indicate
that while faster decreases can further improve performance
Fig. 6.
Impact of KS on SELEX dynamics. Evolution of KD distribution for three
different values of KS. Similar to target concentration, we find an optimal outcome
in the middle range (KS = 10−12M, blue), but the outcome for low KS is not
as adverse as for low T tot, since the distribution still shifts towards low KD with
increasing cycles.
Fig. 7. Optimal KS depends on initial distribution. Plot of mean KD for SELEX pool
at cycle 20 using different values of KS. Reducing KS from its optimal value does
not increase the mean KD as strongly as a reduction of the target concentration from
its optimum, as shown in Fig. 5.
the three investigated drug concentrations that use the initial
Gaussian distribution N(−4, 0.4). To investigate the impact
of T tot more systematically, Fig. 5 shows the mean KD value
of ligands selected after 20 cycles as a function of T tot for
nine different initial distributions. Note that as the mean KD
decreases, the average binding strength of the pool increases.
Fig. 5 confirms that intermediate values of T tot yields optimal
selection. SI Figs. S7 (a)-(c) further show that adding noise to
the initial distributions introduces additional variability, but
provides similar qualitative results. Interestingly, we find that
different initial distributions can have very different optimal
T tot, stressing the importance of devising a strategy to miti-
gate this impact and thereby control the inherent uncertainty
associated with the initial KD distribution.
KS Dependence and Non-specific Selection. Our
hybrid
model has allowed us to explore the impact of the unknown ini-
tial KD distribution and the target concentration T tot, which
are both present in all SELEX protocols. However, our model
additionally introduces a ligand-substrate interaction that has
never before been studied and offers a unique opportunity to
apply it toward more recent selection schemes aimed at small
molecule aptamer development [26 -- 28]. We therefore extend
our analysis to study uncertainties that govern an optimum
KS, and observe how changes in KS impact selection dynamics
for different KD distributions.
Fig. 6 and Movie S3 show the evolution of a single initial
KD distribution for three different values of KS, showing an
optimal outcome for KS = 10−12M (blue). Noting these dy-
namics, we next vary KS systematically and observe the mean
Spill et al.
PNAS June 30, 2021
vol. XXX no. XX 5
Fig. 8. Plots comparing the fraction of high-affinity ligands Φ and speed SC of
SELEX for six different KD distributions. The values are obtained from averaging
50 Monte Carlo simulations. We observe that the protocol with decreasing T tot
and KS over the rounds will lead to a higher fraction of strong binders (here, with
KD < 10−10M), and will reach this fraction faster, than the protocol where T tot
and KS are kept constant.
for some distributions, they may also lead to adverse outcome
for others.
Conclusions and Outlook
Deterministic models for SELEX have shed tremendous insight
on the challenges faced in aptamer selection, but have been un-
able to capture its inherently uncertain nature. Here, we have
presented a hybrid model that captures stochastic binding
and furthermore incorporates non-covalent ligand-substrate
immobilization. Using this framework, we have investigated
previously unexplored questions including the role of the ini-
tial library KD distribution, impact of distribution noise, and
the effect of these factors on the optimization of experimental
parameters such as the total target concentration T tot and
the substrate dissociation constant KS.
The results of our modeling draw striking parallels to out-
comes in evolutionary biology, where environmental parame-
ters define a fitness landscape and competition can change this
landscape to influence survival and reproduction [31]. Within
SELEX, ligands compete for target molecules to ensure sur-
vival into the next cycle, whereas substrate binding traps
the ligands and leads to their removal. Reduction of target
concentration can increase competition, but when few target
molecules are present, even high-affinity binders are unlikely
to find a target. Similar to competition in limited resources
scenarios, we find that the chance of survival for even the
highest affinity ligand strongly depends on the strengths of the
other ligands present in the population. Our surprising find-
ing that a handful of high-affinity ligands can outcompete a
pool of 1015 ligands is also seen in evolutionary biology, where
highly advantageous traits can quickly spread in a population,
given the right conditions. The model enables one to identify
the parameters impacting selection, and can thus be used to
improve selection efficiency. A further important component
of evolution in biological systems is mutations. Mutations in
SELEX can also appear during PCR amplification, but usually
lead to reduced affinities of the strongest aptamers [30], so
we ignored them in our current approach. However, for some
SELEX protocols, mutations can be beneficial to expand the
experimental sampling space [32], and it may be interesting
to extend our model to those protocols.
In summary, our novel model provides a better understand-
ing of the impact of the uncertainties in SELEX, and how
experimental parameters can be tuned to improve outcome
and speed of this expensive and time-consuming protocol. We
have demonstrated how optimization of the parameters can
enhance selection efficiency of one protocol dramatically, and
we envisage that simple adaptations of our model can be used
to improve the many other established protocols, as well as
guide the design of novel protocols, which aim to limit the
impact of uncertainties in selection methods.
ACKNOWLEDGMENTS. We acknowledge the support of the NCI
grant number 5U01CA177799, Saving lives at birth and USP/PQM
cooperative agreement. ZBW is supported by NIGMS Training
Program in Biomolecular Pharmacology T32GM008541.
1. Shangguan D et al. (2006) Aptamers evolved from live cells as effective molecular probes for
cancer study. Proceedings of the National Academy of Sciences 103(32):11838 -- 11843.
2. Ferguson BS et al. (2013) Real-time, aptamer-based tracking of circulating therapeutic agents
in living animals. Science Translational Medicine 5(213):213ra165 -- 213ra165.
3. Keefe AD, Pai S, Ellington A (2010) Aptamers as therapeutics. Nature Reviews Drug Discov-
ery 9(7):537 -- 550.
4. Bunka DH, Stockley PG (2006) Aptamers come of age -- at last. Nature Reviews Microbiology
4(8):588 -- 596.
5. McKeague M, DeRosa MC (2012) Challenges and opportunities for small molecule aptamer
development. Journal of nucleic acids 2012.
6. Blind M, Blank M (2015) Aptamer selection technology and recent advances. Molecular
Therapy Nucleic Acids 4(1):e223.
7. Tuerk C, Gold L (1990) Systematic evolution of ligands by exponential enrichment: Rna lig-
ands to bacteriophage t4 dna polymerase. Science 249(4968):505 -- 510.
8. Ellington AD, Szostak JW (1990) In vitro selection of rna molecules that bind specific ligands.
Nature 346(6287):818 -- 822.
9. Luo X et al. (2010) Computational approaches toward the design of pools for the in vitro
selection of complex aptamers. RNA 16(11):2252 -- 2262.
10. Cho M et al. (2013) Quantitative selection and parallel characterization of aptamers. Pro-
ceedings of the National Academy of Sciences 110(46):18460 -- 18465.
11. Roh C, Kim SE, Jo SK (2011) Label free inhibitor screening of hepatitis c virus (hcv) ns5b
viral protein using rna oligonucleotide. Sensors 11(7):6685 -- 6696.
12. Farokhzad OC et al.
(2006) Targeted nanoparticle-aptamer bioconjugates for cancer
chemotherapy in vivo. Proceedings of the National Academy of Sciences 103(16):6315 --
6320.
13. Sefah K et al. (2014) In vitro selection with artificial expanded genetic information systems.
14.
Proceedings of the National Academy of Sciences 111(4):1449 -- 1454.
Irvine D, Tuerk C, Gold L (1991) Selexion: Systematic evolution of ligands by exponential
enrichment with integrated optimization by non-linear analysis. Journal of molecular biology
222(3):739 -- 761.
15. Levine HA, Nilsen-Hamilton M (2007) A mathematical analysis of selex. Computational biol-
ogy and chemistry 31(1):11 -- 35.
16. Wang J, Rudzinski JF, Gong Q, Soh HT, Atzberger PJ (2012) Influence of target concentration
and background binding on in vitro selection of affinity reagents. PLoS ONE 7(8):e43940.
17. Chen CK, Kuo TL, Chan PC, Lin LY (2007) Subtractive selex against two heterogeneous
target samples: Numerical simulations and analysis. Computers in Biology and Medicine
37(6):750 -- 759.
18. Cherney LT, Obrecht NM, Krylov SN (2013) Theoretical modeling of masking dna application
in aptamer-facilitated biomarker discovery. Analytical Chemistry 85(8):4157 -- 4164.
19. Seo YJ, Nilsen-Hamilton M, Levine HA (2014) A computational study of alternate selex. Bul-
letin of Mathematical Biology 76(7):1455 -- 1521.
20. Van Kampen NG (1992) Stochastic Processes in Physics and Chemistry.
(North holland)
Vol. 1.
21. Gillespie DT (1976) A general method for numerically simulating the stochastic time evolution
of coupled chemical reactions. Journal of Computational Physics 22(4):403 -- 434.
22. Gillespie DT, Hellander A, Petzold LR (2013) Perspective: Stochastic algorithms for chemical
kinetics. The Journal of Chemical Physics 138(17).
23. Sun F, Galas D, Waterman MS (1996) A mathematical analysis of in vitro molecular selection-
amplification. Journal of molecular biology 258(4):650 -- 660.
24. Chen CK (2007) Complex selex against target mixture: Stochastic computer model, simula-
tion, and analysis. Computer methods and programs in biomedicine 87(3):189 -- 200.
25. Gillespie DT (2000) The chemical
langevin equation. The Journal of Chemical Physics
113(1):297 -- 306.
26. Stoltenburg R, Nikolaus N, Strehlitz B (2012) Capture-selex: Selection of dna aptamers for
aminoglycoside antibiotics. Journal of Analytical Methods in Chemistry 2012.
27. Park JW, Tatavarty R, Kim DW, Jung HT, Gu MB (2012) Immobilization-free screening of
aptamers assisted by graphene oxide. Chemical Communications 48:2071 -- 2073.
28. Nguyen VT, Kwon YS, Kim JH, Gu MB (2014) Multiple go-selex for efficient screening of
flexible aptamers. Chemical Communications 50:10513 -- 10516.
29. Vant-Hull B, Payano-Baez A, Davis RH, Gold L (1998) The mathematics of selex against
complex targets. Journal of molecular biology 278(3):579 -- 597.
30. Katilius E, Flores C, Woodbury NW (2007) Exploring the sequence space of a dna aptamer
using microarrays. Nucleic Acids Research 35(22):7626 -- 7635.
31. Nowak MA, Sigmund K (2004) Evolutionary dynamics of biological games.
Science
303(5659):793 -- 799.
32. Hoinka J et al. (2015) Large scale analysis of the mutational landscape in ht-selex improves
aptamer discovery. Nucleic acids research 43(12):5699 -- 5707.
6 www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX
Spill et al.
|
1907.12609 | 1 | 1907 | 2019-07-29T19:25:42 | Uniform intensity in multifocal microscopy using a spatial light modulator | [
"physics.bio-ph",
"q-bio.QM"
] | Multifocal microscopy (MFM) offers high-speed three-dimensional imaging through the simultaneous image capture from multiple focal planes. Conventional MFM systems use a fabricated grating in the emission path for a single emission wavelength band and one set of focal plane separations. While a Spatial Light Modulator (SLM) can add more flexibility, the relatively small number of pixels in the SLM chip, cross-talk between the pixels, and aberrations in the imaging system can produce non-uniform intensity in the different axially separated image planes. We present an in situ iterative SLM calibration algorithm that overcomes these optical- and hardware-related limitations to deliver near-uniform intensity across all focal planes. Using immobilized gold nanoparticles under darkfield illumination, we demonstrate superior intensity evenness compared to current methods. We also demonstrate applicability across emission wavelengths, axial plane separations, imaging modalities, SLM settings, and different SLM manufacturers. Therefore, our microscope design and algorithms provide an alternative to fabricated gratings in MFM, as they are relatively simple and could find broad applications in the wider research community. | physics.bio-ph | physics | Uniform intensity in multifocal microscopy
using a spatial light modulator
M. JUNAID AMIN,1,2,3,4 SABINE PETRY,1 HAW YANG 2 AND JOSHUA W.
SHAEVITZ3,4,*
1Department of Molecular Biology, Princeton University, Princeton, New Jersey 08544, USA
2Department of Chemistry, Princeton University, Princeton, New Jersey 08544, USA
3Lewis-Sigler Institute for Integrative Genomics, Princeton University, Princeton, New Jersey 08544,
USA
4Department of Physics, Princeton University, Princeton, New Jersey 08544, USA
*[email protected]
Abstract: Multifocal microscopy (MFM) offers high-speed three-dimensional imaging
through the simultaneous image capture from multiple focal planes. Conventional MFM
systems use a fabricated grating in the emission path for a single emission wavelength band
and one set of focal plane separations. While a Spatial Light Modulator (SLM) can add more
flexibility, the relatively small number of pixels in the SLM chip, cross-talk between the pixels,
and aberrations in the imaging system can produce non-uniform intensity in the different axially
separated image planes. We present an in situ iterative SLM calibration algorithm that
overcomes these optical- and hardware-related limitations to deliver near-uniform intensity
across all focal planes. Using immobilized gold nanoparticles under darkfield illumination, we
demonstrate superior intensity evenness compared to current methods. We also demonstrate
applicability across emission wavelengths, axial plane separations, imaging modalities, SLM
settings, and different SLM manufacturers. Therefore, our microscope design and algorithms
provide an alternative to fabricated gratings in MFM, as they are relatively simple and could
find broad applications in the wider research community.
Introduction
1.
Multifocal microscopy is a useful method that allows simultaneous imaging of multiple object
planes to realize high-speed 3D imaging. Common implementations of multifocal microscopy
use a fixed, fabricated phase mask optimized for a specific wavelength or object-plane
separation distance [1-3]. The fabricated grating can be replaced with a Liquid Crystal Spatial
Light Modulator (SLM) that can dynamically change the phase mask for use with multiple
wavelengths or object plane separations [4-5]. However, uniform illumination across the
subimages is difficult to achieve using an SLM-based phase mask due to inherent device
characteristics including pixel-to-pixel crosstalk effects [6-7]. We present an in situ iterative
calibration method for the generation of optimized SLM phase patterns that produce multifocal
images with near-uniform subimage brightness.
2. Prior-art method for obtaining uniform subimage brightness in multifocal
microscopes
The phase grating in a multifocal microscope has two tasks: (i) to divide incoming emission
light equally into a 2D array of orders, and (ii) to axially offset the orders to different object
planes separated by a distance ∆z by modifying the phase pattern using a geometric distortion
function. The Pixelflipper algorithm was designed to generate uniformly-illuminated
subimages in existing multifocal microscope systems [2]. This algorithm uses an in situ
iterative procedure that finds the phase pattern of a grating unit cell of size Pu × Pu pixels2 that
gives the highest uniformity among the diffraction orders in the computed Fourier plane. This
optimized unit cell is then repetitively arranged into a grid to provide the phase pattern to be
displayed on the SLM. Pixelflipper often fails to produce adequate results using SLMs as it
assumes an aberration free optical system that takes the Fourier transform precisely.
Furthermore, in addition to other system aberrations, SLMs suffer from pixel-to-pixel crosstalk
effects [6-7] that further alter the resultant diffraction pattern. The same issue occurs when
iterative Fourier transform algorithms [8] are used, particularly when only few SLM pixels
form the repeated grating pattern. These effects are illustrated in the image of a 100 nm Gold
Nanoparticles (AuNPs) under darkfield illumination on an SLM-based multifocal microscope
(see Appendix) using ∆z = 0 nm (Fig. 1(a)). There is an undesirably significant difference in
subimage intensities in the image even though the Pixelflipper algorithm was used to optimize
the SLM pattern with Pu = 4 (Fig. 1(b)).
M
=
b
(a) (b)
Fig. 1. (a) In-focus sub-images cropped from a 9 plane z stack acquired from the camera resulting from deploying a
Pixelflipper algorithm [11] optimized SLM display pattern. The emission light is unevenly distributed in the subimages,
with the central zeroth order subimage receiving most of the emission light, and (b) zoomed-in view of a 64 × 64
pixels2 region of the SLM displayed grating pattern which gives the multifocal subimages in (a).
3.
Here, we propose an in situ iterative calibration method to generate phase patterns which allows
near-uniform illumination in the subimages of an SLM-based multifocal microscope. The
algorithm is based on a feedback loop between the SLM, camera, and the computer, and uses
real-time images from the camera to update the pattern on the SLM until the optimal grating
pattern is acquired. To evaluate the subimage intensity uniformity for our optimization routine,
we modified the metric proposed in ref. [9] and used instead,
In situ iterative calibration routine for subimage intensity uniformity
{
(
min
I
,
im
{
(
max
I
,
im
}
)
}
)
-
-
I
b
I
where !𝐼#,%& is the measured subimage i (i = 1 … N×N) and 𝐼' is a measured background
, (1)
intensity. M ranges from 0 to 1, with M=1 corresponding to completely uniform subimage
intensities. We initially generate 100 random unit cells and start with the one that has the highest
M value. The unit cell is then updated iteratively by sequentially iterating over all graylevel
values for all pixels in the unit cell. This routine is repeated until there is no change in the unit
cell during a complete iteration over all pixel locations.
This calibration method gives visually uniform subimage intensities using ∆z = 0 nm (Fig. 2(a)),
when displaying an in situ iterative optimized pattern on the SLM (Fig. 2(b)). The computed M
value for this pattern is 0.712, much larger than the M value measured when using the
Pixelflipper-based phase mask (M = 0.033, Fig. 1(a)). We compared multiple trials of the
Pixelflipper algorithm, our in situ iterative method, and randomly-generated phase patterns
using Pu = 4 (Fig. 3). The in situ iterative method shows superior performance over both the
Pixelflipper and the randomized methods, realizing multifocal images having large M values,
i.e., near uniform subimage intensities (Fig. 3).
(a) (b)
Fig. 2. (a) In-focus sub-images cropped from a 9 plane z stack acquired resulting from deploying our in situ iteratively
optimized SLM display pattern. The emission light striking the camera is more evenly distributed among the subimages
compared to Fig. 1(a), and (b) zoomed-in view of a 64 × 64 pixels2 region of the SLM displayed grating pattern which
gives the images in (a).
Fig. 3. Boxplots of output M values resulting from grating patterns optimized using Pixflipper (500 iterations,
Randomized (500 iterations) and our in situ iterative calibration algorithm (12 iterations).
4. Multifocal Imaging of Biological Specimen
We used our in situ iteratively optimized SLM phase patterns to acquire 3D images of GFP-
labeled tubulin in MeOH-fixed TPX2 Hela Kyoto cells [10]. The in situ iterative calibration
algorithm was first executed using darkfield imaging under the current settings to obtain the
optimized SLM phase pattern, which was then phase distorted using the algorithm in [2] to
achieve object plane separation in the subimages. Multifocal snapshots of the sample cells using
488-nm laser excitation are obtained (Fig. 4). The sequence of the subimages in the multifocal
image (Fig. 4) follows the sequence shown in Appendix Fig. A1(b), with the top right plane of
the image corresponding to the z = -4∆z plane, where ∆z represents the focal plane separation
in object space. The ∆z values for the images (Fig. 4) are 0.50 µm (Fig. 4(a)) and 1.00 µm (Fig.
4(b)). The different focal cross-sections of the cells can be visibly seen across the subimages,
(a) (b)
Fig. 4. Multifocal Images of different regions of a microtubule stained Hela Kyoto Cells fixed sample with (a) ∆z =
0.50 µm, and (b) ∆z = 1.00 µm.
displaying the structural features in the sample and the 3D imaging power of the multifocal
microscope.
5. Discussion
We demonstrate that our in situ iterative optimization algorithm is effective at generating SLM-
based patterns which allow uniformity across multifocal subimages. The key underlying
concept of the algorithm is the inclusion of real-time experimental variables into the
optimization process. Many of the parameters involved such as Pu, the emission wavelength
and the SLM model are empirical hardware choices. In particular, the Pu = 4 value throughout
the paper is empirically chosen to maximize the imaging field of view without overlapping of
the subimages in the current system optical settings. Other microscope setups may require a
different unit cell size to be optimal.
The method allows comparable MFM imaging performance when using SLMs versus
fabricated gratings. Deploying SLMs as multifocal gratings has numerous advantages: they are
available off the shelf, require no additional investment other than its initial cost, and can
readily be programmed to change any multifocal grating parameter including ∆z values as well
as number of simultaneous imaging planes at high speed, limited by the refresh rate of the SLM
(typically 60 Hz). Fabricated gratings, on the other hand, require access to clean room facilities
having fabrication and lithography tools which require extensive training to use. This process
is expensive, time-consuming and not readily available near many research labs. Using our in
situ iterative algorithm, more researchers can now build SLM based MFMs for investigating
fast microscopic 3D processes in biology, physical chemistry and other domains. The universal
applicability of the calibration routine is demonstrated to account for different SLM
manufacturers, wavelength, unit cell sizes as well as different microscope modalities (see
Appendix), making this method widely applicable to the different imaging requirements. Future
work involves exploring various other optimization techniques and metrics to further improve
subimage intensity uniformity in the SLM based MFM.
6. Appendix
6.1. SLM based Multifocal Microscope (SLM-MF)
6.1.1. Optical Diagram
Fig. A1(a) illustrates the optical design of the home-built multifocal microscope. Three main
segments of the system are highlighted for easier visualization: Darkfield-Brightfield (DF-BF)
illumination, laser illumination and multifocal Imaging. In the DF-BF illumination
(a) (b)
Fig. A1. (a) Optical diagram of the SLM based Multifocal Microscope, and (b) front view of the camera sensor showing
the placement of the simultaneously sub-images, each corresponding to a unique object plane.
module, incoherent light from a white LED passes through a Diffuser before being focused by
a lens L1 onto an iris. Lens L2 then collimates the light and sends it onto the DF Mirror to
realize Darkfield Imaging mode. The DF Mirror is a custom designed mirror with an oval
central transmissive area thereby allowing reflection of the incoming light into a ring-shaped
intensity pattern downwards into the condenser objective. To switch from Darkfield mode to
Brightfield imaging mode, the DF Mirror can be replaced by the BF Mirror, where the BF
Mirror is a conventional fully reflecting optical mirror. Note that when operating in Darkfield
mode, the Detection Objective's Numerical Aperture (NA) is kept smaller than that of the
Condenser Objective.
In the laser illumination module, a laser is spatially filtered and collimated by a combination
of lens L3, a Pinhole and lens L4. Lens L5 is introduced to focus the beam reflecting off a
dichroic to the back focal plane (BFP) of the detection objective for epi-illumination. This beam
will be used for excitation of fluorescent samples. Note that although fluorescence samples can
also be used for the experimental calibration routine, they aren't optimal due to bleaching
induce changing intensities in the images relative to more stable darkfield and brightfield
imaging modes.
In the multifocal imaging module, lens L6 functions as a tube lens and forms an image of
the sample at the plane of the rectangular aperture through a linear polarizer. The rectangular
aperture's function is to control the imaging Field of View (FOV), thereby preventing the
obtained subimages from overlapping at the imaging plane downstream. The polarizer is
necessary for the phase-only function of the deployed SLM. Lens L7, having a focal length f7,
is placed a focal length's distance from the rectangular aperture, while the reflective SLM is
located at the Fourier plane of L7. With the focal lengths of L6 and L7 the same, the SLM is
essentially conjugate to the BFP of the detection objective which is also the Fourier plane of
the sample.
The SLM acts as a multifocal grating in this setup. Each of the diffraction orders emanating
from the SLM displayed multifocus grating pattern is encoded with a unique defocus phase
(except the unaffected zeroth order). These orders when imaged onto the camera sensor, by
Lens L8 (having focal length f8) via an emission filter, form a multifocus image. This image
contains N x N subimages with each subimage represents a unique object plane. Details of the
algorithm used to generate such grating patterns are described later in the paper. The bandwidth
of the emission filter is restricted to around ~15 nm to reduce chromatic dispersion effects
originating from the SLM displayed grating. This emission wavelength band restriction can be
overcome using chromatic correction optics as used in [2], but is not implemented here. Fig.
A1(b) illustration shows the front view of the camera in the multifocus microscope designed
for a 3 x 3 array of subimages, though in theory any N x N number of orders can be obtained.
Different segments of the camera in Fig. A1(b) correspond to different object planes, with each
subimage separated by a distance ∆z in object space. Note that the zeroth order remains
undiffracted and unaffected by the SLM displayed pattern, and thus corresponds to the z = 0
plane. A Personal Computer (PC) running LabVIEW Software interfaces with the SLM and the
camera sensor.
For the SLM-MF, it useful to know the effective lateral FOV as a function of different
microscope parameters involved. FOV in this paper is defined as the field of view for each
subimage in sample space. To find the expression for FOV, denote Mag as the combined
magnification of the detection objective and lens system, Pu the SLM displayed grating period
in pixels units, S the SLM pixel size, the C camera pixel size and λmin the minimum wavelength
in the emission band. The angle θ between the zeroth and 1st orders of a grating is found using
the grating equation: θ = sin-1[λmin/(Pu x S)]. Once θ is known, the FOV can be equated by
finding the distance between the centers of both zeroth and 1st orders on the image plane, before
dividing by the Mag:
FOV
=
( )qtan
8f
, (2)
1
Mag
ae
çç
è
ö
÷÷
ø
6.1.2. Experimental Setup
The SLM-MF microscope is custom built in the lab to test the effectiveness of grating patterns
designed to optimally distribute incoming light equally into the diffraction orders. For the
Darkfield imaging mode, Thorlabs Solis-3C High-Power LED is deployed as the white light
LED, along with the accompanying DC20 driver module for intensity control. The diffuser
used is Thorlabs DG20-1500. Lens L1 is Thorlabs LA1401-A (focal length = 60 mm) and L2
is AC508-150-A-ML (focal length 150 mm). The DF mirror is custom designed to match the
dimensions of the condenser objective used which is the MPLAN BD 50x NA 0.75 objective
from Olympus. The detection objective is a Leica 100x, NA 1.4 - 0.7, where the NA is set to
0.7 during Darkfield imaging. For the sample, a mixture is formed using 10 uL of stock 100
nm Gold nanoparticles (AuNPs) solution from BBI solutions and 10 uL of 1M NACl solution.
10uL is ejected onto a 22 × 22 mm2 coverslip (Fisher scientific) using a pipette before being
covered directly by an 18 × 18 mm2 coverslip. The salt is added to immobilize the AuNP onto
the coverslip surface. Nail Polish from Electron Microscopy Sciences is used to seal the
coverslip edges to avoid leakage of the solution. The sample is mounted onto a P-611.3S
NanoCube XYZ Piezo stage from Physik Instrumente using custom machined mounts. Both
the piezo stage and the detection objective are mounted on a custom designed aluminum block,
which forms the microscopy body. The condenser objective is screwed to a Newport 460A-
XYZ translation stage via a custom aluminum adapter plate, with this stage mounted on an 8
inch high post to position the condenser above the sample. Lens L6 is Thorlabs AC508-200-A-
ML (focal length = 200 mm), while both f7 and f8 are set to 200 mm (Thorlabs AC254-200-A-
ML). The polarizer model is LPVISE100-A, the rectangular aperture is model # 61-1137 from
Ealing Catalog, USA, while the emission filter deployed is Semrock FF01-685/10-25 with a
central wavelength of 685 nm and FWHM ~ 15 nm. For the biological experiment, the emission
filter is changed to one with a 510 nm center wavelength having a bandwidth of ~16 nm. The
camera sensor is Hamamatsu's Orca-Flash4.0 V3 sCMOS with a pixel resolution of 2048 x
2048 pixels and a pixel size of 6.5 μm. The camera exposure time is set to 250 ms for the
duration of the experiment, and all images are stored as raw 16-bit ".tif" format.
The SLM deployed is the reflective Holoeye PLUTO-2-VIS-056 Phase-only spatial light
modulator. It has an 8-bit pixel display resolution, a 1920 x 1080 array of pixels and a pixel
pitch S = 8 μm. It can be calibrated, i.e., mapping of displayed gray-level images to actual phase
imparted to the incoming light, using different methods. One method is to use the manufacturer
provided calibration data specified for certain wavelengths which can be loaded directly onto
the SLM via its USB port. Another method involves an experimental procedure outlined in the
manufacturer's manual which involves striking the two halves of the SLM with circular top-
hat beams originating from the same laser and observing the resulting interference pattern. This
latter method is not easily applicable in multifocal setups since the SLM is used in the emission
path with emission wavelengths over the visible range. Acquiring lasers for each desired
emission wavelength is expensive and impractical, with no guarantee of being effective for a
spread of wavelengths as is the case here. Therefore, for this experiment, a default
manufacturer-provided calibration curve meant for a 2.2π phase cycle corresponding to 0 -- 255
graylevel values for 633 nm laser is uploaded to the SLM firmware. This gives an
approximately 2π phase cycle for 696 nm, closely matching the current emission wavelength
centered at 685 nm. Apart from the calibration, an important step is to ensure that all SLM
displayed patterns have an aperture similar to that of the BFP of the detection objective. Thus,
all displayed patterns are multiplied by an aperture function with the central region size
corresponding to the size of the BFP. Furthermore, the region outside this aperture is set to a
tilted grating with an empirically chosen defocus pattern to steer any stray incoming emission
light striking outside the SLM main aperture area away from the zeroth order. Therefore, any
subsequent grating patterns will only be displayed inside the central region of the aperture
function.
With the experimental setup arranged as described, we implement the prior-art Pixelflipper
algorithm described in Section 3.1 using LabVIEW software. N is chosen to be equal to 3,
giving a target matrix T of size 3 × 3. With 256 SLM displayable graylevels at our disposal,
we limited the graylevel resolution to 80 steps spanning the 0 -- 255 range to reduce the algorithm
run time which empirically provides similar performance to having 256 graylevel steps. The
Pixelflipper algorithm is run by setting Pu = 4 to give an optimized matrix unit cell U of
dimensions 4 × 4. This U is then arranged in a grating format and is phase 'distorted' to give a
∆z value of 0.90 µm to realize multifocus imaging, before being multiplied by the aperture
function described earlier. Once this resulting pattern is displayed on the SLM, a 9 plane z
stack (i.e., images of the sample obtained at multiple z positions by vertical motion of the piezo
stage) of the sample is obtained, where the axial spacing of the z stack is chosen to match the
deployed ∆z = 0.90 µm.
processing only the plane of the z stack when subimage 1 is in focus. Whereas, the mean
To find the value of M due to a given grating pattern using Eqn. (1), the following procedure
is followed: a particle of interest in the field of view is selected and an 80 × 80 pixels2 area
around that particle is chosen as the Region of Interest (RoI). For each subimage i, where i is
an integer between 1 and 9, the mean intensity 𝐼#,% of its respective RoI is calculated from the
image when it is in focus. For example, the mean intensity of subimage 1, 𝐼#,), is found by
intensity of, e.g., subimage 7, 𝐼#,*, is found by processing only that plane of the z stack when
subimage 7 is in focus. Once the mean intensities !𝐼#,%&of all subimages are found, the
minimum and maximum !𝐼#,%& values are selected for use in Eqn. (1). To calculate 𝐼', a uniform
than subimage 5 are calculated and denoted as !𝐼',+& where j is an integer between 1 and 9,
other than 5. 𝐼' is then defined as the minimum value among 𝐼',+. In this step, note that the
minimum of 𝐼',+ is chosen as 𝐼', and not the average of 𝐼',+, to avoid negative M values which
can occur when min(!𝐼#,%&) -- 𝐼' is negative where 𝐼' is
graylevel pattern is displayed on the SLM which results in the SLM not directing light into the
subimages, except into subimage 5 which is the zeroth order and receives most of the incoming
light. In this setting, the mean intensities calculated over the same ROIs for all subimages other
brighter than min(!𝐼#,%&). Therefore, with min(!𝐼#,%&), max(!𝐼#,%&) and 𝐼' at hand, Eqn. (1) is
used to compute M.
6.2. SLM-MF subimage field flatness
The SLM-MF simultaneously images multiple object planes into, e.g., 3 × 3 array of
subimages. In addition to optimizing the intensity distribution among the subimages, it is also
beneficial to characterize the field flatness of the subimages as a result of the phase distortion
implemented in the SLM grating pattern to achieve the 3D imaging capability. An
understanding of the uniformity across each subimage is a necessary step in interpreting and
processing multifocal 3D imaging data acquired from this microscope.
The following procedure is deployed to characterize the field flatness across the subimages
in Darkfield imaging mode: A solution of 200 nm fluorescent beads (660/680) from Life
Technologies Corporation is diluted an empirically chosen 400 times before being mixed with
a solution of 1 Molar NaCl in a 1:1 volume ratio. 10 µL of this mixture is ejected on a coverslip
(22 × 22 mm2 coverslip from Fisher scientific) using a pipette before being covered with
another coverslip (18 × 18 mm2 coverslip (Fisher scientific)). The edges of the smaller coverslip
are sealed with nail polish (Electron Microscopy Sciences). As before, the salt helps to
immobilize the fluorescent beads to the coverslip surface. The sample is mounted on the piezo
stage and a 647 nm Cobolt 130 mW CW laser is used for illumination in an epi-configuration.
On the emission side, a ~3nm emission window is created by inserting both Semrock filters
FF01-685/10-25 and FF01-685/LP-25 into the light path. The ~3 nm bandwidth significantly
minimizes the chromatic dispersion in the subimages (other than the zeroth order subimage 5
which is unaffected). This prevents dispersion related aberrations from negatively affecting the
bead localization process described shortly. For illustration,
(a) (b)
Fig. A2. (a) Images of 100 nm immobilized AuNPs acquired under Darkfield Illumination using emission filter
bandwidths of (a) ~ 15 nm, and (b) ~ 3 nm.
multifocal images acquired using ~3 nm and ~15 nm emission filter bandwidths is shown in
Fig. A2. Z-stacks are acquired using an in situ iteratively optimized pattern displayed on the
SLM having ∆z values of 0 µm and 1.00 µm, using stage stapes of 20 nm. For each z-stack, the
locations of the beads within the field of view are identified in the lateral and axial Cartesian
coordinates. The lateral (xy) positions are identified by identifying the bright regions in a focal
projection of the z-stacks using the imfindcircles() function in MATLAB. Appropriate radii and
intensity thresholds are applied to remove possible bead aggregates. The axial (z) location of
the beads are identified by computing the maximum of the Brenner gradient in a square 16 ×
16 pixels2 region around each identified bead for all images in the z-stack. This localizes the
beads in the axial direction.
Once the lateral and axial positions of the beads are found for each subimage, they are
plotted in 3D and the xz views are displayed in Fig. A3. Fig. A3(a) plot shows displays field
flatness data acquired using ∆z = 0 µm, whereas Fig. A3(b) corresponds to data acquired when
∆z = 1.00 µm. These plots demonstrate near uniform fields across all subimages. Even without
correcting for possible sample tilt inherent to the setup, the average peak-valley (P-V) value
among subimages acquired using ∆z = 0 µm is 0.665 µm while the mean P-V value for ∆z =
1.00 µm is 0.467 µm, both well within 2% variation across the field of view signifying
reasonably flat subimages most practical multifocal imaging purposes.
(a) (b)
Fig. A3. xz view of the xyz localization of 200 nm beads immobilized on a coverslip to demonstrate the field flatness
of the 9 subimages for ∆z values of (a) 0 µm, and (b) 1.00 µm.
6.3. Uniform illumination of orders using Brightfield Imaging mode
To engage the Brightfield imaging mode, the DF Mirror in Fig. A1(a) is replaced by the BF
Mirror to allow full reflection of the light incoming from L2. In this arrangement, we first
remove the sample completely, and allow the unscattered light focused by the condenser passes
through the detection objective and towards the multifocal optics. Another possibility is to have
a cut-out piece of A4 paper as the sample to act as a scattering object. Both methods are tested
to work adequately. Apart from the Imaging mode, other key parameter changes for this
demonstration includes setting ∆z = 0 µm. Since ∆z = 0, there is no need to take z-stacks for M
value calculation, therefore a single image is obtained for each pattern and the sub-images are
processed from it. In addition, M values calculated using the Brightfield imaging mode are
denoted MBF, to discriminate from M calculated in Darkfield imaging mode. Another parameter
change involves setting f8 to 100 mm, which has no effect on the order illumination distribution
characteristics but changes the magnification Mag by half.
This arrangement is first tested for a pattern generated by the Pixelflipper using the
unchanged Pu = 4 and G = 80. An image acquired using this Pixelflipper optimized pattern is
shown in Fig. A4(a). A purple box is shown to annotate the camera region covered spanning
the 3 × 3 orders. Fig. A4(b) shows 64 × 64 pixels2 region in the SLM displayed Pixelflipper
optimized patter. Visually, the illumination spread among the orders, similar to Fig. 1(a) in
terms of contrast, is far from uniform. The computed MBF value for the Fig. A4(a) image is
0.126.
(a) (b)
Fig. A4. Using the Brightfield Imaging mode, (a) image resulting from deploying a Pixflipper optimized SLM
displayed pattern. The pattern is intended to give a 3 x3 array of uniformly illuminated subimages. The MBF value for
(a) is 0.126, and (b) zoomed-in view of the SLM displayed grating pattern which gives the image in (a), showing the
repetitive arrangement of unit cells.
Fig. A5(a) shows an output image due to an in situ iteratively optimized pattern, with the purple
box annotating the relevant 3 × 3 sub-images region. A 64 × 64 region of the pattern is shown
in Fig. A5(b). Visually, Fig. A5(a) shows a higher uniformity of illumination across the orders.
The MBF value for Fig. A5(a) is computed to be 0.712. In terms of statistics, the Pixelflipper
and randomized pattern generation algorithm are repeatedly executed 1000 times each, and MBF
values are computed for each pattern. These MBF values are displayed as a boxplot in Fig. A6
which also shows MBF values from 30 iterations of our algorithm. This plot demonstrates the
high degree of illumination uniformity improvement due to our algorithm.
Note that, in Fig. A5(a), additional subimages apart from the bright 3 × 3 subimage array
are also visible. These are additional orders which receive illumination from the grating pattern.
In the current algorithm framework, only the intensities of the 3 × 3 subimages are optimized,
with no correction for the intensity spilling out into the other orders. In future work, new
optimization will be explored which allow high efficiency illumination of the subimages while
suppressing the unused diffraction orders' intensity to a minimum.
(a) (b)
Fig. A5. (a) Image resulting from deploying an output optimized pattern from the proposed in situ iterative algorithm.
The pattern is intended to give a 3 x3 array of uniformly illuminated subimages. The M value for (a) is 0.712, and (b)
zoomed-in view of SLM displayed grating pattern which gives the image in (a), showing the repetitive arrangement of
unit cells.
Fig. A6. (a) Boxplots of output MBF values resulting from grating patterns optimized using Pixflipper (1000 iterations,
Randomized (1000 iterations) and our in situ iterative algorithm (30 iterations).
6.4. Comparison of in situ iterative calibration routine in Darkfield imaging mode
versus Brightfield imaging mode
To evaluate the performance of Pu = 4 patterns optimized using the in situ iterative
calibration method in Brightfield mode, denoted iterative brightfield, compared to the Pu = 4 in
situ iteratively optimized patterns in Darkfield imaging mode, denoted as iterative darkfield,
the 30 brightfield optimized patterns are implemented in Darkfield imaging of the same sample
used in obtaining the M value data for Fig. 3 (main text). In this demonstration, all the
Brightfield patterns are distorted with ∆z = 0.90 µm. The computed M values resulting from
using the Brightfield patterns (30 iterations) in this Darkfield imaging mode are compiled into
the Fig. A7 boxplot, which also shows the Fig. 3 (main text) M value data (12 iterations)
obtained using our algorithm. This plot indicates that both methods are equally effective in
optimizing illumination uniformity across multifocal images.
Fig. A7. The boxplot shows a comparison of M values due to 12 different in situ iteratively optimized Darkfield patterns
(optimized on AuNP samples under Darkfield imaging) and 30 optimized Brightfield patterns (optimized using
Brightfield illumination with no sample) when implemented on the same AuNP sample with ∆z = 0.90 µm.
6.5. Pixelflipper output using Pu > 4
Our calibration method is demonstrated in this paper to significantly improve the intensity
distribution in the multi-focus subimages. Prior to this method, the Pixelflipper has been widely
applied to fabricated gratings, though with orders of magnitude larger Pu values. Due to the
large pixel sizes of SLMs, larger Pu values equate to large grating periods, which in turn
significantly limit the field of view. Selected Pixelflipper algorithm outputs for Pu values of 4,
16 and 32 are shown in Fig. A8, which is demonstrated using the Brightfield imaging mode
and using f8 = 200 mm. Additionally, G is set to 256 for all generated Pixelflipper patterns in
this demonstration. Fig. A8(a) shows a camera image captured using Pu = 4 using the
Pixelflipper optimized pattern whose zoomed in 64 × 64 pixels2 region is shown in Fig. A8(b).
The FOV for Pu = 4 is calculated to be 42.38 µm using f8 = 200 mm, Mag = 100, λmin = 678 nm
and S = 8 µm. Fig. A8(c) shows a camera image captured using Pu = 16, with the zoomed in 64
× 64 pixels2 region of the corresponding Pixelflipper optimized pattern shown in Fig. A8(d). In
this Pu setting of 16 which realizes larger grating periods displayed on the SLM, the diffraction
angle is decreased and the FOV decreases by a factor of 4 to 10.60 µm. In between Fig. A8(a)
and Fig. A8(c), the Rectangular Aperture is adjusted to prevent overlap between the subimages
on the camera sensor resulting from the Pu increase. Next, Pu is set to 32 and a Pixelflipper
optimized pattern is obtained. Fig. A8(e) shows the resulting camera image captured, while Fig.
A8(f) shows the zoomed in 64 × 64 pixels2 region of the Pixelflipper optimized pattern. As
before, the Rectangular Aperture is adjusted to prevent overlap between the subimages on the
camera sensor resulting from the Pu increase from 16 to 32. The FOV in this Pu = 32 setting is
now 5.30 µm. Qualitatively, according to Fig. A8, the intensity uniformity among the 9
subimages does improve by increasing Pu from 4 to 32. However, this comes at a high cost of
eightfold decrease in the FOV, making Pu = 32. Furthermore, Pu > 32 values are needed to
achieve better uniformity, at the cost of further reduction of the FOV. This shows the power of
our proposed algorithm which allows high illumination even using Pu = 4, without
compromising on the FOV.
(a) (b) (c)
(d) (e) (f)
Fig. A8. In Brightfield Imaging mode, the plots show a comparison of chosen Pixflipper algorithm output patterns
for each Pu = 4, Pu = 16 and Pu = 32. (a) The camera image resulting from a Pu = 4 Pixflipper optimized pattern, (b) a
zoomed in 64x64 pixel2 region of the Pu = 4 resulting pattern displayed on the SLM, (c) the camera image resulting
from a Pu = 16 pixflipper optimized pattern, (d) a zoomed in 64x64 pixel2 region of the Pu = 16 resulting pattern
displayed on the SLM, (e) the camera image resulting from a Pu = 32 pixflipper optimized pattern, and (f) a zoomed in
64x64 pixel2 region of the Pu = 32 resulting optimized pattern displayed on the SLM.
6.6. In situ iterative calibration routine implementation on a different SLM
To demonstrate the universality of our algorithm across SLMs, the Holoeye Pluto-VIS-056
SLM is replaced with a Hamamatsu X10468-07 LCOS-SLM. Pu = 4 is used. The X10468 has
a larger pixel pitch of 20 µm and has a pixel resolution of 800 × 600 pixels. The Brightfield
imaging mode is deployed, and the same algorithm parameters, including Pu = 4, is used for
this demonstration with the results summarized in Fig. A9. Fig. A9(a) is an example output
image due to a Pixelflipper optimized SLM pattern. Note that due to the 20 µm pixel pitch value
of this SLM versus the 8 µm of the SLM used earlier, the diffraction angle is smaller bringing
the orders closer together on the imaging sensor; therefore the Rectangular Aperture in the
optical path is adjusted to prevent the FOVs of the subimages from overlapping with each other.
Fig. A9(b) is an output image due to our algorithm optimized pattern, showing a clear increase
in the illumination uniformity across the central 3 × 3 orders, in comparison to Fig. A9(a). Fig.
A9(c) shows boxplots of MBF values due to the Pixelflipper (90 iterations) and our in situ
iterative calibration routine (40 iterations) algorithms, demonstrating the superior performance
and applicability of the algorithm. Fig. A9 illustrates the effectiveness of our routine in
overcoming the hardware related issues of the Hamamatsu SLM to provide near-uniform
intensity spread across the multifocal subimages.
(a) (b) (c)
Fig. A9. Optimization results after replacing the Holoeye Pluto-VIS-056 SLM with the Hamamatsu X10468-07
SLM using Pu = 4. (a) output image due to a chosen Pixelflipper algorithm optimized SLM pattern, (b) output image
due to our algorithm optimized SLM pattern, and (c) boxplots of M values resulting from 90 iterations of the
Pixelflipper and 40 iterations of the our algorithm output. The plot in (c) demonstrates the superior illumination
intensity distribution performance of our in situ iterative algorithm.
6.7. Wavelength dependence of in situ iteratively optimized patterns
The calibration routine patterns optimized for a specific wavelength band are empirically found
to give different intensity distributions in the subimages at a different wavelength band. This is
demonstrated in Brightfield imaging mode and illustrated in Fig. A10. Fig. A10(a) shows the
image resulting from a 685 nm centered bandpass filter with a pattern optimized for this
wavelength using our method. When the emission filter is changed to be centered at 510 nm
with a bandwidth of 15 nm, the resulting image acquired is shown in Fig. A10(b) using the
same pattern as used for Fig. A10(a). Fig. A10(b) shows an undesirable intensity distribution
among the subimages. Note that whenever the emission filter is changed, the Rectangular
Aperture is adjusted to prevent the FOVs of the subimages from overlapping. For SLM
operations, it is recommended to use an updated calibration curve when switching to different
wavelengths, therefore as a next step, the calibration settings are updated to give a 2π phase
range for 532 nm, which is close to 510 nm for the purpose of this demonstration. In this
updated setting, the same pattern which is optimized for a 685 nm bandpass filter is displayed
on the SLM and the resulting image acquired and shown in Fig. A10(c), still showcasing a far
from ideal intensity spread. Finally, the algorithm is executed using the 532
(a) (b)
(c) (d)
Fig. A10. Demonstration of wavelength dependence of the iterative optimization algorithm. (a) image resulting
from optimized for 685 nm centered emission filter, acquired using 685 nm centered emission filter, with SLM
calibration settings suited to 685 nm, (b) image resulting from our algorithm optimized for 685 nm centered emission
filter, acquired using 510 nm centered emission filter, with SLM calibration settings suited to 685 nm, (c) image
resulting from optimized for 685 nm centered emission filter, acquired using 510 nm centered emission filter, with SLM
calibration settings suited to 510 nm, and (d) image resulting from in situ iteratively optimized for 510 nm centered
emission filter, acquired using 510 nm centered emission filter, with SLM calibration settings suited to 510 nm.
nm based updated SLM calibration settings and deploying the 510 nm centered emission filter,
and the image resulting from this optimized output pattern is shown in Fig. A10(d) which
represents a much more uniform intensity distribution as compared to Fig. A10(b) and Fig.
A10(c). It is recommended to deploy our optimization routine to separately acquire ideal SLM
patterns for each wavelength band desired in the multifocal microscope.
Funding
This work was supported through an Eric and Wendy Schmidt Transformative Technology
Fund award to H.Y., S.P., and J.W.S. and by the National Science Foundation, through the
Center for the Physics of Biological Function (PHY-1734030).
Acknowledgements
We would like to thank Matthew King (Princeton University) for preparation of the cell
samples.
References
1. P. M. Blanchard and A. H. Greenaway, "Simultaneous multiplane imaging with a distorted
diffraction grating," Applied Opt. 38(32), 6692-6699 (1999).
2. S. Abrahamsson, J. Chen, B. Hajj, S. Stallinga, A. Y. Katsov, J. Wisniewski, G. Mizuguchi,
P. Soule, F. Mueller, C. D. Darzacq, X. Darzacq, C. Wu, C. I. Bargmann, D. A. Agard, M.
Dahan and M. G. L. Gustafsson, "Fast multicolor 3D imaging using aberration-corrected
multifocus microscopy," Nat. Methods 10, 60 -- 63 (2013).
3. S. Abrahamsson, M. McQuilken, S. B. Mehta, A. Verma, J. Larsch, R. Ilic, R. Heintzmann,
C. I. Bargmann, A. S. Gladfelter and R. Oldenbourg, "MultiFocus polarization microscope
for 3D polarization imaging of up to 25 focal planes simultaneously," Opt. Express 23(6),
7734 -- 7754 (2015).
4. A. Jesacher, C. Roider and M. Ritsch-Marte, "Enhancing diffractive multi-plane
microscopy using colored illumination," Opt. Express 21(9), 11150-11161 (2013).
5. Q. Ma, B. Khademhosseinieh, E. Huang, H. Qian, M. A. Bakowski, E. R. Troemel and Z.
Liu, "Three-dimensional fluorescent microscopy via simultaneous illumination and
detection at multiple planes," Nat. Sci. Rep. 6(31445), (2016).
6. E. Hällstig , J. Stigwall , T. Martin , L. Sjöqvist and M. Lindgren, "Fringing fields in a
liquid crystal spatial light modulator for beam steering," J. of Modern Opt. 51(8), 1233-
1247 (2004).
7. C. Lingel, T. Haist and W. Osten, "Examination and Optimizing of a Liquid Crystal
Display used as Spatial Light Modulator concerning the Fringing Field Effect," Proc. of
SPIE 8490, (2012).
8. O. Ripoll, V. Kettunen and H. P. Herzig, "Review of iterative Fourier-transform algorithms
for beam shaping applications," SPIE Opt. Eng. 43(11), (2004).
9. M. Persson, D. Engström and M. Goksör, "Reducing the effect of pixel crosstalk in phase
only spatial light modulators," Opt. Express 20(20), 22334-22343 (2012).
10. B. Neumann et al., "Phenotypic profiling of the human genome by time-lapse microscopy
reveals cell division genes," Nature 464, 721 -- 727 (2010).
|
1304.5572 | 2 | 1304 | 2013-05-06T23:28:12 | First Passage Properties of Molecular Spiders | [
"physics.bio-ph"
] | Molecular spiders are synthetic catalytic DNA-based nanoscale walkers. We study the mean first passage time for abstract models of spiders moving on a finite two-dimensional lattice with various boundary conditions, and compare it with the mean first passage time of spiders moving on a one-dimensional track. We evaluate by how much the slowdown on newly visited sites, owing to catalysis, can improve the mean first passage time of spiders and show that in one dimension, when both ends of the track are an absorbing boundary, the performance gain is lower than in two dimensions, when the absorbing boundary is a circle; this persists even when the absorbing boundary is a single site. | physics.bio-ph | physics |
First Passage Properties of Molecular Spiders
Oleg Semenov,∗ David Mohr,† and Darko Stefanovic‡
Department of Computer Science, University of New Mexico, MSC01 1130,
1 University of New Mexico, Albuquerque, NM 87131-0001 and
Center for Biomedical Engineering, University of New Mexico,
MSC01 1141, 1 University of New Mexico, Albuquerque, NM 87131-0001
Molecular spiders are synthetic catalytic DNA-based nanoscale walkers. We study the mean first
passage time for abstract models of spiders moving on a finite two-dimensional lattice with various
boundary conditions, and compare it with the mean first passage time of spiders moving on a one-
dimensional track. We evaluate by how much the slowdown on newly visited sites, owing to catalysis,
can improve the mean first passage time of spiders and show that in one dimension, when both ends
of the track are an absorbing boundary, the performance gain is lower than in two dimensions, when
the absorbing boundary is a circle; this persists even when the absorbing boundary is a single site.
I.
INTRODUCTION
Bound leg
Substrate
(DNA)
Natural molecular motors play an important role in bi-
ological processes that are critical for the functioning of
living organisms; they are the source of most forms of mo-
tion in living beings [1 -- 3]. In addition to naturally occur-
ring molecular motors, several synthetic molecular mo-
tors have been designed [4 -- 12]. Our work is inspired by a
particular type of synthetic molecular motors -- molecular
spiders.
Molecular spiders [13, 14] are synthetic nanoscale walk-
ers which consist of a rigid,
inert chemical body to
which multiple flexible legs are attached. The legs are
deoxyribozymes -- enzymatic sequences of single-stranded
DNA that can bind to and cleave complementary strands
of a DNA substrate. When many such substrates are at-
tached to a surface, a leg can move between substrates,
cleaving them and leaving behind product DNA strands.
Products can be revisited by a leg, but they cannot be
cleaved again (Fig. 1). The leg cleaves, and then de-
taches, more slowly from a substrate than it detaches
from a product. The number of legs, and their lengths,
can be varied, and this defines how a spider moves on the
surface, i.e., its gait.
Mathematical models of molecular spiders at various
levels of abstraction have been proposed and studied.
Antal and collaborators introduced the first abstract
model of molecular spiders, and studied the motion of
a single spider on a one-dimensional track. They investi-
gated the movement of spiders with various numbers of
legs and various gaits over products only [15] and showed
that such spiders are equivalent to a regular diffusion; the
diffusion constants were computed for some gaits. Subse-
quently they introduced substrates and took into account
that cleavage and detachment from substrates together
take more time than the detachment from products [16],
∗Electronic address: [email protected]
†Electronic address: [email protected]
‡Electronic address: [email protected]; to whom correspondence
should be sent
Time
Product
(DNA)
Unbound leg
FIG. 1: (Color online) Molecular spider moves over a surface
of single-stranded DNA substrates. It has several flexible de-
oxyribozyme legs. When a leg detaches, it cleaves a substrate
strand, turning it into a shorter product DNA strand. The
leg can re-attach, but the bond will be weaker.
showing that this difference in residence time and the
presence of multiple legs, together, bias a spider's mo-
tion towards fresh substrates when it is on a boundary
between substrates and products. This important prop-
erty was also observed experimentally [14]. In Ref. [17]
we showed that spiders move superdiffusively for long
periods of time. Samii et al.
investigated various gaits
and numbers of legs [18, 19], emphasizing the possibil-
ity of detachment from the track. In Refs. [20, 21] we
studied the behavior of multiple spiders continuously re-
leased onto a 1D track. In a model with more physical de-
tail [22], we showed that spiders can move against a force
applied to the body. Models of spiders in two dimen-
sions have also been studied. In Ref. [23] we investigated
how fast several spiders with various gaits can locate a
small number of targets placed on a small fixed-size two-
dimensional lattice.
In Ref. [24] Antal and Krapivsky
evaluated the diffusion constant and the amplitude de-
scribing the asymptotic behavior of the number of visited
sites for a single spider with various gaits placed on an
infinite square lattice. Analytical results regarding the
asymptotic behaviors (limit theorems, transience, recur-
rence, and rate of escape) of spiders have been derived in
Refs. [25, 26]. In Refs. [27, 28] the behavior of spiders in
random environments was studied. Rank et al. showed
that several spiders, each placed on a separate 1D track
and connected to a single cargo particle move it faster
and remain superdiffusive longer than a single spider on
a single 1D track [29].
Here we study first passage properties of an abstract
theoretical model of molecular spiders. Our model is a
direct extension of the model introduced in Refs. [15, 16].
Although it is inspired by real molecular spiders, the
model can also be applied to a wider class of random
walkers that exhibit properties similar to spiders. Par-
ticularly, we investigate how the various boundary condi-
tions affect the spider's mean first passage time (MFPT)
when it moves over finite one- and two-dimensional sur-
faces.
First passage properties of regular random walkers
moving over discrete surfaces with various reflecting
and absorbing boundaries have been extensively stud-
ied [30, 31]. Here we show that the difference between the
time the legs spend on visited and on unvisited sites re-
duces the MFPT of two-legged spiders in various surface
settings, and increases the MFPT of one-legged spiders
(which behave like regular random walkers).
We start with the model of a two-legged spider
that moves over a one-dimensional track with absorb-
ing boundaries at both ends of the track. We found that
for this surface the cleavage rate significantly affects the
MFPT, and for any track length there exists an optimal
cleavage rate. Next, we study an extension of the 1D
model to 2D, i.e., the mean first passage time of a two-
legged spider to a circle, where the spider starts in the
center. For this surface we determined that the cleav-
age rate gives the spider an even greater advantage over
a regular random walker. The advantage persists even
when the target is a single site, and thus is much harder
to find. In this second 2D model the circle is a reflecting
boundary, its center is an absorbing boundary, and the
spider starts from various distances from the center.
II. MODEL DETAILS
Our model is a modification of the model we used in
Ref. [23]. It also can been seen as a direct extension of
the AK model of the spiders on a plane [24] that takes
various boundary conditions into account.
A. Motion and Surface
In our model a single spider moves over finite one- or
two-dimensional regular lattices. A spider has k legs. It
moves by detaching a leg from its site on the lattice and
reattaching it to a new site. Only one leg can detach at
any given time, so a spider cannot detach all of its legs
to leave the lattice. Each site can be occupied only by
one leg at a time. There is a restriction on the max-
imum distance between any two legs S (the gait), and
each leg can move to one of the nearest neighboring sites
(2 sites in 1D and 4 sites in 2D; a diagonal step is not
allowed) with equal probability as long as the move does
not violate one of the constraints above. Here we use
only two types of spiders. First, a spider with k = 1;
this spider is equivalent to a regular random walker. The
parameter S does not affect this spider since it has only
one leg. Second, a spider with k = 2 and S = 2; this
spider is exactly the bipedal Euclidean spider with max-
imal separation 2 [24]. When such a spider is placed on
a one-dimensional lattice, the model becomes equivalent
to that of Ref. [17].
Two types of sites can be present on the lattice, sub-
strate and product. A leg detaches at rate 1 from a
product, and at rate r from a substrate, where normally
r ≤ 1. Reattachment is instantaneous. When a leg leaves
a substrate, that substrate is transformed into a product.
Initially all sites are substrates; so any unvisited site is
always a substrate, and a visited site is always a prod-
uct. For r < 1 the legs act differently when they are
on visited versus unvisited sites, as described above. On
the other hand, when r = 1, legs act effectively the same
whether they are on the products or substrates, and it
becomes irrelevant if a site is visited (a product) or un-
visited (a substrate). Thus, spiders with r < 1 can be
seen as having memory, since they react differently to
visited and unvisited sites, and spiders with r = 1 can be
seen as having no memory, since they do not make this
distinction.
B. Boundaries and Starting Positions
We study the first passage time of spiders moving
over one-and two-dimensional regular lattices with var-
ious boundary conditions and initial configurations. In
each case we are interested when the spider reaches the
absorbing boundary.
The boundaries effectively make all our surfaces finite.
In one dimension we use a 1D track of length 2x. The
spider starts its movement from the middle of the track
(the origin), and absorbing boundary sites are located at
each end of the track, i.e., each one is x sites away from
the origin.
In two dimensions the surface is bounded by a circle of
radius x. Here we study three types of boundaries and
initial spider positions. First, the circle is an absorbing
boundary, and the spider starts from the center of the
circle. Second, the circle is a reflective boundary, while
the center is an absorbing boundary, and the spider can
start from any site on the circle. Third, we study the
case when radius x is fixed and the spider starts y sites
away from the target site at the center.
In all these settings x effectively defines the size of the
surface. And in all those settings, except the last, we
study the dependence of MFPT on x, i.e., hτ (x)i. In the
last case we study the dependence of MFPT on y, i.e.,
hτ (y)i.
III. SPIDER ON A 1D TRACK
A. Background: Transient Superdiffusivity of
Spiders
C. Meaningfulness of MFPT in The Studied
Settings
A process to determine if MFPT is a valid character-
istic of the first passage behavior was given in Ref. [32].
Similarly to that, we assess the meaningfulness of the
mean first passage time in all studied settings. For every
setting we estimate the distribution P (ω) of the random
variable ω = τ1/(τ1+τ2), where τ1 and τ2 are first passage
times of two independent spiders. Values of ω close to 1/2
indicate that spiders act similarly in a particular setting.
When ω is close to 0 or 1, the process is not uniform, and
MFPT is not a good measure of actual behavior. Distri-
bution P (ω) can have three distinct shapes: unimodal
bell-shaped, bimodal M-shaped, and plateau-like, almost
uniform behavior. Bell-shaped form with a maximum at
ω = 1/2 indicates that MFPT can be considered as a
valid measure of the first passage times of individual spi-
ders. M-shaped form with two peaks close to 0 and 1,
and local minimum at 1/2 indicates that MFPT is not a
good measure of the first passage time of individual spi-
ders. The plateau-like shape with zero second derivative
at ω = 1/2 separates the two above cases. Just as in
Ref. [32], to quantify the shape of P (ω) we fit P (ω) to
the model χω2 + c1ω + c2 for 0.05 < ω < 0.95. The sign
of χ indicates the shape of P (ω). In cases when χ < 0,
the distribution is bell-shaped; χ > 0 shows that the dis-
tribution is bimodal, M-shaped; and χ = 0 indicates that
the distribution is almost uniform.
D. Simulation
Combining the states of the spider and the surface
gives us a continuous-time Markov process for our model.
We use the Kinetic Monte Carlo method [33] to simulate
many trajectories of the Markov process for every in-
stance of the parameter set. In each case we record the
first passage time to the absorbing boundary.
For Section III we simulated 210 different r rates for up
to a distance of 10000 using 21504 traces of the Markov
process. For Section IV we simulated 100 different r rates
for up to a distance of 1000 using 20000 traces. For
Sections V A we simulated 18 different r rates for up to
a distance of 250 using 20000 traces. For Section V B we
simulated 15 different r rates for up to a distance of 100
using 20000 traces.
Antal and Krapivsky analytically obtain the mean time
hT (n)i to visit n sites in one dimension [15, 16]. Eq. 1
shows how hT (n)i depends on the value of r for a two-
legged spider (k = 2).
hT (n)i =
3
2
1 + r
2 + r
n2 +
1
r
n.
(1)
The parameter r affects the leading asymptotic behav-
ior of hT (n)i. Eq. 2, also from Ref. [16], gives hT (n)i for
a one-legged spider
hT (n)i =
n(n − 1)
4
+
n
2r
.
(2)
Eq. 2 demonstrates that for the one-legged spider
(k = 1), the leading term of hT (n)i does not depend
on the parameter r. The r only slows the one-legged spi-
der down by increasing the sub-leading term. Eq. 2 and
Eq. 1 also show that in the absence of memory (r = 1)
the one-legged spider is faster than the two-legged spi-
der. Thus these two properties, which separately make
the walkers slower, surprisingly improve the performance
of the spiders when combined together. This happens
because difference in residence time between visited and
unvisited sites biases multi-legged spiders towards un-
visited sites when they find themseves on the boundary
between visited and unvisited sites.
Using Kinetic Monte Carlo [33] simulations of the
Markov process we showed the unanticipated result that
spiders of the AK model with k = 2, S = 2, and r < 1
move superdiffusively over a significant span of time and
distance before eventually slowing down to move diffu-
sively [17].
Superdiffusive motion can be described using mean
square displacement of a walker as a function of time.
The mean squared displacement is given by Eq. 3, where
d is the number of dimensions, and D is an amplitude.
hX 2i = 2dDtα
diffusive
(3)
stationary
α = 0
0 < α < 1 subdiffusive
α = 1
1 < α < 2 superdiffusive
α = 2
ballistic or linear
From Eq. 3 we derive the condition for the walker to
be superdiffusive at time t. The walker is moving instan-
taneously superdiffusively [34] at a given time t if
α(t) =
d(ln (hX 2i(t)))
d(ln (t))
> 1.
(4)
In [17], we showed that each spider process goes
through three different phases of motion defined by its
value of α. Initially spiders are at the origin, and must
wait for both legs to cleave a substrate before they start
moving at all. So when t < 1/r the process is essen-
tially stationary (the initial phase). After the spiders
with r < 1 take several steps, they show a sustained pe-
riod of superdiffusive motion over many decades in time
(the superdiffusive phase). Finally, as time goes to in-
finity, all spiders will approach ordinary diffusion with
α ≈ 1 (the diffusive phase); the spiders mainly move
over regions of previously visited sites. which makes the
value of r less relevant.
B. Mean First Passage Time
We measure the mean first passage time, hτ (x)i, where
x is the absolute distance of the walker from the origin
on a one-dimensional track. At that point the walker
is absorbed by the boundary. Fig. 2 shows the initial
configuration of the track where the walker is positioned
in the middle, and the absorbing boundaries are shown
as stars. The length of the track is 2x.
x
FIG. 2: Initial configuration of the track. All sites are initially
substrates. Absorbing boundary is represented by stars.
According to the results of our numerical simulations
the MFPT of both one- and two-legged spiders is pro-
portional to x2. Eq. 5 shows the leading and sub-leading
terms of hτ (x)i.
hτ (x)i ≈ A1(r)x2 + a1(r)x.
(5)
The amplitude A1 of the leading term describes the
asymptotic behavior of the MFPT. For the one-legged
spider we found that A1 does not depend on r. For the
two-legged spider A1 increases with r, and approaches 2
when r = 1. As r approaches zero A1 ≈ 1.4. Fig. 3 shows
the amplitude A1 for one- and two-legged spiders for 182
r values less than 1.
For both one and two-legged spiders the amplitude a1
of the sub-leading term decreases monotonically and ap-
proaches 0 in the absence of memory (r = 1). The ampli-
tudes A1 and a1 show that the parameter r only increases
the MFPT of one-legged spiders. For two-legged spiders
varying r can decrease their MFPT.
In Section III A we recalled that when r < 1, the
AK-model walkers go through three different regimes
of motion -- the initial, superdiffusive, and the diffusive
stage. For lower r values the initial slow period is longer
than for higher r values, but subsequently the superdif-
fusive period is longer and faster. For travel over shorter
2.0
1.8
1.6
1
A
1.4
1.2
1.0
0.8
0.0
k=2
k=1
0.2
0.4
r
0.6
0.8
1.0
FIG. 3: Amplitude A1 from Eq. 5 as a function of the cleavage
and detachment rate r. The data are for one-legged (k = 1)
and two-legged (k = 2) spiders on a one-dimensional track.
distances the initial period is more important and thus
larger r values result in lower first passage times. For
travel over longer distances the superdiffusive period is
important and smaller r values give better results. Thus
for every particular distance there is an optimal value of
r that minimizes the MFPT. For example, for distance
2000 the spider with r = 0.05 is faster than the other
(sampled) r values; but for distances 4000 and longer the
spider with r = 0.01 is faster.
We estimated ropt(x) for tracks of various lengths (up
to 10000) through simulations of 210 r values. The re-
sults are shown in Fig. 4. The figure also shows the hτ (x)i
that corresponds to ropt(x) for each distance, i.e., the
minimum hτ (x)i achievable by varying the parameter r.
The optimum r monotonically decreases with distance.
Smaller r values create a stronger bias towards unvisited
sites, but they make individual steps slower when the
spider is on the boundary. Fig. 4 shows that for better
performance on shorter distances faster steps are more
important than the bias towards unvisited sites, whereas
for longer distances the bias dominates the MFPT.
We also estimate ropt(x) by fitting. First, we assume
that A1(r) and a1(r) have the functional form of Eq. 6,
and find the constants c1 to c8 by fitting Eq. 6 to the
estimates of A1 and a1.
A1(r) = (c1r + c2)/(c3r + c4)
a1(r) = (c5r + c6)/(c7r + c8)
(6)
Then, we substitute the results into Eq. 5 and find
when hτ (x)i is minimized by extracting the derivative
of hτ (x)i with respect to r. The predicted ropt(x) is also
drawn in Fig. 4. The derivation also shows that ropt(x) ∼
x−1/2 (see the Appendix for details).
Our estimates of the indicator χ for various r and x
values show that in the setting of the one-dimensional
optimal r prediction
optimal r data
corresponding t
101
100
10-1
t
p
o
r
10-2
100
101
102
x
103
109
108
107
106
105
104
103
102
101
100
104
e
m
T
i
FIG. 4: Optimal r values (ropt(x)) for various distances, and
the corresponding hτ (x)i. Optimal r values are obtained in
two ways; first, by fitting the data for hτ (x)i into the model of
Eq. 5 and amplitudes A1 and a1 into the model of Eq. 6, and
second, by simulating 210 values of r and choosing those that
correspond to the lowest hτ (x)i. The estimation of ropt(x)
shows that ropt(x) ∼ x−1/2 .
track χ is always negative, independent of the values of
the parameters x and r. Thus the MFPT is a meaning-
ful, reliable measure of the first passage time of spiders
moving on a one-dimensional track.
IV. SPIDER ON A 2D PLANE WITH A
CIRCULAR ABSORBING BOUNDARY
A direct extension of a one-dimensional track with
length 2x to two dimensions is a set of sites bounded
by a circle of radius x. Fig. 5 shows the initial config-
uration of the surface with the spider positioned in the
middle.
FIG. 5: Initial configuration of the 2D surface. All sites are
initially substrates. The spider starts at the center of the
circle. The absorbing boundary is shown as a circle; as soon
as either leg crosses the circle the target is considered to be
found, and the experiment stops.
Using numerical simulations we found that, similarly
to the spiders in one dimension, the MFPT of both one-
and two-legged spiders is proportional to x2. However,
the sub-leading terms are much closer to the leading
terms, and therefore are more important for estimating
the MFPT. This implies that in two dimensions spiders
(especially those with very small cleavage and detach-
ment rate, r < 0.05) approach the asymptotic behavior
especially slowly. Eq. 7 shows the leading and sub-leading
terms of hτ (x)i.
hτ (x)i ≈ (cid:26) A2(r) x2 + a2(r) x2/ ln t
: k = 2
A2(r) x2 + a2(r) x2/(ln t)0.88 : k = 1
(7)
The correlation (ln t)−0.88 for the one-legged spider is
unusual and slower than for the two-legged spider. Some-
what similar effects were observed in Ref. [24] in the es-
timation of the mean squared displacement. As in one
dimension, the amplitude A2 of the leading term does
not depend on r for the one-legged spider and is propor-
tional to r for the two-legged spider. The sub-leading
term's amplitude a2 monotonically decreases with r in
both cases. Fig. 6 shows the amplitude A2 for the one-
and two-legged spiders for 60 r values that are less than
1.
2.0
1.5
2
A
1.0
0.5
0.0
0.0
k=2
k=1
0.2
0.4
r
0.6
0.8
1.0
FIG. 6: Amplitude A2 from Eq. 7 as a function of the cleavage
and detachment rate r. The data is shown for the one-legged
(k = 1) and two-legged (k = 2) spiders moving over two-
dimensional surface with a circular absorbing boundary.
For one-legged spiders, as in one dimension, lower r val-
ues only increase the MFPT. The comparison of Figs. 6
and 3 shows that for two-legged spiders A2 is affected
more strongly by r than A1. The A2 of the two-legged
spider in two dimensions even intersects the A2 of the
one-legged spider. The A2 starts at 2 when r = 1 and
decreases towards ≈ 0.5 as r approaches zero.
Similar to the one-dimensional case, spiders with r < 1
start more slowly than the no-memory spider with r = 1;
then they move faster, and finally they slow down and
approach regular diffusion. But in contrast to spiders on
a 1D track, the r values that correspond to fastest times
study how hτ (x)i is affected by the parameter r, and how
it is affected by x in (1) and y in (2).
A.
2D Circle of Variable Radius With Target in
the Middle and Spider Starting from the Boundary
Fig. 8 shows the initial configuration of the surface
where the spider is positioned on the contour and the sin-
gle target site (absorbing boundary) is shown as a star.
Since the reflecting boundary is a circle and the absorb-
ing boundary is a single site in the center, all starting
positions of the same distance from the target site are
equivalent. In our simulations we pick a fixed position
on the contour as a starting position for the spider. This
position is shown as two small solid circles in Fig. 8.
are higher, and the MFPT of spiders with very small r
values (less than 0.1) approaches the MFPT of spiders
with r = 1 very slowly. However, the transition towards
the diffusive stage happens more slowly compared with
1D.
We estimated ropt(x) for circles of various sizes (up to
1000) through simulations of 100 r values, choosing the
fastest ones for each distance. Comparison of the results
is shown in Fig. 7. The figure also shows the fastest hτ (x)i
that corresponds to ropt(x) for each distance. Similarly
to the 1D case, we also estimate ropt(x) by fitting.
101
t
p
o
r
100
10-1
100
optimal r prediction
optimal r data
corresponding t
107
106
105
104
103
102
101
100
e
m
T
i
101
102
x
103
FIG. 7: Optimal r values (ropt(x)) for various radii. Optimal
r values are obtained in two ways; first, by fitting the data
for hτ (x)i into the model of Eq. 7 and amplitudes A2 and a2
into the model of Eq. 6, and second, by simulating 100 values
of r and choosing those that correspond to the lowest hτ (x)i.
FIG. 8: Initial configuration of the surface. All sites are ini-
tially substrates. The absorbing boundary is shown as a star.
The spider starts at the periphery. The circle is a reflecting
boundary and thus steps outside of the circle are not allowed.
Our estimates of the indicator χ for various r and x
values show that in the setting of the two-dimensional
circle with absorbing boundary at the perimeter χ is al-
ways negative, and this does not change with the values
of the parameters x and r. Just as in 1D, the MFPT
is a meaningful measure of the first passage time of the
individual spiders moving on a two-dimensional lattice.
We measure the first passage time of the walker from
the contour of the surface to its center for various radii
x.
Interestingly, the MFPT asymptotically grows slightly
faster than x2 . The sub-leading term also grows faster
than in the circular absorbing boundary case. Eq. 8
shows the leading and sub-leading terms of hτ (x)i.
V. SPIDER ON A 2D PLANE WITH A
CIRCULAR REFLECTING BOUNDARY AND AN
ABSORBING BOUNDARY IN THE CENTER
When we change the contour of the 2D surface to be
a circular reflecting boundary instead of an absorbing
one, and place a single target site in the center, spiders
with memory (r < 1) still have an advantage over spiders
without memory (r = 1). We consider two cases: (1)
when the radius x of the circle is variable, and spider
starts from any point on the contour, and (2) when the
radius x is fixed, and the distance x between the starting
position and the target is variable. Since both cases are
circularly symmetric, all starting positions with the same
distance from the center are equivalent. In both cases we
hτ (x)i ∼ B2(r)x2 + b2(r)x1.5.
(8)
Fig. 9 shows the amplitude B2 for the two-legged spi-
ders for 20 different r values between 0.1 and 1.0.
Values of r smaller than 1 give the spider an advan-
tage, similar to the 2D configuration discussed above,
but now even for shorter radii x. The typical time-line
of this process can be characterized as follows. First,
the spider starts moving, it eventually visits many sites
without finding the target site, and leaves behind many
smaller regions of substrates. Next, the spider starts to
move only over visited sites more often, and eventually
encounters regions of substrates of various shapes and
sizes. At this stage, the spider with r = 1 will not be
affected by the substrate regions, and will move just as if
they were visited sites. On the other hand, spiders with
38
37
36
35
34
33
32
2
B
31
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
r
FIG. 9: Amplitude B2 from the Eq. 8 as a function of the
cleavage and detachment rate r. The data are shown for two-
legged (k = 2) spiders moving over a two-dimensional surface
with a circular reflecting boundary. The single site in the
middle of the circle is an absorbing boundary.
r < 1 will become biased to stay on the substrates and
explore those regions more thoroughly; this will increase
their chances of finding the target site, since it must be
in one of those unvisited areas. This scenario can explain
how spiders with r < 1 gain an advantage over spiders
with r = 1 on this surface.
The indicator χ, in this setting, grows with the param-
eter x, and for x ' 17 χ becomes close to 0. As in the
previously described settings, χ does not depend on the
parameter r. Fig. 10 shows the dependence of χ on x
for r = 0.5. For small surfaces (small values of x), the
2
0
−2
χ
−4
−6
−8
−10
0
50
100
x
150
200
250
FIG. 10: Indicator χ shows how well the MFPT can describe
the first passage times of the individual spiders to a single tar-
get. The surface is two-dimensional with a circular reflecting
boundary of radius x, and spider starts from the contour.
χ is negative, and thus in those cases MFPT is a good
measure of the first passage time of individual spiders.
However, for larger x, close to 0 values of χ indicate that
the shape of the distribution P (ω) is close to uniform,
and thus the MFPT is not as good a measure of indi-
vidual behavior as it is in the settings when the spider
searches for the contour. It also shows that the possible
paths that the spider can take to locate a single target
are more diverse than paths that lead to the contour.
2D Circle of Fixed Radius With Target in the
B.
Middle and Spider Starting From Various Distances
Fig. 11 shows the initial configuration of the surface
where the walker is positioned at a distance y from the
target site, and the radius x is set to 100 and remains
constant.
FIG. 11:
Initial configuration of the surface. All sites are
initially substrates. The absorbing boundary is shown as a
star. The spider starts y sites away from the center. A circle
of constant radius x = 100 is a reflecting boundary and thus
steps outside of the circle are not allowed.
We measure the first passage time of the walker from
the contour of the surface to its center for various dis-
tances y. Fig. 12 shows a plot of the MFPT.
The shape of the curves is asymptotically logarithmic;
this shows that the initial position of the spider does
not significantly affect hτ (y)i when x is fixed. Even if
the spider starts closer to the target, there are too many
possible paths to the target in 2D, of which one is ran-
domly chosen. Many of them are very long and initially
lead the spider far away from the target.
The indicator χ, in this setting, is positive for y / 80
and decreases with the parameter y. For y ' 70 as the
surface configuration approaches the configuration dis-
cussed in the Section V A, and χ becomes negative, but
as in the case of the Section V A, it still remains close to
0. Fig. 13 shows the dependence of χ on y for r = 0.5.
The positive values of χ for the smaller y values indicate
that MFPT is not a good measure of the first passage
times of the individual spiders, and paths that lead the
spider to a target are very diverse and vary significantly
in their lengths.
×105
2.5
2.0
1.5
)
y
(
m
1.0
0.5
r=0.1
r=0.3
r=0.5
r=0.7
r=1.0
0
20
40
y
60
80
100
FIG. 12: Mean first passage time to a single point in the center
of the circle with a fixed radius. The boundary is reflecting,
and spider starts at various distances from the target.
2.0
1.5
1.0
χ
0.5
0.0
−0.5
−1.0
0
20
40
y
60
80
100
FIG. 13: Indicator χ shows how well the MFPT can describe
the first passage times of the individual spiders to a single
target. The surface is two-dimensional with a circular reflect-
ing boundary of a fixed radius x = 100, and spider starts at
the distance y from the target site.
VI. COMPARISON OF SPIDER
PERFORMANCE IN THE STUDIED SETTINGS
It is interesting to compare the advantage spiders with
r < 1 (i.e., with memory) enjoy over those with r =
1 (i.e., without memory) in the described 1D and 2D
settings. We compute the ratio of hτ (x)i for spiders with
r = 1 and hτ (x)i for spiders with r = ropt(x). This ratio
is plotted in Fig. 14 against the surface size for 1D and
2D.
The plot shows that the cleavage rate r gives spiders
more advantage on a 2D plane searching for a circle than
on a 1D strip searching for its ends. This advantage can
be attributed to the amount of substrates that spiders
1.6
1.5
t
p
o
=
r
r
1.4
t
p
f
m
/
1
=
r
t
p
f
m
1.3
1.2
1.1
1.0
0
1D search for the track's ends
2D search for the circle's contour
2D search for the circle's center
200
400
x
600
800
1000
FIG. 14: Ratio between hτ (x)i of the spiders with r = 1.0
and r = ropt.
leave behind as they move away from the origin. In 1D,
spiders do not leave any substrates behind, so there are
no substrates between the left and the right ends of the
sea of products, and when a spider moves back towards
the origin there are no substrates to bias it towards the
boundary. In 2D, the shape of the product sea can be
very complicated and there can be many substrates left
behind. When a spider moves backward it has a high
probability to still encounter substrates, which can bias
it towards the boundary. The higher ropt(x) values in
2D can be attributed to the direction of the emergent
bias when the spider is on the border between visited
and unvisited sites. In 1D, the border is simple, and its
shape remains the same over time. It is defined by the
two closest unvisited sites to the origin on the right and
on the left side. As a result, when the spider with rate
r < 1 is on the border, it is always biased in the desired
direction -- away from the origin.
In 2D, the shape of
the border can be much more complex and is even not
necessarily connected. This type of border leads to a
much weaker bias towards the edge of the surface.
In
many cases the spider is not biased directly to the edge
in the direction of the shortest path, and sometimes the
spider even is biased back towards the circle's center.
Despite that, the greater amount of substrates accessible
by the spider in 2D overcomes the weaker bias and makes
spiders more efficient at finding the absorbing boundary.
Fig. 15 shows the average density of products; the spider
has a high probability of encountering substrates when it
turns back towards the origin.
VII. DISCUSSION
Our simulations show that the MFPT of two-legged
spiders depends strongly on the kinetic parameter r in
all studied cases. The MFPT is lower for r < 1 (i.e.,
with memory) than for r = 1 (i.e., without memory) in
lands of substrates (i.e., unvisited sites). The presence of
these islands increases the probability that spiders with
r < 1 will find the target site, as opposed to a spider
with r = 1, which would not react to the presence of
those islands. This is a likely explanation for the greater
importance of r for this boundary condition.
Despite the varying importance of the parameter r in
the studied scenarios, in all cases there exists an opti-
mal value of r that minimizes the first passage time to
the absorbing boundary. To the extent that catalysis is
accessible as a design parameter of molecular spider as-
semblies, the results provide a way to optimize spider
system performance in various target search scenarios.
Acknowledgments
The authors would like to thank Paul Krapivsky for
detailed discussions concerning their model and analy-
sis. This material is based upon work supported by the
National Science Foundation under grants 0829896 and
1028238. We would also like to thank NVIDIA Corpora-
tion for a hardware gift that made possible some of the
simulations.
Appendix A: Optimum value of r
First, we find the derivative of the assumed form pf
hτ (x)i (Eqs. 5 and A1) with respect to r.
dhτ (x)i
dr
=
c1(c3r + c4) − c3(c1r + c2)
(c3r + c4)2
x2
+
c5(c7r + c8) − c7(c5r + c6)
(c7r + c8)2
x
=
c1c4 − c3c2
(c3r + c4)2 x2 +
c5c8 − c7c6
(c7r + c8)2 x
(A1)
The fitting of Eq. 6 to the estimates of A1 and a1 yields
estimates of constants c1 to c8. Next, we substitute the
constants into Eq. A1 and find when the derivative is
zero.
3.29r + 3.01
r + 2.15
r =
8.09x − 2.68
x + −4.35r + 1.34
5.75 + √4.05x − 1.34
= 0
r
FIG. 15: Average density of products when spider with r =
0.25 and k = 2 is at distance 1000.
the one-dimensional case when the spider is searching for
the ends of a one-dimensional track. For one-legged spi-
ders the parameter r does not affect leading asymptotic
behavior, however it slows them down by increasing the
constant of the sub-leading term. In the extension of this
problem to two dimensions, when the spider is searching
for the contour of a circle from its center, the advantage
of having r < 1 is even more significant, despite the less
effective bias provided by the shape of the leftover sub-
strates. Here the bias provided by the substrates does not
always direct the spider towards the absorbing boundary.
In contrast, on a 1D track, the substrates always bias the
spider towards the closest end when one of the legs is at-
tached to them. The disadvantage in 2D is overcome by
the greater amount of substrates that are accessible to
spiders. In 1D, the spider (of the studied gait) does not
leave any substrates behind when it progresses towards
the ends of the track.
In 2D, the shape of the sea of
products is complex, and many substrates are left be-
hind. Those substrates can bias the spider towards the
absorbing boundary when it starts to turn back towards
the origin.
When we reverse boundaries in 2D we make the ab-
sorbing boundary the single site in the center of the circle,
and start the spider from the contour, the parameter r
can still improve the MFPT. It is difficult to find the sin-
gle target site in this scenario, and the MFPT increases
significantly for all types of walkers. However, it is plau-
sible that when most of the surface is explored, spiders
with r < 1 are more likely to stick to small remaining is-
[1] M. Schliwa and G. Woehlke, "Molecular motors," Nature,
vol. 422, no. 6933, pp. 759 -- 765, 2003.
Looking under the hood of molecular motor proteins,"
Science, vol. 288, no. 5463, pp. 88 -- 95, 2000.
[2] R. D. Vale and R. A. Milligan, "The way things move:
[3] R. Phillips, J. Kondev, J. Theriot, and H. G. Garcia,
Physical Biology of the Cell. London and New York: Gar-
land Science, 2nd ed., 2013.
[4] P. Yin, H. Yan, X. G. Daniell, A. J. Turberfield, and
J. H. Reif, "A unidirectional DNA walker that moves
autonomously along a track," Angewandte Chemie In-
ternational Edition, vol. 43, pp. 4906 -- 4911, 2004.
[5] J. Bath and A. J. Turberfield, "DNA nanomachines,"
Nature Nanotechnology, vol. 2, no. 5, pp. 275 -- 284, 2007.
[6] R. A. Muscat, J. Bath, and A. J. Turberfield, "A pro-
grammable molecular robot," Nano Letters, vol. 11,
no. 3, pp. 982 -- 987, 2011.
[7] S. F. J. Wickham, J. Bath, Y. Katsuda, M. Endo, K. Hi-
daka, H. Sugiyama, and A. J. Turberfield, "A DNA-based
molecular motor that can navigate a network of tracks,"
Nature Nanotechnology, vol. 7, no. 3, pp. 169 -- 173, 2012.
[8] J.-S. Shin and N. A. Pierce, "A synthetic DNA walker for
molecular transport," Journal of the American Chemical
Society, vol. 126, no. 35, pp. 10834 -- 10835, 2004.
[9] S. Venkataraman, R. M. Dirks, P. W. K. Rothemund,
E. Winfree, and N. A. Pierce, "An autonomous polymer-
ization motor powered by DNA hybridization.," Nature
Nanotechnology, vol. 2, pp. 490 -- 494, 2007.
[10] T. Omabegho, R. Sha, and N. C. Seeman, "A bipedal
DNA brownian motor with coordinated legs," Science,
vol. 324, no. 5923, pp. 67 -- 71, 2009.
[11] M. von Delius, E. M. Geertsema, and D. A. Leigh, "A
synthetic small molecule that can walk down a track,"
Nature Chemistry, vol. 2, no. 2, pp. 96 -- 101, 2010.
[12] E. R. Kay, D. A. Leigh, and F. Zerbetto, "Synthetic
molecular motors and mechanical machines," Ange-
wandte Chemie International Edition, vol. 46, no. 1-2,
pp. 72 -- 191, 2007.
[13] R. Pei, S. K. Taylor, D. Stefanovic, S. Rudchenko, T. E.
Mitchell, and M. N. Stojanovic, "Behavior of polycat-
alytic assemblies in a substrate-displaying matrix," Jour-
nal of the American Chemical Society, vol. 128, no. 39,
pp. 12693 -- 12699, 2006.
[14] K. Lund, A. J. Manzo, N. Dabby, N. Michelotti,
A. Johnson-Buck, J. Nangreave, S. Taylor, R. Pei, M. N.
Stojanovic, N. G. Walter, E. Winfree, and H. Yan,
"Molecular robots guided by prescriptive landscapes,"
Nature, vol. 465, pp. 206 -- 210, May 2010.
[15] T. Antal, P. L. Krapivsky, and K. Mallick, "Molecular
spiders in one dimension," Journal of Statistical Mechan-
ics: Theory and Experiment, vol. 2007, no. 08, p. P08027,
2007.
[16] T. Antal and P. L. Krapivsky, "Molecular spiders with
memory," Physical Review E, vol. 76, no. 2, p. 021121,
2007.
[17] O. Semenov, M. J. Olah, and D. Stefanovic, "Mechanism
of diffusive transport in molecular spider models," Phys-
ical Review E, vol. 83, p. 021117, Feb 2011.
[18] L. Samii, H. Linke, M. J. Zuckermann, and N. R.
Forde, "Biased motion and molecular motor properties
of bipedal spiders," Phys. Rev. E, vol. 81, p. 021106, Feb
2010.
[19] L. Samii, G. A. Blab, E. H. C. Bromley, H. Linke,
P. M. G. Curmi, M. J. Zuckermann, and N. R. Forde,
"Time-dependent motor properties of multipedal molec-
ular spiders," Phys. Rev. E, vol. 84, p. 031111, Sep 2011.
[20] O. Semenov, M. J. Olah, and D. Stefanovic, "Multiple
molecular spiders with a single localized source -- the one-
dimensional case," in DNA 17: Proceedings of The Sev-
enteenth International Meeting on DNA Computing and
Molecular Programming, vol. 6397 of Lecture Notes in
Computer Science, pp. 204 -- 216, Springer, 2011.
[21] O. Semenov, M. Olah, and D. Stefanovic, "Cooperative
linear cargo transport with molecular spiders," Natural
Computing, pp. 1 -- 18, 2012.
[22] M. J. Olah and D. Stefanovic, "Superdiffusive transport
by multivalent molecular walkers moving under load,"
arXiv:1211.3482 [physics.bio-ph], Nov 2012.
[23] O. Semenov, D. Stefanovic, and M. Stojanovic, "The ef-
fects of multivalency and kinetics in nanoscale search by
molecular spiders," in Proceedings of the Italian Work-
shop on Artificial Life and Evolutionary Computation,
pp. 1 -- 12, 2012.
[24] T. Antal and P. L. Krapivsky, "Molecular spiders on a
plane," Physical Review E, vol. 85, p. 061927, Jun 2012.
[25] C. Gallesco, S. Muller, and S. Popov, "A note on spi-
der walks," ESAIM: Probability and Statistics, vol. 15,
pp. 390 -- 401, 0 2011.
[26] I. Ben-Ari, K. Boushaba, A. Matzavinos, and A. Roi-
tershtein, "Stochastic analysis of the motion of DNA
nanomechanical bipeds," Bulletin of Mathematical Biol-
ogy, vol. 73, pp. 1932 -- 1951, 2011.
[27] R. Juhasz, "Anomalous transport in disordered exclusion
processes with coupled particles," Journal of Statistical
Mechanics: Theory and Experiment, vol. 2007, no. 11,
p. P11015, 2007.
[28] C. Gallesco, S. Muller, S. Popov, and M. Vachkovskaia,
"Spiders in random environment," ALEA, Lat. Am. J.
Probab. Math. Stat., no. 8, p. 129147, 2011.
[29] M. Rank, L. Reese, and E. Frey, "Cooperative effects
enhance the transport properties of molecular spider
teams," Phys. Rev. E, vol. 87, p. 032706, Mar 2013.
[30] S. Condamin, O. B´enichou, and M. Moreau, "First-
passage times for random walks in bounded domains,"
Phys. Rev. Lett., vol. 95, p. 260601, Dec 2005.
[31] C. Chevalier, O. B´enichou, B. Meyer, and R. Voi-
turiez, "First-passage quantities of brownian motion in
a bounded domain with multiple targets: a unified ap-
proach," Journal of Physics A: Mathematical and Theo-
retical, vol. 44, no. 2, p. 025002, 2011.
[32] T. G. Mattos, C. Mejia-Monasterio, R. Metzler, and
G. Oshanin, "First passages in bounded domains: When
is the mean first passage time meaningful?," Physical Re-
view E, vol. 86, p. 031143, Sep 2012.
[33] A. B. Bortz, M. H. Kalos, and J. L. Lebowitz, "A new
algorithm for Monte Carlo simulation of Ising spin sys-
tems," Journal of Computational Physics, vol. 17, no. 1,
pp. 10 -- 18, 1975.
[34] A. M. Lacasta, J. M. Sancho, A. H. Romero, I. M.
Sokolov, and K. Lindenberg, "From subdiffusion to su-
perdiffusion of particles on solid surfaces," Physical Re-
view E, vol. 70, p. 051104, Nov 2004.
|
1004.3437 | 1 | 1004 | 2010-04-20T13:14:13 | Dynamic Principles of Center of Mass in Human Walking | [
"physics.bio-ph",
"physics.class-ph"
] | We present results of an analytic and numerical calculation that studies the relationship between the time of initial foot contact and the ground reaction force of human gait and explores the dynamic principle of center of mass. Assuming the ground reaction force of both feet to be the same in the same phase of a stride cycle, we establish the relationships between the time of initial foot contact and the ground reaction force, acceleration, velocity, displacement and average kinetic energy of center of mass. We employ the dispersion to analyze the effect of the time of the initial foot contact that imposes upon these physical quantities. Our study reveals that when the time of one foot's initial contact falls right in the middle of the other foot's stride cycle, these physical quantities reach extrema. An action function has been identified as the dispersion of the physical quantities and optimized analysis used to prove the least-action principle in gait. In addition to being very significant to the research domains such as clinical diagnosis, biped robot's gait control, the exploration of this principle can simplify our understanding of the basic properties of gait. | physics.bio-ph | physics | Dynamic Principles of Center of Mass in Human
Walking
Yifang Fan1∗, Mushtaq Loan2∗, Yubo Fan3, Zhiyu Li4 and
Changsheng Lv1
1Center for Scientific Research, Guangzhou Institute of Physical Education,
Guangzhou 510500, P.R. China
2International School, Jinan University, Guangzhou 510632, P.R. China
3Bioengineering Department, Beijing University of Aeronautics and Astronautics,
Beijing 100191, P.R. China
4College of Foreign Languages, Jinan University, Guangzhou 510632, P.R. China
E-mail: [email protected] and [email protected]
Abstract. We present results of an analytic and numerical calculation that studies
the relationship between the time of initial foot contact and the ground reaction force
of human gait and explores the dynamic principle of center of mass. Assuming the
ground reaction force of both feet to be the same in the same phase of a stride cycle,
we establish the relationships between the time of initial foot contact and the ground
reaction force, acceleration, velocity, displacement and average kinetic energy of center
of mass. We employ the dispersion to analyze the effect of the time of the initial foot
contact that imposes upon these physical quantities. Our study reveals that when the
time of one foot's initial contact falls right in the middle of the other foot's stride cycle,
these physical quantities reach extrema. An action function has been identified as the
dispersion of the physical quantities and optimized analysis used to prove the least-
action principle in gait. In addition to being very significant to the research domains
such as clinical diagnosis, biped robot's gait control, the exploration of this principle
can simplify our understanding of the basic properties of gait.
PACS numbers: 87.85.G, 87.85.gj, 87.55.de.
0
1
0
2
r
p
A
0
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
7
3
4
3
.
4
0
0
1
:
v
i
X
r
a
Dynamic Principles of COM
1. Introduction
2
Gait analysis plays an important role in exploring laws of human motion by gait
parameters via biomechanical methods. Many studies have shown that gait parameters
are significantly symmetric [1, 2] and can be understood in terms of segmental
kinematical and kinetic physical quantities, while walking [3, 4].
It has been shown,
[5, 6, 7, 8] that by developing gait parameters into evaluation indexes, one can assess
the causal relationship between physical injury and gait ability, which in turn has been
applied in rehabilitation therapy with success [9, 6, 10, 11, 12, 13].
The essence of human center of mass (COM) motion is actually the periodical
change which is under the effect of external forces such as, ground reaction force (GRF),
gravity and resistance [14]. Previous studies indicate that the law of COM motion is a
more reliable gait evaluation method [15]. The common features of the frequently-used
gait indexes lie in the fact that normal human gait parameters have been regarded as
the criteria to evaluate rehabilitation. However, the incomplete symmetry of human
shape, coordinate and strength has formed the uniqueness of human gait [17]. This has
brought difficulty to the establishment of standard gait parameter index. Consequently,
exploring a principle that is relevant to gait parameters such as cycle time, cadence
or stride length in a normal human gait has become essential to the study of gait
biomechanics [18]. There are a number of issues we wish to explore in our calculation.
An important question is whether the relationships between the GRF, velocity of COM,
average kinetic energy and the time of initial foot contact (tIFC) in a stride cycle can be
established. In particular we wish to examine if and to what extent such relationships
explore the principle behind the gait characteristics. In the present work, we wish to
advance an attempt towards an approach where the precision of the calculation reaches
new levels.
The discovery of many basic principles originated from the study of animal
movements [14]. By examining the GRF acted on the normal human gait (bare-footed),
we establish the relationships between the GRF, velocity of COM, average kinetic energy
and the tIFC in a stride cycle. In this contribution we present our recent results on
kinetic regularities of COM. Another crucial issue in this attempt concerns the prediction
of action in gait. Using analytic and numerical techniques, we shall identify the action
in gait and explore the possibility of the least-action principle in gait. The rest of the
paper is organized as follows:
In Sec.
II we describe our procedure to establish the relationships between the
GRF and the tIFC and discuss the aspects of determination of the force distribution
in anteroposterior, mediolateral and longitudinal directions. We address the problem
of the force distribution in transverse, sagittal and frontal planes and optimization of
dispersion of force. We present a description of the determination of working parameters,
such as acceleration, velocity and displacement of COM and establish the relationships
between the kinetic regularities and the tIFC in this section. We conclude this section
by addressing the effect of the tIFC upon the average kinetic energy of COM. Our main
Dynamic Principles of COM
3
results are presented and discussed in Sec. III. Finally, we present our conclusions in
Sec. IV.
2. Analysis Strategy - Forward Dynamic Method
2.1. The Effect of tIFC upon GRF
The analysis of COM's dynamic characteristics to evaluate the gait features is rather
non-trivial. The inverse dynamic method [19, 15, 16] has given mixed results with
systematic and statistical errors as major sources of uncertainties in the analysis of COM
dynamics characteristics [20, 21]. An attempt to indicate the dynamics characteristics
of the whole body COM by one particular segment of body will highly underestimate
the results. The target of this work is a calculation that gives more precise estimates
of dynamical measurements in question. We use forward dynamics method [22, 23] to
illustrate on how tIFC determines the force distribution and establish a relationship
between the tIFC and its force dispersion.
Since dynamic characteristics of COM in gait is governed by kinematics and kinetics
of COM, therefore, for constant gravity, the analysis of force upon COM will compliment
the analysis of GRF. Also, GRF is caused by the body segmental movement driven by
the transarticular muscles and eventually by the contact of the foot to the ground, so
one foot's GRF includes the longitudinal GRF, frontal friction and sagittal friction, i.e.,
F (t) = Fx(t)i + Fy(t)j + Fz(t)k,
(1)
i (t) and F r
where F (t) is the GRF of one moment at a stride cycle and Fx(t), Fy(t) and Fz(t)
represent the components in three directions, respectively. If F l
i (t) (i = x, y, z)
represent both feet's sagittal and frontal frictions and longitudinal GRF, respectively,
then the above equation shows a similarity of both feet's GRF distribution when
expressed in biped gait features [18]. Consequently, we assume that in each stride cycle,
the left and right foot have an identical distribution of GRF in all three directions. This
would imply that the GRF variations are only caused by the tIFC of both feet. Setting
T as one foot's stride cycle time, the initial phase of one foot is equal to zero and that
of the other foot to (i.e. tIFC), then corresponding to one foot's GRFs (for example left
foot)
F l
x(t) = Fx(t), F l
y(t) = Fy(t), F l
z(t) = Fz(t),
those of the other foot are
F r
x (t) = Fx(t + to), F r
y (t) = Fy(t + to), F r
z (t) = Fz(t + to)
and Eq. (1) becomes
F (t, to) = Fx(t, to)i + Fy(t, to)j + Fz(t, to)k.
(2)
Since gait is a continuous and periodic movement, therefore while walking at steady
If
speeds, F (t, to) is the GRF when tIFC is to and the stride cycle time is t.
Dynamic Principles of COM
4
F (t, to) = F (t + nT, to + nT ) (n = 1, 2, 3 . . .) holds, then the GRF and gravity (W )
in a stride cycle that act on the impulse of COM will follow the following equation:
I = Ixi + Iyj + Izk ≡ 0.
(3)
To calculate the impulse characteristics of each foot's GRF in gait when both feet have
same GRF, we analyze two phases of one foot's stride cycle. Let Ts and Tw represent the
stance phase and swing phase, respectively, then in the anterposterior direction, Eqs.
(2) and (3) yield
Z T
0
Fx(t)dt = 0 and
Z T
0
Fx(t)dt = Z T
0
Fx(t + t0)dt.
Since T = Ts + Tw and R Ts+Tw
Fx(t)dt = Z Ts
Z Ts
Ts
0
0
Fx(t)dt = 0, therefore
Fx(t + t0)dt = 0,
whereas in the mediplateral direction, we get
Z Ts
0
Fy(t)dt = Z Ts
0
Fy(t + t0)dt.
On the other hand, the calculated contribution in the longitudinal direction gives
Z T
0
Fz(t)dt = Z T
0
Fz(t + t0)dt =
1
2 Z T
0
W dt.
The above equations give the impulse characteristics of each foot's GRF in gait.
2.2. Dispersion of GRF
To analyze the dynamics that emerge in the GRF distribution at t0 = 0, T /2 and T ,
we bring forward the concept of dispersion of GRF. If σi (i = x, y, z) and ¯Fi denote the
GRF dispersions and average values of GRF in three direction, then the corresponding
correlation between the force dispersion and average GRF can be written as
T
T
σi(to) = vuut
Xt=0
1,
(Fi(t, to) − ¯Fi(to))2/T
Xt=0
where Fi(t, t0) is the same as that in Eq . (2) and PT
t=0 1 is the number of △t within
the range [0, T ]. For sequential values of t0 within the range [0, T ], we calculate the
values of σi and use them to evaluate the effect of tIFC upon GRF, velocity, position
and kinetic energy of COM. We use the average value of GRF from 20 subjects to check
the reliability and accuracy of this method. This signature is confirmed in Sec. III of
this study.
(4)
Dynamic Principles of COM
5
2.3. Regularities of velocity and displacement of COM
Ground reaction force‡ is the result of body segments action on the ground via foot
and determines the kinetic regularities such as acceleration, velocity and displacement
of human body COM. Knowing the GRF and weight, the acceleration of COM at a
given instant in a stride cycle using Eq. (2) can be written as
a(t, to) = F (t, to) − W k.
(5)
Expressing the acceleration of COM in component form, Eqs. (1) and (5) reveal that
the tIFC and GRF have the same effect on the acceleration of COM.
The absolute motion of human COM relative to absolute inertia reference frame
is composed of convected motion and relative motion. Following Kokshenev [14], we
define gait in accordance with motion in Eq. (3), as "walking at steady speeds", thus
we regard the convected motion a constant parameter. Using Eq. (5) at tIFC = to, the
COM velocity in a stride cycle is given by
v(t, to) = Z t
0
a(t, to)dt + v0(to),
(6)
where v0(to) is the initial velocity of COM in relative motion at the beginning of a stride
cycle. It seems that we cannot confirm the magnitude of v(t, to) in Eq. (6) by kinetic
method (since everyone's gait speed is different). However, the distinct feature is that
once GRF and tIFC are determined and when it observes the motion in Eq. (3), v0(to)
must have a unique solution. We can then establish the relationship between the initial
velocity and acceleration in relative motion in a stride cycle
v0(to) = −
T
Xλ=1
Z λ
0
a(t, to)dt.
(7)
Eq. (6) reveals that v(T, to) = v(to), which implies that the cycle of velocity of COM
is in accordance of v(T + nT, to + nT )(n = 1, 2, · · ·), that is to say, the end of one
stride cycle marks the beginning of the next stride cycle. This explains the periodical
characteristics of velocity of COM in gait. What needs to be further illustrated is that
the initial velocity calculated using Eq. (7) refers that at the beginning of a stride cycle,
the body is in the state of steady speeds. From Eqs. (5) and (7) we conclude that
tIFC has determined the initial velocity in relative motion, which has nothing to do the
convected velocity.
Just like the velocity of COM, displacement of COM also involves absolute,
convected and relative displacements. We define the COM initial displacement in
relative motion of different tIFC as s0(to). Using Eq. (6), the displacement of COM in
relative motion at any moment has the following form:
s(t, to) = Z t
0
v(t, to)dt + s0(to)
(8)
‡ We assume a constant gravity and a negligible air resistance.
Dynamic Principles of COM
6
and the correlation between the initial displacement of COM and velocity of COM in
relative motion in a stride cycle can be written as
s0(to) = −
T
Xλ=1
Z λ
0
v(t, to)dt.
(9)
Similar to the behaviour of the initial velocity in relative motion, we verified that the
initial displacement seems to be independent of gait velocity, cadence and stride length.
2.4. Average kinetic energy
Nature has always minimized certain important quantities when a physical process takes
place [24]. Bipedal walking has enabled the continuous evolution of human gait [25] and
eventually it has brought about the optimized gait [26] and thus became a behavioral
In this behavior, the force, acceleration and velocity all have their minimal
trait.
values when to = 1
2 T . To understand the physical significance of these gait dynamic
characteristics we explore the issue of energy consumption.
In gait, human segment
movement is a combination of agonist, antagonist and synergist. The movements such
as stretch or flexure all consume mechanical energy. On the other hand, Ek ≥ 0 in COM
1 Ek(t) > 0 in total kinetic energy and ¯Ek(t) > 0 in average kinetic
energy in a stride cycle, whereas the corresponding potential energy counterparts are
zero respectively. Using the COM kinetic energy in relative motion, the description of
mechanical energy consumption in gait simplifies to
kinetic energy, PT
E(t, to) =
1
2
v2(t, to),
(10)
and can be easily resolved in component form. Knowing Ei, we can set up the
relationships between tIFC and COM kinetic energy by defining t and to in the interval
[0, T ].
2.5. Experimental Details
The experimental measurements were carried on the following set of equipments: Simi
Motion 7.0 Three-Dimensional Movement Analysis System; three Kistler 40 × 60 (cm2)
force plates; force plate frequency: 2000Hz, with a systematic uncertainty of ±1%. The
assembled force plate position has been fixed by gradienter to ensure the force plates are
on the same plane. The measurements are taken at the sample frequency of 1000Hz.
Before each measurement, the equipment is examined and returned to zero-position.
Twenty female subjects with mean age 20.63±0.76 years, mean height 163.12±3.72
cm and mean weight 45.74 ± 3.30 kg participated in the study. A few trials of each level
gait item were administered to subjects as they ambulated on instrumented positions.
All the patients agreed to participate in the research, and signed freely an informed
consent form and study was carried out according to the existing rules and regulations
of our institute's Ethnic Committee. Before the test, all the subjects were thoroughly
briefed about the procedures and matters needing attention so that they all understand
Dynamic Principles of COM
7
the purpose and the requirements of the test. Each subject's medical history is inquired
so as to exclude subjects with diseases such as pathological change, deformity or injury
to make sure that their physical conditions would meet the requirements of the test.
When measuring their gaits, we start from the subjects' standing position, bare-footed
(both feet disinfected by 75% of ethanol). A pre-test is to guarantee that after they
walked three steps, the subjects step on the platform and to make sure that each force
plate can measure one foot's stance phase data separately. When the subject's gait is
found to be quite abnormal, for example, it is obviously discontinuous, she would be
asked to perform again so that the recorded data meet the requirements of the test.
Based on the fact that the longitudinal GRF in gait is apparently greater than the
GRFs of the anteroposterior or mediolateral direction, we take the signal of longitudinal
GRF to identify the instants of initial foot contact and of terminal stance. The gait we
study is walking at steady speeds; therefore, we de-noise the signals of the platform and
examine the test signals would meet the requirements in the predetermined accuracy
range. We rate the gait cycle time by percentage, normalize the weight and standardize
the GRF from three directions. We count on average of the processed data of 20 subjects'
three-direction GRF from individual stride cycle to conduct our research.
3. Results and Discussion
3.1. Spatial GRF, COM velocity and displacement
To explore the spatial GRF, we examine Eq.
(2) by using the GRF numeric of 20
subjects. Fig. 1 collects and displays the results of anterposterior, mediolateral and
longitudinal GRF. The GRF has been standardized (Fi(t)/W )) and stride cycle time
is rated by percentage. Fig. 1A - C shows that tIFC has an effect upon the GRF in
three directions. Using three sets of component data from GRF of the subject's gait, we
obtained curved surface effective plots and the quantitative relationship between tIFC
and GRF resultant forces on three planes. Fig. 1D - F indicates that no matter what
changes to undergoes, GRF resultant forces in these three planes remain to be closed
curves. This significantly compliments and confirms the dynamics behind Eq. (3), i.e.,
when the distribution of one foot's GRF is determined and both feet's GRFs remain
the same, the momentum of COM always remains unaffected.
Fig. 1D - F also reveals the geometrical characteristics of the plane resultant force.
On the transverse plane, when to = 0, T , the gait becomes a jump and its frontal
resultant force becomes a straight line. The resultant force changes into a symmetrical
butterfly about Fx, which could imply a possible transformation at to = 1
2T . On the
sagittal plane, when to = 0, T , the resultant force Fxz forms the longest closed curve
(the length of which is calculated by H Fxz(t, to)dFxdFy) and more or less shortest at
to = 1
2T . On the frontal plane, when the resultant force develops into a straight line at
to = 0, T , the resultant force becomes a symmetrical butterfly about Fz at to = 1
2T .
Having developed the relationship between tIFC and GRF's resultant force on three
Dynamic Principles of COM
8
Figure 1. Schematic illustration of the observed spectrum for tIFC and GRF showing
the relationship between tIFC and GRF in three directions (Fig.1A - C), together with
tIFC and GRF resultant force in three planes (Fig.1D - F), compared with a typical
tIFC and dispersion of force correspondence (Fig.1G - I).
planes, we calculate the GRF dispersion in three directions using Eq. (4). To confirm
the results beyond doubt, we identify the GRF dispersion by analyzing the effective
plots shown in Fig. 1G - I. It is clear from the plots that at to = 1
2 T , σx has its
maximum value at the bottom region of "W" shape, σy has a global maxima and σz
a global minima. Using the numerical values, min(σx) = 0.0695, max(σx) = 0.1319,
min(σy) = 0.0000, max(σy) = 0.0355, min(σz) = 0.1086 and max(σz) = 0.9144, of GRF
dispersion obtained in this case study, we find that the largest differences between the
maximal and minimal values are 0.0624, 0.0355 and 0.8058, respectively. This means
that the dominant effect to the global GRF is the longitudinal GRF. The minimal
GRF dispersion on longitudinal direction occurs at to = 1
2T . The analysis of the
distribution of GRF in three directions indicates that when to = 1
2T , the anteroposterior
and longitudinal dispersions of GRF are almost the least and the mediolateral one is the
largest. The closed curve shaped by the two components on the sagittal plane seems to
be the shortest. The GRF resultant force on the transverse and sagittal plane is shown
as symmetric butterfly.
Turning our attention towards the kinetic regularities, we plot the effect of tIFC
on COM velocity in Fig. 2. We can see two platforms in Fig. 2A - C, which emerge
at the two extremes of tIFC rated by percentage forming a concave region around tIFC
whereas the velocity describes a convex region around tIFC. The COM velocity in the
longitudinal direction indicates that tIFC entails the vertical velocity of COM to have
a distribution of tIFC in the form of a saddle. In order to further understand the effect
of tIFC exerting on the velocity of COM, we analyze the variations of the velocity of
Dynamic Principles of COM
9
COM in three planes. Looking at Fig. 2D - F, we notice that in the transverse plane,
Figure 2. Effective plots showing relationship between tIFC and COM velocity in
three directions (Fig.2A - C), tIFC and COM velocity in three planes (Fig.2D - F) and
tIFC and dispersion of COM velocity (Fig.2G - I).
the direction of COM velocity is a straight line at to = 0 and T and the plane velocity
becomes a symmetrical closed curve about vx axis at to = 1
2T . In the sagittal plane
the closed curve formed by the plane velocity is the longest (the length of the closed
curve is calculated by H v(t, to)dvxdvz) at to = 0, T , and is shorter (not the shortest)
at to = 0 and T . In the frontal plane, when to = 0, T , the velocity becomes a straight
line and the resultant velocity becomes a symmetrical closed curve about vz axis at
to = 1
2T . Having obtained the estimated value of COM velocity, we are ready to set
up the dispersion of relationship between tIFC and COM velocity (as shown in Fig.
2G - I) using Eq. (4). We notice that at to = 1
2T , σx and σy have the global minima
and σz has the global maxima. The estimates, min(σx) = 0.0541, max(σx) = 0.1652,
min(σy) = 0.0000, max(σy) = 0.0510, min(σz) = 0.0549 and max(σx) = 1.3502 in our
case study, result in largest differences of maximal and minimal values of 0.1111, 0.0510,
and 1.2953 respectively. This means that the dominant effect upon COM velocity lies in
the longitudinal and anteroposterior directions, where the dispersions have the minimal
values at to = 1
2 T .
The kinematic regularity of the displacement of COM in three directions is shown
in Fig. 3A - C. We see a geometric distribution of convex corners indicating that tIFC
changes the position of COM in anteroposterior and mediolateral direction. The bimodal
plot indicates the complexity of longitudinal direction exerted by tIFC to the position of
COM. The estimated values of the COM displacement components are used to evaluate
the values and directions of COM displacement in three planes. To further analyze the
effect of tIFC upon the position of COM, we explore the changes of COM positions from
Dynamic Principles of COM
10
three planes which are displayed in Fig. 3D - F.
Figure 3. COM displacement as a function of tIFC. The directional and planer
components are shown in Fig.3A - C and Fig.3D - F, respectively. Fig.3G - I shows
the dispersion of COM displacement.
These effective plots show that in the transverse plane the COM displacement is
a straight line at to = 0, T . On the other hand a closed curve is formed by sx and sy
and is symmetric about sx axis at to = 1
2 T . On the sagittal plane, when to = 0, T , the
closed curve shaped by the sagittal displacement is the longest (the length of the closed
2T , the curve is rather shorter (but
not the shortest). On the frontal plane the frontal displacement becomes a straight line
at to = 0, T and a symmetric closed curve about sy axis at to = 1
curve is calculated by H s(t, to)dsxdyz); when to = 1
2 T .
Using the numerical values, min(σx) = 0.0042, max(σx) = 0.0252, min(σy) =
0.0000, max(σy) = 0.0081, min(σz) = 0.0039 and max(σz) = 0.2147, obtained in this
case study, we find that of the GRF dispersion σx and σy have a global minima and σz
has a global maxima at around to = 1
2T , again resulting in largest differences between the
minimal and maximal values with a magnitude of the order 0.0210, 0.0081 and 0.2108,
respectively. This implies that the greatest effect to COM displacement comes from
components on longitudinal and anteroposterior directions while the minimal dispersion
of displacement on the longitudinal and anteroposterior direction emerges at to = 1
2T
(See Fig. 3G - I).
Accordingly, at to = 1
2T , COM velocity in relative motion presents its symmetric
closed curve in transverse and frontal plane and the curve length in the sagittal plane is
approximately the minimal. In a stride cycle, the dispersion of COM velocity in relative
motion has the global minima on the anteroposterior and longitudinal direction while
the global maximal value on the mediolateral direction. The effect of tIFC upon the
Dynamic Principles of COM
11
COM velocity and displacement does not depend on gait velocity, cadence or stride
length whereas the effect of the tIFC upon COM acceleration is in accordance with
the effect it exerts upon the GRF. The COM acceleration and GRF show more or less
identical behaviour. Fig. 4A - C demonstrate the effect of tIFC upon three components
of GRF, COM velocity and COM displacement on three direction. As is clear from Fig.
4D - F that at half stride cycle (t = T /2), the dispersion of GRF and COM kinematics
are minimum. In comparison to GRF and COM velocity, the COM displacement shows
a sharp dip with least minimum.
Figure 4. Relationship between tIFC and dispersion of force.
3.2. Anteroposterior, Mediolateral and longitudinal average kinetic energies
5D - F). Fig.
Finally, we show the COM kinetic energies as a function of stride cycle in anteroposterior,
mediolateral and longitudinal channels in Fig. 5A - C, together with corresponding
transverse, sagittal and frontal energies in three planes (Fig.
5G
- I indicates that tIFC makes COM average kinetic energy the global minima on
anteroposterior and mediolateral direction while that on the mediolateral direction a
global maxima at to = 1
2T . Comparing the relationship between the tIFC's velocity
of COM and the position of COM, tIFC contributes a symmetric distribution of
kinetic energy of COM. Since min(Ex(t, to)) > max(Ey(t, to)) and max(Ey(t, to)) <
min(Ez(t, to)), the anteroposterior and longitudinal average kinetic energies determine
the COM average kinetic energy. when tIFC falls at the 20%, 50%, 60% and 100% of a
stride cycle, the kinetic energy of COM in three planes varies rather considerably. On the
transverse and frontal planes, the effect is similar whereas the sagittal plane undergoes
a non-trivial change. We apply the dispersion of kinetic energy of COM to evaluate
this effect. We notice, from Fig. 5J, that COM average kinetic energy is minimum at
to = 1
2T , which implies a minima for the COM mechanical energy consumption and
signatures the physical significance of gait.
Dynamic Principles of COM
12
Figure 5. Relationship between tIFC and kinetic energy. (Fig.5A - C tIFC and COM
kinetic energy in three directions; Fig.5D - F -tIFC and COM kinetic energy in three
planes; Fig.5G - I - tIFC and average kinetic energy. )
3.3. Least-action principle in gait
The intriguing question that arises from the above discussion if there is a symmetry
or a physical principle that is hidden in the phenomenon that makes the COM kinetic
energy to reach its extremum at to = 1
2 T . In order to explore this issue, we pursue the
technique of optimizing the dispersion objective function [18]
T
X1
(f (t) + f (t + to) − f (to))2 =
T
X1
f (t)2 +
T
X1
f (t + to)2
+2
T
X1
f (t)f (t + to) − 2
T
X1
(f (t) + f (t + to)) + (f (to))2
T
X1
1
(11)
The action of GRF and gravity are the reasons for human COM motion changes. Let's
first analyze the GRF. The average value of the resultant force in the anteroposterior
and mediolateral direction ¯Fx(to) and ¯Fy(to) is zero and in the longitudinal direction
¯Fz(to) is one (when the weight has been normalized). Therefore, for a known gait,
PT
1 (Fx(t))2,PT
1 (Fz(t +
1 (Fy(t + to))2 and PT
1 (Fy(t))2,PT
1 (Fz(t))2,PT
1 (Fx(t + to))2,PT
Dynamic Principles of COM
13
to))2 are all constants and Eq. (11) simplifies the dispersion optimization to
T
min ψFx(t0) =
max ψFy (t0) =
min ψFz (t0) =
T
X1
X1
X1
T
Fx(t)Fx(t + to)
Fy(t)Fy(t + to)
Fz(t)Fz(t + to).
(12)
To find a solution of the optimization problem, we need to set up a function of GRF
that changes with time in three directions. It seems that the pulse periodic GRF fits
the criteria and hence we use segment trigonometric functions
Fx(t) = a sin(2π
t
Ts
),
Fy(t) = b sin(π
t
Ts
) Fz(t) = c sin(π
t
Ts
),
where a, b, c stand for the GRF after normalization in three directions and Ts is the
stance time. For a small enough test frequency, the above equation transforms to
sin(2π
t
Ts
) sin(2π
sin(π
t
Ts
) sin(π
t + to
Ts
t + to
Ts
)dt
)dt
0
min ψFx(t0) = AZ Ts
max ψFy (t0) = − BZ Ts
min ψFz (t0) = C Z Ts
0
0
sin(π
t
Ts
) sin(π
t + to
Ts
)dt.
(13)
Let's take longitudinal GRF in Eq. (13) as an example. In order to get the antiderivative
of the integrand of integral variable t, we detach t and to of trigonometric function to
obtain
Z Ts
0
sin(
t
Ts
π) sin(
t + to
Ts
π)dt = Z Ts
0
(sin2(
tπ
Ts
) cos(
tπ
Ts
)
+ sin(
tπ
Ts
) cos(
tπ
Ts
) sin(
toπ
Ts
))dt,
(14)
Now we can transform Eq. (14) into the integral of integral variable t while regarding
the GRF as segment function. The transformation reduces the pulse periodic GRF
functions into three piecewise functions at three intervals§: [0, Tw], [Tw, Ts] and [Ts, T ]
which, to the confirmed GRF, a, b, c are constants and A, B, C (related to a, b, c) are
§ While walking Ts + Tw = T and Ts > Tw > 0.
Dynamic Principles of COM
14
also constants, yields the following form of longitudinal GRF:
) − (
to − Ts
2Ts
π) cos(π
) − (
to
Ts
to − Tw
to − Ts
2Ts
π) − (
π) cos(π
to
Ts
)
to
Ts
)
sin(π
to
Ts
1
2
0 ≤ to ≤ Tw
1
2
sin(π
1
2
+
sin(
Ts
Tw ≤ to ≤ Ts
1
to − Tw
2
Ts
Ts ≤ to ≤ T
sin(
min ψFz (to) = C
to − Tw
2Ts
π) cos(
to − Tw
Ts
)
(15)
π) − (
to − Tw
2Ts
π) cos(
to − Tw
Ts
π)
−Ts
2Ts
π) sin(π to
Ts
The contribution ( to
which in turn yield (from Eq.
contribution ( to
Ts − to = to − Tw which yields C(cos( Tw
2Ts
we obtain a maximum value Cπ
π) sin(π to
Ts
) + ( to
−Ts
2Ts
in [Ts, T ].
2
) = 0 in the interval [0, Tw], thus giving the to = 0,
(15)) a maxima (= Cπ
2 ). On the other hand, the
−Tw
) = 0 in [Tw, Ts] giving the solution
2Ts
π) sin( Tw
π)) (a minima). Similarly,
2Ts
π) sin(π to
π) − ( Ts
−Tw
Ts
−Tw
2Ts
−Ts
Ts
π) cos( Tw
Ts
π) − sin( Tw
Ts
π)) and a maxima of B(( Ts
it is easy to verify (from Eq.
Following the above procedure,
13) that at
to = T /2 the dispersion of GRF in the anteroposterior direction attains a minima
of A(( Tw
π)).
in the mediolateral direction. Since, in gait, A > B > C, the sum of dispersions in
three directions is minimal at to = 1
2 T . This confirms our results shown in Figs. 4D
- F and 5J. Similarly, the dispersion of COM acceleration, dispersion of COM velocity
and the COM mechanical energy consumption are all the minimal at to = 1
2T . This
phenomenon is independent of the physiological factors such as height, weight and gait
parameters [18].
π) − cos( Tw
2Ts
π) sin( Tw
2Ts
−Tw
2Ts
4. Conclusion
In periodic motion, the foot's stance and swing substitute one another such that one
foot always remains in stance. This allows the human body to be acted by the periodical
GRF, which is related not only to the foot's movement style, but also to the substitution
style of one foot with another. Following the biped movement style in normal gaits and
the similar traits of GRF, we have been able to study the effect of tIFC upon the human
COM dynamic characteristics based on the assumption that both feet's GRFs are the
same in the same phase. Our results suggest that when tIFC falls in the middle of the
other foot's stride cycle, the COM kinematics and dispersions of GRF acted on COM
are the minimal, which has entailed the minimal average MEC of muscles. Our analysis
suggests that it falls into the category of least-action principle and is consistent with
the phenomenon that exists in normal gaits [18].
Based upon the least-action principle, we have observed that tIFC has caused the
GRF and the COM regularities to form a closed-curve on the transverse and frontal
Dynamic Principles of COM
15
plane, which might present a new method for gait evaluation. We advocate the use
of this method to uncover and speed up the diagnosis simply by measuring the GRF
for the patients with foot injuries or arthritis (for such patients, the model for these
quantities is not symmetric [27]). Meanwhile, the patients need only walk a few steps in
their normal gait, their COM dynamic characteristics will be acquired easily and more
accurately. This is exactly what the clinic diagnosis is looking for. In addition to the
human gait's adaptation to natural environment [25, 28], the evolution of gait is the
result of its observation of least-action principle. We believe that precise measurements
of the variations of shear stress in natural gaits (bare-footed) shall profoundly enriched
the content of least-action principle [29]. A further study of least-action principle will be
significant to the domains such as sport rehabilitation, biometric identification [30, 31]
and control of biped robot gaits [32, 33, 34].
Acknowledgments
This project was funded by National Natural Science Foundation of China under the
grant 10772053, 10972061 and by Key Project of Natural Science Research of Guangdong
Higher Education Grant No 06Z019. The authors would like to acknowledge the support
from the subjects.
References
[1] Mitoma H, Hayashi R, Yanagisawa N and Tsukagoshi H 2000 J. Neurol. Sci. 174 22
[2] Kim C M and Eng J J 2003 Gait Posture 18 23
[3] Vaughan C, Davis B, and O'Connor J 1999 Dynamics of Human Gait, second ed., Kiboho, Cape
Town
[4] Yang N F, Wang R C, Jin D w, Dong H, Huang C H, Zhang M 2001 Chinese Journal of
Rehabilitation Medicine 16 336 (in Chinese)
[5] Kondraske G 1994 Proceedings of
the 16th Annual International Conference of
the IEEE
Engineering in Medicine and Biology Society, Baltimore, USA 1 307
[6] Titianova E B and Tarkka I M 1995 J. Rehabil. Res. Dev. 32 236
[7] Liu Y B, Yan N and Yun X P 2000 Modern Rehabilitation 4 28 (in Chinese)
[8] Biswasa A, Lemaire E and Kofman J 2008 J. Biomech. 41 1574
[9] Wall J C and Turnbull G I 1986 Arch. Phys. Med. Rehabil. 67 550
[10] Liu Y B, and Yan N 1996 Chinese Journal of Rehabilitation Theory and Pactice 2 154 (in Chinese)
[11] Sadeghi H, Allard P, Prince F and Labelle H 2000 Gait Posture 12, 34
[12] Wang R C, Zhang M Q, Hua C, Deng X N, Yang N H and Jin D W2005 J. Tsinghua Univ. 45,
190 (in Chinese)
[13] Yang Y Y, Wang R C, Hao Z X, Jin D W, Zhang H 2005 J. Biomed. Eng. 45 1100 (in Chinese)
[14] Kokshenev V B 2004 Physical Review Letters 93 208101
[15] Gutierrez-Farewik E M, Bartonek A and Saraste H 2006 Hum. Movment Sci. 25 238
[16] Winiarski S and Rutkowska-Kucharska A 2009 Acta of Bioengineering and Biomechanics 11 53
[17] Murray M 1967 Am. J. Phys. Med. 46 290
[18] Fan Y F, Loan M, Fan Y B, Li Z Y and Luo D L 2009 Europhys. Lett. 87 58003
[19] Gard S A, Miff S C and Kuo A D 2004 Hum. Movment Sci. 22 597
[20] Fan Y F, Li Z Y and Lv C S 2008 Proc. SPIE 7128 71280I
[21] Ren L, Jones R K and Howard D 2008 J. Biomech. 41 2750
Dynamic Principles of COM
16
[22] Anderson F C and Pandy M G 2001 J. Biomech. Eng-T ASME 123 381
[23] Chau T 2001 Gait Posture 13 49
[24] Marion J 1970 Classical Dynamics of Particles and System (Academic, New York)
[25] Jenkins F A 1972 Science 178 877
[26] Srinivasan M and Ruina A 2006 Nature 439 72
[27] Kfc 2009 Technology Review http://www.technologyreview.com/blog/arxiv/23569
[28] Richmond B G and Jungers W L 2008 Science 319 1662
[29] LIU D W 2007 Science Technology and Engineering 7 2319 (in Chinese)
[30] Boulgouris N V, Hatzinakos D and Plataniotis K N 2005 IEEE Signal Proc. Mag. 22 78
[31] Nixon M S and Carter J N P. IEEE 94 2013
[32] Collins S H, Wisse M and Ruina A 2001 Int. J. Robot Res. 20 607
[33] Collins S, Ruina A, Tedrake R and Wisse M 2005 Science 307 1082
[34] Ohgane K and Ueda K I 2008 Phys. Rev. E 77 051915
|
1702.03997 | 1 | 1702 | 2017-02-13T22:18:20 | The Plasma Membrane is Compartmentalized by a Self-Similar Cortical Actin Meshwork | [
"physics.bio-ph",
"cond-mat.stat-mech"
] | A broad range of membrane proteins display anomalous diffusion on the cell surface. Different methods provide evidence for obstructed subdiffusion and diffusion on a fractal space, but the underlying structure inducing anomalous diffusion has never been visualized due to experimental challenges. We addressed this problem by imaging the cortical actin at high resolution while simultaneously tracking individual membrane proteins in live mammalian cells. Our data confirm that actin introduces barriers leading to compartmentalization of the plasma membrane and that membrane proteins are transiently confined within actin fences. Furthermore, superresolution imaging shows that the cortical actin is organized into a self-similar meshwork. These results present a hierarchical nanoscale picture of the plasma membrane. | physics.bio-ph | physics |
APS/123-QED
The Plasma Membrane is Compartmentalized by a Self-Similar Cortical Actin
Meshwork
Sanaz Sadegh1, Jenny L. Higgins2, Patrick C. Mannion2, Michael M. Tamkun3,4, and Diego Krapf1,2
1Department of Electrical and Computer Engineering,
Colorado State University, Fort Collins, Colorado 80523, USA
2 School of Biomedical Engineering, Colorado State University, Fort Collins, Colorado 80523, USA
3 Department of Biomedical Sciences, Colorado State University, Fort Collins, Colorado 80523, USA
4 Department of Biochemistry and Molecular Biology,
Colorado State University, Fort Collins, Colorado 80523, USA.
(Dated: September 1, 2018)
A broad range of membrane proteins display anomalous diffusion on the cell surface. Different methods
provide evidence for obstructed subdiffusion and diffusion on a fractal space, but the underlying structure
inducing anomalous diffusion has never been visualized due to experimental challenges. We addressed this
problem by imaging the cortical actin at high resolution while simultaneously tracking individual membrane
proteins in live mammalian cells. Our data confirm that actin introduces barriers leading to compartmen-
talization of the plasma membrane and that membrane proteins are transiently confined within actin fences.
Furthermore, superresolution imaging shows that the cortical actin is organized into a self-similar meshwork.
These results present a hierarchical nanoscale picture of the plasma membrane.
PACS numbers: 87.15.K-, 87.15.Vv
I.
INTRODUCTION
The plasma membrane is a complex fluid where
lipids and proteins continuously interact and generate
signaling platforms in order to communicate with the
outside world. One of the key mechanisms by which
membrane molecules search reaction sites is based on
lateral diffusion. Quantitative imaging methods, such
as single-particle tracking [1 -- 4], spatiotemporal image
correlation spectroscopy [5], fluorescence correlation
spectroscopy (FCS) [6, 7], and STED-FCS [8, 9], show
that the dynamics of proteins and lipids in the plasma
membrane often deviate from normal diffusion. In par-
ticular, the mean square displacement (MSD) does not
grow linearly in time as expected for Brownian motion
[10 -- 13]. This behavior suggests processes that hinder
diffusion. Since the formation of protein complexes is
governed by diffusion-mediated encounters, hindered
diffusion plays fundamental roles in cell function.
Unveiling the underlying mechanisms leading to the
observed anomalous diffusion on the cell membrane
is critical to understanding cell behavior. Anomalous
diffusion in the plasma membrane can be caused by
macromolecular crowding [14], transient binding [15],
heterogeneities [16, 17], and membrane compartmental-
ization by the underlying cytoskeleton [2, 9, 18, 19]. In
recent years it has become evident that a single mech-
anism cannot account for the complex dynamics ob-
served in the plasma membrane [13]. We have shown
that interactions with clathrin coated pits (CCPs) cause
anomalous diffusion and ergodicity breaking [15, 20].
However, it was observed that this process coexisted
with a different anomalous diffusion mechanism at-
tributed to diffusion within a fractal topology. Ex-
perimental evidence for the organization of the plasma
membrane by the cortical actin cytoskeleton has been
provided by measurements in cell blebs, spherical
protrusions that lack actin cytoskeleton [21], and in
the presence of actin-disrupting agents [9, 22, 23].
The picket-fence model explains these observations
by postulating that the mobility of membrane-bound
molecules is hindered by the actin-based cytoskeleton
in close proximity to the plasma membrane, leading to
transient confinement [2, 24, 25]. Confinement and seg-
regation of membrane components can have important
physiological consequences by allowing the formation
of functional domains on the cell surface. However, in
spite of the vast evidence that has accumulated over the
last two decades, a direct observation of the dynamic
compartmentalization of membrane proteins by under-
lying actin fences is challenging due to the spatial and
temporal resolutions required for its visualization.
Here we employ superresolution imaging and single-
particle tracking of membrane proteins to elucidate
the compartmentalization of the plasma membrane
by intracellular structures. While tracking individual
voltage-gated potassium channels as described in our
previous studies [15], we found that these membrane
proteins exhibited anomalous diffusion on the cell sur-
face. We now report that the anticorrelated dynam-
ics are best modeled by obstructed diffusion instead
of fractional Brownian motion and we directly visu-
alize the transient confinement of potassium channels
by cortical actin in live cells.
In order to character-
ize the cortical actin meshwork, we employ stochastic
optical reconstruction microscopy (STORM) to obtain
superresolution images in fixed cells. We find a non-
integer fractal dimension for the actin cortex and a
broad distribution of compartment sizes as expected
for a self-similar structure. These observations con-
sistently explain the anticorrelated subdiffusive motion
of membrane proteins and provide new insights on the
hierarchical organization of the plasma membrane.
II. RESULTS
A. Kv1.4 and Kv2.1 ion channels undergo subdiffusion
in the plasma membrane
Voltage-gated potassium channels Kv1.4 and Kv2.1
were expressed in human embryonic kidney (HEK)
cells, labeled with quantum dots (QDs) [15], and im-
aged using total internal reflection fluorescence (TIRF)
microscopy at 50 frames/s, so that individual molecules
could be detected on the cell surface. Kv1.4 and 2.1 are
similar in size, 654 and 853 amino acids, respectively,
but share less than 20% overall amino acid identity [26].
They are placed into distinct gene subfamilies because
of this low identity. They are most similar within a
central core domain composed of six transmembrane
alpha helices and the ion conducting pore. In contrast,
they share no amino sequence identity within the cyto-
plasmic N- and C-terminal regions; each Kv1.4 subunit
has 402 cytoplasmic amino acids while the Kv2.1 sub-
units have 624. Both channels exist as homotetrameric
structures giving the functional channel 24 membrane
spanning domains and a total of either 1608 or 2496
cytoplasmic amino acids. Figure 1(a) shows representa-
tive trajectories of Kv1.4 channels. The motion of the
ion channels was initially evaluated in terms of their
time-averaged MSD,
δ2(∆) =
1
T − ∆ Z T −∆
0
r(t + ∆) − r(t)2dt,
(1)
i.e., δ2(∆) ∼ ∆.
where T is the total experimental time, r the parti-
cle position, and ∆ the lag time, i.e., the time differ-
ence over which the MSD is computed. When a parti-
cle displays Brownian diffusion, the MSD is linear in
In contrast, anomalous
lag time,
diffusion is characterized by a different MSD scaling,
namely MSD ∼ ∆α, where α is the anomalous expo-
nent. Anomalous diffusion is classified as subdiffusion
when 0 < α < 1 and superdiffusion when α > 1.
Figure 1(b) shows the MSD of 20 individual trajecto-
ries. The MSDs of Kv1.4 as well as Kv2.1 channels show
subdiffusive behavior, albeit with large apparent fluctu-
ations. Figures 1(c) and 1(d) show the MSDs averaged
over 1,312 Kv1.4 (n = 10 cells) and 6,385 Kv2.1 (n = 14
cells) trajectories, respectively, Dδ2(∆)E. Throughout
the manuscript we employ overlines to denote time av-
erages and brackets to denote ensemble averages. The
anomalous exponent α of Kv1.4 was found to be 0.89
2
and that of Kv2.1 was 0.74, indicating subdiffusion in
both cases.
Several distinct mathematical models lead to subd-
iffusion [11 -- 13]. Among the most well-accepted types
of subdiffusion in biological systems, we encounter (i)
obstructed diffusion,
(ii) fractional Brownian motion
(fBM), and (iii) continuous time random walks (CTRW).
Both fBM [27, 28] and obstructed diffusion [29 -- 31] are
models for subdiffusive random walks with anticorre-
lated increments that have been extensively used in
live cells.
fBM describes the motion in a viscoelastic
fluid [32, 33], which can be caused by macromolecular
crowding [34, 35]. fBM is a generalization of Brownian
motion that incorporates correlations with power-law
It is characterized by a Hurst exponent H
memory.
that translates into an anomalous exponent α = 2H.
Obstructed diffusion describes the motion of a particle
hindered by immobile (or slowly moving) obstacles, e.g.,
percolation. As the concentration of immobile obsta-
cles increases, the availability of space decreases. Near
a critical concentration known as percolation threshold,
the obstacles form a fractal with dead ends in all length
scales.
In particular, the reduction of the available
space results in anomalous diffusion with a recurrent
exploration pattern. A CTRW is a generalization of a
random walk where a particle waits for a random time
between steps [36]. When the waiting times are asymp-
totically distributed according to a power law such that
the mean waiting time diverges, the CTRW is subdiffu-
sive. These three models describe very distinct physical
underlying mechanisms but they can yield similar sub-
linear MSD scaling, particularly in obstructed diffusion
and fBM models. Thus the MSD analysis is insufficient
to elucidate the type of random walk.
Different tests beyond the MSD have been employed
to distinguish among types of subdiffusive random
walks, including p-variations [37], first passage prob-
ability distribution [38], mean maximal excursion [39],
Gaussianity [40], and fractal dimensions [41]. Here we
employ the distribution of directional changes, i.e., the
turning angles, a tool that probes correlations in the
particle displacements and has been shown to contain
information on the complexity of a random walk [42].
Figure 1(e) illustrates the construction of turning angles
from a particle trajectory. In simple Brownian motion,
the turning angles are uniformly distributed. Contrast-
ingly, when the steps are correlated the distribution of
turning angles is not uniform [42]. Figures 1(f) and 1(g)
show the distribution of turning angles of Kv1.4 and
Kv2.1 for different lag times (1,312 Kv1.4 tracks, 10 cells
and 6,385 Kv2.1 tracks, 14 cells). Both distributions
peak at θ = 180◦ indicating the particles are more
likely to turn back than to move forward.
In other
words, Kv channels have a preference to go in the di-
rection from where they came rather than to persist
moving in the same direction. This property is a fin-
(d)
Kv1.4
~ t
0.1
~ t 0.89
0.01
0.1
Lag time, ∆ (s)
1
(g)
0.012
0.008
0.004
0.02 s
0.08 s
0.2 s
0.4 s
1 s
)
2
m
μ
(
D
S
M
-
A
T
-
A
E
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
P
(a)
(b)
(c)
)
2
m
μ
(
D
S
M
-
A
T
1
Kv1.4
0.1
0.01
0.1
Lag time, ∆ (s)
1
(f )
150 frames
180 frames
(3 s)
(3.6 s)
(e)
1 μm
θ
2
θ
1
220 frames
250 frames
(4.4 s)
(5 s)
(h)
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
P
(i)
fractional BM
0.010
0.005
H = 0.3
H = 0.4
0
0
45 90 135 180
Turning angle
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
P
(j)
percolation
0.010
0.005
c = 41 %
c = 33 %
0
0 45 90 135 180
Turning angle
0
0
Kv1.4
45 90 135 180
Turning angle
(k)
)
2
m
μ
(
D
S
M
-
A
T
-
A
E
1
ΔC318
~ t
0.1
0.01
~ t
0.1
1
Lag time, ∆ (s)
3
)
2
m
μ
(
D
S
M
-
A
T
-
A
E
0.1
Kv2.1
~ t
0.01
~ t 0.74
0.1
Lag time, ∆ (s)
1
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
P
0.012
0.008
0.004
0
0
0.04 s
0.08 s
0.2 s
0.4 s
1 s
Kv2.1
45 90 135 180
Turning angle
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
P
0.012
0.008
0.004
Kv2.1
ΔC318
0
0
45 90 135 180
Turning angle
(a) Four Kv1.4
(b) Time-averaged MSD (TA-MSD) as a function of lag time
(c) Ensemble-averaged time-averaged MSD (EA-TA-MSD) averaged over 1,312 Kv1.4
(d) EA-TA-MSD averaged over 6,385 Kv2.1 trajectories (n = 14 cells). The dashed lines in c and
FIG. 1. Voltage-gated potassium channels Kv1.4 and Kv2.1 undergo subdiffusion in the plasma membrane.
representative trajectories obtained by single-particle tracking.
∆ for 20 individual Kv1.4 trajectories.
trajectories (n = 10 cells).
d are visual guides for linear behavior (free diffusion), i.e., Dδ2(∆)E ∼ ∆. Error bars show standard deviation.
illustrating the construction of turning angles from a particle trajectory.
(f)-(g) Turning angle distributions for Kv1.4 (10 cells,
1,312 trajectories) and Kv2.1 (14 cells, 6,385 trajectories). Turning angle distributions are constructed for lag times between 20
ms and 1 s.
(i)
Turning angle distribution for simulations of obstructed diffusion with obstacle concentrations 33% and 41%, i.e., site percolation.
(j) MSD averaged over 3,114 ∆C318 trajectories (n = 5 cells). (k) Turning angle distributions for Kv2.1 and ∆C318 (5 cells, 3,114
trajectories) measured with lag time of 200 ms.
(h) Turning angle distributions for fractional Brownian motion simulations with Hurst exponents 0.3 and 0.4.
(e) Sketch
gerprint of subdiffusive random walks with anticorre-
lated increments. Besides the shape of the distribution,
the dependence on lag time bears valuable information.
Strikingly, we observe that the distribution is indepen-
dent of lag time, i.e., we measure the same distribution
of directional changes whether the lag time is 20 ms or
1 s.
We examined numerical simulations of fBM and ob-
structed diffusion and found that they have distinctive
attributes in their distribution of directional changes.
Figure 1(h) shows the distribution of directional changes
for subdiffusive fBM simulations with Hurst exponents
H = 0.3 and 0.4. Even though the distributions peak at
180◦, the probability density function is different from
the experimental data [Figs. 1(f) and 1(g)]. In our exper-
imental data, the turning angle distributions increase
sharply as θ approaches 180◦ and most of the devia-
tions from a uniform distribution are above 90◦. How-
ever, fBM gives rise to a gradual increase that takes
place mainly in the range 45◦ < θ < 135◦. Further,
the turning angles of fBM reach a plateau, in contrast
to our measurements. Conversely, obstructed diffusion,
strongly resembles our experimental results. Figure 1(i)
shows the turning angle distribution for obstructed dif-
fusion simulations in a square lattice with obstacle con-
centrations 33% and 41% [31]. Note that 41% is slightly
above the percolation threshold. These results show
that the motion of Kv channels in the plasma mem-
brane is better modeled by percolation, i.e., obstructed
diffusion, rather than motion in a viscoelastic medium,
i.e., fBM.
Potential obstacle candidates for obstructed diffusion
in the plasma membrane are the cortical cytoskeleton,
lipid rafts, and extracellular glycans. By evaluating the
MSD and turning angle distribution of ∆C318, a mu-
tant in which the last 318 amino acids of the C-terminus
of Kv2.1 channel had been deleted [43], we found that
the anticorrelated diffusion originates from interactions
with intracellular structures. We observed that ∆C318
channels diffuse freely in the plasma membrane, α = 1
with a diffusion coefficient D = 0.19 µm2/s [Fig. 1(j),
n=3114 tracks, 5 cells]. Further, the distribution of turn-
ing angles of ∆C318 was flattened, as expected for
Brownian diffusion [Fig. 1(k)], indicating the intracel-
lular C terminal domain of Kv2.1 plays a key role in
the anticorrelations within the particle trajectory. Even
though the distribution of turning angles in the ∆C318
mutant is close to that in Brownian motion, a small
peak is still noticeable at 180◦ suggesting additional
complexities in the plasma membrane.
In contrast to Kv1.4, which is homogeneously dis-
tributed on the cell membrane, a subpopulation of
Kv2.1 channels forms micron-sized clusters that local-
ize to endoplasmic reticulum (ER)-plasma membrane
junctions [44, 45]. Thus, we expect both the ER and
the cortical cytoskeleton introduce intracellular interac-
tions with Kv2.1 channels. To identify the origin of the
observed anticorrelated diffusion, we analyzed the mo-
tion of non-clustered Kv2.1 channels, i.e., the channels
that reside outside ER-plasma membrane junctions. We
labeled Kv2.1 channels both with green fluorescent pro-
tein (GFP) and QDs [45]. While all the channels were
labeled with GFP, only a small fraction included QDs
in order to enable both single-particle tracking and
cluster identification [supplementary Fig. S1]. We ob-
served that the distribution of directional changes of
non-clustered channels is indistinguishable from that of
the overall population [supplementary Fig. S2]. Thus,
we can exclude interactions with the ER as the cause for
anticorrelated subdiffusion. These observations suggest
that diffusion is hindered by intracellular components,
possibly the cortical cytoskeleton, in agreement with a
membrane-skeleton fence model [2].
We observed that the distribution of turning angles
were independent of lag times [Figs. 1(f) and 1(g)] within
the probed spatial and temporal scales. These obser-
vations indicated the anticorrelated subdiffusion of Kv
channels did not have an evident characteristic time
scale. This type of random walk is consistent with diffu-
sion on a self-similar structure, i.e., a fractal subspace.
In order to visualize the difference between diffusion on
a fractal structure, and diffusion on a meshwork with a
characteristic length scale, we performed simulations of
motion of a particle in the presence of permeable fences
that introduce compartments with a well-defined length
scale [supplementary Fig. S3]. In these simulations, we
observed that the distribution of turning angles is not
time invariant; the peak at 180◦ grows as we increase
the lag-time up to a characteristic time, and then it de-
cays when the lag-time increases further [supplemen-
tary Figs. S3(b) and S3(c)]. Thus, hop-diffusion with
a narrow distribution of confinement sizes exhibits a
time-dependent turning angle distribution (with a well-
defined characteristic time scale), in contrast to our ex-
perimental results where the turning angle distribution
is time-invariant.
4
B. Cortical actin transiently confines Kv1.4 and Kv2.1
channels
We observed that Kv channels undergo obstructed
diffusion. The ∆C318 mutant data indicated that hin-
dering of the particle motion originated within cyto-
plasmic structures in close proximity to the plasma
membrane, in agreement with previous experimental
evidence of transient confinement by the actin-based
cytoskeleton [9, 21, 46 -- 48]. Thus we examined the cor-
tical actin as a candidate for the observed obstructed
diffusion in the plasma membrane.
We imaged the cortical actin in live HEK cells us-
ing the photoactivatable probe tdEosFP [49] via an
actin binding peptide (ABP) that reversibly binds to F-
actin [50]. Previous studies showed that expression of
ABP-tdEosFP does not affect the organization of the
cytoskeleton [50, 51]. By activating a sparse subset
of tdEosFP and individually localizing them with high
precision, we generated photoactivated localization mi-
croscopy (PALM) images using localizations from 100
frames (2 s), yielding a smooth video of the dynamic
actin meshwork (supplementary Video S1). Both exci-
tation and photoactivation were implemented in total
internal reflection fluorescence (TIRF) so that only the
actin adjacent to the plasma membrane was imaged.
The dissociation of ABP-tdEosFP occurs with a time
constant on the order of 40 s [50], thus the exchange
within 2-s imaging is negligible.
Figure 2(a) shows a representative PALM reconstruc-
tion of actin. Although the number of localizations in
100 frames is not adequate to fully resolve the cortical
actin and some faint fluorescent single-filament struc-
tures might be missed in the images, we could use the
reconstructed PALM image to study the interactions of
the potassium channels with the actin cortex in live
cells. Previous breakthrough experiments have reported
simultaneous imaging of cortical actin cytoskeleton and
single-particle tracking [48, 52, 53]. Here, to the best of
our knowledge, we perform for the first time simultane-
ous single-particle tracking measurements and imaging
cortical actin with superresolution.
In order to find out whether actin-delimited domains
as identified by PALM hinder diffusion and compart-
mentalize the cell surface, we imaged and tracked Kv1.4
and Kv2.1 channels on the cell surface while simul-
taneously imaging the cortical actin. Channels often
remained confined within the areas enclosed by actin
indicating actin acted as a barrier to channel diffu-
sion. Figures 2(a) and 2(b) show Kv2.1 channel tracks
for one video overlaid on the last reconstructed im-
age of the cortical actin. However, this visualization
method suffers from overlaying long trajectories on a
single reconstruction image of the actin meshwork. In
addition to being constrained by actin structures, some
trajectories exhibit confinement within small nanoscale
(a)
(c)
)
2
m
μ
(
2
r
>
<
Kv1.4
0.04
0.02
(d)
)
2
m
μ
(
2
r
>
<
Kv2.1
0.02
0.01
(b)
5
(e)
)
2
m
μ
(
2
r
>
<
0.03
0.02
0.01
0
0
ΔC318
100 200 300 400
d (nm)
0
0
100 200 300 400
d (nm)
0
0
100 200 300 400
d (nm)
FIG. 2. Cortical actin transiently confines Kv channels.
(a) Trajectories of individual Kv2.1 channels (shown in cyan) overlaid
on actin PALM image (shown in red). Scale bar 2 µm. (b) Enlargements of the areas indicated with yellow arrows in a. Scale
bar is 500 nm. The left trajectory shows confinement in a large compartment, the middle one shows hoping between two
compartments and the right one shows confinement in a nanoscale domain. (c)-(e) Mean square displacements hr2i covered by
Kv1.4 and Kv2.1 and ∆C318 channels in 200 ms as a function of their maximum distance from nearest actin feature. Error bars
indicate standard errors.
domains that do not appear to be enclosed by actin.
We have previously shown that Kv channels exhibit fre-
quent immobilizations when the channels are captured
within clathrin-coated pits [20]. Thus, the cortical actin
cytoskeleton is not the sole mechanism by which the
mobility of Kv channels is hindered. In order to deal
with these complexities, we evaluated the MSDs as a
function of proximity to actin.
Given that actin hinders channel motility, we ex-
pect the particles to explore smaller areas when they
are confined within smaller compartments. To test the
actin fence hypothesis, we overlaid channel trajectories
on the corresponding PALM image of actin obtained in
2 s, a time scale in which the actin structure is fairly
persistent [supplementary video 1], with a sliding time
window 0.2 s. For example, trajectories of the chan-
nels from 0 to 2 s were overlaid on the first recon-
structed actin frame and the trajectories in the interval
200 ms to 2.2 s were overlaid on the second actin PALM
frame. Then we partitioned the trajectories into 200-
ms intervals and classified each segment according to
the maximum distance d of the particle to the nearest
actin feature, calculated using an Euclidean distance
map algorithm. We evaluated the ensemble-averaged
MSD hr2i of all the segments located at a specific dis-
tance away from actin, i.e., we averaged the squared
displacements in 200 ms of the particles transiently lo-
cated a given distance from actin. Figures 2(c) and 2(d)
show the MSD as a function of distance-to-actin for
Kv1.4 and Kv2.1. For both channels we observed that as
molecules dwell closer to actin their MSD decreases.
As a control of our method, we performed the same
analysis for ∆C318 channels. Because of the lack of the
intracellular domain, these channels should not have
any interaction with the cortical cytoskeleton. We ob-
served that the MSD of ∆C318 channels is independent
of distance from actin [Fig. 2(e)], which demonstrates
the effect of intracellular structure in the transient con-
finement of Kv channels.
C. Characterization of cortical actin meshwork
When imaging live cells with PALM, the number of
frames used in the reconstruction is restricted by the
cell dynamic nature. A low number of frames results in
insufficient detected particles to accurately determine
the structure, also having a deleterious effect on res-
olution.
In contrast, very high spatial resolution can
be obtained in fixed cells by collecting data over long
times [54 -- 56]. Therefore, we used TIRF-STORM to
visualize the compartments formed by cortical actin
in fixed cells. Actin was labeled with phalloidin con-
jugated to Alexa Fluor 647, which binds actin fila-
ments with high specificity without significantly enlarg-
ing them [56]. A total of 50,000 frames where used in
the reconstruction. In our STORM reconstructions we
observed both thick and thin actin structures, Fig 3(a).
The finest structures that we observed had a cross sec-
tion standard deviation of 20 nm (FWHM=48 nm). Fig-
(a)
(c)
6
60
(b)
80
60
40
20
t
n
u
o
C
0
-60 - 30
0
30
Position (nm)
(d)
(e)
300
t
n
u
o
C
200
100
0
0.01
0.1
Area (μm2)
1
FIG. 3. Characterizations of actin compartments. (a) Superresolution STORM image of the cortical actin in a HEK cell. The
inset shows the conventional TIRF image. Scale bar is 2 µm.
(b) Average cross-section profile of 20 filaments aligned by the
center of each line. The red line is Gaussian fit with standard deviation σ = 20 nm. (c) Watershed segmentation (shown in green)
of the boxed area overlaid on the STORM image.
(d) Compartments determined by watershed are designated with different
colors. Scale bars in c and d are 1 µm. (e) Distribution of compartment areas for fixed cells (9 cells, n = 2,500 compartments).
Areas are shown in logarithmic scale and the red line is a log-normal distribution with shape parameter σ =0.8 µm.
ure 3(b) shows the average cross section profile of 20
lines aligned by the center of each line. The thickness
of these lines in the reconstruction is governed by the
localization accuracy, 20 ± 8 nm (mean ± SD, supple-
mentary Fig. S4), which sets a lower bound on STORM
resolution. Thus we are unable to determine whether
these structures are individual filaments (10 nm in di-
ameter) or actin bundles.
We employed a watershed segmentation algorithm
[57] to identify actin-delimited compartments in the
STORM reconstructions [Figs. 3(c) and 3(d)] across the
whole cell. The average percentage of the watershed
meshwork covered by actin was 84 ± 4% (mean ± SD,
n=9 cells). Figure 3(e) shows the distribution of com-
partment areas (n = 2, 500 compartments). The areas
of the compartments are fitted well by a log-normal dis-
tribution, which is a subexponential heavy-tailed distri-
bution in the sense that it decays more slowly than any
exponential tail [58]. The log-normal distribution is
in good agreement with Kolmogorov's model for the
distribution of particle sizes after repeated breakage
[59, 60]. When a particle is divided into fragments
in such a way that the fragment proportions are in-
dependent of the original particle size, a log-normal
distribution emerges in the particle sizes after random
repeated fragmentation. Analogously, actin-delimited
compartments are split into smaller compartments by
growing actin filaments and thus their distribution is
predicted to be log-normal.
In addition to the compartment area distribution, the
relation between perimeter and area contains valuable
(a)
)
m
μ
(
r
e
t
e
m
i
r
e
P
10
1
0.1
0.01
0.1
Area (μm2)
1
(b)
s
t
n
u
o
C
106
104
102
1
~ε-1.75
10
1
Box Size, ε (pixels)
100 1000
(a) Log-
FIG. 4. Fractality of the cortical actin meshwork.
log scatter plot of compartment perimeter vs. compartment
area. The fitted line corresponds to L = 4.8A0.55 (Pearson
correlation coefficient ρ = 0.98 in log scales). (b) Representa-
tive example of box counting algorithm in one cell where the
exponent yields df = 1.75.
information. Perimeter-area relations have been exten-
sively used to investigate the properties of complex pla-
nar shapes [61, 62]. As expected, we observe that ar-
eas and perimeters of the different compartments are
highly correlated [Fig.4(a)]. This correlation indicates
shape homogeneity among different compartments [63].
Furthermore, the area exhibits the same scaling over
the whole observed range, A ∼ L1.8, where A and
L are compartment area and perimeter, indicating the
same statistical character at different scales, and sug-
gesting that compartments formed by cortical actin are
scale-invariant in the observed range. Such scale in-
variance is a hallmark of a self-similar fractal struc-
tures.
Fractals are characterized by scaling properties gov-
erned by a non-integer dimension df , i.e., an anoma-
lous dependence of the "mass" on the linear size of the
system with M ∼ ldf .
In a regular object such as a
line, square, or cube, we would refer to its mass M as
the length, area or volume, respectively. In these reg-
ular cases the mass scales as M ∼ ld, where the l is
the linear size and d = 1, 2, 3 is the spatial dimension.
On the other hand, fractals such as a Sierpinsky gasket
or a percolation cluster differ from Euclidean spaces
and display a fractional dimension [61, 64]. Usually the
capacity dimension is obtained using a box-counting
algorithm that quantifies the mass scaling.
In brief,
the structure is placed on a grid, the number of occu-
pied "boxes" are counted, and the process is iterated
for finer grids. The number of occupied boxes scales as
N ∼ ǫ−df , where ǫ is the box length. Figure 4(b) shows
the computation of the fractal dimension of the cortical
actin meshwork from the STORM image in a represen-
tative cell. The box counting analysis shows the actin
structure exhibits statistical self-similarity over more
than three decades. Our data indicate the fractal di-
mension of the meshwork is df = 1.75 ± 0.02 (n = 9
cells).
III. DISCUSSION
Our current understanding of the plasma membrane
is that of a complex partitioned fluid where molecules
often undergo anomalous diffusion and can be segre-
gated according to their function. We observe that K+
channels perform a random walk with antipersistent
nature, i.e., a random walk with an increased probabil-
ity of returning to the site it just left. However elucidat-
ing the mechanisms that cause anomalous diffusion is
not trivial because several different subdiffusion mod-
els lead to similar MSD scaling. The analysis of K+
channel motion is further complicated by the occur-
rence of immobilizations with power law sojourn times,
which introduce deviations from Gaussian functions in
the distribution of displacements [15, 20]. Thus, we
cannot employ Gaussianity-based tests to distinguish
among complex antipersistent random walks. We find
that the distribution of directional changes provides a
robust test for the type of random walk. The measured
channel trajectories are shown to be well described by
obstructed diffusion but not by fBM.
We observed that the Kv2.1 intracellular domain
played a key role in the anomalous diffusion, in agree-
ment with previous observations showing that
the
depth at which a membrane protein extends into the
cytoplasm determined how frequently it encountered
mechanical barriers [47]. The obvious candidate to ob-
struct the motion of proteins with large intracellular do-
mains is the actin cytoskeleton. Thus, we visualized the
7
cortical actin with high temporal and spatial resolution
and evaluated its effect on membrane protein dynam-
ics. Considering that some faint single-filament actin
structures might not be accurately detected by PALM
imaging, we can miss some interactions between actin
and membrane proteins. Notwithstanding, we found
that Kv channels are transiently confined by perme-
able actin fences, confirming existing models for mem-
brane compartmentalization as an organizing principle
of the actin cytoskeleton [2]. By studying the diffusion
of Kv2.1 channels outside ER-plasma membrane junc-
tions, we verified that the observed subdiffusion is not
due to interactions with ER. Other intracellular compo-
nents such as intermediate filaments could also hinder
protein diffusion and further compartmentalize the cell
membrane, but these cytoskeletal filaments were not
studied in the present work. Previous single-particle
tracking works using lipids labeled with 40-nm gold
nanoparticles have observed the compartmentalization
of the plasma membrane of HEK293 cells, with 70-nm
mean compartment size [1]. However, this compartmen-
talization occurs with molecules having virtually no cy-
toplasmic domains and with a residence time close to
3 ms. At time scales above 50 ms, gold-labeled lipids
were found to exhibit Brownian diffusion with an effec-
tive diffusion coefficient D = 0.41 µm2/s [1], similar to
our observations for ∆C318 Kv2.1 mutant.
Ion channels are observed to undergo anticorrelated
anomalous diffusion over at least two orders of magni-
tude in time. In terms of percolation theory, this hints
the cell surface is maintained close to criticality, i.e.,
near the percolation threshold. However, this hypothe-
sis seems highly unlikely. A more feasible explanation
stems from the emergence of a scale-invariant structure
under the plasma membrane. We directly observed
that, within the probed spatial scale, the actin cortex
has in fact a self-similar nature. It is possible to spec-
ulate that actin fractality develops from its branching
structure. Hierarchically branched structures have a
fractal dimension df such that Rb = Rdf
r , where Rb is
the bifurcation ratio and Rr is the length-order ratio
[65]. The bifurcation ratio can be interpreted as the
average number of branches that emerge after a bifur-
cation and the length-order ratio is defined as the ra-
tio between incoming branch length and the length of
the emerging branches until the next bifurcation [66].
Actin branching is driven by the Arp2/3 complex [67]
with a bifurcation ratio Rb = 2. Here we measure
a meshwork fractal dimension df = 1.75, which can
arise from a branching pattern with Rr = 1.5 or, in
other words, the daughter branch being on average 1/3
shorter than the mother branch. The fractal dimension
of the cytoskeleton is in line with a broad range of frac-
tal geometries found in biology ranging from the lung
alveoli to subcellular structures such as mitochondrial
membranes and the endoplasmic reticulum [61].
We propose that the fractal nature of the actin cortex
is employed by the cell to organize the plasma mem-
brane. The complexity of this structure leads to a hier-
archical organization with domains in multiple length
scales and the development of nested compartments.
Such a dynamic hierarchical organization facilitates the
active segregation of domains with different functions
and the maintenance of reactants near reaction cen-
ters. Furthermore, the fractal nature of the cortical
actin has broad implications for anomalous diffusion,
for instance it could bridge the gap between the plasma
membrane hop diffusion models and diffusion in a frac-
tal that leads to anomalous dynamics over broad time
scales. We foresee that a self-similar cytoskeleton struc-
ture also influences active actomyosin-mediated orga-
nization of the plasma membrane [68] in such a way
that these processes can take place over multiple length
scales.
In conclusion, our findings show that the plasma
membrane is compartmentalized in a hierarchical fash-
ion by a dynamic cortical actin fractal. We find that
the anomalous diffusion of potassium channels is best
modeled by obstructed diffusion or diffusion in a frac-
tal. By combining PALM imaging with single-particle
tracking we were able to directly visualize the hindering
effect of cortical actin on the diffusion of the membrane
proteins. We characterized the compartments formed
by cortical actin using superresolution imaging in fixed
cells and found evidence for the self-similar topology
of this structure.
IV. MATERIALS AND METHODS
A. Cell transfection and labeling.
HEK 293 cells (passage 42-49; American Type Cul-
ture Collection) were cultured in phenol red Dulbecco's
Modified Eagle's Medium (DMEM), supplemented with
10% fetal bovine serum (FBS; Gibco) at 37 ◦C. Cells
were transfected to express a Kv2.1 or Kv1.4 construct
with an extracellular biotin acceptor domain that, when
coexpressed with a bacterial biotin ligase, results in bi-
otinylated Kv channels on the cell surface [69]. For
live-cell actin imaging cells were transfected with 3 µg
of ABP-tdEos. ABP is the actin-binding sequence of
ABP140 from S. cerevisiae consisting of 17 amino acids
[50, 51]. Kv2.1-loopBAD-GFP (3 µg) was employed to
identify Kv2.1 clusters on the cell surface.
For single-particle tracking, biotinylated channels
were labeled with QDs. Cells were incubated for 10
minutes in HEK imaging saline with 1 nM streptavidin-
conjugated QD705 or QD655 (Invitrogen) and 10
mg/mL bovine serum albumin (BSA) [15] at 37 ◦C. Fol-
lowing incubation the cells were rinsed again six times
with HEK imaging saline to ensure the removal of any
8
unbound QDs. The diameter of the QDs is in the range
10-15 nm, thus given that Kv channels are similar in
size to the QDs, it is highly unlikely that the Kv:QD
stoichiometry is higher than 1:1. We have previously
shown that QD conjugation does not alter Kv channel
diffusion [15, 20].
B. Live cell imaging.
Imaging was performed in an objective-based TIRF
microscope built around an IX71 Olympus body. Sam-
ple temperature was kept at 37 ◦C using objective and
stage heater (Bioptechs). A 405-nm laser was used to
activate tdEosFP fluorophore, while 473-nm and 561-
nm lasers were used to excite it in its inactive and active
states, respectively.
C. Single-particle tracking.
QD labeling was controlled so that QDs remained
at low density to allow for single-particle tracking
[15]. Particle detection and tracking were performed
in MATLAB using u-track [70].
D. Fixed cell imaging.
Cells were plated on Matrigel-coated 35 mm petri
dishes. After 12 hours the cells were fixed and labeled
with Alexa Fluor 647-phalloidin (Invitrogen).
Image
stacks were obtained using the same setup as live cell
imaging. A continuous illumination of 638-nm laser
was used to excite the Alexa Fluor 647. The laser
power before the objective was 30 mW. To maintain
an appropriate density of activated molecules, 405-nm
laser was used in some experiments. 50,000 frames
were collected to generate a superresolution image.
E. Fractal dimension.
The fractal dimension of the actin cortex was com-
puted using a box-counting algorithm. The thresholded
binary actin image was placed on a grid of square
boxes of size ǫ and the number of occupied boxes was
counted. The process was repeated for grids of boxes
with different sizes. The number of occupied boxes
scales as
N ∼ ǫ−df ,
(2)
where df is the capacity dimension, or simply the frac-
tal dimension.
F. Statistics
Results show mean and s.e.m. unless indicated oth-
erwise. All experimental results were obtained from
multiple different dishes and days. The number of dis-
tinct imaging regions, i.e., different cells is indicated in
the text.
ACKNOWLEDGMENTS
We thank Keith Lidke and Sheng Liu for the codes
for PALM reconstruction, Maxime Dahan for providing
9
the plasmids to express ABP-tdEos, and Aubrey Weigel,
Liz Akin, María Gracia Gervasi, Phil Fox, Ben John-
son, and Xinran Xu for useful discussions. We grate-
fully acknowledge the support of NVIDIA Corporation
with the donation of a GeForce GTX TITAN used for
this research. D.K. and J.L.H. thank Nikki Curthoys for
technical advice with superresolution microscopy dur-
ing the initial stages of the project. This work was sup-
ported by the National Science Foundation under Grant
1401432 (to DK) and the National Institutes of Health
under Grant R01GM109888 (to MMT).
[1] Kotono Murase, Takahiro Fujiwara, Yasuhiro Umemura,
Kenichi Suzuki, Ryota Iino, Hidetoshi Yamashita, Mi-
hoko Saito, Hideji Murakoshi, Ken Ritchie, and Akihiro
Kusumi, "Ultrafine membrane compartments for molec-
ular diffusion as revealed by single molecule techniques,"
Biophys. J. 86, 4075 -- 4093 (2004).
[2] Akihiro Kusumi, Chieko Nakada, Ken Ritchie, Kotono
Murase, Kenichi Suzuki, Hideji Murakoshi, Rinshi S Ka-
and Takahiro Fujiwara, "Paradigm
sai, Junko Kondo,
shift of
from the
two-dimensional continuum fluid to the partitioned
fluid: high-speed single-molecule tracking of membrane
molecules," Annu. Rev. Biophys. Biomol. Struct. 34, 351 --
378 (2005).
the plasma membrane concept
[3] Suliana Manley, Jennifer M Gillette, George H Patter-
and
son, Hari Shroff, Harald F Hess, Eric Betzig,
Jennifer Lippincott-Schwartz, "High-density mapping of
single-molecule trajectories with photoactivated localiza-
tion microscopy," Nat. Methods 5, 155 -- 157 (2008).
[4] Mathias P Clausen and B Christoffer Lagerholm, "Visu-
alization of plasma membrane compartmentalization by
high-speed quantum dot tracking," Nano Lett. 13, 2332 --
2337 (2013).
[5] Carmine Di Rienzo, Enrico Gratton, Fabio Beltram, and
Francesco Cardarelli, "Fast spatiotemporal correlation
spectroscopy to determine protein lateral diffusion laws
in live cell membranes," Proc. Natl. Acad. Sci. USA 110,
12307 -- 12312 (2013).
[6] Pierre-François Lenne, Laure Wawrezinieck, Fabien Con-
chonaud, Olivier Wurtz, Annie Boned, Xiao-Jun Guo,
Hervé Rigneault, Hai-Tao He,
and Didier Marguet,
"Dynamic molecular confinement in the plasma mem-
brane by microdomains and the cytoskeleton meshwork,"
EMBO J. 25, 3245 -- 3256 (2006).
[7] Verena Ruprecht, Stefan Wieser, Didier Marguet, and
Gerhard J Schütz, "Spot variation fluorescence correla-
tion spectroscopy allows for superresolution chronoscopy
of confinement times in membranes," Biophys. J. 100,
2839 -- 2845 (2011).
[8] Alf Honigmann, Veronika Mueller, Haisen Ta, Andreas
Schoenle, Erdinc Sezgin, Stefan W Hell, and Christian
Eggeling, "Scanning STED-FCS reveals spatiotemporal
heterogeneity of lipid interaction in the plasma mem-
brane of living cells," Nat. Comm. 5 (2014).
[9] Débora M Andrade, Mathias P Clausen,
Jan Keller,
Veronika Mueller, Congying Wu, James E Bear, Stefan W
Hell, B Christoffer Lagerholm, and Christian Eggeling,
"Cortical actin networks induce spatio-temporal confine-
ment of phospholipids in the plasma membrane -- a min-
imally invasive investigation by STED-FCS," Sci. Rep. 5
(2015).
[10] Eli Barkai, Yuval Garini, and Ralf Metzler, "Strange ki-
netics of single molecules in living cells," Physics Today
65, 29 -- 35 (2012).
[11] Felix Höfling and Thomas Franosch, "Anomalous trans-
port in the crowded world of biological cells," Rep. Prog.
Phys. 76, 046602 (2013).
[12] Ralf Metzler, Jae-Hyung Jeon, Andrey G Cherstvy, and
Eli Barkai, "Anomalous diffusion models and their prop-
erties: non-stationarity, non-ergodicity, and ageing at
the centenary of single particle tracking," Phys. Chem.
Chem. Phys. 16, 24128 -- 24164 (2014).
[13] Diego Krapf, "Mechanisms underlying anomalous diffu-
sion in the plasma membrane," Curr. Top. Membr. 75,
167 -- 207 (2015).
[14] James A Dix and AS Verkman, "Crowding effects on dif-
fusion in solutions and cells," Annu. Rev. Biophys. 37,
247 -- 263 (2008).
[15] Aubrey V Weigel, Blair Simon, Michael M Tamkun, and
Diego Krapf, "Ergodic and nonergodic processes coexist
in the plasma membrane as observed by single-molecule
tracking," Proc. Natl. Acad. Sci. USA 108, 6438 -- 6443
(2011).
[16] Timothy V Ratto and Marjorie L Longo, "Anomalous
subdiffusion in heterogeneous lipid bilayers," Langmuir
19, 1788 -- 1793 (2003).
[17] Carlo Manzo, Juan A Torreno-Pina, Pietro Massignan,
Gerald J Lapeyre Jr, Maciej Lewenstein, and Maria F
Parajo Garcia, "Weak ergodicity breaking of receptor
motion in living cells stemming from random diffusiv-
ity," Phys. Rev. X 5, 011021 (2015).
[18] Takahiro Fujiwara, Ken Ritchie, Hideji Murakoshi, Ken
Jacobson, and Akihiro Kusumi, "Phospholipids undergo
hop diffusion in compartmentalized cell membrane," J.
Cell Biol. 157, 1071 -- 1082 (2002).
[19] Thorsten Auth and Nir S Gov, "Diffusion in a fluid mem-
brane with a flexible cortical cytoskeleton," Biophys. J.
96, 818 -- 830 (2009).
[20] Aubrey V Weigel, Michael M Tamkun,
and Diego
Krapf, "Quantifying the dynamic interactions between
a clathrin-coated pit and cargo molecules," Proc. Natl.
Acad. Sci. USA 110, E4591 -- E4600 (2013).
[21] David W Tank, En-Shinn Wu,
and Watt W Webb,
"Enhanced molecular diffusibility in muscle membrane
blebs: release of lateral constraints." J. Cell Biol. 92, 207 --
212 (1982).
[22] Yael Lavi, Nir Gov, Michael Edidin, and Levi A Ghe-
ber, "Lifetime of major histocompatibility complex class-
I membrane clusters is controlled by the actin cytoskele-
ton," Biophys. J. 102, 1543 -- 1550 (2012).
[23] Manasa V Gudheti, Nikki M Curthoys, Travis J Gould,
Dahan Kim, Mudalige S Gunewardene, Kristin A Ga-
bor, Julie A Gosse, Carol H Kim, Joshua Zimmerberg,
and Samuel T Hess, "Actin mediates the nanoscale
membrane organization of the clustered membrane pro-
tein influenza hemagglutinin," Biophys. J. 104, 2182 -- 2192
(2013).
[24] MP Sheetz, "Membrane skeletal dynamics: role in mod-
ulation of red cell deformability, mobility of transmem-
brane proteins, and shape." in Seminars in Hematology,
Vol. 20 (1983) pp. 175 -- 188.
[25] Michael J Saxton and Ken Jacobson, "Single-particle
tracking: applications to membrane dynamics," Annu.
Rev. Biophys. Biomol. Struct. 26, 373 -- 399 (1997).
[26] Karen K Deal, Sarah K England,
and Michael M
Tamkun, "Molecular physiology of cardiac potassium
channels," Physiol. Rev. 76, 49 -- 67 (1996).
[27] Benoit B Mandelbrot and John W Van Ness, "Fractional
Brownian motions, fractional noises and applications,"
SIAM REV. 10, 422 -- 437 (1968).
[28] Weihua Deng and Eli Barkai, "Ergodic properties of
fractional Brownian-Langevin motion," Phys. Rev. E 79,
011112 (2009).
[29] Michael J Saxton, "Anomalous diffusion due to obstacles:
a Monte Carlo study." Biophys. J. 66, 394 -- 401 (1994).
[30] Daniel Ben-Avraham and Shlomo Havlin, Diffusion and
reactions in fractals and disordered systems (Cambridge
University Press, Cambridge, 2000).
[31] Aubrey V Weigel, Shankarachary Ragi, Michael L Reid,
Edwin KP Chong, Michael M Tamkun, and Diego Krapf,
"Obstructed diffusion propagator analysis for single-
particle tracking," Phys. Rev. E 85, 041924 (2012).
[32] Jedrzej Szymanski and Matthias Weiss, "Elucidating the
origin of anomalous diffusion in crowded fluids," Phys.
Rev. Lett. 103, 038102 (2009).
[33] Dominique Ernst, Marcel Hellmann, Jürgen Köhler, and
Matthias Weiss, "Fractional Brownian motion in crowded
fluids," Soft Matter 8, 4886 -- 4889 (2012).
[34] Gernot Guigas, Claudia Kalla,
and Matthias Weiss,
"The degree of macromolecular crowding in the cy-
toplasm and nucleoplasm of mammalian cells is con-
served," FEBS Lett. 581, 5094 -- 5098 (2007).
[35] Matthias Weiss, "Single-particle tracking data reveal an-
ticorrelated fractional Brownian motion in crowded flu-
ids," Phys. Rev. E 88, 010101 (2013).
10
[36] Harvey Scher and Elliott W Montroll, "Anomalous
transit-time dispersion in amorphous solids," Phys. Rev.
B: Solid State 12, 2455 (1975).
[37] Marcin Magdziarz, Aleksander Weron, Krzysztof Bur-
necki, and Joseph Klafter, "Fractional Brownian motion
versus the continuous-time random walk: A simple test
for subdiffusive dynamics," Phys. Rev. Lett. 103, 180602
(2009).
[38] Govindan Rangarajan and Mingzhou Ding, "First pas-
sage time distribution for anomalous diffusion," Phys.
Lett. A 273, 322 -- 330 (2000).
[39] Vincent Tejedor, Olivier Bénichou, Raphael Voituriez,
Ralf Jungmann, Friedrich Simmel, Christine Selhuber-
Unkel, Lene B Oddershede, and Ralf Metzler, "Quanti-
tative analysis of single particle trajectories: mean maxi-
mal excursion method," Biophys J. 98, 1364 -- 1372 (2010).
[40] Matthias Weiss, Markus Elsner, Fredrik Kartberg, and
Tommy Nilsson, "Anomalous subdiffusion is a measure
for cytoplasmic crowding in living cells," Biophys. J. 87,
3518 -- 3524 (2004).
[41] Yasmine Meroz, Igor M Sokolov,
and Joseph Klafter,
"Test for determining a subdiffusive model in ergodic
systems from single trajectories," Phys. Rev. Lett. 110,
090601 (2013).
[42] Stanislav Burov, SM Ali Tabei, Toan Huynh, Michael P
Murrell, Louis H Philipson, Stuart A Rice, Margaret L
Gardel, Norbert F Scherer, and Aaron R Dinner, "Dis-
tribution of directional change as a signature of complex
dynamics," Proc. Natl. Acad. Sci. USA 110, 19689 -- 19694
(2013).
[43] Antonius MJ VanDongen, Georges C Frech,
John A
Drewe, Rolf H Joho, and Arthur M Brown, "Alteration
and restoration of K+ channel function by deletions at
the N-and C-termini," Neuron 5, 433 -- 443 (1990).
[44] Philip D Fox, Christopher J Haberkorn, Aubrey V Weigel,
Jenny L Higgins, Elizabeth J Akin, Matthew J Kennedy,
Diego Krapf, and Michael M Tamkun, "Plasma mem-
brane domains enriched in cortical endoplasmic retic-
ulum function as membrane protein trafficking hubs,"
Mol. Biol. Cell 24, 2703 -- 2713 (2013).
[45] Philip D Fox, Christopher J Haberkorn, Elizabeth J Akin,
Peter J Seel, Diego Krapf,
and Michael M Tamkun,
"Induction of stable ER -- plasma-membrane junctions by
Kv2.1 potassium channels," J. Cell Sci. 128, 2096 -- 2105
(2015).
[46] Akihiko Tsuji and Shunichi Ohnishi, "Restriction of the
lateral motion of band 3 in the erythrocyte membrane
by the cytoskeletal network: dependence on spectrin as-
sociation state," Biochemistry 25, 6133 -- 6139 (1986).
[47] Michael Edidin, Martha C Zuniga,
and Michael P
Sheetz, "Truncation mutants define and locate cytoplas-
mic barriers to lateral mobility of membrane glycopro-
teins." Proc. Natl. Acad. Sci. U.S.A. 91, 3378 -- 3382 (1994).
[48] Nicholas L Andrews, Keith A Lidke, Janet R Pfeiffer,
Alan R Burns, Bridget S Wilson, Janet M Oliver, and
Diane S Lidke, "Actin restricts FcεRI diffusion and fa-
cilitates antigen-induced receptor immobilization," Nat.
Cell Biol. 10, 955 -- 963 (2008).
[49] Jörg Wiedenmann, Sergey Ivanchenko, Franz Oswald,
Florian Schmitt, Carlheinz Röcker, Anya Salih, Klaus-
Dieter Spindler, and G Ulrich Nienhaus, "EosFP, a flu-
orescent marker protein with UV-inducible green-to-red
fluorescence conversion," Proc. Natl. Acad. Sci. USA 101,
15905 -- 15910 (2004).
[50] Ignacio Izeddin, Christian G Specht, Mickaël Lelek,
Xavier Darzacq, Antoine Triller, Christophe Zimmer,
and Maxime Dahan, "Super-resolution dynamic imaging
of dendritic spines using a low-affinity photoconvertible
actin probe," PLoS ONE 6, e15611 (2011).
[51] Julia Riedl, Alvaro H Crevenna, Kai Kessenbrock,
Jerry Haochen Yu, Dorothee Neukirchen, Michal Bista,
Frank Bradke, Dieter Jenne, Tad A Holak, Zena Werb,
et al., "Lifeact: a versatile marker to visualize F-actin,"
Nat. Methods 5, 605 -- 607 (2008).
[52] Bebhinn Treanor, David Depoil, Aitor Gonzalez-Granja,
Patricia Barral, Michele Weber, Omer Dushek, An-
dreas Bruckbauer, and Facundo D Batista, "The mem-
brane skeleton controls diffusion dynamics and signal-
ing through the B cell receptor," Immunity 32, 187 -- 199
(2010).
[53] Juan A Torreno-Pina, Carlo Manzo, Mariolina
Salio, Michael C Aichinger, Anna Oddone, Melike
Lakadamyali, Dawn Shepherd, Gurdyal S Besra, Vin-
cenzo Cerundolo,
and Maria F Garcia-Parajo, "The
actin cytoskeleton modulates the activation of iNKT cells
by segregating CD1d nanoclusters on antigen-presenting
cells," Proc. Natl. Acad. Sci. USA 113, E772 -- E781 (2016).
[54] Michael J Rust, Mark Bates, and Xiaowei Zhuang, "Sub-
diffraction-limit imaging by stochastic optical recon-
struction microscopy (STORM)," Nat. Methods 3, 793 --
796 (2006).
[55] Graham T Dempsey, Joshua C Vaughan, Kok Hao Chen,
Mark Bates, and Xiaowei Zhuang, "Evaluation of fluo-
rophores for optimal performance in localization-based
super-resolution imaging," Nat. Methods 8, 1027 -- 1036
(2011).
[56] Ke Xu, Hazen P Babcock, and Xiaowei Zhuang, "Dual-
objective STORM reveals three-dimensional filament or-
ganization in the actin cytoskeleton," Nat. Methods 9,
185 -- 188 (2012).
[57] Fernand Meyer and Serge Beucher, "Morphological seg-
mentation," J. Vis. Commun. Image Represent. 1, 21 -- 46
(1990).
[58] Claudia Klüppelberg, "Subexponential distributions and
integrated tails," J. Appl. Probab. , 132 -- 141 (1988).
11
[59] Andrey N Kolmogorov, "On the logarithmically normal
law of distribution of the size of particles under pulveri-
sation," Dokl. Akad. Nauk SSSR 31, 99 -- 101 (1941).
[60] Benjamin Epstein, "The mathematical description of cer-
tain breakage mechanisms leading to the logarithmico-
normal distribution," J. Franklin Inst. 244, 471 -- 477 (1947).
[61] Benoit B Mandelbrot, The fractal geometry of nature, Vol.
173 (Macmillan, London, 1983).
[62] S Lovejoy, "Area-perimeter relation for rain and cloud
areas," Science 216, 185 -- 187 (1982).
[63] AR Imre and Jan Bogaert, "The fractal dimension as a
measure of the quality of habitats," Acta Biotheor. 52,
41 -- 56 (2004).
[64] Paul S Addison, Fractals and chaos: An illustrated course
(Institute of Physics Publishing, Bristol, UK, 1997).
[65] WI Newman, DL Turcotte, and AM Gabrielov, "Fractal
trees with side branching," Fractals 5, 603 -- 614 (1997).
[66] Robert E Horton, "Erosional development of streams
and their drainage basins; hydrophysical approach to
quantitative morphology," Geol. Soc. Am. Bull. 56, 275 --
370 (1945).
[67] Dominique Pantaloni, Rajaa Boujemaa, Dominique
Didry, Pierre Gounon, and Marie-France Carlier, "The
Arp2/3 complex branches filament barbed ends:
func-
tional antagonism with capping proteins," Nat. Cell Biol.
2, 385 -- 391 (2000).
[68] Kripa Gowrishankar, Subhasri Ghosh, Suvrajit Saha,
C Rumamol, Satyajit Mayor, and Madan Rao, "Active
remodeling of cortical actin regulates spatiotemporal or-
ganization of cell surface molecules," Cell 149, 1353 -- 1367
(2012).
[69] Kristen MS O'Connell, Annah S Rolig, Jennifer D White-
sell, and Michael M Tamkun, "Kv2.1 potassium channels
are retained within dynamic cell surface microdomains
that are defined by a perimeter fence," J. Neurosci. 26,
9609 -- 9618 (2006).
[70] Khuloud Jaqaman, Dinah Loerke, Marcel Mettlen, Hiro-
taka Kuwata, Sergio Grinstein, Sandra L Schmid, and
Gaudenz Danuser, "Robust single-particle tracking in
live-cell time-lapse sequences," Nat. Methods 5, 695 -- 702
(2008).
|
1712.07012 | 1 | 1712 | 2017-12-19T15:59:37 | Multiphasic profiles for the ion activities of NaBr and KBr | [
"physics.bio-ph"
] | The preceding paper (Nissen 2017) shows multiphasic profiles, straight lines separated by discontinuous transitions, for the ion activities of NaCl and KCl at 15, 25, 35 and 45oC based on very precise tabular data of Lee et al. (2002). In addition, references are given to previous findings of multiphasic profiles for a variety of processes and phenomena, starting with the uptake of sulfate by roots and leaf slices of barley (Nissen 1971).
In a continuation, the present paper shows multiphasic profiles for the ion activities of NaBr and KBr for data also from Lee et al. (2002). | physics.bio-ph | physics | Multiphasic profiles for the ion activities of NaBr and KBr
Per Nissen
Norwegian University of Life Sciences
Department of Ecology and Natural Resource Management
P. O. Box 5003, NO-1432 Ås, Norway
[email protected]
Introduction
2
The preceding paper (Nissen 2017) shows multiphasic profiles, straight lines separated by discontinuous
transitions, for the ion activities of NaCl and KCl at 15, 25, 35 and 45oC based on very precise tabular data
of Lee et al. (2002). In addition, references are given to previous findings of multiphasic profiles for a
variety of processes and phenomena, starting with the uptake of sulfate by roots and leaf slices of barley
(Nissen 1971).
In a continuation, the present paper shows multiphasic profiles for the ion activities of NaBr and KBr for
data also from Lee et al. (2002).
Reanalysis
Data for the activity coefficients of Na+ and Br− in aqueous solutions of various concentrations at
288.15, 298.15, 308.15 and 318.15 K (Tables 3 and 7 in Lee et al. 2002) have been plotted against each
other (Figs 1-8). Data for the activity coefficients of K+ and Br− (Tables 4 and 8) have also been plotted
(Figs 9-16).
As before (Nissen 2017), the profiles are very well represented by straight lines (54 of the 58 absolute r
values in Figs 1-16 are 0.999 or higher). The transitions are clearly discontinuous when the lines intersect in
a common point (between lines VIII and IX in Fig. 2, between lines V, VI, VII and VIII in Fig. 3, between
lines I and II in Figs 5 and 6, between lines II and III in Fig. 11). Transitions close to a point are probably
also discontinuous.
When adjacent lines are parallel or nearly so, the transition is necessarily in the form of a jump, as is the
case for the profile for Na+ (lines II and III, and lines VII and VIII) at the lowest temperature (Fig. 1). There
are similar findings also for Figs 2-4, showing that the phases are virtually unaffected by temperature, with
no or only minor changes in the pattern. The profiles for Br− decrease with increasing NaBr concentrations,
but there may be slight and temporary increases at intermediate and high concentrations.
The profiles for K+ decrease consistently with increasing concentrations (Figs 9-12). The profiles for Br-
decrease at first with increasing KBr concentrations, but then increase somewhat (Figs 13-16). Lines II-III
are parallel at all temperatures and are separated by a tiny jump.
The description of this reanalysis is somewhat less complete than in the previous paper (Nissen 2017), but
the figures and their legends provide compelling evidence that the profiles are indeed multiphasic and
cannot be validly represented as curvilinear. The present profiles for Na+ (Figs 1-4) are especially complex,
consisting as they do of four sections: A rapid initial decrease at low concentrations, an increase followed by
a decrease at intermediate concentrations, and a final rapid increase at high concentrations. However, the
data are still represented as curvilinear by the authors (Fig. 4, upper profile, in Lee et al. 2002) using a
modified Pitzer-Debye-Hückel model. This is made possible by their unmentioned omission of the four
circled points in Fig. 17. When these points are included the profile is clearly multiphasic.
3
Fig. 1. Nine phases. Transitions at 0.142, between 0.316 and
0.447 (jump), between 0.548 and 0.632 (jump), and between
0.707 and 0.775 (small jump), at 0.870, between 0.949 and
1.000 (jump), between 1.095 and 1.183 (small jump), and at
1.296. Lines II and III are parallel, as are approximately also
lines VII and VIII. The absolute r value of line I is very high,
that of line VIII is exceedingly high.
Fig. 2. Nine phases. Transitions at 0.155, between 0.316 and
0.447 (jump), between 0.548 and 0.632 (jump), and between
0.707 and 0.775 (small jump), at 0.889, between 0.949 and
1.000 (jump), between 1.095 and 1.183 (small jump), and at
1.265. Lines II and III are about parallel, as are also lines VII
and VIII. The absolute r values of lines I, VIII and IX are very
high.
Fig. 3. Eight phases. Transitions at 0.173, between 0.316 and
0.447 (jump), between 0.548 and 0.632, between 0.632
and 0.707, and at 0.894, 1.000 and 1.265. Insufficiently
detailed data for resolution of phase IV. The absolute r values
of lines I, V, VI, VII and VIII are high to very high.
Fig. 4. Nine phases. Transitions at 0.193, between 0.316 and
0.447 (jump), between 0.548 and 0.632, between 0.632 and
0.707, between 0.837 and 0.894 (jump), between 1.000 and
1.095, between 1.095 and 1.183, and between 1.342 and
1.414. Insufficiently detailed data for resolution of phases IV,
VII and IX. The absolute r values of lines I, V, VI and VIII
are very to exceedingly high.
4
Fig. 5. Eight phases. Transitions at 0.100, between 0.316 and
0.447 (jump), between 0.548 and 0.632, between 0.632 and
0.707, between 0.837 and 0.894 (small jump), between 1.000
and 1.095 small (jump), and at 1.310. Insufficiently detailed
data for resolution of phase IV. Lines VI and VII are about
parallel. The absolute r values of lines I, II, V, VI and VII are
very high.
Fig. 6. Nine phases. Transitions at 0.100, between 0.316 and
0.447 ( jump), between 0.548 and 0.632 (small jump), at
0.775,between 0.894 and 0.949, between 0.949 and 1.000,
and at 1.113 and 1.304. Insufficiently detailed data for
resolution of phase VI. Lines III, V and VII are parallel. The
absolute r values of lines I, II, IV and V are very high.
Fig. 7. Seven phases. Transitions between 0.316 and 0.447
(jump), between 0.548 and 0.632 (tiny jump), at 0.802 and
0.927, between 1.000 and 1.095 (jump), and at 1.302. Lines II
and III are parallel. The absolute r values for lines I, III and
VI are very high.
Fig. 8. Eight phases. Transitions between 0.316 and 0.447
(jump), between 0.548 and 0.632 (tiny jump), at 0.795 and
0.935, between 1.000 and 1.095 (jump), between 1.225 and
1.265 (tiny jump), and between 1.342 and 1.414. Lines II and
III are parallel. The absolute r values for lines I, III, IV and VI
are high to very high.
5
Fig. 9. Seven phases. Transitions between 0.100 and 0.224
(tiny jump), at 0.341, between 0.632 and 0.707 (tiny jump),
between 0.949 and 1.000 (tiny jump), between 1.095 and
1.183 (tiny jump), and between 1.225 and 1.342 (small jump).
Lines V and VI are about parallel. Very high absolute r values
for lines I, III and IV.
Fig. 10. Eight phases. Transitions between 0.100 and 0.224
(tiny jump), at 0.344, between 0.775 and 0.837, between
0.837 and 0.894, between 1.095 and 1.183, between 1.183
and 1.225, and between 1.342 and 1.414. Insufficiently
detailed data for resolution of phases IV, VI and VIII. Lines I
and II are parallel. Very high absolute r values for lines I, III,
V and VII.
Fig. 11. Seven phases. Transitions between 0.224 and 0.316
(jump), at 0.548, between 0.775 and 0.837, between 0.837 and
0.894, between 1.095 and 1.183 (tiny jump), and at 1.275.
Insufficiently detailed data for resolution of phase IV. Lines II
and III are about parallel. Very high absolute r values for lines
I, II, III and V.
Fig. 12. Seven phases. Transitions between 0.224 and 0.316
(jump), between 0.447 and 0.548 (tiny jump), between 0.775
and 0.837, between 0.837 and 0.894, between 1.095 and 1.183
(tiny jump), and at 1.323. Insufficiently detailed data for
resolution of phase IV. Lines V and VI are precisely parallel.
Very high absolute r values for lines I, III and V.
6
Fig. 13. Nine phases. Transitions at 0.132, between 0.316 and
0.447 (tiny jump), between 0.548 and 0.632, between 0.632
and 0.707, between 0.837 and 0.894, between 0.894 and
0.949, between 1.095 and 1.183 (tiny jump), and between
1.225 and 1.342 (jump). Insufficiently detailed data for
resolution of phases IV and VI. Lines II and III are parallel as
are, roughly, also lines VIII and IX. Very high absolute r
values for lines I, V and VII.
Fig. 14. Nine phases. Transitions at 0.140, between 0.316 and
0.447 (tiny jump), between 0.548 and 0.632, between 0.632
and 0.707, between 0.837 and 0.894, between 0.894 and
0.949, at 1.109, and between 1.342 and 1.414. Insufficiently
detailed data for resolution of phases IV, VI and IX. Lines II
and III are parallel. Very high absolute r values for phases I
and V
Fig. 15. Nine phases. Transitions at 0.164, between 0.316 and
0.447 (tiny jump), between 0.548 and 0.632, between 0.632
and 0.707, between 0.837 and 0.894, between 0.894 and
0.949, and at1.139 and 1.327. Insufficiently detailed data for
resolution of phases IV and VI. Lines II and III are parallel.
Very high absolute r values for phases IV and VI.
Fig. 16. Nine phases. Transitions at 0.193, between 0.316 and
0.447 (tiny jump), between 0.548 and 0.632, between 0.632
and 0.707, between 0.837 and 0.894 (tiny jump), between
1.000 and 1.095, between1.095 and 1.183, and between 1.225
and 1.342 (jump). Insufficiently detailed data for resolution of
phases IV and VII. Lines II and III and, approximately, lines
VIII and IX are parallel. Very high absolute r values for
phases I, V and VI.
7
Fig. 17. Data from Table 3 in Lee et al. 2002. The circled points are not included in their Fig. 4.
Conclusions and Questions
The finding of multiphasic profiles for ion activities has far-reaching implications and raises difficult
questions (Nissen 2017). These will not be reiterated here, but complex profiles for activation, such as the
present ones by Na+, are spectacular and may prove useful in further studies.
Acknowledgment – I am very grateful to Bob Eisenberg for his continued interest and encouragement.
References
Lee L-S, Chen T-M, Tsai K-M, Lin C-L (2002) Individual ion and mean activity coefficients in NaCl, NaBr,
KCl and KBr aqueous solutions. J Chin Inst Chem Engrs 33: 267281.
Nissen P (1971) Uptake of sulfate by roots and leaf slices of barley. Physiol Plant 24: 315-324.
Nissen P (2017) Multiphasic profiles for the biologically important ion activities of NaCl and KCl. arXiv. In
press.
|
1012.3879 | 1 | 1012 | 2010-12-17T13:55:55 | Plausibility of Quantum Coherent States in Biological Systems | [
"physics.bio-ph",
"q-bio.NC",
"quant-ph"
] | In this paper we briefly discuss the necessity of using quantum mechanics as a fundamental theory applicable to some key functional aspects of biological systems. This is especially relevant to three important parts of a neuron in the human brain, namely the cell membrane, microtubules (MT) and ion channels. We argue that the recently published papers criticizing the use of quantum theory in these systems are not convincing. | physics.bio-ph | physics | Plausibility of Quantum Coherent States in Biological
Systems
1,2,3V. Salari, 4,5J. Tuszynski, 6,7M. Rahnama, 8G. Bernroider
1Institut de Mineralogie et de Physique des Milieux Condenses, Universite Pierre et Marie Curie-Paris 6, CNRS
UMR7590, France
2Boite courrier 115, 4 place Jussieu, 75252 Paris cedex 05, France
3 BPC Signal, 15 rue Vauquelin, 75005 Paris, France
4Department of Experimental Oncology, Cross Cancer Institute, 11560 University Avenue
Edmonton, AB T6G 1Z2, Canada
5Department of Physics, University of Alberta, Edmonton, AB Canada
6S. B. University of Kerman, Kerman, Iran
7Afzal Research Institute, Kerman, Iran
8Department of Organismic Biology, University of Salzburg, Hellbrunnerstrasse 34, Salzburg, Austria
Abstract:
In this paper we briefly discuss the necessity o f using quantum mechanics as a fundamental
theory applicable to some key funct ional aspects of bio logical systems. This is especially
relevant to three important parts of a neuron in the human brain, namely the cell membrane,
microtubules (MT) and ion channels. We argue that the recent ly published papers crit icizing
the use of quantum theory in these systems are not convincing.
1) Quantum theory and biological systems
Bio logical systems operate within the framework of irreversible thermodynamics and
nonlinear kinet ic theory of open systems, both of which are based on the principles of non-
equilibrium statist ical mechanics. The search for phys ically-based fundamental models in
bio logy that can provide a conceptual bridge between the chemical organizat ion of living
organisms and the phenomenal states of life and experience has generated a vigorous and so
far inconclusive debate [1,2]. Recent ly published experimental evidence has provided support
for the hypothesis that bio logical systems use some type of quantum coherence in their
funct ions. The nearly 100% efficient excitat ion energy transfer in photosynthesis is an
excellent example [3]. Living systems are composed of mo lecules and atoms, and the most
advanced phys ical theory describing interact ions between atoms and mo lecules is quantum
mechanics. For example, making and breaking o f chemical bonds, absorbance of frequency
specific radiat ion (e.g. in photosynthesis and vision), conversion of chemical energy into
mechanical mot ion (e.g. ATP cleavage) and single electron transfer through bio logical
polymers (e.g. in DNA or proteins) are all quantum effects. Regarding the effic ient
funct ioning o f bio logical systems, the relevant question to ask is how can a bio logical system
with billions of semi-autonomous components funct ion effect ively and coherent ly? Why
providing a complete explanation remains a major challenge, quantum coherence is a
plausible mechanism responsible for the efficiency and co-ordinat ion exhibited by bio logical
systems [4].
The hypothesis invoking long-range coherence in bio logical systems was proposed by H.
Fröhlich’s [5-7] and fo llowed by detailed invest igations by Tuszynski et al. [8-22], Pokorny
[23-25], Mesquita et al. [26-28] and others for over three decades. The possible ro le played
by coherent states manifested outside low temperature physics has attracted considerable
interest in both the physics and bio logy communit ies. Despite the potential power of quantum
mechanics to explain coherent phenomena, there are serious challenges invo lved in applying
it in the context of a living system. For instance, in order to have a very high degree o f
coherence between bio-mo lecules, Bose-E instein condensat ion may be a viable mechanism,
but we note that the ambient temperature in a living system is likely to be too high for this
phenomenon to occur. Also, the sizes o f bio-mo lecules are very large by physical standards in
order to be regarded as typical quantum sys tems. Moreover, because of the no isy
environment, according to decoherence theory, quantum states of these mesoscopic bio-
mo lecules would co llapse very rapidly. However, remarkably there is no obvious limitation
placed on the Schrodinger equat ion for its use only in atomic-scale systems. It is a universa l
equation and it can even be used for the ent ire universe (as is the case with quantum gravity
applicat ions). The boundary between quantum theory and classical phys ics is st ill largely
unknown. Quantum theory obviously applies on length scales smaller than atomic radii but
beyond that it is not entirely clear where it should be superseded by Newtonian mechanics.
Superconductors, lasers, superfluids, semiconductors etc., are examples of macroscopic-scale
physical systems that behave quantum mechanically, so it is also possible that bio logica l
systems operate based on quantum principles at least in some o f their funct ions. Here we
argue that recent crit icisms of the use of quantum mechanics in bio logy are not very
convincing since they ignore the already exist ing evidence for quantum effects in bio logica l
systems. In this paper we invest igate three particular systems o f special importance: the cell
membrane, microtubules (MTs), and ion channels which are some of the most important parts
of a neuron in the human brain. We argue that these subsystems are the best candidates for
possible sites of quantum effects.
2) A brief overview of the criticism of coherence and decoherence
2-1) Membrane
The original Fröhlich model was general and did not limit the mechanism o f bio logical
coherence to any part icular cellular structure. In his model, when the energy supply exceeds a
crit ical level, the dipo lar ensemble of bio logically relevant mo lecules populates a steady state
of non-linear vibrat ions characterized by a high degree of structural and funct ional order [29].
This (electrically polarized) ordered state expresses itself in terms of long-range phase
correlat ions, which are physically similar to such phenomena as superconduct ivity and
superfluidity, where the behaviour of particles is collect ive and inseparable. The existence of
very strong static electric fields across the cell membrane led Fröhlich to consider cellular
membranes to be the source of the postulated coherent vibrat ions [29].
Froehlich considered a model for bio logical systems consist ing of three parts:
1) A system o f oscillators
2) A heat bath
3) An external source which pumps energy into the system incoherent ly.
Froehlich [5] derived a formula for the net rate of energy loss
in quanta) which can be written as
i (containing
iL of the mode with frequency
1
L
1
i
enT
(
)(
i
i
kT
1
n
i
)
(1)
where is Planck’s constant (
h ), k is Boltzmann’s constant, T is the temperature of the
2
(T is a funct ion of T. Recently, Reimers et al. [2] have shown
system and the heat bath, and
)
that a very fragile Froehlich coherent state may occur at sufficient ly high temperatures and
concluded that there is no possibility for the existence of Froehlich coherent states in
bio logical systems. However, it should be noted that in equat ion (1) they replaced the
(T with a constant at 0 Kelvin. This replacement causes equat ion (1) to be
funct ion
)
plagued by serious problems in the limit
T
,0
,
i
kT
e
i
kT
L
1
i
(2)
It means that the net rate of energy loss tends to infinity near the abso lute zero, so it makes
this assumpt ion and it s consequences unphys ical. In the second order considered by these
authors, in the case o f a three body system, the same problem emerges, too. Also they have
demonstrated several diagrams in terms of effective temperature which was defined by
T
ST is the temperature of system and T is the temperature of
T
themselves as
, where
S
eff
T
the thermal bath. Their effect ive temperature relat ion is
T
S
T
1
0
kT
2
es
(3)
where s, and
effT is not a well-defined parameter because on both
0 are constants. Here
sides o f the relation in (3) we have temperature of the heat bath T . Therefore, in the limit,
both sides o f the relat ion cannot be reconciled with each other:
0
0
kT
2
kT
2
thus T
eff
0,
1
(4)
while
,
s
0
)
e
(1
0
kT
2
e
T
T
,
0
kT
2
0,
e
0
kT
2
1
thus T
eff
0
while
(1
0
kT
2
s
e
)
1
s
(5)
They have used the effective temperature parameter for the Wu-Aust in Hamiltonian [30-32]
and considered it in the high temperature limit based on the references numbered [49, 50] and
[51] in their paper while there are actually only 49 references cited in this paper. Their
diagrams are most ly based on the effective temperature parameter and hence are, in our
opinion, not acceptable due to the internally contradictory argument used in their derivat ions.
2-2) Microtubules
The crit icism raised by Reimers et al. [2, 37] is mainly directed against the so-called Orch OR
model which was proposed by Penrose and Hameroff to introduce a physical basis for
consciousness. In some formulat ions of the OrchOR model, a manifestation of quantum
coherence invo lved the invo lvement of Froehlich coherent states in MTs [33-35]. MTs are
highly ordered in the neurons o f the brain and can indeed be regarded as good candidates for
supporting Froehlich coherent states. In this context, the conclusions o f sect ion 2 apply to
MTs as well. Therefore, we believe that it is still hypothetically possible to generate
Froehlich coherent states in MTs. However, another issue that arises when considering
quantum states for MTs is the rapid decoherence problem. The quest ion is “how is it possible
for MTs to be in a coherent state while the environment surrounding them is relat ively hot,
wet and no isy?” According to the decoherence theory, while macroscopic objects obey
quantum mechanics, their interactions with the environment cause decoherence, which
destroys quantum effects [36]. For macroscopic particles there are two main 'natural' ways o f
experiencing decoherence, first is due to collisions with other particles and the second is the
thermal emission o f radiation due to the internal heat of an object [38, 41]. In the latter case,
the decoherence t ime of the system is given by
decT
1
a T
6
36
10
9
(
x
)
2
(6)
where a is the size o f the object (diameter or length), T is the abso lute temperature and x is
the superposit ion distance o f the object [38]. At room temperature T=300K and the size
a=8nm for the tubulin dimer (i.e. the elementary const ituent of an MT) and considering
for its superposit ion distance, the decoherence time would be on the order of
8x
nm
decT
s
. This indicates that thermal photons cannot cause decoherence in this case. Hence
610
we concentrate on the first case which is decoherence due to the scattering by environmental
particles. Tegmark [39] has calculated decoherence times for MTs based on the scattering
between MTs and environmental particles. Based on the co llisions o f ions with MTs, he has
obtained the decoherence times on the order of:
D
mkT
2
Ngq
2
10
13
s
(7)
where D is the tubulin diameter, m is the mass of the ion, k is Bo ltzmann’s constant, T is
temperature, N is the number of elementary charges in the microtubule interacting system,
1
g
is the Coulomb constant and q is the charge of an electron. Hagan et al. [40] have
4
0
shown that Tegmark used wrong assumpt ions for his investigation o f MTs. Another main
object ion to the est imate in equat ion (7) is that Tegmark’s formula yields decoherence t imes
that increase with temperature contrary to a well-established physical experience and the
observed behavior of quantum coherent states. In view of these (and other) problems in
Tegmark’s est imates, Hagan et al. [40] assert that the values of quant it ies in Tegmark’s
relat ion are not correct and thus the decoherence time should be approximately
1010 times
larger leading to a ms range of values for typical decoherence times. According to Hagan et
al., MTs in neurons could avo id decoherence via several mechanisms over sufficient ly long
times for quantum processing to occur there.
2-3)
Ion Channel and Selectivity Filter (SF)
Ion channels are proteins in the membranes o f excitable cells that cooperate for the onset and
propagation o f electrical signals across membranes by providing a highly select ive
conduction o f charges bound to ions through a channel like structure. The SF is a part of the
protein forming a narrow tunnel inside the ion channel which is responsible for the select ion
process and fast conduction of ions across the membrane. The determinat ion of atomic
resolut ion structure of the ion channel and select ivity filter by Mac Kinnon led to the award
of the Nobel prize for chemistry in 2003 [42]. The SF is a very narrow region o f the protein.
In the potassium select ive KcsA bacterial channel, that frequent ly serves as a model structure,
this regions extends to about 1.2 nm length and 0.3 nm in diameter. At the atomic scale the
filter region exposes negative charges owing to the lone electron pairs from oxygen ions
bound to 20 carbonyl groups arranged into 5 rings of the lining ’P-loop’ peptide. Alltogether
this structure provides a highly ordered atomic coordination pattern among oxygens and ions.
If a posit ive charged ion, such as sodium or potassium enters the SF, the ion can be
transient ly trapped by Coulombic interact ions with the negat ive charges provided by the
oxygen ions. The physical act ion orders associated with the transient trapping o f ions from
the macroscopic scale down to the quantum scale have been analysed previously [43]. It has
been concluded [43-46] that quantum theory is needed to explain the states of ions in the
select ivity filter and hence the funct ion of an ion channel. In contrast, Tegmark has calculated
decoherence times for the superposit ion of ions crossing the ent ire membrane based on a
simple ion-pore diffusion model [39]. The calculated decoherence times for crossing ions in
this work are derived from the scattering with environmental particles based on the Coulomb
interact ion between ions and particles. He has assumed that ions are in a superposit ion state
of inside and outside of the cell and are separated by a distance of 10 nm. In the view o f
atomic scaled reso lut ion maps and recent mo lecular dynamics studies of the filter region in
this protein, this type o f interaction is oversimplified and the pore-diffusionscattering mode l
is not applicable to describe ion protein interactions. Further, Tegmark introduced a funct ion
for the decoherence rate [47] which is composed of two parts: one for short wavelengths and
the other for long wavelengths. Every scattering calculat ion based on the Coulomb
interact ion and Tegmark’s decoherence rate funct ion leads to decoherence times being
T
T
2
direct ly proportional to temperature according to relat ions such as
,
,
dec
dec
T
3
, etc. Therefore, it can be expected that subsequent calculat ions based on these
dec
criteria are flawed in the high-temperature limit, i.e. as temperature approaches infinity,
decoherence t ime increases too, and if temperature approaches abso lute zero decoherence
time approaches zero, a very unphysical situat ion. In contrast to Tegmarks conclusions, we
suggest that his formulation o f the problem cannot apply to the ion-protein coordinated states
and his calculations do not address the temperature dependence o f decoherence t imes
correctly. Instead it is obvious that this problem remains unreso lved and there is no simple
and general relat ion between decoherence t ime and temperature in ion channels.
Based on a number o f physical examples, we may expect bio logical coherence to occur under
special circumstances at physio logical temperatures. For example, it is well-known that lasers
can maintain their coherence at high temperatures due to external pumping. Moreover,
quantum spin transfer between quantum dots connected by benzene rings (the same structures
that is found in aromatic hydrophobic amino acids) is more efficient at warm temperature
than at abso lute zero [48]. Further, a simple calculat ion o f the translocation t ime o f ions
through an ion channel if based on a particle po int of view, is quite different from what is
obtained from experimental data based on X-ray crystallography and MD simulat ions [49-
54]. The difference in the est imates of these translocation t imes can invo lve a factor up to 102
to 105 which indicates that the previous calculations using the particle point of view (i.e.
classical phys ics) should be corrected when applied to the SF of channel proteins. Ions
coordinated by the filter carbonyls should be considered as wave-packets rather than particles
and their translocat ion as well as filter states must invo lve some interferences that account for
the observed discrepancies between classical estimat ions and real observat ions. Recently,
some of the expected consequences behind quantum interferences in the filter region o f ion
channels have been introduced into the classical equations of motion that lead to Hodgkin-
Huxley type membrane potentials. As a result, these ’semi-classical’ equat ions o f ion mot ion
are shown to carry the signatures of quantum-interferences [45] and can predict signal onset
characterist ics that are in accord with real recordings from central brain neurons [55].
3) Conclusion:
We have argued above that the object ions raised against the feasibility and role of quantum
effects in bio logical systems are not tenable. It is st ill an open question as to how a
macroscopic object is classical while its const ituent atoms and mo lecules are quantum
mechanical in nature. As atomic-scale quantum systems compose into large mo lecules they
become classical objects. In bio logy however, the challenging quest ion is whether and how
the init ial quantum properties extend into the ’funct ional domain’ o f the emerging classical
systems. In the case o f energy transfer in photosynthesis, magnetic compass sensing with
receptor proteins, microtubules and ion conduct ion in channel proteins, transient quantum
coherences may well play a decis ive ro le to explain the observed funct ional states of classical
mo lecular systems. However, it seems that decoherence theory has not solved this problem,
and hence we pose an important question: “What is the meaning of classicality when a large
or complex system (as a quantum system) collapses to become a classical entity while the
components (atoms or mo lecules) are st ill quantum mechanical?”. At least in bio logical
systems, we can expect that the emergence o f classicality will invo lve some quantum
signatures that cannot be ignored in funct ional explanat ions. Answers along this way will
most probably play an increasing role for the understanding of organizational complexity and
funct ions in living systems.
Acknowledgments:
Research support from NSERC (Canada) awarded to J.A.T. is gratefully acknowledged.
Vahid Salari acknowledges BPC Signal for their support. We thank Roman Fuchs from the
Neurosignaling Unit, Univ. Salzburg, Austria for kindly assist ing in preparing the
manuscript.
References:
[1] Abbott D, Gea-Banacloche J, Davies PCW, Hameroff S, Zeilinger A, E isert J,
Wiseman H, Bezrukov SM and Frauenfelder H 2008. Plenary Debate: Quantum
Effects in Bio logy, Trivial or not ? Fluctuation and Noise Letter, Vol. 8, No.1, c5-c26
World Scient ific Publishing Company.
[2] Reimers JR, McKemmish LK, McKenzie RH, Mark AE and Hush, NS 2009. PNAS
0806273106
[3] Cheng, YC and Fleming, GR 2009. Dynamics o f light harvest ing in photosynthesis.
Annu. Rev. Phys. Chem., 60, 241–262.
[4]
[5]
[6]
[7]
[8]
[9]
Salari V, Rahnama M, Tuszynski J, 2008. On the Possibility of Quantum Visual
Informat ion Transfer in the Human Brain, Preprint: arXiv.org 0809.0008. To apprear
in Foundations of Physics.
Fröhlich, H. 1968. Int, J. Quantum Chem., 2, 641
Fröhlich, H. 1970. Nature (London), 228, 1093.
Fröhlich, H. 1975. PNAS, 72, 4211.
Tuszynski, J.A., Dixon, J.M., Oct 2001. Quantitative analysis o f the frequency
spectrum of the radiat ion emitted by cytochrome oxidase enzymes. Physical Review E
64 (5), 051915.
Tuszynski, J.A., Paul, R., Chatterjee, R., Sreenivasan, S.R., Nov 1984. Relationship
between Fröhlich and Davydov models of bio logical order. Physical Review A 30 (5),
2666e2675.
[10] Tuszynski, J.A., Paul, R., Mar 1991. Reply to “comment on ‘relat ionship between
Fröhlich and Davydov models of bio logical order’”. Physical Review A 43 (6),
3179e3181.
[11] Tuszynski, J.A., Bolterauer, H., Sataric, M.V., 1992. Self-organizat ion in bio logical
membranes and the relat ionship between Fröhlich and Davydov theories.
Nanobiology 1, 177e190.
[12] Tuszynski, J.A., Feb. 1985. On the existence of Fröhlich’s long-range interactions in
dipo lar chains. Physics Letters A 107, 225e229.
[13] Tuszynski, J.A., 1985a. Comments on the Davydov Hamiltonian for quasi-one-
dimensionalmo lecular chains. International Journal of QuantumChemistry 29,
379e391.
[14] Tuszynski, J.A., 1985b. Generalization of Fröhlich’s thermodynamic model o f the
living state phase transit ion. Physics Letters A 108 (3), 177e182.
[15] Tuszynski, J.A., 1987. Bio logical membranes as dissipat ive structures. Nuclear
Physics B Proceedings Supplements 2, 618.
[16] Tuszynski, J.A., 1988. On possible mechanisms of rouleau format ion in human
erythrocytes. Journal of Theoretical Biology 132 (3), 369e373.
[17] Bolterauer, H., Tuszynski, J.A., 1989. Fröhlich’s condensat ion in a bio logical
membrane viewed as a Davydov so liton. Journal of Biological Physics 17, 41e50
[18] Bolterauer, H., Tuszynski, J.A., Satari"c, M.V., Jul. 1991. Fröhlich and Davydov
regimes in the dynamics of dipo lar oscillat ions of bio logical membranes. Physical
Review A 44, 1366e1381.
[19] Chatterjee, R., Tuszynski, J.A., Paul, R., 1983. Comments on the Bose condensat ion
of phonons in the bio logical systems. International Journal of Quantum Chemistry
23, 709e712.
[20] Paul, R., Chatterjee, R., Tuszynski, J.A., Fritz, O.G., 1983. Theory of long-range
coherence in bio logical systems. I. The anomalous behaviour of human erythrocytes.
Journal of Theoretical Bio logy 104 (2), 169e185.
[21] Paul, R., Tuszynski, J.A., Chatterjee, R., Nov 1984. Dielectric constant of bio logical
systems. Physical Review A 30, 2676e2685
[22] Portet, S., Tuszynski, J.A., Hogue, C.W.V., Dixon, J.M., 2005. Elast ic vibrations in
seamless microtubules. European Biophysics Journal 34 (7), 912e920.
[23] Pokorný, J., 1982b. Multiple Fröhlich coherent states in bio logical systems: computer
simulat ion. Journal of Theoretical Biology 98 (1), 21e27.
[24] Pokorný, J., 1999. Conditions for coherent vibrat ions in cytoskeleton.
Bioelectrochemistry and Bioenergetics 48 (2), 267e271.
[25] Pokorný, J., 2009. Biophysical cancer transformat ion pathway. Electromagnetic
Biology and Medicine 28 (2), 105e123.
[26] Mesquita MV, Vasconcello s AR, Luzzi R, Mascarenhas S 2005. Large-scale quantum
effects in bio logical systems. Int J Quantum Chem 102:1116–1130.
[27] Mesquita MV,Vasconcellos AR, LuzziR.1996. Near-dissipat ionless coherent
excitat ions in bio systems. Int J Quantum Chem 60:689–697.
[28] Mesquita MV, Vasconcello s AR, Luzzi Roberto, Mascarenhas S. 2005. Systems
Bio logy: An Informat ion-Theoretic-Based Thermo -Statist ical Approach. Brazilian
Journal of Physics. 34, 459-488.
[29] Cifra M, Fields JZ, Farhadi A, 2010. Electromagnet ic cellular interactions, Progress
in Biophysics and Molecular Biology, doi:10.1016/j.pbiomo lbio.2010.07.003.
[30] Wu TM ., Austin S. 1977. Phys Lett A 64:151–152.
[31] Wu TM ., Austin S. 1978. Phys Lett A 65:74–76.
[32] Wu TM Austin S.1981. J Bio Phys 9:97–107.
[33] Penrose, R., Hamero ff, S.R. 1995. J. Conscious. Stud., 2, 98.
[34] Hamero ff, S.R., Penrose, R. 1996. J. Conscious. Stud., 3, 36.
Hameroff, S.R. 1998. Philos. Trans. R. Soc. London Ser. A, 356, 1869.
[35]
[36] Rosa LP, Faber J, 2004. Quantum Models of the Mind: Are they compat ible with
environment decoherence?, Physical Review E, 70, 031902.
[37] McKemmish LK, Reimers JR, McKenzie RH, Mark AE, Hush NS, 2009. Penrose-
Hamero ff orchestrated object ive-reduction proposal for human consciousness is not
bio logically feasible, Physical Review E 80, 021912.
[38] Schlosshauer, 2007. Decoherence and Quantum-to-Classical Transit ion, The frontiers
collect ion, Springer.
[39] Tegmark, M. 2000. Phys. Rev. E 61, 4194-4206.
[40] Hagan, S., Hameroff, S. R., and Tuszynski, J. A. 2002. Quantum computation in brain
MTs: Decoherence and bio logical feasibility. Physical Review E, 65, 061901.
Joos and Zeh, 1985. Z. Phys. B, 59, 223-243.
[41]
[42] MacKinnon et al., 2003. Biophysical Journal, 66: 1061-1067, 1994 and MacKinnon R
FEBS Lett 555, 62-65.
[43] Bernro ider G and Roy S 2004. Quantum-classical correspondence in the brain:
scaling, action distances and predictability behind neural signals FORMA 19 55-68.
[44] Bernro ider G and Roy S. 2005. Quantum entanglement of K ions, mult iple channel
states and the role of no ise in the brain – SPIE Vol. 5841-29, pp. 205–14.
[45] Summhammer J, Bernroider G, 2007. Quantum entanglement in the voltage
dependent sodium channel can reproduce the salient features of neuronal act ion
potential init iat ion. Preprint arXiv: 0712.1474v1.
[46] Vaziri A, Plenio MB, 2010. New Journal of Physics, 12, 085001.
[47] Tegmark M, 1993. Found. Phys. Lett., 6, 571.
[48] Ouyang, M., Awschalom, D. D. 2003. coherent spin transfer between mo leculary
bridged quantum dots. Science, 301, 1074-1078.
[49] Shrivastava et al., 2002. Biophysical Journal, Vol. 83, 633-645.
[50] Shrivastava et al., 2000. Biophysical Journal, vol. 78, 557-570.
[51] Shrivastava et al., 2002. Eur. Biophys. J., vol. 31, 207-216.
[52] Guidoni, L. & P. Carloni, 2002 .Potassium permeation through the KcsA channel: a
density funct ional study., Biochimica et Biophysica Acta, 1563, 1-6.
[53] Guidoni et al. 1999, Biochemistry, 38, 8599-8604.
[54] LeMasurier et al., 2001 C. J. Gen. Physiol. 118, 303–314.
[55] Naundorf B, Wolf F and Vo lgushev M 2006. Nature 440 1060.
|
1811.07045 | 1 | 1811 | 2018-11-16T21:49:58 | Video-rate large-scale imaging with Multi-Z confocal microscopy | [
"physics.bio-ph",
"physics.optics"
] | Fast, volumetric imaging over large scales has been a long-standing goal in biological microscopy. Scanning techniques such as fluorescence confocal microscopy can acquire 2D images at high resolution and high speed, but extending the acquisition to multiple planes at different depths requires an axial scanning mechanism that drastically reduces the acquisition speed. To address this challenge, we report an augmented variant of confocal microscopy where the key innovation consists to use a series of reflecting pinholes axially distributed in the detection plane, each one probing a different depth within the sample. As no axial scanning mechanism is involved, our technique provides simultaneous multiplane imaging over fields of view larger than a millimeter at video-rate. We demonstrate the general applicability of our technique to neuronal imaging of both Caenorhabditis elegans and mouse brains in-vivo. | physics.bio-ph | physics | Video-rate large-scale imaging with Multi-Z confocal microscopy
Amaury Badon,1 Seth Bensussen,1 Howard J. Gritton,1 Mehraj R.
Awal,2 Christopher V. Gabel,2, 3 Xue Han,1, 3 and Jerome Mertz1, 3
1Department of Biomedical Engineering, Boston University, Boston, Massachusetts 02215, USA
2Department of Physiology and Biophysics, Boston University School of Medicine, Boston, Massachusetts 02218, USA
3Boston University Photonics Center, Boston, Massachusetts 02215, USA
(Dated: November 20, 2018)
8
1
0
2
v
o
N
6
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
5
4
0
7
0
.
1
1
8
1
:
v
i
X
r
a
Fast, volumetric imaging over large scales has been a long-standing goal in biological microscopy. Scanning
techniques such as fluorescence confocal microscopy can acquire 2D images at high resolution and high speed,
but extending the acquisition to multiple planes at different depths requires an axial scanning mechanism that
drastically reduces the acquisition speed. To address this challenge, we report an augmented variant of confocal
microscopy where the key innovation consists to use a series of reflecting pinholes axially distributed in the
detection plane, each one probing a different depth within the sample. As no axial scanning mechanism is in-
volved, our technique provides simultaneous multiplane imaging over fields of view larger than a millimeter at
video-rate. We demonstrate the general applicability of our technique to neuronal imaging of both Caenorhab-
ditis elegans and mouse brains in-vivo.
INTRODUCTION
Recently, there has been a trend toward the development
of microscopy systems that can monitor cellular activity over
large spatial scales with high temporal resolution. General
conditions for such imaging are that the microscope provide
enough spatial resolution to distinguish individual cells, and
enough temporal resolution to accurately track the sample dy-
namics of interest. Moreover, to monitor large populations of
cells in-vivo, volumetric imaging is required, ideally encom-
passing hundreds to thousands of individual cells. The conflu-
ence of these conditions poses a challenge [1, 2]. Many previ-
ous approaches have been developed for volumetric imaging.
For example, widefield microscopes benefit from the large
sensor areas and excellent spatiotemporal resolution of mod-
ern cameras, and provide single shot 2D imaging. Volumetric
imaging at high speed can then be achieved with fast axial
scans [3 -- 5] or instantaneously using wavefront coding [6, 7],
diffractive elements [8 -- 10] or light-field approaches [11, 12].
However, lack of optical sectioning limits these approaches
to shallow depths, sparse samples or imposes a reliance on in-
tensive numerical post-processing. Light-sheet [13, 14] or tar-
geted illumination [15] improves sectioning by reducing out-
of-focus excitation, but suffers from an inability to prevent
scattered background light from impinging the camera sensor,
yielding limited image contrast in thick tissue such as mouse
brain. Alternatively, higher contrast in thick tissue can be ob-
tained with scanning techniques involving multiphoton exci-
tation or confocal detection, but these are generally operated
with micron-scale resolution, causing them to be slow and to
produce massive amounts of data when operated over large
fields of view, or to be limited to modest volumes [16]. Ran-
dom access techniques [17, 18] reduce the amounts of data
production, but require scan pre-calibration and are sensitive
to motion artifacts. Alternatively, data production can be re-
duced by purposefully limiting the spatial resolution, either
axially [19, 20] (though see [21 -- 23], where axial resolution
can be recovered in sparse samples by post-processing) or
near-isotropically [24].
In our case, since we are primarily
interested in the segmentation of individual cells in 3D, near-
isotropic resolution is more appropriate.
Here we describe a novel microscopy technique that pro-
vides video-rate, multiplane, optically sectioned imaging over
large FOVs on the millimeter scale. Our technique, called
Multi-Z confocal microscopy, is based on three main ideas.
The first is to combine high-NA detection with low-NA il-
lumination. The former leads to high signal collection effi-
ciency; the latter leads to axially extended illumination over
a large range of Z depths. The second idea is to detect mul-
tiple signals from this extended depth range using multiple
confocal pinholes that are axially distributed (similar to the
technique in [25], though over a much larger scale). The pin-
holes are reflecting, so that signal rejected by one pinhole is
sent to the next pinhole, and so forth. In this manner, no signal
is lost, and signal collection efficiency remains high (a reflect-
ing pinhole has been used previously, though in the context
of two-photon microscopy [26]). The third idea is to exploit
the benefits obtained from our larger confocal probe volumes,
namely larger signals and avoidance of oversampling at cellu-
lar resolution, to scan over much larger FOVs than provided
by standard confocal microscopes, thus optimizing our Multi-
Z microscope for fast, large-scale imaging of cell populations.
MULTI-Z PRINCIPLE
A schematic of our experimental set up is shown in Figure
1a. In practice, our instrument resembles a standard confo-
cal microscope, though with differences in both the illumina-
tion and detection optics. To achieve low-NA illumination,
we significantly underfill the back aperture of our microscope
objective (MO) by inserting an afocal beam compressor in the
excitation laser path, which results in a highly axially elon-
gated illumination beam in the sample. With NAill ≈ 0.1,
the axial extent on the illumination focus is of the order of
100 µm. The fluorescence from the sample, in turn, is epi-
2
FIG. 1: Multi-Z confocal microscopy. (a) Simplified schematic of the experimental setup. The multiplane detection unit is comprised of a
series of axially distributed reflecting pinholes, each probing a different depth within the sample. (b) Axially extended illumination is obtained
by underfilling the back aperture of the MO. The full NA of the MO is used for detection. (c) Transverse (xy) and axial (xz) PSF measured with
a sub-diffraction size bead, and associated x and z profiles. Scale bar, 5 µm. (d) Bead signal recorded by each detection channel at different
z positions of the stage. Continuous lines correspond to Lorentzian fits. (e) Different imaging planes simultaneously acquired of Aspergillus
Conidiophores. Scale bar, 200 µm.
collected through the full aperture of the same MO, which en-
sures both maximized fluorescence collection efficiency and
a detection depth of field much shorter than the illumination
axial depth (see Fig. 1b). The fluorescence is then routed
to not one but several pinholes (here four), each conjugate to
different depths separated by ∆z (here 25 µm) within the il-
lumination profile. The size of the pinholes is adjusted to be
approximately M δx, where δx is the transverse illumination
width and M is the magnification from the sample to the pin-
holes (here M = 62.5). In the image space, the separation
between the pinholes ∆Z is given by approximately M 2∆z.
Because of its quadratic dependence on M, this separation
is in the range of several centimeters even though ∆z is in
the range of microns, making the pinhole layout particularly
convenient to build (as opposed to the more restricted imple-
mentation involving high-NA illumination and close-packed
micro-machined reflectors [25]). We note that downstream
channels are subject to a small amount of screening from the
pinholes in the upstream channels. An evaluation of this mi-
nor effect is discussed in Supplement 4.
To evaluate the resolution of our microscope, we imaged
sub-diffraction-sized beads over a 200 µm axial range about
the MO focal plane. Using a 10× objective, we achieved for
each detection channel a resolution of δx= 4.8 µm and δz=15
µm in the transverse and axial dimensions, respectively (Fig.
1c). Note that this compromise in resolution was purpose-
fully chosen to facilitate high-speed imaging over large FOVs.
Such 3D resolution is easily adequate to resolve individual
cells, such as neurons in brain tissue, while avoiding problems
associated with oversampling and the unnecessary processing
of massive amounts of data [27]. The resultant bead signal si-
multaneously recorded by the four detection channels clearly
demonstrates the multiplane capability of our approach (see
Fig.1d). As expected, the four curves are similar and axi-
ally shifted by ∆z=25 µm, thus illustrating confocal detec-
tion over 100 µm. The slightly broader and lower amplitude
responses associated with the first and last detection channels
arise from the axial roll-off of our Gaussian-Lorentzian illu-
mination beam away from its nominal focus, as detailed in
Supplement 3.
As an example demonstration, we imaged a fixed sample of
Aspergillus Conidiophores, a common mold. Figure 1e shows
the four simultaneously acquired images from the shallow-
est (z = −37.5 µm) to the deepest plane (z = +37.5 µm).
Different features and the branching of the hydrae are clearly
visible in these images, demonstrating the simultaneous mul-
tiplane acquisition and optical sectioning capabilities of our
microscope over a large FOV, here 1.2 × 1.2 mm2.
Microscopeobjective SampleDichroic LenspinholesPhoto-detectorsPupil planeIlluminationDetectionabcd∆zZ-stage position (μm)-50050Intensity (a.u)00.20.40.60.81∆z60070080090010001100100200300400500600700800yx x (um)-10010Intensity (a.u)0Z-stage position (μm)-40040Intensity (a.u)e-37.5 μm-12.5 μm+12.5 μm+37.5 μmScanningheadzxδxδzLaser zOptional ETL 3
FIG. 2: In vivo volumetric imaging of C. elegans. (a) Simultaneous imaging of a single worm at different depths and the resulting volumetric
rendering. Scale bar, 200 µm. (b) Extended depth of field image showing the nuclei marked with NLSmCherry of four different worms in the
FOV. Scale bar, 200 µm. (c,d) Close up view (blue rectangle), showing the nuclei (in red) and the neurons displaying activity (in green). (e)
Activity of the 42 neurons identified in (d). (f) Standard deviation of the calcium activity recorded in the head ganglion of a worm in each
imaging plane. Each channel is displayed with a different color and in log scale. (g) Activity of the 32 neurons identified in (f) and recorded
over 1000 s. Each row corresponds to a time-series of an individual neuron, as shown in the insets.
HIGH-SPEED VOLUMETRIC IMAGING
Because Multi-Z confocal microscopy provides volumet-
ric imaging without the need for axial scanning, its speed is
limited by the mechanism for transverse 2D scanning only.
In our case, we use standard resonant-nonresonant galvano-
metric scanning (see Methods) to achieve video rate imaging,
easily fast enough to monitor the dynamics of cell activity re-
porters such as the calcium sensor GCaMP6. In particular, we
demonstrate the speed capacity of our system by monitoring
large-scale in vivo neuronal activity in different organisms.
Whole animal study : Caenorhabditis elegans
We first performed imaging of C. elegans expressing both
NLSmCherry,
targeted to nuclei, and GCaMP6s. Adult
worms are typically 500-800 µm long and 100 µm wide which
makes them difficult to image entirely with a conventional
confocal microscope. As a result, only young specimens or
the head ganglia are usually imaged. However our microscope
readily captured full 3D volumes of multiple worms embed-
ded in agarose gel, here up to four. Using two-color illumina-
tion and interleaved acquisition (see Methods) we were able to
simultaneously detect the neurons and monitor their activity.
Figure 2a displays the extended depth of field (EDOF) image
of the nuclei obtained by summing the intensity of the four
channels. Numerous neurons of the worms, from the densely
packed head ganglia to the tail are visible. Since the cal-
cium activity is simultaneously recorded, the active neurons
can be spatiotemporally resolved in 4D (see Fig.2b). Such
a rendering would not be possible with conventional confo-
cal microscopy, which is limited to single-plane imaging over
typically much smaller FOVs. Moreover, our approach en-
abled us to perform volumetric imaging of moving specimens
(see Visualization 1). Because of the agarose gel, the worm
was constrained to rotations only but our approach could be
extended to freely moving worms as well. From the same ac-
quisition, the Ca2+ activity of the 42 detected ventral neurons
of a selected worm (blue dashed rectangle) was monitored at
7.5 Hz for 200 s (Fig. 2.d). The corresponding Ca2+ traces
are shown in Fig. 2e, where different subsets of neurons re-
veal clearly correlated and anti-correlated activity associated
to forward and backward motion, in agreement with previous
-0.400.40.81.2∆F/F0abcdeg2004006008001000Time (s)103520180600700800Time (s)Time (s)Neuronsf-37.5 μm+37.5 μm+12.5 μm-12.5 μmEDOF∆F/F0Neurons01∆F/F001-0.400.40.81.2∆F/F0 observations [11, 24].
To image densely packed regions such as head ganglia, a
higher transverse resolution is preferable. This could be ob-
tained while maintaining the same distance ∆z between imag-
ing planes by simply switching the MO (20×) and two re-
lay lenses to maintain a constant M (see Methods). In this
new configuration, we imaged the head ganglion of a worm
(zoomed to 300 × 300 × 100 µm3) at 30 Hz. Figure 2f shows
an overlay of the temporal standard deviation of the Ca2+ sig-
nals, where color corresponds to depth. Neurons from differ-
ent depths are clearly distinguishable in this densely populated
region. Note that due to internal motion, the worm intestine
is also highlighted when we compute the standard deviation.
In total, 32 neurons are identified over a duration of 1000 s,
and their Ca2+ traces are presented as a heat map in Fig. 2g.
Our acquisition speed was amply sufficient to record even the
fastest Ca2+ dynamics (see also Visualization 2).
Mouse brain
More challenging is the demonstration of in vivo neuronal
imaging in mammalian brains. For this, we performed Ca2+
imaging in a mouse brain expressing GCaMP6f, requiring
both fast imaging, at least 15 Hz, and optical sectioning to re-
ject extraneously scattered light. We focused in particular on
the hippocampus, a subcortical region of the brain (see Meth-
ods). During imaging, the mice were awake, head-fixed and
no sensory stimulus was applied.
We recorded the spontaneous Ca2+ activity of neurons
within a 1200 × 1200 × 100 µm3 volume at 30 Hz, the
fastest rate currently achievable with our system for this FOV.
Though the labelling was confined to an area smaller than our
FOV, the EDOF images (i.e. z projections) reveal a large pop-
ulation of neurons (see figure 3a). Higher resolution images of
smaller regions of interest were also obtained by reducing the
FOV while keeping the same acquisition speed and number
of pixels, allowing the details of individual cells to become
clearer (see inset). The benefits of simultaneous, multiplane
acquisition are apparent in Fig. 3b, where different neurons
are revealed in different imaging planes. Using a constrained
non-negative matrix factorization algorithm (CNMF) [28], 90,
264, 312 and 260 neurons were separately identified in each
plane, from deepest to shallowest respectively, resulting in a
total number of 926 neurons. Owing to the intentional partial
overlap of the image planes (see Fig. 1d), some of these iden-
tifications are expected to be redundant. Indeed, similar pro-
cessing applied to the EDOF images pared back these identi-
fications to a total of 826 independent neurons, indicating that
the overlap-induced crosstalk was approximately 12% (Fig.
3a and Visualization 3). The corresponding extracted Ca2+
traces, recorded over 66 s, are shown in Figs. 3c and 3d for
a magnified view. These results illustrate the capacity of our
system to provide adequate SNR for neuronal segmentation in
large, relatively dense, populations at video-rate timescales.
AUGMENTED VOLUMETRIC IMAGING
4
Our microscope was designed to provide multiplane acqui-
sition over an axial range of about 100 µm. In the event that a
larger axial range is desired, several strategies can be consid-
ered. The first, and simplest, is to further extend the range
of the illumination focus by further underfilling the back-
aperture of the MO. In this case, the detection channels can be
spread more sparsely along the illumination profile, reducing
the detection fill factor and hence the spatial overlap of neigh-
boring detection probe volumes. Alternatively, to maintain the
same fill factor, the number of channels (and pinholes) can be
increased. Such a simple solution, however, would incur a
loss in transverse spatial resolution. For example, a doubling
of the axial range would lead to an worsening of the transverse
resolution by a factor
√
2.
Yet another strategy that does not impair transverse resolu-
tion involves the use of a mechanism to rapidly change focal
depths. Example mechanisms are an electrically tunable lens
(ETL) [29], an acoustic gradient lens [30, 31] or a deformable
mirror [5]. In our case, we demonstrated the augmentation
of our axial range with an ETL (Optotune VIS-EL-10-30-C),
which we inserted in the vicinity of the pupil plane between
the scanning mirrors and the MO. By controlling the focal
strength of the ETL, we thus axially translated both the illu-
mination and detection foci by up to 150 µm, while incurring
little change in the magnification and resolution of our system.
In particular, a square-wave control voltage was applied to the
ETL at half the microscope frame rate, leading to interleaved
stack acquisition that doubles the effective axial range of our
microscope, though at the cost of halving our net acquisition
speed.
As an example demonstration, we imaged a fixed mouse
brain with FITC-labeled vasculature. In a two-step acquisi-
tion process, we recorded eight frames in total, four from each
axial position of the ETL, leading to an overall imaging vol-
ume of 1200 × 1200 × 250 µm3 acquired at a net speed of
15 Hz. Figure 4b displays a resulting volume projection with
color-coded depth, revealing both larger vessels that span the
full axial extent of the recorded volume and smaller vessels
visible only in individual frames. These preliminary results
further demonstrate versatility of our microscope in providing
large-scale imaging.
SUMMARY
In summary, we have developed a generalized version of
confocal microscopy that provides simultaneous multiplane
imaging over large FOVs while remaining fast, here video
rate. Advantages of our system are that it is highly light effi-
cient, for two reasons. First, it makes full use of the detection
NA and incurs no loss upon signal detection through reflecting
pinholes, thus maximizing signal collection efficiency. Sec-
ond, whereas in a standard confocal microscopy the illumi-
nation light produces only a single image plane at a time, in
5
FIG. 3: Video-rate volumetric Ca2+ imaging in mouse brain expressing GCaMP6f (note that viral injection here led to a labelled area somewhat
smaller than our FOV) . (a) Extended depth of field image recorded in the hippocampus at 30 frames per second and averaged over 20 seconds.
Scale bar, 100 µm. Inset with 3× zoom illustrates cellular resolution. (b) Identification of neurons in each image plane using constrained
non-negative matrix factorization. (c) Activity of the 826 neurons identified in (a). (d) Magnified view of the neuronal activity traces for the
region indicated by the red rectangle in (c).
Multi-Z confocal microscopy the same illumination light is re-
peatedly utilized to produce multiple image planes at a time,
thus making more efficient usage of the laser excitation power
and, concomitantly, for the acquisition of equal image stacks,
reducing the deleterious effects associated with this excitation
power, such as photobleaching, photodamage, etc.. This sec-
ond excitation efficiency advantage is expected to scale with
the number of pinholes.
Other advantages of Multi-Z microscopy are that the sys-
tem requires no scan pre-calibration and no image post-
processing.
Indeed, the images shown here were not post-
processed in any way, save for image registration to correct for
sample dither, illustrating the robustness of our system against
motion artifacts which can easily be monitored in all three di-
mensions. On the other hand, post-processing could be en-
visaged, such as 3D deconvolution, as facilitated by the direct
acquisition of volumetric image data. Also, our approach is
robust against axial motion artifacts that may arise, for exam-
ple, from blood flow or animal breathing, which is a common
problem in in vivo applications. While in conventional confo-
cal microscopy axial motion can lead to total loss of the signal,
here the signal simply appears in a different channel.
Still other advantages are that our system is readily scalable
to a greater number of imaging planes, with no speed penalty
in principle, by simply adding pinholes and detectors, or to
an augmentation of axial range with no penalty in transverse
z=+37.5 mz=+12.5 mz=-12.5 mab1060Time (s)75450750Neuron index1040Time (s)z=-37.5 me1002002.50∆F/F0Neuron index200 % ∆F/F0c2004006008001000100200300400500600700 6
FIG. 4: Augmented volumetric imaging using an electrically tunable lens controlled by a square-wave voltage VET L. (a) Multi-Z images
obtained for VET L = 0 and VET L = 5V corresponding to a focal shift of 150 µm. (b) Resulting volumetric acquisition of fixed mouse brain
vasculature with structures color-coded by depth. Scale bar 300 µm.
resolution. Moreover, it is versatile in that it is amenable to
multi-color imaging (e.g. two-color here), and M and/or ∆z
can be adjusted with only minor optical modifications. For
example, since ∆Z = M 2 × ∆z, one can adjust the distance
between the imaging planes ∆z by simply tuning the total
magnification of the system M, or alternatively by tuning the
physical distance between the pinholes ∆Z. In our case, ∆z
was adjusted so that the confocal detection volumes filled the
axial extent of the illumination beam and there was no gap
between image planes. Other geometries are, or course, pos-
sible.
The simplicity and ease of use of our system should make
it attractive for general biomedical research applications.
25, Semrock). After being relayed and further magnified, the
fluorescence was detected by four reflecting pinholes (d=300
µm, National Aperture) that were axially distributed. The
separation between consecutive pinholes was 10 cm, which,
given the magnification of the system (M = 62.5×), corre-
sponds to a separation of ∆z = 25 µm in the object space. The
signals from the four photo-detectors (MicroFC-SMA-30020,
SensL) positioned behind the pinholes were amplified by cus-
tom built voltage amplifiers and digitized by a high-speed
4-channel FPGA board (PXI-7961R, National Instruments).
The setup was controlled from a dual-CPU workstation and
the experiments were conducted using Scanimage [32] micro-
scope control software.
METHODS
Hardware setup
Light from a blue (488nm/80mW, Omicron PhoxX) or
yellow-green laser (561nm/50mW, Vortran Stradus) was
scanned by a pair of galvanometric mirrors (CRS 8 kHz,
Cambridge Technology) and relayed by lenses to fill only
a fraction of the back pupil aperture of the MO (Optem
LWD 10X, NA=0.45 or Olympus XLUMPLFLN-W 20X
NA=0.95). The underfilling of the pupil results in low NA il-
lumination (NAill ≈ 0.1) that provides an axially extended fo-
cus in the focal plane of the MO. The fluorescence signal was
epi-detected by the full NA of the same MO, ensuring maxi-
mal collection efficiency and constrained axial resolution. The
fluorescence signal was then de-scanned and isolated with a
dichromatic mirror (403/497/574, Edmund Optics) and two
notch filters (OD4 488NM, Edmund Optics and NF03-561E-
Fixed samples experiments
PSF measurements were performed using 0.2 µm fluores-
cent beads (TetraSpeck, ThermoFisher T14792) mounted on
a glass slide. We used a commercially available Aspergillus
Conidiophores glass slide (Carolina Biological Supply Co.
297872). An average laser output power of 3mW was used
for these experiments.
Caenorhabditis elegans preparation
C. elegans strains were cultivated at 20◦C following stan-
dard procedures (on NGM agar seeded with Escherichia coli
OP50 as a food source). All imaging experiments were per-
formed on young-adult hermaphrodites using the transgenic
strain QW1217 (zfIs124[Prgef-1::GCaMP6s];otIs355[Prab-
3::NLS::tagRFP]) expressing panneuronal GCaMP and
HR 10x/0.45 ∞/0 f=200OPTEMabDepthVETL = 0VVETL = 5VDepth (μm)-2500 nuclear-localized RFP in all neurons in the worm. Worms
were partially immobilized for imaging by following a hy-
drogel encapsulation protocol [33].
In brief, worms were
placed into a droplet of solution consisting of 10 PEG-DA
and 0.1% Igracure. The droplet containing the worms was
then hardened by a minute-long exposure to UV-light, which
crosslinked the solution into a gel, preventing gross movement
of the worms. An averaged laser output power of 5mW was
used for these experiments.
Image processing and data analysis
Motion correction was performed using the Moco plugin
[35] in ImageJ. Data analysis was carried out using custom
scripts in Matlab (MathWorks) for the C. elegans experiments
and using constrained non-negative matrix factorization algo-
rithm for the mice experiments [28]. To extract ∆F /F0, we
computed ∆ F/F0 (t) = 100 Ã U (F(t)-F0)/F0, F0 being the
mean fluorescence intensity of each trace.
7
Mouse preparation and Imaging
Funding Information
All animal procedures were in accordance with the Na-
tional Institutes of Health Guide for the care and use of
laboratory animals and approved by the Boston University
Institutional Animal Care and Use Committee.
Female
C57BL/6 mice (n=2, Taconic; Hudson, NY), 8-12 weeks old
at the time of surgery, were first injected with AAV9-Syn-
GCaMP6f.WPRE.SV40 virus obtained from the University of
Pennsylvania Vector Core (titer ∼6e12 GC/ml). 250 nL of
virus was stereotaxically injected into the CA1 region (AP:
â A¸S2mm, ML: 1.4 mm, DV: â A¸S1.6 mm) using a 10 nL sy-
ringe (World Precision Instruments) fitted with a 33 gauge
needle (NF33BL; World Precision Instruments), at a speed
of 40 µL/min controlled by a micro-syringe pump (UltraMi-
croPump3â A¸S4; World Precision Instruments). Upon com-
plete recovery, animals were surgically implanted with custom
imaging windows, which consisted of a stainless steel cannula
(OD: 0.317 cm., ID: 0.236 cm., height 2 mm) adhered to a
circular coverslip (size 0; OD:3mm) with a UV-curable opti-
cal adhesive (Norland Products). After careful aspiration of
the overlying cortical tissue, using the corpus callosum as an
anatomical guide, the imaging window was placed above the
CA1 viral injection site. During the same surgery, a custom
aluminum head-plate was attached to the skull anterior to the
imaging cannula for head fixation during imaging. Animals
were imaged beginning 4 weeks later to allow for full expres-
sion of GCamp6f. Prior to imaging for the first time, animals
were also habituated to a custom mount head fixation appa-
ratus over several sessions as previously described [34]. An
average laser output power of 5 mW was used for these exper-
iments. Imaging sessions lasted typically 15-20 minutes.
Fixed-brain preparation
Before extracting a mouse brain, a cardiac perfusion with
heparinized saline followed by FITC-albumin-gelatin was
performed. After extraction, the brain was immersed in a 4%
PFA solution for 6 hours and then in a PBS solution for 3
days. Subsequently, the solution was replaced by distilled wa-
ter with 0.5% alpha-thioglycerol and increasing amounts of
fructose, 20, 40, 60, 80 and finally 100%.
This work was partially supported by NIH grant R21-
EY027549 and the Wallace H. Coulter Foundation.
Acknowledgments
We thank Timothy Weber, Jean-Marc Tsang and Sheng
Xiao for technical help and helpful discussions. We thank the
BU Neurophotonics Center for general support, and especially
Kivilcim Kilic for providing the fixed brain sample.
[1] N. Ji, J. Freeman, and S. L. Smith, Nature neuroscience 19,
1154 (2016).
[2] W. Yang and R. Yuste, Nature methods 14, 349 (2017).
[3] E. J. Botcherby, R. Juskaitis, M. J. Booth, and T. Wilson, Optics
[4] T.-H. Chen, J. Ault, H. Stone, and C. Arnold, Experiments in
letters 32 (2007).
Fluids 58, 41 (2017).
[5] W. J. Shain, N. A. Vickers, B. B. Goldberg, T. Bifano, and
J. Mertz, Optics letters 42, 995 (2017).
[6] E. R. Dowski and W. T. Cathey, Applied optics 34, 1859 (1995).
[7] S. Quirin, D. S. Peterka, and R. Yuste, Optics express 21, 16007
[8] P. M. Blanchard and A. H. Greenaway, Applied optics 38, 6692
(2013).
(1999).
[9] P. A. Dalgarno, H. I. Dalgarno, A. Putoud, R. Lambert, L. Pa-
terson, D. C. Logan, D. P. Towers, R. J. Warburton, and A. H.
Greenaway, Optics express 18, 877 (2010).
[10] S. Abrahamsson, J. Chen, B. Hajj, S. Stallinga, A. Y. Katsov,
J. Wisniewski, G. Mizuguchi, P. Soule, F. Mueller, C. D.
Darzacq, et al., Nature methods 10, 60 (2012).
[11] R. Prevedel, Y.-G. Yoon, M. Hoffmann, N. Pak, G. Wetzstein,
S. Kato, T. Schrödel, R. Raskar, M. Zimmer, E. S. Boyden,
et al., Nature methods 11, 727 (2014).
[12] N. C. Pégard, H.-Y. Liu, N. Antipa, M. Gerlock, H. Adesnik,
and L. Waller, Optica 3, 517 (2016).
[13] J. Huisken, J. Swoger, F. Del Bene, J. Wittbrodt, and E. H.
Stelzer, Science 305, 1007 (2004).
[14] M. B. Bouchard, V. Voleti, C. S. Mendes, C. Lacefield, W. B.
Grueber, R. S. Mann, R. M. Bruno, and E. M. Hillman, Nature
photonics 9, 113 (2015).
[15] S. Xiao, H.-a. Tseng, H. Gritton, X. Han, and J. Mertz, Scien-
tific reports 8, 7921 (2018).
8
[26] Q. Hu, P. Li, Y. Xiong, Y. Wang, X. Lv, and S. Zeng, Optics
[27] S. Weisenburger, R. Prevedel, and A. Vaziri, bioRxiv p. 115659
letters 43, 4598 (2018).
(2017).
[24] T. Schrödel, R. Prevedel, K. Aumayr, M. Zimmer, and
[35] A. Dubbs, J. Guevara, and R. Yuste, Frontiers in neuroinformat-
[16] L. Kong, J. Tang, J. P. Little, Y. Yu, T. Lämmermann, C. P. Lin,
R. N. Germain, and M. Cui, Nature methods 12, 759 (2015).
[17] Y. Otsu, V. Bormuth, J. Wong, B. Mathieu, G. P. Dugué,
A. Feltz, and S. Dieudonné, Journal of neuroscience methods
173, 259 (2008).
[18] K. N. S. Nadella, H. Roš, C. Baragli, V. A. Griffiths, G. Kon-
stantinou, T. Koimtzis, G. J. Evans, P. A. Kirkby, and R. A.
Silver, Nature methods 13, 1001 (2016).
[19] W. Yang, J.-e. K. Miller, L. Carrillo-Reid, E. Pnevmatikakis,
L. Paninski, R. Yuste, and D. S. Peterka, Neuron 89, 269 (2016).
[20] R. Lu, W. Sun, Y. Liang, A. Kerlin, J. Bierfeld, J. D. Seelig,
D. E. Wilson, B. Scholl, B. Mohar, M. Tanimoto, et al., Nature
neuroscience 20, 620 (2017).
[21] A. Song, A. S. Charles, S. A. Koay, J. L. Gauthier, S. Y.
Thiberge, J. W. Pillow, and D. W. Tank, Nature methods 14,
420 (2017).
[22] C. Roider, R. Piestun, and A. Jesacher, Optica 4, 1373 (2017).
[23] W. J. Shain, N. A. Vickers, J. Li, X. Han, T. Bifano, and
J. Mertz, Biomedical optics express 9, 1771 (2018).
A. Vaziri, Nature methods 10, 1013 (2013).
[25] C. Yang, K. Shi, M. Zhou, S. Zheng, S. Yin, and Z. Liu, Applied
Physics Letters 101, 231111 (2012).
[28] E. A. Pnevmatikakis, D. Soudry, Y. Gao, T. A. Machado,
J. Merel, D. Pfau, T. Reardon, Y. Mu, C. Lacefield, W. Yang,
et al., Neuron 89, 285 (2016).
[29] B. F. Grewe, F. F. Voigt, M. van t Hoff, and F. Helmchen,
Biomedical optics express 2, 2035 (2011).
[30] N. Olivier, A. Mermillod-Blondin, C. B. Arnold, and E. Beau-
repaire, Optics letters 34, 1684 (2009).
[31] M. Duocastella, G. Vicidomini, and A. Diaspro, Optics express
22, 19293 (2014).
[32] T. A. Pologruto, B. L. Sabatini, and K. Svoboda, Biomedical
engineering online 2, 13 (2003).
[33] K. Burnett, E. Edsinger, and D. R. Albrecht, Communications
Biology 1, 73 (2018).
[34] A. I. Mohammed, H. J. Gritton, H.-a. Tseng, M. E. Bucklin,
Z. Yao, and X. Han, Scientific reports 6, 20986 (2016).
ics 10, 6 (2016).
|
1009.4490 | 1 | 1009 | 2010-09-22T22:08:22 | A closed-form solution of the three-dimensional contact problem for biphasic cartilage layers | [
"physics.bio-ph",
"nlin.SI",
"q-bio.TO"
] | A three-dimensional unilateral contact problem for articular cartilage layers is considered in the framework of the biphasic cartilage model. The articular cartilages bonded to subchondral bones are modeled as biphasic materials consisting of a solid phase and a fluid phase. It is assumed that the subchondral bones are rigid and shaped like elliptic paraboloids. The obtained analytical solution is valid over long time periods and can be used for increasing loading conditions. | physics.bio-ph | physics | Traceback (most recent call last):
File "/usr/bin/pdf2txt.py", line 115, in <module>
if __name__ == '__main__': sys.exit(main(sys.argv))
File "/usr/bin/pdf2txt.py", line 107, in main
caching=caching, check_extractable=True):
File "/usr/lib/python2.6/site-packages/pdfminer/pdfpage.py", line 121, in get_pages
doc = PDFDocument(parser, password=password, caching=caching)
File "/usr/lib/python2.6/site-packages/pdfminer/pdfdocument.py", line 315, in __init__
xref.load(parser)
File "/usr/lib/python2.6/site-packages/pdfminer/pdfdocument.py", line 175, in load
(_, obj) = parser.nextobject()
File "/usr/lib/python2.6/site-packages/pdfminer/psparser.py", line 557, in nextobject
(pos, token) = self.nexttoken()
File "/usr/lib/python2.6/site-packages/pdfminer/psparser.py", line 482, in nexttoken
self.fillbuf()
File "/usr/lib/python2.6/site-packages/pdfminer/psparser.py", line 215, in fillbuf
raise PSEOF('Unexpected EOF')
pdfminer.psparser.PSEOF: Unexpected EOF
|
1108.4802 | 2 | 1108 | 2013-05-09T12:12:52 | Similarity Analysis of Macroecology | [
"physics.bio-ph",
"nlin.AO",
"q-bio.PE"
] | We perform a full similarity analysis of an idealized ecosystem using Buckingham's $\Pi$ theorem to obtain dimensionless similarity parameters given that some (non- unique) method exists that can differentiate different functional groups of individuals within an ecosystem. We then obtain the relationship between the similarity parameters under the assumptions of (i) that the ecosystem is in a dynamically balanced steady state and (ii) that these functional groups are connected to each other by the flow of resource. The expression that we obtain relates the level of complexity that the ecosystem can support to intrinsic macroscopic variables such as density, diversity and characteristic length scales for foraging or dispersal, and extrinsic macroscopic variables such as habitat size and the rate of supply of resource. This expression relates these macroscopic variables to each other, generating commonly observed macroecological patterns; these broad trends simply reflect the similarity property of ecosystems. We thus find that details of the ecosystem function are not required to obtain these broad macroecological patterns this may explain why they are ubiquitous. Departures from our relationship may indicate that the ecosystem is in a state of rapid change, i.e., abundance or diversity explosion or collapse. Our result provides normalised variables that can be used to isolate the trend in one ecosystem variable from another, providing a new method for isolating macroecological patterns in data. A dimensionless control parameter for ecosystem complexity emerges from our analysis and this will be a control parameter in dynamical models for ecosystems based on energy flow and conservation and will order the emergent behaviour of these models. | physics.bio-ph | physics | Similarity analysis of macroecology
S. C. Chapman,1 N. W. Watkins,1, 2, 3 G. Rowlands,1 A. Clarke,3 and E. J. Murphy3
1Physics Department, University of Warwick, CV4 7AL, UK
2London School of Economics and Political Science, London, UK
3British Antarctic Survey, NERC, High Cross,
Madingley Road, Cambridge CB3 0ET, UK
Abstract
We perform a full similarity analysis of an idealized ecosystem using Buckingham's Π theorem
to obtain dimensionless similarity parameters given that some (non- unique) method exists that
can differentiate different functional groups of individuals within an ecosystem. We then obtain
the relationship between the similarity parameters under the assumptions of (i) that the ecosystem
is in a dynamically balanced steady state and (ii) that these functional groups are connected to
each other by the flow of resource. The expression that we obtain relates the level of complexity
that the ecosystem can support to intrinsic macroscopic variables such as density, diversity and
characteristic length scales for foraging or dispersal, and extrinsic macroscopic variables such as
habitat size and the rate of supply of resource. This expression relates these macroscopic variables
to each other, generating commonly observed macroecological patterns; these broad trends simply
reflect the similarity property of ecosystems. We thus find that details of the ecosystem function
are not required to obtain these broad macroecological patterns this may explain why they are
ubiquitous. Departures from our relationship may indicate that the ecosystem is in a state of
rapid change, i.e., abundance or diversity explosion or collapse. Our result provides normalised
variables that can be used to isolate the trend in one ecosystem variable from another, providing a
new method for isolating macroecological patterns in data. A dimensionless control parameter for
ecosystem complexity emerges from our analysis and this will be a control parameter in dynamical
models for ecosystems based on energy flow and conservation and will order the emergent behaviour
of these models.
3
1
0
2
y
a
M
9
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
2
0
8
4
.
8
0
1
1
:
v
i
X
r
a
1
I.
INTRODUCTION
The rapid increase in the availability of large-scale ecological data [1, 2] has increased
knowledge of global patterns and stimulated the search for the underlying processes that
determine them (see [3]). Examples of the large-scale, macroecological, patterns [4] to have
emerged from empirical analyses include the species-area and the species-latitude relation-
ships, and trends in density and diversity with body size. These are broad scale patterns and
generalised rules rather than mechanistic processes, and a range of theories have been pro-
posed to explain why these patterns emerge. This has led to the development of a perspective
in which the detailed biological characteristics of the species (traits) do not determine their
abundance and that the processes affecting community structure can be considered as neu-
tral [5 -- 7], see also [8, 9]. The fact that these patterns are both approximate and ubiquitous,
suggests that they do not reflect the details of how ecosystems operate, rather that they
emerge from underlying general constraints.
The diversity seen throughout ecology has meant that the very possibility of existence
of universal laws remains controversial [10 -- 12].
In stark contrast, universal laws are an
essential feature of physics. One reason for this dichotomy is the covariance principle, which
asserts that the laws of physics can be represented in a form equally valid for all observers
[13 -- 16]. Any physical principles relevant to a classical natural system thus cannot depend
on arbitrarily chosen units [17] i.e.
they must be expressible in terms of dimensionless
parameters which are invariant under a change of fundamental units of measurement. It
implies that the functions describing the behaviour of a system must have arguments which
are dimensionless quantities, known as "similarity parameters" Πi [15, 18]. Underpinning
even this is a deeper classical measurement postulate, that of the possibility of controllable,
repeatable, and observer-independent measurement. Furthermore in physics the existence of
symmetries gives rise to conservation principles governing the relationship between physical
quantities. Such conservation properties will constrain the relationships which can exist
between the Πi, and will therefore give rise to patterns in these quantities. Ecological
patterns are seen and measured in a range of quantities, some of which have physical units.
A natural question [19, 20], that we address here, then arises: to what extent do physical
processes determine the major trends observed in macroecology?
We now briefly review the essential idea of similarity analysis with reference to a well
2
FIG. 1: Two systems with the same similarity properties. Von Karman vortex street seen in (a)
Landsat 7 satellite image of an island in the clouds and (b) laboratory fluid experiment.
known example in fluid dynamics. Figure 1 shows two realizations of a von Karman vortex
"street". These are on quite different scales, they are (a) cloud bearing airflow over an
island at sea and (b) a laboratory based experiment in water (for experiments of this type,
see [21]). The full dynamics is strongly non-linear and the functional form of the observed
vortex street pattern is non-trivial to obtain. Nevertheless, we can see that these patterns
are quite similar and we can characterize the vortex street by two lengthscales, the width
of the pattern, or the obstacle around which the fluid is flowing (Lo), and the distance
downstream between successive vortices (Ld). Figure 1 (a) and (b) look similar because the
ratio of these lengthscales, Lo/Ld is similar, despite the fact that Lo is tens of kilometers for
the island, and is millimeters to centimeters in size in the laboratory. The ratio Lo/Ld is just
the dimensionless Strouhal number, S = f Lo/U in fluid dynamics, since Ld = U/f where U
is the background flow speed and f is the frequency at which vortices are shed at the obstacle
(in air, the "aeolian tone"). Clearly, Lo and Ld, (or L0, U and f ) are governing parameters
3
that characterize the observed vortex street pattern that is formed, with a corresponding
dimensionless similarity parameter that is just the Strouhal number. The question that arises
is, what determines this observed pattern, specifically, why are these two situations at very
different spatial scales, similar? Now, there are other governing parameters for fluid flow.
Let us consider one other, the kinematic viscosity ν (physical dimension, (length)2/time).
A dimensionless parameter involving ν, is Re = ULo/ν, the Reynolds number.
If there
are no other governing parameters, then the observed vortex street pattern, specifically, its
Strouhal number, can only depend upon the Reynolds number, that is, S = F (Re) [22]. This
result can be tested experimentally by plotting S versus Re. Such a plot reveals that the
Strouhal number has only weak dependence on the Reynolds number for Re ∼ 103 − 106, for
which S ∼ 0.2. This is approximately what we see on Figure 1 where the Reynolds numbers
are at either end of this range.
Rayleigh [22] used similarity analysis to obtain the similarity parameters Re and S for
the vortex street without obtaining a model or mathematical form for the detailed nonlinear
function that describes them. The origin of the precise functional relationship between
Re and S (S = F (Re)) remains a topic of current research [23, 24]. Notably, Rayleigh's
procedure [25] was to (i) identify the similarity parameters, Re and S then (ii) note that
these parameters ordered the experimental data and finally (iii) propose an approximate
form for F that modeled the observed dependence of S upon Re. An important corollary
of Rayleigh's result is that the observations, when plotted in dimensional units, will show
patterns or trends, for example, the frequency f will increase linearly with flow speed U,
for a given size obstacle. Furthermore, if data from a number of experiments with different
obstacle size are aggregated, such a plot will still show this trend, but with considerable
scatter. Any organizing principle based on physical properties will thus only clearly emerge
if the dataset is plotted in terms of dimensionless similarity parameters. In this paper we
will perform this important first step, we will obtain similarity parameters for ecosystems.
This will suggest a new method to test this data for such organizing principles, and we
suggest may explain observed macroecological patterns or trends.
Importantly, similarity does not presuppose self-similarity but does encompass it. Self-
similarity is the property of generalized power law dependence where there is no charac-
teristic scale, (e.g [15] pp 86ff; [26]). Scaling relationships between subsets of observed
macroecological variables have been proposed (see e.g. [28, 29]) and self-similarity is intrin-
4
sic to some models of metabolism, dispersal and foraging [30]. Although an important topic,
self-similarity is an extra assumption that we do not need to make here, and are not making.
Although most familiar in physics and engineering [16], the application of dimensional
techniques to biological problems dates back to Galileo's application of similarity arguments
to the load bearing capacity of bones [31]. Dimensional arguments find widespread applica-
tion in the scaling of numerical simulations in biology, as elsewhere. By contrast, as noted
by [19] and more recently [20], only very limited explorations have been made so far of the
use of full similarity analysis to codify the set of possible relationships between macroscopic
variables in ecology. An interesting exception was the conjecture [32] that "life-like processes
[might] require a flux of energy above some minimal value in order to get going and keep
going". Lovelock made an explicit analogy with the observation by "Reynolds that turbu-
lent eddies in gases and liquids could only form if the rate of flow was above some critical
value in relation to local conditions."
In this paper we will perform such a similarity analysis of macroecological variables.
Rather than exploiting statistical constraints as in maximum entropy inference [6, 33 -- 35], we
focus on physical and dimensional constraints and how these in turn constrain the relation-
ships that can exist between intrinsic variables such as density, diversity and characteristic
lengthscales for foraging or dispersal, and extrinsic variables such as habitat size and the
rate of supply of resource.
The starting point in any similarity analysis is to identify the 'governing parameters'
[15], that is, the variables or quantities that are necessary for any description of the sys-
tem. Similarity analysis in essence takes these governing parameters as its input, and as its
output identifies the dimensionless similarity parameters. Here, we will use the extensive
body of observations in macroecology to specify a set of governing parameters. Importantly,
these parameters encompass both size based and species based approaches [36] to describing
an ecosystem. They are intrinsic variables such as the diversity of species (richness), the
density of individuals, characteristic length scales for foraging or dispersal, metabolic rate,
and extrinsic variables such as habitat size and the rate of supply of resource. Provided that
the governing parameters have physical dimension (dimensional units), similarity analysis
constrains the possible relationships that can exist between them. The functional relation-
ships we obtain here are between dimensional quantities, which in practice tend to arise
from approaches which focus on flows of materials and energy through the ecosystem.
5
Similarity analysis does not specify a unique relationship between these parameters. We
can only obtain this if in addition there are known physical constraints, such as conservation
principles. The flow of energy and resource has been one of several key threads of ecosystem
analysis [37] since for example the work of Lotka, Lindeman and the Odums [38 -- 40]. The
physical constraint that is central to this paper is that we consider ecosystems which adapt
dynamically to changes in external parameters to maintain a balance between the rate of
uptake and of utilization of resources taken over the ecosystem as a whole. This is our key
assumption. This does not mean that the system is in a fixed state, only that over the
time scales being considered the inputs/outputs are balanced and this balance constrains
the overall structure and functioning of the ecosystem. The concept of ecosystems which
compensate dynamically to remain in homeostasis has been explored previously [41, 42] but
not developed using dimensional analysis.
In this paper we adopt the formal definition of an ecosystem; that the different functional
groups of individuals within the ecosystem are connected to each other by the flow of re-
source. All individuals within the ecosystem ultimately derive resource from that taken up
by that ecosystem's primary producers. We assume that the ecosystem is in a dynamically
balanced steady state in that the total rate of uptake of resource is just balanced by the
rate summed over the ecosystem at which it is utilized. We will show that this is suffi-
cient to obtain an expression that constrains the relationships that can exist between these
(dimensional) ecosystem variables. We will see that this constraint is reflected in overall
macroecological trends that are observed. Further, we obtain the relationship between these
trends. At minimum this determines for the first time the dimensionless, or normalised
variables that isolate trends in one ecosystem variable from another. This new method for
isolating macroecological trends provides the basis for understanding dependencies between
factors such as size and metabolic rate that quantify the flow of resource, and factors that
categorize individuals into distinct species and types. The more resource that is available,
the more complex an ecosystem can in principle be, in the sense that more distinct species
and types, and relationships between them, can be supported. At maximum, we obtain a
dimensionless control parameter and relate it to this level of complexity that the ecosystem
can support. This points towards 'thresholds for life', that is, a parametrization of the min-
imum level of complexity that can potentially be supported by an ecosystem in dynamic
balance.
6
Process based models of ecosystems that aim to predict macroecological patterns tend to
fall into two approaches. The first of these relates area, diversity and abundance by means
of models for dispersal, occupancy and coexistence/competition and speciation in physical
space (see e.g. [3, 7, 43, 44]). The second relates abundance, body mass and metabolic rate
by means of models and constraints for the availability and flow of resource ([28], see the
review of [29]). The relationship that we obtain here links the key parameters of both these
approaches, suggesting a synthesis of them. It is important to emphasize at the outset that
we do not present a theory of how ecosystems work, nor do we develop an ecosystem model.
Rather we use formal similarity analysis to explore the constraints that must operate on the
variables observed in real ecosystems, given the single key assumption of balance between
the ecosystem summed rates of resource uptake and utilization.
As noted above many of the specific relationships between variables that we examine
are known and been explored [4 -- 7]. These include the species-area, species-latitude and
productivity-diversity relationships. The major insight we provide is to show how these
different relationships are related and under what conditions they will emerge. We also
show that they are an expected consequence of the physical constraints on the system.
II. SIMILARITY ANALYSIS AND A BOTTOM UP-APPROACH TO AN
ECOSYSTEM
The Π theorem formalises the principle of similitude as follows. Any physical system that
depends upon V variables, (the governing parameters), Q1, Q2, ..QV will have a function F
that relates them: F (Q1, Q2, ..QV ) = C where C is some dimensionless constant [15]. The
physical system that we discuss here is that which captures general physical aspects of
ecosystem function, specifically the uptake and utilization of resource. The essential idea
of similarity analysis is that this function can only depend upon dimensionless similarity
parameters Π1, Π2, ..ΠM so that F = F (Π1, Π2, ..ΠM ) only. These dimensionless parameters
Π1..M (Q1..V ) are formed directly from the governing parameters Q1, Q2, ..QV . Similarity
analysis as in Buckingham's Π theorem ([13 -- 15], see also [45]) is simply the process to
obtain the similarity parameters, that is, these dimensionless groups of variables.
If one
then has additional information about the system, such as a conservation property, the
Π1..M (Q1..V ) can be related to each other to make F explicit. If the V governing parameters
7
Q1..V are expressed in W physical dimensions (i.e. mass, length, time) then from [13] there
will be at least M = V − W dimensionless similarity parameters or groups Π1..M (Q1..V )
which we now progressively identify.
We will build our understanding by first considering the simplest possible idealized ecosys-
tem and then successively increasing the level of complexity; at each stage, additional gov-
erning parameters (ecosystem variables) are introduced. Our approach is to use similarity
analysis at each stage to find the constraints that act on this general description of an
ecosystem. Importantly, we seek to describe an 'observed macroscopic ecosystem', that is,
the variables that we will ultimately identify include observed intrinsic properties such as
density, diversity, body size, and metabolic rate.
Introducing progressively more specific
detail inevitably introduces more governing parameters- this procedure could be taken fur-
ther to explore specific detailed ecosystem models by the input of more detailed ecosystem
functional properties into the dimensional analysis. Our aim here is rather to explore the
ecosystem constraints that emerge for the minimum set of assumptions and model inputs.
A. One kind of uniformly distributed single cell organism.
The ecosystem is composed entirely of single cell organisms that are of uniform type:
they have the same function and structure and same typical metabolic rate R, dimensions
of power [M][L]2[T ]−3. They are uniformly distributed in a habitat of size L in D dimensions
with density n, dimensions [L]−D. The available resource (sunlight) is delivered at rate P
per unit area, dimensions power per unit area [M][T ]−3 and the (dimensionless) fraction α
taken up averaged over the ecosystem is a constant. A schematic of such a system is shown
in Figure 2 (a). We recognize that such a simple system does not exist in reality. However,
this simple theoretical construct is an important first step in the analysis, from where we
can move to consider more realistic ecological scenarios.
There are 5 governing parameters (R, L, P, α, n) and 3 physical dimensions (mass [M],
length [L], time [T ]) so we have 2 similarity parameters (dimensionless groups) which are
Π1 =
αP L2
R
, Π2 = nLD
(1)
These are related by the physical constraint that the system is in dynamical balance so that
the rate at which energy is taken up over the ecosystem is the rate at which it is utilized,
8
FIG. 2: Bottom up ecosystem (a) only one type of single cellular organism can be distinguished
(b) 2 types of single cellular organism can be distinguished.
so that:
αP L2
R
= nLD
(2)
which is Π1 = Π2. This expression simply tracks the flow of energy into and through the
ecosystem- it assumes that all other processes necessary for the ecosystem to function, such
as the recycling of resources such as Nitrogen, occur. Introducing a typical metabolic rate for
the single cell organisms has fixed an energetic minimum threshold for life which is Π1 = 1,
that is, one cell in the habitat (one cell ecosystem).
B. More than one kind of single cell organism.
We next consider single cell organisms that can be differentiated from each other, either
by their function or their structure, or both. These would represent distinct types or species
and different methods for categorizing and distinguishing individuals will yield different sets
9
of species. What will follow will be independent of the precise details of this differentiation
method, we only need assume that such a differentiation is now possible. A schematic of
such a system is shown in Figure 2 (b). In the ecosystem there are S types (species) and
there is a species label k = 1..S, the density of the kth species is n(k). We can always define
an average density of the single cell organisms:
¯n =< n(k) >k=
1
S
S
Xk=1
n(k)
(3)
so that the variable n in (1) is now replaced by ¯n and Π2 = ¯nLD. The additional variable
S is dimensionless; so that we now have 6 governing parameters and 3 dimensions, and so
3 dimensionless similarity parameters with Π3 = S. These are again related by the physical
constraint that the system is in dynamical balance:
αP L2
R
=
S
Xk=1
n(k)LD = ¯nSLD
(4)
Expression (4) now encapsulates the idea of primary producers- one or more of the species
is responsible for the uptake of resource with efficiency α. The other species 'feed off' this
primary producer either by grazing, predation or uptake of waste.
C. Multicellular organisms
We now consider more complex organisms that are multicellular. All the organisms
live in an ecosystem and are connected to the primary producers by the flow of resource,
either directly or indirectly by predation, or both. It is now possible to distinguish types
of organism and we will label the different types or categories distinguished in this way
with index p. Again, the results to follow will not depend upon the precise details of
how individuals are assigned to any of the p categories, simply that such an assignment
can be made. Different methods for categorizing the individuals in an ecosystem [46] will
organize individuals into groups or categories of different p, this categorization may focus
on the functional role of individuals in the ecosystem such as niche or trophic level, or
may focus on stage of development, or other factors. Organisms falling into a given pth
category or group will be clustered around an average body size, on average they will be
composed of B(p) cells (this is typically observed [4]), and will have average metabolic rate
RB(p), the per-cell metabolic rate R now corresponds to an ecosystem average over these
10
multicellular organisms. There is a non- trivial correspondence between average size B(p)
and how complex an organism can be. Within each p there will be a number k = 1..S(p)
of differentiable species each with density n(k, p) all with average body size B(p) and with
average density, for that p of
¯n(p) =< n(k, p) >k=
1
S(p)
S(p)
Xk=1
n(k, p)
(5)
Our governing parameters, density and diversity, are now taken to relate to the obser-
vation of a given category p∗ that is embedded in the ecosystem; we observe n∗ = ¯n(p∗),
S∗ = S(p∗).
Individuals in the observed category also have a characteristic average size
B∗ = B(p∗) (number of cells so dimensionless) which introduces an additional governing
parameter giving a total of 7 governing parameters and hence 4 dimensionless similarity
parameters:
Π1 =
αP L2
R
, Π2 = n∗LD, Π3 = S∗, Π4 = B∗
The physical constraint of a dynamically balanced ecosystem is now:
αP L2
R
S(p)
Xk=1
= Xp
n(k, p)B(p)LD = Xp
¯n(p)S(p)B(p)LD
We can write (7) in terms of our observed pth
∗ category:
αP L2
R
= n∗S∗B∗LDXp
¯n(p)S(p)B(p)
¯n(p∗)S(p∗)B(p∗)
We then have an expression of the form:
αP L2
R
= n∗S∗B∗LDΨ(p∗)
(6)
(7)
(8)
(9)
where Ψ(p∗) is a dimensionless function; 1/(Ψ(p∗)) is the fraction of the total rate of resource
supplied to the ecosystem that is utilized by the observed (pth
∗ ) category. Importantly, all of
the species and categories are connected into the same resource flow, so that fundamentally,
(9) will constrain how observed density, diversity and body size are related to each other
across the ecosystem, with consequences for macroecological patterns as we will discuss.
These dimensionless similarity parameters (6), and their relationship (9) express the
following fundamental properties of the idealized ecosystem. There is a building block on
which life is organized- the single cell which has a definable typical (ecosystem average)
metabolic rate, R. There is then the physical property of resource flow: that the single cell
11
metabolic rate, along with the rate of uptake of resource to the ecosystem αP L2 constrains
the number of cells the ecosystem can support, which is Π1. The detailed biological and
ecological properties of the ecosystem then organise these Π1 cells into a complex network
of species and groups of species, observationally these are characterized into functional
units which have an average body size, density and diversity. Hence, the observed ecological
variables are in a macroscopic sense related to each other by the physical property of resource
flow.
D. Non uniform distribution of individuals in space.
Generally, organisms will not be uniformly distributed in space so that the observed
density depends on the length scale r over which it is observed, so that n = n(r, k, p) and
similarly, the efficiency of the primary producers, which depends on their density, is α = α(r).
The lengthscale over which the density varies can either arise from how individuals subdivide
and grow, forage, or other forms of influence they have on each other and on the environment.
This is important since in (7-9) the density refers to that measured over the habitat of the
entire ecosystem on scale L and any observation will be on a more local scale r << L, which
in turn relates to the lengthscale of over which the density varies. This will introduce a
variable for the scale of observation of the pth
∗ category; r∗ (dimension [L]) finally giving 8
governing parameters so 5 dimensionless similarity parameters:
Π1 =
α∗P L2
R
, Π2 = n∗LD, Π3 = S∗, Π4 = B∗, Π5 =
r∗
L
(10)
where α∗ = α(r∗) so that all variables refer to a consistent set of observations on lengthscale
r∗. The density can be generally expressed as n(r, k, p) = n(L, k, p)/g(k, p, r/L) where g
is dimensionless and expresses the spatial variation of the pth category; similarly α(r) =
α(L)/gα(r/L).
If the categorization p is based on function and structure then one can
anticipate that an average of g over the S(p) species in the category is meaningful so that:
1
S(p)
S(p)
Xk=1
n(r, k, p)
= ¯n(L, p)/¯g(p, r/L)
(11)
¯n(r, p) =< n(r, k, p) >k=
=
1
S(p)
S(p)
Xk=1
n(L, k, p)
g(k, p, r/L)
12
If we explicitly reference lengthscale L, expression (8) is:
α(L)P L2
R
= n∗(L)S∗(L)B∗(L)LDXp
¯n(p, L)S(p, L)B(p, L)
¯n(p∗, L)S(p∗, L)B(p∗, L)
(12)
for a consistent set of observations for all p categories on the same lengthscale r∗ this is:
α(r∗)gα(r∗/L)P L2
R
= n∗(r∗)g∗(r∗/L)S∗(r∗)B∗(r∗)LDΨ(p∗, r∗)
(13)
or writing the spatial variation in a single function G(r/L) = ¯g(r/L)/gα(r/L)
α∗P L2
R
= n∗S∗B∗G(
r∗
L
)LDΨ
(14)
Spatial variation in density thus leads to spatial trends in diversity, or species- area rules;
this has arisen quite generally as a consequence of the physical constraint (14) and we will
discuss this next.
Essentially, (14) is:
Π1 = Π2Π3Π4G(Π5)Ψ
(15)
The dimensionless functions Ψ and G contain all the details of the ecosystem function;
Ψ encapsulates the details of how individuals in the ecosystem are categorized and G the
details of how these categories of individuals are dispersed in space. Our expressions (14)
and (15) do not specify a particular model for an ecosystem.
Instead, they specify the
relationship between available resource, and the level of complexity that the ecosystem can
attain. The diversity and complexity possible in an ecosystem is constrained physically by
the total living biomass within it, and this in turn is constrained by the rate of uptake, and
utilization, of resource. We thus identify the ecosystem control parameter Π1 = αP L2/R
namely (productivity) × (habitat size)/(typical metabolic rate) which is just the number
of 'typical' cells the ecosystem can support (we can always define an ecosystem averaged
metabolic rate per cell R). This control parameter constrains the level of complexity that
an ecosystem can support in the sense that it constrains the number of different possible
configurations, or ways that this total living biomass can be arranged into distinct forms
of life. For example, if there is only one distinguishable kind of organism in the ecosystem,
there is only one p and S value and Ψ = 1, we essentially recover (2) where the average
metabolic rate of the organisms is RB∗. The threshold for one such organism of size B∗
to be supported by the ecosystem is Π1/Π4 = 1 or α∗P L2/(RB∗) = 1. As the ecosystem
becomes more complex, Ψ > 1 and each p category utilizes a smaller share of the total
13
resource supplied. The dimensionless function Ψ thus operates as an order parameter of the
ecosystem which reflects the level of complexity.
Importantly, the function Ψ also incorporates the method of categorization. The com-
plexity of an observed ecosystem inevitably depends in part on how the observed data are
categorized. However, the observed values, i.e., the observed density, diversity and so forth
of a given category, will depend on how the observer defines that category, i.e. what organ-
isms are included in it. These ideas can be used to re-order the observations to understand
the role played by how the data is categorized as we discuss in the appendix.
A physical analogy to this is the relationship between the Reynolds number in turbulence
and the number of excited modes or degrees of freedom. The Reynolds number is the ratio
of a rate of energy input on the largest, driving scale to a rate of energy dissipation on the
smallest scale, as is Π1 here, and similarity analysis, along with the assumption of steady
state (no energy pile up) is sufficient (see eg [45]) to constrain the number of degrees of
freedom to grow with increasing Reynolds number.
III. CONSTRAINTS ON MACROECOLOGICAL PATTERNS WITHIN AND
ACROSS ECOSYSTEMS
Observations both within and across ecosystems consist of specifying a method for classi-
fying individuals into particular groups or categories and then for each of the pth categories,
observing the average density, diversity, bodysize, and metabolic rate. From the constraint
(14): we see that these variables are not independent, and (14) suggests relationships be-
tween them which we will now discuss.
Let us consider that a scheme for classifying individuals is consistently adopted, and
observations of average density, diversity, bodysize, and metabolic rate of these categories
are made. These observations simultaneously collect a range of values of n∗, S∗, B∗ for a
given r∗, L and P . We will first consider the case where 'similar' ecosystems are compared,
or where a comparison is made within a single ecosystem, that is, the order parameter Ψ is
not varying. Subsets of the variables in equation (14) will then show functional relationships,
this has been found for example by [47, 48] who demonstrate a relationship between species
richness, area and a measure of productivity.
14
For 'similar' ecosystems then, expression (14) is
α∗P L2
R
= n∗S∗B∗G(
r∗
L
)LD
(16)
This constrains overall patterns or trends, it defines a single multivariate surface in the
variable space of density, diversity and so forth. The observed macroecological patterns are
paths on this surface. The surface then relates these macroecological patterns to each other
through the single expression, equation (16). As a first example, let us consider trends that
can occur as the available resource over the ecosystem αP L2 is varied. From expression (16),
one cannot have arbitrary increase in density and diversity with resource, it is constrained.
Thus (holding all other variables constant) if resource rate of uptake doubles, and density
doubles, diversity cannot increase. If resource is increased by a factor A, and the number
of species doubles, then the density can only increase by factor A/2. This will be the case
for any model which has our assumption of a dynamically balanced steady state. This
constraint on how ecosystem properties such as density and diversity can vary is our main
result. This points to a need to isolate changes in one variable from another and we will
provide a method for this.
We can formalize these constraints as follows. Expression (16) is
α∗P L2 = S∗(cid:20)Rn∗B∗G(
r∗
L
)LD(cid:21)
(17)
An increase in diversity with total net productivity integrated over the habitat is Wright's
Rule (the species-energy relationship) [49]. However, to only see an increase in diversity,
one would also need the contents of [...] to be constant, that is, K constant in:
Rn∗B∗G(
r∗
L
)LD = K
(18)
Thus if a set of observations are indeed across an ecosystem, in the sense that all the observed
categories are linked by the flow of resource, then when Wright's rule is seen, the 'resource
flow constraint' (18) should also be seen.
The general relationship that we have derived (16) presents, for the first time as far as
we are aware, a view of how the major physical and biological variables that determine
key aspects of the structure and functioning of ecosystems are related. It relates the scale
and energy turnover to the variables that describe the internal structure of the ecosystem
(the complexity). The relationship emerges as a result of the dimensions of the underlying
15
variables, but is also constrains and gives the wider context for the relationships between
specific variables. It shows that although specific variables are related, the relationship is
dependent on the modifying effects of other interacting variables. The variables are not
independent but instead can co-vary and equation (16) provides the basis for understanding
how they are expected to interactively affect the relationships between specific variables.
This is a crucial point as it highlights why particular relationships can emerge only under
particular conditions.
In the following we consider a set of well known macro-ecological
relationships to show how they emerge from our analysis and also what equation (16) tells
us about how the relationship is affected by the other key variables. For each of these
macro-ecological relationships there will be a resource flow constraint in the sense of (18)
which we will now identify:
• Diversity and Wright's Rule: S∗ ∝ αP L2 as in (17) so that the number of species
(diversity) increases with total net productivity integrated over the habitat rather
than productivity alone; this is Wright's rule [49]. Whilst Wright's rule is to some
extent ecologically trivial (a greater net energy input allows more individuals, see [50])
the interesting aspect of this result is that it predicts an increase in diversity (richness)
and not just individuals. As we would anticipate from the resource flow constraint(18)
the relationship between productivity and species diversity also varies with spatial
scale as is found [51]. Equation (17) also suggests that the internal configuration of
the ecosystem in terms of density, biomass and metabolic rate of the species present
will affect the relationship.
• Diversity and metabolic rate: S∗ ∝ 1/R since:
S∗ =
1
R " α∗P L2−D
L )#
n∗B∗G( r∗
(19)
so that diversity decreases with increasing metabolic rate: we expect ecosystems domi-
nated by endothermic organisms with high metabolic rate to have lower diversity than
flow constraint is now n∗B∗G( r∗
those dominated by ectothermic, low-metabolism, organisms (e.g. [4]). The resource
L )LD−2/(α∗P ) = K constant, which specifies how vari-
ation in the scale and productivity of the ecosystems considered could mask the effects
of changing metabolic rate.
16
• Latitudinal gradient rule: Diversity will also increase with resource, since:
S∗ =
(20)
L )LD−2i
L )LD−2 = K constant. Provided that
other factors, i.e. α∗, that link resource uptake to available sunlight do not vary [50] our
The resource flow constraint is now Rn∗B∗G( r∗
α∗P
hRn∗B∗G( r∗
general macroecological relationship encapsulates the latitudinal gradient rule. Again,
this trend is present alongside patterns in the other variables when K is not constant,
as discussed by [52]. For example, (20) predicts that low metabolic rate ecosystems
where the rate of resource supply is high will be more diverse that high metabolic rate
ecosystems where the rate of resource supply is low; this may suggest a refined version
of the latitudinal gradient rule and allow comparison of diverse ecosystems.
• Species Area Relationships: A corollary of length-scale dependence of the density is
that diversity will vary with habitat size (which is a function of L) since:
S∗ =
L2−D
G( r∗
L ) (cid:20) α∗P
Rn∗B∗(cid:21)
(21)
More explicitly, diversity will vary both with habitat size (which is a function of L) and
the lengthscale of the observation or characteristic lengthscale of some process (which
is a function of r∗); these are known as Species Area Relationships (SAR) (see eg [53]).
Thus if the individuals grow in clumps, say by division, or live on a fractal structure
(tree, coral, mountain, river) or forage in a random walk pattern (21) will constrain
the resulting ecosystem SAR. For example power law SAR arise if available productive
surface area or volume orders the availability and uptake of resource[54] and that this is
in turn ordered by the roughness of the terrain which can be modeled simply as a fractal
[55, 56], see also [57]. The resource flow constraint is α∗P/(Rn∗B∗) = K constant.
Thus these SAR and the underlying constraint on dispersal and clumping from which
they originate are also found to interact with other variables such as productivity [59]
and bodysize [60]. Observations over the largest regional or continental scales tend
to integrate or aggregate over detailed spatial dependence and over other variable
factors such as metabolic rate. These scales exceed that over which the terrain varies,
and over which processes occur that yield spatial clumping, over such large scales
the effective G → 1 and the landscape is essentially 'flat' so that D = 2. Hence on
17
these largest scales the spatial dependence (21) vanishes, we have that P ∼ S∗, and
positive relationships between productivity and diversity emerge as has been found
[51]. Within a given ecosystem, diversity and abundance will also both vary with the
r∗ over which they are observed, as well as with each other. If all other variables are
not strongly varying, their functional dependence on r∗ can be obtained from the data
by the method of scaling collapse as has been done by [58].
• Abundance, the 'more individuals' hypothesis: The abundance (density of individuals
in each species) increases with productivity integrated over the habitat and decreases
with the typical metabolic rate (e.g. [4]), since:
(22)
n∗ =
α∗P
R
1
hS∗B∗G( r∗
L )LD−2i
As we increase the total ecosystem energy uptake rate, from equation (22) both the
diversity and abundance can increase (for fixed metabolic rate). We will find the above
patterns when one effect does not dominate, for example, Wright's rule will not be seen
if the increase is entirely in abundance, and not in diversity. The relationship between
the number of individuals and the number of species has long intrigued ecologists, and
whilst a positive relationship is sometimes assumed (the 'more individuals' hypothesis
for the increase in richness with overall abundance: see [50]) the detailed mechanism(s)
involved are far from clear.
• Abundance and diversity decrease with increasing with body size: our expression (16)
gives an inverse relationship between abundance, diversity and bodysize
L )LD−2#
L )LD−2) is constant or weakly vary-
ing, indeed the average abundance [61 -- 65] and diversity [4] are found to have statistical
provided the resource flow constraint α∗P/(RG( r∗
B∗ "
n∗S∗ =
(23)
1
α∗P
RG( r∗
trends that decrease with increasing average body size.
• Trends in trophic level and nett productivity: the number of trophic levels that the
ecosystem can support is seen to increase with total net productivity summed over the
habitat and decreases with the typical metabolic rate [4]. This follows since we have
from (14):
B∗Ψ =
α∗P L2
1
hn∗S∗G( r∗
L )LDi
R
18
(24)
Either bodymass relates to trophic level, or the number of trophic levels may be
determined by the level of complexity Ψ which increases with total net productivity
summed over the habitat and decreases with the typical metabolic rate. This pattern
will be seen provided n∗S∗G( r∗
L )LD is constant.
The above relationships (17)-(24) show that well known macroecological relationships
emerge from the single expression (14) that we obtained from formal dimensional analysis.
This provides a clear physical and ecological basis for understanding why particular rela-
tionships exist and have been identified in a wide range of different studies of ecological
systems [1]- [12]. However, our analysis goes further to clarify how these different relation-
ships are related together and are affected by changes in other variables that are crucial for
determining the structure and functioning of ecosystems. It demonstrates how the variables
co-vary and why important macroecological relationships may emerge only under particular
conditions. We thus expect these patterns to emerge most clearly when the corresponding
resource flow constraint is slowly varying and this can be tested for in data. It also suggests
that taking account of that co-variation will allow a much more rigorous basis for analysing
available data to test for the existence of particular macroecological relationships and us-
ing those relationships to test our understanding of the factors that determine ecosystem
structure and function more generally. In the following section we illustrate the power of
this insight by examining how the co-variability can affect the capacity to detect particular
relationships.
A corollary of this is that we can use (14) to identify a method to isolate these patterns.
A particular example of this is testing for SAR. From equation (21), to test for a SAR one
should plot the normalized diversity ¯S versus L:
¯S =
S∗n∗RB∗Ψ(p∗)
α∗P
=
L2−D
G(r∗, L)
(25)
Such a comparison can be made if Ψ(p∗) is not strongly varying ('similar' ecosystems), or if
Ψ(p∗) can be found from the data by the method described in the appendix.
We illustrate this process in Figure 3 where we have modeled synthetic data for a species-
area comparison. We have generated synthetic data in the same manner as described in the
appendix, such that there are trends in abundance, species richness, and body size, and
also random scatter in all variables, constrained such that all the sampled categories of
individuals share the same function Ψ. In addition each group of data is from a different
19
103
102
101
S
y
t
i
s
r
e
v
d
i
100
100
103
S
y
t
i
s
r
e
v
d
i
102
101
100
102
100
10−2
n
y
t
i
s
n
e
d
/
P
α
Ψ
B
R
n
S
y
t
i
s
r
e
v
d
102
104
body size B
i
d
e
z
i
l
a
m
r
o
n
102
100
area L2
10−4
100
102
101
100
10−1
102
104
body size B
100
area L2
102
FIG. 3: Log-log plot of a synthetic dataset generated for categories of individuals from habitats of
different areas. The dataset is constructed with power law dependence of diversity on lengthscale
and trends in abundance, diversity, body size and area with random scatter. This pattern is only
revealed in a plot of dimensionless diversity versus area.
habitat size and has a dependence on area predicted by our result (14) with power law
dependence of G on L. This simple illustrative exercise demonstrates that an underlying
clear pattern emerges in normalized diversity with a corresponding SAR pattern of diversity
versus area which has considerable scatter.
IV. CONCLUSIONS
We have used a 'bottom up' approach to fix the minimum set of governing parameters
needed to specify a generic idealized ecosystem and have used these to perform a similar-
ity analysis of the ecosystem. Physical constraints of energy flow and utilization over the
ecosystem then relate the similarity parameters, which in turn gives an expression which
relate the level of complexity that the ecosystem can support to intrinsic variables such
as density, diversity and characteristic lengthscales for foraging or dispersal, and extrinsic
variables such as habitat size and the rate of supply of resource. These constraints hold
regardless of the details of how a given ecosystem functions and require only the assump-
tion that the ecosystem is in a dynamically balanced steady state, that is, that the total
rate of resource uptake is balanced by the rate of resource utilization summed over the
20
ecosystem. We thus find the constraint on the relationships that can exist between these
(dimensional) ecosystem variables which is reflected in observed macroecological patterns.
Our result may explain why these general, approximate statistical trends appear to be so
ubiquitous in nature: we obtain these patterns without recourse to any detailed information
about the structure or dynamics of ecosystems or indeed how the data are collected. They
simply reflect the underlying similarity properties of the ecosystem and energy conservation
in dynamical steady state.
Our result also shows how these different observed macroecological relationships are re-
lated to each other and how they are affected by changes in other variables, and hence why
particular macroecological relationships may emerge only under particular conditions. This
leads to the dimensionless, or normalised variables that need to be constructed to isolate the
trend in one ecosystem variable from another; we thus provide a new method for isolating
macroecological patterns. Comparisons could thus be made between datasets by controlling
for (normalizing against) characteristic metabolic rate, abundance and diversity in order
to isolate the statistical pattern with respect to one of these variables. In particular this
method isolates a function that expresses how complex the ecosystem is and it would be
intriguing to order the data in this way to determine the level of complexity of ecosystems
that are found in nature, and to what conditions they correspond. An example would be
comparisons across extinct ecosystems, or between extinct and contemporary ecosystems,
provided a comparable sample group could be identified. The fact that Wright's rule, species
area rules and latitudinal gradient rules emerge often, but not always, from the observational
data gathered across ecosystems may reflect varying levels of complexity in these ecosys-
tems, or the effect of different schemes for categorizing individuals within ecosystems and
our results provide a method to control for this.
Departures from these statistical patterns where ecosystems are similar, and consistently
sampled, then may imply that the system is in a state of rapid change, i.e., abundance
or diversity explosion or collapse. Any ecosystem which is dynamically balanced in the
sense discussed above will fall within these macroecological patterns, it does not need to
be a climax or maximum energy utilization system but simply needs to balance the rate of
energy uptake with that of usage integrated over the ecosystem.
Finally, we have identified a dimensionless control parameter for ecosystem density and
diversity, namely (productivity) × (habitat size)/(typical metabolic rate) which emerges
21
quite generally from our dimensional analysis as a similarity parameter. This we suggest
will be a control parameter in dynamical models for ecosystems based on energy flow and
conservation and will order the emergent behaviour of these models. We relate this control
parameter to the level of complexity that a dynamically balanced ecosystem can support (its
order parameter). This control parameter is the ratio of energy input rate to the ecosystem
to the metabolic rate of the smallest possible unit of life, a single cell. If it is reasonable to
identify a smallest possible unit of life then this parameterizes the threshold at which life,
defined in this manner, can occur (c.f. Lovelock's conjecture about dimensionless control
parameters for prebiotic planets [32]).
Acknowledgments
The authors acknowledge the UK EPSRC, STFC and NERC for support. This study
is part of the British Antarctic Survey Polar Science for Planet Earth Programme. NWW
acknowledges valuable interactions with J. E. Lovelock and O. Morton.
Appendix
Any organizing principle based on physical properties will only clearly emerge if the
dataset is plotted in terms of dimensionless similarity parameters. We now demonstrate
the procedure to apply this method to macroecological data. As a starting point say we
have a set of observations based on individuals consistently organized into categories. For
each category we observe on lengthscale r∗ the density n∗, diversity S∗ and body size B∗,
so that for many such categories we have a set of observations of n∗, [n1, n2, ..nj...], of S∗,
[S1, S2, ..Sj..], and of B∗, [B1, B2, ..Bj..] where each n∗ and S∗ refer to size B∗ of the pth
∗
category. We first consider the case where these are all drawn from the same ecosystem and
observed on the same lengthscale so that we have the same α, P, L, R, G and r∗. There will
be a single function Ψ(p∗) which corresponds to this set of observations at different p∗. We
can write
Ψ(n∗S∗B∗) =
n′S ′B ′
n∗S∗B∗
Ψ(n′S ′B ′)
(26)
so that relative to any particular category n′S ′B ′ we can obtain Ψ(n∗S∗B∗) to within a
constant.
22
102
101
S
y
t
i
s
r
e
v
d
i
100
100
102
100
10−2
n
e
c
n
a
d
n
u
b
a
10−4
100
102
104
body size B
102
100
10−2
n
e
c
n
a
d
n
u
b
a
10−4
100
10
5
)
B
n
S
(
/
1
=
Ψ
102
104
body size B
101
102
diversity S
1
100
102
104
body size B
FIG. 4: Log-log plots of synthetic data generated for an ecosystem where all the observed indi-
viduals are constrained to share the same ecosystem complexity function Ψ. The set of data is
constructed to show trends in abundance, diversity and body size with random scatter. The con-
straint can only be discerned by plotting 1/SjnjBj versus body size Bj which to within a constant
is the function Ψ.
This procedure is illustrated in Figure 4 where we have modelled synthetic data. Our
synthetic data are generated such that there are power law trends in abundance, diversity
and body size, and also random scatter in all variables. Importantly this random scatter
is generated to be constrained such that all the sampled categories of individuals share the
same function Ψ. The plots show that the functional dependence of Ψ only emerges in
a plot of 1/S∗n∗B∗ versus body size B∗. This also offers a method to compare different
categorization schemes which for the same ecosystem could yield different Ψ(p); one could
then in principle normalize for (ie compensate for) any 'bias' introduced by a particular
choice of categorization.
[1] Brown, J. H., 1995. Macroecology, Univ. Chicago Press.
[2] Maurer, B. A., 1999. Untangling ecological complexity:the macroscopic perspective, Univ.
Chicago Press.
[3] Marquet, P. A., 2009. Macroecological perspecives on communities and ecosystems, 386-394,
in The Princeton Guide to Ecology, S. A. Levin, Ed., Princeton Univ. Press.
23
[4] Rosenzweig, M. L.,1995. Species diversity in space and time, Cambridge University Press.
[5] Hubbell, S. P., 2001. The unified neutral theory of biodiversity and biogeography, Princeton
University Press.
[6] Harte, J., T. Zillio, E. Conlisk, A. B. Smith 2008. Maximum entropy and the state variable
approach to macroecology, Ecology, 89, 2700-2711.
[7] McGill, B. J. 2010. Towards a unification of unified theories of biodiversity. Ecology Letters
13:627-642.
[8] Chave, J., 2004. Neutral theory and community ecology, Ecology Letters, 7, 241-253.
[9] Liebold, M. A., M. Holyoak, N. Mouquet, P. Amarasekare, J. M. Chase, M. F. Hoopes, R. D.
Holt, J. B. Shurin, R. Law, D. Tilman, M. Loreau, A. Gonzalez, 2004. The metacommunity
concept: a framework for multi-scale community ecology, Ecology Letters, 7, 601-613.
[10] Lawton, J. H., 1999. Are There General Laws in Ecology?, Oikos, 84, 177-192.
[11] Harte, J., 2002. Toward a Synthesis of the Newtonian and Darwinian Worldviews, Physics
Today, October, 29-34.
[12] Dodds, W., 2009. Laws, Theories, and Patterns in Ecology, University of California Press.
[13] Buckingham, E., 1914. On physically similar systems; illustrations of the use of dimensional
equations, Phys Rev., 4, 345.
[14] Buckingham, E., 1915. The principle of similitude, Nature, 96, 396-397.
[15] Barenblatt, G. I.,1996. Scaling, self-similarity, and intermediate asymptotics, Cambridge Uni-
versity Press.
[16] Bolster, D., R. E. Hershberger, R. J. Donelly, 2011. Dynamic similarity, the dimensionless
science, Physics Today, September pp42-47.
[17] Stephens, D. W., S. R. Dunbar, 1993. Dimensional analysis in behavioral ecology, Behav.
Ecol., 4, 172-183
[18] Barenblatt, G. I., 2002. Scaling, CUP.
[19] Rosen, R., 1989. Similitude, similarity and scaling, Landscape Ecology, 3, 207-216.
[20] May, R. M., 2007. Foreword to Scaling Biodiversity, D. Storch, P. A. Marquet and J. H. Brown
(eds.), Cambridge University Press.
[21] Perry, A.E., Chong, M.S., Lim, T.T. 1982. The vortex-shedding process behind two-
dimensional bluff bodies, J, Fluid Mech., 77-90.
[22] Rayleigh (Strutt, J. W., Third Baron Rayleigh), 1915a. The principle of similitude. Nature,
24
95, 66-68.
[23] Ponta, F. L. A. Hassan, 2004. Strouhal-Reynolds number relationships for vortex streets, Phys.
Rev. Lett., 93, 084501.
[24] Roushan, P. X. L. Wu, 2005. Structure-based interpretations of the Strouhal-Reynolds number
relationship, Phys. Rev. Lett., 94, 054504.
[25] Rott, N.,1992. Lord Rayleigh and Hydrodynamic Similarity, Phys. Fluids A4, 2595.
[26] Embrechts, P. and M. Maejima, 2002. Selfsimilar processes, Princeton University Press.
[27] Kolmogorov, A. N.,1941. Local structure of turbulence in an incompressible viscous fluid at
very high Reynolds numbers, C. R. Acad. Sci., 30, 301.
[28] Enquist, B. J., E. P. Economo, T. E. Huxman, A. P. Allen, D. D. Ignace, J. F. Gillooly, 2003.
Scaling metabolism from organisms to ecosystems, Nature, 423, 639-642.
[29] Brown, J. H., J. F. Gillooly, A. P. Allen, Van M. Savage, G. B. West, 2004. Toward a metabolic
theory of ecology, Ecology, 85,1771-1789.
[30] Viswanathan, G. M., M. G. E., da Luz, E. P. Raposo, H. E. Stanley, The physics of foraging,
CUP, (2011)
[31] Thompson D., W. Galileo and the principle of similitude, 1915. Nature, 95, 426
[32] Lovelock, J. E., Gaia, A new look at life on Earth, 1979. Oxford University Press.
[33] Pueyo, S., F. He, T. Zillio, 2007. The maximum entropy formalism and the idiosyncratic
theory of biodiversity, Ecology Letters, 10, 1017-1028.
[34] Banavar, J. R., A. Maritan, I. Volkov, 2010. Applications of the principle of maximum entropy:
from physics to ecology, J. Phys. Cond. Matter, 22, 063103
[35] Haegeman, B., R. S. Etienne, 2010. Entropy Maximization and the Spatial Distribution of
Species. Am. Nat. 175, E74-E90.
[36] Keller, D. R., F. B. Golley, ed. The philosophy of ecology, from science to synthesis, 2000.
University of Georgia press
[37] Hagen, J., 1992. An Entangled Bank: The Origins of Ecosystem Ecology, Rutgers.
[38] Lotka, A.J., 1922. Contribution to the energetics of evolution, Proc Natl Acad Sci, 8, 14751
[39] Lindeman, R. L., 1942. The Trophic-Dynamic Aspect of Ecology Ecology, 23, 399-417
[40] Odum, H.T., 1994. Ecological and General Systems: An Introduction to Systems Ecology,
Colorado University Press.
[41] Ernest, S. K. M., J. H. Brown, 2001. Homeostatis and compensation: the role of species and
25
resources in ecosystem stability. Ecology, 82:2118-2132.
[42] White, E. P., S. K. M. Ernest, K. M. Thibault, 2004. Trade-offs in community properties
through time in a desert rodent community, Am. Nat., 164:670-676.
[43] Ritchie, M. E., 2009. Scale, Heterogeneity, and the Structure and Diversity of Ecological
Communities (Monographs in Population Biology), Princeton University Press.
[44] O'Dwyer, J. P., Green, J. L, 2010. Field theory for biogeography: a spatially explicit model
for predicting patterns of biodiversity, Ecology Letters, 13, 87-95.
[45] Chapman, S. C., G. Rowlands, N. W. Watkins, 2009. Macroscopic control parameters for
avalanche models for bursty transport, Phys. Plasmas, 16, 012303.
[46] Gotelli, N. J., R. K. Colwell, 2001. Quantifying biodiversity: procedures and pitfalls in the
measurement and comparison of species richness, Ecology Letters 4, 379-391.
[47] Storch, D., K. J. Gaston, 2004. Untangling ecological complexity on different scales of space
and time, Basic and App. Ecology, 5, 389-400.
[48] Storch, D., K. L. Evans, K. J. Gaston, 2005. The species-area-energy relationship, Ecology
Letters, 8, 487-492.
[49] Wright, D. H.,1983. Species- energy theory: an extension of species-area theory, Oikos, 41,
496-506.
[50] Clarke, A., K. J. Gaston, 2006. Climate, energy and diversity, Proc. Roy. Soc. B, 273, 2257-
2266.
[51] Gillman, L. N., S. D. Wright, 2006. The influence of productivity on the species richness of
plants: A critical assessment. Ecology 87, 1234-1243.
[52] Hillebrand, H., 2004. On the generality of the latitudinal gradient rule, Am. Nat., 163, 192-211.
[53] Dengler, J., 2009. Which function describes the species-area relationship best? A review and
empirical evaluation. J. Biogeography 36, 728-744.
[54] Haskell, J. P., M. E. Ritchie, H. Olff, 2002. Fractal geometry predicts varying body size scaling
relationships for mammal and bird home ranges, Nature, 418, 527-529.
[55] Ritchie, M. E., H. Olff, 1999. Spatial scaling laws yield a synthetic theory of biodiversity,
Nature, 400, 557-560.
[56] Palmer, M. W., 2007. Species- area curves and the geometry of nature, in Scaling Biodiversity,
ed D. Storch, P. A. Marquet, J. H. Brown, Cambridge University Press.
[57] Milne, B. T. 1992. Spatial aggregation and neutral models in fractal landscapes. Am. Nat.
26
139, 32-57.
[58] Zillio, T. J., R. Banavar, J. L. Green, J. Harte, A. Maritan, 2008. Incipient criticality in
ecological communities, PNAS, 105, 18714-18717.
[59] Chiarucci, A., D. Viciani, C. Winter, M. Diekmann, 2006. Effects of productivity on species-
area curves in herbaceous vegetation: evidence from experimental and observational data.
Oikos 115, 475-483.
[60] Etienne, R. S., H. Olff, 2004, How dispersal limitation shapes species- body size distributions
in local communities, Am. Nat., 163, 69-83.
[61] Damuth, J., 1981. Population density and body size in mammals, Nature, 290, 699-700.
[62] Cohen, T., S. R. Jonsson, J. E. Carpenter, 2002. Ecological community description using the
food web, species abundance and body size, PNAS, 100, 1781-1786.
[63] Schmid, P. E., M. Tokeshi, J. M. Schmid-Araya, 2000. Relation between population density
and body size in stream communities, Science, 289, 1557-1560.
[64] Carbone, C., J. L. Gittleman, 2002. A common rule for the scaling of carnivore density,
Science, 295, 2273-2276.
[65] White, E. P., S. K. M. Ernest, A. J. Kerkhoff, B. J. Enquist, 2007. Relationships between
body size and abundance in ecology, TREE, 22:323-330.
27
|
1511.04936 | 1 | 1511 | 2015-11-16T12:47:44 | Ultrafast Energy Relaxation in Single Light-Harvesting Complexes | [
"physics.bio-ph",
"physics.chem-ph"
] | Energy relaxation in light-harvesting complexes has been extensively studied by various ultrafast spectroscopic techniques, the fastest processes being in the sub-100 fs range. At the same time much slower dynamics have been observed in individual complexes by single-molecule fluorescence spectroscopy (SMS). In this work we employ a pump-probe type SMS technique to observe the ultrafast energy relaxation in single light-harvesting complexes LH2 of purple bacteria. After excitation at 800 nm, the measured relaxation time distribution of multiple complexes has a peak at 95 fs and is asymmetric, with a tail at slower relaxation times. When tuning the excitation wavelength, the distribution changes in both its shape and position. The observed behaviour agrees with what is to be expected from the LH2 excited states structure. As we show by a Redfield theory calculation of the relaxation times, the distribution shape corresponds to the expected effect of Gaussian disorder of the pigment transition energies. By repeatedly measuring few individual complexes for minutes, we find that complexes sample the relaxation time distribution on a timescale of seconds. Furthermore, by comparing the distribution from three long-lived complexes with the whole ensemble, we demonstrate that the ensemble can be considered ergodic. Our findings thus agree with the commonly used notion of an ensemble of identical LH2 complexes experiencing slow random fluctuations. | physics.bio-ph | physics |
Ultrafast Energy Relaxation in Single Light-Harvesting Complexes
Pavel Mal´y[a,b], J. Michael Gruber[a], Richard J. Cogdell[c], Tom´as Mancal[b], and Rienk van Grondelle[a]
[a]Department of Physics and Astronomy, Faculty of Sciences,
Vrije Universiteit Amsterdam, De Boelelaan 1081,
1081HV Amsterdam, The Netherlands, [b]Institute of Physics,
Charles University in Prague, Ke Karlovu 5, 12116 Prague,
Czech Republic, [c]Institute of Molecular, Cellular and Systems Biology,
College of Medical, Veterinary and Life Sciences,
University of Glasgow, Glasgow G128QQ, United Kingdom
Energy relaxation in light-harvesting complexes has been extensively studied by various ultrafast
spectroscopic techniques, the fastest processes being in the sub-100 fs range. At the same time much
slower dynamics have been observed in individual complexes by single-molecule fluorescence spec-
troscopy (SMS). In this work we employ a pump-probe type SMS technique to observe the ultrafast
energy relaxation in single light-harvesting complexes LH2 of purple bacteria. After excitation at
800 nm, the measured relaxation time distribution of multiple complexes has a peak at 95 fs and is
asymmetric, with a tail at slower relaxation times. When tuning the excitation wavelength, the dis-
tribution changes in both its shape and position. The observed behaviour agrees with what is to be
expected from the LH2 excited states structure. As we show by a Redfield theory calculation of the
relaxation times, the distribution shape corresponds to the expected effect of Gaussian disorder of
the pigment transition energies. By repeatedly measuring few individual complexes for minutes, we
find that complexes sample the relaxation time distribution on a timescale of seconds. Furthermore,
by comparing the distribution from three long-lived complexes with the whole ensemble, we demon-
strate that the ensemble can be considered ergodic. Our findings thus agree with the commonly
used notion of an ensemble of identical LH2 complexes experiencing slow random fluctuations.
INTRODUCTION
Time-resolved studies of primary events in photosyn-
thetic light harvesting have a decades-long tradition.
Usually, the fastest processes observed correspond to the
time resolution of the experimental techniques available
at the time. Recently, the most popular tool to study ul-
trafast excitation energy transfer with sub-100 fs resolu-
tion is two-dimensional electronic spectroscopy (2DES).
This technique has been used to study various light-
harvesting complexes such as LH2 and LH1 antennas of
purple bacteria[1, 2], the FMO protein of green sulphur
bacteria[3, 4] and the major antenna complex LHCII of
higher plants[5].
It was shown that after an ultrafast
excitation of photosynthetic light-harvesting complexes
(LHCs) the electronic excitation evolves in a coherent
fashion on a 100 fs timescale. These observations sparked
a still ongoing debate on the role of quantum coherence
in energy transfer in LHCs.
However powerful the ultrafast techniques have be-
come, they are fundamentally limited by ensemble av-
eraging. Although the 2DES can in principle resolve in-
homogeneous and homogeneous lineshapes, the observed
spectra and system dynamics are still averaged over the
whole ensemble of complexes. Another feature of nonlin-
ear spectroscopy such as 2DES is that broadband pulses
are used for excitation, which results in simultaneous ex-
citation of many states. Such pulses inevitably excite
also superpositions of states, which leads to coherent dy-
namics. This can provide useful information about the
system, especially on the electronic coupling between the
pigments and the interplay of electronic and nuclear de-
grees of freedom. On the other hand, it brings with itself
interpretative issues in relation to the relevance of such
coherent dynamics for natural light harvesting under in-
coherent sunlight[6].
At about the same time as ultrafast spectroscopy,
also optical microscopy has seen significant advances[7].
Nowadays it is routinely possible to selectively excite and
observe individual LHCs. This enables us to overcome
the problem of ensemble averaging and observe distribu-
tions of single-molecule properties. However, for prac-
tical reasons only single-molecule emission spectroscopy
has been possible on biological pigment-protein com-
plexes. Photon counting of the weak luminescence signal
becomes a limiting factor for the time resolution, mak-
ing it possible to observe changes only on a timescale of
tens of milliseconds and longer.The standard paradigm is
therefore to think about the ultrafast nonlinear ensem-
ble spectroscopy and single-molecule spectroscopy (SMS)
as complementary methods that access very different
timescales.
In 2005 van Dijk et al. proposed a modification of
SMS called single-molecule pump-probe (SM2P), which
employs excitation by two pulses. This technique visu-
alizes the initial ultrafast excitation relaxation in single
molecules[8]. As they demonstrated on dye monomers[8,
9] and later on dye dimers and trimers[10], it is possible
to observe relaxation rates in the 100 fs range. In this
work we explore the possibility of applying this technique
to light-harvesting complexes.
The LHC of choice for our measurement is the light-
harvesting complex 2 (LH2) of the purple bacterium
Rhodopseudomonas acidophila. LH2 consists of two rings
of bacteriochlorophylls, which result in two distinct ab-
sorption bands at roughly 800 and 850 nm, respectively.
Both the bands and the rings are referred to as B800
and B850 according to their central absorption wave-
length. The pigments in the ring responsible for the
B800 band are relatively weakly coupled, while the pig-
ments from the B850 ring exhibit strong electronic cou-
pling. This strong interaction results in significant ex-
citonic splitting and formation of delocalized excitonic
(vibronic) states[11]. Most of the ultrafast studies of en-
ergy transfer in LH2 were carried out in the late eighties
and nineties using variants of transient absorption (TA)
and fluorescence upconversion[11 -- 14]. From these and
later studies[15, 16] it was concluded that while the en-
ergy transfer from the B800 to the B850 ring is relatively
slow, 1-2 ps, the relaxation dynamics after 800 nm ex-
citation are more complex, including faster components
due to the overlap of the B800 states with high ener-
getic exciton states of the B850 ring (B850*). These
states were found to exhibit ultrafast transfer dynam-
ics on the timescale of hundreds of fs. Recent results
from 2DES spectroscopy furthermore revealed ultrafast
sub-200 fs dynamics[1, 17, 18]. Meanwhile, SMS stud-
ies of LH2 at cryogenic and later at ambient tempera-
tures showed intensity fluctuations and spectral diffusion
on a much slower timescale of seconds[19 -- 22]. By theo-
retical modeling it was shown that most of the spectro-
scopic observations can be explained by dynamic varia-
tions in the realization of the energetic disorder of the
pigments[20, 23]. These findings highlight the dynamic,
fluctuating nature of LHCs. Experimentally, LH2 is a
perfect candidate for our proof-of-principle measurement
for several reasons. The presence of lower B850 states
results in fluorescence emission around 870 nm, which
is sufficiently red-shifted with respect to the absorption
bands to enable easy excitation and detection separation.
Importantly, LH2 shows a high fluorescence yield and sig-
nificant stability in single-molecule conditions, which is
a requirement for our experiment.
RESULTS
The measurement
2
FIG. 1. (A) The three-level scheme used for the data analysis.
kL is the absorption and stimulated emission rate, kF L is the
spontaneous emission rate and kR is the relaxation rate which
is measured. The Gaussian profile represents the laser pulse,
resonant with state 1(cid:105) and off-resonant with state 2(cid:105). (B)
Excited states available in LH2, schematically shown together
with a measured absorption spectrum. The red peak repre-
sents the excitation spectrum at 800 nm, the arrows indicate
possible relaxation channels.
analysis is shown in Fig. 1A. It consists of a ground state
0(cid:105), an excited state 1(cid:105) resonant with the laser excita-
tion and an off-resonant excited state 2(cid:105). This scheme
is universal for the technique and can always be used for
analysis. It then depends on the measured system how
the respective levels should be interpreted. A cartoon of
the actual situation in LH2, together with the measured
absorption spectrum, is presented in Fig. 1B. The main
difference between the isolated molecules studied previ-
ously in Ref.
[8] and LHCs is the dense excited states
manifold in the latter case. However, it can be shown
by numerical simulations, that the three-level description
still holds as effective. In the case of a dense manifold, the
observed relaxation rate is the effective rate with which
the excitation escapes the region resonant with the laser.
For a more detailed description of the SM2P technique
and the analysis procedure we refer the reader to the sup-
porting information (SI) and the original works by van
Dijk et al.[8, 9].
The SM2P principle is based on exciting the system
by a near-saturating laser pulse and giving it a time win-
dow to relax to some off-resonant state before applying
a second pulse. By such relaxation the excitation in the
system can be saved from the stimulated emission caused
by the second pulse, and the overall excitation proba-
bility therefore rises with the pulse delay. The detected
fluorescence signal is proportional to this excitation prob-
ability and therefore depends on the delay between the
two pulses. The excitation relaxation rate can then be
extracted by scanning the pulse delay time and fitting
the resulting change in fluorescence intensity. The effec-
tive three-level scheme which is used for the SM2P traces
Using a confocal microscope, individual complexes are
excited by the two-pulse laser sequence. The pulses with
a center wavelength around 800 nm are 200 - 250 fs long
and about 4 nm wide. Thorough preliminary calcula-
tions, which can be found in the SI, indicated that the
above mentioned laser specifications will work to reveal
ultrafast dynamics in LH2 complexes. The fluorescence
of one complex is collected by the same microscope ob-
jective and recorded by an avalanche photodiode. In this
way the fluorescence intensity traces of multiple individ-
ual LH2 complexes are recorded one by one. The emis-
sion of one complex is measured until it photobleaches,
while simultaneously continuously scanning the delay be-
A0.20.40.60.81.01.2900875850825800775750725B850B800BAbsorption (arb.u.)B800B850Wavelength (nm)kLkFLkR120 tween the two excitation pulses. The first minute of a
typical intensity trace from a stable complex is shown
in Fig. 2. The signal of about 1000 counts per second
is characteristic for the given measurement conditions.
The data are binned into 100 ms bins, which represents
a compromise between the signal-to-noise ratio and the
amount of data points available for fitting. The mea-
sured intensity modulation results from the pulse delay
scanning and each intensity dip can be used to determine
the corresponding relaxation time. The inset in Fig. 2
depicts a single intensity dip from the recorded trace,
with an extracted relaxation time of τR = (89 ± 25) fs.
The present noise can be explained by Poissonian shot
noise. The good sample stability allowed us to perform
multiple pulse delay scanning cycles and therefore to ex-
tract multiple subsequent relaxation times from one com-
plex. In the given example, the complex switches into a
dark state at t = 55 s, a process often called 'blinking'.
This behaviour indicates that the observed signal indeed
arises from a single well connected antenna.
It should
be noted that not all complexes are such stable emitters.
As was observed before (see e.g.
[20, 22]), there can be
a significant amount of blinking with different degrees of
quenching, which results in switching between different
intensity levels. However, no matter what the mecha-
nism of energy dissipation causing these fluctuations is,
the fluorescence intensity is still proportional to the ex-
citation probability. Therefore, whenever the emission is
stable for a sufficiently long time to perform one pulse
delay scan, and the emission intensity is high enough to
provide a reasonable signal-to-noise ratio, the relaxation
time can be measured. In this way we can measure sev-
eral intensity dips for many complexes and extract the
relaxation times by fitting with the three-state model.
Energy relaxation
The distributions of relaxation times obtained for ex-
citation wavelengths of 812 nm, 800 nm, and 780 nm are
shown in Fig. 3A. The average recorded relaxation times
are 92 fs at 812 nm, 106 fs at 800 nm and 139 fs at 780
nm excitation. These measured relaxation times agree
well with the expected ultrafast timescale.
As a result of the already mentioned dense excited
states manifold, there are several differences between the
original work on dye monomers and the LHCs. In the
former case of individual or weakly coupled pigments,
the observed ultrafast relaxation is the intramolecular
vibrational relaxation,
the dynamic Stokes' shift.
By comparing monomers and dimers, van Dijk et al.
showed that this relaxation slows down when the ex-
cited states are delocalized and thus more weakly cou-
pled to the environment[8].
In LHCs the situation is
different. First, the pigments are coupled and thus the
vibrational and electronic states become mixed, resulting
in a vibronic states manifold. The energy transfer be-
tween these states cannot be strictly separated into the
i.e.
3
FIG. 2. First one minute of a measured fluorescence intensity
trace of a single LH2 complex, recorded while continuously
scanning the delay between the two excitation pulses. Red
lines: data fitted with the three level model in Fig. 1A. At
t = 55 s the complex briefly switches to a dark state ('blink-
ing'). Inset: magnification of one intensity dip with a fitted
relaxation time of τR = (89 ± 25) fs. Bottom axis gives the
real recording time, while the top axis denotes the delay be-
tween the two pulses.
intra- or inter- pigment relaxation. Second, unlike the
dye molecules, the bacteriochlorophylls present in LH2
have a much smaller Stokes' shift, typically around 5 nm
(≈ 80 cm−1)[24]. It is thus by itself not enough to es-
≈ 65 cm−1) wide excitation pulse. And
cape the 4 nm (
finally, the measured dependence of the relaxation time
on the excitation wavelength is exactly opposite from
what would be expected from a Stokes' shift.
In our
case we observe the fastest relaxation in the 'red' region
with wavelength longer than 800 nm, where the strongly-
coupled B850* states are present. The relaxation is then
slower when exciting in the 'blue' region at 780 nm, where
the states of weakly-coupled B800 pigments play a larger
role, see Fig. 3A.
Another aspect to consider is that we observe only en-
ergy relaxation and not dynamic localization, because of
excitation with circular polarized light. The latter con-
tributes mainly to absorption depolarization[16]. The
observed relaxation rate then effectively describes how
fast the excitation escapes the resonant laser excitation
range. The next dissimilarity from the case of individ-
ual dye molecules is the possible presence of multiple
excitations and the related singlet-singlet annihilation.
However, because the fluorescence lifetime is orders of
magnitude longer than the singlet-singlet annihilation
time[25, 26], it is precisely the annihilation which ren-
ders the multiply-excited states invisible. The annihila-
tion, always present at near-saturating intensities, thus
effectively ensures that the three state model with a sin-
gle excited state is a good approximation for the observed
fluorescence signal. Another concern are higher excited
states of the pigments, possibly resulting from multiple
050010001500120014001600-600-200200600Intensity (cps) Intensity (cps)0204060Pulse delay (fs)Time (s)Time (s)38404244 4
Relaxation time fluctuations
Having discussed the average observed relaxation time,
we can focus on the true single-molecule measurement
achievements: the relaxation time distribution and fluc-
tuations. We have already mentioned the distributions in
Fig. 3A. Due to the anaerobic conditions which increase
the sample endurance, we were able to follow several
stable complexes for minutes before they photobleached.
The obtained relaxation time trajectories can be found
in Fig. 3B.
Before we start interpreting these results, we need to
make sure that the fluctuations we measure are not just
an artifact of the fitting in the presence of shot noise. To
this end we perform numerical simulations of the SM2P
signal including Poissonian shot noise. In the inset of Fig.
4A we present one of the simulated intensity dips. The
distribution of the fitted relaxation times obtained from
such simulated dips is presented in Fig. 4A. The signal
binning time and the bin size are the same as used in Fig.
3A to illustrate the difference. The calculated relaxation
times are symmetrically distributed around the expected
value of 100 fs and the distribution can be excellently fit-
ted by a Gaussian normal distribution with a FWHM of
33 fs. This distribution is much narrower than the exper-
imentally obtained one and also its shape is completely
different.
In Fig. 4B we show a 'trace' of successively
simulated relaxation times that can be compared to its
experimental counterpart in Fig. 3B. The extent of the
fluctuations caused only by the shot noise is significantly
smaller. Together with a clear wavelength dependence of
the relaxation time distributions, these simulations con-
vince us that the observed fluctuations are real and not
only the result of shot noise.
In order to qualitatively understand the possible origin
of the asymmetric shape of the relaxation time distribu-
tion, we can consider the following simple model. Let
us describe energy transfer between two excitonic states,
originating from two coupled pigments. Using Redfield
theory, the relaxation rate krel between the excitonic lev-
els can be expressed analytically. When we assume, for
the sake of simplicity, that the spectral density of bath
modes is approximately flat in the considered frequency
region, the relaxation rate is proportional to
krel =
∝
1
τrel
(cid:1)2 ,
1 +(cid:0) ∆
1
2J
(1)
where J is the coupling constant between the pigments
and ∆ is the energy difference between the coupled states.
The relaxation time is then determined only by the ra-
tio ∆
2J and the distribution arises from the energetic dis-
order in ∆. We assume a Gaussian distributed disor-
der, as is commonly done in such simulations, with a
FWHM ∆dis = J. This is typical for simulations of light-
harvesting complexes, where all three parameters are ex-
pected to be in the same range, i.e. ∆ ≈ J ≈ ∆dis. In
Fig. 3D the resulting relaxation time distribution is de-
picted for different values of the detuning ∆. We can see
FIG. 3. (A) Relaxation time distribution obtained from many
measured complexes at three different excitation wavelengths
(B) Relaxation time trajectories of three stable complexes,
under 800 nm excitation. The shaded regions indicate the
standard error of the fits. (C) The relaxation time distribu-
tion obtained from these 3 complexes, compared to the ensem-
ble distribution at 800 nm excitation from (A). (D) Modelled
distribution of the relaxation times in a two-state model us-
ing Redfield theory. The distribution shape changes with the
ratio of the coupling and energy gap (see text for description).
excitation of the same pigment. However, these decay
to the lowest excited state much faster than the over-
all excited state lifetime[27]. From the discussion above
we can therefore conclude that we indeed observe energy
relaxation between the singly-excited states within the
complex.
Comparing with the literature, we find that our relax-
ation times are somewhat shorter than those found by
previous time-resolved measurements. As mentioned in
the introduction, excitation at 800 nm results in pop-
ulating states of both the B800 and the B850* bands.
Our experimental results indeed indicate that the B850*
states contribute significantly to the rather fast observed
relaxation rate. The comparably wider excitation pulses
of typically 10-15 nm used in TA measurements fail to re-
solve energy relaxation processes within their bandwidth,
resulting in a slower overall relaxation rate. Further-
more, TA and fluorescence decay kinetics are usually fit-
ted with and resolved into several energy transfer com-
ponents, while this study yields an effective 'escape' rate
comprising all available relaxation channels. As a conse-
quence, the observed relaxation is somewhat faster and
the slow components are not visible in our measurement.
Our results therefore agree with relaxation times of 150-
300 fs reported for 800 nm excitation[13] and furthermore
experimentally validate the faster dynamics determined
by theoretical modeling of the B850* band[16].
501001502002500102030 Frequency (%)Relaxation time (fs) ensemble 3 LH2s1.01.52.02.501020 Frequency (%)Relaxation time (rel.u.) D=J D=1.25J D=1.5JCD050100150200250010203040 Frequency (%)Relaxation time (fs) 812 nm 800 nm 780 nm100200 Trace 3 75175 Trace 2 060120180240100200 Trace 1 Relaxation time (fs)Time (s)AB 5
with the one from the whole ensemble of many complexes.
We find, as is shown in Fig. 3D, that the distributions are
very similar. This indicates that every complex can likely
sample all the possible relaxation times on a timescale
of seconds and that the ensemble can be considered er-
godic. As this argument is not completely conclusive,
further investigation in this direction would certainly be
of interest.
CONCLUSIONS
We have successfully applied the SM2P technique to
LH2 complexes of purple bacteria. We have demon-
strated that it is possible to observe ultrafast energy
relaxation in individual light-harvesting complexes. As
such, our work highlights a new possible way to study
photosynthetic light-harvesting. We have shown how the
relaxation time distribution changes when tuning the ex-
citation wavelength. The observed behaviour can be ex-
plained by varying influence of the B800 and B850* states
of the LH2 rings, in agreement with previous ultrafast
spectroscopy studies. By a numerical calculation we were
able to qualitatively explain the shape of the relaxation
time distribution as a result of the energetic disorder of
the LH2 pigments. The extent of disorder corresponds
to the values commonly used in bulk spectroscopy mod-
elling. Our method can be extended to include a de-
tailed excitation wavelength scan, which would enable us
to study energy transfer dynamics of single LH2 com-
plexes to an extent similar to bulk transient absorption
measurements. Finally, we observed the evolution of the
relaxation rate of individual complexes in time. In accor-
dance with previous SMS studies, we attribute its fluc-
tuations to slow protein motion, based on the relevant
timescale. Our results thus not only serve as a proof-of-
principle measurement for the SM2P technique on pho-
tosynthetic systems, but also present a fitting piece of
evidence to the puzzle of light-harvesting dynamics in
the ever fluctuating antenna complexes.
Materials and methods
Experimental setup The experimental setup is similar
[8], briefly a 76 MHz Ti:Sapphire
to the one in Ref.
laser (Mira 900F, Coherent) is used as a source of 200-
250 fs, 4-5 nm spectrally wide pulses centered at 800 nm.
By tuning the laser cavity the wavelength can be tuned
approx.
from 750 nm to 850 nm. The repetition rate
is decreased to 2 MHz by a pulse-picker (PulseSelect,
APE) to increase the survival time of the complexes and
eliminate long-living dark states such as triplet states.
The absence of the triplet states is verified by checking
the signal drops to half when halving the repetition rate.
The two pulses are produced by a home-built Michelson
interferometer, the delay between them is scanned by a
delay line (Newport) in one of the interferometer arms.
FIG. 4. Testing the effect of Poissonian shot noise. The simu-
lated parameters are: a relaxation time of 100 fs, pulses of 200
fs and a signal of around 1000 cps. (A) The relaxation time
distribution obtained from the simulation, fitted with a Gaus-
sian distribution with a FWHM of 33 fs. Inset: one of the
simulated intensity dips, together with the fitted 3-level model
curve. The recovered relaxation time was τR = (90 ± 17) fs.
(B) A succession of simulated relaxation times that can be
compared with Fig. 3B. The shaded area indicates the stan-
dard error of the fits.
that for strong coupling (or small energy gaps) the relax-
ation is fastest and the distribution is highly asymmetric.
With decreasing coupling (or increasing energy gap) the
distribution maximum shifts to longer relaxation times
and becomes more symmetric. Our experimentally ob-
tained distributions in Fig. 3A seem to follow this trend:
the distribution measured at 812 nm is the most asym-
metric one with the shortest relaxation times, the 800
nm distribution is the intermediate case and the distri-
bution measured at 780 nm excitation is more symmetric
and shifted to longer relaxation times. This fully agrees
with the discussion above, describing the increasing in-
fluence of the strongly-coupled B850* ring states when
tuning the excitation to longer wavelengths. The shape
of the distribution can thus be qualitatively described
as originating from a Gaussian energetic disorder of the
transitions energies of the antenna pigments.
Finally, we want to comment on the relaxation time
trajectories presented in Fig. 3B. The relaxation time
clearly varies on a timescale of seconds, which is in agree-
ment with slow fluctuations observed by SMS on LH2
before[19, 20, 23]. It should be mentioned that fluctua-
tions in LHCs are observed on almost all timescales, from
fast sub-picosecond vibrations of the pigments to slow
protein structural changes in the range of seconds. Our
experiment is able to observe the latter type of flucta-
tions, where slow motion of the protein causes changes in
the local pigment environment resulting predominantly
in a shift of their transition energy [23, 28, 29].
A question arises whether all complexes are identical
and experience the same fluctuations, or whether the en-
semble is heterogeneous. To investigate this we compare
the relaxation time distribution from three long traces
05010015020025002550 Frequency (%)Relaxation time (fs) Shot noise Gaussian fit-75007509001000 Intensity (cps)Pulse delay (fs)05101520253050100150200 Relax. time (fs)Dip number Shot noiseAB The pulse length before the microscope is measured by
fringe-resolved autocorrelation[30] using the same inter-
ferometer and focusing the pulses into a BBO crystal
(Eksma optics). The pulse spectrum is measured by a
spectrometer (OceanOptics). Technical details can be
found in the SI. Due to the narrow bandwidth of the
pulses no significant broadening of the pulses in the mi-
croscope can be expected. Guild et al. measured the
dispersion of common high N.A. objective microscopes,
and for a microscope very similar to ours they find GDD
of around 4000 fs2, including the beam expander [31].
Using a formula for Gaussian pulse second-order disper-
sion, we obtain that our 200 fs (lower limit) pulses stretch
to 208 fs. This is indeed negligible considering the flu-
orescence intensity dip fitting error arising from the sig-
nal to noise ratio. The excitation light is adjusted to
a circular polarization by a Berek compensator (New
Focus) to avoid complex orientation dependence. The
complexes are illuminated and detected by a confocal
microscope with a PlanFluor objective (1.3NA, Nikon)
as described elsewhere[20]. The detected fluorescence is
alternatively dispersed by a grating on a CCD (Prince-
ton Instruments) to measure the emission spectrum or
the intensity is measured by an avalanche photodiode
(Perkin-Elmer). The fluorescence spectrum is used to
check the integrity of the complexes during the course of
the measurement. The excitation intensity is set to be
sufficient to nearly-saturate the complexes. At 800 nm
excitation we used an excitation power of 0.5 pJ/pulse,
focused to a diffraction-limited spot, which is comparable
to previous experiments [9]. For excitation at different
wavelengths the intensity was increased to compensate
for the decreased absorption, see spectrum in 1B. The
measurement is controlled by a custom-made LabView
environment.
Sample preparation The isolated LH2 complexes from
Rhodopseudomonas acidophila are diluted to a concentra-
tion of ∼ 10 pM in a measuring buffer (20 mM Tris, pH 8
and 0.03% (w/v) n-Dodecyl β-D-maltoside) and then im-
mobilized on a PLL (poly-L-Lysine, Sigma) coated cover
glass. The dilution is chosen such as to obtain on average
approximately 10 complexes per 100 µm2. To increase
the survival time of complexes the buffer is deoxygenated
6
by the oxygen-scavenging system PCA/PCD (2.5 mM
protocatechuic acid, 25 nM protocatechuate-3,4- dioxy-
genase, Sigma)[32]. The experiments were conducted at
room temperature.
Relaxation time fitting The detailed description of the
SM2P technique can be found in the SI. When applying
the three-level system description as in Fig. 1A, it can
be shown the intensity dip as a function of pulse delay τ
can be described as
I(τ ) = I∞(cid:26)
(cid:20)
(cid:18) 1
2d
(cid:0)d2k − 2τ(cid:1)(cid:19)
1 − p1
2 − p1
(cid:18) 1
e
4
e−kτ erf c
1
2
k2d2
(cid:0)d2k + 2τ(cid:1)(cid:19)(cid:21)(cid:27)
+ekτ erf c
2d
,
(2)
is the relaxation rate, I∞ is the baseline
where k = 1
τR
intensity, p1 is the probability of excitation by one pulse
( 1
2 for full saturation) and d is the effective pulse width,
related to the pulse full width at half maximum (FWHM)
as d = 1√
dF W HM . We use this formula to fit the
measured dips and extract the relaxation times.
2ln2
ACKNOWLEDGMENTS
P.M., J.M.G. and R.v.G. were supported by the VU
University and by an Advanced Investigator grant from
the European Research Council (no. 267333, PHOT-
PROT) to R.v.G.; R.v.G. was also supported by the
Nederlandse Organisatie voor Wetenschappelijk Onder-
zoek, Council of Chemical Sciences (NWO-CW) via a
TOP-grant (700.58.305), and by the EU FP7 project
PAPETS (GA 323901). R.v.G. gratefully acknowledges
his Academy Professor grant from the Netherlands Royal
Academy of Sciences (KNAW). P.M. and T.M. received
financial support from the Czech Science Foundation
(GACR), grant no. 14-25752S. R.J.C. was supported
as part of the Photosynthetic Antenna Research Center
(PARC), an Energy Frontier Research Center funded by
the U.S. Department of Energy, Office of Science, Basic
Energy Sciences under Award #DE-SC0001035.
[1] Harel, E & Engel, G. S. (2011) Quantum coherence spec-
troscopy reveals complex dynamics in bacterial light- har-
vesting complex 2 (LH2). Proc. Natl. Acad. Sci. U.S.A.
109, 706 -- 711.
[2] Maiuri, M, R´ehault, J, Carey, A.-M, Hacking, K, Gar-
avelli, M, Luer, L, Polli, D, Cogdell, R. J, & Cerullo,
G. (2015) Ultra-broadband 2D electronic spectroscopy
of carotenoid-bacteriochlorophyll interactions in the LH1
complex of a purple bacterium. J. Chem. Phys. 142,
212433.
[3] Brixner, T, Stenger, J, Vaswani, H. M, Cho, M, Blanken-
ship, R. E, & Fleming, G. R. (2005) Two-dimensional
spectroscopy of electronic couplings in photosynthesis.
Nature 434, 625 -- 628.
[4] Engel, G. S, Calhoun, T. R, Read, E. L, Ahn, T.-K,
Mancal, T, Cheng, Y.-C, Blankenship, R. E, & Flem-
ing, G. R. (2007) Evidence for wavelike energy transfer
through quantum coherence in photosynthetic systems.
Nature 446, 782 -- 6.
[5] Schlau-Cohen, G. S, Calhoun, T. R, Ginsberg, N. S,
Read, E. L, Ballottari, M, Bassi, R, van Grondelle, R,
& Fleming, G. R.
(2009) Pathways of energy flow in
LHCII from two-dimensional electronic spectroscopy. J.
Phys. Chem. B 113, 15352 -- 63.
[6] Mancal, T & Valkunas, L. (2010) Exciton dynamics in
photosynthetic complexes: excitation by coherent and in-
coherent light. New J. Phys. 12, 65044.
[7] Brinks, D, Hildner, R, van Dijk, E. M. H. P, Stefani, F. D,
Nieder, J. B, Hernando, J, & van Hulst, N. F. (2014)
Ultrafast dynamics of single molecules. Chem. Soc. Rev.
43, 2476 -- 91.
[8] van Dijk, E, Hernando, J, Garc´ıa-L´opez, J.-J, Crego-
Calama, M, Reinhoudt, D, Kuipers, L, Garc´ıa-Paraj´o,
M. F, & van Hulst, N. F. (2005) Single-Molecule Pump-
Probe Detection Resolves Ultrafast Pathways in Individ-
ual and Coupled Quantum Systems. Phys. Rev. Lett. 94,
078302.
[9] van Dijk, E, Hernando, J, Garc´ıa-Paraj´o, M. F, & van
Hulst, N. F. (2005) Single-molecule pump-probe experi-
ments reveal variations in ultrafast energy redistribution.
J. Chem. Phys. 123, 064703.
[10] Hernando, J, van Dijk, E, Hoogenboom, J, Garc´ıa-L´opez,
J.-J, Reinhoudt, D, Crego-Calama, M, Garc´ıa-Paraj´o,
M. F, & van Hulst, N. F. (2006) Effect of Disorder on
Ultrafast Exciton Dynamics Probed by Single Molecule
Spectroscopy. Phys. Rev. Lett. 97, 216403.
[11] Sundstrom, V, Pullerits, T, & van Grondelle, R. (1999)
Photosynthetic light-harvesting: Reconciling dynamics
and structure of purple bacterial LH2 reveals function of
photosynthetic unit. J. Phys. Chem. B 103, 2327 -- 2346.
[12] Bergstrom, H, Sundstrom, V, van Grondelle, R, Gill-
bro, T, & Cogdell, R. J. (1988) Energy transfer dynam-
ics of isolated B800-850 and B800-820 pigment-protein
complexes of Rhodobacter sphaeroides and Rhodopseu-
domonas acidophila. Biochim. Biophys. Acta - Bioenerg.
936, 90 -- 98.
[13] Hess, S, Feldchtein, F, Babin, A, Nurgaleev, I, Pullerits,
T, Sergeev, A, & Sundstrom, V.
(1993) Femtosecond
energy transfer within the LH2 peripheral antenna of the
photosynthetic purple bacteria Rhodobacter sphaeroides
and Rhodopseudomonas palustris LL. Chem. Phys. Lett.
216, 247 -- 257.
[14] Jimenez, R, Dikshit, S. N, Bradforth, S. E, & Fleming,
G. R. (1996) Electronic Excitation Transfer in the LH2
Complex of Rhodobacter sphaeroides. J. Phys. Chem.
100, 6825 -- 6834.
[15] Wendling, M, Mourik, F. V, van Stokkum, I. H. M,
Salverda, J. M, Michel, H, & van Grondelle, R. (2003)
Low-intensity pump-probe measurements on the B800
band of Rhodospirillum molischianum. Biophys. J. 84,
440 -- 449.
[16] Novoderezhkin, V. I, Wendling, M, & van Grondelle, R.
(2003) Intra- and Interband Transfers in the B800 - B850
Antenna of Rhodospirillum molischianum : Redfield The-
ory Modeling of Polarized Pump - Probe Kinetics. J.
Phys. Chem. B 107, 11534 -- 11548.
[17] Zigmantas, D, Read, E. L, Mancal, T, Brixner, T, Gar-
diner, A. T, Cogdell, R. J, & Fleming, G. R.
(2006)
Two-dimensional electronic spectroscopy of the B800-
B820 light-harvesting complex. Proc. Natl. Acad. Sci.
U. S. A. 103, 12672 -- 12677.
[18] Fidler, A. F, Singh, V. P, Long, P. D, Dahlberg, P. D, &
Engel, G. S. (2014) Dynamic Localization of Electronic
Excitation in Photosynthetic Complexes Revealed with
Chiral Two-Dimensional Spectroscopy. Nat. Commun.
5, 3286.
[19] van Oijen, A. M, Ketelaars, M, Kohler, J, Aartsma,
T. J, & Schmidt, J. (2000) Spectroscopy of individual
7
light-harvesting 2 complexes of Rhodopseudomonas aci-
dophila: diagonal disorder, intercomplex heterogeneity,
spectral diffusion, and energy transfer in the B800 band.
Biophys. J. 78, 1570 -- 1577.
[20] Rutkauskas, D, Novoderezhkin, V. I, Cogdell, R. J, &
Van Grondelle, R. (2004) Fluorescence Spectral Fluctua-
tions of Single LH2 Complexes from Rhodopseudomonas
acidophila Strain 10050. Biochemistry 43, 4431 -- 4438.
[21] Baier, J, Richter, M. F, Cogdell, R. J, Oellerich, S, &
Kohler, J. (2008) Determination of the spectral diffusion
kernel of a protein by single-molecule spectroscopy. Phys.
Rev. Lett. 100, 1 -- 4.
[22] Schlau-Cohen, G. S, Wang, Q, Southall, J, Cogdell, R. J,
& Moerner, W. E. (2013) Single-molecule spectroscopy
reveals photosynthetic LH2 complexes switch between
emissive states. Proc. Natl. Acad. Sci. U. S. A. 110,
10899 -- 903.
[23] Rutkauskas, D, Novoderezhkin, V. I, Gall, A, Olsen, J,
Cogdell, R. J, Hunter, C. N, & van Grondelle, R. (2006)
Spectral trends in the fluorescence of single bacterial
light-harvesting complexes: experiments and modified
Redfield simulations. Biophys. J. 90, 2475 -- 2485.
[24] De Caro, C, Visschers, R. W, van Grondelle, R, &
Voelker, S. (1994) Inter- and Intraband Energy Transfer
in LH2-Antenna Complexes of Purple Bacteria. A Flu-
orescence Line-Narrowing and Hole-Burning Study. J.
Phys. Chem. 98, 10584 -- 10590.
[25] van Grondelle, R, Hunter, C. N, Bakker, J. G. C, &
Kramer, H. J. M.
(1983) Size and structure of an-
tenna complexes of photosynthetic bacteria as studied by
singlet-singlet quenching of the bacteriochlorophyll fluo-
rescence yield. Biochim. Biophys. Acta - Bioenerg. 723,
30 -- 36.
[26] Ma, Y, Cogdell, R. J, & Gillbro, T.
(1997) Energy
transfer and Exciton Annihilation in the B800-850 An-
tenna Complex of the Photosynthetic Purple Bacterium
Rhodopseudomonas acidophila (Strain 10050). A fem-
tosecond Transient Absorption Study. J. Phys. Chem.
B 101, 1087 -- 1095.
[27] Blankenship, R. E. (2002) Molecular Mechanisms of Pho-
tosynthesis. (Blackwell Science).
[28] Kruger, T. P. J, Novoderezhkin, V. I, Ilioaia, C, & van
Grondelle, R. (2010) Fluorescence spectral dynamics of
single LHCII trimers. Biophys. J. 98, 3093 -- 101.
[29] Kruger, T. P. J, Wientjes, E, Croce, R, & van Gron-
delle, R. (2011) Conformational switching explains the
intrinsic multifunctionality of plant light-harvesting com-
plexes. Proc. Natl. Acad. Sci. U. S. A. 108, 13516 -- 21.
[30] Diels, J. C, Fontaine, J. J, McMichael, I. C, & Simoni,
F. (1985) Control and measurement of ultrashort pulse
shapes (in amplitude and phase) with femtosecond accu-
racy. Appl. Opt. 24, 1270 -- 1282.
[31] Guild, J. B, Xu, C, & Webb, W. W. (1997) Measure-
ment of group delay dispersion of high numerical aperture
objective lenses using two-photon excited fluorescence.
Appl. Opt. 36, 397 -- 401.
[32] Swoboda, M, Henig, J, Cheng, H. M, Brugger, D, Hal-
trich, D, Plumer´e, N, & Schlierf, M. (2012) Enzymatic
oxygen scavenging for photostability without ph drop in
single-molecule experiments. ACS Nano 6, 6364 -- 6369.
[33] Loudon, R. (2001) The Quantum Theory of Light. (Ox-
ford University Press, New York), 3rd edition, pp. 68 -- 72.
SUPPLEMENTARY INFORMATION
Theoretical considerations
8
The SM2P technique is based on exciting the molecules by two near-saturating laser pulses, scanning the delay
between them, and observing the correlated change in fluorescence intensity. In order to understand the resulting
signal we study a model three-level system consisting of a ground state 0(cid:105), an excited state 1(cid:105) resonant with the laser
excitation and an off-resonant excited state 2(cid:105). A schematic illustration is shown in Fig. 1A in the main text. During
the interaction with the first pulse the evolution of the system can be described by the Bloch equations[33]. The
external electromagnetic field creates a superposition of the ground and excited state, i.e. an optical coherence, and
the probability of excitation of state 1(cid:105) undergoes Rabi oscillations. Because of the strong coupling of the electronic
degrees of freedom to the fluctuating environment and fast energy transfer from the excited state, the optical coherence
rapidly dephases, typically on the order of (cid:46)50 fs, and the Rabi oscillations become damped. If the duration of the
saturating pulse is longer than the coherence dephasing time, the excitation probability settles at 1
2 . In that case
a description by semiclassical rate equations is appropriate. If the system is excited by the first pulse, it starts to
relax to the off-resonant state 2(cid:105). Now, when the second pulse arrives immediately after the first pulse, the system is
still saturated and the excitation probability remains 1
2 . However, when the delay between the pulses is long enough,
the excitation can relax to state 2(cid:105) and thus escape stimulated emission. The system can then interact with the
second pulse only if it was not excited by the first pulse. It thus gets a second chance to be excited and the overall
probability of excitation rises to 1
4 . Scanning the pulse delay and measuring the emission intensity, which
is proportional to the excitation probability, allows us to observe a dip in fluorescence, see e.g. Fig. S1. The width
of the dip is related to the time it takes for the system to relax to the off-resonant state. In practice the excitation
starts to relax already during interaction with the exciting pulse, and state 2(cid:105) might not be completely off-resonant.
The dynamics of the system can then be described by coupled equations for the state populations
2 · 1
2 + 1
2 = 3
∂P1(t, τ )
∂t
∂P0(t, τ )
∂t
∂P2(t, τ )
∂t
= kL (P0 − P1) IL(t, τ ) − kRP1,
= −kL ((1 + νres) P0 − P1 − νresP2) IL(t, τ ),
= νreskL (P0 − P2) IL(t, τ ) + kRP1.
(3)
Here kL is the absorption/stimulated emission rate, kR is the relaxation rate, IL(t, τ ) is the intensity of the two laser
pulses delayed by τ , and νres is the resonance of state 2(cid:105) (relative to state 1(cid:105), νres = 0 when state 2(cid:105) is completely
off-resonant). Because the spontaneous emission rate is typically orders of magnitude slower than energy relaxation,
the observed fluorescence is proportional to the population of the relaxed state long after interaction with the exciting
pulses, IF L(τ )∝P2(∞, τ ).
Concerning LHCs, several aspects have to be taken into account. Their states, originating from more than one
pigment, form a dense excitonic (vibronic) manifold and the energy transfer is ultrafast. It is thus important to verify
how sensitive the method is to the actual degree of saturation, how much off-resonant the state 2(cid:105) has to be, and how
the fastest observable relaxation rate depends on the duration of the pulse. To address these questions we performed
numerical simulations varying the respective parameters, see Fig. S1. It turns out that the exact saturation level is
not critical (Fig. S1C), in agreement with Ref.
[9]. The off-resonance is, however, very important (Fig. S1B). For
example, already 75% off-resonance, i.e. νres = 0.25, decreases the intensity dip considerably and, in the presence of
Poissonian noise (see below), renders it invisible. Finally, the fastest observable relaxation time is about 5 times faster
than the pulse duration (Fig. S1A). These findings are very encouraging for LHCs. The expected relaxation times
are around 100 fs, which can be well resolved by 200 fs laser pulses. These pulses can then be only about 4 nm wide,
which, together with a high off-resonance demand, ensures high excitation selectivity. The weak dependence on the
degree of saturation furthermore allows to quantitatively measure the relaxation out of the selected narrow excitonic
region.
The findings presented above allow us to use an effective description via Eqs. (3) with νres = 0. Level 1(cid:105) then
represents the states resonant with the excitation and level 2(cid:105) the off-resonant, relaxed states. The effective description
by Eqs. (3) is therefore still valid even for LHCs with dense excited state manifolds.
For Gaussian pulses the solution of Eqs. (3) can be found analytically by convoluting the result for δ-pulse excitation
with both pulse envelopes. This approach, the same as in the original work of van Dijk et al.[9], assumes the pulses
are mixed incoherently. This is true for our setup where the vibrations of the rapidly-moving delay line safely destroy
all phase coherence. We also tried measuring with a vibrating mirror and obtained comparable results. The measured
fluorescence intensity as a function of the pulse delay τ can then be expressed as
I(τ ) = I∞(cid:26)
1 − p1
2 − p1
1
2
k2d2
4
e
e−kτ erf c
(cid:20)
(cid:18) 1
(cid:0)d2k − 2τ(cid:1)(cid:19)
(cid:18) 1
(cid:0)d2k + 2τ(cid:1)(cid:19)(cid:21)(cid:27)
+ ekτ erf c
2d
where I∞ is the baseline intensity, p1 is the probability of excitation by one pulse ( 1
2 for full saturation), k is the
relaxation rate and d is the effective pulse width, related to the pulse full width at half maximum (FWHM) as
d = 1√
dF W HM . Formula (4) is used for fitting the data and extraction of the relaxation time τR = 1
k .
2d
2ln2
9
(4)
Pulse characterization
(cid:90)
To characterize the pulses we use a modified version of fringe-resolved autocorrelation (FRAC)[30]. In this inter-
ferometric method we use the same Michelson interferometer as we use for the pulse delay scanning. For the pulse
characterization the pulses are colinearly focused in a BBO crystal to generate a second harmonic signal (SH). The
fundamental frequency is then filtered out and the SH intensity is detected by a slow detector.
Denoting the laser field E(t, τ ) = E0(t)e−iωt + E0(t + τ )e−iω(t+τ ) + c.c., the second harmonic is proportional to
E2(t, τ ), and the measured SH intensity is proportional to
ISH (τ ) =
dtI 2
0 (t) + I 2
0 (t + τ ) + 2Re{E2
0 (t)E2
0 (t + τ )}cos(2ωτ )
+4 (I0(t) + I0(t + τ )) Re{E0(t)E0(t + τ )cos(ωτ ) + 4I(t)I(t + τ ).
(5)
This signal contains a baseline contribution from both pulses, the SH interferogram, envelope-modulated fundamental
interferogram and the intensity autocorrelation which we want to measure. Using this FRAC signal the pulses can be
characterized, however, for pulses as long as ours, the recording requires an impractically long time. By attaching a
vibrator to one of the interferometer mirrors, the interference-sensitive part of the SH is averaged out on the detector,
which enables us to measure the intensity autocorrelation on a constant background much faster. We checked that this
method works by comparing the obtained pulse width with the one from the full recorded interferogram for several
pulse lengths.
The characterization of the pulses used for the 800 nm excitation measurement, together with their measured
spectrum for control, can be found in Fig. S2.
10
FIG. S1. Numerical simulations of the intensity dip dependence on (A) the relaxation time, (B) the off-resonance of the relaxed
level and (C) the degree of saturation.
√
FIG. S2. Characterization of the pulses used for the 800 nm excitation measurement. (A) Intensity autocorrelation from using
FRAC with vibrating mirror. Gaussian fit yields F W HM = 360 fs, using deconvolution factor
2 then gives 255 fs long
pulses. (B) Measured control spectrum (its parameters are not used for the relaxation time extraction). Gaussian fit gives
F W HM = 3.8 nm spectral width.
-1000010000.60.70.80.91.0 Excitation probabilityPulse Delay (fs) tR=tpulse/6 tR=tpulse/4 tR=tpulse/2 tR=tpulse tR=2 tpulse-1000010000.70.80.91.0 Excitation probability Pulse Delay (fs) nres=0 nres=0.01 nres=0.1 nres=0.25 nres=0.5-1000010000.50.60.70.80.91.0 Excitation probabilityPulse Delay (fs) 95% saturation 90% saturation 75% saturation 50% saturationCBAA760780800820840024Intensity (arb.u.)Wavelength (nm) Pulse control spectrum Gaussian fit-1000-500050010000102030405060Intensity (arb.u.)Delay (fs) Intensity autocorrelation Gaussian fitB |
1908.11144 | 1 | 1908 | 2019-08-29T10:40:58 | Critical threshold for microtubule amplification through templated severing | [
"physics.bio-ph",
"q-bio.SC"
] | The cortical microtubule array of dark-grown hypocotyl cells of plant seedlings undergoes a striking, and developmentally significant, reorientation upon exposure to light. This process is driven by the exponential amplification of a population of longitudinal microtubules, created by severing events localized at crossovers with the microtubules of the pre-existing transverse array. We present a dynamic one-dimensional model for microtubule amplification through this type of templated severing. We focus on the role of the probability of immediate rescue-after-severing of the newly-created lagging microtubule, observed to be a characteristic feature of the reorientation process. Employing stochastic simulations, we show that in the dynamic regime of unbounded microtubule growth, a finite value of this probability is not required for amplification to occur, but does strongly influence the degree of amplification, and hence the speed of the reorientation process. In contrast, in the regime of bounded microtubule growth, we show that amplification only occurs above a critical threshold. We construct an approximate analytical theory, based on a priori limiting the number of crossover events considered, which allows us to predict the observed critical value of the rescue-after-severing probability with reasonable accuracy. | physics.bio-ph | physics | Critical threshold for microtubule amplification through templated severing
Marco Saltini and Bela M. Mulder
AMOLF, Science Park 104 1098XG Amsterdam, Netherlands
The cortical microtubule array of dark-grown hypocotyl cells of plant seedlings undergoes a
striking, and developmentally significant, reorientation upon exposure to light. This process is
driven by the exponential amplification of a population of longitudinal microtubules, created by
severing events localized at crossovers with the microtubules of the pre-existing transverse array.
We present a dynamic one-dimensional model for microtubule amplification through this type of
templated severing. We focus on the role of the probability of immediate rescue-after-severing of
the newly-created lagging microtubule, observed to be a characteristic feature of the reorientation
process. Employing stochastic simulations, we show that in the dynamic regime of unbounded
microtubule growth, a finite value of this probability is not required for amplification to occur, but
does strongly influence the degree of amplification, and hence the speed of the reorientation process.
In contrast, in the regime of bounded microtubule growth, we show that amplification only occurs
above a critical threshold. We construct an approximate analytical theory, based on a priori limiting
the number of crossover events considered, which allows us to predict the observed critical value of
the rescue-after-severing probability with reasonable accuracy.
9
1
0
2
g
u
A
9
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
4
4
1
1
1
.
8
0
9
1
:
v
i
X
r
a
I.
INTRODUCTION
2
Microtubules are a ubiquitous component of the cytoskeleton in eukaryotic cells. They are filamentous aggregates
of tubulin-dimers reaching lengths of several µm's. They are typically part of structures that span the dimensions
of the whole cell, enabling e.g. their major role in intracellular transport by providing "tracks" for cargo carrying
motor proteins and cell division where they form the mitotic spindle responsible for the spatial segregation of the
duplicated chromosomes [1]. The fact that during the cell cycle microtubules can be reassembled into different spatial
structures is due to their intrinsically dynamic nature. They stochastically switch between phases of growth through
polymerization to phases of shrinkage through rapid depolymerization, a mechanism that has been dubbed dynamic
instability [2]. A direct consequence of this mechanism is that microtubules have a finite lifetime, as they can shrink
away, and therefore need to be actively (re)nucleated to sustain their overall number. Cells achieve control over the
microtubule structures they build by manipulating the nucleation and dynamics of microtubules in space and time,
using specific nucleating complexes and a host of microtubule-interacting proteins (MAPS)[3].
Growing plant cells have a unique microtubule structure called the cortical array. The cortical array is an assembly
of mutually aligned microtubules localized close to the cell membrane that almost homogeneously covers the inside
surface of the cell. Generically the preferential direction of the cortical microtubules is transverse to the long axis of
the cell. This is crucial to their function, as they guide the anisotropic deposition of cell-wall building polymers, which
in turn allows the cell to grow along a single expansion axis. In this way the cortical array drives the dominant mode
of morphogenesis in plants, which is the formation of linear extensions, like roots, stems and branches. However, it is
known that this generic growth scenario can be modulated by hormonal, mechanical and other environmental signals
[4]. A striking example of this type of modulation is the reorientation of the cortical array of dark-grown hypocotyl
(stem precursor) cells after exposure to blue light [5]. This effect is highly relevant, as the developmental program of
the plant must change dramatically, once the hypocotyl, which typically emerges from a buried seed, first reaches
the sunlight. It is believed that the observed reorientation from the transverse to the longitudinal orientation of the
cortical array is associated with the arrest of further growth, and the subsequent differentiation of the cells.
The light-induced reorientation of the cortical array is mediated by the microtubule severing protein katanin. It
has been shown to localize at the crossover between differently oriented cortical microtubules and, moreover, to
preferentially sever the overlying microtubule, i.e. the one that crossed over a pre-existing one. As the underlying
microtubule is most likely to be a transversely oriented microtubule from the pre-exposure state and severing effectively
multiplies the number of microtubules, this effect can rapidly create an exponentially growing population of longitudinal
microtubules. In this way the original transverse cortical array serves as a template for the reorientation towards a
longitudinal array.
Recent experiments involving a number of mutants in which the reorientation effect is impaired, have shown that
there is an important role for the propensity of the newly-created plus end of the lagging microtubule to immediately
switch to the growing state, a process which in vivo is mediated by the prominent MAP CLASP [6]. This specific
function of CLASP appears to have evolved, as the default outcome of a severing event is the creation of a shrinking
plus end of the lagging microtubule [7]. Stochastic simulations of a simplified model of the reorientation mechanism,
indeed, showed that the degree of amplification of the numbers of microtubules due to severing increases monotonically
with the probability of the so-called rescue-after-severing of the lagging plus end [6]. However, the dynamic parameters
of the microtubules measured in the experiments suggest that, at least in the initial phase of reorientation, the
microtubules are in the so-called unbounded-growth regime [8]. Since in this regime the microtubules in principle are
very long-lived on the timescale of the reorientation, this raises the question to what extent rescue-after-severing is
in fact a necessary ingredient of the mechanism. Moreover, as the total amount of tubulin in the cell is finite, it is
also clear that unbounded-growth and amplification of microtubules cannot be sustained indefinitely, as the pool of
available tubulin to drive polymerization is inevitably depleted. This will cause the growth speed of the microtubules to
decrease, effectively driving them back to the bounded-growth regime. To fully understand the reorientation process we
thus need to disentangle the role of the microtubule growth state from that of the probability of rescue-after-severing,
and this is the main aim of this paper. We approach this problem using a combination of stochastic simulations and
analytical theory, which together allow us to fully characterize the requirements for the amplification of longitudinal
microtubules to occur.
The structure of the paper is as follows. In Sec. II we introduce our dynamic model of longitudinal microtubules
undergoing dynamic instability and severing in the presence of a grid of stable transverse microtubules. We then
briefly review some of the main features of Dogterom-Leibler model for microtubules undergoing dynamic instability
[8], on which our model is based. In Sec. III we show how, for microtubules in the unbounded-growth regime, while
not a necessary ingredient, the probability of rescue-after-severing does dramatically affect the speed and ultimate
success probability of amplification of the longitudinal microtubule population. Then, we extend our treatment to the
bounded-growth regime, showing that in this case amplification can occur provided that the probability of rescue-after-
severing exceeds a critical threshold. Finally, we calculate the critical value for the probability of rescue-after-severing
using a combination of analytical calculations and computer simulations. To that end, we introduce an approximate
theory in which each microtubule can experience at most two crossovers, allowing an analytical determination of the
contribution of the probability of rescue-after-severing to the probability that the creation of a crossover actually leads
to a severing event, and hence contributes to the amplification. In order to do so, we develop a novel approximate
technique to calculate the first-passage time distribution (hereafter FPTD) for the microtubules to reach relatively
close targets, which has potential application for studying first-passage time problems in other systems as well.
3
II. THE MODEL
A. Dynamic model
Figure 1. Schematic of the model of longitudinal microtubules undergoing dynamic instability in a grid of stable transverse
microtubules. After a crossover creation, a competition between the intrinsic severing waiting time and crossover removal due to
dynamic instability takes place. If severing occurs, the newly-created plus end is rescued with probability p+.
In order to better understand the importance of the probability of rescue-after-severing for the reorganization
mechanism of the cortical microtubule array, we introduce a stochastic model of longitudinal microtubules undergoing
dynamic instability in a one dimensional grid of transverse microtubules. Since before the exposure to light the
transverse array consists of relatively long microtubules, in the initial stage of this process, i.e. the first 500 seconds,
the array can be seen as a constant background. In particular, we focus on this time interval because it is the period
of time in which the amplification takes place.
The model consists of a single longitudinal microtubule undergoing dynamic instability in the one dimensional
grid of stable, transverse microtubules with constant spacing d between neighboring filaments [6]. According to
experimental observations, where the angle between differently oriented microtubules is very close to 90◦, we assume
that all longitudinal microtubules are exactly perpendicular to the transverse. The microtubule is nucleated at position
x = 0 with plus end in the growing state with growing speed v+, and it can switch to shrinking state with constant
catastrophe rate rc. Once it is in the shrinking state, either its plus end shrinks back to position x = 0 and dies, or it
undergoes a rescue with constant rate rr, switching back to the growing state.
Every time the plus end reaches a transverse microtubule - i.e. when its position is x = nd with n ∈ N, it creates a
crossover. This crossover can be resolved in two distinct ways: either it is removed by the shrinkage of the microtubule
due to its dynamic instability, or it survives long enough to lead to a severing event. Whether or not the severing
event occurs is determined not only by the dynamic instability of the longitudinal microtubule, but also by an intrinsic
severing waiting time distribution at the crossover that can be, in principle, arbitrary. Here, however, we choose a
distribution that best fits the biological experimental data about the action of katanin at crossovers [9]. Moreover, we
need the distribution to account the fact that katanin requires a certain amount of time to localize at crossovers before
being able to sever microtubules. For this reason we choose the severing waiting time distribution to be
Wkθ (t) =
tk−1
θkΓ (k)
e− t
θ ,
i.e. Gamma probability density function [10], where
Γ (k) =
ds sk−1e−s,
(cid:90) ∞
0
(1)
(2)
Crossover created withstable transverse MTCompetition:severing vs shrinkingSevering wins:new MT createdShrinking wins:crossover erasedwaiting time distributiontimer startsLongitudinal MTs undergodynamic instabilityNewly-created plus end shrinksAmplification step 4
Parameter
Description
Numerical value
(bounded-growth) (unbounded-growth)
Numerical value
v+
v−
rc
rr
p+
d
θ
k
Growth speed
Shrinkage speed
Catastrophe rate
Rescue rate
0.1
0.25
0.02
0.02
Probability of rescue-after-severing
Tuned
Spacing between neighbors
Scale parameter of Gamma distribution
Shape parameter of Gamma distribution
1.5
8.5
7
Table I. Model parameters.
0.103
0.225
0.0058
0.026
Tuned
1.5
8.5
7
Units
µm s−1
µm s−1
s−1
s−1
-
µm
s
-
is the Euler gamma function, k is the shape, and θ is the scale parameter of the distribution.
When the severing event occurs, the former long microtubule is split in two shorter microtubules, and both of
them keep undergoing dynamic instability and can create new crossovers and being severed again, in order to amplify
the number of longitudinal microtubules. The newly-created plus end of the lagging microtubule either is stabilized
and it enters the growing state with probability p+, or it enters the shrinking state with probability 1 − p+. The
newly-created minus end of the leading microtubule is now positioned at the severing point in a stable state, whilst no
changes are applied to its plus end, see Figure 1.
B. Microtubule behaviour in the interstitial strip
After the creation of a crossover and before the creation of a second one, the dynamics of the plus end of a
microtubule is described by the Dogterom-Leibler model for microtubules undergoing dynamic instability [8]. Notice
that the dynamics of the plus end is not influenced by eventual severing events. Therefore, as long as the plus end is
at x ∈ (nd, (n + 1)d), we can study the property of the correspondent microtubule undergoing dynamic instability in a
strip of width d as if its length is l = x − nd.
In the non-confined-in-a-strip case, the model has two possible solutions for the probability distribution of microtubule
length: in the bounded-growth regime, defined by the relation l > 0, with
(cid:16) rc
v+ − rr
v−
(cid:17)−1
l =
,
(3)
the steady-state solution is reached, and the length distribution is an exponential decay proportional to e−l/l, whilst in
the unbounded-growth regime the average length of microtubules grows linearly in time, with the length distribution
that is well-approximated by a Gaussian-like function [8].
When microtubule dynamics is confined in a strip of a finite width, however, both the bounded and the unbounded-
growth regimes lead to a steady-state solution for the length distribution that is proportional to e−l/l. Notice that,
in the unbounded-growth regime case, l is no-longer positive, and hence the distribution is exponentially increasing
[11]. General features regarding the lifetime distribution and the splitting probabilities of microtubules in the
Dogterom-Leibler model can be found in the Appendix A.
Given our specific interest in studying the properties of the system in both the bounded and the unbounded-growth
regime, we have chosen two sets of dynamic parameters: for the bounded-growth case parameters are chosen accordingly
to previous observations [12], while for the unbounded-growth case, both dynamic parameters and grid parameters are
those that have been directly measured for the WT case by previous experimental works [9], see Table I.
III. RESULTS
A. Amplification in the unbounded-growth regime
The model introduced in the last section had partially been computationally studied for microtubules in the
unbounded-growth regime, and it shows that the factor that influences the most the speed of the amplification of
longitudinal microtubules is the probability of rescue-after-severing p+ rather than the intrinsic rescue rate rr of
microtubules [9].
Here, we want to perform an in-depth study of the response to the system to the change of p+. We will show that,
even though p+ is crucial for the speed of amplification, in the unbounded-growth regime it is not required for the
occurrence of it.
Our simulations consist of N = 105 trials in which a single longitudinal microtubule undergoes dynamic instability
in the whole grid of transverse microtubules. For every trial we keep track of the fate of the initial microtubule and
its offspring until either no more microtubules are present - i.e. they all have shrunk to length zero, and we call
this possible output extinction, or, for every trial that did not result in an extinction, the number of microtubules is
exponentially increasing, and we call this second possible output amplification.
5
Figure 2. (A) Time evolution of the number of longitudinal microtubules for four different values of p+. They all exhibit
amplification. (B) Extinction probability as a function of time. It represents the fraction of trials in which, after a certain
amount of time, all microtubules have completely depolymerized.
Fig. 2A shows that for our choice of dynamic parameters, the speed of amplification increases with p+. Furthermore,
we notice from Fig. 2B that greater values of p+ correspond to lower extinction probabilities, suggesting that a
good rescue-after-severing entails a double effect: not only it increases the speed of amplification, but also makes the
amplification occur more likely.
The interesting result that in the unbounded-growth regime even the p+ = 0 case leads to an overall amplification
can be explained by an intrinsic property of the regime itself. Indeed, although every severing event shortens the
length of the severed microtubule, its plus end is not affected by such an event. Consequently, the dynamic properties
of the leading microtubule are not changed by the severing, and so it applies to the microtubule lifetime as well. Since,
on average, the length of microtubules in the unbounded-growth regime grows as
J =
rrv+ − rcv−
rr + rc
t,
(4)
it follows that the average lifetime of microtubules is infinite [8], and therefore there is no upper bound for the number
of severing events that a microtubule can undergo.
B. Amplification in the bounded-growth regime
In this section, we address the question whether or not the amplification occurs regardless of p+ in the bounded-
growth regime as well as in the unbounded-growth case. To do so we perform computer simulations for microtubules in
the bounded-growth regime (see Table I) to show that p+ needs to be greater that a certain critical value p+
crit in order
have amplification. Moreover, using a combination of computer simulations and analytical calculations we identify
such a critical value as a function of the other model parameters.
1. Critical point in simulations
By tuning the probability of rescue-after-severing p+ from 0 to 1, we observe two different behaviours, see Figure
3A: for lower values of p+ the average number of microtubules exponentially decays in time (extinction), whilst for
higher values of p+ the number of microtubules exponentially increases (amplification). It follows that there exists a
critical threshold for p+ above which the average output is amplification. For our choice of model parameters, the
1101000100200300400500Time [s]Number of microtubules0100200300400500Time [s]Extinction probability0.10.20.30.40.5AB 6
crit (cid:39) 0.36. Therefore, if we define the amplification probability as the
computationally measured critical value is p+
fraction of trials the output of which is amplification, we observe that below the critical threshold the amplification
probability is zero, whilst it is greater than zero otherwise, see Figure 3B.
Figure 3. (A) Time evolution of the number of longitudinal microtubules for three different values of p+. One leads to amplification
(blue line), one to extinction (black line), and one corresponds to the critical behaviour (red line). (B) Amplification probability
as a function of p+. Amplification probability is non-zero for p+ larger than p+
crit (cid:39) 0.36.
2. Calculation of the critical point
When microtubule is created through a severing event, either it shrinks to length zero and dies, or it is severed a
sufficient number of times to create an offspring of new lagging microtubules. If the size of such an offspring is, on
average, greater than one, the output is amplification. In other words, if M is the number of severing events that a
newly-created microtubule undergoes, then amplification occurs if, on average,
M > 1.
(5)
To fix the ideas, suppose that a microtubule is created by severing with initial length x = d. Then, with probability
p+ it is initially created in the growing state, and consequently with probability 1 − p+ in the shrinking state. It
follows that the size of the offspring of the mother microtubule can be written as M = p+M + + (1 − p+) M−, where
M σ is the size of the offspring of a microtubule created in the state σ. However, since a shrinking microtubule with
plus end at x < d cannot be severed, M− equals M + times the probability that the shrinking microtubule recovers the
length at birth d, i.e. M− = R−
d (d) is the splitting probability defined in the Appendix A. Hence,
the condition expressed in Eq. (5) can be rewritten as
d (d) M +, where R−
M =(cid:2)p+ +(cid:0)1 − p+(cid:1) R−
d (d)(cid:3) M + > 1.
(6)
By solving the equality related to Eq. (6) we can find the critical value of p+ = p+
place
crit above which amplification takes
1 − R−
(cid:0)1 − R−
d (d)(cid:1) M +
d (d) M +
p+
crit =
.
(7)
From this equation, we can identify the two extreme scenarios in which amplification never or always occurs,
regardless of p+. In the first case, we state that amplification never occurs if p+
crit > 1, meaning that the maximum
value that p+ can reach is not enough to lead to amplification. We can show that condition p+
crit > 1 is equivalent to
condition M + < 1, i.e. amplification is impossible if the average size of the offspring of mother microtubules born in
growing state is smaller than 1. In the second case, we state that amplification always occurs if p+
crit < 0, meaning
that even without rescue-after-severing, dynamic parameters of the model are such that amplification is still possible.
This condition is equivalent to M +R−
d (d) > 1, or M− > 1, i.e. amplification occurs every time the average size of the
offspring of mother microtubules born in shrinking state is greater than 1.
It is important to underline that, in our discussion, we assumed that all microtubules were born with initial length
d. This choice implies that all severing events occur at the first crossover. However, given the stochastic nature of the
00.20.40.60.810.10.201002003004005000.1110Amplification probabiltyTime [sec]Number of microtubulesAB0.40.50.3 system and of the severing waiting time probability of Eq. (1), it is possible that a severing event occurs further in the
grid than at the first crossover of a microtubule. In other words, the initial length of a newly-created microtubule can
be x = nd, with n > 1.
In this case, we need to add into the count of the size of offspring of a mother microtubule all cases in which a
microtubule that is born with initial length nd, n > 1, it is also severed at (n − 1)d, (n − 2)d, . . . . We consider the
microtubules created by this mechanism as direct daughter microtubules of the mother microtubule we are measuring
the size of the offspring of. To that end, we first define mi via M + = 1
N
mi as the number of microtubules generated
by the mother microtubule labeled by i, and N is the number of microtubules we keep track of the fate. Then, we
denote the number of severing events the microtubule i undergoes with si, and the position of the crossover at which
the first severing takes place with cji, with the rule: cji = n − 1 if the severing occurred at nd. Since after a severing
event at nd the crossovers at kd, k < n, can be removed by either shrinkage or severing, we define bcji
as the number
of crossovers that are resolved by shrinkage. Therefore
N(cid:80)
i=1
si(cid:88)
(cid:2)cji − bcji
(cid:3) .
mi = si +
ji=1
depends on p+, as it depends on the behaviour of the plus end after severing of the newly-created
Notice that bcji
microtubule. Consequently, the r.h.s. of Eq. (7) exhibits a dependency on p+. Hence, we need to find the exact
dependency on p+ of bcji
= 0 for every cji. In this way, we
can computationally measure m(1)
. To avoid the problem, in first approximation we set bcji
cji, and we use Eq. (7) to give a first estimate of the critical probability
i ≡ si +
si(cid:80)
ji=1
p+
crit,(1), see Table II. We refer to this approximation as one-crossover approximation. From the table we notice
that, although this approximation gives a reasonable prediction for the critical probability of rescue-after-severing, it
systematically underestimates it.
only under the condition that a severing event at nd always implies the resolution
of the crossovers at d, 2d, . . . , (n − 2) d through a severing event, whilst the crossover at (n − 1) d can be resolved by
either severing or shrinkage. Therefore, if we denote the probability of resolving this crossover with a shrinkage as
pcr (p+), we have
Analytically, one can calculate bcji
=(cid:0)1 − δcji ,0
(cid:1) pcr
(cid:0)p+(cid:1) .
bcji
The Kronecker function δcji ,0 of Eq. (9) accounts the fact that, if the severing happens at d (i.e. cji = 0), no other
crossovers are removed by either severing or shrinkage. If we plug Eq. (9) into Eq. (8), we can now calculate an
approximate expression for M +, i.e.
7
(8)
(9)
The detailed derivation of Eq. (10) can be found in the Appendix B. Finally, if we define M +
N(cid:88)
i=1
M +
(2) =
1
N
m(2)
i =
1
N
S = 1
N
amplification as
N(cid:80)
i=1
si, and (cid:104)1 − δci,0(cid:105) = 1
(cid:104)
M =
si
si(cid:80)
(cid:0)1 − δcji ,0
p+ +(cid:0)1 − p+(cid:1) R−
ji=1
N(cid:88)
(cid:104)
si(cid:88)
(cid:0)p+(cid:1) si(cid:88)
(cid:105)
.
cji − pcr
(1 − δci,0)
si +
(10)
i=1
ji=1
ji=1
(cid:34)
(cid:35)
(cid:1), (see Appendix B), we can rewrite the condition (6) for microtubule
(cid:105)(cid:34)
(1) = 1
N(cid:80)
si(cid:80)
(cid:35)
si +
ji=1
i=1
cji
,
N(cid:88)
N
(cid:0)p+(cid:1) S
(cid:104)1 − δci,0(cid:105)
> 1.
(11)
d (d)
(1) − pcr
M +
1
N
i=1
The resolution of the equation associated to this inequality provides the critical threshold for the probability of rescue-
after-severing p+
crit,(2) in order to have amplification. We refer to this approximation as two-crossovers approximation.
The expression (11) contains two quantities, M +
(1) and S, that cannot be analytically calculated but can be easily
measured with computer simulations. On the other hand, in the following sections we are going to analytically calculate
(cid:104)1 − δci,0(cid:105). In this way, we will make a better prediction off the critical probability of
the terms pcr (p+) and 1
N
N(cid:80)
rescue-after-severing in order to have amplification.
i=1
C. Analytical approach
8
N(cid:80)
i=1
calculate pcr (p+) and 1
N
In this section we are going to calculate the critical probability of rescue-after-severing. To that end, we first need to
(cid:104)1 − δci,0(cid:105). In order to do so, and because of the complexity of the model, we make the
approximation that the entire grid of transverse microtubules is replaced by just two transverse microtubules. This
reduces the total number of possible crossovers to two. Therefore, we first calculate the FPTD for a longitudinal
microtubule to create a crossover with a transverse, as we will need it for the formulation of our two-crossovers
approximation. Then, we give some analytical results of the one-crossover approximation already introduced in the
previous section. Finally, we present the two-crossovers approximation and we show that we can use it to calculate the
critical probability of rescue-after-severing with a good degree of accuracy.
1. The first passage time distribution
The creation of new crossovers for a microtubule undergoing dynamic instability is intimately linked to a FPTD
problem for the same microtubule to reach a target. Here, we face this problem by making use of an approach where
we consider all possible legal paths to reach the target, given the knowledge of the time needed to reach it.
The first passage time problem for a microtubule to reach length x1 starting from x0 in the absence of severing
can be seen as a reverse lifetime problem, in the sense that in place of studying the reaching of the target at x1 we
study the survival of the microtubule until it arrives at x1, as if it is shrinking from x0 to x1. In this way, the growing
speed of the microtubule acts as its shrinking speed, its catastrophe rate as the rescue rate, and viceversa. However,
with this approach we assume that a microtubule "shrinking" from x0 can undergo a "rescue" and grow beyond the
initial position x0. This means that the microtubule assumes negative length. To avoid this, we need to take into
account only the paths from x0 to x1 that never shrink below x0. Hence, if Lσ (tx1 − x0) is the lifetime distribution
for microtubules with initial length x1 − x0 and initial state σ (see Appendix A), we define
Ltarget
σ
(t, x1 − x0) = Lσ (tx1 − x0)
,
(12)
(cid:12)(cid:12)(cid:12)v± → v∓
rc ↔ rr
as the FPTD to reach the target, including the possibility of assuming negative length. Therefore, this function must
be re-scaled by the number of legal paths Γx0→x1 (t) that reach the target x1 at time t, without ever shrinking back to
x < x0, calculated over all possible paths that arrive at x1 at time t, see Figure 4AB.
For our purpose, the target to reach is a transverse microtubule for the creation of a new crossover, the position of
which is at distance d from the starting point, i.e. the previous transverse microtubule. Typically, for the range of
values of Table I, every plus end that impinges on a transverse microtubule starting from the previous one either it
does it without undergoing any catastrophe, or it undergoes one catastrophe and a subsequent rescue. For the dynamic
parameters we are considering, the occurrence of multiple catastrophe-rescue events is very unlikely. Therefore we
assume that all paths are either direct - no catastrophes, or indirect - one catastrophe and one rescue.
Given the constant growing and shrinking speeds, the amount of time that a microtubule needs to reach the target
at d is given by the time needed to reach it in absence of any catastrophe, added to the time spent from a catastrophe
to the moment when the original length before the catastrophe is restored. Mathematically, if Td is the first-passage
time and ∆x (Td) is the distance walked by the plus end from the catastrophe to the subsequent rescue, the equation
Td =
v+ +
holds. From Eq. (13) we can find the expression for ∆x (T ) = v+v−
v++v−
Since catastrophes are modelled as Poisson events, if a catastrophe occurs, the probability that it occurs does not
depend on the distance from the target. Therefore, the fraction of legal paths can be written as Γ0→d (T ) = 1 − ∆x(T )
,
and, finally, the FPTD as
(13)
v+
d
d
v+ + ∆x (Td)
(cid:18) 1
(cid:20)
,
1
v+
(cid:19)
(cid:0)T − d
(cid:18) 2
v+ +
(cid:1).
(cid:19)
(cid:21)
1
v−
− t
F0d (t) = Ltarget
σ
(t, d) Γ0→d (t) Θ
d
,
(14)
where the Heaviside theta is imposed to allow at most one catastrophe-rescue event.
In order to separate direct paths from indirect paths, it is convenient to split F0d (t) in two parts, and rewrite it as
e−rct + f0d (t) ,
(15)
(cid:18)
(cid:19)
F0d (t) = δ
t − d
v+
9
Figure 4. (A) Legal and (B) illegal path for a microtubule to reach the target at a distance d in a first passage time T . Only
one catastrophe and one rescue are allowed. (C-F) Comparison between simulations (red dots) and theory (blue line) for the
non-direct part fd (t) of the first passage time distribution. (B, C) Our theory nicely fits simulations for relatively close targets
(d = 1.5 µm, d = 6 µm), (D, E) while it is not very good for more distant targets (d = 30 µm, d = 60 µm).
where the term multiplied by the delta function accounts direct paths, while f0d (t) accounts indirect. From Appendix
A that microtubules reach the target with probability R+
d (x), and as a
consequence, the relation
d (x). Therefore F0d (t) is normalized to R+
(cid:90) ∞
dt f0d (t) = R+
d (0) − e
− rc d
v+
(16)
holds.
0
We run N = 106 simulations of microtubules undergoing dynamic instability in a strip of width d, and we create the
histogram of the arrival times for the microtubules that reach the target with an indirect path. Figure 4CD shows
that the approximation of only one catastrophe-rescue event is a good approximation when the target is relatively
close compared to the dynamic parameters of the microtubules, while it apparently fails when the target is more
distant, see Figure 4EF. However, it is convenient to point out that, for d (cid:29) l, we observe a very few arrivals at the
d (0) decays as e−d/l. On the other hand, in the
target, since from Eq. (25) we notice that the arrival probability R+
unbounded-growth regime, since a fraction 1− rcv−
rrv+ of the microtubules always arrives at the target, for distant targets
the approximation of one catastrophe-rescue event is no longer accurate.
2. One-crossover theory
Naıvely, one can think that once a crossover is created, the probability p(1)
sev of resolving it with a severing event is
given by the competition between two independent events: microtubule lifetime, expressed by the random variable
T+ (x) with density function given by Eq. (29), and severing waiting time at the crossover, with random variable
τd and density function defined in Eq. (1). Then, if we define the random variable t = τd − T0, we can calculate
−∞ dz(cid:48) Px (z − z(cid:48)) Py (z(cid:48)) , where Pi is the
its probability density function by using the relation Pz (z = x + y) = (cid:82) +∞
probability density function of the random variable i = x, y, z, and x and y are independent random variables. In our
case, the probability density function is
Pτd−T+(0) (t) =
dt(cid:48) Wk,θ (t + t(cid:48)) L+ (t(cid:48)0) .
(17)
(cid:90) +∞
−∞
10203040506000.51.01.52.02.5x10-3time [sec]timetimeABC[sec-1][sec-1]plus endpositionplus endpositionlegal pathillegal path40801201602002402004006008009001200time [sec]time [sec]DEx10-61x103510152025300.20.40.60.80x00.51.01.52.02.510-300.51.01.52.02.5x10-3time [sec]Fx10-7140600 Hence, the probability that the event "severing" happens before the event "return" is P [t < 0] =(cid:82) 0
−∞ dt Pτd−T+(0) (t).
This probability is not yet the probability of resolving a crossover with a severing event: indeed, microtubules in
the unbounded-growth regime have a finite probability of growing indefinitely. For those, the lifetime T+ (0) → ∞ is
infinite. Therefore, the probability of resolving a crossover with a severing event is
10
sev = S+ (∞0) + [1 − S+ (∞0)]
p(1)
(cid:90) 0
(18)
where S+ (∞0) is the ultimate survival probability defined in the Appendix A. As a consequence the probability of
resolving a crossover with a shrinkage is
dt Pτ0−T+(0) (t) ,
−∞
shrink = 1 − p(1)
p(1)
sev.
(19)
However, with this approach we neglect the number of crossovers removed by shrinkage after the severing at a second
crossover, and, therefore, the dependency on p+. In other words, the one-crossover theory does not take into account
that some microtubules that would have been severed at d are not anymore severed there because an eventual severing
at nd, n > 1, can in principle shorten their lifetimes, and make them shrink below d, resulting in the resolution of the
crossover by a shrinkage induced by the severing at nd, see Figure 5.
3. Two-crossovers theory
Figure 5. Schematic of the full one-crossover theory (A). The newly-created crossover can be resolved either by the shrinkage
of the plus end below the crossover itself (lower blue square), or by the severing at crossover (upper blue square). Schematic
of the full two-crossovers theory (B). The first crossover created can be resolved either by the shrinkage of the longitudinal
microtubule with probability p(2)
shrink (sum of all paths that bring to the lower blue square), or by the severing at crossover with
probability p(2)
sev (sum of all paths that bring to the upper blue square). Whether the severing at the first crossover occurs or not
also depends on what happens at the second crossover: a severing event at the second crossover alters the dynamic instability of
the lagging microtubule, and hence its probability of shrinking before being severed at the first crossover.
In order to take into account the influence of a crossover on the resolution of the previous one, we calculate the
probability of resolving a crossover with a severing event in a scenario in which we have two transverse microtubules,
timer startstimer startstimer startsAB 11
at position d and 2d respectively. We make the further approximation that a microtubule cannot be severed two times
at the same crossover. We denote the probability of having a severing event at d with p(2)
sev, and consequently the
probability of resolving a crossover with shrinkage with p(2)
shrink = 1 − p(2)
sev.
We first notice that p(2)
Figure 5B shows the three distinct ways in which the newly-created crossover at d can be resolved by shrinkage or
severing: 1) the microtubule shrinks (is severed) without reaching 2d, 2) the microtubule shrinks (is severed) after
reaching 2d but without being severed there, 3) the microtubule shrinks (is severed) after reaching 2d and after being
severed there. The third case bears a dependency on p+.
shrink = qsev + qshrink, where qsev is the
probability of shrinkage after severing at 2d (path −→−→(cid:46) of Figure 5B), while qshrink is the probability of
shrinkage without any severing (paths −→↓ or (cid:38) of Figure 5B). Furthermore, since qsev depends on the dynamic
behaviour of the microtubule just after the severing event, it carries a dependency on p+ and can be split again in
qsev (p+) = p+qsev,+ + (1 − p+) qsev,−, where qsev,σ is the probability of shrinkage after being severed at 2d with the
newly-created plus end in the state σ. The derivation of qsev,σ can be found in Appendix B.
shrink can be split in two probabilities, i.e. p(2)
As regards the probability qshrink that microtubules shrink below d without being severed there, we observe that
such a probability accounts all cases in which crossovers are resolved by shrinkage in absence of the crossover at 2d, i.e.
p(1)
shrink, except for those cases in which microtubules that in principle would have shrunk back, do not have enough
shrink − qns.
time to do so because they are severed at 2d. We denote this probability with qns, and hence qshrink = p(1)
The derivation of qns can be found in Appendix B.
Therefore, the final expressions for the probabilities of resolving a crossover with a severing and with a shrinkage are
p(2)
shrink
shrink − qns + qsev
(cid:0)p+(cid:1) = p(1)
(cid:0)p+(cid:1) ,
(cid:0)p+(cid:1) = p(1)
(cid:0)p+(cid:1) .
(cid:0)p+(cid:1) = qsev,− − (qsev,− − qsev,+) p+,
sev + qns − qsev
p(2)
sev
Notice that qsev (p+) can be rewritten as
qsev
(20)
(21)
where the term in the braces is always positive. Indeed, since a microtubule initially in the growing state takes more
time to completely depolymerize than a microtubule in the shrinking state, its probability of resolving the crossover at
d before being severed there is smaller than in the opposite case. Consequently, from Eq. (21) the probability p(2)
sev (p+)
of resolving a crossover with a severing event grows linearly with p+.
severing by calculating the probabilities pcr (p+) and 1
N
By making use of this two-crossovers theory, we finally give a new estimate of the critical probability of rescue-after-
(cid:104)1 − δci,0(cid:105) of Eq. (11). In order to do that, we first define
p2d as the probability to have a severing event at 2d before an eventual severing event at d. The derivation of p2d can
be found in the Appendix B.
i=1
N(cid:80)
We now define the three events A, B, and C as
A = shrinkage of microtubule below d after severing at 2d,
B = severing event at 2d before an eventual severing event at d,
C = severing event at either d or 2d.
The three events are nested as A ⊂ B ⊂ C, and their probabilities are P (A) = qsev (p+), P (B) = p2d, and
P (C) = p(2)
sev (p+) − qsev (p+) + p2d = p(1)
sev − qns + p2d. Thus, it holds
(cid:0)p+(cid:1) = P (AB) =
pcr
P (A ∩ B)
P (B)
qsev (p+)
,
p2d
P (A)
P (B)
=
=
and
N(cid:88)
i=1
1
N
(cid:104)1 − δci,0(cid:105) = P (BC) =
P (B ∩ C)
P (C)
p2d
sev − qns + p2d
p(1)
P (B)
P (C)
=
=
(22)
(23)
.
Dynamic parameters
v−
v+
rc
s−1
µm s−1 µm s−1
0.10
0.08
0.15
0.10
0.10
0.10
0.10
0.10
0.08
0.12
0.08
rr
s−1
0.250 0.020 0.020
0.275 0.016 0.022
0.225 0.020 0.020
0.250 0.030 0.015
0.250 0.015 0.030
0.250 0.030 0.015
0.275 0.020 0.030
0.250 0.010 0.020
0.225 0.015 0.025
0.225 0.020 0.025
0.250 0.002 0.020
12
Critical point Critical point Relative error Critical point Relative error
(simulations) (1-cross. theory)
(2-cross. theory)
p+
crit
-
0.360
0.338
0.142
0.882
0.089
0.800
0.285
0.054
0.208
0.175
0.510
p+
crit,(1)
-
0.316
0.297
0.108
0.819
0.068
0.733
0.240
0.041
0.179
0.140
0.455
∆p+
(1)
-
0.122
0.121
0.239
0.071
0.236
0.084
0.158
0.241
0.139
0.200
0.108
p+
crit,(2)
-
0.361
0.337
0.144
0.864
0.103
0.780
0.285
0.066
0.213
0.179
0.497
∆p+
(2)
-
0.003
0.003
0.014
0.020
0.157
0.025
0.000
0.222
0.024
0.023
0.025
Table II. Comparison p+
crit vs pcrit,(1) vs pcrit,(2) for different sets of dynamic parameters as v+, v−, rc, and rr. All other model
represent the relative error of the one and
parameters are those of Table I. ∆p+
two-crossovers theory to the computationally measured critical value for the probability of rescue-after-severing.
crit−p+
p+
p+
crit
crit−p+
p+
p+
crit
and ∆p+
(2) =
(1) =
crit,(1)
crit,(2)
Plugging pcr and 1
N
threshold for the probability of rescue-after-severing p+
p+
crit,(2) =
(cid:80)N
i=1 (cid:104)1 − δci,0(cid:105) into the equality associated to inequality (11), we finally calculate the critical
2S∆qsevβ(cid:0)1 − R−
d (d)(cid:1)
(cid:26)(cid:0)M +
0 − Sαβ(cid:1)(cid:0)1 − R−
(cid:113)(cid:2)(cid:0)M +
0 − Sαβ(cid:1)(cid:0)1 − R−
d (d)(cid:1) − S∆qsevβR−
d (d)(cid:1) + S∆qsevβR−
d (d)(cid:3)2 − 4S∆qsevβ(cid:0)1 − R−
d (d)(cid:1)(cid:27)
crit,(2), that is
d (d)
(24)
×
+
1
,
where
α = p(1)
sev + qns − qsev,−,
β =
1
sev − qns + p2d
p(1)
,
∆qsev = qsev,− − qsev,+.
Table II shows a very good agreement between our predicted critical probability of rescue-after-severing in the
two-crossovers approximation and the critical probability obtained with our simulations in the whole grid of transverse
microtubules for different choices of dynamic parameters, confirming our hypothesis that, in order to study the critical
properties of the system, we can approximate the entire grid of transverse microtubules with just two of them without
any considerable loss of accuracy.
IV. DISCUSSION
Our aim was to obtain a deeper insight into the conditions under which templated severing of microtubules at
microtubule crossovers can lead to exponential proliferation of a new population of microtubules, as observed in the
recent experiments on the light-induced reorientation of the plant microtubule cortical array. To that end we separately
considered the role of the microtubule growth state, be it bounded or unbounded, and that of the rescue-after-severing
effect previously identified as a key component of the amplification process. Simulations revealed a striking difference
between the unbounded and the bounded microtubule growth regimes. In the unbounded-growth regime, which
13
appears to be salient for the experimental situation, amplification due to templated severing will occur even in the
absence of rescue-after-severing. The reason is that in this growth regime microtubules in principle have infinite
lifetime, allowing them (and their descendants after severing) to be severed without limit, which by itself is sufficient
to drive the amplification. There still is a role for the probability of rescue-after-severing, but only as a moderator for
rate of amplification and the probability of success per microtubule. In contrast, in the bounded-growth regime an
microtubule can in principle only be severed a finite number of times. In this case amplification can only occur if the
process is biased by a sufficiently high probability of rescue-after-severing. When the system is below a critical value
of this parameter, a newly nucleated microtubule, and all of its descendants through severing, is sure to go extinct.
The value of this critical rescue-after-severing probability depends strongly on the probability of a newly-severed
microtubule to cross the interval between neighboring transverse microtubules, so that it can be severed in turn, a
crucial step in the amplification process. This prompted us to develop a novel approach to calculating the appropriate
first passage time distribution, using an approach that may find application in other stochastic systems as well. This
formed the basis of approximate calculation of the critical rescue-after-severing probability, which compares favourably
with the results obtained from simulations.
While our work sheds light on the initial phase of the amplification process, understanding the later stages and the
stability of the final state remains a challenging problem. Here we have neglected a number of important effects. First,
the transverse microtubules were taken to be inert, while in reality they are also dynamic and will tend to be broken
down over time as more and more of the available tubulin is incorporated into the exponentially growing population of
longitudinal microtubules. This will remove opportunities for severing, and therefore tend to dampen the amplification
again. Moreover, as the amplification process develops, the availability of free tubulin dimers, which surely are a
limited resource in the cell, is also bound to decrease, which in turn affects both the growth dynamics and nucleation
rate. Given our results here, the first effect, depression of the growth speed, could in fact switch the microtubules from
the unbounded to the bounded-growth regime, which likely decelerates the amplification process. We are currently
exploring these issue, which will be the subject of a follow-up paper.
The work of MS was supported by the ERC 2013 Synergy Grant MODELCELL. The work of BMM is part of the
research program of the Dutch Research Council (NWO).
ACKNOWLEDGMENTS
V. APPENDIX A: MAIN FEATURES OF THE DOGTEROM-LEIBLER MODEL
A. Splitting probabilities in the interstitial strip
If a microtubule plus end impinges on a transverse microtubule, it creates a crossover. After the creation of the
crossover the plus end is located at x ∈ (nd, (n + 1)d), and, as long as this condition is fulfilled, the dynamics of
microtubules is described by the Dogterom-Leibler model for microtubules with their minus end at nd, regardless the
occurrence of a severing event. Without any loss of generality, we can set n = 0 and length l = x.
Due to the dynamic instability of the plus end, the microtubule either reaches length x = d or shrinks back to length
0 (x)
d (x), that describe the probability that a microtubule with initial state σ and initial length x arrives first at
x = 0. The occurrence probability of either of these events is described by the so-called splitting probabilities Rσ
and Rσ
length 0 or d respectively. Conservation of probability implies that Rσ
0 (x) + Rσ
d (x) = 1.
It is possible to show [13] that
R+
d (x) =
(25)
(26)
(27)
ex/l − rrv+
rcv−
ed/l − rrv+
rcv−
,
(cid:16)
(cid:17)
,
R−
d (x) =
ex/l − 1
rrv+
rcv−
ed/l − rrv+
rcv−
R+
0 (x) =
ed/l − ex/l
ed/l − rrv+
rcv−
,
and
R−
0 (x) =
rcv− ex/l
ed/l − rrv+
ed/l − rrv+
rcv−
.
14
(28)
Interestingly, these expressions hold for both bounded and unbounded-growth case. This is a direct consequence of the
fact that in a strip both regimes produce a steady-state solution [11].
B. Microtubule lifetime and survival probability
distribution of the time needed by microtubules to completely depolymerize.
The lifetime density function Lσ (tx) of a microtubule with initial length x and initial state σ, is defined as the
In the bounded-growth regime all microtubules have a finite lifetime, hence Lσ (tx) is normalized to 1. However,
in the unbounded-growth a fraction of microtubules grows linearly in time. It follows that for unbounded-growth
microtubules, the lifetime density function can be defined only for the fraction of microtubules the lifetime of which is
finite.
In the bounded-growth regime, the lifetime density functions are [14]
(cid:16)
(cid:34)
(cid:17)
(cid:115)
(cid:17)
L+ (tx) = Θ
×
t − x
v−
(cid:18)
x I0
rc
v+t + x
e−[rr(v+t+x)+rc(v−t−x)]
(cid:19)
(cid:112)rrrc (v+t + x) (v−t − x)
2
v+ + v−
+
v+
rc
t − x
v−
(cid:16)
L− (tx) = δ
rc (v−t − x)
rr (v+t + x)
I1
e−rrt + Θ
v+ + v−
2
(cid:17)(cid:114)
(cid:16)
t − x
v−
(cid:18)
× e−[rr(v+t+x)+rc(v−t−x)]I1
2
v+ + v−
(cid:18)
(cid:19)(cid:35)
(cid:112)rrrc (v+t + x) (v−t − x)
rcrr
(v+t + x) (v−t − x)
(cid:19)
(cid:112)rrrc (v+t + x) (v−t − x)
x
,
,
(29)
(30)
where I0 (·) and I1 (·) are the modified Bessel functions of order 0 and 1, respectively.
In order to obtain the densities in the unbounded-growth regime, we need to by divide Eq. (29) and Eq. (30) by
1 − S+ (∞x) and 1 − S− (∞x) respectively, where Sσ (∞x) is the fraction of microtubules with initial length x and
initial state σ that never completely depolymerize. Due to their finite lifetime, in the bounded-growth regime these
fractions are identically 0. The fractions Sσ (∞x) are called ultimate survival probabilities, and they are
(cid:40)
(cid:105)
x
(cid:104)− rrv+−rcv−
(cid:105)
(cid:104)− rrv+−rcv−
v+v−
x
v+v−
1 − rcv−
0
rrv+ exp
(cid:40)
1 − exp
0
S+ (∞x) =
S− (∞x) =
if unbounded-growth regime,
if bounded-growth regime,
if unbounded-growth regime,
if bounded-growth regime.
(31)
(32)
VI. APPENDIX B: DERIVATION OF THE SIZE OF THE OFFSPRING OF A MICROTUBULE
Here, we derive the expression for the size of the offspring of a microtubule in the one and two-crossovers approxima-
tions.
First, we introduce the one-crossover approximation by removing the dependency on p+ from the r.h.s. of Eq. (8).
= 0 for every ji. This implies that, when a severing event occurs at nd, n > 1, then all previous
We assume that bcji
crossovers are resolved by a severing event. With this approximation, we replace mi with
m(1)
i = si +
cji,
(33)
si(cid:88)
ji=1
see Figure 6. Analytically, we cannot calculate neither si nor cji, but these quantity are easily measurable with
computer simulations. We average m(1)
over N = 105 simulations to find the first approximation for M +, i.e.
i
M +
(1) = 1
N
m(1)
i = 2.61. Therefore, from Eq. (7), we can calculate the first estimate of the critical probability of
N(cid:80)
i=1
rescue-after-severing, i.e. p+
crit = 0.360. Table II shows a
comparison between p+
crit for different sets of dynamic parameters. The table shows that, even though our
one-crossover approximation provides a reasonable estimate of the critical probability, we systematically underestimate
it.
crit,(1) = 0.316, against the computationally measured one p+
crit,(1) and p+
15
Figure 6. Schematic of the count of the size of the offspring m(1)
of a microtubule labelled by 0 created by severing in the
growing state. When a crossover is created, the competition severing-shrinking takes place, and if the severing occurs, the
counter for the number of severing events s0 gains one unity, whilst the size of the offspring gains 1 + cs0 . We keep track of
the leading microtubule as it can generate other descendants, further increasing m(1)
0 . We do not keep track of the lagging
microtubules created by severing.
0
Now, we introduce the two-crossovers approximation by assuming that after a severing at nd, n > 1, all crossovers
at d, 2d, . . . , (n − 2) d are resolved by a severing event, while the crossover at (n − 1) d is resolved by a shrinkage with
probability pcr (p+) and by a severing with probability 1 − pcr (p+). Our aim is to give a better estimate of p+
crit then
in the one-crossover approximation. Here
With this definition for bcji
, we approximate mi with
bcji
(cid:0)p+(cid:1) .
(cid:1) pcr
=(cid:0)1 − δcji ,0
(cid:104)
si(cid:88)
cji −(cid:0)1 − δcji ,0
(cid:1) pcr
m(2)
i = si +
ji=1
(cid:0)p+(cid:1)(cid:105)
.
(34)
0001210 From this equation, we can observe that
(cid:0)1 − δcji ,0
(cid:1) pcr
si(cid:88)
ji=1
(cid:0)p+(cid:1) = pcr
= pcr
= pcr
(cid:1)(cid:105)
(cid:0)p+(cid:1)(cid:104)
si −(cid:0)δc1,0 + δc2,0 + ··· + δcsi ,0
(cid:0)p+(cid:1)(cid:104)
si − si (cid:104)δci,0(cid:105)(cid:105)
(cid:0)p+(cid:1) si (cid:104)1 − δci,0(cid:105) ,
16
(35)
where the average value (cid:104)δci,0(cid:105) is calculated over all severing events that a leading microtubule undergoes along its
lifetime. Consequently, (cid:104)1 − δci,0(cid:105) is the fraction of severing events that a leading microtubule undergoes at nd with
n > 1. If we combine Eqs. (33), (34), and (35) together, and we average over N , we obtain
N(cid:88)
N(cid:88)
i=1
i=1
M +
(2) =
=
1
N
1
N
m(2)
i
(cid:104)
si +
si(cid:88)
(cid:0)p+(cid:1) S
ji=1
cji − pcr
N(cid:88)
1
N
i=1
= M +
(1) − pcr
(cid:104)1 − δci,0(cid:105) ,
(cid:0)p+(cid:1) si (cid:104)1 − δci,0(cid:105)(cid:105)
(36)
N(cid:80)
i=1
where S = 1
N
si, and where we assumed that the correlation between the number of severing events that occur
along the lifetime of a microtubule and the fraction of them that occur at nd with n > 1 is neglectable. In this case, if
N (cid:29) 1, for the law of large numbers 1
(cid:104)1 − δci,0(cid:105) is the probability that a microtubule is severed at nd with n > 1,
sampled over all cases in which a severing event has occurred. By replacing M + with M +
(2) in Eq. (6), we obtain the
final amplification condition (11). Table II shows that the two-crossovers reproduces the computationally measured
critical probability of rescue-after-severing with a good degree of accuracy.
i=1
N
N(cid:80)
VII. APPENDIX C: DERIVATION OF SEVERING AND SHRINKAGE PROBABILITIES
In order to calculate qsev,σ we first define the following random variables:
τd = severing waiting time at d,
τ2d = severing waiting time at 2d,
Td = FPT from the first to the second crossover, i.e. from d to 2d,
Tσ (x) = lifetime of a microtubule with initial state σ and initial length x,
(cid:101)τ2d = severing waiting time at 2d given that the severing occurs.
(37)
τd and τ2d have probability density function Wkθ (t) defined in Eq. (1), while the probability density function of Td is
F0d (t) from Eq. (14). The probability density function of Tσ (x) is L+ (tx) defined in the Appendix A. Finally, the
probability density function of(cid:101)τ2d can be calculated by observing that the event "severing" and the event "shrinkage"
are independent. Therefore, the cumulative function Φ(cid:101)τ2d (t) = P [(cid:101)τ2d < t] can be written as
Φ(cid:101)τ2d (t) = P [(τ2d < t) ∩ (τ2d < T+ (0))] =
dt(cid:48) Wk,θ (t(cid:48))
1
ZW
(cid:90) t
(cid:90) ∞
0
(cid:90) ∞
(cid:90) ∞
t(cid:48)
where
ZW =
Thus, the probability density function (cid:102)Wkθ (t) of(cid:101)τ2d is
Φ(cid:101)τ2d (t) =
(cid:102)Wkθ (t) =
0
d
dt
dt Wk,θ (t)
dt(cid:48) L+ (t(cid:48)0) .
t
1
ZW
Wkθ (t)
(cid:90) ∞
t
dt(cid:48) L+ (t(cid:48)0) .
dt(cid:48)(cid:48) L+ (t(cid:48)(cid:48)0) ,
(38)
(39)
(40)
The probability qsev,σ is the probability that a microtubule reaches 2d, it is severed there with newly-created plus
end in the state σ, and finally shrinks back below d before being severed at d. In S+ (∞0) of the cases (i.e.
for
indefinitely growing microtubules), this event occurs if T1 = τd − Td − τ2d − Tσ (d) > 0, with probability density
function of T1 defined by
(41)
In the remaining 1 − S+ (∞0) of the cases (i.e. for microtubules with a finite lifetime), the event occurs if T2 =
τd − Td −(cid:101)τ2d − Tσ (d) > 0, and if it is severed at 2d, i.e. if τd < T+ (0). The probability density function of T2 is
dt(cid:48)dt(cid:48)(cid:48)dt(cid:48)(cid:48)(cid:48) Wkθ (t + t(cid:48) + t(cid:48)(cid:48) + t(cid:48)(cid:48)(cid:48)) F0d (t(cid:48)) Wkθ (t(cid:48)(cid:48)) Lσ (t(cid:48)(cid:48)(cid:48)d) .
PT1 (t) =
R3
dt(cid:48)dt(cid:48)(cid:48)dt(cid:48)(cid:48)(cid:48)(cid:102)Wkθ (t + t(cid:48) + t(cid:48)(cid:48) + t(cid:48)(cid:48)(cid:48)) F0d (t(cid:48)) Wkθ (t(cid:48)(cid:48)) Lσ (t(cid:48)(cid:48)(cid:48)d) .
(cid:27)
(cid:90) ∞
(cid:90) ∞
PT2 (t) =
(cid:26)
Hence, the final expression for qsev,σ is
S+ (∞0)
dt PT1 (t) + [1 − S+ (∞0)]
qsev,σ = R+
d (0)
To calculate qns we first define the random variable (cid:101)T+ as the time that a microtubule initially in the growing state
Similarly to the derivation of (cid:102)Wkθ (t), we can derive the probability density function of (cid:101)T+, that is
and with plus end in 2d needs in order to return in the shrinking state at 2d, given that no severing event occurs at 2d.
0
0
0
dt Pτd−T0 (t)
dt PT2 (t)
[1 − S+ (∞d)] .
(43)
(cid:90)
(cid:90)
(cid:90) ∞
R3
17
(42)
(cid:101)L+ (t) =
(cid:90) ∞
ZL =
L+ (t0)
1
ZL
dt L+ (t0)
0
0
(cid:90) t
(cid:90) t
0
dt(cid:48) Wk,θ (t(cid:48)) ,
dt(cid:48) Wk,θ (t(cid:48)) .
(44)
(45)
with
Therefore, as qns is the probability that a microtubule reaches length 2d and would return to length d but it
cannot because it is severed at 2d, the two conditions that our random variables have to fullfil are τ2d < T+ (0) and
τd > Td + (cid:101)T+ + T− (d). The former condition had already been discussed before, whilst the latter is associated to the
probability density function
Pτd−Td−(cid:101)T+−T−(d) (t) =
Therefore
(cid:90)
R3
qns = R+
0d [1 − S+ (∞0)]
dt(cid:48)dt(cid:48)(cid:48)dt(cid:48)(cid:48)(cid:48) Wkθ (t + t(cid:48) + t(cid:48)(cid:48) + t(cid:48)(cid:48)(cid:48)) F0d (t(cid:48))(cid:101)L+ (t(cid:48)(cid:48)) L− (t(cid:48)(cid:48)(cid:48)d) .
(cid:90) 0
(cid:90) ∞
dt Pτd−Td−(cid:101)T+−T−(d) (t) .
dt Pτd−T+(0) (t)
−∞
0
Finally, in order to calculate the probability p2d to have a severing event at 2d before an eventual severing event at d,
we notice that we have two different cases. In the first case, the microtubule reaches length 2d and it is severed there
before being severed at d, i.e. τd > Td + τ2d. In the second case, the microtubule reaches 2d and it is severed there
before being severed at d, i.e. τd > Td +(cid:101)τ2d, given that the event "severing" wins the competition against the event
"shrinkage" at 2d, or τ2d < T+ (0). The probability density functions associated to these conditions are, respectively,
(cid:90)
(cid:90)
R2
R2
Pτd−Td−τ2d (t) =
Pτd−Td−(cid:101)τ2d (t) =
dt(cid:48)dt(cid:48)(cid:48) Wkθ (t + t(cid:48) + t(cid:48)(cid:48)) F0d (t(cid:48)) Wkθ (t(cid:48)(cid:48)) ,
dt(cid:48)dt(cid:48)(cid:48) Wkθ (t + t(cid:48) + t(cid:48)(cid:48)) F0d (t(cid:48))(cid:102)Wkθ (t(cid:48)(cid:48)) ,
Then
p2d = R+
0d
(cid:26)
S+ (∞0)
(cid:90) ∞
0
Pτ2d−T+(0) (t) = Pτd−T+(0) (t) .
dt Pτd−Td−τ2d (t) + [1 − S+ (∞0)]
dt Pτ2d−T+(0) (t)
(cid:90) 0
−∞
(cid:90) ∞
0
(cid:27)
dt Pτd−Td−(cid:101)τ2d (t)
(46)
(47)
(48)
(49)
(50)
.
(51)
18
[1] Bruce Alberts, Alexander Johnson, Julian Lewis, Martin Raff, Keith Roberts, and Peter Walter. Molecular Biology of the
Cell. Garland Science, 2007.
[2] Tim Mitchison and Marc Kirschner. Dynamic instability of microtubule growth. Nature, 312(5991):237 -- 242, 11 1984.
[3] John C Sedbrook. MAPs in plant cells: delineating microtubule growth dynamics and organization. Current Opinion in
Plant Biology, 7(6):632 -- 640, 12 2004.
[4] Ram Dixit and Richard Cyr. The cortical microtubule array: from dynamics to organization. The Plant cell, 16(10):2546 -- 52,
10 2004.
[5] Jelmer J Lindeboom, Masayoshi Nakamura, Anneke Hibbel, Kostya Shundyak, Ryan Gutierrez, Tijs Ketelaar, Anne Mie C
Emons, Bela M Mulder, Viktor Kirik, and David W Ehrhardt. A mechanism for reorientation of cortical microtubule arrays
driven by microtubule severing. Science (New York, N.Y.), 342(6163):1245533, 12 2013.
[6] Masayoshi Nakamura, Jelmer J Lindeboom, Marco Saltini, Bela M Mulder, and David W Ehrhardt. SPR2 protects minus
ends to promote severing and reorientation of plant cortical microtubule arrays. The Journal of cell biology, 217(3):915 -- 927,
3 2018.
[7] P.T. Tran, R.A. Walker, and E.D. Salmon. A Metastable Intermediate State of Microtubule Dynamic Instability That
Differs Significantly between Plus and Minus Ends. The Journal of Cell Biology, 138(1):105 -- 117, 7 1997.
[8] Marileen Dogterom and Stanislas Leibler. Physical aspects of the growth and regulation of microtubule structures. Physical
Review Letters, 70(9):1347 -- 1350, 3 1993.
[9] Jelmer J Lindeboom, Masayoshi Nakamura, Marco Saltini, Anneke Hibbel, Ankit Walia, Tijs Ketelaar, Anne Mie C Emons,
John C Sedbrook, Viktor Kirik, Bela M Mulder, and David W Ehrhardt. CLASP stabilization of plus ends created by
severing promotes microtubule creation and reorientation. The Journal of cell biology, page jcb.201805047, 10 2018.
[10] Athanasios Papoulis and S. Unnikrishna Pillai. Probability, random variables, and stochastic processes. Tata McGraw-Hill,
2002.
[11] Bindu S. Govindan and William B. Spillman. Steady states of a microtubule assembly in a confined geometry. Physical
Review E, 70(3):032901, 9 2004.
[12] Jan W. Vos, Marileen Dogterom, and Anne Mie C. Emons. Microtubules become more dynamic but not shorter during
preprophase band formation: A possible \"search-and-capture\" mechanism for microtubule translocation. Cell Motility
and the Cytoskeleton, 57(4):246 -- 258, 4 2004.
[13] Bela M. Mulder. Microtubules interacting with a boundary: Mean length and mean first-passage times. Physical Review E,
86(1):011902, 7 2012.
[14] D. J. Bicout. Green's functions and first passage time distributions for dynamic instability of microtubules. Physical Review
E, 56(6):6656 -- 6667, 12 1997.
|
1209.0205 | 1 | 1209 | 2012-09-02T19:31:55 | On correlation between protein secondary structure, backbone bond angles, and side-chain orientations | [
"physics.bio-ph",
"cond-mat.soft",
"q-bio.BM"
] | We investigate the fine structure of the sp3 hybridized covalent bond geometry that governs the tetrahedral architecture around the central C$_\alpha$ carbon of a protein backbone, and for this we develop new visualization techniques to analyze high resolution X-ray structures in Protein Data Bank. We observe that there is a correlation between the deformations of the ideal tetrahedral symmetry and the local secondary structure of the protein. We propose a universal coarse grained energy function to describe the ensuing side-chain geometry in terms of the C$_\beta$ carbon orientations. The energy function can model the side-chain geometry with a sub-atomic precision. As an example we construct the C$_\alpha$-C$_\beta$ structure of HP35 chicken villin headpiece. We obtain a configuration that deviates less than 0.4 \.A in root-mean-square distance from the experimental X-ray structure. | physics.bio-ph | physics | On correlation between protein secondary structure,
backbone bond angles, and side-chain orientations
Martin Lundgren1, ∗ and Antti J. Niemi2, 1, †
1Department of Physics and Astronomy, Uppsala University, P.O. Box 803, S-75108, Uppsala, Sweden
2 Laboratoire de Mathematiques et Physique Theorique CNRS UMR 6083,
F´ed´eration Denis Poisson, Universit´e de Tours, Parc de Grandmont, F37200, Tours, France
We investigate the fine structure of the sp3 hybridized covalent bond geometry that governs the
tetrahedral architecture around the central Cα carbon of a protein backbone, and for this we develop
new visualization techniques to analyze high resolution X-ray structures in Protein Data Bank. We
observe that there is a correlation between the deformations of the ideal tetrahedral symmetry
and the local secondary structure of the protein. We propose a universal coarse grained energy
function to describe the ensuing side-chain geometry in terms of the Cβ carbon orientations. The
energy function can model the side-chain geometry with a sub-atomic precision. As an example we
construct the Cα-Cβ structure of HP35 chicken villin headpiece. We obtain a configuration that
deviates less than 0.4 A in root-mean-square distance from the experimental X-ray structure.
PACS numbers: 87.15.Cc 05.45.Yv 36.20.Ey
I: INTRODUCTION
Protein structure validation is based on various well
tested and broadly accepted stereochemical paradigms.
Methods such as MolProbity [1] and Procheck [2] and
many others help crystallographers to find and fix poten-
tial problems during fitting and refinement. Stereochemi-
cal assumptions are also instrumental to structure predic-
tion packages such as Rosetta and I-Tasser [3]. Likewise,
they form the foundation for parameter determination in
force fields such as Charmm and Amber [4] that aim to
describe protein dynamics at atomic scale.
One of the paradigms is the transfersability assump-
tion. It states that stereochemical restraints are universal
and independent of the environment. Among its conse-
quences are that the covalent bond geometry around the
backbone Cα should seize a very precise tetrahedral sp3
hybridized shape. For example, the backbone
τN C ≡ (N − Cα − C)
bond angle should oscillate around a computable average
value that depends only on the covalent bonds between
the Cα and the N, C, H and Cβ atoms in the trans-
peptide group. In particular, at least to to the leading
order its value should not depend on the character of the
secondary structure environment. Standard molecular
dynamics force fields explicitly assume this to be the case.
These force fields are based on a harmonic approximation
where the bond angles κ oscillate with energy [4]
(cid:88)
Ebond =
ωκ(κ − κ0)2
(1)
bonds
Here ωκ and κ0 are parameters that are in general amino
acid dependent. But these parameters are presumed to
be independent of the geometry of the surrounding sec-
ondary structure. Instead, they are supposed to predict
the local secondary structure environment.
The enormous success that has been enjoyed by the
validation methods and structure prediction programs in
resolving close to 80.000 crystallographic protein struc-
tures that are presently in Protein Data Bank (PDB)
[5] is a clear manifestation that the various paradigms
are valid to a good precision. However, with the ad-
vent of third-generation synchrotron sources of X-rays,
there is now a small but rapidly expanding number of
protein structures that are resolved with an ultrahigh
sub-Angstrom resolution. The present, third-generation
X-ray synchrotron sources such as ESRF in Grenoble and
PETRA at DESY in Hamburg can already produce pho-
tons with wavelengths as short as 10 pico-meters. Thus
it is in principle possible to obtain three dimensional pro-
tein structures with a comparable resolution. The next-
generation sources of high brilliance X-ray beams such
as the European X-Ray Free Electron Laser at DESY,
will push protein X-ray crystallography to its extreme.
These future experimental facilities can reach both ultra-
high spatial and temporal resolutions, with a fully co-
herent peak brightness that is many orders of magnitude
higher than what can be obtained with the present third-
generation synchrotron sources. The on-going experi-
mental revolution in combination with the ever expand-
ing need of higher precision for example in the study of
protein-protein interactions, enzyme catalysis and search
of causes for protein misfolding related diseases, are good
incentives for us to scrutinize the level of precision in
some of the paradigm assumptions on protein backbone
geometry. And, if need be, to try and develop new the-
oretical concepts that aim to describe proteins at a pre-
cision that matches the highest present and near future
experimental standards, in revealing the finer structures
of folded proteins.
In fact, ab initio quantum mechanical calculations [6]
and empirical studies [7]-[9] of protein backbone geome-
try have already disclosed that the backbone bond angle
2
1
0
2
p
e
S
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
5
0
2
0
.
9
0
2
1
:
v
i
X
r
a
τN C ≡ (N-Cα-C) about the Cα carbons might oscillate
quite substantially. The range of variations can be as
large as 8.8o [8]. This corresponds to a shift of ∼ 0.6
A in the relative positioning of two consecutive Cα car-
bons. A deviation of this size from the ideal value can
be subjected to experimental scrutiny in X-ray experi-
ments that reach sub-Angstrom resolution.
Indeed, on
the basis of existing data the authors [7]-[9] have already
reported that the deviations in the values of the τN C an-
gle are systematic, and in particular that these deviations
reflect the local secondary structure.
The τN C angles are primarily affected by the back-
bone. As such, their values relate directly to the two
standard Ramachandran angles, that form the basis for
structure validation. As a consequence, the literature
[6]-[9] has until now mainly concentrated on the effects
that potential deviations of τN C from ideality have on
the backbone geometry. Here we extend this analysis to
the side-chains: The fluctuations in the lengths of the
covalent bonds in the Cα tetrahedron are no more than
around 0.1 A which is much less than the potential ∼ 0.6
A shift in the relative positioning of two consecutive Cα
carbons, due to τN C fluctuations [7]-[9]. This proposes
that any deviation of τN C from its ideal value inevitably
propagates to the side-chain dependent τN β ≡ (N-Cα-
Cβ) and τCβ ≡ (C-Cα-Cβ) bond angles, and this should
lead to observable effects in the angular positions of the
side-chain Cβ atoms.
In this article we first analyze PDB data to find
whether there are experimental variations in the tetra-
hedral angles around the Cα. In particular, we extend
the analysis of [7]-[9] to study correlations between the
side-chain dependent angles τN β and τCβ that determine
the Cβ orientations, and the local secondary structure of
the backbone. Since the side-chain atom positions are not
easily described in terms of the backbone Ramachandran
angles, we start by developing new visualization tools. In
line with [7]-[9] we observe that the local secondary struc-
ture has a systematic effect on the relative tetrahedral
position of the Cβ carbon. We then proceed to utilize
our visualization tools to develop theoretical arguments.
We propose a coarse-grained framework that computes
how the observed direction of the Cβ evolves along the
backbone. In particular, we argue that the direction of
the Cβ can be computed from the soliton solution of a
discrete nonlinear Schrodinger (DNLS) equation. The
DNLS soliton already shares a remarkable history with
protein research [10]. Both the DNLS equation and its
soliton solution were first introduced by Davydov to de-
scribe the propagation of energy along α-helices [11]. He
also proposed that since the propagation leads to a local
deformation of the protein shape, a trapped soliton is a
natural cause for the protein to fold. Here we first ar-
gue on general grounds that the DNLS soliton solution
can be utilized to determine the secondary structure de-
pendence in the relative direction of the Cβ atoms along
2
the backbone. We then consider an explicit example to
illustrate our general arguments. The example we con-
sider is the 35-residue subdomain of the villin headpiece
with PDB code 1YRF. It is a paradigm protein that has
been studied widely in theoretical approaches to protein
folding.
II: VISUALIZATION OF THE Cα
TETRAHEDRON
We start by visual analysis of crystallographic protein
data in PDB. The goal is to reveal any secondary struc-
ture dependence in the values of τN C, and in the adjacent
and
τN β ≡ (N − Cα − Cβ)
τCβ ≡ (C − Cα − Cβ)
bond angles. In order to minimize any bias, we inspect
several subsets of PDB. These include the canonical one
that comprises all PDB configurations with resolution
2.0 A or better, and its two subsets with resolution bet-
ter than 1.5 A, and better than 1.0 A. We also inspect a
subset of the 2.0 A set that contains only those proteins
that have less than 30% sequence similarity. Our conclu-
sions are independent of the data set, and for illustrative
purposes we here use the canonical 2.0 A set. There are
presently over 30.000 such entries in PDB.
The Ramachandran angles are defined in terms of the
backbone amide planes. As such they are not the most
convenient ones for describing the side-chain geometry.
Since both τN β and τCβ relate to the side-chain geome-
try, we prefer to follow [12] and describe the folded pro-
tein structure in terms of the geometrically determined
backbone discrete Frenet frames (DFF). These frames
govern the entire backbone neighborhood, including the
side-chains. But their construction involves only the Cα
coordinates ri where i = 1, ..., N label the residues. As
such, these frames then give a manifestly N, Cβ and C
independent, purely geometric description of the tetra-
hedral sp3 neighborhood of the Cα atoms.
The backbone tangent vectors are
ri+1 − ri
ri+1 − ri
ti =
The unit binormal vectors are
bi =
ti−1 − ti
ti−1 − ti
The unit normal vectors
ni = bi × ti
(2)
(3)
(4)
The orthogonal triplets (ni, bi, ti) determine the discrete
Frenet frame at each of the positions ri of the Cα car-
bons. Note that if the tangent vectors and the distances
between the Cα are known, we can reconstruct the entire
Cα backbone using
k−1(cid:88)
rk =
ri+1 − ri · ti
(5)
i=1
For the initial condition we can utilize Galilean invariance
to take r1 = 0, and t1 to point into the direction of the
positive z-axis.
In particular, (5) does not involve the
vectors ni and bi.
We introduce the backbone bond angles
cos κi+1 ≡ cos κi+1,i = ti+1 · ti
and the backbone torsion angles
cos τi+1 ≡ cos τi+1,i = bi+1 · bi
(6)
(7)
Note that both angles are manifestly independent of the
N, Cβ and C atoms that are covalently bonded to the
Cα atoms, in this sense they provide an unbiased set of
coordinates for describing the positions of these atoms.
If these angles are known, we can use
=
ni+1
bi+1
ti+1
cos κ cos τ cos κ sin τ − sin κ
− sin τ
sin κ cos τ sin κ sin τ
cos κ
cos τ
0
ni
bi
ti
i+1,i
ni
bi
ti
≡ Ri+1,i
(8)
to iteratively construct the Frenet frame at position i + i
from the frame at position i. Once we have all the frames,
we can proceed to construct the entire backbone using
(5).
The bond and torsion angles have a natural interpre-
tation in terms of the canonical latitude and longitude
angles of a two-sphere S2. In the sequel we find it useful
to extend the range of κi to [−π, π] mod(2π). But we in-
troduce no change in the range of τi ∈ [−π, π] mod(2π).
We compensate for this two-fold covering of S2 by the
following discrete Z2 symmetry
κk → −κk
τi → τi − π
for all k ≥ i
(9)
It inverts the directions of the vectors ni and bi but has
no effect on the ti and consequently leaves the backbone
intact. For details we refer to [12].
We use the discrete Frenet frames to display each atom
in the way, how the atom is seen on the surface of a
sphere that surrounds an imaginary observer who roller-
coasts the backbone along the Cα atoms, so that the gaze
direction is always fixed towards the next Cα and with
local orientation determined by the DFF frames [12].
In Figure 1 we show the statistical angular distribution
of the backbone N and C atoms, and in Figure 2 we show
3
the same for the side-chain Cβ atoms in our PDB data set
as seen by the Frenet frame observer who moves through
all the proteins in our data set. The sphere is centered
at the Cα, and its radius coincides with the length of the
(approximatively constant) covalent bond. We take the
vector t that points towards the next Cα to be in the
direction of the positive z-axis, towards the north-pole of
the sphere. With n in the direction of positive x-axis we
have a right-handed Cartesian coordinate system. We
introduce the canonical spherical coordinates (θ, ϕ) to
describe the distributions. The angle θ ∈ [0, π] measures
latitude from the positive z-axis, hence it describes the
distribution of the bond angles κi. The angle ϕ ∈ [−π, π]
measures longitude in a counterclockwise direction from
the x-axis i.e. from the direction of n towards that of b,
with ϕ = 0 at the x-axis. Consequently it describes the
distribution of the torsion angles τi.
FIG. 1:
(Color online) The directions of a) backbone N-
atoms and b) backbone C-atoms as seen by a Frenet frame
observer located at the Cα carbon which is at the center of
the sphere.
In a) the smaller, more point-like direction of
backbone N atoms corresponds to the L-α Ramachandran re-
gion. The larger region forms a segment of the great circle
ϕ ≈ −15o. Loops interpolate latitudinally between α-helices
and β-sheets. In b) the directions of backbone C form a seg-
ment of a small circle around z-axis, with θ ≈ 20o.
FIG. 2:
(Color online) The distribution of the Cβ directions
in the Frenet frame. In a) we have all amino acids (including
proline but excluding glycine that has no Cβ). In b) we show
only proline. Comparison between a) and b) exemplifies how
the Cβ direction can depend on the individual amino acid.
We have chosen proline in b) as it is particularly interesting
due to the way how it appears in Figure 3b).
We find that in the Frenet frame coordinate system,
the N and C oscillations shown in Figures 1a) and 1b)
are fully separated into the locally orthogonal θ and ϕ
directions, respectively. This would certainly not be the
case in a generic coordinate system. Furthermore, sec-
ondary structures such as α-helices, β-sheets, loops and
left-handed α-regions are all clearly identifiable in Fig-
ures 1. Figure 2a) reveals how the N and C oscillations
of Figure 1, through the covalent bonds that form the
sp3 tetrahedron around Cα, combine into a horseshoe
(annulus) shaped nutation of Cβ. As visible in the Fig-
ure, this nutation reflects the local secondary structure
environment in an equally systematic manner as Figures
1.
4
[9]. But the fact that τN C in Figure 3a) displays clearly
visible and systematic secondary structure dependence
makes it plain and clear that the paradigm of transfer-
able tetrahedral covalent symmetry around the Cα car-
bon is absent. Furthermore, the way how transferability
becomes violated reflects the wider secondary structure
environment of the amino acid along the protein back-
bone.
Since the deviation from the ideal tetrahedral symme-
try is organized in the same way how the proteins are
folded, these two must share a common origin. But we
do not have any physical explanation why the lack of ideal
symmetry is only visible in τN C. We suspect this has to
do with existing experimental refinement methods, the
way how refinement tension is distributed between the
backbone and side-chains. The high resolution crystallo-
graphic data which becomes available in future third and
fourth generation experiments should help to clarify this.
III: SOLITONS AND SIDE-CHAINS
FIG. 3:
(Color online) The normalized probability density
angular distribution of the a) τN C , b) τN β and c) τCβ angles in
degrees, with α-helices in red (grey), β-strands in blue (dark
grey), and 3/10 helices in yellow. The secondary peak in b) is
due to cis-peptide prolines. Only in Figure a) are the different
secondary structures visibly separated from each other.
In
Figures b) and c) they are practically fully overlapping, the
only notable effect is the somewhat higher propensity of 3/10
helices in connection of cis-peptide prolines.
The pattern of angular separation in the N and C os-
cillations in Figure 1 reveals that the underlying ideal
tetrahedral covalent symmetry around Cα is not trans-
ferable along the protein backbone [7]-[9].
We proceed to Figures 3 a)-c) where we plot the tetra-
hedral bond angles τN C, τCβ and τN β jointly for the α-
helices, β-strands and 3/10-helices. As in figure 1 the
loops interpolate continuously between these regular sec-
ondary structures. The Figure 3a) clearly reveals that at
the level of the τN C angle the transferability of the tetra-
hedral symmetry is absent in a systematic and secondary
structure dependent manner. But neither τCβ nor τN β
show any sign whatsoever of secondary structure depen-
dence. The distributions are practically the same, inde-
pendently of the secondary structure. (The isolated small
peak in Figure 3b) is due to proline cis-peptide groups.)
As such, the distribution in both Figures 3b) and 3c)
is what we would expect in the case of ideal, transferable
sp3 tetrahedral symmetry.
In each Figure the average
value is around 111o and there are secondary structure
independent fluctuations that are in line with quantum
mechanical estimates and [6] and empirical studies [7]-
Any molecular dynamics approach to protein folding
that we are familiar with, utilizes the harmonic approx-
imation (1) for bond and torsion angles. Here κ0 is in
general amino acid dependent but secondary structure
independent equilibrium value of the bond angle. But
from Figure 3a) we observe, that for the τN C angle there
are three different major equilibrium values. These equi-
librium bond angle values are amino acid independent,
but do depend in a nontrivial manner on the local sec-
ondary structure: The three peaks in Figure 3a) corre-
spond to the α-helix, β-strand, and 3/10-helix while for
a loop the values of the corresponding equilibrium κ0 in-
terpolates between these three ground state values. Since
each of these secondary structures are characterized by a
different equilibrium value of bond angle κ, to the leading
order we may take κ0 to be a function of κ. By expanding
to leading order we get
κ0 → κ0(κ) ≈ κ(0)
· κ2 + O(κ3)
· κ + κ(2)
0 + κ(1)
0
0
where the κ(i) are independent of the local value of κ.
The first two terms simply renormalize the values of ωκ
and κ0 in (1). But the third term is conceptually dif-
ferent. It introduces an anharmonic correction. We con-
clude that after redefinitions of the parameters, in the
leading order the potential obtains the functional form
Ebond ∼ q · (κ2 − m2)2
(10)
We argue that based on the Figure 3a), the anharmonic
corrections are already visible in the existing high res-
olution X-ray data.
In order to answer the theoretical
challenge that this poses, we propose to improve existing
MD force fields, to account for the anharmonic correc-
tions in the bond angle contribution.
A: Backbone energy
The presence of an anharmonic correction in the bond
angle energy has important implications to the way how
proteins fold. For this we start with a simple example.
We consider the anharmonic potential (10) in the pres-
ence of a single coordinate x on a line. The Newton's
equation is
mx = − dV
dx
We take the potential to have the form
V (x) =
σ
4
(x + b)2 · (x − a)2
(11)
We introduce
and define
c = − 1
2
(b + a)
y = x − 1
2
(a − b)
to arrive at the equation of motion
y = − σ
m
y(y2 − c2)
which is essentially the continuum nonlinear Schrodinger
equation (NLSE) [10], [11] Note that the potential has
the symmetric form (10). This equation is solved by
y(t) = c · tanh[ c
(t − t0)]
(cid:114) σ
2m
5
(or x ≡ b) through any kind of continuous finite energy
transformation.
In particular, a soliton configuration
such as (12) can not be obtained from any approach that
only accounts for perturbations that describe small local-
ized fluctuations around the uniform background ground
state.
In [13]-[15], it has been shown that the soliton pro-
file (12) can be used to describe loops in folded proteins.
For this one merges general geometric arguments with
the concept of universality [16]-[19], to arrive at the fol-
lowing simplified, coarse-grained energy function for the
backbone bond and torsion angles [20], [21],
E = − N−1(cid:88)
(cid:26)
N(cid:88)
2 κi+1κi +
i + q · (κ2
2κ2
i=1
+
dτ
2
i=1
i − bτ κ2
i τ 2
κ2
i τi − aτ τi +
cτ
2
τ 2
i
i − m2)2
(cid:27)
(13)
where κi and τi are the backbone bond and torsion angles
(6), (7). Unlike force fields in molecular dynamics, the
energy function (13) does not purport to explain the fine
details of the atomic level mechanisms that give rise to
protein folding. Instead, in line with general principles of
effective Landau-Lifschitz theories it describes the prop-
erties of a folded protein backbone in terms of universal
physical arguments.
Indeed, according to the concept
of universality [16]-[19] the energy function (13) can be
viewed as the universal long distance limit that emerges
from any atomic level energy function when the internal
energy is coarse-grained to include only the backbone
bond and torsion angles.
In order to construct the soliton solution, we start by
introducing the τ -equation of motion
⇒ x(t) = y(t) +
(a − b)
1
2
∂E
∂τi
= dτ κ2
i τi − bτ κ2
i − aτ + cτ τi = 0
√
= − b · ec
σ
2m (t−t0) − a · e
cosh[c(cid:112) σ
√
−c
2m (t − t0)]
σ
2m (t−t0)
(12)
⇒ τi[κ] =
aτ + bτ κ2
i
cτ + dτ κ2
i
(14)
This is the hallmark NLSE soliton configuration, so called
dark soliton solution of the NLSE equation. It interpo-
lates between the two uniform ground states at x = a
and x = −b when t → ±∞. The parameters a, b, t0 and
the combination
(cid:114) σ
c
Notice that even though there are four parameters in (14)
one of them, the overall scale, drops out. We then use
(14) to eliminate the torsion angles, so that the energy
for the bond angles becomes
N(cid:88)
(cid:19)
i=1
·
E[κ] = − N−1(cid:88)
V [κ] = −
2 κi+1κi +
i=1
(cid:18) bτ cτ − aτ dτ
(cid:19)
(cid:18) b2
τ + 8qm2
dτ
2bτ
−
2κ2
i + V [κi]
(15)
1
cτ + dτ κ2
· κ2 + q · κ4
(16)
2m
where
are the canonical ones that characterize the asymptotic
values of x(t) i.e. minima of the potential, and the size
and location of the soliton.
For finite t the soliton (12) describes a configuration
with an energy above the uniform ground state x ≡ a
(or x ≡ b). Nevertheless, it can not decay into x ≡ a
Because the first term contains κ in the denominator, its
variation with κ is not that pronounced as the variation
of the other two terms, which are proportional to the
second and the fourth power of κ, respectively. Moreover,
because κ > 1 radian for proteins, it turns out that the
first term is small in value compared to the other terms.
The second and third terms have then the functional form
of the double well potential (10).
In applications to folded proteins the parameters val-
ues are such, that in the energy ground state both κ and
τ acquire a non-vanishing value. In particular, since the
functional form of (16) is similar to (10), (11) we can
expect that there are soliton solutions:
Geometrically, a uniform constant value of the bond
and torsion angles describes regular protein secondary
structures. For example, the standard α-helix is
α − helix :
τ ≈ 1
and for the standard β-strand we have
2
β − strand :
(cid:26)κ ≈ π
(cid:26)κ ≈ 1
τ ≈ π
(17)
(18)
The additional regular secondary structures including
3/10 helices, left-handed helices etc. are described simi-
larly.
But in addition of constant value configurations, as in
(11) there are also soliton solutions. In particular, since
protein loops are structures that interpolate between dif-
ferent constant values such as (17), (18), this means that
loops correspond to these soliton solutions [13]-[15]. In
order to construct the relevant soliton, we introduce the
generalized discrete nonlinear Schrodinger (DNLS) equa-
tion that derives from the energy (13). Variation of this
energy w.r.t. κi and substitution of (14) gives
κi+1 = 2κi − κi−1 +
dV [κ]
dκ2
i
κi
(i = 1, ..., N )
(19)
where κ0 = κN +1 = 0. The exact soliton solution to
the present discrete nonlinear Schrodinger equation is not
known in a closed form. But numerical approximations
can be easily constructed using the procedure described
in [14]. Furthermore, whenever the first term in (16) is
small as it is in the case of proteins, an excellent approx-
imation [15] is obtained from the naive discretization of
the continuum soliton (12),
µ1 · eσ1(i−s) − µ2 · e−σ2(i−s)
eσ1(i−s) + e−σ2(i−s)
κi =
(20)
Here s is a parameter that determines the backbone site
location of the center of the fundamental loop that is de-
scribed by the soliton. The µ1,2 ∈ [0, π] are parameters,
their values are entirely determined by the adjacent he-
lices and strands: Away from the soliton center we have
(cid:26) µ1
−µ2
κi →
6
and for α-helices and β-strands the µ1,2 values are deter-
mined by (17), (18). We remind that negative values of
κi are related to the positive values by (9). Note that for
µ1 = µ2 and σ1 = σ2 we recover the hyperbolic tangent.
In this case the two regular secondary structures before
and after the loop are the same. Moreover, only the (pos-
itive) σ1 and σ2 are intrinsically loop specific parameters,
they specify the length of the loop and as in the case of
the µ1,2, they are combinations of the parameters in (13).
Similarly, in the case of the torsion angle there is only
one loop specific parameter in (14): The overall, common
scale of the four parameters is irrelevant in (14), and two
of the remaining three parameters characterize the reg-
ular secondary structures that are adjacent to the loop,
as in (17), (18).
Entire protein loops can be constructed by combining
together solitons (19), (14).
In [22] it has been shown
using the Ansatz (20) that over 92% of crystallographic
PDB configurations can be constructed in terms of 200
explicit soliton profiles. The solitons of the DNLS equa-
tion can thus be interpreted as the modular building
blocks of folded proteins.
B: Side-chain energy
We proceed to extend the energy function (13) so that
it models the deviations from the paradigm tetrahedral
symmetry around the Cα atoms: The Figures 1a) and 1b)
reveal that the directions of the backbone N and C atoms
oscillate in the latitudinal (θ) and longitudinal (ϕ) direc-
tions respectively, on the surface of the sphere that sur-
rounds the corresponding Cα atom. The covalent bond
structure that forms the sp3 tetrahedron of the Cα atom
combines these two oscillations into the annulus (horse-
shoe) shaped Cβ nutation of Figure 2a). Consequently
the natural dynamical variable that describes the nuta-
tion of Cβ on the surface of the sphere is the canonically
parametrized three component unit vector
sin θ · cos ϕ
sin θ · sin ϕ
cos θ
u =
(21)
In order to account for the Cβ nutation contribution to
the protein free energy, we then need to augment (13) by
terms that engage the additional variables (θi, ϕi).
The latitude angle θi is counted from the direction of
the corresponding Frenet frame tangent vector ti. Con-
sequently it remains invariant under the rotations of the
local Frenet frames around the direction of ti [12]. Thus
it can only couple to other frame rotation invariant quan-
tities. There are two natural terms,
ti × ui = sin θi
i > s
i < s
and
ti · ui = cos θi
N(cid:88)
In the leading order we only account for local interac-
tions. When we also demand invariance under the Z2
gauge transformation (9) we conclude that to the leading
order the corresponding free energy contribution should
have the form
Eθ =
fi(κ2
i )ti × ui + gi(κ2
i ) ti · ui + . . .
(22)
i=1
According to Figure 2a) the range of variations in θi
are quite small and we estimate that the center of the
annulus-like region is near
< θ > ≈ 113.4o
We Taylor expand (21) around this value so that we have
to the leading order
(cid:27)
i − bθκ2
i θ2
κ2
i θi − aθθi +
cθ
2
θ2
i
+ . . . (23)
N(cid:88)
(cid:26) dθ
2
i=1
Eθ =
The ensuing equation of motion is
θi =
aθ + bθκ2
i
cθ + dθκ2
i
(24)
As in the case of (14) we conclude that the overall scale
of the parameters drops out and this leaves us with three
independent parameters. In the case of a short loop that
we can model in terms of a single soliton like (20), two
of the parameters become determined by the value of
θi in the regular secondary structures that are adjacent
to the loop. This leaves us with only one loop specific
parameter.
The longitude ϕi in (21) is measured from the direction
of the Frenet frame normal vector ni. Under the local
rotations of the Frenet frames
(cid:18)ni
(cid:19)
bi
(cid:18) cos ∆i
→
sin ∆i
− sin ∆i cos ∆i
(cid:19)(cid:18)ni
(cid:19)
bi
around the tangent vectors ti by an angle ∆i we then
have
ϕi → ϕi + ∆i
Thus we may couple ϕ to the torsion angle as follows,
i(cid:88)
ϕi +
τk
k=1
This combination is invariant under the local rotations of
the Frenet frame around ti. Since the τk depend on the
backbone angles according to (14), we can again Taylor
expand the ensuing energy contribution. From Figure
2a) we estimate that for the center of the annulus
< ϕ > ≈ 139.5(o)
7
(cid:27)
N(cid:88)
(cid:26) dϕ
2
i=1
Eϕ =
Following (23) we then Taylor expand the ϕ contribution
to free energy around this value to conclude that to the
leading order we have (in Frenet frames)
i − bϕκ2
i ϕi − aϕϕi +
κ2
i ϕ2
cϕ
2
ϕ2
i
+ . . .
(25)
The equation of motion has the same functional form
with (14), (24)
ϕi =
aϕ + bϕκ2
i
cϕ + dϕκ2
i
(26)
Again only three of the four parameters in ϕ are inde-
pendent, the overall scale drops out.
We confirm that the functional forms (23) and (25) are
in line with the annulus-like (horseshoe-like) form of the
Cβ nutation in Figure 2a). For this we stereographically
project the Cβ distribution in Figure 2a) onto the com-
plex plane. Despite the nonlinear nature of the standard
stereographic projection the annulus-like shape is more
or less retained. Let r be the approximate radius of a thin
annulus on the complex plane and let (θ0, ϕ0) be the lo-
cation of its center. In the limit where the corrections
to the round circular profile of the thin annulus become
small we can determine the approximate form of the Cβ
nutation region from
(tan θeiϕ − tan θ0eiϕ0)(tan θe−iϕ − tan θ0e−iϕ0) = r2
(27)
Note that this is invariant under local frame rotations.
We re-write θ in (24) as follows,
θ = θ0 +
1
c + dκ2
We substitute this into (27) and Taylor expand to find
that to leading order it makes sense to parametrize ϕ by
an expression of the functional form (26).
We make the following remark: When we combine
(13), (23) and (25) we arrive at the total energy
(cid:26)
N(cid:88)
2 κi+1κi +
i + q · (κ2
2κ2
i=1
i − bτ κ2
i τ 2
κ2
i τi − aτ τi +
i − bθκ2
i θ2
κ2
i θi − aθθi +
cθ
2
θ2
i
i − bϕκ2
i ϕi − aϕϕi +
κ2
i ϕ2
cϕ
2
ϕ2
i
+
i=1
E = − N−1(cid:88)
(cid:26) dτ
N(cid:88)
(cid:26) dθ
N(cid:88)
(cid:26) dϕ
N(cid:88)
i=1
i=1
+
2
2
+
2
i=1
(cid:27)
i − m2)2
(cid:27)
cτ
2
τ 2
i
(28)
(29)
(cid:27)
+ . . .
(30)
(cid:27)
(31)
We have already established that protein backbones can
be described in terms of soliton solutions to (28), (29).
According to (14), (24), (26) the presence of (30) and
(31) does not change the functional form of the effective
κi energy (15), (16), all three variables (τi, θi, ϕi) are
similarly slaved to the bond angles κi. In particular, from
Figure 2a) we conclude that the contribution of (30) and
(31) to the full energy must be minuscule: The range
of variations in the variables (θi, ϕi) is relatively small.
(This is not the case with τi, see for example Figure 4
below.) Thus the values of (30) and (31) show very little
variation, and in comparison to (29) these two terms can
be treated as if they were tiny perturbations.
Indeed, in the case of proteins the two terms (30) and
(31) make no contribution to the total energy that we
are able to observe. The variables (θi, ϕi) are entirely
slaved by the DNLS soliton profile of the backbone bond
angles κi. Since the direction of the vector (21) that
specifies the position of the Cβ carbon is slaved to κi,
the deviation from the ideal tetrahedral symmetry in the
Cα covalent bond geometry is determined by the local
secondary structure environment of the amino acid.
C: Comment on parameters
The energy function (28)-(31) introduces eleven essen-
tial parameters, when we account for the overall scales
in (14), (24), (26). According to [22], no more than 200
different parameter sets are needed to describe over 92%
of high resolution structures in PDB with a precision
of around 0.6 A in RMSD for the Cα. The solitons
are like modular components from which the folded pro-
teins are built. At the moment we do not have a method
to compute the parameters directly from the sequence.
However, even in its present form the approach can be
subjected to a stringent experimental scrutiny: A typical
super-secondary structure described by a soliton such as
a helix-loop-helix consists of around 15 amino acids. If
we assume that the bond lengths are fixed, this leaves us
with 60 unknown coordinates for the Cα and Cβ atoms.
Since there are only 11 essential parameters in (28)-(31),
we have a highly under-determined set of equations. Con-
sequently the model is predictive, a comparison with ex-
perimental structures is directly testing the physical prin-
ciples on which (28)-(31) is based, even though we are not
yet able to compute the parameters from the sequence.
IV: EXAMPLE: VILLIN HEADPIECE HP35
As an example we consider the chicken villin headpiece
subdomain HP35. We use the x-ray structure with PDB
code 1YRF. The HP35 is a naturally existing 35-residue
protein with three α-helices separated from each other
by two loops. It continues to be the subject of very ex-
8
tensive studies both experimentally [23]-[25] and in silico
[26]-[29], and [29] reports on a molecular dynamics con-
struction with overall backbone RMSD accuracy around
one Angstrom.
In Figure 4 we have the (κi, τi) spectrum that we com-
pute from the PDB data of 1YRF. In the Figure 4a) we
use the standard convention that bond angles take values
in the range [0, π]. In the Figure 4b) we have extended
the range to [−π, π]. This introduced the Z2 gauge trans-
formation structure (9). In Figure 4b) we have applied
the gauge transformation to disclose the solitons. We
clearly have two solitons with the DNLS profile (20), sep-
arated from each other by regions with κ ≈ ±1.57 and
τ ≈ 1 corresponding to the α-helix (17). Notice the irreg-
(Color online) The profile ofκi(light blue) andτi
FIG. 4:
(dark red)along the 1YRF background. We use PDB indexing
of the sites. In a) the κi are restricted to [0, π] and in b) this
region is extended to [−π, π] using (9).
ular structure of the torsion angle τi in the loop (soliton)
regions. A priori we expect from (14) that the torsion an-
gle should have a regular profile. However, the numerical
values that we compute from (14) are not restricted to
the fundamental range τi ∈ [−π, π], they can take values
beyond this range. The irregular structure of τi follows
when we convert the values to the fundamental range,
using 2π periodicity of τi in the discrete Frenet equa-
tion (8). Similarly we observe slight irregularity in the
κi profile. This can also be removed if we allow κi to take
values beyond [−π, π] and use the 2π periodicity. But in
the case of 1YRF the improvement in the precision turns
out to be very small, and consequently we search for a
solution of (19) by assuming that κi ∈ [−π, π].
In Figure 5 we show the distribution of the side-chain
angles (θi, ϕi) in YRF, by plotting the tips of the unit
vector (21) on the two-sphere of Figure 2. As expected,
they are located in the α-helix region of Figure 2a) except
along the loops, where they are located outside of the
regular structure regions.
9
FIG. 5:
(Color online) The directional distribution of the
side-chain angles (θi, ϕi). The background coincides with the
annulus in Figure 2a).
We start by solving the classical equations of motion
for κi from (19). We then construct the remaining vari-
ables (τi, θi, ϕi) in terms of κi using (14), (24) and (26);
Since the (θi, ϕi) contributions to the κi potential (16)
are minuscule, we ignore the corresponding parameters in
constructing the solution to the DNLS equation for κi.
We use the iterative algorithm and procedure described
in [30], [14], and our results are summarized in Figure
6 and Table 1. We have been able to substantially im-
prove the accuracy reported in [14], in particular for the
first soliton. We now reach a RMSD accuracy less than
0.4 A even when we include the side-chain Cβ atoms.
The result is clearly within the Debye-Waller fluctuation
distance regime that we compute from the experimental
B-factors in the PDB data.
In Figure 6a) we display the distance between the com-
puted and the experimentally measured Cα atoms (ex-
cluding the N and C terminals). The shaded region in
Figure 6a) describes the 0.15 A zero point fluctuations
[22] around our solitons. For comparison, we also dis-
play the experimental Debye-Waller B-factor fluctuation
distances, obtained from the PDB data. Except for the
end point of soliton 1 (residue 58), our soliton solutions
FIG. 6:
(Color online) Comparison between our soliton solu-
tions in red (gray) and the experimental B-factor fluctuation
distance of PDB data for 1YRF (black) along the backbone.
In Figure a) for the Cα, and in Figure b) for the Cβ where the
experimental accuracy is estimated from (32). The shaded re-
gion describes the 0.15 A zero point fluctuations around soli-
tons. The cut in Figure a) at sites 13-14 is where the two
solitons overlap (Phe-58 in PDB), and the empty space in
Figure b) is due to glycine that has no Cβ.
describe the backbone well within the limits of experi-
mental accuracy.
In Figure 6b) we present our results for the Cβ nu-
tation, in comparison with the experimental data. We
also present an estimate for the experimental uncertain-
ties that we estimate as follows: The experimental B-
factors give an estimate for the absolute fluctuation dis-
tance around the measured position. But now we are
interested in estimating the (much smaller) relative error
in the position of Cβ with respect to the position of the
ensuing Cα. For this we introduce the relative B-factor
Brel = Bα − Bβ
(32)
In Figure 6b) we display the ensuing fluctuation distances
that we have computed from the Debye-Waller relation
using (32) in lieu of the B-factor. The precision of our
computed results compare well with these experimental
relative B-factor errors: For most of the sites the differ-
ence is no more than the 0.15 A estimate for zero point
fluctuations.
parameter
q1
q2
m1
m2
aτ
bτ
cτ
dτ
aθ
bθ
dθ
aϕ
bϕ
dϕ
soliton-1
0.459712
4.5533320
1.504535
1.512836
soliton-2
0.995867
9.408796
1.550322
1.535081
9.5752137e-9 7.840467e-6
-676965e-11 -4.973244e-9
4.875744e-9 4.2733696e-6
-2.917129e-9 -2.431388e-6
1.514770
-0.0017952
0.0420877
0.544859
5.66111e-5
-0.1845828
1.322495
-0.018619
6.930946e-8
0.3594184
3.83253e-4
-0.226012
RMSD ( A)
0.38
0.32
TABLE I: Parameter values for the two-soliton solution that
describes the two loops of 1YRF with a combined 0.39 A accu-
racy for both Cα and Cβ atoms. The displayed RMSD values
are for the individual solitons. The soliton-1 is located at
Glu-45 - Phe-58 and the soliton-2 is located at Phe-58 - Lys-
73. We utilize scale invariance to set all cθ = cϕ = 1. The
result has sensitivity to the accuracy of parameters, because
a folded protein is a piecewise linear polygonal chain with a
positive Liapunov exponent.
Finally, Figure 7 shows our soliton solution together
with the 1YRF configuration in PDB.
SUMMARY
In conclusion, the paradigm assumption that the tetra-
hedral covalent symmetry around the backbone Cα car-
bons is transferable, is correct to a good precision. How-
ever, with the advent of third and fourth generation X-
ray sources there is now a rapid growth in the number of
protein structures with sub-Angstrom resolution. This
makes it possible to scrutinize small corrections to this
paradigm. We have found, that the backbone N −Cα−C
bond angle shows systematic deviations from the ideal
value, in a manner that is in direct correspondence with
the corresponding secondary structure environment. We
have investigated how this effect propagates to the orien-
tation of the Cβ carbon. We have found that the angu-
lar orientations of the Cβ carbon similarly deviate from
their ideal values, in a manner which is in a one-to-one
correspondence with the underlying secondary structure
environment.
We have presented a simple energy function that is
based on the concept of universality, to model the sec-
ondary structure dependence in the Cβ orientations. As
an example, we have constructed the Cα-Cβ backbone of
HP35 villin, where we reach an accuracy that matches the
experimental B-factor fluctuation distances. We propose
that our observations and theoretical proposals could
10
FIG. 7: (Color online) A cartoon comparison of HP35 with
our soliton solution summarized in Table 1. The combined
Cα and Cβ root-mean-square distance is 0.39 A which equals
the experimental Debye-Waller B-factor fluctuation distance
for the central carbons.
form a basis for the development of both more accu-
rate refinement tools for experimental data analysis, and
of more precise theoretical and computational MD force
fields, to model the atomic level structure and dynamics
of folded proteins.
∗ Electronic address: [email protected]
† Electronic address: [email protected]
[1] I.W. Davis et.al., Nucl. Acids Res. 35 235 (2007)
[2] R.A. Laskowski et.al., J. Biomol. NMR 8, 477 (1996)
[3] X. Qu, R. Swanson, R. Day, J. Tsai, Curr. Protein Pept.
Sci. 10, 270 (2009)
[4] P.L. Freddolino, C.B. Harrison, Y. Liu, Y. Schulten, Na-
ture Phys. 6, 751 (2010)
[5] H.M. Berman et.al., Nucl. Acids Res. 28, W375 (2007)
[6] L. Schafer, M. Cao, Journ. Mol. Struc. 333, 201 (1995)
[7] P.A. Karplus, Prot. Sci. 5, 1406 (1996)
[8] D.S. Berkholz, M.V. Shapovalov, R.L. Dunbrack Jr., P.
A. Karplus, Structure 17, 1316 (2009)
[9] W.G. Touw, G. Vriend, Acta Cryst. D66, 1341 (2010)
[10] P.G. Kevrekidis, The Discrete Nonlinear Schrodinger
Equation: Mathematical Analysis, Numerical Computa-
tions and Physical Perspectives (Springer-Verlag, Berlin,
2009)
[11] A.S. Davydov, Journ. Theor. Biol. 66, 379 (1977)
11
[12] S. Hu, M. Lundgren, A.J. Niemi, Phys. Rev. E83, 061908
031906 (2012)
(2011)
[23] J. Meng, D. Vardar, Y. Wang, H.C. Guo, J.F. Head, C.J.
[13] M. Chernodub, S. Hu, A.J. Niemi, Phys. Rev. E82,
McKnight, Biochemistry 44, 11963 (2005)
011916 (2010)
[14] N. Molkenthin, S. Hu, A.J. Niemi, , Phys. Rev. Lett. 106,
078102 (2011)
[24] T.K. Chiu, J. Kubelka, R. Herbst-Irmer, W.A. Eaton,
J. Hofrichter, D.R. Davies, Proc. Natl. Acad. Sci. U.S.A
102, 7517 (2005)
[15] S. Hu, A. Krokhotin, A.J. Niemi, X. Peng, Phys. Rev.
[25] L. Wickstrom, Y. Bi, V. Hornak, D.P Raleigh, C. Sim-
E83, 041907 (2011)
[16] B. Widom, J. Chem. Phys. 43, 3892 (1965).
[17] L.P. Kadanoff, Physics 2, 263 (1966).
[18] K.G. Wilson, Phys. Rev. B4, 3174 (1971).
[19] M.E. Fisher, Rev. Mod. Phys. 46, 597 (1974).
[20] A.J. Niemi, Phys. Rev. D67, 106004 (2003)
[21] U.H. Danielsson, M. Lundgren, A.J. Niemi, Phys. Rev.
E82, 021910 (2010)
[22] A. Krokhotin, A.J. Niemi, X.Peng, Phys. Rev. E85,
merling, Biochemistry 46, 3624 (2007)
[26] D.L. Ensign, P.M. Kasson, V.S. Pande, J. Mol. Biol.
374, 806 (2007)
[27] H. Lei, Y. Duan, J. Mol. Biol. 370, 196 (2007)
[28] P.L. Freddolino, K. Schulten, Biophys. Journ. 97, 2338
(2009)
[29] D.E. Shaw et.al., Science 330, 341 (2010)
[30] M. Herrmann, Applic. Anal. 89, 1591 (2010)
|
1904.03925 | 2 | 1904 | 2019-10-14T20:07:33 | Unfolding force definition and the unified model for the mean unfolding force dependence on the loading rate | [
"physics.bio-ph",
"cond-mat.mes-hall",
"cond-mat.soft",
"cond-mat.stat-mech",
"physics.chem-ph"
] | In single-molecule force spectroscopy experiments, the dependence of the mean unfolding force on the loading rate is used for obtaining information about the energetic and dynamic properties of the system under study. However, it is crucial to understand that different dynamic force spectroscopy (DFS) models are applicable in different regimes, and that different definitions of the unfolding force might be used in those models. Here, for the first time, we discuss three definitions of the unfolding force. We carried out Brownian dynamics simulations in order to demonstrate the difference between these definitions and compare DFS models. Importantly, we derive the dependence of the mean unfolding force for the whole range of the loading rates by unifying three previously reported DFS models. Among the currently available models, this unified model shows the best agreement with the simulated data. | physics.bio-ph | physics | Unfolding force definition and the unified model for the mean
unfolding force dependence on the loading rate
Department of Physics, University of Alberta, Edmonton AB, T6G 2E1, Canada
Rafayel Petrosyan*
ABSTRACT
In single-molecule force spectroscopy experiments, the dependence of the mean unfolding force
on the loading rate is used for obtaining information about the energetic and dynamic properties
of the system under study. However, it is crucial to understand that different dynamic force
spectroscopy (DFS) models are applicable in different regimes, and that different definitions of
the unfolding force might be used in those models. Here, for the first time, we discuss three
definitions of the unfolding force. We carried out Brownian dynamics simulations in order to
demonstrate the difference between these definitions and compare DFS models. Importantly, we
derive the dependence of the mean unfolding force for the whole range of the loading rates by
unifying three previously reported DFS models. Among the currently available models, this
unified model shows the best agreement with the simulated data.
*
[email protected]
https://orcid.org/0000-0003-1888-0379
1
INTRODUCTION
In atomic force microscopy (AFM) based dynamic force spectroscopy (DFS)
experiments, the unfolding forces are being measured for a wide range of the loading rates.
Afterwards, the average unfolding force dependence on the loading rate from the appropriate
DFS model is fitted to the corresponding data in order to gain information about the kinetic and
energetic properties of the system[1 -- 4]. For instance, the system could be a protein or a ligand-
receptor bond, but the theory is general and not limited to these examples. Hereafter, we will be
using the proteins' terminology, such as the native and unfolded states.
The first model that was proposed for analyzing DFS data was the phenomenological
model widely referred to as the Bell-Evans model. It assumes first-order irreversible rate
equation for the survival probability[5]. In this model, empirically determined, exponentially
force-dependent unfolding rate[6 -- 8] is used. Evans and Ritchie derived the probability density
function (PDF) of the unfolding force, and they also obtained the most probable unfolding force
depending on the loading rate and the two parameters: the unfolding rate and the distance
between the native and the transition states. The dependence of the mean unfolding force on the
loading rate for this model Eq. (1) was obtained later[9]:
(1)
where
is the mean unfolding force,
is the Boltzmann constant,
is the absolute
temperature,
is the distance between the native state and the transition state,
is the
unfolding rate in the absence of the external force,
is the loading rate, and
is
the exponential integral[10,11].
Here we would like to emphasize that the determination of the mean unfolding force from
the experimentally obtained unfolding force distribution is less error-prone and more
straightforward compared to the determination of the most probable unfolding force. Typically,
the most probable unfolding force is determined by means of fitting the Gaussian to the
unfolding force distribution[12,13] and then taking the maximum of that fit as the most probable
2
‡.B01‡B‡.B0xFTkkEexTkFxFTkkFBkT‡x0kdtdFF.1E
unfolding force, while neglecting the fact that generally the unfolding force distributions are
asymmetric[14 -- 18].
The biomolecular (un)folding could be viewed as diffusion in the multidimensional
potential energy surface. Each configuration of the biomolecule will be represented by a point in
its degree of freedom space. For each configuration, there is a corresponding potential energy,
consequently, the potential energy axis could be added as an additional dimension to the degree
of freedom space. Once the potential energy axis is added as described, in this newly created
space of degrees of freedom and the potential energy, we will have the above-mentioned
potential energy surface. In this framework, the continuous changes of the biomolecular
configuration due to the thermal motion could be viewed as a Brownian motion of the particle in
the degree of freedom space, in the presence of the force: the gradient of the potential energy.
Under certain assumptions, this multidimensional potential energy surface could be coarse-
grained into a one-dimensional free energy profile[19]. The use of Brownian dynamics (BD)
simulations for mimicking dynamics of biomolecules is originates from this framework.
Kramers assumed a simple scenario with a smooth energy profile which had a well and a
barrier. The particle is initially at the well (protein is folded at its native state), then due to the
Brownian motion the particle crosses the barrier (protein unfolds). Kramers addressed the
determination of the rate of this reaction[20]. He managed to solve the problem for large and
small limiting cases of the viscosity (generally, for biomolecular systems the limit of large
viscosity is relevant) assuming that the barrier height is much larger than the thermal fluctuation
energy (kBT).
Kramers' result was used for the development of the next widely used DFS model, i.e.,
the profile assumption model commonly referred to as Dudko-Hummer-Szabo model[15]. Here
two free energy profiles have been assumed: parabolic-cusp[14] and linear-cubic. The force-
dependent unfolding rate was calculated based on Kramers' result[20,21]. Assuming first-order
irreversible rate equation for the survival probability, the approximate unfolding force PDF and
the approximate mean unfolding force Eq. (2) (that was further improved[22] Eq. (3)) were
derived for this model. These were derived depending on the loading rate and three main
parameters: the unfolding rate, the distance between the native state and the transition state, and
the free energy difference between the transition state and the native state. One of the important
achievements of this study was that the authors managed to combine the results of two different
3
free energy profiles into a single expression of the unfolding force PDF (and hence its moments).
As a result of this, the fourth parameter that specifies the free energy profile is added. It is equal
to 1/2 for harmonic-cusp free energy profile and 2/3 for liner-cubic free energy profile. When
that parameter takes value 1, the phenomenological model is recovered. The profile assumption
model is applicable for the loading rates small enough to allow the unfolding to occur before
Kramers' high energy barrier approximation breaks down[20,21].
(2)
(3)
In Eq. (2) and Eq. (3),
is the free energy difference between the transition state and the
native state,
is the free energy profile defining factor, and
is the Euler-Masceroni
constant.
The profile assumption model was further generalized for the case of a larger number of
underlying free energy profiles[23]. In this model, the fourth parameter
(similar to
in the
previous model[15]) determines the free energy profile, and it assumes any value between 0 and
1. It gives the same results as the previous model[15] for values 1/2, 2/3, and 1. The mean
unfolding force dependence on the loading rate for this model is given by Eq. (4):
(4)
where
.
The model that describes relatively well the unfolding force distributions for the low and
high loading rates is commonly referred to as Bullerjan-Sturm-Kroy model. Here we will refer to
it as a rapid model[18]. Authors managed to calculate the PDF for the unfolding forces and the
4
‡.B0‡B‡‡B‡ln1xFTekkGTkxGFTkG‡.B01‡B‡‡‡.B011xFTkkEeGTkxGFxFTkk‡G5772.0.‡32B‡B0B‡FxeTkGTkkTkGlnlnln231ln1‡B‡‡GTkxGF
mean unfolding force dependence on the loading rate Eq. (5). In this model, the mean unfolding
force is simply the sum of the mean unfolding force predicted by the phenomenological model
and the mean unfolding force for the limit of high loading rates when the unfolding is ballistic
and follows the
law[14].
(5)
In Eq. (5),
quantifies the friction. This friction could be the combination of the solvent
viscosity, internal friction due to interatomic interactions, and other factors that are damping the
fluctuations of the molecule[24]. This friction is related to the diffusion coefficient as
. This model assumes harmonic-cusp free energy profile. Using Kramers' high
friction approximation, the intrinsic unfolding rate is expressed through the other parameters of
the system in the following equation:
[18].
For sufficiently low loading rates, in the close to equilibrium regime, before the final
unfolding, the system might hop back and forth between various states[25 -- 30]. All the above-
discussed models neglect the possibility of refolding. If hopping is observed they only use the
first unfolding force. However, there is some information on the equilibrium properties of the
system stored in the data which is obtained from close to equilibrium measurements. This can
give a more complete picture of the system under study.
The model that considers also the refolding rate in the rate equation for the survival
probability is frequently referred to as Friddle-Noy-De Yoreo model. We will refer to it as a
reversible model[31 -- 33]. As the phenomenological model[5], this model assumes exponentially
force-dependent rates[6 -- 8], and it considers that the force is applied by means of the Hookean
spring (henceforth, spring). For this model, the dependence of the mean unfolding force on the
loading rate and on three parameters was obtained Eq. (6)[31 -- 33]. The parameters are the
following: the unfolding rate, the distance between the native state and the transition state, and
the free energy difference between the unfolded state and native state.
5
F‡.‡.B01‡B2‡.B0xFxFTkkEexTkFxFTkkDTkBTkGeTkGxGkB‡B‡2‡‡02
(6)
In Eq. (6),
is the equilibrium force which is close to the force at which unfolding and
refolding rates are having the same value (assuming that in the absence of force the unfolding
rate is lower than the refolding rate). In this model,
, where
is the spring
constant,
is the free energy difference between the unfolded state and the native state, and
. The unfolding force PDF for this model has not been reported yet. As
emphasized in[33], the unfolding force here should be calculated based on the total time
(residence time or occupation time) spent in the native state's well before the final unfolding:
loading rate multiplied by the residence time. In other words, the unfolding force in this model is
the total force applied on the native state while the system is in that state's well.
Some of the above-discussed models, which take only the first unfolding force, have been
extended for the case when the force is applied through the spring. For the phenomenological
model, in the case when the force is applied through the spring, the expression of the mean
unfolding force Eq. (7) could be obtained using the following rate expression
.
(7)
In the case when force is applied through the spring, the dependence of the mean
unfolding force on the loading rate for the profile assumption model is given in Eq. (8)[4,34]:
(8)
6
‡.Beq1‡Beq‡.BeqxFTkFkEexTkFFxFTkFkeqFGF2eqGTkxxFekFkB2‡‡eq20eqTkxxekFkB2‡‡2F0‡.2B01‡BB2‡‡.B22‡B0xFTekkEexTkFTkxxFTekkTkx3B‡31B‡‡.B01‡.B013‡B‡‡11ZTkGexFZTkkexFTkkEeZGTkxZGFZTkG
and
(9)
where
is the curvature of the molecular free energy profile at the native state as defined
in Eq. (9).
is the the free energy profile[15].
For the rapid model, when force is applied through the spring, Eq. (10) is reported in the
original study[18]:
(10)
This model assumes a harmonic-cusp free energy profile, hence, using Eq. (9) we obtain
.
The main aim of this study is to show the importance of the definition of the unfolding
force and, in the context of various DFS models, to compare the predictions of the dependence of
the mean unfolding force on the loading rate. Importantly, we provide a new model (unified
model) based on the profile assumption[15,22] model, the reversible[32,33] model, and the
rapid[14,18] model. The new unified model is applicable for the whole range of the loading rates
from near-equilibrium to ballistic unfolding.
METHODS
Brownian dynamics simulations
Two sets of BD simulations were carried out. In the first set, the quartic free energy
profile was ramped with constant loading rates ranging from 10 pN/s to 1012 pN/s to the final
force, which was 15 pN for low loading rates (Eq. (11)). In the second set of simulations, the
7
min01xGZminmin0202xxxGxxGmin0xG‡11‡0‡‡1xxvGxxvvGxGv‡.‡.1B0un1‡B2B‡‡.1B‡B0unxFxFTekkZEeZxTkFZTkGxFTekkZZTkG‡2‡21GxZ
Morse free energy profile was pulled through a spring with a spring constant of 100 pN/nm with
a constant velocity ranging from 1000 nm/s to 1014 nm/s to the final force, which was 150 pN for
the low loading rates (Eq. (12)). The friction coefficient was chosen to be 10-5 pN s/nm[35] in
both sets of simulations, and the overdamped Langevin equation (Eq. (13)) was solved
numerically using Euler -- Maruyama method.
(11)
(12)
In Eq. (11) and Eq. (12),
is the time and
is the speed with which spring is pulled.
(13)
In Eq. (13),
is a delta-correlated Gaussian white noise with 0 mean. The simulation time
step was chosen to be 10-9 s for the quartic free energy profile and 10-12 s for the Morse free
energy profile, securing less than 10 % of error due to finite time step[36]. For high loading
rates, where the simulation time becomes quite short, the time steps were decreased even more to
ensure that at least half a million steps were in a trajectory (based on random walk as a limit of
Brownian motion, Donker's theorem). For each loading rate, 3000 trajectories were generated.
For each trajectory, the initial value for the reaction coordinate was chosen randomly from the
Boltzmann distribution for the corresponding free energy profile's native state neighborhood.
The first, reversible, and last unfolding forces have been recorded for each trajectory and the
mean unfolding forces were calculated from that data.
Determination of the parameters
To compare how well the models predict the outcome of BD simulations, was necessary
to use the same parameters in the models that were used in the BD simulation. For the BD
8
txFxxxtxG.24455100,2211.021.0241141,xVtetxGxtVtRTkxtxGtxB.2,tR
simulations with the quartic free energy profile, most of the parameters could be determined
straightforwardly. The determination of the more ―complicated‖ parameters is described here.
The energy profile parameter (μ) for the profile assumption (2016) model[23] was determined by
fitting the free energy profile model proposed in[23] to the part of the quartic free energy profile.
The unfolding rate was determined using Eq. (14), Fig. 1[21,37 -- 39].
(14)
Figure. 1. The unfolding rate can be calculated using Eq. (14) where the integration limits are shown in this figure.
The refolding rate can be calculated analogously.
In Eq. (14),
is the free energy profile and a and b are reflecting and absorbing boundaries,
respectively. The refolding rate was calculated using a similar equation. The equilibrium force
was determined by assuming the force dependence of the free energy profile in Eq. (14) and
numerically solving the equation of the unfolding and refolding rates in order to find the force at
which they are equal.
In the case of the Morse free energy profile, the determination of the parameters is less
trivial since this profile does not have any barrier in the absence of the external force. In order to
find the distance between the native and transition states for this model, we choose as a barrier
position 0.262 nm, which is the mean of the barrier positions when it first appears at 0.135 nm
and when it disappears at 0.389 nm. Since the native state is at 0.1 nm, then
≈ 0.162 nm. The
9
bxyaTkzGTkyGdzdyeeTkkminBB1BG‡x
unfolded state distance was chosen at 0.6 nm, which is where the second minimum becomes
noticeable. Therefore, for this case the first unfolding force was taken as the product of the time
at which system visited the unfolded state (0.6 nm) for the first time and the loading rate, the last
unfolding force was taken as the product of the time at which the system was in the native state
(0.1 nm) for the last time and the loading rate, and the reversible unfolding force was taken as the
product of the loading rate and the time the system spent in the native state's well (time spent for
x < 0.262 nm). To determine the distance between the unfolded state and the transition state
and the free energy difference between the transition state and the unfolded state
, for the
Morse free energy profile, we followed the approach described in[40]. To determine the
unfolding and refolding rates we used Eq. (14) and Fig. 1. The equilibrium force was determined
with the same method as in the case of the quartic free energy profile. The determination of the
parameter values for the Morse free energy profile is challenging since some parameters are
virtual. Our choice of the Morse free energy profile with a spring for this study is due to the
historical reasons[32]. We could have chosen a different path for determining the parameters of
the Morse free energy profile, particularly by evaluating the first and third derivatives at the
inflection point as described in[15]. However, parameters determined with this approach might
put profile assumption models and related models in a favorable position. Ideally
should
be used to calculate the loading rate[4,34]. However, we have used
since in our case
which means that the soft spring approximation is applicable.
R2 and χ2 Criterions
As a goodness of prediction criterion, we used R2 as defined by Eq. (15).
(15)
In Eq. (15),
is the observed value of the ith data point, i.e. the value we get from the data,
is
the expected value for the ith data point, i.e. the value predicted by the model, and
is the
10
‡refx‡refGZVV00305.1ZniiniiiOOEOR121221iOiEO
mean of the data (see Fig. 2a). Clearly, the closer the R2 value is to 1, the better the model
predicts the data. A model with more parameters has an advantage when models are being fitted.
In such a case the adjusted R2 might be used, which is defined by Eq. (16).
(16)
In Eq. (16),
is defined by Eq. (15),
is the sample size (number of data points), and
is the
number of parameters in the model.
Another criterion used in this study for the goodness of the prediction is χ2 defined by Eq.
(17).
(17)
In Eq. (17),
is the standard deviation for each data point and the summation goes through all
the data points from first to the last (nth) (see Fig. 2b). It is important to note that instead of the
standard deviation, the standard error or other types of errors can also be used to define χ2.
However, since all our data points are obtained from the averaging of the same number of
unfolding forces, we use the standard deviation. Clearly, the closer χ2 is to 0, the better the model
predicts the data. Similar to adjusted R2, there is also a reduced χ2 that takes into account the
degrees of freedom of the model, as defined by Eq. (18).
(18)
In Eq. (18),
is given by Eq. (17). In our analyses we have not used the adjusted R2 or reduced
χ2 since we did not fit models to the data, i.e., we did not change the parameters of the models to
optimize R2 or χ2. By using the values in Table 1 and Table 2 and by determining the number of
the data points for each range from Fig. 5 and Fig. 6, the adjusted R2 and reduced χ2 can easily be
calculated for all models discussed in this work using Eq. (16) and Eq. (18).
11
111122pnnRR2RnpniiiiEO1222i122pn2
Figure. 2. Visualization of the goodness of prediction criteria R2 Eq. (15) (a) and χ2 Eq. (17) (b) used in this study.
Error bars in (b) show the standard deviations. The closer the R2 is to 1, the better the prediction. The closer the χ2
is to 0, the better the prediction.
RESULTS
Definition of the unfolding force
The determination of the unfolding force for the case of a single transition (Fig. 3) is
straightforward. However, it is unclear what force should be taken as an unfolding force if the
system hops back and forth between various states before the final unfolding (Fig. 4). For such
situations we define three unfolding forces. The first unfolding force is the product of the loading
rate and the time when the system visits the unfolded state for the first time. The last unfolding
force is the product of the loading rate and the time when the system was in the native state for
the last time. The reversible unfolding force is the product of the loading rate and the time that
the system spent in the native state's well (see yellow stripes in Fig. 4).
12
Figure. 3. An example of a trajectory from BD simulation with a single transition. The quartic free energy profile,
which is under constantly increasing external force (with 1000 pN/s loading rate), was used. There is a single
transition here, and, hence, no ambiguity in the determination of the unfolding force. The graphs of the free energy
profile on the top are given at different times (0.005 seconds apart) showing that the first minimum (blue dot) and
the maximum (green dot) are getting closer with time and eventually merge (cyan dot). The blue, green, and red
lines are showing the positions of the first minimum, maximum, and the second minimum respectively as functions of
time.
13
Figure. 4. An example of a trajectory from BD simulation with multiple transitions between the native and unfolded
states. The quartic free energy profile that is under constantly increasing external force (with 100 pN/s loading rate)
was used. We define three unfolding forces. The first unfolding force is the product of the loading rate and the time
when system visits the unfolded state (x = 3.25 nm) for the first time (red dot). The last unfolding force is the product
of the loading rate and the time when the system was in the native state (x = -3.76 nm) for the last time (blue dot).
The reversible unfolding force is the product of the loading rate and the time that the system spent in the native
state's well (the time spent below the green line x < 0.51 nm, yellow strip). The blue, green, and red lines are
showing the positions of the native, transition, and unfolded states respectively.
As can be seen from Fig. 3, the positions of the extrema of the free energy profile change
with time if the external, time-dependent force is applied. However, the native, transition, and
unfolded states are specified as particular values of the reaction coordinate and are independent
of the externally applied force. In the above-stated example the native, transition, and unfolded
states coincide in the absence of the external force with the first minimum, maximum, and the
second minimum of the free energy profile. The state is specified by a particular value (or range
of values) of the reaction coordinate, and it does not depend on the external force. What depends
on the external force is the likelihood of finding the system in a particular state.
14
In order to see how the means of these three unfolding forces depend on the loading rate,
we carry out BD simulations using quartic free energy profile under the external force that
increases with a constant loading rate. We generate 3000 trajectories for each loading rate, for
the loading rates that span 12 orders of magnitude (see Fig. 5). For further details, see the
methods section.
The surprising observation is that the mean last unfolding force shows non-monotonic
dependence on the loading rate. At sufficiently low loading rates, the mean last unfolding force
starts to increase with the decrease of the loading rate, a phenomenon that was not observed
previously. This means that the time when the system was in the native state for the last time
increases faster than the loading rate decreases, as a result the product of two increases. The
behavior of the mean last unfolding force at even lower loading rates will be discussed at the end
of this section (see Fig. 8). The mean reversible unfolding force at the low loading rate regime
with the decrease of the loading rate tends to the constant value of Feq. This means that the mean
time spent in the native state's well increases to the same extent as the loading rate decreases.
The decrease in the mean first unfolding force at the low loading rate regime is evident. As a
reminder, the first unfolding force is the time when the first unfolding occurred multiplied by the
loading rate. In the absence of the external force, there is a constant mean time (mean first
passage time) when the system visits the unfolded state for the first time. In the presence of the
slowly changing external force, the time when the first unfolding occurred tends to the mean first
passage time. Hence, one of the two terms in the product tends to the constant while the other
one decreases. This results in the decrease of the first unfolding force with the decreasing loading
rate at the near-equilibrium regime.
As can be seen in Fig. 5, three differently defined mean unfolding forces converge with
the increase of the loading rate. This is the result of moving away from the equilibrium and
entering the regime where there are mostly single transitions in trajectories. In the case of a
single transition, the difference between three unfolding forces is negligible as long as the time
spent during the transition (transition path time) is negligible compared to the time spent before
the transition (residence time of the native state). This is true for low loading rates for which also
the hopping is common. This is not true for very high loading rates. For these rates only the
mean reversible unfolding force is shown as the vast majority of trajectories at high loading rates
15
have a single transition. For the single transition the determination of the unfolding force is
straightforward: it is the time spent in the native state's well multiplied by the loading rate.
The next point that can be noticed from Fig. 5b is the kink around 5 106 pN/s of the
loading rate value. After the critical value of the loading rate, the data points lie on a line with a
higher slope. This is the indication of the change in the regime of unfolding. At such high
loading rates, the unfolding mostly occurs after the disappearance of the barrier.
Hereby, we have specified three regimes: the low loading rate regime (close to
equilibrium), the intermediate loading rate regime (non-equilibrium), and the high loading rate
regime (non-equilibrium, ballistic unfolding). The smoothness of the kinks between these
regimes (Fig. 5b) is due to the fact that the mean unfolding force for the given loading rate
contains data from many trajectories. These trajectories can show multiple transitions between
states, or can have a single transition to the unfolded state before or after the disappearance of the
barrier.
A unified model for the dependence of the mean unfolding force on the loading rate
To accurately predict the behavior of the dependence of the mean unfolding force on the
loading rate in all the three regimes, we developed a unified model. We used the force-dependent
unfolding rate expression proposed in[15], and, following the method proposed in[22], we
calculated the mean first unfolding force for the low and intermediate loading rates. Next, we
extended that result for the case of high loading rates, when the unfolding is happening after the
disappearance of the barrier[14,18]. This results in Eq. (19), the derivations of which can be
found in Appendix A.
(19)
16
cFirst..‡.‡cFirst.‡‡.cFirst..‡B01‡B‡‡First2211.‡B0FFxFxFxGFFFxTkkEeGTkxGFFxTkk
In Eq. (19),
,
is not an independent parameter, and it can be
calculated using the Kramers' approximation if the free energy profile is known. For example,
for the linear-cubic profile (v = 2/3), we will have
in contrast to Eq. (5), where
the harmonic-cusp profile is assumed. Generally using the linear-cubic free energy profile is
more reasonable as discussed in[15,34], and we will assume that profile in the future, i.e. v = 2/3,
if not specified. Here
is the Heaviside step function:
. Hereby, the
dependence of the mean first unfolding force on the loading rate is determined by 3 independent
parameters:
,
, and
.
Following the same procedure, except this time calculating the mean differently[32], we
obtained the mean reversible unfolding force: Eq. (20) (see Appendix A for details).
(20)
In Eq. (20),
. Hereby, the dependence of the mean
reversible unfolding force on the loading rate is determined by 4 independent parameters:
,
,
, and
.
Notice that the critical loading rate depends on the definition of the unfolding force.
Generally, the mean reversible unfolding force riches the given force at the lower loading rate
than the mean first unfolding force since by definition for the same trajectory the first unfolding
17
TkGTkGexTekkFB‡B‡1‡B0cFirst.0kTkGexGkB‡2‡‡030,00,1xxx‡x‡GcRev..‡.‡cRev.‡‡.cRev.11.‡B01‡B1‡‡eq‡‡Rev22111‡‡eqB‡1‡‡eq11B‡.‡B0FFxFxFxGFFeFxTkkEeGTkGxFxGFGxFTkGeFxTkkGxFTkG1‡‡eqB‡B‡1‡B0cRev.1GxFTkGTkGexTekkF‡x‡GeqF
force is not higher than the reversible unfolding force. With the increase of the loading rate, the
number of trajectories with hopping decreases, and, due to this, the first and reversible mean
unfolding forces are getting closer, but eventually the mean reversible unfolding force riches the
critical force at a slightly smaller loading rate than the mean first unfolding force.
The mean last unfolding force can be determined based on the symmetry arguments. In
Fig. 4 we observe that
where
is the mean force due to the time spent
in the unfolded state's well (above 0.51 nm), before the last unfolding.
equals the time
spent in the unfolded state's well before the last unfolding multiplied by the loading rate.
Interestingly, at close to equilibrium regime, it is possible to calculate
from the relaxation
protocol, namely,
where
is the man first refolding force, and
is the mean reversible refolding force. Hence, the mean last unfolding force can be
calculated through Eq. (21):
(21)
At the non-equilibrium regime, the symmetry will break, but the Eq. (21) will still hold since in
the case of a single transition the last two terms will cancel each other.
and
can be calculated following the same path as for
and
only using the force
dependent refolding rate as in[41] and different integration limits (see Appendix A for details).
This results in Eq. (22) and Eq. (23):
(22)
18
xRevLastFFFxFxFxFRevRefFirstRefxFFFFirstRefFRevRefFRevRefFirstRefRevLastFFFFFirstRefFRevRefFFirstFRevFcFirstRef..‡ref.‡refcFirstRef.‡ref‡ref.cFirstRef.11.‡refB0ref1‡refB1‡ref‡reffin‡ref‡refFirstRef22111‡ref‡reffinB‡ref1‡ref‡reffin11B‡ref.‡refB0refFFxFxFxGFFeFxTkkEeGTkGxFxGFGxFTkGeFxTkkGxFTkG
In Eq. (22),
,
,
is the
distance between the unfolded state and the transition state,
is the refolding rate in the
absence of the external force,
is the free energy difference between the transition state and
the unfolded state, and
is the final force till which the system is pulled and from which the
(hypothetical) refolding protocol starts. The final force is set by the experimental design and is
not an intrinsic parameter describing the system.
(23)
In Eq. (23),
. In this case,
is not an
independent parameter since in Eq. (22) and Eq. (23) there are parameters involved which
describing the unfolded state's well. The equilibrium unfolding force can be calculated using the
probability
of
having
the
system
folded
at
a
given
force[33]:
. This
formula of
is model
independent in the sense that we do not assume force dependent unfolding and refolding rates.
By inserting Eq. (20), Eq. (22), and Eq. (23) into Eq. (21), the dependence of the mean last
unfolding force on the loading rate is determined by 5 independent parameters:
,
,
,
, and
. Independently of applying or not applying relaxation protocol, Eq. (21) (which
19
1‡ref‡reffinB‡refB‡ref1‡refB0refcFirstRef.1GxFTkGTkGexTekkFTkGexGkB‡ref2‡ref‡ref0ref3‡refx0refk‡refGfinFcRevRef..‡ref.‡refcRevRef.‡ref‡ref.cRevRef.11.‡refB0ref1‡refB1‡ref‡refeq‡ref‡refRevRef22111‡ref‡refeqB‡ref1‡ref‡refeq11B‡ref.‡refB0refFFxFxFxGFFeFxTkkEeGTkGxFxGFGxFTkGeFxTkkGxFTkG1‡ref‡refeqB‡refB‡ref1‡refB0refcRevRef.1GxFTkGTkGexTekkFeqFTkGGTkGGxxFexxTkeFB‡ref‡B‡ref‡‡ref‡1ln11‡ref‡B0eqeqF‡x‡G‡refx‡refG
includes Eq. (20), Eq. (22), and Eq. (23)) is applicable for predicting the dependence of the mean
last unfolding force on the loading rate.
Exponential integral and Heaviside step function can be approximated using other well-
known functions, for example,
and
where q is a large
positive number[11]. However, for maximum precision, the fitting function files provided in the
supplementary material exclude these approximations. These files can be directly used for data
analysis.
Comparing various models' predictions using simulated data
To show how well all above-described models predict the dependence of the mean
unfolding force on the loading rate, the models' predictions and the data from the BD
simulations are presented on the same graph (Fig. 5). The parameters in these models are the
same as in the BD simulations. To further quantify how well the models predicted the BD
simulated data, R2 and χ2 were calculated (see Table 1). For further details, see the methods
section.
We observe that the phenomenological model[5,9] fails at high and intermediate loading
rates. The main disadvantage of the profile assumption models[15,22,23] is that they fail at high
loading rates. The rapid model[18], which combines the result of the phenomenological model[9]
with the high loading rate result[14,18], arguably provides the best prediction among the
previously proposed models. The reason for this could be that the intermediate loading rate
regime range is small compared to the high loading rate regime, and more data is generated in
the high loading rate regime. The loading rate range for the BD simulations is chosen based on
the previous experimental and molecular dynamics simulations studies[13]. The unified first
model clearly gives the overall best prediction. According to unified last model, the mean last
unfolding force tends to the constant at very low loading rates as it will be shown later in Fig. 8.
However, conducting simulations at very low loading rates is challenging due to long simulation
times and large memory required to deal with simulated data. Hence, in the case of very low
loading rates, the prediction of the unified last model needs further verification.
20
xexEex1ln1qxex11
21
Figure. 5. Dependence of the mean unfolding force on the loading rate, from BD simulations with quartic free
energy profile, and from the predictions from the DFS models. (b) shows the log-log plot for the data in (a), but for
a larger range of loading rates. Three regimes can be observed from (b). In the low loading rate regime the three
differently defined mean unfolding forces diverge. This regime ends roughly where the mean last unfolding force has
a minimum, the mean reversible unfolding force rises from the constant Feq value, and the mean first unfolding force
follows a different slope, around 1000 pN/s. The intermediate loading rate regime starts where the low loading rate
regime ends, and ends roughly where the change in the slope can be noticed around 106 pN/s indicating that
unfolding occurs after the barrier has disappeared. The models are plotted using the parameters from the BD
simulation (hence these are not fits). Phenomenological (1997, 2000) corresponds to Eq. (1), profile assumption
(2006) corresponds to Eq. (2), profile assumption (2008) corresponds to Eq. (3), profile assumption (2016)
corresponds to Eq. (4), rapid (2014) corresponds to Eq. (5), unified first (2019) corresponds to Eq. (19), unified
reversible (2019) corresponds to Eq. (20), unified last (2019) corresponds to Eq. (21) where Eq. (20), Eq. (22), and
Eq. (23) were used. Table 1 shows how well these models predict the data from this figure. The parameters used in
the models are kBT = 4.1 pN nm,
= 4.27 nm, k0 = 9.33 s-1,
= 40.62 pN nm,
= 10-5 pN s/nm, Feq = 4.02
pN, Ffin = 15 pN, v = 0.66667, μ = 0.5872,
=2.74 nm, k0ref = 6495 s-1 , and
= 12.41 pN nm. The error
bars for the standard error of the mean are smaller than the symbols.
22
‡x‡G‡refx‡refG
Model
Low
loading
rates
Intermediate loading
rates (5000 pN/s
(
< 5000 pN/s)
High loading rates
All loading rates
(
> 106 pN/s)
106 pN/s)
R2
χ2
R2
χ 2
R 2
χ 2
R2
χ 2
Phenomenological
(1997)
Profile assumption
(2006)
Profile assumption
(2008)
Profile assumption
(2016)
0.9695
0.1809
0.0275
12.6468
-0.4443
1882.94
-0.1878
1895.77
0.8446
6.3365
1.0414-
-0.432+
-0.4441
1836.41+
-0.1876-
1842.31+
0.1402i
1.4711i
-0.0072i
88.0905i
0.0059i
89.5616i
0.9889
0.2245
1.0415-
-0.4327+
-0.4441
1836.41+
-0.1876-
1836.2+
0.1402i
1.4711i
-0.0072i
88.0905i
0.0059i
89.5616i
0.842
6.2111
0.883+
1.598-
-0.4446
1824.61+
-0.188-
1832.42+
0.028i
0.2846i
-0.0088i
107.478i
0.0073i
107.193i
Rapid (2014)
0.9968
0.0187
0.9727
0.3191
0.9997
4.145
0.9997
4.4828
Unified first (2019) 0.9987
0.0093
0.997
0.0693
0.9998
0.4946
0.9998
0.5732
Unified
(2019)
reversible
0.6972
0.5088
0.9977
0.0553
0.9998
0.4961
0.9998
1.0602
Unified last (2019)
0.5613
2.2316
0.9952
0.1886
0.9998
0.4974
0.9998
2.9176
Table 1. To quantify how well the models predict the simulated data we calculate R2 and χ2 for each model plotted
in Fig. 5 at three regimes separately, and then all together (last column). Some of the models predict complex values
for the mean unfolding force at high loading rates. For the completeness of the table, for these models, we also
present R2 and χ2 . The perfect prediction would result in R2 = 1 and χ2 = 0.
Unified model when the force is applied through the spring
We further extend the unified model for the case when the system is being pulled through
the spring. The dependence of the mean unfolding forces on the loading rate is derived using the
same methods as when the force is applied directly. The only difference here is that different
force dependent unfolding and refolding rates are used (see Appendix B for derivations)[34]. The
calculations result in Eq. (24), Eq. (25), Eq. (26), and Eq. (27) for the mean first unfolding force,
mean reversible unfolding force, mean first refolding force, and mean reversible refolding force
respectively. The last three are used for calculating the mean last unfolding force in Eq. (21).
23
.F.F.F
(24)
In Eq. (24),
, and Z is given by Eq. (9). For the linear-cubic
free energy profile
.
(25)
In Eq. (25),
.
could be calculated using Eq. (21) and Eq. (25), Eq. (26), and Eq. (27) for
,
and
respectively.
24
cFirst..‡.‡cFirst.‡‡.cFirst.1.‡B01‡B22‡‡First222B‡21B‡.‡B0FFxFxFxZGFFeZFxTkkEeGTkZZZxGFvvZTkGZTkGeZFxTkkvTkZGTkGveZxTekkFB2‡B‡1‡B0cFirst.‡2‡61GxZcRev..‡.‡cRev.‡‡.cRev.1.‡B01‡B1‡‡eq22‡‡Rev221‡‡eq2B‡1‡‡eq21B‡.‡B0FFxFxFxZGFFeZFxTkkEeGTkGZxFZZZxGFGZxFZTkGeZFxTkkGZxFZTkG1‡‡eq2B‡B‡1‡B0cRev.GZxFZTkGTkGeZxTekkFLastFRevFFirstRefFRevRefF
In Eq. (26),
, and
(26)
where
is the curvature of the free energy profile in the unfolded state, and
is the model for the free energy profile. For the linear-
cubic free energy profile
.
(27)
In Eq. (27),
. As previously
(in the case of
the mean last unfolding force) is not an independent parameter and can be expressed through
other parameters of the system[33]
25
cFirst..‡ref.‡refcFirst.‡refref‡ref.cFirst.2ref1ref.‡refB0ref1‡refB1‡refref‡reffin2refref‡ref‡refFirstRef221‡refref‡reffin2refB‡ref1‡refref‡reffin2ref1B‡refref.‡refB0refFFxFxFxZGFFZeZFxTkkEeGTkGZxFZZxGFGZxFZTkGeZFxTkkGZxFZTkG1‡refref‡reffin2refB‡refB‡ref1ref‡refB0refcFirst.GZxFZTkGTkGeZxTekkFminGref0ref1xZmin0ref20ref2minGxxxxGx‡refref11‡refref0ref‡‡1xxvGxxvvGxGv‡ref2‡refref61GxZcRevRef..‡ref.‡refcRevRef.‡refref‡ref.cRevRef.2ref1ref.‡refB0ref1‡refB1‡refref‡refeq2refref‡ref‡refRevRef221‡refref‡refeq2refB‡ref1‡refref‡refeq2ref1B‡refref.‡refB0refFFxFxFxZGFFZeZFxTkkEeGTkGZxFZZxGFGZxFZTkGeZFxTkkGZxFZTkG1‡refref‡refeq2refB‡refB‡ref1ref‡refB0refcRevRef.GZxFZTkGTkGeZxTekkFeqF
where
is the polylogarithm[11].
In order to see how well different models predict the dependence of the mean unfolding
force on the loading rate when a pulling force is applied through a spring, we carried out BD
simulations using Morse free energy profile which was pulled with constant speed through a
spring. The spring constant was 100 pN/nm and the Morse free energy profile parameters were
chosen as in[32]. We generated 3000 trajectories using loading rates, for the loading rates
ranging through 11 orders of magnitude, as presented in Fig. 6 (see methods section for details).
To show how well above discussed models predict the dependence of the mean unfolding
force on the loading rate, we plot the models' predictions and the data from the BD simulations
on the same graph (Fig. 6). The parameters for plotting the models were the same as those used
in the BD simulations. To quantify how well the models predict the BD simulation data, we have
calculated R2 and χ2 (see methods section for details), which are presented in Table 2. Unified
models generally give the best predictions. It appears that the soft spring approximation works
well in this case because Z is close to 1, and hence we could have used models that do not take
into account the correction for the spring. However, since the reversible model[32,33] was
originally developed assuming pulling through the spring, we used models that take into account
the effect of the spring in order to have all models on equal ground.
26
TkGGTkGGFeTkeFB‡ref‡B‡ref‡221B02eqLi21121Li
27
Figure. 6. Dependence of the mean unfolding force on the loading rate, from BD simulations with Morse free energy
profile, and from the predictions from the DFS models. (b) shows the log-log plot for the data in (a), but for a larger
range of loading rates. Three regimes can be observed in (b). In the low loading rate regime the three differently
defined mean unfolding forces diverge. This regime ends roughly where the mean last unfolding force has a
minimum, the mean reversible unfolding force rises from the constant Feq value, and the mean first unfolding force
follows a different slope, around 5 106 pN/s. The intermediate loading rate regime starts where the low loading rate
regime ends and ends roughly where the change in the slope can be noticed around 1010 pN/s, indicating that
unfolding occurs after the barrier has disappeared. The models are plotted using the parameters from the BD
simulation (hence these are not fits). Phenomenological (1997, 2000, 2012) corresponds to Eq. (7), profile
assumption (2006, 2008, 2010) corresponds to Eq. (8), reversible (2012, 2016) corresponds to Eq. (6), rapid (2014)
corresponds to Eq. (10), unified first (2019) corresponds to Eq. (24), unified reversible (2019) corresponds to Eq.
(25), unified last (2019) corresponds to Eq. (21) where Eq. (25), Eq. (26), and Eq. (27) were used. Table 2 shows
how well models predict the data from this figure. The parameters used in the models were kBT = 4.1 pN nm, =
100 pN/nm,
= 0.162 nm, k0 = 5002.51 s-1,
= 37.85 pN nm,
= 10-5 pN s/nm, Feq = 76.4 pN, Ffin = 150
pN, v = 0.66667,
= 0.338 nm, k0ref = 550713 s-1 , and
= 8.4 pN nm. The error bars for the standard
error of the mean are smaller than the symbols.
Model
Low loading rates
Intermediate
loading
High loading rates
All loading rates
(
< 107pN/s)
rates (107 pN/s
1010 pN/s)
(
> 1010 pN/s)
R2
χ2
R2
χ 2
R 2
χ 2
R2
χ 2
Phenomenological
(1997, 2000, 2012)
Profile
(2006, 2008, 2010)
assumption
-0.8937
8.5561
0.7867
8.6606
-0.4415
2534.49
-0.1848
2550.6
0.6993
2.5944
1.0765+
0.6246-
-0.4428-
2513.46+
-0.1859-
2517.91+
111.189i
0.0716i
1.0892
0.0096i
112.249i
0.0079i
Rapid (2014)
0.6861
2.726
0.9396
1.2329
0.9995
4.8118
0.9996
8.459
Unified first (2019)
0.6993
2.5944
0.8855
3.0696
0.9996
1.2131
0.9997
6.5727
Reversible
2016)
Unified
(2019)
(2012,
0.3416
1.2726
0.6724
4.6319
-0.4426
2554.35
-0.1858
2560.25
reversible
-0.6159
2.7631
0.8735
4.1891
0.9996
1.2143
0.9997
8.1665
Unified last (2019)
-3.0559
1.5944
0.8723
4.275
0.9996
1.2144
0.9997
7.8419
Table 2. To quantify how well the models predict the simulated data we calculate R2 and χ2 for each model plotted in
Fig. 6 at three regimes separately and then all together (last column). The profile assumption (2006, 2008, 2010)
model predicts complex values for the mean unfolding force at high loading rates. For the completeness of the table
we present R2 and χ2 for this model as well. Perfect prediction would result in R2 = 1 and χ2 = 0.
28
‡x‡G‡refx‡refG.F.F.F
Distributions of differently defined unfolding forces
The distributions of differently defined unfolding forces in the low loading rate regime
are vastly different (Fig. 7). The reversible unfolding force distribution is symmetric and narrow.
The last unfolding force is more widely distributed and has a right tail. Generally at intermediate
loading rates, in cases of single transition, the unfolding force distribution has a left tail[14 -- 18].
When there is hopping, the first unfolding force distribution also shows a left tail, but as the
loading rate decreases the overall distribution moves closer to 0 and the left tail disappears, as
can be seen in Fig. 7a. The evolution of the differently defined unfolding force distributions over
the loading rate is an interesting question since the distributions are very different in the near-
equilibrium regime but eventually coincide as the system moves further from equilibrium. A
subject of future research could be the derivation of a differential equation describing these
―dynamics‖.
Figure. 7. Distributions of the first, reversible, and last unfolding forces obtained from BD simulations on quartic
free energy profile with a loading rate of 100 pN/s (a) and on a Morse free energy profile where the force is applied
through a spring with 100 pN/nm stiffness that was pulled at a speed of 1000 nm/s (b).The last unfolding force
distribution shows a right tail. The reversible unfolding force distribution is symmetric around its mean and is the
most compact. The characteristic left tail of the first unfolding force distribution disappears at low loading rates.
29
Mean last unfolding force behavior at low loading rates
As described above, in the low loading rate regime the mean last unfolding force starts to
increase with the decrease of the loading rate. Performing BD simulations at lower loading rates
is challenging because it requires long simulation times and large memory. However, we could
see the prediction of the unified last model at lower loading rates (Fig. 8).
Figure. 8. The dependence of the mean last unfolding force on the loading rate for very low loading rates from the
unified last model. For very low loading rates, the mean last unfolding force approaches toward the constant value
of the final force (Ffin) until which the system is driven. The final force is not an intrinsic parameter of the system but
is set by the experimenter and can be higher than the critical force (force at which barrier disappears). These curves
were generated using Eq. (21) with Eq. (20), eq (22), and Eq. (23), with parameter values kBT = 4.1 pN nm,
=
4.27 nm, k0 = 9.33 s-1,
= 40.62 pN nm,
= 10-5 pN s/nm, Feq = 4.02 pN, v = 0.66667,
=2.74 nm, k0ref =
6495 s-1, and
= 12.41 pN nm.
We first observe an increase and then a plateau of the mean last unfolding force at low
loading rates. This means that the time for the last unfolding increases with decreasing loading
rate, such that the product of that time and the loading rate still increases. When the loading rate
30
‡x‡G‡refx‡refG
goes to 0,
(Eq. (22)) tends toward
, i.e. to the final force until which the system is
pulled or the force from which the relaxation could start. At the same limit,
(Eq. (20)) and
(Eq. (23)) both tend toward the same constant
, hence
(Eq. (21)) will tend
toward
as
. The plateau is an indication that at a sufficiently low loading rates
hopping will continue until the end of the process, i.e. until the final force to which the system is
driven. This occurs because there is always a finite probability that the system will refold back
and unfold again, even if the energy of the native state has become much higher than that of
unfolded state during the pulling process. Over sufficiently long time periods, such as those
needed to accomplish the entire pulling process at very low loading rates, such improbable
events are likely to occur.
DISCUSSION
It is crucial to understand that different models might use different definitions of the
unfolding forces. As was shown in Fig. 5 and Fig. 6, differently defined mean unfolding forces
are very different in the low loading rate regime. Nevertheless, the reversible model[32,33] has
been widely used for fitting the dependence of the mean unfolding force on the loading rate
without specifying how the unfolding forces were determined[13,42 -- 44]. It is possible that data
was collected from the medium and high loading rate regimes where hopping is not observed.
However, if there is a lack of data from all three regimes some of the parameters cannot be
estimated reliably. For example, if there is no data from the low loading rate regime (where the
hopping occurs) then the parameters describing the equilibrium properties of the system (such as
,
and
) cannot be reliably estimated. The detection of hopping is not
only related to lower loading rates but also requires a higher signal to noise ratio and temporal
resolution. For example in mechanical unfolding studies of bacteriorhodopsin through AFM
from the freely spanning membrane which lowers the loading rates considerably, hopping was
not observed[45]. However, using modified cantilevers which significantly increase the temporal
resolution hopping and new
intermediates
in bacteriorhodopsin unfolding have been
observed[30,46]. Therefore, we encourage researchers[28,30,46] who possess the data or
31
FirstRefFfinFRevFRevRefFeqFLastFfinFFirstRefFeqF‡ref‡xx‡ref‡GG
capability of acquiring data where hopping is detected for a wide range of loading rates to test
the predictions of the unified model for the mean last unfolding force.
The little cusp in the unified models at the transition from intermediate to high loading
rates (see Fig. 5a and Fig. 6a) is a consequence of using the profile assumption model[15,22],
which in turn uses Kramers' high barrier approximation[20]. The barrier gets lower under the
externally applied force and the approximation breaks down before but close to the critical force.
In the current study, the unification of the dependence of the mean unfolding force on the
loading rate in the low and intermediate loading rate regimes with the dependence in the high
loading rate regime was done using a mathematical trick. In order to have a full physical
description, we need an analytical and analytically integrable expression for the dependence of
the unfolding rate on the force. The expression must avoid the Kramers' high barrier
approximation and be valid for the whole range of the forces, even above the critical force.
CONCLUSIONS
The determination of the unfolding force from the simulated or measured data (trajectory
or force-extension curve) is highly important for obtaining accurate results for the energetic and
dynamic parameters of the system under study. The determination of the unfolding force should
be done according to the definition of the unfolding force used in the DFS model in question,
according to which dependence of the mean unfolding force on the loading rate is analyzed. This
is especially important in the low loading rate regime where the three unfolding force definitions,
discussed in this work, differ significantly.
The mean last unfolding force is more instructive than, for example, the mean first
unfolding force. The unified last model has 5 independent parameters, and the curve describing
the dependence of the mean last unfolding force on the loading rate shows more features than the
curve describing the dependence of the mean first unfolding force on the loading rate.
The unified models (unified first, unified reversible, unified last) proposed in this work
give better overall predictions for the BD simulated data compared to other currently available
models. Many of the previously proposed models fail in the high loading rate regime. The
decisive test for these models would be DFS measurements on a system for which all the
32
energetic and dynamic parameters are known, either through the design of the system (protein)
or by different direct measurements. Then, the models' predictions should be fitted to the data
for the dependence of the mean unfolding force on the loading rate. The parameters obtained
from those fits should be compared to the actual values of the parameters.
APPENDIX A
Mean reversible rnfolding force dependence on the loading rate for the unified model
The derivation of Eq. (20) will be shown in two steps. First, the dependence of the mean
unfolding force on the loading rate for the low and intermediate loading rates is calculated.
Second, that result is united with the result for high loading rates.
For the first part, we follow the approach of[32] with a difference that we use the profile
assumption rate expression[15] (Eq. (A1)) which is more general than the phenomenological
rate[6 -- 8] used in[32] and reduces to it when v = 1.
(A1)
(A2)
Using the approach of[32], by substituting Eq. (A2) into Eq. (A3) we get Eq. (A4).
(A3)
33
1‡‡B‡1111‡‡01GFxTkGeGFxkFk1‡‡B‡11‡0GFxTkGBexTkkdFFkFdFkFppdFFpeq.11
By solving Eq. (A4) for F we find Eq. (A5).
(A4)
(A5)
The advantage of using the rate in Eq. (A1)[15] versus the rate used in[23], where the first power
is
instead of
, is that the indefinite integral in Eq. (A2) will contain an En-
function[10,11] which complicates the later step of solving Eq. (A4) for F.
At this step, we follow the approximate averaging procedure as in[22].
From Eq. (A5) and Eq. (A6), we obtain Eq. (A7) for the mean reversible unfolding force.
(A6)
34
1‡‡eqB‡1‡‡B‡1111.‡B0lnGxFTkGGFxTkGeeFxTkkpTkkpFxeGTkxGFGxFTkGB0.‡11‡B‡‡lnln111‡‡eqB‡/121/1TkkpFxeGxFTkGB0.‡11lnln1‡‡eqB‡FFBeAAABEedpBpAABlnlnlnln1101‡‡eqB‡11.‡B01‡‡‡Rev11111‡‡eqB‡1‡‡eq11B‡.‡B0GxFTkGeFxTkkEeGTkxGFGxFTkGeFxTkkBGxFTkG
Using the approximation from Eq. (A6) in Eq. (A7), we get Eq. (A8).
(A7)
(A8)
Note that Eq. (A7) and Eq. (A8) are valid until the disappearance of the barrier. Mathematically,
this is demonstrated by the negative values taken to the non-integer power. Next, we determine
the minimum loading rate at which this occurs.
At the critical loading rate, the mean reversible unfolding force is approximately
according to Eq. (A7) and Eq. (A8). After the barrier disappearance, the mean unfolding force
35
TkkeFxeGTkxGFGxFTkGB0un.‡11‡B‡‡Rev1‡‡eqB‡ln110ln1B0.‡11‡B1‡‡eqB‡TkkeFxeGTkGxFTkGTkGGxFTkGeTkkeFxeB‡1‡‡eqB‡B0.‡111‡‡eqB‡B‡11‡0.GxFTkGTkGBeexTekkF1‡‡eqB‡B‡1‡B0.1GxFTkGTkGexTekkF‡‡/xG
increases as the square root of the loading rate
[14,18]. In order to combine the two
functions that are valid for the different loading rate regimes, we take the absolute value of the
expression that is raised to the non-integer power in Eq. (A7) and use the Heaviside step
function:
(A9)
where
and
.
Eq. (A9) can be further simplified to Eq. (20), or using the approximation from Eq. (A6) to Eq.
(A10).
(A10)
36
‡.2xFcRev..‡.‡cRev.‡‡.cRev.1‡‡eqB‡11.‡B01‡B‡‡Rev2211111‡‡eqB‡1‡‡eq11B‡.‡B0FFxFxFxGFFGxFTkGeFxTkkEeGTkxGFGxFTkGeFxTkkGxFTkG1‡‡eqB‡B‡1‡B0cRev.1GxFTkGTkGexTekkFTkGexGkB‡2‡‡03cRev..‡.‡cRev.‡‡.cRev.B0un.‡11‡B‡‡Rev22ln111‡‡eqB‡FFxFxFxGFFTkkFxeeGTkxGFGxFTkG
Dependence of the mean first unfolding force on the loading rate for the unified model
The dependence of the mean first unfolding force on the loading rate Eq. (19) can be
obtained by starting the integration in Eq. (A3) from 0 and following the same steps, or by
directly setting
in Eq. (A9) and further simplifying.
Dependence of the mean last unfolding force on the loading rate for the unified model
The mean first refolding force Eq. (22) and the mean reversible refolding force Eq. (23)
that are used for calculating the mean last unfolding force in Eq. (21) can be obtained using the
same methods as described above, though by using Eq. (A11) for the refolding rate[41] and by
using limits of integration for
from
to
in Eq. (A13) and for
from
to
in a similar equation.
(A11)
(A12)
By substituting Eq. (A12) into Eq. (A13) we get Eq. (A14).
(A13)
(A14)
By solving Eq. (A14) for F we find Eq. (A15).
37
0eqFFirstRefFfinFFRevRefFeqFF1‡ref‡refB‡ref1111‡ref‡ref0refref1GFxTkGeGFxkFk1‡ref‡refB‡ref11‡refB0refrefGFxTkGexTkkdFFkFdFkFppdFFpfinref.111‡ref‡reffinB‡ref1‡ref‡refB‡ref1111.‡refB0reflnGxFTkGGFxTkGeeFxTkkp
(A15)
At this step, we follow the approximate averaging procedure as in[22].
From Eq. (A15) and Eq. (A6), we obtain Eq. (A16) for the mean first refolding force.
Using the approximation from Eq. (A6) in Eq. (A16), we get Eq. (A17).
(A16)
(A17)
Next, we find the critical loading rate in the same way as previously.
38
1lnln1B0ref.‡ref11‡refB‡ref‡ref1‡ref‡reffinB‡refTkkpFxeGTkxGFGxFTkGTkkpFxeGxFTkGB0ref.‡ref11lnln1‡ref‡reffinB‡refFF11111‡ref‡reffinB‡ref11.‡refB0ref1‡refB‡ref‡refFirstRef1‡ref‡reffinB‡ref1‡ref‡reffin11B‡ref.‡refB0refGxFTkGeFxTkkEeGTkxGFGxFTkGeFxTkkGxFTkG1ln1B0ref.‡ref11‡refB‡ref‡refFirstRef1‡ref‡reffinB‡refTkkFxeeGTkxGFGxFTkG0ln1B0ref.‡ref11‡refB1‡ref‡reffinB‡refTkkFxeeGTkGxFTkG
By combining the two functions that are valid for the different loading rate regimes as
previously, using the Heaviside step function we obtain
(A18)
where
and
.
Eq. (A18) can be further simplified to Eq. (22), or by using the approximation from Eq. (A6) in a
similar equation to Eq. (A10).
The dependence of the mean reversible refolding force on the loading rate Eq. (23) can be
obtained by starting the integration in Eq. (A13) from
and following the same steps, or by
directly setting
to
in Eq. (A18) and further simplifying.
APPENDIX B
Dependence of the mean reversible unfolding force on the loading rate for the unified
model when the force is applied through the spring
39
1‡ref‡reffinB‡refB‡ref1‡refB0ref.1GxFTkGTkGexTekkFcFirstRef..‡ref.‡refcFirstRef.‡ref‡ref.cFirstRef.1‡ref‡reffinB‡ref11.‡refB0ref1‡refB‡ref‡refFirstRef2211111‡ref‡reffinB‡ref1‡ref‡reffin11B‡ref.‡refB0refFFxFxFxGFFGxFTkGeFxTkkEeGTkxGFGxFTkGeFxTkkGxFTkG1‡ref‡reffinB‡refB‡ref1‡refB0refcFirstRef.1GxFTkGTkGexTekkFTkGexGkB‡ref2‡ref‡ref0ref3eqFfinFeqF
When the force is being applied through the spring, we need to use an appropriate
expression for the force-dependent unfolding rate Eq. (B1)[34], which then can be plugged into
Eq. (B2).
(B1)
(B2)
Using the approach of[32], by substituting Eq. (B2) into Eq. (B3) we get Eq. (B4).
(B3)
By solving Eq. (B4) for F, we find Eq. (B5).
(B4)
(B5)
At this step, we follow the approximate averaging procedure as in[22].
40
1‡‡2B‡111‡‡20GZFxZTkGeGZFxZkFk1‡‡2B‡1‡B0GZFxZTkGeZxTkkdFFkFdFkFppdFFpeq.111‡‡eq2B‡1‡‡2B‡11.‡0lnGZxFZTkGGZFxZTkGBeeFZxTkkpTkkpFZxeGTkZZxGFGZxFZTkGB0.‡1‡B2‡‡lnln11‡‡eq2B‡TkkpZFxeGZxFZTkGB0.‡1lnln1‡‡eq2B‡
From Eq. (B5) and Eq. (A6) we obtain the mean reversible unfolding force:
Using the approximation from Eq. (A6) in Eq. (B6), we get Eq. (B7).
(B6)
Next, we find the critical loading rate in the same way as in Appendix A.
(B7)
By combining the two functions that are valid for the different loading rate regimes as was done
in Appendix A, using the Heaviside step function we obtain
41
FF1‡‡eq2B‡1.‡B01‡B2‡‡Rev111‡‡eq2B‡1‡‡eq21B‡.‡B0GZxFZTkGeZFxTkkEeGTkZZxGFGZxFZTkGeZFxTkkGZxFZTkGTkkZeFxeGTkZZxGFGZxFZTkGB0.‡1‡B2‡‡Rev1‡‡eq2B‡ln10ln1B0.‡1‡B1‡‡eq2B‡TkkZeFxeGTkGZxFZTkG1‡‡eq2B‡B‡1‡B0.GZxFZTkGTkGeZxTekkF
(B8)
where
Eq. (B8) can be further simplified to Eq. (25), or by using the approximation from Eq. (A6) to a
similar equation to Eq. (A10).
Dependence of the mean first unfolding force on the loading rate for the unified model
when the force is applied through the spring
The dependence of the mean first unfolding force on the loading rate Eq. (24) can be
obtained by starting the integration in Eq. (B3) from 0 and following the same steps, or by
directly setting
in Eq. (B8) and further simplifying.
Dependence of the mean last unfolding force on the loading rate for the unified model when
the force is applied through the spring
The mean first refolding force Eq. (26) and the mean reversible refolding force Eq. (27)
that are used for calculating the mean last unfolding force when the force is applied through the
spring in Eq. (21) can be obtained using the same methods as described above, using Eq. (B9) for
the refolding rate[41] and limits of integration for
from
to
in Eq. (B11) and for
from
to
in a similar equation.
42
cRev..‡.‡cRev.‡‡.cRev.1‡‡eq2B‡1.‡B01‡B2‡‡Rev22111‡‡eq2B‡1‡‡eq21B‡.‡B0FFxFxFxZGFFGZxFZTkGeZFxTkkEeGTkZZxGFGZxFZTkGeZFxTkkGZxFZTkG1‡‡eq2B‡B‡1‡B0cRev.GZxFZTkGTkGeZxTekkF0eqFFirstRefFfinFFRevRefFeqFF
By substituting Eq. (B10) into Eq. (B11) we get Eq. (B12).
(B9)
(B10)
(B11)
(B12)
By solving Eq. (B12) for F, we find Eq. (B13).
(B13)
At this step, we follow the approximate averaging procedure as in[22].
From Eq. (B13) and Eq. (A6), we obtain for the mean first refolding force Eq. (B14).
43
1‡refref‡ref2refB‡111‡refref‡ref2ref0refrefGZFxZTkGrefeGZFxZkFk1‡refref‡ref2refB‡ref1ref‡refB0refrefGZFxZTkGeZxTkkdFFkFdFkFppdFFpfinref.111‡refref‡reffin2refB‡ref1‡refref‡ref2refB‡ref11ref.‡refB0reflnGZxFZTkGGZFxZTkGeeZFxTkkp2refB0refref.‡ref1‡refBref‡ref‡reflnln11‡refref‡reffin2refB‡refZTkkpZFxeGTkZxGFGZxFZTkGTkkpZFxeGZxFZTkGB0refref.‡ref1lnln1‡refref‡reffin2refB‡refFF
Using the approximation from Eq. (A6) in Eq. (B14), we get Eq. (B15).
(B14)
(B15)
Next, we find the critical loading rate in the same way as previously.
By combining the two functions that are valid for the different loading rate regimes as
previously, using the Heaviside step function we obtain Eq. (B16):
(B16)
44
2ref1‡refref‡reffin2refB‡ref1ref.‡refB0ref1‡refBref‡ref‡refFirstRef111‡refref‡reffin2refB‡ref1‡refref‡reffin2ref1B‡refref.‡refB0refZGZxFZTkGeZFxTkkEeGTkZxGFGZxFZTkGeZFxTkkGZxFZTkG2refB0refref.‡ref1‡refBref‡ref‡refFirstRef1‡refref‡reffin2refB‡refln1ZTkkeZFxeGTkZxGFGZxFZTkG0ln1B0refref.‡ref1‡refB1‡refref‡reffin2refB‡refTkkeZFxeGTkGZxFZTkG1‡refref‡reffin2refB‡refB‡ref1ref‡refB0ref.GZxFZTkGTkGeZxTekkFcFirst..‡ref.‡refcFirst.‡refref‡ref.cFirst.2ref1‡refref‡reffin2refB‡ref1ref.‡refB0ref1‡refBref‡ref‡refFirstRef22111‡refref‡reffin2refB‡ref1‡refref‡reffin2ref1B‡refref.‡refB0refFFxFxFxZGFFZGZxFZTkGeZFxTkkEeGTkZxGFGZxFZTkGeZFxTkkGZxFZTkG
where
Eq. (B16) can be further simplified to Eq. (26) or using the approximation from Eq. (A6) to a
similar equation to Eq. (A10).
The dependence of the mean reversible refolding force on the loading rate when the force
is applied through a spring, given by Eq. (27), can be obtained by starting the integration in Eq.
(B11) from
and following the same steps, or by directly setting
to
in Eq. (B16) and
further simplifying.
REFERENCES
[1]
Merkel R, Nassoy P, Leung A, Ritchie K and Evans E 1999 Energy landscapes of
receptor -- ligand bonds explored with dynamic force spectroscopy Nature 397 50 -- 3
[2]
Noy A 2011 Force spectroscopy 101: how to design, perform, and analyze an AFM-based
single molecule force spectroscopy experiment Curr. Opin. Chem. Biol. 15 710 -- 8
[3]
Noy A and Friddle R W 2013 Practical single molecule force spectroscopy: How to
determine fundamental thermodynamic parameters of intermolecular bonds with an atomic
force microscope Methods 60 142 -- 50
[4]
Arya G 2016 Models for recovering the energy landscape of conformational transitions
from single-molecule pulling experiments Mol. Simul. 42 1102 -- 15
[5]
Evans E and Ritchie K 1997 Dynamic strength of molecular adhesion bonds Biophys. J. 72
1541 -- 55
[6]
Busse W F, Lessig E T, Loughborough D L and Larrick L 1942 Fatigue of Fabrics J. Appl.
Phys. 13 715 -- 24
[7]
Regel' V R, Slutsker A I and Tomashevskii É E 1974 The kinetic nature of the strength of
solids (Moscow: Nauka)
[8]
Bell G 1978 Models for the specific adhesion of cells to cells Science 200 618 -- 27
45
1‡refref‡reffin2refB‡refB‡ref1ref‡refB0refcFirst.GZxFZTkGTkGeZxTekkFeqFfinFeqF
[9]
Gergely C, Voegel J-C, Schaaf P, Senger B, Maaloum M, Horber J K H and Hemmerle J
2000 Unbinding process of adsorbed proteins under external stress studied by atomic force
microscopy spectroscopy Proc. Natl. Acad. Sci. 97 10802 -- 7
[10] Abramowitz M and Stegun I A 1972 Handbook of Mathematical Functions With
Formulas, Graphs and Mathematical Tables (Washington D. C.: National Bureau of
Standards)
[11] Weisstein E W 2002 CRC Concise Encyclopedia of Mathematics (Boca Raton: CRC Press)
[12] Zocher M, Zhang C, Rasmussen S G F, Kobilka B K and Müller D J 2012 Cholesterol
increases kinetic, energetic, and mechanical stability of the human 2-adrenergic receptor
Proc. Natl. Acad. Sci. 109 E3463 -- 72
[13] Rico F, Russek A, González L, Grubmüller H and Scheuring S 2019 Heterogeneous and
rate-dependent streptavidin -- biotin unbinding revealed by high-speed force spectroscopy
and atomistic simulations Proc. Natl. Acad. Sci. 116 6594 -- 601
[14] Hummer G and Szabo A 2003 Kinetics from Nonequilibrium Single-Molecule Pulling
Experiments Biophys. J. 85 5 -- 15
[15] Dudko O, Hummer G and Szabo A 2006 Intrinsic Rates and Activation Free Energies from
Single-Molecule Pulling Experiments Phys. Rev. Lett. 96
[16] Greenleaf W J, Frieda K L, Foster D A N, Woodside M T and Block S M 2008 Direct
Observation of Hierarchical Folding in Single Riboswitch Aptamers Science 319 630 -- 3
[17] Yu H, Gupta A N, Liu X, Neupane K, Brigley A M, Sosova I and Woodside M T 2012
Energy landscape analysis of native folding of the prion protein yields the diffusion
constant, transition path time, and rates Proc. Natl. Acad. Sci. 109 14452 -- 7
[18] Bullerjahn J T, Sturm S and Kroy K 2014 Theory of rapid force spectroscopy Nat.
Commun. 5
[19] Makarov D E 2016 Perspective: Mechanochemistry of biological and synthetic molecules
J. Chem. Phys. 144 030901
[20] Kramers H A 1940 Brownian motion in a field of force and the diffusion model of
chemical reactions Physica 7 284 -- 304
[21] Hänggi P, Talkner P and Borkovec M 1990 Reaction-rate theory: fifty years after Kramers
Rev. Mod. Phys. 62 251 -- 341
[22] Friddle R W 2008 Unified Model of Dynamic Forced Barrier Crossing in Single Molecules
Phys. Rev. Lett. 100
[23] Cossio P, Hummer G and Szabo A 2016 Kinetic Ductility and Force-Spike Resistance of
Proteins from Single-Molecule Force Spectroscopy Biophys. J. 111 832 -- 40
46
[24] Chung H S and Eaton W A 2013 Single-molecule fluorescence probes dynamics of barrier
crossing Nature 502 685 -- 8
[25] Manosas M and Ritort F 2005 Thermodynamic and Kinetic Aspects of RNA Pulling
Experiments Biophys. J. 88 3224 -- 42
[26] Mossa A, Manosas M, Forns N, Huguet J M and Ritort F 2009 Dynamic force
spectroscopy of DNA hairpins: I. Force kinetics and free energy landscapes J. Stat. Mech.
Theory Exp. 2009 P02060
[27] Manosas M, Mossa A, Forns N, Huguet J M and Ritort F 2009 Dynamic force
spectroscopy of DNA hairpins: II. Irreversibility and dissipation J. Stat. Mech. Theory Exp.
2009 P02061
[28] Junker J P, Ziegler F and Rief M 2009 Ligand-Dependent Equilibrium Fluctuations of
Single Calmodulin Molecules Science 323 633 -- 7
[29] He C, Hu C, Hu X, Hu X, Xiao A, Perkins T T and Li H 2015 Direct Observation of the
Reversible Two-State Unfolding and Refolding of an α/β Protein by Single-Molecule
Atomic Force Microscopy Angew. Chem. Int. Ed. 54 9921 -- 5
[30] Yu H, Siewny M G W, Edwards D T, Sanders A W and Perkins T T 2017 Hidden
dynamics in the unfolding of individual bacteriorhodopsin proteins Science 355 945 -- 50
[31] Evans E 2001 Probing the Relation Between Force -- Lifetime -- and Chemistry in Single
Molecular Bonds Annu. Rev. Biophys. Biomol. Struct. 30 105 -- 28
[32] Friddle R W, Noy A and De Yoreo J J 2012 Interpreting the widespread nonlinear force
spectra of intermolecular bonds Proc. Natl. Acad. Sci. 109 13573 -- 8
[33] Friddle R W 2016 Analytic descriptions of stochastic bistable systems under force ramp
Phys. Rev. E 93
[34] Maitra A and Arya G 2010 Model Accounting for the Effects of Pulling-Device Stiffness
in the Analyses of Single-Molecule Force Measurements Phys. Rev. Lett. 104
[35] Neupane K, Foster D A N, Dee D R, Yu H, Wang F and Woodside M T 2016 Direct
observation of transition paths during the folding of proteins and nucleic acids Science 352
239 -- 42
[36] Schulten K and Kosztin I 2000 Lectures in Theoretical Biophysics (Urbana IL: University
ofIllinois at Urbana -- Champaign)
[37] Szabo A, Schulten K and Schulten Z 1980 First passage time approach to diffusion
controlled reactions J. Chem. Phys. 72 4350 -- 7
[38] Reimann P, Schmid G J and Hänggi P 1999 Universal equivalence of mean first-passage
time and Kramers rate Phys. Rev. E 60 R1 -- 4
47
[39] Hyeon C and Thirumalai D 2007 Measuring the energy landscape roughness and the
transition state location of biomolecules using single molecule mechanical unfolding
experiments J. Phys. Condens. Matter 19 113101
[40] Tshiprut Z and Urbakh M 2009 Exploring hysteresis and energy dissipation in single-
molecule force spectroscopy J. Chem. Phys. 130 084703
[41] Pierse C A and Dudko O K 2013 Kinetics and Energetics of Biomolecular Folding and
Binding Biophys. J. 105 L19 -- 22
[42] Alsteens D, Pfreundschuh M, Zhang C, Spoerri P M, Coughlin S R, Kobilka B K and
Müller D J 2015 Imaging G protein -- coupled receptors while quantifying their ligand-
binding free-energy landscape Nat. Methods 12 845 -- 51
[43] Alsteens D, Newton R, Schubert R, Martinez-Martin D, Delguste M, Roska B and Müller
D J 2016 Nanomechanical mapping of first binding steps of a virus to animal cells Nat.
Nanotechnol. 12 177 -- 83
[44] Reiter-Scherer V, Cuellar-Camacho J L, Bhatia S, Haag R, Herrmann A, Lauster D and
Rabe J P 2019 Force Spectroscopy Shows Dynamic Binding of Influenza Hemagglutinin
and Neuraminidase to Sialic Acid Biophys. J. 116 1037 -- 48
[45] Petrosyan R, Bippes C A, Walheim S, Harder D, Fotiadis D, Schimmel T, Alsteens D and
Müller D J 2015 Single-Molecule Force Spectroscopy of Membrane Proteins from
Membranes Freely Spanning Across Nanoscopic Pores Nano Lett. 15 3624 -- 33
[46] Yu H, Heenan P R, Edwards D T, Uyetake L and Perkins T T 2019 Quantifying the Initial
Unfolding of Bacteriorhodopsin Reveals Retinal Stabilization Angew. Chem. 131 1724 -- 7
48
|
1305.1897 | 1 | 1305 | 2013-05-07T14:30:42 | Coherent diffraction and holographic imaging of individual biomolecules using low-energy electrons | [
"physics.bio-ph",
"physics.ins-det",
"q-bio.BM"
] | Modern microscopy techniques are aimed at imaging an individual molecule at atomic resolution. Here we show that low-energy electrons with kinetic energies of 50-250 eV offer a possibility of overcome the problem of radiation damage, and obtaining images of individual biomolecules. Two experimental schemes for obtaining images of individual molecules, holography and coherent diffraction imaging, are discussed and compared. Images of individual molecules obtained by both techniques, using low-energy electrons, are shown. | physics.bio-ph | physics | Coherent diffraction and holographic
imaging of individual biomolecules using
low-energy electrons
Tatiana Latychevskaia*, Jean-Nicolas Longchamp, Conrad Escher,
Hans-Werner Fink
Physics Institute, University of Zurich
Winterthurerstrasse 190, 8057 Zurich,
Switzerland
*Email: [email protected]
Abstract Modern microscopy techniques are aimed at imaging an in-
dividual molecule at atomic resolution. Here we show that low-energy
electrons with kinetic energies of 50-250 eV offer a possibility of over-
come the problem of radiation damage, and obtaining images of indi-
vidual biomolecules. Two experimental schemes for obtaining images
of individual molecules – holography and coherent diffraction imaging
– are discussed and compared. Images of individual molecules ob-
tained by both techniques, using low-energy electrons, are shown.
Keywords: Phase problem, Individual molecule, Coherent diffraction
imaging, Holography, Low-energy electrons
Introduction
Investigating the structure of biomolecules at the atomic scale has al-
ways been of utmost importance for healthcare, medicine and life sci-
ence in general, since the three-dimensional shape of proteins, for ex-
ample, relates to their function. At the moment, these structural data
are predominantly obtained by X-ray crystallography, cryo-electron
microscopy and NMR. Despite there being an impressive database
(www.pdb.org) obtained with these methods, they all require large
2
quantities of a particular protein. This leads to averaging over fine
conformational details in the recovered structure. The goal of modern
imaging techniques is to visualize an individual biomolecule at atomic
resolution.
Imaging an individual molecule: choice of radiation
A direct visualization of an individual molecule at Ångstrom resolution
can be achieved using electron or X-ray waves which have a wave-
length of about 1 Å, see Fig. 1. Although both, X-rays and high-energy
electrons possess sufficiently short wavelengths to resolve the indi-
vidual atoms constituting a protein, the resolution achieved is mainly
limited by radiation damage inherent to both types of radiation.
Fig. 1. Spectrum of radiation used for imaging. The bars show the range of the sizes of the
objects which can be imaged with the assigned radiation.
Imaging with high-energy electrons (80-200 keV)
In cryo-electron microscopy (Adrian et al. 1984), cooling the sample to
the temperature of liquid nitrogen allows a higher electron dose to be
used for the same amount of radiation damage. Depending on the
3
resolution required, typical electron exposures vary between 5 and
25 e/Å2 (Henderson 2004). Due to the very low signal-to-noise ratio in
the images obtained, over 10,000 images of individual molecules typi-
cally need to be collected and averaged to arrive at the reconstruction
of the structure (van Heel et al. 2000).
Imaging with X-rays
Visualization of an individual molecule at atomic resolution by em-
ploying X-rays is planned at the X-ray Free Electron Lasers (XFELs) fa-
cilities which are being developed worldwide. Here the radiation
damage problem (Howells et al. 2009) is circumvented by employing
ultra-short X-ray pulses, which allow the diffraction pattern of an indi-
vidual molecule to be recorded before it decomposes due to the strong
inelastic scattering (Neutze et al. 2000; Bergh et al. 2008). High inten-
sity X-ray laser pulses will provide the intensity in the diffraction pat-
tern detected at the high scattering angles, which is required for fur-
ther numerical recovery of the molecular structure at high resolution.
Imaging with low-energy electrons (50-250 eV)
Low-energy electrons (with kinetic energies of 50-250 eV, corre-
sponding to wavelengths of 0.78-1.73 Å) can be employed to visualize
individual biomolecules directly. It has been shown (Germann et al.
2010) that individual DNA molecules can withstand low-energy elec-
tron radiation having energy of 60 eV (corresponding to a wavelength
of 1.58 Å) for at least 70 minutes. This in total amounts to a radiation
dose of 106 e/Å2, which is at least six orders of magnitude larger than
the permissible dose in high-energy electron microscopy or X-ray im-
aging.
4
Imaging an individual molecule: the phase problem
The principle of lensless imaging of an individual molecule is as fol-
lows: when a coherent wave is scattered by a molecule, it carries both,
amplitude and phase information imposed by the scattering event. The
phase distribution is especially important since it carries information
about the position of the atoms constituting the molecule. However,
detectors are not sensitive to the phase information; instead they just
record the intensity which is the square of the wave amplitude. Hence,
the recovery of the complex-valued scattered wave requires a solution
to the so-called phase problem. Today there are two known solutions
to the phase retrieval problem: holography and coherent diffraction
imaging (CDI), both schematically shown in Fig. 2. Their proper im-
plementation would ultimately allow the atomic mapping of an indi-
vidual molecule in three dimensions.
Fig. 2. Schematics of the lensless imaging of an individual molecule.
In Fig. 3, the experimental set-ups for both, holographic and CDI re-
cording with low-energy electrons designed and built in our labora-
tory (Fink et al. 1990; Steinwand et al. 2011) are sketched.
5
Fig. 3. Low-energy electron microscopes. Coherent low-energy electrons are extracted
from the electron point source (EPS) by field emission. A biological molecule is fixed in the
sample holder (SH). In the holographic microscope, the interference between the scattered
(object) wave and unscattered (reference) wave is recorded by the detector unit (consist-
ing of a micro-channel plate (MCP) followed by a phosphor screen (PS)). In the CDI micro-
scope, the electron beam is collimated by a microlens (ML) and the detector unit consists of
an MCP followed by a fibre optic plate (FOP) with a thin phosphorous layer (PS).
6
Holography
In holography the unknown wave that is scattered by an object is su-
perimposed with a known reference wave. A hologram is the interfer-
ence pattern formed by constructive and destructive interference be-
tween these two waves (Gabor 1949) and is illustrated in Fig. 4. The
holography technique uniquely solves the phase problem in one step
because of the presence of the reference wave. However, it lacks high
resolution due to the higher-order scattered signal being buried in the
experimental noise of the reference wave.
Fig. 4. Illustration of recording an inline hologram and reconstructing the object. Re-
cording: the superposition of the reference wave and the wave scattered by the object is re-
corded; a magnified region of the hologram shows the fringes of the interference pattern.
Reconstruction: back propagation of the recovered object wave from the hologram plane to
the planes of the object’s location (analogous to optical sectioning) results in a three-
dimensional reconstruction.
7
Hologram reconstruction includes two steps: (1) illumination of the
hologram with the reference wave and (2) backward propagation of
the wavefront to the position of the object. In numerical reconstruc-
tion, the complex-valued reference wave at the hologram plane is
simulated and the propagation from the hologram back to the object is
calculated using Huygens’ principle and Fresnel formalism:
, (1)
where H(r) is the hologram’s transmission function distribution, r and
are defined as illustrated in Fig. 4, and the integration is performed
over the hologram’s surface s. The result of this integral transform is
a complex-valued distribution of the object wavefront at any coordi-
nate , and, hence, a three-dimensional reconstruction.
Fig. 5. Low-energy electron hologram of an individual ssDNA molecule stretched over a
hole in a carbon film (sample courtesy by Michael and William Andregg,
www.halcyonmolecular.com).
An example of an inline hologram of an individual DNA molecule and
its reconstruction is shown in Fig. 5. The successful trials during the
last decade of imaging individual biological molecules by low-energy
electron holography include the imaging of: DNA molecules (Fink et al.
1997; Eisele et al. 2008), phthalocyaninato polysiloxane molecules
Sexp()exp()()()dikriikrUHrrr 8
(Golzhauser et al. 1998), the tobacco mosaic virus (Weierstall et al.
1999), a bacteriophage (Stevens et al. 2011) and ferritin (Longchamp
et al. 2012). Despite a very short wavelength (1-2 Å) of the probing
electron wave, the resolution in the reconstructed molecular struc-
tures remains in the order of a few nanometres. The reason is that the
resolution in inline holography is limited by the detectability of the in-
terference fringes at high diffraction angles (Spence et al. 1994;
Latychevskaia et al. 2011) (such as, for instance, the fringes shown in
the magnified region in Fig. 4). The pattern of these fine fringes is very
sensitive to the object’s lateral movements and can be destroyed by
the object shifting even by just the distance corresponding to the
wavelength. In addition, these fine fringes are often buried in the ex-
perimental noise of the reference wave.
Coherent diffraction imaging
CDI is a relatively modern technique which combines the recording of
a far-field diffraction pattern of a non-crystalline object and the nu-
merical recovery of the object structure. In 1952, Sayre proposed that
it was possible to recover the phase information associated with scat-
tering off a non-crystalline specimen by sampling its diffraction pat-
tern at a frequency higher than twice the Nyquist frequency (over-
sampling) (Sayre 1952). In 1972, Gerchberg and Saxton proposed an
iterative algorithm to recover the phase distribution from two ampli-
tude measurements taken: at the object plane and at the far-field plane
(Gerchberg and Saxton 1972). In 1998, Miao et al. combined these two
ideas and successfully recovered an object from its oversampled dif-
fraction pattern (Miao et al. 1998). They demonstrated that the phase
retrieval algorithm converges if the initial conditions are such that the
surrounding of the molecule (“support”) is known. The concept of
knowing the support of the molecule is analogous to the solvent flat-
tening technique in the phasing methods. The known surrounding of a
molecule is usually mathematically described by zero-padding the ob-
ject, which in turn leads to oversampling of the spectrum in the Fou-
rier domain. Thus, reconstruction becomes possible if the diffraction
pattern is recorded under the oversampling condition (Miao et al.
1999; Miao and Sayre 2000; Miao et al. 2003b); this is also illustrated
in Fig. 6.
9
Fig. 6. Sampling at the Nyquist frequency (upper row) and twice the Nyquist frequency
(lower row). The Fourier transform of the spectrum sampled at the Nyquist frequency re-
sults in the object distribution filling the entire reconstructed area. The Fourier transform
of the spectrum sampled at twice the Nyquist frequency results in the zero-padded object
distribution (Miao and Sayre 2000).
The basic iterative reconstruction loop (Fienup 1982) is shown in Fig.
7. It begins with the complex-valued wave distribution at the detector
plane which is formed by the superposition of the square root of the
measured intensity and a random phase distribution. In the object
domain various constraints are applied. For instance, the electron
density reconstructed from the X-ray diffraction images must be real
and positive.
10
The resolution in CDI is defined by the outermost detected signal in
the diffraction pattern, R=/sin, where is the scattering angle. The
resolving power of the CDI technique has already been demonstrated
by the reconstruction of a double-walled carbon nanotube at a resolu-
tion of 1 Å from a coherent diffraction pattern recorded using a 200
keV electron microscope exhibiting a nominal conventional TEM reso-
lution of 2.2 Å (Zuo et al. 2003).
Fig. 7. Iterative reconstruction of a coherent diffraction pattern. Left column: amplitude
and phase distributions at the detector plane initiating the iterative loop. Right: the steps
(i)-(iv) showing the flow of the iterative loop.
The X-ray diffraction pattern of a crystal, unlike that of an individual
molecule, displays a strong signal due to the periodicity of the crystal.
Obtaining an X-ray diffraction pattern of an individual molecule in turn
requires a much more intense X-ray beam. As a consequence, the reso-
lution is limited by radiation damage and remains very moderate. A
few biological specimens have been imaged by CDI using X-rays at a
resolution of a few nanometres: E.coli bacteria (Miao et al. 2003a), an
11
unstained yeast cell (Shapiro et al. 2005), single herpes virions (Song
et al. 2008), malaria-infected red blood cells (Williams et al. 2008), a
frozen hydrated yeast cell (Huang et al. 2009), human chromosomes
(Nishino et al. 2009), unstained and unsliced freeze-dried cells of the
bacterium Deinococcus radiodurans by ptychography (Giewekemeyer
et al. 2010), and labelled yeast cells (Nelson et al. 2010). Ultra-short
and extremely bright coherent X-ray pulses from XFEL allow the re-
cording of a high-resolution diffraction pattern before the sample ex-
plodes (Neutze et al. 2000; Bergh et al. 2008). The first results from
the first XFEL facility to be operational (the Linac Coherent Light
Source) reported imaging an individual unstained mimivirus at 32
nanometre resolution (Seibert et al. 2011); in this experiment an X-ray
pulse of 1.8 keV (6.9Å) energy and 70 fs duration was focused to a spot
10m in diameter with 1.6x1010 photons per 1 m2. A sub-nanometre
resolution could be achieved by employing shorter pulses and a higher
photon flux (Bergh et al. 2008; Seibert et al. 2011); at present this is
beyond the capabilities of the XFELs but might be realized with the
next generation of XFELs.
Comparing holography and CDI
Each of the two techniques has its pros (+) and cons (-), which are
summarized below:
Holography
• Requires well-defined reference wave over entire detector area (-)
• Non-iterative reconstruction by calculating back-propagation inte-
gral (+)
• Three-dimensional reconstruction (+)
• Low resolution, due to high sensitivity of the interference pattern to
object shifts and experimental noise in the reference wave (-)
12
CDI
• No reference wave is required (+)
• Reconstruction is done by an iterative procedure and does not al-
ways converge to a uniquely defined outcome (-)
• Reconstruction is not three-dimensional, it is limited to one plane (-)
• High resolution provided by stability of diffraction pattern being in-
sensitive to shifts of the object (+)
HCDI: Combining holography and CDI
Recently, we revealed the relationship between the hologram and the
diffraction pattern of an object, which allows holography and CDI to be
combined into a superior technique: holographic coherent diffraction
imaging (HCDI). HCDI inherits fast and reliable reconstruction from
holography
from
resolution
possible
highest
the
and
CDI(Latychevskaia et al. 2012).
The Fourier transform of an inline hologram is proportional to the
complex-valued object wave in the far-field, as is illustrated for ex-
perimental images in Fig. 8. Thus, the phase distribution of the Fou-
rier transform of the inline hologram provides the phase distribution
of the object wave in the far-field and hence the solution to the “phase
problem” in just one step. The diffraction pattern is then required to
refine the reconstruction of the high-resolution information by a con-
ventional iterative procedure. In addition, the central region of the dif-
fraction pattern, which is usually missing, can be adapted from the
amplitude of the Fourier transform of the hologram; see Fig. 8d.
The hologram and the diffraction pattern of a bundle of carbon nano-
tubes recorded with the coherent low-energy electron diffraction mi-
croscope (Steinwand et al. 2011) are shown in Fig. 9. The HCDI tech-
nique was applied to reconstruct these images and the result is shown
in Fig. 9d.
13
Fig. 8. Experimental verification of the relationship between a hologram and a diffraction
pattern. (a) Reflected-light microscopy image of two twisted tungsten wires. (b) Inline
hologram recorded with laser light. (c) Diffraction pattern. (d) The amplitude of the Fourier
transform of the hologram is displayed using a logarithmic and inverted intensity scale. The
diffraction pattern provides the same resolution as the hologram - namely 6 m – but it is
recorded while fulfilling the oversampling condition (Latychevskaia et al. 2012).
14
Fig. 9. (a) Hologram of carbon nanotubes acquired with 51 eV electrons. (b) Coherent dif-
fraction pattern of the same nanotubes recorded with electrons of 145 eV kinetic energy.
The dashed circle shows the highest order signal detected corresponding to 9.2 Å resolu-
tion. (c) TEM image of the very same area obtained with 80 keV electrons. (d) Reconstruc-
tion obtained by the HCDI method (Latychevskaia et al. 2012).
Because the phase distribution stored in a holographic image is
uniquely defined and is associated with the three-dimensional object
distribution, HCDI may offer the possibility of retrieving a three-
dimensional object distribution from its diffraction pattern.
Acknowledgments We would like to thank the Swiss National Science Foundation for its
financial support.
References
15
Adrian M, Dubochet J, Lepault J, McDowall AW (1984) Cryo-electron microscopy of
viruses. Nature 308 (5954):32-36
Bergh M, Huldt G, Timneanu N, Maia F, Hajdu J (2008) Feasibility of imaging living cells at
subnanometer resolutions by ultrafast x-ray diffraction. Q Rev Biophys 41 (3-4):181-
204
Eisele A, Voelkel B, Grunze M, Golzhauser A (2008) Nanometer resolution holography
with the low energy electron point source microscope. Z Phys Chem 222 (5 -6):779-
787
Fienup JR (1982) Phase retrieval algorithms - a comparison. Appl Optics 21 (15):2758-
2769
Fink H-W, Schmid H, Ermantraut E, Schulz T (1997) Electron holography of individual
DNA molecules. J Opt Soc Am A 14 (9):2168-2172
Fink H-W, Stocker W, Schmid H (1990) Holography with low-energy electrons. Phys Rev
Lett 65 (10):1204-1206
Gabor D (1949) Microscopy by reconstructed wave-fronts. Proc R Soc A 197 (1051):454-
487
Gerchberg RW, Saxton WO (1972) A practical algorithm for determination of phase from
image and diffraction plane pictures. Optik 35 (2):237-246
Germann M, Latychevskaia T, Escher C, Fink H-W (2010) Nondestructive imaging of
individual biomolecules. Phys Rev Lett 104 (9):095501
Giewekemeyer K, Thibault P, Kalbfleisch S, Beerlink A, Kewish CM, Dierolf M, Pfeiffer F,
Salditt T (2010) Quantitative biological imaging by ptychographic x-ray diffraction
microscopy. Proc Natl Acad Sci USA 107 (2):529-534
Golzhauser A, Volkel B, Jager B, Zharnikov M, Kreuzer HJ, Grunze M (1998) Holographic
imaging of macromolecules. J Vac Sci Technol A - Vac Surf Films 16 (5):3025-3028
Henderson R (2004) Realizing the potential of electron cryo-microscopy. Q Rev Biophys
37 (1):3-13
Howells MR, Beetz T, Chapman HN, Cui C, Holton JM, Jacobsen CJ, Kirz J, Lima E,
Marchesini S, Miao H, Sayre D, Shapiro DA, Spence JCH, Starodub D (2009) An
assessment of the resolution limitation due to radiation-damage in x-ray diffraction
microscopy. J Electron Spectrosc 170 (1-3):4-12
Huang X, Nelson J, Kirz J, Lima E, Marchesini S, Miao H, Neiman AM, Shapiro D, Steinbrener
J, Stewart A, Turner JJ, Jacobsen C (2009) Soft x-ray diffraction microscopy of a frozen
hydrated yeast cell. Phys Rev Lett 103 (19):198101
Latychevskaia T, Longchamp J-N, Fink H-W (2011) When holography meets coherent
diffraction imaging arXiv:1106.1320
Latychevskaia T, Longchamp J-N, Fink H-W (2012) When holography meets coherent
diffraction imaging Opt Express 20 (27):28871-28892
Longchamp J-N, Latychevskaia T, Escher C, Fink H-W (2012) Non-destructive imaging of
an individual protein. Appl Phys Lett 101 (9):093701
Miao JW, Charalambous P, Kirz J, Sayre D (1999) Extending the methodology of x-ray
crystallography to allow imaging of micrometre-sized non-crystalline specimens.
Nature 400 (6742):342-344
Miao JW, Hodgson KO, Ishikawa T, Larabell CA, LeGros MA, Nishino Y (2003a) Imaging
whole Escherichia coli bacteria by using single-particle x-ray diffraction. Proc Natl
Acad Sci USA 100 (1):110-112
Miao JW, Ishikawa T, Anderson EH, Hodgson KO (2003b) Phase retrieval of diffraction
patterns from noncrystalline samples using the oversampling method. Phys Rev B 67
(17)
16
Miao JW, Sayre D (2000) On possible extensions of x-ray crystallography through
diffraction-pattern oversampling. Acta Crystallogr A 56:596-605
Miao JW, Sayre D, Chapman HN (1998) Phase retrieval from the magnitude of the Fourier
transforms of nonperiodic objects. Journal of the Optical Society of America A - Optics
Image Science and Vision 15 (6):1662-1669
Nelson J, Huang XJ, Steinbrener J, Shapiro D, Kirz J, Marchesini S, Neiman AM, Turner JJ,
Jacobsen C (2010) High-resolution x-ray diffraction microscopy of specifically labeled
yeast cells. Proc Natl Acad Sci USA 107 (16):7235-7239
Neutze R, Wouts R, van der Spoel D, Weckert E, Hajdu J (2000) Potential for biomolecular
imaging with femtosecond x-ray pulses. Nature 406 (6797):752-757
Nishino Y, Takahashi Y, Imamoto N, Ishikawa T, Maeshima K (2009) Three -dimensional
visualization of a human chromosome using coherent x-ray diffraction. Phys Rev Lett
102 (1):018101
Sayre D (1952) Some implications of a theorem due to Shannon. Acta Crystallogr 5
(6):843-843
Seibert MM, Ekeberg T, Maia FRNC, Svenda M, Andreasson J, Jonsson O, Odic D, Iwan B,
Rocker A, Westphal D, Hantke M, DePonte DP, Barty A, Schulz J, Gumprecht L, Coppola
N, Aquila A, Liang M, White TA, Martin A, Caleman C, Stern S, Abergel C, Seltzer V,
Claverie J-M, Bostedt C, Bozek JD, Boutet S, Miahnahri AA, Messerschmidt M,
Krzywinski J, Williams G, Hodgson KO, Bogan MJ, Hampton CY, Sierra RG, Starodub D,
Andersson I, Bajt S, Barthelmess M, Spence JCH, Fromme P, Weierstall U, Kirian R,
Hunter M, Doak RB, Marchesini S, Hau-Riege SP, Frank M, Shoeman RL, Lomb L, Epp
SW, Hartmann R, Rolles D, Rudenko A, Schmidt C, Foucar L, Kimmel N, Holl P, Rudek B,
Erk B, Homke A, Reich C, Pietschner D, Weidenspointner G, Struder L, Hauser G, Gorke
H, Ullrich J, Schlichting I, Herrmann S, Schaller G, Schopper F, Soltau H, Kuhnel K-U,
Andritschke R, Schroter C-D, Krasniqi F, Bott M, Schorb S, Rupp D, Adolph M,
Gorkhover T, Hirsemann H, Potdevin G, Graafsma H, Nilsson B, Chapman HN , Hajdu J
(2011) Single mimivirus particles intercepted and imaged with an x-ray laser. Nature
470 (7332):78-81
Shapiro D, Thibault P, Beetz T, Elser V, Howells M, Jacobsen C, Kirz J, Lima E, Miao H,
Neiman AM, Sayre D (2005) Biological imaging by soft x-ray diffraction microscopy.
Proc Natl Acad Sci USA 102 (43):15343-15346
Song CY, Jiang HD, Mancuso A, Amirbekian B, Peng L, Sun R, Shah SS, Zhou ZH, Ishikawa T,
Miao JW (2008) Quantitative imaging of single, unstained viruses with coherent x-
rays. Phys Rev Lett 101 (15):158101
Spence JCH, Qian W, Silverman MP (1994) Electron source brightness and degeneracy
from Fresnel fringes in-field emission point projection microscopy. J Vac Sci Technol A
- Vac Surf Films 12 (2):542-547
Steinwand E, Longchamp J-N, Fink H-W (2011) Coherent low-energy electron diffraction
on individual nanometer sized objects. Ultramicroscopy 111 (4):282-284
Stevens GB, Krüger M, Latychevskaia T, Lindner P, Plückthun A, Fink H -W (2011)
Individual filamentous phage imaged by electron holography. Eur Biophys J 40:1197-
1201
van Heel M, Gowen B, Matadeen R, Orlova EV, Finn R, Pape T, Cohen D, Stark H, Schmidt R,
Schatz M, Patwardhan A (2000) Single-particle electron cryo-microscopy: towards
atomic resolution. Q Rev Biophys 33 (4):307-369
Weierstall U, Spence JCH, Stevens M, Downing KH (1999) Point-projection electron
imaging of tobacco mosaic virus at 40 eV electron energy. Micron 30 (4):335 -338
Williams GJ, Hanssen E, Peele AG, Pfeifer MA, Clark J, Abbey B, Cadenazzi G, de Jonge MD,
Vogt S, Tilley L, Nugent KA (2008) High-resolution x-ray imaging of plasmodium
falciparum-infected red blood cells. Cytom Part A 73A (10):949-957
www.pdb.org Protein Data Bank.
Zuo JM, Vartanyants I, Gao M, Zhang R, Nagahara LA (2003) Atomic resolution imaging of
a carbon nanotube from diffraction intensities. Science 300 (5624):1419-1421
17
|
1209.6523 | 1 | 1209 | 2012-09-28T14:05:06 | On Surface Structure and Friction Regulation in Reptilian Limbless Locomotion | [
"physics.bio-ph"
] | One way of controlling friction and associated energy losses is to engineer a deterministic structural pattern on the surface of the rubbing parts (i.e., texture engineering). Custom texturing enhances the quality of lubrication, reduces friction, and allows the use of lubricants of lower viscosity. To date, a standardized procedure to generate deterministic texture constructs is virtually non-existent. Many engineers, therefore, study natural species to explore surface construction and to probe the role surface topography assumes in friction control. Snakes offer rich examples of surfaces where topological features allow the optimization and control of frictional behavior. In this paper, we investigate the frictional behavior of a constrictor type reptile, Python regius. The study employed a specially designed tribo-acoustic probe capable of measuring the coefficient of friction and detecting the acoustical behavior of the skin in vivo. The results confirm the anisotropy of the frictional response of snakeskin. The coefficient of friction depends on the direction of sliding: the value in forward motion is lower than that in the converse direction. Detailed analysis of the surface metrological feature reveal that tuning frictional response in snakes originates from the hierarchical nature of surface topology combined to the profile asymmetry of surface micro-features, and the variation of the curvature of the contacting scales at different body regions. Such a combination affords the reptile the ability to optimize the frictional response. | physics.bio-ph | physics | JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
On Surface Structure and Friction Regulation in Reptilian Limbless Locomotion
H. A. Abdel-Aal
Arts et Métier ParisTech, Rue Saint Dominique BP 508,
51006 Chalons-en-Champagne, France
[email protected]
Abstract
One way of controlling friction and associated energy losses is to engineer a deterministic
structural pattern on the surface of the rubbing parts (i.e., texture engineering). Custom texturing
enhances the quality of lubrication, reduces friction, and allows the use of lubricants of lower
viscosity. To date, a standardized procedure to generate deterministic texture constructs is
virtually non-existent. Many engineers, therefore, study natural species to explore surface
construction and to probe the role surface topography assumes in friction control. Snakes offer
rich examples of surfaces where topological features allow the optimization and control of
frictional behavior. In this paper, we investigate the frictional behavior of a constrictor type
reptile, Python regius. The study employed a specially designed tribo-acoustic probe capable of
measuring the coefficient of friction and detecting the acoustical behavior of the skin in vivo.
The results confirm the anisotropy of the frictional response of snakeskin. The coefficient of
friction depends on the direction of sliding: the value in forward motion is lower than that in the
converse direction. Detailed analysis of the surface metrological feature reveal that tuning
frictional response in snakes originates from the hierarchical nature of surface topology
combined to the profile asymmetry of surface micro-features, and the variation of the curvature
of the contacting scales at different body regions. Such a combination affords the reptile the
ability to optimize the frictional response.
Nomenclature
Apl
Areal
Ff
Fs
Fpl
H
R
Ra
Rku
Rq
Rsk
RT
Directions
AE-PE
RL-LL
LR
LL
LF
cross sectional area that the counter face material
established upon indenting the skin
real area of contact between the contacting region of the
reptile and the substrate
Friction force
shear component of friction force
ploughing component of friction force
Hardness
Radius of curvature
Mean arithmetic value of roughness (µm)
Profile Kurtosis parameter
Root mean square average of the roughness profile
ordinates (μm)
Profile skewness parameter
Radius of curvature in transverse direction
Anterior Posterior
Lateral Axis
Lateral right hand side
Lateral left hand side
Lateral forward,
1
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
Lateral backward.
Straight backward
Straight forward
Coefficients of friction
Differential Friction Effect
Fibril Tip Asymmetry Ratio,
Leading Body Half
Mid Trunk Section
Ratio of frictional anisotropy,
Trailing Body Half
White light Interferograms
Fibril-tip leading edge apex slant
Fibril-Tip trailing edge apex slant
Shear strength of the skin
Coefficient of friction in backward motion
Coefficient of friction in forward motion
Coefficient of friction for the trailing half of the skin
Coefficient of friction for leading half of the skin
LB
SB
SF
Acronyms
COF
DFE
FTAR
LBH
MTS
RFA
TBH
WLI
Greek symbols
ΘL
ΘT
τ
μB
μF
μT.H
μL.H
Introduction
The ultimate goal of surface customization for rubbing applications is to improve lubrication,
reduce friction losses, and to minimize (or eliminate if possible) mass loss due to wear and
friction-induced structural degradation in general. The design considerations for a surface
depend on the particular tribological situation. In a lubricated surface, for example, it is desirable
to alter the topography of the surface so that a full hydrodynamic regime is established within a
short distance from the entrance of the lubricant to the rubbing interface (Ferguson and
Kirkpatrick 2001). This leads to establishing complete separation of the rubbing surfaces early
on in rubbing. Controlled adhesion may be a goal of surface structuring. Additional tribological
design targets may be to establish anisotropic friction for motion control (e.g. for reduction of
locomotion costs in rescue robots), or to control the wettability of a surface for enhanced
lubricity or self-cleaning purposes (Thor et al., 2011).
A structured surface for enhanced tribo-performance should posses several advanced features.
One principal feature is the ability to tune the frictional response upon rubbing. That is, the
surface should be able to adapt its’ frictional profile in response to sensed changes in sliding
conditions (e.g. changes in texture of the mating surface, variation in contact pressure, etc.). The
tuning requirement may stem from geometry of the surface topographical building blocks, their
distribution and placement within the surface, presence of embedded sensory, or a combination
of all these factors. One of the difficulties in engineering such a surface, despite the availability
of several enabling technologies, is the current limited understanding of the interaction between
deterministic surface textures and frictional response. This, in turn, is due to the relatively recent
history of deterministic surface texturing in human engineering.
While the technical world lacks diverse examples of functional-self adapting tribo-surfaces, our
surroundings contain an abundance of examples of hierarchically structured naturally occurring
surfaces capable of delivering super functionality. These may provide inspiration for surface
2
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
designers. The richness and diversity of the examples provided by natural surfaces are worthy of
study to extract viable solutions for surface design problems encountered in the technical world.
This is particularly feasible since the natural world obeys the same physical laws that govern the
behaviour of engineering systems. As such, any extracted design rules should, in principle, be
valid across both realms: the natural and the technological.
An order of species that manifest an interesting interaction between micro-structural surface
features and frictional requirements for locomotion is that of snakes. Snakes belong to the
serpents order within the Squamate Reptiles clade. Squamata (scaled reptiles) is a large order of
reptiles of relatively recent origin. The order is distinguishable through the scales that are born
on the skin of members of the order. Squamata comprises two large clades: Iguania and
Scleroglossa. The later comprises 6,000 known species, 3100 of which are “lizards,” and the
remaining 2,900 species as “snakes” (Vitt et al., 2003). Snakes contain diverse examples where
surface structuring, and modifications through submicron and nano-scale features, achieve
frictional regulation manifested in: reduction of adhesion (Arzt., et al 2003) abrasion resistance
(Rechenberg, 2003), and frictional anisotropy (Hazel et al, 1999). They are found almost
everywhere on earth. Their diverse habitat presents a broad range of tribological environments.
Diversity in habitat requires adaptable features capable of efficient performance within the
particular environment. Thus, a snake species particular to the desert, for example, would entail
distinct features tailored to function within an abrasive sliding environment (Klein, et. al., 2010).
The same would apply to a snake that roams a tropical forest where essential functional
requirements differ from those dominant in a desert environment (Jayne and Herrmann, 2011).
Function specialization requires analogous specialization in the composition, shape, geometry
and mechanical properties of the skin. However, since the chemical compositional elements of
reptilian skin are almost invariant within the particular species the study of functional
specialization within a given species becomes more intriguing. This is because; invariance of
chemical composition implies that functional adaptation takes place through adaptation of form,
geometry and metrology of the skin building blocks. Such implications provide a venue to scour
the customized surface features within the particular species to extract surface design lessons
suitable for the technical world. Many, therefore, studied appearance and structure of skin in
snakes Squamata (Hazel et al, 1999, Vitt et al 2003, Chang et al, 2009, Alibardi and Thompson,
2002, Ruibal 1968, Chiasson et al, 1989, Jayne, 1988, Scherge and Gorb 2001, Rivera et al,
2005). Furthermore, attracted by legless locomotion, others studied the tribological performance
of snakes (Berthe et al, 2009, Saito et al, 2000, Shafei and Alpas, 2008, 2009). The results
emphasized the role that diverse ornamentation actively contributes in the dynamic control of
friction and regulation of locomotory energy consumption (Shafei and Alpas, 2008, 2009, Abdel-
aal et al, 2010, Abdel-aal and El Mansori, 2011, Gray 1946).
In previous work Abdel-aal and co-workers (Abdel-aal, et al, 2011, 2012), reported the dynamic
friction coefficient for the skin of a Python regius. The results confirmed the anisotropy of the
friction of the reptile. The COF in forward motion (i.e., with the grain of the skin (caudal
direction)) was less than that measured in backward motion (against the grain of the skin (cranial
direction)). A similar trend emerged from measurements obtained in diagonal motion in both the
forward and the backward directions while not reflected in measurements pertaining to lateral
motion. The data suggested that such a friction differential effect stems from the geometry of the
surface. In particular, the asymmetric profile of the individual micro-fibrils present on the
ventral scales correlated to the anisotropy of the COF. Moreover, the metrological parameters of
3
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
the surface (both macro and micro-scaled) showed a non-uniform distribution along the Anterior-
Posterior axis of the reptile.
The variation in the metrological and geometrical parameters of the surface, in theory, affects the
mechanics of contact between the ventral surface of the reptile and the substrate. Consequently,
different locations on the body of the reptile will show varying frictional profiles. Additionally,
due to the stocky build of the reptile, the distribution of body mass per-unit-length along the AE-
PE axis of the body is non-uniform. Now, if the skin of the snake has a constant COF, then the
non-uniform mass distribution will affect the frictional tractions and the friction-induced losses
accommodated through the skin. That is, the irregular mass distribution will induce an
analogous, locally variable, friction force distribution along the AE-PE axis. Such an irregularity
should affect the structural integrity of the skin and on the energy consumed by the reptile to
generate and maintain motion. However, observations in nature and in experimental work
indicate that there is a distribution to the frictional tractions along the body of snakes (Berthe, et.
al., 2009). In addition, the irregular distribution of frictional forces does not compromise the
structural integrity of the skin. This is partly due to the ply-like skin structure (Klein, et. al.,
2010). Interestingly moreover, the energetic cost of legless locomotion is found to be equivalent
to that of running by limbed animals of similar size (Walton et., al., 1990). This implies that the
COF of the skin varies locally and that an analogous distribution of the metrological parameters
of the skin compensates the variation in the distribution of the frictional forces. In other words,
the hierarchy of the textural features of the skin act as a control mechanism that “fine-tunes” the
frictional response of the skin through modifying the contact between the reptile and substrate. In
this sense, the skin of a snake not only would accommodate tractions (through its mechanical
response) but also would actively control friction through texturing (micro-scale fibril elements).
This is in contrast to recent explanations (Goldman and Hu, 2010; Hu et. al., 2009) that the
overlapping arrangement of the ventral scales is the origin of frictional control and locomotion.
This hypothesis, if validated, should contribute to linking the various textural elements (observed
on snakeskins) to their tribological function. This, in turn, will help correlating the various
shapes distributions, arrangements, of the micro-fibrils often observed on ventral scales of all
snakes (Schmidt and Gorb, 2012) to their tribological environments and frictional response of the
particular species. Such a correlation should enhance our understanding of the interaction
between surface texturing and friction control in particular environments, thereby advancing our
knowledge of surface engineering. To date, however, a study that examines such a hypothesis is
non-existent despite its direct relation to many tribological problems, of fundamental nature, that
relate to intrinsic control of friction and surface engineering.
The goal of this study is to compare local frictional behaviour of the skin to textural make up.
Therefore, we investigate the validity of our hypothesis concerning the relation of surface
texturing to frictional control in snakes. Namely, we attempt to answer the question of whether
the COF for snakeskin is a property of the skin (whence a constant as implied in classical
tribology) or rather it is a consequence of skin composition and particular texturing of the
surface. Further, we investigate the correlation between surface geometry and the local variation
in the COF along the body of the reptile. In principle, the present study is an extension of earlier
studies by the author and co-workers. However, the current work comprises some fundamental
differences that distinct the findings. Our earlier work (Abdel-aal, et al, 2012) stemmed from the
premise that the COF of the skin is material property and therefore is a constant. Such an
assumption implied that the geometry of the skin does not contribute to any functional
adaptation. In addition, no attempt was made previously to link the frictional behaviour to the
4
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
metrological features of the surface. Consequently, extrapolation of the findings to deduce
design rules for technological surfaces was rather difficult. In the current work, however, the
major assumption is that the skin of the snakes contributes to local adaptation through variation
in micro-geometry. A consequence of such an assumption is that observed frictional behaviour
of the snake, from a tribological point of view, is no longer a mere function of morphological
traits (i.e., muscular activity). Rather, morphology and surface micro-design features form an
integrated system of optimized, and adaptable, tribological function. Moreover, linking the
frictional response to geometry should facilitate the deduction of design rules for technological
surfaces especially that the description of surface topography is based on technological
standards. This should facilitate the transfer of design ideas from the biological domain to the
technological domain.
To simplify the presentation, without losing generality, we compare the metrological
characteristics and the frictional behaviour for two locations on the skin of the reptile. The first
location is representative of the leading half of the reptile, whereas, the second represents the
trailing half. For brevity, we focus on presenting the metrological parameters that directly affect
the frictional behaviour of the skin as implied from our preliminary study (Abdel-aal et al., 2012,
2011).
The manuscript comprises two parts. The first provides a comparison between the surface
structure, geometrical and textural metrology on both sides of the skin. The second part,
meanwhile, presents data pertaining to the frictional behaviour of each half of the skin and a
comparison of the general trends emerging for each half. In addition, part two of the manuscript
provides a correlation between the metrological parameters of the surface and the frictional
behaviour determined in this work.
2. Anatomy of Snake skin
The skin of a reptile comprises two basic strata: the “dermis” and the “epidermis”. The dermis is
deeper than the epidermis. It is composed mainly of connective tissue. The epidermis contains
an abundance supply of blood vessels and nerves. However, it does not have blood supply of its
own. This renders the living cells, contained within this layer, depending in their nourishment on
diffusion from capillaries in the dermis layer.
The epidermal layer in a snake entails seven layers. These are organized in plies of cells with
tight packing. The epidermal layer (sub-layers included) encases the body of the retile to form an
outer shield. Figure one details the seven layers present within the epidermis. Described from
the inside of the skin, the first layer is the “stratum germinativum”. This is the deepest layer of
the skin. It is lined with rapidly dividing cells. Six additional sub-layers, again from the inside
of the skin, follow. Together the six sub-layers form a so-called “epidermal generation” (old and
new skin layers). Thus, stacked above the stratum germinativum, there exists: the clear layer and
the lacunar layer. The lacunar layer matures in the old skin layer as the new skin is growing
beneath. Following there is the α−layer, the mesos layer and the β-layer. The mesos layer is
similar to the human stratum corneum (Fraser and Macrae, 1973) and contains several layers of
flat and extremely thin cells surrounded by intercellular lipids (Lillywhite and Maderson, 1982).
These three layers consist of cells that become keratinized with the production of two types of
keratin: α (hair-like) and β (feather like). Keratinisation continuously transforms these cells into
a hard protective layer. Finally, there is the “oberhautchen” layer, which forms the toughest outer
most layer of keratinized dead skin cells.
The oberhautchen layer contains the fine surface structure known as the micro-ornamentation
(Meyers et al, 2008). Before the molt, a new layer of epidermis forms under the currently
5
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
existing one; the two layers are zipped together by a spinulae structure (Alibardi and Toni,
2007). During the molt, the reptile sheds the outer (older) layer of the epidermis. The principle
constituents of snakeskin are keratin fibres (Toni et al; 2007). Keratinized-cells constitute the
outer part of the skin. The process of keratinisation consists in synthesizing keratins that will
potentially form the keratin fibres. The keratinisation brings an increase of keratin production
from the cells that start to begin platter before dying (Ripamonti et al., 2009).
Two types of keratins form the epidermis: the α-keratin (which in a snake is acid or neutral) and
the β keratin (which in a snake is basic). The β-layer consists mainly of β-keratin; this type of
keratin is not present in other layers of the skin. Alpha α-keratin constitutes most of the
epidermal layers and it contributes to the mechanical properties of skin cells (Maderson 1985).
Obe rhau tchen
Be ta ( ) -laye r
β
mesos laye r
a lpha ( ) laye r
α
lacuna r t issue
clea r laye r
oute r gene rati on
laye rs
Dermis
Ep i-derm is
Obe rhau tchen
Be ta (
- laye r
β)
mesos laye r
a lpha laye r
stra tum
ge rmina t ivum
in ne r gene ra tio n
laye rs
Figure 1: General structure of the epidermis of a squamate reptile. The figure depicts the
“outer” generation layer which is the layer about to be shed; and the “inner”
generation layer which is the new replacement skin layer.
The oberhautchen consists mainly of β-keratin. The presence of the two different types of
keratinaceous protein α and β distinguishes the reptilian epidermis from its’ mammalian
counterpart (Fraser and Macrae, 1973). The shed epidermis of snakeskin consists of four layers:
the outermost Oberhautchen, the β-layer (mainly protein), the mesos layer (lipid-rich), and the
inner α-layer (mainly protein). The oberhautchen consists of a particular type of β-cells that play
a major role in the shedding process. Together, this layer and the β-layer, both containing β-
keratins, are considered as a unique β-layer in the mature epidermis. The oberhautchen layer is
the outermost ply within the epidermal layers. It contains the micro-textural ornamentation. It is
also the layer which in direct contact with the surroundings. That is it is the most active layer of
the skin in the sense that it simultaneously accommodates contact and frictional effects.
3. Materials and Methods
3.1 Skin treatment
All observations reported herein pertain to shed skin obtained from five male Ball pythons
(Python regius). All the received shed skin was initially soaked in distilled water kept at room
temperature for two hours to unfold. Following soaking, the skin was dried using compressed air
and stored in sealed plastic bags. Note that the exuvium surface geometry of shed epidermis
does not differ from that of a live animal (Klein et al, 2010, Klein and Gorb, 2012). Therefore,
using shed skin to characterize of the skin contribution to the frictional response should not
affect the quality of the results.
6
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
3.2 surface texture metrology
Evaluation of surface texture metrology utilised a white light interferometer (WYKO 3300 3D
automated optical profiler system). Analysis of all resulting White light Interferograms, WLI, to
extract the surface parameters used two software packages: Vision ®v. 3.6 and Mountains® v 6.0.
To determine the metrological features of the skin, we identified three major regions on the hyde
of each of the studied snakes, these are shown in figure 2. The first region is located at the mid-
section of the reptile. It is about 20 cm long and is the stockiest portion of the trunk (contained
in the dashed rectangle in the figure). This was termed the Mid Trunk Section (MTS).
Trailing Body half
Leading Body half
Last Vent ra l
LV
Tail
m id trunk section
FV-LV= 1220 mm
Fi rst ventral
FV
Head
Active Body Length
Figure 2 positions chosen on the snake shed skin for metrological characterization.
The remaining portion of the active length of the skin was then divided in two parts roughly
equal in length (L= 47 cm). The portion of the skin extending from the first ventral scale (point
FV in figure 3) to the right hand side boundary of the MTS was labeled as the Leading Body
Half (LBH). The portion of the skin extending from the left hand side of the MTS to the Last
Ventral scale (point LV in figure 2) was labeled the Trailing Body Half (TBH).
For each of the skin halves, we recorded fifty WLIs at randomly selected points within the
particular half of the skin. These were further analyzed to extract the textural metrological
parameters. In this work, we did not examine the MTS since we considered its geometry an
anomaly with respect to the rest of the body. However, work currently in progress is comparing
the makeup and friction behavior of this section to the rest of the body.
3.3 Friction measurements
All friction measurements utilized a tribo-acoustic probe, which is described elsewhere All
measurements utilized a patented bio-tribometer (Zahouani et al., 2009). The device includes a
tribo-acoustic probe that is sensitive to the range of friction forces and the acoustic emission
generated during skin friction. It is also capable of measuring normal and tangential loads and of
detecting sound emission due to sliding. The probe comprises a thin nitrocellulose spherical
membrane, 40 mm in diameter, with a thickness of 1 mm. The probe material has a Young’s
modulus of 1 GPa. The roughness of the probe, Ra, is 4 μm and the mean value between peak to
valley, Rz, is 31 μm.
In all frictional tests, the skin was stationary and the tribo-probe was moving at an average speed
of 40 mm/s using a normal force of 0.4 (±0.05) N. The skin used in measurements consisted of
150 mm long patches taken from four locations on the ventral side of the shed skin. Skin
samples did not receive any chemical or physical treatment beyond the water-assisted unfolding
procedure described in the previous section.
The skin used in measurements consisted of 150 mm long patches taken from four locations on
the ventral side of the shed skin, two from the leading half of the skin and two from the trailing
7
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
half of the skin. Skin samples did not receive any chemical or physical treatment beyond water
unfolding.
Norma l load
Ft
Fn
measu remen t
p robe
examined sk in
Sil icone Substrate
Figure 3: sample setup of the skin sample and the tribo-acoustic probe used in measurements .
To mimic the effect of the body of the snake on the skin, before starting an experiment, the
particular skin patch was placed on a rectangular elastic pad of dimensions length L= 200 mm
,width W=100 mm and thickness of approximately 4 mm (figure 3). The pad is made of silicone
rubber (Silflo®™, Flexico Developments Ltd., Potters Bar, UK). Table 2 provides a summary
of the pad material properties.
Measurement of the friction forces proceeded along the two major body axes: the anterior-
posterior axis (AE-PE) and the lateral axis (LL-RL) (see figure 4-a). In addition, we performed
measurements along the diagonal directions shown in figure 4-a. For each direction,
measurements were taken in the forward and backward directions. Figure 4, depicts the sense of
forward and backward in relation to the motion of the reptile first in a global sense (figure 4-b),
and second as it applies locally n the ventral scale (figure 4-c). To facilitate the description of
the results we provide table three, which describes the measurement directions on each of the
examined skin halves in vector form.
Table 2: Summary of geometric dimensions and mechanical properties of elastic pads used to
cushion skin in experiments.
Geometry
Rectangle
Length (mm) x Width (mm)
Material
Silicone Rubber (Silflo ®, Flexico Developments
Potters Bar, UK)
Young’s Modulus (E) MPa
Poissons Ratio
Stiffness (K)
Mechanical Properties
Table 3: Summary of vectors representing direction of friction measurements
Vector Designation
Measurement Direction
Trailing half
Principal Directions
PE→C.O
Caudal
Cranial
PE→C.O
Leading Half
C.O →AE
C.O →AE
200 x 100
2 @20 C
0.3
300 N/m
8
LL → RL
Dextral
RL → LL
Sinistral
Diagonal Directions
C.O → AE-LL
Dextro-Caudal
AE-LL → C.O
Sinistro-Cranial
C.O → AE-RL
Sinistro-Caudal
Dextro-Cranial
AE-RL→ C.O
D irection of Frictiona l
Mesaurements
AE
Anter ior E nd
AE-LL
F
r
o
n
t
h
a
l
f
DX
AE-RL
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
RL → LL
RL → LL
PE-RL→ C.O
C.O → PE-RL
PE-RL→ C.O
C.O → PE-LL
Snake Body Movement
re lative to Substrate
CD
SN
CR
)
D
C
(
l
a
d
u
a
C
F
r
o
n
t
h
a
l
f
Lef t
(later a l s id e)
LL
C.O
R ig ht
(Later a l S id e)
RL
Sin istra l (SN)
C.O
Dextra l (DX)
PE- LL
R
e
a
r
h
a
l
f
PE-RL
R
e
a
r
h
a
l
f
C
r
a
n
i
a
l
(
C
R
)
PE
P os ter ior E nd
Figure 4 Description of the axes-used to define directions of frictional measurements on the skin
of the reptile (axes are defined in table 3in vector form)
4. Metrological Characterization
The results of our preliminary study (Abdel-aal et al., 2012) identified parameters pertaining to
surface asperity height, asperity distribution and form as primary metrological quantities. In this
work, therefore, we will limit the presentation of the metrological aspects of the shed skin to
those parameters.
4.1 Small Scale Metrology
As described earlier, in section 3, initial metrological characterization of the skin took place by
generating White Light Interferograms (WLI) of selected patches within the ventral scales of the
shed skin. Figure 5 depicts two of such WLIs. The interferogram shown as figure (5-a) depicts
the overall topography of a ventral skin patch located within the leading half of the skin, whereas
that shown in figure 5-b, details a patch located within the trailing half of the skin. Processing
each of the interferograms provided profile data along the directions used for frictional
9
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
measurements. The remaining plots, within figure 5, depict the extracted roughness profiles in
the following directions: AE-PE-Axis (figure 5-c, d), RL-LL-Axis (figure 5-e, f), AE-RL-PE-
LL-Axis (figure 5-g, h), and AE-LL-PE-RL-Axis (figures 5-i, j). Note that the profiles presented
in figures 5-c: 5-j) represent roughness along a line and not an area. The scale to the left of
figures 5 c-j differs from that to the right of figures (5-a and 5-b). The numbers on the former
represent the heights and depths of the surface protrusions with respect to a reference line
(whence the positive and negative values). The numbers to the right of figures 5-a and 5-b
represent absolute height of surface points (i.e., height is referred to the lowest point on the
surface). The maximum peaks and valleys of the surface roughness, irrespective of the direction
of profile extraction, does not exceed two microns (i.e.,-2 μm ≤ h ≤ 2μm). Additionally, on
average, the differences in heights between the leading and the trailing halves of the skin are not
pronounced. This observation, however, is rather deceptive as the statistical roughness height
parameters, Ra (average roughness) and Rq (root mean square height parameter) shows some
variation both with respect to the direction of profile extraction and with respect to the location
of the examined patch on the ventral side (leading or trailing).
μm
3.5
100
120 μm
60
80
20
40
0
0
20
40
60
80
100
120 μm
0
10
20
30
40
50
60
70
80
90
μm
0
10
20
30
40
50
60
70
80
90
μm
3
2.5
2
1.5
1
0.5
0
NM
3 0 μ m
3 0 μm
μm
3
2 .75
2 .5
2 .25
2
1 .75
1 .5
1 .25
1
0 .75
0 .5
0 .25
0
µm
1
0. 5
0
-0. 5
-1
-1. 5
µm
c
0
10
20
30
40
50
60
70
80
9 0 μm
20
30
40
50
60
70
80
90
100
110 120 µm
20
30
40 50
60
70 80
90 100 110 120 130 140 µm
1
0. 5
0
-0.5
-1
-1.5
0
µm
e
10
g
10
0
1
0. 5
0
-0. 5
-1
-1. 5
µm
1
0. 5
0
-0. 5
-1
µm
1.5
1
0.5
0
-0.5
-1
µm
1. 5
1
0. 5
0
-0. 5
-1
d
0
10
20
30
40
50
60
70
80
90 µm
f
0
10
20
30
40
50
60
70
80
90
100 110
120 µm
h
10
0
20 30
40 50 60 70
80 90 100 110 120 130 140 µm
10
µm
2
1.5
1
0.5
0
-0. 5
-1
-1. 5
I
0
10 20 30
40 50 60 70 80 90 100 110 120 130 140 150 µm
µm
1.5
1
0.5
0
-0.5
-1
-1.5
-2
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
j
0
10 20 30
40
50
60
70
80 90 100 110 120 130 140 µm
Figure 5 Roughness profiles along the various axes of frictional mesaurements. The left hand
side (figures 5-a, c, g, and i) depicts profiles representative of the leading half of the
skin and the right hand side (figures 5-b, d, h, and j) depicts profiles representative of
the trailing half of the skin.
This variation is illustrated in figure 6 (a-d). The figure depicts bar plots of two roughness
parameters extracted for each of the profiles shown in figure 5 (c through j). The plotted
parameters are the mean arithemetic roughness parameter Ra (figure 6- a and 6-b); and the root
mean square average roughness parameter Rq (figures 6-c and 6-d). The extracted parameters are
plotted for the leading and the trailing half of the skin for ease of comparison. Examination of
the data reveal both parameters, Ra and Rq, have similar magnitude ranking. Namely, the highest
values of these parameters pertain to the profiles located along the pricncipal axes (the Anterior
poterior, AE-PE, and the lateral, LL-RL, axes). Values for profiles located along the diagonal
axes are smaller than those along the principal axes. The values of the roughness parameters
vary by location on the body of the reptile. For example, values of the roughness parameter Ra
on the leading half of the skin are, in general, smaller than their counterparts on the trailing half
of the skin (compare the values in figure 6-a to those in figure 6-b). Values of the mean square
roughness Rq reflect a trend similar to that of Ra (compare values in figure 6-c to values in
figure 6-d).
4.2 Shape of fibril tips
The microstructure of the ventral scales constitutes waves of micron-sized fibrils (Abdel-Aal et
al., 2011). The shape of the tips of individual fibrils influences the frictional behavior of the skin.
Hazel and co-workers (Hazel et al., 1999) suggested that the spherically asymmetric shape of the
tips is the origin of the anisotropic frictional behavior they observed in their investigation of the
skin of a Boa constrictor. This suggestion highlights the importance of characterizing the shape
of the fibrils and the relation of that shape to frictional behavior. In this work, we use two
metrics to characterize the shape of the fibril tips. The first is the extraction of the projection of
the topography of a single fibril row in all directions of interest from WLI. The second is to map
the profile kurtosis parameter Rku in all directions of interest.
4.2.1 Projection of Fibrils
Figure 7 (a-d) shows the extracted profiles of a single fibril row. The orientation of all figures is
inversed with respect to the natural position of the fibrils on the ventral scales and the position of
the ventral side of the reptile during motion. Each profile shown in figure 7 is a plot of the
projection of the fibril-tip in the respective direction. The apex of the profile is the point (or arc)
that will contact the substrate when the reptile moves. Thus, in each of the figures, the top of the
plot represents the relative position of the plane containing the contacting terrain with respect to
the fibril plane. Based on this orientation the edges of a fibril are designated as “leading” or
“trailing”, and the motion is designated as “forward” and “backwards”.
Figure 8 shows that fibril tips have an asymmetric profile. Moreover, a common theme to all
examined fibril rows is asymmetry of slopes. The slope of the trailing edge of a fibril is steeper
11
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
than the slope of the leading edge. The degree of profile asymmetry, however, is not constant.
Rather, it varies with the direction of examining the profile. The variation of asymmetry seems
to play an important role in determining the frictional profile of the reptile. To this effect, it is
necessary to develop a quantitative measure to characterize the extent of fibril tip asymmetry.
For such purpose, we define two angles ΘL and ΘT. These angles denote the leading and the
trailing edge apex angles of the fibril tips. Table three, presents a summary of the trailing and
leading apex angles in all examined directions. The Values are in degrees. The last column
within the table gives the ratio of the leading to the trailing apex angles ΘL /ΘT. The values of
the angles confirm the observation that the leading angle is greater than the trailing angle in all
directions.
0 .25
a
0 .25
b
Tra i ling Ha l f o f Skin
Front Ha lf of Skin
a
R
r
e
t
e
m
a
r
a
P
s
s
e
n
h
g
u
o
R
e
g
a
r
e
v
A
e
c
a
f
r
u
S
0 .20
0 .15
0 .10
0 .05
0 .00
0 .25
E
P
-
E
A
L
L
-
E
A
-
L
R
-
E
P
L
L
-
E
P
-
L
R
-
E
A
Fron t Hal f of Skin
L
L
-
L
R
c
a
R
r
e
t
e
m
a
r
a
P
s
s
e
n
h
g
u
o
R
e
g
a
r
e
v
A
e
c
a
f
r
u
S
0 .20
0 .15
0 .10
0 .05
0 .00
0 .25
L
L
-
L
R
E
P
-
E
A
L
L
-
E
A
-
L
R
-
E
P
L
L
-
E
P
-
L
R
-
E
A
d
Rear Hal f of Skin
s
s
e
n
h
g
u
o
q
R
R
e
r
e
r
a
t
e
u
m
q
S
a
r
n
a
a
P
e
t
M
h
g
t
i
o
e
o
H
R
e
c
a
f
r
u
S
0 .20
0 .15
0 .10
0 .05
L
L
-
L
R
0 .20
0 .15
0 .10
0 .05
E
P
-
E
A
L
L
-
E
P
-
L
R
-
E
A
L
L
-
E
A
-
L
R
-
E
P
s
s
e
n
h
g
u
o
q
R
R
e
r
e
r
a
t
e
u
m
q
S
a
r
n
a
a
P
e
t
M
h
g
t
i
o
e
o
H
R
e
c
a
f
r
u
S
0 .00
0 .00
Figure 6 Distribution of the average and root mean square roughness along the axes of
measurements for the leading and the trailing halves of the skin. Figures 6-a and 6-b
depict distribution of Ra representative of the leading and the trailing halves of the
skin respectively. Figures 6-c and 6-d depict distributions of Rq representative of the
leading and the trailing halves of the skin respectively (error bars are ± SD, values
are significantly different One way ANOVA, p< 0.001)
4.2.2 Profile kurtosis
The kurtosis is a measure of the “peakedness” or “roundness” of the distribution of the asperity
heights (the fibril tips in this work). It measures the number of surface peak measurements that
significantly vary from the mean of the heights. High kurtosis values (Rku > 3) indicate a surface
12
L
L
-
E
P
-
L
R
-
E
A
L
L
-
E
A
-
L
R
-
E
P
E
P
-
E
A
L
L
-
L
R
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
with a very wide distribution of surface heights, with many high peaks and low valleys (a so-
called spiky surface). A low value (Rku < 3) meanwhile implies a surface that is relatively flat,
with the majority of the asperity heights close to the mean (a so-called bumpy surface). For a
Gaussian (perfectly random) surface, the kurtosis parameter Rku is equal to three (Whitehouse,
1994). The value of the kurtosis of a surface affects the friction force developed during sliding.
In particular, the contact loading is directly proportional to the value of the kurtosis (Tayebi and
Polycarpou, 2004). When the kurtosis of a surface is high, more asperities establish contact with
the counter-face body. This increases the real area of contact and, in turn, increases the frictional
force. At very low values of the kurtosis, adhesion dominates especially in relatively smooth
surfaces (Liu et al., 1998). Figure 8, presents a plot of the kurtosis parameter Rku in the directions
examined in figure 5. Figure (8-a) presents the kurtosis values for the profiles located within the
leading half of the skin whereas, figure 8-b depicts the Rku values for the trailing half of the skin.
µm
b
µm
0.4
0.4
Trailing edge
Leading Edge
le ading edge
Fo rwa rd m ot ion
trail ing e dge
Backwa rd mo ti on
AE-RL
a
0
1
2
3
4 µm
Leading edge
Trailing edge
Forwa rd m otion
Backwa rd m otion
0.35
0.3
0.25
0.2
0.15
0.1
0.05
µm
0 .5
0 .4
0 .3
0 .2
0 .1
Backward mot ion
Forward mo tion
AE-LL
0
d
1
Leading edge
2 µm
Trailing edge
Forwa rd m otion
Backward mo tion
0.3
0.2
0.1
µm
0. 6
0. 5
0. 4
0. 3
0. 2
0. 1
RL-LL
c
AE-PE
0
0
3 µm
2
1
4 .5 µm
4
3 .5
3
2 .5
2
1 .5
1
0 .5
0
Figure 7 Fibril tip profile along the different axes of measurements. Note that “forward” and
“backward” directions of motion depend on the projection of the fibril raw with respect
to the reptile axis (refer to table 3 for detailed description of directions of motion)
13
JMBBM-D-12-00373R2
Journal of the Mechanical Behavior of Biomedical Materials
Table 3 Definition and summary of fibril tip angles.
Leading edge
Angle
lΘ
Trailing edge
Angle ΘΤ
Forwa rd mo tion
Bac kward mo tion
e
g
d
g e
din
a
e
Fibril L
F
i
b
r
i
l
T
r
a
i
l
i
n
g
e
d
g
e
Direction
AE-RL
AE-LL
RL-LL
AE-PE
ΘL (deg)
21.65
16.4
14.4
21.25
ΘT (deg)
14.8
17.2
11.6
13.3
ΘL /ΘT
1.46
0.95
1.24
1.6
The plots imply that values of the Rku parameter depend on the orientation of the particular
profile. Moreover, the magnitude ranking of the values does not display a consistent order. For
example, within the leading half of the skin (figure 8-a), the smallest kurtosis value pertains to
the profile along te AE-PE-Axis. This, however, is not the case within the trailing half of the
skin, figure 8-b, where the smallest value pertains to one of the diagonal profiles (along the
dextro-cranial direction AR-PL). Similarly, the largest kurtosis value within the leading half is
that of the diagonal profile AL-PR (sinistro-caudal axis), whereas, the largest value within the
trailing half is that of the lateral profile (LR-LL).
Values of the kurtosis within the leading half of the skin are, in general, less than three (the cut-
off value for complete Gaussian height distribution) except for the sinistro-cranial profile. In
contrast, within the trailing half of the skin, the kurtosis values are greater than three (Rku > 3)
except for one diagonal profile for which the value is very close to three. The data of figure 8
imply that fibril-tip heights within the trailing half have a random distribution. For the leading
half, moreover, the kurtosis values fall within the interval (2.25 ≤ Rku ≤ 3.25). This implies that
the leading half of the skin is generally more flat than trailing half. As such, other factors being
the same, friction of the leading half would entail higher adhesion contribution than the trailing
half.
4.3 Form and curvature
Surface topography profiles presented in figure 5 are a superposition of two components: the
basic form of the skin surface (the so-called deterministic component of roughness), and the
rugosity (the so-called stochastic component of roughness). Separation of the form component
provides information about the periodicity of the basic surface constituents and about the
symmetry of the surface structural elements. This information, in turn, contributes toward
understanding the kinematics of the surface during locomotion. One way of looking at form
information, is considering contact of mating surfaces. Form establishes the global geometry of
the contact area (ellipse, circle, etc.,). Rugosity, however, modifies this shape (e.g, by inducing
deviation from basic shapes or causing the discontinuity of the contact spot). Figure 10 depicts
14
|
1709.04019 | 2 | 1709 | 2018-04-18T13:50:21 | Sound emitted by some grassland animals as an indicator of motion in the surroundings | [
"physics.bio-ph"
] | It is argued based on the results of both numerical modelling and the experiments performed on an artificial substitute of a meadow that the sound emitted by animals living in a dense surrounding such as a meadow or shrubs can be used as a tool for detection of motion. Some characteristics of the sound emitted by these animals, e.g. its frequency, seem to be adjusted to the meadow density to optimize the effectiveness of this skill. This kind of sensing the environment could be used as a useful tool improving detection of mates or predators. A study thereof would be important both from the basic-knowledge and ecological points of view (unnatural environmental changes like increasing of a noise or changes in plants species composition can make this sensing ineffective). | physics.bio-ph | physics |
Sound emitted by some grassland animals as an indicator of
motion in the surroundings
M. Pietrow∗1, P. S(cid:32)lomski2, B. Tkaczyk3, G. Grzywaczewski4, J. Malinowski3, P. Konkol1
1Institute of Physics, M. Curie-Sk(cid:32)lodowska University, ul. Pl. M. Curie-Sk(cid:32)lodowskiej 1,
20-031 Lublin, Poland
2Geographic Information Systems Development Company Martinex, ul. Me(cid:32)lgiewska 95,
21-040 ´Swidnik, Poland
3Olympic special-interest group for young physicists in Lublin of P. Kononowicz,
ul. B. Pa´snikowskiego 6, 20-707 Lublin, Poland
4Department of Zoology, Animal Ecology and Wildlife Management, University of Life
Sciences in Lublin, Akademicka 13, 20-950 Lublin, Poland
October 1, 2018
1 Abstract
It is argued based on the results of both numerical modelling and the experiments performed on an
artificial substitute of a meadow that the sound emitted by animals living in a dense surrounding such
as a meadow or shrubs can be used as a tool for detection of motion. Some characteristics of the sound
emitted by these animals, e.g. its frequency, seem to be adjusted to the meadow density to optimize the
effectiveness of this skill.
This kind of sensing the environment could be used as a useful tool improving detection of mates or
predators. A study thereof would be important both from the basic-knowledge and ecological points of
view (unnatural environmental changes like increasing of a noise or changes in plants species composition
can make this sensing ineffective).
2 Motivation
A sound emitted by most animals can be heard by them [1] and could be used in many ways [2 -- 7].
In particular, it is interesting to consider the significance of a voice emitted by creatures living in a grass-
land or shrubbery. An intriguing question arises whether a long-lasting and monotonous sound emitted by
some orthopterans like crickets, cicadas or some birds like a corncrake could be designed for any purpose
and, if yes, whether it requires that the sound should have some special characteristics.
The range of vision is restricted and not effective in the grass environment. Receiving a sound feed-back
from the surroundings can potentially play an important role in environment sensing.
Orthopterans are one of the noisiest animals living in grass and shrubs. Studies on their acoustic skills
(both emission of the sound and hearing) are divided into at least into two branches: neurobiological
[8 -- 12]
and the behavioural ones. From the latter view point, cricket's sound is identified mostly as mate attrac-
tion (phonotaxis) [13 -- 16]. It is known that crickets can recognize the direction from which the sound
comes [17]. It also seems that some performance is preferred as a steering signal
[18, 19]. Crickets are
able to adjust their own sound to the sound emitted by neighbours [20]. It is not clear yet if the sound
emission by crickets has other behavioural meaning as, for example, the predator detection [21]. To our
knowledge, the only known auditory behaviour of crickets against predators is some reaction to stimula-
tion by ultrasounds emitted by insectivorous bats [22, 23]. A wide review of the subject of hearing and
∗corresponding author; email: [email protected]
1
Figure 1: Frequency spectrum of sound reflected from a 1×1 m2 wall made of cane blades exposed to one
of the given tones from a range of 0.5 kHz -- 15 kHz (the abscissa in a logarithmic scale). The frequency of a
stimulated tone is indicated by a tag near a respective peak present also in this reflected sound spectrum.
emission of the sound by insects can be found in [1, 13, 23 -- 25]. In all described cases, the sound of crickets
is loosely related to the physical properties of the environment.
The second large group of animals living in grasses are birds. Although the main role of the songs of most
singing birds is to attract a mate [26], the mostly exposed behavioural reason for sound emission in the
case of a corncrake [27] living in grass is territorial marking [28]. Results of studies on variation of the
corncrake songs show that they are connected with the degree of aggression. No reference of the sound to
the properties of the environment can be found in these studies.
From an acoustic view point, blades of grass in a meadow can be regarded as penduli that emit sound
if they are stimulated by an acoustic wave. In this way, the waves emitted from blades spread and can
interfere. Fig. 1 shows a sound spectrum reflected from a wall made of dry cane Phragmites australis
blades placed one next to each other. The blades were stimulated by one of tones from the range of
0.5 kHz -- 15 kHz and a reflected sound was recorded. Its spectrum is peaked at the frequency of stimula-
tion (however, some of these peaks are broadened considerably). Additionally, in the case of stimulation
with any tone, there is visible a non-specific continuous spectrum with low frequencies in the reflected
sound.
The role of interference of the sound is secondary in the known mechanisms of communication used by
animals, e.g. echolocation, and can be considered as a disturbance. It is negligible in open space com-
munication and in spaces scattering sound in random directions. For example, in closed spaces of caves
where bats forage, the sound is reflected from distant walls which are inclined in an irregular way to each
other; interference is negligible.
In the case of a dense environment of uniformly located blades of grass vibrating in a common plane the
interference is much more important. Here, its account is amplified by a regular arrangement of reflecting
objects that are sufficiently close to each other (compared to the wave length) to be able to interchange
the signal with a large amplitude and thus to form nearly a collective pattern of a reflected signal.
In particular, interference is not to neglect for a sound propagating through a field of blades for a typical
wavelength of sound emitted by animals inhabiting grass fields with a typical density. The typical distance
between grass blades on meadows in middle Europe is some centimeters. In turn, the sound spectra of a
corncrake Crex Crex and a cricket Gryllus campestris are shown in fig. 2.
f [kHz]
λ [cm]
0.5
69
1
34
3
11
5
7
10
3
15
2
Table 1: Wavelength λ as a function of frequency f of a tone for several frequencies covering an acoustic
spectrum. λ(f ) = c/f , where c=343 m/s is the speed of sound in the air at 20◦C.
2
(a)
(b)
Figure 2: Sound frequency spectrum of a corncrake (a) and a cricket (b).
Although these spectra are wide to some extent, they both are concentrated at about 5 kHz. Tab. 1
presents the wavelength for some tones as a function of their frequency. The wavelength for 5 kHz is
7 cm, which seems to be in good correspondence to the mean nearest neighbour distance for blades in a
meadow. Thus, the influence of interference should not be neglected in the case of a sound emitted into
blades. The composition of the interfered waves from all blades forms a final wave picture in a space.
Especially, the outgoing sound can be detected at the place of the emitting source (e.g. a cricket).
Both in optics and in acoustics, interference fringes are analyzed mostly in a spatial domain. However,
interference can be observed in a time domain, too, forming an acoustic picture of a meadow dynamically.
Although a fixed position of blades could form only a static interference pattern, the moving objects (like
a mate or a predator) within a field start to make this pattern a dynamical one.
In this paper, we present an analysis of changes in the loudness (forming interference fringes in time) of
the sound perceived by an animal emitting it and hearing it at a fixed position. It is shown here that a
sound with a frequency adjusted to the density of blades could be used as a sensitive tool indicating a
movement nearby.
3 Acoustic experiments
Description of the experiments. The experiment analysed here consisted in emission of sound toward
an artificial meadow and registering the sound returning to the place of emission thereof. To measure
changes in the loudness of the sound reflected from a meadow, we constructed an experimental setup
shown in fig. 3a.
Hereafter, we shall use the notion artificial meadow for a circular surface with several blades scattered
randomly throughout this surface (fig. 3a). In the case of our acoustic experiments, these blades were dry
blades of cane stuck into a Styrofoam board.
Furthermore, a newcomer denotes an animal (a mate or a predator) which can move through the meadow.
Its move is supposed to be detected by an animal emitting the sound by registration of loudness fluctuations
of the sound reflected from the meadow. In the acoustic experiments, the newcomer was substituted by a
20×30 cm metallic or Plexiglas plate moving 15 cm above the ground level between blades -- denoted as
S in the fig. 3a.
To calculate the number of blades needed to construct a field with a given nearest neighbour distance α,
we used formula
√
√
R
2
π
N
α =
,
(1)
where R -- the radius of the meadow, and N -- the number of blades [29]. It is clear from eq. (1) that the
surface density is an unambiguous function of α.
All reliable recordings that we performed were taken to a recording room of a local radio station. The
walls in this room were padded with mats absorbing some fraction of sound. In our further experiments,
we measured the propagation of the following sounds through the meadow:
• a set of pure tones F={0.5 kHz, 1 kHz, 3 kHz, 5 kHz, 10 kHz, 15 kHz},
3
(a)
(b)
Figure 3: (a) -- Scheme of the experimental setup.
'S' denotes a moving shutter whereas 'MIC' -- the
position of the microphone to measure the loudness of the reflected sound from the field almost in the
place of the singer represented by a loudspeaker 'L' here. (b) -- An artificial meadow with a radius R=1 m
made of N =200 randomly placed blades. Red points denote an initial newcomer position. The source
emitting the sound is placed at (0,0).
• a set of modulated tones Υ = Fm × F={0.5 Hz, 1 Hz, 2 Hz, 5 Hz} × {0.5 kHz, 1 kHz, 3 kHz, 5 kHz,
10 kHz, 15 kHz}, where F is a chosen set of carrier frequencies f , whereas Fm is a set of modulation
frequencies fm,
• prepared samples of corncrake's and cricket's sounds.
The samples of tones and modulated tones were generated in Mathematica software. They lasted 15 s and
were sampled with a frequency of 44100 Hz (CD-quality).
The prepared samples of natural sounds of a corncrake and a cricket were transformed in the Audacity
software, where a given sound was cropped to one period of a chirp, and next this part was multiplied.
This preparation allowed avoiding some natural irregularities present during a few-minute call.
The samples were emitted by a JBL EON10 G2 loudspeaker into the artificial meadow. The reflected
sound was recorded with an Audio-technica AT2031 microphone coupled with a computer and with
a Sennheiser ME67 microphone coupled with a Zoom H4n handy recorder. The recorded sound was
analysed by the Audacity software and programs written in Mathematica, C++, and Python.
Results of the experiments. A spectral decomposition of a received sound reveals interesting changes
in time. We used wavelet analysis as a comprehensive tool to show the frequency spectrum in time. Fig. 4
shows a Morlet wavelet [30] decomposition of the prepared signal of the cricket. The sub-figures show
the intensity of the signal (in colours; represented by power) at a given time (on the abscissa) and for
a given frequency for three values of the mean nearest distance α (the period is shown instead of the
frequency on the ordinate of the plot). During these recordings, the newcomer travelled across the arti-
ficial meadow with a velocity of 0.15 m/s. A comparison shows that if blades are present, there appear
additional frequencies in a signal (a period 100 -- 10−3 s) whose amplitude is less homogeneous in time
for a denser meadow. The amplitude of these frequencies is the largest for an intermediate density of the
meadow (fig. 4b). The appearance of this frequency band is characteristic for the sound of a cricket and
a corncrake as well as modulated tones. Intensive fluctuations of these frequencies with time are visible.
The origin of these additional frequencies is acoustic beats, which are a function of the modulation fre-
quency, the density of blades, and the velocity of the newcomer. They can possibly be used by animals
for identification of changes in the position of objects. Furthermore, some of these additional frequencies
4
(a) original sound (no blades)
(b) α=0.06 m
(c) α=0.04 m
Figure 4: Morlet wavelet decomposition of the prepared cricket's sound reflected from the artificial
meadow. At each time (on the abscissa), the oscillation period spectrum of the sound is shown. Their
intensity (represented by power) is denoted by colorus. An additional chart on the left side of each plot
shows the mean relative amplitude as a function of the period of the oscillations. The time series is
registered while the newcomer is passing across the artificial meadow with a velocity of 0.1 m/s.
come from the effect of generation of some spectrum of frequencies by blades when they are stimulated
by sound (as demonstrated in fig. 1).
Furthermore, we performed some measurements of the loudness changes in time for the reflected sound
for some simple tones and the velocity of the newcomer w=0.1 m/s. For low frequency f =0.5 kHz, the
loudness changes by 1-2 dB depended on the density (fig. 5a).
In this case, much of the intensity is
reflected from the wall of the meadow and no interference effects are visible. For greater frequencies,
the sound can penetrate the inside of the meadow and it is partially dispersed. However, in this case,
interference inside is possible. For example, for a tone 3 kHz (the wavelength is λ=11 cm), the condition
α (cid:39) λ is true. It suffices to produce an interference pattern -- fig. 5b. A similar effect is observed for
greater frequencies -- fig. 5c.
Similar experiments were repeated for modulated tones Υ. For these tones, we performed recordings where
the newcomer (a shutter) was at rest during the first 3 s; next, it moved with a velocity of w=0.1 m/s
from the 3rd to the 10th second, and then was at rest again from the 10th to the 15th second.
The result for both f and fm, which have relatively small values, is shown in fig. 6a. For the meadow with
sparse blades (a blue line), some waving of loudness is observed but it is inadequate to the moments when
the shutter appears in front of the microphone. For a greater density (in red there), a more adequate
increase in the amplitude is visible. Some shift of the amplitude in time, compared to the signal in blue,
is an interference effect.
The waving of the amplitude is present for both meadow densities if greater modulating frequency fm
is applied -- fig. 6b. For the dense meadow (red), the waving reduces the amplitude very slowly -- some
waving is visible even after moving out of a meadow by the shutter. The waving is the greatest when the
newcomer passes nearby but it is impossible to reconstruct its position from this pattern.
It is interesting to check the variation of loudness in time for a carrier frequency f comparable to those
which are characteristic for corncrakes or crickets. In fig. 7, the signal is compared for three densities of
the meadow. This signal seems to rebuild better with the increasing density of the meadow. The more
regular the signal is the less information it carries about the passage of the newcomer. However, in the
case of small density (blue), the amplitude fluctuation is significant and adequate to the position of the
newcomer. These fluctuations are not as great as those giving separate chirps so they are possibly not
much informative. In this case, the small density is inadequate for realistic meadows so these fluctuations
possibly do not correspond to realistic case. For these meadows, a signal in red is rather expected which
allows a conclusion that such a small value of the fm as in this case would be inadequate for signaling in
realistic meadows. Thus, the small fm values are not appropriate to give information about the movement
in the case of the dense meadow.
In contrast, when the modulation frequency fm is greater, the interference picture starts to be more in-
formative -- fig. 8. The maximum value of fluctuations in this case precedes the moment of passing the
newcomer at the nearest distance to the cricket. Furthermore, a fluctuation of the amplitude of the signal
is greater in the case of greater meadow density (red). For more quantitative comparison of these signals
fluctuations a distortion function was introduced -- see the Appendix.
5
(a)
(b)
(c)
Figure 5: Relative changes in loudness in time of the reflected sound for the artificial meadow penetrated
by pure tones: (a) -- 0.5 kHz, (b) -- 3 kHz, and (c) -- 10 kHz. The velocity of the newcomer was w=0.1 m/s.
In sub-figure (a), the colors denote various numbers of blades (N =0, 100, ..., 500). In (b): α=0.3 m for
the blue series, and α=0.04 m for the red series. In (c): blue -- α=0.1 m, red -- α=0.04 m (here, to show
differences more accurately, the spectrum in blue was shifted up by 5 dB).
6
(a)
(b)
Figure 6: A time-line of loudness for a low carrier frequency tone f =1 kHz scattered from the artificial
meadow. The modulation frequencies are: (a) fm=0.5 Hz and (b) fm=5 Hz. In part (a), α is 0.1 m (blue)
and 0.04 m (red). In (b), α is 0.1 m (blue) and 0.06 m (red). The gray vertical lines denote the moment
of starting and stopping moving by the newcomer (w=0.1 m/s). In the case of the greater fm, (b), more
intensive fluctuations are visible during the passage of the newcomer.
Figure 7: Low modulation frequency case: a time-line of loudness of a modulated tone fm=0.5 Hz f =5 kHz
scattered from the meadow with α=0.1 m (blue), α=0.06 m (orange), and α=0.04 (red). A meaning of
gray vertical lines is the same as in fig. 6.
Figure 8: High modulation frequency case: a time-line of a reflected signal loudness of a modulated tone
fm=5 Hz f =5 kHz from the meadow with α=0.1 m (blue) and with α=0.06 m (red). The red part was
shifted by 5 dB to see the differences more accurately. The meaning of the gray vertical lines is the same
as in fig. 6. Considerable fluctuations of the sound loudness are visible when the newcomer is moving
through the area with blades.
7
Figure 9: Loudness of a reflected signal with time for the voice of a corncrake emitted from the loudspeaker;
blue points -- the original recording of the voice, red points -- the signal emitted into the field with α=0.06 m.
Some fluctuations of the amplitude of peaks are visible during the passage of the newcomer.
Figure 10: Loudness of a reflected signal with time of the cricket's voice reflected from a meadow with
α=0.06 m (red). The blue points -- an original signal. The lower part of the figure is a magnification of
the region of the peaks. Fluctuations of the amplitude of peaks are visible.
Finally, we checked amplitude fluctuations with time for the sound of a corncrake emitted from a loud-
speaker toward the artificial meadow -- fig. 9. The greatest changes in the amplitude (red), compared to
the original one (blue), were observed for the moment when the shutter passed at the nearest distance from
the emitting center. However, the main feature of the spectrum is that the signal is shifted in time and
its base is elongating, i.e. the duration of each signal is prolonged; furthermore, its amplitude increases
due to the interference effect.
Contrary to this, in the case of the cricket, the amplitude of the reflected sound rebuilds poorly com-
pared to the level of the original signal and never reaches its maximum value. Nevertheless, similarly to
the case of the corncrake, the amplitude of the signal fluctuates during the transition of the newcomer
(fig. 10). This spectrum reveals some more interesting findings. Some peaks here are not reconstructed
and some irregular shifts in time are observed. The differences in the position of the peaks vary between
the different phases of the experiment. We checked numerically, for example, that for a given density of
blades, the shifts for the beginning and the middle part of the experiment differ by tenths of milliseconds.
This interesting feature shows that the sound peaks are reconstructed by interference and are not a simple
reflection of the original signal with the echo effect delay. Such an effect was observed many times and
seems to be very common -- see fig. 6a as an example (here, positions of minimum values remain the same,
only maximum values change their positions).
We were unable to reconstruct the parameters of a real meadow that is characteristic for a corncrake or a
cricket in the studio. Thus, our artificial meadow made of dry cane would perhaps be inadequate to some
extent for the sound of real animals that we used and the position of the peak of the sound is somewhat
not reliable. However, the general finding of the dependence of the position of the peak on density is
8
correct and should be explored in detail in the future.
4 Numerical experiments
Theoretical model. By similarity to the voice of crickets and a corncrake, we assumed that a represen-
tative of an animal voice can be a tone of frequency f modulated by frequency fm (amplitude modulation).
Numerically, the f value is a representation of the pitch of the voice (e.g. 5 kHz), whereas fm represents
the frequency of repetitions of the sound (chirps; fm=2 Hz).
The wave emitted by the i-th singer is
Ψi(r, t) = sin [Ω · t − k ◦ (r − r(i))] × sin [Ωm · t − km ◦ (r − r(i))] =
= sin
2πf · t − 2πf
c
(rj − r(i)
j )
2πfm · t − 2πfm
c
(rj − r(i)
j )
,
(2)
3(cid:88)
j=1
(cid:17) × sin
(cid:16)
(cid:17)
3(cid:88)
j=1
where c is the speed of sound and r(i) is the position vector of the i-th singer.
Furthermore, a wave incoming to a blade stimulates it to vibrate. The wave emitted by the j-th blade is
a composition of spherical waves with frequency f1 (as the simplest case, only one frequency was taken
into account) and with an amplitude generated each time as a sum of i amplitudes of incoming waves
(cid:16)
0 (t)
ψij
r − rj sin
2πf1 (t − tij
0 ) − 2πf1
c
(rk − r(j)
k )
3(cid:88)
k=1
(cid:17) · θ(t − tij
0 ),
where ψij
i (rj, t) is the wave amplitude at time t generated by the i-th cricket in the place of
the j-th blade, N -- is the number of crickets, r(j) is the position vector for the j-th blade, θ(t − tij
0 ) is
the step function, and tij
0 is the time at which the stimulating wave from the i-th cricket arrives at the
j-th blade. The latest term is added because the j-th blade is stimulated to vibration by the i-th wave,
when it arrives at it (but not before). The waves emitted from crickets influence not only the amplitude of
vibration of a given blade. Also other blades which scatter spherical waves are important here. Therefore,
in our code, the sum in eq. (3) does take into account other blades as a source of stimulating waves.
Waves defined by expression (3) propagate and reach back the k-th cricket. The wave at this place is
described by
(cid:16)
N(cid:88)
i
ψj(r, t) =
0 (t) =(cid:112)Ψ2
(3)
(4)
(5)
(cid:88)
j
ψk(rk, t) =
ψj(rk, t).
Ik(t) =(cid:112)ψ2(rk, t).
This provides the following expression for the sound intensity at the k-th cricket
Description of the numerical experiments. Loudness as a function of time for a reflected sound from
a meadow could be better understood by analyzing the numerical experiments that we conducted. Here,
the meadow was simulated as a set of 2-dimensional points scattered through a circular area (fig. 3b).
Each of the blades of this artificial meadow is a center of circular waves emitted as a consequence of
stimulation thereof by an incoming acoustic wave. This reduction of a problem to a 2-dimensional case is
not primarily important to the main ideas presented here but does reduce time and resources consumed
by numerical calculations considerably.
The newcomer was simulated here as a set of points (segments) which can reflect the sound waves as the
blades can. The coordinates of the newcomer's segments were chosen randomly along the direction of its
future movement. The distance between the segments was short, compared to the radius of the meadow
(less than 1% of R; fig. 3b), and adequately to simulate an elongated and irregular shape of a worm.
We made programs allowing us to simulate the crucial acoustic features of a meadow. First, an artificial
meadow was created. To do this, input parameters such as meadow's radius and the number of blades were
set. Based on these parameters, the coordinates of blades and a newcomer were chosen randomly inside
the radius by the software. The formula (1) was used to calculate the number of randomly scattered blades
required to have given α for a meadow of a given radius R. As a result, we obtained a set of blades scattered
quasi-uniformly within this circular field. In these numerical calculations, sound reflection coefficients for
9
Figure 11: Time-line of a numerically calculated relative intensity of a reflected 3.4 kHz tone emitted
toward the meadow (R=1 m) for different values of the nearest neighbor distance α, which are in meters:
α=0.3 (blue; almost no blades of the meadow), 0.13 (yellow), 0.07 (green), and 0.03 (orange). The velocity
of the newcomer is 0.1 m/s, whereas the number of its compartments is 5 here. The red dots at the abscissa
denote the time when the newcomer entered and left the meadow ring.
Figure 12: Numerically calculated values of the relative intensity of the sound returning to the singer as
a function of the radius of the meadow. α is kept at the value of 0.07 m.
both the newcomer and the blades were set as input parameters. Finally, the blades' and newcomer's
coordinates were used to calculate the intensity of sound as a function of time in the place of an emitting
singer. To calculate this intensity, we used the formula (5). As a simplification, only one singer was
considered (placed at point (0,0) -- see fig. 3b).
Results of the numerical experiments. The importance of the presence of blades and the interference
effect is clear when the intensity for different densities of the meadow is compared. The blades reflect
sound and thus they increase the intensity of the interfered sound returning to the singer. The left red
dot on the abscissa in fig. 11 means a moment when the newcomer started to move, while the right one
means the time when the newcomer left the area with blades and continued to move way. The data series
in this figure show the results for different nearest neighbor distances α. When the density of blades is
much lesser than the wavelength (blue and yellow lines in this figure), the signal has a small amplitude.
For α equal or less than the wavelength (α ≤ 0.07 m, in this case), the intensity of the reflected signal
and its fluctuation amplitude increase considerably due to the interference.
Two preliminary numerical experiments should be mentioned. First, we checked the variation of the
loudness at the same α when the position of particular blades is randomly rearranged. Some trials of the
positions of the blades were made for a meadow radii R=3 m and 8 m. In these experiments, we recorded
a maximum of the intensity when the newcomer was absent. The conclusion is that, for both radii, the
loudness of the back signal differs by 2-5 dB for different randomly chosen positions of blades.
The next preliminary experiment consisted in recording of the sound for some radii R of the meadow in
a range of 0.5 m -- 6 m with a step of 0.5 m (keeping the positions of blades, density of the meadow, and
parameters of the sound unchanged). It is expected that the influence of distant blades for a sufficiently
wide meadow is negligible. In this case, the variation of loudness was 2-3 dB over all radii. It seems that
the loudness of the sound for R=3 m is comparable to that for R=10 m -- fig. 12. Thus, next experiments
with changing other parameters of sound were performed for a meadow with R=6 m.
The main goal of our numerical simulations was to check how the periodic signal emitted by a cricket
10
changes if some of the crucial parameters such as f , fm, or w vary. For each signal, we calculated
and plotted some secondary functions of the intensity with time. They are chosen as intuitive functions
represented changes in time. The most useful picture of loudness fluctuation is given by the first derivative
of the intensity and the local distortion of the loudness. The distortion function ∆ is a measure of changes
of the intensity -- for details see the Appendix. The larger is the distortion the greater are fluctuations of
the intensity.
The α to λ adjustment (it can be translated to the adjustment of the density of a meadow to the wave
length) seems to be crucial when one considers a time function of the sound intensity. For given α, a given
set of sound parameters may be regarded as well or more badly adjusted to signalling a move. We define
a sound wave as a well-adjusted one to α if the wavelength λ = c/f value of the sound is comparable to
α. Oppositely, when the wavelength and the α diverge considerably, the sound is more badly adjusted.
All the results below were obtained for α=0.07 m (quite a realistic value). This distance equals to
the wavelength of the tone f =f0 ≡5 kHz characteristic for a corncrake or a cricket. The results were
compared to frequencies f << f0 and f >> f0. In these experiments, the newcomer is represented by 20
compartments. The sound reflection coefficient of the newcomer compartments was set equal to that for
the blades of grass and set to 0.8 (its value does not change the results qualitatively). In each cases, the
newcomer started to move 2 s after starting sound emission. The sampling step was 0.05 s. For simplicity,
the intensity rather than the loudness is presented1.
Fig. 13 presents changes in the sound from the meadow with time for modulated tones: f =1 kHz, 5 kHz,
and 10 kHz, whereas fm=2 Hz and w=1 m/s. Based on this example, a general regularity is shown -- for a
badly-adjusted frequency of f > f0=5 kHz, the intensity of the signal decreases (the green line). That is
the reason that the sound with a great frequency does not suffice to serve as a tool for indicating changes.
In contrast, the intensity and its fluctuations for a sound with f < f0 seem to be more adequate from this
point of view. However, for such a small frequency f (gray line), the fluctuations are spread in time and
are generally lesser than the fluctuations for f = f0. The distortion function (see the Appendix) for these
cases are ∆=7.20, 7.24, and 6.84 for f =1, 5, and 10 kHz, respectively, which means that the greatest
amplitude variation is for f ∼ f0.
The modulation frequency fm seems to play a non-trivial role in the intensity variation -- fig. 14. Although
the distortion varies slowly with fm, its maximum value is at some intermediate fm (fm=2 Hz here). Its
value for fm=0.5, 2, and 10 Hz are 7.27, 7.33, and 7.16, respectively. The maximum of the amplitude
fluctuation as a function of fm is probably related to the velocity of the newcomer. This issue, however,
has not been explored systematically yet.
Finally, we checked the influence of the velocity w of the newcomer. It is clear that there are no interference
changes in time if w=0. In general, the distortion is a weakly changing function of the velocity value in a
wide range. For example, for f =5 kHz, fm=2 Hz, the ∆ is 7.44, 7.45, and 7.45, for w=0.1 m/s, 0.5 m/s,
1 m/s, respectively.
The number of the compartments of the newcomer is relatively small compared to the number of blades
(ca. 6×103 here). Thus, the amplitudes of peaks are not large. By adding the compartments (or, to
some extent, increasing the reflection coefficient), one can obtain much greater intensity fluctuations.
The fluctuations of loudness that we observed in the acoustic experiments are easy to be simulated by
increasing the size of the newcomer. For example, let us assume that ca. 400 blades form a meadow
with R=2 m and α=0.07 m. If the newcomer is made of Nc=10 compartments, the maximum intensity is
3×105, in turn, if Nc=50, the maximum value is 8×107.
To generalize, our numerical simulations prove the existence of amplitude (intensity) fluctuations during
a passage of the object reflecting sound throughout the meadow. The greatest fluctuations are obtained
for a sound with the frequency well-adjusted to α.
5 Conclusions
We have shown both experimentally and by numerical experiments that the sound emitted by an
animal living in grass or shrubs returns to this animal after reflections from blades of the meadow and
interference and forms a complex pattern of changes in the intensity with time. This picture of changes
is sensitive to some crucial parameters of the sound and the meadow like the velocity of the newcomer or
1In the loudness definition L[dB] = 10 log(I/I0), the I0 could be replaced for our purposes by the intensity of some
referred tone instead of I0 known for a standard definition of loudness related to human hearing capabilities, for example,
the mean intensity of a sound reflected when a newcomer is at rest.
11
Figure 13: Numerically calculated values of the intensity of the reflected sound during passage of the
newcomer across the meadow for several frequencies: f =1 kHz -- gray dotted line, 5 kHz (optimal for the
density of a meadow) -- red line, 10 kHz -- green thick line. The red points at the abscissa indicate the
time of starting to move by the newcomer and the time of moving out of the meadow. In these numerical
calculations, R=6 m, fm=2 Hz, and w=1 m/s. The figure in the middle shows a local distortion function
of the adequate signals, whereas the figure at the bottom is the 1st derivative of the intensity.
Figure 14: Intensity of the reflected signal as a function of the modulation frequency fm: 0.5 Hz -- gray
dotted line, 2 Hz -- green line, 10 Hz -- red dashed line. The red points at the abscissa indicate the time
of starting to move by the newcomer and the time of moving out of the meadow. Other parameters are
R=6 m, f =5 kHz, and w=1 m/s.
12
the relation between the sound frequency and the density of the meadow. The relative variation in the
loudness is some decibels. Such a variation of the intensity should be distinguishable and allows regarding
these fluctuations as a possible tool that could be used by animals for detection of movement.
Hearing distortions seem to be stable within the range of natural changes in temperature -- when the
temperature changes by ten degrees or more, the sound wavelength changes by millimeters, and the
interference conditions remain almost unchanged. Also, the changes of the rate of chirp with temperature
(Dolbear's law [31]) does not affect the quality of this tool (fig. 14).
Sensing the environment by analyzing sound amplitude changes seems to be an interesting issue (according
to our knowledge, not reported until now) from both the basic-knowledge and ecological points of view.
For example, if the mechanism of such detection of a mate or a predator is really used by living creatures,
some modification of the environment by human activity could change the size of a population. However,
one should be aware that sound communication in real biological systems depends on biocoenotic and
geological conditions. Both of them should be incorporated as parameters in a mathematical model of the
acoustic environment of this system. Here, the basic version of such a model is presented.
In this paper, we limited the investigations to the case when only one sound-emitting animal is taken into
account. Nonetheless, it would be interesting to consider a net of animals which cooperate with each other
in producing a field of sound for detection the motion in the surroundings.
6 Acknowledgement
The authors want to thank A. Nieoczym, the manager of 'Chatka Zaka' and M. Skowronek, the manager
of Radio Centrum for enabling us to perform the acoustic experiments in a recording studio.
Thanks for Prof. B. Bernatowicz (Department of Art of UMCS University) for enabling us to perform our
first experiments in an acoustically controlled space.
We (B.T. and J.M., especially) would like to thank Mr. P. Kononowicz (Olympic special-interest group
for young physicists in Lublin) for developing passion for discovering Nature's marvels in young people.
7 Appendix -- Distortion function
(i + T )-th samples, which are distanced by one T period, i.e. ∆ = (cid:80)
Because of the overlap of the position of many samples in the figures used in this paper (CD-quality
sampling rate), real differences cannot be presented graphically. Therefore, we introduced a numerical
measure for a periodic signal to quantify distortions made by a disturbing factor (like a shutter entering
the meadow). We define distortion ∆ as a sum of absolute differences in loudness Li for the i-th and
i=1 Li+T − Li. The summation
is taken over all samples that have their counterpart at a period forward in the recording. Finally, to
normalise the expression, this sum is divided by the number of samples2. For example, for a periodic signal
without disturbances, ∆=0. The distortion from periodicity contains information about the presence of a
disturbing factor.
For instance, in the case of modulated (fm=5 Hz) tones: f =0.5 kHz, 5 kHz, and 10 kHz, the value of
∆ is 0.15, 4.1, 0.6 for α=0.06 m and 0.17, 5.4, and 0.48 for α=0.04 m, respectively. This means that
the highest distortion for quite realistic densities of a meadow is noted in the case of a sound with an
intermediate frequency.
Similarly, the sound amplitude fluctuations for our representative animals were quantified by us using the
∆. In these experiments, the starting and stopping time of the newcomer move was normalized (to tenths
of a second) and the prepared sounds samples were used. For example, the ∆ for the sound of a corncrake
as a function of α is shown in fig. 15. The peak of this experimental function shows that meadows with
moderate density (the peak is at ca. α=0.05 m) are preferable to serve as a medium providing information
about changes therein.
References
[1] R. Hoy, A. Popper, and R. Fay, eds., Comparative Hearing: Insects. Springer, 1998.
2In this text, we also used a notion of the local distortion, which is a distortion function in time by means of the time
series of the absolute difference in the intensity at given time and that for a period forward.
13
Figure 15: Distortion function ∆ of the prepared sound of a corncrake as a function of the nearest neighbor
distance α.
[2] D. Blumstein, "The evolution, function, and meaning of marmot alarm communication," Adv. Study
Behav., vol. 37, pp. 371 -- 400, 2007.
[3] M. Fenton, "Echolocation:
implications for ecology and evolution of bats," Q. Rev. Biol., vol. 59,
pp. 33 -- 53, 1984.
[4] D. Kroodsma and B. Byers, "The function(s) of bird song," American Zoologist, vol. 31, pp. 318 -- 328,
1991.
[5] A. Michelsen and H. Nocke, "Biophysical aspects of sound communication in insects," Advances in
Insect Physiology, vol. 10, pp. 247 -- 296, 1974.
[6] J. Podos, S. Huber, and B. Taft, "Bird song: The interface of evolution and mechanism," Annual
Review of Ecology and Systematics, vol. 35, pp. 55 -- 87, 2004.
[7] R. Cocroft and R. Rodr´ıguez, "The behavioral ecology of insect vibrational communication," Bio-
Science, vol. 55, pp. 323 -- 334, 2005.
[8] G. Pollack, "Who, what, where? Recognition and localization of acoustic signals by insects," Curr.
Opin. Neurobiol., vol. 10, pp. 763 -- 767, 2000.
[9] B. Geurten, C. Spalthoff, and M. Gopfert, "Insect Hearing: Active Amplification in Tympanal Ears,"
Curr. Biol., vol. 23, pp. R950 -- R952, 2013.
[10] A. Moiseff and R. Hoy, "Sensitivity to ultrasound in an identified auditory interneuron in the cricket:
a possible neural link to phonotactic behavior," J. Comp. Physiol., vol. 152, pp. 155 -- 167, 1983.
[11] F. Huber, H. Kleindienst, T. Weber, and J. Thorson, "Auditory behavior of the cricket," J. Comp.
Physiol. A, vol. 155, pp. 725 -- 738, 1984.
[12] N. Fletcher and S. Thwaites, "Acoustical analysis of the auditory system of the cricket Teleogryllus
commodus (Walker)," J. Acoust. Soc. Am., vol. 66, pp. 350 -- 357, 1979.
[13] D. Robinson and M. Hall, "Sound signalling in orthoptera," Adv In Insect Phys., vol. 29, pp. 151 -- 278,
2002.
[14] J. Poulet and B. Hedwig, "Auditory orientation in crickets: Pattern recognition controls reactive
steering," PNAS, vol. 102, no. 43, pp. 15665 -- 15669, 2005.
[15] N. Mhatre and R. Balakrishnan, "Phonotactic walking path of field crickets in closed-loop conditions
and their simulation using a stochastic model," J. Exp. Biol., vol. 210, pp. 3661 -- 3676, 2007.
14
[16] A. Witney and B. Hedwig, "Kinematics of phonotactic steering in the walking cricket Gryllus bimac-
ulatus (de Geer)," J. Exp. Biol., vol. 214, pp. 69 -- 79, 2011.
[17] G. Wandler, "Pattern recognition and localization in cricket phonotaxis," in Sensory systems and
Communication in Arthropods (F. Gribakin, K. Wiese, and A. Popov, eds.), p. 387, Birkhauser
Verlag, 1990.
[18] G. Meckenhauser, R. Hennig, and N. M.P., "Critical song features for auditory pattern recognition
in crickets," PLoS ONE, vol. 8, no. 2, p. e55349, 2013.
[19] J. Stabel, G. Wendler, and H. Scharstein, "Cricket phonotaxis: localization depends on recognition
of the calling song pattern," J. Comp. Physiol. A, vol. 165, pp. 165 -- 177, 1989.
[20] T. Walker, "Acoustic synchrony: Two mechanisms in the snowy tree cricket," Science, vol. 166,
no. 3907, pp. 891 -- 894, 1969.
[21] A. Covey, "Information content of house cricket (acheta domesticus) songs and the evo-
thesis, State University of New York at Fredonia,
lution of multiple signals," master's
http://hdl.handle.net/1951/58372, 2012.
[22] J. Doherty and R. Hoy, "The auditory behavior of crickets: some views of genetic coupling, song
recognition, and predator detection," Q. Rev. Biol., vol. 60, no. 4, pp. 457 -- 472, 1985.
[23] P. Gullan and P. Cranston, The insects: an outline of entomology. Wiley-Blackwell, 2010.
[24] G. Pollack, A. Mason, A. Popper, and R. Fay, eds., Insect Hearing. Springer, 2016.
[25] G. Pollack, "Insect bioacoustics," Acoustics Today, vol. 13, pp. 26 -- 34, 2017.
[26] P. R¸ek and T. Osiejuk, "Nonpasserine bird produces soft calls and pays retaliation cost," Behavioral
Ecology, vol. 22, pp. 657 -- 662, 2011.
[27] G. Tyler, R. Green, T. Stowe, and A. Newton, "Sex differences in the behaviour and measurements
of corncrakes crex crex in scotland," Ringing & Migration, vol. 17, pp. 15 -- 19, 1996.
[28] P. R¸ek and T. Osiejuk, "Temporal patterns of broadcast calls in the corncrake encode information
arbitrarily," Behavioral Ecology, vol. 24, pp. 547 -- 552, 2013.
[29] S. Torquato, Random Heterogeneous Materials: Microstructure and Macroscopic Properties. Springer
Science & Business Media, 2002.
[30] C. Torrence and G. Compo, "A practical guide to wavelet analysis."
[31] A. Dolbear, "The cricket as a thermometer," The American Naturalist, vol. 31, p. 970 -- 971, 1897.
15
|
1712.02666 | 1 | 1712 | 2017-12-07T15:23:10 | CRC 1114 - Report Membrane Deformation by N-BAR Proteins: Extraction of membrane geometry and protein diffusion characteristics from MD simulations | [
"physics.bio-ph"
] | We describe simulations of Proteins and artificial pseudo-molecules interacting and shaping lipid bilayer membranes. We extract protein diffusion Parameters, membrane deformation profiles and the elastic properties of the used membrane models in preparation of calculations based on a large scale continuum model. | physics.bio-ph | physics |
CRC 1114 -- Report
Membrane Deformation by N-BAR Proteins:
Extraction of membrane geometry and protein
diffusion characteristics from MD simulations
J. H. Peters, C. Graser, R. Klein
September 10, 2018
1 Membrane Shaping Proteins (N-BAR domain)
A good overview of mechanisms shaping biological membranes can be found
in [1].
In a previous study ([2]), membrane bending by a BAR domain was
investigating using four models of different resolutions.
1.1 Simulation Setup
Membrane protein structures are notoriously difficult to determine and hence
only a small number of structures are available, most of which are of lower qual-
ity. The N-BAR system used here was based on two crystallographic structures,
a lower quality dimer structure (PDB id 4I1Q [3]) and a higher quality monomer
structure (PDB id: 4AVM [4]). While the monomer structure is of higher quality
than the dimer one, it still is missing a number of atoms, which were using PDB-
Fixer (https://github.com/pandegroup/pdbfixer accessed 30.10.2017) but also
a N-Terminal alpha-helix which is assumed to be essential in membrane binding
([5, 2]) and which had to be modelled using PyMOL (https://pymol.org). Two
copies of the corrected monomer structure were then fitted on the two chains of
the dimer structure to produce a fixed dimer structure to be used in simulation.
To reduce interaction between periodic images, the dimer was placed in
contact with a big (600Ax 600A) membrane patch. As in previous work [2],
the membrane was composed of 70% DOPC and 30% DOPS which carries a
net charge of −1e per molecule, as BAR-domain membrane interactions are
likely electrostatic in nature ([2]). To reduce computational cost, the MAR-
TINI coarse-grained forcefield [6, 7] was used. The system was set up using
CHARMM-GUI (http://www.charmm-gui.org) [8, 9, 10]. And simulated using
GROMACS 5.1.2 [11].
After equilibration, a total of 300 ns were simulated using suggested sim-
ulation parameters from CHARMM-GUI: A time step of 20 fs was used and
1
neighbour searching was performed every 10 steps. The Shift algorithm was
used for electrostatic interactions with a cut-off of 1.2 nm. A single cut-off
of 1.4 was used for Van der Waals interactions. The V-rescale temperature
coupling algorithm was used to keep the system at 303.15K and semi-isotropic
pressure coupling was done with the Berendsen algorithm.
1.2 Simulation results
1.2.1
2D diffusion of N-BAR dimer
One possible long-term goal of the collaboration with project C07 of CRC 1114
was the design of a dynamic model of proteins diffusing (under consideration of
membrane-induced interaction potentials) on a membrane. To this end, the 2D
diffusion coefficient of a N-BAR Dimer on a membrane is of interest.
The trajectory of the centre of mass of the N-BAR dimer from 1 ns snapshots
(figure 1) corresponds to a random walk on a 2D plane defined by the membrane.
Calculating from this the mean squared displacement over time (figure 2),the
diffusion constant can be approximated using
Figure 1: Trajectory of the center of mass of the Dimer projected to the XY
plane.
(cid:104)x2(cid:105) = dDt
2
Figure 2: Mean squared displacement (in the XY-plane) of the dimer center of
mass for different time differences. For dt ≤ 100ns statistics are good enough
to be described as normal diffusion.
Based on dt < 100ns (the regime for which (cid:104)x2(cid:105) increases linear over time
as would be expected in normal diffusion, figure 2), the diffusion coefficient is
D = (0.047±0.002)nm2/ns. Using all data (including the low statistics dt > 250
ns), D = (0.039 ± 0.008)nm2/ns
1.2.2 Induced curvature
During the whole simulation time, the N-BAR dimer remains attached to the
membrane (figure 4).
To quantify the effect of the N-BAR dimer, the curvature of the membrane at
the location of the protein has to be calculated. Since the membrane fluctuates
over time (figure 6), an average over a long time (200 snapshots representing 200
ns simulation time) results in a clear average membrane deformation (figure 5).
To calculate the average, translational movement in the xy plane and rotational
movement around the z-axis of the N-BAR dimer had to be removed.
3
Figure 3: 2D diffusion coefficient calculated for different maximal time differ-
ences.
Figure 4: Membrane deformation by N-BAR Dimer.
4
Figure 5: Average deformation (for trajectory from 100 ns to 300 ns) of the
membrane due to interaction with an N-BAR dimer (indicated in black).
Figure 6: Average height profile of the membrane in x-direction and y-direction
through the N-BAR dimer position. The shaded areas describe the standard
deviation of membrane fluctuations over time.
5
2 Geometric Membrane Insertions
There are only a small number of fully resolved structure of molecules known
to shape biological lipid membranes, and even fewer of transmembrane proteins
known to have this effect. For a more systematic investigation of the effect
cone-shaped membrane insertions, we are using geometrically defined pseudo-
molecules.
As protein domains that are inserted into biological membranes usually have
they have charged or hydrophilic
corresponding electrostatic properties (i.e.
residues where the protein is in contact with charged or hydrophilic lipid head
groups and hydrophobic where the protein is in contact with hydrophobic lipid
tails), such insertions ideally should have similar properties.
One simple way to construct such domains is to select atoms from an existing
lipid bilayer structure selected to be located along the desired geometry (figure
7).
Figure 7: Cone shaped membrane insertion with an opening angle of 30◦. When
position restraints are used, the cone does retain its initial shape (left). By using
nearest neighbour bonds and bond angles, the cone can be kept at approximately
the same angle even after longer simulation time (right). In this case, the cone
can move through the simulation box.
In a simulation, these artificial trans-membrane domains are inserted into
a lipid bilayer domain consisting of DPPC lipids. To have a mobile but stable
cone structure, bonds and bond angles between the atoms have been introduced
to form cone-shaped "pseudo molecules" (figure 7).
As there is no explicit influence of water in these interactions, we have chosen
DryMARTINI [12], the implicit solvent variant of the MARTINI coarse grained
forcefield for these simulations.
A total of 500 ns were simulated with a time step of 20 fs. Neighbour search-
ing was performed every 10 steps. The Shift algorithm was used for electrostatic
interactions with a cut-off of 1.2 nm and a single cut-off of 1.4 was used for Van
6
der Waals interactions. Semi-isotropic pressure coupling was done with the
Berendsen algorithm. As is common for simulations using the DryMARTINI
forcefield, the leap-frog stochastic dynamics integrator implemented in GRO-
MACS was used with a reference temperature of 323K (DPPC transitions into
the gel phase below 314K) and an inverse friction constant of 4.0 ps.
Membrane insertions are known to have a weaker effect on membrane cur-
vature than memberane-attached proteins the BAR domain ([1]). Again, sys-
tematic membrane deformation can best be observed by averaging over longer
trajectories. Also, since the membrane insertions are rotationally symmetric,
the height profile was averaged radially (figure 8).
Figure 8: Average deformation of the membrane due to interaction with a cone-
shaped insertion. the shaded areas describe the standard deviation of the mem-
brane fluctuations.
3 Mapping from Molecular Dynamics to Con-
tinuum Model
Eliott Et. Al.
[13] have developed a variational approach to model the in-
teraction between membranes and membrane-shaping particles based on the
Canham-Helfrich continuum model [14, 15, 16], which requires the bending
rigidity to describe the response of the membrane to interaction.
Both for membrane stiffness calculations (section 4) and for comparison with
continuum models, it is helpful to map the particle representation of the mem-
brane to a continuous model of the form h(x, y) = z. Radial basis functions are
a mesh-free model of the form
N(cid:88)
cnf ((cid:112)(x − xn)2 + (y − yn)2)
h(x, y) =
By using the coordinates of N atoms (xn, yn, zn) as anchor points of the rbf,
the function can be constructed in a way to satisfy
n
h(xn, yn) = zn
7
for all points by selecting the correct cn through a simple set of linear equa-
tions.
To account for statistical fluctuations in atom positions, we divide the mem-
brane into a number of layers (using the same atom from each lipid molecule),
construct an rbf for each of them (after centring them in z-direction) and then
average over these models.
8
4 Bending Stiffness Calculations
An approach to calculate the membrane bending stiffness from equilibrium sim-
ulations (in contrast to non-equilibrium simulations like vesicle closing simu-
laitons) has been developed by Khelashvili Et. Al. ([17, 18]). While the origi-
nal implementation ([18] requires a solvent density to calculate the membrane
surface (and from it the surface normal which is required for the calculation of
the bending stiffness), our Rbf-based membrane fitting approach can also be
applied to implicit solvent systems where no water density is available.
Following the method described in [18], we have calculated bending rigidity
of an unperturbed DOPC/DOPS membrane using the explicit solvent MARTINI
coarse grained forcefield (corresponding to the membrane used in the N-BAR
domain simulations, section 1) and a DPPC membrane using the implicit solvent
DryMARTINI forcefield (corresponding to the membrane used in the geometric
insertion simulations, section 2 ). The resulting bilayer bending moduli are
listed in table 1.
membrane system bending rigidity
(16.8 ± 0.8)kBT
DOPC/DOPS
(20.2 ± 0.8)kBT
DPPC
Table 1: Bending rigidities calculated for unperturbed membrane systems
9
5 Conclusions
Acknowledgements
This research has been funded by Deutsche Forschungsgemeinschaft (DFG)
through grant CRC 1114, project C01. The authors acknowledge the North-
German Supercomputing Alliance (HLRN) for providing HPC resources that
have contributed to the research results reported in this paper. We would like
to thank Prof. Roland Netz for his suggestions leading to the simulation of
cone-shaped membrane inclusions.
10
References
[1] I. K. Jarsch, F. Daste, and J. L. Gallop, "Membrane curvature in cell
biology: An integration of molecular mechanisms," J Cell Biol, vol. 214,
pp. 375 -- 387, Aug. 2016.
[2] A. Arkhipov, Y. Yin, and K. Schulten, "Four-Scale Description of Mem-
brane Sculpting by BAR Domains," Biophysical Journal, vol. 95, pp. 2806 --
2821, Sept. 2008.
[3] M. J. Snchez-Barrena, Y. Vallis, M. R. Clatworthy, G. J. Doherty, D. B.
Veprintsev, P. R. Evans, and H. T. McMahon, "Bin2 Is a Membrane Sculpt-
ing N-BAR Protein That Influences Leucocyte Podosomes, Motility and
Phagocytosis," PLOS ONE, vol. 7, p. e52401, Dec. 2012.
[4] C. K. Allerston, T. Krojer, C. D. O. Cooper, M. Vollmar, C. H. Arrowsmith,
A. Edwards, C. Bountra, F. Von Delft, and O. Gileadi, "Crystal Structure
of the N-Bar Domain of Human Bridging Integrator 2.," To be Published.
[5] J. L. Gallop, C. C. Jao, H. M. Kent, P. J. G. Butler, P. R. Evans, R. Langen,
and H. T. McMahon, "Mechanism of endophilin NBAR domainmediated
membrane curvature," The EMBO Journal, vol. 25, pp. 2898 -- 2910, June
2006.
[6] L. Monticelli, S. K. Kandasamy, X. Periole, R. G. Larson, D. P. Tieleman,
and S.-J. Marrink, "The MARTINI Coarse-Grained Force Field: Extension
to Proteins," Journal of Chemical Theory and Computation, vol. 4, pp. 819 --
834, May 2008.
[7] D. H. de Jong, G. Singh, W. F. D. Bennett, C. Arnarez, T. A. Wassenaar,
L. V. Schfer, X. Periole, D. P. Tieleman, and S. J. Marrink, "Improved
Parameters for the Martini Coarse-Grained Protein Force Field," Journal
of Chemical Theory and Computation, vol. 9, no. 1, pp. 687 -- 697, 2012.
[8] S. Jo, T. Kim, V. G. Iyer, and W. Im, "CHARMM-GUI: A web-based
graphical user interface for CHARMM," Journal of Computational Chem-
istry, vol. 29, pp. 1859 -- 1865, Aug. 2008.
[9] Y. Qi, H. I. Inglfsson, X. Cheng, J. Lee, S. J. Marrink, and W. Im,
"CHARMM-GUI Martini Maker for Coarse-Grained Simulations with the
Martini Force Field," Journal of Chemical Theory and Computation,
vol. 11, pp. 4486 -- 4494, Sept. 2015.
[10] P.-C. Hsu, B. M. H. Bruininks, D. Jefferies, P. Cesar Telles de Souza, J. Lee,
D. S. Patel, S. J. Marrink, Y. Qi, S. Khalid, and W. Im, "CHARMM-GUI
Martini Maker for modeling and simulation of complex bacterial mem-
branes with lipopolysaccharides," Journal of Computational Chemistry,
vol. 38, pp. 2354 -- 2363, Oct. 2017.
11
[11] M. J. Abraham, T. Murtola, R. Schulz, S. Pll, J. C. Smith, B. Hess, and
E. Lindahl, "GROMACS: High performance molecular simulations through
multi-level parallelism from laptops to supercomputers," SoftwareX, vol. 12,
pp. 19 -- 25, Sept. 2015.
[12] C. Arnarez, J. J. Uusitalo, M. F. Masman, H. I. Inglfsson, D. H. de Jong,
M. N. Melo, X. Periole, A. H. de Vries, and S. J. Marrink, "Dry Martini, a
Coarse-Grained Force Field for Lipid Membrane Simulations with Implicit
Solvent," Journal of Chemical Theory and Computation, vol. 11, pp. 260 --
275, Jan. 2015.
[13] C. M. Elliott, C. Grser, G. Hobbs, R. Kornhuber, and M.-W. Wolf, "A Vari-
ational Approach to Particles in Lipid Membranes," Archive for Rational
Mechanics and Analysis, vol. 222, pp. 1011 -- 1075, June 2016.
[14] P. B. Canham, "The minimum energy of bending as a possible explanation
of the biconcave shape of the human red blood cell," Journal of Theoretical
Biology, vol. 26, pp. 61 -- 81, Jan. 1970.
[15] W. Helfrich, "Elastic Properties of Lipid Bilayers: Theory and Possible
Experiments," Zeitschrift fr Naturforschung C, vol. 28, pp. 693 -- 703, Dec.
1973.
[16] E. A. Evans, "Bending Resistance and Chemically Induced Moments in
Membrane Bilayers," Biophysical Journal, vol. 14, pp. 923 -- 931, Dec. 1974.
[17] G. Khelashvili, B. Kollmitzer, P. Heftberger, G. Pabst, and D. Harries,
"Calculating the Bending Modulus for Multicomponent Lipid Membranes
in Different Thermodynamic Phases," Journal of Chemical Theory and
Computation, vol. 9, pp. 3866 -- 3871, Sept. 2013.
[18] N. Johner, D. Harries, and G. Khelashvili, "Implementation of a method-
ology for determining elastic properties of lipid assemblies from molecular
dynamics simulations," BMC Bioinformatics, vol. 17, p. 161, 2016.
12
|
1910.13512 | 1 | 1910 | 2019-05-15T08:10:50 | MnO2-gated Nanoplatforms with Targeted Controlled Drug Release and Contrast-Enhanced MRI Properties: from 2D Cell Culture to 3D Biomimetic Hydrogels | [
"physics.bio-ph",
"physics.chem-ph"
] | Multifunctional nanomaterials combining diagnosis and therapeutic properties have attracted a considerable attention in cancer research. Yet some important challenges are still to be faced, including an optimal coupling between these two types of properties that would be effective within complex biological tissues. To address these points, we have prepared novel nanoplatforms associating controlled drug delivery of doxorubicin and Magnetic Resonance Imaging (MRI) contrast-enhancement that exhibit high specificity towards cancer cells compared to normal cells and evaluated them both in 2D cultures and within 3D tissue-like biomimetic matrices. Methods: Nanoplatforms were prepared from hollow silica nanoparticles coated with MnO2 nanosheets and conjugated with the AS1411 aptamer as a targeting agent. They were fully characterized from a chemical and structural point of view as well as for drug release and MRI signal enhancement. Standard two-dimensional monolayer cultures were performed using HeLa and Normal Human Dermal Fibroblasts (NHDF) cells to testify targeting and cytotoxicity. Cellularized type I collagen-based hydrogels were also used to study nanoparticles behavior in 3D biomimetic environments. Results: The as-established nanoplatforms can enter HeLa cells, leading to the dissociation of the MnO2 nanosheets into Mn 2+ that enhanced T1 magnetic resonance signals and concomitantly release doxorubicin, both effects being markedly more significant than in the presence of NHDFs. Moreover, particles functionality and specificity were preserved when the cells were immobilized within type I collagen-based fibrillar hydrogels. Conclusion: The use of MnO2 nanosheets as glutathione-sensitive coatings of drug loaded nanoparticles together with surface conjugation with a targeting aptamer offers an effective strategy to obtain efficient and specific nanotheranostic systems for cancer research, both in 2D and 3D. The here-described tissue-like models should be easy to implement and could constitute an interesting intermediate validation step for newly-developed theranostic nanoparticles before in vivo evaluation. | physics.bio-ph | physics | Nanotheranostics 2018, Vol. 2
Ivyspring
International Publisher
Research Paper
403
NNaannootthheerraannoossttiiccss
2018; 2(4): 403-416. doi: 10.7150/ntno.28046
MnO2-gated Nanoplatforms with Targeted Controlled
Drug Release and Contrast-Enhanced MRI Properties:
from 2D Cell Culture to 3D Biomimetic Hydrogels
Yupeng Shi, Flavien Guenneau, Xiaolin Wang#, Christophe Hélary, Thibaud Coradin
Sorbonne Université, CNRS, Collège de France, Laboratoire de Chimie de la Matière Condensée de Paris, 75005 Paris, France.
#Current address: State Key Laboratory of Quality Research in Chinese Medicine and School of Pharmacy, Macau University of Science and Technology, Macau.
Corresponding author: Thibaud Coradin E-mail: [email protected]; Tel: +33-1-44274018
© Ivyspring International Publisher. This is an open access article distributed under the terms of the Creative Commons Attribution (CC BY-NC) license
(https://creativecommons.org/licenses/by-nc/4.0/). See http://ivyspring.com/terms for full terms and conditions.
Received: 2018.06.24; Accepted: 2018.09.08; Published: 2018.09.21
Abstract
Multifunctional nanomaterials combining diagnosis and therapeutic properties have attracted a
considerable attention in cancer research. Yet some important challenges are still to be faced, including
an optimal coupling between these two types of properties that would be effective within complex
biological tissues. To address these points, we have prepared novel nanoplatforms associating controlled
drug delivery of doxorubicin and Magnetic Resonance Imaging (MRI) contrast-enhancement that exhibit
high specificity towards cancer cells compared to normal cells and evaluated them both in 2D cultures
and within 3D tissue-like biomimetic matrices.
Methods: Nanoplatforms were prepared from hollow silica nanoparticles coated with MnO2 nanosheets
and conjugated with the AS1411 aptamer as a targeting agent. They were fully characterized from a
chemical and structural point of view as well as for drug release and MRI signal enhancement. Standard
two-dimensional monolayer cultures were performed using HeLa and Normal Human Dermal
Fibroblasts (NHDF) cells to testify targeting and cytotoxicity. Cellularized type I collagen-based hydrogels
were also used to study nanoparticles behavior in 3D biomimetic environments.
Results: The as-established nanoplatforms can enter HeLa cells, leading to the dissociation of the MnO2
nanosheets into Mn2+ that enhanced T1 magnetic resonance signals and concomitantly release
doxorubicin, both effects being markedly more significant than in the presence of NHDFs. Moreover,
particles functionality and specificity were preserved when the cells were immobilized within type I
collagen-based fibrillar hydrogels.
Conclusion: The use of MnO2 nanosheets as glutathione-sensitive coatings of drug loaded nanoparticles
together with surface conjugation with a targeting aptamer offers an effective strategy to obtain efficient
and specific nanotheranostic systems for cancer research, both in 2D and 3D. The here-described
tissue-like models should be easy to implement and could constitute an interesting intermediate
validation step for newly-developed theranostic nanoparticles before in vivo evaluation.
Key words: Silica nanoparticles, manganese dioxide, drug release, MR imaging, collagen hydrogels
Introduction
Despite the rapid improvement of modern
medicine, the early diagnosis and therapy of cancer is
still a challenge. The continuous development of
nanotechnology and the emergence of targeted
treatments provide an inequivalent opportunity in
this area. The past decade has witnessed the
engineering of many theranostic nanosystems, where
the integration of different imaging agents and
therapeutic drugs into a single nanoparticle (NP) has
made it possible to exhibit multiple functionalities
[1-4]. Extensive efforts have also allowed for the
identification of active targeting reagents, such as folic
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
404
acid, hyaluronic acid, aptamers, or transferrin, that
bind with high specificity to the cancerous cell
membrane [5-8]. At the same time, strategies were
developed to achieve a controlled drug release
triggered by
intrinsic physiological microenvi-
ronment changes (pH, redox, enzyme, heat, etc.)
and/or
light,
magnetic/electric field, ultrasound, etc.) [9-11]. All
those developments taken together,
it becomes
possible to simultaneously reduce the side-effects of
anticancer agents to normal tissues and enhance their
therapeutic efficiency [12]. Indeed, there still is some
room for large improvement, in terms of drug loading
and targeting efficiency.
(including
external
stimuli
limited
intrinsic
nanoparticles, Mesoporous
From an imaging perspective, signals of most
previously-reported nanosystems are "always on"
regardless of the absence or presence of the target
cells, resulting in a low contrast detection [13]. Among
available
Silica
Nanoparticles (MSNs) have been widely considered
for the delivery of anticancer drugs [11,14,15]. To
tackle MSNs
loading capacity,
Hollow MSNs (HMSNs) with rattle or hollow
structure have been prepared as they can efficiently
accommodate drugs not only
into mesoporous
channels of their shell but also within their internal
cavity [16-18]. Moreover, recent progress in the design
of gated HMSNs has shown some promise in the
theranostic
development
nanosystems. Different "gatekeepers",
such as
organic molecules,
and
nanoparticles, have been used, allowing pore-opening
under the stimulation of pH change, temperature,
nucleotides,
enzymes, glucose, or
photoirradiation
these gated
nanosystems usually require a highly-sophisticated
design that has a strong impact on synthesis time and
development cost, ultimately hindering their transfer
to
their demonstrated
functionality.
the market, despite
[19-21]. However,
biomacromolecules
controlled-release
antibodies,
Recently, MnO2 nanosheets, a 2D ultrathin
semiconductor material with wide applications in the
energy field [22], has attracted large attention as a
gatekeeper for drug nanocarriers [23-25]. They exhibit
a broad and intense absorption band at around 374
nm, making MnO2 nanosheets
efficient
broad-spectrum quencher [26]. Moreover, MnO2 can
be converted to Mn2+ via reduction by endogenous
glutathione (GSH), leading to decomposition of the
MnO2 nanosheets [27]. While manganese(IV) ions in
the nanosheets are shielded from water due to their
6-fold coordination with oxygen atoms, free solvated
Mn(II) ions are efficient T1 contrast agents for
Magnetic Resonance Imaging (MRI) [28]. GSH is an
essential endogenous antioxidant that has many
an
of
on
the
cellular functions and high GSH levels are implicated
in many diseases typically associated with cancer,
liver damage, or heart problems [29]. Therefore, much
attention has been paid to the use of GSH to design
intracellular controlled release systems [30]. In
addition, whereas the extracellular pH of normal
tissues and blood is constant at 7.4, the measured
extracellular values of most solid tumors range from
pH 6.5 to 7.2, such lower pH values being more
favorable for the reaction between MnO2 and GSH
[31].
several
MnO2@SiO2-based controlled-release nanosystems
with GSH-induced
contrast-enhanced magnetic
resonance signal have already been described [23,
32,33]. However, none of them took advantage of the
specificity of HMSNs so far.
principle,
Based
this
and
efficiency
control of
the drug
systems.
Another key challenge in this area lies in the
release using
precise
stimuli-responsive
tumor
Thus,
microenvironment, that has a crucial role in cancer
progression, exhibit several specific biological and
physico-chemical features that can be exploited to
trigger the drug release at the tumor site [34].
However,
stimulation of
nanosystems are often tested in two-dimensional (2D)
systems,
in which seeded cancer cells do not
experience the real microenvironment they find in
vivo, where physiological fluids, tissues as well as
interactions between cancer and stromal cells may
impair the drug delivery or its functionality. The
transfer from 2D data to in vivo during preclinical
studies may therefore be extremely time-consuming
and costly. Therefore, the design of three-dimensional
(3D) environments exhibiting some features of in vivo
tumors, such as
three-dimensional architecture,
cell-cell interaction and hypoxia should provide
highly useful tumor tissue in vitro models for testing
anticancer therapeutics [35,36]. In this context, the use
of type I collagen, the major protein in most animal
tissues, to prepare biomimetic constructs
is of
particular relevance [37-39]. Most importantly, it has
been shown that they could act as models to study the
interactions between nanoparticles and cells in a 3D
environment [40,41].
Herein, we propose, for the first time, a novel
and
for
the
facile strategy
fabrication of a
multifunctional nanoprobe
for contrast-enhanced
bimodal cellular imaging and targeted therapy, as
depicted in Figure 1. The nanoprobe consists of three
components, including (1) doxorubicin(DOX)-loaded
hollow mesoporous silica nanoparticles, (2) MnO2
nanosheets that act as both gatekeeper for DOX
release from HMSNs and contrast agent for MRI and
(3) cancer cell-targeting aptamers (AS1411), that bind
to nucleolin, a nucleolar phosphoprotein which is
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
overexpressed on the surface of certain cancer cells
[28, 42, 43]. We demonstrate that this multifunctional
HMSNs@MnO2(DOX)/apt
theranostic nanosystem
exhibits a synergistic delivery/imaging effect in
standard 2D conditions. We also show that the
functionality and specificity of these nanoplatforms
are preserved within 3D cellularized collagen
hydrogels, that can be useful biomimetic models for
evaluating nanotheranostic systems performance
before in vivo studies.
Methods
Materials and reagents
hydrate
(MES),
sodium
carbonate
All chemical reagents were analytical grade and
used without further purification. Cetyltrimethyla-
mmonium bromide (CTAB), tetraethyl orthosilicate
(Na2CO3), absolute
(TEOS),
ethanol, concentrated hydrochloric acid, ammonium
aqueous solution (25-28%), triethanolamine (TEA),
3-Triethoxysilylpropylamine
(APTES), potassium
(KMnO4), 2-(N-Morpholino)ethane-
permanganate
sulfonic
acid
doxorubicin
hydrochloride (DOX) and fluorescein isothiocyanate
(FITC) were purchased from Sigma-Aldrich. MilliQ
water (18 MΩ, Millipore, France) was used for the
preparation of the solutions and for all rinses.
4,6-diamidino-2-phenylindole (DAPI). Fetal bovine
serum, Dulbecco's Modified Eagle's Medium
(DMEM), trypsin, glutamine, penicillin-streptomycin
solution and AS1411 aptamer (5′-NH2-GGTGGTGGT
GGTTGTGGTGGTGGTGG-3′) were purchased from
ThermoFisher.
Preparation of HMSNs@MnO2/apt
Bare HMSNs were prepared using previously
405
reported methods [44,45]. To functionalize the particle
surface with amine groups, as-synthesized HMSNs
were first dispersed in 20 mL of toluene, followed by
addition of 1 mL of APTES. The system was sealed
and refluxed at 120 °C in oil bath for 12 h. Afterward,
the mixture was centrifuged and washed with ethanol
for several times to remove the residual APTES. Then,
20 mg of HMSNs-NH2 was dispersed in 4.2 mL MES
buffer (0.1M, pH 6.0) and then 0.8 mL of 5 mM
KMnO4 in water was added to the mixture under
ultrasonic condition. The resulting mixture was
sonicated
for another one hour during which
brown-black colloids were observed. Subsequently,
the raw product was collected by centrifugation,
washed several times with deionized water and
alcohol to remove any possible residual reactants, and
redispersed in 2 mL PBS solution (pH 7.4). The
physical adsorption of aptamer on HMSNs@MnO2
was carried out by mixing 0.8 mL of HMSNs@MnO2
(20 mg/mL) in PBS and 200 µL of AS1411 (1 µM).
After 3 h of incubation, the solution was centrifuged
at 6000 rpm, washed several times, and then
dispersed in PBS (pH 7.4) for further application. Full
protocol is available in Supplementary Material. The
preparation
hollow
mesoporous silica nanoparticles was performed by
adding FITC to the starting solution used for bare
HSMNs preparation.
Particles characterization
fluorescently-labeled
of
Transmission Electron Microscopy
(TEM)
studies were performed on a FEI Tecnai G2 Spirit
microscope operating at 120 kV. Scanning electron
microscopy (SEM) images were obtained on a Hitachi
S-3400N scanning electron microscope with a field
emission electron gun. Nitrogen sorption-desorption
Figure 1. Schematic illustration of the preparation of HMSN@MnO2(DOX)/apt and the drug release mechanism.
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
(DLS) was used
the hydrodynamic diameter of
isotherms were measured at 77 K with a
Micromeritics ASAP2010 analyzer. Fluorescence
images were recorded on a LEICA microscope.
to
Dynamic Light Scattering
determine
the
nanoparticles in Milli-Q water or in culture medium
with Mastersizer 3000 Particle Size Analyzer. The
reading was carried out at an angle of 90° to the
incident beam (632 nm). The Contin algorithm was
used to analyze the autocorrelation functions. X-ray
Photoelectron Spectroscopy (XPS) analyses were
performed with a PHOIBOS 100 spectrometer from
SPECS GmbH. Inductively Coupled Plasma Mass
Spectrometry (ICP-MS) was used to determine the Mn
content of the nanoparticles.
Drug loading and release in vitro
the
10 mg of HMSNs@MnO2 nanoparticles were
mixed with 1.5 mg of DOX in 1.5 mL of PBS/DMSO
(1:1) mixture solution and then stirred under dark
conditions for 24 h. Subsequently, the product was
collected by centrifugation and washed several times
with PBS to remove the free DOX. Then 100 µL of
AS1411 (1 µM) in PBS was added, and after 3 h of
resulting HMSNs@MnO2(DOX)/apt
stirring,
nanoparticles were collected by centrifugation and
re-dispersed in PBS solution for subsequent use. To
evaluate the DOX loading, the remaining DOX
molecules
solutions were
determined by fluorescence spectroscopy (λex= 500
nm, λem = 590 nm). The loading of the nanoparticles
was expressed as the mass percentage of DOX with
respect to the total mass of HMSNs@MnO2(DOX)/apt
nanoparticles.
supernatant
the
in
In
vitro
from
DOX
release
the
HMSNs@MnO2(DOX)/apt nanoparticles was studied
in PBS buffer in absence or presence of GSH at pH
values of 7.4 and 5.5. For each release study, 1.0 mL of
(1.0 mg.mL-1) were
DOX-loaded nanoparticles
dispersed inside a dialysis bag that was soaked in 9.0
mL PBS and shaken at room temperature. At selected
time intervals, the sample was collected and 2 mL of
solution outside the dialysis bag was removed. Then,
2 mL of fresh PBS buffer was added. The removed
solution was properly diluted and the amount of DOX
molecules present was measured by fluorescence
spectroscopy. The same amount of the DOX solution
was introduced in the dialysis bag and its diffusion
kinetics was used as control for drug release studies.
Cell culture
Normal Human Dermal Fibroblasts (NHDF)
and HeLa cells were cultured in complete cell culture
medium (DMEM with GlutaMAX™, without phenol
red supplement, with 10% fetal serum, 100 U.mL-1
406
penicillin, 100 μg.mL-1 streptomycin). Tissue culture
flasks (75 cm2) were kept at 37 °C in a 95% air: 5% CO2
atmosphere. Before confluence,
the cells were
removed from culture flasks by treatment with 0.1%
trypsin and 0.02% EDTA. Cells were rinsed and
resuspended in the above culture medium before use.
In vitro cytotoxicity assays
First, HeLa were seeded in a 24-well plate at a
density of ~5×104 cells/well overnight to allow cell
attachment onto the surface of the wells. Then, 0.5 mL
of fresh medium containing various concentrations of
HMSNs@MnO2(DOX)/apt was added into the wells.
After incubation for 24 h, cell activity was evaluated
by the Alamar Blue assay. Control experiments were
performed by incubating HeLa cells with free DOX
and HMSNs@MnO2(DOX) at various drug contents or
HMSNs@MnO2(DOX)/apt at equivalent particle
concentration for 24 h. To confirm the cytotoxicity of
DOX released from the nanoplatform under GSH,
similar control experiments were performed with
NHDF cells incubated in presence of 5 mM GSH.
Cell uptake studies by fluorescence imaging
NHDFs and HeLa cells were seeded in a round
glass disk that was inserted in 24-well plate (~5×104
cells per well) and cultured for 24 h. The cell medium
was removed, and then cells were incubated with 0.5
mL of fresh cell medium containing FITC-labeled
HMSNs@MnO2(DOX)/apt nanoparticles for another
12 h. After medium removal, the cells were washed
with PBS for several times, and then 0.5 mL of fresh
cell medium with or without 5 mM GSH was added
and incubated for another 3 hours. Fixation of the cells
was carried out using 4% (v/v) paraformaldehyde
(PFA) for 2 h and rinsed three times using PBS. After
washing, cellular nuclei were stained with 1% (v/v)
solution of DAPI in PBS buffer for 20 min and washed
with PBS three times. Last, cell imaging was carried
out by fluorescence microscopy. For the HeLa cells
imaging, the procedure was similar, except that
HMSNs@MnO2(DOX) or HMSNs@MnO2(DOX)/apt
particles were added and no GSH treatment was
performed.
Measurement of MRI relaxation properties
Imaging of MRI phantoms measurement was
performed with a Bruker Avance III HD 300
spectrometer, equipped with a 10 mm micro imaging
probe, having a maximum gradient capacity of 3 T.m-1
in the x, y and z directions. The Multi Slice Multi Echo
(MSME) pulse sequence was used, acquiring 1 echo
but 8 slices with a slice thickness of 1 mm. The size of
the images was: 128×128 with a Field of View of
9.5×9.5 mm, resulting in a voxel size of 74×74 µm.
Images were acquired with 1 scan and T1 weighting
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
was obtained by using a short repetition time (TR) of
200 ms and using the shortest possible echo time (TE),
namely 4.92 ms, in order to minimize the effect of T2
relaxation.
Preparation of 3D collagen-based tumor
models
estimated
concentration was
Type I collagen was purified from rat tails and
the
final
by
hydroxyproline titration, as previously described [37].
Tubes separately filled with collagen solution (2
mg.mL-1 in 17 mM acetic acid), whole cell culture
medium, and 0.1 M NaOH were kept in ice bathes for
1 h before preparation to slow down the gelling
kinetics of collagen. First, 500 μL of collagen solution
and 400 μL of culture medium were added to a 1.5 mL
tube and vortexed vigorously. After addition of 30 μL
of 0.1 M NaOH and strong vortexing, 125 μL of the
NHDF or HeLa cell suspension at a density of 106
cells/mL was added and mixed homogeneously.
Then 0.9 mL was sampled from the mixture and
deposited into a 24-well plate. The plate was then
incubated at room temperature for 10 min for
complete gelling of collagen.
Cytotoxicity and functionality within 3D
models
Two sets of experiments were designed [41]. In a
first approach, selected particles were introduced in
the cell suspension prior to their addition to the
collagen solution, at a final concentration of 100
µg.mL-1. Then collagen gels were formed by pH
increase to 7. After 48 h of incubation, the cell viability
was assessed in the same way as for 2D tests except
that 800 μL of water were first added to the collagen
gel, left for 0.5 h at room temperature in order to
extract the Alamar Blue solution trapped in the gel,
and then collected for the absorbance measurements.
Particle internalization was also studied in the same
configuration. For this, after the 48 h incubation
period, the gels were rinsed 3 times with PBS and
fixed with 4% paraformaldehyde overnight. Next, the
fixed samples were dehydrated in ethanol and
butanol, and incorporated in paraffin to be able to
obtain 10 μm histological sections with a manual
microtome. Before observation,
the as-obtained
samples were immersed in toluene, ethanol, and then
water for rehydration. The cell nuclei were stained
with DAPI for 10 min and rinsed with PBS before
observation. The second approach consisted in adding
1.0 mL of a 0.2 mg.mL-1 suspension of
the
HMSNs@MnO2(DOX)/apt particles onto the surface
of particle-free cellularized collagen gels. After 3
407
hours of contact, the MR imaging experiments were
performed as described above.
Statistical Analysis
Graphical results are presented as mean ± SD
(standard deviation). Statistical significance was
assessed using one-way analysis of variance
(ANOVA) followed by Tukey (compare all pairs of
groups) or Dunnett (compare a control group with
other groups) post-hoc test. The level of significance
in all statistical analyses was set at a probability of P <
0.05.
Results and discussion
Preparation and characterization of the
nanoplatforms
sSiO2.
The
pre-coated
The HMSN@MnO2/apt nanoparticles were
prepared according to the process depicted in Figure
1. In brief, monodisperse solid SiO2 nanoparticles
(sSiO2) were first prepared using a modified Stöber
method. The prepared sSiO2 were then coated with a
CTAB/SiO2 shell via base-catalyzed hydrolysis of
TEOS and condensation of silica onto the surface of
resulting
CTAB
sSiO2@CTAB/SiO2 particles were simultaneously
treated with Na2CO3 to remove sSiO2, and NH4NO3 to
remove CTAB, resulting in HMSNs with hollow cores
and penetrating pore channels [44]. APTES was
grafted on the surface of nanoparticles to get
HMSN-NH2. Finally, ultrathin MnO2 nanosheets were
formed onto the surface of HMSN-NH2 thanks to the
reaction with MES and KMnO4
to obtain
HMSN@MnO2. DOX was then loaded inside or on the
surface of the nanoparticles by the impregnation
scheme described above. To end, AS1411 aptamers
were attached on the surface of MnO2 nanosheets
thanks to nucleobase-mediated physisorption [28].
As shown in the TEM image of Figure 2a, the
prepared hollow mesoporous HMSN nanoparticles
exhibit a uniform diameter of ~140 nm and form a
colorless well-dispersed suspension in deionized
water (DI) (Figure 2d). The diameter measured by
DLS, Dm, is about 200 nm in DI, which is slightly
larger than the TEM data and may reflect some
aggregation (Table 1). The contrast between shell and
core of the silica nanospheres confirms their hollow
structure. The shell of the HMSN with a thickness of
~25 nm displays an obvious mesoporous silica
structure generated by the removal of pore templates.
In DI water, the obtained HMSNs exhibited a negative
zeta potential ζ of -18.4 mV, as expected for silica
surfaces (Table 1).
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
408
Table 1. Mean diameter Dm from DLS with corresponding
polydispersity index (PDI) and Zeta potential ζ in deionized water
Sample
HMSNs
HMSNs-NH2
HMSNs@MnO2
HMSNs@MnO2(DOX)
HMSNs@MnO2(DOX)/apt
Dm DLS (nm)
200 ± 18
256 ± 27
319 ± 41
316 ± 26
248 ± 15
PDI
0.254
0.317
0.579
0.500
0.279
ζ (mV)
-18.4 ± 0.5
+10.4 ± 0.3
-15.4 ± 0.2
-7.3 ± 0.8
-16.1 ± 0.3
type
typical
The N2 adsorption-desorption isotherm of the
HMSN showed a
IV curve of
surfactant-assisted mesoporous silica with a double
strong and sharp adsorption step at intermediate
relative partial pressure values around 0.4 (Figure S1).
This feature is associated with nitrogen condensation
inside the mesopores by capillarity. The application of
the Brunauer-Emmett-Teller (BET) model resulted in
a high value for the specific surface of 840 m2.g-1. In
parallel, the Barrett-Joyner-Halenda (BJH) model
applied on the adsorption branch of the isotherms led
to an average pore diameter of 3.89 nm (Figure S1).
The absence of a hysteresis loop in this interval of
P/P0, as well as the narrow BJH pore size distribution,
also suggested the existence of uniform cylindrical
channels throughout the material.
The successful surface functionalization of the
hollow mesoporous silica nanoparticles using APTES
could be checked by the increase of the zeta potential
that reached a positive value (+10.4 mV) (Table 1).
Upon contact of these HMSN-NH2 particles with MES
and KMnO4, a visual color change from white to
brown was observed (Figure 2e). After extensive
to
-15.4 mV after
functionalization,
washing, UV-vis spectra clearly evidenced a strong
absorption band at 400 nm, that can be attributed to
the MnO2 nanosheets, whose intensity increase with
the initial KMnO4 concentration (Figure S2). The
formation of composite nanoparticles (HMSN@MnO2)
with a rough capping layer was confirmed by TEM
images (Figure 2b). Moreover, the average diameter of
HMSN@MnO2 was changed from 200 nm to ca. 320
nm, accompanied by a polydispersity index (PDI)
increase from 0.254 to 0.579, which suggests that the
MnO2 coating induces a slight tendency for the
particles
to aggregate. The zeta potential of
nanoparticles was also modified, from +10.4 mV
down
in
agreement with values in the literature [46]. After
coating, FTIR spectra also evidenced a peak at 1413
cm-1 characteristic of Mn-O vibrations in MnO2
(Figure S3). Sorption isotherms also confirmed the
coating of the nanoparticles. HMSN@MnO2 presented
flat sorption curves when compared to those of
HMSN (at P / P0 = 0.4) (Figure S1), thus indicating a
significant pore blocking. Nonetheless, HMSN@MnO2
showed an additional inflection at P/P0 > 0.8, that can
correspond to secondary macropores resulting from
the MnO2 capping
layer with rough surface.
HMSN@MnO2 showed a weak peak at pore diameter
of 2.35 nm, confirming the partial pore blocking effect
layer. Finally, when
of
HMSN@MnO2 were put in contact with GSH (10 mM),
the color of the suspension changed back from brown
to white, indicative of the disappearance of the MnO2
layer (Figure 2f). TEM images clearly confirmed the
regeneration of
the smooth surface and clear
mesoporous structure of HMSN (Figure 2c).
the MnO2
capping
Figure 2. TEM images of (a) HMSNs, (b) HMSN@MnO2/apt, (c) HMSN@MnO2/apt after GSH treatment and the corresponding digital images from left to right (d,
e, f).
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
409
DOX was then selected as a guest molecule to
confer anticancer properties to the nanoparticles. The
loading was performed by a simple impregnation
route in a mixed aqueous/organic medium. The
resulting DOX-loaded HMSN@MnO2 particles had a
size distribution similar to unloaded particles but the
zeta potential changed from -15.4 mV to -7.3 mV,
suggesting that at least part of the drug is adsorbed on
the MnO2 coating. The UV-vis spectra of the loaded
particles showed an additional absorption band at ca.
480 nm similar to the one of DOX (Figure S2) and two
peaks at 1624 cm-1 and 1580 cm-1 belonging to the
benzene ring structure of DOX could be observed on
the FTIR spectra of the HMSNs@MnO2(DOX) sample
(Figure S3). The DOX loading of HMSN@MnO2, as
estimated by fluorescent emission measurements of
the supernatant after impregnation, reached ca. 80
µg.mg-1.
As a final step, the anti-nucleolin AS1411 was
the HMSNs@MnO2(DOX) particles
adsorbed on
which resulted in a significant decrease in the average
particle diameter and of the PDI (Table 1). This can be
correlated with the more negative value of the zeta
potential (from -7.6 mV without aptamer to -16.1 mV
after AS1411 adsorption), that results in an enhanced
particle colloidal stability thanks to electrostatic
repulsion. FTIR spectra showed an additional peak at
1742 cm-1 that can correspond to the C=O stretching
vibration of the aptamer (Figure S3).
To characterize the chemical compositions of the
final HMSNs@MnO2(DOX)/apt nanoparticles, their
element composition was determined by XPS (Figure
3a). Mn, Si, C, O and N peaks were observed. Indeed,
considering that the analysis depth of XPS analysis is
10 nm at maximum, it only probes the outer surface of
the particles. It is therefore not surprising that the Si
At% is rather low (19.8 %), corresponding to ca. 60
wt% of SiO2. In contrast, the contribution of organic
molecules (DOX and AS1411) (> 35 % C) is enhanced
compared to the whole particle volume. The Mn
relative amount is 0.9 At% corresponding to ca. 5 wt %
MnO2. Considering that manganese oxide nanosheets
have been reported to be ca. 1 nm thick, a ca. 10 wt %
amount would have been expected, suggesting that
Figure 3. XPS analysis of the HMSNs@MnO2(DOX)/apt nanoparticles: (a) full XPS spectrum with atomic analysis and deconvoluted signals with proposed
attributions at the (b) Mn2p, (c) Si2p and (d) O1s levels.
http://www.ntno.org
410
that
the
it can be suggested
% of the DOX content was released within 10 h and
then a slower release, up to ca. 35-40 %, occurred in
the next 40 h. When GSH was added in acidic
conditions, the shape of the release curve was quite
similar (i.e. fast release during 10 h followed by
slower release) but the amount of released DOX at the
end of the first phase was much higher (ca. 50 %) than
in the two previous conditions and the total release
after 50 h was larger than 60 %. As pointed out earlier,
DOX molecules may be located on the MnO2 surface,
inside the porous silica shell and within the particle
empty core. Assuming that the GSH-free neutral pH
conditions induce no significant dissolution of the
MnO2 nanosheets, then the measured low release (less
than 10 % of the initial dose) should correspond to
surface-adsorbed DOX molecules. Upon GSH
addition or acidification, much more DOX molecules
are released following a two-step process. It is worth
emphasizing that silica nanoparticles are less soluble
in acidic than in neutral pH conditions and not
sensitive to GSH unless specifically modified [50].
Thus
first step
corresponds to the release of DOX molecules both
adsorbed on the surface and trapped in the porous
shell due to the MnO2 coating dissolution, while the
second one would sign for the slow diffusion of the
drug located within the particles. However, when
both acidic conditions and GSH addition are
combined, the overall release profile, and especially
the duration of the fast-release period,
is not
significantly modified but more DOX is initially
released. This may be interpreted considering that, in
the MnO2
these
degradation is efficient enough to
uncap all pores before diffusion of
DOX molecules from the inside of
the particle starts. In contrast, in
GSH-only
pH-only
conditions, the MnO2 degradation
process is less efficient and only part
of the pores are uncapped before
diffusion
the
inside
compartment becomes
effective.
This points out that the combination
of reductive conditions due to GSH
presence and acidic media where
MnO2 is less stable is the most
favorable for DOX release.
conditions,
or
acidic
from
Nanotheranostics 2018, Vol. 2
the MnO2 coating is not homogeneous over the whole
particle surface. At the Mn2p level (Figure 3b), the two
peaks at 642 eV and 653.8 eV unambiguously sign for
the presence of MnO2, as also confirmed by the peak
at 530.1 eV on the O1s spectrum [47] (Figure 3d). The
peaks at 103.2 eV on the Si2p spectrum (Figure 3c) and
532.4 eV on the O1s spectrum are characteristic of
hydrated silica [48]. Deconvoluted spectra at the C1s
and N1s levels evidenced the presence of C-C, C-O,
C=O, C-N and N-C=O groups (Figure S4). The latter is
a clear indication of the presence of the aptamer on
the particle surface [49], while the others can belong to
both AS1411 and DOX.
Drug release and MRI imaging properties of
the nanoplatforms
The feasibility of MnO2 degradation when
treated with GSH and pH was first investigated. As
shown in the Figure S5, the HMSNs@MnO2/apt
nanoplatforms were stable at neutral pH but, after
interaction with GSH, the color of the system quickly
changed from brown to colorless, with a more
pronounced effect when GSH concentration was
increased from 5 mM to 10 mM. Acidic pH slightly
destabilized the nanoplatform, due to the intrinsic
instability of MnO2 in these conditions.
These observations were quantitatively assessed
by monitoring DOX release in different conditions
(Figure 4). In the absence of GSH, the DOX release
was less than 10 % over 50 h at pH 7.4. When the
MnO2 coating was destabilized, either in acidic
conditions or by addition of GSH at neutral pH, ca. 20
Figure 4. Cumulative release of the HMSNs@MnO2(DOX)/apt in different conditions: (blue) pH 7.4,
(black) pH 5.5, (green) pH 7.4+ 5 mM GSH, (red) pH 5.5 + 5 mM GSH. Control samples: (pink) free DOX
diffusion from dialysis bag and (pale green) uncoated HSMNs(DOX) in PBS solution (pH =7.4).
Because the reduction of MnO2
by intracellular GSH could produce
a large amount of Mn2+, which are
efficient T1 contrast agents,
the
nanoplatform could also afford
GSH/pH-activated detection by
MRI. To verify this hypothesis,
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
411
In
with
improved
T1-weighted MR
images of HMSN@MnO2/apt
aqueous suspensions under different conditions were
first performed. As shown in Figure 5a, the MRI
signal of the HMSN@MnO2/apt sample in PBS at pH
7.4 didn't exhibit an obvious T1-MRI signal difference
compared to that of deionized water. A significant
T1-MRI signal enhancement could be observed at pH
5.5. Adding 5 mM GSH in neutral and then acidic
the T1-MRI signal
conditions further
intensity. The r1 relaxivity of the nanoplatform
suspensions was also measured (Figure 5b) for
HMSN@MnO2/apt at pH 5.5 in the presence of GSH,
reaching a value of 9.25 mM-1.s-1 which is remarkably
higher than that of the GSH-free solution at pH 7.4 (r1
= 1.68 mM-1.s-1). Importantly such r1 values are higher
than those reported for other Mn-based nanocontrast
8.26 mM−1.s−1;
agents
Mn-MSNs, 2.28 mM−1.s−1 and MnO nanoplates, 5.5
mM−1.s−1) [51-53]. The above results suggest that the
MnO2 nanosheets can efficiently be converted into
Mn2+ in a mildly acidic/GSH environment, explaining
the rapid enhancement of the longitudinal relaxation
rate
conclusion,
HMSNs@MnO2/apt present a clear pH/GSH-
responsive T1-MRI performance.
concentration.
nanoparticles,
(Mn3O4
checked
(Figure S6). Then
In a second step, it was necessary to check that
the designed nanoplatforms could efficiently kill
cancer cells while being safe for normal cells. First, it
that drug-free HMSNs@MnO2
was
nanoparticles had no impact on HeLa cells viability
after 24 h of contact, as monitored by the Alamar Blue
test
the cytotoxic effect of
HMSNs@MnO2(DOX) and HMSNs@MnO2(DOX)/apt
particles on HeLa cells viability was studied with the
same method and compared to free DOX at the same
drug concentrations. As shown on Figure 6a, no
significant difference in cytotoxicity between free and
encapsulated drug could be observed, with a
continuous decrease of HeLa cells activity with
increasing DOX dose, down to ca. 50 % for the highest
investigated drug dose (16 µg.mL-1). The presence of
the aptamer had no influence either. In a step further,
the effect of HMSNs@MnO2(DOX)/apt particles
concentration on HeLa and NHDF cells was
monitored (Figure 6b). In the 5-160 µg.mL-1 range,
HMSNs@MnO2(DOX)/apt particles have no obvious
impact on NHDF while they are toxic for HeLa cells
(i.e. viability below 80 %) for particle concentrations of
10 µg.mL-1 or more. Importantly, if GSH (5 mM) was
present in the NHDF culture medium, then high
cytotoxicity was also measured, confirming its ability
to trigger DOX release.
To confirm this hypothesis, the internalization of
the HMSNs@MnO2(DOX)/apt was studied. As
For
In
this
localization.
(red signal)
in Figure 7a-d, accumulation of
shown
the
nanoplatforms, visualized by the green fluorescence
of encapsulated FITC, in contact or within the cells
could be observed for NHDF, without or with GSH,
and HeLa cells. In the latter case, bright green
aggregates are located near the nuclei, which appear
blue after DAPI staining, strongly supporting their
intracellular
sample,
colocalization of the nanoplatform (green signal) and
the DOX molecules
is evidenced.
Noticeably, when no aptamer was present on the
surface, low particle accumulation was observed and
no clear red signal could be imaged (Figure 7d). For
NHDF cells without GSH, the red areas are hardly
seen whereas they are more clearly visible/detectable
when GSH was present. It is important to point out
that MnO2 nanosheets have previously been shown to
efficiently quench the fluorescence of dyes in the
UV-visible range [54]. Thus the observation of red
signals on the fluorescence images indicates that DOX
is no longer interacting with the manganese oxide
coating, probably because the MnO2 nanosheets were
reduced to Mn2+ by GSH. Noticeably, in the case of
HMSNs@MnO2(DOX)/apt interacting with HeLa, the
strong co-localization of FITC and DOX inside the
cells would suggest that the MnO2 layer has been
degraded but that part of the drug remains inside the
particles
for
HMSNs@MnO2(DOX)/apt interacting with NHDF in
the presence of GSH, many particles visualized by
monitoring the green fluorescence do not exhibit a red
fluorescence (Figure 7b). This can be attributed to the
fact that the MnO2 layer is degraded by GSH leading
to DOX release before the internalization process.
While viability tests have evidenced NHDF cells
death in these conditions (blue bars on Figure 6b),
indicating that released drug could be uptaken, the
dilution of DOX within the cytoplasm may lead to a
low fluorescence intensity, precluding its detection.
(Figure 7c).
contrast,
sharp
Altogether,
HMSNs@MnO2(DOX)/apt
nanoplatforms have the ability to be internalized by
HeLa cells with a clear benefit of the presence of the
targeting aptamer on the efficiency of this process.
After degradation of the MnO2 coating, that may be
due to both acidic pH of the lysosome and enhanced
GSH expression, the particles can thereby release
DOX molecules at the vicinity of the nucleus,
resulting in an efficient cancer cell killing. At the same
time, the formation of Mn2+ ions should allow for the
detection of the cancer tissue via T1-weighted MRI.
Very importantly, our experiments also indicate that
these nanoplatforms are safe for normal NHDF cells
in the absence of GSH.
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
412
Figure 5. (a) T1-weighted solution MR images of HMSN@MnO2/apt under different conditions. (i: H2O; ii: pH 7.4; iii: pH 5.5; iv: pH 7.4 + GSH 5 mM and v: pH 5.5
+ GSH 5 mM) after incubation at 37 °C for 1 h. (b) 1/T1 versus Mn concentrations for HMSN@MnO2/apt at (black) pH 7.4 and (red) pH 5.5 + 5 mM GSH.
Figure 6. (a) HeLa cell viability after incubation with DOX, HMSNs@MnO2(DOX) and HMSNs@MnO2(DOX)/apt for 24 h as a function of DOX dose. (b) Cell
cytotoxicity of HMSNs@MnO2(DOX)/apt after 24 h of incubation with NHDF, HeLa and NHDF cells treated with GSH as a function of particle concentration.
Figure 7. Fluorescent images of (a) NHDF cells incubated with HMSNs@MnO2(DOX)/apt nanoparticles, (b) NHDF cells incubated with HMSNs@MnO2(DOX)/apt
nanoparticles and 5 mM GSH, (c) HeLa cells incubated with HMSNs@MnO2(DOX)/apt nanoparticles, (d) HeLa cells incubated with HMSNs@MnO2(DOX)
nanoparticles.
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
413
Figure 8. (a) Viability of NHDF and HeLa cells within different types of 3D collagen/silica nanocomposites after 24 h of incubation. (b) SEM image of nanoparticles
(red arrow) interacting with HeLa cells (white arrow) within collagen hydrogels.
Functionality of nanoplatforms in 3D
biomimetic collagen matrices
At
are
with
usually
this stage of similar studies,
undertaken
in vivo
evaluations
using
tumor-bearing animals. However, in this work, we
chose instead to go deeper into the understanding of
the impact of a 3D environment on the functionality of
the nanoplatform, a point that was not previously
assessed in the literature. For this we prepared type I
collagen matrices in conditions compatible with cell
immobilization and particle
encapsulation or
diffusion [40,41].
the control group,
types of cells. After DOX-loading,
To check whether the nanoparticles preserved
their efficiency and selectivity as cancer cell killing
agents, the viability of NHDF and HeLa cells
HMSNs@MnO2,
co-encapsulated
HMSNs@MnO2(DOX), and HMSNs@MnO2(DOX)/
apt at a dose of 100 μg.mL-1 within the collagen
matrices was studied
(Figure 8a). After 24 h,
compared
the unloaded
HMSNs@MnO2/apt showed very little toxicity for
both
the
HMSNs@MnO2(DOX)/apt
neither
showed much effect the viability of NHDF cells but
exhibited a clear cytotoxic effect on HeLa cells, that
was independent of the presence of the aptamer on
the surface. Importantly, the decrease of HeLa cell
viability was more significant than
in the 2D
experiments (see Figure 6b above), suggesting that the
confinement of the cells and particles within the
matrices strongly favors their interaction. As a matter
of fact, SEM imaging allowed for the observation of
nanoparticle aggregates at the close vicinity of the
HeLa cells (Figure 8b). In such conditions, our results
suggest that it may be no longer necessary to
nanoparticles
to
specifically target the cells by using the AS1411
aptamer.
Fluorescence imaging of the encapsulated HeLa
cells also evidenced the co-localization of DOX with
the nuclei-staining DAPI, signing for their successful
internalization (Figure 9). However red fluorescence
signals could also be visualized outside the cells,
that some MnO2 degradation also
suggesting
occurred outside the cells.
suggested partial destabilization of
Then, to evaluate whether such an ultrasensitive
GSH-responsive is suitable for in vivo tumor imaging
by magnetic resonance, another configuration was
used where selected nanoparticles were placed onto
cellularized collagen hydrogels and left to diffuse for
3 h. As shown in Figure 10, a weak MRI signal was
obtained for all samples prepared in the absence of
HeLa cells or in the presence of NHDF cells, in
agreement with the fluorescence image in Figure 9
that
the
nanoplatforms in contact with the collagen network.
While comparable results were obtained when the
aptamer-free particles were added to HeLa cells,
HMSNs@MnO2(DOX)/apt particles
a
led
significant
intensity of
the
T1-weighted images, that was increased when the
initial particle concentration was increased. These
results clearly demonstrate that the functionality and
the HMSNs@MnO2(DOX)/apt
specificity
nanoplatforms is preserved within the 3D models.
Interestingly, in this configuration, particles are not
initially confined at the proximity of the cells but must
find their way through the collagen network and bind
to the HeLa cells. This explains why, in contrast to the
previous system, the presence of the aptamer does
have a significant beneficial effect on the MRI signal.
In these 3D models, determination of the r1 value is
http://www.ntno.org
enhancement of
of
to
Nanotheranostics 2018, Vol. 2
difficult as it would require to determine the precise
amount of HMSNs@MnO2(DOX)/apt
that have
diffused inside the matrices and interacted with the
cells. However, when comparing Figure 5a and Figure
10, it quite clear that the T1 relaxation time is lower in
the matrix than in suspension and that the decreased
stability of the nanoplatforms in the collagen gels
ultimately impacts on its contrasting efficacy, i.e. its
ability to differentiate safe and tumor tissues.
Conclusion
As a conclusion, a multicomponent hybrid
nanoplatform
combined MRI
contrast-enhanced imaging and targeted therapy
through controlled drug delivery was designed,
allowing
for
414
prepared and fully characterized in vitro in 2D and 3D
configurations. Of particular importance is the fact
that their high selectivity towards HeLa cancer cells
compared to normal human cells was preserved when
they were able to diffuse through cellularized
biomimetic collagen hydrogels. This strongly suggests
that such hydrogels, that are (i) simple to prepare, (ii)
suitable for traditional cell biology studies, (iii)
amenable to MRI measurements and (iv) can easily be
implemented in terms of structure, composition and
type of immobilized cells, should be useful for a
preliminary
newly-developed
nanotheranostics before entering in vivo studies that
are time-consuming, costly and are raising increasing
ethical and public concern.
screening
of
Figure 9. Fluorescent images of HeLa cells incubated with HMSNs@MnO2(DOX)/apt within a collagen hydrogel for 24 h, (a) bright field; (b) blue channel (DAPI);
(c) red channel (DOX) and (d) merge.
Figure 10. T1-weighted MR images of collagen gels (a) without cells, (b,c) with NHDF cells and (d-f) with HeLa cells after 3h diffusion of (b,d) 0.2 mg.mL-1
HMSNs@MnO2(DOX) (c,e) 0.2 mg.mL-1 HMSNs@MnO2(DOX)/apt and (f) 0.4 mg.mL-1 HMSNs@MnO2(DOX)/apt.
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
415
Acknowledgment
Y.S. PhD grant was funded by the China
Scholarship Council.
Supplementary Material
Full protocol for particle synthesis and supplementary
figures. http://www.ntno.org/v02p0403s1.pdf
Competing Interests
The authors have declared that no competing
interest exists.
References
1. Xie J, Lee S, Chen X. Nanoparticle-based theranostic agents. Adv Drug Deliv
Rev. 2010; 62: 1064-79.
2. Kelkar SS, Reineke TM. Theranostics: Combining Imaging and Therapy.
Bioconjugate Chem. 2011; 22:1879-1903.
3. Muthu MS, Leong DT, Mei L, Feng S-S. Nanotheranostics Application and Further
Development of Nanomedicine Strategies for Advanced Theranostics. Theranostics.
2014; 4:660-77.
4. Han L, Zhang X-Y, Wang Y-L, Li X, Yang X-H, Huang M, et al.
inorganic
Redox-responsive
nanoplatforms
nanomaterials. J Control Release. 2017; 259: 40-52.
theranostic
based
5. Otsuka H, Nagasaki Y, Kataoka K. PEGylated nanoparticles for biological and
on
pharmaceutical applications. Adv Drug Deliv Rev. 2003; 55: 403-19.
7.
6. Xiang D, Shigdar S, Qiao G, Wang T, Kouzani AZ, Zhou S-F, et al. Nucleic acid
aptamer-guided cancer therapeutics and diagnostics: the next generation of
cancer medicine. Theranostics. 2015; 5: 23-42.
Jiang T, Zhang Z, Zhang Y, Lv H, Zhou J, Li C, et al. Dual-functional
liposomes based on pH-responsive cell-penetrating peptide and hyaluronic
acid for tumor-targeted anticancer drug delivery. Biomaterials. 2012; 33:
9246-58.
8. Dixit S, Novak T, Miller K, Zhu Y, Kenney ME, Broome A-M. Transferrin
receptor-targeted theranostic gold nanoparticles for photosensitizer delivery
in brain tumors. Nanoscale. 2015; 7: 1782-90.
9. Mura S, Nicolas J, Couvreur P. Stimuli-responsive nanocarriers for drug
delivery. Nature Mater. 2013; 12: 991-1003.
10. Cheng R, Meng F, Deng C, Klok H-A, Zhong Z. Dual and multi-stimuli
responsive polymeric nanoparticles for programmed site-specific drug
delivery. Biomaterials. 2013; 34: 3647-57.
11. Tarn D, Ashley CE, Xue M, Carnes EC, Zink JI, Brinker CJ. Mesoporous silica
nanoparticle nanocarriers: biofunctionality and biocompatibility. Acc Chem
Res. 2013; 46: 792-801.
12. Fan Z, Fu FP, Yu H, Ray PC. Theranostic nanomedicine for cancer detection
and treatment. J Food Drug Anal. 2014, 22:3-17
13. Hu X, Liu G, Li Y, Wang X, Liu S. Cell-penetrating hyperbranched
polyprodrug amphiphiles for synergistic reductive milieu-triggered drug
release and enhanced magnetic resonance signals. J Am Chem Soc. 2015; 137:
362-8.
14. Tang F, Li L, Chen D. Mesoporous silica nanoparticles: synthesis,
biocompatibility and drug delivery. Adv Mater. 2012; 24: 1504-34.
from
15. Kim J, Kim HS, Lee N, Kim T, Kim H, Yu T, et al. Multifunctional uniform
nanoparticles composed of a magnetite nanocrystal core and a mesoporous
silica shell for magnetic resonance and fluorescence imaging and for drug
delivery. Angew Chem Int Ed. 2008; 47: 8438-41.
16. Chen Y, Chen H, Zeng D, Tian Y, Chen F, Feng J, et al. Core/shell structured
hollow mesoporous nanocapsules: a potential platform for simultaneous cell
imaging and anticancer drug delivery. ACS Nano. 2010; 4: 6001-13.
Controlled Drug Release
Silica
Sphere/Polyelectrolyte Multilayer Core -- Shell Structure. Angew Chem Int Ed.
2005; 44: 5083-7.
17. Zhu Y, Shi J, Shen W, Dong X, Feng J, Ruan M, et al. Stimuli‐Responsive
18. Zhao W, Chen H, Li Y, Li L, Lang M, Shi J. Uniform rattle‐type hollow
sustained‐release property. Adv Funct Mater. 2008; 18: 2780-8.
19. Cheng Y-J, Luo G-F, Zhu J-Y, Xu X-D, Zeng X, Cheng D-B, et al.
Enzyme-induced and tumor-targeted drug delivery system based on
multifunctional mesoporous silica nanoparticles. ACS Appl Mater Interfaces.
2015; 7: 9078-87.
magnetic mesoporous spheres as drug delivery carriers and
a Hollow Mesoporous
20. Zhang B, Luo Z, Liu J, Ding X, Li J, Cai K. Cytochrome c end-capped
mesoporous silica nanoparticles as redox-responsive drug delivery vehicles
for liver tumor-targeted triplex therapy in vitro and in vivo. J Control Release.
2014; 192: 192-201.
21. Tang Y, Hu H, Zhang MG, Song J, Nie L, Wang S, et al. An aptamer-targeting
photoresponsive drug delivery system using "off -- on" graphene oxide
wrapped mesoporous silica nanoparticles. Nanoscale. 2015; 7: 6304-10.
their
22. Huang M, Li F, Dong F, Zhang YX, Zhang LL. MnO2-based nanostructures for
high-performance supercapacitors. J Mater Chem A. 2015; 3: 21380-423.
23. Chen Y, Chen H, Zhang S, Chen F, Sun S, He Q, et al. Structure-property relationships
in manganese oxide - mesoporous silica nanoparticles used for T1-weighted MRI and
simultaneous anti-cancer drug delivery. Biomaterials. 2012; 33:2388-98.
24. Zhang M, Xing L, Ke H, He YJ, Cui PF, Zhu Y, et al. MnO2-based Nanoplatform
Serves as Drug Vehicle and MRI Contrast Agent for Cancer Theranostics. ACS Appl
Mater Interfaces. 2017; 9: 11337−44.
25. Jin L, Liu J, Tang Y, Cao L, Zhang T, Yuan Q, et al. MnO2‐Functionalized Co−P
Nanocomposite: A New Theranostic Agent for pH-Triggered T1/T2 Dual-Modality
Magnetic Resonance Imaging-Guided Chemo-photothermal Synergistic Therapy. ACS
Appl Mater Interfaces. 2017; 9:41648−58.
26. Zhang X-L, Zheng C, Guo S-S, Li J, Yang H-H, Chen G. Turn-on fluorescence
sensor for intracellular imaging of glutathione using g-C3N4 nanosheet -- MnO2
sandwich nanocomposite. Anal Chem. 2014; 86: 3426-34.
27. Liu J, Meng L, Fei Z, Dyson PJ, Jing X, Liu X. MnO2 nanosheets as an artificial
enzyme to mimic oxidase for rapid and sensitive detection of glutathione.
Biosens Bioelectron. 2017; 90: 69-74.
28. Zhao Z, Fan H, Zhou G, Bai H, Liang H, Wang R, et al. Activatable
fluorescence/MRI bimodal platform for tumor cell imaging via MnO2
nanosheets -- aptamer nanoprobe. J Am Chem Soc. 2014; 136: 11220-3.
29. Wu G, Fang Y-Z, Yang S, Lupton JR, Turner ND. Glutathione metabolism and
its implications for health. J Nutr. 2004; 134: 489-92.
30. Meng F, Cheng R, Deng C, Zhong Z. Intracellular drug release nanosystems.
Mater Today. 2012; 15: 436-42.
31. Gerweck LE, Seetharaman K. Cellular pH gradient in tumor versus normal
tissue: potential exploitation for the treatment of cancer. Cancer Res. 1996; 56:
1194-8.
32. Yang X, He D, He X, Wang K, Zou Z, Li X, et al. Glutathione-Mediated
Degradation of Surface-Capped MnO2 for Drug Release from Mesoporous
Silica Nanoparticles to Cancer Cells. Part Part Syst Charact. 2015; 32: 205-12.
33. Li X, Zhao W, Liu X, Chen K, Zhu S, Shi P, et al. MnO2‐Based Mesoporous
manganese silicate coated silica nanoparticles as multi-stimuli-responsive T1-MRI
contrast agents and drug delivery carriers. Acta Biomater. 2016; 30 :378 -- 87.
34. Bertrand N, Wu J, Xu X, Kamaly N, Farokhzad OC. Cancer nanotechnology:
the impact of passive and active targeting in the era of modern cancer biology.
Adv Drug Deliv Rev. 2014; 66: 2-25.
35. Edmondson R, Broglie JJ, Adcock AF, Yang L. Three-dimensional cell culture
systems and their applications in drug discovery and cell-based biosensors.
Assay Drug Dev Technol. 2014; 12: 207-18.
36. Huang H, Zhang P, Chen H, Ji L, Chao H. Comparison between polypyridyl
and cyclometalated ruthenium (II) complexes: anticancer activities against 2D
and 3D cancer models. Chem Eur J. 2015; 21: 715-25.
37. Hélary C, Ovtracht L, Coulomb B, Godeau G, Giraud-Guille MM. Dense
fibrillar collagen matrices: a model to study myofibroblast behaviour during
wound healing. Biomaterials. 2006; 27:4443-52.
38. Giraud-Guille MM, Mosser G, Hélary C, Eglin D. Bone matrix like assemblies
of collagen: from liquid crystals to gel and biomimetic materials. Micron. 2005;
36:602-8.
39. Tidu A, Ghoubay-Benallaoua D, Lynch B, Haye B, Illoul C, Allain J-M, et al.
Development of human corneal epithelium on organized fibrillated transparent
collagen matrices synthesized at high concentration. Acta Biomater. 2015; 22:50-8.
40. Quignard S, Hélary C, Boissière M, Fullana J-M, Lagrée P-Y, Coradin T.
in dermis-like cellularized collagen
Behaviour of silica nanoparticles
hydrogels. Biomater Sci. 2014; 4:484-92.
41. Wang X, Hélary C, Coradin T. Local and Sustained Gene Delivery in
Silica-Collagen Nanocomposites. ACS Appl Mater Interfaces. 2015; 7:2503-11.
42. Bates P, Laber D, Miller D, Thomas S, Trent JO. Discovery and development of
the G-rich oligonucleotide AS1411 as a novel treatment for cancer. Exp Mol
Pathol. 2009; 86: 151-64.
43. Reyes-Reyes EM, Teng Y, Bates PJ. A new paradigm for aptamer therapeutic
AS1411 action: uptake by macropinocytosis and its stimulation by a
nucleolin-dependent mechanism. Cancer Res. 2010; 70:8617-29.
44. Fang X, Zhao X, Fang W, Chen C, Zheng N. Self-templating synthesis of
hollow mesoporous silica and their applications in catalysis and drug
delivery. Nanoscale. 2013; 5: 2205-18.
45. Chen F, Hong H, Shi S, Goel S, Valdovinos H, Hernandez R, et al. Engineering
of hollow mesoporous silica nanoparticles for remarkably enhanced tumor
active targeting efficacy. Sci Rep. 2014; 4: 5080.
46. He X, Yang X, Hai L, He D, He X, Wang K, et al. Single-layer MnO2 nanosheet
quenched fluorescence ruthenium complexes for sensitive detection of ferrous
iron. RSC Adv. 2016; 6: 79204-8.
47. Militello MC, Gaarenstroom SW. Manganese Dioxide (MnO2) by XPS. Surf Sci
Spectra. 2001; 8:200-6.
48. Soulé S, Bulteau A-L, Faucher S, Haye B, Aimé C, Allouche J, et al. Design and
Cellular Fate of Bioinspired Au-Ag Nanoshells@Hybrid Silica Nanoparticles.
Langmuir. 2016; 32:10073-82.
49. Petrovykh DY, Kimura-Suda H, Whitman LJ, Tarlov MF. Quantitative
Analysis and Characterization of DNA Immobilized on Gold. J Am Chem Soc.
2003; 125:5219-26.
50. Quignard S, Masse S, Laurent G, Coradin T. Introduction of disulfide bridges
within silica nanoparticles to control their intra-cellular degradation. Chem
Commun. 2013; 49:3410-2.
51. Huang C-C, Khu N-H, Yeh C-S. The characteristics of sub 10 nm manganese
oxide T1 contrast agents of different nanostructured morphologies.
Biomaterials. 2010; 31: 4073-8.
http://www.ntno.org
Nanotheranostics 2018, Vol. 2
416
52. Peng Y-K, Tsang SCE, Chou P-T. Chemical design of nanoprobes for
T1-weighted magnetic resonance imaging. Mater Today. 2016; 19: 336-48.
53. Lee SH, Kim BH, Na HB, Hyeon T. Paramagnetic inorganic nanoparticles as T1
MRI contrast agents. Wiley Interdiscip Rev Nanomed Nanobiotechnol. 2014; 6:
196-209.
54. He D, He X, Wang K, Yang X, Yang X, Zou Z, Li X. Redox-responsive
degradable honeycomb manganese oxide nanostructures as effective
nanocarriers for intracellular glutathione-triggered drug release. Chem.
Commun. 2015; 51: 776-9.
http://www.ntno.org
|
1608.07356 | 2 | 1608 | 2017-07-15T23:41:14 | Mode selection in compressible active flow networks | [
"physics.bio-ph",
"cond-mat.soft",
"physics.flu-dyn"
] | Coherent, large scale dynamics in many nonequilibrium physical, biological, or information transport networks are driven by small-scale local energy input. Here, we introduce and explore an analytically tractable nonlinear model for compressible active flow networks. In contrast to thermally-driven systems, we find that active friction selects discrete states with a limited number of oscillation modes activated at distinct fixed amplitudes. Using perturbation theory, we systematically predict the stationary states of noisy networks and find good agreement with a Bayesian state estimation based on a hidden Markov model applied to simulated time series data. Our results suggest that the macroscopic response of active network structures, from actomyosin force networks to cytoplasmic flows, can be dominated by a significantly reduced number of modes, in contrast to energy equipartition in thermal equilibrium. The model is also well-suited to study topological sound modes and spectral band gaps in active matter. | physics.bio-ph | physics | Mode selection in compressible active flow networks
Aden Forrow,1 Francis G. Woodhouse,2 and Jorn Dunkel1, ∗
1Department of Mathematics, Massachusetts Institute of Technology,
77 Massachusetts Avenue, Cambridge MA 02139-4307, U.S.A.
2Department of Applied Mathematics and Theoretical Physics, Centre for Mathematical Sciences,
University of Cambridge, Wilberforce Road, Cambridge CB3 0WA, U.K.
(Dated: September 12, 2018)
Coherent, large scale dynamics in many nonequilibrium physical, biological, or information trans-
port networks are driven by small-scale local energy input. Here, we introduce and explore an ana-
lytically tractable nonlinear model for compressible active flow networks. In contrast to thermally-
driven systems, we find that active friction selects discrete states with a limited number of oscillation
modes activated at distinct fixed amplitudes. Using perturbation theory, we systematically predict
the stationary states of noisy networks and find good agreement with a Bayesian state estimation
based on a hidden Markov model applied to simulated time series data. Our results suggest that the
macroscopic response of active network structures, from actomyosin force networks to cytoplasmic
flows, can be dominated by a significantly reduced number of modes, in contrast to energy equipar-
tition in thermal equilibrium. The model is also well-suited to study topological sound modes and
spectral band gaps in active matter.
PACS numbers: 47.63.-b, 05.70.Ln, 05.65.+b 05.40.-a
Active networks constitute an important class of
nonequilibrium systems spanning a wide range of scales,
from the intracellular cytoskeleton [1, 2] and amoeboid
organisms [3 -- 8] to macroscopic transport networks [9 --
12]. Identifying generic self-organization principles [13 --
15] that control the dynamics of these biological or ar-
tificial far-from-equilibrium systems remains one of the
foremost challenges of modern statistical physics. De-
spite promising experimental [4 -- 6, 16 -- 18] and theoret-
ical [1, 7, 19 -- 21] advances over the past decade, it is
not well understood how the interactions between local
energy input, dissipation and network topology deter-
mine the coordinated global behaviors of cells [16], plas-
modia [4 -- 6] or tissues [22]. Further progress requires ana-
lytically tractable models that help clarify the underlying
nonequilibrium mode selection principles [23, 24].
We inroduce here a generic model for active flows on
a network, motivated by recent experimental studies of
bacterial fluids [20, 25] and ATP-driven microtubule sus-
pensions [26] in microfluidic channel systems. Building
on Rayleigh's work [27] on driven vibrations and the
Toner-Tu model of flocking [28], the theory accounts
for network activity through a nonlinear friction [28 --
31]. We work in a fully compressible framework allow-
ing accumulated matter at vertices to affect flow through
network pressure gradients, generalizing previous work
on incompressible pseudo-equilibrium active flow net-
works [32, 33], as suited to the many biological systems
exhibiting flexible network geometry [4 -- 6] or variations
in the density of active components [15]. Although in-
herently nonlinear, the model can be systematically ana-
lyzed through perturbation theory. Such analysis shows
how slow global dynamics emerge naturally from the fast
local dynamics, enabling prediction of the typical states
in large noisy networks; these states have significantly
We
consider
fewer active modes than for energy equipartition [34] in
thermal equilibrium. More broadly, our model provides
an accessible framework for investigating generic physi-
cal phenomena in active systems, including topologically-
protected sound modes [15] and the influence of spectral
band gaps (SM [35]).
activity-driven mass flow on an
arbitrarily-oriented graph G = (V, E) with V = V ver-
tices and E = E edges. The elements of the V × E
gradient (incidence) matrix ∇ are ∇ve = −1 if edge e
is oriented outwards from vertex v, ∇ve = +1 if e is
oriented inwards into v, and ∇ve = 0 otherwise. The dy-
namical state variables are the deviations from the mean
mass ¯ = M/V on the nodes, (1(t), . . . , V (t)), and the
mass fluxes on the edges, (φ1(t), . . . , φE(t)), governed by
the non-dimensionalized (SM [35]) transport equations
(cid:88)
φe = −(cid:88)
v =
e
∇veφe,
∇(cid:62)
evv +
µ − φ2
e
1 + φ2
e
√
φe +
2Dξe(t),
(1b)
v
(1a)
where ξe(t) is standard Gaussian white noise. Equa-
tion (4a) ensures mass conservation. The first term
on the r.h.s. of Eq. (4b) represents the gradient of an
ideal gas-type node pressure pv ∝ v, corresponding to
the leading term in a virial expansion; the second term
is a Toner-Tu type (SM [35]) active friction force de-
rived from a depot model [29, 36] with coupling > 0
and active -- passive control parameter µ, which drives
µ when
µ > 0. Many networks have non-uniform edge and ver-
tex weights, which can be incorporated into equations of
identical form to Eqs. (4) with appropriate rescaling of
, φ, and ∇ (SM [35]).
the edge fluxes φe towards preferred values ±√
7
1
0
2
l
u
J
5
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
6
5
3
7
0
.
8
0
6
1
:
v
i
X
r
a
graphs by choosing appropriate edge weights.
2
v =(cid:80)E
n=1 rn(t)vn and flux φe =(cid:80)E
The complex active flow dynamics encoded by Eqs. (4)
can be understood analytically by considering the basis
of oscillation modes of the network, as we illustrate now
in the fully deterministic case (D = 0). To progress,
we adopt a Rayleigh [27] approximation (µ − φ2
e)φe for
the active friction (SM [35]). Now, expand the pressure
n=1 fn(t)φen in the
right and left singular vectors n = (vn) and φn = (φen)
of ∇(cid:62) corresponding to the E = V − 1 non-zero singular
values λn. (On a tree, there is a single zero eigenvalue
of ∇∇(cid:62) yielding an additional right singular vector for
the pressure, but this corresponds to a constant mass
shift and so can be safely neglected.) Defining mode am-
n + f 2
plitudes A2
n, the network energy then takes
the simple form H = 1
n (SM [35]). When
2
is small there are two distinct timescales, namely the
fast oscillation timescale t and the slow friction timescale
τ = t, which we separate in the perturbation ansatz
σ=0 σfσn [39]. Active
friction does not contribute at lowest order, so the O(1)
contribution to each mode (rn, fn) is an uncoupled har-
monic oscillator r0n(t) = A0n(τ ) cos[λnt − δn(τ )] and
f0n(t) = −A0n(τ ) sin[λnt−δn(τ )] with t-independent am-
plitude A0n and phase δn (SM [35]).
σ=0 σrσn and fn = (cid:80)∞
rn = (cid:80)∞
n = r2
(cid:80)
n λ2
nA2
The influence of activity becomes apparent at first or-
der in , introducing couplings between mode amplitudes
whose dynamics encode the state selection behavior of
the active network. Requiring that the O() amplitudes
r1n and f1n remain small relative to the leading terms
implies that the secular (unbounded) terms in the first
order equations must vanish [39]. Assuming negligible
mode degeneracies, the slow dynamics of the O(1) mode
amplitudes A0n(τ ) are found to obey (SM [35])
(cid:32)
µ − E(cid:88)
k=1
0n)
d(A2
dτ
=
PnkA2
0k
A2
(2)
(cid:33)
2 δnk)(cid:80)
0n,
2 (1 − 1
e φ2
k=1 PnkA2
be zero and solving (cid:80)E
where the overlap matrix Pnk = 3
enφ2
ek
encodes the network topology. Fixed points of Eq. (16)
can then be found by choosing a subset of the A0n to
0n over the
remaining non-zero modes. If all the non-zero solutions
for A2
0n are positive, then there is a stationary point with
those modes activated (SM [35]).
Activity-driven fixed points with exactly one mode ac-
tive always exist. If only mode p is active at leading or-
der, then A0n =(cid:112)µ/Ppp δnp is a fixed point of Eq. (16).
0k = µ for A2
a supercritical Hopf bifurcation with A0n ∼ √
These amplitudes, which closely match both those calcu-
lated with the full unapproximated active friction force
and those from averages computed over fully nonlinear
simulations (SM [35]), show that as µ crosses 0 there is
µ. How-
ever, the stability of such a single-mode state depends on
topology: our simulations suggest that activity always se-
lects exactly one oscillation mode in simple path graphs,
FIG. 1.
Activity can select a single dominant oscillation
mode on hierarchically weighted networks. (a) The edges in
the graph simulated in (b) and (c) are given weights decreas-
ing exponentially with their distance from the central red
path. (b) Oscillations in pressure and flux develop primar-
ily along the central high-weight path (Movie 1). (c) Edge
fluxes φe settle into steady synchronized oscillations as exem-
plified for two edges indicated in (b), one on (φ17) and one off
(φ59) the path. (d) Plotting the time-dependent amplitude
of each analytically-determined flow eigenmode confirms se-
lection of a single oscillatory mode. The ten modes with the
highest average amplitude in this simulation run are pictured;
the marked top two rows are oscillatory modes, while the re-
maining rows are cyclic modes. See Fig. S6 for all modes.
Simulation parameters were = 0.1, µ = 1, and D = 10−4.
Active flow networks described by Eqs. (4) exhibit rich
oscillatory transport behavior, including the mode selec-
tion illustrated in Movie 1 and Fig. 1 for a hierarchically-
weighted network with vertex degrees at most 3 as is typ-
ical of Physarum polycephalum [37]. When this network
is initialized with zero pressure variation and flux, it typ-
ically settles into a quasi-steady state with a single dom-
inant oscillation frequency on the highest-weight path.
This is a manifestation of the fact that single-frequency
selection is the norm on actively driven path graphs, as
we shall show analytically below.
Generally, the features of the steady-state attractor
will be determined by the topology of the subgraph of
high-weight edges, which may be much sparser than the
original network. For this reason, as well as for ease
of analysis and illustration, we will henceforth assume
G to be a tree, as realized in certain peripheral sensory
neurons [38], though in general the full model in Eqs. (4)
is not restricted to any particular class of graph. The
behaviors observed on trees can be extended to denser
3
FIG. 2. First order perturbation theory accurately predicts the stable states on small trees. (a) A five vertex tree possessing
four nontrivial modes, as illustrated. (b) On the tree in (a), mode amplitudes settle into one of two stable stationary states,
as seen in simulations for three different initial conditions. Modes are ordered by frequency from high (top) to low (bottom).
(c) Simulated mode trajectories (rainbow) in (b) match analytic predictions (blue streamlines) in the subspaces of activated
modes. There are three possible arrangements of nonzero critical points in each 2D subspace: a saddle point on one axis and
a stable node on the other axis (left), a stable node on each axis and a saddle point in the middle (center), or a saddle point
on each axis and a stable node in the middle (right; Movie 2). Higher order effects cause both the convergence to a point with
A2 > 0 in the left and middle plots and the oscillations in the trajectories. Parameters used are = 0.5, µ = 1, D = 0.
whereas single-mode states are typically unstable in net-
works with complex topologies. We can use this observa-
tion to model more complex active networks with single
mode selection by appropriately weighting the edges: if
the edge weights for a path are large enough compared to
the weights elsewhere in the network, the path behavior
dominates (Fig. 1).
Insight into stability is provided by the case with up
to two modes active. Writing A0n = A0pδnp + A0qδnq,
Eq. (16) yields
d(A2
0p)/dτ = (µ − PppA2
0p − PpqA2
0q) A2
0p,
(3)
pq
0p, A∗
(cid:1) with A∗
and, potentially, a mixed state (A∗
and symmetrically for A2
0q. Depending on the topology-
this gives up to
encoding overlap coefficients Pnk,
four fixed points:
the zero state A0p = A0q = 0,
which is always linearly unstable; the single-mode state
(A0p, A0q) = ((cid:112)µ/Ppp, 0), which is stable if Ppq > Ppp
and a saddle if not, plus analogously for (0,(cid:112)µ/Pqq);
(cid:113)
µ (Pqq − Ppq)/(cid:0)PppPqq − P 2
0q) where A∗
0p =
0q defined sym-
metrically. When it exists, the mixed state is either sta-
ble (if P 2
pq > PppPqq), but
if one of the single-mode states is stable and one is un-
stable, then one of A∗
0q is imaginary and there
is no mixed state. Hence, we have three possible sce-
narios (Fig. 2): one stable single mode and the other
a saddle with no mixed state (Fig. 2b,c; left); two sta-
ble single-mode states with a mixed saddle in-between
(Fig. 2b,c; center); and two single-mode saddles with a
stable mixed state in-between (Fig. 2b,c; right). These
predictions match simulations quantitatively even for rel-
atively large beyond the small- perturbation regime
pq < PppPqq) or a saddle (if P 2
0p and A∗
(Fig. 2). In fact, simulations show the same qualitative
behavior for = 2, suggesting perturbation analysis re-
mains predictive at high activity.
This two-mode analysis yields a simple topological
heuristic for the stability of single-mode states. Since
φp = 1, Ppp is small when φp is spread over many edges
and large when φp is localized to a few edges. If φq is
localized to the same edges as φp, Ppq will also be large
and mode p will be stable to perturbations in mode q.
However, if φq is localized to a disjoint set of edges, Ppq
will be a scaled inner product of near-orthogonal vectors
(φ2
eq) and will be small. Thus localized modes
will be unstable to modes in other regions, while con-
versely if a mode is to be stable alone then it will be
spread out across the entire network. Therefore, a stable
combination of modes will possess significant flows on all
edges of the network.
ep) and (φ2
Biological systems exhibit vastly different macroscopic
and microscopic time scales [40 -- 43]. This phenomenon
is present in our compressible active flow network, where
higher-order nonlinear effects induce slow global time
scales from faster small-scale dynamics. When the
zeroth-order amplitudes A0n are at a fixed point, the
first-order corrections r1n and f1n are harmonic oscilla-
tors with natural frequency λn driven at linear combina-
tions of the frequencies active at zeroth order (SM [35]).
For instance, if two modes p and q are active at zeroth
order, the driving frequencies are 3λp − k(λp ± λq) for
k = 0, . . . , 3. This introduces new, slower timescales
into the dynamics, including oscillations in the energy
n) with frequency λp − λq. Their
H = 1
2
magnitude depends on the difference in frequency: slower
(cid:80)
n + f 2
n(r2
n λ2
01230123Mode amplitude A1Mode amplitude A301230123Mode amplitude A3Mode amplitude A401230123Mode amplitude A1Mode amplitude A2TimeA1A4A3A210Edge flux1-1Vertex pressure03Mode amplitude0002002002002000TimeTimeTime(a)(b)(c) 4
FIG. 3. States on larger trees possess surprisingly few active modes, which can be inferred from time series with non-zero noise.
(a) The mean number of stationary states of Eq. (16) grows exponentially with edges E as 1.77E ≈ (2E)4/5 (solid orange line),
close to the upper bound of 2E states (dashed black line), while the mean number of stable states grows as 1.21E ≈ (2E)1/4
(solid blue line). We counted states on all nonisomorphic trees with E ≤ 14 edges (filled circles) and on a random sample of
∼ 175 trees per point for 15 ≤ E ≤ 24 (open circles). Averages are over trees with a fixed number of edges. (b) As E increases,
both the mean and the variance of the distribution of trees with each number of stable states increase rapidly. (c) Distribution
of the average number of modes active in a stable state. The mean over trees scales like 0.26E ≈ E/4 (solid line), significantly
below E/2 expected if modes were selected randomly. (d) Two example trees indicated in (a-c) by the corresponding colored
symbols. Stable states on paths (×) always only activate one mode; complex trees (+) have more modes active. (e) Noisy
networks (D > 0) transition stochastically between stable states, exemplified by an amplitude-time trace for the tree shown.
Modes are ordered by frequency from high (top) to low (bottom). Simulation parameters are = 0.5, µ = 1, D = 5 × 10−3.
(f) States found by vbFRET from simulations on the tree in (e) (SM [35]). The second, first, and fifth columns are states seen
in (e), indicated by the colored bars above. (g) States predicted by Eq. (16) for the tree in (e). The first five states in (f) match
those in (g); the sixth column in (f) is likely a transient combination of analytically stable states.
oscillations, driven by modes with similar frequencies
λp ≈ λq, have higher amplitudes (SM [35], Fig. S7).
The number of activated modes in an arbitrary com-
pressible active network depends on intricate interac-
tions between local activity and global flow configura-
tions. The total number of available modes is equal to
the number of edges E, meaning that, were each com-
bination of modes to be a fixed point, a tree could have
up to 2E stationary states. To see how the true num-
ber of stationary and stable states depends on tree size,
we performed an exhaustive numerical fixed point search
of Eq. (16) over a large sample of trees with E ≤ 24
(Fig. 3a-d). The naive upper bound of 2E suggests expo-
nential growth of the mean number of steady states with
edges E; this is indeed what we see, going as ∼ (2E)4/5.
However, though still exponential in E, the mean number
of stable states is much smaller at ∼ (2E)1/4 (Fig. 3a).
Remarkably, these stable states have only ∼ E/4 modes
active on average (Fig. 3c) in stark contrast to the acti-
vation of all E modes under thermal equipartition [34].
Path-like topologies lead to even more dramatic reduc-
tions in the number of modes active (Fig. 3c), suggesting
that a biological system can further reduce the number
of active modes through an optimal choice of topology;
moreover, hierarchically tuned edge capacities as realized
in Physarum [5, 6, 37] can further enhance mode selection
even in non-tree topologies (Fig. 1).
Real active transport networks will have some nonzero
level of thermal or athermal noise [44 -- 46]. Provided the
noise is not too large, it will render previously stable
states now only metastable, with flow patterns exhibiting
small fluctuations around these metastable states punc-
tuated by noise-driven stochastic transitions between
them [32, 46]. Long-time simulations of Eqs. (4) with
D > 0 therefore offer an independent numerical way
to find stable fixed points of the amplitude dynamics.
We use vbFRET [47], a variational Bayesian analysis
of a continuous time hidden Markov model, to iden-
tify states from simulated time series. Almost all of the
states discovered by vbFRET match stable states pre-
dicted by Eq. (16) even in the presence of non-negligible
noise (Fig. 3e-g), justifying the simplifications used in
deriving Eq. (16). This also promises that Bayesian
methods like vbFRET will function as reliable inference
tools for experimental data from real-life active flow net-
works [4, 5, 18].
Beyond active density oscillations [20], the above the-
oretical framework can be used to probe the effects of
topology on the physical properties of complex active sys-
tems. For instance, it was recently shown that continuum
Toner -- Tu systems in finite lattice confinement possess
topologically protected edge-localized sound modes [15].
Similar edge modes can be reproduced in our coarse-
grained model through a simplified network represen-
tation of complex channel geometries (SM [35] and
Movie 3). In addition, generalizing to allow different ef-
fective weights at vertices opens up band gaps, reflected
in the excitation spectrum of spontaneous activity modes
(SM [35]). As we focus on phenomenological properties
shared by many active systems, akin to the Toner -- Tu
approach [28], the results and techniques presented here
promise insights into the mode selection mechanisms gov-
erning a wide range of non-equilibrium transport and
force networks.
This work was supported by NSF Award CBET-
1510768 (A.F. and J.D.), Trinity College, Cam-
bridge (F.G.W.), and an Alfred P. Sloan Research Fel-
lowship (J.D.). The authors thank Martin Zwierlein for
stimulating discussions on band gaps.
∗ [email protected]
[1] C. Broedersz and F. MacKintosh, Rev. Mod. Phys. 86,
995 (2014).
[2] V. Ruprecht, S. Wieser, A. Callan-Jones, M. Smutny,
H. Morita, K. Sako, V. Barone, M. Ritsch-Marte, M. Sixt,
R. Voituriez, and C.-P. Heisenberg, Cell 160, 673 (2015).
[3] A. Takamatsu, R. Tanaka, H. Yamada, T. Nakagaki,
T. Fujii, and I. Endo, Phys. Rev. Lett. 87, 7 (2001).
[4] A. Tero, S. Takagi, T. Saigusa, K. Ito, D. P. Bebber,
M. D. Fricker, K. Yumiki, R. Kobayashi, and T. Naka-
gaki, Science 327, 439 (2010).
[5] K. Alim, G. Amselem, M. P. Brenner, and A. Pringle,
Proc. Natl. Acad. Sci. U.S.A. 110, 13306 (2013).
[6] K. Alim, N. Andrew, A. Pringle, and M. P. Brenner,
Proc. Natl. Acad. Sci. USA 114, 5136 (2017).
[7] V. Bonifaci, K. Mehlhorn, and G. Varma, J. Theor. Biol.
309, 121 (2012).
[8] C. R. Reid, H. Macdonald, R. P. Mann, J. A. R. Marshall,
T. Latty, and S. Garnier, J. R. Soc. Interface 13, 44
(2016).
[9] G. Coclite, M. Garavello, and B. Piccoli, SIAM J. Math.
Anal. 36, 1862 (2005).
[10] B. Piccoli and M. Garavello, Traffic Flow on Networks:
Conservation Laws Models (AIMS, Springfield, MO,
2006).
[11] S. Hata, H. Nakao, and A. S. Mikhailov, Phys. Rev. E
5
arXiv:1610.06873.
[16] N. Fakhri, A. D. Wessel, C. Willms, M. Pasquali, D. R.
and C. F. Schmidt,
Klopfenstein, F. C. MacKintosh,
Science 344, 1031 (2014).
[17] P. Ronceray, C. P. Broedersz, and M. Lenz, Proc. Natl.
Acad. Sci. U.S.A. 113, 2827 (2016).
[18] S. Marbach, K. Alim, N. Andrew, A. Pringle, and M. P.
Brenner, Phys. Rev. Lett. 117 (2016).
[19] C. P. Broedersz, X. Mao, T. C. Lubensky, and F. C.
MacKintosh, Nat. Phys. 7, 983 (2011).
[20] M. Paoluzzi, R. Di Leonardo, and L. Angelani, Phys.
Rev. Lett. 115, 188303 (2015).
[21] A. Bressan, S. Cani´c, M. Garavello, M. Herty,
and
B. Piccoli, EMS Surv. Math. Sci. 1, 47 (2014).
[22] C. G. Vasquez and A. C. Martin, Dev. Dynam. 245, 361
(2016).
[23] W. Ebeling, U. Erdmann, J. Dunkel, and M. Jenssen, J.
Stat. Phys. 101, 443 (2000).
[24] J. Dunkel, W. Ebeling, U. Erdmann, and V. A. Makarov,
Int. J. Bifurcat. Chaos 12, 2359 (2002).
[25] H. Wioland, E. Lushi, and R. E. Goldstein, New J. Phys.
18, 075002 (2016).
[26] K.-T. Wu, J. B. Hishamunda, D. T. N. Chen, S. J. De-
Camp, Y.-W. Chang, A. Fern´andez-Nieves, S. Fraden,
and Z. Dogic, Science 355, eaal1979 (2017).
[27] J. W. S. B. Rayleigh, The Theory of Sound vol. 1, 2nd
ed. (Macmillan, New York, 1894) p. 81.
[28] J. Toner, Y. Tu, and S. Ramaswamy, Annals of Physics
318, 170 (2005).
[29] F. Schweitzer, W. Ebeling, and B. Tilch, Phys. Rev.
Lett. 80, 5044 (1998).
[30] P. Romanczuk, M. Bar, W. Ebeling, B. Lindner, and
L. Schimansky-Geier, Eur. Phys. J. Spec. Top. 202, 1
(2012).
[31] P. S. Burada and B. Lindner, Phys. Rev. E 85, 032102
(2013).
[32] F. G. Woodhouse, A. Forrow, J. B. Fawcett,
and
J. Dunkel, Proc. Natl. Acad. Sci. U.S.A. 113, 8200
(2016).
[33] F. G. Woodhouse and J. Dunkel, Nat. Commun. 8, 15169
(2017).
[34] A. I. Khinchin, Mathematical Foundations of Statistical
Mechanics (Dover, New York, 1949).
[35] See Supplemental Material, which includes Refs. [48].
[36] P. Romanczuk, W. Ebeling, U. Erdmann,
L. Schimansky-Geier, Chaos 21, 047517 (2011).
and
[37] W. Baumgarten, T. Ueda, and M. J. B. Hauser, Phys.
Rev. E 82, 046113 (2010).
[38] J. Kromer, A. Khaledi-Nasab, L. Schimansky-Geier, and
A. B. Neiman, ArXiv:1701.01693.
[39] S. H. Strogatz, Nonlinear Dynamics and Chaos (West-
view Press, Boulder, CO, 2015).
[40] J. Halatek and E. Frey, Cell Rep. 1, 741 (2012).
[41] R. A. Kerr, H. Levine, T. J. Sejnowski, and W. Rappel,
Proc. Natl. Acad. Sci. U.S.A. 103, 347 (2006).
[42] I. H. Riedel-Kruse, C. Muller, and A. C. Oates, Science
317, 1911 (2007).
89, 020801 (2014).
[43] A. Varma, K. C. Huang, and K. D. Young, J. Bacteriol.
[12] L. L. Heaton, E. L´opez, P. K. Maini, M. D. Fricker, and
190, 2106 (2008).
N. S. Jones, Phys. Rev. E 86, 021905 (2012).
[13] H. Nakao and A. S. Mikhailov, Nat. Phys. 6, 544 (2010).
[14] A. S. Mikhailov and K. Showalter, Chaos 18, 026101
[44] B. Lindner, J. Garcıa-Ojalvo, A. Neiman,
and
L. Schimansky-Geier, Phys. Rep. 392, 321 (2004).
[45] L. Gammaitoni, P. Hanggi, P. Jung, and F. Marchesoni,
(2008).
Rev. Mod. Phys. 70, 223 (1998).
[15] A. Souslov, B. C. van Zuiden, D. Bartolo, and V. Vitelli,
[46] P. Hanggi, P. Talkner,
and M. Borkovec, Rev. Mod.
Phys. 62, 251 (1990).
[47] J. E. Bronson, J. Fei, J. M. Hofman, R. L. Gonzalez, and
C. H. Wiggins, Biophys. J. 97, 3196 (2009).
[48] J. W. S. B. Rayleigh, Proc. Lond. Math. Soc. s1-10, 4
(1878); P. Misra, Physics of Condensed Matter (Elsevier
Science, 2011).
6
Supplemental material: Mode selection in compressible active flow networks
7
Aden Forrow, Francis G. Woodhouse, and Jorn Dunkel
Nondimensionalization of governing equations
We can define the model in terms of the dimensional quantities v, φe, and t; global dimensional parameters ,
β, and D; dimensionless edge conductances γe and vertex volumes mv; and a dimensionless global parameter µ and
function g as
(cid:88)
=
e
= −γe
dv
dt
d φe
dt
φe,
∇ve
(cid:88)
v
∇(cid:62)
evm−1
v v + g
(cid:18)
µ,
(cid:19)
φe
βγe
(cid:112)
φe +
2 D ξe(t).
The scaling by conductance in the argument of g is chosen to match the phenomenology observed in dense bacterial
suspensions, where activity selects a characteristic velocity φe/γe and not a fixed flux φe. If we choose a conductance
scale γ and volume scale m and insert the rescaled, nondimensional parameters
γe = γ−1γe,
mv = m−1 mv,
= γ− 1
2 ,
De = β−2γ− 1
2 γ−1
e
D
and variables
we are left with
v = m
1
2
v γ
dv
dt
dφe
dt
1
2 β−1 v,
(cid:88)
= −(cid:88)
=
e
v
− 1
φe = γ
e
2
β−1 φe,
t = γ
1
2 t,
ξe(t) = γ− 1
4 ξe(t),
v ∇veγ1/2
m−1/2
e φe,
(cid:18)
(cid:19)
φe +
(cid:112)
φe√
γe
2Deξe(t).
evm−1/2
v
v + g
µ,
With constant conductances γe = 1 and volumes mv = 1, we recover the model introduced in the main text, namely
with nonzero entries of the gradient matrix equal to ±1. All of our analysis applies equally well to the varying weights
case: the only substantive change is replacing ∇ve with the weighted gradient ∇∗
−1/2
v ∇veγ1/2
.
e
ve = m
We can combine Eqs. (4a) and (4b) into one second order equation for the pressure dynamics reading
In the absence of friction, when g(µ, φe) = 0, the dynamics are Hamiltonian with energy
The energy is particularly simple when written in the basis of singular vectors of ∇(cid:62) with non-zero singular values,
giving
(cid:88)
(cid:0)r2
(cid:1) ≡(cid:88)
H =
1
2
λ2
n
n + f 2
n
Hn.
n
n
e ∇(cid:62)
γ1/2
(cid:88)
= −(cid:88)
=
e
v
∇veφe,
(cid:32)
−(cid:88)
(cid:88)
u
dv
dt
dφe
dt
(cid:88)
e
∇(cid:62)
evv + g(µ, φe)φe +
√
2Dξe(t),
(cid:33)
v =
∇ve
∇(cid:62)
euu + g(µ, φe)φe +
√
2Dξe(t)
.
H =
1
2
v,e,u
v∇ve∇(cid:62)
euρu +
φe∇(cid:62)
ev∇vf φf .
(cid:88)
e,v,f
1
2
(4a)
(4b)
(5)
(6)
Relation to physical flow systems
8
We chose to explore a minimal model coupling local active energy input to network structure, rather than capture
the details of any particular model system. Nevertheless, the key features of our model, namely mass conservation
and a polynomial expansion of the active term, are generic enough to be straightforwardly adapted to a range of
applications.
Mass conservation and pressure driven flow are likely to remain in any active flow model; the form of the active
term may change in different contexts.
In our case, staying close to examples of bacterial suspensions, we model
activity as driving spontaneous flow on all edges. An alternative option, more closely related to shuttle streaming in
networks, would be to apply an active force fv that compresses or expands each vertex and drives flow in or out, with
modified dynamics
(cid:88)
= −(cid:88)
=
e
v
dv
dt
dφe
dt
∇veφe,
∇(cid:62)
ev(v + fv) +
√
2Dξe(t).
The correct form of the active force depends on the microscopic details of the driving. Some generic features, however,
will not depend on the exact form of fv and will be discoverable by choosing a simple function of local quantities
(v, v, etc.) as an approximate driving force.
The same method is used to derive the Toner-Tu equations for continuous active flows [28]; our model can be
understood as a discrete version of a special case of these equations. If advective and diffusive terms are rendered
negligible in favor of pressure-driven and activity-driven flow by geometric effects or otherwise, and we take only the
linear term in the virial expansion of the active pressure, the general Toner -- Tu model simplifies to
In a limit where deviations from the mean density are small, so ρ = ρ0 + η for some η (cid:28) 1, we can further reduce to
= α(cid:126)v − β(cid:126)v2(cid:126)v − σ1 (cid:126)∇(ρ − ρ0) + (cid:126)f ,
∂(cid:126)v
∂t
+ (cid:126)∇ · ((cid:126)vρ) = 0.
∂ρ
∂t
= α(cid:126)v − β(cid:126)v2(cid:126)v − ησ1 (cid:126)∇ + (cid:126)f ,
∂(cid:126)v
∂t
η
∂
∂t
+ (ρ0 + η)(cid:126)∇ · (cid:126)v + η(cid:126)v · (cid:126)∇ = 0.
Then on short time scales τ = t/η, we have
= ηα(cid:126)v − ηβ(cid:126)v2(cid:126)v − η2σ1 (cid:126)∇ + η (cid:126)f ,
∂(cid:126)v
∂τ
≈ −ρ0 (cid:126)∇ · (cid:126)v,
∂
∂τ
where we neglect terms that must be of order η: if the coefficients α, β, and σ1 are sufficiently large, their terms will
remain relevant. The scaling of τ ensures that t is small when τ is order one or smaller. Discretizing the velocity and
ev or −∇ve as appropriate
density fields as well as the noise (cid:126)f and replacing the continuous gradient with either ∇(cid:62)
yields Eqs. (4).
Compressibility
Compressibility as included in our model is intended to describe changes in density or volume of the active com-
ponent, not the underlying fluid. For example, variations in may be interpreted as variations in the density of
swimmers in a bacterial system or variations in the tube volume in Physarum polycephalum. Such systems may be
effectively compressible even though the solvent fluid (e.g. water) is incompressible.
In some cases, compressibility is the primary object of interest. For example, a recent preprint [15] discusses
sound in active fluids in a network using a continuous wave equation derived from the Toner-Tu model. On top of
a background flow taking the form of a lattice of counter-rotating cycles, they find modes confined to the edges of a
Lieb lattice, which we can reproduce in our discretized setting (Fig. S1 and Movie 3). In both their setting and ours,
these edge modes decay over time without propagating into the bulk (cf. discussion in App. I.B of Ref. [15]).
We can recover an incompressible limit of our model by first extending it to include damping on the vertices:
(cid:88)
=
e
= −γ
(cid:88)
v
dv
dt
dφe
dt
∇veφe − ηv,
∇(cid:62)
evv + g(µ, φe)φe +
√
2Dξe(t).
(11a)
(11b)
9
Fig. S1. Our active network model exhibits behavior similar to the topological edge modes of [15]. (a) A discretized version of
the Lieb lattice considered in [15]. Edges shared by adjacent 8-cycles have weight γe = 2 to account for the additional width of
√
the corresponding channels. The most stable flow on this network consists of a lattice of counter-rotating cycles, in which both
γe) and the pressure variations v are everywhere zero. (b) This lattice has modes confined
the active friction term g(µ, φe/
to the edges of the domain, allowing sound waves to propagate and decay without scattering into the bulk (cf. discussion in
App. I.B of Ref. [15]); one such mode is pictured. Simulations started in this mode as a perturbation to the most stable flow
pattern do not cause density changes in the center (Movie 3). The network model allows study of such phenomena without
resorting to full scale simulation of the flow patterns.
This paper examines the limit η → 0 where total mass is exactly conserved. Previous work [32] has looked at the
opposite limit, η → ∞, where Eq. (11a) can only be balanced if v → 0 and
(cid:88)
e
v =
1
η
∇veφe.
Substituting this into Eq. (11b) gives
dφe
dt
= − γ
η
(cid:88)
v
∇(cid:62)
ev∇vaφa + g(µ, φe)φe +
√
2Dξe(t).
e), this is equivalent to the model discussed in [32]. If γ → ∞ so that γ/η is constant, small
With g(µ, φe) = φ2
deviations from incompressibility are allowed; if γ/η → ∞, incompressibility is fully enforced. However, compressibility
is a necessary ingredient for sound waves [15] and density oscillations [20].
e(1 − φ2
While choosing the friction function to be [29]
Rayleigh friction approximation
g(µ, φe) =
µ − φ2
e
1 + φ2
e
has convenient theoretical properties, namely that it gives a passive constant friction coefficient for µ = −1 and
for φ → ∞, it is analytically difficult. To simplify the analysis, we approximate this g(µ, φe) with a symmetric
quadratic [27]
g(µ, φe) = a − bφ2
e,
(12)
where a = µ and b = 1 are chosen so that g(µ, 0) = g(µ, 0) and g(µ, φe) has the same zeros as g(µ, φe). This ensures
that the two functions approximately match when they are both negative, that is, when activity is putting energy
into the flow. The large difference between g(µ, φe) and g(µ, φe) when the flux is large is less important, as the flow
will be damped down in either case. The larger damping in g(µ, φe) does result in slightly lower steady amplitudes,
both analytically and in simulations.
Perturbation expansion
10
If is small, there will be two widely separated timescales: the fast oscillation timescale t and the slow friction
timescale τ = t. After writing v and φe in the mode basis, we can further expand in as
∞(cid:88)
∞(cid:88)
k=0
rn(t) =
fn(t) =
krkn(t, τ ),
kfkn(t, τ ),
(13a)
(13b)
where we explicitly separate the dependence on the two timescales. Then
k=0
rkn(t, τ ) = ∂2
fkn(t, τ ) = ∂2
t rkn + 2∂t∂τ rkn + 2∂2
t fkn + 2∂t∂τ fkn + 2∂2
τ rkn,
τ fkn.
At zeroth order in , with D = 0, Eq. (5) becomes
V(cid:88)
t r0nvn = − V(cid:88)
∂2
λ2
nr0nvn.
The modes vn are orthonormal, so the terms decouple into separate harmonic oscillators; fkn can be found from rkn
using Eq. (4a). The leading order solution is then
n=1
n=1
r0n(t) = A0n(τ ) cos(λnt − δn(τ )),
f0n(t) = −A0n(τ ) sin(λnt − δn(τ )).
At first order in , with g(µ, φe) = (µ − φ2
e),
V(cid:88)
t r1n + 2∂t∂τ r0n)vn = − V(cid:88)
(∂2
n=1
n=1
Multiplying by vm and summing over v, we find
t r1m + 2∂t∂τ r0m = −λ2
∂2
mr1m + λm
λ2
nr1nvn +
f0nφen
∇ve
µ −
(cid:88)
µf0m −(cid:88)
e
(cid:32) E(cid:88)
(cid:32) E(cid:88)
n=1
φem
f0nφen
(cid:33)2 E(cid:88)
(cid:33)3 .
l=1
f0lφel.
(14)
e
n=1
Leading order amplitude dynamics
In order for the expansion in Eqs. (13a) and (13b) to make sense, the magnitudes of the summands rkn and fkn
must remain bounded. From Eq. (14), r1m is a harmonic oscillator with natural frequency λm driven by the zeroth
order oscillations. It will have bounded oscillations only if the resonant terms in Eq. (14), those that drive r1m at
its natural frequency, are zero. Finding the resonant terms and setting them to zero will fix the leading order mode
amplitudes An(τ ).
Expanding the cube in Eq. (14) gives
t r1m + 2∂t∂τ r0m = − λ2
∂2
mr1m + λm
= − λ2
mr1m + λm
(cid:20)
µf0m −(cid:88)
(cid:20)
E(cid:88)
e
µf0m +
φem
E(cid:88)
(cid:32)(cid:88)
k,(cid:96),n=1
f0kφekf0nφe(cid:96)f0nφen
(cid:33)
φemφekφe(cid:96)φen
(cid:21)
k,(cid:96),n=1
e
× A0kA0(cid:96)A0n sin(λkt − δk) sin(λ(cid:96)t − δ(cid:96)) sin(λnt − δn)
(cid:21)
.
(15)
Now, the product of sines can be expanded into
sin(λkt − δk) sin(λ(cid:96)t − δ(cid:96)) sin(λnt − δn) =
1
4
(cid:104)
sin(δk − δ(cid:96) − δn − λkt + λnt + λ(cid:96)t)
− sin(δk − δ(cid:96) + δn − λkt − λnt + λ(cid:96)t)
− sin(δk + δ(cid:96) − δn − λkt + λnt − λ(cid:96)t)
+ sin(δk + δ(cid:96) + δn − λkt − λnt − λ(cid:96)t)
(cid:105)
11
.
We seek only resonant terms, which only occur when ±λk, ±λ(cid:96), and ±λn sum to λm. This happens most often
in one of two ways. First, we might have k = (cid:96) and n = m or similar. Alternatively, we might have degenerate
modes, λk = λ(cid:96) and λn = λm. However, we ignore the latter possibility because degeneracies add significant analytic
complications, including nontrivial dynamics of their relative phases. We also ignore the rare possibility of resonant
terms arising from interactions of modes with three or four distinct singular values. The results we get with these
assumptions closely match simulated time series (Fig. 3e-g), suggesting that the existence of degeneracies has little
impact on the dynamics of nondegenerate modes.
The remaining resonant terms in Eq. (15) must cancel so that r1m is not an oscillator of frequency λm driven at
frequency λm. Thus,
µf0m +
(cid:32)(cid:88)
e
1
4
(cid:33)
E(cid:88)
(cid:32)(cid:88)
k=1,k(cid:54)=m
e
(cid:33)
2∂t∂τ r0m = λm
φ4
em
0m(3 sin(λmt − δm)) + 3
A3
φ2
emφ2
ek
A2
0kA0m
(2 sin(λmt − δm))
1
4
.
Substituting in r0m and f0m,
(cid:34)
−2A(cid:48)
(cid:33)
0mλm sin(λmt − δm) + 2λ2
(cid:32)(cid:88)
1
4
φ4
em
e
m cos(λmt − δm)δ(cid:48)
m =
0m(3 sin(λmt − δm))+ 3
A3
E(cid:88)
k=1,k(cid:54)=m
(cid:32)(cid:88)
e
(cid:33)
φ2
emφ2
ek
A2
0kA0m
1
4
,
(2 sin(λmt − δm))
−µA0m sin(λmt − δm) +
λm
where primes denote differentiation with respect to τ . For this to hold for all t we need the coefficients of the sine
and cosine terms to separately cancel. From the cosine term, δ(cid:48)
m = 0; from the sine term,
A(cid:48)
0m =
1
2
A0m
µ − 3
4
(cid:32)(cid:88)
e
(cid:33)
φ4
em
0m − 3
A2
2
k=1,k(cid:54)=m
e
E(cid:88)
2 δmk)(cid:80)
2 (1 − 1
(cid:32)(cid:88)
(cid:33)
φ2
emφ2
ek
A2
0k
A0m
PmkA2
0k
,
(cid:32)
µ − E(cid:88)
k=1
(cid:33)
where the matrix P has entries Pmk = 3
e φ2
emφ2
ek. Rewriting in terms of the squared amplitudes,
≡ 1
(cid:33)
2
(A2
0m) = 2A0mA(cid:48)
0m = A2
0m
d
dτ
(cid:32)
µ − E(cid:88)
k=1
PmkA2
0k
.
(16)
(17)
As a matrix equation, with xm = A2
0m, this reads
x(cid:48) = x (cid:12) (µ1 − Px),
where 1 denotes the vector of ones and (cid:12) is the component-wise product.
To find stationary points, we set x (cid:12) (µ1 − Px) = 0. The obvious way to solve Eq. (17) for all stationary points is
to exhaustively search over combinations of active modes: on picking certain elements of x to be zero, the remaining
nonzero entries x are found by solving Px = µ1, where P is P restricted to those modes chosen to be nonzero.
Stability of a fixed point x0 then follows by standard perturbation analysis: inserting a small perturbation x0 + δx(τ )
into Eq. (17) gives
δx(cid:48) = δx − x0 (cid:12) (Pδx) − (Px0) (cid:12) δx + O(δx2) ≡ Mδx + O(δx2),
where I denotes the identity matrix, and the eigenvalues of M then determine stability in the usual fashion.
12
Fig. S2. Steady state amplitudes Ai as a function of activity µ for the tree pictured undergo a Hopf bifurcation as µ crosses 0.
Dots are long-time root-mean-square amplitudes from simulations started in each mode; lines are numerical solutions of Eq. (18).
Mode A2 is too unstable to reliably observe in simulations, so it is omitted. For µ < 0, all amplitudes go to zero in simulations;
the dot included in that region is at µ = −1 where the friction is purely passive. Some deviations between simulation and
analytics are expected because the simulations do not use the Rayleigh friction approximation and (cid:54)= 0. Parameters were
= 0.5 and D = 0.
Accuracy of Rayleigh friction approximation
To verify that the Rayleigh friction approximation does not significantly impact the results, we check the amplitude
e) on all edges. Here setting the first
and stability of single modes for the full model with g(µ, φe) = (µ − φ2
order secular terms to zero in a perturbation expansion with A0n = A0pδnp leads to
e)/(1 + φ2
A2
0p = (µ + 1)
2 − 2
1
1 + A2
0pφ2
ep
Numerically solving Eq. (18) for µ = 1 yields solutions within a few percent of the Rayleigh approximation solution
1/(cid:112)Ppp which additionally match numerical simulations of the full model even for as large as 0.5 (Fig. S2).
If we assume µ (cid:28) 1 (so Ap (cid:28) 1) and expand the square root to order A4
matching the Rayleigh friction result. The scaling Ap ∼ √
When the system transitions from no energy input to active flow, the steady state amplitudes will grow with µ.
p) = µ/Ppp, exactly
µ is typical of a supercritical Hopf bifurcation.
e
p, we find A2
p + O(A4
(cid:32)
(cid:88)
(cid:115)
(cid:33)
.
(18)
Attractor characteristics on tree networks
The mode interactions of Eq. (17) can lead to complex oscillation patterns dependent on global, not local, topology,
as shown for a 127-vertex complete binary tree in Movie 4 and Fig. S3. After initializing with zero pressure variation
and flux, the system settles into quasi-steady states with dramatically different dynamics in separate regions of the
tree (Fig. 3a,b). Flux in edges near the leaves of the tree tends to oscillate rapidly, driving large pressure fluctuations
in nearby vertices, whereas flux oscillations near the root are comparatively slow with nearly constant pressure in the
vertices (Fig. 3b,d). Since, apart from the root and leaves, each vertex has the same local topology, the different time
scales emerge from the interaction of the local active friction with the global structure of the tree.
A comprehensive and precise characterization of the relative lifetimes of different attractors in large active flow
networks remains out of reach with current numerical methods, in part because the range of noise levels low enough
to observe state selection and high enough to observe transitions is quite small. Such a fine-tuning between thermal
and active transport processes is a characteristic feature of many, if not all, biological systems that function optimally
in a narrow temperature range: bacterial flagellar motors are designed to barely beat Brownian diffusion at room
temperature, ATP-driven intracellular transport is tuned such that it improves moderately over thermal diffusion,
and so-on. Another well-known example in this context is stochastic resonance in driven multistable systems [45].
However, as all these systems typically exhibit exponential Arrhenius-type waiting times, it is practically impossible
to completely explore their attractor statistics in the moderate-to-weak noise regime, except for the simplest two-state
systems [46].
Nevertheless, long simulation runs as shown in Fig. S4 offer some insight into the qualitative behavior of attractors
in active flow networks. Specifically, our simulations suggest that, while there is considerable variation in the relative
occupancy of different attractors, stable states can be approximately divided in two classes: (1) states with one high
energy mode at high amplitude and a few low energy modes at low amplitude and (2) states with multiple low-energy
modes active at moderate amplitude, some of them degenerate. States of type (2) tend to quickly transition to other
states of type (2) (Fig. S4); states of type (1) have a wide range of lifetimes but no obvious transition patterns.
13
Fig. S3. Activity causes depth-dependent separation of time scales on a large tree. (a) Most pressure variation occurs near
the leaves on large binary trees (Movie 4). (b) The tree in (a) develops an activity-driven steady state with slow oscillations in
the center and fast oscillations near the edges, as illustrated by the flux φe on the three edges labelled in (a). (c) Unnormalized
correlations between the Fourier transforms of the flux through the edges of the tree in (a), with phases ignored. Colors indicate
the tree level of the tail vertex of the edge. There are strong correlations within each level and between neighboring levels,
but low correlations for edges in widely-separated levels. (d) Frequency spectra of each tree level, computed by taking Fourier
transforms of the edge fluxes as in (c) and averaging the magnitudes across all edges at each level. A distinct primary oscillation
frequency for each level can be seen, which increases with distance from the tree center. Simulation parameters in all panels
are = 0.5, µ = 1, and D = 10−3. (e-h) While adding edges in the center leads to steady flow on cycles there, frequency still
increases with distance from the center in the outer, tree-like sections.
14
Fig. S4. Lower energy modes transition more often for the graph in Fig. 3e of the Main Text. Modes are ordered by frequency
from high (top) to low (bottom). Simulation parameters are = 0.5, µ = 1, D = 5 × 10−3, identical to those in Fig. 3. Note
that rows 7 and 8, the two modes that switch on and off most, are degenerate.
Networks with cycles
We focus on tree networks in this paper as they allow substantial analytical progress. However, Eqs. (4) can be
applied without modification to networks with cycles. Cycles correspond to right singular vectors φn of ∇(cid:62) with
singular value zero. As these are always degenerate, we expect the conclusions of Section to be most accurate when
there are few or no cycles. Alternatively, on a weighted graph where the edges of high conductance form a tree, the
attractor characteristics will be similar to the attractors on that tree (Fig. 1; all modes pictured in Fig. S6).
Qualitatively, we find the same stochastic switching between states with subsets of modes active in simulations of
Eqs. (4) on cyclic graphs even with equal weights, with the additional feature that cyclic modes are particularly stable
and take longer to transition on average (Fig. S5). For further discussion of similar dynamics on cycles, see [32].
Fig. S5. States on graphs with cycles, like the one shown, tend to be more stable. Modes are ordered by frequency from high
(top) to low (bottom). Note that the eight modes at the bottom, which are the only ones active in the lower half of the trace,
are all cycles. Simulation parameters are = 0.5, µ = 1, D = 5 × 10−3.
15
Fig. S6.
Including all of the modes from the simulation in Fig. 1 of the Main Text shows clear single mode selection on this
weighted network. Edges a distance d from the central red path were given weight e−d. Modes are ordered by frequency from
high (top) to low (bottom); the last thirty modes, marked in red, are cycles. The modes pictured in Fig. 1 are marked in black.
Higher order oscillations
Before, by setting resonant terms to zero, we found the slow dynamics of An. Now we look at the non-resonant
terms driving r1m to find higher order effects. If we let
Sin1
nk
1 ...i
k
=
(cid:88)
k(cid:89)
φnj
eij
,
e
j=1
16
Fig. S7. Slow global oscillations emerge from the fast active dynamics. (a) First order considerations fix a constant mean flow
energy; higher order effects cause significant slow oscillations about that mean. Simulation parameters were µ = 1, = 0.5,
and D = 0; the tree used is inset. (b) The mode amplitudes A2 and A3, like the energy, oscillate much more slowly than the
harmonic oscillations of f2 and f3. All other mode amplitudes (unlabelled traces) are close to zero. (c) Frequency spectra of
the two active modes and the energy H for the simulation in (a) and (b). The energy oscillates due to higher-order interactions
between modes at frequencies that are linear combinations of active mode frequencies, not the harmonic frequencies alone
(dashed lines).
assume the resonant terms are zero, and assume A0m = A0pδmp + A0qδmq, the remainder of Eq. (14) is
∂2
t r1m + λ2
mr1m =
1
4
λm
Smp3A3
0p sin(3λpt) + 3Smq2pA0pA2
0q
(cid:2) sin((2λq − λp)t) − sin((2λq + λp)t)(cid:3)
(cid:2) sin((2λp − λq)t) − sin((2λp + λq)t)(cid:3)
(cid:111)
(cid:110)
0pA0q
0q sin(3λqt)
.
+ 3Smqp2A2
+ Smq3A3
(cid:8)3Sq2p2 A0pA2
Setting m = p and only looking at the terms closest to resonance, we obtain
0q sin((2λq − λp)t) + 3Sp3qA2
∂2
t r1p + λ2
λp
pr1p ≈ 1
4
Thus
where
r1p ≈ c1 cos((2λq − λp)t − δ1) + c2 cos((2λp − λq)t − δ2),
f1p ≈ −c1 sin((2λq − λp)t − δ1) − c2 sin((2λp − λq)t − δ2),
(19)
0pA0q sin((2λp − λq)t)(cid:9).
3
c1 =
c2 =
4((2λq − λp)2 − λ2
p)
4((2λp − λq)2 − λ2
p)
3
λpSq2p2A0pA2
0q,
λpSqp3A2
0pA0q.
The energy in this mode to first order in is
(cid:0)(r2
(cid:110)
A2
λ2
p
2
λ2
p
2
Hp =
=
0p + f1p)2(cid:1) + O(2)
(cid:2)c1 cos((2λq − 2λp)t) + c2 cos((λp − λq)t)(cid:3)(cid:111)
0p + r1p)2 + (f 2
0p + 2A0p
+ O(2),
exhibiting an order time dependence. The coefficients c1 and c2 are small unless λp ≈ λq. If we kept the frequency
3λp, 3λq, 2λp + λq, and 2λq + λp terms from Eq. (19), we would find energy oscillations with frequencies 2λp, 2λq,
3λq − λp, and λp + λq (Fig. S7); those oscillations have smaller amplitudes as the driving is farther from resonance.
Noise and thermalization
In Eqs. (4a) and (4b) we add Gaussian white noise only to the flux as a physically intuitive source of random
fluctuations that preserve mass conservation. However, even with purely passive friction, this does not lead to
equipartition of energy as seen in thermal systems.
Written as stochastic differential equations with g(µ, φ) = −1, Eqs. (4a) and (4b) become
(cid:88)
dφe = −(cid:88)
dv =
e
∇veφedt,
∇(cid:62)
evvdt − φedt +
√
2Dd Be(t),
17
(20a)
(20b)
where each Be(t) is standard Brownian motion. The components of the E-dimensional Brownian motion B(t) =
( B1, . . . , BE(t)) in any orthonormal basis are also standard Brownian motions, so we can rewrite the system in the
mode basis as
v
drn = λnfndt,
dfn = −λnrndt − fndt +
√
2DdBn(t).
The associated Fokker-Planck equation for the probability distribution p(r, f , t) is
∂tp =
− ∂
∂rn
(λnfnp) +
∂
∂fn
(λnrnp) +
∂
∂fn
(fnp) + D
(cid:20)
(cid:88)
n
(21a)
(21b)
(cid:21)
∂2p
∂f 2
n
(fnp) + D
0 =
∂
∂fn
(cid:88)
p(H1, . . . , HM ) ∝ M(cid:89)
n
∂2p
∂f 2
n
,
− Hn
kTn ,
e
n=1
(cid:112)
2DndBn(t) ≡
dBn(t).
√
2D
λn
with p → 0 as rn, fn → ∞ and p integrating to 1. Now, without friction or noise, the dynamics are governed by the
Hamiltonian
H =
1
2
v∇ve∇euu +
φe∇ev∇vaφa =
1
2
1
2
λ2
n(r2
n + f 2
Hn.
n
n
(cid:88)
n) ≡(cid:88)
If p is a function of the Hn alone, the Fokker-Planck equation in steady state reduces to
which has solution
nD/.
where kTn = λ2
equipartition of amplitude, not energy: the long-time average (cid:104)A2
between modes by making µ > −1 does not change this.
Loosely, adding noise this way couples each mode to a heat bath with a distinct temperature. The result is
n(cid:105) is independent of n. Adding weak coupling
To get equipartition of energy one could change the coupling to noise, replacing the final term in Eq. (21b) with
This is only possible for λn (cid:54)= 0, which precludes cyclic modes. Equation (4b) becomes
dφe = −(cid:88)
∇(cid:62)
evvdt + g(µ, φe)φedt +
v
n
(cid:88)
√
1
λn
φen
2DdBn(t).
The previous analysis goes through identically, leading to kTn = λ2
nDn/ = D/.
Differential growth rates
While the E/4 active modes per state that we observe is significantly reduced relative to the total number of modes
available, it is still a not insignificant fraction of E. There are, however, several straightforward generalizations of our
model that may lead to more strict mode selection. We discuss two possibilities in this and the subsequent section:
variations in activity across the network and variations in weights of vertices or edges.
For simplicity, we introduced Eqs. (4) with a uniform activity level µ across the entire network. This leads to equal
driving on all modes: if Eq. (16) is initialized near zero, it can be linearized to
d
dτ
(A2
0m) = µA2
0m,
where all modes grow at the same rate. Mode selection occurs in this system only because of interactions between
modes.
In many physical systems, however, differences in growth rate between modes are important for mode selection. For
example, the Rayleigh-Plateau instability [48] causes fluid jets to break apart into droplets whose size is determined
by the fastest growing unstable perturbation to the jet radius. Nonlinear mode competition akin to that in Eqs. (4)
may only act on the subset of modes that grow quickly.
18
We can add this effect to our model by replacing µ in Eqs. (4) with edge-dependent parameters µe. With the
v
√
∇(cid:62)
dφe
dt
= −(cid:88)
Following through the previous calculations with this change, Eq. (14) becomes
quadratic driving of Eq. (12), Eq. (4b) becomes
(cid:1) φe +
φemµef0lφel −(cid:88)
µe. However, if we again ignore degeneracies, the only resonant term is (cid:80)
defining νm =(cid:80)
(cid:33)
evv + (cid:0)µe − φ2
(cid:88)
t r1m + 2∂t∂τ r0m = −λ2
∂2
emµe, Eq. (16) then reads
mr1m + λm
(cid:32)
e φ2
e,l
e
φem
e
2Dξe(t).
(cid:33)3 .
f0nφen
(cid:32) E(cid:88)
n=1
The first term inside the square brackets no longer simplifies, since the φen are not orthonormal with the weighting
emµef0m from l = m. In this case,
e φ2
d
dτ
(A2
0m) = A2
0m
νm − E(cid:88)
k=1
PmkA2
0k
,
(22)
where modes have distinct growth rates independent of their interactions. Alternatively, one could specify νm arbi-
trarily in Eq. (22), though this would require more complex changes in Eq. (4b) coupling activity across edges.
Band gaps
3
In addition to distinct activity levels µe across edges, we can also introduce edge weights γe or vertex weights
mv that vary across the network. Changing the conductances γe and volumes mv changes our system in two ways:
first, by changing the modes to the singular vectors of γ∗
ve; and second, by changing the coupling matrix to Pmk =
2 (1 − 1
ek, which depends explicitly on the edge weights.
2 δmk)(cid:80)
Such changes are known to cause qualitative changes in the physics of classical spring-mass networks, including the
introduction of band gaps. In an infinite one-dimensional line of beads of equal mass m connected by springs with
equal spring constant f , for example, the dispersion relation between frequency ω and wavenumber q is
e γ−1
emφ2
e φ2
(cid:114)
(cid:12)(cid:12)(cid:12)sin
(cid:16) qa
2
(cid:17)(cid:12)(cid:12)(cid:12) ,
f
m
ω(q) = 2
Fig. S8. The emergence of an activity-driven spectral band gap is exhibited by a simulation on a 14-vertex path with (a)
all weights equal to 1 and (b) alternating vertex weights 1 and 5. Modes are ordered by frequency from high (top) to low
(bottom). Note that in (b) the central n = 7 mode is always active and the low energy states on the right half of the plot
are significantly more suppressed than they ever are in (a). The qualitative difference is due to the presence of vertices with
unequal weights, not the overall scale of the vertex weights; changing vertex weights uniformly is equivalent to rescaling other
parameters. Parameters were µ = 1.2, D = 5 × 10−3, and = 0.5. Both simulations used the same random seed.
19
(cid:18) 1
(cid:19)
(cid:115)(cid:18) 1
(cid:19)2 − 4
sin2(cid:16) qa
(cid:17)
.
where a is the size of the unit cell, in this case equal to distance between adjacent beads [48]. If instead of equal
masses the beads alternate between a smaller mass m1 and larger mass m2, the dispersion relation splits into two
branches,
+
1
m2
± f
+
1
m2
ω(q)2± = f
Here a unit cell has two beads, so the distance between beads is a/2. At q = π/a, there is a gap between ω+ =(cid:112)2f /m1
and ω− = (cid:112)2f /m2. This band gap shows up in a finite system as a large difference in frequency between modes
m1m2
m1
m1
2
above and below the gap.
Since varying what are effectively vertex weights causes such a clear qualitative change in behavior in the spring
system, we can reasonably expect similar changes in our model. Simulations on paths with alternating vertex weights
show a distinct separation of of low- and high-energy states not present with uniform weights (Fig. S8), with stronger
and more consistent suppression of the low-energy states and few transitions across the band gap created by nonuniform
weights. Band gaps in more realistic topologies may have similar effects, allowing for enhanced control of the large-scale
behavior.
|
1503.01808 | 1 | 1503 | 2015-03-05T22:31:54 | Detecting temperature fluctuations at equilibrium | [
"physics.bio-ph",
"cond-mat.stat-mech"
] | Gibbs and Boltzmann definitions of temperature agree only in the macroscopic limit. The ambiguity in identifying the equilibrium temperature of a finite sized `small' system exchanging energy with a bath is usually understood as a limitation of conventional statistical mechanics. We interpret this ambiguity as resulting from a stochastically fluctuating temperature coupled with the phase space variables giving rise to a broad temperature distribution. With this ansatz, we develop the equilibrium statistics and dynamics of small systems. Numerical evidence using an analytically tractable model shows that the effects of temperature fluctuations can be detected in equilibrium and dynamical properties of the phase space of the small system. Our theory generalizes statistical mechanics to small systems relevant to biophysics and nanotechnology. | physics.bio-ph | physics | Detecting temperature fluctuations at equilibrium
Department of System Biology, Columbia University∗
Purushottam D. Dixit
Gibbs and Boltzmann definitions of temperature agree only in the macroscopic limit. The ambi-
guity in identifying the equilibrium temperature of a finite sized 'small' system exchanging energy
with a bath is usually understood as a limitation of conventional statistical mechanics. We interpret
this ambiguity as resulting from a stochastically fluctuating temperature coupled with the phase
space variables giving rise to a broad temperature distribution. With this ansatz, we develop the
equilibrium statistics and dynamics of small systems. Numerical evidence using an analytically
tractable model shows that the effects of temperature fluctuations can be detected in equilibrium
and dynamical properties of the phase space of the small system. Our theory generalizes statistical
mechanics to small systems relevant to biophysics and nanotechnology.
5
1
0
2
r
a
M
5
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
8
0
8
1
0
.
3
0
5
1
:
v
i
X
r
a
Introduction: Equilibrium properties of a macro-
scopic system exchanging energy with a bath can be de-
scribed by a single intensive paramter, its temperature,
with remarkable accuracy; independently of the chemi-
cal nature of the bath and system-bath interactions ow-
ing to weak coupling between the system and the bath.
On the other hand, it is unlikely that a bath couples
weakly to a system with small number of degrees of free-
dom; consequently, small systems including biophysical
polymers (1) and nanomagnets (2) show considerable de-
viations from the traditional statistical mechanical de-
scription (3). Mathematically, no inverse temperature β
exists such that the exponential canonical ensemble dis-
tribution accurately predicts equilibrium properties of a
small system solely dependent on its Hamiltonian. Alter-
natively, Gibbs' definition of temperature which depends
on the typical value of energy and the Boltzmann's defini-
tion of temperature which depends on the mean value of
energy differ substantially from each other in the case of
small systems systems (4, 5). Traditionally, this ambigu-
ity is interpreted as an inevitable statistical uncertainty
in parameter estimation or a limitation of statistical me-
chanics (4, 6 -- 9).
In this communication, instead of treating the ambigu-
ity in identifying a unique temperature as a limitation,
we let go of the notion of a unique temperature, espe-
cially for small systems. We identify the ambiguity as
a consequence of a broad distribution peq(β). Further-
more, we identify the broad temperature distribution as
the ¯r−marginalization of the joint equilibrium distribu-
tion peq(¯r, β) of the stochastic variable (¯r(t), β(t)) where
¯r is the phase space of the system.
Using maximum entropy arguments, we first estimate
the joint equilibrium distribution peq(¯r, β) by introduc-
ing two new intensive parameters in the hyperensemble.
We then show how our theory reduces to traditional sta-
tistical mechanics of macroscopic systems in the suitable
limit. We illustrate a connections with non-extensive sta-
tistical mechanics of Tsallis (10, 11) and our theory at
thermodynamic equilibrium. Then, we propose Fokker-
Planck and Langevin equations for the time evolution
of the instantaneous distribution p(¯r, β; t). Finally, us-
ing realistic all atom molecular dynamics simulations, we
present numerical evidence to support our framework and
discuss its limitations.
Statitical mechanics of small systems: Consider
a small system as above. Due to possible non-weak
coupling between system and the bath, the equilibrium
phase space distribution of the system peq(¯r) will depend
on the nature of system-bath interactions (12 -- 14). Let
us work with the ansatz that the non-canonical behav-
ior arises because the temperature of the system fluc-
tuates (9, 11, 15). The joint equilibrium distribution is
simply
peq(¯r, β) = peq(¯rβ) × peq(β).
In Eq. 1,
peq(¯rβ) = eβ(F (β)−H(¯r))
(1)
(2)
is the usual Boltzmann distribution and peq(β) needs to
be determined. Since there are no conservation laws for
temperature, Gibbs' ensemble picture is inapplicable. We
resort to an equally valid alternative. We employ the
maximum entropy (maxEnt) framework (16, 17). We
maximize the entropy of the joint distribution p(¯r, β) =
p(¯rβ)×p(β) subject to suitable constraints. The entropy
of the joint distribution is given by
¯r,β
S [p(¯r, β)] = −(cid:88)
= −(cid:88)
s(β) = −(cid:88)
(cid:88)
β
¯r
where
(cid:88)
(3)
s(β)p(β)
(4)
p(¯r, β) log p(¯r, β)
p(β) log p(β) +
β
p(¯rβ) log p(¯rβ),
(5)
(6)
p(β) =
p(¯r, β)
¯r
is the ¯r−marginal of p(¯r, β), and p(¯rβ) is given by Eq. 2.
When determining peq(β) the choice of constraints is
important. Since the temperature of the system is not
fixed, we choose (cid:104)β(cid:105) as a constraint. Also, while the en-
tropy of the composite macroscopic system comprising
the system and the surrounding bath is maximized, the
entropy of the small system itself not. Consequently, we
choose the average entropy (cid:104)s(β)(cid:105) as an additional con-
straints and maximize S [p(¯r, β)] using Lagrange multi-
pliers. The constraint of average entropy is a common in
statistical physics and Bayesian statistics of hyperensem-
bles. See (18 -- 21) for different motivations behind this
choice. After maximization, we find that the equilibrium
distribution peq(β) is estimated by
peq(β) =
eλs(β)−ζβ
Z(λ, ζ)
(7)
In Eq. 7, Z is a generalized partition function and λ
and ζ are Lagrange multipliers that determine the shape
of peq(β).
If entropy s(β) is a unitless number, then
λ is unitless and ζ has the units of 1/β. The physical
interpretation of these Lagrange multipliers will become
clearer below.
joint
peq(¯rβ) × peq(β) is
equilibrium distribution peq(¯r, β) =
The
2
scales proportional to the temperature, U (β) = U0/β,
when coupled to a bath of ideal gas particles at in-
verse temperature β. These are excellent assumptions for
bound systems where density of states increases mono-
tonically with energy. Examples include ideal gas in a
container and a collection of harmonic oscillators. From
Eq. 7, we have
peq(β) =
e−βζβλs0 ζ λs0+1
Γ(λs0 + 1)
.
(9)
Eq. 22 is a Gamma distribution also known as the gen-
eralized χ−squared distribution. Interestingly, a gamma
distributed inverse temperature is very commonly used
in a superstatistical explanation of non-extensive statis-
tics (22). Marginalizing over gamma distributed inverse
temperature in Eq. 1 results in the so called "Tsallis
statistics" for the phase space. We have
peq(¯r, β) =
=
eβ(U (β)−s(β)/β)−βH(¯r)+λs(β)−ζβ
eU0−β(H(¯r)+ζ)+(λ−1)s0 log(β)
Z(λ, ζ)
Z(λ, ζ)
(10)
.
(11)
peq(¯r, β) =
eβF (β)−βH(¯r)+λs(β)−ζβ
Z(λ, ζ)
(8)
Integrateing over β, we have
Thus, instead of describing a thermally equilibrated small
system with one intensive parameter, its inverse temper-
ature β, our framework requires two intensive parameters
λ and ζ whose meaning will become clear below.
Connections to traditional statistical mechan-
ics: Assume that the entropy s(β) is monotonically de-
creasing in β, a reasonable assumption for systems with
monotonically increasing density of states. A straightfor-
ward calculation shows that the maximum of peq(β) is sit-
uated at β = β0 where β0 is such that ζ/λ = −c(β0)/β0.
Here, c(β0) is the heat capacity of the system when in-
teracting with an ideal gas at inverse temperature β0.
In the limiting case when λ → ∞ and ζ → ∞ such
that their ratio is constant, non-negligible contribution
to peq(β) comes only from near β = β0 and peq(β) ≈
δ(β − β0) where δ(x) is the Dirac Delta function. This is
exactly the traditional canonical ensemble picture where
the system is assigned the temperature of the surround-
ing thermal bath. It is clear that the magnitudes of λ
and ζ dictate the breadth of the peq(β) distribution and
hence the deviation from canonical ensemble. The ratio
λ/ζ dictates the most likely tempearture of the system.
Connection to non-extensive statistical me-
chanics: Systems that do not obey the conventional dis-
tributions from statistical mechanics are sometimes en-
tertained within a framework called non-extensive statis-
tical mechanics (10). Though not commonly invoked for
small systems at equilibrium, here, we will demonstrate
that non-extensive statistical mechanics can be arrived at
by marginalization over temperature in a hyperensemble.
Consider a system whose entropy scales as logarithm
of temperature, s(β) = s0 log β, and the internal energy
peq(¯r) ∝ (1 − β0(q − 1)H(¯r))
(12)
Eq. 12 is the q−generalized canonical ensemble distribu-
tion in Tsallis statistics where
1
q−1 .
q =
s0 − λ
s0 − λ − 1
and β0 =
.
(13)
λ − s0 + 1
ζ
In the framework of non-extensive statistical mechan-
ics, one arrives at Eq. 12 by maximizing Tsallis' q entropy
with respect to p(¯r) by constraining an unnatural escort
expectation of energy (10).
In this work, in contrast to deriving peq(¯r) by maximiz-
ing the non-extensive Tsallis entropy by constraining an
unnatural expectation value, we derive it from a super-
statistical distribution Eq. 8 and additional assumptions
about peq(β) and system behavior.
In our derivation,
the gamma distribution peq(β) arises in a context specific
manner i.e. through the logarithmic dependence of the
entropy on the inverse temperature and by constraining
average inverse temperature. Therefore, starting from
the extensive Gibbs-Shannon entropy, maxEnt can act
as a predictive framework for constructing non-extensive
effective entropies (23) of which the Tsallis entropy is a
particular example.
Previously, non-extensive entropies have been criti-
cized from an Occam's razor point of view (16, 24 -- 26)
when compared to the Gibbs-Shannon entropy. Our work
suggests that non-extensive entropies may arise as 'ef-
fective entropies' when considering extensive entropies
in a hyperensemble. Nevertheless, there is a potential
loss of information when marginalizing over the temper-
ature β in the hyperensemble that is inherent to con-
structing these effective entropies. We believe that the
above demonstration argues in favor the extensive Gibbs-
Shannon entropy, albeit in a hyperensemble, even when
the observable phase space may show non-extensive be-
havior.
in r, we get
(cid:90) t
0
ds · l12 · e−l22s
r = l11r + l12el22t
(cid:90) t
+ l12el22t
ds · ηβ · e−l22s + ηr
⇒ r = (l11 + l22) r + (l12l21 − l11l22) r
0
+ (l12ηβ − l22ηr) + ηr
3
(19)
(20)
Stochastic Dynamics: For simplicity of notation,
let us consider a one dimensional system. The simplest
time evolution of the instantaneous distribution p(r, β; t)
of the extended phase space that relaxes to a prescribed
equlibrium distribution peq(r, β) can be modeled by an
over damped Smoluchowski equation. We have
(cid:18) 1
(cid:19)
∂p(r, β; t)
∂t
= −
[fr · p] +
∂
∂r
1
γβ
∂
∂β
[fβ · p]
γr
∂2p
∂r2 + Dβ
+ Dr
∂2p
∂β2
(14)
where the 'forces' fr and fβ are defined as
fr =
∂
∂r
log peq(r, β) and fβ =
∂
∂β
log peq(r, β). (15)
By construction, Eq. 14 will relax to the equilibrium dis-
tribution peq(r, β) if Dr = 1/γr and Dβ = 1/γβ. Note
that the statistical properties of (r(t), β(t)) can also be
estimated by an overdamped Langevin equation (Brown-
ian dynamics) that is equivalent to Eq. 14. The Langevin
equation reads
(cid:112)
β = Dβfβ +(cid:112)2Dβηβ
r = Drfr +
2Drηr
(16)
Here, ηr and ηβ are usual uncorrelated Gaussian random
variables with unit variance.
Linear analysis: It is instructive to study a linear
system before analyzing realistic molecules. Consider a
one dimensional harmonic oscillator interacting with a
thermal bath. If the deviations from a canonical distri-
bution are negligible, we can treat Eq. 16 in the linear
regime by expanding fr and fβ to the first order in r and
β. In the linear approximation, the joint equilibrium dis-
tribution peq(r, β) will be described by a joint normal dis-
tribution. The simplest coupled system of overdamped
Langevin equations for r(t) and β(t) that relaxes to to a
joint normal distribution is given by
r ≈ l11r + l12β + ηr
β ≈ l21r + l22β + ηβ
(17)
(18)
We have assumed that the variables r and β are appro-
priately scaled by absorbing the diffusion constants Dr
and Dβ, lij are the scaled linear expansion coefficients
of fr and fβ, and ηr and ηβ are the usual uncorrelated
Gaussian noises. Integrating over β(t) and substituting
The time derivative of white noise ηr is a purple noise
which has quadratically increasing power spectrum. The
dynamics of temperature fluctuations are governed by the
linear terms l12, l21, l22, and the white noise ηβ. These
terms also appear in the effective Langevin equation for
r(t). The linear analsysis suggests that one can infer the
of dynamics of β(t) by observing the dynamics of r(t).
The dynamics of r(t) is governed by a much richer
equation than the usual overdamped Langevin equation.
A one dimensional small linear harmonic oscillator ex-
changing energy with a thermal bath can be modeled by
a second order Langevin equation with a combination of
white and purple noise. These predictions can be tested
by observing dynamical properties of a small colloidal
particle trapped in a harmonic well using optical traps.
A 'small' harmonic oscillator: How do we verify
the effects of temperature fluctuations on the phase space
of a small system? We resort to realistic molecular dy-
namics simulations of an analytically tractable system
viz. a harmonic oscillator.
Consider a three dimensional dumbell shaped Lennard-
Jones harmonic oscillator interacting non-weakly with a
bath. Realistic examples include colloidal beads tied to
each other by a biopolymer or linear molecules such as
CO2. The canonical ensemble distribution for the Har-
monic oscillator is given by
peq(rβ) =
4β3/2r2√
π
× e−βr2
(21)
where r is the displacement of the oscillator. Without
loss of generality, we have assumed that the spring con-
stant of the oscillator is k = 2. If the system-bath inter-
actions are non-negligible, we expect that the equilibrium
phase space distribution of the oscillator will deviate con-
siderably from the Boltzmann distribution.
The entropy of the oscillator scales as s(β) ∼ log β and
from Eq. 7, we know that the equilibrium distribution
peq(β) will be governed by a Gamma distribution
e−βζβλζ λ+1
Γ(λ + 1)
peq(β) =
(22)
The joint equilibrium distribution peq(r, β) = peq(rβ) ×
peq(β) on the other hand is obtained by multiplying
Eq. 21 and Eq. 22
.
peq(r, β) =
2 ζ λ+1e−β(ζ+r2)
4r2βλ+ 3
√
πΓ(λ + 1)
.
(23)
4
II). In Fig. 2 we compare the numerically estimated au-
tocorrelation function
C(τ ) = (cid:104)r(τ )r(0)(cid:105)eq − (cid:104)r(cid:105)2
eq
(26)
from MD simulation (black squares) and the prediction
from the 2-d Langevin equation (red). The predictions
from an analogous 1-d Langevin equation that relaxes to
peq(r) of Eq. 24 are shown in blue. While the dynam-
ics observed in the MD simulation has two time scales
resulting in a double exponential decay in the autocorre-
lation function, the 1-d Langevin equation is only able to
capture one effective time scale. On the other hand, the
2-d Langevin equation has two natural time scales gov-
erned by Dr and Dβ respectively. The coupled Langevin
equation equivalent to Eq. 14 (see appendix II) with
dt = 5×10−8 and Dβ ≈ 50×Dr does indeed captures the
autocorrelation function while an analogous 1-d equation
fails to do so (see appendix II for details of the fit).
In appendix III we show that the theoretical predic-
tions are valid over a range of bath temperatures and
system-bath interactions. In this work, we study a sys-
tem whose canonical ensemble distribution can be analyt-
ically computed and the entropy analytically estimated.
This allowed us to compute peq(β) and peq(r, β) analyt-
ically. For more realistic systems with multiple degrees
of freedom, peq(rβ) needs to be estimated numerically
along with peq(β).
In summary, the mesoscopic harmonic oscillator inter-
acting with a thermal bath of water molecules shows sig-
nificant deviation from the canonical ensemble descrip-
tion. We can correctly predict both equilibrium and dy-
namical properties of the oscillator by allowing its tem-
perature to vary as a stochastic variable which is coupled
with the phase space variable r(t).
Discussion:
It is known that, at equilibrium, mesoscopic systems
have larger fluctuations compared to a macrosopic sys-
tem. We have argued that these enhanced fluctuations
be understood as arising from a dynamically fluctuating
temperature.
How do we reconcile a time dependent temperature,
a non-equilibrium phenomena prima facie, in an equilib-
rium setting? Even though the temperature is chang-
ing, the extended phase space (¯r(t), β(t)) is still gov-
erned by a detailed-balanced Markov process.
It's an
easy calculation to show that the entropy production,
as defined in stochastic thermodynamics (27), is indeed
zero for the hyperensemble. Nevertheless, there are mul-
tiple questions which need resolution. For example, How
do we formulate non-equilibrium phenomena in the hy-
perensemble setting? For example, how do we modify
non-equilibrium fluctuation relationships (28) for small
systems? We leave this to future work.
FIG. 1. We study the equilibrium properties of a 3D dumbbell
shaped harmonic oscillator comprising of Lennard Jones par-
ticles interacting with a bath of water molecules at 300 K (see
appendix I for details) using all atom MD simulations. The
numerically obtained marginal distribution p(r) of the oscilla-
tor separation r (black squares) is better captured by Eq. 24
(red line) than the usual canonical ensemble distribution of
Eq. 21 (blue line).
Integrating over all values of β, we obtain the marginal
r distribution
4r2ζ λ+1Γ(cid:0)λ + 5
2
(cid:1)(cid:0)ζ + r2(cid:1)−λ− 5
2
peq(r) =
.
(24)
√
πΓ(λ + 1)
Moreover, we can also model the dynamics of the oscil-
lator by the coupled Langevin equation of Eq. 16. From
Eq. 23, the "forces" fr and fβ are given by
fr =
2
r
− 2rβ and fβ =
3 − 2βr2 − 2βζ + 2λ
2β
(25)
Eq. 24 along with Eq. 16 where the forces fr and fβ
are given by Eq. 25 are our predictions for the Harmonic
oscillator regardless of the bath that is interacting with.
These predictions can be tested experimentally or in a
realistic numerical simulation.
Numerical validation: With the aid of MD simu-
lations of a dumbbell shaped Lennard-Jones harmonic
oscillator coupled to a bath of water molecules at 300 K
(see appendix I for details), we confirmed the numerical
superiority of Eq. 24 compared to Eq. 21 and estimated
the parameters λ ≈ 2.19 and ζ ≈ 0.34. Fig. 1 shows
that Eq. 24 which allows for a broad temperature distri-
bution indeed fits the numerically estimated distribution
much better than the usual canonical ensemble distri-
bution of Eq. 21. It is clear that by allowing the inverse
temperature to have a broad distribution, the equilibrium
properties of the harmonic oscillator interacting with its
thermal surroundings are captured correctly.
The dynamics of r(t) can be predicted using Eq. 14 by
studying the equivalent Langevin equation (see appendix
□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□0.00.51.01.52.010-310-20.11rP(r) 5
Gibbs statistics. Journal of Statistical Physics 52:479 --
487.
[11] Wilk, G., and Z. W(cid:32)lodarczyk, 2000.
Interpretation of
the nonextensivity parameter q in some applications of
Tsallis statistics and L´evy distributions. Physical Review
Letters 84:2770.
[12] Dixit, P. D., and D. Asthagiri, 2011. The role of bulk
protein in local models of ion-binding to proteins. Com-
parative study of KcsA, its semi-synthetic analog with
a locked-in binding site, and Valinomycin. Biophys. J.
100:1542 -- 1549.
[13] Dixit, P., and D. Asthagiri, 2011. An elastic-network-
based local molecular field analysis of zinc finger proteins.
J. Phys. Chem. B 115:7374 -- 7382.
[14] Dixit, P., and D. Asthagiri, 2012. Role of Local Metal-
Site Interactions and Bulk Protein Restraints in the
Thermodynamics of Zinc Binding to a Zinc Finger Pro-
tein. Biophys. J. 102:457.
[15] Touchette, H., 2004. Temperature fluctuations and mix-
tures of equilibrium states in the canonical ensemble.
Nonextensive EntropyInterdisciplinary Applications 159 --
176.
[16] Press´e, S., K. Ghosh, J. Lee, and K. A. Dill, 2013. The
principles of Maximum Entropy and Maximum Caliber
in statistical physics. Rev. Mod. Phys. 85:1115 -- 1141.
[17] Jaynes, E. T., 1957. Information theory and statistical
mechanics I. Phys. Rev. 106:620 -- 630.
[18] Caticha, A., and R. Preuss, 2004. Maximum entropy
and Bayesian data analysis: Entropic prior distributions.
Phys. Rev. E 70:046127.
[19] Dixit, P. D., 2013. Quantifying Extrinsic Noise in Gene
Expression Using the Maximum Entropy Framework.
Biophysical Journal 104:2743 -- 2750.
[20] Crooks, G. E., 2008. Beyond Boltzmann-Gibbs statistics:
Maximum entropy hyperensembles out-of-equilibrium.
Phys. Rev. E 75:041119.
[21] Dixit, P. D., 2013. A maximum entropy thermodynamics
for small systems. J. Chem. Phys. 138:184111.
[22] Beck, C., and E. G. D. Cohen, 2003. Superstatistics.
Physica A 322:267 -- 275.
[23] Hanel, R., S. Thurner, and M. Gell-Mann, 2011. Gener-
alized entropies and the transformation group of super-
statistics. Proceedings of the National Academy of Sci-
ences 108:6390 -- 6394.
[24] Press´e, S., K. Ghosh, J. Lee, and K. A. Dill, 2013. Non-
additive entropies yield probability distributions with bi-
ases not warranted by the data. Physical review letters
111:180604.
[25] Peterson, J., P. D. Dixit, and K. A. Dill, 2013. A max-
imum entropy framework for nonexponential distribu-
tions. Proceedings of the National Academy of Sciences
110:20380 -- 20385.
[26] Press´e, S., 2014. Nonadditive entropy maximization is
inconsistent with Bayesian updating. Physical Review E
90:052149.
[27] Seifert, U., 2008. Stochastic thermodynamics: princi-
ples and perspectives. The European Physical Journal
B-Condensed Matter and Complex Systems 64:423 -- 431.
irre-
versibility and the second law of thermodynamics at the
nanoscale. In Time, Springer, 145 -- 172.
[28] Jarzynski, C., 2013. Equalities and inequalities:
[29] Jorgensen, W., J. Chandrasekhar, J. D. Madura, R. W.
Impey, and M. L. Klein, 1983. Comparison of simple
potential functions for simulating liquid water. J. Chem.
FIG. 2. We study the autocorrelation function of the har-
monic oscillator interacting with a bath of water molecules.
We model the dynamics of the extended phase space
(r(t), β(t)) using a simple coupled Langevin equation (see ap-
pendix II). We find that the 2 dimensional Langevin equation
(red line) captures the two time scales inherent to the dynam-
ics of r(t) as observed in MD simulations (black squares). On
the other hand, an analogous 1-d Langevin equation can only
capture one effective time scale (blue line).
Acknowledgment: We thank Ken Dill, Steve Press´e,
Dilip Asthagiri for a critical reading of the manuscript.
We thank Sumedh Risbud, Karthik Shekhar, Manas
Raach, and Anjor Kanekar for fruitful discussions.
∗ email:[email protected]
[1] Pohorille, A., C. Jarzynski, and C. Chipot, 2010. Good
practices in free-energy calculations. J. Phys. Chem. B
114:10235 -- 10253.
[2] Chamberlin, R. V., J. V. Vermaas, and G. H. Wolf, 2009.
Beyond the Boltzmann factor for corrections to scaling
in ferromagnetic materials and critical fluids. The Euro-
pean Physical Journal B-Condensed Matter and Complex
Systems 71:1 -- 6.
[3] Hill, T. L., 2002. Thermodynamics of small systems.
Dover publications, Mineola, New York.
[4] Mandelbrot, B. B., 2008. Temperature Fluctuation: A
Well-Defined and Unavoidable Notion. Physics Today
42:71 -- 73.
[5] Dunkel, J., and S. Hilbert, 2013. Consistent thermo-
statistics forbids negative absolute temperatures. Nature
Physics .
[6] Mandelbrot, B., 1962. The role of sufficiency and of esti-
mation in thermodynamics. The Annals of Mathematical
Statistics 1021 -- 1038.
[7] Uffink, J., and J. van Lith, 1999. Thermodynamic uncer-
tainty relations. Foundations of physics 29:655 -- 692.
[8] Schlogl, F., 1988. Thermodynamic uncertainty relation.
Journal of Physics and Chemistry of Solids 49:679 -- 683.
[9] van Hemmen, J. L., and A. Longtin, 2013. Temperature
Fluctuations for a System in Contact with a Heat Bath.
Journal of Statistical Physics 153:1132 -- 1142.
[10] Tsallis, C., 1988. Possible generalization of Boltzmann-
□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□01002003004005000.020.030.040.050.06timeτ(fs)C(τ) Phys 79:926 -- 935.
[30] Neria, E., S. Fischer, and M. Karplus, 1996. Simulation
of activation free energies in molecular systems. J. Chem.
Phys 105:1902 -- 1921.
[31] Phillips,
J., R. Braun, W. Wang,
J. Gumbart,
E. Tajkhorshid, E. Villa, C. Chipot, R. Skeel, L. Kale,
and K. Schulten, 2005. Scalable Molecular Dynamics
with NAMD. J. Comp. Chem. 26:1781 -- 1802.
[32] MacKerell, Jr., A. D., D. Bashford, M. Bellott, R. L.
Dunbrack, Jr., J. D. Evanseck, M. J. Field, S. Fis-
cher, J. Gao, H. Guo, S. Ha, and D. Joseph-McCarthy,
1998. All-atom empirical potential for molecular model-
ing and dynamics studies of proteins. J. Phys. Chem. B
102:3586 -- 3616.
APPENDIX I: MD SIMULATIONS
A harmonic dumbbell oscillator consisting of two
Lennard-Jones particles was immersed in a bath of 333
TIP3 (29, 30) water molecules. NVT molecular dynam-
ics simulations were run with NAMD (31) at 300K with
a box size of 19.12A. The CHARMM (32) forcefield was
used to describe the interaction between the harmonic os-
cillator particles and surrounding water molecules. The
spring constant for the dumbell was chosen to be k =
0.25 kcal/mol·A2, the parameter was set at = −20.0
kcal/mole and the size parameter was set at r = 1A.
The systems were minimized for 2000 steps followed by
an equilibration of 1 nanosecond and a production run of
2 nanosecond. The integration time step was 0.25 fem-
toseconds and the trajectory was saved every 2.5 fem-
toseconds.
APPENDIX II: FITTING LANGEVIN
DYNAMICS TO DATA
6
The coupled Langevin equation corresponding to
Eq. 14 where the equilibrium distribution peq(r, β) is
given by Eq. 22 is given by
(cid:18) r(t + dt)
(cid:19)
β(t + dt)
(cid:19)
(cid:18) Drfr
(cid:18) r(t)
(cid:19)
(cid:18) ηr
(cid:19)
(cid:112)Dβ
+ dt
√
Dβfβ
β(t)
√
2dt
Dr
≈
+
ηβ
(27)
Here, ηr and ηβ are uncorrelated Gaussian random vari-
ables with unit variance, dt is a small time step, Dr and
Dβ are diffusion coefficients for the phase space coordi-
nate r and the temperature β.
From the MD simulation, we first estimated the auto-
correlaion function C(τ ). The Langevin equation can be
scaled in time by multiplying the diffusion constants and
dividing the time step dt by the same number. In order
to ensure smooth integration, we first set the integration
time step to a very small value; dt = 5 × 10−8. Every
pair (Dr, Dβ) of diffusion constants predicted an auto-
correlation function that had two inherent time scales
manifested in a double exponential decay. We man-
ually scanned the (Dr, Dβ)−space to match the MD-
autocorrelation function. We found that Dr = 1 and
Dβ = 50 gave reasonable fits (red curve).
We also wrote down a 1-d Langevin equation analogous
to Eq. 27,
r(t + dt) ≈ r(t) + Drfrdt +
2Drdtηr
(28)
(cid:112)
where fr = d
dr log peq(r) (see Eq. 24). This equation
had only one diffusion constant Dr. A one dimensional
scan of Dr suggested that the autocorrelation function
predicted using the 1-d Langevin equation always had a
single exponential decay. We found the best fit to the
autocorrelation function at Dr ≈ 50 (blue curve).
|
1804.01696 | 1 | 1804 | 2018-04-05T07:19:36 | Threshold Response to Stochasticity in Morphogenesis | [
"physics.bio-ph",
"cond-mat.other",
"q-bio.MN"
] | During development of biological organisms, multiple complex structures are formed. In many instances, these structures need to exhibit a high degree of order to be functional, although many of their constituents are intrinsically stochastic. Hence, it has been suggested that biological robustness ultimately must rely on complex gene regulatory networks and clean-up mechanisms. Here we explore developmental processes that have evolved inherent robustness against stochasticity. In the context of the Drosophila eye disc, multiple optical units, ommatidia, develop into crystal-like patterns. During the larva-to-pupa stage of metamorphosis, the centers of the ommatidia are specified initially through the diffusion of morphogens, followed by the specification of R8 cells. Establishing the R8 cell is crucial in setting up the geometric, and functional, relationships of cells within an ommatidium and among neighboring ommatidia. Here we study a mathematical model of these spatio-temporal processes in the presence of stochasticity, defining and applying measures that quantify order within the resulting spatial patterns. We observe a universal sigmoidal response to increasing transcriptional noise. Ordered patterns persist up to a threshold noise level in the model parameters. As the noise is further increased past a threshold point of no return, these ordered patterns rapidly become disordered. Such robustness in development allows for the accumulation of genetic variation without any observable changes in phenotype. We argue that the observed sigmoidal dependence introduces robustness allowing for sizable amounts of genetic variation and transcriptional noise to be tolerated in natural populations without resulting in phenotype variation. | physics.bio-ph | physics |
Threshold Response to Stochasticity in
Morphogenesis
George Courcoubetis1, Sammi Ali 2, Sergey V Nuzhdin2, Paul Marjoram3, and
Stephan Haas1
1Department of Physics & Astronomy, University of Southern California, Los
Angeles, California, United States of America
2Department of Molecular and Computational Biology, University of Southern
California, Los Angeles, California, United States of America
3 Department of Preventative Medicine, Keck School of Medicine of USC, Los
Angeles, California, United States of America
April 4, 2018
Abstract
During development of biological organisms, multiple complex structures are formed.
In many instances, these structures need to exhibit a high degree of order to be func-
tional, although many of their constituents are intrinsically stochastic. Hence, it has
been suggested that biological robustness ultimately must rely on complex gene regula-
tory networks and clean-up mechanisms. Here we explore developmental processes that
have evolved inherent robustness against stochasticity. In the context of the Drosophila
eye disc, multiple optical units, ommatidia, develop into crystal-like patterns. During
the larva-to-pupa stage of metamorphosis, the centers of the ommatidia are specified
initially through the diffusion of morphogens, followed by the specification of R8 cells.
Establishing the R8 cell is crucial in setting up the geometric, and functional, relation-
ships of cells within an ommatidium and among neighboring ommatidia. Here we study
a mathematical model of these spatio-temporal processes in the presence of stochas-
ticity, defining and applying measures that quantify order within the resulting spatial
patterns. We observe a universal sigmoidal response to increasing transcriptional noise.
Ordered patterns persist up to a threshold noise level in the model parameters. As the
noise is further increased past a threshold point of no return, these ordered patterns
rapidly become disordered. Such robustness in development allows for the accumula-
tion of genetic variation without any observable changes in phenotype. We argue that
the observed sigmoidal dependence introduces robustness allowing for sizable amounts
of genetic variation and transcriptional noise to be tolerated in natural populations
without resulting in phenotype variation.
1
Author summary
The development of biological organisms requires the formation of highly ordered struc-
tures. These structures are created utilizing cell signaling through morphogens, a pro-
cess that is inherently stochastic. In this paper, we apply rigorous measures to quantify
order and disorder in the context of the development of the Drosophila eye disc, R8
pattern formation in the larva-to-pupa stage. Specifically, we turn to mathematical
formulations of the mechanism, and we illustrate that introducing noise below a cer-
tain threshold does not affect the final outcome at all. Furthermore, we show that after
this threshold is crossed, there is rapid deterioration of the pattern that intensifies as
the size of the eye disc is increased. Through quantitative analysis, we explain the
origin of both the robustness and rapid deterioration of the pattern. These findings
highlight generic characteristics that are needed for developmental models to repro-
duce experimentally observed collective mechanisms, such as cleanup of superfluous
cells and cell threshold responses to stochasticity. We discuss the connections of in
these findings with canalization and cryptic genetic variation, i.e. the fact that envi-
ronmental differences and stochastically perturbed genetic makeups result to identical
phenotypes.
Introduction
Deterministic Outcomes from Inherently Stochastic Components
Biological systems are intrinsically noisy but nonetheless produce deterministic out-
comes. During development, organisms utilize signaling molecules, i.e. morphogens,
to generate a body plan and differentiate cells. With modern experimental techniques,
it is possible to measure temporal concentrations of selected morphogens in each cell.
The expression of a gene depends on the probabilistic outcomes of several factors,
such as molecular binding affinities, processivity, and regulatory sequence interactions.
Specifically, genetically identical cells produce morphogen transcripts in asynchronous
bursts and in varying quantities [1]. It is therefore essential to investigate how ordered
deterministic structures are formed from underlying stochastic components, such as
gene expression. Early studies have addressed robustness of developmental processes,
also termed as canalization [2], which remain subject of great interest today [3, 4, 5, 6].
Developmental robustness has been highlighted as a necessary condition that narrows
down dramatically the search for plausible models and one that can even "predict key
mechanistic and molecular properties of the associated biochemical circuits" [3]. For
example, the Turing mechanism has been deemed inapplicable in many instances due
to its sensitivity to noise common in developmental processes [6]. In this paper, we an-
alyze the response of a developmental system to noise quantitatively from a statistical
physics perspective.
2
Pattern Formation in the Drosophila Eye Disc is an Ideal Model
System to Study the Effects of Stochastic Transcription on De-
terministic Development
The compound eye, found primarily in insects, consists of an array of repeating visual
units. The Drosophila compound eye is made up of approximately 800 unit eyes, known
as ommatidia. In wild-type Drosophila, the structure of the ommatidia resembles a
near-perfect hexagonal lattice. This highly structured pattern is developed via the
delicate coordination of cell signaling, proliferation, movement and apoptosis [7]. Some
of these cellular processes are guided through the communication of a few conserved
molecules, known as morphogens, resulting in tissue morphogenesis.
While the resulting functional eye emerges in the adult, the role of each cell within
repeating ommatidial arrays, and other head structures, are specified during larval de-
velopment [7]. The larval eye-antenna imaginal disc, hereafter referred to as the eye
disc, contains numerous cells that produce various morphogens. The Drosophila eye
disc is one of the simplest systems for studying morphogenesis since this tissue develops
in a mostly two-dimensional fashion. This is because the eye-antenna disc is derived
from an epithelial monolayer [8]. Furthermore, this system is an ideal one in which to
model morphogenesis, since cell lineage has a minimal effect on pattern formation [9],
and differentiation relies primarily on cell-to-cell signaling. The underlying develop-
mental mechanisms that guide the formation of the eye disc from larva to adult have
been studied extensively, allowing for the generalization and analysis of biologically
realistic models.
To investigate the stability of the developmental pathways in the Drosophila eye
disc, mathematical modeling and numerical simulations are used and the results are
analyzed here from a physics perspective. Mathematical modeling of the Drosophila eye
disc pattern-formation mechanism has been the focus of previous investigations,
[10,
11] which proposed a mathematical model that reproduces the triangular lattice pattern
of differentiated R8 cells and that is robust enough to be biologically plausible. Using
this latest model as a basis, we examine and quantitatively test the robustness of the
emerging spatially ordered patterns of differentiated R8 cells when transcriptional noise
is introduced. To this end, we implement various appropriate measures of spatial order,
testing the functional relationship of R8 cell pattern order with increasing stochasticity.
In this model, morphogens are produced and emitted by individual cells in quan-
tities that are determined by morphogen inputs from other cells in their vicinity. In
addition, the production and diffusion of morphogens is non-uniform, i.e. a given mor-
phogen concentration in a cell does not result in a single production or diffusion rate.
These processes are thus inherently noisy. Therefore, the highly ordered structure of
the ommatidia in the Drosophila eye disc, emerging from such stochastic constituents,
requires robustness in the underlying gene regulatory networks. Our study aims to
quantify eye disc developmental robustness. We focus on one specific step out of many
involved in eye disc development, in which additional robustness is introduced [12].
However, the regulation of R8 cell distances is a crucial developmental step in precisely
placing each repeating eye unit, thereby strongly influencing the organism's overall
visual acuity.
3
Mechanism Leading to R8 Cell Specification
Each ommatidium is made up of 14 cells; 8 photoreceptors (R1-R8), 4 cone cells, and
2 primary pigment cells. The specification of the first photoreceptor, R8, guides the
specification and orientation of the remainder of the cells within the ommatidium.
Thus, the spacing between neighboring R8 cells is pivotal for positioning and refining
geometric relationships in resulting ommatidia. Cellular defects that misplace a single
ommatidium will influence the position of neighboring ommatidia, thus propagating
further flaws in the lattice. If even a single gene involved in ommatidia formation is
perturbed, all unit eyes can be affected, since the regulatory logic that generates each
ommatidium is repeated using the same set of morphogens and gene networks [7].
Prior to R8 cell specification, the eye disc is composed of tightly-packed undiffer-
entiated cells dividing asynchronously. Differentiation starts with the initiation of the
morphogenetic furrow (MF), a physical indentation in the eye disc that sweeps through
the tissue to dictate the pace at which ommatidia are specified and positioned. The MF
advances anteriorly via the communication of several morphogens, primarily Hedgehog
(Hh) and Decapentapegic (Dpp). It is initiated in the posterior end of the eye disc and
sweeps through the tissue towards the anterior end. As the MF sweeps through the
eye disc, the cells posterior to the MF differentiate and commit to their respective role
within the ommatidium. Anterior to the MF, the cells are asynchronously proliferat-
ing. At the anterior interface of the MF, cells are arrested in G1 to allow synchronous
divisions [13]. Within the MF, the central R8 photoreceptor cells are specified via the
expression of atonal (ato). The spacing between ato-expressing cells is refined poste-
riorly within the MF. This process starts with 5-7 ato-expressing cells (intermediate
groups), followed by 3 ato-expressing cells (equivalence groups) and finally a single
ato-expressing cell (R8) [14]. Controlling the spacing of R8 cells within and between
columns during MF progression is a crucial step in establishing the spacing pf omma-
tidia within and between visual columns.
Determining the central R8 cell drives the subsequent specification of the other cells
within the ommatidium. Hedgehog (Hh) induces long-range and short-range secondary
signals that control the precise position of the MF. Hh acts over a short range to induce
the expression of Decapentapegic (Dpp), which in turn diffuses over a long range to
turn off homeothorax (hth) and turn on hairy (h), establishing a pre-proneural domain
(PPN). Hh and Dpp also induce the expression of Delta (Dl), a trans-membrane ligand
that acts on adjacent cells to turn off hairy. DI also acts on the intermediate groups
to help refine the ato-expressing cluster, ultimately resulting in a single R8 cell per
equivalence group [15]. The single ato-expressing R8 cell produces morphogens that
inhibit nearby cells from further expressing ato. This ensures regular spacing between
R8 cells. The progression of the morphogenetic furrow allows R8 specification to be
repeated unidirectionally from posterior to anterior, while the regulation of the distance
between neighboring R8 cells generates the remarkable hexagonal array.
4
Process Leading to Pattern Formation in the Drosophila Eye
Disc
Here we provide a simplified description of the rational behind the current model,
describing the basic mechanism of pattern formation in the eye disc. This process
is described using a system of coupled differential equations intended to provide the
reader with an intuitive understanding of the foundations of the mathematical model.
The details of the full model are elaborated on in the methods section below.
The cells of the eye disc are assumed to lie on a N × N hexagonal grid, as appro-
priate for a hard sphere tight packing. After the MF passes, a portion of these cells
differentiate to become activated R8 cells, and the remainder will be left undifferenti-
ated. At the end of the process, activated R8 cell centers are positioned on a triangular
lattice. This process is illustrated in Fig 1.
Some morphogens are diffusible, while others are cell-specific and stay within the
boundaries of the cell. In this study, ato is the only non-diffusing, cell-specific mor-
phogen, while all others can diffuse between cells. The various morphogens either have
inhibitory or activating properties. The inhibitors are morphogens that, when present
in a cell, decrease the rate of production of another morphogen, and vice-versa for the
activators.
Initially, differentiated R8 cells are defined as cells that express any non-zero level
of ato . There are two main ingredients that make the pattern of R8 cells form. First,
there is an inhibition signal that blocks atonal. When a cell becomes differentiated, it
instantly produces an inhibitory signal that completely blocks the production rate of
ato of all cells within its vicinity, a fixed circular region. Second, there is an inductor
(or activator) signal that causes undifferentiated cells that do not receive the inhibition
signal to produce atonal. The activator signal represents the ultimate function of the
MF in the mathematical model. It is simply a rectangular wave form, with one edge
expanding with a constant velocity. It eventually moves through the entire hexagonal
lattice. Cells within this area that are not receiving inhibitory signals will differentiate.
As an initial condition, a periodic array of differentiated cells are placed as in the
posterior region. The spacing of these cell clusters is chosen such that it will allow for
the pattern to propagate. In this simplified setting, the initial configuration is defined
by differentiated cells separated evenly in a single row in the posterior-most region.
The MF then starts to propagate from posterior to anterior, and it activates cells that
do not receive the inhibition signal. This causes the pattern to propagate as shown in
Fig 1.
The actual model that leads to pattern formation is more elaborate. Now that there
is a basic understanding of how the pattern propagates, some more intricate aspects
can be introduced. In order for the model to be biologically plausible, cell clusters are
formed instead of single cells. Even in this illustrative context, as it can be inferred
schematically from Fig 1, the propagation of the pattern is very sensitive to the shape
of the inhibitory regions. A recent refinement of the model eliminates this flaw by
including a diffusible activator. In simple terms, this forces the first uninhibited cell to
receive the linearly propagating activator signal to create a circular activator region.
This activator region defines the cluster size and shape and makes the model robust,
preventing the creation of catastrophes, as explained in Ref. [10].
5
(a)
(b)
Fig 1. Visualization of the foundations of the R8 cell specification mechanism
in the Drosophila eye disc. Simplified illustration of pattern formation mechanism in
the Drosophila eye disc as a result of the competition between short-range inhibitor and
long-range activator morphogens. Posterior region of the eye disc with an initial row of
differentiated precursor cells, denoted by red dots. The morphogenetic furrow (MF),
modeled by a plane wave front, moves to the right towards the anterior region. The gray
circles represent the boundaries of regions affected by the short-range inhibitor, where the
morphogenetic furrow (MF) will not initiate production of atonal. Therefore
differentiation can only occur in those locations which are not affected by the short-range
inhibitor, leading to the hexagonal supper-lattice of differentiated R8 cells, which is
experimentally observed.
6
Methods
Mathematical Formulation
We study models of coupled differential equations that describe morphogenetic pattern
formation, relying on the simplifying assumption that the number of cells in the eye
disc is fixed [10, 11]. The ommatidia are arranged in a hexagonal grid, treating it as an
underlying two-dimensional structure. Every ommatidium, due to its hexagonal geom-
etry, has six adjacent neighbors. The resulting coupled ordinary differential equations
are of the form:
τa
dai
dt
= Paθ(ai−aa)−λaai+Gθ(hi−h1)(1−θ(ui−u1))+Sθ(si−s1)(1−θ(ui−u1)), (1)
τu
dui
dt
dsi
τs
dt
1 − ( vτh+c1
(cid:16)
2c1
Ph
Ph(−vτh+c1
2c1
(cid:40)
hi(t) =
= Puθ(ai − au) − λuui + Du∆ui,
= Psθ(ai − aa) − λssi + Ds∆si,
(cid:17)
) exp [−vτh+c1
(y − vt)]
) exp [−vτh−c1
(y − vt)]
2Dh
2Dh
y > vt
y ≤ vt
(2)
(3)
(4)
These represent the spatio-temporal evolution of the morphogen concentrations a,
u, s, and h, responsible for pattern formation in the drosophila eye disc. Here, the
upper case indexes label each cell. The pro-neural transcription factor marking the
center of the future ommatidia is a (for atonal ), u represents all diffusible inhibitors, s
denotes diffusible activator (scabrous), and h (for hairy) describes the morphogenetic
furrow (MF) which activates production of atonal along a propagating wave front. MF
propagation is mathematically derived to have this functional form from the underlying
morphogen differential equations [10]. The constant v is the velocity of the MF, D h is
the diffusivity, P h the production rate, and τh is the reaction time scale of hairy.
The mathematical model contains multiple parameters, operators and functions
that play important roles in the dynamics of the system. First, the characteristic
reaction time scale of each morphogen is given by corresponding τ's on the left hand
side of the equations. In addition, each morphogen's differential equation contains a
term with a −λ coefficient that incorporates the spontaneous decay of the morphogens.
This term, combined with τ, determines the mean lifetime of the morphogen via the
ratio τ
λ . In addition, for the morphogens that diffuse, there are Laplacian operators
associated with diffusion, Ɗ, which are discretized at the cell level (see S1 Appendix
for exact expression). For cells on the boundary, reflective boundary conditions are
used. Note that although only the equations for u and s have this diffusion term, the
differential equations that lead to the functional form of h are also based on diffusion.
Each Laplacian operator is multiplied by a term D, the value of which (when divided
by the respective τ) sets the scale how fast the morphogen diffuses.
The theta functions are essential, as they set the morphogen concentration thresh-
olds that start, accelerate and stop production. They are a simplified version of the
7
more biologically plausible sigmoid functions. They are defined by
θ(x) = 0 x < 0.
θ(x) = 1 x > 0,
(5)
The theta function θ(hi−h1) in Eq (1) is responsible for initiating production of atonal,
a, in the target cell with a rate G
, once the concentration of h crosses the threshold h 1.
τa
This term is multiplied by (1 − θ(ui − u1)), a function that goes to zero when ui > u1.
This term introduces the functionality of u as a direct inhibitor of atonal in the model.
The last term in Eq (1) introduces s as an activator of hairy. Furthermore, the first
term in Eq (1), Paθ(ai−aa), sets a threshold for atonal at which its production becomes
refractive to inhibitory signals. Finally, the two remaining differential equations for the
morphogens u and s in Eqs (2) and (3) have similar functional forms. Here, the theta
functions specify the threshold value for atonal in the target cell that is needed for
initialization of production.
In this model parameter sets have to be tuned for the experimentally observed
ordered pattern to emerge. Furthermore, as the placement of the R8 cells can vary in
cluster size and separation, we use the biologically most plausible and stable parameter
set, determined by comparison of numeric simulations to experimental imaging data of
developing drosophila ommatidia adopted from [10, 11], shown in S1 Table. In Fig 2(a),
we show simulation results for this ideal parameter set at long times after the MF has
passed.
8
(a)
(b)
(c)
Fig 2. Pattern formation simulation results in the Drosophila eye disc, the
morphogenetic furrow moves from left to right. Cluster positions and sizes of
differentiated R8 cells are shown for increasing noise σ in the diffusivity of the inhibitor
Du. (a)-(c) are the final patterns for σ/µ = 0, σ/µ = 30%, and σ/µ = 40% respectively.
Here, µ is the mean of the corresponding normal distribution and the value that is used
to generate the perfect pattern.
Placement of initial clusters
In order for the pattern to propagate, one must place preexisting differentiated cells
as an initial condition. Without proper placement, the pattern does not propagate
periodically. In this model, a differentiated cell is defined as a cell that expresses a
concentration of a above a threshold value aa. Prior work has not specified a method
for determining the initial differentiated cell placement other than trying all possible
configurations.
To determine the spacing of the differentiated cells, an isolated differentiated cluster
was simulated. After the cell cluster reached a steady state, the shape of the resulting
inhibition region was recorded. Then, the spacing between clusters or single cells was
determined by requiring an inhibition region that resembles Fig 1. Note that the size
of a cluster is determined from the properties of the local activator in Eq 3, and in its
absence single cells are placed instead.
9
Introducing Noise
Starting from validated simulations of the noiseless case [10, 11], we are now well posi-
tioned to study the effects of parametric disorder on pattern formation. The motivation
for this is that even in quasi-identical cells, it is experimentally observed that genes
are expressed at appreciably different levels due to stochasticity.
In addition, cells
produce morphogens in bursts that randomly diffuse and bind, contributing further to
the stochasticity of the system. Here we investigate how such variations affect the out-
come of pattern formation in the Drosophila eye disc. Generalizing the model for the
clean system to include the effects of disorder, each model parameter that appears in
the differential equations Eqs 1-4 is chosen from a normal distribution centered at the
mean value that produces the ordered pattern of differentiated R8 clusters in Fig 2(a).
They are then kept constant during the time evolution. The width and center of each
Gaussian distribution is chosen to be the same for all cells. When more noise is in-
troduced, the width of the distribution is increased accordingly. More specifically, in
the numerical data discussed below the widths are tuned from 0% to 60% of the mean.
As an example, introducing noise in the production rate of u corresponds to picking
values from the probability distribution:
1√
2πσ2
exp (Pu − µ)2/2σ2,
P (Pu) =
(6)
where µ = (cid:104)Pu(cid:105). Introducing noise in the diffusivity of the morphogens is a bit more
intricate and is explained in S1 Appendix. The effect of introducing noise in the
In this paper, σ will refer to the standard
diffusivity of u is presented in Fig 2.
deviation of the normal distribution, as shown for P u in Eq 6, and therefore will define
the degree of noise. In the unlikely instance that a negative value is drawn from the
distribution, the absolute value is considered.
Cluster Refinement
Until now, in the parameter regime deemed biologically plausible, the mathematical
model in Eqs 1-4 produces patterns containing clusters. However, the actual develop-
mental process includes an extra step that reduces the clusters to single activated R8
cells, which then become the center of the future ommatidia. This process is associated
with the Notch-Delta pathway, whereby the cell that produces the most delta inhibits
its nearest neighbors [15]. In order to reproduce this refinement process, a single cell for
each cluster is kept as the R8 cell. The cell is chosen to be the most central cell in the
most populated row in each cluster, an approximation to the non-trivial pathways and
models identified in the literature [10]. This cluster refinement process is illustrated
graphically in Fig 3. This extra step, of cluster refinement, has been analytically shown
to increase robustness of the pattern [16].
10
(a)
(b)
Fig 3. Visualization of the cluster refinement step. The code takes the R8 cell
clusters, as shown in (a), as an input and determines the single R8 cell that will become
the center of the future ommatidium, as shown in (b). This example is shown for a noise
level of σ = 40%, applied to the diffusivity of u, Du.
Numerical Evaluation
Numerical forward integration on a 120×44 hexagonal lattice with temporal step size
10−2 was used to evaluate the spatio-temporal evolution of the coupled differential
equations (1)-(4). To obtain sufficient data for statistical analysis, the code was paral-
lelized and run on USC's high-performance supercomputer cluster center. Each com-
putation was assigned to multiple processors at once, running for around 10 hours on
each node.
To obtain reliable error bars, at least 50 realizations of each noise level were simu-
lated. The total eye disc size width was chosen such to eliminate edge effects, allowing
sufficient space for clusters to form on the edges. Also, a relatively long eye disk length
of 120 lattice sites was used is to allow investigation of error propagation of clusters
as a function of position from posterior (P) to anterior (A). The simulations were
terminated once the morphogenetic furrow had passed throughout all of the eye disc.
Quantitative Measures of Structural Order
While previous quantitative studies have concentrated on characterizing the disorder
of the point patterns of experiments and creating minimal underlying models, [17, 18]
in this work we focus on the robustness of the mathematical model and its structural
characteristics in the presence of stochasticity. This analysis will provide insight into
pathways with which developmental systems either cope with or succumb to stochas-
ticity beyond a threshold, ultimately leading to malformation.
11
We wish to analyze the emerging point pattern using appropriate measures of struc-
tural order, including translational, bond orientation and variance of nearest neighbor
distributions. Details of how these measures are designed and applied are discussed in
this section.
Traditional measures for determining structural order have been developed in solid
state physics [19, 20, 21, 22]. However, in the context of the spatio-temporal formation
of patterns in biological developmental models there are additional issues to address.
Foremost, the degree of structural order in the final pattern is generally not homo-
geneous in the presence of stochasticity. While, per initial conditions, the posterior
point pattern starts off close to an ordered state, it may result in a strongly disordered
anterior region at later stages of development, depending on the degree of stochasticity.
This type of disorder is correlated between rows because of the manifestly Markovian
mechanism that produces the patterns, i.e. the points on a given row are specified
based on the geometry of the points previous row [23].
This observation deems traditional solid state measures used to evaluate the struc-
ture of homogeneous crystals insufficient [20]. Furthermore, the type of correlated
disorder exhibited in the point patterns of the eye disc is unlike random thermal dis-
placements found in atoms or infrequent impurities. Instead, such correlated disorder
causes Bragg's law to become inapplicable, since it is based on the assumption that the
underlying unperturbed lattice is periodic [19]. Specifically, the model system is far
away from the two instances where a corrected Bragg's law for imperfect lattices could
still be applied. The Debye-Waller approximation [24, 25] applies only for uncorrelated
deviations from a perfectly periodic lattice, and the available corrections for correlated
deviations [26] apply only in the limit of small deviations and short-ranged Gaussian
correlations. In order to study the structure of amorphous solids and general highly
disordered materials, as in the case of developmental models, it is therefore necessary
to resort to the pair distribution function as a local measure of structural order [20].
The definition of the pair distribution function is given by
N(cid:88)
j=1
g(r) =
1
N
δ(r − rj).
(7)
In the context of developmental models, this function describes the probability of
finding an activated cell at position r, given that another activated cell is located at
position rj. The pair distribution when treated as function of the scalar distance,
as in the case of isotropic patterns [20], is denoted by g2(r) and is called the radial
distribution function.
In the following analysis, we primarily use the radial distribution function. Prac-
tically, in numerical simulations of eye disk development one can calculate the radial
pair distribution function only for small distances. Each simulation of the pattern only
contains around 40 activated cells. Moreover, the calculation of g2(r) is limited by the
open boundary conditions, because the application of periodic boundary conditions to
a single realization and the combination of multiple realizations is unphysical. There-
fore, here we focus on the radial distribution function for small radii, containing only
nearest neighbor activated cells.
Furthermore, we analyze a scalar measure of the spatial order of the patterns, the
12
translational order parameter defined as
(cid:82) ηc
0
T =
1 − g2(r)dr
ηc
,
(8)
[21] where ηc is a cutoff limited by the simulation size. This order parameter can be
interpreted as the Kolmogorov probability distance [27] between the radial distribution
function of the target pattern and the Poisson random point process.
(The radial
distribution function of a Poisson random point process is the uniform distribution,
which is equal to 1.) The translational order parameter, T, is a general order metric
used to describe systems independently of the underlying crystal structure.
In this
setting, we examine how fast the pattern deteriorates compared to the perfect pattern,
so we replace the uniform probability distribution with that of the perfect pattern. In
addition, we set the cutoff ηc equal to the furthest nearest neighbor distance.
Finally, we study the bond angle order parameter. Contrary to the translational
measure, this order parameter evaluates the spatial orientation of vectors connecting
the nearest neighbors of all points. It is defined by
N(cid:88)
Nn(cid:88)
j=1
k=1
q6 = 1
N
exp (6iθjk),
(9)
[22], taking a value of 1 for a perfect hexagonal point pattern and 0 for a completely
random pattern. Here Nn is the total number of nearest neighbors in the point pattern,
the j's sum over all lattice points, and the k's sum over all nearest neighbors of a given
reference point. Lastly, θjk is the angle of the vector connecting each point with its
nearest neighbor with respect to a fixed axis.
Probability Distribution Functions of Nearest-Neighbor Distances
and Angles
To quantitatively determine the degree of disorder, post-processing code is used to
record the activated R8 cell positions in the final patterns and generate the probability
distributions of nearest-neighbor angles and distances. For every R8 cell, all the neigh-
boring distances are calculated. Then, the nearest neighbors of a cluster are identified
as all R8 cells within 1 to 1.5 times the nearest R8 cell distance.
Next, nearest-neighbor distances and angles of each cluster are calculated and a fil-
ter is applied to correct for boundary effects for the nearest neighbor angles. Combining
this information for all random realizations, the probability distribution functions of
nearest-neighbor distances and angles are determined. Since the cells in the simula-
tion are positioned on a hexagonal lattice, the nearest-neighbor distances and angles
take discrete values. As a consequence, the probability distributions are also discrete.
They are shown in Fig 4 for the ordered case. These types of probability distribution
functions are used to quantify the order of the R8 cell point patterns by calculating
the variance and the distance between probability distributions. The same approach
can be taken when interpreting experimental data obtained from imaging.
13
(a)
(b)
Fig 4. Nearest neighbor (a) distance and (b) angle probability distributions of
R8 cell point pattern in the Drosophila eye disc in the absence of
stochasticity. Distances are naturally binned by the available sites on the underlying
triangular lattice, whereas angular positions are collected in bins of 0.05 radians.
Variance as a Measure of Order
The nearest-neighbor angles and distances can be thought of as samples of an under-
lying distribution. If this distribution has non-vanishing higher-order moments, it is
not trivial to produce a measure of disorder. However, in the simple case where the
distribution can be approximated by a Gaussian, its variance, can be used as a reliable
measure of the spread of the distribution. The variance used in this context is the
sample variance of the nearest neighbor distances and angles. It is calculated from N
values, xi, using
s2 =
1
N − 1
(x − xi)2,
(10)
N(cid:88)
i=1
where x is the sample mean. This variance can be used to quantify the noise level of the
pattern. Since Gaussian distributions have only non-zero first and second moments,
checking whether the higher moments of the histograms vanish with increasing number
of realizations can be used to verify that approximating the distribution as Gaussian
is reasonable.
Probability Distribution Distance as a Measure of Order: Fidelity
and Kolmogorov Distance
Here we introduce two new and useful measures of order which best address the needs
of this study: fidelity and Kolmogorov distance. Both of these rely on the concept of
distance between probability distributions, and are generalizations of the translational
order parameter T defined in Eq 8. To define these measures of order, we use the zero
noise case probability distribution as a reference and calculate its distance from each
of the noisy probability distributions. This choice is intuitive, since we are addressing
the question of how disordered is the pattern relative to the perfect pattern.
The fidelity of two discrete probability distributions p(xi) and q(xi) is defined as
F (p(xi), q(xi)) =
(11)
(cid:88)
(cid:112)p(xk)q(xk),
k
14
67891011120.00.10.20.30.4distanceProbability0.60.81.01.21.41.61.82.00.000.050.100.150.200.25angleProbability where the sum runs over all the bins of the discrete distributions [27]. The fidelity
between two distribution falls into the range 0 ≤ F (p(xi), q(xi)) ≤ 1, where 1 is
attained only when p(xi) = q(xi) ∀ xi. The Kolmogorov distance, in this context, is
defined as
D(p(xi), q(xi)) =
(12)
(cid:88)
k
1
2
p(xk) − q(xk),
where the sum again extends over all the bins of the discrete distributions [27]. Akin
to the fidelity, the Kolmogorov distance is bounded by zero and one. This measure
represents the maximum deviation of the two probability distribution functions given
that a collection of events occur.
Results and Interpretation
Threshold Response
To quantify disorder in the emerging activated R8 cell patterns, the nearest neighbor
distance and angle probably distribution functions were computed for various noise
levels in the diffusion coefficient Du. (Later we will discuss the effect of stochasticity
on other model parameters.) Their histograms are shown in Fig 5. The binning used
for the distance histograms reflects the discreteness of the underlying triangular lattice
on which the cell centers are placed. For the distance diagrams, all possible nearest
neighbor distances on the triangular lattice are used as bins. For the angle histograms,
since there are many more possibilities, a constant bin size of 0.05 radians was chosen
in order to appropriately resolve the distribution.
15
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Fig 5. Nearest neighbor distance (left panels) and angle (right panels)
probability distributions generated from the point-pattern in the Drosophila
eye disc R8 cell specification for increasing noise levels. Stochasticity was
introduced in the differential equations by drawing the value of the diffusion coefficient
Du from a normal distribution with mean µ and standard deviation σ. The nearest
neighbor distance histograms in (a) to (d) and nearest neighbor angle histograms in (e)
to (h) correspond Gaussian stochasticity with standard deviations
σ/µ = 0%, 20%, 40%, 60%. While the first two histograms are almost identical, as noise in
Du increases, the peaks get smeared out and the histograms appear wider.
The angle histograms, unlike the distance histograms, are skewed towards larger
angles. This asymmetry makes the analysis of the distributions more complicated, as in
this case the variance, paints an incomplete picture. As discussed below, the probability
distance order measure eliminates this issue. The angle histograms are skewed as a
direct consequence of the cluster refinement process. As seen in the plots of atonal
16
0.00.10.20.30.40.00.10.20.30.40.00.10.20.30.4Probability678910110.00.10.20.30.4distance0.0.10.20.0.10.20.0.10.2Probability0.60.81.01.21.41.61.82.00.0.10.2angle patterns (Fig. 3), the onset of disorder is signaled by elongated clusters of activated
R8 cells. Since these elongated clusters are brought down to a single cell in the cluster
refinement step, and they cast large inhibition radii, the number of nearest neighbors
found for the elongated clusters is less than the six found in the perfect pattern. A
lesser number of neighbors trivially leads to higher angles in this case. Both histograms,
distances and angles, start out sparse for the case without any stochasticity in Du, with
empty bins between peaks, indicative of the highly ordered repeating lattice structure
produced by the simulations in this case. As stochasticity increases, the resulting point
patterns become aperiodic, leading to denser histograms. Furthermore, the histograms
become wider with increasing stochasticity. This is captured in the variance, shown in
Fig 6.
(a)
(b)
Fig 6. Disorder in the R8 cell point pattern of the developing Drosophila eye
disc as a function of stochasticity. Here, the variance of the distributions of (a)
nearest neighbor distances and (b) nearest neighbor angles is used to quantify the
amount of disorder. As in the previous figure, stochasticity is introduced in the diffusivity
Du, and the x-axis represents the noise level in terms of the standard deviation of the
underlying Gaussian distribution from which this parameter is drawn. As the noise level
in the simulation is increased beyond a threshold, the variance grows. The slight
irregularity of the angle plot between 20% and 30% is a result of the skewness of the
nearest neighbor angle distributions. The angle variance saturates after 40%.
The variance plots in Fig 6 illustrate the generic nature of pattern formation in
stochastic developmental models. They show that for small levels of stochasticity
(here σ/µ < σc/µ = 20%) the ordered pattern remains basically unaffected. However,
beyond this threshold the variance starts increasing linearly, both in the angle and
distance histograms. This threshold response is the central finding of this study. As
discussed below, this is a generic phenomenon, largely independent of which model
parameters experience stochasticity and which measures of order are used to assess the
system response. It implies that these stochastic models capture biological resilience
against stochastic variation up to a certain point, resulting in a regime of deterministic
outcomes in spite of stochastic input.
Let us now discuss how this type of threshold behavior is also picked up by the more
general measures of order given by Eqs 11,12. The results of applying these probabil-
17
01020304050600.60.70.80.9σ/μ(%)Variance01020304050600.0500.0550.0600.0650.0700.075σ/μ(%)Variance ity distance measures is shown in Fig 7. Note that saturation observed in the nearest
angle distribution variance plot is apparent in both cases, and the responses of both
observables (distances and angles) to stochasticity exhibit a universal sigmoidal func-
tional form, implying threshold behavior. These measures reveal the same interesting
quality as the variance: the pattern order exhibits an initial resistance to weak input
stochasticity that eventually gives way to structural malformation at larger noise levels.
The sigmoidal functional form confirms that the mechanism does exhibit robustness
for σ < σc. The fact that this is observed in the Kolmogorov measure, the fidelity,
the bond orientation and variance measures supports the notion that such threshold
response to stochasticity is a generic feature (see supporting information, S1 Fig, for
bond orientation sigmoidal response).
(a)
(c)
(b)
(d)
Fig 7. Probability distance measures applied to analyze response to
stochasticity in R8 cell point pattern formation. (a)-(b) are generated using the
fidelity, F, and (c)-(d) using the Kolmogorov distance, K, for nearest neighbor distances
and angles respectively. Stochasticity was introduced to the model through drawing the
parameter Du in the differential equations Eqs 1-4 from a normal distribution with
varying standard deviation, σ, and mean value, . The simulation results were compiled,
and nearest R8 cell neighbor angle and distance distributions were obtained. Kolmogorov
and fidelity of the resulting distributions where computed with reference to the perfect
pattern case. Here, a value of 1 corresponds to a perfectly ordered pattern, and a value of
0 corresponds to a completely irregular pattern. This figure illustrates that the functional
form is independent of the measure used.
18
01020304050600.60.70.80.91.0σ/μ(%)1-F01020304050600.750.800.850.900.951.001.05σ/μ(%)1-F01020304050600.40.50.60.70.80.91.0σ/μ(%)K01020304050600.50.60.70.80.91.0σ/μ(%)K Next, we turn to the question of universality with respect to how stochasticity
is introduced. We observe that stochasticity in other model parameters also produces
threshold behavior, indicating that the observed threshold response is a generic feature
of this class of evolutionary models. Specifically, in Fig 8 we show the response to
increasing stochasticity levels in the production rate of atonal Pa, the production rate
due to the activator S, and the production rate caused by the morphogenetic furrow
G. In all these cases, there is also a regime σ < σc where the ordered pattern remains
intact.
(a)
(c)
(b)
(d)
Fig 8. Universal threshold response to stochasticity in different model
parameters. In each sub-plot, stochasticity is introduced in a single parameter of the
underlying differential equations: (a) Du, (b) Pa, (c) S, and (d) G. Introduction of
parametric noise in all of the parameters, produces a sigmoid response.
Next let us address the origin of the observed resilience of this model towards weak
noise with σ < σc, signified by the initial plateaus observed in Fig 8. The main reason
for these wide regimes, exhibiting robustness with respect to parameter stochasticity,
is the presence of Scabrous as a diffusible activator [10]. While this ingredient is not
needed for the pattern to propagate, without this element even low stochasticity lev-
els of Du cause it to deteriorate, making the model of pattern formation biologically
implausible [10]. This effect is quantitatively demonstrated in Fig 9, which compares
the response to stochasticity with and without Scabrous as a diffusible activator. In
the absence of Scabrous, with the same parameter set and with optimized initial con-
ditions, there is an immediate deterioration of the patterned order upon introduction
of infinitesimal noise.
In contrast, the presence of Scabrous changes this functional
relationship to the observed sigmoid response, thus introducing a resilience scale to
19
01020304050600.60.70.80.91.0σ/μ(%)1-F01020304050600.70.80.91.0σ/μ(%)1-F01020304050600.70.80.91.0σ/μ(%)1-F01020304050600.700.750.800.850.900.951.001.05σ/μ(%)1-F the pattern formation mechanism set by the Scabrous production rate Ps and the
activation threshold s1.
Fig 9. Robustness of Drosophila eye disc R8 cell pattern order versus
stochasticity, with and without the diffusible activator Scabrous in the model.
The analysis is applied in the same setting as Fig 7, i.e. stochasticity is introduced in the
parameter Du. In the absence of the diffusible activator s, there is a much increased
sensitivity to stochasticity for infinitesimal non-vanishing σ.
In addition to the robustness introduced by Scabrous, the threshold response of
the order of the pattern to stochasticity in the various parameters shown in Fig 8 is
a consequence of three further elements: (i) the spatial discreteness of the underlying
lattice, (ii) the threshold activated response of cells to morphogen levels, and (iii) prop-
agation of avalanches from anterior to posterior. As we now discuss, the robustness
for low stochasticity levels is a consequence of the lattice discreteness and the thresh-
old response of the cells to morphogen concentrations, whereas the sharp decay is a
consequence of propagation of avalanches.
To understand how the discreteness of the underlying lattice contributes to robust-
ness for low noise levels, consider the effect of altering the diffusivity of a morphogen.
Since the local morphogen concentration drops off at distances on the order of a few
cells, and the cells lie on a discrete triangular lattice, a small change in the diffusivity
radius will not result in any difference unless it leads to the inclusion or exclusion of
additional cells. A continuous change in the radius of inhibition will exclude or include
new cells when it increases by a factor comparable to the lattice spacing, thus creating
a threshold response. Alternatively, consider the effect of the morphogenetic furrow,
mathematically denoted by Eq 4, which propagates with a constant speed through
the lattice of cells. Two consecutive cells on along the direction of propagation of the
morphogenetic furrow will receive the activation signal with a delay lag, δt. Only if the
20
01020304050600.60.70.80.91.0σ/μ(%)1-Fwithswithouts production rates, denoted by G in Eg. 1, picked from the normal distributions differ
above a finite threshold one will result to a misplaced cell. An analytical approach to
this effect and its contribution in error is carried out in [16].
Second, the threshold response of pattern formation to parameter stochasticity is
linked to the a threshold activation of cells in response to morphogen levels. The
turn-on and turn-off of morphogens in the differential equations is governed by step
functions with threshold values, aa, au, u1, h1, s1. This has the consequence that
the effect of morphogen concentration is binary, i.e. the parameter variation must be
sufficient such that the concentration threshold is crossed for activation or inhibition
to occur.
Finally, to explain the fast decay of the pattern beyond threshold stochasticity, the
propagation of avalanches comes into play, an effect also identified by [10]. Beyond σc,
R8 clusters start to get significantly elongated and misplaced. The first such elongated
R8 cell cluster causes the next row of R8 cell clusters to be misplaced, leading to a
cascade onset of pattern disorder. This avalanche effect is quantitatively discussed
next, where we analyze the order of the pattern as a function of position.
Increasing Disorder of the Pattern from Posterior to Anterior
Here we use the fidelity measure to evaluate error propagation as a function of position
on the spatial posterior-to-anterior axis. The mechanism of pattern formation relies
delicately on carefully spaced precursor R8 cells that are placed in the posterior region
before the growth process described by the differential equations is initiated. Their
positions define the subsequent cluster spacing of activated R8 cells for the entire eye
disc. In the presence of stochasticity, this information dissipates during the propagation
of the morphogenetic furrow from posterior to anterior, since the ommatidial pattern on
a column is defined by the shape of the inhibition signals emitted by the ommatidial row
before it. It is thus expected that as the morphogenetic furrow travels from posterior
to anterior, errors in the pattern order grow in an avalanche fashion. This effect is
observed in the simulations, as shown in Fig 10, where the disorder of the pattern
clearly increases from posterior to anterior.
21
Fig 10. Decay of R8 pattern order as a function of position on the eye disc
(same parameter set as in Fig. 5). The local order measure is calculated as a function of
position, from posterior to anterior. The x-axis refers to regions of the simulated eye disc
0-40,40-80,80-120 respectively, whereas the y-axis refers to the fidelity probability
distance measure applied to nearest neighbor distance distributions. The three colored
lines correspond to various stochasticity levels in Du, from σ/µ = 15% to σ/µ = 25% in
steps of 5%. Local order generally decreases from posterior to anterior. Once the
threshold of σc/µ ≈ 20% is crossed, there is a very sharp spatial decay of the order of the
pattern, signaling the underlying avalanche effect.
Previous studies have not, so far, quantitatively analyzed the consequences of finite
system sizes on the order of the pattern. To illustrate these finite size effects on the
threshold response quantitatively, in Fig 11 we plot the order parameter obtained
from numerical simulations for different eye disc sizes, illustrating the effect of size on
the robustness of the system. Here one clearly observes that the larger eye-disc, the
sharper the threshold response to stochasticity. Since our numerical simulations were
performed on lattices much smaller than the actual eye disc, the observed threshold
effect can be extrapolated to be much more pronounced for realistic eye disc sizes.
22
0-4040-8080-1200.60.70.80.91.0Position1-Fσ/μ=15%σ/μ=20%σ/μ=25% Fig 11. Pattern order as a function of eye disc size (same parameter sets as in Fig
8). The threshold response is more pronounced as the length of the eye disc in the
direction of the morphogenetic furrow is increased. The order parameter is calculated for
cell number 40, 80 and 120 while keeping the size perpendicular to the propagation
direction of the morphogenetic furrow at 44 cells.
This variation in the degree of dependency of the pattern order measure upon
stochasticity levels across different eye disk sizes is analogous to the finite size scaling
of order parameters commonly observed in interacting physical many body systems,
such as Ising models on lattices. In analogy to phase transitions in the thermodynamic
limit of such models, i.e. for infinite system sizes, here we observe precursor sigmoidal
response to increasing stochasticity levels that becomes more pronounced for larger
system (eye disk) sizes, ultimately culminating into a true phase transition as Lx → ∞.
The structure of the underlying system of coupled dynamical equations suggests that
this transition belongs to the universality class of directed percolation [28]. While a
complete scaling analysis is beyond the scope of this study, it will be provided in a
future publication.
Conclusions
The model of coupled differential equations in Eqs 1-4, describing eye disc R8 cell dif-
ferentiation, provides robust developmental mechanisms that lead to resilience against
stochastic perturbations. This robustness can be attributed to the discrete nature of
the underlying lattice and the threshold activation of cells to morphogens. In addi-
tion, the introduction of Scabrus, as rigorously illustrated in [10], leads to increased
resilience against stochasticity, especially in stochasticity in the diffusivity of the in-
hibitor. Beyond a critical noise threshold σc, there is an acute loss of order with
increasing stochasticity due to avalanches of misplaced differentiated cells.
23
01020304050600.60.70.80.91.0σ/μ(%)1-F40v4480v44120v44 This threshold response to noise is an essential characteristic of developmental
systems. A lack of resilience would deem biological systems unable to produce complex,
nearly perfect structures in an inherently noisy background. This is especially true in
eye disc formation with large numbers of 800 ommatidia. The introduction of Scabrus
creates a redundancy that supports the threshold response. The model system studied
here exhibits biologically plausible robustness, identifying the sigmoid functional form
as the generic response to noise of developmental models.
The expression of a gene depends on a probabilistic outcome determined by the
upstream regulatory sequences that act in cis, and the binding and processivity of other
molecules that act in trans. Thus, mutational changes in cis-regulatory elements and
trans-regulatory factors can affect transcription levels and transcriptional noise [29].
These changes in transcriptional noise reveal some of the evolutionary constraints on
cis-regulatory elements [30]. The sigmoid functional response buffers the system against
perturbations. This functional form indicates the effect of Cryptic Genetic Variation
(CGV). CGV is largely neutral until it is exposed in certain genetic backgrounds.
Robustness in development allows for the accumulation of CGV without any observable
changes in phenotype. Once the genetic background changes, for example through
mutational perturbation, this previously-neutral bottled-up CGV can be released to
produce strikingly different phenotypes [31].
Our analysis further illustrates that as the R8 clusters are specified, errors propa-
gate, leading to increased irregularity from posterior to anterior. This rate is important
as it suggests that there is an interplay between how large an eye disc is with how per-
fectly ordered it can be, with larger eye discs being more likely to accumulate some
positional errors. This has interesting implications on limiting eye disc size, a hypoth-
esis that can be investigated experimentally.
The structural order measures outlined in this paper can be used to quantify order
of patterns in developmental systems containing local point patterns. The application
of the radial pair correlation function for nearest neighbors, as specifically performed
in this study, is useful for data sets that have non trivial periodic structure and are
small in size. Specifically, the measure is useful for analysis of experimental data of
the eye-disc, since the effect of curvature of the eye-disc is eliminated and restrictive
boundary effects are resolved. The point-pattern order measure used here, based on
probability distance, is independent of the functional form of the underlying nearest
neighbor distributions and boundary conditions, thus making it applicable independent
of the form of the point pattern.
24
Supporting Information
S1 Appendix. Exact expression of the Laplace operator and noise.
cells, is given by: Du∆ui = (cid:80)
The discretized form of the Laplace operator, in the triangle lattice arrangement of
u and < ij >
denotes all six nearest neighbors, rigorously defined by via a Voronoi diagram of the
triangle lattice. Only in the case where there is no noise, Dij
u = Du and the expression
simplifies.
u (ui − uj). Where Dij
u = Dji
<ij> Dij
S1 Table. Numerical parameters of complete mathematical model. This
parameter set is taken from previous studies, plugging it in the mathematical equations
produces the desired pattern in a biologically plausible regime [10].
τa = 1.2
τh = 100
τu = 0.015
Pa = 3 Dh = 600
Ph = 2.4 Du = 0.1
Pu = 10
G = 0.9
aa = 0.4
au = 1.55
v=0.75
c1 =(cid:112)(vτh)2 + 4Dh
h1 = 1
u1 = 2 · 10−6
λa = λh = λu = 1
Ps = 10
λs = 1
τs = 0.01
S = 2.5
Ds = 0.016
s1 = 0.01
S1 Fig. Sigmoid response of bond orientation order parameter. The bond
orientation order parameter also exhibits the sigmoid response to noise. This further
supports the fact that the response of the system to noise is physical. To calculate the
bond orientation order, the Voronoi diagram method was used to precisely determine
the nearest neighbors in the point pattern.
25
Acknowledgements
We thank Avishai Gavish, first author of [10], for answering multiple questions regard-
ing the mathematical model adopted in this paper. In addition, Alexander Samsonov,
Simon Restreppo and Georgios Styliaris for valuable discussions. Computation for the
work described in this paper was supported by the University of Southern California's
Center for High-Performance Computing (hpc.usc.edu).
References
[1] Chubb JR, Trcek T, Shenoy SM, Singer RH. Transcriptional pulsing of a devel-
opmental gene. Current biology. 2006;16(10):1018–1025.
[2] Waddington CH. Canalization of development and genetic assimilation of acquired
characters. Nature. 1959;183(4676):1654.
[3] Averbukh I, Gavish A, Shilo BZ, Barkai N. Dealing with noise: The challenge of
buffering biological variability. Current Opinion in Systems Biology. 2017;1:69–74.
[4] Barkai N, Shilo BZ. Robust generation and decoding of morphogen gradients.
Cold Spring Harbor perspectives in biology. 2009;1(5):a001990.
[5] Gursky VV, Surkova SY, Samsonova MG. Mechanisms of developmental robust-
ness. Biosystems. 2012;109(3):329–335.
[6] Keller EF. Developmental robustness. Annals of the New York Academy of
Sciences. 2002;981(1):189–201.
26
01020304050600.20.30.40.50.60.70.8σ(%μ)q6 [7] Kumar JP. Building an ommatidium one cell at a time. Developmental Dynamics.
2012;241(1):136–149.
[8] Haynie JL, Bryant PJ. Development of the eye-antenna imaginal disc and mor-
phogenesis of the adult head in Drosophila melanogaster. Journal of Experimental
Zoology Part A: Ecological Genetics and Physiology. 1986;237(3):293–308.
[9] Lawrence PA, Green SM. Cell lineage in the developing retina of Drosophila.
Developmental biology. 1979;71(1):142–152.
[10] Gavish A, Shwartz A, Weizman A, Schejter E, Shilo BZ, Barkai N. Periodic
patterning of the Drosophila eye is stabilized by the diffusible activator Scabrous.
Nature Communications. 2016;7:10461 EP –.
[11] Lubensky DK, Pennington MW, Shraiman BI, Baker NE. A dynamical model of
ommatidial crystal formation. Proceedings of the National Academy of Sciences.
2011;108(27):11145–11150. doi:10.1073/pnas.1015302108.
[12] Wolff T, Ready DF. Cell death in normal and rough eye mutants of Drosophila.
Development. 1991;113(3):825–839.
[13] Wolff T, Ready DF. The beginning of pattern formation in the Drosophila com-
pound eye: the morphogenetic furrow and the second mitotic wave. Development.
1991;113(3):841–850.
[14] Sun Y, Jan LY, Jan YN. Transcriptional regulation of atonal during development
of the Drosophila peripheral nervous system. Development. 1998;125(18):3731–
3740.
[15] Baker NE, Yu S, Han D. Evolution of proneural atonal expression during
distinct regulatory phases in the developing Drosophila eye. Current Biology.
1996;6(10):1290–1302.
[16] Gavish A, Barkai N. A two-step patterning process increases the robustness of
periodic patterning in the fly eye. Journal of biological physics. 2016;42(3):317–
338.
[17] Jiao Y, Lau T, Hatzikirou H, Meyer-Hermann M, Corbo JC, Torquato S. Avian
photoreceptor patterns represent a disordered hyperuniform solution to a multi-
scale packing problem. Physical Review E. 2014;89(2):022721.
[18] Kim S, Cassidy JJ, Yang B, Carthew RW, Hilgenfeldt S. Hexagonal patterning of
the insect compound eye: Facet area variation, defects, and disorder. Biophysical
journal. 2016;111(12):2735–2746.
[19] Kittel C.
Introduction to Solid State Physics. Wiley; 2004. Available from:
https://books.google.com/books?id=kym4QgAACAAJ.
[20] Egami T, Billinge SJ. Underneath the Bragg peaks: structural analysis of complex
materials. vol. 16. Newnes; 2012.
[21] Truskett TM, Torquato S, Debenedetti PG. Towards a quantification of disorder
in materials: Distinguishing equilibrium and glassy sphere packings. Physical
Review E. 2000;62(1):993.
[22] Halperin B, Nelson DR. Theory of two-dimensional melting. Physical Review
Letters. 1978;41(2):121.
27
[23] Markov A. The theory of Algorithms. Academy of Sciences of the USSR.
1954;42:3–375.
[24] Debye P.
Interferenz von rontgenstrahlen und warmebewegung. Annalen der
Physik. 1913;348(1):49–92.
[25] Waller I. Zur frage der einwirkung der warmebewegung auf die interferenz von
rontgenstrahlen. Zeitschrift fur Physik. 1923;17(1):398–408.
[26] Lindenmeyer P, Hosemann R. Application of the theory of paracrystals to
the crystal structure analysis of polyacrylonitrile. Journal of Applied Physics.
1963;34(1):42–45.
[27] Rachev ST, Klebanov L, Stoyanov SV, Fabozzi F. The methods of distances in
the theory of probability and statistics. Springer Science & Business Media; 2013.
[28] Chowdhury D, Stauffer D. Statistical Mechanics of Interacting Systems: Theory
of Phase Transitions. Principles of Equilibrium Statistical Mechanics. 2005; p.
253–255.
[29] Metzger BP, Duveau F, Yuan DC, Tryban S, Yang B, Wittkopp PJ. Contrast-
ing frequencies and effects of cis-and trans-regulatory mutations affecting gene
expression. Molecular biology and evolution. 2016;33(5):1131–1146.
[30] Metzger BP, Yuan DC, Gruber JD, Duveau F, Wittkopp PJ. Selection on noise
constrains variation in a eukaryotic promoter. Nature. 2015;521(7552):344.
[31] Gibson G, Dworkin I. Uncovering cryptic genetic variation. Nature Reviews
Genetics. 2004;5(9):681.
28
|
1310.1482 | 1 | 1310 | 2013-10-05T14:46:58 | Effective viscosity of non-gravitactic Chlamydomonas Reinhardtii microswimmer suspensions | [
"physics.bio-ph",
"cond-mat.soft",
"physics.flu-dyn"
] | Active microswimmers are known to affect the macroscopic viscosity of suspensions in a more complex manner than passive particles. For puller-like microswimmers an increase in the viscosity has been observed. It has been suggested that the persistence of the orientation of the microswimmers hinders the rotation that is normally caused by the vorticity. It was previously shown that some sorts of algaes are bottom-heavy swimmers, i.e. their centre of mass is not located in the centre of the body. In this way, the algae affects the vorticity of the flow when it is perpendicular oriented to the axis of gravity. This orientation of gravity to vorticity is given in a rheometer that is equipped with a cone-plate geometry. Here we present measurements of the viscosity both in a cone-plate and a Taylor-Couette cell. The two set-ups yielded the same increase in viscosity although the axis of gravitation in the Taylor-Couette cell is parallel to the direction of vorticity. In a complementary experiment we tested the orientation of the direction of swimming through microscopic observation of single \textit{Chlamydomonas reinhardtii} and could not identify a preferred orientation, i. e. our specific strain of \textit{Chlamydomonas reinhardtii} are not bottom-heavy swimmers. We thus conclude that bottom heaviness is not a prerequisite for the increase of viscosity and that the effect of gravity on the rheology of our strain of \textit{Chlamydomonas reinhardtii} is negligible. This finding reopens the question of whether origin of persistence in the orientation of cells is actually responsible for the increased viscosity of the suspension. | physics.bio-ph | physics | epl draft
Effective viscosity of non-gravitactic Chlamydomonas Reinhardtii microswimmer suspensions
3
1
0
2
t
c
O
5
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
2
8
4
1
.
0
1
3
1
:
v
i
X
r
a
Matthias Mussler1 , Salima Rafaï2 , Philippe Peyla2 and Christian Wagner1
1 Experimentalphysik, Saarland University, Germany
2 Laboratoire Interdisciplinaire de Physique - UJF-CNRS, UMR 5588, Grenoble, France
PACS 47.63.Gd – Swimming microorganisms
PACS 47.50.-d – Non-Newtonian fluid flows
PACS 47.57.-s – Complex fluids and colloidal systems
Abstract – Active microswimmers are known to affect the macroscopic viscosity of suspensions in
a more complex manner than passive particles. For puller-like microswimmers an increase in the
viscosity has been observed. It has been suggested that the persistence of the orientation of the
microswimmers hinders the rotation that is normally caused by the vorticity. It was previously
shown that some sorts of algaes are bottom-heavy swimmers, i.e.
their centre of mass is not
located in the centre of the body. In this way, the algae affects the vorticity of the flow when it
is perpendicular oriented to the axis of gravity. This orientation of gravity to vorticity is given
in a rheometer that is equipped with a cone-plate geometry. Here we present measurements of
the viscosity both in a cone-plate and a Taylor-Couette cell. The two set-ups yielded the same
increase in viscosity although the axis of gravitation in the Taylor-Couette cell is parallel to the
direction of vorticity. In a complementary experiment we tested the orientation of the direction
of swimming through microscopic observation of single Chlamydomonas reinhardtii and could not
identify a preferred orientation, i. e. our specific strain of Chlamydomonas reinhardtii are not
bottom-heavy swimmers. We thus conclude that bottom heaviness is not a prerequisite for the
increase of viscosity and that the effect of gravity on the rheology of our strain of Chlamydomonas
reinhardtii is negligible. This finding reopens the question of whether origin of persistence in the
orientation of cells is actually responsible for the increased viscosity of the suspension.
Introduction. – Colloidal suspensions are a classi-
cal example of complex fluids and many of the rheologi-
cal properties are now theoretically described quite accu-
rately. However, the contribution of active particles to the
effective viscosity of a suspension is more complex. Re-
cently, interest has arisen in the theoretical description
of microswimmer suspensions [1–6] but experimental data
on their rheology remain limited [5, 7, 8]. Motile micro-
organisms, either unicellular or multicellular, can be found
throughout our eco system, and there is interest in the
use of active fluids for many applications, including devel-
opment of bio fuels, such as cyanobacteria that produce
ethanol [10] and microalgae that produce hydrogen [11].
Two types of microswimmers can be distinguished:
pushers – which push the fluid behind their body - such
as Escherichia coli and pullers – which pull the fluid in
front of their body - such as Chlamydomonas reinhardtii
(CR). Here the notion of "behind" and "in front of" refer
to the motion direction of the cells. It has been recently
predicted that slender pushers decrease the effective vis-
cosity of a suspension that is submitted to shear, whereas
slender pullers have the opposite effect [1, 3, 4, 12]. Rhe-
ological characterisations of CR in a cone-plate geometry
have been previously performed by two of the authors of
the present study [8]. Despite their nearly spherical shape,
the viscosity of a suspension of living CR is significantly
increased compared with a suspension of fixed CR, which
can be accurately modelled as a suspension of passive mi-
crospheres of the same diameter. The increase in the vis-
cosity of living CR was attributed to the anisotropy in the
orientation of the cells [8]. Due to their nearly spherical
shape, this anisotropy has been suspected to result either
from an external torque due to gravity or from an effec-
tive large hydrodynamic aspect ratio of the cell. A similar
comparison was presented in Ref. [5] in which the bacteria
were immobilized by restricting oxygen.
In ref. [4] a set of coarse-grained equations that gov-
ern the hydrodynamic velocity, concentration and stress
fields in a suspension of active and energy-dissipating so-
lution were analyzed and several predictions for the rheol-
p-1
Matthias Mussler1, Salima Rafaï2, Philippe Peyla2 and Christian Wagner1
Figure 1: Left: Schematic sketch of the Taylor-Couette Sys-
tem. The outer cylinder is fixed and the stress is exerted by
the inner cylinder, which is driven by the rheometer. Right:
The orientation of the shear gradient in the Taylor-Couette
geometry. The arrows show the shear in the x-y layer for a
Taylor-Couette geometry in which the axis of vorticity is par-
allel to gravity
ogy of such systems were given. This model was developed
further in another study [12] in which a 2D model for a
suspension of microswimmers in a fluid was analytically
and numerically analysed in the dilute regime.
In Ref.
[12] the analysis has been performed both in dilute regime
(no swimmer-swimmer interactions) and numerically in a
moderate concentration regime (with swimmer-swimmer
interactions). The effect of decreasing viscosity for pusher
particles has been described, e.g., in ref. [2] and the theo-
retical results were compared with experiments with sus-
pensions of Bacillus subtilis [3]. In ref. [1] an orientation
distribution within a shear flow was used to characterise
the configuration of a suspension of microswimmers. The
bodies of the microswimmers were modelled by a force
dipole model that is carried by a slender body.
In ref. [9] it was shown that the algae Chlamydomonas
nivalis are bottom-heavy and orient in the gravitational
field.
In this article, we provide complementary experi-
ments that aim to determine the effect of gravity on the
effective viscosity of a Chlamydomonas reinhardtii suspen-
sion. Our specific strain of CR had a nearly spherical
shape and was not explicitly known as a bottom-heavy
swimmer. We measured the effective viscosity of a CR
suspension in a Taylor-Couette geometry (fig. 1) at differ-
ent concentrations. In this geometry, gravity is oriented
in the same direction as the vorticity of the flow. However,
we found the same changes in viscosity that were observed
in a cone-plate geometry (fig. 2). This result shows that
gravity does not play a role in the effective viscosity of our
Chlamydomonas reinhardtii suspensions.
Theoretical description. – The Einstein relation
gives the effective viscosity ηef f of a dilute suspension of
rigid spherical particles:
ηef f = η0(1 + αφ)
(1)
for a solvent viscosity η0, a volume fraction φ and an
Figure 2: Left: Schematic sketch of the cone-plate system. The
plate is fixed and the stress is exerted by the cone, which is
driven by the rheometer. Right: The orientation of the shear
gradient in the cone plate geometry. The arrows show the
direction of the shear in the y-z layer. The axis of vorticity
(here x-axes) is perpendicular to gravity.
intrinsic viscosity α. The latter is defined as the contribu-
tion of the solute to the viscosity of the suspension:
α = lim
φ→0
ηef f − η0
η0φ
.
(2)
For passive rigid spherical particles, the intrinsic viscos-
ity is α = 2.5 [13]. Suspensions that have a given particle
size distribution and a fixed volume fraction of the particle
phase behave quantitatively similarly to the monodisperse
suspension described above. Eq. 1 is exact in the limit of
dilute suspensions (i.e. for φ → 0) [14]. For more concen-
trated suspensions, Krieger and Dougherty's law
ηef f = η0(cid:18)1 −
φ
φm(cid:19)−αφm
(3)
has been shown to describe the rheology of rigid sphere sus-
pensions fairly accurately [15] (φm is the maximal packing
fraction).
Our measurements were performed using a rotational
rheometer for both the cone-plate and the Taylor-Couette
geometries. The effective viscosity was obtained by the
following equation:
ηef f =
τ
γ
(4)
where τ is the stress and γ the shear rate. The shear
rate can be chosen and the stress is measured using the
controlled shear rate mode of the rheometer. In the cone-
plate geometry, gravity is perpendicular to the direction
of the flow and the vorticity, whereas gravity is parallel to
the vorticity in a Taylor-Couette geometry, as shown in
fig. 1and 2.
Materials and Methods. – CR can be found world-
wide in soil or brick water and can grow in a medium
that lacks carbon and energy sources, if it is illuminated
through a day-night cycle [16]. This organism has a nearly
spherical shape with a diameter of 5 − 10µm, a large cup-
shaped chloroplast, a large pyrenoid, and an eyespot that
p-2
Effective viscosity of non-gravitactic Chlamydomonas Reinhardtii microswimmer suspensions
Re =
ρvd
η0
(5)
0.0
0.00
senses light and causes phototactic behaviour. CR has
two flagella and propels itself using a breast-stroke like
movement at a frequency of approximately 30Hz for the
strain used in this study, which results in a velocity of
approximately 50µm/s as measured in Ref [19]. Because
of its small size and weight and the low propulsion ve-
locity force compared to the viscous forces, CR is a self-
motile low-Reynolds-number swimmer (Re = 2.5 × 10−4);
hence, inertia is negligible compared to viscous forces. The
Reynoldsnumber is
where d is the cell diameter, v the velocity and η the dy-
namic viscosity of the surrounding fluid (η0 = 1.15mP as).
Wild-type strains were obtained from the IBPC Lab
(Physiologie Membranaire et Molculaire du Chloroplaste,
UMR 7141, CNRS et Université Pierre et Marie Curie
(Paris VI)). Synchronous cultures of CR were grown in a
Tris-acetate phosphate medium (TAP) using a 12/12 hr
light/dark cycle at 25◦C [16]. The fluid cultures were typ-
ically grown for three days under fluorescent light before
the cells were harvested for the experiments. These cul-
tures were concentrated to a typical volume fraction of
20-40% by centrifugation for 5 minutes at 200g and resus-
pended in fresh culture medium to achieve the required
volume fractions. The volume fractions were measured us-
ing a Neubauer counting chamber. Samples of different
volume fractions were prepared and the effective viscosity
was measured for both live and fixed cells. The fixed cells
were treated with formaldehyde (2%) to disable the flag-
ella but to preserve the protein structure. The cell motility
was checked before and after the rheological measurements.
We used a Haake Mars 2 rheometer (Thermo Scientific)
equipped with a cone-plate geometry (cone angle α = 2◦,
radius R = 30mm, maximum distance between the cone
tip and the plate h = 105µm) or a Taylor-Couette geome-
try (inner cylinder radius r1 = 10mm, d = r2 − r1 = 1mm,
height L = 72mm)) (fig. 1 and and 2). The measurements
were obtained at T = 25◦C. The viscosity of the diluted
suspensions of CR was on the order of the viscosity of
water, which means that that the measurements were ob-
tained at the resolution limits of the rheometer. Therefore,
the viscosity data were integrated over one complete rev-
olution to prevent the influence of any imbalances in the
air bearing of the rheometer.
For the microscopic measurements, the cell suspension
was placed in a sealed capillary of size 0.2 × 2 × 5 mm.
The observations were performed using a bright-field mi-
croscope with a magnification of ×1.25. The microscope
was tilted down on the side to ensure that the direction
of gravity was in the observation plane. A red filter was
used to avoid any bias due to phototaxis. IDL (Interac-
tive Data Language)-based tracking routines were used to
locate the cells in time and to measure their velocities.
Results. –
p-3
1.8
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0
/
)
-
f
f
e
(
Alive, equ. 3: 4.5(+-0.17)
Fixed, equ. 3: 2.4(+-0.15)
0.04
0.08
0.12
0.16
0.20
volume fraction
Figure 3: The normalised effective viscosity of suspensions of
different volume fractions of CR measured in a cone-plate ge-
ometry. The squares and triangles indicate the living and fixed
CR, respectively. The straight lines are fits using eq. 3 to ob-
tain the intrinsic viscosity, α.
Cone-plate geometry. We first repeated the experi-
mental results that were previously obtained by Rafaï et
al. [8]. In agreement with this preceding investigation a
shear thinning of ηef f can be observed. As discussed in
ref. [8], it reveals competition between the shear rate and
cells orientation in the flow. In order to investigate the
dependence of effective viscosity on the volume fraction of
the suspension, we choose a shear rate of γ = 5 1
s . Fig. 3
shows the relative effective viscosity (ηef f − η0) /η0 of live
and fixed cells as a function of the volume fraction φ mea-
sured in a cone-plate geometry. In both cases, the relative
viscosity increases with increasing volume fraction, but the
effect is more pronounced with living cells. More quanti-
tatively, the data were fitted with eq. 3. We maintained
α as a free parameter and chose φm = 0.63. We found
values of α = 4.5 ± 0.17 for living and α = 2.4 ± 0.15 for
fixed CR. Moreover, the shearthinning behaviour is also
retreived. These results are in agreement with previous
work [8].
Taylor-Couette geometry.
To study the influence of
gravity on the swimming behaviour of CR, measurements
were obtained using a Taylor-Couette geometry. Fig. 4
shows the relative effective viscosity (ηef f − η0) /η0 of live
and fixed cells as a function of the volume fraction φ mea-
sured at a shear rate of γ = 5 1
s . The data obtained with
both the fixed and living CR were fitted with eq. 3. As in
the previous analysis, we maintained α as a free parame-
ter and chose φm = 0.63. We found α = 4.8 and α = 2.5
for live and fixed CR suspensions, respectively. Thus, for
both geometries, the intrinsic viscosity of the living and
dead CR were identical, within the experimental error.
Microscopic measurements.
Some type of algae mi-
croswimmers are known to be bottom-heavy, i.e.
their
mass is not homogeneously distributed. This results in a
biased swimming of the cell because the flagella are likely
Matthias Mussler1, Salima Rafaï2, Philippe Peyla2 and Christian Wagner1
1.8
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0.00
0
/
)
-
f
f
e
(
Alive, equ. 3: 4.8(+-0.5)
Fixed, equ. 3: =2.5(+-0.2)
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
y
t
i
l
i
0.04
0.08
0.12
0.16
0.20
volume fraction
b
a
b
o
r
p
0.015
0.012
0.009
0.006
0.003
0.000
direction of
gravity
direction of
the horizontal
distributions
difference
-0.003
-100 -80 -60 -40 -20 0
20 40 60 80 100
velocity (µm/s)
Figure 4: The normalized effective viscosity of suspensions
of different volume fractions of CR measured in the Taylor-
Couette geometry. The symbols are the same as those used in
fig. 3.
Figure 5: The microscopic measured distribution of the veloc-
ities of the swimmers along the axis of gravity and along the
horizontal axis.
to point upward. However, we tested whether this was
the case for the strain used in this study (as well as in Ref.
[8]).
We measured the distribution of velocities of hundreds
of cells over twenty minutes both in the direction of gravity
and parallel to the horizontal (designated x-direction) (see
fig. 5). We observed no significant difference between the
two distributions and we found only a slight asymmetry,
but it was negligible when compared to real taxis. Thus,
the swimming of the cells does not exhibit any bias toward
the upward direction, which suggests that this strain is not
bottom-heavy. For comparison, when a light was placed on
the right side (x>0) of the capillary (fig. 6), we observed
a pronounced biased swimming behaviour, which resulted
in an asymmetric distribution of the velocities toward the
x>0 region.
Discussion. – The difference between a passive parti-
cle that rotates with Ωy = γ/2 and an active particle that
maintains its orientation is shown in fig. 7. It is obvious
that the persistence of orientation leads to an additional
shear stress. Previous studies [17,18] theoretically showed
that if the cells are sensitive to gravity (named bottom-
heavy squirmers due to the off-centred centre of mass), the
orientation of the vorticity to the gravity field would in-
fluence the rheological measurements. If the gravitational
field is perpendicular to the vorticity, the effective viscos-
ity is increased; however, if the gravity and vorticity fields
are parallel, the effective viscosity should not be affected.
This is not in agreement with our observations since the
strain of cells used in our work is not sensitive to gravity.
Complementary experiments have shown that the random
walk of the strain of CR used in this study is indeed not
determined by gravity. The velocity distribution of the mi-
croswimmers is similar in the flow and gradient directions,
as measured by particle tracking microscopy. Therefore,
our results cannot be explained with this model. These
0.025
0.020
0.015
0.010
0.005
0.000
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
p
direction of gravity
direction of the horizontal
100
150
-100
-50
0
50
velocity (µm/s)
Figure 6: The microscopic measured distribution of the veloc-
ities of the swimmers along the axis of gravity and along the
horizontal axis when the sample was illuminated from the right
side.
Figure 7: The left sphere exhibits the behaviour of a passive
particle in a shear flow. The particles roll around an axis
through its centre with the half shear rate. The right sphere is
an example of the behaviour of active particles that maintain
a constant orientation in the shear flow.
p-4
Effective viscosity of non-gravitactic Chlamydomonas Reinhardtii microswimmer suspensions
data confirm that the effect of motility on the rheology of
microswimmer suspensions is not a simple consequence of
the orientation of the gravity field.
The breast-stroke-like propulsion of CR leads to the de-
velopment of several models to describe the behaviour of
such microswimmers, especially in shear flow.
Since CR have a slight nonspherical shape, it is first
instructive to evaluate the contribution of ellipticity of ac-
tive swimmers to the intrinsic viscosity α = 2.5 + δ. The
analytical correction δ is due to the presence of the non-
spherical swimmers (δ < 0 for pushers, and δ > 0 for
pullers); it has been derived analytically by Haines et al.
[3] for a shear flow and is such that:
directions confirmed these findings. These results led to
the conclusion that the pure geometrical aspects and the
inner structure of CR [its centre of mass eccentricity (fig.
7)] i.e., gravitational effects, are not sufficient to describe
this viscosity increase from a macroscopic point of view.
The change in viscosity may or may not be explained by
preferential orientation of swimmers in shear flow. But
it is clear that if there is a preferential orientation in our
case it is not due to bottom-heaviness nor due to elonga-
tion. We think that the motility generated by the breast
strokes of the flagella have to be taken into account to de-
velop a realistic model for the description of CR in shear
flow; this model is currently being developed by our teams.
δ =
27vbDN (5λ2 + 10λ + 2)
10a2(36D2 + γ 2)(1 + λ)4 ǫ + O(ǫ2)
(6)
where v ≈ 50µms−1 is the velocity of the CR, a ≈ 7 µm
and b ≈ 7.7 µm the ellipsoidal
lengths, yielding ǫ =
b/a − 1 ≈ 0.1. (1 + λ)b is the typical distance from the
body centre where the volume forces are applied and we
choose λ ≈ 0.75. The force that is exerted by the flag-
ella is f = 6πηbN v, N being a scalar function which
rescales the drag coefficient and following ref. [3] we use
N = 0.5. The shear rate is γ = 5s−1. Finally the rota-
tional diffusion for algae has been estimated by Pedley and
Kessler [17] D ≈ 0.067 s−1 (for comparison the thermal ro-
tational diffusion coefficient for a 10µm diameter sphere
is 0.0013 s−1). With these parameters, we find δ ≈ 0.0037
which is much below our experimental value δ ≈ 2.0. We
can thus conclude that the geometry cannot explain the
observed changes in effective viscosity.
Another approach is the force dipole model, which pro-
poses that the force vector, F, points along the direction of
movement and that each flagellum exerts a force –F/2. If
a fixed CR is introduced into shear flow, a tumbling move-
ment will be observed; however, a living CR is known to
withstand shear flow below 10Hz [8] and to align its force
vector in a constant angle to the shear stress. In a previous
study [7], the flow field of a CR was measured directly in a
fluid at rest and compared to the computed flow field of a
three- or two-Stokeslet model. These results indicate that
the simple puller-type description for CR is only valid at
distances larger than seven times the diameter of the CR.
In addition, shear flow might affect the flow field around
a swimmer; thus, it is also important to study the micro-
scopic behaviour under shear.
Conclusion. – We have studied the macroscopic ac-
cessible effective viscosity of an active suspension com-
posed of living and fixed Chlamydomonas reinhardtii. It
has been shown that for both cone-plate and Taylor-
Couette geometries, the effective viscosity of suspensions
of CR increases identically as a function of the volume
fraction. The dependency of the viscosity on the concen-
tration can be described by the Krieger-Dougherty law for
both geometries with identical intrinsic viscosities. Mi-
croscopic measurements of the distribution of swimming
∗ ∗ ∗
This work was funded by the Graduate School 1276
(DFG). SR and PP are funded by ANR MICMACSWIM.
References
[1] Saintillan D., Physical Review E , 81 (2010) 056307.
[2] Giomi L., Liverpool T. B. and Marchetti M. C.,
Physical Review E , 81 (2010) 051908.
[3] Haines B. M., Sokolov A., Aranson I. S., Berlyand
L. and Karpeev D. A., Physical Review E , 80 (2009)
041922.
[4] Hatwalne Y., Ramaswamy S., Rao M. and Simha
R. A., Physical Review Letters , 92 (2004) 118101.
[5] Sokolov, Andrey and Aranson, Igor S., Physical Re-
view Letters , 103 (2009) 148101.
[6] Ramaswamy, S., Annual Review of Condensed Matter
Physics , 1 (2012) 323-345.
[7] Drescher K., Goldstein R. E., Michel N., Polin
M. and Tuval I., Physical Review Letters , 105 (2010)
168101.
[8] Rafaï S., Jibuti L. and Peyla P., Physical Review Let-
ters , 104 (2010) 098102.
[9] Durham M. D., Kessler J.O. and Stocker R., Science,
323 (2009) 1066.
[10] Deng M. D. and Coleman, Applied and Environmental
Microbiology , 65 (1999) 523.
[11] Ghirardi M. L., Zhang L., Lee J. W., Flynn T.,
Seibert M., Greenbaum E. and Melis A., Trends in
Biotechnology , 18 (2000) 506.
[12] Gyrya V., Lipnikov K., Aranson I. and Berlyand L.,
Journal of Mathematical Biology , 62 (2011) 707.
[13] W. Pabst, Silikáty , 2004 (2004) 6.
[14] Dörr A., Sadiki A. and Mehdizadeh A., ArXiv e-
prints, 2012 (.)
[15] Krieger I. M., Trans. Soc. Rheol , 3 (1959) 137.
[16] Gorman D. S. and Levine R. P., Proceedings of the
National Academy of Sciences , 54 (1965) 1665.
[17] Pedley T. J. and Kessler J. O., Annual Review of Fluid
Mechanics , 24 (1992) 313.
[18] Ishikawa, Takuji and Pedley, T. J., Journal of Fluid
Mechanics , 588 (2007) 399.
[19] Garcia, M. and Berti, S. and Peyla, P. and Rafaï,
S., Physical Review E , 83 (2011) 3.
p-5
|
1305.5428 | 2 | 1305 | 2013-05-27T18:52:30 | Influence of swimming strategy on microorganism separation by asymmetric obstacles | [
"physics.bio-ph",
"cond-mat.soft"
] | It has been shown that a nanoliter chamber separated by a wall of asymmetric obstacles can lead to an inhomogeneous distribution of self-propelled microorganisms. Although it is well established that this rectification effect arises from the interaction between the swimmers and the non-centrosymmetric pillars, here we demonstrate numerically that its efficiency is strongly dependent on the detailed dynamics of the individual microorganism. In particular, for the case of run-and-tumble dynamics, the distribution of run lengths, the rotational diffusion and the partial preservation of run orientation memory through a tumble are important factors when computing the rectification efficiency. In addition, we optimize the geometrical dimensions of the asymmetric pillars in order to maximize the swimmer concentration and we illustrate how it can be used for sorting by swimming strategy in a long array of parallel obstacles. | physics.bio-ph | physics |
Influence of swimming strategy on microorganism separation by asymmetric obstacles
I. Berdakin1, Y. Jeyaram2, V. V. Moshchalkov2, L. Venken3, S. Dierckx3,
S. J. Vanderleyden3, A. V. Silhanek4, C. A. Condat1, and V. I. Marconi1
1Facultad de Matem´atica, Astronom´ıa y F´ısica,
Universidad Nacional de C´ordoba and IFEG-CONICET,
X5000HUA C´ordoba, Argentina.
2INPAC -- Institute for Nanoscale Physics and Chemistry,
Nanoscale Superconductivity and Magnetism Group,
K.U.Leuven, Celestijnenlaan 200D,
B -- 3001 Leuven, Belgium.
3Faculty of Bioscience Engineering - Department of Microbial and Molecular Systems,
Katholieke Universiteit Leuven, Belgium.
4D´epartement de Physique, Universit´e de Li`ege,
B-4000 Sart Tilman, Belgium.
It has been shown that a nanoliter chamber separated by a wall of asymmetric obstacles can
lead to an inhomogeneous distribution of self-propelled microorganisms. Although it is well es-
tablished that this rectification effect arises from the interaction between the swimmers and the
non-centrosymmetric pillars, here we demonstrate numerically that its efficiency is strongly de-
pendent on the detailed dynamics of the individual microorganism. In particular, for the case of
run-and-tumble dynamics, the distribution of run lengths, the rotational diffusion and the partial
preservation of run orientation memory through a tumble are important factors when computing
the rectification efficiency. In addition, we optimize the geometrical dimensions of the asymmetric
pillars in order to maximize the swimmer concentration and we illustrate how it can be used for
sorting by swimming strategy in a long array of parallel obstacles.
PACS numbers: 87.17.Jj, 87.18.Hf, 05.40.-a
I.
INTRODUCTION
In recent years there has been an increasing inter-
est in the dynamics of microscopic agents confined to
nanoliter-volume chambers [1, 2]. These agents range
from cancer [3 -- 5], fibroblast [6] and stem cells [7] to self-
propelled bacteria [8, 9], human spermatozoa [10], and
microbots [11]. Each self-propelled agent has a specific
propulsion mechanism and interacts in its own character-
istic way with the confining walls [12 -- 14]. A strong moti-
vation for the study of these systems is the possibility to
sort out, concentrate and manipulate the movement and
distribution of the swimmers, or even to harvest their en-
ergy by using suitably designed micro-architectures [2, 4 --
10, 14 -- 16].
In a pioneer work, Galajda et al. [8] experimentally
demonstrated the possibility of achieving inhomogeneous
bacterial concentrations by inducing an asymmetric av-
erage bacterial displacement with a micro-fabricated wall
of funnel-shaped openings. This mechanically driven seg-
regation seems to be threatened at sufficiently high bac-
terial concentrations when they can collectively migrate
against the confining barriers by creating a chemoat-
tractant gradient [17]. Very recently a counter-intuitive
symmetry breaking of the strong bacterial concentra-
tion was observed under controlled flow as well, using a
micro-fluidic channel with a symmetric single funnel [18].
Clearly, the ability to design these structures is highly
dependent on our understanding of the key biophysical
concepts: the motility mechanisms, the interactions be-
tween the agents and the walls, and the hydrodynamic
agent-agent interactions.
Considerable attention has been devoted to motile bac-
teria propelled by rotary motors and exhibiting run-and-
tumble dynamics. During the run mode, the flagella
rotate counterclockwise and the microorganism moves
in a forward, relatively straight direction, whereas dur-
ing the tumble mode, one or more flagella rotate clock-
wise and the bacterium is reoriented towards a new di-
rection [19, 20].
In the case of the paradigmatic bac-
terium Escherichia coli, the dynamics of its wild-type
and two mutants has been previously studied by Berg
and Brown [21] using a three-dimensional tracking micro-
scope. These authors found that (i) the run length is not
a constant, but follows an exponential distribution, (ii)
the runs do not consist of strictly straight displacements
but the cell meanders due to rotational diffusion, and (iii)
the distribution of changes of direction from the end of
one run to the beginning of the next has a maximum at
a direction making an acute angle with the trajectory of
the precedent run.
The first numerical attempt to describe the observed
rectification of bacterial displacement [8] taking into ac-
count some of the above mentioned ingredients was car-
ried out by Wan and collaborators [22]. These authors
considered point-like swimming bacteria following a run-
and-tumble dynamics with a constant motor force mag-
nitude and thermal fluctuations without taking into ac-
count hydrodynamic effects. This model is able to repro-
duce the most important experimental findings, i.e., the
accumulation of swimmers next to the boundaries and
the ratchet-like effects of the asymmetric wall of funnels,
although it ignores some important details of the swim-
mer dynamics, whose consideration, as we will show in
this work, leads to a more accomplished and quantitative
description of the observed phenomena.
In several previous simulations bacteria were assumed
to have fixed run lengths, to emerge from each tumble in a
completely random direction, and to move all with equal
speed. Although these hypotheses capture the essential
micro-swimmers mechanisms, they are rather artificial or
inaccurate when addressing particular species [21, 23 -- 25].
The question now arises as to whether the particularities
of the dynamics of each species have an impact on the
efficiency of a mechanical sorter of the type mentioned
above.
In this paper we address this question by investigat-
ing how different swimming strategies influence both the
rectification and the separation of micro-swimmers. In
particular, we find that the residual memory of run ori-
entation that remains after a tumble, leading to a persis-
tent random walk trajectory, substantially increases the
rectification efficiency, thanks to an enhancement of the
effective diffusion coefficient and a longer dwell time near
the walls. Taking this effect into account drastically im-
proves the quantitative agreement with previous experi-
ments, in time and magnitude of the rectification. This
refined model allows us to determine the optimum geo-
metrical parameters that maximize rectification, which is
essential for optimizing future applications in sorting, fil-
tering, or harvesting mixed populations of self-propelled
agents.
II. MODEL
We model the dynamics of Ns self-propelled swimmers
at low Reynolds numbers, confined to a micro-patterned
two-dimensional box of size Lx × Ly, as schematically
shown in Fig. 1. The swimmer density is ρs = Ns/(Lx ×
Ly) and each individual swimmer, si, is represented by a
soft disk of radius rs, obeying the following overdamped
equation of motion:
γ
dri
dt
= Fm
i + Fsw
i + Fs
i ,
(1)
where γ is the damping constant associated to a given
aqueous medium with viscosity η, ri is the position of
the swimmer center of mass, Fm
i accounts for the inter-
nal motor of the swimmer, i.e. the driving force. Fsw
is
the interaction force between a swimmer and the solid
boundaries, and Fs
i represents the steric force due to
the swimmer-swimmer interaction. In the following the
forces will be given in units of γ. We simulate diluted
systems with Ns between 1000 and 3000 swimmers of
i
2
the same type inside a rectangular box divided by a cen-
tral line of asymmetric obstacles, Fig. 1(a), Fs
i being the
smallest contribution to Eq. (1). The line of obstacles
consists of Nf funnels of identical geometry, lg being the
gap width between obstacles (opening of the funnel), lf
the funnel wall length, and θ the angle of the aperture
(see Fig. 1(b)).
(a)
(b)
Ly
Lx
lg
lf
θ
FIG. 1. (a) Two dimensional reservoir with a central column
of 10 asymmetric funnel-shaped obstacles separating chamber
1 (left) from chamber 2 (right). Box size: Lx ×Ly. (b) Details
of the ratchet geometry:
lf , θ, and lg, the relevant control
parameters for a single funnel.
i . Run lengths, lrun,i = τ ∆tFm
We assume for the run-and-tumble dynamics that each
microorganism swims in an almost straight trajectory
during a run, at a constant speed vi for a period of
time τ (see scheme in Fig.S1(a) in the Supporting Mate-
rial [27] ) and that the population of Ns swimmers has a
normal velocity distribution with mean ¯v. The swimmer
speed vi is directly proportional to the driving force
magnitude Fm
i , with
∆t the numerical integration step, are not constant in
each swimmer trajectory, but have an exponential distri-
bution and their direction is changed only after a tum-
ble or due to the noise affecting its rotational dynamics.
The stochasticity in the motion is introduced here. We
model the velocity during each run as vi = vivi, with
vi = cos(Φi)i + sin(Φi)j. The change in direction during
a run, δΦ, in a simulated time interval ∆t, is taken to be
proportional to a gaussian-distributed random variable
ν: δΦ = ν√2Dr∆t [26]. Dr should be the coefficient
of rotational diffusion obtained experimentally for each
swimmer species. Then we take into account rotational
diffusion through the motor force term. In particular to
simulate wild-type E. coli we use Dr= 0.1 rad2/sec ap-
proximating the measured Dr ∼ 0.062 rad2/sec under
tracking conditions of T = 305◦K and η = 0.027 g/cm
sec [26].
The run-and-tumble dynamics was experimentally an-
alyzed by three-dimensional tracking microscopy by Berg
and Brown, who found that the mean change in direc-
Mean
direction
change,
¯φ [◦]
s1 33 ± 15
s2 0 ± 36
s3 68 ± 36
s4 74 ± 33
s5 180 ± 36
s6 68 ± 36
s7 90(srw)
Mean run
duration,
τ [s]
Mean
speed,
¯v[µm/s]
D
[µm2/s]
Rectification
time [min]
6.3
100
0.86
0.42
0.86
0.01
0.86
20 ± 4.9
1788
14.2 ± 3.4 1293
14.2 ± 3.4 128
14.4 ± 3.9 48
14.2 ± 3.4 51
14.2 ± 3.4 1.6
14.2 ± 3.4 87
7.1
7.4
21.4
32.7
33.5
95.9
22.4
TABLE I. Motility parameters associated to different swim-
mers. s1, s3 and s4 are the mutants of E. coli used in [21];
s2, s5, s6 and s7 are artificial swimmers (srw = symmetric
random walk). The last two columns show calculated values.
The rectification time was determined as the time needed to
get 99 % of the final rectification r(∞) = α/β (see Eq. 3 and
Fig. 6)
tion for wild-type E. coli is ¯φ = 68◦ after a tumble,
with a standard deviation of 36◦ [21].
In other words,
E. coli keeps some memory of the original direction of
motion. Lovely and Dahlquist [29] studied this effect
for cells swimming at a constant speed along a trajec-
tory comprising a sequence of exponentially distributed
straight runs of mean duration τ . They showed that this
memory increases the effective diffusion coefficient as:
D =
v2τ
d(1 − hcos(φ)i)
(2)
where φ is the angle between the incoming and the out-
going directions at a tumble (See Fig.S1 at [27]), h...i
indicates the mean value, d is the system dimension, and
v is the cell speed. If the change in direction from run
to run is random, D = v2τ /d [26]. The mean square dis-
placement of the cell is given by the standard expression
as a function of time: h∆r2(t)i = 2dDt.
In our simulations we take memory effects into ac-
count, and the direction of motion after a tumble, φ,
is chosen from a normal probability distribution with the
parameters associated to a specific species. In Fig.S1(c)
(see [27]) we show a sample trajectory obtained with a
mean change in direction ¯φ = 60◦ with standard devi-
ation 25◦. The calculations were performed using both
mean direction changes during runs and changes in direc-
tion after tumbles distributed in accordance with obser-
vations [21]. Table I lists the different swimmers consid-
ered in the present work, with s1, s3 and s4 corresponding
to mutants of E. coli used in [21], CheC497 (long runs),
wild-type AW405 (intermediate runs) and Unc602 (short
runs) respectively, whereas s2, s5, s6 and s7 represent
idealized swimmers.
3
It is interesting to compare a real single trajectory un-
der no-confinement with paths obtained from different
models. Fig.S1(b), at [27], shows a trajectory obtained
using the model of Ref. [22] with a constant driving force
applied in a randomly chosen direction after each tumble
and a small thermal force term. Notice that the runs con-
sist of apparently straight displacements, which in that
model could be made more curvilinear by adding an un-
realistically high temperature. Fig.S1(b) is at odds with
the experimental trajectory shown in Fig.S1(d), obtained
in Ref. [21], where clearly each run consists of a series of
small reorientations around a mean value. This observed
effect due to rotational diffusion has been incorporated
in our model, which includes the exponential distribu-
tion of runs, rotational diffusion and persistence in the
motor force term (see Fig.S1(c)). Even though our path
is in two dimensions, it still resembles the particular track
called "a flamenco dancer"; simulated by E.M. Purcell in
1975 and shown in Ref. [26].
It is worth noting that up to this point, we need only
the first term of the r.h.s. of Eq. 1, to model a single free
swimmer trajectory. Let us now describe in detail how
the interaction forces and confinement walls are included
in our model.
i
Interaction swimmer-wall. The second term in the
r.h.s. of Eq. 1, Fsw
, accounts for the interactions with
confining walls and obstacles that swimmers could find
in their paths. Therefore, it is precisely through this
term that the detailed asymmetric geometry shown in
Fig. 1(b) manifests itself. Wall widths (funnels and box)
are taken to be similar, w ∼ 2 − 5 µm, and the disk ra-
dius representing the swimmer, si, is in all cases rs = 0.5
µm. Each k wall is interacting repulsively with each si
i = F sw(1 − rik/a)0.1 nk, for rik < a and
as follows: Fsw
Fsw
i = 0 otherwise. Here a = w/2 + rs, rik is the dis-
tance between the center of mass of si and the middle of
the k-th wall, and nk is a versor normal to the wall. This
force term prevents the crossing of the swimmers over the
walls. After a swimmer, si, hits the k-th wall at an an-
gle Θi formed by the incidence direction vi and the wall,
its trajectory becomes parallel to the wall. The velocity
component in this direction is assumed to be preserved
and its value is given by vw = vi cos(Θi)tk where tk is a
vector tangent to the k-th wall. This is unchanged until
either the next tumble or rotational diffusion deflects the
direction of motion. We always use F sw/F m ≫ 1, F sw
being one order of magnitude larger than F m.
In this
way we phenomenologically introduce the swimming and
accumulation along walls. Note that our model adds the
possibility of tumble-mediated wall detachment, which is
strongly dependent on swimming strategy.
of Eq. 1 is the repulsive steric force, Fs
Interaction among swimmers. The last term in the
r.h.s.
i =
F s PNs
j6=i (cid:0)1 − rij/2rs(cid:1)rij , which represents the inter-
action between bacterium i and all other bacteria j,
each separated by a distance rij in the direction rij =
(ri − rj)/rij. This is a short range force approximated
by a linearly decreasing force, having a maximum Fs
i
at rij = 0 and a zero at rij = rs;
it is nonzero for
rij < rs and zero otherwise. We use F s/F m ≫ 1 and
F s of the same order of magnitude as F sw. Very recent
measurements reported in Ref. [30] for cell-cell and cell-
wall interactions using E. coli show that the thermal and
intrinsic stochasticity wash out the effects of long-range
fluid dynamics, implying that physical interactions be-
tween bacteria are basically determined by steric colli-
sions and near-field lubrication forces. These results jus-
tify neglecting long-range hydrodynamic forces. In addi-
tion, we are interested in studying very diluted systems
and rectified motion under confinement, situations where
we expect only a small contribution of these forces.
〉
)
t
(
v
.
)
0
(
v
〈
1
0.8
0.6
0.4
0.2
0
0
s1
s2
s3
s4
s5
s7
5
10
15
t (s)
20
25
30
FIG. 2. (color online) Velocity-velocity correlation function
for different self-propelled swimmers. Motility parameters for
each si were taken from Table I. The dot-dashed, dotted, and
dashed lines correspond, respectively, to swimmers s1, s2, and
s7 when rotational diffusion is not taken into account.
III. RESULTS
A. Swimmers in unbounded environments
First, we use our model to describe the motility prop-
erties of run-and-tumble swimmers in the absence of con-
fining surfaces. For free swimmers and very diluted sys-
tems, our results will be dominated by the first force term
Fm in Eq. 1. There are two independent processes that
degrade the orientational correlation: tumbling and ro-
tational diffusion. Their combined effect is reflected in
the velocity correlation function, Cvv(t) = hv(t).v(0)i,
shown in Fig. 2 for the swimmers listed in Table I. Note
that most of the group of seven swimmers si, used for
comparing different swimming strategies, have the same
mean speed and that the parameters used for compari-
son are parameters not often included in previous numer-
ical simulations. The comparative parameters used here
characterize each species: the mean change in direction
4
after a tumble, ranging from ¯φ = 0◦(high persistence) to
¯φ = 180◦ (run and reverse), and the exponentially dis-
tributed run lengths, which are proportional to the mean
duration of the runs, ranging from very short (τ = 0.01
s) to long (τ = 100 s) runs.
Fig. 2 shows clearly that longer runs and a stronger
persistence (corresponding to a smaller average direc-
tional change ¯φ at each tumble) yield longer correlation
times. The correlation for the run-and-reverse case, s5,
decays faster than that for s3, which has the same run
duration but less built-in persistence. Note also that s7
decays faster than two real examples of E. coli mutants,
s1 and s3, because the direction of s7 is completely ran-
domized at each tumble. A simple exponential fit (not
shown) to the s7 curve yields a decorrelation time of
0.8 sec, barely shorter than the mean run duration of
0.86 sec, indicating the relatively weak influence of ro-
tational diffusion. Further evidence for the lack of rel-
evance of rotational diffusion for short-run swimmers is
the very small difference between the results for s7 with
(solid line) and without (dashed line) rotational diffu-
sion. The curves for s1 and s2, on the other hand, ex-
hibit long correlations with decorrelation times equal to
τvv,s1 = 7.9 s and τvv,s2 = 8.9 s respectively. Once rota-
tional diffusion is removed, both s1 (dot-dashed line) and
s2 (dotted line) decay much more slowly, in τvv,s1 = 43.2
s and τvv,s2 = 138.7 s respectively, and the difference
between their correlations is strongly enhanced, showing
that, for long-run swimmers, rotational diffusion is the
fastest path to memory loss. From the precedent dis-
cussion it is clear that self-propelled swimmers following
run-and-tumble dynamics can be classified into two cat-
egories: in the first group, typified by s1 and s2, the loss
of correlation is mostly due to rotational diffusion; in the
second group, s3 to s7, decorrelation is controlled by the
specific swimming strategy.
In Fig. 3 we present the mean square displacement
(MSD) as a function of time for all si considered, tak-
ing averages over a population of Ns = 1000 swimmers.
In all cases, except for the short-run swimmer s6 (orange
pentagons), for which the change occurs at very short
times, the slope of the time evolution of the MSD is seen
to change from 2 (dashed line), during ballistic motion,
to 1 in the diffusive regime (dotted line). This transition
occurs later for those swimmers with longer runs and
smaller average directional changes. For swimmers s3 to
s4, the diffusion coefficient is approximately given by its
value in the absence of rotational diffusion (see Eq. 2).
Note, for instance, the strong reduction in the transla-
tional diffusion coefficient resulting from changing the
mean angle from 68◦ , s3, to 180◦ , s5. The results for
s7 swimmers, for which ¯φ = 90◦ and whose τ is equal to
that of s3, are intermediate between those for s3 and s5,
in reasonable agreement with Eq. 2. The diffusion coeffi-
cients for s1 and s2 are controlled by rotational diffusion
instead. As a consequence, a hundredfold increment in
the run duration from s3 to s2 generates only a tenfold in-
crement in D. The simulation values of D3=128.3 µm2/s
s1
s2
s3
s4
s5
s6
s7
107
105
103
101
)
2
m
µ
(
D
S
M
10-1
100
slope 2
e 1
p
o
l
s
10
t (s)
102
103
FIG. 3. (color online) Mean square displacement for the swim-
mers of Table I. In all cases except for the Brownian-like swim-
mer, s6, the time evolution of the MSD changes from a slope
2 during ballistic motion (dashed line) to slope 1 during the
diffusive regime (dotted line). Diffusion increases for larger
τ ′s and decreases for larger ¯φ′s. The self-diffusion coefficients
obtained for the different swimmers are given in Table I. D3
is in good agreement with experiments [19].
and D4=47.94 µm2/s are also in good agreement with the
values obtained by applying Eq. 2 to the experimental re-
sults reported in Ref. [21]. These results show clearly the
importance of taking into account the detailed motility
properties (specific swimming strategies) for making ac-
curate predictions, for instance, of swimmer separation in
mixed species systems. Both precise numerical calcula-
tions and accurate experimental data characterizing the
swimmer dynamics are therefore of paramount impor-
tance. Better motility statistical characterization with
improved techniques [31] is also needed for each species.
B. Swimmers in confined environments
Wall accumulation. The specific parameters charac-
terizing each swimmer motility have a strong influence
on the time that swimmers spend near the walls, a fact
crucial to characterize and to use in applications aim-
ing to control and direct swimmer motion with micro-
geometrical confinements. The near-wall accumulation
effect is illustrated in Fig. 4. These 2d simulations were
performed for all si under the same conditions: a popula-
tion of swimmers, Ns, is introduced in a square obstacle-
free box, i.e without the funnels shown in Fig. 1. The
density distribution versus distance to the walls is mea-
sured, averaging over the four walls and time. Figure 4
shows that a larger τ leads to an increment in the time of
permanence close to the walls, which, in turn, enhances
the swimmer accumulation near the walls. This explains
the large accumulation of swimmers s1 and s2 as com-
pared with that of other swimmers. In particular, if we
0.05
)
1
-
m
µ
0.04
(
y
t
i
s
n
e
D
y
t
i
l
i
b
a
b
o
r
P
0.03
0.02
0.01
0
0
40
s1
s2
s3
s4
s5
s6
s7
80
120
Wall Distance (µm)
5
160
200
(color online) Accumulation of seven different swim-
FIG. 4.
mers (see Table I) near the walls of a 2d container. Swimmers
were confined to a 200 × 200 µm2 box.
compare s1, a real non-tumbling E. coli, with s3, a tum-
bling mutant strain, we observe that s1 has a near-wall
probability density more than twice as large as s3, in
good agreement with recent measurements reported in
Ref. [14]. The role played by the mean angle ¯φ turned
after tumbles can also be analyzed following the progres-
sive decrease in the accumulation of swimmers s3, s4 and
s5, for which the mean speed, ¯v, and mean run duration,
τ , are kept on the same order of magnitude (see table
I). This can be understood as a competition between the
diverting effect caused by tumbles and the directional
persistence due to the small rotation angle around the
mean direction of motion. The accumulation for swim-
mers s3 is consistent with experimental results shown in
Ref. [12, 13] for E. coli.
Single wall of asymmetric funnels between two cham-
bers. We next investigate the dependence of rectification
efficiency on both, geometrical and dynamical parame-
ters. We use a box divided into two equal chambers,
1 (left) and 2 (right), by a wall of asymmetric micro-
designed funnels, as originally used in Ref. [8] and il-
lustrated in Fig. 1.
In Ref. [8] it is reported that the
easy direction of motion for this kind of geometrical rec-
tifier is from chamber 1 to chamber 2. Knowing that
bacteria swim along walls it is possible to gain control
on their motion by optimizing the funnel geometry ac-
cording to the swimming strategy. Our aim is to de-
sign the most efficient rectifier to concentrate different
self-propelled swimmers in chamber 2, and later use this
knowledge to separate efficiently mixed populations. Let
us start by showing the final results (Fig. 5) of a rec-
tification measurement, defined as the ratio of densities
in each chamber, r(t) = ρ2(t)/ρ1(t), after reaching their
steady state.
In simulations this quantity is obtained
after inoculating the swimmers with the same concen-
tration in both chambers ρ2(t = 0) = ρ1(t = 0). The
initial state of the system is prepared as a population
uniformly distributed over the box and then r(t) is mea-
6
(a)
15
12
9
6
3
ρ1
/
ρ2
=
r
0
30
θ (°)
60
90
6
5
4
3
2
1
(b)
(c)
8
6
4
2
0
10
20
lf (µm)
30
40
2
4
6
lg (µm)
8
FIG. 5.
(color online) Rectification efficiency, r = ρ2/ρ1, vs geometrical parameters: (a) Effect of changing the funnel angle,
θ, for swimmers s2 (upper curve, circles) and s3 (lower curve, triangles). (b) Funnel length dependence: r vs lf for s3. (c) Gap
width dependence, r vs lg for s3. After a threshold equal to swimmer diameter plus wall width, rectification decreases with a
wider funnel gap as 1 + a/(lg − b)γ with γ = 0.783 for s3. In (a): Nf = 10, lf = 15 µm, lg = 2 µm. While changing one of
the parameters the box width is scaled to keep the other two parameters unchanged. In (b) and (c), θ = 65◦. The number of
swimmers in the box is also scaled to keep a constant area fraction As/Abox = 0.0785.
sured. Each point in Fig. 5 corresponding to a different
geometrical parameter, θ, lf , lg, has been obtained taking
special care to keep all other parameters fixed, including
the swimmer density, by scaling the box width. We stud-
ied these dependences for all swimmers si in Table I. The
dependence of rectification on the angle of the funnels,
θ, both for an artificial swimmer s2 (upper curve) and
the wild type E. coli s3 (lower curve), presents a well-
defined peak around 65◦ independently of the swimming
strategies (not all shown in Fig. 5(a)), but with markedly
different maximum values. Meanwhile the directing pro-
cess becomes more efficient for longer funnel walls until
saturating for walls longer than 40 µm for the real swim-
mer s3, as shown in Fig. 5(b). Two points need to be
highlighted in Fig. 5(c). First, swimmers begin to pass
through the funnels for a gap of size lg = 1.7 µm, which
is slightly smaller than the sum of swimmer diameter and
funnel wall width, w = 2 µm, due to the softness of the
interactions. Second, once lg is above this threshold, the
rectification decreases as an inverse power law with ex-
ponent γ = 0.783, which is slightly smaller than γ = 1 as
proposed previously with a different model [22]. These
two results together imply an optimum value of lg = 1.8
µm for our system, always in the range of the cell size.
Regarding strictly the "single funnel geometry", it is im-
portant to emphasize that the angular dependences for
designing an optimal rectifier device with a single wall
of funnels are shared by all swimming strategies stud-
ied. The optimal values of the main parameters (θ, lf
or lg) agree with those reported in previous studies that
consider a single strategy to swim [22] but there are im-
portant changes in the magnitude of rectification and on
the time required to achieve it (see Fig. 6). These results
are, for s3, in good quantitative agreement with exper-
iments [8]. The geometrical results obtained show that
neither small angle apertures, corresponding to obstacles
nearly perpendicular to the rectification direction nor an-
gles close to 90◦, representing obstacles closely parallel to
the current direction are profitable (Fig. 5(a)). Regard-
ing the length of the funnels, a short lf is expected not to
be optimal for rectifying because the bacterium cannot
take advantage of the near-to-the-wall swimming for di-
recting its motion, but a long lf implies a longer time of
trapping by the walls in the desired direction until reach-
ing a constant r value. Naturally, the optimal opening
width, lg, will be close to the cell size, preventing reen-
trance or back current swimming, which is easier in wider
gap designs. Small gaps do not exhibit jamming draw-
backs due to the very low concentrations considered in
this work, for which the probability of clogging is negligi-
ble. In conclusion, the best single funnel geometry could
be predicted in advance with accurate simulations that
introduce the specific swimmer motility parameters from
good measurements. This interdisciplinary approach is
very useful for experimentalists during the initial stage
of designing the masks for the lithography process. For
example, if the swimmer is s3, we are sure that the best
geometrical parameters to be chosen are θ ∼ 65◦, lf ∼ 30
µm and lg ∼ the cell size, for an expected efficiency of
r ∼ 5.
In Fig. 6 we show how the swimming strategies of dif-
ferent swimmers influence the time dependence of the ef-
ficiency of the mechanical rectifier. Swimmers with long
runs, such as s1 and s2, achieve almost twice the recti-
fication of s3 [8] and very quickly, in 10 minutes. Note
that even though s1 has higher persistence and larger
mean speed than s2, both are rectified similarly, show-
ing that the considerably longer runs of s2 compensate
for its slower motion and lower persistence. A similar
analysis could be done for s4 and s5. Since both have
the same mean speed and s5 has twice the run duration
of s4, a quicker rectification for s5 could be expected.
However, this advantage is lost because the s5 strategy
involves reversals in the direction of motion. The results
for s6, which represents a typical Brownian swimmer with
very short runs, show that it is not possible to obtain
directed motion if the mean run duration is negligible:
these swimmers cannot profit from swimming along the
walls as in the previous cases. Finally the rectification
time for the s7 swimmer is very similar to the one of the
ρ1
/
ρ2
=
r
7
5
3
1
0
10
20
t (min)
30
40
FIG. 6. (color online) Rectification vs time for swimmers from
Table I: s1(red square), s2(light blue circle), s3(blue up tri-
angle), s4(green down triangle), s5(grey rhombus), s6(orange
pentagon), s7 (filled purple circle). Parameters: Nf = 13,
lf = 27 µm, lg = 3.8 µm, θ = 60◦ and Lx = Ly = 400 µm.
The curve for wild type E. coli, s3, is in excellent agreement
with the experimentally determined time to reach the steady
state and the final rectification value found in [8]. Data are
fitted by Eq. 3. The end of the transitory period (black filled
squares) is calculated as the time needed to get to 99% of the
final rectification r(∞) = α/β.
10
ρ1
/
ρ2
=
r
8
6
4
2
10
ρ1
/
ρ2
=
r
8
6
4
0
4
τ (s)
0
2
80
120
40
φ (°)
6
8
FIG. 7. Dynamical parameters dependences. Rectification,
r, vs τ with persistence ¯φ = 60◦. Inset: r vs ¯φ for τ =1.0 s.
Geometrical parameters used: Nf = 10, lf = 15 µm, lg = 2
µm, θ = 60◦ and Lx = Ly = 136 µm.
wild type strain because they have the same run time du-
ration and only a 22◦ difference in the mean persistence.
A simple model for the rectification by a single barrier
can be obtained by considering two identical chambers
connected by an asymmetric channel [8].
If the initial
concentrations in both chambers are the same, and bac-
teria cross from chamber 1 (2) to chamber 2 (1) at a rate
7
α (β), it is easy to see that the rectification evolves in
time as,
r(t) =
ρ2(t)
ρ1(t)
=
2α + (β − α)e−(α+β)t
2β − (β − α)e−(α+β)t
(3)
In all cases, a very good fit is obtained using Eq. 3.
The fitting parameters are presented in Table S1 at [28].
In general, there is a one-to-one correlation between
how much swimmers get rectified by the funnels and
how much they diffuse (Fig. 3) or get trapped near walls
(Fig. 4). The same is also true for the time elapsed to
reach the steady state (see the values in Table I). This
time was determined as the necessary time for the system
to present 99% of the final rectification. The specific de-
pendence of rectification on the strategy of motion was
further investigated due to its importance, varying the
mean run time, τ , and persistence, ¯φ, continuously, as it
can be seen in Fig. 7. In coincidence with Fig. 6, these
results demonstrate that swimmers with longer runs can
be rectified more efficiently. At very short runs, we ob-
serve a rapid growth of the final rectification, which goes
from r = 1 to r = 5 when τ goes from zero to 2 sec-
onds. The value of r saturates for τ larger than about
5 s. This last result agrees with Fig. 6, where we see
that two swimmers with remarkably different run dura-
tions, s1 with τ = 6.3 s and s2 with τ = 100 s, show the
same r value at long times, even though s1 has higher
persistence, ¯φ = 33◦.
The inset shows that r( ¯φ) follows the opposite trend,
reducing the rectification, as ¯φ increases. For example,
compare s1 and s3, both E. coli mutants. Swimmer s1,
with ¯φ = 33◦, is twice more efficient than s3, which has
a much lower memory of direction after tumbles, ¯φ =
68◦. The most inefficiently rectified swimmers are those
following true random walks paths, as s7 with ¯φ = 90◦,
and the most efficient micro-swimmers are those with the
highest memory of the previous direction of motion.
In short, it is important to note here that after opti-
mizing the geometrical parameters,
the dynamical pa-
rameters have a strong influence on the efficiency of the
rectifiers, as it is evidenced by the results shown in Fig. 6
and Fig. 7. Then we emphasize again that for designing
optimal devices a previous and accurate motility char-
acterization of each species is needed [31] with better
details in runs distribution and persistence. It is impor-
tant to characterize each mutant motility not only in free
swimming but close to surfaces [14] and inside constric-
tions/channels with dimensions comparable to the cell
sizes.
Array of parallel funnel walls. For extra corrobora-
tion of our model, using a geometry consistent of a large
box with five equally distanced and parallel columns of
Nf =10 asymmetric funnels (six chambers from left to
right, ranging from 1 to 6) we inoculated in one extreme,
in chamber 1, a homogeneous population of s3 swimmers.
The simulations reproduce the exponential increase of
the swimmer concentration per consecutive chamber as
it was found experimentally [8] (see Fig. S2 in the Sup-
porting Material [27]). This exponential distribution of
swimmers along the box is due to the geometric progres-
sion of the rectification provided by the successive walls.
We studied the dependence on the number of funnels in
each wall, Nf , concluding that it is very weak. Then, as
we aimed to separate mixed swimmers populations us-
ing their swimming strategies, we simulated larger and
narrower boxes instead, designs closer to tubes or chan-
nels geometries, using the most efficient geometrical pa-
rameters for each opening as obtained previously. We
simulated a mixture of swimmers in a 630× 64 µm2 con-
finement box separated in 21 identical chambers divided
by 20 parallel columns, now with only 2 funnels each.
For the initial state, the swimmers s1, s3 and s4, E. coli
mutants used in Ref. [21], were inoculated in equal pro-
portions in the first chamber. From the precedent analy-
sis of the effects of a single-wall array, we expect that the
various populations will begin to move away from the
inoculation chamber, even without chemical attraction
or nutrient or temperature gradients, simply by physi-
cal guidance and motility. Without the funnel array, the
expected final state at long time, determined by active
diffusion as in section III.A, is a homogenous distribu-
tion of the mixed swimmers over the whole system. But
we observe that the micro-swimmer mixture moves along
the array in the easy ratchet direction with the typical
rectification parameters associated to their respective dy-
namical parameters, until arriving at the opposite end of
the box, the fastest arriving before 2 min (see second
snapshot in Fig. 8). In the first panel of Fig. 8 we show
the initial state of the system, with all swimmers accumu-
lated at the left end of chamber 1, from where they start
moving through the different chambers of the box with
a mean speed that is strongly dependent on the strategy
of motion of the swimmers.
If the three different strains of E. coli are inoculated at
the same time, only part of the s4 population remains in
the first chamber 5 minutes after the inoculation, while
none of the other two strains can be found there. See
these details in Fig. 9, where the numbers of swimmers
at the first and last chambers are plotted vs time. The
times chosen for the snapshots in Fig. 8 are represented
here by vertical dotted lines. The three decreasing curves
in Fig. 9 represent the time dependence of the various
populations in the first chamber and the three increasing
curves are the results measured in the last chamber. In
chamber 21, mutants of the strain s1 begin to arrive just
a minute after inoculation, about 4 minutes before the
arrival of the s3 strain, opening a window of time large
enough as to permit the extraction of purified s1 swim-
mers. Notice in Fig. 8 that during the first two minutes s1
could be extracted alone, in pure concentration, from the
last five chambers, offering also a wide spatial window.
The same behavior can be noticed in Fig 10 which shows
the number of swimmers of each kind in the middle of
the box (chamber number 11) as a function of time. The
strain s1 has its highest concentration there 2 minutes
8
t0 = 0 min
= 2 min
t
1
t2 = 5 min
t3 = 12 min
FIG. 8. (color online) Snapshots of simulated bacteria distri-
bution vs time for a mixture including three E. coli mutants:
(a) CheC497, s1, represented by a red (black) symbol. (b)
AW405 wild-type, s3, blue (dark-grey) and (c) Unc602, s4,
green (light-grey), being the slowest.
3000
t1
t2
t3
s
r
e
m
m
w
S
f
o
i
r
e
b
m
u
N
2250
1500
750
0
0
s1
s3
s4
5
10
t (min)
15
20
FIG. 9.
(color online) Concentration of swimmers in the
first/last chambers (decreasing and increasing curves, respec-
tively) as a function of time, for three E. coli strains s1, s3
and s4. A box Lx = 630 µm, Ly = 64 µm was divided into
equal chambers by 20 columns of funnels, with 2 funnels each,
of lf = 30 µm, lg = 2 µm and θ = 60◦.
after inoculation, whereas s3 needs more than twice that
time to get to the same place. Around 5 minutes after in-
oculation, we could see that all chambers present at least
a small percentage of s4. Then this kind of chamber by
chamber analysis, with our more accurate model, is help-
ful to design optimal separators and density-controlled
mixers according to the swimming strategy and the re-
quired percentage of purity. Properly adjusting the ge-
ometry of the confinement box and the separation of the
funnels walls, it is possible to obtain results in accordance
300
200
100
s
r
e
m
m
w
S
f
o
i
r
e
b
m
u
N
0
0
t1
t2
t3
s1
s3
s4
5
10
t (min)
15
20
FIG. 10. (color online) Populations of swimmers s1, s3 and
s4 passing through the middle chamber, chamber number 11,
vs time. All geometrical parameters are the same as in Fig. 9.
with specified application requirements.
IV. SUMMARY
Arguably among the highest accomplishments in the
separation of motile cells are the sperm sorting achieved
in a micro-fluidic system in 2003 by Cho and cowork-
ers [32] and the sorting of E. coli cells by age/length
reported by Hulme and collaborators in 2008 [33]. There
are many other very recent achievements in sorting dif-
ferent (not self-propelled) particle species using electroki-
netic effects or magnetic fields (Ref. [34] and references
therein). Motivated by this growing technological inter-
est in finding efficient mechanisms able to separate mix-
tures of cell populations, we have studied numerically
both the role of the free-space microorganism dynamics
and the importance of the particular architecture of the
asymmetric obstacles.
We have considered the case of swimming bacteria,
looking at the most relevant parameters characterizing
the interaction between the microorganisms and their
confining walls and searching for their optimal values.
We have extended previous investigations by including
such so far neglected experimental properties as the dis-
tribution of run lengths, the rotational diffusion, and the
directional persistence. Starting by examining the ve-
locity correlation and the mean square displacement of
unconfined self-propelled swimmers, we have shown how
these properties are strongly dependent on run length
and persistence angle. A smaller persistence angle im-
plies a stronger memory of the initial direction and in-
creases both the velocity correlation and the diffusion co-
efficient. Its effect is therefore similar to that of a longer
run length. In the case of short runs, the diffusion coef-
ficient is approximately proportional to the run duration
9
whereas for the long-run swimmers, the effective diffu-
sivity is controlled by the rotational diffusion, Dr. Con-
sequently, the effective translational diffusion coefficient,
D, turns out to be much smaller than what would result
from a simple rescaling of the run length.
When the dynamics of free swimmers is incorporated
into a spatially constrained environment, new effects
emerge. Long run lengths and small tumble emergence
angles lead to high wall accumulation and, consequently,
to fast net displacement in the ratchet direction, which is
clearly confirmed by the numerical calculations. In gen-
eral, long permanence near the walls and suitable wall
architecture favor rectification.
It is important to point out that these results are rather
robust and should remain valid for organisms such as
sperm cells (50 µm), algae (10 µm), bacteria (1 µm) or
even viruses (100 nm) shaken by a drive of zero mean,
such as pressure oscillation [35]. In addition, the concept
of mechanical reorientation of swimmer displacement via
suitable wall architecture would allow one to envisage the
design of surfaces with -phobic or -philic properties for
adsorption or repulsion of microorganisms, respectively.
It would be also of high interest to perform separation
measurements on swimmer mixtures with more complex
motility patterns, such as the recently studied flicking
marine bacteria Vibrio alginolyticus [25]. These mea-
surements could also be performed during the growing
stage of a heterogeneous wild-type E. coli population in
order to select the spherical baby cells by physical guid-
ance with a simple and easy-to-fabricate geometrical ar-
ray (Fig. 8). Such a process would eliminate the stress
and damage associated with centrifugation techniques,
a quality shared with the harder-to-build heart-shaped
channel array proposed originally by Hulme and collab-
orators [33]. This efficient separation by shape during
the cell cycle is very important to study the cell motility
properties by age both in free swimming and close to the
walls of confinement (see Ref. [14]).
In order to guarantee the applicability of our method,
we have to ensure that the smallest constriction remains
larger than the smaller dimension of the swimmer. Re-
cently in Ref. [36, 37], the authors addressed the extreme
case of growth and motion of bacteria in ultra-narrow
constrictions with a size even smaller than their diame-
ter. In this case, free-space swimmer strategy is clearly of
no relevance although the architecture of the constraints
and the particular form of the asymmetric walls can be
of paramount importance.
We
support
acknowledge financial
from CON-
ICET, MINCyT, SeCyT-UNC, and FWO-MINCyT and
KULeuven-UNC bilateral projects and the uncondi-
tional support of F.A. Tamarit on this interdisciplinary
project at Famaf, UNC. This work was partially sup-
ported by the Fonds de la Recherche Scientifique-
FNRS, the Methusalem Funding of the Flemish Govern-
ment, the Fund for Scientific Research-Flanders (FWO-
Vlaanderen).
10
[1] T.M. Squires and S.R. Quake, Rev. Mod. Phys. 77, 977
[20] C. A. Condat, J. Jackle, and S. A. Mench´on, Phys. Rev.
(2005).
[2] K. Leung, H. Zahn, T. Leaver, K.M. Konwar, N.W. Han-
son, A.P. Pag´e, Ch. Lo, P.S. Chain, S.J. Hallam, and C.L.
Hansen, Proc. Natl. Acad. Sci. USA 109, 20 (2012).
[3] J. Voldman, Nature Phys. 5, 536 (2009).
[4] G. Mahmud, C. J. Campbell, K. J. M. Bishop, Y. A.
Komarova, O. Chaga, S. Soh, S. Huda, K. Kandere-
Grzybowska, and B.A. Grzybowski, Nature Phys. 5, 606
(2009).
E 72, 021909 (2005).
[21] H. C. Berg and D.A. Brown, Nature 239, 500 (1972).
[22] M. B. Wan, C. J. Olson Reichhardt, Z. Nussinov, and C.
Reichhardt, Phys. Rev. Lett. 101, 018102 (2008); C. J.
Olson Reichhardt, J. Drocco, T. Mai, M. B. Wan, and
C. Reichhardt, SPIE proc. 8097, doi: 10.1117/12.897424
(2011).
[23] J. Liu and R. S. Ford, Environ. Sci. Technol. 43, 8874
(2009).
[5] K. Konstantopoulus, P. Wu, and D. Wirtz, Biophys. J.
[24] K. Erglis, Q. Wen, V. Ose, A. Zeltins, A. Sharipo, P. A.
104, 279 (2013).
[6] S. Chang, W. Guo, Y. Kim, and Y. Wang, Biophys. J.
104, 313 (2013).
[7] R. Peng, X. Yao, and J. Ding, Biomaterials 32, 8048
(2011).
Janmey, and A. Cebers, Biophys. J. 93, 1402 (2007).
[25] R. Stocker, Proc. Natl. Acad. Sci. USA 108, 2635 (2011).
L. Xie, T. Altindal, S. Chattopadhyay, and X.L. Wu,
Proc. Natl. Acad. Sci. USA 108 2246 (2011).
[26] H. C. Berg, Random walks in biology (Princeton Univer-
[8] P. Galajda, J. Keymer, P. Chaikin, and R. Austin, J.
sity Press, New Jersey 1993).
Bacteriol. 189, 8704 (2007).
[27] See supplemental material at [http:...] for : FigS1 and
[9] S. Y. Kim, E. S. Lee, H. J. Lee, S. Y. Lee, S. K. Lee, and
FigS2.
T. Kim, J. Micromech. Microeng. 20, 0950061 (2010).
[10] P. Denissenko, V. Kantsler, D.J. Smith, and J. Kirkman-
[28] See supplemental material at [http:...] for : TableS1.
[29] P. S. Lovely and F.W. Dahlquist, J. Theor. Biol. 50, 477
Brown, Proc. Natl. Acad. Sci. USA 109, 8007 (2012).
(1975).
[11] S. S´anchez, A.A. Solovev, S.M. Harazim, and O.G.
Schmidt, J. Am. Chem. Soc. 133, 701 (2011).
[12] A.P. Berke, L. Turner, H. C. Berg, and E. Lauga, Phys.
Rev. Lett. 101, 038102 (2008).
[13] G. Li, and J.X. Tang, Phys. Rev. Lett. 103, 078101
(2009).
[14] G. Mino, T.E. Mallouk, T. Darnige, M. Hoyos, J.
Dauchet, J. Dunstan, R. Soto, Y. Wang, A. Rousselet,
and E. Clement, Phys. Rev. Lett. 106, 048102 (2011).
[15] L. Angelani, R. Di Leonardo, and G. Ruocco, Phys. Rev.
Lett. 102, 048104 (2009).
[30] K. Drescher, J. Dunkel, L.H. Cisneros, S. Ganguly, and
R.E. Goldstein, Proc. Natl. Acad. Sci. USA 108, 10940
(2011).
[31] L. G. Wilson, V. A. Martinez, J. Schwarz-Linek, J.
Tailleur, P. N. Pusey, W. C. K. Poon, and G. Bryant,
Phys. Rev. Lett. 106, 018101 (2011).
[32] B. S. Cho, T. G. Schuster, X. Zhu, D. Chang, G. D.
Smith, and S. Takayama, Anal. Chem. 75, 4671 (2003).
[33] S. E. Hulme, W. R. DiLuzio, S. S. Shevkoplyas, L.
Turner, M. Mayer, H. C. Berg, and G. M. Whitesides,
Lab on a Chip 8, 1888 (2008).
[16] A. Sokolov, M. M. Apodaca, B. A. Grzybowski, and I. S.
[34] L. Bogunovic, R. Eichhorn, J. Regtmeier, D. Anselmetti,
Aranson, Proc. Natl. Acad. Sci. USA 107, 969 (2010).
and P. Reimann, Soft Matter 8, 3900 (2012)
[17] G. Lambert, D. Liao and R. H. Austin, Phys. Rev. Lett.
104, 168102 (2010).
[18] E. Altshuler, G. Mino, C. P´erez-Penichet, L. del Rio, A.
Lindner, A. Rousselet, and E. Cl´ement, Soft Matter 9,
1864 (2013).
[19] H. C. Berg, E. coli in motion (Springer, New York, 2004).
[35] S. Matthias and F. Muller, Nature 424, 53 (2003).
[36] J. Mannik, R. Driessen, P. Galajda, J.E. Keymer, and C.
Dekker, Proc. Natl. Acad. Sci. USA 106, 14861 (2009).
[37] Q. Guo, S. M. McFaul, and H. Ma, Phys. Rev. E. 83,
051910 (2011).
|
1504.03984 | 1 | 1504 | 2015-04-14T10:43:58 | From Animaculum to single-molecules: 300 years of the light microscope | [
"physics.bio-ph",
"physics.optics",
"q-bio.BM"
] | Although not laying claim to being the inventor of the light microscope, Antonj van Leeuwenhoek, (1632-1723) was arguably the first person to bring this new technological wonder of the age properly to the attention of natural scientists interested in the study of living things (people we might now term biologists). He was a Dutch draper with no formal scientific training. From using magnifying glasses to observe threads in cloth, he went on to develop over 500 simple single lens microscopes with which he used to observe many different biological samples. He communicated his finding to the Royal Society in a series of letters including the one republished in this edition of Open Biology. Our review here begins with the work of van Leeuwenhoek before summarising the key developments over the last ca. 300 years which has seen the light microscope evolve from a simple single lens device of van Leeuwenhoek's day into an instrument capable of observing the dynamics of single biological molecules inside living cells, and to tracking every cell nucleus in the development of whole embryos and plants. | physics.bio-ph | physics | From Animaculum to single-molecules: 300 years of the light microscope
Adam J. M. Wollman, Richard Nudd, Erik G. Hedlund, Mark C. Leake*
Biological Physical Sciences Institute (BPSI), Departments of Physics and Biology, University
of York, York YO10 5DD, UK.
*[email protected]
Abstract
Although not laying claim to being the inventor of the light microscope, Antonj van
Leeuwenhoek, (1632 –1723) was arguably the first person to bring this new technological
wonder of the age properly to the attention of natural scientists interested in the study of
living things (people we might now term ‘biologists’). He was a Dutch draper with no formal
scientific training. From using magnifying glasses to observe threads in cloth, he went on to
develop over 500 simple single lens microscopes1 with which he used to observe many
different biological samples. He communicated his finding to the Royal Society in a series of
letters2 including the one republished in this edition of Open Biology. Our review here
begins with the work of van Leeuwenhoek before summarising the key developments over
the last ca. 300 years which has seen the light microscope evolve from a simple single lens
device of van Leeuwenhoek’s day into an instrument capable of observing the dynamics of
single biological molecules inside living cells, and to tracking every cell nucleus in the
development of whole embryos and plants.
1.
Antonj van Leeuwenhoek and invention of the microscope
Prior to van Leeuwenhoek, lenses had existed for hundreds of years but it was not until the
17th century that their scientific potential was realized with the invention of the light
microscope. The word ‘microscope’ was first coined by Giovanni Faber in 1625 to describe
an instrument invented by Galileo in 1609. Gailieo’s design was a compound microscope - it
used an objective lens to collect light from a specimen and a second lens to magnify the
image, but this was not the first microscope invented. In around 1590 Hans and Zacharias
Janssen had created a microscope based on lenses in a tube3. No observations from these
microscopes were published and it was not until Robert Hooke and Antonj van
Leeuwenhoek that the microscope, as a scientific instrument, was born.
Robert Hooke was a contemporary of van Leeuwenhoek. He used a compound microscope,
in some ways very similar to those used today with a stage, light source and three lenses. He
made many observations which he published in his Micrographia in 1665.4 These included
seeds, plants, the eye of a fly and the structure of cork. He described the pores inside the
cork as ‘cells’, the origin of the current use of the word in biology today.
Unlike Hooke, van Leeuwenhoek did not use compound optics but single lenses. Using only
one lens dramatically reduced problems of optical aberration in lenses at the time, and in
fact van Leeuwenhoek’s instruments for this reason generated a superior quality of image to
those of his contemporaries. His equipment was all handmade, from the spherical glass
lenses to their bespoke fittings. His many microscopes consisted mainly of a solid base, to
hold the single spherical lens in place, along with adjusting screws which were mounted and
glued in place to adjust the sample holding pin, and sometimes an aperture placed before
the sample to control illumination1 (see Figure 1 for an illustration). These simple
instruments could be held up to the sun or other light source such as a candle and did not
themselves have any light sources inbuilt. His microscopes were very lightweight and
portable, however, allowing them to be taken into the field to view samples as they were
collected. Imaging consisted of often painstaking mounting of samples, focussing and then
sketching, with sometimes intriguing levels of imagination, or documenting observations.
Van Leeuwenhoek’s studies included the microbiology and microscopic structure of seeds,
bones, skin, fish scales, oyster shell, tongue, the white matter upon the tongues of feverish
persons, nerves, muscle fibres, fish circulatory system, insect eyes, parasitic worms, spider
physiology, mite reproduction, sheep foetuses, aquatic plants and the ‘animalcula’ – the
microorganisms described in his letter.2 As he created the microscopes with the greatest
magnification of his time he pioneered research into many areas of biology. He can arguably
be credited with the discovery of protists, bacteria, cell vacuoles and spermatozoa.
2.
The development of the microscope and its theoretical underpinnings
It was not until the 19th century that the theoretical and technical underpinnings of the
modern light microscope were developed. Most notably diffraction-limit theory, but also
aberration-corrected lenses and an optimized illumination mode called Köhler illumination.
There is a fundamental limit to the resolving power of the standard light microscope; these
operate by projecting an image of the sample a distance of several wavelengths of light
from the sample itself, known as the ‘far-field’ regime. In this regime the diffraction of light
becomes significant, for example through the circular aperture of the objective lens. This
diffraction causes ‘point sources’ in the sample which scatter the light to become blurred
spots when viewed through a microscope, with the level of blurring determined by the
imaging properties of the microscope known as the point spread function (PSF). Through a
circular aperture, such as those of lenses in a light microscope, the PSF can be described by
a mathematical pattern called an Airy disk, which contains a central peak of light intensity
surrounded by dimmer rings moving away from the centre (see Figure 2a). This
phenomenon was first theoretically characterized by George Airy in 1835.5 Later, Ernst Abbe
would state that the limit on the size of the Airy disk was roughly half the wavelength of the
imaging light6 which agrees with the so-called Raleigh criterion for the optical resolution
limit7, which determines the minimum distance between resolvable objects (see Figure 2b).
This became the canonical limit in microscopy for over a hundred years, with the only
attempts to improve spatial resolution through the use of lower wavelength light or using
electrons rather than photons, as in electron microscopy, which have a smaller effective
wavelength by ~4 orders of magnitude.
Ernst Abbe also helped solve the problem of chromatic aberration. A normal lens focuses
light to different points depending on its wavelength. In the 18th century, Chester Moore
Hall invented the achromatic lens which used two lenses of different materials fused
together to focus light of different wavelengths to the same point. In 1868 Abbe invented
the apochromatic lens, using more fused lenses, which better corrected chromatic and
spherical aberrations.8 Abbe also created the world's first refractometer and we still use the
‘Abbe Number’ to quantify how diffraction varies with wavelength.9 He also collaborated
with Otto Schott, a glass chemist, to produce the first lenses which were engineered with
sufficiently high quality to produce diffraction-limited microscopes.10 Their work in 1883 set
the limits of far-field optics for over a century, until the advent of the 4π microscope in
1994.11
Another eponymous invention of Abbe was the Abbe condenser – a unit which focuses light
with multiple lenses which improved sample illumination but was quickly superseded by
Köhler Illumination, the modern standard for ‘brightfield’ light microscopy. August Köhler
was a student of many fields of the ‘natural sciences’. During his PhD studying limpet
taxonomy he modified his illumination optics to include a field iris and also an aperture iris
with a focusing lens to produce the best illumination with the lowest glare which aided in
image collection using photosensitive chemicals.12 Due to the slow nature of photography of
the period good images required relatively long exposure times and Köhler Illumination
greatly aided in producing high-quality images. He joined the Zeiss Optical Works in 1900,
where his illumination technique coupled with the optics already developed by Abbe and
Schott went on to form the basis of the modern brightfield light microscope.
3.
Increasing optical contrast
One of the greatest challenges in imaging biological samples is their inherently low contrast,
due to their refractive index being very close to water and thus generating little scatter
interaction with incident light. A number of different methods for increasing contrast have
been developed including imaging phase and polarization changes, staining and
fluorescence, the latter being possibly the most far-reaching development since the
invention of the light microscope.
Biological samples generate contrast in brightfield microscopy by scattering and absorbing
some of the incident light. As they are almost transparent, the contrast is very poor. One
way around this, is to generate contrast from phase (rather than amplitude) changes in the
incident light wave. Fritz Zernike developed phase contrast microscopy in the 1930s13 while
working on diffraction gratings. Imaging these gratings with a telescope, they would
‘disappear’ when in focus.14 These observations led him to realize the effects of phase in
imaging, and their application to microscopy subsequently earned him the Nobel prize in
1953. Phase contrast is achieved by manipulating the transmitted, background light
differently from the scattered light, which is typically phase-shifted 90 degrees by the
sample. This scattered light contains information about the sample. A circular annulus is
placed in front of the light source, producing a ring of illumination. A ring-shaped phase
plate below the objective, shifts the phase of the background light by 90 degrees such that it
is in phase (or sometimes completely out of phase, depending on the direction of the phase
shift) with the scattered light producing a much higher contrast image.
An alternative to phase contrast is Differential Interference Contrast (DIC). It was created by
Smith15 and further developed by Georges Nomanski in 1955.16 It makes use of a Normanski-
Wollaston Prism through which polarized light is sheared into two beams polarized at 90
degrees to each other. These beams then pass through the sample and carry two brightfield
images laterally displaced a distance equal to the offset of the two incoming beams at the
sample plane. Both beams are focussed through the objective lens and then recombined
through a second Normanski-Wollaston prism. The emergent beam goes through a final
analyser emerging with a polarization of 135 degrees. The coaxial beams interfere with each
other owing to the slightly different path lengths of the two beams at the same point in the
image, giving rise to a phase difference and thus a high contrast image. The resultant image
appears to have bright and dark spots which resemble an illuminated relief map. This faux
relief map should not be interpreted as such, however, as the bright and dark spots contain
information instead about path differences between the two sheared beams. The images
produced are exceptionally sharp compared to other transmission modes. DIC is still the
current standard technique for imaging unstained microbiological samples in having an
exceptional ability to reveal the boundaries of cells and subcellular organelles.
Contrast can also be improved in biological samples by staining them with higher contrast
material, for example dyes. This also allows differential contrast, where only specific parts of
a sample, such as the cell nucleus, are stained. In 1858 came one of the earliest documented
staining in microscopy when Joseph von Gerlach demonstrated differential staining of the
nucleus and cytoplasm in human brain tissue soaked in the contrast agent carmine.17 Other
notable examples include silver staining introduced by Camillo Golgi in 1873, which allowed
nervous tissue to be visualized, 18 and Gram staining invented by Hans Christian Gram in
1884,19 which allowing differentiation of different types of bacteria. Sample staining is still
widely in use today, including many medical diagnostic applications. However, the advent of
fluorescent staining would revolutionize contrast enhancement in biological samples.
The word ‘fluorescence’ to describe emission of light at a different wavelength to the
excitation wavelength was first made by Stokes in 1852.20 Combining staining with
fluorescence detection allows for enormous increases in contrast, with the first fluorescent
stain fluorescein being developed in 1871.21 In 1941, Albert Coons published the first work
on Immunofluorescence. This technique uses fluorescently-labelled antibodies to label
specific parts of a sample. Coons used a fluorescein derivative labelled antibody and showed
that it could still bind to its antigen. 22 This opened the way to using fluorescent antibodies
as a highly specific fluorescent stain.
Green fluorescent protein (GFP) was first isolated from the jellyfish, Aequorea victoria, in
196223 but it was not until 1994 that Chalfie et al.24 showed that it could be expressed and
fluoresce outside of the jellyfish. They incorporated it into the promoter for a gene that
encoded β-tubulin and showed that it could serve as a marker for expression levels. The
discovery and development of GFP by Osamu Shimomura, Martin Chalfie and Roger Tsien
was recognised in 2008 by the Nobel prize in chemistry.
By mutating GFP, blue, cyan and yellow derivatives had been manufactured25 but orange
and red fluorescent proteins proved difficult to produce until the search for fluorescent
proteins was expanded to non-bioluminescent organisms. This led to the isolation of dsRed
from Anthozoa, a species of coral.26 Brighter and more photostable fluorescent proteins
were subsequently produced by directed evolution.27 The discovery of spectrally distinct
fluorescent proteins allowed multichannel (dual and multi-colour) fluorescence imaging and
opened the way to studying the interaction between different fluorescently-labelled
proteins.
Early work with fluorescent proteins simply co-expressed GFP on the same promoter as
another gene to monitor expression levels. Proteins could also be chemically labelled
outside of the cell and then inserted using microinjection.25,28 A real breakthrough, with the
discovery of GFP, was optimizing a method to fuse the genes of a protein of interest with a
fluorescent protein and express this in a cell - thus leaving the cell relatively unperturbed.
This was first demonstrated29 on a GFP fusion to the bcd transcription factor in Drosophila.30
Fluorescent dyes have been used, not just high-contrast markers, but as part of molecular
probes, which can readout dynamics between molecules and also environmental factors
such as pH. In 1946, Theodore Förster posited that if a donor and acceptor molecule were
sufficiently close together, non-radiative transfer of energy could occur between the two,
now known as Förster resonance energy transfer (FRET), with efficiency proportional to the
sixth power of the distance between them.31 If such molecules are themselves fluorescent
dyes then fluroescenbce can be used as a metric of putative molecular interaction through
FRET. In 1967, Stryer and Haugland showed this phenomenon could be used as a molecular
ruler over a length scale of ~1-10 nm.32 Since then, FRET is used routinely to image
molecular interactions and the distances between biological molecules, and also in
fluorescence lifetime imaging (FLIM).33 Fluorescent probes have also been developed to
detect cell membrane voltages, local cellular viscosity levels and the concentration of
specific ions, with calcium ion probes, for example, first introduced by Roger Tsien in 1980.34
4. The Fluorescence Microscope
The fluorescence microscope has its origins in ultraviolet (UV) microscopy. Abbe theory
meant that better spatial resolution could be achieved using shorter wavelengths of light.
August Köhler constructed the first UV microscope in 1904.35 He found that his samples
would also emit light under UV illumination (although he noted this as an annoyance). Not
long after, Oskar Heimstaedt realized the potential for fluorescence and had a working
instrument by 1911.36 These transmission fluorescence microscopes were greatly improved
in 1929 when Philipp Ellinger and August Hirt, placed the excitation and emission optics on
the same side as the sample and invented the ‘epifluorescence’ microscope.37 With the
invention of dichroic mirrors in 196738, this design would become the standard in
fluorescence microscopes. Several innovative illumination modes have also been developed
for the fluorescence microscope, which have allowed it to image many different samples
over a wide range of length scales. These modes include confocal, FRAP, TIRF, two-photon
and light-sheet microscopy.
In conventional fluorescence microscopy, the whole sample is illuminated and emitted light
collected. Much of the collected light is from parts of the sample which are out of focus. In
confocal microscopy, a pinhole is placed after the light source such that only a small portion
of the sample is illuminated and another pinhole placed before the detector such that only
in-focus light is collected (see Figure 1). This can reduce the background in a fluorescence
image and allow imaging further into a sample. The latter even enables optical sectioning
and 3D reconstruction. The first confocal microscope was patented by Marvin Minsky in
1961.39 This instrument preceded the laser so the incident light was not bright enough for
fluorescence. With laser-scanning confocal microscopes40 much better fluorescence
contrast is achievable, as explored by White who compared the contrast in different human
and animal cell lines.41
Fluorophores only emit light for a short time before they are irreversibly photobleached,
and so microscopists must limit their sample’s exposure to excitation light. Photobleaching
can be used to reveal kinetic information about a sample by fluorescence recovery. In the
earliest fluorescence recovery study, in 1974, Peters et al. bleached one half of fluorescein-
labelled human erythrocyte plasma membranes and found that no fluorescence returned,
indicating no observable mean diffusive process of the membrane over the experimental
time scales employed.42 Soon after, analytical work by Axelrod et al.43 (on what they termed
Fluorescence Photobleaching Recovery) allowed them to characterize different modes of
diffusion in intracellular membrane trafficking. The term Fluorescence Recovery After
Photobleaching (FRAP) appears to have been coined by Jacobson, Wu and Poste in 1976.44
With FRAP capabilities commercially available on confocal systems, it is now widely used for
measuring turnover kinetics in live cells.
When imaging features that are thin or peripheral such as cell membranes and molecules
embedded in these, a widely used method is Total Internal Reflection Fluorescence (TIRF)
microscopy. This technique uses a light beam introduced above the critical angle of the
interface between the (normally) glass microscope coverslip and the water-based sample.
The beam itself will be reflected by total internal reflection due to the differences in
refractive index between the water and the glass, but at the interface an evanescent wave
of excitation light is generated which penetrates only ~100 nm into the sample, thus only
fluorophores close to the coverslip surface are excited, producing much higher signal-to-
noise than conventional epifluorescence microscopy. It was first demonstrated on biological
samples by Axelrod in 1981 to image membrane proteins in rat muscle cells and lipids in
human skin cells.45
In conventional epifluorescence or even confocal, there is a limit to how far into the sample
it is possible to image because of incident light scattering from the sample, creating a
fluorescent background. This is particularly problematic when imaging tissues. Longer
wavelength light scatters much less but few fluorophores can be excited by this with
standard single photon excitation. In her doctoral thesis, in 1931, Maria Gopport-Mayer
theorized that two photons with half the energy needed, can excite emission of one photon
whose energy was the sum of the two photons during a narrow time window for absorption
of ~10-18 s. 46 The phenomenon of two-photon excitation (2PE) was not observed
experimentally for another 30 years, until Kaiser and Garrett demonstrated it in CaF
crystals.47 The probability of 2PE occurring in a sample is low due to the very narrow time
window of coincidence with respect to the two excitation photons, so high intensity light
with a large photon flux is required to use the phenomenon in microscopy. In 1990 Denk
used a laser in a confocal scanning microscope to image human kidney cells w 2PE.48 Since
then, it has become a powerful technique for observing molecular processes in live tissues,
particularly in neuroscience, where the dynamics of neurons within a live rat brain were first
observed by Svoboda et al.49
Another method of reducing background in fluorescent samples is to only illuminate the
sample through the plane which is in focus. This can be achieved by shining a very flat
excitation beam through the sample perpendicular to the optical axis. A. Voie et al. first
demonstrated this, using light-sheet microscopy (LSM) in 1993.50 LSM can be used to take
fluorescence images through slices of a sample, allowing a stack of images to build a 3D
reconstruction. One caveat of LSM is that samples need to be specially mounted to allow an
unobstructed excitation beam as well as a perpendicular detection beam so a bespoke
microscope is required. The technique was pioneered and developed by Ernst Stelzer in
2004, and termed Selective Plane Illumination Microscopy (SPIM), it was used to image live
embryos in 3D.51 Stelzer’s group went on to image and track every nucleus in a developing
zebrafish over 24 hours52 and also the growth of plant roots at the cellular level in
Arabidopsis.53 LSM has proven itself a powerful tool for developmental biology, the
potential of which is only now being realized.
5.
Improving resolution in length and time
Fluorescence microscopy set new standards of contrast in biological samples that have
enabled the technique to achieve possibly the ultimate goal of microscopy in biology and
visualize single molecules in live cells. The Abbe diffraction limit, thought unbreakable for
over one hundred years, has been circumvented by ever more inventive microscopy
techniques which are now extending into three spatial dimensions.
The first single biological molecules detected were observed by Cecil Hall in the
1950s,54using electron microscopy of metallic fibres shadowed replicas of large, filamentous
molecules including DNA and fibrous proteins, with dried samples in a vacuum. The very
first detection of a single biological molecule in its functional aqueous phase was made by
Boris Rotman, his seminal work published in 1961 involving the observation of
fluorescently-labelled substrates of beta-galactosidase suspended in water droplets. The
enzyme catalysed the hydrolysis of galactopyranose labelled with fluorescein to the sugar
galactose plus free fluorescein, which had a much greater fluorescence intensity than when
attached to the substrate. He could detect single molecules because each enzyme could
turnover thousands of fluorescent substrate.55 A more direct measurement was made by
Thomas Hirschfield, in work published in 1976, who managed to see single molecules of
globulin, labelled with ~100 fluorescein dyes, passing through a focused laser.56 Single dye
molecules were not observable directly until the advent of Scanning Near-field Optical
Microscopy (SNOM) developed by Eric Betzig and Robert Chichester allowing them to image
individual cyanine dye molecules in a sub-monolayer.57 SNOM uses an evanescent wave
from a laser incident on a ~100 nm probe aperture which illuminates a small section and
penetrates only a small distance into the sample. Images are generated by scanning this
probe over the sample. This is technically challenging as the probe must then be very close
to the sample.
Single molecules were shown to be observable with less challenging methods when, using
TIRF microscopy, single ATP turnover reactions in single myosin molecules was observed in
1995.58 Other studies observed single F1-ATPase rotating using fluorescently labelled actin
molecules in 199759 and the dynamics of single cholesterol oxidase molecules.60 In a
landmark study, the mechanism and step size of the myosin motor was determined by
labelling one foot, observing and using precise Gaussian fitting to obtain nanometre
resolution (termed ‘fluorescence imaging with one nanometer accuracy’ - FIONA).61 This
localization microscopy could effectively break the diffraction limit by using mathematical
fitting algorithms to pinpoint the centre of a dye molecule’s PSF image, as long as they are
resolvable such that the typical nearest-neighbour separation of dye molecules in the
sample is greater than the optical resolution limit. These techniques were soon applied to
image single molecules in living cells62,63 and now it is possible to count the number single
molecules in complexes inside cells.64,65
Stefan Hell showed that it was possible to optically break the diffraction limit with a more
deterministic technique which modified the actual shape of the PSF, called Stimulated
Emission Depletion microscopy (STED), which he proposed with Jan Wichmann in 199466
and implemented with Thomas Klar in 1999.67 STED works by depleting the population of
excited energy state electrons through stimulated emission. Fluorescence emission only
occurs subsequently from a narrow central beam inside the deactivation annulus region
which is scanned over the sample. The emission region is smaller than the diffraction limit
(~100 nm in the original study), thus allowing a superresolution image to be generated.
The development of STED showed that the diffraction limit could be broken and many new
techniques followed. In 2002, Ando et al.68 isolated a fluorescent protein from the stony
coral, Trachyphyllia geoffroyi, which they named Kaede. They found that if exposed to UV
light, its fluorescence would change from green to red and demonstrated this in Kaede
protein expressed in HeLa cells. Photoactivatable proteins, such as this were used in 2006 by
Hess et al. in Photo-Activated Localization Microscopy (PALM) using TIRF69 and by Betzig et
al. in Fluorescence Photo-Activated Localization Microscopy (FPALM) using confocal. Both
methods use low intensity long UV laser light to photoactivate a small subset of sample
fluorophores then another laser to excite them to emit and photobleach. This is repeated to
build a superresolution image. A related method utilizes stochastic photoblinking of
fluorescent dyes, which for example can be used to generate superresolution structures of
DNA.70
Other notable superresolution techniques include Structured Illumination Microscopy
(SIM).71 In 1993, B. Bailey et al showed that structured stripes of light could be used to
generate a spatial ‘beat’ pattern in the image which could be used to extract spatial features
in the underlying sample image which had a resolution of ~2 times that of the optical
resolution limit. In 2006, Xiaowei Zhuang et al demonstrated Stochastic Optical
Reconstruction Microscopy (STORM)72 which used a Cy5/Cy3 pair as a switchable probe. A
red laser keeps Cy5 in a dark state and excites fluorescence, while a green laser brings the
pair back into a fluorescent state. Thus, similarly to PALM, a superresolution image can be
generated.
Improvements in dynamic fluorescence imaging have been significant over the past few
decades. For example, using essentially the same localization algorithms as developed for
PALM/STORM imaging, fluorescent dye tags can be tracked in a cellular sample in real-time
for example tracking of membrane protein complexes in bacteria to nanoscale
precicsion,73which have been extended into high time resolution dual-colour microscopy in
vivo to monitor dynamic co-localization over with a spatial precision of ~10-100 nm.74
Modifications to increase the laser excitation of several recent bespoke microscope systems
have also improved the time resolution of fluorescence imaging down to the millisecond
level, for example using narrowfield and Slimfield microscopy.75
3D information can be obtained in many ways including using interferometric methods76 or
multiplane microscopy77 which image multiple focal planes simultaneously. Another method
of encoding depth information in images is to distort the PSF image in an asymmetrical but
measureable way as the light source moves away from the imaging plane. Astigmatism and
double-helix microscopy accomplish this using different methods and are compatible with
many modes of fluorescence illumination as the equipment used is placed between the
objective lens and the camera. As such, it is a viable way to extract 3D data from many
currently developed fluorescence microscopes.
Astigmatism microscopy is a simple 3D microscopy technique, first demonstrated by Kao
and Verkman in 1994.78 An asymmetry is introduced in the imaging path by placing a
cylindrical lens before the camera detector. The introduced astigmatism offsets the focal
plane along one lateral axis slightly, resulting in a controlled image distortion. When imaging
singular or very small aggregates of fluorophores, the distortion takes the form of an ellipse,
extending along either the x or y axis in the lateral plane of a camera detector conjugate to
the microscope focal plane, depending on whether the fluorophore is above or below the
focal plane. Values of 30 nm resolution in the lateral plane and 50 nm in the axial dimension
have been reported using astigmatism with STORM.79
Double-helix PSF (DH-PSF) microscopy is a similar 3D microscopy technique using controlled
PSF distortion. It exploits optical vortex beams, beams of light with angular momentum and
works by placing a phase mask – an object which modifies the phase of the beam differently
at different points along a cross-section – between the camera and the objective lens to
turn the laser beam intensity profile from a Gaussian beam to a mixture of higher-order
optical vortex beams - a superposition of two so-called Laguerre-Gauss (LG) beams. These
two beams interfere with each other at the point that the light hits the camera creating two
bright lobes.80 The fields rotate as a function of distance propagated. As the two beams are
superposed the distance is the same; if the two LG beams are slightly different the electric
fields will rotate at different rates thanks to different so-called ‘Gouy Phase’ components.
This means that the interference pattern produced rotates as a function of the distance of
the point source from the image plane only.81 The distance from the focal plane can be
determined by measuring the rotation angle of the two lobes.
The phase mask can be created using transparent media such as etched glass or using a
Spatial Light Modulator (SLM). An SLM is a 2D array of microscale bit components, each of
which can be used to change the phase of the incident light across a beam profile. A liquid-
crystal-on-silicon SLM retards light as a function of the input voltage to each bit. As such, a
phase mask can be applied and changed in real-time using computer control. One major
drawback is that they are sensitive to the polarization of light82 limiting the efficiency of light
propagation through the SLM. Alternatively a fixed glass phase plate can be etched using
nanolithography. This is phase-independent and much more photon efficient. The phase is
retarded simply by the thickness of the glass at each point in the beam. However, glass
phase plates are less precise than SLM due to limitations in the lithography. Still, these are
much easier to implement and can be purchased commercially or custom-built and used
with almost any microscope setup with minimal detrimental impact. DH-PSF microscopy has
been shown to have some of the smallest spatial localization errors of any 3D localization
mode in high signal to noise systems.83
The power of beam-shaping combined with light-sheet illumination has been recently used
to create lattice light-sheet microscopy.84 Using Bessel beams, which focused laser profiles
with minimal divergence due to diffraction, they create different bound optical lattices with
different properties allowing them to image across four orders of magnitude in space and
time and in diverse samples including diffusing transcription factors in stem cells, mitotic
microtubules and embryogenesis in Caenorhabditis elegans.
6.
The future
Although over 300 years since the pioneering work of van Leeuwenhoek, many of the major
developments in light microscopy have occurred in just the past few decades and their full
impact may not yet be felt. There are several technologies currently in development which
may have a profound impact on microscopy. These include, for example, adaptive optics,
lens-free microscopy, super lenses, miniaturization and combinational microscopy
approaches.
A biological sample itself adds aberration through spatial variation in the refractive index.
This is even more of a problem when imaging deep into tissues. Adaptive optics uses so-
called dynamic correction elements such as deformable mirrors or SLMs to correct for this
aberration, increasing spatial resolution and contrast. There have been many recent
developments, reviewed comprehensively by Martin Booth,85 but the technology is still yet
to be widely adopted.
The archetypal lens used in light microscopy is made of glass, however this is not the only
type of lens available. Optical diffraction gratings (optical gratings) can be used to focus,
steer and even reflect light. Recognizing the need for miniaturization, researchers have been
investigating the use of diffraction gratings in place of glass to help reduce the necessary
size of optical components. While glass is great for large applications, it is extremely bulky
when compared to the minimum size of a diffraction grating.86 Optical gratings can be used
as equivalent to lenses under some circumstances, for example a Fresnel Zone Plate can be
used to focus light to a point as a convex lens does. Optical gratings all rely on the
interaction of electromagnetic waves as they pass through the spaces in the gratings. This is
fundamentally linked to the wavelength of propagating light making achromatic optical
gratings very difficult to achieve in practice. Only recently have scientists been able to
produce achromatic glass analogues such as an achromatic grating quarter-wave plates, for
example, with good operational ranges.87
Ptychography completely removes the need for imaging optics, lenses or gratings, and
directly reconstructs real-space images from diffraction patterns captured from a beam
scanned over a sample. In many cases, this allows higher contrast images than DIC or phase
contrast and 3D reconstruction.88,89
All optics currently used in microscopes are diffraction-limited but it is theoretically possible
to construct, using so-called ‘metamaterials’, a perfect lens or super lens which could image
with perfect sharpness. This was thought to require a material with negative refractive
index90 but it has now been shown that ordinary positive refractive index materials can also
be used.91 Even if super lenses are not achievable, new materials may revolutionize
microscope lenses, still mostly composed of the same materials used by van Leeuwenhoek.
It is interesting to note the return of microscopes such as van Leeuwenhoek’s which use
only a single lens, in the foldascope developed by Manu Prakash at Stanford University92
Using cardboard (an essential and surprisingly cheap component of some of the most
advanced bespoke light microscopes found in our own laboratory) and simple filters and
lenses, a near indestructible microscope with both normal transmission modes and
fluorescence modes has been created that can be used by scientists and physicians working
in areas far from expensive lab equipment.
Combinatorial microscopy is an interesting recent advance, which shows significant future
potential. Here, several different microscopy methods are implemented on the same light
microscope device. Many advances are being made at the level of single-molecule
biophysics coupled to light microscopy in this regard. For example, methods being
developed which can permit simultaneous superresolution imaging of DNA coupled to
magnetic tweezers manipulation.93
The ultimate practical limits at the other end of the length scale for imaging tissues and
whole organisms in the future are difficult to determine. Recent technological
developments, such as the light-sheet imaging of Arabidopsis or lattice light-sheet
microscopy discussed previously have enabled imaging of ever larger samples in greater
detail. What limits the largest possible sample and to what level of detail it can be imaged is
unknown. And, just as importantly, is computing technology used to store and analyse these
data up to the challenge?
It is unquestionable that light microscopy has advanced enormously since the days of Antonj
van Leeuwenhoek. The improvements have been, in a broad sense, twofold. Firstly, in
length scale precision. This has been a ‘middling-out’ improvement, in that superresolution
methods have allowed unprecedented access to nanoscale biological features, whereas
light-sheet approaches and multi-photon deep imaging methods in particular have allowed
incredible detail to be discerned at the much larger length scale level of multicellular tissues.
Secondly, there has been an enormous advance, almost to the level of a paradigm shift,
towards faster imaging in light microscopy, to permit truly dynamic biological processes to
be investigated, right down to the millisecond level. Not only can we investigate detailed
biological structures using light microscopy, but we can watch them change with time.
And yet, equally so, the basic principles of light microscopy for the study of biology remain
essentially unchanged. These were facilitated in no small part by the genius and diligence of
van Leeuwenhoek. It is perhaps the finest legacy for a true pioneer of light microscopy (.
Author’s contributions
All authors contributed to the drafting and revision of the manuscript, and gave their final
approval for publication.
Acknowledgements
M.C.L. is supported by a Royal Society University Research Fellowship (UF110111).
E.G.H. was supported by Marie Curie EU FP7 ITN ‘ISOLATE’ ref 289995.
R.N. was supported by the White Rose Consortium.
The work was supported by the Biological Physics Sciences Institute (BPSI).
Figures
Figure 1
Optical microscope designs through the ages. a) One design of a simple compound microscope used
by Hooke while writing Micrographia. b) An example of the single spherical lens mount system that
van Leeuwenhoek used, approximately 5 cm in height. c) A simple epi-fluorescence system. d) A
simple modern-day confocal microscope.
Figure 2
Mathematically generated Point Spread Function (PSF) images from in different light microscope
designs. a) The Airy pattern, a disc and one of the rings produced by a point source emitter imaged
using a spherical lens. b) Two such Airy discs separated by less than the Abbe limit for optical
resolution. c) The lateral xy stretching exhibited in astigmatic imaging systems when the height z of a
point source emitter is above or below the focal plane, the degree of stretching a metric for z. d)
Expected pattern when a point source emitter is defocused. e) Two-lobed PSF used in Double-Helix
PSF techniques, where the rotation of the lobes about the central point is used to calculate z.
Figure 3
By chance, in the last days of finishing this review the corresponding author was staying ~100m from
Leeuwenhoek's final resting place in the Oude Kerk, Delft, and captured these images.
References
1.
Baker, H. & Leeuwenhoek, M. An Account of Mr. Leeuwenhoek’s Microscopes; By Mr.
Henry Baker, F. R. S. 1739 Philos. Trans. R. Soc. London 41, 503–519.
2.
Leeuwenhoek, A. van & Hoole, S. 1800 The Select Works of Antony Van Leeuwenhoek,
Containing His Microscopical Discoveries in Many of the Works of Nature, Volume 1 .
3.
Orchard, G. & Nation, B. 2014 Cell Structure & Function. 448 Oxford University Press,.
4.
Hooke, R. 1665 Micrographia: Or Some Physiological Descriptions of Minute Bodies
Made by Magnifying Glasses, with Observations and Inquiries Thereupon. 273 Courier
Corporation,.
5.
Airy, G. B. 1835 On the Diffraction of an Object-glass with Circular Aperture. Trans.
Cambridge Philos. Soc. 5,.
6.
Abbe, E. 1873 Beiträge zur Theorie des Mikroskops und der mikroskopischen
Wahrnehmung. Arch. für Mikroskopische Anat. 9, 413–418.
7.
F.R.S., L. R. XXXI. 1879 Investigations in optics, with special reference to the
spectroscope. Philos. Mag. Ser. 5 8, 261–274.
8. Mansuripur, M. 2002 Classical Optics and Its Applications. 502 Cambridge University
Press,.
9.
Abbe, E. 1881 VII.-On the Estimation of Aperture in the Microscope. J. R. Microsc. Soc.
1, 388–423.
10.
Abbe, E. 1883 XV.-The Relation of Aperture and Power in the Microscope. J. R.
Microsc. Soc. 3, 790–812.
11. Hell, S. W., Stelzer, E. H. K., Lindek, S. & Cremer, C. 1994 Confocal microscopy with an
increased detection aperture: type-B 4Pi confocal microscopy. Opt. Lett. 19, 222.
12.
Köhler, A. 1893 Ein neues Beleuchtungsverfahren für mikrophotographische Zwecke.
Z. Wiss. Mikrosk. 10, 433–440.
13.
Zernike, F. 1942 Phase contrast, a new method for the microscopic observation of
transparent objects. Physica 9, 686–698.
14.
Zernike, F. 1955 How I Discovered Phase Contrast. Science (80-. ). 121, 345–349.
15.
Smith, F. 1955 Microscopic interferometry. Res. 8, 385–395.
16. Nomarski, G. & Weill, A. 1955 Application à la métallographie des méthodes
interférentielles à deux ondes polarisées. Rev. Met. 2, 121–128.
17. Gerlach, J. von. 1858 Mikroskopische Studien aus dem Gebiete der menschlichen
Morphologie 72. .
18. Golgi, C. 1873 Sulla struttura della grigia del cervello. Gaz. Med. Intalianna Lomb. 6,
244–246.
19. Gram, H. 1884 Über die isolierte Färbung der Schizomyceten in Schnitt- und
Trockenpräparaten. Fortschr. Med. 2,.
20.
Stokes, G. G. 1852 On the Change of Refrangibility of Light. Philos. Trans. R. Soc.
London 142, 463–562.
21.
Baeyer, A. 1871 Über eine neue Klasse von Farbstoffen. Berichte der Dtsch. Chem.
Gesellschaft 4, 555–558.
22.
Coons, A. H., Creech, H. J. & Jones, R. N. 1941 Immunological Properties of an
Antibody Containing a Fluorescent Group. Exp. Biol. Med. 47, 200–202.
23.
Shimomura, O., Johnson, F. H. & Saiga, Y. 1962 Extraction, Purification and Properties
of Aequorin, a Bioluminescent Protein from the Luminous Hydromedusan, Aequorea.
J. Cell. Comp. Physiol. 59, 223–239.
24.
Chalfie, M., Tu, Y., Euskirchen, G., Ward, W. & Prasher, D. 1994 Green fluorescent
protein as a marker for gene expression. Science (80-. ). 263, 802–805.
25. Heim, R., Prasher, D. C. & Tsien, R. Y. 1994 Wavelength mutations and
posttranslational autoxidation of green fluorescent protein. Proc. Natl. Acad. Sci. U. S.
A. 91, 12501–4.
26. Matz, M. V et al. 1999 Fluorescent proteins from nonbioluminescent Anthozoa
species. Nat. Biotechnol. 17, 969–73.
27.
Shaner, N. C. et al. 2004 Improved monomeric red, orange and yellow fluorescent
proteins derived from Discosoma sp. red fluorescent protein. Nat. Biotechnol. 22,
1567–72.
28.
Taylor, D., Amato, P., Luby-Phelps, K. & McNeil, P. 1984 Fluorescent analog
cytochemistry. Trends Biochem. Sci. 9, 88–91.
29. Gerdes, H.-H. & Kaether, C. 1996 Green fluorescent protein: applications in cell
biology. FEBS Lett. 389, 44–47.
30. Wang, S. & Hazelrigg, T. 1994 Implications for bcd mRNA localization from spatial
distribution of exu protein in Drosophila oogenesis. Nature 369, 400–03.
31.
Förster, T. 1946 Energiewanderung und Fluoreszenz. Naturwissenschaften 33, 166–
175.
32.
Stryer, L. & Haugland, R. P. 1967 Energy transfer: a spectroscopic ruler. Proc. Natl.
Acad. Sci. U. S. A. 58, 719–726.
33.
Clegg, R. M., Feddersen, B. A., Gratton, E. & Jovin, T. M. 1992 Time-resolved imaging
fluorescence microscopy. Proc. SPIE 1640, 448–460.
34.
Tsien, R. Y. 1980 New calcium indicators and buffers with high selectivity against
magnesium and protons: design, synthesis, and properties of prototype structures.
Biochemistry 19, 2396–2404.
35.
Köhler, A. & Rohr, M., J. 1905 Photomicrography with Ultra-violet Light. J. R. Microsc.
Soc. 25, 513.
36. Heimstadt O. 1911 Das Fluoreszenzmikroskop. Z Wiss Mikrosk 28, 330–7.
37.
Ellinger, P. & Hirt, A. 1929 Mikroskopische Beobachtungen an lebenden Organen mit
Demonstrationen (Intravitalmikroskopie). Arch. für Exp. Pathol. und Pharmakologie
147, 63–63.
38.
Ploem, J. S. 1967 The use of a vertical illuminator with interchangeable dichroic
mirrors for fluorescence microscopy with incidental light. Z. Wiss. Mikrosk. 68, 129–
142.
39. Minsky, M. 1961 Microscopy apparatus.
40. Wilke, V. 1985 Optical scanning microscopy-The laser scan microscope. Scanning 7,
88–96.
41. White, J. G. 1987 An evaluation of confocal versus conventional imaging of biological
structures by fluorescence light microscopy. J. Cell Biol. 105, 41–48.
42.
Peters, R., Peters, J., Tews, K. H. & Bähr, W. 1974 A microfluorimetric study of
translational diffusion in erythrocyte membranes. Biochim. Biophys. Acta - Biomembr.
367, 282–294.
43.
Axelrod, D., Koppel, D. E., Schlessinger, J., Elson, E. & Webb, W. W. 1976 Mobility
measurement by analysis of fluorescence photobleaching recovery kinetics. Biophys.
J. 16, 1055–69.
44.
Jacobson, K., Wu, E. & Poste, G. 1976 Measurement of the translational mobility of
concanavalin A in glycerol-saline solutions and on the cell surface by fluorescence
recovery after photobleaching. Biochim. Biophys. Acta 433, 215–22.
45.
Axelrod, D. 1981 Cell-substrate contacts illuminated by total internal reflection
fluorescence. J. Cell Biol. 89, 141–145.
46. Göppert-Mayer, M. 1931 Über Elementarakte mit zwei Quantensprüngen. Ann. Phys.
401, 273–294.
47.
Kaiser, W. & Garrett, C. 1961 Two-Photon Excitation in CaF_{2}: Eu^{2+}. Phys. Rev.
Lett. 7, 229–231.
48. Denk, W., Strickler, J. & Webb, W. 1990 Two-photon laser scanning fluorescence
microscopy. Science (80-. ).
49.
Svoboda, K., Denk, W., Kleinfeld, D. & Tank, D. W. 1997 In vivo dendritic calcium
dynamics in neocortical pyramidal neurons. Nature 385, 161–5.
50.
Voie, A. H., Burns, D. H. & Spelman, F. A. 1993 Orthogonal-plane fluorescence optical
sectioning: Three-dimensional imaging of macroscopic biological specimens. J.
Microsc. 170, 229–236.
51. Huisken, J., Swoger, J., Del Bene, F., Wittbrodt, J. & Stelzer, E. H. K. 2004 Optical
sectioning deep inside live embryos by selective plane illumination microscopy.
Science 305, 1007–9.
52.
Keller, P. J., Schmidt, A. D., Wittbrodt, J. & Stelzer, E. H. K. 2008 Reconstruction of
zebrafish early embryonic development by scanned light sheet microscopy. Science
322, 1065–9.
53. Maizel, A., von Wangenheim, D., Federici, F., Haseloff, J. & Stelzer, E. H. K. 2011 High-
resolution live imaging of plant growth in near physiological bright conditions using
light sheet fluorescence microscopy. Plant J. 68, 377–85.
54. HALL, C. E. 1956 Method for the observation of macromolecules with the electron
microscope illustrated with micrographs of DNA. J. Biophys. Biochem. Cytol. 2, 625–8.
55.
ROTMAN, B. 1961 Measurement of activity of single molecules of beta-D-
galactosidase. Proc. Natl. Acad. Sci. U. S. A. 47, 1981–91.
56. Hirschfeld, T. 1976 Optical microscopic observation of single small molecules. Appl.
Opt. 15, 2965–6.
57.
Betzig, E. & Chichester, R. J. 1993 Single molecules observed by near-field scanning
optical microscopy. Science 262, 1422–5.
58.
Funatsu, T., Harada, Y., Tokunaga, M., Saito, K. & Yanagida, T. 1995 Imaging of single
fluorescent molecules and individual ATP turnovers by single myosin molecules in
aqueous solution. Nature.
59. Noji, H., Yasuda, R., Yoshida, M. & Kinosita, K. 1997 Direct observation of the rotation
of F1-ATPase. Nature 386, 299–302.
60.
Lu, H. P. 1998 Single-Molecule Enzymatic Dynamics. Science (80-. ). 282, 1877–1882.
61.
Yildiz, A. et al. 2003 Myosin V walks hand-over-hand: single fluorophore imaging with
1.5-nm localization. Science 300, 2061–5.
62.
Byassee, T., Chan, W. & Nie, S. 2000 Probing single molecules in single living cells.
Anal. Chem..
63.
Sako, Y., Minoghchi, S. & Yanagida, T. 2000 Single-molecule imaging of EGFR signalling
on the surface of living cells. Nat. Cell Biol. 2, 168–72.
64.
Reyes-Lamothe, R., Sherratt, D. J. & Leake, M. C. 2010 Stoichiometry and architecture
of active DNA replication machinery in Escherichia coli. Science 328, 498–501.
65.
Leake, M. C. et al. 2006 Stoichiometry and turnover in single, functioning membrane
protein complexes. Nature 443, 355–8.
66. Hell, S. W. & Wichmann, J. 1994 Breaking the diffraction resolution limit by
stimulated emission: stimulated-emission-depletion fluorescence microscopy. Opt.
Lett. 19, 780.
67.
Klar, T. A. & Hell, S. W. 1999 Subdiffraction resolution in far-field fluorescence
microscopy. Opt. Lett. 24, 954.
68.
Ando, R., Hama, H., Yamamoto-Hino, M., Mizuno, H. & Miyawaki, A. 2002 An optical
marker based on the UV-induced green-to-red photoconversion of a fluorescent
protein. Proc. Natl. Acad. Sci. U. S. A. 99, 12651–6.
69. Hess, S. T., Girirajan, T. P. K. & Mason, M. D. 2006 Ultra-high resolution imaging by
fluorescence photoactivation localization microscopy. Biophys. J. 91, 4258–72.
70. Miller, H., Zhaokun, Z., Wollman, A. J. M. & Leake, M. C. 2015 Superresolution
imaging of single DNA molecules using stochastic photoblinking of minor groove and
intercalating dyes. Methods In press. DOI: 10.1016/j.ymeth.2015.01.010
71.
Bailey, B., Farkas, D. L., Taylor, D. L. & Lanni, F. 1993 Enhancement of axial resolution
in fluorescence microscopy by standing-wave excitation. Nature 366, 44–8.
72.
Rust, M. J., Bates, M. & Zhuang, X. 2006 Sub-diffraction-limit imaging by stochastic
optical reconstruction microscopy (STORM). Nat. Methods 3, 793–5.
73.
Leake, M. C. et al. 2008 Variable stoichiometry of the TatA component of the twin-
arginine protein transport system observed by in vivo single-molecule imaging. Proc.
Natl. Acad. Sci. U. S. A. 105, 15376–81.
74.
Llorente-Garcia, I. et al. 2014 Single-molecule in vivo imaging of bacterial respiratory
complexes indicates delocalized oxidative phosphorylation. Biochim. Biophys. Acta
1837, 811–24.
75.
Reyes-Lamothe, R., Sherratt, D. J. & Leake, M. C. 2010 Stoichiometry and architecture
of active DNA replication machinery in Escherichia coli. Science (80-. ). 328, 498–501.
76.
Shtengel, G. et al. 2009 Interferometric fluorescent super-resolution microscopy
resolves 3D cellular ultrastructure. Proc. Natl. Acad. Sci. U. S. A. 106, 3125–30.
77.
Juette, M. F. et al. 2008 Three-dimensional sub-100 nm resolution fluorescence
microscopy of thick samples. Nat. Methods 5, 527–9.
78.
Kao, H. P. & Verkman, A. S. 1994 Tracking of single fluorescent particles in three
dimensions: use of cylindrical optics to encode particle position. Biophys. J. 67, 1291–
300.
79. Huang, B., Wang, W., Bates, M. & Zhuang, X. 2008 Three-dimensional super-
resolution imaging by stochastic optical reconstruction microscopy. Science 319, 810–
3.
80.
Pavani, S. R. P. et al. 2009 Three-dimensional, single-molecule fluorescence imaging
beyond the diffraction limit by using a double-helix point spread function. Proc. Natl.
Acad. Sci. U. S. A. 106, 2995–9.
81.
Schechner, Y., Piestun, R. & Shamir, J. 1996 Wave propagation with rotating intensity
distributions. Phys. Rev. E 54, R50–R53.
82.
Carbone, L. et al. 2013 The generation of higher-order Laguerre-Gauss optical beams
for high-precision interferometry. J. Vis. Exp.. (doi:10.3791/50564)
83.
Badieirostami, M., Lew, M. D., Thompson, M. A. & Moerner, W. E. 2010 Three-
dimensional localization precision of the double-helix point spread function versus
astigmatism and biplane. Appl. Phys. Lett. 97, 161103.
84.
Chen, B.-C. et al. 2014 Lattice light-sheet microscopy: Imaging molecules to embryos
at high spatiotemporal resolution. Science (80-. ). 346, 1257998–1257998.
85.
Booth, M. J. 2014 Adaptive optical microscopy: the ongoing quest for a perfect image.
Light Sci. Appl. 3, e165 .
86.
Chrostowski, L. 2010 Optical gratings: Nano-engineered lenses. Nat. Photonics 4,
413–415.
87.
Passilly, N. & Ventola, K. 2008 Achromatic phase retardation by subwavelength
gratings in total internal reflection. J. Opt. A: Pure Appl. Opt..
88. Marrison, J., Räty, L., Marriott, P. & O’Toole, P. 2013 Ptychographya label free, high-
contrast imaging technique for live cells using quantitative phase information. Sci.
Rep. 3, 2369 .
89.
Rodenburg, J. M., Hurst, A. C. & Cullis, A. G. 2007 Transmission microscopy without
lenses for objects of unlimited size. Ultramicroscopy 107, 227–31 .
90.
Pendry, J. 2000 Negative refraction makes a perfect lens. Phys. Rev. Lett. 85, 3966–9 .
91.
Leonhardt, U. 2009 Perfect imaging without negative refraction. New J. Phys. 11,
093040 .
92.
Cybulski, J. S., Clements, J. & Prakash, M. 2014 Foldscope: origami-based paper
microscope. PLoS One 9, e98781 .
93. Wollman, A. J. M., Miller, H., Zhaokun, Z. & Leake, M. C. 2015 Probing DNA
interactions with proteins using a single-molecule toolbox: inside the cell, in a test
tube, and in a computer. Biochem. Soc. Trans. 43, 139-145.
|
0911.5541 | 2 | 0911 | 2010-03-08T18:53:00 | Theoretical Study of Physisorption of Nucleobases on
Boron Nitride Nanotubes: A New Class of Hybrid Nano-Bio Materials | [
"physics.bio-ph",
"cond-mat.mtrl-sci",
"physics.ins-det"
] | We investigate the adsorption of the nucleic acid bases, adenine (A), guanine (G), cytosine (C), thymine (T) and uracil (U) on the outer wall of a high curvature semiconducting single-walled boron nitride nanotube (BNNT) by first principles density functional theory calculations. The calculated binding energy shows the order: G>A\approxC\approxT\approxU implying that the interaction strength of the (high-curvature) BNNT with the nucleobases, G being an exception, is nearly the same. A higher binding energy for the G-BNNT conjugate appears to result from a stronger hybridization of the molecular orbitals of G and BNNT, since the charge transfer involved in the physisorption process is insignificant. A smaller energy gap predicted for the G-BNNT conjugate relative to that of the pristine BNNT may be useful in application of this class of biofunctional materials to the design of the next generation sensing devices. | physics.bio-ph | physics | Theoretical Study of Physisorption of Nucleobases on
Boron Nitride Nanotubes: A New Class of Hybrid Nano-Bio Materials
Saikat Mukhopadhyay1, S. Gowtham1, Ralph H. Scheicher2, Ravindra Pandey1* and Shashi P. Karna3
1Department of Physics, Michigan Technological University. Houghton, MI 49931, USA.
2Department of Physics and Materials Science, Uppsala University, SE-751 21 Uppsala, Sweden.
3US Army Research Laboratory, Weapons and Materials Research Directorate, ATTN: RDRL-WM,
Aberdeen Proving Ground, MD 21005-5069, U.S.A.
March 8, 2010
To whom correspondence should be addressed: E-mail: [email protected]
Accepted in Nanotechnology (March 01, 2010)
1
Abstract
We investigate the adsorption of the nucleic acid bases - adenine (A), guanine (G), cytosine (C),
thymine (T) and uracil (U) - on the outer wall of a high curvature semiconducting single-walled
boron nitride nanotube (BNNT) by first-principles density functional theory calculations. The
calculated binding energy shows the order: G>A≈C≈T≈U implying that the interaction strength of
the high curvature BNNT with the nucleobases, G being an exception, is nearly the same. A higher
binding energy for the G-BNNT conjugate appears to result from a hybridization of the molecular
orbitals of G and BNNT. A smaller energy gap predicted for the G-BNNT conjugate relative to that
of the pristine BNNT may be useful in application of this class of biofunctional materials to the
design of the next generation sensing devices.
2
1. Introduction:
Boron nitride nanotubes (BNNTs) have been the focus of several experimental and theoretical
studies [e.g. References 1-3] due to their potential applications in high speed electronics. BNNTs
are a typical member of III-V compound semiconductors with morphology similar to that of carbon
nanotubes (CNTs) but with their own distinct properties. A tubular structure of BN can be formed
by rolling up a sheet of hexagonal rings, with boron and nitrogen in equal proportions possessing
peculiar electrical [4], optical [5] and thermal [6] properties, which drastically differ from those of
CNTs.
The nucleic acid bases, on the other hand, being the key components of the genetic macromolecules
- deoxyribonucleic acid (DNA) and ribonucleic acid (RNA) - play a central role in all biological
systems and thus have been a focus of intense research activities over the past five decades.
Recently, there has been a keen interest in understanding the interaction between nucleobases and
matter, especially nanostructured materials such as carbon nanotubes [7-16] due to the potential
application of the unique signature of the latter in probing the structural and conformational changes
[17, 18] of the former, and hence leading to new detection mechanism [19] and medical diagnostic
tools.
Very recently, a thiol-modified DNA was used to obtain high concentration BNNT aqueous
solutions assuming the interaction between DNA and multi-walled BNNT to be strong [20].
Analysis of the transmission electron microscopy measurements showed that the thiol-modified
DNA wraps around the tubular surface of BN. The tubular surface of BN consists of dissimilar
atoms and, thus, its interaction with the nucleobases may show different characteristics as compared
to that observed in case of either graphene or CNTs.
Previously, the interaction of nucleobases with graphene [7, 11] and CNTs [12] was predicted to be
dominated by van der Waals (vdW) forces as the binding energy is seen to increase with the
polarizability of the nucleobases. The charge transfer between the nucleobases and CNTs was found
to be negligible. In the present study, our motivation is to systematically investigate the self -
organization of the nucleobases onto the tubular surface of BN and identify factors playing a role on
the differences in the interaction for different base molecules. Wherever possible, we compare the
results of our study with the previous studies on CNTs.
2. Methodology:
We consider a high curvature (5, 0) single-walled BNNT of diameter of 0.416 nm, which has been
predicted to be stable by theoretical calculations [21]. All calculations were performed by
employing the plane wave pseudopotential approach within the local density approximation (LDA)
[22] of density functional theory (DFT) [23, 24]. The Vienna ab initio Simulation Package (VASP)
3
was used [25, 26] with the energy cut off of 850 eV and 0.03 eV/Å for its gradient. The periodically
repeated BNNT units were separated by 15 Å of vacuum to avoid interaction between them. The
(1x1x3) Monkhorst Pack grid [27] was used for k-point sampling of the Brillouin zone. The average
B-N bond length in the optimized configuration of the pristine BNNT is 1.44 Å, consistent with
previously reported DFT calculations [31 and references therein].
In the calculations of the energy surface describing the interaction of the nucleobases with BNNT,
the nucleobases were allowed to approach the tubular surface in the direction perpendicular to the
axis of the tube. In order to simulate an electronic environment resembling more closely the
situation in DNA and RNA, the N atom of the base molecules linked to the sugar ring in nucleic
acid was terminated with a methyl group. There is an additional benefit of introducing the small
magnitude of steric hindrance due to the attached methyl group. It will help us to imitate a more
probable situation in which a nucleobase in a strand would interact with the surface of the BNNT
rather than an isolated nucleobase interacting with BNNT. For simulations based on force fields [9-
14], it is certainly possible to include all constituents of DNA, including the sugar-phosphate
backbone. In the present first-principles study however, simulation of the nucleobases attached to
the backbone (i.e. sugar + phosphate group) is computationally rather expensive.
The optimized configurations of the nucleobases-BNNT conjugate systems were obtained following
a similar scheme as employed in the previous study of BNNT-CNT complex [12]. It consisted of (i)
an initial force relaxation calculation step to determine the preferred orientation and optimum height
of the planar base molecule relative to the surface of the BNNT (ii) calculations of the potential
energy surface [Figure 1] for nucleobase-BNNT interaction by translating the relaxed base
molecules parallel to the BNNT surface covering a surface area 4.26 Å in height, 700 in width, and
containing a mesh of 230 scan points. The separation between base molecule and the surface o f the
BNNT was held fixed at the optimum height determined in step (i). (iii) a 3600 rotation of the base
molecules in steps of 50 to probe the energy dependence on the orientation of the base molecules
with respect to the underlying BNNT surface; and (iv) a full optimization of the conjugate system in
which all atoms were free to relax.
Certainly, in a potential energy surface scan, some lateral restriction against sliding must be applied.
However, it is true that, in principle, reorientation through rotations should be considered. This was
not done here, for the following two reasons: first, regarding rotations around any axis that lies in
the plane of the nucleobase (comparable to “roll” and “pitch” for airplanes), the preferred
orientation of the nucleobase plane relative to the tube surface is parallel in order to maximize the
attractive van der Waals interaction while minimizing the repulsive interaction from overlapping
electron clouds. Second, regarding rotations around the axis that goes perpendicular through the
plane of the nucleobase (comparable to “yaw” for airplanes), at least for the larger purine base
molecules, the preferred orientation is such that their longer dimension is aligned with the tube axis
(as seen in the equilibrium geometries shown in Figure 2) again in order to maximize the attractive
interaction with the tube surface.
4
It should be pointed that LDA due to a lack of the description of dispersive forces is, in principle,
not the most optimal choice for calculating interaction energies of systems governed by vdW forces.
However, more sophisticated methods, such as many-body perturbation theory, which are more
suitable for describing long range forces, become prohibitively expensive for complex systems as
considered here. Earlier studies [28, 29] have shown that, unlike the generalized gradient
approximation (GGA) [30] for which the binding for vdW bound systems does not exist, the LDA
approximation does indeed provide reasonably good description of the dispersive interactions. Also
a recent study [7] on the adsorption of adenine on graphite suggests that the potential energy surface
obtained by using LDA and GGA with a modified version of the London dispersion formula for
vdW interactions is effectively indistinguishable. Additionally, the LDA equilibrium distance
between adenine and graphene obtained by LDA is found to be equal to that obtained using GGA +
vdW level of theory. This gives us confidence in the results obtained in the present study to be
reasonably accurate in describing nucleobase-BNNT interaction.
3. Results and Discussion:
The calculated base-BNNT binding energy, Eb, the equilibrium base-BNNT distance, and the band
gap of the corresponding base-BNNT complex are listed in Table 1. The optimized configuration of
the nucleobase-BNNT conjugates are shown in Figure 2. It should be noted that the base molecule
was allowed to approach the tubular surface along the axes perpendicular to that of the tube while
obtaining the potential energy surface. In addition to that as mentioned in the methodology section,
we have scanned the surface of the tube [step (ii)] shown in Figure 1. After that we have rotated the
base molecule to check if any particular orientation of the base molecule is preferred [optimization
step (iii)] and at the end the whole system was optimized relaxing BNNT and the nucleobases
[optimization step (iv)]. The equilibrium configurations shown in Figure 2 were the energetically
most favorable one.
None of the nucleobases show a perfect Bernal’s AB stacking. This feature matches with what was
found in the interaction of the nucleobases and SWCNT. There is, however, a slight difference in
the stacking of the nucleobases between BNNT and SWCNT, because BNNT possess a hetero -
nucleic surface unlike SWCNT. The partially negatively charged oxygen atom in guanine can
interact electrostatically with the polar network of this heteronucleic BNNT surface, specifically in
an attractive manner with the partial positive charges on boron, and repulsively with the partial
negative charges on nitrogen. This could help to explain our theoretical observation that G+BNNT
differs in several ways from A+BNNT, since in the latter combination, adenine lacks the oxygen
atom. Thus, the deviation in the stacking arrangement, the higher binding energy (see Table 1), and
the slight tilting angle could all be consequences of that interaction between the oxygen atom of
guanine and the polar network of the BNNT.
5
The nearest-neighbor distance (Rbase-BNNT) of the individual atoms of the nucleobases from the
tubular surface atoms is plotted in Figure 3 which is found to depend on the nucleobases. We note
that Rbase-BNNT is comparable to the average distance of organic molecules including amino
functional groups and 2, 4, 6-trinitrotoluene physisorbed on BNNTs [31, 32].
Figure 4 shows the energy surface representing the interaction of nucleobases with BNNT. Here, the
distance is taken to be the separation from the center of mass of the tubular configuration to the
center of mass of the nucleobases. The asymptotic limit of the energy surface is used to calculate the
binding energy (Eb) of the system (Table 1) in which the base molecule is moved away from BNNT
along the direction perpendicular to the tubular axis. The binding energy data is presented in Table
1.
Table 1: Binding energy (Eb), band gap, and nearest-neighbor distance (Rbase-BNNT) of
nucleobase conjugated BNNT. The calculated LDA band gap of the pristine (5, 0) BNNT
is 2.2 eV.
System
Eb (eV)
Rbase-BNNT (Å)
Band gap (eV)
G+BNNT
A+BNNT
C+BNNT
T+BNNT
U+BNNT
0.42
0.32
0.31
0.29
0.29
2.49
3.06
2.96
2.55
2.86
1.0
1.7
1.8
2.0
2.1
The magnitude of the calculated binding energy exhibit the following order: G>A≈C≈T≈U. It is
worth noting that the binding energy of the nucleobases interacting with a high curvature CNT [12]
followed the order of G>A>T>C≈U. Since Eb associated with CNTs was found to be correlated with
the polarizability of individual bases, it was suggested that the interaction of nucleobases with CNTs
was governed by the dispersive force like vdW which varies with the polarizability of the
interacting entity. The calculated polarizability values of G, A, C, T, U are 131.2, 123.7, 111.4,
108.5 and 97.6 e2a0
2 Eh
−1, respectively, at the Hartree-Fock level of theory together with the second
order Møller–Plesset corrections [12].
Figure 5 shows the total charge density plot of the representative conjugate system of G physisorbed
on BNNT. The Bader charge analysis does not show a noticeable charge transfer in the conjugate
system relative to the pristine BNNT and individual nucleobases; change in the total charge of the
nucleobases being quite small (< 10-2 e). This is in contrast to the cases of covalent functionalized
6
BNNTs [32-34] where a significant charge transfer of the order of 0.36 e from the organic molecule
such as NH3 and amino functional groups to BNNT was reported. Our results are consistent with
the case of 2, 4, 6-trinitrotoluene physisorbed on BNNTs reporting a very small charge transfer in
the system [31].
In order to further understand the underlying interaction between the nucleobases and BNNT, we
also calculated the polarizability of a BN sheet which comes out to be 265.7 e2a0
2 Eh
−1 at the LDA
level of theory. The polarizability of a BN sheet is therefore significan tly smaller than 402.2 e2a0
2
−1 calculated for graphene at the same level of theory. This suggests that the tubular surface of a
Eh
BNNT can be expected to be less polarizable than that of a CNT which, in turn, would lead to
relatively weaker vdW interactions between BNNT and nucleobases. This is reflected in the
calculated binding energy values of physisorbed nucleobases on BNNT which are lower in
magnitude as compared to those associated with CNTs. For example, the calculated binding energy
of G+BNNT conjugate is 0.4 eV while the corresponding value for the G+CNT conjugate is 0.5 eV.
A comparison of the present results with those from a previous study, it is found that the binding
energy of nucleobases with (7, 7) BNNT is significantly higher than that with CNTs [34]. This LDA
study using numerical atomic orbitals reported a binding energy of about 1 eV for the G+BNNT
conjugate system. This clearly suggests that the lower surface curvature of the (7, 7) BNNT (with a
diameter of 9.60 Å) leads to a stronger interaction with the nucleobases than a large surface
curvature for the (5, 0) BNNT (with a diameter of 4.16 Å) considered in the present study. A similar
trend in the effect of the surface curvature on the binding energy between nucleobase and carbon
nanostructures, graphene [11] and CNT [12] was noted in previous studies.
The semiconducting nature of BNNT with a band gap of about 2.2 eV can be seen in the calculated
density of states shown in Figure 6. This is in agreement with the recent LDA calculations on the
pristine (4, 0) BNNTs reporting a direct band gap of about 2.0 eV [36]. Both the top of the valence
edge and bottom of the conduction edge of BNNT are associated with the N p-orbitals. The
asymmetry in DOS appears to be due to the difference in coupling strength between the π-orbitals of
BNNT and the base molecule in the valance band compared to the conduction band. In the former
case contribution from the nucleobases dominates and contributions from the BNNT dominate in
the latter case. The appearance of the mid gap states (Figure 6) in the conjugated BNNT represents a
mixing of electronic states of the nucleobases and BNNT separated at about at about 2.5Å. It may
be noted that the covalent interaction at the separation of 2.5Å will be very weak [37]. On the other
hand a very small charge transfer from BNNT to oxygen of guanine may indicate the presence of a
relatively weaker electrostatic interaction between BNNT and guanine. The interaction between
BNNT and nucleobases is however essentially dominated by the vdW forces.
7
4. Summary and Conclusion.
In summary, we have investigated the interaction of the nucleobases on a high curvature, zigzag
(5,0) BNNT by first-principles DFT method. Our calculations show that, except G, the base
molecules (A, C, T, U) of DNA and RNA exhibit almost similar interaction strengths when
physisorbed on BNNT. It is also observed that the binding energy of the base molecules not only
depends upon their individual polarizability but also marginally depends on the degrees of mixing of
electronic states with the tubular surface of BNNT. The strong binding of the BNNT with G
compared to the other nucleobases suggests that this interaction can be used in sensing and also for
distinguishing this base molecule from other nucleic acid bases.
Acknowledgments
Helpful discussions with Haiying He are thankfully acknowledged. The work at Michigan
Technological University was performed under support by Army Research Office through contract
number W911NF-09-1-0221. RHS acknowledges financial support from Wenner-Gren Foundations
in Stockholm.
8
References:
[1] A. Rubio, J. L. Corkill, M. L. Cohen “Theory of graphitic boron nanotube” Phys. Rev. B 49,
5081, 1994.
[2] N. G. Chopra, R. J. Luyken , K. Cherrey , V. H. Crespi , M. L. Cohen , S. G. Louie , and A.
Zettl “Boron Nitride Nanotubes” Science 269, 966, 1995.
[3] S. Riikonen, A. S. Foster, A. V. Krasheninnikov, R. M. Nieminen “Computational study of
boron nitride nanotube synthesis: How catalyst morphology stabilizes the boron nitride bond”
Phys. Rev. B 80, 155429, 2009.
[4] R. Pati, P. Panigrahi, P. P. Pal, B. Akdim, R. Pachter “Gate field induced electronic current
modulation in a single wall boron nitride nanotube: Molecular scale field effect transistor” Chem.
Phys. Lett. 482, 312 , 2009.
[5] J. S. Lauret, R. Arenal, F. Ducastelle, A. Loiseau, M. Cau, B. Attal-Tretout, E. Rosencher, L.
Goux-Capes “Optical Transitions in Single-Wall Boron Nitride Nanotubes” Phys. Rev. Lett. 94,
037405,2005.
[6] Y. Xiao, X. H. Yan, J. Xiang, Y. L. Mao, Y. Zhang, J. X. Cao, J. W. Ding “Specific heat of
single-walled boron nitride nanotubes” Appl. Phys. Lett. 84, 4626, 2004.
[7] F. Ortmann, W. G. Schmidt, F. Bechstedt “Attracted by Long-Range Electron Correlation:
Adenine on Graphite” Phy. Rev. Lett 95, 186101, 2005.
[8] E. S. Jeng, A. E. Moll, A. C. Roy, J. B. Gastala, M. S. Strano “Detection of DNA Hybridization
Using the Near-Infrared Band-Gap Fluorescence of Single-Walled Carbon
Nanotubes” Nano Lett. 6, 371, 2006.
[9] S. Meng, P. Maragakis, C. Papaloukas, E. Kaxiras “DNA Nucleoside Interaction and
Identification with Carbon Nanotubes” Nano Lett. 7, 45, 2007.
[10] A. N. Enyashin, S. Gemming, G. Seifert “DNA-wrapped carbon nanotubes” Nanotechnology
18, 245702, 2007.
[11] S. Gowtham, R. H. Scheicher, R. Ahuja, R. Pandey, S. P. Karna “Physisorption of nucleobases
on graphene: Density-functional calculations” Phys. Rev. B 76, 033401, 2007.
[12] S. Gowtham, R. H. Scheicher, Ravindra Pandey, S. P. Karna, R. Ahuja “First-principles study
of physisorption of nucleic acid bases on small-diameter carbon nanotubes” Nanotechnology 19,
125701 , 2008.
[13] R. R. Johnson, A. T. C. Johnson, M. L. Klein “Probing the Structure of DNA-Carbon
Nanotube Hybrids with Molecular Dynamics” Nano Lett. 8, 69, 2008.
9
[14] A. Das, A. K. Sood, P. K. Maiti, M. Das, R. Varadarajan, C. N. R. Rao “Binding of
nucleobases with single-walled carbon nanotubes: Theory and experiment” Chem. Phys. Lett. 453,
266, 2008.
[15] R. R. Johnson, A. Kohlmeyer, A. T. C. Johnson, M. L. Klein “Free Energy Landscape of a
DNA-Carbon Nanotube Hybrid Using Replica Exchange Molecular Dynamics” Nano Lett. 9, 537,
2009.
[16] X. M. Tu, S. Manohar, A. Jagota, M. Zheng “DNA sequence motifs for structure-specific
recognition and separation of carbon nanotubes” Nature 460, 250, 2009.
[17] V. G. Vaidyanathan, B. U. Nair “Synthesis, characterization and binding studies of
chromium(III) complex containing an intercalating ligand with DNA” J. Inorg. Biochemistry 95,
334, 2003.
[18] A. Rajendran , C. J. Magesh, P. T. Perumal “DNA–DNA cross-linking mediated by bifunctional
[SalenAlIII]+ complex” Biochimica et Biophysica Acta, 1780, 282,2008.
[19] D. A. Heller, E. S. Jeng, T. K. Yeung, B. M. Martinez, A. E. Moll, J. B. Gastala, M. S. Strano
“Optical Detection of DNA Conformational Polymorphism on Single-Walled Carbon Nanotubes”
Science 311, 508, 2006.
[20] C. Zhi, Y. Bando, W. Wang, C. Tang, H. Kuwahara, D. Golberg “DNA-Mediated Assembly of
Boron Nitride Nanotubes” Chem. Asian J. 2, 1581 (2007).
[21] H. J. Xiang, J. Yang, J. G. Hou, Q. Zhu “First-principles study of small-radius single-walled
BN nanotubes” Phys Rev B 68, 035427, 2003.
[22] J. P. Perdew, A. Zunger “ Self-interaction correction to density-functional approximations for
many-electron systems” Phys. Rev B 23, 5048, 1981.
[23] P. Hohenberg, W. Kohn “Inhomogeneous Electron Gas” Phys. Rev. 136, B865, 1964.
[24] W. Kohn, L. J. Sham “ Self-Consistent Equations Including Exchange and Correlation Effects”
Phys. Rev. 140, A1133, 1965.
[25] G. Kresse, J. Furthmüller “Efficiency of ab-initio total energy calculations for metals and
semiconductors using a plane-wave basis set” Comput Mat Sci 6 , 15 1996.
[26] G. Kresse, D. Joubert “From ultrasoft pseudopotentials to the projector augmented-wave
method”Phys. Rev. B 59, 1758, 1999.
[27] H. J. Monkhorst, J. D. Pack “ Special points for Brillouin-zone integrations” Phys. Rev. B 13,
5188, 1976.
10
[28] M. Simeoni, C. D. Luca, S. Picozzi, S. Santucci, B, Delley “Interaction between zigzag single-
wall carbon nanotubes and polymers: A density-functional study” J. Chem. Phys. 122, 214710,
2005.
[29] F. Tournus, S. Latil, M. I. Heggie, J. C. Charlier “π- stacking interaction between carbon
nanotubes and organic molecules” Phys. Rev. B 72, 075431, 2005.
[30] J. P. Perdew, J.A, Chevary, S. H. Vosko, K. A. Jackson, M. R. Pederson, D. J. Sing, C. Fiolhais
“Atoms, molecules, solids, and surfaces: Applications of the generalized gradient approximation for
exchange and correlation” Phys. Rev. B 46, 6671,1992.
[31] B. Akdim, S. N. Kim, R. R. Naik, B. Maruyama, M. J. Pender, R Pachter “Understanding
effects of molecular adsorption at a single-wall boron nitride nanotube interface from density
functional theory calculations” Nanotechnology 20, 355705, 2009.
[32] X. Wu, W. An, X. C. Zeng “Chemical Functionalization of Boron-Nitride Nanotubes with
NH3 and Amino Functional Group ” J. Am. Chem. Soc. 128, 12001.2006.
[33] C. Zhi, Y. Bando, C. Tang, S. Honda, K. Sato, H. Kuwahara, D. Golberg “Covalent
Functionalization: Towards Soluble Multiwalled Boron Nitride Nanotubes” Angew. Chem., Int. Ed.
44, 7932, 2005.
[34] J. Zheng, W. Song, L. Wang, J. Lu, G. Luo, J. Zhou, R. Qin, H. Li, Z. Gao, L. Lai, G. Li, W. N.
Mei “Adsorption of Nucleic Acid Bases and Amino Acids on Single-Walled Carbon and Boron
Nitride Nanotubes: A First-Principles Study ” J. Nanoscience and Nanotechnology 9, 6376, 2009.
[35] C. Zhi, Y. Bando, C. Tang, D. Golberg “Engineering of electronic structure of boron-nitride
nanotubes by covalent functionalization” Phys. Rev. B 74, 153413 (2006).
[36] Z. Zhang, W. Guo, Y. Dai “Stability and electronic properties of small boron nitride
nanotubes.” J. Appl. Phys 105, 84312 ,2009.
[37] S. J. Grabowski, W. A. Sokalski, E. Dyguda, J. Leszczyński “Quantitative Classification of
Covalent and Noncovalent H-Bonds” J. Phys. Chem. B 110, 6444, 2006.
11
Figure 1: Potential energy surface plot (eV) for guanine scanning the surface of BNNT. The
specifications of the Scanned area (rectangle highlighted by dark blue) are shown above. The energy
barrier between two adjacent global minima is 0.15 eV as indicated by the arrows. Qualitatively similar
features were found in case of the other four nucleobases.
12
Figure 2: Equilibrium geometry of physisorbed nucleobases on the surface BNNT. (a) guanine,
(b) adenine, (c) cytosine, (d) thymine and (e) uracil. The pink, blue, green, red and white colors
represent B, N, C, O and H, respectively.
13
Figure 3: The distance between the nucleobases atoms and the tubular surface atoms in the
equilibrium configurations of BNNT conjugates.
14
23456 Distance (Å) GACTU
Figure 4: (color online) The potential energy variation of the nucleobases interacting with BNNT
as a function of the distance. The distance represents the separation between the center of the
mass of the tubular surface and that of the base. A, G, C, T, U are represented by black, red, blue,
green and pink lines, respectively. Zero of the energy is aligned to the non-interacting regime of
the surface.
15
Figure 5: Total charge density of guanine (above) and thymine (below) conjugated BNNT. The
isosurface levels were set at 0.08 bohr -3. The pink, blue, green, red and white sticks represent B,
N, C, O and H, respectively.
16
Figure 6: Density of states of a pristine BNNT and guanine and thymine conjugated BNNT.
Zero of the energy is aligned to the top of the valence band.
17
-5050204060BNNTGBNNTTBNNTEnergy(eV)DOS(arb. unit) BNNT G+BNNT T+BNNT 5 0 -5
18
|
1902.02296 | 1 | 1902 | 2019-02-06T17:41:01 | Stick-Slip Dynamics of Migrating Cells on Viscoelastic Substrates | [
"physics.bio-ph",
"cond-mat.soft",
"nlin.AO",
"q-bio.CB"
] | Stick-slip motion, a common phenomenon observed during crawling of cells, is found to be strongly sensitive to the substrate stiffness. Stick-slip behaviours have previously been investigated typically using purely elastic substrates. For a more realistic understanding of this phenomenon, we propose a theoretical model to study the dynamics on a viscoelastic substrate. Our model based on a reaction-diffusion framework, incorporates known important interactions such as retrograde flow of actin, myosin contractility, force dependent assembly and disassembly of focal adhesions coupled with cell-substrate interaction. We show that consideration of a viscoelastic substrate not only captures the usually observed stick-slip jumps, but also predicts the existence of an optimal substrate viscosity corresponding to maximum traction force and minimum retrograde flow which was hitherto unexplored. Moreover, our theory predicts the time evolution of individual bond force that characterizes the stick-slip patterns on soft versus stiff substrates. Our analysis also elucidates how the duration of the stick-slip cycles are affected by various cellular parameters. | physics.bio-ph | physics |
Stick-Slip Dynamics of Migrating Cells on Viscoelastic Substrates
Partho Sakha De and Rumi De∗
Department of Physical Sciences,
Indian Institute of Science Education and Research Kolkata,
Mohanpur 741246, West Bengal, India
Stick-slip motion, a common phenomenon observed during crawling of cells, is found to be strongly
sensitive to the substrate stiffness. Stick-slip behaviours have previously been investigated typically
using purely elastic substrates. For a more realistic understanding of this phenomenon, we pro-
pose a theoretical model to study the dynamics on a viscoelastic substrate. Our model based on
a reaction-diffusion framework, incorporates known important interactions such as retrograde flow
of actin, myosin contractility, force dependent assembly and disassembly of focal adhesions coupled
with cell-substrate interaction. We show that consideration of a viscoelastic substrate not only
captures the usually observed stick-slip jumps, but also predicts the existence of an optimal sub-
strate viscosity corresponding to maximum traction force and minimum retrograde flow which was
hitherto unexplored. Moreover, our theory predicts the time evolution of individual bond force that
characterizes the stick-slip patterns on soft versus stiff substrates. Our analysis also elucidates how
the duration of the stick-slip cycles are affected by various cellular parameters.
I.
INTRODUCTION
Cell motility plays a key role in many important bi-
ological processes such as wound healing, morphogene-
sis, embryonic development, tissue regeneration to name
a few [1 -- 5]. Though motility is expressed in multiple
ways, crawling happens to be the most common form
of movement for eukaryotic cells. During crawling, cell
forms protrusions at the leading edge which pushes the
membrane forward and as a consequence the membrane
exerts a backward force on the polymerising actin fila-
ments, resulting in them 'slipping' rearward towards the
cell center, in a process known as retrograde flow [6 -- 8].
This process is accompanied by growth and strengthen-
ing of focal adhesions between the cell and the substrate
which slow down the actin retrograde flow [9]. Thus,
it allows actin polymerization to advance at the leading
edge and in turn, the rate of translocation of the cell
increases [1, 8, 10, 11]. The dynamic variation in retro-
grade flow coordinated with assembly and disassembly of
focal adhesions lead to stick-slip motion.
Stick-slip is a kind of jerky motion that has been found
not only in living systems but also in passive systems such
as peeling of scotch tapes [12 -- 14], earthquakes [13, 15, 16]
to name a few. Stick-slip behaviour is characterised by
the system spending most of it's time in the 'stuck' state
and comparatively a short time in the 'slip' state. During
crawling of cells, stick-slip dynamics has been experimen-
tally observed on multiple occasions. Experiments on mi-
grating epithelial cells showed that in the lamellipodium
region the traction force decreases with increasing ve-
locity inferring a stick-slip regime of the actin-adhesion
interaction [17]. Stick slip motion has also been observed
in embryonic chick forebrain neurons [18], migrating hu-
man glioma cells [19] and also in human osteosarcoma
∗ Corresponding author: [email protected]
cells [20]. In case of fish keratocytes, stick-slip kind of be-
haviours have been found in modulation of the cell shape
during crawling [21 -- 23]. Moreover, recent experiments
show that the stick-slip dynamics is strongly affected by
altering the substrate stiffness. For example, a close
inspection of the leading-edge motion of crawling and
spreading mouse embryonic fibroblasts revealed that the
periodic lamellipodial contractions are vastly substrate-
dependent [24]. Further, in recent studies, cell motility
found to be maximum and actin flow rate minimum at
an optimal substrate stiffness [18, 19, 25 -- 27].
There are many theoretical studies that have con-
tributed significantly to understand the cell migration
process [28 -- 35]. Quite a few models have also been pro-
posed to unravel the stick-slip mechanisms. Such as,
Barnhart et. al. deveoped a mechanical model to pre-
dict periodic shape change during keratocyte migration
caused by alternating stick-slip motion at opposite sides
of the cell trailing edge [23]. Also, leading edge dynam-
ics, spatial distribution of actin flow, and demarkation
of lamellipodium-lamellum boundary have been studied
[31, 32]. Besides, simple stochastic models have provided
a great deal of information on cell crawling. Stochastic
bond dynamics integrated with traction stress dependent
retrograde actin flow could capture the biphasic stick-slip
force velocity relation [36, 37]. There are other models
based on stochastic motor-clutch mechanisms which have
provided many insights into substrate stiffness dependent
migration process [18, 19, 27, 38, 39]. As observed in
experiments, these studies reveal the existence of an op-
timal substrate stiffness which found to be sensitive to
cell motor-clutch parameters. However, most of these
studies on rigidity sensing so far are either focused on
purely elastic substrate [19, 38] or on purely viscous sub-
strate [40]; whereas, physiological extracellular matrix is
viscoelastic in nature. Also, recent experimental studies
further reveal that the dynamics is greatly sensitive to
the substrate viscosity [40 -- 42].
In this paper, we present a theoretical model of the
leading edge dynamics of crawling cells on a viscoelastic
substrate. Our theory based on a framework of reaction-
diffusion equations takes into account the retrogate flow
of actin, myosin contractility, force dependent assembly
and disassembly of focal adhesions integrated with cell-
substrate interaction. The model predicts how these cel-
lular components work together to give rise to the sponta-
neous emergence of stick-slip jumps as observed in experi-
ments. More importantly, it elucidates the effect of varia-
tion of substrate viscoelasticity on the 'stick-slip' dynam-
ics. Interestingly, it predicts, the existence of an optimal
substrate viscosity corresponding to maximum traction
force and minimum retrograde flow as observed in case
of elastic substrate [18, 19, 38]. Moreover, our contin-
uum model framework captures the time evolution of in-
dividual bond force that has remained unexplored so far.
These findings suggest that the nature of non-trivial force
loading rate of individual bonds play a crucial role in de-
termining the stick-slip jumps and thus explain the dis-
tinctive patterns on varying substrate compliance. Our
theory also provides an analytical understanding of how
the cellular parameters such as substrate stiffness, myosin
activity, retrograde flow affect the duration of the stick-
slip cycles.
II. THEORETICAL MODEL
Our theory is based on the molecular clutch mecha-
nisms proposed to describe the transmission of force from
actin cytokeleton to extra cellular matrix [10, 18, 39].
The clutch module or the 'connector' proteins provide a
dynamic link between F-actin and adhesion complexes
and slow down the F-actin retrograde flow [1, 6, 8]. Our
model consists of free receptors representing these 'con-
nector' proteins diffusing in the actin cytoplasm. The
substrate consists of a large number of adhesive ligands
which can bind with these free receptors to form closed
bonds as illustrated in Fig. 1. Thus, the receptors are
considered to be either in free or bound states denot-
ing open or closed ligand-receptor adhesion bonds. The
ligand-receptor bonds are modelled as Hookean springs
of stiffness Kc. As the F-actin bundle (modelled as a
rigid rod), pulled by the myosin motors [9], moves with
the retrograde velocity, vm, the spring gets stretched
and thus, the force on a single bond is given by mul-
tiplying the spring stiffness with the bond elongation as,
f = Kc (xb − xsub); where xb is the displacement of one
end of the bond attached to the actin bundle and xsub is
the displacement of the substrate (where the other end
of the bond is attached). The retrograde flow velocity
of the F-actin bundle slows down with increase in the
force on the closed bonds and is given by the relation,
F total
b
vm = v0(cid:16)1 −
Fstall(cid:17); where v0 is the unloaded velocity,
F total
is the total traction force due to all closed bonds,
and Fstall is the total force exerted by myosin motors.
b
2
free receptor
bound receptor
motor protein
ligand
actin bundle
v
m
koff
xb-xsub
kon
Ksub
γ
FIG. 1. Schematic diagram of the model: free receptors (de-
noted by light circles) diffuse within the actin cytoplasm. The
free receptors bind with ligands on the substrate to form the
bound receptors (dark circles), that forms the ligand-receptor
bonds. The F-actin filament is pulled by myosin motors (ma-
roon structures) with the retrograde flow velocity, vm. Vis-
coelastic substrate is modelled by a spring and a dashpot.
(color online)
Here, Fstall = nm ∗Fm, where nm is the number of myosin
motors present and Fm is the force exerted by an indi-
vidual myosin motor [9, 17, 18, 39].
We express the reaction between the free receptors and
the ligands to form the bound receptors as
Rf + L
kon(f )
−−−−−⇀↽−−−−−
koff (f )
Rc
[ρf ]
[ρL]
[ρc]
where ρf , ρc, and ρL denote the densities of free re-
ceptors, bound receptors, and ligands on the substrate
respectively. (The total number of free and bound recep-
0 (ρc + ρf ) dx = Nt
(constant). Now, the time evolution of the density of
the free and the bound receptors are described by the
following coupled reaction-diffusion equations,
tors is taken to be constant, i.e., R L
∂ρf
∂t
∂ρc
∂t
= D
= vm
∂ 2ρf
∂x2
∂ρc
∂x
− kon(f )ρf + koff (f )ρc
+ kon(f )ρf − koff (f )ρc.
(1)
(2)
Here, the first term on the R.H.S of Eq. 1 repre-
sents the diffusion of free receptors in the actin cyto-
plasm. The last two terms are the reaction terms of
formation of bound receptors and free receptors with re-
spective reaction rates. For Eq. 2, the first R.H.S term
denotes the drift of the bound receptors with the retro-
grade flow velocity, vm, as they are attached to F-actin
bundle, and the other two terms are the reaction terms
as in Eq. 1. In our model, motivated by the experimen-
tal findings, the association and the dissociation rates,
kon(f ) and koff (f ), are considered to be force dependent
[43 -- 45]. We have taken, kon(f ) = k0
on + g(f )ρc; where,
k0
on is the rate constant, and g(f ) is a function of the
bond force, f ; for sake of simplicity it is taken to be lin-
ear, g(f ) = ξf . Thus, the force increases the binding
rate and allows for the formation of new bonds; thus, ef-
fectively strengthening the adhesion cluster. Moreover,
it has been observed that the force, upto an optimal
value, helps strengthen the existing focal adhesion bonds
and these force-strengthening molecular bonds are called
catch bonds [46, 47]. In our model, the dissociation rate
of the closed bonds is considered to demonstrate catch
behavior as [48],koff = ksef /fs + kce−f /fc ., here k0, ks,
and kc are the rate constants.
Moreover, in our theory, the substrate is considered
to be viscoelastic in nature and has been modelled as
a spring-dash pot system with spring stiffness Ksub and
viscosity γ as shown in Fig. 1. Now, the equation of
motion for the substrate is obtained by balancing the
total force experienced by all the bonds with the sum of
the elastic force (Ksubxsub) and the viscous drag (γ xsub)
of the substrate,
γ xsub + Ksubxsub = F total
b
;
here the total traction force is given as, F total
0 f ρc (x) dx.
b
R L
III. DIMENSIONLESS FORMULATION
(3)
=
, nc = ρc
ρ0
LR L
We study the dynamics in dimensionless units. The
densities have been scaled as: nf = ρf
, where
ρ0
ρ0 is the average density of the receptors and is defined
as, ρ0 = 1
0 (ρc + ρf ) dx. The dimensionless time is de-
fined as τ = k0t, where k0
on is expressed as αk0. Thus,
the dimensionless binding and unbinding rates are, kon =
α + ξ f nc and koff = ks
, f = f
where ξ = ξfsρ0
k0
fs
and fc = fc
. The position coordinate is scaled as
fs
X = x
x0
are: D = D
k0x2
Ksub = Ksub
Kc
. Other dimensionless variables
; Xsub = xsub
;
x0
off exp(cid:16)− f / fc(cid:17),
, where x0 = fs
Kc
, vm = vm
x0k0
, γ = γ k0x0
fs
off exp(cid:16) f(cid:17) + kc
, kc
off = kc
off = ks
; and F total
b
= F total
b
fs
.
off
k0
, ks
off
k0
0
Thus, the scaled equations of motion turn out to be,
∂nf
∂τ
∂nc
∂τ
and
= D
= vm
∂ 2nf
∂X 2
∂nc
∂X
− konnf + koff nc
+ konnf − koff nc
γ Xsub + Ksub Xsub = F total
b
.
(4)
(5)
(6)
3
IV. RESULTS
off = 120, ks
We have investigated the stick-slip dynamics by solving
the coupled reaction-diffusion Eqs. 4-6 numerically. The
equations are discretized using finite difference method
on a grid of size N and then solved by fourth order
Runge-Kutta method. The boundary conditions are
taken to be such that the total number of free and bound
receptors present in the system is conserved. We have
studied the dynamics for a wide range of parameter val-
ues by varying system size, number of myosin motor, ret-
rograde flow velocity, binding rates, substrate stiffness
and viscosity. Here, we present the result for a represen-
tative parameter set, where values of the force dependent
rate constants are kept at α = 2, ξ = 1, kc
off =
0.25 and fc = 0.5; also, the unloaded velocity v0 = 120,
the diffusion constant D = 5, the stall force Fm = 2,
number of myosin motors nm = 100, and the system size
N = 100. The substrate viscosity γ and rigidity Ksub
remain as variable parameters. We have also varied the
diffusion constant, D; the system reaches to the steady
state faster with a higher value of diffusion constant, how-
ever, the stick-slip dynamics remain the same. More-
over, we note that even though the model parameters
are scaled and dimensionless, nonetheless their values are
taken from experiments, e.g., the dissociation rate con-
stants are taken as kc
off = 0.25 s−1 [48],
whereas, unloaded myosin motor stall force is Fm = 2 pN
and unloaded retrograde flow velocity is v0 = 120 nm s−1
[18, 38]. Also, the variation of substrate elastic stiffness
is considered as Ksub ∼ 0.01 − 100 pN nm−1 [18, 38], and
the range of substrate viscosity is γ ∼ 0.01−10pN.s nm−1
as observed in experiments [40].
off = 120 s−1 and ks
A. Stick-slip dynamics: dependence on substrate
stiffness
Figures 2 a-d show the time evolution of single bond
force, total number of bonds, total traction force and cor-
responding retrograde flow velocity on a soft substrate.
Soft substrates are very compliant and deform easily,
thus, the build up of force, f , on an individual bond is
also slow as shown in Fig. 2 a. Now, the increase in bond
force, f , increases the binding rate, kon, of the free re-
ceptors and at the same time, decreases the dissociation
rate, koff , of the bound receptors upto an optimal force
value due to catch bond behaviour. As a result, initially a
large concentration of receptors are bound to the ligands
on the substrate (shown in Fig. 2 b). As the density of
the bound receptors increases, they share the total trac-
tion force exerted by the substrate. Thus, the traction
force slowly grows with time (Fig. 2 c) as the substrate
gets deformed. The growth of traction force in turn slows
down the retrograde flow of actin (Fig. 2 d). This gives
rise to the 'stuck' state which allow actin polymerization
to advance the leading edge of the cell. But as the force
increases even further, the dissociation rate starts to in-
50
40
30
20
10
s
d
n
o
B
d
e
s
o
l
C
20
40
Time
60
0
0
(b)
120
20
40
Time
60
e
c
r
o
F
d
n
o
B
8
6
4
2
0
(a)
40
0
0.1
0.2
0.3
Time
w
o
l
F
e
d
a
r
g
o
r
t
e
R
100
80
60
e
c
r
o
F
n
o
i
t
c
a
r
T
30
20
10
4
s
d
n
o
B
d
e
s
o
l
C
8
6
4
2
0
(b)
120
w
o
l
115
0
0.1
0.2
0.3
Time
F
e
d
a
r
g
o
r
t
e
R
110
105
100
20
40
Time
60
40
0
(d)
20
40
Time
60
0
0
(c)
0.1
0.2
0.3
Time
95
0
(d)
0.1
0.2
0.3
Time
12
e
c
r
o
F
d
n
o
B
8
4
0
0
(a)
120
e
c
r
o
F
n
o
i
t
c
a
r
T
80
40
0
0
(c)
FIG. 2. Stick-slip motion on soft substrate. (a) Time evolu-
tion of single bond force, f . (b) time evolution of the total
number of bonds, Nc = R L
0 ncdX. (c) corresponding evolution
of the traction force, F total
, and (d) the retrograde flow ve-
locity, vm, mirrors the effect of the slowing down of the actin
flow with increase in the traction force. (Keeping Ksub = 0.1
and γ = 0.01, all values are dimensionless.)
b
FIG. 3. Stick-slip motion on stiff substrate. (a) Evolution of
single bond force, f , as a function of time. (b) Time evolu-
tion of the total number of bonds, Nc = R L
0 ncdX, (c) corre-
sponding evolution of the traction force, F total
, and (d) the
b
retrograde flow velocity. ( Keeping Ksub = 100 and γ = 0.01.)
crease. Then, the linearly growing binding rate can no
longer keep up and falls below the much faster growing
dissociation rate and thus, the adhesion cluster dissociate
very quickly and the number of closed bonds decreases.
The dissociation of bonds increases the effective force on
the remaining bound receptors/bonds as less number of
bonds have to share the currently high value of traction
force. This increases the dissociation rate even further
and as a result of this feedback cycle, the bound recep-
tors dissociate very quickly. The quick dissociation of
bonds means that there is nothing to holding on or an-
chor to the substrate and thus, the retrograde velocity
increases rapidly during what is known as the 'slip' state
as could be seen from Fig. 2 d, and the stick-slip cycle
thus repeats.
On the other hand, in case of stiff substrate, as the
substrate does not deform easily, force on an individual
bond increases very quickly due to actin retrograde ve-
locity as shown in Fig. 3 a. However, within this little
time, due to lack of sufficient binding time, the forma-
tion of bound receptors/bonds happens to be very small
as could be seen from Fig. 3 b. Now, as these small num-
ber of closed bonds are to share the total traction force
exerted by the substrate, this makes the force on a sin-
gle bond to increase even faster within a short time (Fig.
3 a). As a result, the exponentially varying dissocia-
tion rate dominates over linearly increasing binding rate;
hence, the adhesions start dissociating even before the
substrate has a substantial deformation and the traction
force has attained a high value (Fig. 3 c). Lower traction
force also results in a higher retrograde flow rate (shown
in Fig. 3 d) as the actin filament slips backward faster
and the cell thus is unable to effectively transmit forces
to the stiff substrate compared to a soft substrate.
We further investigate the stick-slip behaviours as a
function of varying substrate stiffness. As observed in
earlier studies [18, 19, 38], our simulations show that
there exists an optimal substrate stiffness, where the
mean value of the total traction force is maximum and
the retrograde flow is minimum as shown in Fig. 4. This
could be attributed to the difference in the nature of in-
crease of force of an individual bond depending on the
substrate stiffness as evidently seen from Fig 2 a and Fig
3 a. On a stiff substrate, since the single bond force in-
creases rapidly, only limited number of bound receptors
could form within that short time. Moreover, fast in-
creasing force shortened the life time of the bonds; thus,
resulting in lower total traction force and higher retro-
grade flow. However, on a softer substrate, as the sub-
strate deforms easily, bond force increases slowly which
allows for the formation of more bound receptors. Thus,
the higher density of adhesion bonds results in higher
value of traction force that resists the actin flow and thus,
decreases retrograde velocity. As the substrate is made
even more softer, the traction force and consecutively the
force on individual bonds grows very slowly, due to the
extreme compliance of the substrate. As a result, the
system spends a large amount of time in a state of expe-
w
o
l
F
e
d
a
r
g
o
r
t
e
R
n
a
e
M
110
100
90
80
70
v-
m
total
b
-
F
0.1
1
10
100
Substrate Elastic Stiffness
80
60
40
20
0
1000
e
c
r
o
F
n
o
i
t
c
a
r
T
n
a
e
M
5
Numerical
Solution
Analytical
Prediction
n
o
i
t
a
r
u
D
e
l
c
y
C
p
i
l
s
-
k
c
i
t
S
20
15
10
5
0
0.1
10
1
100
Substrate Rigidity
1000
b
FIG. 4. Mean value of the retrograde flow velocity ¯vm ( ) and
thetraction force ¯F total
( ) avaraged over the stick-slip cycle
as a function of substrate stiffness, Ksub (keeping γ = 0.01
to a constant value). Average traction force is maximum and
retrograde flow is minimum for an optimal value of substrate
stiffness as observed in experiments [18, 19, 38].
FIG. 5. Duration of the stick-slip cycle as a function of sub-
strate stiffness, Ksub . Numerical results agree well with the
analytical prediction of the cycle duration, given by the char-
acteristic time, τc, obtained from Eq. 9 (parameters are kept
at same value as of numerical simulation.)
riencing low traction force and higher retrograde velocity.
This reduces the mean value of the traction force for very
compliant substrates to some extent and thus increases
the mean retrograde velocity.
b
Our theory further elucidates how the variation of sub-
strate stiffness affects the duration of a stick-slip cycle.
We obtain an analytic expression of the time evolution of
the total traction force, F total
(τ ) through a rudimentary
calculation along with a few approximations (see sup-
porting material). At any instant, the total traction force
due to the deformation of the substrate must be balanced
by summing over forces of all bound receptors/bonds.
Following, the evolution of the total traction force dur-
ing a stick-slip cycle (starting from a value, F total
= 0 at
time τ = 0) is given by
b
F total
b
(τ ) = Fstall
Eq. 7 can be rewritten as,
F total
b
KcNcv0 Ksub
Fstall(cid:16) Ksub + KcNc(cid:17)
1 − exp
−
(τ ) = Fstall(cid:20)1 − exp(cid:18)−
τ
τc(cid:19)(cid:21) ;
where the time constant, τc, is of the form,
τc =
Fstall(cid:16) Ksub + KcNc(cid:17)
KcNcv0 Ksub
= A +
B
Ksub
.
τ
(7)
.
(8)
(9)
The characteristic time, τc, is a measure of the growth
rate of the total traction force and consequently, slow-
ing down of actin retrograde flow. Therefore, it denotes
the time scale corresponding to 'stuck' state which occu-
pies the majority of the stick-slip cycle duration. As the
slip state duration is very small compared to stuck state,
the variation of τc provides some insights into the dura-
tion of the stick-slip cycle on various system parameters,
namely, substrate stiffness, bond stiffness, retrograde ve-
locity, myosin activity, and the system size as described
by Eq. 9. Moreover, we also numerically compute the
duration of the stick-slip cycle as a function of substrate
stiffness, Ksub, as shown in Fig. 5. Our theoretical pre-
diction (approximated by Eq. 9) matches quite well with
the stick-slip cycle duration obtained from simulation re-
sults, as seen from the figure Fig. 5.
B. Effect of the substrate viscosity on the dynamics
Recent experiments have shown that how cells spread,
adhere, migrate or modulate their contractile activity
vary with the extracellular matrix depending on whether
it is elastic or viscoelastic in nature [40 -- 42, 49, 50]. More-
over, the substrate stress relaxation is controlled by the
viscosity; thus, it alters the cellular force transmission
process and so the overall response of cells. Here, we in-
vestigate how the presence of substrate viscosity affects
the stick-slip behaviour of crawling cells.
As shown in Fig. 6 a, increase in viscosity, γ, increases
the effective traction force of the compliant substrate and
shifts the optimal substrate stiffness for which the max-
imum traction occurs to a lower value towards softer,
Ksub. However, for stiff substrates as observed, the pres-
ence of viscosity does not cause any noticeable difference
in the dynamics because the individual bonds already
experience a rapid building of tension due to high elas-
tic stiffness and therefore, destabilize quickly resulting in
higher retrograde flow and low stick-slip cycle duration.
Thus, at higher stiffness regime, the stick-slip dynamics
remain unaltered for elastic and viscoelastic substrate.
0.1
0.2
0.3
0.4
100
e
c
r
o
F
n
o
i
t
c
a
r
T
n
a
e
80
60
40
20
M
0
0.1
100 1000
Substrate Elastic Stiffness
(a)
10
1
110
w
o
l
100
F
e
d
a
r
g
o
r
t
e
R
n
a
e
M
60
(b)
90
80
70
0.001
0.01
0.1
1
10
100
200
500
1000
w
o
l
F
e
d
a
r
g
o
r
t
e
R
n
a
e
M
110
100
90
80
70
100
200
500
1000
w
o
l
110
100
F
e
d
a
r
g
o
r
t
e
R
n
a
e
M
90
80
70
60
(b)
6
10
0.01
0.1
1
Substrate viscosity
10
(a)
0.1
100 1000
Substrate Elastic Stiffness
10
1
0.01
0.1
1
Substrate viscosity
(a) Variation of mean traction force, F total
FIG. 6.
, as a
function of substrate stiffness, Ksub, for different values of
substrate viscosity: γ = 0.1 ( ), 0.2 ( ), 0.3 ( ), 0.4 ( ) .
(b) Effect of mean retrograde velocity on varying substrate
viscosity keeping the elastic stiffness, Ksub, at a fixed value for
different values of elastic stiffness: Ksub = 0.001 ( ), 0.01( ),
0.1( ), 1( ), 10( ).
b
Interestingly, our theory predicts the existence of an
optimal substrate viscosity corresponding to maximum
traction force and minimum retrograde flow as observed
in case of purely elastic substrate. As seen from Fig. 6
b, keeping the substrate elastic stiffness, Ksub, at a con-
stant value and increasing viscosity, γ, increases effective
traction force, thus, reducing the actin retrograde flow.
However, increasing the viscosity further beyond an op-
timal value, traction force start decreasing and hence,
increases the retrograde velocity. As the viscous drag
resists the deformation of the substrate, the overall stiff-
ness of the substrate increases with increasing viscosity.
At a lower substrate viscosity, since the substrate relaxes
faster, force on individual bonds grow slowly allowing
longer interaction time with the substrate and formation
of more bonds. As the adhesion cluster grows total trac-
tion force increases and hence slows down the retrograde
flow. On the other hand, at higher substrate viscosity,
as the substrate relax slowly, that causes the bond force
to rise quickly without providing enough time for the
formation of new bonds to share the traction force. As
the bond force increases faster, it destabilize the adhe-
sion cluster resulting in lower traction force and higher
retrograde velocity as seen from Fig. 6 b.
C. Shifting of optimal stiffness
We now study how the system parameters such as
binding rates, size of the adhesion patch, number of
myosin motors affect the optimal substrate viscosity and
also compare with that of elastic substrate. It is observed
that to exhibit stick-slip behaviour, the pulling force on
the F-actin bundle by the myosin motors must be bal-
anced by the total force of the ligand-receptor bonds.
However, if the total number of available receptors is too
less compared to the number of myosin motors, then it
FIG. 7. Variation in the total number of myosin motors and
the size of the adhesion cluster causes shift in (a) the opti-
mal substrate stiffness and also in (b) the optimal viscosity.
(Number of motors, nm and total number of receptors, Nt
have been varied equally.) nm = Nt = 100( ), 200( ), 500( ),
1000( )
results in perpetually 'slipping' mode, with F-actin fil-
ament moving at near it's unloaded velocity, v0. The
reverse scenario can also take place where the number of
bonds is much higher compared to the number of myosin
motors, so that it slows down the retrograde velocity to
zero, thus resulting in a permanently 'stuck' state. How-
ever, if the number of myosin motors and the number
of bonds, i.e., the size of the ligand-receptor adhesive
patch are varied appropriately, the stick-slip dynamics is
restored.
In our simulation, the simultaneous change in number
of motors and the size of the adhesive patch is taken into
consideration by changing the grid size N and also mod-
ifying Fstall = nm ∗ Fm, where nm denotes the number of
myosin motors pulling the F-actin filament. Fig 7 a and
Fig 7 b show the mean retrograde flow for varying system
size N and nm; where in (a), it is plotted as a function of
substrate elastic stiffness, Ksub, keeping the viscosity, γ,
at a constant value and in (b), as a function of substrate
viscosity, γ, keeping Ksub constant. As seen from Fig.
7 a, with increase in the system size, as more number
of receptors/bonds can bind with the ligand on the sub-
strate, the total bond force increases which balances the
traction force by the substrate; thus, the optimal elastic
stiffness for minimum retrograde flow shifts to a higher
value. Fig. 7 b shows that increase in the system size
also brings similar shift in the optimal substrate viscosity
towards a higher value. This is because, higher viscos-
ity results in slow stress relaxation, thus, increases the
effective combined stiffness of the substrate and allows
transmission of larger traction force as observed in case
of pure elastic stiffness.
We have, further, compared the results for varying
binding rates and also studied without the force induced
adhesion reinforcement.
It is found that the force de-
pendent rate provides an added flexibility for changing
the optimal substrate stiffness or the optimal viscosity to
match the experimentally observed values for different
cell types. This is achieved because changing the force
induced rate results in variation of the total number of
closed bonds, thus, changes the total traction force and
the retrograde flow velocity and consequently shifts the
optimal stiffness.
the responses to modulate the behaviour in order to
fine-tune the biophysical applications such as cancer
research [39], tissue engineering, regenerative medicine
etcetera.
ACKNOWLEDGMENTS
7
V. DISCUSSIONS
the computational
We have developed a theoretical model to study the
'stick-slip' motion at the leading edge of a crawling
cell. The extracellular matrix has been modelled as
a viscoelastic system to better mimic the biological
substrates, as opposed to the generally modelled pure
elastic substrates. Our continuum model framework
comprising of coupled reaction-diffusion equations pre-
dicts the time evolution of force on an individual bond
during a stick-slip cycle, that could not be captured
in existing stochastic model frameworks. Our study
reveals that the loading rate of single bond force is
distinctively different on soft substrate compared to
stiff substrate.
It plays a crucial role in determining
the pattern of the stick-slip jumps on varying substrate
rigidity.
It is also worth noting that our continuum
model description reduces
time
required for averaging of the dynamical quantities as
compared to stochastic models where the averaging
needs to be done over a large number of trajectories
to extract useful statistical information. Also, in our
model, motivated by the experimental findings, the
bond association and dissociation rates are considered
to be force dependent. Experimentally observed force
induced reinforcement of adhesion complexes has been
incorporated in the binding rate as well as through the
'catch bond' behaviour of adhesion complexes [46 -- 48].
Moreover, our analysis elucidates the dependence of
the duration of the stick-slip cycle on various cellular
parameters, for example, how it is affected by myosin
activity, retrograde flow, or substrate stiffness. Our
theory further suggests that the viscoelasticity of the
substrate plays a central role in driving the cell migration
process. It reveals the existence of an 'optimal' substrate
viscosity where the traction force is maximum and the
retrograde flow is minimum similar to the variation of
elastic substrate stiffness. This indicates the importance
of substrate stress relaxation process in cell motility.
As in experiments, cell crawling has been found to
be most efficient on an optimal substrate stiffness,
it
could further be tested by altering the viscosity or a
combination of both viscosity and elasticity to see how
cells respond to viscoelastic tissues and interpreting to
The authors acknowledge the financial support from
Science and Engineering Research Board (SERB), Grant
No. SR/FTP/PS-105/2013, Department of Science and
Technology (DST), India.
Appendix A: Analytical estimation of the duration
of a stick-slip cycle:
b
Considering an elastic substrate, the traction force
due to the deformation of the substrate will be given
= Ksub Xsub, where Xsub denotes the displace-
by F total
ment of the substrate and Ksub is the substrate stiffness.
At any instant, the traction force must be balanced by
summing over forces of all ligand-receptor bonds. Now,
the total bond force could be calculated as, F total
=
b
KcNc(cid:16) Xb − Xsub(cid:17), where Nc is the total number of
closed bonds at any instant, Nc = R L
tion of the bond is given by(cid:16) Xb − Xsub(cid:17). Here Xb is the
filament, thus, Xb =R vmdτ . Here, vm is the dimension-
/ Fstall).
Now, the expression for the traction force at any time
less retrograde flow given by vm = v0(1 − F total
displacement of one end of the bond attached to the actin
0 ncdX, the elonga-
b
τ can be rewritten as,
F total
b
= KcNc Xb −
b
F total
Ksub !
(A1)
Differentiating Eq. A1 w.r.t. the dimensionless time τ
and using the relation, d Xb
dτ = vm, we obtain,
b
d F total
dτ 1 +
KcNc
Ksub ! = KcNcv0 1 −
b
F total
Fstall ! (A2)
Thus, the evolution of the total traction force during a
stick-slip cycle starting from a value 0 at time τ = 0 can
be given by the solution of Eq. A2,
F total
b
(τ ) = Fstall
1 − exp
−
KcNcv0 Ksub
Fstall(cid:16) Ksub + KcNc(cid:17)
(A3)
τ
.
[1] Gardel, M. L., Schneider, I. C., Aratyn-Schaus, Y., and
C. M. Waterman. 2010. Mechanical integration of actin
and adhesion dynamics in cell migration. Annu. Rev. Cell
Dev. Biol. 26:315-333.
[2] Ridley, A. J., Schwartz, M. A., Burridge, K., Firtel, R.
A., Ginsberg, M. H., Borisy, G., Parsons, J. T., and A. R.
Horwitz. 2003. Cell migration:
integrating signals from
front to back. Science. 302(5651):1704-1709.
[3] Friedl, P., and D. Gilmour. 2009. Collective cell migration
in morphogenesis, regeneration and cancer. Nat. Rev.
Mol. Cell Biol. 10:445-457
[4] Dahlin, C., Linde, A., Gottlow, J., and S. Nyman. 1988.
Healing of Bone Defects by Guided Tissue Regeneration.
Plastic and Reconstructive Surgery. 81(5):672-676
[5] Krawczyk, W. S. 1971. A pattern of epidermal cell mi-
gration during wound healing. J Cell Biol. 49:247-263
[6] Ananthakrishnan, R. and A. Ehrlicher. 2007. The Forces
Behind Cell Movement. Int. J. Biol. Sci. 3(5):303-317.
[7] Lin, C. H., and P. Forscher. 1995. Growth cone advance is
inversely proportional to retrograde F-actin flow. Neuron.
14(4):763-771.
[8] Jurado, C., Haserick, J. R., and J. Lee. 2005. Slipping or
gripping? Fluorescent speckle microscopy in fish kerato-
cytes reveals two different mechanisms for generating a
retrograde flow of actin. Mol Biol Cell. 16(2):507-518.
[9] Greenberg M.J., Arpag G., Tzel E., and E. M.
Ostap. 2016. A perspective on the role of myosins as
mechanosensors. Biophys. J. 110:2568-2576.
[10] Mitchison, T.J., and M. Kirschner. 1988. Cytoskeletal
dynamics and nerve growth. Neuron. 1:761-772.
[11] Hu, K., Ji, L., Applegate, K. T., Danuser, G. and C. M.
Waterman-Storer. 2007. Differential transmission of actin
motion within focal adhesions. Science. 315:111-115
[12] De, R., Maybhate, A., and G. Ananthakrishna. 2004. Dy-
namics of stick-slip in peeling of an adhesive tape Phys.
Rev. E. 70:046223
[13] G. Ananthakrishna and R. De. 2007. Dynamics of Stick-
Slip: Some Universal and Not So Universal Features. Lec-
ture Notes in Physics, Springer, Berlin. vol. 705:423-457
[14] De, R. and G. Ananthakrishna. 2005. Missing physics in
stick-slip dynamics of a model for peeling of an adhesive
tape. Phys. Rev. E vol 71:055201(R)-055204(R)
[15] De. R., Maybhate, A., and G. Ananthakrishna, 2004.
Power laws, precursors and predictability during failure
Europhys. Lett. 66 (5):715-721
[16] Burridge, R. and L. Knopoff. 1967. Model and Theoret-
ical Seismicity Bulletin of the Seismological Society of
America. Vol. 57, 3:341-371
[17] Gardel, M. L., Sabass, B., Ji, L., Danuser, G., Schwarz,
U. S. and C. M. Waterman. 2008. Traction stress in focal
adhesions correlates biphasically with actin retrograde
flow speed. J. Cell Biol. 183:999-1005
[18] Chan, C. E., and D. J. Odde. 2008. Traction dynamics
of filopodia on compliant substrates. Science. 322:1687-
1691.
[19] Bangasser, B.L., Shamsan, G. A., Chan, C. E., Opoku,
K. N., Tuzel, E., Schlichtmann, B. W., Kasim, J. A.,
Fuller, B. J., McCullough, B. R., Rosenfeld, S. S. and D.
J. Odde. Shifting the optimal stiffness for cell migration.
Nat. Commun. 2017 8:15313
[20] Aratyn-Schaus Y and Gardel M L. 2010. Transient fric-
tional slip between integrin and the ECM in focal adhe-
sions under myosin II tension Curr. Biol. 20:1145-1153.
[21] Lacayo, C. I., Pincus, Z., VanDuijn, M. M., Wilson, C.
A., Fletcher, D. A., Gertler, F. B., Mogilner, A. and J. A.
Theriot. 2007. Emergence of large-scale cell morphology
and movement from local actin filament growth dynam-
ics. PLoS Biol. 5:e233.
8
[22] Lee, J., Ishihara, A., Theriot, J. and K. Jacobson. 1993.
Principles of locomotion for simple-shaped cells. Nature.;
362, 167-171.
[23] Barnhart,E. L., Allen, G. M., Julicher, F. and J. A. The-
riot. 2010. Bipedal Locomotion in Crawling Cells. Bio-
phys. J. 98:933-942.
[24] Giannone, G., Dubin-Thaler, B. J., Doebereiner, H. G.,
Kieffer, N., Bresnick, A. R. and M. P. Sheetz. 2004. Pe-
riodic lamellipodial contractions correlate with rearward
actin waves. Cell. 116:431-443.
[25] Stroka, K. M., and H. Aranda-Espinoza. 2009. Neu-
trophils display biphasic relationship between migration
and substrate stiffness. Cell Motil. Cytoskeleton. 66:328-
341.
[26] Peyton, S. R., and A. J. Putnam. 2005. Extracellular
matrix rigidity governs smooth muscle cell motility in a
biphasic fashion. J. Cell. Physiol. 204:198-209.
[27] Elosegui-Artola, A., Trepat, X. and P. Roca-Cusachs.
2018. Control of Mechanotransduction by Molecular
Clutch Dynamics. Trends in Cell Biology. 28:356-367
[28] Rangarajan, R., and M. H. Zaman. 2008. Modeling cell
migration in 3D: status and challenges. Cell Adhes. Migr.
2:106-109.
[29] DiMilla, P. A., Barbee, K. and D. A. Lauffenburger. 1991.
Mathematical model for the effects of adhesion and me-
chanics on cell migration speed. Biophys. J. 60:15-37.
[30] Danuser, G., Allard, J., and A. Mogilner 2013. Mathe-
matical modeling of eukaryotic cell migration:
insights
beyond experiments. Annual Review of Cell and Devel-
opmental Biology. 29:501-528.
[31] Shemesh, T.,Bershadsky A. D., and M. M. Kozlov. 2012.
Physical model for self-organization of actin cytoskele-
ton and adhesion complexes at the cell front. Biophys. J.
102:1746-1756.
[32] Craig, E. M., Stricker, J., Gardel, M. and A. Mogilner.
2015. Model for adhesion clutch explains biphasic rela-
tionship between actin flow and traction at the cell lead-
ing edge. Phys. Biol. 12:035002
[33] Kabaso, D., Shlomovitz, R., Schloen, K., Stradal, T. and
N. S. Gov, 2011. Theoretical model for cellular shapes
driven by protrusive and adhesive forces. PLoS Comput.
Biol. 7:e1001127.
[34] Rubinstein, B., Fournier, M. F.,
Jacobson, K.,
Verkhovsky, A. B. and A. Mogilner. 2009. Actin-myosin
viscoelastic flow in the keratocyte lamellipod. Biophys.
J. 97:1853-1863.
[35] Kruse, K., Joanny, J. F., Julicher, F. and J Prost. 2006.
Contractility and retrograde flow in lamellipodium mo-
tion. Phys Biol.3(2):130-137.
[36] Sabass, B. and U .S. Schwarz, 2010. Modeling cy-
toskeletal flow over adhesion sites: competition between
stochastic bond dynamics and intracellular relaxation. J.
Phys. Condens. Matter. 22:194112.
[37] Li, Y., Bhimalapuram, P. and A. R. Dinner. 2010. Model
for how retrograde actin flow regulates adhesion traction
stresses. J. Phys. Condens. Matter. 22:194113.
[38] Bangasser, B. L., Rosenfeld, S. S. and D. J. Odde. 2013.
Determinants of maximal force transmission in a motor-
clutch model of cell traction in a compliant microenvi-
ronment. Biophys. J. 105:581-592.
[39] Elosegui-Artola, A., Bazellires, E., Allen, M. D., Andreu,
I., Oria, R., Sunyer, R., Gomm, J. J., Marshall, J. F.,
Louise Jones, J., Trepat. X and P. Roca-Cusachs. 2014.
Rigidity sensing and adaptation through regulation of
integrin types. Nature Mater. 13:631-637
[40] Bennett, M., Cantini, M., Reboud, J., Cooper, J.
M., Roca-Cusachs, P. and M. Salmeron Sanchez. 2018.
Molecular clutch drives cell response to surface viscos-
ity,Proc. Natl. Acad. Sci. U.S.A. 115(6):1192-1197.
[41] Chaudhuri, O., Gu, L., Darnell, M., Klumpers, D.,
Bencherif, S. A., Weaver, J. C., Huebsch, N. and D. J.
Mooney. 2015. Substrate stress relaxation regulates cell
spreading. Nature Commun. 6:6364.
[42] Lautscham, L. A., Lin, C. Y., Auernheimer, V., Nau-
mann, C.A., Goldmann, W. H. and B. Fabry, 2014.
Biomembrane-mimicking lipid bilayer system as a me-
chanically tunable cell substrate, Biomaterials. 35:3198-
3207
[43] Balaban, N.Q., Schwartz, U.S., Riveline, D., Goichberg,
P., Tzur, G Sabanay, I., Mahalu, D., Safran, S., Ber-
shadsky, A., Addadi, L., and B. Geiger, 2001. Force and
focal adhesion assembly: a close relationship studied us-
ing elastic micropatterned substrates. Nat. Cell Biol.
3:466-472.
[44] Tan, J. L., Tien, J., Pirone, D. M., Gray, D. S., Bhadri-
raju, K. and C. S. Chen. 2003. Cells lying on a bed of
9
microneedles: an approach to isolate mechanical force.
Proc. Natl Acad. Sci. U.S.A. 100:1484-1489.
[45] De, R. 2018. A general model of focal adhesion orien-
tation dynamics in response to static and cyclic stretch,
Communications Biology. 1: 81
[46] Thomas, W. E., Vogel, V. and E. Sokurenko, 2008. Bio-
physics of catch bonds. Annu. Rev. Biophys. 37:399-416
[47] Marshall, B. T., Long, M., Piper, J. W., Yago, T.,
McEver, R. P. and C. Zhu. 2003. Direct observation of
catch bonds involving cell-adhesion molecules. Nature.
423:190-193.
[48] Pereverzev, Y., Prezhdo, O. V., Forero, M., Sokurenko,
E. and W. Thomas. 2005. The two-pathway model for
the catch-slip transition in biological adhesion. Biophys.
J. 89:1446-1454
[49] De, R., Zemel, A. and S. A. Safran. 2010. Theoretical
Concepts and Models of Cellular Mechanosensing. Meth-
ods Cell Biol. 98:143-175.
[50] De, R., Zemel, A. and S. A. Safran. Dynamics of cell
orientation. 2007. Nature Physics. 3:655-659.
|
1709.03752 | 2 | 1709 | 2018-01-17T16:22:49 | Microstructurally-based constitutive modelling of the skin - Linking intrinsic ageing to microstructural parameters | [
"physics.bio-ph",
"q-bio.TO"
] | A multiphasic constitutive model of the skin that implicitly accounts for the process of intrinsic (i.e.\ chronological) ageing via variation of the constitutive parameters is proposed. The structurally-motivated constitutive formulation features distinct mechanical contributions from collagen and elastin fibres. The central hypothesis underpinning this study is that the effects of ageing on the mechanical properties of the tissue are directly linked to alterations in the microstructural characteristics of the collagen and elastin networks. Constitutive parameters in the model, corresponding to different ages, are identified from published experimental data on bulge tests of human skin. The identification procedure is based on an inverse finite element method. The numerical results demonstrate that degradation of the elastin meshwork and variations in anisotropy of the collagen network are plausible mechanisms to explain ageing in terms of macroscopic tissue stiffening. Whereas alterations in elastin affect the low-modulus region of the skin stress-strain curve, those related to collagen have an impact on the linear region. | physics.bio-ph | physics | Microstructurally-based constitutive modelling of the skin –
Linking intrinsic ageing to microstructural parameters
D. Ponda, A.T. McBridea,b,1, L.M. Davidsc, B.D. Reddya, G. Limbertd,e
aCentre for Research in Computational and Applied Mechanics, University of Cape Town, South
Africa
bSchool of Engineering, University of Glasgow, G12 8QQ, United Kingdom
cDepartment of Medical Biosciences, University of the Western Cape, South Africa
dnational Centre for Advanced Tribology at Southampton (nCATS) / Bioengineering Science Research
Group, Faculty of Engineering and the Environment, University of Southampton, United Kingdom
eLaboratory of Biomechanics and Mechanobiology, Department of Human Biology, Faculty of Health
Sciences, University of Cape Town, South Africa
Abstract
A multiphasic constitutive model of the skin that implicitly accounts for the process of
intrinsic (i.e. chronological) ageing via variation of the constitutive parameters is pro-
posed. The structurally-motivated constitutive formulation features distinct mechanical
contributions from collagen and elastin fibres. The central hypothesis underpinning this
study is that the effects of ageing on the mechanical properties of the tissue are directly
linked to alterations in the microstructural characteristics of the collagen and elastin
networks. Constitutive parameters in the model, corresponding to different ages, are
identified from published experimental data on bulge tests of human skin. The identi-
fication procedure is based on an inverse finite element method. The numerical results
demonstrate that degradation of the elastin meshwork and variations in anisotropy of
the collagen network are plausible mechanisms to explain ageing in terms of macroscopic
tissue stiffening. Whereas alterations in elastin affect the low-modulus region of the skin
stress-strain curve, those related to collagen have an impact on the linear region.
Keywords:
continuum mechanics; finite element method
soft tissue; microstructure; anisotropy; collagen; elastin; bulge test;
1. Introduction
The skin acts as a biophysical interface between the human body's internal and ex-
ternal environments. Due to this critical role, significant research has been undertaken
to elucidate the complex processes underlying the skin's physiology and biophysics [1–4].
Of particular relevance to the present study, is the search for plausible microstructural
mechanobiological mechanisms responsible for alterations of the skins macroscopic me-
chanical properties with age.
∗Corresponding author.
Email address: [email protected] (A.T. McBride)
1
8
1
0
2
n
a
J
7
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
2
5
7
3
0
.
9
0
7
1
:
v
i
X
r
a
Between 2001 and 2011, the UK population aged 65 and over increased by 0.92 million
[5]. The UK population aged 85 and over increased by almost 25% (from 1.01 million
to 1.25 million). By 2039, almost a quarter of the UK population will be 65 and over
[6]. As the aged population continues to increase, two main issues emerge. First, the
need for medical treatment tends to increase with longevity. Second, the ageing process
itself results in significant degradation of the biophysical properties of skin. This leads to
further, and potentially life-threatening, complications such as skin tears and associated
chronic wounds [7].
The skin accounts for up to 16% of an adults total body weight and covers an average
surface area of about 1.6 m2 [8]. Given this, and the role of the skin as the prime line
of defence against the external environment, it is clear that any significant alteration of
mechanical properties that could compromise the structural integrity and barrier func-
tion, has the potential to cause serious detrimental health issues. Moreover, both in
the UK and globally, the ageing population is rapidly becoming a significant market seg-
ment across many industrial sectors, including medical devices, consumer goods, personal
care products, sports equipment and consumer electronics. It is therefore important to
understand how age-related alterations of skin biophysics can be accounted for in the
development of new or improved products.
In accordance with the universal principle of close integration of structure and func-
tion in biological tissues [9], the human skin features a complex multiscale structural
architecture. For simplicity's sake, the skin is often mesoscopically described as a three-
layer structure made of an epidermis, dermis and hypodermis (see Fig. 1(b)) [8, 10, 11].
Figure 1: (a) Histological section of haematoxylin and eosin stained back human skin sample obtained
from a 30 years-old healthy white caucasian female volunteer following biopsy (10 x magnification, image
resolution: 1600 x 1200 pixels, imaged using a modified Nikon E950 camera). Image courtesy of Dr.
Maria-Fabiola Leyva-Mendivil (University of Southampton, UK); (b) Schematic representation of the
mesoscopic multi-layer structure of the skin.
The dermis can be decomposed into three main layers: the papillary layer juxtaposed
to the epidermis, the sub-papillary layer underneath, and the reticular layer which is
connected to the underlying subcutaneous tissue. The subcutaneous tissue is the layer
between the dermis and the fascia which is a band of connective tissue, primarily collagen,
that attaches, stabilises, encloses, and separates muscles and other internal organs. The
thickness of the subcutaneous tissue is highly variable, both intra- and inter-individually.
This layer is mainly composed of adipocytes. Its role is to provide mechanical cushioning,
2
(a)(b)figure 1 heat generation, insulation, and as a reserve of nutrients.
Like any other organ of the human body, the skin inexorably undergoes what is
termed chronological ageing, or more correctly, intrinsic ageing: a series of biochemical
molecular degenerative changes occurring as the indirect result of the passage of time
and progression into older age. These alterations involve decreased proliferative capacity
which leads to cellular senescence and altered biosynthetic activity of skin-derived cells.
Deterioration of the skin barrier function, a reduced rate of turnover of epidermal cells
(e.g. keratinocytes), alteration of the biochemical and mechanical properties of the colla-
gen network in the dermis, macroscopic stiffening, loss of elasticity and a poor vascular
network around hair bulbs and glands are manifestation of these intrinsic ageing-induced
physiological changes [12]. The general consensus is that intrinsic ageing is triggered by
two main mechanisms which can operate in concert: DNA damage and chromosomes
telomere shortening [13–17]. An important aspect to consider is that intrinsic ageing
typically occurs in combination with extrinsic ageing which is the result of external envi-
ronmental and lifestyle factors, particularly exposure to ultraviolet (UV) radiations from
sun light [18–20], and increasingly, sunbeds [21, 22], as well as smoking [23, 24] and air
pollution [24, 25]. Extrinsic ageing due to UV radiation exposure is called photoageing
[19, 20].
A popular view on skin ageing is embodied by the microinflammatory model advo-
cated by Giacomoni and Rein [17]. These authors proposed a mechanistic biochemical
model of ageing which is not rooted in any mathematical formulation but is rather a
descriptive qualitative model integrating information about the ageing of the extracellu-
lar matrix and that of the dermis. The model accounts for both intrinsic and extrinsic
ageing effects through loss of elasticity, resilience and flexibility of the dermis, and also in-
tegrates the formation of wrinkles and the thinning of the epidermis. Ageing affects both
structure and function of the skin and these alterations have important consequences for
its physiology [13, 14, 26, 27], rheology [28, 29], tribology [30–32] and the likelihood of
developing particular age-related skin disorders such as xerosis and pruritus [33] or skin
tears [7].
As a result of biochemical changes associated with ageing, the mechanical behaviour
of the skin is significantly altered through time-dependent variations in the structural and
mechanical properties of its elemental constituents (e.g. proteoglycans, collagen, elastin
and keratin). The effects of skin ageing become apparent in the form of skin wrinkles -
particularly on the face and hands - which are often unconsciously or consciously used
as criteria to evaluate age. These morphological changes of the skin surface are manifes-
tations of complex coupled biophysical phenomena where mechanics is believed to play
a critical role as in many other remodelling and morphomechanical processes [34, 35].
Unravelling the inherent complexity of the skin ageing process, firstly by identifying its
biophysical drivers, underlying factors and effects, and secondly, by gaining a mechanistic
insight into their interplay, is a formidable challenge at both experimental and modelling
levels. However, building upon the extensive body of work on the constitutive modelling
of the skin [1–3], there are opportunities to extend existing models and develop mecha-
nistic constitutive formulations that could first describe, and ultimately, predict ageing
effects. These hypothesis-driven research tools have the potential to unveil the biophys-
ical complexity of skin physiology as well as mechanobiological aspects associated with
diseases and the ageing process.
Various phenomenological and structurally-based mechanical models have been pro-
3
posed and adopted to model the skin at finite deformations. For a good coverage of the
relevant literature see the monograph by Xu and Lu [36], review papers by Jor et al.
[1], Li [2], Limbert [3] or book chapters by Limbert [10], Flynn [37]. Structurally-based
models attempt to reflect the contributions and mutual interactions from the primary
constituents of the dermal layer, i.e. collagen, elastin and ground substance. Developing
these type of models is desirable as they offer the ability to link the microscopic con-
stituent characteristics (i.e. materials and structures) to the macroscopic response of the
tissue [see 38–40]. Mazza et al. [41, 42] developed a non-linear constitutive model to
simulate ageing of the human face. The elasto-visco-plastic constitutive model is based
on the formulation by Rubin and Bodner [43]. Mazza et al. [41, 42] extended the model
of Rubin and Bodner [43] by including an ageing parameter equipped with its own time
evolution equation. This ageing-driven parameter was a modulator of tissue stiffness. A
four-layer model of facial skin combined with a face-like geometrical base was developed.
The work highlighted the utility of the model to study the effects of skin ageing on facial
appearance. Maceri et al. [44] proposed an age-dependent multiscale mechanical model
for arterial walls that effectively coupled elastic nanoscale mechanisms, linked to molec-
ular and cross-link stretching, to micro- and macroscopic structural effects. The model
successfully captured the age-dependent evolution of arterial wall mechanics through al-
terations of the constitutive parameters including geometric characteristics of collagen
fibres, cross-link stiffness of collagen fibrils and volume fraction of constituents.
To date, however, no mechanistic constitutive model for the skin has been devel-
oped that is capable of simultaneously capturing intrinsic ageing through evolution of
material and structural constitutive parameters whilst being embedded in the rigorous
framework of non-linear continuum mechanics. The goal of the research presented here
is to develop an experimentally-based mathematical and computational model of the
skin to study the interplay of its material and structural properties and their evolution
as a result of the intrinsic ageing process. Here, rather than proposing evolution laws
governing the alteration of constitutive parameters as a function of ageing, we conduct
an experimentally-based parametric study of skin's mechanics to investigate the corre-
lation between the process of ageing and the constitutive model's parameters. Plausible
mechanisms associated with ageing-induced material and microstructural evolution are
explored in an attempt to explain observed effects associated with ageing (e.g. macro-
scopic stiffening of the tissue).
The paper is organised as follows. The general structural and material properties
of the skin are discussed in Sec. 2. Special focus is placed on the evolution of these
characteristics with intrinsic ageing. A continuum-mechanics description of the skin is
presented in Sec. 3. Sec. 4 is concerned with the formulation of hypotheses on the link
between ageing and the evolution of material and structural properties and their effects
on the observed macroscopic mechanical response of the skin. Experimental bulge test
data [45, 46] are used in Sec. 5 to calibrate the finite element approximation of the
continuum model. The proposed ageing model is then used to predict the influence of
ageing on the mechanical response of the skin in Sec. 6. Sec. 7 includes a discussion of
the results and conclusions.
4
2. Structural and material properties of the ageing skin
2.1. Structural aspects
It is generally accepted that the bulk of the skins response to loading is due to the
dermal layer [28, 47]. The dermis, is composed of papillary, sub-papillary and reticular
layers, and is comprised of a matrix of collagen, elastin, other elastic fibres and ground
substance. The expression of the macroscopic mechanical properties of the skin is due
to these components, their structural organisation and their mutual interactions. The
ground substance is a gel-like amorphous phase mainly constituted of proteoglycans
and glycoproteins (e.g. fibronectin) as well as blood and lymph-derived fluids which
are involved in the transport of substances crucial to cellular and metabolic activities.
Proteoglycans are composed of multiple glycosaminoglycans (i.e. mucopolysaccharides)
interlaced with back bone proteins. Dermal fibroblasts produced glycosamine which is
rich in hyaluronic acid and therefore plays an essential role in moisture retention.
Collagen has been found to make up approximately 66-69% of the fractional volume
of the dermis [28], and approximately 34% [48] and 70-80% [49] of the wet weight and
dry weight of the skin, respectively. Experiments where collagen was isolated through
enzymatic treatment [50–52], conclude that collagen is responsible for the tensile strength
of the skin. The papillary layer (Fig. 1(a)) is defined by the rete ridges protruding into
the epidermis and contains thin collagen fibres, sensory nerve endings, cytoplasms and a
rich network of blood capillaries. The sub-papillary layer - the zone below the epidermis
and papillary layer - features similar structural and biological components to those of
the papillary layer.
Besides the dominant content of types I and III collagen (respectively 80% and 15% of
total collagen content), the reticular layer is innervated and vascularised, contains elastic
fibres (e.g. elastin) and the dermal matrix made of cells in the interstitial space. Cells
present in the reticular dermis include fibroblasts, plasma cells, macrophages and mast
cells. Collagen fibres in the papillary and sub-papillary dermis are thin (because of their
low aggregate content of fibrils) and sparsely distributed while reticular fibres are thick,
organised in bundles and densely distributed. Fibrils are typically very long, 100 to 500
nm in diameter featuring a cross striation pattern with a 60 to 70 nm spatial periodicity.
The diameter of thick collagen bundles can span 2 to 15 µm. Birefringence techniques
have been used to characterise the orientation of a supramolecular organisation of collagen
bundles in skin [53].
Contributing approximately 2-4% of the dry weight of skin [45], elastin fibres are
highly compliant with the ability to stretch elastically to twice their original length [54].
Their diameter ranges between 1 and 3 µm. Their mechanical intertwining with the
collagen network of the dermis is what gives the skin its resilience and recoil ability.
This is evidenced by the correlation between degradation of elastin/abnormal collagen
synthesis associated with ageing and the apparent stiffening of the dermis [27]. This
aspect will be shown to be particularly important for the mechanistic model linking
ageing to the alteration of the skin's constitutive model parameters presented in Sec. 4.
The diameter of elastic fibres in the dermis is inversely proportional to their proximity
to the papillary layer where they tend to align perpendicular to the dermal-epidermal
junction surface.
5
2.2. Material aspects
At a macroscopic level, the mechanical properties of the skin are anisotropic and
inhomogeneous. This is due to the complex hierarchical structure and materially non-
linear constituents and residual tension lines in the skin (i.e. the so-called "Langer lines",
first recognised by the Austrian anatomist Karl Langer in his seminal study [55]).
The angular distribution of collagen fibres and the non-uniform fibre geometry means
that under stretch not all fibres straighten and stretch equally. This accounts for the
anisotropic stiffness response when load is applied either along or across the preferential
fibre direction. The magnitude of these directional effects has been the subject of several
modern studies. The Youngs modulus parallel to the Langer lines was found to be greater
than that perpendicular by a ratio of approximately 2.21:1 [56]. Similarly, Reihsner et al.
[49] found that the degree of anisotropy differs across anatomical site, an observation also
made by Langer in his original study. In addition, they found that in situ stress ranges
from 0.2-1.6 N mg−1 along the Langer lines and 0.1-1.3 N mg−1 perpendicularly to them,
with the degree of anisotropy differing between principle stress components by 0.1-0.3
N mg−1.
When the skin is stretched, the elastin fibres are the first to take on stretch [57],
indicating that this contribution is important at low strain levels. Stress-strain curves
of elastin-free skin [51] show that elastin supports the entire load up to 50% strain after
which the strength rapidly increases due to the collagen. The elastic modulus of elastin
has been found to be around 1 MPa. This agrees with the Youngs modulus of skin at low
strain. In addition, elastin is not strong enough to provide much tensile strength at higher
strains [58]. Reihsner et al. [49] state that elastin is responsible for the recoiling of the
skin and collagen after stress is applied. Following degradation of the elastin through the
use of elastase, Oxlund et al. [51] found that the large strain response occurs sooner for
a given tensile load. This seems to suggest that, in the absence of elastin, collagen fibres
take on load at lower strain levels than when elastin is present. The ground substance
has been shown to influence mainly the viscoelastic properties of the skin because of
high-water content and complex time-dependent interstitial fluid motion. Upon removal
of various macromolecules within the ground substance, Oxlund and Andreassen [50]
showed that there was no effect on the mechanical response of rat skin, while Oomens
et al. [59] suggests that ground substance probably only plays a major role when soft
tissue is subject to compression.
Fig. 2 highlights the typical strain hardening characteristics of skin subjected to
uniaxial extension where each portion of the stress-strain curve can be related to the
dermal constituents as described in the next paragraph.
Low modulus portion of the stress-strain curve: this is associated with the gradual
straightening of crimped collagen fibres. During this stage, the greatest resistance to
loading is generated by the elastin and ground substance, with collagen fibres offering
very little resistance. The low modulus portion can be further divided into two phases:
Phase 1: wavy collagen fibres are still relaxed and elastin fibres take on the
majority of the load.
Phase 2: collagen fibres start to uncrimp, then elongate and eventually start to
bear load.
Linear region of the stress-strain curve: collagen fibres straighten and align with the
load direction. Straightened collagen fibres strongly resist loading. This results in the
6
Figure 2: Typical stress-strain curve for skin subjected to uniaxial extension.
rapid stiffening of the skin. The steep linear stress-strain relation is due to stretching
and slippage between fibrils and molecules.
Final yield region of the stress-strain curve:
tensile strength of collagen fibres is
reached and fibres begin to sequentially break (not shown in Fig. 2).
2.3. Variations of skin structural and material properties with age
It is generally accepted that skin thickness decreases with age [52, 60]. Pawlaczyk
et al. [26] found that there is an overall loss of 0.7-0.8 mm of thickness in older skin. It
has been found that skin thickness reaches a maximum around the fourth decade for men
and third decade for women after which there is a gradual decrease [48], a result similar
to that established by Diridollou et al. [61] who found that after an initial increase during
maturation (0-20 years), thickness remains constant to about the age of 60 followed by
a decrease as follows
t = −6 × 10
−3 [mm/year] × age [years] + 1.3 [mm] .
(1)
Moreover, the rate of decrease is more significant in female subjects. It is important to
highlight that these observations on skin thickness apply to the skin as a whole composite
structure as individual skin layers may follow different trends depending on whether
intrinsic and extrinsic ageing effects are considered separately or assumed to be combined.
After 20 years of age, across all layers, the skin thickness starts to diminish at a rate that
increases with age [62]. Between 30 and 80, the unexposed skin can lose up to 50% of its
thickness and this effect is accentuated in zones exposed to sunlight such as the face or
neck. Overall, epidermal thickness drops by about 6.4% per decade faster in women than
men. It is generally believed that the reported reduction in dermal thickness is mainly
caused by the loss of dermal collagen and elastin in elderly adults [63]. It was shown
that in post-menopausal women a 1.13% per year skin thickness reduction is correlated
7
stressstrainlow modulus regionlinear region30%figure 2 with a 2% decrease per year in collagen content [64]. It is also accepted that skin ageing
is characterised by an increase in macroscopic or apparent stiffness [26, 36, 65], although
there is little agreement on the magnitude of the ground state's Youngs modulus, or the
age of onset of stiffening.
As mentioned by Xu and Lu [36], there is a sudden increase in the Youngs modulus
of the skin at age 30 of around 50%, whereas others quote an increase from the age
of 45. It was observed by Escoffier et al. [60] that there is an increase of around 20%
after the age of 70, which is backed by the findings of L´eveque et al. [48]. Furthermore,
Alexander and Cook [65] found that the stiffness of skin starts increasing from the age
of 25 but noted that the variation in results increases with age. This suggests that the
process of skin ageing is a highly patient specific and may explain the large variation in
results reported in the literature. It was found by Escoffier et al. [60] and L´eveque et al.
[48] that intrinsic skin extensibility (i.e. a standardised mean extensibility to account for
varying skin thickness) decreases with age, while Alexander and Cook [65] found that
intrinsic extensibility decreases by around 35% after the age of 65 which agrees with the
findings of Xu and Lu [36] in that maximum skin elongation is found between ages 35
and 55. Similarly, skin elasticity (i.e. recoil ability) decreases with age [36, 60], which
was similarly found by Henry et al. [66], but this may include the effects of UV radiation.
It has been widely observed that, with age, there is a decrease in the initial portion
of the elongation-stress curve [49, 65, 67] which means that the onset of stiffening asso-
ciated with recruitment of collagen fibres occurs at lower stretch. As elastin is primarily
responsible for the low stretch response, it is clear that there must be some modification
to the elastin network with age. Elastin is a highly stable protein, the biosynthesis of
which remains steady for the first 40 to 50 years, after which there is a decline. Various
authors [28, 49, 65, 67] attribute the decrease in the low-modulus portion to the gradual
degradation of the elastin network, as well as the deposition of amorphous elastin. This
statement is backed by the observation that, upon enzymatic removal of elastin from skin
samples, similar effects were obtained to those of ageing.
In the linear region of the stress-strain curve (see Fig. 2), collagen fibres take on
tension which results in the typical nonlinear response. It has been reported that the slope
of the linear portion of the elongation-stress curve tends to increase with age [65]. This
suggests an apparent stiffening of the collagen fibre network with age but not necessarily a
stiffening of the collagen fibres themselves. This aspect is supported by the experimental
evidence obtained by Daly and Odland [67] and Reihsner et al. [49] who note that the final
slope of the strain-stress curve of skin in tension remains constant with age, and that the
stiffness of the collagen remains constant, possibly due to the reduction in the collagen
content in the dermal layer [68, 69]. There is therefore suggestion that the alteration of
the skin apparent stiffness is not due to a stiffening of the collagen fibres, but rather to an
alteration in the structure of the collagen network. It was found that alterations in total
skin thickness were proportional to the alterations in dermal thickness and that, due
to the compacting of the dermis with age, collagen bundles tend to flatten and unravel
[48]. Coupled with the loss of elastic integrity, Diridollou et al. [61] postulate that, due
to the dermal thinning, collagen fibres may tend to unfold to some extent, which could
further explain the reduction in skin extensibility and the smaller low-modulus portion,
as collagen is activated sooner.
Reihsner et al. [49], Escoffier et al. [60], Agache et al. [70] mention that there is in-
creased crosslinking with age, which would reduce any slippage between neighbouring
8
fibres and stiffen up the collagen network. Anisotropy tends to increase with age sug-
gesting that there is an increase in alignment along Langer lines with age [29, 71], or at
least, a correlation between age, microstructurally-induced anisotropy and Langer lines,
although Tonge et al. [45] found a decrease in overall anisotropy with age.
In summary, three factors that certainly influence biomechanical behaviour of skin
with age are:
1. Increased cross-linking between collagen fibres - collagen network density increases
with age as well as a decrease in fibre free length. Collagen fibres are thus more
compact and appear to unravel.
2. The loss of elasticity at low strain is attributed to the destruction of the elastin
network. Elastic fibres become thinner and fractionated with age. The decreased
extensibility properties of skin with age are generally attributed to the loss of elastin
in the upper dermis with age.
3. Water content tends to decrease with age, which alters the overall viscoelastic
properties.
3. Continuum model of skin
3.1. Notation
Direct notation is adopted throughout. The scalar product of two vectors a and b is
denoted by a· b. The scalar product of two second-order tensors A and B is denoted by
A : B. The composition of two second-order tensors A and B is denoted by AB. The
action of a second-order tensor A on a vector b is a vector denoted by Ab. The unit
basis vectors in the Cartesian (standard-orthonormal) basis are {e1, e2, e3}.
3.2. Finite strain kinematics
Consider a continuum body, here representing skin, defined in relation to the Carte-
sian reference frame. The reference configuration is defined as the placement of this
body at time t = 0, with that region denoted by Ω0 and boundary ∂Ω0 with outward
unit normal N . As the body deforms, this region takes on subsequent configurations. At
a current time t, the body occupies the region Ω referred to as the current configuration
and current boundary surface ∂Ω with outward unit normal n. For an extensive overview
of continuum mechanics the reader is referred to [72, 73], among others.
It is assumed that there exists a mapping χ such that each material point with
position X ∈ Ω0 uniquely maps to a spatial point with position x ∈ Ω at time t, i.e.
x = χ(X, t)
∀X ∈ Ω0 .
The motion χ is assumed to possess continuous derivatives both in space and time.
The deformation of the body can be characterized by the deformation gradient F , i.e.
the material gradient of the motion, defined by
F (X, t) :=
∂χ(X, t)
∂X
= Gradχ(X, t) .
The determinant of F is defined by J := det F > 0. The right Cauchy–Green tensor C,
defined by
C := F T F ,
9
provides a stretch measure in the reference configuration. Additionally, the principal
scalar invariants of C are defined by
(cid:0)tr(C)2 − tr(C2)(cid:1) ,
I1(C) := tr(C) = I : C ,
I2(C) :=
1
2
I3(C) := det(C) .
(2)
Consider now the case of transverse isotropy where the material properties depend on
a single given direction (which is a defining feature of the present constitutive formulation
for skin). The preferred material direction is given by v0. The fabric tensor is defined
by
A0 := v0 ⊗ v0 .
An additional invariant characteristic of transverse isotropic symmetry is defined by
I4(C, v0) = v0 · Cv0 = λ2,
(3)
where λ is the principal stretch along vector v0 defined in the reference configuration.
Note, five scalar-valued tensor invariants are necessary to form the irreducible in-
tegrity bases of the tensors C and v0 ⊗ v0 [74, 75]. Thus, to fully describe transverse
isotropic symmetry, an additional invariant I5 must be introduced [72, 74–76], where
I5 := v0 · C2v0 .
The fifth invariant is often discarded for mathematical convenience and because of the
intrinsic difficulty of providing it a direct physical interpretation that can be linked to
measurement [77]. Destrade et al. [78] demonstrated that excluding I5 as an argument
of the strain energy function of a transversely isotropic hyperelastic material could result
in non-satisfaction of kinematic constraints present in physical tests. As is often the case
in constitutive models of biological soft tissues, and for simplicity's sake with regards
to the interpretation of the results of our study, the fifth invariant I5 is not considered
here. For a general discussion on the inclusion of additional invariants to capture the
mechanical response of transversely isotropic materials, see [78].
3.3. Constitutive relations for invariant-based transversely isotropic hyperelasticity
A hyperelastic material is one for which a free energy ψ, defined per unit reference
volume, acts as a potential for the stress. For homogeneous materials, the free energy
is purely a function of the deformation gradient, i.e. ψ = ψ(F ), and of any additional
tensor agency (e.g. structure tensors). Note that, from standard objectivity arguments
arguments, ψ depends on F through C [see e.g. 72, 73]. As a general procedure to for-
mulate constitutive equations for hyperelastic materials, one can postulate the existence
of a strain energy density ψ, that is an isotropic function of its deformation or strain-
invariant arguments. The first and second Piola–Kirchhoff stress measures, P and S,
are defined by
P =
∂ψ(F )
∂F
and
10
S = 2
∂ψ(C)
∂C
,
where P = F S. Under the assumption that the material modelled exhibits some form
of material symmetry, the dependence of the free energy on C can be be expressed in
terms of the ninv invariants of C. The second Piola–Kirchhoff stress is thus given by
ninv(cid:88)
S = 2
∂ψ(C)
∂Ii
∂Ii
∂C
.
(4)
3.4. Balance relations
i=1
The balance of linear momentum, in the absence of inertial or body forces, and the
natural boundary condition are given by
DivP = 0
T = P N = T
in Ω ,
on ∂Ω0,N ,
(5)
(6)
where Div is the material divergence operator. The Piola traction T is prescribed on
the Neumann part of the boundary ∂Ω0,N ⊂ ∂Ω0. Dirichlet boundary conditions on the
motion χ = χ are prescribed on ∂Ω0,D where ∂Ω0 = ∂Ω0,D∪∂Ω0,N and ∂Ω0,D∩∂Ω0,N = ∅.
3.5. Constitutive assumptions for the dermis
Based on the experimental evidence reported in Sec. 2, it is assumed that elastin and
ground substance are the main contributors to the low modulus portion of the stress-
strain curve (see Fig. 2) which is largely linear and isotropic. The second region of the
loading curve is dominated by the mechanical response of collagen fibres. As the collagen
fibres straighten and take on load, they exponentially resist further stretch which results
in a rapid nonlinear strain stiffening or locking behaviour. This behaviour is also strongly
anisotropic due the inherent preferred alignment of collagen fibres along the direction of
extension. As the stiffening response of the skin is dominated by the response of the
collagen network, the formulation of an effective microstructurally-motivated model is
essential. Hence, it is postulated that the free energy describing the overall mechani-
cal behaviour of the dermis, assumed to be that of the skin because of the negligible
contributions of the epidermis in tension, is given by
Ψ = Ψgs + Ψelastin + Ψcollagen,
(7)
where Ψgs, Ψelastin and Ψcollagen represent the free energy contributions from the ground
substance, elastin and collagen, respectively.
To capture the behaviour at low stretches, the elastin and ground substance energies
are given by the following compressible neo-Hookean type free energy
(cid:18)
(cid:16)
(cid:17)(cid:19)
I1 − 3 +
1
β
I
−β
3 − 1
Ψgs + Ψelastin = (αgs + αelastin)
where
α(•) =
µ(•)
2
,
β =
ν
1 − 2ν
11
(8)
(9)
are constitutive parameters, respectively associated with the shear and volumetric re-
sponse, and (•) is either gs or elastin. The shear modulus is denoted by µ, and ν is
the ground state Poissons ratio of the composite material represented by the ground
substance and elastin phases. Here, the nonlinear, anisotropic and network nature of
the collagen response is captured through a transversely isotropic network eight-chain
model proposed by Kuhl et al. [38]. This model is based on theories related to the
micromechanics of macromolecule mechanical networks [79, 80] and is a particularisa-
tion of the orthotropic eight-chain model developed by Bischoff et al. [81].
In these
mechanical descriptions of macromolecular networks, long molecular chains are assumed
to rearrange their conformation under the influence of random thermal fluctuations (i.e.
entropic forces). Unlike polymer chains in rubber [38] which have conformations reminis-
cent of a random walk, biopolymer chains such as collagen assemblies feature smoothly
varying curvature and are therefore considered correlated. This type of idealised molec-
ular chains is best described using the concept of wormlike chains of Kratky and Porod
[82]. This approach was successfully applied to model the mechanics of DNA molecules
by Marko and Siggia [83]. Since then, wormlike chain models have been used to de-
scribe the structure and mechanical behaviour of collagen assemblies in the context of
skin modelling [40, 81, 84–87] and other fibrous biological soft tissues [38, 39, 88, 89].
It is important to note that, in the development of microstructurally-based constitutive
theories, these macromolecular chains could also be defined or interpreted as tropocol-
lagen molecules, collagen micro-fibrils, fibrils, fibres or fibre bundles.
If considering
supra-molecular scales, it is clear that the wormlike chain energy is no longer associ-
ated with the notion of entropic elasticity and true molecular behaviour but is rather
a microstructurally-motivated macroscopic phenomenological energy that capture well
the strain-stiffening behaviour of collagenous structures. In the present approach, it is
assumed that correlated chains represent collagen fibres (Fig. 3). This assumption is
reasonable and sufficient to demonstrate how microstructural alterations of the dermis
as a function of ageing could lead to macroscopic stiffening of the dermis. In molecular
network theories, molecular chains are assumed to be made of beads connected by N
rigid links of equal length d, the so-called Kuhn length [90], so that the maximum length
of a chain, the contour length, is L = N d.
The mechanics of macromolecular polymer structures is not only governed by the
mechanical properties of individual chains but also by their electromagnetic and mechan-
ical interactions which can take the form of covalent bonds, entanglement and physical
cross-links. These combined effects give rise to strong isotropic and anisotropic network
properties which can be implicitly and effectively captured by network models such as the
eight-chain model of Arruda and Boyce [79] or Kuhl et al. [38]. The essential assumption
underpinning these formulations is that there exists a representative microscopic unit cell
able to capture network properties. The original eight-chain model of Arruda and Boyce
[79] assumes that the unit cell is made of eight entropic chains of equal lengths con-
nected from the centre of the cell to each of its corners (Fig. 3), each equipped with their
own entropic energy Ψwormlike. For correlated chains, one would consider the following
wormlike chain energy defined by
(cid:18)
(cid:19)
Ψwormlike = Ψ0 +
kθL
4A
2
r2
L2 +
1
(1 − r
L ) −
r
L
,
(10)
where L, A, r0 and r and are respectively the contour, persistence, initial end-to-end
12
Figure 3: Wormlike chain assembly. The successive chain link is correlated to the chain before it. A is
the persistence length, r is the effective end-to-end length and L is the chain contour length.
length and the current end-to-end length of the chain (see Fig. 3), and Ψ0 is the wormlike
chain energy in the unperturbed state. The wormlike chain has the defining characteristic
that the chain segments are correlated and exhibit a smooth curvature along the contour.
This correlated form is captured by the persistence length A. The persistence length can
be viewed as a measure of stiffness. Garikipati et al. [40] refers to it as a measurement
of the degree to which a chain departs from a straight line, while Marko and Siggia
[83] interpret it as the characteristic length over which a bend can be made with energy
cost kθ, where k = 1.380 648 52 J K−1 is the Boltzmann constant and θ the absolute
temperature.
In order to incorporate such chain models into an invariant-based constitutive frame-
work it is necessary to relate the individual chain stretch to the overall or macroscopic
deformation. To that end, the principle of affinity is invoked so that the macroscopic and
unit cell principal directions are identical. Due to the symmetry of the chain structure,
the stretch of each chain can be found as a function of the principle stretches. In Fig. 4
a unit cell arrangement of dimensions a × b × b is depicted. For the case of anisotropy
a (cid:54)= b. Additionally, the unit cell is characterised by the unit vector v0 that corresponds
to the (local) preferred orientation of the collagen network. The undeformed end-to-end
length of each of the individual chains is given by
(cid:112)
a2 + 2b2 .
r0 =
1
2
(cid:112)
r =
1
2
I4a2 + (I1 − I4)b2 .
The deformed end-to-end length r can be expressed using the invariants defined in Eq. (2)
and Eq. (3) by
The invariants I1, I3 and I4 are characteristic of the macroscopic deformations of the
polymer. It is assumed that the microscopic stretch along the direction v0 is captured
13
collagen fibresLrfigure 4ArL Figure 4: Schematic representation of the transversely isotropic eight chain network model of [91]. The
eight polymer chains are governed by a wormlike energy function ψwormlike (see Eq. (10)). The unit cell
dimensions are a and b while unit vector v0, aligned with one of the principal direction of the unit cell,
is the unit vector corresponding to the preferred orientation of the collagen network in the undeformed
configuration.
by the macroscopic term I4. This effectively couples the microscopic and macroscopic
length scales giving rise to a multiscale model. The relative stretch of a collagen fibre is
defined by λr := r/r0.
The final term of Eq. (7) is further decomposed into the additive form as
Ψcollagen = Ψchain + Ψrepulsive ,
where Ψchain reflects the effective assembly of the eight chain energies, i.e. Ψchain :=
γchainΨwormlike and Ψrepulsive is a repulsive energy that ensures the initial configuration
is stress free and that the chain does not collapse. Ψchain denotes the chain density per
unit cell. S´aez et al. [89] interprets γchain as the number of molecules within a collagen
fibril, while Bischoff et al. [81] describes it as the collagen fibre density which in the
context of collagen, this corresponds to the number of fibres within a bundle. It follows
that
(cid:18)
2
(cid:19)
Ψchain =
(cid:18) 1
γchainkLθ
4A
r2
L2 +
1
4r0
r
L
(cid:19)(cid:18)
1
(cid:16)
L −
1 − r
I [a2−b2]/2
ln
,
4
(cid:17)
+
3
2
ln
(cid:16)
I b2
1
(cid:17)(cid:19)
.
Ψrepulsion = −
γchainkθ
4A
L
+
1
4r0
1
1 − r0
L −
All length quantities have been normalised by the link length L which is the length
of a single link along the wormlike chain assembly. A description of the constitutive
parameters of Ψcollagen is given in Table 1.
The expression for the first Piola–Kirchhoff stress P then follows directly from Eq. (4).
14
Parameter
Poisson's ratio
Shear modulus
Boltzmann constant
Absolute temperature
Chain density
Contour length
Persistence length
Unit cell dimensions
Symbol Units
-
ν
N m−2
µ
J K−1
k
K
θ
m3
γchain
-
L
A
-
-
a, b
Table 1: Constiutive parameters of Ψcollagen.
3.6. Weak form of the governing relations
The weak form of the governing and the accompanying Neumann boundary condition
in Eqs. (5) and (6) is essential for establishing the approximate solution using the finite
element method (FEM). Following the Principle of Virtual Working, multiplying Eq. (5)
by an vector-valued test function δu, where δu = 0 on ∂Ω0,D, and integrating by parts
the result over the reference configuration Ω0 yields the expression of the residual for the
equilibrium equation as(cid:90)
(cid:90)
Gradδu : P dV −
Ω0
δu · T dA = 0 .
(11)
∂Ω0,N
4. Microstructurally-based mechanistic modelling of ageing
The central hypothesis underpinning this study is that the effects of ageing on me-
chanical properties of the tissue are directly linked to alterations in the microstructural
characteristics of the collagen and elastin networks. In this section, some of the constitu-
tive parameters introduced in Sec. 3 are motivated as plausible mechanistic descriptions
of the intrinsic ageing process as evidenced by experimental observations (see Table 2).
It is worthy to point out that, in reality, the material and structural effects of ageing
are likely to be coupled and would therefore lead to a dependency between constitutive
parameters. For sake of simplicity, and in the absence of relevant experimental data, this
interplay is not accounted for here. The effect of intrinsic ageing on the ground substance
manifests itself as a more pronounced viscoelastic response with age. As the constitu-
tive model for the ground substance only capture the elastic response it is assumed that
ageing does not affect the ground substance related parameters in Eq. (8).
The elastin network is observed to degrade with age. The loss of elastic integrity would
logically lead to a reduced contribution from the elastin component to the composite
strain energy function (Eq. (7)) of the constitutive model. One approach could have been
to introduce a volume fraction for elastin, ground substance and collagen. Degradation
of the elastin network would have then been modelled as a reduction in volume fraction
of elastin. Here, instead of using the latter approach, the value of αelastin is assumed to
decrease with age. It should be noted again that elastin is a highly-stable protein which
undergoes little turnover before the age of 40. Thus αelastin would remain constant for
the first four decades.
15
Constituent
Ground substance
Ageing effect
Increased water content results in more pro-
nounced viscoelasticity
Elastin
Collagen
Loss of elasticity
Destruction of elastin network
Loss of mature collagen
Increased crosslinking
Flattening and unravelling of collagen network
Alterations in anisotropy
Parameter
n/a
αelastin
αelastin
γchain
γchain
a, b
a, b
Table 2: Description of the constitutive parameters the evolution of which is associated with ageing.
Within the current model, there is no natural parameter that explicitly describes
the level of collagen crosslinking at any point in time. It is believed that the increased
crosslinking with age is one factor responsible for the increased stiffness of the skin.
Through increasing γchain, it is expected that a stiffer response will be elicited. Although
γchain describes the number of fibres within a bundle, it represents the closest link between
the observed increase in crosslinking with age and the model. It would be expected that
a decrease in collagen density would contribute to a decrease in the overall stiffness of
the skin as the collagen network loses integrity. Thus it is reasonable to assume that a
decrease in γchain with age could describe a loss of collagen. It is impossible to decode
the complex interplay between these two factors.
As discussed previously, the slope of the linear region of the stress-strain curve for skin,
as a composite structure, remains constant with age but the onset of strain hardening
occurs for lower strains. This suggests that there might be a mechanism that simply alters
the structural characteristics of the collagen network and not the mechanical properties
of collagen microfibrils. In their unloaded states collagen fibres are crimped. It is believed
that this conformational state is due to the presence of highly cross-linked elastic fibres
mainly elastin fibres [92] - and is also the result of active tensions exerted by fibroblasts.
Any reduction in fibroblast density in the collagen dermal network would have an effect
on these active forces and on the rest state of collagen fibres. Such a reduction would tend
to relax the residual strain/tension in the dermis. This would make collagen fibres less
crimped and therefore their apparent length would be closer to their fully taut length.
Thus, when the skin is macroscopically loaded in tension, the dermal tissue will not
exhibit a very pronounced toe region corresponding to the progressive uncrimping of
collagen fibres but will rather reach the stiffer linear region much quicker. The corollary
of this observation is that for a given applied macroscopic strain to the skin, the response
of the intrinsically-aged dermal collagen network will be stiffer. With age, due to dermal
flattening and the loss of elastic recoil, the collagen fibres are observed to uncrimp [27].
In the chain network structure of the model, this is captured by adjustments to the
end-to-end length of the collagen fibres through the parameters a and b. By increasing
a and b there will be a reduction in the range of strain of the low modulus portion
prior to the onset of the collagen response at lower stretches. Thus by increasing a and
b the model should be capable of capturing the stiffening response. Additionally, the
anisotropic response of the skin is observed to change with age. Tonge et al. [46] reports
16
a loss of anisotropy through observations of the bulge test. As the model developed here
will be compared to the bulge tests at various ages, it is this behaviour that is intended
to be captured. By reducing the ratio of a/b , either by increasing b or decreasing a or
a combination of both, this should be obtained. Undeniably, there is a reduction in the
thickness of the skin with age. Such an observation is backed by numerous sources in
the literature [52, 60, 61], as mentioned in the previous section, but there is still much
debate over the exact relationship between age and skin thinning. The dermal thickness
changes with age which will be captured directly in the geometry.
5. Numerical simulations of the bulge test
5.1. Details of the finite element implementation
The highly-nonlinear system of governing relations (11) are solved using the finite
element method. The system of equations is linearised using a global Newton–Raphson
procedure. The finite element library AceGen / AceFEM [93, 94] is used to implement the
solution scheme. AceGen facilitates the generation of the finite element interpolation and
constitutive model by exploiting symbolic and automatic differentiation, and automated
code generation. AceFEM provides a flexible environment for the iterative solution of the
nonlinear finite element problem. An adaptive incremental loading scheme is employed
to ensure convergence of the numericlal scheme.
5.2. Parameter identification
In the literature, multiple mechanical skin tests have been proposed to characterise
the skin response to loading. The variety of skin tests and different methods to perform
them, as well as the natural skin variation that exists through factors such as ethnicity,
gender, age and anatomical site has resulted in a broad and varying characterisation
In general, in vivo, in both physiological and supra-physiological
of skin behaviour.
conditions, the skin exhibits material nonlinearity, anisotropy, viscoelasticity and near
incompressibility and can also sustain large deformations [3]. However, elicitation and
relevance of these characteristics is highly variable. Although there are many experi-
mental techniques to characterise certain aspects of skin mechanics [1], bulge tests are
considered here. Performed in vitro, the bulge test applies a positive pressure to the
underside of an excised skin sample. The skin sample is fixed at a specified diameter
aperture which allows for "bulge-like" deformation of the sample under pressure. Defor-
mations are measured and linked to the state of stress so that constitutive parameters
can be identified [45, 46]. The choice of the bulge experiment as a test-bed for the con-
stitutive and computational and models developed here is motivated by the following
factors:
• As the experiments are performed in vitro, the boundary conditions of the numerical
model are simpler to define and impose.
• The large deformations induced in the physical tests ensure that the various com-
ponents of the constitutive model would contribute. At small deformations the
elastin and ground substance will assume the dominant role, and ultimately, as
macroscopic strains increase, the collagen fibres will become active and dictate the
majority of the response at large deformations.
17
• The specimens were taken from the back of the patients, thus minimising any
photoageing effects such as those observed on sun-exposed regions (e.g. the face).
• The experimental results published by Tonge et al. [45, 46] provided a broad range
of kinematic measurements with which to compare the models.
• Crucially, for the present study, the experiments by Tonge et al. [45, 46] were
conducted on skin samples harvested from donors featuring a broad range of ages.
This allows the correlation between age and mechanical properties (i.e. constitutive
parameters) to be studied.
The procedure for the bulge test follows that detailed in [45] where 10 × 10 cm skin
specimens were procured from the back torso of donors, ages ranging from 43 to 83
years. After excision, the adipose tissue was removed and tissue thicknesses measured.
The thickness, gender, age and anatomical site of the samples are listed in Table 3.
Thickness was measured at the centre of each edge and an average taken. This average
was used as the uniform thickness of the sample in the numerical model.
Age Gender Site
43
44
59
61
62
83
Male
Male
Female
Male
Female
Male
Lower back
Lower back
Unknown
Left upper back
Unknown
Unknown
Thickness [mm]
4.86
4.38
5.18
2.01
2.95
2.43
Table 3: Donor and specimen information from Tonge et al. [45].
The specimens were glued to a 7.5 cm diameter ring. The ring serves to constrain
the skin specimen. The skin interior to the ring is subject to pressure loading while that
exterior is fully constrained. The coordinate system for the samples was set such that the
y-axis corresponded to the vertical body axis and the x-axis to the horizontal body axis,
as shown in Fig. 5. Fibre and perpendicular directions can thus be defined by the angle Φ
from the horizontal axis. Controlling for relative humidity and temperature, samples were
inflated through a controlled applied pressure, with a maximum pressure of 5.516 kPa.
Upon inflation the skin samples deformed to an elliptical dome. The dimensions of the
ellipse are used to determine the dominant fibre direction with Φ defined accordingly.
In this part of the study, the objective is to use a finite element updating technique [see
e.g. 1, 95, 96] to identify the constitutive parameters of the transversely isotropic eight-
chain model of the skin for various ages. Fig. 6 shows the discretisation of the skin sample.
The skin thickness is set to the dermal thicknesses as specified in Table 3. The geometry
is discretised using 8-noded hexahedral trilinear elements, with 5 elements through the
thickness. Through a mesh sensitivity study, it was found that 46 416 nodes ensured
a sufficiently converged solution. The nodes on the upper surface outside the bulge
diameter of 7.5 cm and the outer edges are held fixed. The pressure loading condition
is applied to the bottom surface. As proposed in the previous Section, only the original
network dimensions a and b, γchain and αelastin are assumed to be variable. The rest of the
parameters are assumed age invariant and are given by L = 2.125, A = 1.82, γgs = 100 Pa,
18
Figure 5: Body axes: the y-axis corresponds to the vertical body axis and the x-axis to the horizontal
body axis. Recreated from Tonge et al. [46].
β = 4.5, θ = 310 K. Furthermore, only the tests for male specimens are simulated in order
to exclude the influence of variations between genders. The age-dependent parameters
that gave the best fit to the published bulge test-data were identified using an iterative
manual process.
Figure 6: Finite element mesh used for the bulge tests.
In Fig. 7, the profile of the skin obtained from the finite element simulation of the
bulge test on the 44 years-old male specimen is given at monotonously-increasing pres-
sure loading from 34.37 Pa to 5.52 kPa. In each figure, the profiles are given along and
perpendicular to the fibre direction v0 as illustrated by the arrows in each subfigure.
The bulge specimen undergoes a rapid initial displacement while the stress state is
19
fibre figure 6xytboundary ring 10cm10cm Figure 7: Displacement profiles for the 44 years-old male skin specimen: (a) pressure = 34.37 Pa; (b)
pressure = 206.4 Pa; (c) pressure = 502.1 Pa; (d) pressure = 1840 Pa; (e) pressure = 5520 Pa.
still within the low modulus portion of the stress-strain curve. This can be seen in the plot
of the pressure versus the in-plane stretch where a small increment in pressure induces
a large change in stretch (Fig. 8). During this phase, the elastin and ground substance
contributions dominate the mechanical response and offer little resistance to inflation.
The divergence of the pressure-stretch lines indicates the activation of the anisotropic
collagen network. In terms of the material model, the collagen energy contribution is
no longer negligible as the end-to-end chain length r sustains significant stretch and
approaches the contour length L. For the age 44 specimen, above 400 Pa further stretch
parallel to the fibre direction no longer occurs whereas perpendicular stretch continues to
increase with pressure. This results in the response shown in Fig. 7(c) where at 502 Pa the
profiles along the perpendicular and parallel directions differ. Mechanical anisotropy is
amplified with increased pressure as the perpendicular stretch increases. As the specimen
is deformed, exponentially more pressure is needed to attain further deformation which
is characterised by the locking type limit for many biological soft tissues and engineered
polymers. In Fig. 7(d) and Fig. 7(e) displacement at high pressure occurs laterally in
the perpendicular fibre direction.
The 8-chain model is now used to simulate the stretch behaviour of the skin at
increasing pressure at increasing age, and the results summarised in Fig. 8. The stretch
is measured parallel and perpendicular to the dominant fibre direction. In general, the
stretch behaviour with increasing age is reasonably captured. The fit at age 43 is notably
poor which suggests that there is an additional directional dependence not captured by
the model or a missing constituent contribution (i.e. an incomplete constitutive model),
or experimental error.
20
Figure9:Displacementprofile,pressure=206.4Pa.Figure10:Displacementprofile,pressure=502.1Pa.23ParameterAge43Age44Age61Age83a3.583.453.553.65b0.50.81.251.287 chn(⇥1022)6m 30.856m 30.0856m 30.856m 3 elas1000Nm 21300Nm 21000Nm 2500Nm 2Table4:Wormlike8-chainmodelparametervaluesforage43,44,61and83bulgetest.Figure8:Displacementprofile,pressure=34.37Pa.InFig.8-Fig.12,theprofileoftheskinobtainedfromthefiniteelementsimulationoftheage44bulgetestisgivenatincreasingstagesofthepressureloading.Ineachfigure,theprofilesaregivenalongandperpendiculartothe22Figure9:Displacementprofile,pressure=206.4Pa.Figure10:Displacementprofile,pressure=502.1Pa.23Figure11:Displacementprofile,pressure=1.84kPa.Figure12:Displacementprofile,pressure=5.52kPa.24Figure11:Displacementprofile,pressure=1.84kPa.Figure12:Displacementprofile,pressure=5.52kPa.24Figure11:Displacementprofile,pressure=1.84kPa.Figure12:Displacementprofile,pressure=5.52kPa.24uv0XYv0XYv0XYv0XYv0XYv0XYv0XYv0XYv0XYv0XY(a) 34.37 Pa(b) 206.4 Pa(c) 502.1 Pa(d) 1840 Pa(e) 5520 Pa Figure 8: Plots of in-plane stretch as a function of inflation pressure for the experimental measurements
of Tonge et al. [46] overlaid over the theoretical curves obtained from the identification of constitutive
parameters.
Excluding the response of the age 43 test, the identified parameters shown in Table 4
are in agreement with the hypotheses advocated in Sec. 2. Note, γchain does not vary
monotonically with age, with the same values obtained for age 44 and 83. This suggests
that γchain is not age dependent. The difference in the initial end-to-end lengths as
dictated by the increased values of a and b allows for the difference in maximum stretches
observed in the experiments. Additionally the ratio between the two changes with age
with a : b = 4.313 : 1, 2.84 : 1, 2.836 : 1 at age 44, 61 and 83 respectively. This reduction,
especially between age 44 and 61 is indicative of the expected reduction in anisotropy.
The value of αelastin decreases almost linearly with age, with an approximate reduction
of 100 N m−2 every five years.
6. Modelling of ageing
The factors discussed in the previous sections provide a means by which the skin
constitutive model (and its associated finite element model replicating bulge tests) can
21
(a)Pressurevsstretch,age43.(b)Pressurevsstretch,age44.(c)Pressurevsstretch,age61.(d)Pressurevsstretch,age83.Figure13:Pressurevsin-planestretchfitstoexperimentaldataatvaryingageextractedfrom[47]fibredirectionasillustratedineachsubfigure.Thebulgespecimenunder-goesarapidinitialdisplacementwhilethestressstateisstillwithinthelowmodulusportion.Thiscanbeseenintheplotofthepressurevsthein-planestretchinFig.13whereasmallincrementinpressureinducesalargechangeinstretch.Duringthisphase,theelastinandgroundsubstancecontribu-tionsdominatetheresponseando↵erlittleresistance.Thedivergenceofthepressure-stretchlinesindicatestheactivationoftheanisotropiccollagennet-work.Intermsofthematerialmodel,thecollagenenergycontributionisnolongernegligibleastheend-to-endparameterrtakesonsignificantstretchandapproachesthecontourlengthL.Fortheage44specimen,above400Pafurtherstretchparalleltothefibredirectionnolongeroccurswhereasperpendicularstretchcontinuestoincreasewithpressure.ThisresultsintheresponseshowninFig.10,whereat502Patheprofilesalongtheperpendic-25(a)age 43 (b) age 44(b) age 61(d) age 83experimental experimental 8-chain 8-chain(a)Pressurevsstretch,age43.(b)Pressurevsstretch,age44.(c)Pressurevsstretch,age61.(d)Pressurevsstretch,age83.Figure13:Pressurevsin-planestretchfitstoexperimentaldataatvaryingageextractedfrom[47]fibredirectionasillustratedineachsubfigure.Thebulgespecimenunder-goesarapidinitialdisplacementwhilethestressstateisstillwithinthelowmodulusportion.Thiscanbeseenintheplotofthepressurevsthein-planestretchinFig.13whereasmallincrementinpressureinducesalargechangeinstretch.Duringthisphase,theelastinandgroundsubstancecontribu-tionsdominatetheresponseando↵erlittleresistance.Thedivergenceofthepressure-stretchlinesindicatestheactivationoftheanisotropiccollagennet-work.Intermsofthematerialmodel,thecollagenenergycontributionisnolongernegligibleastheend-to-endparameterrtakesonsignificantstretchandapproachesthecontourlengthL.Fortheage44specimen,above400Pafurtherstretchparalleltothefibredirectionnolongeroccurswhereasperpendicularstretchcontinuestoincreasewithpressure.ThisresultsintheresponseshowninFig.10,whereat502Patheprofilesalongtheperpendic-25experimental experimental 8-chain 8-chain(a)Pressurevsstretch,age43.(b)Pressurevsstretch,age44.(c)Pressurevsstretch,age61.(d)Pressurevsstretch,age83.Figure13:Pressurevsin-planestretchfitstoexperimentaldataatvaryingageextractedfrom[47]fibredirectionasillustratedineachsubfigure.Thebulgespecimenunder-goesarapidinitialdisplacementwhilethestressstateisstillwithinthelowmodulusportion.Thiscanbeseenintheplotofthepressurevsthein-planestretchinFig.13whereasmallincrementinpressureinducesalargechangeinstretch.Duringthisphase,theelastinandgroundsubstancecontribu-tionsdominatetheresponseando↵erlittleresistance.Thedivergenceofthepressure-stretchlinesindicatestheactivationoftheanisotropiccollagennet-work.Intermsofthematerialmodel,thecollagenenergycontributionisnolongernegligibleastheend-to-endparameterrtakesonsignificantstretchandapproachesthecontourlengthL.Fortheage44specimen,above400Pafurtherstretchparalleltothefibredirectionnolongeroccurswhereasperpendicularstretchcontinuestoincreasewithpressure.ThisresultsintheresponseshowninFig.10,whereat502Patheprofilesalongtheperpendic-25experimental experimental 8-chain 8-chain(a)Pressurevsstretch,age43.(b)Pressurevsstretch,age44.(c)Pressurevsstretch,age61.(d)Pressurevsstretch,age83.Figure13:Pressurevsin-planestretchfitstoexperimentaldataatvaryingageextractedfrom[47]fibredirectionasillustratedineachsubfigure.Thebulgespecimenunder-goesarapidinitialdisplacementwhilethestressstateisstillwithinthelowmodulusportion.Thiscanbeseenintheplotofthepressurevsthein-planestretchinFig.13whereasmallincrementinpressureinducesalargechangeinstretch.Duringthisphase,theelastinandgroundsubstancecontribu-tionsdominatetheresponseando↵erlittleresistance.Thedivergenceofthepressure-stretchlinesindicatestheactivationoftheanisotropiccollagennet-work.Intermsofthematerialmodel,thecollagenenergycontributionisnolongernegligibleastheend-to-endparameterrtakesonsignificantstretchandapproachesthecontourlengthL.Fortheage44specimen,above400Pafurtherstretchparalleltothefibredirectionnolongeroccurswhereasperpendicularstretchcontinuestoincreasewithpressure.ThisresultsintheresponseshowninFig.10,whereat502Patheprofilesalongtheperpendic-25experimental experimental 8-chain 8-chain(a)Pressurevsstretch,age43.(b)Pressurevsstretch,age44.(c)Pressurevsstretch,age61.(d)Pressurevsstretch,age83.Figure13:Pressurevsin-planestretchfitstoexperimentaldataatvaryingageextractedfrom[47]fibredirectionasillustratedineachsubfigure.Thebulgespecimenunder-goesarapidinitialdisplacementwhilethestressstateisstillwithinthelowmodulusportion.Thiscanbeseenintheplotofthepressurevsthein-planestretchinFig.13whereasmallincrementinpressureinducesalargechangeinstretch.Duringthisphase,theelastinandgroundsubstancecontribu-tionsdominatetheresponseando↵erlittleresistance.Thedivergenceofthepressure-stretchlinesindicatestheactivationoftheanisotropiccollagennet-work.Intermsofthematerialmodel,thecollagenenergycontributionisnolongernegligibleastheend-to-endparameterrtakesonsignificantstretchandapproachesthecontourlengthL.Fortheage44specimen,above400Pafurtherstretchparalleltothefibredirectionnolongeroccurswhereasperpendicularstretchcontinuestoincreasewithpressure.ThisresultsintheresponseshowninFig.10,whereat502Patheprofilesalongtheperpendic-25Pressure (kPa)0123451.01.11.21.3StretchPressure (kPa)0123451.01.11.21.3StretchStretchStretch1.01.11.051.151.01.041.08Pressure (kPa)0123450123456Pressure (kPa)figure 131.051.151.251.051.151.251.351.021.061.1 Parameter Age 43
a
b
γchain(×1022)
αelastin
3.58
0.5
6 m−3
1000 N m−2
Age 44
3.45
0.8
0.856 m−3
1300 N m−2
Age 61
3.55
1.25
0.0856 m−3
1000 N m−2
Age 83
3.65
1.287
0.856 m−3
500 N m−2
Table 4: Wormlike 8-chain model parameter values for the age 43, 44, 61 and 83 bulge test.
capture ageing effects in a continuous sense through the modification of a few selected
parameters. Using the a and b parameter values as found from the age 44, 61 and 83
fits, general ageing trends were established. It is clear that a mere three data points is
insufficient in order to conclusively determine an ageing trend, but they are adequate
for the current proof of concept. Shape-preserving fitting algorithms in Matlab (The
MathWorks Inc., Natick, MA, USA) were used to find a continuous evolution of a and
b as shown in Fig. 9(a) and Fig. 9(b). As expected, there is a general increase in both
parameters with age. a evolves almost linearly with age whereas b undergoes a large
increase between age 44 and 61 followed by a plateau. This suggests that between age 44
and 61 there is a more significant loss of anisotropy due to a realignment or redistribution
of the collagen fibres as a network as compared to later in life.
Figure 9: Network dimension and skin thickness trends with age as found through bulge test fits.
The skin thickness data as presented in Table 3 was used to determine a linear equa-
tion for the skin thickness as a function of age. From Fig. 9(c) it can observed that a very
22
(a)Agevsa.(b)Agevsb.(c)Agevsskinthickness.Figure14:Networkdimensionandskinthicknesstrendswithageasfoundthroughbulgetestfits.28(a)Agevsa.(b)Agevsb.(c)Agevsskinthickness.Figure14:Networkdimensionandskinthicknesstrendswithageasfoundthroughbulgetestfits.284550556065707580859045505560657075808590Age (years)Age (years)(a)(b)3.453.53.553.63.653.70.80.911.11.21.31.4abbulge datafit(a)Agevsa.(b)Agevsb.(c)Agevsskinthickness.Figure14:Networkdimensionandskinthicknesstrendswithageasfoundthroughbulgetestfits.2845505560657075808540Age (years)(c)0.150.20.250.30.350.40.45t(mm)figure 14 rough fit has been established due to highly irregular data. The linear fit is sufficient
and is given by
t = −0.0047[mm/year] × age [years] + 0.59 [mm] .
(12)
The overall decrease in skin thickness is consistent with experimental values found
in the literature. With reference to Eq. (1), the gradient of the proposed relation is
comparable. The difference in the y-intercept value is due to Eq. (1) accounting for
thinning from the age of 60 while Eq. (12) does not account for the fact that thinning
does not occur earlier in life.
With the established trends, bulge test simulations were run at 5 year increments.
Table 5 contains the adjustments to the parameters of interest. As mentioned, αelastin
decreases from 1300 N/m2 at age 44 by 100 N/m2 every 5 years.
Age
49
54
59
64
69
74
79
84
89
a
3.483
3.513
3.54
3.564
3.587
3.613
3.634
3.654
3.671
b
0.984
1.14
1.236
1.26
1.273
1.282
1.287
1.288
1.285
αelastin
1200
1100
1000
900
800
700
600
500
400
Table 5: Modification of age dependent model parameters.
In Fig. 10 the simulated evolution of the stretch behaviour with age is given, and
compared to the experimental data at ages 44, 61 and 83. As expected, the general
ageing-induced trend is captured as the skin stiffens with age. The maximum stretch ob-
tained drops significantly between ages 44 and 59, with less significant drops thereafter.
This is to be expected primarily due to the trend in parameter b as given in Fig. 10(b),
where there is an initial sharp increase followed by a plateau. The convergence of the
stretch values between ages 59 and 89 is due to the initial end-to-end length r0 approach-
ing the contour length L. As r0 increases with age, the amount of allowable macroscopic
stretch is reduced.
The contour plots of the displacement components for the age 44 specimen is dis-
played in Fig. 11(I), with the fibre direction along the x-axis. A comparison is given
with the experimental results obtained by Tonge et al. [46]. As expected, the contour
profiles for each displacement component are comparable and the overall behaviour of
the simulations matches the experimental results. The age 44 simulations accurately
capture the displacement along the x-axis Fig. 11(Ia), but the displacements along the y-
and z-axes Fig. 11(Ib-Ic) are slightly overestimated by approximately 0.5 mm and 2 mm,
respectively. This overestimation arises from a constituent contribution not captured
in the model.
It is also possible that the effect of the rig set-up is not appropriately
captured in the simulation. The guard ring may introduce a compressive force through
the thickness that would limit the extent of the deformation.
23
Figure 10: Stretch at the apex of the pressurised skin as a function of applied pressure for 9 age values
(see Table 5); (a) parallel to the fibre direction; (b) normal to the fibre direction.
Figure 11: Plots of deformed profile for the bulge test coloured by the various componenets of the
displacement [mm]. The simulations are compared to the experimental results of Tonge et al. [46] for
various different ages (I–III).
Fig. 11(II-III) show the simulated contour profile of the bulge test for samples accord-
ing to the proposed "ageing" of the various parameters. These profiles are compared to
experimental results at approximately the same age, as extracted from the data obtained
by Tonge et al. [46].
It is important to note that in the experimental specimens the
Langer lines were not aligned with the body axes. The Langer lines were oriented at
64◦ and −24.6◦ degrees to the x-axis for the age 61 and 83 specimens, respectively, as
indicated in the figures. This alteration in the orientation is reflected in the simulated
results.
The profile comparison between the age 64 simulated and age 61 experimental results
are very similar to those of age 44. The alteration in the fibre orientation is captured quite
24
Withtheestablishedtrends,bulgetestsimulationswererunat5yearincrements.Table5containstheadjustmentstotheparametersofinterest.Asmentioned, elasdecreasesfrom1300Nm 2atage44by100Nm 2every5years.ParametersAgeab elas493.4830.9841200543.5131.141100593.541.2361000643.5641.26900693.5871.273800743.6131.282700793.6341.287600843.6541.288500893.6711.285400Table5:Modificationofagedependentmodelparameters.Figure15:Pressurevsin-planestretchsimulationsatincreasingageaccordingtoproposedageingmodel.Left:Parallelstretch.Right:Perpendicularstretch.InFig.15thesimulatedevolutionofthestretchbehaviourwithageisgiven,andcomparedtotheexperimentaldataatages44,61and83.As29Withtheestablishedtrends,bulgetestsimulationswererunat5yearincrements.Table5containstheadjustmentstotheparametersofinterest.Asmentioned, elasdecreasesfrom1300Nm 2atage44by100Nm 2every5years.ParametersAgeab elas493.4830.9841200543.5131.141100593.541.2361000643.5641.26900693.5871.273800743.6131.282700793.6341.287600843.6541.288500893.6711.285400Table5:Modificationofagedependentmodelparameters.Figure15:Pressurevsin-planestretchsimulationsatincreasingageaccordingtoproposedageingmodel.Left:Parallelstretch.Right:Perpendicularstretch.InFig.15thesimulatedevolutionofthestretchbehaviourwithageisgiven,andcomparedtotheexperimentaldataatages44,61and83.As29Stretch at apex (mm)Pressure (kPa)0123451.01.11.051.15increasing ageStretch at apex (mm)Withtheestablishedtrends,bulgetestsimulationswererunat5yearincrements.Table5containstheadjustmentstotheparametersofinterest.Asmentioned, elasdecreasesfrom1300Nm 2atage44by100Nm 2every5years.ParametersAgeab elas493.4830.9841200543.5131.141100593.541.2361000643.5641.26900693.5871.273800743.6131.282700793.6341.287600843.6541.288500893.6711.285400Table5:Modificationofagedependentmodelparameters.Figure15:Pressurevsin-planestretchsimulationsatincreasingageaccordingtoproposedageingmodel.Left:Parallelstretch.Right:Perpendicularstretch.InFig.15thesimulatedevolutionofthestretchbehaviourwithageisgiven,andcomparedtotheexperimentaldataatages44,61and83.As29increasing age1.01.11.21.31.051.151.251.35Pressure (kPa)012345(a)(b)figure 15(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure17:Displacementcomparison.Left:Experimentaldata,age61.Right:8-chainsimulationage6432figure 17-2002040-40-20020-2002040-40-20020-202-201-124681012(a)(b)(c)-20020-2002040-40YXYXYX(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure17:Displacementcomparison.Left:Experimentaldata,age61.Right:8-chainsimulationage6432(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure17:Displacementcomparison.Left:Experimentaldata,age61.Right:8-chainsimulationage6432(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure17:Displacementcomparison.Left:Experimentaldata,age61.Right:8-chainsimulationage6432-3.3-1.6501.653.3-2-10120481216uzuxuy(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure16:Displacementcomparison.Left:Experimentaldata,age44.Right:8-chainsimulation,age44.31(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure16:Displacementcomparison.Left:Experimentaldata,age44.Right:8-chainsimulation,age44.31Y-20020X-20020-2.5-1.2501.252.5Y-20020X-20020Y-20020X-20020(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure16:Displacementcomparison.Left:Experimentaldata,age44.Right:8-chainsimulation,age44.31-8-4048uzuxuy(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure16:Displacementcomparison.Left:Experimentaldata,age44.Right:8-chainsimulation,age44.3105.7511.517.2523(a)(b)(c)-202-50551015figure 16figure 18(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure18:Displacementcomparison.Left:Experimentaldata,age83.Right:8-chainsimulation,age84.33-20020-2002040-40YX-20020-2002040-40YX-20020-2002040-40YX-1021-102100.51.512(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure18:Displacementcomparison.Left:Experimentaldata,age83.Right:8-chainsimulation,age84.33(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure18:Displacementcomparison.Left:Experimentaldata,age83.Right:8-chainsimulation,age84.33(a)Ucomponentofdisplacement.(b)Vcomponentofdisplacement.(c)Wcomponentofdisplacement.Figure18:Displacementcomparison.Left:Experimentaldata,age83.Right:8-chainsimulation,age84.33-1-0.500.51-2.5-1.2501.252.5-2.5-1.2501.252.5(a)(b)(c)uzuxuyExperiment Age 44Simulation Age 44Experiment Age 61Simulation Age 64Experiment Age 83Simulation Age 84IIIIII satisfactorily, with the resulting axial assymmetry in the components of displacement
along the x- and y-axes represented accurately. The z component of displacement in
the out-of-plane direction is acceptably captured, although there seems to be an over
estimation in the level of anisotropy in the simulated results. In terms of the magnitude of
the displacements, the y-axis components of the age 61 experiment and age 64 simulation
are similar, but the simulations overestimate the other two components. This is not too
surprising considering the unsatisfactory fit with the pressure-stretch data at age 61 as
shown in Fig. 10.
The age 84 simulations were orientated at −24.6◦ to the positive x-axis in order
to replicate the age 83 experimental contour plots. There is sufficient agreement in the
general profile of the simulated results when compared to the experimental data. The only
discrepancy lies in the component along the x-axis of the experimental data as shown in
Fig. 11(IIIa), where the lack of symmetry suggests a possible defect in the skin specimen,
such as a non-uniform skin thickness or an irregularity on a constituent level, such as
an inconsistent collagen distribution or dispersion. Despite this, the displacements along
the x- and y-axes of the age 84 simulated results are comparable to the experimental
results, but there is again a minor overestimation in the z-axis component.
7. Discussion
The impact of ageing on the mechanical response of human skin was investigated us-
ing a microstructurally-motivated constitutive continuum description featuring distinct
phases representing the ground substance, elastin and collagen. The ground substance
and elastin contributions were modelled by an isotropic, coupled, and compressible neo-
Hookean free energy. For collagen, the primary load-bearing constituent of skin in ten-
sion, a wormlike eight-chain model motivated by the fibrous and intertwined nature of
the collagen network was selected. This approach captures well the inherent nonlinearity
and anisotropy of the skin under finite deformations.
A subset of the constitutive parameters whose age-dependent variation has the most
physically justified influence on the response was identified. The parameters controlled
age-related structural changes such as loss of integrity within the elastin network, dermal
thinning and unravelling of the collagen network. The continuum model was then used to
predict the response of bulge tests which, compared to uniaxial extension tests, provide
a richer set of deformations and stress states to evaluate the model performance. The
proposed constitutive model was found to adequately capture the mechanical response.
By experimentally-informed modification to the constitutive parameters the ageing
process was successfully replicated. The reduction in stiffness at low stretches (low
modulus portion) through degradation of the elastin network was well captured through
modification of the elastin free energy. Results show that degradation of the elastin
meshwork and variations in anisotropy of the collagen network are plausible mechanisms
to explain ageing in terms of macroscopic tissue stiffening. Whereas alterations in elastin
affect the low-strain region of the skin stress-strain curve, those related to collagen have
an impact on its (large strain) linear region.
As with any modelling study, and despite the ability of the proposed constitutive
model to capture ageing effects, some limitations can be identified and form the basis for
future improvements. In order to incorporate an age-based modification to the model
microstructural constitutive parameters, experimental bulge test data was used at ages
25
44, 61 and 83. At only three sample points, this does not represent an adequate range of
data with which to make a conclusive parameter fit. Therefore, for a more comprehensive
ageing model to be developed, it would be necessary to include data from a statistically
significant number of specimens over a large range of age groups. Additionally, experi-
mental samples would need to be controlled for several factors so as to limit variability,
or at least, to characterise it so it could be accounted for in the modelling. Of significant
importance would be control over anatomical site. Not only should skin from the same
site of the patients be tested, but also controlled for segregating the respective influence
of intrinsic ageing and photoageing.
The model does not explicitly account for crosslinking within the collagen network,
and therefore, cannot provide a mechanistic description of this effect. Crosslinking pre-
vents slippage between fibres and accordingly contributes toward the stiffness elicited by
the collagen network under stretch [97]. With age, crosslinking has been observed to
increase which has often been attributed as a possible source of the increase in macro-
scopic stiffness [98]. Here, more complex chemomechanobiological processes are at play
and increase in crosslinking might rather be a by-product of UV radiation exposure and
glycation associated with diabetes [27] rather than the consequence of intrinsic ageing
per se. Modifications of crosslinking properties are implicitly captured in the eight-chain
model. For a more explicit description of fibre-to-fibre and matrix-to-fibre mechani-
cal/physical interactions, other approaches are possible. One could use dedicated tensor
invariants [99, 100] that segregate deformation modes associated with such interactions or
one could use multiscale mechanistic micromechanical constitutive models that explicitly
describe these interactions [101–103].
The mechanical response of the collagen meshwork was captured using an eight-chain
entropic model based on a wormlike chain energy. This type of model is rooted in
statistical mechanics where the assumption is that long chain macromolecular networks
can take a very large number of possible conformations. This assumption might be
questionable as it is unclear whether the highly-bonded and hierarchical structure of
collagen assemblies allows degrees of freedom for all states to be sampled in an entropy-
dominated regime. Moreover, the wormlike chain is assumed to represent an individual
fibre or a bundle of collagen fibres. At that length scale, it is expected that enthalpic
effects would be dominant over entropic effects. However, energetic elasticity models can
also model the typical strain hardening behaviour of biological soft tissues by uncoiling
of crimps [104]. Other types of microstructurally-motivated strain energy function for
the individual chains of the unit cell could also be used [77, 105].
The skin was modelled here as single homogeneous layer as this was supported by
the reasonable assumption that the dermal layer is the main mechanical contributor of
the skin under states of tension. In order to develop a more realistic model it would be
relevant to incorporate a multi-layer model with realistic structural geometries [3, 30].
It is worth highlighting that, with age, although the skin undergoes drastic alterations
of its own biophysical properties, particularly mechanical properties, distinct ageing pro-
cesses occur in other body tissues, structures and organs. They have the potential to alter
the mechanical environment experienced by the skin. For example, thinning/atrophy of
adipose and muscular tissues [98], or even bone resorption, are factors that can mod-
ify the complex mechanobiology and associated residual strains of the skin as an organ
covering the entire body. This is consistent with observations that the magnitude and
directions of Langer lines [55] vary with age. These aspects were not directly accounted
26
for in the present constitutive model and could explain the inability to obtain an ex-
cellent fit between experimental and model data for the inflation tests eliciting the skin
response along the main fibre direction (Fig. 9). Also, with age, the skin tends to lose
its in-plane isotropy [29] because of the strong mechanical effects introduced by der-
mal collagen realignment arising in combination with collagen cross-linking and density
alteration.
In intrinsic skin ageing, the dermal-epidermal junction (DEJ) which is a 0.5-1 µm thick
three-dimensional interlocking wavy interface basement membrane [106] connecting the
dermis to the epidermis, tends to flatten out [107–111]. This reduction in the amplitude
of the papillae which are the protrusions of the dermis into the epidermis (see Fig. 1),
is also accompanied by a decrease in their density (unit per surface area) [108, 109].
This alters the mechanical interactions between the epidermis and dermis resulting in
decreased resistance to shear deformation at the DEJ. In turn, this is likely to modulate
the macroscopic mechanical response of the skin as evidenced by the increasing frailty
of skin with age [7]. Besides these important structural effects, the flattening of the
DEJ has important metabolic consequences as it reduces the surface area for nutritional
exchange and metabolic by-products evacuation between the dermis and epidermis. As
a result, epidermal cell turnover is slowed down and free radicals accumulate.
At this stage, the model does not account for ageing-triggered enzymatic degradation
of elastic fibres, abnormal collagen deposition and remodelling. At the sub-cellular and
cellular levels, there exists a variety of enzymes and cytokines that are active in the
ageing of the skins primary constituents. Intrinsic ageing is accompanied by a reduction
of the fibroblast population [112] which induces a decrease in collagen production and is
associated with an increased build-up of matrix metalloproteinases (MMP) [113]. MMP
are enzymes that can cleave elastic fibre molecules [113, 114] and are believed to play
an active role in degrading important structural proteins in the skin such as collagen,
elastin, fibronectin and laminin. Disruption of the homeostatic state between activation
and inhibition of MMPs is a key ingredient in the pathophysiology of intrinsic skin
ageing and photoageing by controlling the degradation of collagen and elastin in the
skin. Reactive oxygen species (ROS) are known to play a crucial part in driving the
ageing process. Such chemical species could form the building blocks of mixture type
models [115–118] or models based on open-system thermodynamics [91, 119].
The number of processes, their interplay, and the scales at which these occur make
deconstructing skin ageing a very challenging task at both the experimental and mod-
elling levels. Additionally, the effects of intrinsic ageing are generally compounded by
the influence of extrinsic ageing (e.g. photoageing), which only serves to add complexity
to the problem. In summary, much work remains to be done to develop a continuum
description of skin capable of incorporating the full ageing process. These aspects are
currently being implemented in a new chemomechanobiological constitutive model which
will be the subject of subsequent publications.
8. Acknowledgements
This work was funded through the award of a Royal Society Newton Fund grant
(2014-2016) between the Universities of Southampton and Cape Town. The authors
would like to gratefully acknowledge this financial support as well as the logistic and
27
infrastructure support provided by their respective institutions for research visits of DP,
AM and GL.
9. References
References
[1] J. W. Y. Jor, M. D. Parker, A. J. Taberner, M. P. Nash, and P. M. F. Nielsen. Computational
and experimental characterization of skin mechanics:
identifying current challenges and future
directions. Wiley Interdisciplinary Reviews: Systems Biology and Medicine, 5(5):539–556, 2013.
[2] Wenguang Li. Modelling methods for in vitro biomechanical properties of the skin: A review.
Biomedical Engineering Letters, 5(4):241–250, 2015.
[3] G. Limbert. Mathematical and computational modelling of skin biophysics-a review. Proceedings
of the Royal Society A-Mathematical Physical and Engineering Sciences, 473(2203):1–39, 2017.
[4] F. H. Silver, L. M. Siperko, and G. P. Seehra. Mechanobiology of force transduction in dermal
tissue. Skin Research and Technology, 9(1):3–23, 2003.
[5] www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/ageing/articles/
characteristicsofolderpeople/2013-12-06, 2010.
[6] www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/
articles/overviewoftheukpopulation/february2016, 2016.
[7] P. Morey. Skin tears: a literature review. Primary Intention, 15(3):122–129, 2007.
[8] H. Shimizu. Shimizu's Textbook of Dermatology. Hokkaido University Press - Nakayama Shoten
Publishers, 2007.
[9] Y. C. Fung. Biomechanics: Mechanical Properties of Living Tissues. Springer-Verlag, New York,
1981.
[10] G. Limbert. Chapter 4: State-of-the-art constitutive models of skin biomechanics, chapter 4, pages
95–131. Pan Stanford Publishing Pte. Ltd, Singapore, 2014.
[11] Y. Lanir. Skin mechanics. McGraw-Hill, New York, 1987.
[12] M. Ramos-e Silva and S. C. da Silva Carneiro. Elderly skin and its rejuvenation: products and
procedures for the aging skin. Journal of Cosmetic Dermatology, 6(1):40–50, 2007.
[13] H. Assaf, M. A. Adly, and M. R. Hussein. Aging and Intrinsic Aging: Pathogenesis and Manifes-
tations, book section 13, pages 129–138. 2010.
[14] E. C. Naylor, R. E. B. Watson, and M. J. Sherratt. Molecular aspects of skin ageing. Maturitas,
69(3):249–256, 2011.
[15] B. A. Gilchrest and M. Yaar. Ageing and photoageing of the skin: observations at the cellular and
molecular level. British Journal of Dermatology, 127 Suppl 41:25–30, 1992.
[16] D. Goukassian, F. Gad, M. Yaar, M. S. Eller, U. S. Nehal, and B. A. Gilchrest. Mechanisms and
implications of the age-associated decrease in dna repair capacity. FASEB J, 14(10):1325–34, 2000.
[17] P. U. Giacomoni and G. Rein. A mechanistic model for the aging of human skin. Micron, 35(3):
179–84, 2004.
[18] G. J. Fisher, Z. Wang, S. C. Datta, J. Varani, S. Kang, and J. J. Voorhees. Pathophysiology of
premature skin aging induced by ultraviolet light. New England Journal of Medicine, 337(20):
1419–1429, 1997.
[19] M. Berneburg, H. Plettenberg, and J. Krutmann. Photoaging of human skin. Photodermatol
Photoimmunol Photomed, 16(6):239–44, 2000.
[20] A. M. Kligman. Early destructive effects of sunlight on human skin. Journal of the American
Medical Association, 210:2377–2380, 1969.
[21] G. J. Fisher, S. Kang, J. Varani, Z. Bata-Csorgo, Y. Wan, S. Datta, and J. J. Voorhees. Mechanisms
of photoaging and chronological skin aging. Archives of Dermatology, 138(11):1462–70, 2002.
[22] B. L. Diffey. A quantitative estimate of melanoma mortality from ultraviolet a sunbed use in the
uk. British Journal of Dermatology, 149(3):578–581, 2003.
[23] J. Sandby-Moller, T. Poulsen, and H. C. Wulf. Epidermal thickness at different body sites: rela-
tionship to age, gender, pigmentation, blood content, skin type and smoking habits. Acta Dermato
Venereologica, 83(6):410–3, 2003.
[24] A. Vierktter and J. Krutmann. Environmental influences on skin aging and ethnic-specific mani-
festations. Dermato-endocrinology, 4(3):227–231, 2012.
[25] A. V. Benedetto. The environment and skin aging. Clinical Dermatology, 16(1):129–139, 1998.
28
[26] M. Pawlaczyk, M. Lelonkiewicz, and M. Wieczorowski. Age-dependent biomechanical properties
of the skin. Postepy Dermatologii i Alergologii, 30(5):302–306, 2013.
[27] M. J. Sherratt. Age-related tissue stiffening: Cause and effect. Advances in Wound Care, 2(1):
11–17, 2013.
[28] F. H. Silver, G. Seehra, J. W. Freeman, and D. DeVore. Viscoelastic of young and old human
dermis: a proposed moelcular mechanism for elastic energy storage in collagen and elastin. Journal
of Applied Polymer Science, 86:1978–1985, 2002.
[29] E. C. Ruvolo Jr, G. N. Stamatas, and N. Kollias. Skin viscoelasticity displays site- and age-
dependent angular anisotropy. Skin Pharmacology and Physiology, 20(6):313–321, 2007.
[30] M. F. Leyva-Mendivil, J. Lengiewicz, A. Page, N. W. Bressloff, and G. Limbert. Skin microstruc-
ture is a key contributor to its friction behaviour. Tribology Letters, 65(1):12, 2017.
[31] M. F. Leyva-Mendivil, J. Lengiewicz, A. Page, N. W. Bressloff, and G. Limbert. Implications of
multi-asperity contact for shear stress distribution in the viable epidermis - an image-based finite
element study. Biotribology, page in press, 2017.
[32] M. F. Leyva-Mendivil, A. Page, N. W. Bressloff, and G Limbert. A mechanistic insight into the
mechanical role of the stratum corneum during stretching and compression of the skin. Journal
of the Mechanical Behavior of Biomedical Materials, 49:197–219, 2015.
[33] E. Hahnel, A. Lichterfeld, U. Blume-Peytavi, and J. Kottner. The epidemiology of skin conditions
in the aged: A systematic review. Journal of Tissue Viability, 26(1):20–28, 2017.
[34] A. Goriely and M. Ben Amar. On the definition and modeling of incremental, cumulative, and
continuous growth laws in morphoelasticity. Biomechanics and Modeling in Mechanobiology, 6(5):
289–296, 2007.
[35] D. Ambrosi, G. A. Ateshian, E. M. Arruda, S. C. Cowin, J. Dumais, A. Goriely, G. A. Holzapfel,
J. D. Humphrey, R. Kemkemer, E. Kuhl, J. E. Olberding, L. A. Taber, and K. Garikipati. Per-
spectives on biological growth and remodeling. Journal of the Mechanics and Physics of Solids,
59(4):863–883, 2011.
[36] F. Xu and T. Lu. Introduction to Skin Biothermomechanics and Thermal Pain. Springer, Hei-
delberg Dordrecht London New York, 2011.
[37] C. Flynn. Fiber-matrix models of the dermis, book section 5, pages 133–159. Pan Stanford
Publishing Pty Ltd, Singapore, 2014.
[38] E. Kuhl, K. Garikipati, E. Arruda, and K. Grosh. Remodeling of biological tissue: Mechanically
induced reorientation of a transversely isotropic chain network. Journal of the Mechanics and
Physics of Solids, 53:1552–1573, 2005.
[39] E. Kuhl and G. A. Holzapfel. A continuum model for remodeling in living structures. Journal of
Materials Science, 42(21):8811–8823, 2007. ISSN 0022-2461.
[40] K. Garikipati, E. M. Arruda, K. Grosh, H. Narayanan, and S. Calve. A continuum treatment
of growth in biological tissue: the coupling of mass transport and mechanics. Journal of the
Mechanics and Physics of Solids, 52(7):1595–1625, 2004.
[41] E. Mazza, O. Papes, M. B. Rubin, S. R. Bodner, and N. S. Binur. Nonlinear elastic-viscoplastic
constitutive equations for aging facial tissues. Biomechanics and Modeling in Mechanobiology, 4
(2-3):178–189, 2005.
[42] E. Mazza, O. Papes, M. B. Rubin, S. R. Bodner, and N. S. Binur. Simulation of the aging face.
Journal of Biomechanical Engineering-Transactions of the Asme, 129(4):619–623, 2007.
[43] M. B. Rubin and S. R. Bodner. A three-dimensional nonlinear model for dissipative response of
soft tissue. International Journal of Solids and Structures, 39(19):5081–5099, 2002.
[44] F. Maceri, M. Marino, and G. Vairo. Age-dependent arterial mechanics via a multiscale elastic ap-
proach. International Journal for Computational Methods in Engineering Science and Mechanics,
14(2):141–151, 2013.
[45] T. K. Tonge, L. S. Atlan, L. M. Voo, and T. D. Nguyen. Full-field bulge test for planar anisotropic
tissues: Part I experimental methods applied to human skin tissue. Acta Biomaterialia, 9(4):
5913–5925, 2013.
[46] T. K. Tonge, L. M. Voo, and T. D. Nguyen. Full-field bulge test for planar anisotropic tissues: Part
II a thin shell method for determining material parameters and comparison of two distributed
fiber modeling approaches. Acta Biomaterialia, 9(4):5926–5942, 2013.
[47] F. H. Silver, J. W. Freeman, and D. DeVore. Viscoelastic properties of human skin and processed
dermis. Skin Research and Technology, 7(1):18–23, 2001.
[48] J. L. L´eveque, J. de Rigal, P. G. Agache, and C. Monneur. Influence of ageing on the in vivo
extensibility of human skin at a low stress. Archives of Dermatological Research, 269(2):127–135,
1980.
29
[49] R. Reihsner, B. Balogh, and E. J. Menzel. Two-dimensional elastic properties of human skin in
terms of an incremental model at the in vivo configuration. Medical Engineering and Physics, 17
(4):304–313, 1995.
[50] H. Oxlund and T. T. Andreassen. The roles of hyaluronic acid, collagen and elastin in the me-
chanical properties of connective tissues. Journal of Anatomy, 131(4):611–620, 1980.
[51] H. Oxlund, J. Manschot, and A. Viidik. The role of elastin in the mechanical properties of skin.
Journal of Biomechanics, 21(3):213–218, 1988.
[52] C. Pailler-Mattei, S. Bec, and H. Zahouani.
In vivo measurements of the elastic mechanical
properties of human skin by indentation tests. Medical Engineering & Physics, 30(5):599–606,
2008.
[53] J. F. Ribeiro, E. H. M. dos Anjos, Maria Luiza S. Mello, and B. de Campos Vidal. Skin collagen
fiber molecular order: a pattern of distributional fiber orientation as assessed by optical anisotropy
and image analysis. PLOS ONE, 8(1):e54724, 2013.
[54] J. Gosline, M. Lillie, E. Carrington, P. Guerette, C. Ortlepp, and K. Savage. Elastic proteins:
biological roles and mechanical properties. Philosophical Transactions of the Royal Society of
London. Series B: Biological Sciences, 357(1418):121–132, 2002.
[55] K. Langer. On the anatomy and physiology of the skin: II. Skin tension. British Journal of Plastic
Surgery, 31(2):93–106, 1978.
[56] R. J. Lapeer, P. D. Gasson, and V. Karri. Simulating plastic surgery: From human skin tensile
tests, through hyperelastic finite element models to real-time haptics. Progress in Biophysics and
Molecular Biology, 103(2-3):208–216, 2010.
[57] F. H. Silver, Y. P. Kato, M. Ohno, and A. J. Wasserman. Analysis of mammalian connective
tissue: relationship between hierarchical structures and mechanical properties. Journal of long-
term effects of medical implants, 2(2-3):165–198, 1992.
[58] R. T. Tregear. The mechanical properties of skin. Journal of the Society of Cosmetic Chemists,
20:467–477, 1969.
[59] C. W. Oomens, D. H. van Campen, and H. J. Grootenboer. In vitro compression of a soft tissue
layer on a rigid foundation. Journal of Biomechanics, 20(10):923–935, 1987.
[60] C. Escoffier, J. de Rigal, A. Rochefort, R. Vasselet, J. L. Lvque, and P. G. Agache. Age-related
mechanical properties of human skin: an in vivo study. The Journal of Investigative Dermatology,
93(3):353–357, 1989.
[61] S. Diridollou, V. Vabre, M. Berson, L. Vaillant, D. Black, J. M. Lagarde, J. M. Gregoire, Y. Gall,
and F. Q. Patat. Skin ageing: changes of physical properties of human skin in vivo. International
Journal of Cosmetic Science, 23(6):353–62, 2001.
[62] H. A. Oriba, D. A. Bucks, and H. I. Maibach. Percutaneous absorption of hydrocortisone and
testosterone on the vulva and foreaarm: effect of the menopause and site. British Journal of
Dermatology, 134:229–233, 1996.
[63] K. O. Duncan and D. J. Lefell. Preoperative assessment of eth elderly patient. Dermatologic
Clinics, 15:583–593, 1997.
[64] M. P. Brincat, S. Kabalan, J. W. Stud, C. F. Moniz, J. de Trafford, and J. Montgomery. A study of
the decrease of skin collagen content, skin thickness, and bone mass in the postmeopausal women.
Obstretic Gynecology, 70(6):840–845, 1987.
[65] H. Alexander and T. Cook. Variations with age in the mechanical properties of human skin in
vivo. Journal of Tissue Viability, 16(3):6–11, 2006.
[66] F. Henry, C. Pi´erard-Franchimont, G. Cauwenbergh, and G. E. Pi´erard. Age-related changes in
facial skin contours and rheology. Journal of the American Geriatrics Society, 45(2):220–222,
1997.
[67] Colin H. Daly and George F. Odland. Age-related changes in the mechanical properties of human
skin. Journal of Investigative Dermatology, 73(1):84–87, 1979.
[68] J. H. Chung, S. Kang, J. Varani, J. Lin, G. J. Fisher, and J. J. Voorhees. Decreased extracellular-
signal-regulated kinase and increased stress-activated map kinase activities in aged human skin in
vivo. Journal of Investigative Dermatology, 115(2):177–182, 2000.
[69] G. Jenkins. Molecular mechanisms of skin ageing. Mechanisms of Ageing and Development, 123
(7):801–10, 2002.
[70] P. G. Agache, C. Monneur, J. L. L´eveque, and J. de Rigal. Mechanical properties and Young's
modulus of human skin in vivo. Archives of Dermatological Research, 269:221–232, 1980.
[71] A. Vexler, I. Polyansky, and R. Gorodetsky. Evaluation of skin viscoelasticity and anisotropy by
measurement of speed of shear wave propagation with viscoelasticity skin analyzer. Journal of
Investigative Dermatology, 113(5):732–739, 1999.
30
[72] G. A. Holzapfel. Nonlinear Solid Mechanics. A Continuum Approach for Engineering. John Wiley
& Sons, Chichester, UK, 2000.
[73] J. E. Marsden and T. J. R. Hughes. Mathematical Foundations of Elasticity. Dover, New-York,
1994.
[74] J. P. Boehler. Lois de comportement anisotrope des milieux continus. Journal de M´ecanique,
(17):153–190.
[75] A. J. M. Spencer. Continuum theory of the mechanics of fibre-reinforced composites. Springer-
Verlag, New York, 1992.
[76] G. Limbert and M. Taylor. On the constitutive modeling of biological soft connective tissues.
a general theoretical framework and tensors of elasticity for strongly anisotropic fiber-reinforced
composites at finite strain. International Journal of Solids and Structures, 39(8):2343–2358, 2002.
[77] J. Schroder and P. Neff. Invariant formulation of hyperelastic transverse isotropy based on poly-
convex free energy functions. International Journal of Solids and Structures, 40(2):401–445, 2003.
[78] M. Destrade, B. MacDonald, J. G. Murphy, and G. Saccomandi. At least three invariants are
necessary to model the mechanical response of incompressible, transversely isotropic materials.
Computational Mechanics, 52(4):959–969, 2013.
[79] E. M. Arruda and M. C. Boyce. A three-dimensional constitutive model for the large stretch
behavior of rubber elastic-materials. Journal of the Mechanics and Physics of Solids, 41(2):389–
412, 1993.
[80] P. J. Flory. Statistical mechanics of chain molecules. John Wiley & Sons, Chichester-New York,
1969.
[81] J. E. Bischoff, E. A. Arruda, and K. Grosh. A microstructurally based orthotropic hyperelastic
constitutive law. Journal of Applied Mechanics-Transactions of the ASME, 69(5):570–579, 2002.
[82] O. Kratky and G. Porod. Rontgenuntersuchungen geloster fadenmolekule. Recueil des Travaux
Chimiques des Pays-Bas et de la Belgique, 68:1106–1122, 1949.
[83] J. F. Marko and E. D. Siggia. Stretching dna. Macromolecules, 28(26):8759–8770, 1995.
[84] A. Buganza Tepole, A. K. Gosain, and E. Kuhl. Stretching skin: The physiological limit and
beyond. International Journal of Non-Linear Mechanics, 47(8):938–949, 2012.
[85] C. Flynn, A. J. Taberner, P. M. F. Nielsen, and S. Fels. Simulating the three-dimensional defor-
mation of in vivo facial skin. Journal of the Mechanical Behavior of Biomedical Materials, 28(0):
484–494, 2013.
[86] C. Flynn and B. A. O. McCormack. A simplified model of scar contraction. Journal of Biome-
chanics, 41(7):1582–1589, 2008.
[87] C. O. Flynn and B. A. O. McCormack. A three-layer model of skin and its application in simulating
wrinkling. Computer Methods in Biomechanics and Biomedical Engineering, 12(2):125–134, 2009.
[88] J. E. Bischoff, E. M. Arruda, and K. Grosh. A rheological network model for the continuum
anisotropic and viscoelastic behavior of soft tissue. Biomechanics and Modeling in Mechanobiology,
3(1):56–65, 2004.
[89] P. S´aez, E. Pea, M. A. Martnez, and E. Kuhl. Computational modeling of hypertensive growth in
the human carotid artery. Computational Mechanics, 53(6):1183–1196, 2013.
[90] W. Kuhn. Beziehungen zwischen Molekhlgrsse, statistischer Molekulgestalt und elastischen eigen-
schaften hochpolymerer Stoffe. Kolloid-Zeitschrift,, 76:258–271, 1936.
[91] E. Kuhl, A. Menzel, and P. Steinmann. Computational modeling of growth - A critical review, a
classification of concepts and two new consistent approaches. Computational Mechanics, 32(1-2):
71–88, 2003.
[92] C. M. Kielty, M. J. Sherratt, and C. A. Shuttleworth. Elastic fibres. Journal of Cell Science, 115
(14):2817–2828, 2002.
[93] J. Korelc. Multi-language and multi-environment generation of nonlinear finite element codes.
Engineering with Computers, 18(4):312–327, 2002.
[94] J. Korelc and P. Wriggers. Automation of Finite Element Methods. Springer, first edition, 2016.
[95] J. W. Y. Jor, M. P. Nash, P. M. F. Nielsen, and P. J. Hunter. Estimating material parameters
of a structurally based constitutive relation for skin mechanics. Biomechanics and Modeling in
Mechanobiology, 10(5):767–778, 2011.
[96] Y. A. Kvistedal and P. M. F. Nielsen. Estimating material parameters of human skin in vivo.
Biomechanics and Modeling in Mechanobiology, 8(1):1–8, 2009.
[97] Y. Tang, R. Ballarini, M. J. Buehler, and S. J. Eppell. Deformation micromechanisms of collagen
fibrils under uniaxial tension. Journal of the Royal Society Interface, 7(46):839–850, 2010.
[98] P. Quatresooz, L. Thirion, C. Pi´erard-Franchimont, and G. E. Pi´erard. The riddle of genuine skin
microrelief and wrinkles. International Journal of Cosmetic Science, 28(6):389–395, 2006.
31
[99] G. Limbert. A mesostructurally-based anisotropic continuum model for biological soft tissues–
decoupled invariant formulation. Journal of the Mechanical Behavior of Biomedical Materials, 4
(8):1637–1657, 2011.
[100] X. Q. Peng, Z. Y. Guo, and B. Moran. An anisotropic hyperelasticconstitutive model with fiber-
matrix shear interaction for the human annulus fibrosus. Journal of Applied Mechanics, 73(5):
815–824, 2005.
[101] F. Maceri, M. Marino, and G. Vairo. A unified multiscale mechanical model for soft collagenous
tissues with regular fiber arrangement. Journal of biomechanics, 43(2):355–363, 2010.
[102] M. Marino and G. Vairo. Multiscale Elastic Models of Collagen Bio-structures: From Cross-Linked
Molecules to Soft Tissues, pages 73–102. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013.
[103] M. Marino, G. Vairo, and P. Wriggers. Multiscale hierarchical mechanics in soft tissues. Proceedings
in Applied Mathematics and Mechanics, 15(1):35–38, 2015.
[104] K. Garikipati, S. Goktepe, and C. Miehe. Elastica-based strain energy functions for soft biological
tissue. Journal of the Mechanics and Physics of Solids, 56(4):1693–1713, 2008.
[105] G. A. Holzapfel, T. C. Gasser, and R. W. Ogden. A new constitutive framework for arterial wall
mechanics and a comparative study of material models. Journal of Elasticity, 61:1–48, 2000.
[106] L. S. Chan. Human skin basement membrane in health and autoimmune diseases. Frontiers in
Bioscience, 2(July 15):343–352, 1997.
[107] R. M. Lavker. Structural alterations in exposed and unexposed aged skin. Journal of Investigative
Dermatology, 73:59–66, 1979.
[108] S. G. Lagarrigue, J. George, E. Questel, C. Lauze, N. Meyer, J. M. Lagarde, M. Simon, A. M.
Schmitt, G. Serre, and C. Paul.
In vivo quantification of epidermis pigmentation and dermis
papilla density with reflectance confocal microscopy: variations with age and skin phototype.
Experimental dermatology, 21(4):281–286, 2012.
[109] K. Sauermann, S. Clemann, S. Jaspers, T. Gambichler, P. Altmeyer, K. Hoffmann, and J. Ennen.
Age related changes of human skin investigated with histometric measurements by confocal laser
scanning microscopy in vivo. Skin Research and Technology, 8(1):52–56, 2002.
[110] B. Querleux, T. Baldeweck, S. Diridollou, J. de Rigal, E. Huguet, F. Leroy, and V. H. Barbosa.
Skin from various ethnic origins and aging: an in vivo cross-sectional multimodality imaging study.
Skin Research and Technology, 15(3):306–313, 2009.
[111] A. M. Kligman, P. Zheng, and R. M. Lavker. The anatomy and pathogenesis of wrinkles. British
Journal of Dermatology, 113:37–42, 1985.
[112] N. A. Fenske and C. W. Lober. Structural and functional changes of normal aging skin. Journal
of the American Academy of Dermatology, 15(4(1)):571–585, 1986.
[113] J. Varani, D. Spearman, P. Perone, S. E. Fligiel, S. C. Datta, Z. Q. Wang, Y. Shao, S. Kang,
G. J. Fisher, and J. J. Voorhees. Inhibition of type i procollagen synthesis by damaged collagen in
photoaged skin and by collagenase-degraded collagen in vitro. Am J Pathol, 158(3):931–42, 2001.
[114] J. L. Ashworth, G. Murphy, M. J. Rock, M. J. Sherratt, S. D. Shapiro, C. A. Shuttleworth, and
C. M. Kielty. Fibrillin degradation by matrix metalloproteinases: implications for connective tissue
remodelling. Biochemical Journal, 340(Part 1):171–181, 1999.
[115] J. D. Humphrey and K. R. Rajagopal. A constrained mixture model for arterial adaptations to a
sustained step change in blood flow. Biomechanics and Modeling in Mechanobiology, 2(2):109–126,
2003.
[116] G. Rouhi, M. Epstein, L. Sudak, and W. Herzog. Modeling bone resorbtion using mixture theory
with chemical reactions. Mechanics of Materials and Structures, 2(6):1141–1155, 2007.
[117] A. Valent´ın and J. D. Humphrey. Evaluation of fundamental hypotheses underlying constrained
mixture models of arterial growth and remodelling. Philosophical Transactions of the Royal Society
A - Mathematical Physical and Engineering Sciences, 367(1902):3585–3606, 2009.
[118] A. Valent´ın and G. A. Holzapfel. Constrained mixture models as tools for testing competing
hypothesis in arterial biomechanics: Survey. Mechanics Research Communications, 29:126–133,
2012.
[119] A. Menzel and E. Kuhl. Frontiers in growth and remodeling. Mechanics Research Communications,
42:1–14, 2012.
32
|
1204.0751 | 2 | 1204 | 2013-01-13T01:00:17 | Optimal Channel Efficiency in a Sensory Network | [
"physics.bio-ph",
"q-bio.NC"
] | We show that the entropy of the distribution of avalanche lifetimes in the Kinouchi-Copelli model always achieves a maximum jointly with the dynamic range. This is noteworthy and nontrivial because while the dynamic range is an equilibrium average measure of the sensibility of a sensory system to a stimulus, the entropy of relaxation times is a purely dynamical quantity, independent of the stimulus rate, that can be interpreted as the efficiency of the network seen as a communication channel. The newly found optimization occurs for all topologies we tested, even when the distribution of avalanche lifetimes itself is not a power-law and when the entropy of the size distribution of avalanches is not concomitantly maximized, strongly suggesting that dynamical rules allowing a proper temporal matching of the states of the interacting neurons is the key for achieving good performance in information processing, rather than increasing the number of available units. | physics.bio-ph | physics | Optimal Channel Efficiency in a Sensory Network
Thiago S. Mosqueiro∗ and Leonardo P. Maia†
Instituto de F´ısica de Sao Carlos, Universidade de Sao Paulo, 13560-970 Sao Carlos, SP, Brasil
(Dated: June 29, 2018)
Abstract
We show that the entropy of the distribution of avalanche lifetimes in the Kinouchi-Copelli model always
achieves a maximum jointly with the dynamic range. This is noteworthy and nontrivial because while
the dynamic range is an equilibrium average measure of the sensibility of a sensory system to a stimulus,
the entropy of relaxation times is a purely dynamical quantity, independent of the stimulus rate, that can be
interpreted as the efficiency of the network seen as a communication channel. The newly found optimization
occurs for all topologies we tested, even when the distribution of avalanche lifetimes itself is not a power
law and when the entropy of the size distribution of avalanches is not concomitantly maximized, strongly
suggesting that dynamical rules allowing a proper temporal matching of the states of the interacting neurons
is the key for achieving good performance in information processing, rather than increasing the number of
available units.
PACS numbers: 64.60.av, 89.70.Cf, 87.85.dq
3
1
0
2
n
a
J
3
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
1
5
7
0
.
4
0
2
1
:
v
i
X
r
a
1
In the end of the last century appeared the first claims on the criticality hypothesis, stating
that some biological systems could evolve towards the edge of chaos [1–6], the critical surface
separating two phases in an abstract space of parameters. The heuristic justification for this hy-
pothesis is that adaptation of biological systems would be guided by selective pressures favoring
the optimization of some key attributes, e.g.
the capacity of sensing environments. No matter
how appealing was the proposal, earlier work [1–3] relied only on simulations and the lack of
experimental validation combined with some criticisms [7] turned down the theory for some time.
However, captivating researchers on brain dynamics, the idea acquired a wholly new motivation
[8]. Indeed not only new theoretical evidence of increased computational performance at criti-
cality showed up recently [9–12] but mainly the observation of power-law behavior of neuronal
avalanches in cortical networks both in vitro [13, 14] and in vivo [15–17] constituted stronger-
than-ever evidence of the relevance of the edge of chaos to the operation of biological systems.
There was some debate regarding the proper characterization of the recorded power laws [18, 19]
and whether criticality would be the sole explanation for the observed scale invariance [20] but
this time the criticality hypothesis is standing up criticisms [5, 6]. Remarkably, some experiments
have explicitly revealed maximized quantities as information transmission [13, 21], information
capacity [21], synchronizability [22] and dynamic range [23] in cortical networks at a critical
condition.
The dynamic range is a sensibility measure associated to the high-slope region of a tuning
curve (also response function, the stimuli-response relationship characterizing a sensory network),
where nearby stimuli can be most easily discriminated since small changes in stimulus lead to high
changes in the firing response. Theoretical work [11] indicated that the dynamic range should be
maximized in sensory systems when the topology of the network was set in a specific condition,
critical for the signal propagation among interacting excitable neurons. A beautiful marriage of
theoretical prediction with experimental confirmation happened when an optimal dynamic range
was found [23] in cortex slice cultures with a proper balance between excitatory and inhibitory
interactions achieved through pharmacological manipulation.
In this study, we employ simulations to show that, when the Kinouchi-Copelli (KC) model [11]
is tuned at the edge of chaos, the Shannon entropy of its avalanche lifetime statistics (hereafter
information efficiency) is always jointly maximized with the dynamic range, both in the original
[11] random graph topology and alternatives. We note that the information capacity (the entropy
of avalanche size distribution) does not always exhibit such a critical optimization. Indeed, we
2
believe that information efficiency rules the behavior of the dynamic range and outweighs by far
the relevance of information capacity in determining the information processing properties of a
sensory network. Two previous works discussed critical optimization of entropies of avalanches
sizes. The first one [24] was a theoretical study of measures of information propagation in Boolean
networks and the second one was the discovery of optimal information capacity in critical cortical
networks [21]. None of them have raised the possibility that the information efficiency could be
maximized instead of capacity.
The model.- We start giving a concise description of the KC model [11]: each of the N neurons
is a cellular automaton that can be at states 0 (quiescent/excitable), 1 (excited) or 2, · · · , m − 1
(refractory states). The neurons are arranged as a weighted undirected graph with mean degree K.
The sequential transitions 1 → 2, 2 → 3, ..., m − 2 → m − 1 and m − 1 → 0 are deterministic. On
the other hand, the transition 0 → 1 happens for neuron k if either (i) a spark of any of its excited
neighbors j reaches it, with probability pk j drawn from an uniform distribution in [0, pmax], or (ii)
if k gets an external stimulus, modeled by a Poisson process with rate r, resulting in an excitation
probability l = 1 − exp (rD
t. All other transitions are forbidden.
t) at each time interval D
By setting pmax = 2s /K, the mean number of excitations an excited neuron could generate
, namely, the average branching ratio.
in one time step if all its neighbors were quiescent is s
Given the fraction of excited nodes r
t at time t, the proper psychophysical response of the system
is the average activity F = T −1 (cid:229) T
t. Consequently, the response is a function F = F(r) of
the stimulus rate r. The dynamic range D
is defined in decibels as D = 10 log(r0.9/r0.1), where
F(rx) = Fmin +x[Fmax −Fmin], x ∈ [0, 1], Fmax = F(¥ ) is the satured response and Fmin = F(r → 0)
is the spontaneous activity. Kinouchi and Copelli showed that self-sustained activity is possible if
is greater than s c = 1 [11], so that Fmin plays the role of the order parameter in a phase transition
in the neural activity with s as a control parameter. They have also found a critical optimization
for D
t=1
.
In [11] the authors studied only the Erdos-R´enyi topology (ERT) with a fixed number NK/2 of
connections and focused on characterizing the maximization of the dynamic range, as did further
works on alternative topologies [25–27]. They did not dwell on exploring the bursts of activity
(avalanches) generated by their model, although stating that critical networks exhibit both large
variance of avalanche lifetimes and a power-law distribution for avalanche sizes with the classical
exponent −3/2 [28, 29].
In this work, we study the avalanches exhibited by the KC model implemented on both the ERT
3
r
s
and the Barab´asi-Albert [30] topology (BAT). Unless explicitly stated otherwise, the simulations
were performed with N = 105 and K = 10. Given a randomly generated representant of a topology,
with chosen average connectivity K and average branching ratio s
, we randomly choose a neuron
of the network to be initially excited while all others are quiescent and record both the number s
of neurons that get excited due to that single spark and the number t of consecutive generations
the network remained active. We repeat this procedure a large number of times in order to get
the distributions {ps} and {pt } of the size and the lifetime of an avalanche, respectively. In this
setting, there is no role at all for a stimulus rate.
Size and lifetime distributions.- Fig. 1 (a) illustrates the critical (at s c = 1) emergence of
power-law scaling in the bursts of activity, ps ∼ s−1.5 and pt ∼ t−1.9, for the ERT (exponents
estimated with standard techniques [31, 32]). Since the ERT allows the propagation of almost
independent branches of activity, this behavior is perfectly compatible with the predictions from
the theory of branching processes [28, 29]. Indeed, we will describe elsewhere how that formalism
predicts the solid lines in the bottom of Fig. 1 (a).
In Fig. 1 (b) we exhibit the avalanche distributions in a BAT. It is harder to estimate these
distributions and there is much uncertainty in the size distribution, but the bottom of Fig. 1 (b)
shows clearly the presence of a bump in the lifetime distribution for s = 0.4 (ps ∼ s−2 and pt ∼
t−2.9 in the first decades) and strongly suggests power-law behavior for slightly smaller values
of s
. A naıve analysis based on the criterium of a power law as a signature of critical behavior
would favor the latter against the former, but we remark that the observation of power laws in
the KC model in this topology demands subsampling to a 2-10% level (not shown, but see also
[16, 33]) and below we will present results supporting the "bumpy" curve as the critical one. We
start discussing the relation among the dynamic range and the entropies of the distributions just
described.
Information efficiency.- The Shannon entropy H of a distribution {pn},
H({pn}) = −(cid:229)
pn log pn,
n
is a standard measure of the uncertainty of a stochastic observable and so it is quite obvious it
should be applied to the analysis of avalanches. We will set n as either s or t to indicate which
distribution we are talking about. In Fig. 2, we jointly illustrate the behavior of the dynamic range
D and of the entropies as functions of s
, given K and a topology.
It was not surprising to observe in Fig. 2 (a) H({ps}) getting maximized jointly with D
in
4
Barab´asi-Albert topologies.
Figure 1. Distributions of (top) burst size and (bottom) activity lifetimes for (a) Erdos-R´enyi and (b)
In (a), power laws emerge for s c = 1 and the solid lines in the bottom re-
sult from the theory of branching process (see text).
In (b), especially in the bottom, huge avalanches
become more frequent when s = 0.4. We have evidence that this is the critical condition, despite the lack
of a power law (see text).
the ERT, since two previous studies [21, 24], in different contexts, reported such an effect for the
information capacity. Moreover, it seems natural to associate the dynamic range to that entropies
(as measures of the flexibility of the system in dealing with signals), so that heuristically H({pt})
should exhibit critical optimization following D
.
Thus, since the critical optimization of D
in BAT's has already been stablished [25, 26], the
lack of a corresponding peak for H({ps}) in Fig. 2 (b) came as a complete surprise. Later we will
discuss how that happens while H({pt}) and D keep getting optimized jointly, this time at s = 0.4.
Is this optimization really critical? As did the authors of [25], we claim so invoking as first
evidence the plot of the order parameter Fmin of the KC model against the control parameter s
in
the inset of Fig. 2 (b). However, our results suggest s c = 0.4, while s c ≈ 0.5 was estimated in [25].
Despite the much smaller networks simulated in that work, such a divergence demands an even
more stringent analysis regarding the critical nature of the position of the peaks in this topology,
namely, a tentative of data collapse into a scaling function. We emphasize we are not going to take
the abscissa of the peak as the location of a critical point a priori, since there is no support for
such a procedure. Quite the opposite, in [24], for instance, stochasticity makes H({ps}) exhibit a
5
Figure 2. Dynamic range (D ) and entropies of avalanche size and lifetime distributions as functions of
the average branching ratio s
optimization emerges clearly in (a), but only the lifetime entropy gets optimized jointly with D
green dashed lines are guides for the eyes. Inset: Fmin vs. s
in (b). The
, clearly indicating the critical value s c ≈ 0.4.
in (a) Erdos-R´enyi and (b) Barab´asi-Albert topologies. Concomitant critical
peak away from criticality.
We also remark that (i) the deviation of s c from 1 in alternative topologies has been recently
explained in terms of a spectral analysis [34, 35] and nowadays is not surprising at all, (ii) there
is no critical optimization for BAT's grown node by node (i.e., the information efficiency and
the dynamic range keep behaving jointly even in such a "charmless" scenery) and (iii) whatever
be the critical point, it is very clear from Fig. 2 (b), mainly from the plot of the information
efficiency H({pt}), that the joint optimization of H({pt}) and D does not happen at the values of
resembling power-law behavior in Fig. 1 (b).
Critical optimization.- Despite the inset in Fig. 2 (b) being a classical signature of a phase
transition, the aforementioned displacement of the critical point prompted us to perform a scaling
analysis in order to confidently assess the issue of criticality. Let r N(t) be the complementary
cumulative distribution function (CDF) of the lifetime of an avalanche when there are N neurons
in the KC model (i.e., r N(t) is the probability of the duration of an avalanche surviving for at
least t given a network with N units). Given a topology and a value of s
thought to be a critical
point, we have looked for exponents g and D able to make all the transformed distributions xg
r N(x)
collapse into a single universal curve F when plotted against x/ND, regardless the value of N. Fig.
3 illustrates the success of such endeavour, considering the distributions obtained with s c = 1 in
6
s
Figure 3. Instances of (top) CDF's for several system sizes and (bottom) their collapse into a scaling function
for both (a) Erdos-R´enyi and (b) Barab´asi-Albert topologies. To collapse the CDF's at the proposed critical
conditions, we have fit g ER = 0.9, DER = 0.5, g BA = 1.25 and DBA = 1 (see text). The distributions in (b)
closely resemble the ones presented recently by Dehghani et al [36].
the ERT and s c = 0.4 in the BAT.
There are two more observations supporting our claim of criticality in the absence of pure power
laws in the avalanches (data not shown). Both preliminary results on the collapse of avalanche
shapes, as advocated by [37] and implemented in [38], and an analysis of the extinction probability
(a branching-process-like study to be discussed elsewhere) point to s c = 0.4 in the BAT.
We also report complementary observations regarding data not shown in this paper. In contrast
with other contexts [9, 39], our results are qualitatively robust to changes in the distribution of
edge weights (even for constant weigths). Most alternative topologies [40, 41] leading to more
realistic degree distributions [42] are qualitatively equivalent to the BAT and exhibit even more
pronounced bumps. In the uncorrelated version [43] of the configuration model [44], bumps are
still present in the avalanche distributions but both information capacity and efficiency exhibit
critical optimization. This behavior fits in the phenomenology we describe here, but also suggests
that the existence of degree correlations in the BAT may be the reason why H({ps}) does not
maximize in that topology.
Discussion.- Neuroscientists regularly employ the mutual information as a measure of the sta-
tistical dependence between stimulus and response, taking into account both specificity (change in
stimulus implying change in response) and fidelity (low variability given a stimulus) of responses.
Notwithstanding the relevance of such studies, indispensable for the comprehension of neural cod-
ing, we remark our aim in this paper is the study of intrinsic, stimulus-free, behavior of sensory
systems rather than of patterns of response variability.
7
The KC model [11] revealed that the stimulus-dependant measure D
is optimized precisely at
s c, a critical point of the self-sustained activity Fmin, that does not depend at all on the nature of the
stimulus. If not a theoretical artifact, that phenomenology suggests that generally the behavior of
actual excitable networks (e.g., sensory systems) could be strongly determined by its stimulus-free
properties. Accordingly, an optimal dynamic range was observed experimentally at a condition de-
termined in the absence of stimuli [23] and there is further robust evidence that stimulus-evoked
activity is strongly dependent on the spontaneous (r = 0) firing patterns in the cortex (see refer-
ences in [23]). Therefore, despite it may seem quite plausible per se that the sensibility measured
by the dynamic range gets optimal when signals do not either fade out fastly (subcritical) or fre-
netically superimpose themselves (supercritical), there is enough motivation for scrutiny on the
microscopic mechanisms leading to improvements in information processing capabilities even in
stimulus-free conditions.
We firmly believe to have discovered one such mechanism. It seems fairly intuitive to expect
that a sensory system be more flexible if it disposes of greater variability of the duration of the
bursts of neural activity it must process. Indeed the support of the avalanche lifetime distribution
(the interval of lifetimes with positive probabilities) achieves a maximum jointly with the dynamic
range, while noncritical distributions hold for no longer than two decades (see Fig. 1). We have
chosen the information efficiency as the proper measure of that effect because it takes into account
not only the magnitude of the reportoire of lifetimes but also a balanced utilization [45] of that
resource for information transmission (from the olfactory bulb to the cortex, for instance).
It is not evident why such arguments are always valid for the information efficiency but not
for the capacity. We speculate that an explanation must rely on the relationship between the mi-
crostructure of the network (motifs) and the dynamics of the excitable units. Progress beyond the
initial studies on the synchronizability properties of the KC model [22, 27, 46] will be probably
achieved by deciphering such relationship. As a rule of thumb, greater values of the clustering co-
efficient [41] should lead to stronger deviations from the −3/2 law. It is also worth mentioning that
it has already been suggested [47] that "the lifetime distributions of neuronal avalanches may carry
rich information about the local cortical circuit structure" and may exhibit consistent deviations
from power-law scaling, while the size distribution would be much more well-behaved. Anyway,
despite our praise of the intrinsic activity, it is definitively worth studying stimuli-dependant fea-
tures of excitable networks by information-theoretic tools like mutual information and transfer
entropy.
8
The notion of criticality without power laws may have wide implications in the interpretation
of observations of neuronal avalanches. Recent experiments exhibiting critical optimization [21–
23] have described the phase transitions in terms of a control parameter k
resembling s but based
on the tacit assumption that neuronal avalanches are pure power laws. Further investigations are
necessary to reveal eventual consequences of the breakdown of that hypothesis. Likewise, Ref.
[36] employed robust statistical techniques to analyze neuronal avalanches in vivo and stand up
against critical dynamics. However, the CDF's they present are very similar to Fig. 3 (b), so that
probably their data is ruling out power-law scaling, but not criticality. Finally, distributions pretty
much like the ones in Fig. 1 (b) have been recently observed in high-resolution experiments in vitro
[38] and the bumps were no obstacle for a remarkable data collapse constituting very compelling
evidence of critical behavior in brain dynamics.
Summarizing, we studied the avalanches in Kinouchi-Copelli model in a first attempt to figure
out detailed mechanisms of information transmission in cortical networks. We discovered that,
in a critical point, the entropy of avalanche lifetime statistics (information efficiency) is always
maximized jointly with the dynamic range, an important measure of information transmission
extracted from the psychophysical tuning curves. Our findings fit in the discussions regarding the
role of criticality in information processing [5, 6] and the relationship of long bursts of activity
with the dynamic range [23], specially because they suggest critical behavior without pure scale
invariance.
Acknowledgements. We acknowledge the contribution of an anonymous referee that brought
reference [38] to our attention and gave suggestions that significantly improved the paper. Thiago
S. Mosqueiro acknowledges CAPES for financial support. The work of Leonardo P. Maia was
supported by FAPESP grant No. 2010/20446-5.
∗ [email protected]
† [email protected]
[1] N. Packard, in Dynamic patterns in complex systems, edited by J. A. S. Kelso, M. F. Shlesinger, and
A. J. Mandell (World Scientific, Singapore, 1988) pp. 293–301.
[2] C. G. Langton, Physica D 42, 12 (1990).
[3] S. A. Kauffman and S. Johnsen, J. Theor. Biol. 149, 467 (1991).
9
[4] P. Bak, How nature works: the science of self-organized criticality, 1st ed. (Springer, 1999).
[5] J. M. Beggs, Philos. T. Roy. Soc. A 366, 329 (2008).
[6] T. Mora and W. Bialek, J. Stat. Phys. 144, 268 (2011).
[7] M. Mitchell, P. T. Hraber, and J. P. Crutchfield, Complex Syst. 7, 89 (1993).
[8] D. R. Chialvo, Nat. Phys. 6, 744 (2010).
[9] N. Bertschinger and T. Natschlager, Neural Comput. 16, 1413 (2004).
[10] C. Haldeman and J. M. Beggs, Phys. Rev. Lett. 94, 058101 (2005).
[11] O. Kinouchi and M. Copelli, Nat. Phys. 2, 348 (2006).
[12] T. Tanaka, T. Kaneko, and T. Aoyagi, Neural Comput. 21, 1038 (2008).
[13] J. M. Beggs and D. Plenz, J. Neurosci. 23, 11167 (2003).
[14] J. M. Beggs and D. Plenz, J. Neurosci. 24, 5216 (2004).
[15] T. Petermann, T. C. Thiagarajan, M. A. Lebedev, M. A. L. Nicolelis, D. R. Chialvo, and D. Plenz,
Proc. Natl. Acad. Sci. USA 106, 15921 (2009).
[16] T. L. Ribeiro, M. Copelli, F. Caixeta, H. Belchior, D. R. Chialvo, M. A. L. Nicolelis, and S. Ribeiro,
PLoS ONE 5, e14129 (2010).
[17] G. Hahn, T. Petermann, M. N. Havenith, S. Yu, W. Singer, D. Plenz, and D. Nikolic, J. of Neurophys-
iol. 104, 3312 (2010).
[18] J. Touboul and A. Destexhe, PLoS ONE 5, e8982 (2010).
[19] A. Klaus, S. Yu, and D. Plenz, PLoS ONE 6, e19779 (2011).
[20] M. Benayoun, J. D. Cowan, W. van Drongelen, and E. Wallace, PLoS Comput. Biol. 6, e1000846
(2010).
[21] W. L. Shew, H. Yang, S. Yu, R. Roy, and D. Plenz, J. Neurosci. 31, 55 (2011).
[22] H. Yang, W. L. Shew, R. Roy, and D. Plenz, J. Neurosci. 32, 1061 (2012).
[23] W. L. Shew, H. Yang, T. Petermann, R. Roy, and D. Plenz, J. Neurosci. 29, 15595 (2009).
[24] P. Ramo, S. Kauffman, J. Kesseli, and O. Yli-Harja, Physica D 227, 100 (2007).
[25] M. Copelli and P. R. Campos, Eur. Phys. J. B 56, 273 (2007).
[26] A. Wu, X. Xu, and Y. Wang, Phys. Rev. E 75, 032901 (2007).
[27] M. Chavez, M. Besserve, and M. Le Van Quyen, Prog. Biophys. Biophys. Chem. 105, 29 (2011).
[28] T. E. Harris, The Theory of Branching Processes (Springer-Verlag, Berlin, 1963).
[29] S. Zapperi, K. B. Lauritsen, and H. E. Stanley, Phys. Rev. Lett. 75, 4071 (1995).
[30] A. Barab´asi and R. Albert, Science 286, 509 (1999).
10
[31] M. E. J. Newman, Contemporary Physics 49, 323 (2005).
[32] A. Clauset, C. R. Shalizi, and M. E. J. Newman, SIAM Review 51, 661 (2009).
[33] V. Priesemann, M. Munk, and M. Wibral, BMC Neurosci. 10, 40 (2009).
[34] D. B. Larremore, W. L. Shew, and J. G. Restrepo, Phys. Rev. Lett. 106, 058101 (2011).
[35] D. B. Larremore, W. L. Shew, E. Ott, and J. G. Restrepo, Chaos 21, 025117 (2011).
[36] N. Dehghani, N. G. Hatsopoulos, Z. D. Haga, R. A. Parker, B. Greger, E. Halgren, S. S. Cash, and
A. Destexhe, "Avalanche analysis from multi-electrode ensemble recordings in cat, monkey and hu-
man cerebral cortex during wakefulness and sleep," (2012), arXiv:1203.0738v2 [q-bio.NC].
[37] J. P. Sethna, K. A. Dahmen, and C. R. Myers, Nature 410, 242 (2001).
[38] N. Friedman, S. Ito, B. A. W. Brinkman, M. Shimono, R. E. L. DeVille, K. A. Dahmen, J. M. Beggs,
and T. C. Butler, Phys. Rev. Lett. 108, 208102 (2012).
[39] W. Chen, J. Hobbs, A. Tang, and J. Beggs, BMC Neurosci. 11, 3 (2010).
[40] S. N. Dorogovtsev, J. F. F. Mendes, and A. N. Samukhin, Phys. Rev. Lett. 85, 4633 (2000).
[41] D. J. Watts and S. H. Strogatz, Nature 393, 440 (1998).
[42] V. M. Egu´ıluz, D. R. Chialvo, G. A. Cecchi, M. Baliki, and A. V. Apkarian, Phys. Rev. Lett. 94,
018102 (2005).
[43] M. Catanzaro, M. Bogun´a, and R. Pastor-Satorras, Phys. Rev. E 71, 027103 (2005).
[44] M. Molloy and B. Reed, Random Structures & Algorithms 6, 161 (1995).
[45] As an entropy, the information efficiency gets higher the more spread is the avalanche lifetime distri-
bution on its support. Also on this issue, we remark that the breakdown of power-law scaling in the
critical condition occurs by increased concentration of probability mass in the tails of the distributions
instead of throwing away large and durable excursions, see Fig. 1.
[46] F. Rozenblit and M. Copelli, J. Stat. Mech.: Theory Exp. 2011, P01012 (2011).
[47] J.-n. Teramae and T. Fukai, J. Computat. Neurosci. 22, 301 (2007).
11
|
1408.2817 | 2 | 1408 | 2015-01-03T13:46:33 | Elasticity of 3D networks with rigid filaments and compliant crosslinks | [
"physics.bio-ph",
"cond-mat.soft"
] | Disordered filamentous networks with compliant crosslinks exhibit a low linear elastic shear modulus at small strains, but stiffen dramatically at high strains. Experiments have shown that the elastic modulus can increase by up to three orders of magnitude while the networks withstand relatively large stresses without rupturing. Here, we perform an analytical and numerical study on model networks in three dimensions. Our model consists of a collection of randomly oriented rigid filaments connected by flexible crosslinks that are modeled as wormlike chains. Due to zero probability of filament intersection in three dimensions, our model networks are by construction prestressed in terms of initial tension in the crosslinks. We demonstrate how the linear elastic modulus can be related to the prestress in these network. Under the assumption of affine deformations in the limit of infinite crosslink density, we show analytically that the nonlinear elastic regime in 1- and 2-dimensional networks is characterized by power-law scaling of the elastic modulus with the stress. In contrast, 3-dimensional networks show an exponential dependence of the modulus on stress. Independent of dimensionality, if the crosslink density is finite, we show that the only persistent scaling exponent is that of the single wormlike chain. We further show that there is no qualitative change in the stiffening behavior of filamentous networks even if the filaments are bending-compliant. Consequently, unlike suggested in prior work, the model system studied here cannot provide an explanation for the experimentally observed linear scaling of the modulus with the stress in filamentous networks. | physics.bio-ph | physics | Elasticity of 3D networks with rigid filaments and compliant crosslinks
Knut M. Heidemann,1 Abhinav Sharma,2 Florian Rehfeldt,2 Christoph F. Schmidt,2, ∗ and Max Wardetzky1, †
1Institute for Numerical and Applied Mathematics, Georg-August-Universitat, Gottingen, Germany
2Third Institute of Physics -- Biophysics, Georg-August-Universitat, Gottingen, Germany
(Dated: January 22, 2020)
5
1
0
2
n
a
J
3
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
7
1
8
2
.
8
0
4
1
:
v
i
X
r
a
Disordered filamentous networks with compliant crosslinks exhibit a low linear elastic shear modulus at small
strains, but stiffen dramatically at high strains. Experiments have shown that the elastic modulus can increase by
up to three orders of magnitude while the networks withstand relatively large stresses without rupturing. Here,
we perform an analytical and numerical study on model networks in three dimensions. Our model consists of
a collection of randomly oriented rigid filaments connected by flexible crosslinks that are modeled as wormlike
chains. Due to zero probability of filament intersection in three dimensions, our model networks are by con-
struction prestressed in terms of initial tension in the crosslinks. We demonstrate how the linear elastic modulus
can be related to the prestress in these network. Under the assumption of affine deformations in the limit of
infinite crosslink density, we show analytically that the nonlinear elastic regime in 1- and 2-dimensional net-
works is characterized by power-law scaling of the elastic modulus with the stress. In contrast, 3-dimensional
networks show an exponential dependence of the modulus on stress.
Independent of dimensionality, if the
crosslink density is finite, we show that the only persistent scaling exponent is that of the single wormlike chain.
We further show that there is no qualitative change in the stiffening behavior of filamentous networks even if the
filaments are bending-compliant. Consequently, unlike suggested in prior work, the model system studied here
cannot provide an explanation for the experimentally observed linear scaling of the modulus with the stress in
filamentous networks.
I.
INTRODUCTION
The mechanical properties of biological cells are governed
by the cytoskeleton, a viscoelastic composite consisting of
three main types of linear protein polymers: actin, micro-
tubules, and intermediate filaments. These filamentous poly-
mers are crosslinked by various binding proteins and consti-
tute a dynamic complex network that maintains the structural
integrity of the cell with the capacity for dynamic reorganiza-
tion needed for active processes. Many in vitro studies have
focused on reconstituted networks with rigid crosslinks [1 --
12].
In the cytoskeleton, however, many of the crosslinks
are themselves extended and highly compliant proteins. Such
flexible crosslinks can strongly affect the macroscopic net-
work elasticity [13 -- 21]. Indeed, experimental studies show
that composite networks can have a linear modulus as low as
∼ 1Pa, while being able to stiffen by up to a factor of 1000
[11, 14].
Here we analyze 3-dimensional (3D) composite networks
theoretically, and we offer physical simulations thereof. Our
networks are composed of randomly oriented rigid filaments
that are connected by highly flexible crosslinks, each of which
is modeled as a wormlike chain (WLC) [22, 23], which has
been shown to accurately describe flexible crosslinkers, such
as filamin [24, 25]. In our approach we assume that the fila-
ments are much more rigid than the crosslinks, meaning that
the network elasticity is dominated by the entropic stretching
resistance of the crosslinks.
In our theoretical analysis we adopt the widely employed
assumption of affine deformations [16, 19, 26]. Under this
premise, the network is assumed to deform affinely on the
∗ [email protected]
† [email protected]
length scale of the filaments, which in turn is assumed to
be much longer than the contour length of the crosslinks.
Using a single filament description in the limit of a contin-
uous distribution of crosslinks along the filament, we ob-
tain the asymptotic scaling behavior of the elastic modulus
with the stress in the nonlinear regime. We show that in 1-
dimensional (1D) networks, the elastic modulus scales with
the second power of the stress, whereas it scales with the third
power in 2-dimensional (2D) networks. Remarkably, there
is no power law scaling in 3D -- in fact, the elastic modu-
lus of a 3D composite network increases exponentially with
the stress. Numerical evaluation of the affine theory at finite
crosslink densities -- as opposed to a continuous distribution
of crosslinks -- shows that (i) the only asymptotic scaling is
that of the modulus scaling with an exponent 3/2 with the
stress and that (ii) the dependence on dimensionality of the
system is limited to an intermediate-stress regime. These find-
ings are in agreement with our extensive physical simulations
of 3D composite networks. For all cases, the elastic modulus
diverges at a finite strain.
Our theoretical analysis is inspired by the mean-field model
proposed by Broedersz et al. [16, 26]. In sharp contrast to
our theoretical analysis and to the results of our physical sim-
ulations, however, these authors predict linear scaling of the
elastic modulus with applied stress. In particular, for any finite
strain, the elastic modulus remains finite in their model. While
this linear scaling of the elastic modulus is in accordance with
what has been observed experimentally [13, 20, 21], we here
argue that this model does not adequately capture the elastic
response of networks with rigid filaments and permanent (i.e.,
non rupturing or rebinding) crosslinks of finite length.
In Ref. [19], the authors ruled out that the experimentally
observed approximate linear scaling of the modulus with the
stress might be be due to enthalpic (linear) stretching com-
pliance of the crosslinks or filaments. Here, we complement
their analysis by physical simulations that take into account
bending of filaments. Our results empirically show that the
inclusion of bending rigidity does not impact the nonlinear
stiffening behavior of composite networks either. We there-
fore conclude that the theoretical explanation for the linear
scaling of the modulus with stress in experiments remains an
challenging open problem.
By physical simulations, we also study the role of prestress.
We show that in contrast to 1D and 2D networks, 3D networks
experience an initial tension due to non-intersecting filaments
resulting in initially stretched crosslinks, and are therefore
prestressed. The modulus in the linear deformation regime
is then governed by this prestress; indeed, it is higher than the
modulus expected from the affine theory. Our simulations ad-
ditionally indicate that if the network is allowed to relax initial
tension by unbinding and rebinding of crosslinks, the impact
of prestress on the elastic modulus in the linear regime be-
comes insignificant, although the prestress does not relax all
the way to zero.
The remainder of the article is organized as follows. First,
we present the affine theory of composite networks in Sec-
tion II. Under the assumption that deformations of the net-
work are affine on the length scale of the filaments, we derive
expressions for the stress and modulus in 1D, 2D, and 3D.
We then present our physical simulation model and describe
our network generation procedure in Section III. We expand
on the implications of our 3D simulation procedure in Sec-
tion IV; in particular, we explain the emergence of prestress.
We then discuss the results of our simulations in the linear
deformation regime in Section V and indicate which results
from the affine theory are still valid. Finally, we analyze the
simulation results in the nonlinear regime in Section VI.
II. THEORY
In this section we analytically derive the stress and mod-
ulus of a composite network under the assumption of affine
deformations on the length scale of the filaments. We con-
sider a collection of N rigid filaments of length L that are per-
manently connected by nN/2 flexible crosslinks of contour
length l0, where n is referred to as the crosslink density, i.e.,
the number of crosslinks per filament. The filaments are as-
sumed to be perfectly rigid, i.e., they neither bend nor stretch,
and the crosslinks are modeled via the WLC interpolation for-
mula [23]
(cid:32)
(cid:33)
fcl(u) =
kBT
lp
1
4(1− u
l0
1
4
+
u
l0
)2 −
,
(1)
where kBT is the thermal energy, lp the persistence length and
u ≥ 0 the end-to-end distance of the crosslink. Assuming
l0 (cid:29) lp this force-extension relation implements a crosslink
rest-length of zero and shows a characteristic stiffening with
divergence of force as u → l0. Equation (1) can be integrated
2
∆L
L
0
x
(a)
(b)
γ
2 sin2θ
ε ≈
ε = ∆L/L
u = εx
∆x
γ = ∆x
h
h
ϑ
θ
FIG. 1. Sketch of the assumptions of the affine theory: (a) 1D: A
filament (green) of length L is connected to its surrounding through
n crosslinks (blue) that have zero extension at zero strain. The sur-
rounding of the filament is subject to a uniform extensional strain
ε = ∆L/L. Since the filament itself is assumed to be perfectly rigid,
all deformation goes into the crosslinks (drawn in y-direction for bet-
ter visualization). The deformation of a crosslink at distance x from
the center of the filament is given by u = εx (deformation field de-
picted by the horizontal gray arrows). (b) For a 2D system, the exten-
sional strain on a filament at angle θ with the axis in shear direction is
given by ε ≈ (γ/2)sin2θ, for a small shear strain γ = ∆x/h = tanϑ.
(cid:32)
to yield the energy [27] (up to a constant)
Ecl(u) =
kBT
lp
l0
4(1− u
l0
l0
4 −
u
4
) −
+
u2
2l0
(cid:33)
.
(2)
Imposing affine deformations on the filament level fully
determines the deformation field u on the subfilament level.
In the following analysis, we consider a single representative
filament subject to an extensional strain of the surrounding
medium that it is embedded in and crosslinked to.
A.
1D network calculation
We start with a one dimensional system, i.e., 1D extensional
strain ε, and advance in dimensionality by converting an ap-
plied shear strain γ to the orientation dependent extensional
strain ε(γ) felt by the filament.
In the rest frame of the filament, the end-to-end distance
of a crosslink at distance x from the center of the filament is
given by u(x,ε) = εx (see Fig. 1 (a)). For notational con-
venience, we consider positive ε only. Under the assumption
that the crosslink density is high enough that one can consider
the associated distribution as uniformly continuous, the total
energy of a filament in 1D is given by
(cid:90) L/2
E1D(ε) = 2
n
L
0
Ecl(εx)dx .
(3)
Substituting Eq. (2) into Eq. (3), this expression can be inte-
grated analytically (see Appendix A 1).
Following the described approach for the linear regime of
the WLC force-extension relation, i.e., for u (cid:28) l0, the lin-
0 = 2E
ear modulus may be extracted as Gaff
Vε2 , where E/V is
the energy per unit volume V in the network and ε is a small
0 = 1
8ρnkclL, with
strain [28]. For a 1D system this yields Gaff
kcl = 3
being the linear spring constant of a crosslink
2
and ρ := NL/V the total length of filaments per unit vol-
ume. The same holds for the modulus in 2D and 3D, but with
different numerical prefactors: 1/96 and 1/192, respectively
[16, 19, 26].
kBT
lpl0
Next we show that one can extract a functional relation be-
tween nonlinear modulus and stress in the nonlinear regime,
based on simple asymptotic scaling analysis. It follows from
above that there is a strain εd := l0/(L/2) at which the outer
most crosslink (at x = L/2) reaches maximum extension. For
ε → εd the energy diverges as
1
Ediv
1D (ε) ∼ −
ε ln
(cid:18)
1−
(cid:19)
(4)
,
ε
εd
with '∼' defined via E ∼ f ⇔ E/ f → const. The upper index
'div' always indicates that we are only taking into account the
diverging part of the 1D filament energy. We express stress
d2E
dε and K = 1
dE
and differential elastic modulus via σ = 1
dε2 ,
V
V
respectively, in order to obtain σ1D ∼ 1/(1−ε/εd), and K1D ∼
1/(1− ε/εd)2. We arrive at the asymptotic scaling relation
K1D ∼ (σ1D)2 .
(5)
This scaling relation between modulus and stress in 1D has
also been derived in previous work [19]. Next we consider
scaling relations in 2D and 3D.
B.
2D network calculation
We perform similar calculations as in 1D, while taking into
account that the extensional strain ε, which results from a
shear strain γ on a 2D system, depends on the orientation of
the filament under consideration. In the small-strain limit one
thus obtains
ε(γ,θ ) = (γ/2)sin2θ ,
(6)
where θ ∈ [0,π] is the angle between the filament and the
shear direction (see Fig. 1(b)).
Substituting this expression into Eq. (4) and averaging over
all orientations leads to
(cid:104)Ediv
2D(cid:105)θ (γ) ∼
0
−ln(1−
γL
4l0
(γ/2)sin2θ
sin2θ )
dθ ,
(7)
(cid:90) π/2
3
where we assume γ ≥ 0 for notational convenience; the upper
integration limit is reduced to π/2 because sin2θ is π/2-
periodic. Note that we do not take into account a redistribu-
tion of filament orientations under the shear transformation.
This approach, as well as the small-strain approximation for
ε(γ,θ ), are justified if L (cid:29) l0, since then the strain γd := 4l0/L
at which the integrand diverges is small.
Differentiating Eq. (7) with respect to γ and neglecting the
weaker (logarithmically) diverging part of the integrand leads
(cid:90) π/2
to an expression for the stress, as γ → γd:
(cid:17)
(cid:16) γ
dθ
(cid:112)1− (γ/γd)2
sin2θ
π − arccos(1− γ/γd)
(cid:104)σ2D(cid:105)θ (γ) ∼
1−
(9)
(8)
=
γd
0
,
.
The divergence of the stress is of the form σ2D ∼ 1/(1 −
(γ/γd))1/2 and hence K2D ∼ 1/(1− γ/γd)3/2. Therefore, the
asymptotic scaling behavior for the differential modulus in
two dimensions is given by
K2D ∼ (σ2D)3 .
(10)
Note the difference of the scaling relations to the ones in the
1D scenario. Stress shows a weaker divergence with strain
than in 1D but a stronger dependence on the differential mod-
ulus.
Integration of the diverging part of the stress further
shows that the energy at maximum strain is finite -- in con-
trast to the 1D setting, where the energy diverges at the critical
strain. This is an effect introduced by orientational averaging
only.
C.
3D network calculation
For a 3D network, the extensional strain on a filament in the
small-strain limit is given by
ε(γ,θ ,φ ) = (γ/2)sin2θ cosφ ,
(11)
in the usual spherical coordinates. In direct analogy to the 2D
case (see Eq. (9)), the averaged stress close to γd = 4l0/L can
be written as
(cid:104)σ3D(cid:105)θ ,φ (γ) ∼
π/2(cid:90)
π/2(cid:90)
0
0
1−
sinθ dφdθ
(cid:17)
(cid:16) γ
γd
sin2θ cosφ
,
(12)
with γ ≥ 0; the upper integration limit for the φ integration
is reduced to π/2 because cosφ is π-periodic and sym-
metric about π/2 on [0,π].
If we carry out the φ integral
and expand the integrand around θ = π/4, in order to inte-
grate over θ (see Appendix A 2 for details), we obtain σ3D ∼
−ln(1− γ/γd) and hence K ∼ 1/(1− γ/γd).Consequently, K
does not scale with σ as a power law; instead, one obtains
K3D ∼ ecσ3D ,
(13)
4
oretically derived scaling of K is exponential in σ. Such an
exponential increase is quantified by an (in principle) indefi-
nitely increasing maximal slope with increasing n in the lnK
versus lnσ plots; e.g., for n = 60 the maximal slope is 3.49,
for n = 3000 it is 5.82. However, for any finite n, eventually
there is always a universal scaling of K ∼ σ 3/2, resulting from
the single WLC force-extension relation, independent of the
dimensionality of the system. Indeed, for any given n, the in-
tegral representation Eq. (3) becomes invalid close to γ = γd
due to the divergence of the WLC energy.
The numerical results in Fig. 2 have been obtained without
the small-strain approximation for the extension of the fila-
ments. However, redistribution of the filament orientations
under shear has not been taken into account in Fig. 2. Calcula-
tions including this effect show that it may both decrease and
increase the maximum intermediate slope in the lnK versus
lnσ plot and shift the peak to larger stress values depending
on the maximum strain γd. In any case, the asymptotic scaling
regime remains unchanged.
In the next section we introduce the simulation framework
that we use to study 3D networks consisting of many fila-
ments and crosslinks, relaxing the assumption of affine de-
formations.
III. SIMULATION MODEL
We perform quasistatic simulations of 3D networks that
consist of N rigid filaments of
length L, permanently
crosslinked by a collection of nN/2 crosslinks of length l0.
All lengths are measured in units of the side length of the cu-
bic periodic simulation box. A typical set of parameters is
N = 3000, L = 0.3, n = 60, l0 = 0.03.
Each filament is modeled as perfectly rigid, implying that
its configuration can be described by its two endpoints only,
which are constraint to stay at distance L. The flexible
crosslinks are modeled as a central force acting between the
two binding sites. In particular, we use the WLC interpola-
tion formula (Eq. (1)) and the corresponding energy (Eq. (2)).
In all data that is presented, forces are measured in units of
(kBT )/lp. There are no additional degrees of freedom in-
troduced through the crosslinks, since their configuration is
represented via the endpoints of the filaments, in terms of
barycentric coordinates.
In order to generate an initial network configuration we
proceed as follows. We generate N randomly distributed fil-
aments by first randomly choosing their centers of mass in
our simulation box and by then picking a random orienta-
tion for each filament. For crosslinking we apply the fol-
lowing iterative procedure. We randomly choose two points
on distinct filaments and insert a crosslink if the correspond-
ing point-to-point distance is shorter than a certain threshold
αl0. Here α ∈ [0,1) serves as a parameter to vary the ini-
tially allowed crosslink lengths in the system. This procedure
is repeated until the desired number of crosslinks is reached;
see Fig. 3 for an illustration of the final configuration. Since
we perform quasistatic simulations, the system must be at
static equilibrium at all times. As practically all crosslinks
FIG. 2. Differential modulus K as a function of shear stress σ in the
affine limit, with finite number of crosslinks (n = 60), rescaled by the
linear elastic modulus G0 := Kγ=0 and critical stress σc := σ (γc),
respectively, where γc is defined via K(γc) = 3G0. Straight line
indicates power law scaling K ∼ σ 3/2.
Inset shows local slope
d lnK/d lnσ; dotted lines indicate power law scaling with expo-
nents from affine theory {2,3} and single WLC scaling {3/2}. In-
dependent of dimensionality, the asymptotic large stress scaling is
K ∼ σ 3/2. In an intermediate-stress regime, the theoretical values
for infinite crosslink densities are approached.
with a real constant c. The absence of asymptotic power law
scaling sets 3D networks apart from 1D and 2D networks. In
3D, we observe the weakest (logarithmic) divergence of stress
with strain. Integrating the diverging part of the stress shows
that the energy again remains finite for γ → γd.
By considering the limit of in-
Finite crosslink density.
finite crosslink density, we have derived theoretical scaling
relations for strain stiffening by integrating along a filament's
backbone (see Eq. (3)). For any real system, however, the
crosslink density is finite and Eq. (3) turns into a sum
E =
n
∑
i=1
Ecl(εxi) ,
(14)
where {xi} are the crosslink binding sites along the filament.
Fig. 2 shows numerical results for the behavior of the cor-
responding differential modulus K for finite n, obtained by
numerical evaluation of Eq. (14) and proper orientational av-
eraging. Note that the asymptotic scaling behavior of K in
the limit of infinite crosslink density influences a finite net-
work's behavior in the intermediate-stress regime (see inset
of Fig. 2); however, near the critical strain, the differential
modulus scales as K ∼ σ 3/2, i.e., like the response of a sin-
gle WLC. Furthermore, for 1D and 2D systems the theoret-
ical scaling exponents in the limit of infinite crosslink den-
sities can (in the intermediate regime) indeed be approached
by increasing n. In contrast, as shown above, in 3D the the-
10−210−1100101102103104σ/σc100101102103104105106107108109K/G0K∼σ3/21D2D3D10−2100102104σ/σc01234dln(K)/dln(σ) 5
convergence because it allows us to use a warm-start proce-
dure that reuses Lagrange multipliers from one minimization
as initial guesses for the next one. Moreover, the application
of small shear steps reduces the likelihood of discontinuously
jumping between local energy minima.
We stop shearing when the achievable increment in shear
strain becomes smaller than a chosen threshold due to
crosslinks that are very close to their maximum extension.
During the entire simulation process, we record network pa-
rameters in the equilibrated states -- in particular, the energy
E as a function of shear strain γ. This allows us to extract the
shear stress σ = 1
dE
dγ as well as the differential shear elastic
V
d2E
modulus K = dσ
dγ = 1
dγ . Derivatives are taken by first inter-
V
polating E(γ) with a cubic spline. We define the linear shear
elastic modulus as
G0 := Kγ=0 .
(15)
In the following section we discuss the implications of our
specific simulation model, in particular with respect to net-
work structure, and contrast it with previous studies that have
been carried out mostly in 2D.
IV.
INITIAL TENSION AND PRESTRESS
As mentioned in Section III, our network generation results
in a non-zero initial energy E0 at zero strain. Indeed, by ran-
domly placing (zero-diameter) filaments in a 3D container,
filaments have zero probability to intersect; thus, crosslinks
have finite initial extension with probability one. This is dif-
ferent from 2D, where randomly placed filaments mutually in-
tersect with a probability approaching one as their number in-
creases. Indeed, so-called Mikado models [19, 32 -- 34], where
filaments are crosslinked at their intersection sites only, ex-
hibit no forces at zero strain.
In contrast, the initial stretching of crosslinks in our net-
works results in an initial tension before any deformation. For
a quantitative analysis we measure a global variant of this ef-
fect by what we call total prestress σ0, which measures the
normal stress [35] component orthogonal to the shear planes
[36]. More precisely, we measure the single sided (e.g., up-
ward) normal component of the force that is acting on a given
shear plane, by summing up the normal components of the
forces exerted by each crosslink and filament passing through
the given shear plane, see Fig. 4 (a). The normal stress is then
given by dividing by the surface area of the shear plane. Note
that σ0 does not depend on the choice of a particular shear
plane; indeed, if the total stress was changing during verti-
cal movement of a shear plane, then this would immediately
contradict force balance in the system.
Intuitively, one might expect negative normal stresses
(pulling down on the upper face of the simulation box), since
crosslinks are contractile. However, since filaments withstand
compression, it is possible to construct systems that exhibit
positive normal stress. This suggests the existence of configu-
rations with zero normal stress [38]. Indeed, so-called tenseg-
rity structures [39], which are in static equilibrium in the ab-
sence of boundary conditions satisfy this criterion -- while still
FIG. 3. Example of an initially generated network that has not
been relaxed into static equilibrium yet. Rigid filaments are shown
in green, flexible crosslinks in blue. Short crosslink or filament
fragments correspond to filaments/crosslinks that cross the periodic
boundaries of the simulation box. For the sake of visual appearance,
the network is much sparser than the systems that are studied in the
remainder of this article, and the ratio of filament to crosslink length
is much smaller, N = 300, n = 10, L = 0.3, l0 = 0.1, α = 0.9.
will be stretched beyond their rest-length after the initial net-
work generation, we minimize the energy (of the crosslinks)
before subjecting the simulation box to any deformation [29].
For energy minimization we use the freely available external
library IPOPT [30], which requires the gradient and the Hes-
sian of the system's energy function. It might happen during
the optimization process, that individual crosslinks reach ex-
tensions u larger than their contour length l0. Acceptance of
these solutions is prohibited by setting the energy to infinity
(1019) for u ≥ l0 in Eq. (2); without this modification it would
become negative in that regime. The length constraints for the
filaments are realized via Lagrange multipliers.
In order to extract elastic properties of the network we per-
form quasistatic shearing by applying an affine incremental
shear strain δγ to the network, with subsequent rescaling of
filaments to length L (see Fig. 1). We apply Lees-Edwards
shearing periodic boundary conditions [31]. The magnitude of
δγ is determined by calculating the maximum affine shear that
leaves all crosslinks below their contour length. Due to the
rescaling of filament lengths, a nonaffine deformation com-
ponent is introduced. This nonaffinity may lead to crosslinks
being overstretched after all. In this case, we iteratively halve
the shear strain until the length of all crosslinks remains below
their contour length. After each shear increment, the energy
is minimized. We apply a fixed upper bound of 1% strain
on δγ in order to stay reasonably close to the previous so-
lution. This increases numerical efficiency and accelerates
(b)
f4
n
(a)
f1
f2
f1 + f3
n
f3
FIG. 4. (a) Measuring the total prestress σ0 by extracting the nor-
mal component of the total force acting on a shear plane. We sum
up all the forces acting on one side of the plane exerted by (i) the
crosslinks passing through (here f2 and f4) and (ii) the filaments pass-
ing through (here f1 +f3) -- then we project onto the normal vector n.
(b) A tensegrity structure (here: Snelson's X [37]) remains in static
equilibrium without application of boundary conditions. The forces
acting on any plane add up to zero, i.e. no plane carries any total
prestress although it is under tension locally.
being able to store arbitrary amounts of energy (see Fig. 4 (b)).
Empirically, our simulations show that the random networks
generated by the procedure described in Section III exhibit
negative initial normal stresses throughout. Their integrity is
provided through the application of periodic boundary condi-
tions. Note in particular, that our setup enforces conservation
of volume of the simulation box. In general, it would be pos-
sible to relax the prestress by letting the volume of the simu-
lation box change. However, we did not follow this approach
in the study presented here, in order to ensure that the fila-
ment length remains significantly smaller than the size of the
simulation box.
In the following, we relate total prestress to the linear elastic
response of our networks.
V. LINEAR REGIME
In previous work [16, 19, 26], an expression for the linear
modulus in 3D was derived under the assumption of affine
deformations and in absence of any initial tension in the net-
work. Our simulations show that the linear elastic modulus
depends on the initial tension in the network.
One scenario that clearly demonstrates the dependence of
the linear modulus G0 (defined in Eq. (15)) on the initial ten-
sion is illustrated in Fig. 5 where the admissible maximum
initial crosslink length was varied.
For a more quantitative analysis we have designed a method
that allows us to change initial tension for a network with a
fixed set of simulation parameters. We first randomly gener-
ate a network as described above and let it relax into static
equilibrium. We then remove a given amount (5%) of the
most-stretched crosslinks in the system. Then we reconnect
those crosslinks randomly again, and let the network relax.
This procedure is repeated Nrel times. Thereby, we succes-
sively decrease the system's initial tension, and therefore also
its total energy, see inset of Fig. 6. Not only does the total
energy decrease, we also observe a change in the distribution
of forces (see Fig. 6). As long as one performs the crosslink
6
FIG. 5. Differential elastic modulus K as a function of strain γ for
different levels of initial tension. The initial tension in the network is
varied by changing the initially admissible maximal crosslink length
αl0. The linear modulus G0 = Kγ=0 increases with the initial ten-
sion in the network (initial tension increases with α). It is also evi-
dent that the divergence of K occurs at a strain γd that decreases with
increasing α. Here: N = 3000, n = 60, L = 0.3, l0 = 0.03.
FIG. 6. Distribution of forces in crosslinks for a system without or
with Nrel = 100 relaxation steps. The relaxation procedure cuts the
large force tail of the initial distribution and establishes a sharper
peak at small forces. The inset shows the total energy E in the sys-
tem, normalized by the initial energy E0, as a function of number of
relaxation steps Nrel.
10−210−1100γ103104105106107108109Kα=0.9α=0.8α=0.7α=0.6α=0.50.00.51.01.52.02.53.0fcl012345p(fcl)Nrel=0Nrel=100100101102103Nrel0.00.20.40.60.81.0E/E0 7
γ = 0
σ0
γ = tanϑ
σS
σ
σN
ϕ
ϑ
FIG. 8. The initial network carries a total prestress σ0. After a small
shear γ = tanϑ has been applied it exhibits a shear stress σS and
normal stress σN, with tanϕ = σS/σN.
FIG. 7. Linear elastic modulus G0 normalized by the affine predic-
tion Gaff
0 as a function of total prestress σ0 normalized by the total
prestress σ∗0 immediately after initial network generation. The to-
tal prestress is reduced via the procedure described in Section V.
For small total prestress, G0 exhibits superlinear dependence on σ0.
Up to σ0 = σ∗0 , we observe linear scaling G0 ∝ σ0, as predicted by
the model. The straight line is drawn as a guide to the eye, rep-
resenting linear scaling. Parameters: N = 3000, n = 60, L = 0.3,
l0 = 0.06, α = 0.5. The inset shows differential elastic modulus K
versus shear strain γ for systems with varying number of relaxation
steps Nrel ∈ {0,50,100,150}. G0 goes down with increasing Nrel.
Parameters: N = 3000, n = 60, L = 0.3, l0 = 0.03, α = 0.5.
binding-unbinding procedure over a small enough fraction of
crosslinks, the network remains nearly isotropic.
It is apparent from the inset of Fig. 7 that the linear elas-
tic modulus is reduced by increasing the number of relaxation
steps, as expected. Fig. 7 also shows the dependence of linear
modulus G0 on the total prestress σ0, which has been intro-
duced in Section IV. We varied σ0 via the above described
procedure, and measured G0 with the shearing protocol de-
scribed in Section III. After a certain number of relaxation
steps the empirical value for G0 equals the value Gaff
0 expected
from affine theory (see Section II A). Relaxing initial tension
further, we reach moduli even below Gaff
0 . This is possible be-
cause the network can rearrange nonaffinely, thereby soften-
ing its response. Over a certain range of total prestresses, we
observe linear scaling of G0 with σ0, a phenomenon, which
has been discussed in other contexts before (see for exam-
ple Ref. [40]). We explain the linear regime as follows. For
small strains the normal component σN of the stress acting on
shear planes is close in magnitude to the total prestress σ0,
i.e., σN ≈ σ0. For small strains given by shear angles ϑ ≈ 0,
total forces acting on the shear planes make an angle ϕ with
the direction normal to the shear planes (see Fig. 8). Our sim-
ulations show that tanϕ ∝ tanϑ and that the constant of pro-
portionality remains unchanged in the linear scaling regime.
FIG. 9. Linear elastic modulus G0 versus crosslink density n for sys-
tems with different number of relaxation steps: Nrel = 0 (diamonds)
and Nrel = 50 (squares). Solid line indicates values expected from
affine theory: Gaff
0 = ρnkclL/192. Parameters: N = 3000, L = 0.3,
l0 = 0.06, α = 0.5.
Therefore, shear satisfies
γ = tanϑ ∝
σS
σ0
,
(16)
where σS is the component of the stress acting on shear planes
in the shear direction, see Fig. 8. Hence, the linear elastic
shear modulus G0 defined via σS = G0γ is proportional to
the total prestress σ0 via Eq. (16). However, for very small
total prestresses, i.e., after many relaxation steps, the modu-
lus shows a steeper than linear dependence on σ0. Indeed, in
this regime the aforementioned constant of proportionality be-
comes larger. This effect might be attributed to the fact that for
small σ0, tensegrity type elements (see Fig. 4 (b)), which do
not contribute to the total prestress but carry energy, contribute
significantly to the measured shear stress, thereby increasing
ϕ (see Fig. 8).
Furthermore, affine theory predicts linear scaling of the
modulus G0 with crosslink density n. Fig. 9 shows that this
linear scaling is indeed reproduced in our simulations, inde-
0.00.20.40.60.81.0σ0/σ∗00.60.81.01.21.4G0/Gaff0G0∼σ010−210−1100γ10−1100101102K/Gaff0Nrel020406080100n020406080100120140G0/(ρkclL/192)Nrel=0Nrel=50 8
(a) Critical strain γc versus inverse filament length 1/L
FIG. 10.
for Nrel = 0 and Nrel = 50. Other parameters: N = 5000, n = 60,
l0 = 0.04, α = 0.7. We observe linear scaling γc ∝ 1/L for Nrel = 0;
systems in which relaxation has been applied show deviations from
(b) Critical strain γc versus
this behavior (see Nrel = 50 here).
crosslink contour length l0 for a system with N = 3000, n = 50,
L = 0.3, α = 0.5.
pendent of the prestress. Moreover, by changing the prestress
via our relaxation procedure it is possible to reach comparable
slopes to what is predicted by the affine theory.
The next section deals with the nonlinear elastic response of
the simulated networks, and relates it to the theoretical results
that were derived in Section II.
FIG. 11. Differential modulus as a function of shear stress, rescaled
by linear modulus and critical stress σc = σ (γc), respectively. Pa-
rameters: N = 3000, n = 60, L = 0.3, l0 = 0.06, α = 0.5, with
(Nrel = 150) and without (Nrel = 0) relaxation. Inset shows the lo-
cal slope d ln(K)/d ln(σ ) from the main plot. For large stresses, we
observe power law scaling K ∼ σ 3/2 (solid straight line). For inter-
mediate stresses we recover slopes in the range of those derived from
affine theory.
VI. NONLINEAR REGIME
A. Critical strain
The networks that we study are inherently nonlinear be-
cause crosslinks are WLCs with finite length l0 (see Eq. (1)),
resulting in pronounced strain stiffening at a critical strain γc.
Stress diverges at a higher strain γd. In our simulations, we
define the critical strain γc to be the strain where K/G0 ≈ 3.
In the affine theory, γd and γc scale linearly with the ratio of
crosslink to filament length l0/L. In our simulations, we can-
not conclusively report on this dependence because the ac-
cessible ranges for l0 and L are quite limited. On the one
hand, there exists an upper limit for L (therefore also for l0,
since l0/L (cid:28) 1 should hold) to be significantly smaller than
the simulation box. On the other hand, L and l0 are bounded
from below due to computational limitations -- this is because
we need to increase the number of filaments in order to keep
networks homogenous.
For ranges that are accessible to our simulations, we ob-
tain the following results. If we fix l0, then we observe linear
scaling γc ∝ 1/L for systems where no relaxation procedure
has been applied (see Fig. 10 (a)). Relaxed systems, however,
sometimes show a less than linear dependence. This effect
might be due to anisotropies induced by the relaxation proce-
dure. If we fix L, then the dependence of γc on l0 is slightly
less than linear (see Fig. 10 (b)).
B. Differential modulus
It remains to discuss the dependence of the differential
modulus on stress, the affine theory of which has been de-
rived in Section II. For finite crosslink densities, the only per-
sistent scaling behavior is K ∼ σ 3/2, as γ approaches γd --
due to the fact that eventually single WLC response dom-
inates.
In an intermediate regime, above the critical stress
σc = σ (γc), we observe slopes (d lnK/d lnσ ) > 3/2. The ma-
jority of the simulations shows intermediate slopes around 2
or slightly above, mostly independent of simulation parame-
ters, but there are also realizations that show maximum slopes
up to 3.5 (see Fig. 11). These higher slopes and the final scal-
ing K ∼ σ 3/2 are in accordance with the predictions of affine
theory. Indeed, a slope of 3.5 is the maximum slope predicted
by the affine theory when using the same crosslink density as
in the simulation (Fig. 2). There are, however, differences be-
tween theory and simulation in terms of slope profiles since
various assumption are made by the theory that do not hold in
the simulations: A randomly generated network does not have
a uniform crosslink density along the filaments, these systems
are prestressed, and there is no perfect isotropy. Moreover, the
networks do not deform perfectly affinely.
2345671/L0.10.20.30.40.5γc(a)Nrel=0Nrel=500.030.050.070.09l0(b)10−410−310−210−1100101102103104105106σ/σc10−1100101102103104105106107108109K/G0K∼σ3/2K∼σ3/2Nrel=0Nrel=15010−2100102104σ/σc01234dln(K)/dln(σ) 9
FIG. 13. Average tension ¯τ as a function of position x along the
filament for various strain values. Tension ¯τ is normalized by its
maximum absolute value ¯τ0. Dashed curves correspond to theoreti-
cal results for n = 60 at γ = γc (blue), γ (cid:39) γd (green). Solid curves
show simulation data, with N = 3000, n = 60, L = 0.3, l0 = 0.06,
α = 0.5. Inset shows a snapshot of the same system at maximum
strain γd (cid:39) 0.6 where only the 15 most stretched crosslinks and the
corresponding filaments are shown. They form singular paths that
span the whole system, thereby preventing further stress reduction
via nonaffine rearrangements in these finite systems.
both, theoretical and simulated systems at various strains. In
the simulations there is non-zero tension at zero strain due to
prestress. With increasing γ, the simulations resemble the pro-
files expected from affine theory. However, when approach-
ing the maximum strain γd, the emergence of selective paths
(force chains) that carry most of the tension becomes evident.
The highly stretched crosslinks dominate the averaged tension
profiles and therefore lead to jumps in the tension curves at the
respective binding sites along the filament (green solid curve
in Fig. 13).
D. Bending
Thus far we have restricted our theory and simulations to
rigid filaments that can neither bend nor stretch. In Ref. [19],
the authors considered finite stretching compliance of fila-
ments, while bending compliance was assumed to be zero.
They report that finite stretching stiffness does not impact
the nonlinear stiffening regime of a composite network apart
from the expected convergence (to some constant value) of the
modulus at high strains. Here we complement this analysis by
considering filaments that have finite bending but no stretch-
ing compliance. We performed simulations on a 2D network
because of the relative computational ease compared to the
3D case. In addition to the energy stored in the crosslinks, we
consider bending energy of the form Eb = κθ 2/(2lav), where
FIG. 12. Differential nonaffinity δΓ as a function of scaled shear
strain γ/γc for a system with N = 3000, n = 60, L = 0.3, l0 = 0.06,
α = 0.05.
C. Nonaffinity
In order to study to what extent simulation results deviate
from affine theory, apart from prestress, nonuniform crosslink
density, and anisotropy, we analyze the nonaffinity of the net-
work deformation under shear. For a single filament, we de-
fine its differential nonaffinity with respect to the center of
mass by
(cid:107)δ raff − δ r(cid:107)2
(cid:107)δγ(cid:107)2
,
(17)
where δ raff and δ r are the 3D coordinates of a filament's cen-
ter of mass after applying an incremental shear strain δγ with-
out and with relaxation, respectively.
We let δΓ denote the average of the differential nonaffini-
ties over all filaments. Affine approximations imply δΓ = 0.
Fig. 12 shows that center of mass deformations are mostly
affine for small strains. However, the differential nonaffinity
increases starting at a strain around γc and eventually diverges
as γ → γd. This can be understood, since the networks are
strain stiffening, such that small incremental strain can induce
large increase in the forces of individual crosslinks, thereby
inducing large local rearrangements during energy minimiza-
tion.
While increasing shear strain, there are force chains [41 --
43] developing in the network, which carry most of the ten-
sion, and which cannot reduce their strain due to the fact
that they span the entire system (see inset of Fig. 13). We
quantify this effect by considering tension profiles along fila-
ments. The tension τ at position x along a filament is given
via τ(x) = ∑xi>x fcl(ui), where {xi} are the crosslink bind-
ing sites and {ui} their extensions (ui = εxi in affine theory).
Fig. 13 shows tension profiles averaged over all filaments for
0.00.20.40.60.81.01.21.41.61.8γ/γc10−410−310−210−1100101102103δΓ−1.0−0.50.00.51.0x/(L/2)0.00.20.40.60.81.0¯τ/¯τ0γ'γdγ=γc l1
l2
π − θ
merically. We modeled such networks as a collection of rigid
filaments connected by WLC crosslinks.
10
FIG. 14. Sketch of the local bending geometry of a filament (green)
with crosslinks attached (blue). The local bending energy is given
by Eb = κθ 2/(2lav), with κ being the bending rigidity and lav =
(l1 + l2)/2.
FIG. 15.
(a) Differential modulus K as a function of shear stress
σ, rescaled by linear modulus G0 and critical stress σc = σ (γc), re-
spectively, for various bending rigidities κ. Solid straight line indi-
cates power law scaling K ∼ σ 3/2. (b) Differential modulus K as a
function of shear strain γ, rescaled by linear modulus G0 and crit-
ical strain γc, respectively. Parameters: N = 800, L = 1, l0 = 0.1,
system-size Lx = Ly = 6.
κ is the bending rigidity, θ is the angle through which the
filaments bend locally, and lav = (l1 +l2)/2 is the average dis-
tance between two adjacent pairs of crosslinks. We show the
results in Fig. 15. The range of bending rigidity was chosen
such that the linear modulus was still determined by the soft
stretching modes of the crosslinks, so that bending did not
impact the linear regime. As can be seen from these plots,
bending compliance does not impact the nonlinear stiffening
regime either -- since bending modes are geometrically pro-
hibited for large strains.
Thus, in isolation, neither bending nor stretching compli-
ance of filaments impacts the nonlinear stiffening regime of
composite networks. These findings suggest that the theoret-
ical models at present cannot explain the K ∼ σ scaling ob-
served in experiments.
VII. CONCLUSIONS
We have studied the elastic properties of composite
crosslinked filamentous networks in 3D analytically and nu-
Based on the affine theory introduced in Ref. [19] we de-
rived asymptotic power law scaling exponents for the dif-
ferential elastic modulus with stress, in the limit of infinite
crosslink density. In this case, the scaling exponents depend
on the dimensionality of the system. In particular, we showed
that 3D systems no longer exhibit a power law. Furthermore,
we showed that for finite crosslink densities, the only per-
sistent regime (over several orders of magnitude of stress) is
the σ 3/2 scaling, as it is derived from the single WLC force-
extension relation Eq. (1). This is in sharp contrast with the
model proposed in Ref. [16, 26], where linear scaling was
suggested, independent of the dimensionality of the system.
There model implies finite stress at any strain and therefore
does not apply to composite networks of rigid filaments with
flexible crosslinks of finite length.
We further developed a simulation framework that allows
us to measure the elastic response of random filamentous net-
works with WLC crosslinks. One important property of these
3D networks is that, by construction, they are prestressed due
to initial extensions of the crosslinks. In addition to geomet-
rical constraints, active elements such as motors can induce
prestress as well [44]. We showed that the prestress in a net-
work can dominate the linear response and might therefore
be a feature that is worthwhile analyzing in experimental sys-
tems.
Regarding nonlinear response, we observed divergence of
stress (and differential modulus) at finite strain. Close to this
strain we measured a power law scaling of the differential
modulus with stress, with an exponent 3/2, just as expected for
a single WLC. In an intermediate-stress regime we observed
local exponents that span the entire range of theoretically de-
rived values for systems of differing dimensionality. The fact
that our simulation results do not always resemble the predic-
tions of a 3D affine theory, in this intermediate regime, may be
attributed to nonaffine deformations. Extracting the exact set
of assumptions -- such as uniform crosslink density, isotropy,
or zero prestress -- that are responsible for these discrepancies
is left for future investigation.
Experiments (see, e.g., [13, 20, 21]) have shown that in
the nonlinear regime the differential modulus scales approxi-
mately linearly with the shear stress. We did not find such a
regime in our simulations -- neither when working with rigid
filaments nor when incorporating finite bending stiffness (or
enthalpic stretching as done in Ref. [19]). Therefore, we
argue that none of the currently available theories can ade-
quately explain the linear scaling of the differential modulus
observed experimentally. It could possibly be that the WLC
model does not accurately describe the elastic response of a
single crosslink throughout the whole experimentally acces-
sible regime. We speculate, however, that the linear scaling
might be due to thermal fluctuations of the filaments, which
have not been considered so far.
10−1100101102σ/σc100101102103104K/G0(a)K∼σ3/2κ=10−2κ=10−1κ=100κ=101κ=102100γ/γc(b)κ=10−2κ=10−1κ=100κ=101κ=102 ACKNOWLEDGMENTS
The authors would like to thank Fred MacKintosh for fruit-
ful discussions. This work was funded by the Deutsche For-
schungsgemeinschaft (DFG) within the collaborative research
center SFB 755, project A3.
Appendix A: Derivation of scaling relations for the shear
modulus
1. 1D network
11
can approximately consider tan−1(cid:104)(cid:113) 1+(γ/γd)sin2θ
The integral diverges for γ = γd due to a pole at θ = π/4. We
× sinθ as
a constant because it takes finite values around the pole. Since
we are interested in the regime close to the divergence of the
integrand, we expand sin2 2θ up to second order in ν := θ −
π/4. We arrive at
1−(γ/γd)sin2θ
(cid:105)
(cid:90) π/4
−π/4
(cid:112)1− (γ/γd)2(1− 4ν2)
dν
.
(A3)
(cid:18)
1−
εL
2l0
(cid:19)(cid:21)
.
The integral Eq. (3) for the total energy of a single filament
can be solved to give
E1D(ε) = 2
L2ε
32 −
l0L
8 −
l2
0
4ε ln
48l0 −
(A1)
The divergence of the energy for ε → εd = 2l0/L stems from
the term ∼ 1
, which is therefore the only one that
ε ln
we need to consider for the asymptotic scaling analysis in 2D
and 3D.
(cid:17)
n
L
(cid:20)L3ε2
(cid:16)
1− ε
εd
2. 3D network
To approximate the solution of the integral in Eq. (12) we
first carry out the φ integration analytically and obtain
(cid:104)σ3D(cid:105)θ ,φ (γ) ∼
(cid:104)(cid:113) 1+(γ/γd)sin2θ
(cid:105)
1−(γ/γd)sin2θ
1− (γ/γd)2 sin2 2θ
arctan
(cid:90) π/2
(cid:113)
× sinθ dθ .
0
(cid:18)
(cid:113)
∼ ln
2
η2(1− δ )2 + δ (1− δ ) + 2(1− δ )η
∼ −lnδ ,
∼ −ln(1− γ/γd) ,
(A2)
which is what has been proposed in Section II C.
Approximation errors close to the boundary of the interval of
integration that are made by expanding sin2 2θ are negligi-
ble, regarding the asymptotics, because the integrand diverges
right at the center of the interval. Now we define µ := 1−γ/γd
and drop all terms of higher than first order in µ, since we are
interested in the behavior close to γ = γd. With η2 := 4ν2 and
δ := 2µ, we obtain
(cid:90) π/2
−π/2
(cid:112)η2(1− δ ) + δ
dη
.
(A4)
This can be integrated, with the diverging part being
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)π/2
−π/2
,
(A5)
(A6)
(A7)
[1] P. A. Janmey, U. Euteneuer, P. Traub, and M. Schliwa, J. Cell
Biol. 113, 155 (1991).
[2] F. C. MacKintosh and P. A. Janmey, Current Opinion in Solid
State and Materials Science 2, 350 (1997).
[3] J. Xu, D. Wirtz, and T. D. Pollard, J. Biol. Chem. 273, 9570
(1998).
[4] M. L. Gardel, J. H. Shin, F. C. MacKintosh, L. Mahadevan,
P. Matsudaira, and D. A. Weitz, Science 304, 1301 (2004).
[5] M. L. Gardel, J. H. Shin, F. C. MacKintosh, L. Mahadevan,
and D. A. Weitz, Phys. Rev. Lett. 93, 1
P. A. Matsudaira,
(2004).
[6] C. Storm, J. J. Pastore, F. C. Mackintosh, T. C. Lubensky, and
P. A. Janmey, Nature 435, 191 (2005).
[7] A. R. Bausch and K. Kroy, Nature Physics 2, 231 (2006).
[8] G. H. Koenderink, M. Atakhorrami, F. C. MacKintosh, and
C. F. Schmidt, Phys. Rev. Lett. 96, 138307 (2006).
[9] O. Chaudhuri, S. H. Parekh, and D. A. Fletcher, Nature 445,
295 (2007).
[10] P. A. Janmey, M. E. McCormick, S. Rammensee, J. L. Leight,
P. C. Georges, and F. C. MacKintosh, Nature Materials 6, 48
(2007).
[11] K. E. Kasza, A. C. Rowat, J. Liu, T. E. Angelini, C. P. Brang-
wynne, G. H. Koenderink, and D. A. Weitz, Curr. Opin. Cell
Biol. 19, 101 (2007).
[12] J. Liu, G. H. Koenderink, K. E. Kasza, F. C. MacKintosh, and
D. A. Weitz, Phys. Rev. Lett. 98, 198304 (2007).
[13] M. L. Gardel, F. Nakamura, J. H. Hartwig, J. C. Crocker, T. P.
Stossel, and D. A. Weitz, Proc. Natl. Acad. Sci. U.S.A. 103,
1762 (2006).
[14] M. L. Gardel, F. Nakamura, J. Hartwig, J. C. Crocker, T. P.
Stossel, and D. A. Weitz, Phys. Rev. Lett. 96, 088102 (2006).
[15] B. A. DiDonna and A. J. Levine, Phys. Rev. E: Stat., Nonlinear,
Soft Matter Phys. 75, 041909 (2007).
[16] C. P. Broedersz, C. Storm, and F. C. MacKintosh, Phys. Rev.
[17] P. Dalhaimer, D. E. Discher,
and T. C. Lubensky, Nature
Lett. 101, 118103 (2008).
Physics 3, 354 (2007).
[18] H. Lee, B. Pelz, J. M. Ferrer, T. Kim, M. J. Lang, and R. D.
Kamm, Cellular and Molecular Bioengineering 2, 28 (2009).
[19] A. Sharma, M. Sheinman, K. M. Heidemann, and F. C. MacK-
intosh, Phys. Rev. E: Stat., Nonlinear, Soft Matter Phys. 88,
052705 (2013).
[20] K. E. Kasza, G. H. Koenderink, Y. C. Lin, C. P. Broedersz,
W. Messner, F. Nakamura, T. P. Stossel, F. C. MacKintosh, and
D. A. Weitz, Phys. Rev. E: Stat., Nonlinear, Soft Matter Phys.
79, 041928 (2009).
[21] K. E. Kasza, C. P. Broedersz, G. H. Koenderink, Y. C. Lin,
W. Messner, E. A. Millman, F. Nakamura, T. P. Stossel, F. C.
Mackintosh, and D. A. Weitz, Biophys. J. 99, 1091 (2010).
[22] C. Bustamante, J. F. Marko, E. D. Siggia, and S. Smith, Science
265, 1599 (1994).
[23] J. F. Marko and E. D. Siggia, Macromolecules 28, 8759 (1995).
[24] I. Schwaiger, A. Kardinal, M. Schleicher, A. A. Noegel, and
M. Rief, Nature structural & molecular biology 11, 81 (2004).
[25] S. Furuike, T. Ito, and M. Yamazaki, FEBS letters 498, 72
[26] C. P. Broedersz, C. Storm, and F. C. MacKintosh, Phys. Rev.
E: Stat., Nonlinear, Soft Matter Phys. 79, 61914 (2009).
[27] More precisely, it is a free energy, which includes both, ener-
getic (bending) and entropic terms for the crosslinks (not for
the filaments).
[28] L. Landau and E. Lifshitz, Elasticity theory (Pergamon Press,
(2001).
1975).
[29] We do neither take into account fluctuations of the filaments nor
excluded-volume effects.
[30] A. Wachter and L. T. Biegler, Math. Prog. 106, 25 (2005).
[31] A. W. Lees and S. F. Edwards, J. Phys. C: Solid State Phys. 5,
1921 (1972).
12
[32] J. Wilhelm and E. Frey, Phys. Rev. Lett. 91, 108103 (2003),
[33] D. A. Head, A. J. Levine, and F. C. MacKintosh, Phys. Rev.
arXiv:0303592 [cond-mat].
Lett. 91, 2 (2003).
[34] P. R. Onck, T. Koeman, T. van Dillen, and E. van der Giessen,
Phys. Rev. Lett. 95, 19 (2005).
[35] Note that our notion of prestress is not to be confused with the
constant prestress externally applied in bulk rheology experi-
ments, which is a shear stress in general.
[36] Although we could in principle define total prestress as the nor-
mal component of the stress acting on any plane in our system
we prefer to use shear planes as this simplifies the forthcoming
analysis.
[37] R. Connelly and A. Back, American Scientist 86, 142 (1998).
[38] Note that individual crosslinks are still under tension; however,
the total normal force acting on the shear plane vanishes.
[39] A. Pugh, An introduction to tensegrity (University of California
Pr, 1976).
[40] S. Alexander, Physics Reports 296, 65 (1998).
[41] C. Heussinger and E. Frey, The European Physical Journal E ,
1 (2007), arXiv:arXiv:0705.1425v2.
[42] E. M. Huisman, T. van Dillen, P. R. Onck, and E. Van der
Giessen, Phys. Rev. Lett. 99, 2 (2007).
[43] G. Zagar, P. R. Onck, and E. Van der Giessen, Macromolecules
44, 7026 (2011).
[44] G. H. Koenderink, Z. Dogic, F. Nakamura, P. M. Bendix, F. C.
MacKintosh, J. H. Hartwig, T. P. Stossel, and D. A. Weitz,
Proc. Natl. Acad. Sci. U.S.A. 106, 15192 (2009).
|
1010.4288 | 2 | 1010 | 2011-07-14T17:10:33 | Modeling the Effects of Drug Binding on the Dynamic Instability of Microtubules | [
"physics.bio-ph"
] | We propose a stochastic model that accounts for the growth, catastrophe and rescue processes of steady state microtubules assembled from MAP-free tubulin. Both experimentally and theoretically we study the perturbation of microtubule dynamic instability by S-methyl-D-DM1, a synthetic derivative of the microtubule-targeted agent maytansine and a potential anticancer agent. We find that to be an effective suppressor of microtubule dynamics a drug must primarily suppress the loss of GDP tubulin from the microtubule tip. | physics.bio-ph | physics | Modeling the Effects of Drug Binding on the
Dynamic Instability of Microtubules
Peter Hinow1, Vahid Rezania2, Manu Lopus3, Mary Ann
Jordan3 and Jack A. Tuszy´nski4
1Department of Mathematical Sciences, University of Wisconsin -- Milwaukee,
P.O. Box 413, Milwaukee, WI 53201, USA
E-mail: [email protected]
2Department of Physical Sciences, Grant MacEwan University, Edmonton AB, T5J
4S2, Canada
3Department of Molecular, Cellular, and Developmental Biology and the
Neuroscience Research Institute, University of California, Santa Barbara, CA 93106,
USA
4Cross Cancer Institute and Department of Physics, University of Alberta, Edmonton
AB, T6G 2J1, Canada
Abstract.
We propose a stochastic model that accounts for the growth, catastrophe and
rescue processes of steady state microtubules assembled from MAP-free tubulin in the
possible presence of a microtubule associated drug. As an example for the latter, we
both experimentally and theoretically study the perturbation of microtubule dynamic
instability by S-methyl-D-DM1, a synthetic derivative of the microtubule-targeted
agent maytansine and a potential anticancer agent. Our model predicts that among
drugs that act locally at the microtubule tip, primary inhibition of the loss of GDP
tubulin results in stronger damping of microtubule dynamics than inhibition of GTP
tubulin addition. On the other hand, drugs whose action occurs in the interior of the
microtubule need to be present in much higher concentrations to have visible effects.
Keywords: microtubules, dynamic instability, stochastic modeling
Submitted to: Phys. Biol.
PACS numbers: 87.10.Mn
1
1
0
2
l
u
J
4
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
8
8
2
4
.
0
1
0
1
:
v
i
X
r
a
Drug Binding and Dynamic Instability
2
1. Introduction
Microtubules are hollow and flexible cylindrical polymers of the protein tubulin that
form a major component of the cytoskeleton of eukaryotic cells. They play a central
role in maintenance of structural stability of the cell, intracellular vesicle transport and
chromosome separation during mitosis. The polymerization of tubulin into microtubules
and the subsequent catastrophic depolymerization have been studied extensively both
experimentally and theoretically, see [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14] for just
a few examples. The prevailing model to explain dynamic instability, the lateral cap
model, is that a cap of GTP tubulin at the growing tip is required for stability of the
polymer and that a loss of this GTP cap results in dramatic shortening of microtubules
[15].
In the paper [14] we proposed a partial differential equation model inspired by
dynamics of size-structured populations. The variables of that continuous model are
length distributions of microtubules and the amounts of free tubulin. The major
reactions are the polymerization of free GTP tubulin, the hydrolysis of assembled GTP
tubulin to GDP tubulin, the decay and rescue of microtubules without a GTP cap
and the recycling of free GDP tubulin to GTP tubulin. The model conserves the
total amount of tubulin in all its forms.
In addition, we allowed for the nucleation
of fresh microtubules at certain specified (short) lengths. With a small number of
parameters that have clear biochemical interpretations, we were able to reproduce
commonly observed experimental behaviors, such as oscillations in the amount of tubulin
assembled into microtubules. Obviously, the continuous model is not expected to
reproduce the lengths of individual microtubules. To this end, in this paper we present a
stochastic discrete model of microtubule dynamic instability and compare its predictions
to observations of microtubule lengths from in vitro experiments that also include the
presence of a dynamic instability suppressing drug.
Stochastic discrete models of biopolymer dynamics have been investigated, among
others, in [4, 5, 16, 17, 18, 12, 13, 19, 20]. Bolterauer et al. [4, 5] introduced a stochastic
model to study the dynamics of free microtubules using the master equation approach.
They showed that in the continuum limit the microtubule length distribution follows a
bell-shaped curve. Mishra et al. [20] applied a similar technique to elaborate the effect
of catastrophe-suppressing drugs on the dynamic instability of microtubules. They
assumed that drug molecules bind rapidly to free tubulin in the solution. Then, the
drug-tubulin complexes bind to the growing tips of the microtubules and reduce the
catastrophe frequency by stabilizing the microtubule caps. They also assumed that
the drug-tubulin complex has a lower attachment rate than drug-free tubulin. As a
result, one would expect to observe shorter drug-treated microtubules in the steady
state. While they found a qualitative reduction in catastrophe frequency, surprisingly,
they found that the drug-treated microtubules have the same length distribution as free
microtubules.
Here we present a stochastic model that represents the same set of reactions as our
Drug Binding and Dynamic Instability
3
continuous model in [14], except for the nucleation of fresh microtubules which is an
unnecessary complication for the case studied here. Similar to other stochastic models
[16, 17, 12, 13, 19, 20], our underlying model that describes microtubule dynamics
without drugs is based on an effective two-state model, switching between a growing
and shortening state. As in [14] and in contrast to other works such as [16, 20],
we couple the growth velocity of microtubules to the amount of free GTP tubulin
and the model is constructed so as to preserve the total amount of tubulin in all its
forms. Following [8], we include hydrolysis events of two kinds, namely scalar hydrolysis
(conversion of bound GTP tubulin to GDP tubulin with equal probability) and vectorial
hydrolysis (the probability is enhanced by a GDP tubulin neighbor). Furthermore, in
contrast to previous models that simulate one microtubule at a time, we simulate several
microtubules at the same time. This enables us to capture a more realistic situation
similar to one in experiments. Last but not least, a great advantage of our model is
that it can be implemented straightforwardly with the help of the Gillespie Algorithm
[21], an exact simulation method for stochastic chemical reactions. Our goal is to
explain individual length observations of microtubules growing without microtubule
associated proteins (MAPs) in vitro or length constraints (as in [17]). Moreover, we
investigate the effects of tubulin-binding drugs on microtubule dynamic instability.
These drugs belong roughly to one of two classes, namely those that bind to assembled
microtubules and those that bind to free tubulin [22]. This difference in behaviors
may be due to conformational changes of the α/β-tubulin heterodimer [23] (the tubulin
unit from now on) upon incorporation into the microtubule lattice that expose or hide
the binding site for the drug molecule. Apart from these different binding modes, the
effects on the microtubule reactions can also differ [24]. Some drugs mainly slow down
microtubule formation while others mainly prevent microtubule decay. This is not a
strict dichotomy in that some drugs can have multiple actions, depending on their
concentration. For example, vinblastine inhibits microtubule formation at high drug
concentrations, and inhibits microtubule decay at low concentrations [22]. Moreover,
there are multiple action mechanisms of drugs, for example either by preventing addition
of GTP tubulin,loss of GDP tubulin from an exposed tip or by preventing the hydrolysis
of GTP tubulin incorporated in the microtubule. In any case, within a living cell exposed
to anti-mitotic agents, mitosis cannot be completed and the cell dies. This property of
tubulin-binding drugs leads to many successful anti-cancer chemotherapeutics such as
paclitaxel, vincristine, vinblastine to name but a few. Maytansine and its derivatives
are known to suppress microtubule dynamics in vitro and in cells [25]. As an example,
we focus on the potential anticancer agent S-methyl-D-DM1, a synthetic derivative of
the microtubule-binding agent maytansine. Antibody-DM1 conjugates are currently
under clinical trials with promising results [26]. In our model, drugs act by decelerating
(or accelerating) binding or release reactions by a certain factor. Thus the strength
of the action can be quantified and linked to the binding energy through the standard
Arrhenius relation.
Drug Binding and Dynamic Instability
4
2. The stochastic model
1 , vk
2 , vk
We consider a linear model of the microtubule and disregard the fact that it actually
consists of 13-17 protofilaments arranged in a helical lattice. Tubulin can be added to
the microtubule in form of small oligomers of varying sizes [27].
Let m ≥ 1 be the number of dynamic microtubules. The state of each microtubule
is represented as a word vk = (. . . , vk
0 ), k = 1, . . . , m on the binary alphabet
{0, 1} where the letter 1 stands for a position occupied by a GTP tubulin unit and 0
stands for a position occupied by a GDP tubulin unit. The length of the microtubule
vk is denoted by vk. The number of GTP tubulin units within vk is denoted by I(vk).
The "tip" of the microtubule is the letter vk
0 and this is the only position where growth
or shrinkage can occur. Consecutive strings of 0s and 1s are called GDP zones and
GTP zones, respectively. The number of boundaries between such zones is denoted by
B(vk). The numbers of free GTP tubulin and GDP tubulin are denoted by N T and
N D, respectively. These numbers can be converted to and from concentrations, if the
volume in which the reactions take place is given. The following reactions occur (see
also Figure 1).
(i) Growth by attachment of GTP tubulin(s)
vk (cid:55)→ [vk 1 . . . 1
], N T (cid:55)→ N T − l,
(cid:124) (cid:123)(cid:122) (cid:125)
l
at rate (derived from mass action kinetics)
λN T (vk
0 + p(1 − vk
0 )).
(1)
Notice that every microtubule has its own growth reaction so that there are m
of them. The number of added GTP tubulin units l can be set to a fixed value
(say, 1) or drawn from a Poisson distribution with parameter L. The dimensionless
parameter p ≥ 0 is the propensity of a rescue event when a GDP tubulin unit at
the tip of the microtubule is exposed. In the simplest case, p = 1, attachment of a
new GTP tubulin is independent of the tip status.
(ii) Loss of a GDP tubulin (when the tip is in the state 0)
vk (cid:55)→ (. . . , vk
1 ), N D (cid:55)→ N D + 1,
2 , vk
at rate µGDP (1 − vk
0 ). Again there are m such shrinking reactions.
(iii) Loss of a GTP tubulin (when the tip is in the state 1)
vk (cid:55)→ (. . . , vk
at rate µGT P vk
0 .
1 ), N T (cid:55)→ N T + 1,
2 , vk
(iv) Hydrolytic conversion of a bound GTP tubulin vk∗ (cid:55)→(cid:101)vk∗
k=1 I(vk).
The index k∗ is chosen uniformly in the set {1, . . . , m} and a random position of
vk∗
that is occupied by 1 is changed to 0 (this hydrolysis mechanism is called scalar
hydrolysis in [8]).
at rate δsc
(cid:80)m
Drug Binding and Dynamic Instability
(v) Hydrolytic conversion of a bound GTP tubulin vk∗ (cid:55)→(cid:101)vk∗
Again, the index k∗ is chosen uniformly in the set {1, . . . , m} and the word (cid:101)vk∗
is created by selecting randomly a position of vk∗
where a 1 neighbors a 0 and
changing that 1 to 0 (this hydrolysis mechanism is called vectorial hydrolysis in [8],
see Figure 2, left panel).
k=1 B(vk).
at rate δvec
5
(cid:80)m
(vi) Recycling of free GDP tubulin to GTP tubulin
N D (cid:55)→ N D − 1, N T (cid:55)→ N T + 1,
at rate κN D. It is assumed that a sufficient amount of chemical energy in the form
of free GTP is always present.
This scheme can be simplified by setting some parameter values to zero. For
example, one may disregard the possibility of a bound GTP tubulin to be lost again
(µGT P = 0, cf. [8]), although other authors argue that this may take place in up to 90%
of all binding events [28]. The addition size l can be selected to be constant 1. The
hydrolysis reaction (iv) picks any bound GTP tubulin and changes it to a GDP tubulin,
thereby creating islands of GTP tubulin within the length of the microtubule. That
this is possible and important for the rescue process was recently shown by Dimitrov
et al. [29]. Both hydrolysis mechanisms in concert provide an indirect coupling of the
hydrolysis reaction to the addition of new GTP tubulin units [8].
Tubulin-binding drugs can bind to tubulin in one of two states, whether it is free
or bound within a microtubule. Here, we consider drugs that suppress microtubule
dynamic instability by specifically binding to microtubules. For every microtubule
encoded by a word v, we introduce a second word w = (. . . , w2, w1, w0) over the alphabet
{0, 1} (the drug state), of equal length as v. Here wi = 1, if the tubulin unit at position
vi is occupied by a drug molecule and wi = 0 otherwise. There are binding events of
drug molecules to unoccupied sites and release of drug molecules from the microtubule.
Let E(w) be the number of available sites for drug binding and let F (w) be the number
of drug occupied sites. The latter is always the sum of the entries 1 in w while the
former may be only a subset of entries 0 in w. We have the association and dissociation
events
(vii) Binding of drug to tubulin units within the microtubule
at rate ρD(cid:80)m
wk∗ (cid:55)→ (cid:101)wk∗
k=1 E(wk). The new word (cid:101)wk∗
, D (cid:55)→ D − 1,
is obtained by selecting randomly one
letter 0 among the sites available for binding and changing it to 1. This set may be
the set of all entries 0 or the entries 0 that are within a certain distance from the
tips or the unoccupied tips alone.
(viii) Release of drug from the occupied sites of the microtubule
at rate σ(cid:80)m
wk∗ (cid:55)→ (cid:101)wk∗
k=1 F (wk). The new word (cid:101)wk∗
, D (cid:55)→ D + 1,
selected letter 1 in any of the drug words to 0.
is obtained by changing one randomly
Drug Binding and Dynamic Instability
6
The reactions (i) -- (vi) have the same outcomes as far as changes in numerical quantities
are concerned, however the rates of reactions (i), (ii) and (iii) have a more complicated
dependence upon the status of the microtubule tip. Since the tip can now have four
different states, the attachment process (i) to microtubule k occurs at rate
0 + p(1 − vk
0 ))(1 − wk
0 ) + r(vk
0 + p(1 − vk
0 ))wk
0
λN T(cid:0)(vk
(cid:1) ,
(2)
where the dimensionless non-negative constant r modulates the attachment propensity
compared to the drug-free tip, see Equation (1). Small values of r would mean that
attachment of new GTP tubulin units is hindered by drug molecules bound to the
tip. On the other hand, values r > 1 would increase microtubule polymerization. The
shrinking reactions (ii) and (iii) occur at rates
µGDP (1 − vk
0 )((1 − wk
0 ) + qwk
0 ),
and µGT P vk
(3)
where a small value of q ≥ 0 implies a high level of protection afforded by a drug molecule
bound to the tip. If a drug bound tubulin can fall off a microtubule (i.e. if q > 0), then
the drug-tubulin compound is assumed to dissociate immediately. We refer to Figure 1
for the possible interactions of the drug with the microtubules (as implemented in this
paper).
0 ) + qwk
0 ),
0 ((1 − wk
There are also possible drug actions beyond the microtubule tips. As was first
discovered by Lin and Hamel [30], a drug can also inhibit the hydrolysis of bound GTP
tubulin. This can be implemented by "splitting" the scalar and vectorial hydrolysis
reactions (iv) and (v). More precisely, let I0(vk) and I1(vk) be the number of GTP
(cid:80)m
tubulin units within the microtubule word vk that are unoccupied respectively occupied
by a drug molecule. Then the scalar hydrolysis reactions (iv0) and (iv1) occur at rates
k=1 I1(vk) (with 0 ≤ s ≤ 1). A similar splitting is
δsc
used for the vectorial hydrolysis reaction (v).
k=1 I0(vk) (as before) and sδsc
(cid:80)m
3. Materials and Methods
Tubulin (15 µM ), phosphocellulose purified, MAP-free, was assembled on the ends of
sea urchin (Strongylocentrotus purpuratus) axoneme fragments at 30◦C in 87 mmol/L
Pipes, 36 mmol/L Mes, 1.4 mmol/L MgCl2, 1 mmol/L EGTA, pH 6.8 (PMME
buffer) containing 2 mmol/L GTP for 30 min to achieve steady state. We used
a 100 nmol/L concentration of S-methyl-D-DM1 (N 2(cid:48)
-(3-thiomethyl-1-
oxopropyl)-D-maytansine [31], see Figure 2, right panel), which had no considerable
effect on microtubule polymer mass, to analyze their individual effects on dynamic
instability. Time-lapse images of microtubule plus ends were obtained at 30◦C by
video-enhanced differential
interference contrast microscopy at a spatial resolution
of 0.3 µm using an Olympus IX71 inverted microscope with a 100 × oil immersion
objective (NA = 1.4). The end of an axoneme that possesses more, faster growing, and
longer microtubules than the other end was designated as the plus end as described
previously [32, 33]. Microtubule dynamics were recorded for 40 min at 30◦C, capturing
∼ 10 min long videos for each area under observation. The microtubules were tracked
-deacetyl-N 2(cid:48)
Drug Binding and Dynamic Instability
7
approximately every 3 s using RTMII software, and the life-history data were obtained
using IgorPro software (MediaCybernetics, Bethesda, MD) [34].
We have programmed the reactions (i) -- (viii) using the Gillespie Algorithm [21] (in
Java, available from the corresponding author upon request). This algorithm simulates
the chemical reactions as collisions of particles in real time and its parameters are the
actual reaction rates, not probabilities.
If the empty word is reached, then a new
microtubule is created (a word containing a single GTP tubulin unit). The gain in
length due to addition of a single tubulin unit is taken to be (cid:96) = 8 nm/13 = 0.6 nm; see
e.g. [8, Equation (2)].
4. Results
Results of the in vitro experiments are shown in Figures 3 and 7. During periods of
growth, the untreated microtubules grow at a velocity of ≈ 3 µm min−1 while during
periods of shrinkage, microtubules shrink at ≈ 20 µm min−1. While they are infrequent,
we attribute occasional periods of stagnation to the spatial resolution of the microscope
of approximately 400 tubulin dimer units.
The left panels of Figures 4, 5 and 6 show possible simulations of the control
scenario in the absence of drugs. The choice of appropriate parameter values for the
simulations is a difficult problem, since different values have been reported in different
literature sources and some parameters have only been estimated on the basis of the
Arrhenius Equation (see [14, Table 1] for some ranges). While there is no unique
choice of parameter values, we observe a good agreement of the simulated and observed
growth and shrinking velocities, for the choice of parameter values in Table 1, see Figure
4. To quantify the agreement between the simulations and the experimental data we
used the absolute Fourier spectra [35] of the length time series. We first re-sampled
both the experimental data and the numerical simulations on equispaced time grids at
approximately 0.4 Hz. In order to make different simulations comparable, we subtracted
the mean from each length time series so that the resulting normalized lengths have mean
zero. If ln, n = 0, . . . , N − 1 are the re-sampled normalized lengths, then the discrete
Fourier transform is given by the absolute values of
(cid:18)
(cid:19)
N−1(cid:88)
n=0
Lk =
ln exp
−2πikn
N
k = 0, . . . , N − 1.
,
This is conveniently done with the help of the fast Fourier transform routine fft in
scilab‡. Results are shown in the right panels of Figures 3 and 4, showing a good
agreement among experimental and simulated data of the location and the height of the
peak of the averaged spectra. In Figure 6, a sensitivity check is run with significantly
larger parameter values λ = 1.0 ((cid:96)s)−1, µGDP = 2000 s−1, δsc = δvec = 3.0 ((cid:96)s)−1 and
κ = 0.5 s−1, see Table 1 for comparison. As a result, higher oscillation frequencies
would be expected in the Fourier spectra due to higher rate of polymerization,
‡ Open source software; available at www.scilab.org.
Drug Binding and Dynamic Instability
8
depolymerization, hydrolysis and recycling events. This is clearly shown in the right
panel of Figure 6, as the spectrum visibly shifts towards higher frequencies.
The presence of the drug clearly decelerates the dynamic activity of the
microtubules as can be seen from Figure 7. This is also visible in the reduced peak
heights of the Fourier spectra, while the position of the peak, i.e. the main frequency
of the oscillations remains unchanged, at least not discernible on the 0.4 Hz frequency
grid, see also Figure 9. With the same parameter values as in the control in Figure 4,
we simulated the presence of a microtubule-binding drug that suppresses strongly the
addition of new GTP tubulin (r = 0.01), completely inhibits the loss of tubulin at the
tip (q = 0) and does not affect the hydrolysis of bound GTP tubulin (s = 1). The
open drug binding sites are all entries 0 in the drug words. The results are shown in
Figure 8. There are 400 drug molecules present in the simulation (compared to ≈ 105
tubulin units). This amount of roughly 100 drug molecules per microtubule matches
approximately the concentration of S-methyl-D-DM1 in the experiments.
The nonlocal drug action mechanism where the bound drug inhibits hydrolysis of
GTP tubulin is much less effective in suppressing microtubule dynamic instability. In
Figure 10 we show simulations of 5 microtubules built of approximately 105 tubulin
units in the presence of a drug that completely inhibits hydrolysis reactions (iv) and
(v), i.e. s = 0 without affecting the binding and loss reactions, r = q = 1. We observe
that a much higher amount of drug, namely of the order of tubulin units, is required to
have any visible effect, while at the same time, long periods of shortening still occur.
In order to better understand the influence of a drug acting at the microtubule
tip, we systematically varied the parameters r and q, keeping all other parameters and
the drug concentration constant. We consider the reduction of the peak height of the
absolute Fourier spectra relative to the control scenario. If the drug molecules are free to
bind any of the open drug binding sites, then a complete repression of the loss reactions
(ii) and (iii), i.e. q = 0 is required to efficiently suppress the dynamic instability of
microtubules, see Figure 11, left panel. An alternative is to allow drug molecules to
bind only at the tip [33]. In that case, the suppression effect persists for weaker drug
effects (Figure 11, right panel). However, it is still more important to suppress the loss
reactions (small value of q) than to suppress the growth reaction (r ≈ 1 is admissible).
Notice that these predictions of the model are drawn from the shapes of the surfaces in
Figure 11, not from particular values.
5. Discussion
Spatial and temporal regulation of the dynamic instability of microtubules is essential
to carry out several vital cellular functions. In cells, the dynamicity of microtubules
is regulated by a number of proteins such as MAPs [11], G proteins [36], and the
plus end tracking proteins, including EB1 [11]. Perturbations in the innate dynamicity
of microtubules induce cell cycle arrest and thereby inhibit cell proliferation. Thus,
compounds that target microtubules are potential anticancer drugs. Our recent studies
Drug Binding and Dynamic Instability
9
have found synthetic derivatives of maytansine such as DM1 (for drug maytansinoid 1)
that can be conjugated to tumor-specific antibodies as potent suppressors of microtubule
dynamics [25, 33]. Antibody-DM1 Conjugates are under clinical evaluation, and they
show promising early results. Thus, the synthetic derivative of maytansine S-methyl-D-
DM1 is a good example to complement our modeling studies.
catastrophic shrinking,
In this paper we have presented a detailed reaction scheme for microtubule
polymerization, GTP tubulin hydrolysis,
recycling and
interaction with tubulin-binding drugs. Our in silico simulations show good agreement
of Fourier spectra with experimental data of growing microtubules under control
conditions and treated with the maytansine derivative S-methyl-D-DM1. Our model
allows to accommodate a wide variety of drug binding mechanisms and interactions of
the drug with the normal microtubule polymerization and depolymerization processes.
We find that drugs that act at the microtubule tip by inhibiting addition of GTP tubulin
and loss of GDP tubulin are effective suppressors of microtubule dynamic instability.
This would also be the action mechanism of S-methyl-D-DM1. Among these two actions,
the inhibition of the loss of GDP tubulin is more important than the inhibition of the
growth reaction. A localized binding to the tip inhibits dynamic instability even more.
On the other hand, drugs whose action is to inhibit hydrolysis of bound GTP tubulin
need to be present at numbers comparable to the number of tubulin units to have a
visible effect on microtubule dynamic instability.
Our assumption in modeling the drug binding reaction (vii) has been that either
the drug molecule binds to any open binding site with equal probability or that it binds
only at an open site at the tip. The former is the binding mode of a drug like paclitaxel
that stabilizes a microtubule along its entire length. Other drugs, such as vinblastine
bind with high affinity only at the microtubule plus end [22]. In future work we will
implement a probability of drug binding that decreases with increasing distance from
the tip.
taxol
(paclitaxel)
It is well known that some drugs have different effects at different concentrations.
For example,
increases microtubule polymerization at high
concentrations (50 taxol molecules per 100 tubulin molecules), while it reduces the
rate of shortening at low concentrations (1 taxol molecule per 100 tubulin molecules)
[22]. This suggests the need for a "nonlocal" generalization of the perturbation of the
binding and shrinking processes, in contrast to the present choices in Equations (2) and
(3).
The design of therapeutic drugs requires long and tedious searches among all
possible binding sites on the target. Identifying the most favorable binding site (with the
lowest binding energy), however, is needed to design or discover a drug compound with
the highest possible efficacy. Information on drug-ligand binding affinities (for tubulin
and tubulin-binding drugs in particular) can be extracted by careful comparison of
simulation results and experimental data. Such estimations of binding energies will be
addressed in a future study.
Drug Binding and Dynamic Instability
10
Acknowledgments
Part of this research was carried out during visits at the University of Wisconsin
-- Milwaukee and the University of Alberta, Edmonton. We thank our respective
institutions for their warm hospitality. PH is partially supported by NSF grant DMS-
1016214. Financial support was partially supplied by grants from NSERC, Alberta's
Advanced Education and Technology, the Allard Foundation and the Alberta Cancer
Foundation to JAT and from grant NIH CA57291 and a gift from ImmunoGen Inc. to
MAJ. We thank Dr. Ravi Chari (ImmunoGen Inc., Waltham, MA) for providing us with
the drug S-methyl-D-DM1 and Bruce Fenske and Philip Winter (University of Alberta)
for implementing the Gillespie algorithm. We are much indebted to two anonymous
readers whose comments greatly helped to improve the paper.
References
[1] Jackson, M. B. and S. A. Berkovitz. Nucleation and the kinetics of microtubule assembly. Proc.
Natl. Acad. Sci. USA, 77:7302 -- 7305, 1980.
[2] Mandelkow, E., E. M. Mandelkow, H. Hotani, B. Hess, and S. C. Muller. Spatial patterns from
oscillating microtubules. Science, 246:1291 -- 1293, 1989.
[3] Walker, R. A., E. T. O'Brien, N. K. Pryer, M. E. Soboeiro, W. A. Voter, H. P. Erickson, and E. D.
Salmon. Dynamic instability of individual microtubules analyzed by video light microscopy:
rate constants and transition frequencies. J. Cell Biol., 107:1437 -- 1448, 1988.
[4] Bolterauer, H., H. -- J. Limbach, and J. A. Tuszy´nski. Microtubules: strange polymers inside the
cell. Bioelectrochem. Bioenerg., 48:285 -- 295, 1999.
[5] Bolterauer, H., H. -- J. Limbach, and J. A. Tuszy´nski. Models of assembly and disassembly of
individual microtubules: stochastic and averaged equations. J. Biol. Phys., 25:1 -- 22, 1999.
[6] Mitchison, T. and M. Kirschner. Dynamic instability of microtubule growth. Nature, 312:237 -- 242,
1984.
[7] Flyvbjerg, H., T. E. Holy, and S. Leibler. Stochastic dynamics of microtubules: A model for caps
and catastrophes. Phys. Rev. Lett., 73:2372 -- 2375, 1994.
[8] Flyvbjerg, H., T. E. Holy, and S. Leibler. Microtubule dynamics: caps, catastrophes, and coupled
hydrolysis. Phys. Rev. E, 54:5538 -- 5560, 1996.
[9] Jobs, E., D. E. Wolf, and H. Flyvbjerg. Modeling microtubule oscillations. Phys. Rev. Lett.,
79:519 -- 522, 1997.
[10] Rezania, V., O. Azarenko, M. A. Jordan, H. Bolterauer, R. F. Luduena, J. T. Huzil, and J. A.
Tuszy´nski. Microtubule assembly of isotypically purified tubulin and its mixtures. Biophys. J.,
95:1 -- 16, 2008.
[11] Lopus, M., M. Yenjerla, and L. Wilson. Microtubule Dynamics, volume 3 of Wiley Encyclopedia
of Chemical Biology, pages 153 -- 160. 2009.
[12] van Buren, V., D. J. Odde, and L. Cassimeris. Estimates of lateral and longitudinal bond energies
within the microtubule lattice. Proc. Natl. Acad. Sci. USA, 99:6035 -- 6040, 2002.
[13] van Buren, V., L. Cassimeris, and D. J. Odde. Mechanochemical model of microtubule structure
and self-assembly kinetics. Biophys. J., 89:2911 -- 2926, 2005.
[14] Hinow, P., V. Rezania, and J. A. Tuszy´nski. A continuous model for microtubule dynamics with
collapse, rescue and nucleation. Phys. Rev. E, 80:031904, 2009. arXiv:0811.2245.
[15] Bayley, P. M., M. J. Schilstra, and S. R. Martin. A lateral cap model of microtubule dynamic
instability. FEBS Letters, 259:181 -- 184, 1989.
Drug Binding and Dynamic Instability
11
[16] Antal, T., P. L. Krapivsky, S. Redner, M. Mailman, and B. Chakraborty. Dynamics of an
idealized model of microtubule growth and catastrophe. Phys. Rev. E, 76:041907, 2007.
arXiv:q-bio/0703001.
[17] Gregoretti, I. V., G. Margolin, M. S. Alber, and H. V. Goodson. Insights into cytoskeletal behavior
from computational modeling of dynamic microtubules in a cell -- like environment. J. Cell Sci.,
119:4781 -- 4788, 2006. arXiv:q-bio/0604023.
[18] Matzavinos, A. and H. G. Othmer. A stochastic analysis of actin polymerization in the presence
of twinfilin and gelsolin. J. Theor. Biol., 249:723 -- 736, 2007.
[19] Brun, L., B. Rupp, J. Ward, and F. Nedelec. A theory of microtubule catastrophes and their
regulation. Proc. Natl. Acad. Sci. USA, 106:21173 -- 21178, 2009.
[20] Mishra, P. K., A. Kunwar, S. Mukherji, and D. Chowdhury. Dynamic instability of
microtubules: Effect of catastrophe-suppressing drugs. Phys. Rev. E, 72:051914, 2005.
arXiv:cond-mat/0310546.
[21] Gillespie, D. T. Exact stochastic simulation of coupled chemical reactions. J. Phys. Chem.,
81:2340 -- 2361, 1977.
[22] Jordan, M. A. and K. Kamath. How do microtubule-targeted drugs work? An overview. Current
Cancer Drug Targets, 7:730 -- 742, 2007.
[23] Bennett, M. J., J. K. Chik, G. W. Slysz, T. Luchko, J. A. Tuszy´nski, D. L. Sackett, and D. C.
Schriemer. Structural mass spectrometry of the αβ-tubulin dimer supports a revised model of
microtubule assembly. Biochemistry, 48:4858 -- 4870, 2009.
[24] Correia, J. J. and S. Lobert. Molecular mechanisms of microtubule acting cancer drugs. In T. Fojo,
editor, Microtubules in Health and Disease, pages 21 -- 46. 2008.
[25] Oroudjev, E., M. Lopus, L. Wilson, C. Audette, C. Provenzano, H. Erickson, Y. Kovtun,
R. Chari, and M. A. Jordan. Maytansinoid-antibody conjugates induce mitotic arrest by
suppressing microtubule dynamic instability. Mol. Cancer Therapeutics, 9:2700 -- 2713, 2010.
doi:10.1158/1535-7163.MCT-10-0645.
[26] Lopus, M. Antibody-DM1 conjugates as cancer therapeutics. Cancer Lett., 307:113 -- 118, 2011.
[27] Mozziconacci, J., L. Sandblad, M. Wachsmuth, D. Brunner, and E. Karsenti. Tubulin dimers
oligomerize before their incorporation into microtubules. PLoS ONE, 3:e3821, 2008.
[28] Howard, J. and A. A. Hyman. Growth, fluctuation and switching at microtubule plus ends. Nat.
Rev. Mol. Cell Biol., 10:569 -- 574, 2009.
[29] Dimitrov, A., M. Quesnoit, S. Moutel, I. Cantaloube, C. Pous, and F. Perez. Detection of GTP-
Tubulin conformation in vivo reveals a role for GTP remnants in microtubule rescues. Science,
322:1353 -- 1356, 2008.
[30] Lin, C. M. and E. Hamel. Effects of inhibitors of tubulin polymerization on GTP hydrolysis. J.
Biol. Chem., 256:9242 -- 9245, 1981.
[31] Widdison, W. C., S. D. Wilhelm, E. E. Cavanagh, K. R. Whiteman, B. A. Leece, Y. Kovtun,
V. S. Goldmacher, H. Xie, R. M. Steeves, R. J. Lutz, R. Zhao, L. Wang, W. A. Blattler, and
R. V. J. Chari. Semisynthetic maytansine analogues for the targeted treatment of cancer. J.
Med. Chem., 49:4392 -- 4408, 2006.
[32] Yenjerla, M., M. Lopus, and L. Wilson. Analysis of Dynamic Instability of Steady-State
Microtubules In Vitro by Video-Enhanced Differential Interference Contrast Microscopy,
volume 95 of Methods in Cell Biology, chapter 11, pages 189 -- 206. 2010.
[33] Lopus, M., E. Oroudjev, L. Wilson, S. Wilhelm, W. Widdison, R. Chari, and M. A. Jordan.
Maytansine and cellular metabolites of antibody-maytansinoid conjugates strongly suppress
microtubule dynamics by binding to microtubules. Mol. Cancer Therapeutics, 9:2689 -- 2699,
2010. doi:10.1158/1535-7163.MCT-10-0644.
[34] Yenjerla, M., N. E. LaPointe, M. Lopus, C. Cox, M. A. Jordan, S. C. Feinstein, and L. Wilson.
The neuroprotective peptide NAP does not directly affect polymerization or dynamics of
reconstituted neural microtubules. J. Alzheimer Dis., 19:1377 -- 1386, 2010.
[35] Odde, D. J., H. M. Buettner, and L. Cassimeris. Spectral analysis of microtubule assembly
Drug Binding and Dynamic Instability
12
dynamics. AIChE Journal, 42:1434 -- 1442, 1996.
[36] Dav´e, R. H., W. Saengsawang, S. Dav´e, M. Lopus, L. Wilson, and M. M. Rasenick. A molecular
and structural mechanism for G-protein mediated microtubule destabilization. J. Biol. Chem,
286:4319 -- 4328, 2011.
[37] Carlier, M. F., R. Melki, D. Pantaloni, T. L. Hill, and Y. Chen. Synchronous oscillations in
microtubule polymerization. Proc. Natl. Acad. Sci. USA, 84:5257, 1987.
6. Tables and Figures
parameter
λ
p
µGDP
µGT P
δsc
δvec
κ
L
ρ
σ
value
0.4 ((cid:96)s)−1
0.05
800 s−1
1.5 s−1
1.2 ((cid:96)s)−1
1.2 ((cid:96)s)−1
0.1 s−1
6
1.0s−1
0.1 s−1
remark
addition rate of GTP tubulin to GTP tip
reduction of GTP addition to GDP tip
loss of GDP tubulin from tip
loss of GTP tubulin from tip
rate of scalar hydrolysis
rate of vectorial hydrolysis
rate of GDP tubulin recycling
average GTP tubulin addition size
rate of drug binding
rate of drug release
reference
[16]
[16]
[3]
[8, 28]
[8]
[8]
[37]
[27]
Table 1. Baseline values of parameters used in the stochastic simulations. Here
(cid:96) = 0.6 nm is the gain in length by addition of a single tubulin unit. The references
provide further discussion and sometimes comparable values.
Figure 1. The reactions and possible drug interactions implemented in our stochastic
model. A drug may also promote polymerization and depolymerization.
Drug Binding and Dynamic Instability
13
Figure 2. (Left panel) Schematic depiction of the two hydrolysis mechanisms. The
scalar hydrolysis reaction (p. 4 (iv), top) picks a random bound GTP tubulin and
changes it into a bound GDP tubulin. The vectorial hydrolysis reaction (p. 5 (v),
bottom) occurs at a boundary between a GDP zone and a GTP zone. (Right panel)
Structural formula of the maytansine analog S-methyl-D-DM1.
Figure 3.
(Left panel) Length time series of 10 microtubules in absence of drug.
Notice that life histories from several experiments are plotted in the same diagram.
(Right panel) Absolute Fourier spectra of the control experimental data that were
normalized to mean length zero. The thick blue line is the average of the 10 individual
spectra.
05101520253035024681001002003004005006007008000.00.51.01.52.02.53.03.5 Drug Binding and Dynamic Instability
14
Figure 4.
(Left panel) Simulation of m = 5 microtubules starting from random
initial states, with a total of ≈ 105 tubulin units. The parameter values are as in
Table 1. GTP tubulin is added in the form of oligomers whose length is Poisson
distributed with with mean L = 6. The resulting average growth velocity during
periods of growth is approximately 2 µm min−1 while the resulting shrinking velocity is
approximately 20 µm min−1. (Right panel) Absolute Fourier spectra of the simulation
data, normalized to mean length zero. The thick blue line is the average of the 5
individual spectra.
Figure 5. (Left panel) Simulation of m = 5 microtubules with parameter values as
in Table 1 and Figure 4 except that GTP tubulin is added in units of fixed length
L = 1. (Right panel) Absolute Fourier spectra of the simulation data, normalized to
mean length zero and their average.
024681012141618012345678901002003004005000.00.51.01.52.02.53.03.502468101214161801234567890501001502000.00.51.01.52.02.53.03.5 Drug Binding and Dynamic Instability
15
Figure 6.
(Left panel) Simulation of m = 5 microtubules with parameter values
λ = 1.0 ((cid:96)s)−1, µGDP = 2000 s−1, δsc = δvec = 3.0 ((cid:96)s)−1 and κ = 0.5 s−1, all of which
are larger than those in Table 1. (Right panel) The corresponding absolute Fourier
spectra show a visible shift towards higher frequencies.
Figure 7. (Left panel) Length time series of 10 microtubules in presence of the drug S-
methyl-D-DM1. (Right panel) The corresponding absolute Fourier spectra, normalized
to mean length zero. The thick blue line is the average of the 10 individual spectra.
Figure 8. (Left panel) Simulation of m = 5 microtubules in the presence of 400 drug
molecules, with a total of ≈ 105 tubulin units. The parameter values are as in Figure
4, in addition r = 0.01, q = 0 and s = 1. Here the drug molecules bind to any open
site with equal probability. (Right panel) The corresponding absolute Fourier spectra.
This simulation suggests a possible action mechanism for the drug S-methyl-D-DM1.
051015200123456780501001502002503000.00.51.01.52.02.53.03.50510152025303502468100501001502002503000.00.51.01.52.02.53.03.502468101214161801234567890204060801001201401601800.00.51.01.52.02.53.03.5 Drug Binding and Dynamic Instability
16
Figure 9. Detail of the average absolute Fourier spectra from Figures 3, 4, 7 and 8.
Within the resolution of the frequency grid ≈ 0.11 min−1, there is no discernible shift
of the position of the peaks.
Figure 10. Simulation of m = 5 microtubules with a total of ≈ 105 tubulin units
in the presence a drug that completely inhibits GTP tubulin hydrolysis. The relevant
parameter values are r = q = 1 and s = 0. The total amount of drug is 20000 (left
panel) respectively 40000 (right panel).
0501001502002500.00.20.40.60.81.002468101214161801234567890246810121416180123456789 Drug Binding and Dynamic Instability
17
Figure 11. The reduction of microtubule dynamic instability relative to the untreated
case as the drug effects vary. Shown are the relative peak heights of the averaged
absolute Fourier spectra of 20 microtubules at a concentration of 100 drug molecules
for every microtubule. The drug molecules bind either at any open site along the
microtubule (left panel) or at the tip only (right panel). The drug does not affect the
hydrolysis of bound GTP tubulin (s = 1).
0.200.40.160.60.200.80.120.161.00.120.081.20.080.041.40.040.000.001.00.10.80.21.00.90.60.30.80.70.40.60.40.50.40.50.20.30.20.60.10.00.0 |
1811.12242 | 3 | 1811 | 2019-01-25T09:12:17 | Theory of mechano-chemical patterning in biphasic biological tissues | [
"physics.bio-ph",
"q-bio.CB",
"q-bio.TO"
] | The formation of self-organized patterns is key to the morphogenesis of multicellular organisms, although a comprehensive theory of biological pattern formation is still lacking. Here, we propose a minimal model combining tissue mechanics to morphogen turnover and transport in order to explore new routes to patterning. Our active description couples morphogen reaction-diffusion, which impact on cell differentiation and tissue mechanics, to a two-phase poroelastic rheology, where one tissue phase consists of a poroelastic cell network and the other of a permeating extracellular fluid, which provides a feedback by actively transporting morphogens. While this model encompasses previous theories approximating tissues to inert monophasic media, such as Turing's reaction-diffusion model, it overcomes some of their key limitations permitting pattern formation via any two-species biochemical kinetics thanks to mechanically induced cross-diffusion flows. Moreover, we describe a qualitatively different advection-driven Keller-Segel instability which allows for the formation of patterns with a single morphogen, and whose fundamental mode pattern robustly scales with tissue size. We discuss the potential relevance of these findings for tissue morphogenesis. | physics.bio-ph | physics | Theory of mechano-chemical patterning in biphasic biological tissues
Pierre Recho∗§, Adrien Hallou†§and Edouard Hannezo‡§
9
1
0
2
n
a
J
5
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
3
v
2
4
2
2
1
.
1
1
8
1
:
v
i
X
r
a
The formation of self-organized patterns is key to the
morphogenesis of multicellular organisms, although a
comprehensive theory of biological pattern formation
is still lacking. Here, we propose a minimal model com-
bining tissue mechanics to morphogen turnover and
transport in order to explore new routes to patterning.
Our active description couples morphogen reaction-
diffusion, which impact on cell differentiation and tis-
sue mechanics, to a two-phase poroelastic rheology,
where one tissue phase consists of a poroelastic cell
network and the other of a permeating extracellular
fluid, which provides a feedback by actively transport-
ing morphogens. While this model encompasses previ-
ous theories approximating tissues to inert monopha-
sic media, such as Turing's reaction-diffusion model,
it overcomes some of their key limitations permitting
pattern formation via any two-species biochemical ki-
netics thanks to mechanically induced cross-diffusion
flows. Moreover, we describe a qualitatively different
advection-driven Keller-Segel instability which allows
for the formation of patterns with a single morphogen,
and whose fundamental mode pattern robustly scales
with tissue size. We discuss the potential relevance of
these findings for tissue morphogenesis.
How symmetry is broken in the early embryo to give
rise to a complex organism, is a central question in de-
velopmental biology. To address this question, Alan Tur-
ing proposed an elegant mathematical model where two
reactants can spontaneously form periodic spatial pat-
terns through an instability driven by their difference in
diffusivity [1]. Molecular evidence of such a reaction-
diffusion scheme in vivo remained long elusive, until pairs
of activator-inhibitor morphogens were proposed to be re-
sponsible of pattern formation in various embryonic tis-
sues [2 -- 9]. Interestingly, these studies also highlight some
theoretical and practical limitations of existing reaction-
diffusion models, including the fact that Turing patterns
require the inhibitor to diffuse at least one order of magni-
tude faster than the activator (DI /DA > 10) [3], although
most morphogens are small proteins of similar molecular
weights, implying that DI /DA ≈ 1. As a consequence,
the formation of Turing patterns in vivo should result
from other properties of the system such as selective mor-
phogen immobilisation [10 -- 12] or active transport [13] as
demonstrated in synthetic sytems. Moreover, reaction-
∗LIPHY, CNRS UMR 5588 & Universit´e Grenoble Alpes, F-
†Cavendish Laboratory, Department of Physics, University of
Cambridge, Cambridge, CB3 0HE, UK; Wellcome Trust/CRUK
Gurdon Institute, University of Cambridge, Cambridge, CB2 1QN,
UK; Wellcome Trust/MRC Stem Cell Institute, University of Cam-
bridge, Cambridge, CB2 1QR, UK.
38000 Grenoble, France.
‡IST Austria, Am Campus 1, 3400 Klosterneuburg, Austria.
§A.H, P.R and E.H contributed equally to this work.
E-
[email protected], [email protected] or
mail:
[email protected]
diffusion models of pattern formation entail a number of
restrictions regarding the number and interactions of mor-
phogens, and pattern scaling with respect to the tissue
size, which have been all limiting their quantitative appli-
cability in vivo. While the genetic and biochemical aspects
of developmental pattern formation have been the focus of
most investigations, the interplay between mechanics and
biochemical processes in morphogenesis started to unfold
following some pioneering contributions [14]. The crucial
role played by multiphasic tissue organisation and active
cell behaviours in biological pattern formation is now an
active field of research [15 -- 18].
In this article, we derive a general mathematical for-
mulation of tissues as active biphasic media coupled with
reaction-diffusion processes, where morphogen turnover
inside cells, import/export at the cell membrane and ac-
tive mechanical transport in the extracellular fluid are cou-
pled together through tissue mechanics. While encompass-
ing classical reaction-diffusion results [1 -- 4], for instance
allowing import-export mechanisms to rescale diffusion
coefficients and to form patterns with equally diffusing
morphogens [11], this theory provides multiple new routes
to robust pattern formation.
In particular, assuming a
generic coupling between intracellular morphogen concen-
tration and poroelastic tissue mechanics, we demonstrate
the existence of two fundamentally different non-Turing
patterning instabilities, respectively assisted and driven
by advective extracellular fluid flows, explaining pattern
formation with only a single morphogen with robust scal-
ing properties, and how patterning can be independent
of underlying morphogen reaction schemes. Finally, we
discuss the biological relevance of such a model, and in
particular its detailed predictions that could be verified in
vivo.
Derivation of the model
As sketched in Fig. 1(a), we model multicellular tissues as
continuum biphasic porous media of typical length l, with
a first phase consisting of a poroelastic network made of
adhesive cells of arbitrary shape and typical size lc (with
local volume fraction φ), and a second phase of aqueous
extracellular fluid permeating in-between cells in gaps of
a characteristic size li. These two internal length scales
disappear in the coarse-graning averaging over a repre-
sentative volume element of typical lengthscale lr satis-
fying li,c (cid:28) lr (cid:28) l. Both phases are separated by cell
membranes, actively regulating the interfacial exchange of
water and other molecules thanks to genetically controlled
transport mechanisms [19,20]. At the boundary of the do-
main, no-flux boundary conditions are imposed such that
the system is considered in isolation. We present below
the main steps of the model derivation, which are detailed
in SI Appendix.
1
Figure 1: Model for pattern formation in active biphasic tissues. (a) Schematic of the model: (Left) Cells form a
poroelastic network, permeated by extracellular fluid, where three natural length scales can be defined: the interstitial space
size (li), the characteristic cell size (lc) and the tissue size (l). (Right) Biochemical interactions between morphogens, A and I,
take place inside the cell and are described by their respective turnover rate functions f (A, I) and g(A, I). A and I are exported
across the cell membrane at rates λA,I and imported at rates γA,I , respectively. In the extracellular space, both A and I spread
freely by diffusion at the same rate D, or can be advected by the fluid at velocity ve. (b) Evolution of the effective diffusion
coefficient as a function of time and space scales. At shorter distances and times, diffusive behaviour of morphogens is described
by a molecular diffusion coefficient, DFick. At intermediate scales, the diffusive motion of morphogens starts to be hindered by
cells and the global diffusion coefficient, D, depends of the tissue spatial organisation through φ∗. At larger scales, morphogen
diffusion is controlled by dynamic interactions with cells (import/export, adsorption/desorption,) and an effective coefficient
DKA,I [9].
Intracellular morphogen dynamics
Morphogens enable cell-cell communication across the tis-
sue and determine cell fate decisions. Importantly, most
known morphogens cannot directly react together and as
such, have to interact "through" cells (or cell membranes)
where they are produced and degraded [20]. Concentra-
tion fields of two morphogens, Ai,e((cid:126)r, t) and Ii,e((cid:126)r, t), are
thus defined separately in each phase of the system, indices
(i, e) denoting intra- and extra-cellular phases, respec-
tively. The conservation laws of the intracellular phase,
which cannot be transported, read:
∂t(φAi) = f (Ai, Ii) + γAAe − λAAi
∂t(φIi) = g(Ai, Ii) + γI Ie − λI Ii
(1)
where ∂t denotes the partial derivative with respect to
time and γA,I (resp. λA,I ) the import (resp. export) rates
of morphogens (which can also describe immobilization
rates at the cell membrane). We also introduce f and g,
the non-linear morphogen turnover rates describing their
production and degradation by cells, with a single stable
equilibrium solution f (A∗
i ) = 0. Finally,
we introduce the transmembrane transport equilibrium
constants by KA = λA/γA and KI = λI /γI . Although the
import/export coefficients KA,I could in principle depend
on morphogen concentrations, this constitutes a non-linear
effect that we ignore in our linear theory.
i ) = g(A∗
i , I∗
i , I∗
Extracellular fluid dynamics
Next, we write a mass conservation equation for the in-
compressible fluid contained in the tissue interstitial space
between cells:
∂tφ − ∇.((1 − φ)ve) = φh(Ai,Ii)−φ
τ
(2)
where ve is the velocity of the extracellular fluid. The
right-hand side of this equation describes the fact that
cells actively regulate their relative volume fraction to an
homeostatic value φh(Ai, Ii) at a timescale τ [21]. Note
that, (2) with ve (cid:54)= 0 implies a recirculation of internal
fluid, via gap junctions [22] (SI Appendix, Sec. 1.A.3).
i )/I∗
i +χI (Ii−I∗
i where we denote φ∗ = φh(A∗
As detailed below, we assume that local cellular mor-
phogen concentrations have an influence on the volume
fraction φ which couples tissue mechanics to local mor-
phogens concentration in our theory. At linear order,
this coupling generically reads φh(Ai, Ii) = φ∗ + χA(Ai −
A∗
i )/A∗
i , I∗
i ),
the equilibrium cell volume fraction, and the χA,I terms
account for the sensitivity of cell volume to intracellular
morphogen concentrations. Such a mechano-chemical ef-
fect on the tissue packing fraction, φ, can occur either via
the active control of individual cell volume [21] or through
the active balance between cell proliferation and loss (SI
Appendix, Sec. 1.A.4), with χA,I > 0 for morphogens
acting as growth factors and χA,I < 0 for morphogens
working as growth inhibitors. This is a reasonable as-
sumption, as a number of morphogens involved in cell fate
decisions can act as growth factor/inhibitors [23, 24], and
in vitro experiments have shown that cells, upon exposure
to factors such as FGF or EGF, elicit a series of signal-
ing mediated responses involving an increase in transmem-
brane ion flux, cell volume changes [21] and subsequent cell
growth/division [25]. Moreover, during digits pattern for-
mation in the limb bud, which has been proposed to rely
on a Turing instability, morphogens such as BMP par-
ticipate in both the reaction-diffusion scheme [8] and in
morphogenetic events such as cell condensation [26], with
skeletal formation being associated with large cell volume
fraction changes [27]. The cell volume fraction is thus
highly modulated in space and time, concomitantly with
morphogen pattern formation [26], advocating for the need
of a global mechano-chemical theory taking into account
both effects.
Extracellular morphogen dynamics
Morphogens, once secreted by cells, are transported by
diffusion and advection in the extracellular fluid:
∂t((1 − φ)Ae) + ∇. ((1 − φ)Aeve − D∇Ae) = −γAAe + λAAi
∂t((1 − φ)Ie) + ∇. ((1 − φ)Ieve − D∇Ie) = −γI Ie + λI Ii
(3)
2
TimeDfickDli2/Dfickli1/γA,IDKA,IMoleculardiffusionEffectivediffusion(activetransport)Hindered diffusion (tissue structure) lclDiffusion coefficientbaBiphasic tissuelExtracellular fluidCelllilcRVEAdherens junctionMorphogens importexport Gap junctionAquaporinsIon pumpsAI where D is the global Fickian diffusion coefficient of both
morphogens depending on tissue packing and tortuosity
[9, 28, 29]. As we are interested in a linear theory, we con-
sider here D = D(φ∗) as a constant. We neglect here,
for the sake of simplicity, phenomena such as extracel-
lular morphogen degradation or the influence of extra-
cellular morphogen concentrations on reaction terms, as
they do not modify qualitatively the dynamics (SI Ap-
pendix, Sec. 1.C). Note that one could also take into ac-
count, at the mesoscopic level, some effective non-local in-
teractions such as cell-cell communication via long-ranged
cellular protrusions [30]. This may require to consider spa-
tial terms in (1) to introduce an additional characteristic
lengthscale from non-local cell-cell transport.
Mechanical behaviour of the cellular phase
To complete our description, we need to specify a relation
linking cell volume fraction to interstitial fluid velocity.
For this, we use a poroelastic framework, whose appli-
cability to describe the mechanical response of biological
tissues has been thoroughly investigated in various con-
texts [31, 32]. Taking an homogeneous tissue as reference
state, poroelastic properties imply that a local change of
the cell volume fraction creates elastic stresses in the cellu-
lar phase which translate to gradients of extracellular fluid
pressure p. Such gradients of pressure in turn drive extra-
cellular fluid flows, which can advect morphogens, and we
show (SI Appendix, Sec. 1.A.7) that this effects results
in a simple Darcy's law between cell volume fraction and
fluid flow [29]:
(1 − φ)ve = − κ
η∇p = Dm∇φ.
(4)
√
This relation introduces the hydrodynamic diffusion coef-
ficient of the extracellular fluid, Dm = Kκ/η, a key me-
chanical parameter of the model which feeds back on the
reaction diffusion dynamics (3), with κ the tissue perme-
ability, K the elastic drained bulk modulus and η the fluid
viscosity. The hydrodynamic length scale lm =
Dmτ is
associated to such fluid movement. Importantly, we only
explore here the simplest tissue rheology for the sake of
simplicity and concision. Nevertheless, we also investigate
(SI Appendix, Sec. 1.H) the role of growth and plastic cell
rearrangements and show that they can be readily incorpo-
rated in our model, leading to different types of patterning
instabilities. However, we would like to highlight here that
the results presented thereafter are all robust to small to
intermediate levels of tissue rearrangements.
Model of an active biphasic tissue
Eqs.(1-4) define a full set of equations describing the
chemo-mechanical behaviour of an active biphasic mul-
ticellular tissue (SI Appendix, Sec. 1.B). To provide clear
insights on the biophysical behaviour of the system, we
focus on a limit case where γA,I (cid:29) λA,I (cid:29) f, g such that
KA,I (cid:28) 1. This corresponds to an ubiquitous biologi-
cal situation where rates of membrane transport are order
of magnitudes faster than transcriptionaly controled mor-
phogen turnover rates, and where endocytosis occurs at a
much faster rate than exocytosis. In that case, the rela-
tions Ae (cid:39) KAAi and Ie (cid:39) KI Ii always hold and even
if a significant fraction of morphogens is immobilized in-
side the cells [9], the import/export terms cannot be ne-
glected as γA,I are very large, so that γA(Ae − KAAi) and
γI (Ie − KI Ii) are indeterminate quantities (SI Appendix,
Sec. 1.C). Summing both internal (1) and external (3) con-
servation laws, we obtain a simplified description of the
system (SI Appendix, Sec. 1.C):
∂t(φAi) + ∇. (AiKADm∇φ − KAD∇Ai) = f (Ai, Ii)
∂t(φIi) + ∇. (IiKI Dm∇φ − KI D∇Ii) = g(Ai, Ii)
−l2
m∆φ + φ = φh(Ai, Ii).
√
(5)
Non-dimensionalizing times with τA associated with the
degradation of Ai in the morphogen turnover functions f
KADτA we find that (5) is
and g and lengths with lA =
controlled by a few non-dimensional parameters: KI /KA
describes the mismatch of morphogen membrane trans-
port, Dm/D compares the global hydrodynamic and Fick-
ian diffusion of the morphogens, τ /(KAτA) compares the
response time of cell volume fraction to the effective mor-
phogen turnover rate, and χA and χI account for the sen-
sitivity of φ to morphogen levels. Using this restricted set
of parameters encapsulating the behaviour of the model,
we investigate several of its biologically relevant limits,
demonstrating that they provide independent routes to-
wards tissue patterning.
Orders of magnitude on morphogen trans-
port
In the simplest limit of the model, the cell fraction re-
mains constant, φ = φ∗, which is valid if the effect of the
morphogens on φ is very small compared to the restor-
ing mechanical forces (i.e. χA,I = 0). The model then
reduces to Turing's original system, with diffusion coeffi-
cients being renormalised by morphogens transmembrane
transport equilibrium constants, KA,I D, similar to results
obtained in [9, 11]. This implies that even species with
similar D, can exhibit effective diffusion coefficients widely
differing from each other on longer timescales and produce
Turing patterns when KI (cid:29) KA (SI Appendix, Sec. 1.F).
In Fig. 1(b), we depict scaling arguments for the changes
in effective diffusion coefficient at various time/length
scales, associated both with tissue structure and im-
port/export kinetics [11]. At small timescales, diffusion
is characterised by a local Fickian diffusion coefficient,
theoretically expected to be of the order of DFick ≈
10−11m2s−1,
in line with fluorescence correlation spec-
troscopy (FCS) measurements [7,9,20]. This occurs across
a typical cell-to-cell distance of li ≈ 10−7 − 10−9m [33], so
i /DFick ≈
that this regime is valid for time scales below l2
10−2 − 10−6s, which is much faster than the typical im-
port/export kinetics of 1/γA,I ≈ 101 − 102s [34]. At in-
termediate timescales, the diffusion coefficient needs to be
corrected for volume exclusion effects due to the porous
nature of the tissue, an effect which can be very large for
cell volume fraction close to one [35]. An upper bound
(Hashin-Shtrikman) for global diffusion can be computed,
irrespective of the microscopic details of tissue geometry,
3
Figure 2: Linear stability analysis and numerical simulations of pattern formation in active biphasic tissues. (a)
Phase diagram of (5) in the (KI /KA, Dm/D) parameter space for τ /(KAτA) = 0.01 and τ /(KAτA) = 0.1 (inset). The red
and blue dashed lines correspond to analytical thresholds of instability (given in the text) for Turing and Keller-Segel patterns
respectively. The black dashed line is the analytical phase boundary between both regimes in the limit KI (cid:29) KA given by
χA = D/Dm + τ /(τAKA). This limit is shifted up when the ratio τ /τAKA is increased, while a pronounced notch appears in
the "Keller-Segel patterns" domain (see inset). Other parameters are set to χA = 0.25, χI = 0, τI /(KAτA) = 0.2, KAτAρ = 1,
φ∗ = 0.85 and large tissue size (lA/l (cid:28) 1). (b) 1D numerical simulations of (5) with random initial conditions for several choices
of parameters identified by letters A, B, C & D. lA/l = 0.1.
as D(φ∗) ≤ DFick(1 − φ∗)/(1 + φ∗/2) [28], which would
suggest, in the case of φ∗ ≈ 0.8 − 0.9, that it should be
around an order of magnitude smaller than local diffusion,
D(φ∗) ≈ 10−12m2s−1. Finally, at the time scales larger
than 1/γA,I described by the present model, the diffusion
is decreased further by a factor KA,I , i.e. by the relative
concentrations of morphogens "trapped" cellularly (i.e. a
1−10 ratio) such that D(φ∗)KA,I ≈ 10−12−10−13 m2s−1.
This is consistent with effective diffusion coefficients mea-
sured from tissue-wide fluorescence recovery after pho-
tobleaching (FRAP) over minutes to hours time scales
[7, 9, 20, 35]. Note here, that the respective contributions
of volume exclusion and import/export effects on FRAP
measured diffusion coefficients are non-trivial and are de-
tailed in SI Appendix, Sec. 1.H. Overall, although our
model in its simplest limit (φ = φ∗) relaxes the classical
Turing condition DI (cid:29) DA, it still implies quite stringent
conditions on the ratio of intracellular and extracellular
morphogens (Ie/Ii (cid:29) Ae/Ai). Exploring further the effect
of a variable cell volume fraction φ, we demonstrate that
coupling morphogen dynamics and tissue mechanics sup-
presses this limitation via active transport of morphogens.
bility threshold given by KI τI − KAτA > 2
τAτI KAKI
for lA/l (cid:28) 1(dashed red line on Fig. 2 (a)) which, as
expected, is always true regardless of the value of τA,I
if KI (cid:29) KA. However, another generic pattern form-
ing instability driven by active transport phenomena is
present in the phase diagram, labelled "Keller-Segel pat-
terns" [36]. The physical origin of the resulting pattern
is here similar to active fluid instabilities [15, 17, 37 -- 40]:
if stochastic local changes in morphogen concentration re-
sult in an increase in cell volume fraction, fluid must be
pumped inside cells. This causes local elastic deforma-
tions in the tissue which generate large-scale extracellular
fluid flows from regions of low to high morphogen con-
centration, resulting in a positive feedback loop of mor-
phogens enrichement (Fig. 3 (a)), and steady-state pat-
terns.
Interestingly, such an instability can even occur
√
for a single morphogen. In this limit, patterning occurs if
the volume fraction sensitivity χA is above a critical value
(dashed blue line in Fig. 2 (a), which captures well the
phase boundary in the limit KA (cid:29) KI , although the in-
stability occurs generically for any value of KA,I ). The
number of patterns displayed by the profiles shown on
Fig. 2 (b) can be predicted by linear analysis (See Ap-
pendix, Sec. 1.D) because they are chosen close to the
onset of instability.
χA > (cid:112)D/Dm +(cid:112)τ /(τAKA) when lA/l (cid:28) 1 so that
√
Turing-Keller-Segel instabilities
To assess the regions in parameter space where stable
patterns can form in our mechano-chemical framework,
we perform a linear stability analysis on (5). Here, we
consider a classical Gierer-Meinhardt activator-inhibitor
scheme [2]: f (A, I) = ρA2/I − A/τA and g(A, I) =
ρA2 − I/τI , where ρ is the rate of activation and inhibi-
tion and τA,I the timescales of degradation of A and I [2]
and the particular case of a single morphogen capable of
increasing φh (χA > 0, χI = 0).
In the phase diagram in Fig. 2 (a), we show that two
distincts instabilities can be captured by this simplified
theory. The first instability, identified here as "Turing pat-
terns", corresponds to a classical Turing instability, where
diffusive transport of morphogens dominates over their ad-
vection by interstitial fluid (Dm (cid:28) D) and with insta-
Thus, coupling tissue mechanical behaviour to mor-
phogen reaction-diffusion provides, via the generation of
advective fluid flows, a new route to stable pattern forma-
tion with a single morphogen. Moreover, this instability
has two remarkable features. First, it only requires the
presence of a single morphogen (SI Appendix, Sec. 1.G)
which could correspond to many practical situations where
a pair of activator/inhibitor has not been clearly identified,
for instance the role of Wnt in the antero-posterior pat-
tern of planarians [41]. Second, it possesses spatial scal-
ing properties regarding to its fundamental mode, as com-
pared to a Turing instability.
Indeed, when morphogen
turnover rate is small compared to its effective hydrody-
4
like sensitivities χA,I would need to be better assessed in
vivo in future works.
Cross-diffusion Turing instabilities
Finally, we investigate the behaviour of our model ((5)),
when cell fraction sensitivity to morphogen concentration
is negative (χA,I < 0), eliminating the possibility of up-
hill morphogen diffusion at the origin of the Keller-Segel
instability. We also consider that f and g do not necessar-
ily follow an activator-inhibitor kinetics, but any possible
interaction scheme between two morphogens. For math-
ematical clarity on the physical nature of the instability
studied here, we make the simplifying assumptions that
τ = 0 and χA,I (cid:28) 1, with D ∼ DmχA,I in (5). This re-
lates to a realistic biological situation, where cell volume
fraction relaxes rapidly after perturbation and depends
weakly on morphogen levels, yielding:
φ∗∂tAi + ∇. (AiKADm∇φh − KAD∇Ai) = f (Ai, Ii)
φ∗∂tIi + ∇. (IiKI Dm∇φh − KI D∇Ii) = g(Ai, Ii).
(6)
In this limit, the conditions for linear stability of the
homogeneous solution are exactly the ones of a classi-
cal Turing system but with cross-diffusion terms (SI Ap-
pendix, Sec. 1.E). Such a scenario has been studied in
the framework of monophasic reaction-diffusion systems
with ad hoc cross-diffusion terms [43], which arise generi-
cally in various chemical and biological systems [44]. Our
work thus provides a particular biophysical interpretation
of these terms in multicellular tissues, which we show to
originate from intrinsically mechano-chemical feedbacks
between morphogen dynamics and tissue mechanics.
Figure 4: Pattern formation for cross-diffusion Turing
instabilities. (a) Phase diagram of (5) in the (χA, χI ) space
obtained by numerical linear stability analysis. Parameters are
τ /(KAτA) = 0.01, Dm/D = 10, KI /KA = 10, τI /(KAτA) =
0.9, φ∗ = 0.85 and lA/l (cid:28) 1.
(b) 1D numerical simulation
of (5) using a simple inhibitor-inhibitor reaction scheme (SI
Appendix, Sec. 1.B).
As shown in [43], such cross diffusion terms result in
a dramatic broadening of the phase space for patterns.
In particular, any two-morphogen reaction scheme can
now generate spatial patterns and not just the classi-
cal activator-inhibitor schemes. For instance, it becomes
possible to obtain patterns with activator-activator or
inhibitor-inhibitor kinetics similar to those observed in
numerous gene regulatory networks or signaling path-
ways involved in cell fate decisions [45]. We illustrate
this result by considering an inhibitor-inhibitor kinetic
scheme, which cannot yield patterns in the classical Tur-
ing framework, wandemonstrate analytically and numer-
ically the existence of a region of stable patterns (from
5
Figure 3: Scaling properties of the Keller-Segel instabil-
ity with one morphogen.(a) Schematic of the Keller-Segel
instability in a 1D tissue. Morphogens gradients generate cell
volume fraction gradients (via local fluid exchanges, blue ar-
rows in inset), which in return cause mechanically-induced self-
amplifying extracellular flows that advect morphogens from
morphogen-poor to morphogen-rich regions (green arrow). (b)
Normalized pattern size as a function of system size in the
single morphogen case with f = 0.
(c) Morphogen concen-
tration and cell packing fraction (inset) profiles remain quasi-
stationary as system size increases. Parameters are χA = 0.25,
Dm/D = 10 and φ∗ = 0.85
namic and Fickian diffusion (f → 0), the fundamental
mode, i.e. a single two-zones pattern, is the most unsta-
ble in a robust manner, given that morphogen turnover f
stabilises specifically this mode (SI Appendix, Sec. 1.G.2),
whereas in the case of a Turing instability, this would re-
quire fine-tuning and marginally stable reaction kinetics.
We illustrate such a scaling property in Fig. 3. This mech-
anism could potentially apply to situations where a bi-
nary spatial pattern is independent of system size such as
dorso-ventral or left-right patterns in early vertebrate em-
bryos [7, 9], or planarian antero-posterior pattern [41, 42].
If so, it could provide a simpler alternative to previously
proposed mechanisms involving additional species or com-
plex biochemical signaling pathways [7, 42].
Importantly, simple estimates can be used to demon-
strate the biological plausibility of such mechanical effects
during morphogenetic patterning. A key parameter driv-
ing Keller-Segel instabilities is the hydrodynamic diffusion
coefficient Dm, which can be estimated from values of the
drained bulk modulus K ≈ 104 Pa [31] and the tissue
permeability upper bound [28] κ ≈ l2
i (1 − φ∗)/(1 + φ∗/2)
with li ≈ 10−7 − 10−9m and φ∗ ≈ 0.85 as above. Us-
ing η ≈ 10−3 Pa.s (water viscosity), we obtain Dm ≈
10−12 − 10−8 m2s−1, showing that the hydrodynamic dif-
fusion can be similar or even much larger than Fickian
diffusion.
In agreement with typical timescales involved
in regulatory volume increase or decrease of cells follow-
ing an osmotic perturbation [21], we estimate that τ ≈ 102
s, while morphogen turnover time scale has been measured
as τA ≈ 104 − 105 s [9]. With KA ≈ 0.1 as above, we ob-
tain τ /(KAτA) ≈ 0.01 − 0.1, which is used in Fig. 2, and
displays broad regions of instability, although parameters
abVolume fractionMorphogens00120.51010.5 (5)), where a cross-diffusion driven Turing instability can
develop (Fig. 4).
Discussion
In this paper, we have introduced a generalisation of
Turing's work on pattern formation in biological tissues
by coupling equations describing the structure and me-
chanical properties of multicellular tissues with a classical
reaction-diffusion scheme.
In particular, our work high-
lights two important features of multicellular tissues, as of
yet largely unexplored in this context: their biphasic na-
ture, i.e. the fact that morphogen production/degradation
is controlled by cells while transport takes place extracel-
lularly requiring active membrane exchanges (effectively
rescaling diffusion [9, 11]), and the possibility for active
large scale flows to develop within the tissue interstitial
space. We demonstrate that coupling tissue cell volume
fraction to local morphogen levels (based on the dual role
of morphogens in patterning and cell growth/volume reg-
ulation [23,24]) provides a biophysically realistic route to-
wards two qualitatively different modes of patterning in-
stability. Extracellular fluid flows can have two important
consequences on patterning. Firstly, as the Turing insta-
bility is rooted in the cross-effects between a stable chem-
ical reaction of two morphogens and their diffusion, the
conditions of such instability are deeply affected by active
hydrodynamic transport which can create cross terms into
the effective diffusion matrix. This causes a drastic widen-
ing of the phase space of Turing patterning, rendering it
robust and only weakly dependent on morphogens reac-
tion scheme. Secondly, extracellular fluid flows can also
create an instability of a different nature (Keller-Segel),
when these flows have an anti-diffusive structure, spon-
taneously creating morphogens gradients. Here, chemical
reactions between morphogens are only setting the num-
ber of patterns, and if such reactions are sufficiently slow,
the spatial pattern of morphogen always coarsens to the
fundamental mode of instability, and has robust scaling
properties compared to conventional Turing models. This
could have interesting implications concerning recent ex-
perimental evidences for robust scaling of the Nodal/Lefty
pattern in the early zebrafish embryo [46].
In this respect, our approach, which has the advantage
of parsimony, taking into account the manifest biphasic
nature of multicellular tissues, is complementary to oth-
ers which have been proposed to solve limitations of Tur-
ing's model by introducing additional morphogen regu-
lators [42, 47], and also displays connections with recent
development in the mechano-chemical descriptions of ac-
tive fluids such as the cell cytoskeleton [15, 16]. Never-
theless, although our hypothesis of cell volume fraction
gradients driving large-scale flows is generic to biphasic
tissues, further quantitative experiments would be needed
to test the relationship between morphogen concentra-
tion and cell volume fraction, as well as probe the role
of transmembrane import/export kinetics or similar phe-
nomena such as transmembrane signaling [11], morphogen
adsorption/desorption on cell surface [9] and long-distance
cellular protrusions [30], on effective morphogen diffusion
Interestingly,
rates. Systems such as digits patterning, where cell vol-
ume fraction spatial pattern appears concomitant to mor-
phogen patterns [26], or planarian antero-posterior pat-
terning, where pairs of activator/inhibitor have not been
clearly identified [41], provide possible testing grounds for
our model.
large-scale extracellular fluid
flows have been increasingly observed during embryo de-
velopment, not only in the classical case of cilia driven
flows [48], but also due to mechanical forces arising from
cellular contractions as well as osmotic and poro-viscous
effects [49,50], calling for a more systematic understanding
of passive vs. active transport mechanisms during embry-
onic pattern formation. Whether biological examples of
Turing patterning instabilities, such as left-right or dorso-
ventral patterning, digits pattern formation or skin ap-
pendages patterns are causally associated with concomi-
tant changes in cell volume and/or cell packing remains a
result to be experimentally investigated.
Methods
Linear stability analysis was performed numerically using
Mathematica, while numerical integrations of the model
equations were performed using a custom-made Matlab
code.
Acknowledgments
A.H acknowledges the University of Cambridge for a David
Crighton Fellowship and the Wellcome Trust for an In-
terdisciplinary Research Fellowship. P.R. acknowledges a
CNRS-Momentum grant and E.H from the Austrian Sci-
ence Fund (FWF) [P 31639]. The authors warmly thank
Benjamin Simons, Anna Kicheva, David Jorg, Tom His-
cock, Pau Formosa-Jordan, Erik Clark and Lev Truski-
novsky for useful discussions.
References
[1] Turing AM (1952) The chemical basis of morphogenesis.
Philosophical Transactions of the Royal Society B: Biolog-
ical Sciences 237(641):37-72.
[2] Gierer A and Meinhardt H (1972) A theory of biological
pattern formation. Kybernetik 12(1):30-39.
[3] Murray JD, Mathematical Biology (Springer-Verlag,
Berlin, 2003), third edition.
[4] Kondo S and Miura T (2010) Reaction-diffusion model as a
framework for understanding biological pattern formation.
Science 329(5999):1616-1620.
[5] Sick S, Reinker S, Timmer J and Schlake T (2006)
WNT and DKK determine hair follicle spacing through
a reaction-diffusion mechanism. Science 314(5804):1447-
1450.
[6] Economou AD et al. (2012) Periodic stripe formation by a
Turing mechanism operating at growth zones in the mam-
malian palate. Nature Genetics 44(3):348-351.
6
[7] Inomata H, Shibata T, Haraguchi T and Sasai Y (2013)
Scaling of dorsal-ventral patterning by embryo size-
dependent degradation of Spemann's organizer signals. Cell
153(6):1296-311.
[8] Raspopovic J, Marcon L, Russo L and Sharpe J (2014)
Modeling digits. Digit patterning is controlled by a Bmp-
Sox9-Wnt Turing network modulated by morphogen gra-
dients. Science 345(6196):566-570.
[9] Muller P et al. (2012) Differential diffusivity of Nodal and
Lefty underlies a reaction-diffusion patterning system. Sci-
ence 336(6082):721-724.
[10] Lengyel I and Epstein IR (1991) Modeling of Turing struc-
tures in the chlorite -- iodide -- malonic acid -- starch reaction
system. Science 251(4994): 650-2
[11] Rauch EM and Millonas MM (2004) The role of trans-
membrane signal transduction in turing-type cellular pat-
tern formation. Journal of Theoretical Biology 226(4):401-
407.
[12] Castets V, Dulos E, Boissonade J and De Kepper P (1990)
Experimental evidence of a sustained standing Turing-type
nonequilibrium chemical pattern. Physical Review Letters
64(24):2953-2956.
[13] Rovinsky AB and Menzinger M (1993) Chemical instabil-
ity induced by a differential flow. Physical Review Letters
69(8):1193-1196.
[14] Oster GF, Murray JD and Harris AK (1983) Mechani-
cal aspects of mesenchymal morphogenesis. Development
78:83-125.
[15] Bois JS, Julicher F and Grill SW (2011) Pattern formation
in active fluids. Physical Review Letters 106:028103.
[16] Howard J, Grill SW and Bois JS (2011) Turing's next
steps: the mechanochemical basis of morphogenesis. Na-
ture Review Molecular and Cellular Biology 12(6):392-398.
[17] Weber CA, Rycroft CH and Mahadevan L (2011) Differ-
ential activity-driven instabilities in biphasic active matter.
Physical Review Letters 120:248003.
[18] Hiscock TW and Megason SG (2015). Mathematically
guided approaches to distinguish models of periodic pat-
terning. Development 142:409-419.
[19] Bokel C and Brand M (2014) Endocytosis and signaling
during development. Cold Spring Harbor Perspectives in
Biology 6:a017020.
[20] Kicheva A, Bollenbach T, Wartlick O, Julicher F and
Gonzalez-Gaitan M (2012) Investigating the principles of
morphogen gradient formation: from tissues to cells. Cur-
rent Opinion in Genetics & Development 22(6):527-532.
[21] Hoffmann EK, Lambert IH and Pedersen SF (2009) Physi-
ology of cell volume regulation in vertebrates. Physiological
Reviews 89(1):193-277.
[22] Zehnder SM, Suaris M, Bellaire MM and Angelini TE
(2015) Cell Volume Fluctuations in MDCK Monolayers.
Biophysical Journal 108(2):247250.
[23] Smith JC (1981) Growth factors and pattern forma-
tion, Journal of Embryology and Experimental Morphology
65:187-207.
7
[24] Ginzberg MB, Kafri R and Kirschner M (2015) On being
the right (cell) size. Science 348(6236):1245075.
[25] Zetterberg A, Engstrom W and Dafgird E (1984) The rel-
ative effects of different types of growth factors on DNA
replication, mitosis, and cellular enlargement. Cytometry
5(4):368-375.
[26] B´enazet JD et al (2012) Smad4 is required to induce digit
ray primordia and to initiate the aggregation and differ-
entiation of chondrogenic progenitors in mouse limb buds.
Development 139(22):4250-4260.
[27] Cooper KL et al (2013) Multiple phases of chondrocyte en-
largement underlie differences in skeletal proportions. Na-
ture 495(7441):375.
[28] Crank J, The Mathematics Of Diffusion (Oxford Univer-
sity Press, Oxford, 1979), first edition.
[29] Bear J, Dynamics of Fluids in Porous Media (Dover, New-
York, 1989), new edition.
[30] Kondo S (2017) An updated kernel-based Turing model for
studying the mechanisms of biological pattern formation.
Journal of Theoretical Biology 414:120-127.
[31] Netti PA, Berk DA, Swartz MA, Grodzinsky AJ and Jain
RK (2000) Role of extracellular matrix assembly in intersti-
tial transport in solid tumors. Cancer Research 60(9):2497-
2503.
[32] Fraldi M and Carotenuto AR (2018) Cells competition in
tumor growth poroelasticity. Journal of the Mechanics and
Physics of Solids 112:345-367.
[33] Barua D, Parent SE and Winklbauer R (2017) Mechan-
ics of Fluid-Filled Interstitial Gaps. II. Gap Characteris-
tics in Xenopus Embryonic Ectoderm. Biophysical journal
113(4):923-936.
[34] Smith CB and Betz WJ (1996) Simultaneous indepen-
dent measurement of endocytosis and exocytosis. Nature
380(6574): 531-534.
[35] Blassle A et al (2018) Quantitative diffusion measurements
using the open-source software PyFRAP. Nature Commu-
nications 9(1):1582.
[36] Keller EF and Segel LA (1970) Initiation of slime mold
aggregation viewed as an instability. Journal of theoretical
biology 26(3):399-415.
[37] Recho P, Putelat T and Truskinovsky L (2015) Mechanics
of motility initiation and motility arrest in crawling cells.
Journal of the Mechanics and Physics of Solids84:469-505.
[38] Recho P, Putelat T and Truskinovsky L(2013)
Contraction-driven cell motility. Physical Review Letters
111(10):108102.
[39] Hannezo E et al (2015) Cortical instability drives pe-
riodic supracellular actin pattern formation in epithelial
tubes. Proceedings of the National Academy of Sciences
112(28):8620-8625.
[40] Aguilar-Hidalgo D et al
in
Self-Organized Tissue Growth. Physical Review Letters
120(19):198102.
(2018) Critical Point
[41] Stuckemann T et al (2017) Antagonistic self-organizing
patterning systems control maintenance and regeneration
of the anteroposterior axis in planarians. Developmental
Cell 40(3):248-263.
[42] Werner S et al (2015) Scaling and regeneration of self-
organized patterns. Physical Review Letters 114(13):
138101.
[43] Madzvamuse A, Ndakwo HS and Barreira R (2015) Cross-
diffusion-driven instability for reaction-diffusion systems.
Journal of Mathematical Biology70(4):709-743.
[44] Vanag VK and Epstein IR (2009) Cross-diffusion and
pattern formation in reaction-diffusion systems. Physical
Chemistry Chemical Physics 11:897-912.
[45] Zhou JX and Huang S (2011) Understanding gene circuits
at cell-fate branch points for rational cell reprogramming.
Trends in Genetics 27(2):55-62.
[46] Almuedo-Castillo et al( 2018) Scale-invariant patterning
by size-dependent inhibition of Nodal signalling. Nature
cell biology, 20(9), 1032.
[47] Diego X, Marcon L, Muller P and Sharpe J (2018) Key
Features of Turing Systems Are Determined Purely by Net-
work Topology. Physical Review X 8(2):21071.
[48] Freund JB, Goetz JG, Hill KL and Vermot J (2012) Fluid
functions, features and
flows and forces in development:
biophysical principles. Development 139(7):1229-1245.
[49] Krens SG et al (2017) Interstitial fluid osmolarity modu-
lates the action of differential tissue surface tension in pro-
genitor cell segregation during gastrulation. Development
144(10):1798-1806.
[50] Ruiz-Herrero T, Alessandri K, Gurchenkov BV, Nassoy P
and Mahadevan L (2017) Organ size control via hydrauli-
cally gated oscillations. Development 144(23):4422-4427.
8
|
1911.09499 | 1 | 1911 | 2019-11-21T14:44:00 | Collective olfactory search in a turbulent environment | [
"physics.bio-ph",
"cond-mat.stat-mech"
] | Finding the distant source of an odor dispersed by a turbulent flow is a vital task for many organisms, either for foraging or for mating purposes. At the level of individual search, animals like moths have developed effective strategies to solve this very difficult navigation problem based on the noisy detection of odor concentration and wind velocity alone.
When many individuals concurrently perform the same olfactory search task, without any centralized control, sharing information about the decisions made by the members of the group can potentially increase the performance. But how much of this information is actually valuable and exploitable for the collective task ? Here we show that, in a model of a swarm of agents inspired by moth behavior, there is an optimal way to blend the private information about odor and wind detections with the publicly available information about other agents' heading direction. At optimality, the time required for the first agent to reach the source is essentially the shortest flight time from the departure point to the target. Conversely, agents who discard public information are several fold slower and groups that do not put enough weight on private information perform even worse. Our results then suggest an efficient multi-agent olfactory search algorithm that could prove useful in robotics, for instance in the identification of sources of harmful volatile compounds. | physics.bio-ph | physics |
Collective olfactory search in a turbulent environment
Mihir Durve∗,1, 2 Lorenzo Piro∗,3, 4 Massimo Cencini,5 Luca Biferale,3 and Antonio Celani6
2Quantitative Life Sciences, The Abdus Salam International Centre for Theoretical Physics - ICTP, 34151,Trieste, Italy
1Department of Physics, Universit`a degli studi di Trieste, Trieste, 34127 Italy
3Department of Physics and INFN, University of Rome Tor Vergata,
Via della Ricerca Scientifica 1, 00133, Rome, Italy
4Max Planck Institute for Dynamics and Self-Organization, Am Fassberg 17, 37077 Goettingen, Germany
5Istituto dei Sistemi Complessi, CNR, via dei Taurini 19,
00185 Rome, Italy and INFN, sez. Roma Tor Vergata
6Quantitative Life Sciences, The Abdus Salam International Centre for Theoretical Physics - ICTP, Trieste, 34151, Italy
(Dated: November 22, 2019)
Finding the distant source of an odor dispersed by a turbulent flow is a vital task for many
organisms, either for foraging or for mating purposes. At the level of individual search, animals like
moths have developed effective strategies to solve this very difficult navigation problem based on the
noisy detection of odor concentration and wind velocity alone. When many individuals concurrently
perform the same olfactory search task, without any centralized control, sharing information about
the decisions made by the members of the group can potentially increase the performance. But how
much of this information is actually valuable and exploitable for the collective task ? Here we show
that, in a model of a swarm of agents inspired by moth behavior, there is an optimal way to blend
the private information about odor and wind detections with the publicly available information
about other agents' heading direction. At optimality, the time required for the first agent to reach
the source is essentially the shortest flight time from the departure point to the target. Conversely,
agents who discard public information are several fold slower and groups that do not put enough
weight on private information perform even worse. Our results then suggest an efficient multi-agent
olfactory search algorithm that could prove useful in robotics, for instance in the identification of
sources of harmful volatile compounds.
Animals are often on the move to search for some-
thing: a food source, a potential mate or a desirable site
for laying their eggs. In many instances their navigation
is informed by airborne chemical cues. One of the best
known, and most impressive, olfactory search behavior
is displayed by male moths [1 -- 4]. Males are attracted
by the scent of pheromones emitted in minute amounts
by calling females that might be at hundreds of meters
away. The difficulty of olfactory search can be appreci-
ated by realizing that, due to air turbulence, the odor
plume downwind of the source breaks down into small,
sparse patches interspersed by clean air or other extra-
neous environmental odors [5, 6]. The absence of a well-
defined gradient in odor concentration at any given loca-
tion and time greatly limits the efficiency of conventional
search strategies like gradient climbing. Experimental
studies have in fact shown that moths display a differ-
ent search strategy composed of two phases: surging, i.e.
sustained upwind flight, and casting, i.e. extended alter-
nating crosswind motion. These phases occur depending
on whether the pheromone signal is detected or not. This
strategy and others have inspired the design of robotic
systems for the identification of sources of gas leaks or
other harmful volatile compounds [7 -- 11]. Albeit the ef-
fectiveness of individual search is already remarkable in
itself, the performance can be further boosted by cooper-
ation among individuals, even in absence of a centralized
control [12 -- 17].
In this Letter, we tackle the problem of collective olfac-
tory search in a turbulent environment. When the search
takes place in a group, there are two classes of informa-
tive cues available to the agents. First, there is private
information, such as the detection of external signals --
odor, wind velocity, etc -- by an individual. This percep-
tion takes place at short distances and is not shared with
other members of the group. Second, there is public in-
formation, in the form of the decisions made by other in-
dividuals. These are accessible to (a subset of) the other
peers, usually relayed by visual cues, and therefore with
a longer transmission range. Since the action taken by
another individual may be also informed by its own pri-
vate perception of external inputs, public cues indirectly
convey information about the odor distribution and the
wind direction at a distance. However, the spatial and
temporal filtering that is induced by the sharing of public
cues may in principle destroy the relevant, hidden infor-
mation about the external guiding signals.
These considerations naturally lead to the question if
the public information is exploitable at all for the collec-
tive search process. And if it is, how should the agents
combine the information from private and public cues to
improve the search performances ? Below, we will ad-
dress these questions by making use of a combination of
models for individual olfactory search and for flocking
behavior, in a turbulent flow.
A model for collective olfactory search. The setup for
our model is illustrated in Fig 1A. Initially, N agents are
randomly and uniformly placed within a circle of radius
Rb at a distance Lx from the source S. The odor source
S emits odor particles at a fixed rate of J particles per
unit time. The odor particles are transported in the sur-
rounding environment by a turbulent flow u with mean
wind U (details are given below). Notice that the odor
particles are not to be understood as actual molecules,
but rather represent patches of odor with a concentration
above the detection threshold of the agents. The entire
system is placed inside a larger square box of size bLx
with reflecting boundary conditions for the agents. A
complete list of parameters with their numerical values
is given in the Supplementary Material.
FIG. 1. Collective olfactory search. (A) Odor particles dis-
persed by the turbulent environment are shown by semi-
transparent blue dots emitted by the source S. Agents (red)
are initially placed far from the source in a packed configura-
tion. (B) Perception of an agent (red). Detected odor parti-
cles by the agent are shown as darker blue dots and neighbors
of the agent are shown in green. Arrows indicate the instan-
taneous moving direction of agents. We set Lx = 250Rd,
Rb = 25Rd, Ra = 5Rd, Rd = 0.2 b = 2.5 (C) Trajectory of
an isolated agent performing the cast-and-surge program (see
text). The locations where the agent detects the presence of
odor particles are shown as blue crosses.
Response to private cues. The behavior elicited by
private cues such as odor and wind speed is in-
spired by the cast-and-surge strategy observed in moths.
We adopted a modified version of the "active search
model" [18] that works as follows.
0
agent along its trajectory: u(t) = λR t
We assume that the agents have access to an estimate
of the mean velocity of the wind as moths actually do via
a mechanism named optomotor anemotaxis [19]. In the
model this estimate u(t) is an exponentially discounted
running average of the flow velocity u perceived by the
u(s) exp[−λ(t −
s)]ds. The parameter λ is the inverse of the memory time:
for λ → 0 the estimate converges to the mean wind, while
for λ → ∞ it reduces to the instantaneous wind velocity
at the current location of the agent. In the following we
have taken λ = 1 which is of the same order of magnitude
of the inverse correlation time of the flow.
It is worth
pointing out that the only effect of the wind is to provide
contextual information about the location of the source.
Indeed, in our model the agents are not carried away
by the flow, an assumption that is compatible with the
fact that the typical airspeed of moths and birds largely
exceeds the wind velocity.
2
At each time interval ∆t, the agent checks if there are
odor particles within its olfactory range Rd (see Fig. 1B).
If this is the case, then it moves against the direction of
the current estimated mean wind at a prescribed speed
v0. When the agent loses contact with the odor cue, it
starts the "casting" behavioral program: it proceeds by
moving in a zig-zag fashion, always transversally to the
current estimated mean wind, with turning times that
increase linearly with the time from the last odor de-
tection (a sample trajectory is shown in Fig 1C, see the
Supplementary Material for details about the implemen-
tation). We denote by vpriv
(t) the instantaneous veloc-
ity of agent i prescribed by this cast-and-surge program.
This is uniquely based on private cues and would indeed
be the actual velocity adopted by the agent if it were
acting in isolation.
i
Response to public cues. To describe the interactions
among agents we have drawn inspiration from flocking
and adopted the Vicsek model to describe the tendency
of agents to align with their neighbors (see [20, 21] and
references therein). We assume that an individual can
perceive the presence of its peers within a visual range Ra
(see Fig. 1B) and actually measure their mean velocity.
According to this model, the behavioral response elicited
in agent i by its neighbors is
vpub
i
(t) = v0 Xj∈Di
vj(t),(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj∈Di
vj(t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
,
(1)
where Di is the disk of radius Ra centered around the
position of the i−th individual. In order to account for
errors in the sensing of the velocities of the neighbors
we have added, as is customarily done, a noise term in
the form of a rotation by a random angle vpub
(t) ←
R(θ)vpub
(t). Here θ is independently sampled for each
agent and at each decision time from a uniform distri-
bution in [−ηπ, ηπ]. The strength of the noise η may
range from zero (no noise) to unity (only noise): in the
following we set η = 0.1.
i
i
In the absence of external cues, and for small enough
noise, the group of agents described by this dynamics dis-
plays collective flocking and moves coherently in a given
direction -- totally unrelated with the source location,
however.
Combining private and public information. To study
collective olfactory search we then merged the two models
above as follows. The velocity of the i−th agent is a linear
combination of the two prescriptions arising from private
and public cues, resulting in the update rule
vi(t) = (1 − β)vpriv
i
(t) + βvpub
ri(t + ∆t) = ri(t) + v0
(t),
∆t.
(2)
i
vi(t)
vi(t)
The parameter β, that we have dubbed "trust", mea-
sures the balance between private and public informa-
tion. For β = 0 the agents have no confidence in their
peers, they ignore the suggestion to align and behave in-
dependently by acting on the basis of the cast-and-surge
program only. Conversely, for β = 1 agents entirely fol-
low the public cues and discard the private information.
While it is reasonable to expect that for β = 1 the
unchecked trust in public cues leads to poor performances
in olfactory search, the nontrivial question here is rather
if there is any value at all in public information; that is,
in other words, if the best results are obtained for a finite
β strictly larger than zero.
Modeling the turbulent environment. To complete the
description of our model, we have to specify the under-
lying flow and the ensuing transport of odor particles.
In our simulations the flow is given by an incompress-
ible, two-dimensional velocity field, u(x, t) with a con-
stant, uniform mean wind U and statistically stationary,
homogeneous and isotropic velocity fluctuations. The
odor particles are considered as tracers whose position, x,
evolves according to x = u(x, t). For the velocity fluc-
tuations we first considered a stochastic flow and then
moved to a more realistic dynamics where the flow obeys
the Navier-Stokes equations.
Results for the stochastic flow This model flow is
characterized by a single length and time scale and is
obtained by superimposing a few Fourier modes whose
Gaussian amplitudes evolve according to an Ornstein-
Uhlenbeck process with a specified correlation time. The
resulting flow is spatially smooth, exponentially corre-
lated in time and approximately isotropic (see Supple-
mentary Material for details).
We studied the performance of collective search as a
function of the trust parameter β while keeping the other
parameters fixed to the values detailed in the Supplemen-
tary Material. Initially, the agents are waiting in place
without any prescribed heading direction until one of the
agents detects the odor particles carried by the flow. Af-
ter this event, agents move as per the equations of motion
Eq. (2). Since the search task is a stochastic process, we
run many episodes for each value of β to compute the
average values of several observables of interest. A given
episode is terminated when at least one of the agents is
within a distance Ra from the source. At this stage we
say that the search task is accomplished and agents have
(collectively) found the odor source.
We focused our attention on four key observables: (i)
the mean time needed to complete the task which mea-
sures the effectiveness of the search; (ii) the average frac-
tion of agents that, at the time of completion, are close to
the source; (iii) the order parameter which measures the
consensus among members of the group about their head-
ing direction; (iv) the degree of alignment of the agents
against the mean wind.
In Fig. 2A we show the average time T for the search
completion in units of the shortest path time Ts = Lx/v0,
which corresponds to a straight trajectory joining the
target with the center of mass of the flock at the initial
time. We observe that there exists an optimal value of
the trust parameter β ≈ 0.85 for which agents find the
3
FIG. 2. Collective olfactory search in a stochastic flow. (A)
Average search time T for the first agent that reaches the
target normalized to the straight-path time, Ts = Lx/v0. The
inset shows a blow-up of region close to the minimum. (B)
Fraction of agents within a region of size Rb around the source
at the time of arrival of the first agent reaching the target. (C)
Averaged order parameter ψ (D) Average alignment against
the mean wind M . For all data, the error bars denote the
upper and lower standard deviation with respect to the mean
value. Statistics is over 103 episodes. The parameters were
set as λ = 1, N = 100, J = 1, η = 0.1, v0 = 0.5, ∆t = 1,
Lx = 50.
odor source in the quickest way. Remarkably, for this
value we obtain T ≃ 1.03 Ts: this means that the agent
which arrives first is actually behaving almost as if it had
perfect information about the location of the source and
were able to move along the shortest path (see movie
Beta=0.85.mp4).
This result has to be contrasted with the singular case
of independent agents who act only on the basis of pri-
vate cues (β = 0) which display a significantly worse
performance (the time to complete the task is more than
threefold longer) and move in a zig-zagging fashion (see
movie Beta=0.00.mp4). It is also important to remark
that the average time grows very rapidly as β increases
above the optimum. As β approaches unity, agents are
dominated by the interactions with their neighbors and
pay little attention to odor and wind cues. As a result,
they form a flock which moves coherently in a direction
that is essentially taken at random.
If by chance this
direction is aligned against the wind, the task will be
completed in a short time. However, in most instances
the flock will miss the target and either turn because of
the noise η or bounce on the boundaries until, again by
sheer chance, some agent will hit the target (see movie
Beta=0.95.mp4). This behavior results in a very long
average time accompanied by very large fluctuations. As
the outer reflecting boundaries are moved away by in-
creasing b, this effect becomes more and more prominent.
Since we focused on the time of arrival for the first
(A)
0 0.2 0.4 0.6 0.8 1
(C)
(B)
4
0.4 0.5 0.6 0.7 0.8 0.9
t / Ts=0.00
t / Ts=0.00
t / Ts=0.00
t / Ts=0.00
t / Ts=0.00
t / Ts=0.40
t / Ts=0.40
t / Ts=0.40
t / Ts=0.40
s
T
T
/
1.8
1.7
1.6
1.5
1.4
1.3
1.2
1.1
1.0
0.9
(D)
M
t
n
e
m
n
g
i
l
a
i
d
n
w
p
U
0.4
0.2
0.0
-0.2
-0.4
0 0.2 0.4 0.6 0.8 1
Trust parameter β
0 0.2 0.4 0.6 0.8 1
Trust parameter β
t / Ts=0.80
t / Ts=0.80
t / Ts=0.80
t / Ts=0.80
t / Ts=1.04
t / Ts=1.04
t / Ts=1.04
t / Ts=1.04
12
10
8
6
4
2
s
T
T
/
Ψ
r
e
t
e
m
a
r
a
P
r
e
d
r
O
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
FIG. 3. Collective olfactory search in a turbulent flow. A: Search time T for the first agent reaching the target normalized by
the shortest-path time Ts. B: An enlargement of A that highlights the region close to the minimum. C: Mutual alignment order
parameter ψ averaged over time and episodes. D: Average wind alignment M . Right panels: snapshots of the velocity field
(grey arrows) at four different times t. The agents (red arrows) navigate in the turbulent flow with the optimal trust parameter
β = 0.8. Blue dots represent odor particles dispersed by the flow, while the large blue circle corresponds to the source.
agent that reaches the source, it is natural to ask what
has happened to the other members of the group that
have been trailing behind. In Fig. 2B we show the aver-
age fraction of agents that are within a distance Rb (the
initial size of the group) when the first agent reaches the
target and the task is completed. This quantity is an
indicator of the coherence of the group at the time of
arrival. It turns out that this fraction has a maximum
value ≈ 0.3 at about the same value of β ≈ 0.85 that
gives the best performance in terms of time. This means
that on average about one third of the group has been
moving coherently along the straight path that connects
the initial center of mass of the flock to the target.
To quantify the consensus among agents about which
direction they have to take, it is customary to introduce
the order parameter
ψ(t) =
.
(3)
1
N v0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
N
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xi=1
vi(t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
When all the agents move in the same direction,
whichever it may be, then ψ = 1. Conversely, if the
agents are randomly oriented then ψ ≃ N −1/2 ≪ 1. In
Fig. 2C we show the order parameter averaged over all
agents and all times along the trajectories. As in the
previous case we observe a maximum around the range
of values of β where performance is optimal.
Another parameter of interest is the upwind alignment
of the agents
M (t) = 1 −
1
N
N
Xi=1
U + vi(t) .
(4)
When all the agents move upwind one has M = 1 whereas
if they all move downwind M = −1. As shown in Fig. 2D
the upwind alignment, averaged over time, has a maxi-
mum around β = 0.85 which again confirms that a large
fraction of the group is heading against the mean wind
even if it has access only to a local running time aver-
age (the memory time is λ−1 = 1, much shorter than
Ts = 100).
The previous results point to the conclusion that there
is a relatively narrow range of the trust parameter β,
around 0.85, for which the collective olfactory search pro-
cess is nearly optimal, i.e. the time to reach the target is
close to the shortest possible one, and takes place with a
remarkable coherence of the group.
Results for a turbulent flow. To test the robustness
of our findings in a somewhat more realistic situation
we also considered the case where the wind velocity is
obtained from a direct numerical simulation (DNS) of
2D Navier-Stokes equations
∂tω + u · ∇ω = ν∆ω − αω + f ,
(5)
where ω = ∇ × u, the forcing f acts at small scales
so to generate an inverse kinetic energy cascade, that is
stopped at large scales by the Ekman friction term with
intensity α. In order to attain a statistically steady state,
the viscous term with viscosity ν dissipates enstrophy at
small scales.
In this way we obtain a multiscale flow
which is non-smooth above the forcing scale and smooth
below it (see [22 -- 24] for phenomenological and statisti-
cal flow properties). DNS have been carried out using
a standard 2/3 dealiased pseudo-spectral solver over a
bi-periodic 2π × 2π box with 2562 collocation points and
2nd order Runge-Kutta time stepping, see Supplementary
Material for technical details. In this flow the large scale
of the velocity field is about half the size of the simulation
box. The numerically obtained velocity field, for a dura-
tion of about 10 eddy turnover times, was then used to
integrate the motion of odor particles in the whole plane
exploiting the periodicity of the velocity field. Finally,
the mean wind is then superimposed. Further details
about the simulations are available in the Supplementary
Material.
Fig. 3 summarizes the main results obtained with the
turbulent flow. As shown in the left panel, the average
time taken by the first agent to reach the source is very
similar to the one obtained for the stochastic flow.
It
displays a minimum time close to the shortest-path time
Ts = Lx/v0 at values of the trust parameter β ≈ 0.8. The
other observables display very similar features as the ones
observed with the stochastic flow.
In the right panels we show four snapshots of the agents
at different times during the search process, for β = 0.8,
i.e. close to optimality. The flock appears to be moving
coherently in the upwind direction and the task is com-
pleted in a time 1.04 Ts just a few percent in excess of
the nominal minimal time.
Conclusions and discussion. We have shown that
there is an optimal way of blending private and pub-
lic information to obtain nearly perfect performances in
the olfactory search task. The first agent that reaches
the target completes the task by essentially moving in
a straight line to the target. This behavior is striking,
since in isolation agents move in a zig-zagging fashion
(see Fig.1C). Interestingly, the information about odor
and wind is essential to achieve this behavior, but its
weight in the decision making is numerically rather small,
about 20%. Although we do not expect that this number
stays exactly the same upon changing the various param-
eters of the model, we suspect that there is a common
trend for having optimal values of the trust parameter β
at the higher end of its spectrum, that is, closer to unity.
This may reflect the existence of a general principle of a
"temperate wisdom of the crowds" by which public infor-
mation must be exploited -- but only to a point. In the
present case, one way of summarizing our findings would
be the following rule: follow the advice of your neighbors
but once every four or five times ignore them and act
based on your own sensations.
With reference to the remarkable similarity between
5
searching in stochastic and turbulent flows shown by
Figs. 2 and 3, we stress that this is likely due to the spe-
cific sensing mechanisms that we have chosen, which is es-
sentially based on single-point single-time measurements.
If private cues included consecutive inputs along the
agent's trajectory and/or on spatially coarse-grained sig-
nals we expect that the results could have been more sen-
sitive to the structure of small-scale and high-frequency
turbulent fluctuations.
Our results suggest how to build efficient algorithms for
distributed search in strongly fluctuating environments.
It is important to point out, however, that our construc-
tion is inherently heuristic. Our model heavily draws
inspiration from animal behavior, combining features of
individual olfactory search in moths and collective navi-
gation in bird flocks. A more principled way of attacking
the collective search problem would be to cast it in the
framework of Multi Agent Reinforcement Learning [25]
and seek for approximate optimal strategies under the
same set of constraints on the accessible set of actions
and on the available private and public information. It
would then be very interesting to see if the strategy dis-
covered by the learning algorithms actually resembles the
one proposed here, or points to other known behavior
displayed by animal groups, or perhaps unveil some yet
unknown way of optimizing the integration of public and
private cues for collective search.
ACKNOWLEDGMENTS
M. D. and L. P. contributed equally to this work. L. B.
acknowledges partial funding from the European Union
Programme (FP7/2007-2013) grant No.339032. M. D.
is grateful for the graduate fellowship by the ICTP and
University of Trieste. M. D. acknowledges the kind hospi-
tality and support from University of Rome Tor Vergata.
M. D. acknowledges fruitful discussions with A. Pezzotta,
M. Adorisio, A. Roy and A. Mazzolini.
[1] C. David, J. Kennedy, and A. Ludlow, Nature 303, 804
[10] F. Grasso and J. Atema, J. Environmental Fluid Mechan-
(1983).
ics 2, 95 (2002).
[2] J. Kennedy, Physiol. Entomol. 8, 109 (1983).
[3] J. Elkinton, C. Schal, T. Ono, and R. Card´e, Physiol.
[11] T. Lochmatter and A. Martinoli, Experimental Robotics.
Springer Tracts in Advanced Robotics 54, 473 (2009).
Entomol. 12, 399 (1987).
[4] T. Dekker, M. Geier, and R. Card´e, J. Exp. Biol. 208,
2963 (2005).
[5] J. Murlis and C. Jones, Physiol. Entomol. 6, 71 (1981).
[6] A. Celani, E. Villermaux, and M. Vergassola, Phys Rev
[12] E. Bonabeau, G. Theraulaz, and M. Dorigo, Swarm In-
telligence: From Natural to Artificial Systems ( (Oxford
University Press, New York, USA, 1999).
[13] T. Pitcher, A. Magurran, and I. Winfield, Behav Ecol
Sociobiol 10, 149 (1982).
X 4, 041015 (2014).
[14] C. Torney, Z. Neufeld, and I. Couzin, Proceedings of the
[7] A. Lilienthal, D. Reimann, and A. Zell, Autonome Mo-
National Academy of Sciences 106, 22055 (2009).
bile Systeme 2003 , 150 (2003).
[15] C. Ioannou, M. Singh,
and I. Couzin, The American
[8] G. Ferri, E. Caselli, V. Mattoli, A. Mondini, B. Mazzolai,
and P. Dario, IEEE/RAS-EMBS Conf. Proc. BioRob-
2006 , 573 (2006).
[9] H. Ishida, H. Tanaka, H. Taniguchi, and T. Moriizumi,
Autonomous Robots 20, 231 (2006).
Naturalist 186, 000 (2015).
[16] A. Berdahl, C. Torney, C. Ioannou, J. Faria,
and
I. Couzin, Science 339, 574 (2013).
[17] N. Miller, S. Garnier, A. Hartnett, and I. Couzin, Pro-
ceedings of the National Academy of Sciences 110, 5263
(2013).
[22] G. Boffetta, A. Celani, and M. Vergassola, Phys. Rev. E
[18] E. Balkovsky and B. Shraiman, Proceedings of the Na-
61, R29 (2000).
tional Academy of Sciences 99, 12589 12593 (2002).
[23] G. Boffetta and R. E. Ecke, Annu. Rev. Fluid Mech. 44,
[19] T. Baker, M. Willis, and P. Phelan, Physiological Ento-
427 (2012).
mology 9, 365 (1984).
[24] A. Alexakis and L. Biferale, Phys. Rep. 767-769, 1
[20] T. Vicsek and A. Zafeiris, Phys. Rep. 517, 71 (2012).
[21] F. Ginelli, Europ. Phys. J. Special Topics 225, 2099
(2016).
(2018).
[25] L. Bu¸soniu, R. Babuska, and B. Schutter, IEEE trans-
actions on systems, man, and cybernetics-part C: appli-
cation and reviews 38 (2008).
6
|
1302.2961 | 1 | 1302 | 2013-02-13T00:38:23 | Detecting Lyme Disease Using Antibody-Functionalized Single-Walled Carbon Nanotube Transistors | [
"physics.bio-ph",
"cond-mat.mtrl-sci",
"q-bio.BM"
] | We examined the potential of antibody-functionalized single-walled carbon nanotube (SWNT) field-effect transistors (FETs) for use as a fast and accurate sensor for a Lyme disease antigen. Biosensors were fabricated on oxidized silicon wafers using chemical vapor deposition grown carbon nanotubes that were functionalized using diazonium salts. Attachment of Borrelia burgdorferi (Lyme) flagellar antibodies to the nanotubes was verified by Atomic Force Microscopy and electronic measurements. A reproducible shift in the turn-off voltage of the semiconducting SWNT FETs was seen upon incubation with Borrelia burgdorferi flagellar antigen, indicative of the nanotube FET being locally gated by the residues of flagellar protein bound to the antibody. This sensor effectively detected antigen in buffer at concentrations as low as 1 ng/ml, and the response varied strongly over a concentration range coinciding with levels of clinical interest. Generalizable binding chemistry gives this biosensing platform the potential to be expanded to monitor other relevant antigens, enabling a multiple vector sensor for Lyme disease. The speed and sensitivity of this biosensor make it an ideal candidate for development as a medical diagnostic test. | physics.bio-ph | physics | Detecting Lyme Disease Using Antibody-Functionalized Single-
Walled Carbon Nanotube Transistors
Author names and affiliations:
Mitchell B. Lerner*1, Jennifer Dailey*1, Brett R. Goldsmith1#, Dustin Brisson2, A.T. Charlie
Johnson1
* These authors contributed equally to this work
1 Department of Physics and Astronomy, University of Pennsylvania, 209 South 33rd Street ,
Philadelphia PA 19104
2 Department of Biology, University of Pennsylvania, 3740 Hamilton Walk, Philadelphia PA
19104
# Current address SPAWAR Systems Center Pacific, 53560 Hull St., San Diego CA 92152
Corresponding author: A.T. Charlie Johnson, [email protected]
Present/permanent address: Department of Physics and Astronomy, University of
Pennsylvania, 209 South 33rd Street, Philadelphia, PA 19104
Abstract:
We examined the potential of antibody-functionalized single-walled carbon nanotube
(SWNT) field-effect transistors (FETs) for use as a fast and accurate sensor for a Lyme
disease antigen. Biosensors were fabricated on oxidized silicon wafers using chemical
vapor deposition grown carbon nanotubes that were functionalized using diazonium salts.
Attachment of Borrelia burgdorferi (Lyme) flagellar antibodies to the nanotubes was
verified by Atomic Force Microscopy and electronic measurements. A reproducible shift in
the turn-off voltage of the semiconducting SWNT FETs was seen upon incubation with
Borrelia burgdorferi flagellar antigen , indicative of the nanotube FET being locally gated by
the residues of flagellar protein bound to the antibody. This sensor effectively detected
antigen in buffer at concentrations as low as 1 ng/ml, and the response varied strongly
over a concentration range coinciding with levels of clinical interest. Generalizable binding
chemistry gives this biosensing platform the potential to be expanded to monitor other
relevant antigens, enabling a multiple vector sensor for Lyme disease. The speed and
sensitivity of this biosensor make it an ideal candidate for development as a medical
diagnostic test.
Keywords: carbon nanotubes Lyme FET
1. Introduction
Lyme disease is a tick-borne illness caused by the bacteria Borrelia burgdorferi,
generating at least 30,000 new cases in the United States each year, although there are
likely many more cases that go undetected or misdiagnosed due to generality of symptoms
(Centers for Disease Control and Prevention 2011). Of the patients diagnosed with Lyme
disease, many are originally misdiagnosed due to the general symptoms of the disease
(Williams et al. 1990), inconsistent disease presentation in patients (Aguerorosenfeld et al.
1993), and lack of sensitive testing available for early stages of the infection (Bakken et al.
1997). Late detection of Lyme disease can result in further complications including
arthritis and permanent neurological disorders (Marques 2008). Diagnosis of Lyme disease
is severely hindered by the lack of reliable diagnostic tools despite its importance to
treatment success (Murray and Shapiro 2010; O'Connell 2010). A reliable and rapid
laboratory diagnostic tool is crucial to reducing the number of misdiagnosed patients and
for investigating appropriate treatment protocols for chronic Lyme disease.
In recent years, great progress has been made in the field of carbon nanotube field
effect transistor (CNT FET)-based biosensors (Allen et al. 2007). Benefits of nanotube-
based sensors include the speed and reliability obtained from performing multiple assays
in parallel (Chikkaveeraiah et al. 2009). Protein-functionalized nanotube-based FETs are of
great research and clinical interest for several reasons. Their nanometer size is
comparable to the size of many biomolecules of interest , suggesting a unique biocompatible
platform (Harrison and Atala 2007; Lerner et al. 2011; Sudibya et al. 2009). Additionally,
since every atom of a carbon nanotube is located on its surface, in direct contact with the
environment, they are a clear choice for direct environmental sensing . Commercially
available assays for Lyme-specific antigens in urine and cerebrospinal fluid have a limit of
detection of 12-15 ng/mL (Shah et al. 2004). We hypothesized that an antibody-
functionalized SWNT FET immunosensor would be able to detect the small amount of Lyme
antigen that is present in bodily fluids at very early stages of Borrelia infection (Harris and
Stephens 1995) since protein-functionalized nano-enabled sensors have demonstrated
very low detection limits, on the order of fM (Duan et al. 2012; Lerner et al. 2012a). Direct
detection of the antigen provides earlier test results because it eliminates the delay
required for the immune system to produce sufficient quantities of antibodies to be
detected via Western Blot or ELISA, which can improve patient prognosis (Ma et al. 1992).
Here we demonstrate that antibody-functionalized SWNT FET devices are effective
biosensors for rapidly detecting Lyme flagellar antigen at clinically relevant concentrations,
as low as 1 ng/ml, with negligible response to negative-control proteins and to pure buffer
solution.
2. Materials and Methods
2.1 Device Fabrication:
Carbon nanotube transistors were fabricated using previously described methods (Khamis
et al. 2011). Briefly, a solution of Iron (III) Nitrate dissolved in isopropanol was spin coated
onto a p++ doped Si/SiO2 wafer to give an iron catalyst layer. Single-walled nanotubes were
grown by catalytic chemical vapor deposition (CVD) with methane (2500 sccm) as the
carbon source in a background of forming gas (600 sccm Ar, 320 sccm H2) at 900°C for 2
mins. Following the growth, an optimized bilayer photolithography process using PMGI
and Shipley 1813 was used to pattern source and drain electrodes with 2.5 um channel
length that were subsequently metallized (3 nm Ti/40 nm Pd) using thermal evaporation.
The doped silicon substrate served as a global backgate to complete the three terminal
field-effect transistor geometry. Devices were individually characterized by sweeping the
back gate voltage from -10 V to 10 V while holding the bias voltage fixed at 100 mV.
Approximately 80 high quality semiconducting SWNT devices with ON/OFF ratios >1000
were selected for use in experiments.
2.2 Protein functionalization
Monoclonal Lyme antibodies specific for Borrelia Burgdorferi flagellar antigen (p41) were
obtained commercially (antibodies-online.com). Histidine-tagged Lyme antigen containing
the p41 flagellar immunodominant region (also known as flagellin) was obtained from
ProSpec. Antigen was diluted with Tris-HCl buffer (pH 7.5), aliquoted to several
concentrations and stored at 4° C. Antibodies were aliquoted at a concentration of 1 μg/ml
in Tris-HCl buffer and stored at -20° C.
Nanotube functionalization followed previously documented procedures (Lerner et al.
2012a) adapted from (Strano et al. 2003) (see Figure 1). Carbon nanotubes were
functionalized using diazonium salts synthesized according to a published recipe (Saby et
al. 1997). Samples were immersed in a solution of 4-carboxybenzene diazonium
tetrafluoroborate at a concentration of 2.5 mg/1 mL deionized (DI) water for 1 hr at 40°C
to create sp3 hybridization sites ending in carboxylic acid groups, then rinsed for 1 min
each in acetone, methanol, and deionized water baths. The carboxylic acid groups were
then activated with EDC and stabilized with sulfo-NHS at concentrations of 6 mg and 16 mg
per 15 mL MES buffer (pH 6.0) respectively for 15 minutes at room temperature, followed
by a DI water rinse. A solution of antibodies at a concentration of 1 μg/mL was pipetted
onto the nanotube devices in a humid environment to prevent the solution from
evaporating, causing primary amines on the antibody to displace stabilized sulfo -NHS sites
over a period of one hour. The devices were washed thoroughly by rinsing in two DI water
baths for 2 mins each and dried with gentle (less than 20 psi) nitrogen flow in order to
minimize the amount of salts and non-specifically bound proteins on the device.
Figure 1 Functionalization chemistry for Lyme antibody and binding of flagellar antigen. First, a pristine
nanotube is treated with diazonium salts to create sp3 hybridization sites ending in carboxylic acid moieties. The
carboxylic acid is activated with EDC and stabilized with sulfo-NHS. The sulfo-NHS is displaced by the primary
amine in a surface lysine residue (depicted in red) on the anti-p41 antibody. The flagellar antigen (depicted
according to secondary structure in purple, cyan, and yellow) then binds to the epitope on the antibody during
the exposure step.
Antibody-functionalized SWNT-FET devices were similarly exposed to droplets of antigen
at a known concentration for 20 minutes in a humid environment to prevent the droplet
from evaporating. Exposure to Lyme flagellar antigen occurred over a sufficiently long time
scale to ensure adequate time for the proteins to diffuse to the sensor and establish
equilibrium between bound and unbound species. This is known to be a critical
consideration for detection of biomolecular analytes at pM or lower concentration (Squires
et al. 2008). The devices were then washed in two DI water baths for 2 mins each to
remove non-specifically bound antigen and dried with gentle nitrogen flow. Each device
was exposed to only one concentration of antigen in order to avoid contamination of
samples, and each concentration of antigen was tested on 5-10 functionalized devices to
ensure reproducibility of the results.
3. Results and Discussion
In order to verify the validity of the attachment chemistry, Atomic Force Microscopy (AFM)
data were gathered on an Asylum AFM in tapping mode following both antibody
attachment and exposure to antigen. Figure 2a-b are AFM images that show the
effectiveness of the functionalization procedure, which results in 3-5 attachment sites per
micrometer of nanotube length. Statistical analysis shows that the antibodies are 2.78 nm ±
0.22 nm in size, slightly smaller than expected for a complete IgG. This is likely due to the
protein being slightly distorted during tapping mode AFM in air. We conducted a control
experiment to establish that nanotubes and the nanotube/SiO2 interface show very low
affinity for nonspecific binding of the Lyme antigen; details are provided in Supplementary
Materials Figure S1. When exposed to 1 µg/mL Lyme antigen, there was minimal non-
specific binding to the nanotubes or the surrounding substrate .
Figure 2 a) AFM image of anti-p41 antibody proteins attached to carbon nanotubes. Z scale is 6 nm, average
protein feature is ~2.8 nm in size. b) AFM image after incubation in a solution of FLA antigen at a concentration of
400 ng/mL shows antigen attached to bound antibodies on a nanotube. Z scale is 8 nm . c) Histogram of feature
heights following antibody attachment (red data) and exposure to antigen (black data).
After exposure to flagellar antigen at a concentration of 400 ng/mL, the histogram of
feature heights shows a shift towards larger proteins complexes. There are two main peaks
at feature height of 4.7 ± 0.6 nm and 6.5 ± 0.4 nm. These are associated with binding of one
and two antigen proteins, respectively, to a bound antibody, consistent with the fact that
the IgG used in the experiments has two binding sites (Talwar and Srivastava 2006). Each
added antigen increases the feature height by approximately 1.8 nm. There also appears to
be an additional peak at ~3 nm that represents unreacted antibodies; this peak accounts
for ~20% of the total features measured. These data suggest that after exposure to antigen
at 400 ng/mL, approximately 80% of the antibodies have bound at least one antigen, in
good agreement with the electronic response data presented in Figure 3b, below.
Electronic measurements of the current as a function of the backgate voltage (I-Vg
characteristics) for individual NT FET devices were taken following each chemical
modification to monitor the effect of chemical functionalization and to confirm attachment
of antibodies (Figure 3a). Parameters of interest included ION, the ON state current of the
device, and VTH, the threshold voltage, where the line tangent to the I-Vg curve intersects
the gate voltage axis. A 50-90% drop in ION as well as a 3-4 V decrease in the threshold
voltage resulted from diazonium oxidation, which is associated with the creation of sp3
hybridization sites terminated with a carboxylic acid group. EDC/NHS treatment resulted
in a slight decrease in the ON state current and no statistically significant shift in the
threshold voltage. Following antibody attachment, there was typically an increase in ON
state current, suggesting that attachment led to a decrease in carrier scattering . Upon
antigen exposure, a shift in the threshold voltage towards more negative values was
consistently observed. There was no statistically significant change in ON state current
following antigen exposure.
Figure 3 a) Current vs. Gate Voltage characteristics at subsequent stages of nanotube functionalization. The
sensing response is the 2.3 V decrease of the threshold voltage upon exposure to antigen (green to purple curves).
b) Threshold voltage shift as a function of antigen concentration can be fit with a model based on the Hill -
Langmuir equation. These data indicate a limit of detection of ~ 1 ng/mL and non -cooperative antigen binding .
Inset shows structure of anti-p41 Borrelia antibody with lysine (red) , histidine (orange) and arginine (green)
residues highlighted. Any of these basic residues could become protonated and generate a local gating effect if
they come into closer contact with the nanotube upon antigen binding.
There was a systematic dependence of the threshold voltage shift with varying antigen
concentration in the range 0.1 ng/ml to 3 µg/ml, with each concentration tested on 5-10
functionalized devices. The variation of the average measured shift in the threshold voltage
as a function of antigen concentration is displayed in Figure 3b. Error bars shown are the
standard error of the mean. The sensor responses agree with a model based on the Hill-
Langmuir equation for equilibrium protein binding (see Figure 3b) (Hill 1910; Lehninger et
al. 2008).
Here c is the Lyme antigen concentration, A is the sensor response at saturation when all
binding sites are occupied, Z is an overall offset to account for the response to pure buffer,
Kd is the dissociation constant describing the concentration at which half of available
binding sites are occupied, and n is the Hill coefficient describing cooperativity of binding.
The best fit to the data yielded a maximum response A = -2.52 V ± 0.32 V, offset parameter
Z = -0.12 V ± 0.16 V, dissociation constant Kd = 87 ng/mL ±27 ng/mL, and n = 1.04 ± 0.08.
The best fit value of the offset parameter Z = -0.12 V ± 0.16 V was statistically
indistinguishable from the experimentally measured responses of seven devices to pure
buffer as a negative control (ΔV=0.11 V ± 0.15 V). The value of Kd = 87 ± 27 ng/mL
determined from the fit describes the concentration of antigen at which half the receptors
are occupied; this regime coincides with antigen levels of diagnostic significance, 12-15
ng/mL (Shah et al. 2004). It is notable that the AFM image in Figure 2, taken after exposure
to antigen at 400 ng/mL, showed occupation of at ~80% of the attached antibody sites, in
good agreement with the prediction of the Hill-Langmuir fit that 95% of the active sites
were occupied at this concentration. The slight discrepancy could be due to antibodies that
were bound to the nanotube in such an orientation that their epitope is obstructed or were
otherwise non-functional. The Hill-Langmuir model combined with the AFM data suggested
that the number of such antibodies unavailable for binding is no more than 20% of the total
bound antibodies.
The best fit value of the cooperativity parameter, n = 1.04 ± 0.08, indicates independent
binding of Lyme antigen to the anti-p41 in the context of the NT-FET biosensor. The data
presented in Figure 3b show that the measured responses from a collection of 5 -10 devices
could be used to differentiate between pure buffer solution (ΔV = 0.11 V ± 0.15 V) and
buffer containing Lyme antigen at a concentration of at a concentration of 1 ng/mL (ΔV = -
0.31 V ± 0.24 V). Previous work suggests that this limit of detection may be lowered by as
much as a factor of 1000 by replacing the complete IgG with its single chain variable
fragment, thus allowing the binding event to occur closer to the nanotube where the
electrostatic influence on transport in the NT transistor will be more pronounced (Lerner
et al. 2012a).
The observed reduction in turnoff voltage is consistent with a gating effect due to the
presence of positive charges in the local electrostatic environment of the carbon nanotube
(Heller et al. 2008). The proposed mechanism responsible for introducing these positive
local charges is a conformational change in the antibody protein upon binding the flagellar
antigen, which results in the carbon nanotube being exposed to different amino acids on
the antibody exterior. Charges in close proximity to carbon nanotubes have been shown to
shift the transistor threshold voltage by local gating (Lerner et al. 2012b). Lysine, histidine,
and arginine (highlighted in the inset of Figure 3b) are common amino acids, a portion of
which will be protonated when the pH is less than 7 and could thus be candidates for
locally gating the nanotube as observed. Experimental conditions may be as low as pH 5
due to deprotonation of silanol groups in a thin water layer that forms on the hydrophilic
SiO2 surface (O'Reilly et al. 2005). Even in the absence of a quantitative understanding of
the device response, the results provide strong evidence that the methods used here
enabled attachment of antibody proteins to a NT-FET while maintaining both the high
quality electronic characteristics of the NT device and the chemical recognition
functionality characteristic of the protein.
As a further control experiment, antibody-functionalized sensors were incubated in a
solution of bovine serum albumin (BSA) at high concentration (1 μg/mL) to approximate
the effect of non-specific proteins present in patient samples (human serum albumin is a
large component of blood plasma proteins and represents a potential interfering agent
(Peters 1996)). A sample I-Vg characteristic is presented in Supplementary Materials Figure
S2. The sensor response to BSA, averaged over eight devices, was a shift of -0.13 V ± 0.18 V,
similar to the response produced by buffer alone. We thus concluded that the Lyme
flagellar antibody-functionalized carbon nanotube transistors exhibited a high level of
specificity for the flagellar protein target antigen.
4. Conclusions
We demonstrated that antibody-functionalized SWNT FET devices are effective
biosensors for rapidly detecting Lyme disease antigen at diagnostically relevant
concentrations. The sensor shows a decrease in the threshold voltage in minutes rather
than days, as required for a traditional clinical assay. We achieved detection of flagellar
antigen protein at a concentration of 1 ng/ml with negligible response to control proteins
and to pure buffer solution. The experiments showed an antigen-specific, concentration-
dependent sensor response over a wide range of concentrations (1 ng/mL to 3000 ng/mL)
that was in excellent quantitative agreement with a model based on the Hill-Langmuir
equation of equilibrium thermodynamics. Future work includes functionalizing samples
with single chain variable fragments to increase sensitivity and the use of a sensor array
based on multiple antibodies in order to capture several types of proteins indicative of
Lyme disease for a multiplexed biosensor platform. The rapidity and ease of use of this
sensor is superior to traditional immunoassays, suggesting its utility as a point-of-care
diagnostic tool.
Acknowledgements
This research was supported by the Department of Defense US Army Medical
Research and Materiel Command through grants W81XWH-09-1-0205 and W81XWH-09-1-
0206 (M.L, J.D. A.T.C.J) and by the National Institutes of Health grant R01AI076342 (D.B.).
We acknowledge use of facilities associated with the Nano/Bio Interface Center, National
Science Foundation NSEC DMR08-32802. J.D. acknowledges support of the REU program of
the Laboratory for Research on the Structure of Matter (NSF REU Site Grant DMR-
1062638) and of the Penn University Scholars program.
References
Aguerorosenfeld, M.E., Nowakowski, J., Mckenna, D.F., Carbonaro, C.A., Wormser, G.P., 1993.
J. Clin. Microbiol. 31(12), 3090-3095.
Allen, B.L., Kichambare, P.D., Star, A., 2007. Adv Mater 19(11), 1439-1451.
Bakken, L.L., Callister, S.M., Wand, P.J., Schell, R.F., 1997. J. Clin. Microbiol. 35(3), 537-543.
Centers for Disease Control and Prevention, N.C.f.E.a.Z.I.D.N., Division of Vector -Borne
Diseases (DVBD), 2011. Chikkaveeraiah, B.V., Bhirde, A., Malhotra, R., Patel, V., Gutkind, J.S.,
Rusling, J.F., 2009. 81(21), 9129-9134.
Duan, X.X., Li, Y., Rajan, N.K., Routenberg, D.A., Modis, Y., Reed, M.A., 2012. Nat Nanotechnol
7(6), 401-407.
Harris, N., Stephens, B.G., 1995. Harris, N., and B. G. Stephens. Journal of Spirochetal and
Tick-borne Diseases 2, 37-41.
Harrison, B.S., Atala, A. , 2007. Biomaterials 28(2), 344-353.
Heller, I., Janssens, A.M., Mannik, J., Minot, E.D., Lemay, S.G., Dekker, C., 2008. Nano Lett.
8(2), 591-595.
Hill, A., 1910. The Journal of Physiology 40, 4-7.
Khamis, S.M., Jones, R.A., Johnson, A.T.C., 2011. Aip Adv 1(2).
Lehninger, A.L., Nelson, D.L., Cox, M.M., 2008. W.H. Freeman, New York.
Lerner, M.B., D'Souza, J., Pazina, T., Dailey, J., Goldsmith, B.R., Robinson, M .K., Johnson, A.T.C.,
2012a. Acs Nano 6(6), 5143-5149.
Lerner, M.B., Goldsmith, B.R., McMillon, R., Dailey, J., Pillai, S., Singh, S.R., Johnson, A.T.C.,
2011. Aip Adv 1(4).
Lerner, M.B., Resczenski, J.M., Amin, A., Johnson, R.R., Goldsmith, J.I., Johnson, A.T., 2012b. J.
Am. Chem. Soc. 134(35), 14318-14321.
Luger, S.W., Krauss, E., 1990. Archives of internal medicine 150(4), 761-763.
Ma, B., Christen, B., Leung, D., Vigo-Pelfrey, C., 1992. J. Clin. Microbiol. 30(2), 370-376.
Marques, A., 2008. 22(2), 341-360, vii-viii.
Murray, T.S., Shapiro, E.D., 2010. Clinics in laboratory medicine 30(1), 311-328.
O'Connell, S., 2010. Current opinion in infectious diseases 23(3), 231 -235.
O'Reilly, J.P., Butts, C.P., I'Anson, I.A., Shaw, A.M ., 2005. J Am Chem Soc 127(6), 1632-1633.
Peters, T., 1996. Academic Press, San Diego.
Saby, C., Ortiz, B., Champagne, G.Y., Belanger, D., 1997.Langmuir 13(25), 6805-6813.
Shah, J., Harris, N., Bai, H., Remollo, G., 2004. In: Office, U.S.P.a.T. (Ed.). IGeneX, Inc. (Palo
Alto, CA), United States of America.
Squires, T.M., Messinger, R.J., Manalis, S.R., 2008. Nat Biotechnol 26(4), 417-426.
Strano, M.S., Dyke, C.A., Usrey, M.L., Barone, P.W., Allen, M.J., Shan, H.W., Kittre ll, C., Hauge,
R.H., Tour, J.M., Smalley, R.E., 2003. Science 301(5639), 1519-1522.
Sudibya, H.G., Ma, J.M., Dong, X.C., Ng, S., Li, L.J., Liu, X.W., Chen, P., 2009. Angew Chem Int
Edit 48(15), 2723-2726.
Talwar, G.P., Srivastava, L.M., 2006. Prentice-Hall of India Private Limited, New Delhi, India.
Williams, C.L., Strobino, B., Lee, A., Curran, A.S., Benach, J.L., Inamdar, S., Cristofaro, R., 1990.
The Pediatric infectious disease journal 9(1), 10-14.
|
1210.7897 | 1 | 1210 | 2012-10-30T03:34:25 | Diffusion wave and signal transduction in biological live cells | [
"physics.bio-ph"
] | Transduction of mechanical stimuli into biochemical signals is a fundamental subject for cell physics. In the experiments of FRET signal in cells a wave propagation in nanoscope was observed. We here develop a diffusion wave concept and try to give an explanation to the experimental observation. The theoretical prediction is in good agreement to result of the experiment. | physics.bio-ph | physics | Diffusion wave and signal transduction in biological live cells
Tian You Fan and Lei Fan
Department of Physics, School of Physics, Beijing Institute of
Technology, Beijing 100081, China
Department of Biology, School of Life Science, Beijing Institute of
Technology, Beijing 100081, China
[email protected]; [email protected]
Abstract
Transduction of mechanical stimuli into biochemical signals is a
fundamental subject for cell physics. In the experiments of FRET signal
in cells a wave propagation in nanoscope was observed. We here develop
a diffusion wave concept and try to give an explanation to the
experimental observation. The theoretical prediction is in good agreement
to result of the experiment.
Mechanical effects of environment influence many essential processes of
gene expression, cell survival and growth etc [1], among them the
transduction of mechanical stimuli into biochemical signals is a
fundamental subject for cell physics. However the mechanism on the
transduction remains unsolved mystery so far [2-7].
Combination of
fluorescent
resonance energy
transfer
(FRET),
1
1212 genetically encoded biosensor and
laser-tweezer [8-11] makes a
revolutionary progress of our understanding on the transduction above
mentioned in the recent two decades. This experimental technology
enables
the
imaging
and
quantification
of
spatio-temporal
characterization of mechanotransduction to biochemical signals in live
cells.
Wang et al. [9] carried out a systematical study by using fluorescent
resonance energy transfer and developed a genetically encoded biosensor
approach. They observed a series of signal transduction events at the cell
membrane, in which the long-range wave propagation with speed
along the cell membrane is especially interesting.
They further measured the wave length in a certain extent, this provides
the most important information for theoretical analysis.
Na et al. [2] summarized the relevant studies up to 2008 and developed
the stress-induced Src activation, and analyzed some mechanism on the
transduction, e.g. stress effect,
interaction between
integrins and
cytoskeleton, deformation of mircotubules, effect of stiffness of
substrates, conformational changes or unfolding of membrane-bound
proteins, and diffusion of the proteins, and so on.
There is no possibility to explore all aspects concerning the mechanism, it
is difficult even if revealing each among them. In this note we focus on
the dynamic process explored by Wang et al. [9], and suggest a
2
(18.11.7)/Unms physical-mathematical model
to explain
the
long-range wave
propagation.
Experiment [9] is carried out in a fibronectin-integrin-cytoskeleton-cell
memebrane of human umbilical vein endothelial cells (HUVEs), this is a
complex of proteins. A local mechanical stimulation is implemented by
the laser-tweezer traction (The experimental detail can be referred to
Wang et al. [9]). This aroused a motion of the protein complex, and led to
a kind of wave propagation of Src activation with new feature. This new
feature should reflect the dynamic nature of interaction between proteins
and cell membrane. Due to the unclearness of mechanism of the wave
propagation,
the
following
study
has
to
undertake
some
phenomenological analyses.
One-dimensional analysis. Relative to slow wave propagation, the
dimensional of cell body (
)
is quite
large, we
approximately assume that the cell membrane is a large plane for a
certain time interval, we take the plane is
plane. Furthermore, for
simplicity we consider the one-dimensional case first. If the protein
concentration of a cell is denoted by
, then the equation of motion
of the membrane-bound proteins is
(1)
in which
represents diffusion coefficient of the membrane-bound
proteins,
the spatial coordinate,
the time.
3
111mmmxy(,)Cxt22CCDtxDxt The local mechanical stimulation
(induced by external force
, or displacement
, or stress
) is not, in general, a
periodic function, but which may be expanded by a Fourier series as
follows
in which each Fourier component is a periodic function. For simplicity,
we consider a periodic mechanical stimulation. Equation (1) is diffusion
equation, for the periodic stimulation, it can have a wave propagation
solution for long-range wave propagation such as
(2)
the wave is propagating along the cytoplasm membrane, in which
is
a constant,
the wave number,
the circular frequency (or angular
frequency) ,
, respectively, and the wave speed
is unknown at
moment.
Substituting (2) into (1) yields
(3)
This is a new relation—dispersion relation between
and
, in which
the wave number is connected to diffusion coefficient
, this is quite
different from those based on traditional wave theory. The wave
described by (2) and (3) may be viewed as diffusion wave. According to
the well-known wave law one can find the wave speed as
4
(,)fxt(,)Fxt(,)uxt(,)xt()(,)nnnixktnnofxtAe()0(,)ikxtCxtCe0Ck1iU12kDkD
(4)
and substituting (3) into (4), we determine the wave speed as follows
(5)
where
is the wave length and there is relation between the wave
length
and the wave number
as below
(6)
Formula (5) is also a new result, in which the wave speed depends upon
not only quantity concerning wave (e.g. wave length) but also diffusion
coefficient.
With the above theoretical results, we can analysis on experimental
phenomena. As reported in Ref [9], the observed values of wave length
for the signal propagation are
.
In the following calculation one can take an average value, e.g.
Furthermore we analyze the proteins are membrane-bound membrane
proteins of cells, the diffusion coefficient of the proteins of cell
membrane is in the order of magnitude (refer to Fan and Fan [12])
As an average value we take an average value, e.g.
Substituting above data into (5), one finds that the wave propagation
speed of PRET signal is in the range
5
Uk48DUkDk2k470550nm500nm91321010/Dcms102(35)10/Dcms
which is in good agreement to the experimental measurement value
that was measured by Wang et al. [9]. This has examined the efficiency of
our physical-mathematical model.
Two- and
three-dimensional analysis. For simplicity only
the
one-dimensional analysis is discussed in the above section. However the
two- and three-dimensional analyses have no difficulty and are given in
the following.
For two- and three-dimensional cases the equation of motion (1) will be
replaced by
(7)
in which
represents the concentration of proteins of a cell, and
for two-dimensional case, and
for
three-dimensional case, respectively. The wave solution of equation (7) is
(8)
where
is a constant,
,
the wave vector,
the radius vector, and
for two-dimensional
case, and
for three-dimensional case, respectively.
Furthermore the dispersion relation is similar to (3), i.e.,
(9)
6
(16.828.6)/Unms(18.11.7)/Unms2CDCt(,)Ctr22222xy2222222xyz()0(,)itCxtCekr0C1i(,,)xyzkkkk(,,)xyzrxykxkykrxyzkxkykzkr22212xyzkkkkDk The wave velocity formula is also similar to (4), i.e.,
(10)
but here the wave velocity
is a vector rather than a scalar. According
to (10) any wave velocity components can be evaluated. Up to now we
have not found the experimental data concerning the two- and
three-dimensional wave propagation in the mechanical activation in cells ,
the further calculation cannot be offered yet.
Although the present study is phenomenological, the solution especially
the new dispersion relation (3) are very simple and effective, which can
serve a mathematical basis to explain the long-range wave propagation
induced by external force stimulation to the living cells.
Furthermore there is another example in physics (the phason dynamics of
quasicrystals) supporting the diffusion wave theory, we will report the
relevant results elsewhere.
Acknowledgement The first author thanks the National Natural
Science Foundation of China for financial support. We also thank Dr. S.
Chien of the University of California at San Diego and Dr. Y. Wang of
University of Illinois at Urban for their helps, so that the figures of Ref [9]
with high resolution are available.
7
UkU
[1] E B Singer, K D Puri, R Alon, M B Lawrences , U H von Andrian and
H A Springer, Nature, 349, 266(1996).
[2] S Na, D Collin, F Chowdhury, B Tay, M Ouyan, Y Wang and N Wang,
Proc. Natl. Acad. Sci., USA, 105, 6626(2008).
[3] D E Ingber and H Tensegrity, J. Cell. Sci., 116, 1397(2003).
[4] D P Felsenfeld, P L Schwarzberg, A Venegas, R Tse and M P Sheetz,
Nature Cell Biol. 1, 200(1999).
[5] D R Critchley, Curr. Opin. Cell Biol. 12, 133(2000).
[6] M E Chicurel, R H Singer, C J Meyer and D E Ingber, Nature 392,
730 (1998).
[7] E G Arias-Salgado, S. Lizano, S Sarker, J S Brugge, M. H. Ginsberg
and S J Shattil, Proc. Natl. Acad. Sci. USA 100, 13298(2003).
[8] M Puig-de-Moreles, E Millet, B Fabry, D Navajas, N Wang, J P Butler
and J J Fredberg , Am J. Physiol. 287, C643(2004)
[9] Y Wang, E L Botvinck, Y Zhaol, Berns M W, Usami S, Tsien R Y and
S Chien , Nature 434, 1040(2005).
[10] Y Sawada, M Tamada, B J Dubin-Thaler, O Chemiavskaya, R Sakai,
S Tanaka and M P Sheetz , Cell 127, 1015(2006).
[11] C P Johnson, H Y Tang, C Carag, D W Speicher and D E Discher ,
Science 317, 663(2007).
[12] L Fan and T Y Fan, Modern Physics Letters B 24, 1533(2010).
8
|
1506.02884 | 2 | 1506 | 2015-06-11T13:36:58 | To understanding of the mechanisms of DNA deactivation in ion therapy of cancer cells | [
"physics.bio-ph"
] | Changes in the medium of biological cell nucleus under ion beam action is considered as a possible cause of cell functioning disruption in the living body. As the most long-lived molecular product appeared in the cell after the passage of high energy ions, the hydrogen peroxide molecule is picked out. The possibility of the formation of stable complexes of hydrogen peroxide molecules with active sites of DNA nonspecific recognition (phosphate groups of the double helix backbone) is studied, and the formation of stable DNA-peroxide complexes is considered. Due to the negative charge on the oxygen atoms of DNA phosphate group in solution the counterions that under natural conditions neutralize the double helix have been also taken into consideration. The complexes consisting of oxygen atoms of DNA phosphate group, H$_2$O$_2$ and H$_2$O molecules, and Na$^{+}$ counterion have been considered. Energy of the complexes have been determined based on the electrostatic and van der Waals interactions within the approach of atom-atom potential functions. The stability of various configurations of molecular complexes has been estimated. It has been found that hydrogen peroxide molecules can form the stable complexes with phosphate groups of DNA and counterions which are no less stable than the complexes with water molecules. It is shown that the formation of stable complexes of H$_2$O$_2$--Na$^{+}$--PO$_{4}^{-}$ can be detected experimentally by the observation of specific DNA vibrations in the low-frequency Raman spectra. The interaction of H$_2 $O$_2$ molecule with phosphate group of the double helix backbone can block the processes of DNA biological functioning and induce the deactivation of the genetic apparatus of the cell. Thus, the new channel of high-energy ions action on living cell has been proposed. | physics.bio-ph | physics |
To understanding of the mechanisms of DNA deactivation
in ion therapy of cancer cells
D.V. Piatnytskyi, O.O. Zdorevskyi, S.M. Perepelytsya, S.N. Volkov
Bogolyubov Institute for Theoretical Physics, NAS of Ukraine,
14-b Metrolohichna Str., Kyiv, 03680, Ukraine
[email protected]
July 3, 2021
Abstract
Changes in the medium of biological cell nucleus under ion beam action is consid-
ered as a possible cause of cell functioning disruption in the living body. As the most
long-lived molecular product appeared in the cell after the passage of high energy ions,
the hydrogen peroxide molecule is picked out. The possibility of the formation of stable
complexes of hydrogen peroxide molecules with active sites of DNA nonspecific recogni-
tion (phosphate groups of the double helix backbone) is studied, and the formation of
stable DNA-peroxide complexes is considered. Due to the negative charge on the oxygen
atoms of DNA phosphate group in solution the counterions that under natural conditions
neutralize the double helix have been also taken into consideration. The complexes con-
sisting of oxygen atoms of DNA phosphate group, H2O2 and H2O molecules, and Na+
counterion have been considered. Energy of the complexes have been determined based
on the electrostatic and van der Waals interactions within the approach of atom-atom
potential functions. The stability of various configurations of molecular complexes has
been estimated. It has been found that hydrogen peroxide molecules can form the stable
complexes with phosphate groups of DNA and counterions which are no less stable than
the complexes with water molecules. It is shown that the formation of stable complexes
of H2O2 -- Na+ -- PO−
4 can be detected experimentally by the observation of specific DNA
vibrations in the low-frequency Raman spectra. The interaction of H2O2 molecule with
phosphate group of the double helix backbone can block the processes of DNA biological
functioning and induce the deactivation of the genetic apparatus of the cell. Thus, the
new channel of high-energy ions action on living cell has been proposed.
1
Introduction
Invention of new instruments for cancer therapy is known to be one of the most important
challenges of modern science. In spite of the increase of funding for the development of new
anticancer drugs the illness rate is still very high [1]. In this regard, the elaboration of non
pharmacological methods of cancer therapy are stimulated. The last decades, use of high energy
ion beams has established itself as an effective treatment of cancer disease. This method is based
on well-known Bragg effect [2 -- 5]. Heavy ions (usually protons or 12C6+), accelerated to the
1
energies of hundreds of MeV, are targeted into tumor. Within organism tissue the ions lose
their initial energy mainly at the end of their trajectory of motion, and the energetic peak is
formed (the Bragg peak). The initial beam energy is chosen in such a way that the position
of the Bragg peak coincide with the location of tumor in human body (it can be 5-10 cm from
body surface). The energy which is lost in living tissues destroys cancer cells and consequently
the whole tumor. The method of ion beam therapy is much more effective than X-ray therapy,
because of local action in the Bragg peak, especially in the case of places of difficult access,
such as human brain [3 -- 5]. In spite of great practical interest and large number of ion therapy
facilities already built, there is no clear understanding of the molecular mechanism of action of
high energy ions.
Starting with the early study [6] the action of radiation on living organisms is known to be
related with the DNA damage. The breaking of DNA double helix affects the mechanism of
storage and transfer of genetic information which leads to cell death. Under heavy ion irradia-
tion the breaks of hydrogen bounds in nucleotide pairs occur, which cause DNA melting, single
and double strand breaks in double helix [3 -- 5]. Single strand breaks in DNA and melting can
be repaired due to the complementary structure of double helix and cell reparation mechanisms.
The double strand breaks in DNA lead to damage of macromolecule, which can not be repaired
[7]. Thus the formation of large number of double strand breaks of double helix is necessary
for deactivation and destruction of DNA molecule.
The double strand breaks in DNA macromolecule occur due to the large quantity of sec-
ondary electrons born in the medium under the action of ion beam [8 -- 10, 13]. Some molecules
decay into atoms and radicals that can break chemical bonds in biological macromolecules.
At the same time, the calculations show that the number of the double strand breaks, made
by secondary electrons and radicals, is insufficient for cell death [10]. Therefore, additional
channels of ion beam action on living tissues should exist.
To describe DNA damage the formation of the shockwave in the medium after passage
of high energy ions is suggested [11, 12]. Within the framework of shockwave mechanism the
destruction of DNA double helix may occur even in the case of macromolecules situated far from
ion hit point in the cell. The molecular dynamics simulations shows that shockwave cause large
number of double strand breaks in DNA double helix [13]. However, in spite of high efficiency
of proposed mechanism, there are processes which can prevent DNA damage in the living cell,
particularly, the histone proteins restraining the macromolecule in frames of chromatin.
The deactivation of DNA macromolecule may be induced by changes of medium that occur
after the passage of high energy ions. The goal of the present work is to consider the changes
of medium in the Bragg peak region and to pick out new molecular compounds that can
influence the structure and dynamics of DNA double helix. The analysis of changes in water
medium induced by ion beams is carried out in second section and the most probable molecular
compounds formed in the Bragg peak are analyzed. The third section is dedicated to the
description of calculation method of interaction of appeared molecules and atoms of DNA double
helix. The results of energy calculations are presented in fourth section and the most stable
complexes are selected. In the fifth section the low-frequency vibrations of formed complexes
are studied, and the specific modes of the complexes are found in the Raman spectra of DNA.
The possible action of medium changes on DNA biological functioning is discussed in the sixth
section.
2
l
d
e
Y
i
t
c
u
d
o
r
P
l
)
s
e
u
a
v
e
v
i
t
a
e
r
(
l
H O2
2
H2
HO
+
H O3
H
Products
e -
-
HO
Figure 1: Products of water radiolysis after 10−6 s of ion passage. Bar graph is built using the
Monte Carlo simulation data [20].
2 Medium change under ion beam radiation
The ions of high energy passing through intracellular medium induce the process of decay of
water molecules (water radiolysis), different chemical and nuclear reactions [14 -- 21]. The pro-
cesses of nuclear fragmentation are showed be not significant for environment [15, 16], whereas
the fragmentation of water molecules change essentially the composition and properties of water
medium [3 -- 5,10]. In the process of water radiolysis the secondary electrons (e−), radicals (OH•
2), molecules (H2O2 and H2O), and other products (H•, OH−, H3O+) are produced
and HO•
[17 -- 19].
Quantity and composition of radiolysis products depend on absorbed radiation dose. At
small doses the number of secondary electrons and radicals is the largest, while the increase
of the dose energy leads to the increase of molecular products yields. Due to the high energy
transfer the number of molecular products is expected to be the largest in the Bragg peak
region. The Monte Carlo simulation data for water radiolysis [20, 21] show that in the Bragg
peak the secondary electrons and hydroxyl radicals yields prevail at early stage of radiolysis
(time 10−12 s). After the time 10−6 s the reactions OH•+OH•
→H2O2+O2
occur, and the yields of molecular products become larger than other yields [20]. Taking into
consideration that the microsecond time scale is typical for biological processes, the equilibrium
composition of the intracellular medium after the passage of high energy ions must be enriched
by the molecular products of water radiolysis.
→H2O2, HO•
2+HO•
2
To analyze the composition of the radiolysis products at the Bragg peak the Monte Carlo
simulation data [20] are used. The distribution of radiolysis products at the time 1 µs after ion
passage are showed in the Figure 1. It is seen that the number of hydrogen peroxide molecules
in intracellular medium is larger than other products of water radiolysis. The number of H2
molecules is also considerable, but these molecules don't have high reactivity and as a result do
not threaten biological macromolecules. Radicals which also have substantial yields after some
time transform into neutral water molecules or hydrogen peroxide. Thus, the composition of
intracellular medium changes after ion beam irradiation and significant amounts of hydrogen
peroxide molecules appear.
3
Table 1: Parameters of potentials.
Atomic charges in molecules
H
O
Molecule
H2O
H2O2
PO−
4
+0.33 (e)
+0.41 (e)
-
-0.66 (e) q (e)
-0.41 (e) r0 (A)
b (A)
-0.50 (e)
Potential (4) [29]
Parameter Na+ Parameter
Potential (3) [24, 25]
+1.00 Aij (A6kcal/mol)
2.35 Bij (A12kcal/mol) 12900 31300
0.3
86
O -- O O -- H
200
The concentration of hydrogen peroxide in the living cell is kept constant by special fer-
ments [22], but it may be increased greatly by the action of ion beams. Increasing the amount
of hydrogen peroxide molecules raises the probability of their interaction with DNA. The inter-
action of large number of H2O2 molecules and DNA can deactivate the cell genetical apparatus
and induce cell death. Therapeutic effect of the hydrogen peroxide has been found under the
treatment of cancer disease (see for example [23]). Ion therapy allows to increase the concen-
tration of hydrogen peroxide in tumor cells with high precision and with minimal damage of
healthy tissues.
The hydrogen peroxide molecule can interact with DNA molecule via charged atomic groups
of the double helix. It is well known that the phosphate groups of double helix backbone bear
the highest charge [7]. Under the natural conditions it is neutralized by metal ions (for example,
Na+ or K+). Thus, the interaction of H2O2 molecule with the phosphate groups of the double
helix and tethered metal counterions is expected to be the most probable.
3 Model and method of energy calculation
The capability of hydrogen peroxide to interact with DNA is studied by calculating the energy
of different complexes. The considered complexes consist of DNA phosphate group, hydrogen
peroxide molecule, water molecule and Na+ counterion. For the determination of relative
stability of different complexes two oxygen atoms of the phosphate group are sufficient to
consider. Other atoms of PO−
4 are essential for accurate calculations of energy. Schematic
constructions and structural parameters of molecules in complexes are showed at Figure 2.
The method of atom-atom potential functions is used for calculation of energies of complexes
[24, 25]. Within the framework of this method the energy is presented as the sum of pair
interactions between atoms which belong to different structure elements. The energy of complex
O
0.9584 Å
H
°
5
.
4
0
1
H
H
0.95 Å
2.26 Å
O
94.8o
1.474 Å
O
111.5 o
H
O
Å
1
5
.
2
O
P
O
O
Figure 2: Schematic structure of hydrogen peroxide molecule, water molecule, and phosphate.
Characteristic values of bond lengths, valence angle and dihedral angles are showed.
4
can be presented as the sum of two components:
U = Xij
[Uel(rij) + Uvdw(rij)] ,
(1)
where Uel(rij) and Uvdw(rij) are the potential energy of electrostatic and van der Waals inter-
actions between i and j atoms at the distance rij.
Electrostatic interaction is defined as follows:
Uel(rij) =
qiqj
4πεε0rij
,
(2)
where qi and qj are the charges of i and j atoms, ε0 is the dielectric constant of vacuum, ε is the
dielectric constant of medium. Charges of atoms of hydrogen peroxide and water molecules are
determined using the values of dipole moments 2,1 D (H2O2) and 1,86 D (H2O) [26]. Charges
of atoms in phosphate group are taken only for oxygen atoms in order to equate total charge
with -e (Table 1). The value of dielectric constant ε = 1 is used in calculations.
In the case of van der Waals interaction the Lennard-Jones potential is used:
Uvdw(rij) = −Aijr−6
ij + Bijr−12
ij
,
(3)
where Aij and Bij are the constants (Table 1). The first and the second terms in (3) describe
the van der Waals repulsion and short range repulsion between atoms, respectively. To take
into consideration the hydrogen bonds (O -- H· · ·O) the interaction between respective oxygen
atoms are calculated using the modified potential
[25], where exponent in the first term of
equation (3) is changed to -10. Such exponent describes the short-range character of hydrogen
bonds.
For the description of interaction between ion and charged oxygen atoms the Born-Mayer
potential is used:
UBM (rij) =
qiqj
4πεε0rij (cid:20)1 −
brij
r2
0
exp(cid:18) r0 − rij
b (cid:19)(cid:21) ,
(4)
where b is the repulsion constant, r0 is the equilibrium length (Table 1). This potential was
developed to describe the ion crystal energy, and it takes into account electrostatic attraction
and short-range repulsion [27]. The potential (4) has been adapted in [28,29] for the description
of counterion interactions with the phosphate groups of DNA double helix backbone.
Using the formulae (1) -- (4) with parameters in the Table 1 the energies of different com-
plexes may be calculated.
4 Complexes of hydrogen peroxide with DNA
Within the framework of proposed method the energies of two-component (H2O2 -- PO−
PO−
PO−
4 , Na+ -- PO4, H2O2 -- Na+, H2O -- Na+) and three-component (H2O2 -- Na+ -- PO−
4 ) complexes are calculated. The most stable configurations for each complex are found.
Two-component complexes. The most stable state for the complex Na+ -- PO−
4 is obtained in
case when sodium ion is localized at the distance 2 A from the center of O -- O segment of the
phosphate group. The calculated anergy is the lowest among two-component complexes (Fig.
3). This complex corresponds to neutralization of DNA phosphate group by counterion under
the natural conditions.
4 , H2O --
4 , H2O -- Na+ --
5
+
Na
2 Å
O
O
P
O
O
-122 kcal/mol
2,7 Å
Na+
O
O
H
HH
O
O
O
H
2,9 Å
HH
O
P
-14 kcal/mol
-11 kcal/mol
2,7 Å
Na+
O
H
H
2,9 Å
O
H
H
O
O
P
-12 kcal/mol
-12 kcal/mol
O
O
O
O
Figure 3: The structures of two-component complexes: Na+ -- PO4, H2O2 -- Na+, H2O -- Na+,
H2O2 -- PO4, H2O -- PO4 .
The most stable complex of sodium ion with hydrogen peroxide is obtained in the case of
ion localization at distance 2.66 A to the center of O -- O bond with equal distances to H atoms
on the opposite side of the molecule (Fig. 3). In the case of Na+ -- H2O the most stable complex
is obtained for the Na+ -- O distance 2.74 A. Under the natural conditions water molecules in
complex with Na+ are the part of ion hydration shell. The hydrogen peroxide molecule may
substitute the water in the ion hydration shell.
The equilibrium configurations of H2O2 -- PO−
4 complex is obtained in the case of symmetrical
localization of H atoms of hydrogen peroxide molecule with respect to the O -- O segment in the
phosphate group (Fig. 3). The complex has minimal energy for the distance 2.85 A between
centers of O -- O segments in the molecule and in the phosphate group. In the case of H2O -- PO−
4
complex the equilibrium configuration is obtained for the distance 2.9 A of water oxygen to
O -- O segment in the phosphate group (Fig. 3).
The hydrogen bonds, which can be formed between H (water or hydrogen peroxide) and O
(phosphate group), do not influence significantly the energy values. In the case of the strongest
hydrogen bond (O -- H -- O atoms are localized on the line) the calculated energies are lower
than energies of H2O2 -- Na+ -- PO4 and H2O -- Na+ -- PO4 complexes. In configurations with two
hydrogen bonds between atoms of H2O2 (H2O) molecule the O -- H -- O segment is bent more than
30◦, that makes such hydrogen bonds very weak.
The equilibrium configurations of H2O2 -- PO−
4 and H2O -- PO4 complexes depend on mutual
disposition of molecules. Rotating H2O2 molecule with respect to the axis passing through the
center of O -- O segment, the most stable complex is found for the angle about 60◦ (energy -12.8
kcal/mol) (Fig. 4a). In the case of molecule rotates around O -- O bond, the energy minimum is
obtained for the angle 0◦, while at 180◦ the repulsion occurs (Fig. 4b). Some small minimums
at about the angles 70◦ and 290◦ appear due to the van der Waals repulsion and Coulomb
attraction (dash lines at figure). In the case of rotation of water molecule the obtained results
are qualitatively similar.
Three-component complexes. The calculations are carried out for H2O2 -- Na+ -- PO−
4 and
H2O -- Na+ -- PO−
4 complexes. The energy of the complexes are calculated for different distances
between hydrogen peroxide (water) molecule and ion-phosphate complex. The ion-phosphate
6
a
)
O
O
H
HH
O
O
P
O
O
-10
-11
)
l
o
m
/
l
a
c
k
(
y
g
r
e
n
E
-12
-13
b
)
15
10
5
0
-5
-10
-15
O
O
H
HH
O
O
P
O
O
vdW
Coulomb
Total
0 60 120 180 240 300 360
0 60 120 180 240 300 360
Rotation angle (degree)
Rotation angle (degree)
4 complex as a function of rotation angle of H2O2 molecule.
Figure 4: Energy of H2O2 -- PO−
a) The dependence of energy on rotation angle with respect to the axis passing through the
centers of O -- O segments in H2O2 and PO4 (is showed in inset). b) The dependence of energy
on rotation angle with respect to the O -- O bond of H2O2 molecule (is showed in inset). The
van der Waals and Colomb contributions to the energy are showed by dashed lines.
distance is taken equal to 2 A, correspondingly to the result for two-component complex Na+ --
PO4 (Fig. 3). Such ion localization is the most energetically favorable, therefore other variants
of ion localization are not considered in the present work.
The configurations where hydrogen atoms of H2O2 (H2O) molecule are turned toward and
outward phosphate group are considered (Fig. 5a). In the first case there is no energy minimum
due to the electrostatic repulsion between hydrogen atoms and sodium ion.
Increasing the
distance to H2O2 (H2O) the energy of the complex monotonously decreases to the value of
Na+ -- PO4 complex (Fig. 5b). In the case when H atoms of H2O2 (H2O) molecule are turned
toward the phosphate group the energy minimum 130.3 kcal/mol (129.2 kcal/mol) appear at the
distance 2.75 A (2.82 A). The minimum is conditioned by the electrostatic attraction between
ion and oxygen atoms of molecule. Rotation of hydrogen peroxide or water molecules around
axis which passes through center of O -- O segment plays insignificant role because the total
energy depends on interaction between ion and oxygen atoms mainly. Thus, the stability of
H2O2 -- Na+ -- PO4 and H2O -- Na+ -- PO4 complexes are almost equal.
5 Vibrations of hydrogen peroxide in DNA low-frequency
Raman spectra
The formation of stable complexes of H2O2 molecule with DNA phosphate groups should be
accompanied by the appearance of specific modes of vibrations of hydrogen peroxide molecule in
the vibrational spectra. The frequencies of vibrations of H2O2 -- Na+ -- PO−
4 complex are expected
to be in the low-frequency spectra range (<200 cm−1), analogically to the modes of counterion
vibrations with respect to DNA phosphate groups (ion-phosphate modes) [28 -- 31]. The low-
frequency spectra of DNA is known to be characterized by the conformational vibrations of the
double helix structural elements as whole (nucleosides, nucleotides, phosphate groups) [32 -- 35].
7
a)
(1)
H
HH
(2)
r
2 Å
Na
O
O
r
2 Å
Na
O
H
H
O
O
O
O
P
P
O
O
O
O
r
2 Å
Na
(3)
O
H
H
(4)
O
O
H
r
HH
2 Å
Na
O
O
O
O
P
P
O
O
O
O
b)
-100
l
o
m
/
l
a
c
k
,
U
-110
-120
-130
-1
822.
29 2
. kcal/mol
Å
-130.3 kcal/mol
2.75 Å
2
4
6
r, Å
(1)
(2)
(3)
(4)
8
10
Figure 5: Energy of three-component complexes. a) Position of H2O2 and H2O molecules in
the considered complexes. b) Energy of the complex as a function of distance between H2O2
or H2O molecules to the ion-phosphate complex.
8
The vibrations of hydrogen peroxide molecule with respect to the DNA phosphate groups should
disturb the internal dynamics of the double helix changing the low-frequency spectra shape.
To determine the influence of H2O2 molecules on DNA conformational vibrations the phe-
nomenological model describing the Raman low-frequency spectra of DNA with counteri-
ons [28 -- 30] is used. Within the framework of this model the DNA double helix is repre-
sented as the double chain of nucleotides with counterions tethered to the phosphate groups
of the macromolecule backbone. To take into consideration the influence of hydrogen peroxide
molecule this model is modified by increasing the mass of tethered counterions for the mass
of one H2O2 molecule. The low-frequency vibration spectra of DNA with counterions is also
calculated for the case of H2O molecule attached to the ion. The equations, necessary for the
calculations of the mode frequencies and Raman intensities, are the same as in the original
works [29, 30]. The calculation parameters are taken for the case of B -form of the double helix.
As the result the low-frequency Raman spectra of DNA with attached H2O2-Na+ complex
is calculated as well as the spectra of DNA with H2O-Na+ complex (Fig. 6). The internal
vibrations of the double helix (lower than 120 cm−1) are connected with the vibrations of H-
bond stretching in nucleic bases and the vibrations in nucleosides due to the flexibility of sugar
ring. These modes do not sense essentially the presence of hydrogen peroxide molecule. At the
higher frequencies the modes, characterizing the vibrations of the complexes with respect to
the phosphate groups, are determined about 155 cm−1 and 140 cm−1, respectively. The mode
intensity is higher in the case of complex with hydrogen peroxide molecule. Thus, our estimates
show that hydrogen peroxide molecule, attached to DNA phosphate groups, may induce the
visible changes in the low-frequency Raman spectra.
DNA + Na + H O2
+
DNA + Na + H O2
2
+
O
Na+
H
H
O
O
P
O
O
H
HH
Na+
O
O
O
O
P
O
O
DNA modes
H-bond
H-bond
+
Sugar
Sugar
H O2
2+ion
H O2
+
Ion
2
1
y
t
i
s
n
e
t
n
i
e
v
i
t
a
l
e
R
50 100 150
Frequency (cm )-1
Figure 6: The low-frequency Raman spectra of DNA with H2O2 -- Na+ complex (solid line) and
with H2O -- Na+ complex (dashed line). In the insets the structure of respective complexes are
showed.
9
6 Discussion
The calculations show that the complexes including hydrogen peroxide and water molecules
are the most stable under the presence of sodium counterion. Without Na+ the stability of
the complexes decreases for about one order. That is due to the dominant role of electrostatic
interactions in the complex formation. The stability of H2O2 -- Na+ -- PO−
4 is about the same
as H2O -- Na+ -- PO−
4 . Therefore, hydrogen peroxide molecule may be expected to incorporate
into the hydration shell of sodium counterion already tethered to the oxygen atoms of DNA
phosphate group.
The formation of stable complex of H2O2 -- Na+ with the phosphate group may be detected
by the low-frequency Raman spectra of DNA. According to our estimations the vibrations of
counterion with hydrogen peroxide molecule with respect to the phosphate group of the double
helix backbone should be about 10 cm−1 lower than the frequencies of vibrations of the complex
with water molecule. Taking into consideration the determined frequency shift the lifetimes of
H2O2 -- Na+ -- PO−
4 complexes may be compared. According to the Arrhenius
equation the ratio between characteristic lifetimes may be written in the following form:
4 and H2O -- Na+ -- PO−
τ1
τ2
=
ω2
ω1
exp(cid:18) E1 − E2
kT (cid:19) ≈
ω2
ω1
,
(5)
Due to the larger lifetime of H2O2 -- Na+ -- PO−
where τ1, ω1 and τ2, ω2 are the lifetimes and frequencies of the complex with water and hydrogen
peroxide molecules, respectively; E1 and E2 are the energies of the complexes with hydrogen
peroxide and water molecules, respectively. The energies of the considered complexes are rather
close (Fig. 5), therefore the value of the energy difference can be equal to zero. As the result the
expected lifetime of the complex with H2O2 is higher than in the complex with H2O molecule.
4 complex the hydrogen peroxide molecules may
be accumulated near DNA double helix. At significant amounts the H2O2 molecules can influ-
ence DNA biological functioning by two mechanisms. Firstly H2O2 can induce double strand
breaks in double helix backbone by decaying into radicals, and secondly H2O2 molecules may
break nucleic-protein recognition, stopping the processes of translation of genetic information.
Thus, the hydrogen peroxide molecule, interacting with DNA double helix, can deactivate cell
genetic apparatus and induce cell death.
7 Conclusions
The hydrogen peroxide molecules appear in the cell medium after the passage of high-energy
ions. To study the possible effect of H2O2 molecule in the functioning of the living cell the
complexes of hydrogen peroxide molecules with the phosphate groups of DNA backbone has
been considered. Using the method of atom-atom potential functions the stability of different
complexes, consisting of DNA phosphate group, H2O2 (or H2O) molecule, and Na+ metal ion,
have been studied. The results show that the complexes with H2O2 molecule and phosphate
group of DNA are no less stable than respective complexes with H2O molecule. The sodium
counterion neutralizing the charge of phosphate group plays the key role in stabilization of these
complexes. For the case of most stable complexes of hydrogen peroxide with the phosphate
groups the low-frequency Raman spectra of DNA have been calculated. The vibrations of
the complex as whole with respect to DNA phosphate group has been found near 150 cm−1.
The observation of this mode may be considered as a test revealing the complex formation.
10
4 and H2O -- Na+ -- PO−
Taking into consideration the determined frequencies of vibrations the lifetimes of H2O2 -- Na+ --
PO−
4 complexes have been compared, and the lifetime of the complex with
hydrogen peroxide has been found higher than in the case of water molecule. Taking this
into consideration the hydrogen peroxide molecules are expected to be accumulated near DNA
macromolecule. The action of H2O2 on DNA structure and dynamics may be sufficient for
the braking the processes of DNA biological functioning. The influence of hydrogen peroxide
molecules on DNA can be regarded as an individual channel in heavy ion therapy.
References
[1] L. Gravitz, Nature Outlook 491, (2012) 7425.
[2] A. Brown, S. Herman, Radiology and Oncology 73, (2004) 265-268.
[3] G. Kraft, Progress in Particle and Nuclear Physics 45, (2000) S473-S544.
[4] H. Suit, et all., Radiotherapy and Oncology, 95, (2010), 3-22.
[5] C.D. Schlaff, A. Krauze, A. Belard, J.J. O'Connell, K.A. Camphausen, Radiation Oncology
9, (2014) 88.
[6] N. W. Timofeeff-Ressovsky, A. V. Savich, and M. I. Shal'nov, Introduction to Molecular
Radiobiology: Physico-Chemical Foundations (Medicina, Moscow, 1981) 320. [In Russian]
[7] W. Saenger, Principles of Nucleic Acid Structure (Springer, New York, 1984) 584.
[8] B. Boudaoiffa, P. Cloutier, D. Hunting, M.A. Huels, L. Sanche, Science, 287, (2000) 1658-
1660.
[9] N. Hamada, J. Radiat. Res., 50, (2009) 1-9.
[10] A.V. Solov'yov, E. Surdutovich, E. Scifoni, I. Mishustin, W. Greiner, Phys. Rev. E, 79,
(2009) 011909.
[11] A.V. Yakubovich, E. Surdutovich, A.V. Solov'yov, Nuclear Instruments and Methods in
Physical Research B, 279, (2012) 135-139.
[12] E. Surdutovich, A.V. Yakubovich, A.V. Solov'yov, Scientific Reports, 3, (2013) 1289.
[13] E. Surdutovich, A.V. Solov'yov, J. Phys.: Conf. Series, 438, (2013) 012014.
[14] I. Pshenichnov, A. Botvina, I. Mishustin, W. Greiner, Nuclear Instruments and Methods
in Physics Research, B 268, (2010) 604-615.
[15] E. Haettner, H. Iwase, D. Schardt, Radiation Protection Dosimetry 122 (2006) 485-487.
[16] J. Soltani-Nabipour, M A. Popovici, Gh. Cata-Danil , 62, (2010) 37-46.
[17] B. Pastina, J.A. LaVerne, J. Phys. Chem. A 103, (1999) 1592-1597.
[18] S. Le Caer, Water 3, (2011) 235-253.
11
[19] V. Wasselin-Trupin, G. Baldacchino, S. Bouffard, B. Hickel, Radiation Physics and Chem-
istry, 65, (2002) 53-61.
[20] M.S. Kreipl, W. Friedland, H.G. Paretzke, Radiat Environ Biophys, 48, (2009) 11-20.
[21] S. Uehara, H. Nikjoo, J. Radiat. Res., 47, (2006) 69-81.
[22] J.D. Watson, Molecular Biology of the Gene (W.B. Benjamin. Inc., Menlo Park, 1978)
706.
[23] G. Manda, M.T. Nechifor, T.M. Neagu, Current Chemical Biology, 3, (2009) 342-366.
[24] V.V. Zhurkin, V.I. Poletaev, V.L. Florentiev, Molecular Biology (Moscow). 14, (1980)
1116-1130.
[25] V.I. Poltev, N.V. Shuliupina, Molecular Biology (Moscow). 18, (1984) 1549-1561.
[26] Brief Chemical Encyclopedia. V. 1. (Soviet Encyclopedia, Moscow, 1961). [In Russian]
[27] C. Kittel, Introduction to Solid State Physics (John Wiley and Sons, Inc., New York, 1954)
696.
[28] S.M. Perepelytsya and S.N. Volkov, Ukr. J. Phys., 49, (2004) 1074-1080.
[29] S.M. Perepelytsya and S.N. Volkov, Eur. Phys. J. E, 24, (2007) 261-269.
[30] S.M. Perepelytsya and S.N. Volkov, Eur. Phys. J. E, 31, (2010) 201-205.
[31] L.A. Bulavin, S.N. Volkov, S.Yu. Kutovy and S.M. Perepelytsya, Reports of National
Academy of Science of Ukraine, No. 11, (2007) 69-73. arXiv:0805.0696.
[32] S.N. Volkov, A.M. Kosevich, Mol. Biol. 21, (1987) 797-806. [Moscow]
[33] S.N. Volkov, A.M. Kosevich, G.E. Weinreb, Biopolimery i Kletka 5, (1989) 32-39. [Kiev]
[34] S.N. Volkov, A.M. Kosevich, J. Biomolec. Struct. Dyn. 8, (1991) 1069-1083.
[35] A.M. Kosevich, S.N. Volkov,
in Nonlinear Excitations in Biomolecules, edited by M.
Peyrard (Springer, 1995) 118-128.
12
|
1501.05886 | 1 | 1501 | 2015-01-23T17:40:15 | The fractal structure of the ventral scales in legless reptiles | [
"physics.bio-ph"
] | Surface constructs in snakes reflect desirable design traits for technical surface engineering. Their micro-textural patterns, however, do not lend themselves easily to unified analysis due to species-specific variations in surface geometry and topology. Fractal description is useful in this context since it accentuates the correspondence between patterns especially when responding to tribological phenomena. In this work we examine the surface construction of 14 snake species, representing five families, and evaluate the fractal dimension for each of the skins (both the dorsal and ventral sides) using three different computational algorithms. Our results indicate first that all of the examined species share a common fractal dimension (with a very small variation between species in the order 4-5%). This finding implies that despite the different micro-geometry of texture among species, the skin as a unit responds in a similar manner to many interfacial influences. | physics.bio-ph | physics | The fractal structure of the ventral scales in legless reptiles
H A Abdel-Aal1 M El Mansori2
1Department of Mechanical Engineering, Drexel University, 3141 Chestnut St, Philadelphia,
Pennsylvania, USA, 19144.
E-mail: [email protected]
2Mechanics, Surfaces and Materials Processing Laboratory (MSMP), Arts et Métiers
ParisTech, 2 cours des Arts et Métiers, 13617, Aix-en-Provence, France
Abstract. Surface constructs in snakes reflect desirable design traits for technical surface
engineering. Their micro-textural patterns, however, do not lend themselves easily to unified
analysis due to species-specific variations in surface geometry and topology. Fractal
description is useful in this context since it accentuates the correspondence between patterns
especially when responding to tribological phenomena. In this work we examine the surface
construction of 14 snake species, representing five families, and evaluate the fractal dimension
for each of the skins (both the dorsal and ventral sides) using three different computational
algorithms. Our results indicate first that all of the examined species share a common fractal
dimension (with a very small variation between species in the order 4-5%). This finding
implies that despite the different micro-geometry of texture among species, the skin as a unit
responds in a similar manner to many interfacial influences.
Introduction
In the natural world, emphasis on micro-texture arises in response to the severe demands on
efficiency, economy of effort, and on survival. In legless locomotion, not only the reptile has to
economize its effort when moving, but also it has to utilize friction (dynamic or static) to generate
tractions, balance motion, and to initiate propulsion [1-7]. These demands lead to a skin made
essentially of the same chemical elements, yet with different local concentrations to meet local
functional requirements (i.e., a skin with inherited compositional heterogeneity and added
complexity). Frictional surfaces in snakes employ micro-texture as a means to compensate for this
heterogeneity, and thereby manipulate system complexity, to optimize function. In this context pattern
and distribution of topographical features work to facilitate the manipulation of system complexity.
General composition of surfaces within legless locomotors, entails textural elements of primitive
geometry. The pattern, size, structural density, and spatial arrangement of these elements work
holistically to facilitate multitasking. Deliberate spatial arrangement rather than the geometry of the
textural elements appear to be of more prominence in engineering of these surfaces. Reduction of
1 To whom any correspondence should be addressed.
complexity through micro-texture takes place through two main mechanisms. The first is repetition, of
micro-textural features of simple geometries and strategic spatial placement of features in a manner
compatible with the multifunctional requirements of the surface. Adopting this elementary design
philosophy is clearly desirable within the technological realm. However, there is a need to understand
the sense of pattern formation, and topological distribution, within natural surfaces in light of the
design and functional imperatives on the reptile. A point of entry to this endeavour is to study the
unifying characteristics among the various textural patterns observed in snake surfaces. Micro-textural
patterns however, do not lend themselves easily to the proposed unified analysis due to species-
specific variations in surface geometry and topology.
Due to the variations in geometry and other parameters, a unifying design concept may not be
directly apparent. This hinders probing the surface to deduce a unified design principle. One remedy to
this problem is to view the geometry, and topology, of a natural surface within a different frame of
reference. This frame of reference should allow the abstract description of the surface (thus
neutralizing the effect of variation in conventional geometry and associated effects). One powerful
method useful in this context is fractal description.
Fractals efficiently describe the geometry of regular shapes that are not amenable to conventional
description through Euclidean geometry. In particular, fractals describe how much space an irregular
object occupies [8]. The structure of the fractal number is indicative of the nature of the object
described. The higher the fractal dimension, the more space the curve occupies. That is, the fractal
dimension (FD) indicates how densely a phenomenon occupies the space in which it is located.
Furthermore, a fractal dimension is a unique number that is independent of the alteration of space
through manipulation by stretching or compacting.
Two surfaces of the same fractal dimension do not necessarily manifest identical microstructure.
Rather, they display a common attitude in their dependence on their roughness to occupy a three
dimensional space. By focusing on the growth of form, and its evolution in space, fractal description
of an object shifts the focus from devising the Euclidean geometry of an object to describing the
evolution of its growth into its prescribed space. That is, fractal description of a surface deals with the
evolution of the process of surface growth into the space rather than finding its boundaries with
respect to an origin. As such, a fractal frame of reference projects differences within the micro-textural
elements, comprising the surface, in a view that unmasks their common phenomenal traits. This
accentuates the correspondence between behavioural patterns, of apparently dissimilar Euclidean
form, especially when responding to tribological phenomena (e.g. contact forces or diffusion through
surface topography). It follows that two surfaces sharing the same fractal dimension, in principle,
should manifest common dynamic responses to interfacial events. To reveal the sense of pattern in
micro-texture of reptilian skin, it is, therefore, essential to explore the fractal structure of the respective
surface construction.
In this work, we attempt to find the common attributes of micro-textural patterns in snakeskin
through examination of topology within a fractal frame of reference. To this effect, we examine the
surface construction of 14 snake species representing five families of snakes to determine their fractal
dimension (both the dorsal and ventral sides). We further, use the information extracted from the skins
to point out the common behavioural aspects that should persist across species (especially with respect
to locomotion and frictional response).
2.0 Materials and methods
Species
2.1
Allometry of size guided the choice of the species examined in this work. In choosing the species, we
attempted to diversify our choice of species to reflect the wide variation of size and length in snakes.
To this end, we examined the exuviae of the snakes shown in Table 1 (n=5). Table 2, meanwhile,
provides a summary of the basic dimensions (mass and length) and ventral scale counts of the species.
Bitis
Echis
Table-1 Summary classification and taxonomy of species examined in the current work
Mass
(g)
1306
142
256
227
950
755
1413
8144
Species:
Subfamily: Genus:
Family:
B. gabonica
Bitis
Viperinae
Bittis Gabonica
Viperdae Viperinae
E. Carinatus
Echis Carinatus
Viperidae Crotalinae Agkistrodon A. bilineatus
Agkistrodon Taylori
Cerrophidion Godmani Viperidae Crotalinae
Cerrophidion C. godmani
Agkistrodon Contortrix Viperidae Crotalinae Agkistrodon A. contortrix
Trimeresurus tokarensis Viperdae Crotalinae
Trimeresurus T. Tokarensis
Agkistrodon piscivorus Viperidae Crotalinae Agkistrodon A. piscivorus
Pseudechis Austarlis
Gongylophis colubrinus Boidae
Epicrates Cenchria
Boidae
Pituophis Melanoleucus
Thamnophis Sirtalis
L G. Californiae
Erycinae
Boidae
Colubrinae
Colubridae Natricinae
Colubridae Lampropeltis L. Getula
Elapidae
P. australis
Pseudechis
Gongylophis G. colubrinus
Epicrates
Pituophis
Thamnophis
E. Cenchria
P. melanoleucus
Length
cm
150
72
87
84
135
125
154
275
91
92
154
130
135
200
150
562
1414
848
950
3113
1306
Morelia Virdis
Python Regius
Pythonidae
Pythonidae
Morelia
Python
M.Virdis
P. regius
2.2 Characterization and computation of the fractal dimension
All examined skin underwent three imaging procedures to extract detailed information about general
dimensions, geometrical proportions, and essential metrological parameters. The procedure started by
pre-identifying twelve equally spaced locations along the Anterior Posterior (AP) axis of each skin.
SEM images, WLI-interferograms, and AFM scans were obtained for the dorsal and ventral scales at
each location.
This work used three methods to compute the fractal dimension of the examined skins: Cube
counting [9-11], Triangulation [9], and the Variance method [11, 12]. For each species, we extracted
the fractal dimension from SEM images AFM scans in topography mode, and WLI-images taken at
twelve predetermined areas on the Dorsal and Ventral sides of the skins. Five sets of images were
taken randomly at different spots within each area for each skin side (dorsal and ventral). An image set
consisted of an array of SEM imagery at various magnifications (250, 1000, 2500, 5000, 15000, and
50000 X), AFM scans of square areas (of sides 50, 25, and 10 m respectively), and WLI-imagery that
examined areas of dimensions 1250m by 500 m. To each of the image sets we applied the three
computational algorithms to extract the fractal dimension. In all, for each species, we analyzed 150
SEM images, 15 AFM scans, and 25 WLI-images for each predetermined spot using each of the three
computational algorithms to estimate the respective fractal dimension.
20
40
60
80
100
120
b
P.R.V
Trailing half
30 µm
0
0
50
µm
3.5
3
2.5
2
100
1.5
1
150
0.5
200
0
100
200
300 µm
75 µm
b
0
10
20
30
40
50
60
70
80
90
µm
35
30
25
20
15
10
5
0
N M
Figure 1 Representative imagery used to determine the fractal dimension of the skin. First raw (left to
right) shows SEM imagery of the ventral scales of three snakes: C. Godmani, P. regius, and A.
Piscivorus (X=5000). Second raw depicts white light interferograms for the same species. Third raw
depicts AFM scans for the same species at different scan sizes.
3.0 Results and discussion
3.1 Results
Figure 2 (a and b) examines the variation in the fractal dimension (FD) at different zones within the
body of two snake species P. regius and B. gabonica.
3.0
3.0
Python Rigius
a
b
Bittis Gabonica
2.5
2.0
1.5
i
n
o
s
n
e
m
d
i
l
a
t
c
a
r
F
1.0
0.5
f
l
a
h
t
n
o
r
f
e
l
a
c
s
l
a
r
t
n
e
V
k
c
a
b
s
e
l
a
c
s
l
a
r
t
n
e
V
d
a
e
H
i
f
o
e
d
s
d
n
a
H
t
h
g
R
i
i
n
k
s
f
o
e
d
s
n
i
l
I
s
e
a
c
s
l
a
r
t
n
e
V
e
d
s
i
l
a
r
t
n
e
V
l
i
a
T
i
n
o
s
n
e
m
d
i
l
a
t
c
a
r
F
2.5
2.0
1.5
1.0
0.5
f
l
a
h
t
n
o
r
f
l
a
r
t
n
e
V
f
l
a
h
t
n
o
r
f
l
a
s
r
o
D
f
l
a
h
g
n
i
l
i
a
r
t
l
a
s
r
o
D
f
l
a
h
g
n
i
l
i
a
r
t
l
a
r
t
n
e
V
f
l
a
h
g
n
i
l
i
a
r
t
l
a
s
r
o
D
f
l
a
h
t
n
o
r
F
l
a
s
r
o
D
)
n
o
i
t
c
e
s
-
d
m
i
(
k
n
u
r
T
l
a
s
r
o
D
0.0
Figure 2 Comparison between local fractal numbers (dimensions) of the snakeskin examined in the
current work. a- local fractal dimension of P. regius, b comparison of the fractal dimension of
representative spots located on the ventral and dorsal sides of B. gabonica.
Figure 2-a, depicts the fractal dimension for four ventral spots and for dorsal spots on the Python
species. The locations of these skin patches yield a fair representation of the fractal dimension within
the leading and the trailing halves of the reptile. A remarkable observation of the plot is that, on
average, the FD of all examined locations is almost equal (error percentage, however, differs by
location). This implies that the surface of the species has a unique fractal value. Figure 2-b, which
plots the fractal dimensions of four locations on the skin of B. Gabonica implies a different
arrangement. Here the FD shows visible variation between examined body zones (and not sides i.e
dorsal Vs ventral). The FD is equal for the ventral and dorsal skins on the front half of the body
(DFH≈2.5). A similar trend is observed for the trailing half of the body, however the FD differs from
that of the leading body half (DTH ≈ 2.765). For comparison, DFH for B. Gabonica is almost equal to
the average for P. regius (arithmetic mean of all values shown in figure 2-a).
3.0
2.5
2.0
1.5
s
u
t
a
n
i
r
a
C
s
i
h
c
E
Vipers
i
n
a
m
d
o
G
n
o
d
i
i
h
p
o
r
r
e
C
l
l
s
u
c
u
e
o
n
e
M
s
h
p
o
u
t
i
i
P
3.0
2.5
2.0
1.5
Pythons
Boidae
elapidae
a
i
r
h
c
n
e
C
a
i
r
h
c
n
e
C
s
e
t
a
r
c
i
p
E
l
s
u
n
i
r
b
u
o
c
s
i
h
p
o
l
y
g
n
o
G
s
i
d
r
e
V
a
i
l
e
r
o
M
s
i
l
i
a
r
t
s
u
A
s
h
c
E
u
d
e
s
P
i
s
u
g
e
R
n
o
h
t
y
P
i
D
n
o
s
n
e
m
D
i
l
a
t
c
a
r
F
i
s
s
n
e
r
a
k
o
T
s
p
o
r
h
t
o
b
o
t
o
r
P
i
r
o
l
y
a
T
n
o
d
o
r
t
s
i
k
g
A
x
i
t
r
o
t
n
o
C
n
o
d
o
r
t
s
k
g
A
i
s
u
o
r
i
i
v
c
s
P
n
o
d
o
r
t
s
i
k
g
A
a
c
i
n
o
b
a
g
s
i
t
i
B
Colubridae
n
e
i
n
r
o
f
i
l
a
C
.
G
.
L
Figure 3 Fractal Dimension for all snake
species examined in the current work grouped
by family. Figure 3-a depicts the FD for
Vipers, Figure 3-b depicts
the FD for,
Pythonidae, Elapidae, and Boidae, whereas,
figure 3-c plots the FD for members of the
family Colubridae.
)
e
k
a
n
S
r
e
t
r
a
G
(
s
i
l
i
a
t
r
i
S
s
h
p
o
n
m
a
h
T
i
D
n
o
s
n
e
m
D
i
l
a
t
c
a
r
F
3.0
2.5
2.0
1.5
i
D
n
o
s
n
e
m
D
i
l
a
t
c
a
r
F
Figure 3 (a-c) presents the average (arithmetic mean of the FD for all examined spots on a species)
for all snakes examined in the current work. Each of the figures, when possible, presents the FD for a
clad or family of snakes. Figure 3-a depicts the FD for Vipers, Figure 3-b depicts the FD for,
Pythonidae, Elapidae, and Boidae, whereas, figure 3-c plots the FD for members of the family
Colubridae. The data show that, in general, the FD is above 2.5 and close to three (2.55 ≤ D ≤ 2.75.
Two species, P. Australis (Colubridae and P. Tokarensis (Viperdae), deviate from this value (D ≤ 2.5).
The general trend of the data reveals a characteristic FD that distinguishes each genus examined.
For example, within the genus Agkistrodon (A. Taylori, A. Piscvirous, and A Contortex) the FD is
approximately 2.7, whereas for the genus Crotilanae, the FD falls around 2.55. This trend is more
visible among Colubrides which manifest a FDas high as 2.75. This value is also shared by the family
Boidae and the two python species examined.
3.2 Discussion
Discussion of the results will focus on implications for performance and surface construction rather
than metrological aspects. The later discussion will be a subject of future work.
Fractals stand for a method to describe the geometry of regular shapes that are not amenable to
conventional description through Euclidean geometry. In particular, fractals describe how much space
an irregular object occupies [13]. A typical description of a surface is a so called Fractal number
(dimension D). This number is a positive non-integer that varies between 1 ≤ D ≤ 3. The structure of
the fractal number is indicative of the nature of the object described. The most significant digit of the
parameter D, stands for the topological dimension of the object, whereas the second part, the fractional
part that varies from 0.0 to 0.999, is a so-called fractal increment. The higher the fractal dimension, the
more space the curve occupies. That is, the fractal dimension indicates how densely a phenomenon
occupies the space in which it is located.
A surface of which FD is close to 3 implies roughness that densely occupies its volume, and by so
doing, the roughness manifests complex branching. The complexity resulting from the manner of
filling the limited volume maximizes the pathways necessary for many functions (e.g. thermal
regulation). Thus in essence, a high FD enhancing efficiency of particular skin functions. A high FD,
close to three, may also imply a high ratio of surface area to volume. However the increase in area
does not take place through extending a feature but through fractal branching of bends and folds of the
skin. Physically, this indicates that the increased surface area is subdivided into quasi-independent
hierarchical domains, which potentially can display swarm-like behavior. Hierarchical construction of
textural features integrates the ability of the reptile to vary the contraction strength of ventral muscles
to form an active control mechanism of friction tractions. This mechanism essentially can vary the
propulsive forces necessary for motion and therefore controls the power requirements of the reptile.
An interesting finding of the analysis, figure 2-a, is the higher FD for the trailing body half of B.
Gabonica than that of the leading half. Such an increase indicates that the ratio o f the volume
occupied by the branching textural features is higher for the trailing half of the body which in turn
implies functional adaptation. B. Gabonica is a heavy ground dweller that inhabits regions with high
rain fall. Defense and feeding requirements mandate rigidity of the trailing half of the body. More
roughness is required, whence the higher fractal branching of texture features. On the other hand, the
almost uniform FD for P. regius implies uniform density of roughness branching which helps function
of the reptile within its arboreal habitat that may require the snake to cling to a tree trunk for example
with the entire body stretched. Uniform fractal branching contributes even loading on ventral muscles
(which contract to provide gripping forces) in resistance to gravitational tractions acting at the
interface of the body and the clinging medium.
Furthermore, a fractal dimension is a unique number that is independent of the alteration of space
through manipulation by stretching or compacting. Therefore, in essence, two surfaces that share a
fractal dimension, share also several intrinsic behavioral characteristics. Mainly, the two surfaces
share the manner by which they occupy their space (even if at first glance they appear fictitiously
different) along with the consequences for accommodation of loads and transportation flux.
The skin being a visco-elastic material exhibits stress relaxation (i.e., an applied stress will vanish
asymptotically to a zero state if the time duration of applying that stress is almost infinite (or
sufficiently long)). Roughness of the surface influences relaxation behavior of the skin. It is possible
to express the topology of the roughness in terms of the fractal dimension D. Within such formulation,
the value of the FD being indicative of roughness density, may also indicate whether bulk deformation
or roughness deformation would dominate the stress relaxation of the material and vice-versa. Any
time stress is greater for the smaller FD, at the same temperature and same strain, bulk deformation
will dominate. This is implied from the definition of the FD itself; lower fractal dimensions mean less
roughness density; consequently, the bulk material dominates, while a higher FD indicates denser
roughness and consequently asperity deformation dominates [14]. As such, for all snakes having close
values of the FD imply also that they also share similar stress relaxation behavior and whether or not
roughness deformation dominates such a process. The FD being very close to three implies the
dominance of roughness deformation. This result is important as it points at the similarity of
accommodating contact stresses applied to the skin (despite the apparent differences in ornamentation,
size, and geometry of textural features).
Fractal surface geometry significantly impacts functional performance of materials. Friction of
snakeskin is a function of contact stiffness. Bauem et. al., [15] reported that the coefficient of friction
(COF) of the California king snake L. G. Californae depends on the roughness of the substratum as
well as on the stiffness of the contact. In their experiments they reported that the COF of a rough glass
ball on the skin in the cranial direction was some 30% lower than the COF of a smooth sphere. The
variation was directional (for example, the lateral COF was some 12% higher for a rough sphere than
for a smooth sphere sliding at identical conditions). Stiffness of the underlying skin layers also affects
friction un-cushioned samples had considerably higher COF. Contact stiffness is a strong function of
the surface roughness of the skin, so the density of roughness can explain that counter-intuitive
behavior.
The FD of L. Californae is approximately 2.75 which implies dense roughness within the volume
allowed for the skin. Rougher solid surfaces are more compliant and penetrable, their properties
depend on surface morphological parameters. Denser roughness, implied by a high FD, causes a
decrease of stiffness. Buzio et. al., [16] report that an increase of the FD by 10% for example causes a
decrease in stiffness by one order of magnitude. The presence of denser roughness causes the
reduction of the contact stress as well, which reduces the area of true contact and causes a lower
friction force (thereby lower COF). The results of Pohrt and Popov [17] who studied the role of
fractal roughness on contact stiffness support the proposed role of roughness in relation to the fractal
structure of the studied skins.
Within classical contact models [18, 19] the apparent area of contact grows larger with added
roughness compared to a smooth Hertzian case. The contact radius is proportional to contact stiffness
in concentrated contacts. As such, within a classical formulation, one could expect the contact stiffness
to grow. However, Pohrt and Popov point that this behavior is true beyond a critical load (which is a
function of the Holder exponent and thereby the FD). For loads smaller than this transitional limit,
added roughness leads to a decrease in contact stiffness (compared to a smooth surface), whence larger
contact radius would lead to a lower contact area and thereby lower friction force.
A significant consequence of the common fractal dimension pertains to transdermal diffusion of
particles. Pazkossy and Nykos [20] studied the decay of the diffusion of a current of particles diffusing
from an initially homogenous medium to a completely absorbing fractal boundary. They found that
such a process, which is important in studying of drug administration and absorbance, exhibits a time
dependency of the form t-b as compared to the conventional time factor t1/2. The exponent in this
case is a direct function of the FD, and assumes the form FDThis corresponds to a
generalized Cottrell equation and may be used to describe the frequency dispersion induced by
roughness. Sharing a fractal dimension, therefore, indicates that all skins examined types have similar
diffusion decay kinetics. This typically explains the clinically observed permeation compatibility
between human and snake skins [21-23].
Conclusions
This work presented a comprehensive study of the fractal structure for snakeskin. Results indicated
that the fractal dimension of snakes occupies a narrow band (around 2.75 ± 0.05). This implies that the
skin, despite differences in shape and geometry of the micro-texture elements responds in a similar
manner to interfacial phenomena (e.g., tangential and contact loading).
References
[1] Arzt, E., Gorb, S., Spolenak, R. (2003). From Micro to Nano Contacts in Biological Attachment
Devices, Proceedings of National Academy of Sciences (PNAS), 100 (19), 10603-10606.
[2] Hazel, J., Stone, M., Grace, M.S., Tsukruk, V. V. 1998 Tribological design of biomaterial
surfaces for reptation motions. Polymer Preprints, 39, 1187-1188.
[3] Berthé, R. A., Westhoff, G., Bleckmann, H., Gorb, S. N., 2009. Surface structure and frictional
properties of the skin of the Amazon tree boa Corallus hortulanus (Squamata, Boidae) Journal of
Comparative Physiology A: Neuroethology, Sensory, Neural, and Behavioral Physiology, 195,
311–318.
[4] Abdel-Aal, H. A., El Mansori, M. 2013. Tribological analysis of the ventral scale structure in a
Python regius in relation to laser textured surfaces, Surface Topography and Metrological
Properties. 1 015001. doi:10.1088/2051-672X/1/1/015001
[5] Abdel-Aal, H. A. 2013. On Surface Structure and Friction Regulation in Reptilian Locomotion,
the Mechanical Behavior of Biomedical Materials 22:115-135, DOI
Journal of
10.1016/j.jmbbm.2012.09.014
[6] Gray, J., Lissemann, H. W., 1950. The kinetics of locomotion of the grass-snake. J. exp. Biol. 26:
354–367
American Zoologist, 23, 443–454.
[7] Pough, F. H., Groves, J. D. 1983. Specialization of the body form and food habits of snakes.
[8] D.S. Ebert, F. K. Musgrave, D. Peachey, K. Peril, S. Worley. 1998. Texturing and Modeling: A
Procedural Approach. Second Ed. Academic Press, San Diego.
[9] W. Zahn, A. Zösch: 1999 The dependence of fractal dimension on measuring conditions of
scanning probe microscopy. Fresenius J Analen Chem, 365: 168-172
[10] W. Zahn, A. Zösch: Characterization of thin film surfaces by fractal geometry. Fresenius J Anal
1064.
Chem (1997) 358: 119-121
[11] C. Douketis, Z. Wang, T. L. Haslett, M. Moskovits 1995 Fractal character of cold-deposited
silver films determined by low-temperature scanning tunneling microscopy. Physical Review B,
51, 16,
[12] P. G. de Gennes, Physique des surfaces et des interfaces, C. R. Acad. Sc. Paris 295 (1982) 1061–
[13] Osama M. Abuzeid and Taher A. Alabed 2009: Mathematical modeling of the thermal relaxation
of nominally flat surfaces in contact using fractal geometry: Maxwell type medium, Trib.
International, 42(2), 206-212.
[14] Baum M. J., Kovalev A. E., Michels J. and Gorb S. N. 2014 Anisotropic Friction of the Ventral
Scales in the Snake Lampropeltis getula californiae Tribology Letters 54, 2, 139-150.
[15] Renato Buzio, Corrado B, Biscarini F, Buatier De Mongeot F, Valbusa U, 2003, The contact
mechanics of fractal surfaces, Nature Materials 2, 233 - 236
[16] Pohrt R, Popov V, 2013 Contact Mechanics of Rough Spheres: Crossover from Fractal to
Hertzian Behavior Advances in Tribology, 2013. doi:10.1155/2013/974178
[17] J. A. Greenwood and J. B. P. Williamson, Contact of nominally flat surfaces, Proc of the Royal
[18] J. A. Greenwood, J. H. Tripp, 1967 The elastic contact of rough spheres, ASME J. Applied
Society A,. 295, 300–319, 1966
Mechanics, 34, 1, 153–159.
[19] Tamás Pajkossy, Lajos Nyikos, 1989 Diffusion to fractal surfaces—II. Verification of theory,
Electrochimica Acta, 34, 2, 171-179.doi./10.1016/0013-4686(89)87082-3.
[20] Bhatt, P.P., J.H. Rytting, E.M. Topp, 1991. Influence of azone and lauryl alcohol on the
transport of acetaminophen and ibuprofen through shed snake skin., Int. J. Pharm., 72: 219-226.
[21] Godin B , Touitou E 2007 Transdermal skin delivery: predictions for humans from in vivo, ex
vivo and animal models Adv Drug Deliv Rev 59 (11):1152-61
[22] Harada, K., T. Murakami, E. Kawasaka, Y. Higashi, S. Yamanoyo and N. Yata, , 1993. In vitro
permeability to salicylic acid of human, rodent and shed snake skin. J. Pharm. Pharmacol., 45:
414-418.
[23] Haigh, J.M., E. Beyssac, L. Chanet and J.M. Aiache, 1998. In vitro permeation of progesterone
from a gel through the shed skin of three different snake species. Int. J. Pharm., 170: 151-156.
|
1903.10800 | 2 | 1903 | 2019-06-13T12:27:27 | Hole-closing model reveals exponents for nonlinear degenerate diffusivity functions in cell biology | [
"physics.bio-ph",
"math.AP",
"q-bio.CB"
] | Continuum mathematical models for collective cell motion normally involve reaction-diffusion equations, such as the Fisher-KPP equation, with a linear diffusion term to describe cell motility and a logistic term to describe cell proliferation. While the Fisher-KPP equation and its generalisations are commonplace, a significant drawback for this family of models is that they are not able to capture the moving fronts that arise in cell invasion applications such as wound healing and tumour growth. An alternative, less common, approach is to include nonlinear degenerate diffusion in the models, such as in the Porous-Fisher equation, since solutions to the corresponding equations have compact support and therefore explicitly allow for moving fronts. We consider here a hole-closing problem for the Porous-Fisher equation whereby there is initially a simply connected region (the hole) with a nonzero population outside of the hole and a zero population inside. We outline how self-similar solutions (of the second kind) describe both circular and non-circular fronts in the hole-closing limit. Further, we present new experimental and theoretical evidence to support the use of nonlinear degenerate diffusion in models for collective cell motion. Our methodology involves setting up a 2D wound healing assay that has the geometry of a hole-closing problem, with cells initially seeded outside of a hole that closes as cells migrate and proliferate. For a particular class of fibroblast cells, the aspect ratio of an initially rectangular wound increases in time, so the wound becomes longer and thinner as it closes; our theoretical analysis shows that this behaviour is consistent with nonlinear degenerate diffusion but is not able to be captured with commonly used linear diffusion. This work is important because it provides a clear test for degenerate diffusion over linear diffusion in cell lines. | physics.bio-ph | physics | Hole-closing model reveals exponents for nonlinear degenerate diffusivity
functions in cell biology
Scott W McCuea,∗, Wang Jina, Timothy J Moroneya, Kai-Yin Lob, Shih-En Choub, Matthew J Simpsona
aSchool of Mathematical Sciences, Queensland University of Technology, Brisbane QLD 4000, Australia.
bDepartment of Agricultural Chemistry, National Taiwan University, Taipei 10617, Taiwan
9
1
0
2
n
u
J
3
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
0
0
8
0
1
.
3
0
9
1
:
v
i
X
r
a
Abstract
Continuum mathematical models for collective cell motion normally involve reaction-diffusion equations, such as
the Fisher-KPP equation, with a linear diffusion term to describe cell motility and a logistic term to describe cell
proliferation. While the Fisher-KPP equation and its generalisations are commonplace, a significant drawback for
this family of models is that they are not able to capture the moving fronts that arise in cell invasion applications
such as wound healing and tumour growth. An alternative, less common, approach is to include nonlinear degenerate
diffusion in the models, such as in the Porous-Fisher equation, since solutions to the corresponding equations have
compact support and therefore explicitly allow for moving fronts. We consider here a hole-closing problem for the
Porous-Fisher equation whereby there is initially a simply connected region (the hole) with a nonzero population
outside of the hole and a zero population inside. We outline how self-similar solutions (of the second kind) describe
both circular and non-circular fronts in the hole-closing limit. Further, we present new experimental and theoretical
evidence to support the use of nonlinear degenerate diffusion in models for collective cell motion. Our methodology
involves setting up a two-dimensional wound healing assay that has the geometry of a hole-closing problem, with
cells initially seeded outside of a hole that closes as cells migrate and proliferate. For a particular class of fibroblast
cells, the aspect ratio of an initially rectangular wound increases in time, so the wound becomes longer and thinner
as it closes; our theoretical analysis shows that this behaviour is consistent with nonlinear degenerate diffusion but
is not able to be captured with commonly used linear diffusion. This work is important because it provides a clear
test for degenerate diffusion over linear diffusion in cell lines, whereas standard one-dimensional experiments are
unfortunately not capable of distinguishing between the two approaches.
Keywords: nonlinear degenerate diffusion; Porous-Fisher equation; hole-closing problem; cell migration assays;
collective cell motion; wound healing; self-similarity of the second kind
1. Introduction
The use of reaction-diffusion equations in mathematical biology is widespread [1, 2], especially for models of
collective cell motion in applications of wound healing [3 -- 5] and tumour invasion [6 -- 8]. The most commonly used
type of diffusion in these models is Fick's first law, which gives rise to a linear diffusion term. However, a deficiency
in this approach is that such governing equations do not allow for explicit descriptions of invading fronts.
In an
attempt to address this deficiency, some literature covering mathematical models of cell migration and proliferation
has included studies of reaction-diffusion equations with nonlinear degenerate diffusion [9 -- 16]. These models allow
for solutions with a well-defined moving boundary at the front of the invading cell population. An ongoing challenge
in this area of research is to identify the most appropriate choice of nonlinear diffusion. We continue this work in the
present paper by applying analytical and numerical techniques to study a hole-closing problem in the plane, focussing
on the role of nonlinear degenerate diffusion and supporting our study with experimental results from an in vitro cell
migration assay (that complements our recent study [17]).
∗Corresponding author: [email protected]
∗∗SWM and WJ are joint first authors.
Preprint submitted to Physica D: Nonlinear Phenomena.
June 14, 2019
In this paper we study the most simple nontrivial reaction diffusion equation with nonlinear degenerate diffusion
that describes cell proliferation and migration, namely the Porous-Fisher equation
= D∇ ·(cid:18)(cid:18) u
(cid:19)
(cid:19)n ∇u
K
∂u
∂t
(cid:19)
(cid:18)
1 − u
K
+ λu
.
(1)
This is a parabolic partial differential equation with nonlinear degenerate diffusion and a logistic growth term [18 -- 27].
In cell biology, u represents the density of a particular cell type, while x ∈ RN, where N = 1, 2 or 3. The diffusion term
in (1) describes migration (or cell motility) of the cell population, with n > 0 corresponding to a scenario in which
cells are more likely to migrate if crowded, while the logistic growth component of (1) models cell proliferation with
a carrying capacity K. The limiting case n = 0 reduces (1) to the well-known the Fisher-Kolmogorov-Petrovski-
Piskunov (Fisher-KPP) equation
,
(2)
(cid:19)
(cid:18)
1 − u
K
= D∇2u + λu
∂u
∂t
for which cell migration is due to random cell motility that is independent of cell density. As mentioned above, there
are a plethora of studies in cell biology and ecology that use (2) or related models. In much of this vast literature,
the Fisher-KPP equation has proved effective, and can successfully reproduce experimental behaviours; however, the
choice of linear Fickian diffusion (or n = 0 in (1)) is often also made because of its simplicity, not necessarily because
of its biological relevance.
The main goal of the present study is to explore the role of nonlinear degenerate diffusion in a simple model of
collective cell motion. As such, it is important to emphasise the key feature of the Porous-Fisher equation (1) with
n > 0 is that, unlike the Fisher-KPP equation (2) with linear diffusion, it allows solutions with compact support. This
is a direct consequence of the term (u/K)n, which vanishes in the limit u → 0+ (the diffusion is said to be degenerate
in the sense that as u → 0+, the diffusion itself vanishes, and the equation changes from being of parabolic type
to elliptic; to see this, write v = un to give the eikonal equation ∂v/∂t ∼ ∇v2 in the limit u → 0+). Therefore,
it is possible to use the Porous-Fisher equation (1) to model well-defined fronts of cell populations advancing on a
region of zero population. On the other hand, even with initial conditions that have compact support, solutions to the
Fisher-KPP equation (2) have u > 0 for all x and t > 0, meaning information is travelling infinitely fast, which strictly
speaking is not biologically realistic.
The application we have in mind is a two-dimensional cell migration assay (x ∈ R2), with an invading population
of cells moving over a substrate with a sharp front, ahead of which the cell population is essentially zero. In particular,
we focus on the geometry in which there is initially a simply connected region devoid of cells, which we refer to as a
hole or a wound. The resulting hole-closing problem is (1) subject to initial conditions
u(x, 0) = I(x), with
I(x) = 0
for
x ∈ Ω(0),
(3)
where the simple closed curved ∂Ω(0) describes the initial shape of the hole (or wound). The challenge is to solve
for u, but also to track the shape and speed of the boundary of the hole ∂Ω(t), especially in the hole-closing limit
t → t−
c . We are motivated by our recent ([17]) and new experimental data from a wound healing (sticker) assay, as
illustrated in Figure 1(a)-(b). In this particular example, the initial wound is circular in shape; however, a feature of
our experimental design is that we are able to make the initial wounds any shape we choose. More generally, the study
of hole-closing problems in cell biology has applications to a range of experimental wound healing scenarios, such
as the circular wounds created by barrier assays [28 -- 30], in a mammal's ear [31] or cornea [32, 33] or a human skin
equivalent construct [34], for example; similar mechanisms and geometry are involved in cell bridging experiments in
tissue engineering [35]. Further motivation for our study arises from the lack of consensus when using (1) to model
cell migration on the appropriate choice of the diffusion exponent n; we aim to shed light on this issue.
The outline of our paper is as follows. In the following section we set up our subsequent analysis by nondimen-
sionaling the problem and presenting illustrative numerical solutions for the radially symmetric case of a circular hole
(the numerical scheme is summarised in the Supplementary Material). In section 3, we argue that for times leading
up to the hole-closing time tc, the asymptotic behaviour of (1)-(3) in the neighbourhood of the point at which the hole
closes xc is equivalent to that for the Porous Medium equation (which is (1) with λ = 0). For the radially symmetric
case, we revisit the role of self-similarity of the second kind for solutions to the Porous Medium equation, reproducing
some existing results by using an approach which is in some sense more transparent than that documented in the liter-
ature [36 -- 38]. Numerical solutions are presented that illustrate how solutions to (1)-(3) with k-fold symmetry may or
2
Figure 1: Experimental images of hole-closing and numerical solutions. (a) -- (b) Experimental images from a sticker assay with the position
of the leading edge highlighted by yellow circles. The circular holes at t = 0 and 48 h are shown. The scale bar corresponds to 500 µm. (c) -- (f)
Density profiles for the dimensionless Porous-Fisher equation with n = 0.5, 1, 2, and 3, respectively. The red solid line corresponds to the density
profile at focusing time tc. The black arrow indicate increasing time. (g) -- (h) Time evolution of the position of the leading edge for n = 0.5, 1, 2,
and 3. Dashed lines in (h) correspond to straight lines that merge with the time evolution of the position of the leading edge. All the numerical
solutions are obtained on 0 < r < 2 with a nonuniform mesh of 4001 nodes, with an initial circular-hole at R(0) = 0.5. Zero net flux boundary
conditions are imposed at both boundaries r = 0 and 2.
3
00.511.52r00.20.40.60.81u(c)(d)(e)n = 0.5n = 1(f)n = 0.5n = 1n = 2n = 3t = 0 t = 0.01 t = 0.02 t = 0.03t = 0.0465t = 0 t = 0.05 t = 0.07 t = 0.09t = 0.1087t = 0 t = 0.12 t = 0.16 t = 0.20t = 0.2550t = 0 ht = 48 h(a)(b)t = 0 t = 0.10 t = 0.20 t = 0.30t = 0.415100.10.20.30.40.5R(g)n = 2n = 300.511.52r00.20.40.60.81u00.511.52r00.20.40.60.81u00.511.52r00.20.40.60.81u00.10.20.3t0.40.5-6-5-4-1ln(tc-t)-3-2-5.5-4.5-2.5-1.5ln(R)-3.5-0.5(h)n = 0.5n = 1n = 2n = 3 may not evolve to a circle in the hole-closing limit, depending on the diffusion exponent n, as per the analogous case
of the Porous Medium equation [38, 39]. In section 4, we study solutions that are not self-similar in the hole-closing
limit. In particular, we choose initial conditions for which the hole is initially rectangular and show how the hole
becomes long and thin in the hole-closing limit, again following the behaviour of the Porous Medium equation [40].
By comparing with new experimental data from recently developed sticker assays (described in the Supplementary
Material), we demonstrate how to estimate the exponent n in (1). Finally, we close the paper in section 5 with a sum-
mary and a discussion about how our results provide support for the use of nonlinear degenerate diffusion in models
for collective cell motion.
2. Governing equations and preliminary numerical results
2.1. Nondimensionalisation
As discussed in the Introduction, the model we focus on is the Porous-Fisher equation (1), which is the simplest
model for collective cell motion in a two-dimensional assay that includes nonlinear degenerate diffusion. The only
dependent variable is the cell density and, as a consequence, this model implicitly assumes that an excess of nutrients
is available so that cell migration and cell proliferation are not affected by any lack of nutrient availability [41].
tative length and time scales to be
We nondimensionalise (1) by scaling cell density with respect to carrying capacity K and choosing the represen-
D/λ and λ−1, respectively. As such, our dimensionless Porous-Fisher equation,
= ∇ · (un∇u) + u (1 − u)
in x ∈ R2 \ Ω(t),
(4)
√
∂u
∂t
does not involve any parameters apart from the diffusion exponent n > 0. The dimensionless version of the initial
condition (3) will, however, depend implicitly on the dimensional parameters. For example, as we discuss shortly, for
a wound that is initially circular, the density u = u(r, t) is a function of r and t, and the relevant initial condition is
u(r, 0) = H(r − R(0)),
(5)
where H(r) is the Heaviside function and R(0) is the initial radius of the wound. In this case, the dimensionless
quantity R(0) is the dimensional initial radius scaled by
D/λ.
We may formulate our hole-closing problem as a moving boundary problem by coupling (4) with the conditions
√
where here ∂u/∂ν is a normal derivative. The second condition in (6) enforces conservation of mass at the interface.
u = un ∂u
∂ν
= 0
on ∂Ω(t),
(6)
2.2. Numerical results for radially symmetric problem
We now provide numerical results for radially symmetric hole closing, which is governed by
(cid:32)
(cid:33)
∂u
∂t
=
run ∂u
1
∂
r
∂r
∂r
u = un ∂u
∂r
+ u (1 − u) ,
r > R(t),
= 0
on r = R(t),
(7)
(8)
subject to the initial condition (5). The numerical approach we use, based on applying a straight-forward finite-
difference scheme on an uneven grid, is summarised in the Supplementary Material.
In Figure 1(c)-(f) we present density profiles for the four exponents n = 0.5, 1, 2 and 3, each with R(0) = 0.5. In
all four cases we see that the solutions have compact support so that u > 0 for r > R(t). Further, we note that the radius
of the hole, R(t), decreases in time until R → 0 as t → t−
c , where tc is the hole-closing time. The dependence of R(t)
on t is shown in Figure 1(g). We see here that the contact line r = R(t) moves more slowly as the diffusion exponent
n increases, which is to be expected, since, for a fixed u, the nonlinear diffusion term un decreases as n increases.
4
One interesting observation is that the slope of the density profiles at the contact line r = R(t) vanishes for n < 1,
is a nonzero constant for n = 1, and is infinite for n > 1 [15, 24]. Indeed, a simple leading order balance near the
contact line for R(t) > 0 suggests that u ∼ B(t) (r − R(t))1/n, which explains these qualitative behaviours.
Another observation about the propagation of the contact line is that R(t) appears to follow a power-law R(t) ∼
µ(tc − t)β in the hole-closing limit t → t−
c . This point is demonstrated in Figure 1(h), where the solid curves are
numerical results corresponding to Figure 1(g). As these curves appear to approach a line on the log-log, a power-law
behaviour is anticipated. The slope β is dependent on n, as we discuss in detail in the following section.
3. Self-similar solutions
3.1. Similarity solutions for radially symmetric geometry
In the limit the hole closes, t → t−
c , we look for a similarity solution
u ∼ (tc − t)αU(ρ), where ρ =
where at this stage the exponents α and β are unknown. To proceed we require the partial derivatives
(cid:32)
(tc − t)β ,
(cid:33)
r
,
∂u
∂t
−αU + βρ
= (tc − t)α−1
(cid:33)
= (tc − t)α−β dU
∂u
∂r
dρ
= (tc − t)α(n+1)−2β 1
,
dU
dρ
(cid:32)
(cid:32)
1
r
∂
∂r
run ∂u
∂r
d
dρ
ρUn dU
dρ
(cid:33)
(9)
(10)
(11)
(13)
(14)
(15)
(16)
(12)
We see that ∂u/∂t = O((tc − t)α−1), u = O((tc − t)α) and u2 = O((tc − t)2α). Therefore, the source terms u and u2 in
(7) do not contribute to leading order, and instead the self-similar behaviour is driven by nonlinear diffusion via the
Porous Medium equation
ρ
.
(cid:32)
(cid:33)
run ∂u
∂r
(cid:33)
Furthermore, we even can keep the first source term u so that
run ∂u
∂r
∼ 1
r
∂u
∂t
∂u
∂t
∂
∂r
(cid:32)
1
r
∂
∂r
=
in this case, the change of variables
u = e(tc−t)u,
leads again to the Porous Medium equation
∂u
∂t
=
1
r
∂
∂r
(cid:32)
t =
1
n
(cid:33)
,
r > R(t).
+ u,
r > R(t);
(cid:16)
(cid:17)
e−n(tc−t) − 1
run ∂u
∂r
,
r > R(t).
We shall continue with (13) but simply note that our analysis of the Porous Medium equation holds exactly for (14).
By substituting (10) and (12) into (13), we find that
(cid:32)2β − 1
(cid:33)
n
−
U + βρ
dU
dρ
=
2β − 1
α =
(cid:33)
,
(cid:32)
ρUn dU
dρ
,
ρ > µ.
(17)
(18)
n
1
ρ
d
dρ
Here the similarity exponent β cannot be determined by dimensional analysis of the governing equation (13) or by
applying global conservation of mass; instead, it acts as an eigenvalue for a boundary-value problem associated
with (18), which we discuss below in some detail. This self-similarity of the second kind makes the problem rather
challenging.
5
3.2. Relationship to other studies
At this point it is worth making two comments on this formulation. First, the hole-closing problem for the Porous
Medium equation (13) has been studied by a number of authors [36 -- 39] (including for the special case n = 3 [42, 43],
which corresponds to an inwardly-filling viscous gravity current). In all of these studies, the governing equation (13)
is rewritten using the so-called pressure variable v = un and the subsequent analysis involves similarity solutions of
the form v = (tc − t)2β−1V(ρ). Further changes of variables are required to complete the analysis, which to a certain
extent has the effect of hiding the physical interpretation.
Second, the hole-closing problem for the fourth-order analogue of (13), the thin film equation
(cid:33)
(cid:32)
run ∂3u
∂r3
∂u
∂t
= −1
r
∂
∂r
,
r > R(t),
(19)
has been studied recently by Zheng et al. [44, 45]. This analogue shares some features of the hole-closing problem for
(13). In the spirit of [44], and in an attempt to provide a more transparent analysis than that which uses the pressure
variable v = un, we shall concentrate on the dependent variable u (and not v) and note some similarities and differences
with [44] later.
To formulate the appropriate boundary-value problem associated with (18), we first derive the far-field condition
c for a fixed r > 0 (in other words, our solution cannot blow up at
3.3. Far-field conditions
as ρ → ∞. As we do not want ∂u/∂t → ∞ as t → t−
(cid:33)
any point other than r = 0), we conclude from (10) and (17) that
∼ 0
(cid:32)2β − 1
U + βρ
−
n
dU
dρ
as ρ → ∞,
(20)
which implies
U ∼ aρ(2β−1)/nβ
as ρ → ∞.
(21)
This argument is used extensively in dealing with similarity solutions [46, 47]. The constant a in (21) is arbitrary;
however, the differential equation (18) is invariant under the transformation
thus, without loss of generality, we can set a = 1 and recover any solution we like by stretching U and ρ appropriately.
Therefore we have as our far-field condition,
U → U,
ρ → n/2ρ,
(22)
(23)
Numerically, we will need to truncate 0 < ρ < ∞ to 0 < ρ < ρ∞, and then interpret (23) as two boundary conditions
(24)
U = ρ(2β−1)/nβ
on ρ = ρ∞,
,
U ∼ ρ(2β−1)/nβ
as ρ → ∞.
ρ(2β−1−nβ)/nβ
∞
,
on ρ = ρ∞.
(25)
(cid:32)2β − 1
∞
(cid:33)
nβ
dU
dρ
=
Thus we can treat (18) with (24)-(25) as an initial-value problem starting at ρ = ρ∞.
3.4. Shooting method and near-field conditions
For a fixed n, we have a one-parameter family of initial-value problems, each for a different value of β. Our
strategy is to interpret these via a shooting method, where we start at ρ = ρ∞ and shoot backwards until
We wish to determine the appropriate value of β for which (8) is satisfied, or alternatively,
U = 0
on ρ = µ.
Un dU
dρ
→ 0
as ρ → µ+.
6
(26)
(27)
In addition to determining the appropriate value of β, the parameter µ must also be computed as part of the numerical
solution. It will turn out that for each value of n, there are infinitely many pairs (µ, β) that satisfy (24)-(25) and (26),
but only one pair (µc, βc) which also satisfies (27). In order to proceed, we need to analyse possible behaviours of
solutions to (18) as U → 0+.
Case 1, β < βc. By considering (18) directly, we see that, provided µ (cid:44) 0, for (26) to hold we must have
(cid:32)
(cid:33)
βµ
dU
dρ
∼ d
dρ
Un dU
dρ
as ρ → µ+.
(28)
(29)
(30)
Integrating, we find
βµU ∼ Un dU
dρ
+ ¯C as ρ → µ+,
where ¯C is a constant. In the generic case where ¯C (cid:44) 0, the two terms on the right-hand side balance, giving
U ∼ C(ρ − µ)1/(n+1)
as ρ → µ+.
Following [44], we call this case generic touch-down.
Case 2, β = βc. For the special case in (29) in which ¯C = 0, the left-hand side must balance the first term on the
right-hand side, giving
(31)
Note in this case, (27) is satisfied, so this is the physically relevant solution we are after. We call this nongeneric
touch-down [44].
Case 3, β > βc. Here we now have µ = 0, which means that all the terms in (18) balance in the limit. By employing a
power-law ansatz, we find that
U ∼ (βcµcn)1/n(ρ − µc)1/n
as ρ → µ+
c .
(cid:32)
(cid:33)1/n
U ∼
We call this touch-down at the origin [44].
n
4(n + 1)
ρ2/n
as ρ → 0+.
(32)
In summary, the numerical task is to vary β until the special borderline case 2 (nongeneric touch-down) is deter-
mined for β = βc and µ = µc. To ensure the borderline case is accurately identified, we utilise a high-order implicit
finite-difference scheme to discretise (18), with local error tolerance set to near machine precision. The contrasting
touch-down behaviours of cases 1 and 3 provide a convenient means of bracketing the critical value βc, and hence
convergence to this value is achieved iteratively through repeated bisection of the interval.
Some results of this task are presented in Figure 2. In Figure 2(a) the plots are for n = 0.5. We can see clearly
see the two representative profiles for β < βc (generic touch-down) have infinite slope at ρ = µ; indeed, for this value
of the diffusion exponent n, we have U ∼ C(ρ − µ)2/3. On the other hand, the two representative profiles for β > βc
(touch-down at the origin) are very flat in the limit (here U ∼ ρ4/144). The borderline case β = βc has U scaling
like (ρ − µ)2, which also has a zero slope, but does not approach the origin. The other three examples of n, shown in
Figure 2(b)-(d), also show profiles for each of the three cases described above. We see that qualitatively the behaviour
changes, depending on the value of the exponents in (30)-(32). In Figure 2(e)-(f) we show only the physically relevant
similarity solutions with β = βc. Again, the qualitative behaviour as these curves intersect the ρ-axis depends on the
exponent in (31). In particular, we observe that the similarity solutions for n = 0.5 and 0.75 have zero slope at the
contact line, while the solutions for n = 1.5, 2 and 3 have infinite slope. The borderline case is n = 1, where the slope
is finite.
As a check on our similarity solutions, we present in Figure 3 profiles of U versus ρ (where we recall that U =
u(r, t)/(tc − t)(2β−1)/n and ρ = r/(tc − t)β) that are computed with our numerical scheme for various times. The figure
includes plots for four values of n, namely n = 0.5, 1, 2 and 3. For each value of n, we see the initial condition is a
scaled version of the Heaviside function (5), while as time increases, the numerical solutions of (7) (blue solid lines)
approach the similarity solution (red dashed line) in the limit t → t−
c . This comparison between our numerical solution
and our similarity solution provides confidence that our analysis in Section 3 is correct. Note we had to rescale our
similarity solutions using (22) in order to match the point at which U = 0 with the numerical solution.
7
Figure 2: Similarity solutions generated by the shooting method. (a) is for n = 0.5. Here the dashed (green) curves, for β = 0.6 and 0.75, are
examples of generic touch-down; the solid (blue) curve, for β = βc = 0.909, is the physically relevant nongeneric touch-down solution; while the
dot-dashed (magenta) curves, for β = 1.05 and 1.2, are examples of touch-down at the origin. Using the same convention: (b) is for n = 1 and
β = 0.6, 0.7, 0.856, 1 and 1.5; (c) is for n = 2 and β = 0.58, 0.65, 0.796, 0.9 and 1.5; while (d) is for n = 3 and β = 0.67, 0.7, 0.762, 0.85 and 1.5.
In (e)-(f), the physically relevant nongeneric touch-down solutions are shown for n = 0.5 (β = βc = 0.909), 0.75 (0.880), 1 (0.856), 1.5 (0.822), 2
(0.796) and 3 (0.762), with the black arrow black arrows indicating increasing n.
8
Uρ2.50051243120.51.5n = 0.5n = 1n = 2n = 3(a)(b)(c)(d)Uρ2.50051243120.51.5Uρ2.50051243120.51.5Uρ2.50051243120.51.5Uρ2.50051243120.51.5ρ0102486U1004826(e)(f)touch-down at the origingeneric touch-downnongeneric touch-downnnn Figure 3: Scaled solutions of the dimensionless Porous-Fisher equation for n = 0.5, 1, 2, and 3. In all cases, numerical solutions of (7)-(8) with
(5), computed via the numerical scheme summarised in the Supplementary Material, are represented by (blue) solid curves. Numerical solutions
shown for: (a) t = 0, 0.01, 0.02, 0.03, 0.04, and 0.046; (b) t = 0.05, 0.07, 0.09, 0.10, and 0.108; (c) t = 0.12, 0.16, 0.20, 0.24, and 0.254; and (d)
t = 0.10, 0.20, 0.30, 0.40, and 0.41. In each of the four cases, solutions are obtained on 0 < r < 2 with a nonuniform mesh of 4001 nodes, with
an initial circular hole with R(0) = 0.5. Zero net flux boundary conditions are imposed at both boundaries r = 0 and 2. Also included in these
results are the similarity solutions computed using the algorithm described in Section 3, represented by (red) dashed curves. We see the numerical
solutions approach the similarity solutions in the limit t → t−
c .
9
050103020400600040002000n = 0.5030504010200642(a)(b)(c)n = 1n = 2Numerical solutionSimilarity solution010302001.522.50.51n = 3(d)UρρUUUtttt0501030204005010302040ρρ We close this subsection by recalling the analogue problem for the thin film equation (19). As explained by
Zheng et al. [44], the (circular) hole-closing problem for this fourth-order equation is similar to our problem (which
is essentially for the Porous Medium equation (13)) in that there are self-similar solutions of the second kind. Further,
our procedure for computing these similarity solutions is based on that presented in [44]. In particular, like in [44], we
use a type of shooting method to explore numerical solutions for a range of parameter values and choose the relevant
solution by ensuring the near-field limiting behaviour matches a physical constraint. The main difference between our
study and that in [44] is that we have one free parameter β and only three near-field options (which we call case 1, 2
and 3). On the other hand, the thin film equation is higher order and so Zheng et al.'s study is more complicated. They
have a free parameter in the far-field condition as well as the similarity exponent β and, consequently, they have more
than three options for their near-field behaviour. Another further complication for the thin film model in [44] is that
there needs to be a prewetting film which regularises a well-known singularity in stress at the moving contact line.
3.5. Similarity solutions with k-fold symmetry
For the Porous Medium equation, the stability of the radially symmetric similarity solutions (9) is known to depend
on the diffusion exponent n. In particular, if we consider stability of the interface R(t) = µ(tc − t)β by adding a small
perturbation γ(t) cos kθ, then for each k ≥ 3, the perturbed solution is stable for n > nk and unstable for n < nk, where
nk is some borderline exponent that can be computed numerically [38, 39]. By stable, we mean that γ/R → 0 as
t → t−
c .
We postulate that the same type of stability holds for the full Porous Fisher equation (4), and now consider a
relevant example for k = 4. Suppose the initial condition is a square-shaped hole with a side of length unity. Here
the solution has a 4-fold symmetry, so there are two options for the shape of the interface in the limit it closes. The
first is for n > n4, where we assume n4 ≈ 0.32 [39]. For example, in Figure 4(a) we show the shape of the interface
for a numerical solution with n = 0.5. Since the radially symmetric similarity solution (9) is stable for n = 0.5 to a
4-fold perturbation, we expect the interface for the full solution to approach a circle at extinction. While it is difficult
to compute such two-dimensional solutions accurately near extinction, our numerical results in Figure 4(a) appears
to show the hole becoming more circular as it closes. To support this idea, we have in Figure 4(b) plotted a kind of
aspect ratio Ad, which is the ratio of the diagonal of the hole to the x-intercept. Again, it is difficult to tell given the
scales involved, but it is not unreasonable to believe that Ad is tending to unity at extinction, which should happen if
it is becoming more circular.
Also shown in Figure 4 is an example for n < n4, which demonstrates the second option for the shape of the hole
as it closes. In Figure 4(c)-(d) the results are for n = 0.2, for which we postulate the interface is unstable. Given
the solution is rotationally symmetric (with 4-fold symmetry), this instability manifests itself by forcing the hole to
approach a (noncircular) 4-fold symmetric shape at extinction, which is like a square with rounded corners. This type
of evolution is suggested in Figure 4(c). Further, such a 4-fold symmetric shape will have a value of Ad which is not
unity. The time-dependence of Ad in Figure 4(d) appears to support the idea that Ad (cid:54)→ 1 in the hole-closing limit.
We expect that qualitatively similar results could be presented for any k-fold symmetric initial condition, where
k ≥ 3, with the hole approaching a circle in shape for n > nk and a k-fold symmetric shape (which is like a regular
k-sided polygon with rounded corners) for n < nk. On the other hand, for k = 2 we expect the similarity solution
to unstable for all n, so that solutions with 2-fold symmetric initial conditions are no longer self-similar in the hole-
closing limit [38, 40]; instead, the interface becomes oval in shape with an increasingly large aspect ratio which scales
like A = O((tc−t)−1/2). We discuss this possibility further below in Section 4 where we focus on a rectangular-shaped
initial hole.
All of these stability results are analogous to those for contracting bubbles in a Hele-Shaw cell for which the
viscous fluid is of a power-law type [48, 49] or for which there is a competition between surface tension and a
kinetic-type boundary condition [50, 51]. For these problems, there are also linear stability results that show a circular
interface may be stable or unstable, depending on a parameter value, leading to the existence of noncircular self-
similar solutions (that evolve to shapes which appear like k-sided polygons with rounded corners). Further, 2-fold
symmetric perturbations are unstable for these Hele-Shaw problems, leading to interfaces that approach a slit in the
hole-closing limit.
10
Figure 4: Time evolution of the hole shape for n = 0.2 and 0.5. (a) Evolution of the hole shape during closing for n = 0.2 from t−tc = −5.05×10−4
to −2.5 × 10−5. (b) Time evolution of the diagonal aspect ratio for n = 0.2. (c) Evolution of the hole shape during closing for n = 0.5 from
t − tc = −1.09 × 10−3 to −9.35 × 10−5. (d) Time evolution of the diagonal aspect ratio for n = 0.5. Both of the numerical solutions are obtained
by solving the two-dimensional (dimensionless) Porous-Fisher equation on 0 < x < 2 and 0 < y < 2 with a 401 × 401 nonuniform mesh, with an
initial square hole of length 0.5 centred at origin.
11
00.010.020.030.040.05x00.010.020.030.040.05yn = 0.500.030.0611.11.21.31.41.5(c)(d)(a)(b)n = 0.2tdn = 0.2n = 0.5A00.010.020.030.040.05x00.010.020.030.040.05y00.010.0211.11.21.31.41.5tdA 4. Rectangular shaped wounds
As just discussed, linear stability analysis of the radial similarity solutions shows that 2-fold symmetric perturba-
tions grow in time [40], which suggests that an initially rectangular hole will become longer and thinner as it closes.
We use this property to demonstrate how the diffusion exponent n can be determined by fitting to aspect ratio data
from a rectangular wound healing assay.
4.1. Sticker assays
We briefly summarise the sticker assays reported in [17]. These two-dimensional wound healing assays were
performed using (NIH 3T3) fibroblast cells. A double-sided sticker was cut to a particular wound shape (circle,
triangle or square) using a laser scribe, and then attached to a cell culture dish. The fibroblast cells were placed in the
dish and incubated overnight. The sticker was removed to reveal the wound area of the required shape. As the cells
migrated into the vacant space, images were taken at various discrete times and subsequently analysed using ImageJ
[52] to determine the wound area at each time point (see Supplementary Material).
In [17], we simulated the sticker assays (with circular, triangular and square wounds), using a discrete random walk
model on a hexagonal lattice incorporating crowding effects via an exclusion process [53 -- 55] (whose continuum-limit
description is the two-dimensional Fisher-KPP equation). We estimated the cell proliferation rate λ = 0.036 /h and
the carrying capacity K = 1.4 × 10−3 cells/µm2 by counting cells in sample regions of a corresponding proliferation
assay and calibrating to the logistic growth model. The random walk model was then used to estimate the diffusion
coefficient D = 1200 ± 260 µm2/h which can be used for the Fisher-KPP equation.
4.2. New experimental results
We now report on new sticker assays performed for rectangular-shaped wounds using the same protocols as in [17].
For example, in Figure 5(a) we show experimental images for an initially rectangular wound whose aspect ratio is 2.
These three representative images are taken for times t = 24, 48 and 57 h. By approximating the wound boundary by
a rectangle at each time step (see Supplementary Material), we are able to record the aspect ratio versus time for 3
replicates, and plot the result in Figure 5(d).
4.3. Fitting for the diffusion exponent n
To calibrate our experimental data with the Porous-Fisher equation (1), we first take the values λ = 0.036 /h and
K = 1.4 × 10−3 cells/µm2 from our previous study [17] described above. Then, for each fixed value of n we choose
to deal with, we fit for D by comparing our numerical results of (7)-(8) with initial condition (5) with our previously
obtained experimental data for the wound area with an initially circular wound [17]. We use a simple least-squares
error to identify the most appropriate choice for D. Our results give estimates of D and an interval of uncertainty in
our estimates (Supplementary Material).
For each pair of n and D (together with λ = 0.036 /h and K = 1.4 × 10−3 cells/µm2), we then solve the Porous-
Fisher equation (1) numerically with a rectangular-shaped hole that represents the new experimental results shown in
Figure 5(a). As time increases, the hole begins to close, as expected. Representative numerical results are shown in
Figure 5(b)-(c) for n = 1 and 2 (note that the initial condition for these numerical solutions involves setting u = 1
outside of the hole, truncated for numerical purposes at finite values of x and y). Crucially, we plot the aspect ratio of
the closing hole in Figure 5(e) for various values of n and also include the experimental data in the same image. This
strategy allows us to choose the value of n which best matches the experimental data. A least-squares error between
the experimental aspect ratio and the numerical results is shown in Figure 5(f). We see that, of the values of n we have
simulated, the choice n = 1 appears to provide the closest match. Further details of the strategies employed in this
subsection are provided in the Supplementary Material.
5. Discussion
In this paper we have studied various properties of solutions to the so-called hole-closing problem for the Porous-
Fisher equation (1). We have summarised self-similar solutions of the second kind, which apply in the neighbourhood
of the moving front in the limit the hole-closes.
It turns out that the dynamics in this limit are governed by the
12
Figure 5: Comparison of the aspect ratio profile between experimental data and simulation results. (a) Experimental image of wound-healing
assay at t = 24, 48, and 57 h with an initial aspect ratio of 2. (b) -- (c) Numerical simulation of the dimensional Porous-Fisher equation for n = 1
and 2, with the same initial wound size at t = 24, 48, and 57 h. The white area indicates the vacant space where cell density is zero. (d) Time
evolution of the aspect ratio from experimental data. (e) Comparison of the aspect ratio profile obtained by solving the Porous-Fisher equation for
n = 1, 2, and 3, and the experimental data. (f) The least-squares difference of the aspect ratio profile between the experimental data and simulation
results for n = 1, 2, and 3. Black arrow indicates the direction of increasing n. All the numerical solutions are obtained by solving two-dimensional
Porous-Fisher equation on a domain of the same size as the experimental field of view (6 mm × 4 mm) with uniform meshes. The initial rectangular
hole size is 4.20 mm × 2.22 mm. For each choice of n, the value of D is estimated by calibrating the Porous-Fisher equation to the experimental
data of circular wound [17]. For all the solutions λ = 0.036 /h, K = 1.4 × 10−3 cells/µm2. δx = 3.75 and δy = 2.5 for n = 0.5, 1, and 2. δx = 7.5
and δy = 5 for n = 3.
Figure 6: Aspect ratios of level sets u = ¯u for the Fisher-KPP equation. the Fisher-KPP equation (2) is simulated with the same initial condition
as the simulation in Figure 5. Here we have used D = 1000, λ = 0.036 /h, K = 1.4 × 10−3cells/µm2. δx = 7.5, δy = 5, and δt = 5 × 10−3.
13
(a)t = 48 h(d)Data(e)(b)(c)t = 48 ht = 96 ht = 24 ht = 24 ht = 48 ht = 24 ht = 48 ht [h]1001.50204060802.53.54.55.5t [h]1000204060801.52.53.54.5t = 57 ht = 57 ht = 57 hn = 1n = 2SimulationE041230.250.400.300.35n(f)AAt [h]01206018234567204080100Au Porous Medium equation and therefore many results carry over from previous studies of that equation [36 -- 39]. In
contrast to those studies, we have formulated our analysis in terms of the original dependent variable and not the
so-called pressure variable v = un; in this way our approach is more like the recent analogous study of the thin
film equation [44]. Stability analysis [40] shows that k-fold (with k ≥ 3) symmetric initial hole-shapes (like regular
polygons) may become circular in the hole-closing limit, or may evolve to other non-circular shapes that keep the
symmetry, all depending on the value of the exponent n. Further, 2-fold symmetric initial conditions (like rectangles)
will become long and thin in the limit. Our numerical solutions confirm these predictions.
The present study is inspired by our recently published experimental data from a two-dimensional sticker assay
with circular, square and triangular shaped wounds [17]. These experiments are modelled by hole-closing problems
as there is an initially vacant wound area which is ultimately closed up as the cells migrate inwards and proliferate
to occupy the initial wound space. In this paper we have presented new experimental results from sticker assays
with rectangular-shaped wounds. We find the aspect ratio of the wounds increases in time in a way that agrees with
our model using the Porous-Fisher equation (1). For various values of n, we are able to fit for the diffusivity D by
minimising the error between the numerically computed aspect ratio and the corresponding experimental data. Our
results suggest that, for this cell line, a reasonable estimate for the exponent n is n = 1. This result is compatible with
other studies of cell migration using the Porous-Fisher equation [10, 15]. Furthermore, this estimate for n is consistent
with our previous theoretical prediction [14], which suggests that for cells that themselves have an aspect ratio of N,
the appropriate choice of n is n = N − 1. While NIH 3T3 fibroblast cells are not at all the same shape, many have an
aspect ratio of roughly two.
To put these results in context, we recall how difficult it has been to identify an appropriate choice of the diffusion
exponent n in (1) when fitting with data from wound healing assays or experiments with traditional wound shapes. For
example, for (approximately) one-dimensional fronts that arise from scratch assays with PC-3 prostate cancer cells,
Jin et al. [10] calibrated the Porous-Fisher equation to the data for the examples n = 0.5, 1, 2, 3 and 4. While it was
concluded that the choice n = 1 outperforms the others, the evidence was not straightforward as the corresponding
estimates for the diffusion coefficient D varied greatly over a range of initial conditions. In their study of circular
wounds, Sherratt & Murray [12] briefly compared numerical simulations of the Porous-Fisher (1) with n = 4 with
experimental results from rabbit ears [31] and attempted to fit the data for D. While they observed the fit was not
impressive, they did not attempt to vary n but instead added other features of the model (to account for biochemical
mediators). A further example is the study of Sengers et al. [13], who concluded that MG-63 bone cancer cells spread
out radially in a way that was well represented by the Porous-Fisher with n = 1, but again they did not attempt to
fit the data with other values of the diffusion exponent. We conclude that our approach of using rectangular-shaped
wounds in a sticker assay has the attractive feature of providing an additional means to fit for the diffusion exponent n
which has thus far been missing in the literature.
It is worth reflecting on how poorly the Fisher-KPP equation (2) performs at identifying certain properties of
moving fronts at they evolve. As discussed above (and is well known), the Fisher-KPP equation does not allow
solutions with compact support and so struggles to model scenarios with well-defined fronts. In particular, we are
concerned with experiments where fronts of cell populations invade an empty space, such as in our two-dimensional
wound healing assays. For this type of experiment, in order to apply the Fisher-KPP equation one must arbitrarily
nominate a level set u = ¯u as representing the moving front and then track properties of that level set as it evolves. For
example, we have solved (2) numerically for a case with a 2× 1 rectangular shaped wound and plotted in Figure 6 the
aspect ratio of a number of different level sets versus time. In all cases, the aspect ratios start at approximately A = 2
and then increase monotonically with time which is generally consistent with the experimental data (also included in
this figure for reference). However, the actual values the aspect ratios take are very different for each level set and
clearly there is no obvious choice as to which level set is most appropriate. As such, as a predictive tool, solutions to
the Fisher-KPP equation (2) are not at all useful for describing this particular experimental property.
One possible feature of our experiments that is not included in the model (1) is chemotaxis, whereby cells produce
a chemical signal, or chemoattractant, with concentration g, which can promote directional motion as the cells prefer-
entially move up or down a gradient of g [56]. The inclusion of chemotaxis in mathematical models for wound healing
and tumour growth is commonplace [4, 5, 57, 58]. An extension of our model (1) which incorporates chemotaxis could
be
= ∇ ·(cid:18)
D
(cid:18) u
K
∂u
∂t
(cid:19)n ∇u − χu∇g
(cid:19)
(cid:18)
(cid:19)
+ λu
1 − u
K
14
,
(33)
= Dg∇2g + k1u − k2g,
∂g
∂t
(34)
where χ is the chemotactic sensitivity coefficient, Dg is the diffusivity of the chemotactic chemical, k1 is the rate at
which cells produce the chemotactic chemical, and k2 is the rate at which the chemotactic chemical undergoes natural
decay [59, 60]. One significant challenge would be to obtain reasonable estimates for the four additional parameters
Dg, χ, k1 and k2 by calibrating the solution of the coupled system (33)-(34) with the experimental data. We have
not pursued this approach for two main reasons. First, we have already been able to obtain a good match with the
data using the simpler model (1) and we were able to use this calibrated model to draw conclusions about the use
of degenerate diffusion for cell migration. Second, as is often the case with two-dimensional wound healing assays,
we have not taken any measurements of concentrations of chemoattractants; without such measurements, there is
obvious metholodogy for estimating the four additional parameters in the extended model (33)-(34). For this more
complicated modelling to be useful, we suggest experimentalists make such measurements.
In summary, we have provided a range of evidence to support the use of nonlinear degenerate diffusion via the
Porous-Fisher equation (1) for problems involving invading fronts. In particular, by comparing simulations with aspect
ratio data taken from sticker assays with rectangular wounds, our methodology provides a clear test for degenerate
diffusion over linear diffusion in two-dimensional cell migration experiments.
Acknowledgements
This work is supported by the Australian Research Council (DP140100249, DP170100474) and the Taiwan Min-
istry of Science and Technology (MOST 106-2313-B-002-031-MY3). WJ is supported by a QUT Vice Chancellor's
Research Fellowship. SWM acknowledges many useful discussions with David Frances and Sean McElwain. The
authors are grateful for the computational resources and technical support provided by QUT's High Performance
Computing and Research Support group. Finally, the authors appreciate the supportive comments of the anonymous
referees.
References
[1] Murray J.D., 2002. Mathematical Biology I: An Introduction, 3rd Ed, Springer.
[2] Britton N.F., 1986. Reaction-diffusion equations and their applications to biology, Academic Press, London.
[3] Tranquillo R.T., Murray J.D., 1992. Continuum model of fibroblast-driven wound contraction: inflammation-mediation. J Theo Biol. 158:
135 -- 172.
[4] Olsen L., Sherratt J.A., Maini P.K., 1995. A mechanochemical model for adult dermal wound contraction and the permanence of the contracted
tissue displacement profile. J Theor Bio. 17: 113 -- 128.
[5] Murphy, K.E., Hall, C.L., Maini, P.K., McCue, S.W., McElwain, D.L.S., 2012 A fibrocontractive mechanochemical model of dermal wound
closure incorporating realistic growth factor kinetics. Bull Math Bio. 74: 1143 -- 1170.
[6] Gatenby R.A., Gawlinski E.T., 1996. A reaction-diffusion model of cancer invasion. Cancer Res. 56: 5745 -- 5753.
[7] Roose T., Chapman S.J., Maini P.K., 2007. Mathematical models of avascular tumor growth. SIAM Rev. 49: 179 -- 208.
[8] Sherratt J.A., Chaplain, M.A.J., 2001. A new mathematical model for avascular tumour growth. J Math Biol. 43: 291 -- 312.
[9] Baker R.E., Simpson M.J., 2012. Models of collective cell motion for cell populations with different aspect ratio: Diffusion, proliferation and
travelling waves. Physica A. 391: 3729-3750.
[10] Jin W., Shah E.T., Penington C.J., McCue S.W., Chopin L.K., Simpson M.J., 2016. Reproducibility of scratch assays is affected by the initial
degree of confluence: Experiments, modelling and model selection. J Theor Biol. 390: 136-145.
[11] Maini P.K., McElwain D.L.S., Leavesley D., 2004. Travelling waves in a wound healing assay. Appl Math Lett. 17: 575-580.
[12] Sherratt J.A., Murray J.D., 1990. Models of epidermal wound healing. Proc Roy Soc Lond B. 241: 29-36.
[13] Sengers B.G., Please C.P., Oreffo R.O.C., 2007. Experimental characterization and computational modelling of two-dimensional cell spread-
ing for skeletal regeneration. J Roy Soc Interface 4: 1107 -- 1117.
[14] Simpson M.J., Baker R.E., McCue S.W., 2011. Models of collective cell spreading with variable cell aspect ratio: A motivation for degenerate
diffusion models. Phys Rev E. 83: 021901.
[15] Warne D.J., Baker R.E., Simpson M.J., 2019. Using experimental data and information criteria to guide model selection for reactiondiffusion
problems in mathematical biology. Bull Math Biol 81: 1760 -- 1804.
[16] Simpson M.J., Landman K.A., Hughes B.D., Fernando A.E., 2010. A model for mesoscale patterns in motile populations. Physica A. 389:
[17] Jin W., Lo K.Y., Chou S.E., McCue S.W., Simpson M.J., 2018. The role of initial geometry in experimental models of wound closing. Chem
1412 -- 1424.
Eng Sci. 179: 221-226.
12: 880 -- 892.
[18] Atkinson C., Reuter G.E.H., Ridler-Rowe C.J., 1981. Traveling wave solutions for some nonlinear diffusion equations. SIAM J Math Anal.
[19] De Pablo A., S´anchez A., 1998. Travelling wave behaviour for a porous-Fisher equation. Euro J Appl Math. 9: 285 -- 304.
15
[20] Harris S., 2004. Fisher equation with density-dependent diffusion: special solutions. J Phys A. 37: 6267 -- 6268.
[21] Hilhorst D., Kersner R., Logakc E., Mimura M., 2008. Interface dynamics of the Fisher equation with degenerate diffusion. J Diff Eq. 244:
[22] King J.R., McCabe P.M., 2003. On the Fisher-KPP equation with fast nonlinear diffusion. Proc Roy Soc Lond A. 459: 2529 -- 2546.
[23] Medvedev G.S., Ono K., Holmes P.J., 2003. Travelling wave solutions of the degenerate Kolmogorov-Petrovski-Piskunov equation. Euro J
[24] Newman, W.I., 1980. Some exact solutions to a non-linear diffusion problem in population genetics and combustion. J Theor Biol. 85:
[25] S´anchez-Garduno F., Maini P.K., 1995. Travelling wave phenomena in some degenerate reaction-diffusion equations. J Diff Eqns. 117: 281 --
[26] Sherratt J.A., Marchant, B.P., 1996. Nonsharp travelling wave fronts in the Fisher equation with degenerate nonlinear diffusion. Appl Math
Appl Math. 14: 343 -- 367.
2870 -- 2889.
325 -- 334.
319.
Lett. 9: 33 -- 38.
[27] Witelski T.P., 1995. Merging traveling waves for the Porous-Fisher equation. Appl Math Lett. 8: 57 -- 62.
[28] Ashby W.J., Zijlstra A., 2012. Established and novel methods of interrogating two-dimensional cell migration. Integr Biol. 4: 1338 -- 1350.
[29] Kramer N., Walzl A., Unger C., Rosner M., Krupitza G., Hengstschlager M., Dolznig H., 2013. In vitro cell migration and invasion assays.
[30] Treloar K.K., Simpson M.J., McElwain D.L.S., Baker, R.E., 2014. Are in vitro estimates of cell diffusivity and cell proliferation rate sensitive
Mutat Res. 752: 10 -- 24.
to assay geometry? J Theor Bio. 356: 71 -- 84.
[31] van den Brenk H.A.S., 1956. Studies in restorative growth processes in mammalian wound healing. Exp Surgery 43: 525 -- 550.
[32] Buck R.C., 1979. Cell migration in repair of mouse corneal epithelium. Invest Ophthalmol Vis Sci 18: 767 -- 784.
[33] Sheardown H., Cheng Y.L., 1996. Mechanisms of corneal epithelial wound healing. Chem Eng Sci. 51: 4517-4529.
[34] Xie Y., Rizzi S.C., Dawson R., Lynam E., Richards S., Leavesley D.I., Upton Z., 2010. Development of a three-dimensional human skin
equivalent wound model for investigating novel wound healing therapies. Tissue Eng Part C: Methods 16: 1111 -- 1123.
[35] Totti S., Allenby M.C., Brito Dos Santos S., Mantalaris A., Velliou E.G., 2018. A 3D bioinspired highly porous polymeric scaffolding system
for in vitro simulation of pancreatic ductal adenocarcinoma. RSC Adv. 8: 20928 -- 20940.
[36] Aronson D.G., Graveleau J., 1993. A selfsimilar solution to the focusing problem for the porous medium equation. Euro J Appl Math. 4:
[37] Angenent S.B., Aronson D.G., 1995. The focusing problem for the radially symmetric porous medium equation. Comm Partial Diff Eq. 20:
[38] Aronson D.G., van den Berg J.B., Hulshof J., 2003. Parametric dependence of exponents and eigenvalues in focusing porous media flows.
[39] Betel´u S.I., Aronson D.G., Angenent S.B., 2000. Renormalization study of two-dimensional convergent solutions of the porous medium
Euro J Appl Math. 14: 485 -- 512.
equation. Physica D. 138: 344-359.
[40] Angenent S.B., Aronson D.G., Betel´u S.I., Lowengrub J.S., 2001. Focusing of an elongated hole in porous medium flow. Physica D. 151:
[41] Simpson M.J., Landman K.A., Hughes B.D., 2010. Cell invasion with proliferation mechanisms motivated by time-lapse data. Physica A.
[42] Angenent S.B., Aronson D.G., 1995. Intermediate asymptotics for convergent viscous gravity currents. Phys Fluids 7: 223-225.
[43] Diez J.A., Thomas L.P., Betel´u, S., Gratton, R., Marino B., Gratton J., Aronson D.G., Angenent S.B., 1998. Noncircular converging flows in
viscous gravity currents. Phys Rev E. 58: 6182 -- 6187.
[44] Zheng Z., Fontelos M.A., Shin S., Dallaston M.C., Tseluiko D., Kalliadasis S., Stone H.A., 2018 Healing capillary films. J Fluid Mech. 838:
65-81.
1217-1240.
228-252.
389: 3779-3790.
404 -- 434.
[45] Zheng Z., Fontelos M.A., Shin S., Stone H.A., 2018 Universality in the nonlinear leveling of capillary films. Phys Rev Fluids 3: 032001(R).
[46] Eggers J., Fontelos M.A. 2009 The role of self-similarity in singularities of partial differential equations. Nonlinearity 22: R1 -- R44.
[47] Eggers J., Fontelos M.A. 2015 Singularities: Formation, Structure, and Propagation, Cambridge University Press.
[48] King J.R., McCue S.W. 2009 Quadrature domains and p-Laplacian growth. Complex Anal Oper Th. 3:453 -- 469.
[49] McCue S.W., King J.R. 2011 Contracting bubbles in Hele-Shaw cells with a power-law fluid. Nonlinearity 24: 613 -- 641.
[50] Dallaston M.C., McCue S.W. 2013 Bubble extinction in Hele-Shaw flow with surface tension and kinetic undercooling regularisation. Non-
[51] Dallaston M.C., McCue S.W. 2016 A curve shortening flow rule for closed embedded plane curves with a prescribed rate of change in
linearity. 26:1639 -- 1665.
enclosed area. Proc R Soc A. 472: 20150629.
[52] Treloar K.K., Simpson M.J., 2013 Sensitivity of edge detection methods for quantifying cell migration assays. PLOS ONE, 8: e67389.
[53] Fernando, A.E., Landman, K.A., Simpson, M.J., 2010 Nonlinear diffusion and exclusion processes with contact interactions. Phys Rev E. 81:
011903.
857 -- 900.
[54] Jin W., Penington C.J., McCue S.W., Simpson M.J., 2016 Stochastic simulation tools and continuum models for describing two-dimensional
collective cell spreading with universal growth functions. Phys Biol. 13: 056003.
[55] Jin W., McCue S.W., Simpson M.J., 2017 Extended logistic growth model for heterogeneous populations. J Theo Biol. 445: 51 -- 61.
[56] Keller E.F., Segel L.A., 1971 Models for chemotaxis. J Theo Biol. 30: 225 -- 234.
[57] Anderson, A.R.A., Chaplain M.A.J., 1998 Continuous and discrete mathematical models of tumor-induced angiogenesis. Bull Math Bio. 60:
[58] Menon S.N., Flegg J.A., McCue S.W., Schugart R.C., Dawson R.A., McElwain D.L.S., 2012 Modelling the interaction of keratinocytes and
fibroblasts during normal and abnormal wound healing processes. Proc R Soc B. 279: 3329 -- 3338.
[59] Johnston S.T., Shah E.T., Chopin L.K., McElwain, D.L.S., Simpson M.J., 2015 Estimating cell diffusivity and cell proliferation rate by
interpreting IncuCyte ZOOMTM assay data using the Fisher-Kolmogorov model. BMC Syst Biol. 9:38.
[60] Simpson M.J., Landman, K.A., Newgreen, D.F., 2006 Chemotactic and diffusive migration on a nonuniformly growing domain: numerical
16
algorithm development and applications. J Comp Appl Math. 192: 282 -- 300.
17
|
1802.02903 | 1 | 1802 | 2018-02-08T14:54:06 | Nonspecific biological effects of weak magnetic fields depend on molecular rotations | [
"physics.bio-ph"
] | The radical pair mechanism is a leading hypothesis in animal magnetic navigation. This mechanism associates the magnetic sense with the visual system, the radical pairs in cryptochromes of the eye retina being specialized magnetic receptors that modulate rhodopsin-mediated photoreception. There are also nonspecific magnetic effects in biology, which occur mostly by chance and originate from the interaction of weak magnetic fields with the magnetic moments dispersed all over the organism at the microscopic level. The radical pair mechanism cannot explain this type of response for many reasons. We have previously shown that the above interaction has a finite probability of resulting in an observable. Here, we develop our physical model of nonspecific magnetic effects for the case of magnetic moments located in rotating molecules. We generalize the results of recent experiments on gene expression in plants in a constant magnetic field, and show that the precession of the magnetic moments that reside on rotating molecules can be slowed relative to the immediate biophysical structures. In quantum mechanical language, the crossing of the quantum levels of magnetic moments conjointly with molecular rotations explain nonspecific magnetic effects and leads to magnetic field-dependences that are in good agreement with the experiment. | physics.bio-ph | physics |
Nonspecific biological effects of weak magnetic fields
depend on molecular rotations
Vladimir N. Binhi1,2,# & Frank S. Prato3,4
1A.M. Prokhorov General Physics Institute, Moscow, Russia
2M.V. Lomonosov Moscow State University, Moscow, Russia
3Lawson Health Research Institute, Ontario, Canada
4University of Western Ontario, Ontario, Canada
#[email protected]
Abstract
The radical pair mechanism is a leading hypothesis in animal mag-
netic navigation. This mechanism associates the magnetic sense with the
visual system, the radical pairs in cryptochromes of the eye retina being
specialized magnetic receptors that modulate rhodopsin-mediated pho-
toreception. There are also nonspecific magnetic effects in biology, which
occur mostly by chance and originate from the interaction of weak mag-
netic fields with the magnetic moments dispersed all over the organism
at the microscopic level. The radical pair mechanism cannot explain this
type of response for many reasons. We have previously shown that the
above interaction has a finite probability of resulting in an observable.
Here, we develop our physical model of nonspecific magnetic effects for
the case of magnetic moments located in rotating molecules. We gen-
eralize the results of recent experiments on gene expression in plants in
a constant magnetic field, and show that the precession of the magnetic
moments that reside on rotating molecules can be slowed relative to the
immediate biophysical structures. In quantum mechanical language, the
crossing of the quantum levels of magnetic moments conjointly with molec-
ular rotations explain nonspecific magnetic effects and leads to magnetic
field-dependences that are in good agreement with the experiment.
Keywords: hypomagnetic field, magnetic moment precession, magnetoreception,
radical pair mechanism, gene expression
Introduction
A large number of different biological effects can be observed at weak magnetic
fields (MFs) in the range 0.1–100 µT, e.g. [1]. For migratory animals that have
formed their specific magnetic sense in the course of evolution, the initial trans-
duction mechanism is apparently associated with spin-correlated radical pairs
in retina cryptochromes, or with the radical pair mechanism (RPM), e.g. [2, 3].
In contrast, there is a so-called nonspecific response to magnetic fields that is
characteristic of all organisms and manifests itself only occasionally, at random
effective combinations of electromagnetic and biochemical/physiological con-
ditions [4]. Nonspecific response differs from magnetoreception in navigating
1
animals, is of fundamental importance, and has its own biophysical mechanisms
that remains largely unknown.
We have recently shown that the effects of a hypomagnetic field (HMF) oc-
cupy a special place among the nonspecific effects due to their higher magnitude
and reproducibility and their capability of providing more precise information
on the origin of the MF effects. An accurate definition of HMF is given in [4]; it
reads mainly that both the static H and the rms value of the variable component
h should be much less than the geomagnetic field (GMF) Hg ∼ 50 µT. Special
terms have also been introduced for those micro objects that react with MF at
the beginning of a signal transduction path and thus can be referred to as the
targets of the MF. These are primary physical targets, i.e., magnetic moments,
and molecular or biophysical MF targets, or sensors that carry the moments
and can change depending on the state of the moments. The existence of HMF
effects can be illustrated in terms of quantum mechanics. The Zeeman sublevels
of a magnetic moment that are split in a MF, degenerate, or cross, provided
their width, of the order of /τ , where τ is the thermal relaxation time and
is the reduced Plank constant, becomes comparable with the Zeeman splitting
γH, where γ is the gyromagnetic ratio. Then a critical MF H ∼ 1/γτ defines
the MF magnitude below which changes should occur at the quantum level.
A physical mechanism for HMF effects has been proposed [5, 6], which con-
siders the dynamics of non-uniformly precessing magnetic moments in biophys-
ical targets, or MF sensors that are not specialized MF receptors. Magnetic ef-
fects occur as the consequence of a significant slowing the precession of magnetic
moments, or, in a quantum picture, their quantum levels crossing. Therefore,
in what follows, the mechanism is referred to as the "level crossing" mechanism
(LCM).
Recent experiments show that a weak static MF affects the expression of
some genes in A. thaliana [7], where a few well-resolved peaks can be observed
in MF-dependences. This work has also shown that a MF reversal produces an
asymmetric response. Similar effects on gene expression have been previously
reported [8].
It seems unlikely that the RPM could explain these results due to experimen-
tal [9, 8, 7, 10] and theoretical [4] reasons. At the same time, similar multi-peak
magnetic response in E. coli [11] had been well described by the interference
mechanism [12] extended to the case of molecular rotations. The interference
mechanism and LCM are cognate: both predict preferred angular positions of
their objects. For this reason, it was relevant to explain the spectral character
of gene expression in A. thaliana and asymmetric response at the MF reversal
in terms of molecular rotations using the LCM theory.
Rotations are ubiquitous at the level of molecular processes, particularly in
many of those related to DNA, RNA, and ATPases. As previously suggested
[13, 11], rotations of the molecules that carry the precessing magnetic moment
can significantly affect the process in which the magnetic moment initiates sub-
sequent biophysical events.
Below we extend the LCM to the case where the molecular surrounding of
magnetic moments rotates with a natural biological speed. The probability of
2
secondary biophysical events in the MF sensors that reside on rotating molecules
is estimated and H-dependences of this probability is calculated. Based on the
comparison with the multi-peak H-dependence [7] for gene expression in A.
thaliana growing in static MF (SMF), a conclusion will be made that rotating
macromolecules are involved in nonspecific response to MF in organisms, and
the molecular MF sensors reside on such macromolecules.
Rotations
It is interesting to note, that reproducible and large nonspecific magnetic effects
are observed in systems with pronounced processes involving gene expression:
neurite outgrowth [14], cephalic regeneration in planarians [15], morphological
changes during embryogenesis [16], response to heat shock [17], cell growth and
gene expression [8] in plants, the proliferation of neuroblastoma cells [18] and
of nerve stem cells [19], gravitropism [20]. The combined action of MF and the
X-ray [21] and of MF and heavy ions [22] can be seen as the interference between
DNA repair and nonspecific magnetic effects. There is also strong dependence
of magnetic effects on the genetic modification of organisms [23, 24, 25, 26, 27].
All the above suggests that gene expression could be a prerequisite for MFs to
elicit an effect.
Transcription and translation are known to be accompanied by the rotation
of macromolecules. For example, ribosome and its parts rotate in the process
of translation [28]. Motion of RNA polymerase and helicase along the DNA
helix is accompanied by relative rotation of the enzyme and DNA. Several ex-
amples of the macromolecular rotations in E. coli cells with speeds from a few
to a few hundred rps are listed in [11]: DNA topoisomerase, DNA and RNA
polymerases, FOF1-ATPases, the flagellar motor. Other contenders for rotating
macromolecules could be: myosin rotation about an actin filament at 1.5–2.5 rps
[29], rotation of microtubules induced by dynein at 1–4 rps [30].
Galland [7] considers chloroplast FOF1 ATPases with a concomitant func-
tioning of V-ATPases most likely MF sensors in plants. Rotation of the chloro-
plast FOF1-ATPase motor, through the redox state of plastids, modulates gene
expression and depends on light, which is in agreement with the observed light-
dependence of magnetic response.
Thus, when modeling magnetic biological effects one must take into account
rotations of the immediate environment of magnetic moments, i.e., the rotations
of the MF sensor itself.
If an organism is experiencing a spatial shift along any axis in the laboratory
frame, then shifted are all its elements such as atoms and molecules together
with their magnetic moments. It might seem that if the body rotated, all of its
elements would rotate with it also. However, this is not the case with respect
to elementary precessing angular momenta and their magnetic moments. Due
to physical laws, rotational motion of the MF sensor body does not transmit
the torque to the precessing angular momentum. For example, while holding
a gyroscope, it is easy to move its body linearly, but it is difficult to roll out
or slow down its rotor - this would require specific movements. Similarly, a
3
magnetic moment precesses mostly independently of the molecular enclosure,
although the moment's thermal relaxation is due to the interaction with it.
If the angular velocity Λ of the rotating body is close to that of the mag-
netic moment precession γH, there will be a situation similar to the temporary
slowing of the moment, in the rotating frame of the body, or a level crossing. A
HMF effect arises, although the MF is not a HMF. This mechanism resembles
a roulette - the ball falls in the cell when the angular velocities of the ball
and of the roulette wheel coincide. Similarly, HMF effect occurs when there is
a coincidence of the angular velocities of the precession and of the sensor body.
In general, an organism may contain various types of sensors that rotate with
several different speeds. In this case, when scanning the dc MF magnitude, the
effect like a HMF effect will occur sequentially for sensors rotating at different
speeds. A quantum interference mechanism of nonspecific magnetic effects that
considers processes involving rotations has been previously proposed in [13] and
developed in [31] p. 266–275. This mechanism predicted that rotations affect the
H-dependences. The a posteriori prediction was in agreement with experiment
[11], where a complex multi-peak H-dependence of a cell culture response to
MF has been observed.
Essentially, it has been shown that the HMF effect at h = H = 0 decreases
for rotating MF sensors. It is easy to deduce that the characteristic feature of
the HMF effect - an abrupt change in the measured value with reducing H to
zero - is merely shifted to some other value of H that depends on the target
rotation speed. In the H-dependence, this generates a kind of "window," where
the magnetic effect can be observed. The ion interference mechanism's [13]
capability of explaining multi-peak SMF dependences is reproduced in the LCM.
At the same time, the LCM allows one to determine the thermal relaxation time
of the primary MF target (magnetic moments) from experimental data, which
is a significant advantage.
Recently we have shown that the magnitude of a hundred different HMF
effects correlate neither with the HMF value, nor with the period of the expo-
sure to HMF, nor with their product, or "dose" [4]. The lack of correlation
strongly suggests that there is no MF sensor with the same magnetic properties
for all organisms. The lack of a general MF sensor and the necessity of gene
expression for nonspecific effects suggest their random nature and their link to
the varying rotations that bring about the dispersal of magnetic properties of
the MF sensors.
Thus, H-dependences of nonspecific effects in the SMF could be a new form
of magnetic spectroscopy capable of measuring the physical characteristics of
molecular processes of transcription and replication.
Important information
about the molecular processes involving rotations can be extracted from the
experiment supported by a quantitative theory capable of explaining such de-
pendences.
4
Mechanism
In [6] we proposed a physical mechanism of nonspecific response to MF in or-
ganisms, which considers a nonuniform precession and thermal relaxation of a
magnetic moment in the MF of parallel dc and ac components - the LCM.
There are no quantum transitions caused by such a MF, and therefore it is suf-
ficient to use the classical model of the Larmor precession. Here we extend the
model to the case of rotating molecular structure that encloses the precessing
magnetic moment.
The LCM assumes that a biological response occurs where, within the re-
laxation time period, the MF disturbs the dynamics of the magnetic moment so
that the deviation from the state of the undisturbed uniform precession becomes
significant. In [6] we derive the following equation of motion in spherical co-
ordinates for a precessing magnetic moment under applied dc and ac magnetic
fields,
ϕ(t) = γHt +
sin (Ωt)
(1)
γh
Ω
where ϕ is the precession phase, or an azimuth angle, γ is the gyromagnetic
ratio, H is the dc MF magnitude, and h and Ω are the amplitude and frequency
of the ac MF. In the absence of the ac MF, at h = 0, a uniform precession takes
place: ϕ(t) = γHt, where γH is the Larmor frequency. For a weak MF effect to
cause a biological response, this background precession should be disturbed. As
shown, an effective disturbance can be achieved either by a significant decrease
in the dc MF or by modulating the dc field with an ac one. In the first case,
the precession stops:
ϕ = 0.
Further, a concept of "reaction" was defined:
it is the change in the state
of the biophysical environment that immediately surrounds the precessing mo-
ment. The probability of the event to occur in the time interval [t− τ /2, t + τ /2]
was assumed to be e.g. [32]
p (t, τ ) = 1 − exp
−
λ (u) du
(2)
(cid:34)
(cid:90) t+τ /2
t−τ /2
(cid:35)
where λ is the density of the Poisson process and τ is the thermal relaxation
time. If magnetic moments precess, their oscillations transform to those in the
rate of downstream events λ. The simplest idealization of this fact is
λ = β [1 + cos (ϕ − ξ)]
(3)
where the proportionality factor β with the dimension of frequency is introduced,
and ξ denotes the random magnetic moment direction that maximizes λ. The
density is minimal when the directions ϕ and ξ are opposite.
Based on Eqs. 1–3, the probability of biophysical events P (H, h, Ω, γ, τ, β) ≡
(cid:104)p(cid:105)t,ξ averaged over t and ξ was derived which depends on six quantities. First
three of these are MF variables, and three others are MF sensor parameters:
the gyromagnetic ratio γ and the parameters τ (relaxation time) and β (mean
temp of the biophysical events initiated by the precessing moments). Difference
5
P (H, h, Ω, ...)− P (H, 0, Ω, ...) describes the probability change under ac/dc MF
exposure and shows the maximum effect at h = 1.8H and Ω = γH, which is in
agreement with many experiments, see a review in [31] p. 307–314.
At h = 0, under a dc MF that is decreasing from the geomagnetic field Hg
to an HMF H, the probability change ∆P ≡ P (H, 0, ...) − P (Hg, 0, ...) equals
∆P (H, γ, τ, β) ≈ − 1
4
(4)
where P (Hg, 0, ...) is assumed to be P (∞, 0, ...). This equation explains the
HMF effect, that occurs when the dc MF is decreased from Hg to HMF H, see
also Fig. 6 of [6].
β2τ 2e−βτ sinc2 (γHτ /2)
Below, in order not to complicate the model by including all possible types
of MF exposure, only this case of a constant MF will be considered. For conve-
nience in what follows, the HMF effect will be associated with −∆P , i.e., with a
positive magnitude. It will be shown that rotations of MF sensors shift the MF
at which a biological effect gets an abrupt change from the range of MFs close
to zero to higher MFs. We will refer to this as a static magnetic field (SMF)
effect. Extension of the theory above explained to a rotating MF sensor requires
modification of Eq. 3.
First, besides β - the undisturbed rate of biophysical events generation -
another parameter, the depth of modulation of β should be introduced. How-
ever, the value of this additional parameter would be determined by a specific
biophysical mechanism indicating the type and characteristics of the MF sensor,
whereas our model, being a physical one, should be formulated independently
of biophysical details. Therefore, we leave this parameter equal to unity and
note that the absolute value of the actual magnetic effects can be either larger
or smaller than those calculated. Thus, the purpose of the theory below is to
calculate the MF-dependencies rather than the magnitude of a magnetic effect.
Only the specific form of such dependencies can be compared with experiment.
Second, let m be a unit vector in the direction of a precessing magnetic
moment and b one in the direction that is associated with the biophysical en-
vironment. This vector is defined so that λ acquires a maximum value when m
points along b. Then, a generalization of Eq. 3 includes a scalar product of m
and b:
λ = β(1 + mb)
(5)
Let now n ≡ Λ/Λ be the unit vector of the MF sensor rotation, and the
Cartesian coordinates are chosen so that axis z is directed along H and axis x
is in the plane formed by vectors H and n, Fig. 1. Let the target rotate about
n with angular frequency Λ, which means a rotation of vector b so that ξ = Λt.
Note that due to subsequent time averaging and practical incommensurability
of the rates of precession and rotation, their phases are not significant, and we
assume the vectors m and b at t = 0 be in the xz plane. Note also that for
a convenient definition of b, we use angle α as one between b at t = 0 and n
rather than its polar angle. Vector m of the magnetic moment precesses about
H so that ϕ = γHt, Eq. 1.
6
Figure 1: Unit vectors and their polar angles in a spherical coordinate system.
Dash-line circles are the tracing of the precession of vector m and the rotation
of vector b around n; vector b is shown at t = 0.
With the above notation, one can find the Cartesian components of
vectors m and b: mx = sin(θ) cos(γHt), my = sin(θ) sin(γHt), mz =
cos(θ), bx = cos(α) sin(η) − sin(α) cos(η) cos(Λt), by = − sin(α) sin(Λt), bz =
cos(α) cos(η) + sin(α) sin(η) cos(Λt). Then the density of biophysical events
λ = λ(t, β, θ, η, α, γ, H, Λ) initiated by precession can be calculated from Eq. 5.
Substitution of λ into Eq. 2 gives a result that needs to be further averaged over
time and random variables. One should keep in mind that random variables θ
and η, α are of different type with regard to averaging, for the following reasons.
The polar angle θ is that of the magnetic moment vector m at t = 0. The
orientation of this vector should be considered random for each MF sensor sep-
arately. This means the averaging over θ should be performed unconditionally.
The initial orientation of m is arbitrary in a full solid angle, hence the result gen-
erally must be averaged over the azimuth and polar angles. However, the result
does not depend on the azimuth angle, and only the unconditional averaging
over θ remains.
In contrast, the positions of the rotation vector n and target vector b at
t = 0 given by angles η and α respectively, Fig. 1, has a definite value for
each specific target. Averaging over these angles makes sense only if they have
variable random values in different targets. Note that this is not the case if
the targets of the same type have a predominant orientation. For example,
many plant cells are oriented in a certain way relative to the gravity vector.
Consequently, rotations of macromolecules carrying MF sensors could inherit
this preferred orientation in the form of a more or less definite orientation of
the vector n and vector b at t = 0. Thus, it is relevant to study two cases, (i)
deterministic and (ii) stochastic with uniformly distributed random values of η
and α.
Substituting Eq. 5 in Eq. 2 and performing averaging over time and θ, one
can derive the probability of the secondary events initiated by a precessing
7
magnetic moment, i.e., P = P (H, γ, τ, β, Λ, η, α),
(cid:32)
(cid:90) t+τ /2
t−τ /2
(cid:33)(cid:35)
λ du
dt sin(θ)dθ
(cid:90) π
(cid:90) T
(cid:34)
0
0
P =
1
πT
1 − exp
−
where T ≡ 2π/γH − Λ is the period of a two-frequency oscillating process.
Finally, we arrive at an expression for the probability change that is suitable for
comparison with experiment,
∆P (H, γ, τ, β, Λ, η, α) ≡ P (H, γ, ...) − P (∞, γ, ...)
(6)
As is seen, three new variables are added to ∆P as compared with the case of
fixed MF sensors, see Eq. 4 - it is the speed Λ of the MF sensor rotation and
the angles that define axis of rotation n and vector b of the sensor.
As an analytical evaluation of Eq. 6 would be too cumbersome, a few different
cases have been studied numerically. SMF effect has been calculated in different
modes at β = 2.
Dependences on γH at η = 0 and α = π/2 show: a) a resonance-like re-
sponse, Λ = 0; b) the asymmetry of the response regarding a MF reversal,
Λ = 10 and 15, τ = 1. Dependences at τ = 1, γH = 10, and Λ = 10 on α with
η = 0 and on η with α = π/2 show relatively wide ranges of angles α and η
where magnetic effect exist.
If Λ = 0 or the vector b is parallel to n, i.e., α = 0, this obviously comes to
the case without rotation. The half-width of peak at H = 0 is consistent with
Eq. 4: half-width is defined by the argument of the cardinal sine function where
it rapidly changes, i.e., ∆H ∼ 1/(γτ ).
In case of rotation, the most pronounced result occurs where the axis of
rotation coincides with the z-axis and the MF sensor vector b is perpendicular
to the z-axis, that is η = 0 and α = π/2. Then, the probability of biophysical
events that are caused by precessing magnetic moments has a resonance-like
peak provided the angular velocity of rotation is in certain relation with the MF
vector: Λ ∼ −γH. The position of the peak shifts proportionate to Λ. This
means that the level crossing, or a slow precession, occurs at values H ∼ Λ/γ.
Note that the effect is not symmetric with respect to the MF reversal, which
can be directly tested in experiment.
The γH-position of the peak of magnetic responce is independent of the
values of all variables other than Λ. This enables one to study the η- and α-
dependences that are mentioned above. While the effect is in its maximum, one
can examine how it depends on the orientation of the axis of rotation n and
that of the sensor vector b. These dependences are rather smooth. This leaves a
chance to observe the SMF effect even for arbitrary or unknown values of these
angles.
As was said above, the rotation axes of macromolecules can inherit a pre-
ferred orientation of cells; then the values of η and α can be considered definite.
How could the result change, where these values are random rather than defi-
nite? Let both angles, as polar ones, be distributed in the range [0, π). Since
8
the position of the peak in ∆P (γH) does not depend on these variables, it is
sufficient to average the effect magnitude over these angles only in the peak.
The result shows that the peak height decreases by more than an order of
magnitude, from 7% to about 0.5%. The latter is an order of magnitude larger
than the RPM effect observed in the GMF-like MFs. However, this value is still
small for reliable explanation of the SMF biological effects. This means that for
the SMF effect to occur, some kind of rotation ordering is desirable. However it
is not necessary for non-rotating MF sensors that show their 14 % HMF effect
(as follows from Eq. 4 at τ β = 2) independently of molecular rotations.
Discussion
Although there are more than two hundred articles, documenting HMF effects in
organisms [4], these data have not, in general, investigated the MF-dependence
needed for comparison to the predictions of the LCM or LCMr - its "rotation"
extension. However, recently, a set of MF stimulus-response curves for gene ex-
pression has been obtained in a study of plant germination in MFs ranging from
about 0.5 to 188 µT [7], making it possible to compare theory and experiment.
The experimental data demonstrate at least two well-resolved peaks. In order
to fit these data, we have assumed that the MF sensors of the same type ro-
tate at two different speeds, thus forming two peaks in the H-dependence. The
sum of their equal contributions has been calculated from Eq. 6 with the values
of variables that provide maximum effect and assuming magnetic moments of
the same type precess inside biophysical structure/structures rotating at two
speeds, 46 and 116 µT/γ.
The result shows a good ageement with experimental H-dependence of the
relative transcript amount of rbcl (large subunit of ribulose bisphosphate car-
boxylase/oxygenase) in seedlings of A. thaliana raised for 120 h under broad-
band blue light. Experimental data were kindly provided by P. Galland; article
[7] contains information on the methods, magnetic exposure, and conditions for
this and similar experiments with other genes and strains.
Two general features of gene expression that have been observed experimen-
tally in germinating plants are essential for a theoretical discourse. This is 1) the
presence of resonance-like peaks in H-dependences of the SMF effect, and 2) the
absence of a symmetry in the response with regard to SMF reversal [7]. These
features are observed in the H-dependences of the expression of a few genes in
a few A. thaliana strains. As our results demonstrate, the LCM modified to
molecular rotations is able to describe, if not to explain, these key features.
Based on the above H-dependence and the LCMr predictions, one might
speculate on what is the primary MF target.
As follows from the above formulae, two relations should be satisfied in order
for this theory to be consistent with the experiment:
γHp ∼ Λ, γτ ∆H ∼ 1
(7)
where Hp is the location and ∆H is the half-width of a peak in the H-dependence
of the probability of biophysical events initiated by the precessing magnetic
9
moment at the sensor rotation with angular speed Λ. As follows from the
comparison of experiment with theory, exemplary experimental values for Hp
and ∆H are about 0.5 and 0.1 G, or 50 and 10 µT, respectively. Could this
connection of theory with experiment reveal characteristics or the nature of the
primary physical target? To do this, one can test Eqs. 7, while substituting γ
and calculating τ and Λ. The choice is spinning or orbiting electrons, protons
or magnetic isotopes, bound ions, and molecular gyroscopes [33].
For an electron spin magnetic moment (γ = 1.76× 107 rad G−1s−1), thermal
relaxation time, as follows from Eqs. 7, is a few times greater than 100 ns -
a maximum expected for relevant electrons in a wet tissue under physiological
temperature. However, this would require the MF sensor to rotate at about
1.5 × 106 rps, which is not realistic.
A proton (γ = 2.68 × 104 rad G−1s−1) and proton-like spin magnetic mo-
ments would require rotation speeds of about 2×103 rps, which is more realistic,
and relaxation time of about 0.5 ms, which is acceptable for the spin of bound
protons. However, a proton's influence on the immediate environment is ex-
pected to be weaker than that of electrons, as protons are less mobile and carry
much smaller magnetic moment.
The involvement of orbiting bound ions, e.g., Ca2+ (γ = 241 rad G−1s−1),
as a primary physical target is not realistic. Although in this case Eqs. 7 give
Λ ∼ 20 rps, τ should be too large, about 30 ms, which is impossible due to the
fact that this time is mostly picoseconds in the order of magnitude.
Finally, a big molecular rotor, like a GLU residue (γ ≈ 70 rad G−1s−1),
would require the speed of rotation of about 6 rps and the thermal relaxation
time 0.1 s. This is more likely, although not without difficulty, as a relatively
large cavity of 1.5-nm radius and free of water molecules is needed to house the
rotating group inside the folding protein [31, 33].
Evidently, H-dependencies are not yet sufficient to identify the nature of the
primary targets. Probably, the theory should take into account the often inter-
mittent character of molecular rotations, like in RNA and ATPase. In addition,
Ω-dependencies obtained in the same organism under the ac/dc MF exposure,
as explained in [4], could provide direct information on the gyromagnetic ratio
of the primary targets.
LCMr is a general mechanism that explains nonspecific response to MF
regardless of the biophysical construct that hosts a precessing magnetic moment.
Other mechanism that take into account the biophysical medium is molecular
gyroscope mechanism [33] that could explain the above discussed effects [7].
Future experiments are needed to discriminate between two molecular rotor
mechanisms - LCMr and the gyroscopic one - regarding their involvement in
formation of the multi-peak H-dependence of the nonspecific response.
Studies on the HMF effects are important for future space flights that are
featured by MFs more than thousand-fold smaller than the GMF. For this rea-
son, when studying the magnetic effects on Earth, researchers model the space
conditions by correcting for gravity. This is usually achieved with clinostats
that rotate samples so that the gravity vector in the frame of the sample is
averaged to zero. Due to the influence of rotations on the nonspecific magnetic
10
effects, the widely accepted interpretation of the results obtained in clinostats
should be revised.
In summary: (i) molecular rotations are a significant factor affecting nonspe-
cific magnetic effects in organisms, (ii) the Level Crossing Mechanism as applied
to rotations, or LCMr, that we have introduced here explains key features of the
observed MF-dependences including resonance-like peaks and asymmetry with
regard to the SMF reversal, (iii) further insight into the nature of nonspecific re-
sponse to MF can be provided by multi-peak H-dependences of biological effects
under controlled outer rotations and, if in plants, under exposure to HMF/SMF
of different orientation with respect to the gravity vector, and (iv) fundamental
biological information on the molecular rotations can be obtained from the shift
of the peaks. The non-invasive extraction of such fundamental information on
the rotation of sub-cellular structures introduces the potential to use LCMr as
a new biological spectroscopy.
References
[1] T. K. Breus, V. N. Binhi, and A. A. Petrukovich. Magnetic factor of
the solar terrestrial relations and its impact on the human body: physical
problems and prospects for research. Physics–Uspekhi, 59(5):502–510, 2016.
[2] K. Schulten, C. Swenberg, and A. Weller. A biomagnetic sensory mecha-
nism based on magnetic field modulated coherent electron spin motion. Z.
Phys. Chem., 111(1):1–5, 1978.
[3] P. J. Hore and H. Mouritsen. The radical-pair mechanism of magnetore-
ception. Annual Review of Biophysics, 45(1):299–344, 2016.
[4] V. N. Binhi and F. S. Prato. Biological effects of the hypomagnetic field:
An analytical review of experiments and theories. PLoS ONE, 12(6):1–51,
2017.
[5] V. N. Binhi. A primary physical mechanism of the biological effects of weak
magnetic fields. Biophysics, 61(1):170–176, 2016.
[6] V. N. Binhi and F. S. Prato. A physical mechanism of magnetoreception:
extension and analysis. Bioelectromagnetics, 38(1):41–52, 2017.
[7] S. K. Dhiman and P. Galland. Effects of weak static magnetic fields on
seedlings of Arabidopsis thaliana: Gene expression: cryptochromes versus
molecular rotors as potential magnetoreceptors. To be published, 2018.
[8] C. M. Bertea, R. Narayana, C. Agliassa, C. T. Rodgers, and M. E. Maffei.
Geomagnetic field (Gmf) and plant evolution: Investigating the effects of
Gmf reversal on Arabidopsis thaliana development and gene expression.
Journal of Visualized Experiments, 105(e53286), 2015.
11
[9] S. R. Harris, K. B. Henbest, K. Maeda, J. R. Pannell, C. R. Timmel,
P. J. Hore, and H. Okamoto. Effect of magnetic fields on cryptochrome-
dependent responses in Arabidopsis thaliana.
J. R. Soc. Interface,
6(41):1193–1205, 2009.
[10] Anne Hoyto, Mikko Herrala, Jukka Luukkonen, Jukka Juutilainen, and
Jonne Naarala. Cellular detection of 50 hz magnetic fields and weak blue
light: effects on superoxide levels and genotoxicity. International Journal
of Radiation Biology, 93(6):646–652, 2017.
[11] V. N. Binhi, Ye. D. Alipov, and I. Ya. Belyaev. Effect of static magnetic
field on E. coli cells and individual rotations of ion-protein complexes. Bio-
electromagnetics, 22(2):79–86, 2001.
[12] V. N. Bingi. Mechanism of magnetosensitive binding of ions by certain
proteins. Biofizika, 42(2):338–342, 1997.
[13] V. N. Binhi. Amplitude and frequency dissociation spectra of ion-protein
complexes rotating in magnetic fields. Bioelectromagnetics, 21(1):34–45,
2000.
[14] C. F. Blackman, S. G. Benane, and D. E. House. Evidence for direct effect
of magnetic fields on neurite outgrowth. The FASEB Journal, 7(9):801–806,
1993.
[15] K. A. Jenrow, C. H. Smith, and A. R. Liboff. Weak extremely-low-frequency
magnetic fields and regeneration in the planarian Dugesia tigrina. Bioelec-
tromagnetics, 16(2):106–112, 1995.
[16] J. M. R. Delgado, J. Leal, J. L. Monteagudo, and M. G. Gracia. Embry-
ological changes induced by weak, extremely low frequency electromagnetic
fields. Journal of Anatomy, 134(3):533–551, 1982.
[17] F. S. Prato. Non-thermal extremely low frequency magnetic field effects
on opioid related behaviors: snails to humans, mechanisms to therapy.
Bioelectromagnetics, 36(5):333–348, 2015.
[18] W.-C. Mo, Y. Liu, P. F. Bartlett, and R.-Q. He. Transcriptome profile of
human neuroblastoma cells in the hypomagnetic field. Science China – Life
Sciences, 57(4):448–461, 2014.
[19] J.-P. Fu, W.-C. Mo, Y. Liu, and R.-Q. He. Decline of cell viability and
mitochondrial activity in mouse skeletal muscle cell in a hypomagnetic field.
Bioelectromagnetics, 37(4):212–222, 2016.
[20] N. A. Belyavskaya. Biological effects due to weak magnetic field on plants.
Advances in Space Research, 34(7):1566–1574, 2004.
12
[21] P. Politanski, E. Rajkowska, M. Brodecki, A. Bednarek, and M. Zmyslony.
Combined effect of X-ray radiation and static magnetic fields on reactive
oxygen species in rat lymphocytes in vitro. Bioelectromagnetics, 34(4):333–
336, 2013.
[22] V. M. Lebedev, G. V. Maksimov, E. G. Maksimov, V. Z. Paschenko, A. V.
Spassky, K. A. Trukhanov, and G. V. Tsoraev. Using a 120-cm cyclotron to
study the synchronous effects of ionizing radiation and hypomagnetic con-
ditions on the simplest biological objects. Bulletin of the Russian Academy
of Sciences: Physics, 78(7):626–629, 2014.
[23] Ye. D. Alipov and I. Ya. Belyaev. Difference in frequency spectrum
of extremely-low-frequency effects on the genom conformational state of
AB1157 and EMG2 E. coli cells. Bioelectromagnetics, 17(5):384–387, 1996.
[24] C. Celestino, M. L. Picazo, M. Toribio, J. A. Alvarez-Ude, and J. L. Bar-
dasano. Influence of 50 Hz electromagnetic fields on recurrent embryogen-
esis and germination of cork oak somatic embryos. Plant Cell, Tissue and
Organ Culture, 54(1):65–69, 1998.
[25] W.-C. Mo, Y. Liu, H. M. Cooper, and R.-Q. He. Altered development
of Xenopus embryos in a hypogeomagnetic field. Bioelectromagnetics,
33(3):238–246, 2012.
[26] W.-C. Mo, Z.-J. Zhang, Y. Liu, P. F. Bartlett, and R.-Q. He. Magnetic
shielding accelerates the proliferation of human neuroblastoma cell by pro-
moting G1-phase progression. PLoS One, 8(1):e54775, 2013.
[27] C. Xu, Y. Yu, Y. Zhang, Y. Li, and S. Wei. Gibberellins are involved in
effect of near-null magnetic field on arabidopsis flowering. Bioelectromag-
netics, 38(1):1–10, 2017.
[28] P. Xie. Model of ribosomal translocation coupled with intra- and inter-
subunit rotations. Biochemistry and Biophysics Reports, 2:87–93, 2015.
[29] I. Sase, H. Miyata, S. Ishiwata, and K. Kinosita. Axial rotation of sliding
actin filaments revealed by single-fluorophore imaging. Proceedings of the
National Academy of Sciences, 94(11):5646–5650, 1997.
[30] R. D. Vale and Y. Y. Toyoshima. Rotation and translocation of mi-
crotubules in vitro induced by dyneins from Tetrahymena cilia. Cell,
52(3):459–469, 1988.
[31] V. N. Binhi. Magnetobiology: Underlying Physical Problems. Academic
Press, San Diego, 2002.
[32] S. M. Ross. Stochastic Processes. Wiley, New York, 1996.
[33] V. N. Binhi and A. V. Savin. Molecular gyroscopes and biological ef-
fects of weak extremely low-frequency magnetic fields. Physical Review E,
65(5):051912, 2002.
13
|
1708.08887 | 1 | 1708 | 2017-08-23T22:54:52 | Are there optical communication channels in the brain? | [
"physics.bio-ph",
"physics.optics",
"q-bio.NC",
"quant-ph"
] | Despite great progress in neuroscience, there are still fundamental unanswered questions about the brain, including the origin of subjective experience and consciousness. Some answers might rely on new physical mechanisms. Given that biophotons have been discovered in the brain, it is interesting to explore if neurons use photonic communication in addition to the well-studied electro-chemical signals. Such photonic communication in the brain would require waveguides. Here we review recent work [S. Kumar, K. Boone, J. Tuszynski, P. Barclay, and C. Simon, Scientific Reports 6, 36508 (2016)] suggesting that myelinated axons could serve as photonic waveguides. The light transmission in the myelinated axon was modeled, taking into account its realistic imperfections, and experiments were proposed both in-vivo and in-vitro to test this hypothesis. Potential implications for quantum biology are discussed. | physics.bio-ph | physics |
Are there optical communication channels in the brain?
Parisa Zarkeshian1, Sourabh Kumar1, Jack Tuszy´nski2,3, Paul Barclay1,4, and Christoph
Simon 1
1 Institute for Quantum Science and Technology and Department of Physics and Astronomy, University of Calgary, Calgary T2N
1N4, Alberta, Canada
2 Department of Oncology, University of Alberta, Cross Cancer Institute, Edmonton T6G 1Z2, Alberta, Canada
3 Department of Physics, University of Alberta, Edmonton T6G 2E1, Alberta, Canada
4 National Institute for Nanotechnology, Edmonton T6G 2M9, Alberta, Canada
TABLE OF CONTENTS
1 ABSTRACT
2
INTRODUCTION
3 Results
3.1 Optical transmission in myelinated axons
3.1.1
3.1.2
3.1.3
3.1.4
3.1.5
Transmission in nodal and paranodal regions
Transmission in bends
Transmission in presence of varying cross-sections
Transmission in non-circular cross sections
Transmission in presence of other imperfections
3.2 Absorption
3.3 Attainable transmission
3.4 Attainable communication rates
Proposals to test the hypothesis
3.5
4 Discussion
5 Acknowledgment
6 References
1. ABSTRACT
Despite great progress in neuroscience, there are still fundamental unanswered questions about the brain, including the
origin of subjective experience and consciousness. Some answers might rely on new physical mechanisms. Given that biophotons
have been discovered in the brain, it is interesting to explore if neurons use photonic communication in addition to the well-studied
electro-chemical signals. Such photonic communication in the brain would require waveguides. Here we review recent work [S.
Kumar, K. Boone, J. Tuszynski, P. Barclay, and C. Simon, Scientific Reports 6, 36508 (2016)] suggesting that myelinated axons
could serve as photonic waveguides. The light transmission in the myelinated axon was modeled, taking into account its realistic
imperfections, and experiments were proposed both in vivo and in vitro to test this hypothesis. Potential implications for quantum
biology are discussed.
2.
INTRODUCTION
Over the past decades a substantial number of facts has been discovered in the field of brain research. However, the
fundamental question of how neurons, or more specifically all particles involved in the biological processes in the brain, contribute
to mental abilities such as consciousness is still unanswered. The true explanation to this question might rely on physical processes
other than those that have been discovered so far. One interesting candidate to focus on is biophotons, which might serve as
supplementary information carriers in the brain in addition to the well established electro-chemical signals.
Biophotons – which are photons ranging from near-IR to near-UV frequency and emitted without any enhancement or
excitation– have been observed in many organisms such as bacteria (1), fungi (2), germinating seeds (3), plants (4), animal tissue
1
Are there optical communication channels in the brain?
cultures (5), and different parts of the human body (6–9), including the brain (10–15). These biophotons are produced by the decay
of electronically excited species which are created chemically during oxidative metabolic processes (16, 17) and can contribute to
communication between cells (18). Moreover, several experimental studies show the effects of light on neurons' and, generally, the
brain's function (19–21). The existence of biophotons and their possible effects on the the brain along with the fact that photons
are convenient carriers of information raises the question whether there could be optical communication in the brain.
For the sources and detectors of the optical communication process in the brain, mitochondrial respiration (22, 23) or
lipid oxidation (24), and centrosomes (25) or chromophores in the mitochondria (26) have been proposed, respectively. It has
also been observed that opsins, photoreceptor protein molecules, exist in the brains of birds (27, 28), mammals (29–32), and more
general vertebrates (33) and even in other parts of their bodies (34, 35) as well.
Another essential element for this optical communication, which is not well established yet, is the existence of physical
links to connect all of these spatially separated agents in a selective way. In the dense and (seemingly) disordered environment of
the brain, waveguide channels for traveling photons would be the only viable way to achieve the targeted optical communication
processes. Mitochondria and microtubules in neurons have been introduced as the candidates for such waveguides (36–39). How-
ever, they are not suitable in reality due to their small and inhomogeneous structure for light guidance over proper distances in the
brain.
Ref. (40) proposed myelinated axons as potential biophoton waveguides in the brain. The proposal is supported by a
theoretical model and numerical results taking into account real imperfections. Myelin sheath (formed in the central nervous system
by a kind of glia cell called oligodendrocyte) is a lamellar structure surrounding the axon and has a higher refractive index (41)
than both the inside of the axon and the interstitial fluid outside (see Fig. 1a) which let the myelin sheath to guide the light inside
itself for optical communications. This compact sheath also increases the propagation speed of an action potential (via saltatory
conduction) based on its insulating property (42). There has been a few indirect experimental evidence for light conduction by
axons (12, 43, 44). Another related and interesting experiment has shown that a certain type of glia cells, known as Muller cells,
guide light in mammalian eyes (45, 46). Ref. (40) also proposed experiments to test the existence of the optical waveguides in the
brain.
One interesting property of optical communication channels is that they can also transmit quantum information. Quan-
tum effects in biological systems are being studied in different areas such as photosynthesis (47, 48), avian magnetoreception
(49, 50), and olfaction (51, 52). There is an increasing number of conjectures about the role of quintessential quantum features such
as superposition and entanglement (53) in the brain (15, 38, 54–56). The greatest challenge when considering quantum effects in
the brain or any biological system in general is environmentally induced decoherence (57), which leads to the suppression of these
quantum phenomena. However, some biological processes can be fast and may show quantum features before they are destroyed
by the environment. Moreover, nuclear spins can have coherence times of tens of milliseconds in the brain (58, 59). A recent
proposal on "quantum cognition" suggests even longer coherence times of nuclear spins (56), but relies on quantum information
transmission via molecule transport, which is very slow. In contrast, photons are the fastest and most robust carriers for quantum
information over long distances, which is why currently man-made quantum networks rely on optical communication channels
(typically optical fibers) between spins (60, 61).
3. Results
To show that myelinated axons could serve as the waveguides for traveling biophotons in the brain, Ref. (40) solved the
three dimensional electromagnetic field equations numerically in different conditions, using Lumerical's software packages FDTD
(Finite Difference Time Domain) Solutions and MODE Solutions. These software packages solve Maxwell's equations numerically,
allowing the optical properties of dielectric structures defined over a mesh with subwavelength resolution to be simulated.
The refractive indices of the fluid outside of the axon, the axon, and the myelin sheath were taken close to 1.34, 1.38
and 1.44 respectively (see Fig. 1a), which are consistent with their typical values (41, 62, 63). These indexes let the myelin sheath
guide the light inside itself. The ratio of the radius of the axon, r to the outer radius of the myelin sheath r(cid:48) (g-ratio) is taken equal
to 0.6 for the most of the simulations, close to the experimental values (64). In reality, the radius of the myelinated axons in the
brain changes from 0.2 microns to close to 10 microns (65). For the purpose of guiding light inside the myelin sheath, Ref. (40)
considered the wavelength of the observed biophotons in the brain which is from 200 nm to 1300 nm. Since several proteins in
the environment of the axons strongly absorb at wavelengths close to 300nm, a wavelength range of the transmitted light from the
shortest permissible wavelength, λmin = 400nm, to the longest one, λmax, was chosen to avoid the absorption and confine the light
well in the myelin sheath. λmax is chosen to the upper bound of the observed biophoton wavelength (1300 nm) or the thickness of
the myelin sheath (denoted by d), whichever is smaller. Besides λmin and λmax, an intermediate wavelength was considered, denoted
by λint, corresponding to the central permissible frequency (mid-frequency of the permissible frequency range) in the simulations.
In the following section we discuss the guided modes in the myelinated axons and their transmissions in nodal and para-
nodal regions and even in the presence of the imperfections such as bends, varying cross-sections, and non-circular cross-sections.
2
Are there optical communication channels in the brain?
Fig. 1: Simplified depiction of a segment of a neuron, and the cylindrically symmetric eigenmode of a myelinated axon. (a)
Structure of a segment of a neuron which is cut longitudinally near the end of the segment. Each layer of the compact myelin sheath
(shown in red) ends in the cytoplasm filled loops (shown in light red) in the paranodal region close to a Node of Ranvier. The inset
indicates the cross section in the transverse plane with r and r(cid:48) as the inner and outer radii of the myelin sheath. d is the thickness
of the myelin sheath, and nmy, nax, and next are the refractive indices of the myelin sheath, the inside of the axon, and the interstitial
fluid outside, respectively. (b) Electric field magnitude of a cylindrically symmetric eigenmode (with wavelength λ = 0.612 µm) for
a myelinated axon (with r=3 µm, and r(cid:48)=5 µm). (c) Electric field vector at different points for displaying the azimuthal polarization
of the input mode. The adjacent color bar shows the field magnitude. Figure from Ref. (40).
3.1. Optical transmission in myelinated axons
Within the neuron, one can identify numerous intra-cellular structures that can function as potential scatterers, i.e.
sources of waveguide loss. They are located both inside the axon and outside of the axon. Intra-cellular structures include cell
organelles, for example, mitochondria, the endoplasmic reticulum, lipid vesicles, as well as the many filaments of the cytoskeleton,
namely microtubules, microfilaments and neurofilaments. Extra-cellular structures include microglia, and astrocytes. However,
the electromagnetic modes which are spatially confined within the myelin sheath, should not be affected by the presence of these
structures. These biophoton modes considered here would be able to propagate in a biological waveguide provided its dimension
is close to or larger than the wavelength of the light. Fig. 1b shows the numerically calculated magnitude of the electric field
of a cylindrically symmetric eigenmode of an axon with radius r = 3µm and myelin sheath radius r(cid:48) = 5µm for the wavelength
0.612µm. This electric field is azimuthally polarized as depicted is Fig. 1c and it is similar to the TE01 mode of a conventional fiber
(66) which has higher refractive index of the core than that of the cladding. It is important to note that azimuthal polarization would
prevent modal dispersion in the birefringent myelin sheath. Importantly, its optical axes are oriented radially (67). It can be readily
established that there are hundreds of potential guided modes allowed to exist given the thickness of myelin sheath. Consequently,
biophotons that could be generated by a source in the axons (e.g. mitochondria or recombination of reactive oxygen species) could
readily interact with these modes as determined by mode-specific coupling coefficients. While we lack detailed knowledge of
the particulars for these interactions, for the sake of simplicity and ease of illustration we select a single mode and examine its
transmission. It is interesting to analyze transmission in the presence of optical imperfections such as discontinuities, bends and
varying cross-sectional diameters. In this connection, we simulated short axonal segments due to computational limitations and
extrapolated the results for the full length of an axon.
3.1.1. Transmission in nodal and paranodal regions
A myelinated axon has periodically unmyelinated segments, called Nodes of Ranvier, which are approximatly 1µm long
(68) (while the whole axon length varies from 1 mm to the order of a meter). Here, we discuss the transmission in the Ranvier nodes
3
Are there optical communication channels in the brain?
Fig. 2: Nodal and paranodal regions and their transmissions.
(a) Longitudinal cross-section of the nodal and paranodal
regions in the model of Ref. (40). Here, the ratio of the radius of the axon and the outer radius of the myelin sheath, g-ratio, is
equal to 0.6, where the outer radius of the myelin sheath is chosen r(cid:48) = 5 µm, and the length of the paranode, lparanode = 5 µm.
(b) Magnitude of the electric field profile in the longitudinal direction (EFPL) when a cylindrically symmetric input mode with
wavelength 0.612 µm passes through the nodal and paranodal region. (c) Transmission percentage of an axon with the outer radius
r(cid:48) = 5 µm, as a function of the p-ratio (the ratio of the paranodal length and the thickness of the myelin sheath). (d)-(f) Transmission
percentage as a function of the axon radius for three different wavelengths of the input mode and three different lengths of paranode.
Figure from Ref. (40).
and at the edges of the nodes, the paranodes. The configuration of myelin sheath is special in the paranodal regions (see Fig. 2a).
There are many layers making up the compact myelin sheath and at the edge of each node, almost all of the layers are in contact
with the core (bared axon) with a small pocket of cytoplasm. That's because each layer moving from the innermost outward is
longer than the one below. However, for thick myelin sheaths, many cytoplasmic pockets cannot reach the surface of the bare axon,
but end on inner layers. Thus, the length of paranodal regions is dependent on the thickness of the myelin sheath. We call the ratio
of the length of paranode, lparanode, to the thickness of the myelin sheat, d, p-ratio and take its value close to 5 in our simulations
based on the realistic values (69).
Fig. 2a displays the model of Ref. (40) for two adjacent paranodal regions with the node in between, and Fig. 2b shows
the magnitude of the electric field profile in the longitudinal direction (along the length of the axon), EFPL, as a cylindrically
symmetric input mode crosses this region. Fig. 2c shows the power transmission in the guided modes as a function of p-ratio for
three wavelengths, 0.40 µm, 0.61 µm, and 1.30 µm. For the transmission, there are two main losses: divergence or scattering of the
light beam. Shorter wavelengths scatter more but diverge less. Thus, in Fig. 2c, for small p-ratios, shorter wavelengths have higher
transmission and as the effect of divergence is dominant in this region and the shorter wavelengths diverge less. However, for the
large p-ratios, the effect of scattering is dominant and since the higher wavelengths scatter less, and have a higher transmission.
Fig. 2d–f compares the transmission percentage for different axon radii, wavelengths, and p–ratios. Although in Fig. 2d,
the behavior of the transmission as a fuction of axon radius is independent of p–ratios for the longest permissible wavelength, it
4
bcMyelin sheathAxon5 mParanodal regionNode of RanvierCytoplasmic loops (with microtubules)Myelin sheathAxon5 mParanodal regionNode of Ranvier10 m6 mCytoplasmic loops (with microtubules)Incident lighta paranodel=longestpermissibleTransmission(%)■■■■■◆◆◆◆◆1234575808590lparanode/d=2.5■lparanode/d=5◆lparanode/d=7.5Radiusoftheaxonincludingthemyelinsheath(r'inµm)=correspondingtothecentralpermissiblefrequencyTransmission(%)■■■■■◆◆◆◆◆1234520406080100lparanode/d=2.5■lparanode/d=5◆lparanode/d=7.5Radiusoftheaxonincludingthemyelinsheath(r'inµm)=0.4mTransmission(%)■■■■■◆◆◆◆◆1234520406080100lparanode/d=2.5■lparanode/d=5◆lparanode/d=7.5Radiusoftheaxonincludingthemyelinsheath(r'inµm)de fTransmission(%)■■■■◆◆◆◆123456720406080100=1.30m■=0.612m◆=0.40mp-ratio=lparanode/d Are there optical communication channels in the brain?
can be concluded that for the most loosely confined modes (λmax) transmission increases in thicker axons. It's also possible that
for long wavelengths, a fraction of the light diverging into the axon comes back into the myelin sheath at the end of the paranodal
region and not all the light that diverges is lost. This can be an explanation for not well-defined dependency of the transmission on
the paranodal lengths (see Fig. 2d).
In, Fig. 2e, and Fig. 2f, for p-ratio = 2.5, based on our intuition from Fig. 2c, the divergence is dominant. Here, the
thickness of the axon plays a role in the transmission such that the thicker the axon the divergence is less and the light is transmitted
more. However, for larger p–ratios, the scattering is dominant and the light scatters more in thick axons.
To summarize, for small p–ratios (∼2.5), the well confined modes (shorter wavelengths) transmit better while for large
p–ratios (∼5 or greater), the loosely confined ones (longer wavelengths) transmit better. Thicker axons yield higher transmission
for all wavelengths with small p–ratios while it's inverse only for the shorter wavelengths with large p–ratios. The transmission
after several paranodal regions can be roughly estimated by following the intuition of exponentiating the transmission through one
(see Supplementary Information of Ref (40)).
3.1.2. Transmission in bends
Fig. 3: Bends in the waveguides and their transmission. (a) a schematic of a sinusoidally bent waveguide in our model. Here,
r(cid:48) = 1 µm, the wavelength of the cosine function is 100 µm, and its amplitude (A) is 5 µm. (b) Magnitude of the EFPL when the
input mode with wavelength 0.4 µm passes through the bent waveguide. (c) Transmission percentage as a function of the change of
curvature, ∆κ, for three different wavelengths where r(cid:48) = 5 µm and A is being changed to vary ∆κ. (d)-(f) Transmission percentage
as a function of the myelinated axon radius for three different wavelengths with four different ∆κ. Figure from Ref. (40).
Power transmission of a straight waveguide has loss on encountering bends in the waveguide. Although this type of loss
5
fabc100 m10 m2 mIncident lightTransmission(%)■■■■■■■■■■■◆◆◆◆◆◆◆◆◆◆◆0.020.040.060.080.100.1220406080100=1.3m■=0.612m◆=0.4mChangeofcurvature(inµm-1)=longestpermissibleTransmission(%)■■■■■◆◆◆◆◆▲▲▲▲▲1234520406080100=0.024m-1■=0.039m-1◆=0.079m-1▲=0.126m-1Radiusoftheaxonincludingthemyelinsheath(r'inµm) =correspondingtothecentralpermissiblefrequencyTransmission(%)
■■■■■◆◆◆◆◆▲▲▲▲▲1234520406080100
=0.024
m-1■ =0.039
m-1◆ =0.079
m-1▲ =0.126
m-1Radiusoftheaxonincludingthemyelinsheath(r'inµm)=0.4µmTransmission(%)■■■■■◆◆◆◆◆▲▲▲▲▲1234520406080100=0.024m-1■=0.039m-1◆=0.079m-1▲=0.126m-1Radiusoftheaxonincludingthemyelinsheath(r'inµm)deAxonMyelin Are there optical communication channels in the brain?
can be minimized by propagation of the eigenmodes of circular bends of constant curvature along the waveguide, one cannot use
them for axons. For the varying curvature of axons, these modes are more lossy than eigenmodes of a straight waveguide. So, to
verify the bend losses in the axons, Ref. (40) considered the straight–mode in a sinusoidal waveguide with changing curvature and
obtained the transmission in the myelin sheath at the other end. Fig. 3a shows one example of S-bend in an axon with radius 0.6 µm,
and Fig. 3b shows the EFPL as a straight–mode passes through the axon.
The loss in sinusoidal S-shaped bends (shown in Fig. 3a) is highly dependent on the varying curvature (70). Thus, Ref.
(40) calculated the total power transmission up to a wavelength away from the myelin sheath boundaries as a function of the change
of curvature, ∆κ = 4Ak2 (k is the wavenumber and A is the amplitude of the cosine function) and plotted it for 3 different wave-
lengths in Fig. 3c (r(cid:48)=5 µm). The shorter wavelength gives higher transmission and the more the curvature changes the smaller the
transmission. Fig. 3d–f compares the transmission of the 3 different wavelengths for different radii of axon and different curvatures.
For ∆κ ∼0.024 µm−1, all the permissible wavelengths are guided with transmission percentage close to 100% for all axon radii.
Note that here we consider the change of curvature, ∆κ, of the curve passing through the central axis of the axon. However, the
inner part of the bent has the most curvature in comparison with the center or outer part at each point. These differences are more
noticeable for thicker axons since they have more change of curvature than the thiner ones and therefore experience more loss for
the same ∆κ. For the typical axons in the brain similar to those in the images of Ref (71) (the relatively straight axons with length
of ∼1 mm and radius of ∼1 µm), ∆κ < 0.05 µm−1 which results in transmission of over 90 %.
Fig. 4: Varying cross-sections and the transmission in such areas. (a) a schematic of a myelinated axon with longitudinally
varying cross sections. Here, the mean radius of the myelinated axon (outer radius of the axon) is 3 µm and the standard deviation
(s.d.) of the variation of radius for the myelin sheath is 0.36 µm. (b) Magnitude of the EFPL when the input mode with wavelength
0.48 µm passes through the region. (c) Transmission percentage as a function of the s.d. of the variation in radius for the myelin
sheath for three different wavelengths (the mean radius of the myelinated axon is chosen as 5 µm). (d)-(f) Transmission percentage
as a function of the mean radius of the myelinated axon for three different wavelengths with three different s.d. of the variation of
radius for the myelin sheath. Figure from Ref. (40).
6
abIncident lightAxonMyelin sheathcTransmission(%)■■■■■■■■◆◆◆◆◆◆◆◆05101520253060708090100=1.30m■=0.612m◆=0.40mS.D.oftheinhomogeneity(%oftheaveragethicknessofthemyelinsheath)=longestpermissibleTransmission(%) ■■■■■◆◆◆◆◆1234560708090100 s.d=10%■s.d=20%◆s.d=30%Meanradiusoftheaxonincludingthemyelinsheath(µm)=correspondingtothecentralpermissiblefrequencyTransmission(%)■■■■■◆◆◆◆◆12345708090100s.d=10%■s.d=20%◆s.d=30%Meanradiusoftheaxonincludingthemyelinsheath(µm)=0.4mTransmission(%)■■■■■◆◆◆◆◆1234580859095100s.d=10%■s.d=20%◆s.d=30%Meanradiusoftheaxonincludingthemyelinsheath(µm)def Are there optical communication channels in the brain?
3.1.3. Transmission in presence of varying cross-sections
The thickness of the myelin sheath, d, is not uniform all along the length of the axon. Ref. (40) varied d according to
an approximate normal distribution. The correlation length in the roughness of the myelin sheath – the width of bumps or valleys
in the outer surface of myelin sheath– was taken to be 5 µm to 10 µm. The mean value of the distribution (r(cid:48)) is chosen based
on the value of g-ratio, and the standard deviation (s.d.) of d is varied. Fig. 4a shows an example of the simulation for an axon
with length of 50 µm, r=2.4 µm, and the s.d. of 30 % of the average d. In Fig. 4b, the EFPL for input light with λ = 0.612 µm is
calculated. Fig. 4 c shows that in general, more deviation of thickness of the myelin sheath results in less transmission and shorter
wavelengths transmit better than longer ones. In Fig. 4d–f, we see the behavior of the transmission as a function of mean radius
of the myelinated axon with 3 different deviation and for 3 different wavelengths. For variations less than 10%, almost all of the
wavelengths can pass through with efficiency over 90%. These results were obtained for the 50 µm segment of the axon and was
extrapolated to the case of a longer segment transmission along the axon. In particular, the transmission fraction was exponentiated
by the number of 50 µm segments formed in the axon. In general, thicker axons are more sensitive to the large deviations and suffer
more loss. Note that longer correlation lengths lead to better transmission for the same s.d, while significantly shorter correlation
lengths are known to strongly scatter the mode. Some of the axonal segments (length ∼5 µm) of thin axons (r ∼1 µm) are within
this type of inhomogeneity, as represented in the images of (72). However, we were unable to find appropriate images of thicker
myelinated axons and longer segments required for a more realistic estimate of this type of inhomogeneity.
3.1.4. Transmission in non-circular cross sections
Axons and their myelin sheaths can have different cross-sectional shapes (72). As an example, Ref. (40) simulated the
cross-section of a myelinated axon shown in Fig. 5. The points along the axons cross-sectional boundary in Fig. 5a follow a normal
distribution whose mean value is 3µm and standard deviation 0.4µm. The myelin sheath is approximated as a parallel curve drawn
at a perpendicular distance of 2µm giving an average g-ratio = 0.6. Fig. 5b displays the magnitude of the EFPL for the incident
light of λ = 612nm. As Fig. 5c shows, the total power transmission decreases for all wavelengths while the cross-section becomes
more random. But the effect of this loss is small in reality, as many axons have less than 10% inhomogeneity in the cross-sectional
shape. Therefore, if the axon and myelin sheath change their cross-sectional shape slightly along the length of the axon, the primary
source of loss would be the coupling loss (72). However, for a substantial change of circular cross-section along the length, there
will be propagation loss as well (40).
Fig. 5: Non-circular cross sections of the axon and myelin sheath and their transmission. (a) A model of the cross-section
of an axon with the myelin sheath. The mean distance of the points along the cross-sectional boundary of the axon from its center
is 3 µm with s.d. of 0.4 µm. A parallel curve with an approximate of 2 µm apart from the axon's boundary is taken as the myelin
sheath's boundary. (b) Magnitude of the EFPL when a cylindrically symmetric input mode for a circular cross-section passes along
the waveguide with the modeled non-circular cross-section. Here, the input mode wavelength is 0.612 µm, the axon radius is 3 µm,
and the myelinated axon's radius is 5 µm. (c) Transmission percentage as a function of the s.d. of the distance between the points
on the boundary of the axon and a circle of radius 3 µm for three different wavelength. Figure from Ref. (40).
3.1.5. Transmission in presence of other imperfections
In addition to the sources of loss considered so far, there can be more imperfections. One of the notable ones is the
cross-talk between axons. The conducted light can leak from one axon to the other one if they are placed so close to each other.
7
AxonMyelina b cTransmission(%)■■■■■■◆◆◆◆◆◆01020304050859095100=1.30μm■λ=0.612μm◆λ=0.40μmS.D.oftheinhomogeneity(%ofthemeanradiusoftheaxon) Are there optical communication channels in the brain?
To avoid this loss, the distance between two adjacent axon should be at least a wavelength, which happens in most of the realistic
cases based on the images in Ref. (72).
The other imperfection we discuss is the varying refractive indexes of the axon and the myelin sheath which have been
assumed constant until now. Changing the refractive index both transversely and longitudinally with a s.d. of 0.02 (typical variation
as expected from (41, 62)), while keeping the mean the same as the one used so far, Ref. (40) observed no considerable difference
in the transmission (typically less than 1 %). Furthermore, Ref. (40) considered the effect of the glia cells next to the internodal
segment. These cells were modeled as spheres with radii varying from 0.1 µm to 0.3 µm, and refractive index 1.4, filling up one
third of the volume of the nodal region outside the axon as expected from the images of the Ref. (72). For the thickest axons,
transmission through a node of Ranvier was not affected significantly while for the thinnest one, the transmission increased slightly
(∼2 %). There would be also additional scattering losses in the areas that myelin sheath is inhomogeneous.
3.2. Absorption
In biological tissues, and more so in the brain, scattering of light, rather than absorption, is the main source of attenuation
of optical signals (73). To our knowledge, the absorption coefficient of the myelin sheath has not been measured experimentally.
One can only infer it indirectly with limited accuracy. The average absorption coefficient in the white matter decreases almost
monotonically from ∼0.3 mm−1 to ∼0.07 mm−1 for wavelengths 0.4 µm to 1.1 µm (74). But myelin can not be responsible for the
majority of the absorption since grey matter (which is almost devoid of myelin) has comparable absorption coefficients (74). It is
likely that light sensitive structures (e.g. chromophores in the mitochondria) are the main contributors to the absorption. Another
way to infer myelin's absorption coefficient is to look at the absorption of its constituents, i.e. lipids, proteins and water. Mammalian
fat shows an absorption coefficient less than 0.01 mm−1 for the biophotonic wavelength range (75). Water has similar absorption
coefficients. Most proteins have a strong resonance peak close to 0.28 µm with almost negligible absorption above 0.34 µm, and the
proteins in the myelin (e.g. myelin protolipid protein, and myelin basic protein) behave similarly (76). Thus, absorption in myelin
for the biophotonic wavelengths seems negligible (over a length scale of ∼1 cm), based on the data of its constituents.
3.3. Attainable transmission
Ref. (40) estimated the attainable transmission percentage after 1cm length of the axon in different examples, taking
the axon diameter as 100–150 times less than its internodal length according to the realistic values (64, 77). For an axon with
r = 3 µm, r(cid:48) = 5 µm, p-ratio = 7.5, internodal length = 1 mm, wavelength of input light = 1.3 µm, s.d. for varying area = 2.5 %,
∆κ = 0.039 µm−1, s.d. for non-circularity in cross-section shape = 13.33 %, and separation from the nearby axons = 1 µm, the
transmission after 1 cm would be ∼31 %. This transmission can be increased to ∼82 % if we take the wavelength of input light
= 0.61 µm, p-ratio = 2.5, and keep all the other parameters the same. For a thinner axon with r = 1.8 µm, r(cid:48) = 3 µm, p-ratio = 7.5,
internodal length = 500 µm, wavelength of light = 1.2 µm, s.d. for varying area = 20 %, separation from other axons = 1.2 µm, and
∆κ = 0.039 µm−1 would yield ∼3 % transmission after 1 cm.
If one chooses the shorter length of 2 mm for the axons (as there are axons with ∼1 mm length in the brain(78)), then the
transmission for the 3 examples discussed above would be ∼78 %, ∼96 %, and ∼46 %, respectively. The main source of loss for
these examples is the coupling in the paranodal regions. By locating the sources and receivers close to the ends of the myelinated
sections of the axon, one can reduce coupling losses. It is worth noting that photons can propagate in either directions: from the
axon terminal up to the axon hillock or in the opposite direction along the axon.
3.4. Attainable communication rates
To estimate the biophoton emission rate per neuron, one can use the experimental data of Ref. (12) in which the number
of biophotons emitted per minute by a slice of mouse brain is counted while the neurons are excited with the neurotransmitter
glutamate. Ref. (40) calculated this biophoton emission rate as about 1 photon per neuron per minute. This rate is about one to two
orders of magnitude lower than the the average rate of electrochemical spikes (79). But there is a considerable uncertainty on this
number and it might be higher in reality since it only takes into account biophotons scattered outside, while most of them would
likely be absorbed in the brain itself rather than being scattered (if they were propagating inside the waveguides in the brain). On
the other hand, one can argue that the estimate could also be too high because the brain slice was stimulated with glutamate. One
should also notice that this rate can be different depending on the neuron type.
If one takes such low rate of biophoton emission and consider the fact that there are about 1011 neurons in a human brain,
there would still be over a billion photon emission per second. This mechanism appears to be sufficient to facilitate transmission of
a large number of bits of information, or even allow the creation of a large amount of quantum entanglement. Note that the behavior
of about one hundred photons can already not be simulated efficiently with classical computers (80). It is also worth to mention
that psychophysical experiments performed in the past indicate that the bandwidth of conscious experience is below the range of
8
Are there optical communication channels in the brain?
100 bits per second (81, 82).
3.5. Proposals to test the hypothesis
Although there is already some experimental evidence of biophoton propagation in the brain and axons (12, 43, 44), it
would nevertheless be very interesting to test the light guidance of axons directly in-vitro and in-vivo. For testing in-vitro, one way
is to light up one end of a thin brain slice (with proper homogeneous myelinated axons) and look for the bright spots related to
the open ends of the myelinated axons at the other end. To get more accurate results, one can isolate a suitable myelinated axon
and, while keeping it alive in the proper solution, couple into one of guided modes of the axon, similar to what was done for light
guidance by Muller cells in Ref. (45). Since the real light sources in the brain may be close to axons terminals (12), one can cut
the axon near the terminal and hillock regions, couple the mode directly to the myelin sheath and observe the light intensity from
the other end of the axon. This test can verify the guidance of the myelin sheath. Evanescent coupling and readout of light offers
another possibility.
For an in-vivo test of light guidance, first one needs to prove the existence of biophotons in the myelin sheath. To do so,
one can add light-sensitive chemicals like AgNO3 into the cytoplasmic loops in the paranodal region or in oligodendrocytes, which
will be passed around to the myelin sheath. Then, the light-activated decomposition of AgNO3 to Ag leaves Ag atoms as the dark
insoluble grains. The idea of this test is similar to the technique involved in the development of photographic films, and the in-situ
biophoton autography (IBA) technique (43).
An additional type of tests that can be done in-vivo is to insert fluorescent molecules or nano-particles as the sources.
They could also serve as the detectors if their fluorescence emission is due to the absorption of photons emitted from the sources
(83). One can also employ optogenetics (84) to create artificial detectors. Optogenetics uses neurons which are genetically modified
to produce light-sensitive proteins operating as ion-channels (like channel rhodopsin). If one places these light-sensitive proteins
into the axonal membrane at the end of myelin sheath close to an axon terminal or into the the membranes of the cytoplasmic loops
in the paranodal region, one can detect light by observing the operation of the ion-channels. It is worth noting that the rate of
oligodendrogenesis increases and myelin sheath becomes thicker in neighboring of light-illuminated genetically modified neurons
(85). The question comes up whether the neurons adjust themselves for a better light guidance by forming more layers.
Besides the artificial sources and detectors, it is also interesting to use natural ones to show the photon guidance from the
sources to the detectors through the axons. To do so, one needs to first understand well the photon sources, and their emission rates
and wavelength. One can make use of nanoantennas to raise the emission rates (86) and get a better knowledge about the sources
and emissions. Although, some measurements on the number of scattered photons from the axons (not guided or absorbed ones)
(10–13) have been performed, photon emission from the individual neurons has not yet been analyzed. One also needs to measure
photon detection capabilities of the natural candidate detectors such as opsins, centrosomes (25) and chromophores in mitochondria
(26).
4. Discussion
In this review of Ref. (40) we have discussed how light conduction in a myelinated axon is feasible even in the presence
of realistic imperfections in the neuron. We have also described future experiments that could validate or falsify this model of
biophoton transmission (40). It is also worth addressing a few related questions. It is of interest to identify possible interaction
mechanisms between biophotons and nuclear spins within the framework of quantum communication. Spin chemistry research
(87) determined effects whereby electron and nuclear spins affect chemical reactions. These effects can also involve photons. In
particular, a class of cryptochrome proteins can be photo-activated resulting in the production of a pair of radicals per event, with
correlated electronic spins. This effect has been hypothesized to explain bird magnetoreception (49). It has been recently shown
by theoretical considerations that interactions between electron and nuclear spins in cryptochromes are of critical importance to
the elucidation of the precision of magnetoreception effects (50). Importantly for this topic, cryptochrome complexes are found in
the eyes of mammals and they are also magnetosensitive at the molecular level (88). Therefore, if similar proteins can be found
in the inner regions of the human brain, this could provide the required interface between biophotons and nuclear spins. However,
for individual quantum communication links to form a larger quantum network with an associated entanglement process involving
many distant spins, the nuclear spins interfacing with different axons must interact coherently. This, most likely, requires close
enough contact between the interacting spins. The involvement of synaptic junctions between individual axons may provide such a
proximity mechanism.
We should also address the question of the potential relevance of optical communication between neurons with respect
to consciousness and the binding problem. A specific anatomical question that arises is whether brain regions implicated in con-
sciousness (89) (e.g. claustrum (90, 91), the thalamus, hypothalamus and amygdala (92), or the posterior cerebral cortex (89)) have
myelinated axons with sufficient diameter to allow light transmission.
A major role of the myelin sheath as an optical waveguide could provide a better understanding of the causes of the
various diseases associated with it (e.g. multiple sclerosis (93)) and hence lead to a design and implementation of novel therapies
9
Are there optical communication channels in the brain?
for these pathologies.
Let us note that, following Ref. (40), we have focused our discussion here on guidance by myelinated axons. However,
light guidance by unmyelinated axons is also a possibility, as discussed in more detail in the supplementary information of Ref.
(40).
Finally, with the advantages optical communication provides in terms of precision and speed, it is indeed a wonder why
biological evolution would not fully exploit this modality. On the other hand, if optical communication involving axons is har-
nessed by the brain, this would reveal a remarkable, hitherto unknown new aspect of the brains functioning, with potential impacts
on unraveling fundamental issues of neuroscience.
5. Acknowledgment
P.Z. is supported by an AITF scholarship. S.K. is supported by AITF and Eyes High scholarships. J.T., P.B. and C.S.
are supported by NSERC.
6. References
1. R Vogel and R Sussmuth: Weak light emission patterns from lactic acid bacteria. Luminescence 14(2), 99-105 (1999)
2. TI Quickenden and RN Tilbury: Luminescence spectra of exponential and stationary phase cultures of respiratory deficient
Saccharomyces cerevisiae . J Photochem Photobiol B 8(2), 169-174 (1991).
3. CM Gallep and SR Dos Santos: Photon-counts during germination of wheat (Triticum aestivum) in wastewater sediment
solutions correlated with seedling growth. Seed Sci Technol 35(3), 607-614 (2007)
4. A Prasad and P Posp´ısil: Towards the two-dimensional imaging of spontaneous ultra-weak photon emission from microbial,
plant and animal cells. Sci Rep 3, 1211 (2013)
5. M Hossu, L Ma, X Zou and W Chen: Enhancement of biophoton emission of prostate cancer cells by Ag nanoparticles.
Cancer Nanotech 4(1), 21-26 (2013)
6. A Rastogi and P Posp´ısil: Spontaneous ultraweak photon emission imaging of oxidative metabolic processes in human skin:
effect of molecular oxygen and antioxidant defense system. J Biomed Opt 16, 9 (2011)
7. M Cifra, E Van Wijk, H Koch, S Bosman and R Van Wijk: Spontaneous ultra-weak photon emission from human hands is
time dependent. Rarioeng 16(2), 15-19 (2007)
8. M Kobayashi, D Kikuchi and H Okamura: Imaging of ultraweak spontaneous photon emission from human body displaying
diurnal rhythm. PLoS One 4, 7 (2009)
9. A Prasad and P Pospisil: Two-dimensional imaging of spontaneous ultra-weak photon emission from the human skin: role of
reactive oxygen species. J Biophotonics 4(11-12), 840849 (2011)
10. Y Isojima, T Isoshima, K Nagai, K Kikuchi and H Nakagawa: Ultraweak biochemiluminescence detected from rat hippocam-
pal slices. Neuroreport 6(4), 658-660 (1995)
11. M Kobayashi, M Takeda, T Sato, Y Yamazaki, K Kaneko, KI Ito, H Kato and H Inaba: In vivo imaging of spontaneous
ultraweak photon emission from a rats brain correlated with cerebral energy metabolism and oxidative stress. J Neurosci Res
34(2), 103-113 (1999)
12. R Tang and J Dai: Spatiotemporal imaging of glutamate-induced biophotonic activities and transmission in neural circuits.
PLoS One 9, 1 (2014)
13. Y Kataoka, Y Cui, A Yamagata, M Niigaki, T Hirohata, N Oishi and Y Watanabe: Activity-dependent neural tissue oxidation
emits intrinsic ultraweak photons. Biochem Biophys Res Commun 285(4), 1007-1011 (2011)
14. V Salari, F Scholkmann, I Bokkon, F Shahbazi and J Tuszynski: The physical mechanism for retinal discrete dark noise:
Thermal activation or cellular ultraweak photon emission? PLOS One 11, 3 (2016)
15. Z Wang, N Wang, Z Li, F Xiao, and J Dai: Human high intelligence is involved in spectral redshift of biophotonic activities
in the brain. Proc Natl Acad Sci 113, 31 (2016)
16. J Chang, J Fisch and FA Popp: Biophotons. Springer Science & Business Media, (1998)
17. M Cifra and P Posp´ısil: Ultra-weak photon emission from biological samples: definition, mechanisms, properties, detection
and applications. J Photochem Photobiol B 139, 2-10 (2014)
18. D Fels: Cellular Communication through Light. PLoS One 4, 4 (2009).
10
Are there optical communication channels in the brain?
19. PD Wade, J Taylor and P Siekevitz: Mammalian cerebral cortical tissue responds to low-intensity visible light. Proc Natl
Acad Sci 85(23), 9322-9326 (1988)
20. T Starck, J Nissil, A Aunio, A Abou-Elseoud, J Remes, J Nikkinen, M Timonen, T Takala, O Tervonen and V Kiviniemi:
Stimulating brain tissue with bright light alters functional connectivity in brain at the resting state. World J Neurosci 2(02), 81
(2012)
21. DN Leszkiewicz, K Kandler and E Aizenman: Enhancement of NMDA receptormediated currents by light in rat neurones in
vitro. J Physiol 524(2), 365-374 (2000)
22. AI Zhuravlev, OP Tsvylev and SM Zubkova: Spontaneous endogenous ultraweak luminescence of rat liver mitochondria in
conditions of normal metabolism. Biofizika 18, 1037-1040 (1973)
23. JA Tuszy´nski and JM Dixon: Quantitative analysis of the frequency spectrum of the radiation emitted by cytochrome oxidase
enzymes. Phys Rev E 64, 051915 (2001)
24. VM Mazhul and DG Shcherbin: Phosphorescent analysis of lipid peroxidation products in liposomes. Biofizika 44, 676-681
(1999)
25. G Albrecht-Buehler: Cellular infrared detector appears to be contained in the centrosome. Cell Motil Cytoskeleton 27, 262-271
(1994)
26. M Kato, K Shinzawa and S Yoshikawa: Cytochrome oxidase is a possible photoreceptor in mitochondria. Photobiochem
Photobiophys 2, 263-269 (1981)
27. T Okano, T Yoshizawa and Y Fukada: Pinopsin is a chicken pineal photoreceptive molecule. Nature 372, 94-97 (1994)
28. Y Nakane, K Ikegami, H Ono, N Yamamoto, S Yoshida, K Hirunagi, S Ebihara, Y Kubo and T Yoshimura: A mammalian
neural tissue opsin (Opsin 5) is a deep brain photoreceptor in birds. Proc Natl Acad Sci 107, 34 (2010)
29. S Blackshaw and SH Snyder: Encephalopsin: a novel mammalian extraretinal opsin discretely localized in the brain. J
Neurosci 19, 10 (1999)
30. J Nissila, S Manttari, T Sarkioja, H Tuominen, T Takala, M Timonen and S Saarela: Encephalopsin (OPN3) protein abundance
in the adult mouse brain. J Comp Physiol A 198, 11 (2012)
31. D Kojima, S Mori, M Torii, A Wada, R Morishita and Y Fukada: UV-sensitive photoreceptor protein OPN5 in humans and
mice. PLoS One 6(10) e26388 (2011)
32. I Provencio, IR Rodriguez, G Jiang, WP Hayes, EF Moreira and MD Rollag: A novel human opsin in the inner retina. J
Neurosci 20(2), 600-605 (2000)
33. D Kojima, H Mano and Y Fukada: Vertebrate ancient-long opsin: a green-sensitive photoreceptive molecule present in
zebrafish deep brain and retinal horizontal cells. J Neurosci 20, 8 (2000)
34. EE Tarttelin, J Bellingham, MW Hankins, RG Foster and RJ Lucas: Neuropsin (Opn5): a novel opsin identified in mammalian
neural tissue. Febs Letters 554(3), 410-416 (2003)
35. Y Shichida and T Matsuyama: Evolution of opsins and phototransduction. Phil Trans R Soc B 364(1531), 2881-2895 (2009)
36. R Thar and M Kuhl: Propagation of electromagnetic radiation in mitochondria? J Theor Biol 230(2), 261-270 (2004)
37. M Rahnama, J A Tuszynski, I Bokkon, M Cifra, P Sardar and V Salari: Emission of mitochondrial biophotons and their effect
on electrical activity of membrane via microtubules. J. Integr. Neurosci. 10(01), 65-88 (2011)
38. M Jibu, S Hagan, SR Hameroff, KH Pribram and K Yasue: Quantum optical coherence in cytoskeletal microtubules: impli-
cations for brain function. Biosystems 32(3), 195-209 (1994)
39. F Scholkmann: Long range physical cell-to-cell signalling via mitochondria inside membrane nanotubes: a hypothesis. Theor.
Biol. Med. Model. 13(1), 16 (2016)
40. S Kumar, K Boone, J Tuszynski, P Barclay and C Simon: Possible existence of optical communication channels in the brain.
Sci Rep 6, 36508 (2016)
41. IP Antonov, AV Goroshkov, VN Kalyunov, IV Markhvida, AS Rubanov and LV Tanin: Measurement of the radial distribution
of the refractive index of the Schwanns sheath and the axon of a myelinated nerve fiber in vivo. J Appl Spectrosc 39(1), 822-
824 (1983)
42. D Purves, GJ Augustine, D Fitzpatrick, LC Katz, AS LaMantia, JO McNamara and SM Williams (editors): Neuroscience.
Sinauer Associates, (2001)
43. Y Sun, C Wang and J Dai: Biophotons as neural communication signals demonstrated by in situ biophoton autography.
Photochem Photobiol Sci 9(3), 315-322 (2010)
44. KM Hebeda, T Menovsky, JF Beek, JG Wolbers, MJ van Gemert: Light propagation in the brain depends on nerve fiber
orientation. Neurosurgery 35(4), 720-724 (1994)
11
Are there optical communication channels in the brain?
45. K Franze, J Grosche, SN Skatchkov, S Schinkinger, C Foja, D Schild, O Uckermann, K Travis, A Reichenbach and J Guck:
Muller cells are living optical fibers in the vertebrate retina. Proc Natl Acad Sci USA 104(20), 8287-8292 (2007)
46. AM Labin, SK Safuri, EN Ribak and I Perlman: Muller cells separate between wavelengths to improve day vision with
minimal effect upon night vision. Nat Commun 5, 4319 (2014)
47. GS Engel, TR Calhoun, EL Read, TK Ahn, T Mancal, YC Cheng, RE Blankenship and GR Fleming: Evidence for wavelike
energy transfer through quantum coherence in photosynthetic systems. Nature 446, 782-786 (2007)
48. E Romero, R Augulis, VI Novoderezhkin, M Ferretti, J Thieme, D Zigmantas and RV Grondelle: Quantum coherence in
photosynthesis for efficient solar-energy conversion. Nat Phys 10(9), 676-682 (2014)
49. T Ritz, S Adem, and K Schulten: A model for photoreceptor-based magnetoreception in birds. Biophys J 78, 707-718 (2000)
50. H G Hiscock, S Worster, D R Kattnig, C Steers, Y Jin, D E Manolopoulos, H Mouritsen, and P J Hore: The quantum needle
of the avian magnetic compass. Proc Natl Acad Sci USA 113(17), 4634-4639 (2016)
51. L Turin: A Spectroscopic Mechanism for Primary Olfactory Reception. Chem Senses 21, 773-791 (1996)
52. M I Franco, L Turin, A Mershin, and E M C Skoulakis: Molecular vibration-sensing component in Drosophila melanogaster
olfaction. Proc Natl Acad Sci USA 108(9), 3797-3802 (2011)
53. R Horodecki, P Horodecki, M Horodecki, K Horodecki: Quantum entanglement. Rev Mod Phys 81, 865 (2009)
54. GA Mashour: The cognitive binding problem: from Kant to quantum neurodynamics. NeuroQuantology 1, 29-38 (2004)
55. S Hameroff, R Penrose: Consciousness in the universe: A review of the 'Orch OR' theory. Phys Life Rev 11, 39-78 (2014)
56. M Fisher: Quantum cognition: The possibility of processing with nuclear spins in the brain. Ann Phys 61, 593-602 (2015)
57. M Tegmark: Importance of quantum decoherence in brain processes. Phys Rev E 61, 4194-4206 (2000)
58. W S Warren, S Ahn, M Mescher, M Garwood, K Ugurbil, W Richter, R R Rizi, J Hopkins, J S Leigh: MR Imaging contrast
enhancement based on intermolecular zero quantum coherences. Science 281(5374), 247-251 (1998)
59. H Lei, X H Xhu, X L Zhang, K Ugurbil, and W Chen, W: In vivo 31P magnetic resonance spectroscopy of human brain at 7
T: an initial experience. Magn Reson Med 49, 199-205 (2003)
60. H J Kimble: The quantum internet. Nature 453, 1023-1030 (2008)
61. N Sangouard, C Simon, H de Riedmatten, and N Gisin: Quantum repeaters based on atomic ensembles and linear optics. Rev
Mod Phys 83(33), (2011)
62. Z Wang, I S Chun, X Li, Z Ong, E Pop, L Millet, M Gillette, and G Popescu: Topography and refractometry of nanostructures
using spatial light interference microscopy. Opt Lett, 35(2), 208210 (2010)
63. V V Tuchin, I L Maksimova, D A Zimnyakov, I L Kon, A H Mavlyutov, A A Mishin: Light propagation in tissues with
controlled optical properties. J Biomed Opt 2(4), 401-417 (1997)
64. R L Friede, and R Bischhausen: How are sheath dimensions affected by axon caliber and internode length? Brain Res 235,
335-350 (1982)
65. D Liewald, R Miller, N Logothetis, H J Wagner, and A Schuz: Distribution of axon diameters in cortical white matter: an
electron-microscopic study on three human brains and a macaque. Biol Cybern 108, 541-557 (2014)
66. C Jocher, C Jauregui, C Voigtlander, F Stutzki, S Nolte, J Limpert, and A Tunnermann: Fiber based polarization filter for
radially and azimuthally polarized light. Opt Express 19(20), 19582-19590 (2011)
67. P Chinn and F O Schmitt: On the birefringence of nerve sheaths as studied in cross sections. J Cell Biol 9(2), 289-296 (1937)
68. E Kandel, T Jessell, J Schwartz, S Siegelbaum, and A J Hudspeth: Principles of Neural Science. McGraw-Hill Medical,
(2012)
69. J Zagoren and S Fedoroff (eds): The Node of Ranvier. Academic Press, Inc, Orlando, (1984)
70. A Syahriar: A simple analytical solution for loss in S-bend optical waveguide. RF and Microwave Conference, 2008. RFM
2008. IEEE International, 10.1109/RFM.2008.4897467, 357-360 (2008)
71. A J Schain, R A Hill, and J Grutzendler: Label-free in vivo imaging of myelinated axons in health and disease with spectral
confocal reflectance microscopy. Nat Med 20(4), 443-449 (2014)
72. A Peters, S L Palay, H deF. Webster: Fine Structure of the Nervous System: Neurons and Their Supporting Cells. In: Journal
of Neuropathology & Experimental Neurology. Oxford University Press, (1991)
73. W Cheong, S Prahl, and A J Welch: A review of the optical properties of biological tissues. IEEE J Quantum Electron 26,
2166-2185 (1990)
12
Are there optical communication channels in the brain?
74. A N Yaroslavsky, P C Schulze, I V Yaroslavsky, R Schober, F Ulrich, H J Schwarzmaier: Optical properties of selected native
and coagulated human brain tissues in vitro in the visible and near infrared spectral range. Phys Med Biol 47(12), 2059-2073
(2002)
75. R L P van Veen, H J C M Sterenborg, A Pifferi. A Torricelli, E Chikoidze, R Cubeddu: Determination of visible near-IR
absorption coefficients of mammalian fat using time- and spatially resolved diffuse reflectance and transmission spectroscopy.
J Biomed Opt 10, (054004) (2005)
76. P Facci, P Cavatorta, L Cristofolini, M P Fontana, A Fasano, and P Riccio: Kinetic and structural study of the interaction of
myelin basic protein with dipalmitoylphosphatidylglycerol layers. Biophys J 78, 14131419 (2000)
77. M Ibrahim, A M Butt, and M Berry: Relationship between myelin sheath diameter and internodal length in axons of the
anterior medullary velum of the adult rat. J Neurol Sci 133, 119-127 (1995)
78. J Nolte: The Human Brain: An Introduction to its Functional Anatomy. 3rd ed. St. Louis: Mosby-Year Book, (1993)
79. G Buzs´aki, and K Mizuseki: The log-dynamic brain: how skewed distributions affect network operations. Nat Rev Neurosci
15, 264-278 (2014)
80. S Lloyd: Universal Quantum Simulators. Science 273, 1073-1078 (1996)
81. M Zimmermann: The Nervous System in the Context of Information Theory. In: Human Physiology. Eds: R F Schmidt and
G Thews, Springer, (1989)
82. T Norretranders: The User Illusion: Cutting Consciousness Down to Size. Penguin Press Science, (1999).
83. X Yang: Nanotechnology in Modern Medical Imaging and Interventions. Nova Science Pub Inc, (2013).
84. K Deisseroth: Optogenetics. Nat Methods 8, 26-29 (2011)
85. E M Gibson, D Purger, C W Mount, A K Goldstein, G L Lin, L S Wood, I Inema, S E Miller, G Bieri, J Bradley Zuchero,
B A Barres, P J Woo, H Voge, M Monje: Neuronal activity promotes oligodendrogenesis and adaptive myelination in the
mammalian brain. Science 344, (1252304) (2014)
86. S Kuhn, U Hakanson, L Rogobete, and V Sandoghdar: Enhancement of Single-Molecule Fluorescence Using a Gold Nanopar-
ticle as an Optical Nanoantenna. Phys Rev Lett 97, (017402) (2006)
87. N J Turro: Fun with Photons, Reactive Intermediates, and Friends. Skating on the Edge of the Paradigms of Physical Organic
Chemistry, Organic Supramolecular Photochemistry, and Spin Chemistry. J Org Chem 76, 9863-9890 (2011)
88. E L Foley, R J Gegear, and S M Reppert: Human cryptochrome exhibits light-dependent magnetosensitivity. Nat Commun 2,
(356) (2011)
89. C Koch, M Massimini, M Boly, and G Tononi: Neural correlates of consciousness: progress and problems. Nat Rev Neurosci
17, 307321 (2016)
90. M Z Koubeissi, F Bartolomei, A Beltagy, and F Picard: Electrical stimulation of a small brain area reversibly disrupts
consciousness. Epilepsy Behav 37, 32-35 (2014)
91. F C Crick and C Koch: What is the function of the claustrum? Philos Trans R Soc Lond B Biol Sci 360, 1271-1279 (2005)
92. W Loewenstein: Physics in Mind: A Quantum View of the Brain. Basic Books, ( 2013)
93. J Nakahara, M Maeda , S Iso, and N Suzuki: Current concepts in multiple sclerosis: autoimmunity versus oligodendrogliopa-
thy. Clin Rev Allergy Immunol 42, 26-34 (2012)
Key Words: Brain, Biophotons, Myelinated axons, Waveguides, Quantum, Review
Send correspondence to: Christoph Simon, Department of Physics and Astronomy, University of Calgary, Calgary T2N 1N4,
Alberta, Canada, Tel: +1(403)220-7007, Fax: +1(403)210-8876, E-mail: [email protected]
13
|
1312.6708 | 1 | 1312 | 2013-12-23T21:53:08 | Development of Morphogen Gradient: The Role of Dimension and Discreteness | [
"physics.bio-ph",
"q-bio.SC"
] | The fundamental processes of biological development are governed by multiple signaling molecules that create non-uniform concentration profiles known as morphogen gradients. It is widely believed that the establishment of morphogen gradients is a result of complex processes that involve diffusion and degradation of locally produced signaling molecules. We developed a multi-dimensional discrete-state stochastic approach for investigating the corresponding reaction-diffusion models. It provided a full analytical description for stationary profiles and for important dynamic properties such as local accumulation times, variances and mean first-passage times. The role of discreteness in developing of morphogen gradients is analyzed by comparing with available continuum descriptions. It is found that the continuum models prediction about multiple time scales near the source region in two-dimensional and three-dimensional systems is not supported in our analysis. Using ideas that view the degradation process as an effective potential, the effect of dimensionality on establishment of morphogen gradients is also discussed. In addition, we investigated how these reaction-diffusion processes are modified with changing the size of the source region. | physics.bio-ph | physics | Development of Morphogen Gradient: The Role of Dimension
and Discreteness
Hamid Teimouri, Anatoly B. Kolomeisky
Department of Chemistry and Center for Theoretical Biological Physics,
Rice University, Houston, TX 77005-1892
The fundamental processes of biological development are governed by multiple sig-
naling molecules that create non-uniform concentration profiles known as morphogen
gradients. It is widely believed that the establishment of morphogen gradients is a
result of complex processes that involve diffusion and degradation of locally produced
signaling molecules. We developed a multi-dimensional discrete-state stochastic ap-
proach for investigating the corresponding reaction-diffusion models. It provided a
full analytical description for stationary profiles and for important dynamic proper-
ties such as local accumulation times, variances and mean first-passage times. The
role of discreteness in developing of morphogen gradients is analyzed by comparing
with available continuum descriptions. It is found that the continuum models pre-
diction about multiple time scales near the source region in two-dimensional and
three-dimensional systems is not supported in our analysis. Using ideas that view
the degradation process as an effective potential, the effect of dimensionality on es-
tablishment of morphogen gradients is also discussed. In addition, we investigated
how these reaction-diffusion processes are modified with changing the size of the
source region.
3
1
0
2
c
e
D
3
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
8
0
7
6
.
2
1
3
1
:
v
i
X
r
a
I.
INTRODUCTION
1
One of the most important biological phenomena is a development of multi-cellular or-
ganisms from embryos with initially finite number of genetically identical cells. It is now
well established that such spatial patterning and tissue formation is controlled by multiple
signaling molecules that are called morphogens.1 -- 5 These signaling molecules create non-
uniform concentration profiles, which are also known as morphogen gradients, that serve as
local dose-dependent gene regulators for embryo cells. Depending on the concentration of
morphogens different genes are turned on or turned off, producing morphologically different
cells.2,3 The original ideas of reaction-diffusion control in biological development have been
proposed by more than 60 years ago by A. Turing,6 and in recent years a significant number
of quantitative studies that uncovered important properties of morphogen gradients in vari-
ous biological systems have been presented.7 -- 17 However, our understanding of mechanisms
of formation of signaling molecules profiles in embryos is still quite limited.18
To explain complex processes that determine the establishment of morphogen gradients
several ideas have been discussed.18 But the most popular proposed mechanism is based
on a so-called synthesis-diffusion-degradation (SDD) model.7,19 It assumes that morphogen
molecules are synthesized in some localized region of the embryo, then they diffuse along the
cellular lines, and with some probability they can associate to receptors on cells surfaces,
from which they are eventually degraded and removed from the system.5,9,10 This model
has been widely used, although with a variable success, for analyzing dynamics of signaling
molecules in different biological systems.7 -- 10,12,14
In most cases, theoretical analysis of the morphogen gradient formation employs one-
dimensional continuum versions of the SDD model.20 -- 24 A more realistic description of sig-
naling processes that takes into account the structure of embryo requires the application
of multi-dimensional models.25 The importance of dimensionality has been also pointed out
in recent experiments on diffusion of morphogens in extracellular space where geometric
obstacles strongly affect trajectories of signaling molecules.18,26 Recently, the spherically-
symmetric continuum SDD model has been investigated for multi-dimensional situations.27,28
This elegant theoretical method analyzed the kinetics of formation of morphogen gradients,
and it also provided analytical expressions for stationary concentration profiles and for local
accumulation times (LAT), which are defined as average times to reach locally the steady-
2
state concentrations. One of the most surprising observations of this work is a prediction that
for two-dimensional and three-dimensional systems, in contrast to one-dimensional models,
there are multiple time scales for dynamics of formation of signaling molecules concentration
profiles for the spatial region near the source.27 It suggested that the dimensionality might
play an important role in dynamics of these reaction-diffusion processes.
Analyzing the application of the results from continuum SDD models one has to remember
that the continuum picture is an approximation which does not work for all set of parameters.
The discrete-state stochastic version of the SDD model on semi-infinite interval has been
recently introduced by one of us.29 The advantage of this approach is that it is valid for all
ranges of the parameter space, and the continuum version is just a special limiting case. It
has been also shown that the local accumulation times can be well approximated via the
corresponding mean first-passage times, and the degradation process can be viewed as an
effective potential that drives morphogens away from the source.29
In this paper we develop a discrete-state stochastic framework for description of SDD
models with strongly localized sources in all spatial dimensions. This analysis allows us
to compute stationary-state density profiles for morphogens as well as transient dynamic
properties such as local accumulations times, mean first-passage times and variance of the
local accumulation times. It provides a direct way for measuring the effect of dimensionality
in these reaction-diffusion processes. By comparing the obtained results with predictions
from continuum models the role of discreteness is also investigated. The existence of multiple
times scales in dynamics of morphogen gradients formation for two-dimensional and three-
dimensional cases is critically tested, and it is found that there is only one time scale for all
distances from the source. In addition, we generalized our approach for the systems with
extended source regions, which allowed us to analyze the effect of the producing region size
on dynamics of morphogen gradients formation.
The paper is organized as follows.
In Sec.
II the general d-dimensional discrete-state
stochastic SDD model is presented and analyzed. The extended version of the model for
source regions of variable size is given in Sec. III. Summary and concluding remarks are
made in the final Sec.IV, while some detailed calculations are presented in Appendices A
and B.
3
II. A DISCRETE-STATE STOCHASTIC SDD MODEL IN GENERAL
DIMENSIONS
Let us consider a general discrete-state stochastic SDD model in d spatial dimensions as
illustrated in Fig. 1 for d = 2. Any lattice site in the d-dimensional space is characterized
by d coordinates, namely ~n = (n1, n2, ..., nd). We assume that morphogens are produced
at the origin ~n0 = (0, 0, ..., 0) with a time-independent rate Q. Then, from any lattice
site ~n = (n1, n2, ..., nd) they can jump to any nearest neighbor site with a diffusion rate
u. The particle can also be degraded at any position with a rate k: see Fig. 1. The
continuum limit is realized when u ≫ k, i.e., when the diffusion rate is much larger than
the degradation rate. For convenience, we adopt here a single-molecule view of the process,
in which the concentration of signaling molecules is equivalent to the probability of finding
a single morphogen particle at given site.29 We define a function P (n1, n2, ..., nd; t) as the
probability to find the particle at the position ~n = (n1, n2, ..., nd) at time t. The temporal
evolution of these probabilities is governed by a set of master equations,
dP (n1, n2, ..., nd; t)
dt
= uXnn
P (n1, n2, ..., nd; t) − (2ud + k)P (n1, n2, ..., nd; t),
(1)
wherePnn is an operator corresponding to summing over all nearest neighbors, namely:
P (n1, n2, ..., nd; t) = P (n1 − 1, n2, ..., nd; t) + P (n1 + 1, n2, ..., nd; t)
+P (n1, n2 − 1, ..., nd; t) + P (n1, n2 + 1, ..., nd; t) + ....
Xnn
For the origin site we have a slightly different master equation,
dP (0, 0, ...; t)
= Q + uXnn
P (0, 0, ...; t) − (2du + k)P (0, 0, ...; t)
(2)
(3)
(4)
with
dt
Xnn
P (0, 0, ...; t) = P (−1, 0, ..., 0; t) + P (1, 0, ..., 0; t)
+P (0,−1, ..., 0; t) + P (0, 1, ..., 0; t) + ...
At large times, these equations can be solved exactly, producing the stationary density
profiles,
P (s)(n1, n2, ..., nd) =
2Qxn1+n2+...+nd
√k2 + 4duk
=
2Q
√k2 + 4duk
exp (− n1 − n2 −...− nd
λ
),
(5)
where
x = (2du + k −
√k2 + 4duk)/(2du), λ = −1/ ln x,
4
(6)
and λ is a decay length. For d = 1 these expressions reduce, as expected, to already known
results.29
One can see that that in the steady-state the density profile is the exponentially decaying
function with the decay length being independent of the source production rate Q. Similar
behavior has been observed in multi-dimensional continuum SDD models.27,28 In our ap-
proach, the continuum limit corresponds to the case when the diffusion rate is much larger
than the degradation rate, u ≫ k. In this case we have λ ≃ p(du/k). At another limit,
for fast degradation rates, k ≫ u, the decay length is equal to λ ≃ 1/ ln (k/2du). The
analysis of Eqs. (5) and (6) suggests that increasing d leads to lower probability to find the
signaling molecules at the origin, while at the same time the decay in the density profile is
also slower. There is an important difference between the predictions for the decay length
in the continuum and discrete SDD models. We argue that λ is generally larger (∼ √d in
the continuum limit), and it might be important for interpretation of experimental results
in the formation of morphogen gradients.25
A. Local Accumulation Times
An important dynamic property of morphogen gradients formation are local accumulation
times. They are defined as average times at which the stationary density profile is achieved
at given spatial position. Berezhkovskii and coworkers20 have introduced a method of calcu-
lating explicitly these quantities by using local relaxation functions R(n1, n2, ..., nd; t), which
can be written as
R(n1, n2, ..., nd; t) =
P (n1, n2, ..., nd; t) − P (s)(n1, n2, ..., nd)
P (n1, n2, ..., nd; t = 0) − P (s)(n1, n2, ..., nd)
P (n1, n2, ..., nd; t)
P (s)(n1, n2, ..., nd)
= 1 −
(7)
for the discrete-state multi-dimensional SDD models. The physical meaning of the local re-
laxation function is that it gives a measure of how close the system to the steady-state
conditions.
It ranges from R = 1 at t = 0 to R = 0 when the system reaches the
stationary state at given location.
Introducing the Laplace transform of this function,
eR(n1, n2, ..., nd; s) = R ∞
by20
0 R(n1, n2, ..., nd; t)e−stdt, it can be shown that the LAT are given
5
(8)
t(n1, n2, ..., nd) = −
∞Z0
t
∂R(n1, n2, ..., nd; t)
∂t
dt =
R(n1, n2, ..., nd; t)dt
∞Z0
= eR(n1, n2, ..., nd; s = 0).
From this relation the explicit expressions for the local accumulation times can be found:
t(n1, n2, ..., nd) =
(2du + k)
(k2 + 4duk)
+ n1 + n2 +...+ nd
√k2 + 4duk
.
(9)
To compare our results with continuum SDD models (which were analyzed for spherically
symmetric conditions),27,28 it is convenient to consider a specific direction in space along a
radial vector ~r = (n1, n2, ..., nd) where n1 = n2 = ... = nd . One can easily show that
n1 = r√d
where r is the radius of hypersphere enclosing the hypercube of volume (2 n1 )d.
This corresponds to a line of length 2 n1 , a square of area 4 n1 2 and a cube of volume
8 n1 3 in one, two and three dimensions respectively. Therefore, the equivalent expression
for the LAT at the distance r from the origin is equal to
√d
t(r) =
(2du + k)
(k2 + 4duk)
+ (
√k2 + 4duk
)r.
(10)
In the fast degradation limit, k ≫ u, this equation simplifies into
t(r) ≃
1
k
+
r√d
k
.
In the fast diffusion case, u ≫ k, we obtain
t(r) ≃
1
2k
+
r
2√uk
.
(11)
(12)
The dependence of the LAT on the radial distance r for 1D, 2D and 3D systems for
various sets of parameters is illustrated in Figs. 2-4. In one dimension, the expression for
the local accumulation time derived in the discrete-state SDD model reads as
√k2 + 4uk
while in the continuum SDD model it was shown that27,28
t(r) ≃
+
2u + k
k2 + 4uk
r
t(r) ≃
1
2k
+
r
2√uk
.
,
(13)
(14)
6
The last expression could also be obtained in the limit of very large diffusion, u ≫ k, from Eq.
(13). These results are plotted in Fig. 2. For fast diffusion rates the predictions from discrete
and continuum calculations, as expected, fully agree (see Fig. 2c). The deviations between
discrete and continuum models start to appear for comparable diffusion and degradation
rates (Fig. 2b), and for fast degradation rates (Fig. 2a) the local accumulation times for
discrete case is smaller for all range of distances except very close to the origin.
In this
regime the continuum model cannot be applied, but the discrete-state approach is valid for
analyzing reaction-diffusion processes of morphogen gradients formation.
Similar calculations in two dimensions yield the following expression for the LAT,
t(r) ≃
4u + k
k2 + 8uk
+
√2r
√k2 + 8uk
.
The 2D continuum SDD model predicts the following result,28
(15)
(16)
t(r) ≃
r
2√uk
K1(rpk/u)
K0(rpk/u)
,
where Km(x) is the m-th order modified Bessel function of the second kind. Note that
taking the limit of u ≫ k in our theoretical approach in Eq. (15), which supposed to be
corresponding to the continuum limit, produces a different expression,
t(r) ≃
1
2k
+
r
2√uk
.
(17)
Fig. 3 presents these functions for different sets of parameters. We can see that even for
large diffusion rates (Fig. 3c) the predictions of discrete and continuum models do not fully
agree, but for large distances from the source the differences are small. Again, as for 1D
case, the deviations between two approaches start to build up with decreasing the diffusion
rate (Fig. 3b), and for large degradation rates the LAT for the discrete-state model are
significantly smaller for most distances, except for very small r (Fig. 3a).
For 3D systems the expressions for the local accumulation times in the discrete SDD
model is given by
t(r) ≃
6u + k
k2 + 12uk
+
√3r
√k2 + 12uk
.
The continuum description of the same reaction-diffusion processes yields,27,28
t(r) ≃
r
2√uk
.
(18)
(19)
7
For this case, the LAT are presented in Fig. 4. The observed trends are very similar
to 2D systems, but with stronger deviations between discrete and continuum predictions.
Again, even in the continuum limit our theoretical predictions for the LAT do not agree
with calculations from continuum SDD models,27,28 although for large r the differences are
not significant.
Comparing local accumulation times for discrete-state and continuum SDD models, the
important observation can be made that for all regimes the continuum models in both 2D
and 3D predict t(r = 0) = 0, while in the discrete-state analysis this time is always finite.
Since at t = 0 there are no morphogens in the system and the LAT is associated with the
time to reach the stationary density at given position, it is expected that this quantity to be
finite even at the origin. It seems that predictions of the continuum models do not satisfy
this requirement for d > 1, suggesting that they cannot properly describe reaction-diffusion
processes of formation of signaling molecules profiles close to the origin, even for conditions
when the continuum approximation should hold. No such problems exist for the discrete-
state approach. This is the main reason for predicting multiple time scales in the continuum
description (for d > 1) of the development of signaling molecules profiles. In the discrete
model there is one time scale, given by the LAT, at all distances. The possible reason for this
discrepancy might be special boundary conditions utilized in solving differential equations
that describe these reaction-diffusion processes in the continuum approach.27,28
It is interesting also to investigate the role of dimensionality in the establishment of
morphogen gradients. For fast degradation the discrete model predictions are given by Eq.
(11), while in the fast diffusion limit they are given by Eq. (12). The dependencies of the
local accumulation times on dimension d for the discrete and continuum SDD models are
plotted in Fig. 5 for the sites that are far away from the source (r ≫ 0), and in Fig. 6 for the
sites that are close to the origin (r = 0). Surprisingly, the results are quite different. For fast
degradation rates, the LAT is increasing with d for the sites not so close to the source, while
at the origin and closest sites the LAT is slowly decreasing (compare upper plots in Figs.
5 and 6). A similar behavior is observed for comparable diffusion and degradation rates,
although the effect is getting weaker (see middle plots in Figs. 5 and 6). For continuum
limit, u ≫ k, the LAT in both positions become independent of the dimension, as correctly
predicted by Eq. (12).
The following arguments can be given to understand this behavior in the discrete SDD
8
model. At t = 0 the signaling molecules start at the origin, r = 0. The local accumulation
time is the average time to reach the steady-state density at a given position, so it depends on
possible pathways connecting the origin and any site at r > 0. Increasing the dimensionality
produces more pathways so it takes longer time if the diffusion is the rate limiting step. For
this reason, the LAT depends on d for diffusion rates comparable or smaller the degradation,
while for u ≫ k there is no dependence on the dimension - the degradation is a rate-limiting
step in this case. At sites close to the origin these diffusion pathways do not play any role.
But the stationary density at these sites is also smaller for larger d, so it is faster to reach
the steady-state concentration with increasing d when the diffusion is rate limiting.
B. Mean First-Passage Times
It has been argued before that in order to understand mechanisms of formation of mor-
phogen gradients it is useful to consider mean first-passage times (MFPT) to reach specific
locations for molecules starting from the origin.29 The reason for this is the fact that first-
passage events are the dominating factors determining the local accumulation times at large
distances, at least for one-dimensional systems,29; the explicit connections between these
quantities have been recently studied for d = 1.22 It is important to understand if first-
passage processes describe the morphogen gradient formation in higher dimensions.
To compute MFPT we define f (n1, n2, ..., nd; t) to be a first-passage probability to reach
for the first time the site ~n = (n1, n2, ..., nd) at time t if at t = 0 the particle started at the
origin. The temporal evolution of this function follows a backward master equation,30
df (n1, n2, ..., nd; t)
dt
= uXnn
f (n1, n2, ..., nd; t) − (2ud + k)f (n1, n2, ..., nd; t),
(20)
transformations, we obtain
where Pnn is the operator that sums over all nearest neighbors. Utilizing the Laplace
ef (n1, n2, ..., nd; s) =
(a − 2du + √a2 − 4d2u2)y2n1+2n2+...+2nd − (a − 2du − √a2 − 4d2u2)
2√a2 − 4d2u2yn1+n2+...+nd
.
(21)
where
a = s + 2du + k,
y =ha + √a2 − 4d2u2i /2ud.
(22)
The conditional mean first-passage time to reach the site ~n = (n1, n2, ..., nd) can be found
from the following expression,
9
τ (n1, n2, ..., nd) = −
d ef (n1,n2,...,nd;s)
ds
s=0
ef (n1, n2, ..., nd; s)s=0
After some algebra, the corresponding expression for the MFPT is derived,
(23)
(24)
τ (n1, n2, ..., nd) =
1
√k2 + 4dku
[−
2du + k
√k2 + 4dku
+
+
+
(k + √k2 + 4dku)zn1+n2+...+nd − (k − √k2 + 4dku)z−n1−n2−...−nd
(2du + k + √k2 + 4dku)zn1+n2+...+nd + (2du + k − √k2 + 4dku)z−n1−n2−...−nd
( n1 + n2 +...+ nd )(2du + k + √k2 + 4dku)zn1+n2+...+nd
(k + √k2 + 4dku)zn1+n2+...+nd − (k − √k2 + 4dku)z−n1−n2−...−nd
( n1 + n2 +...+ nd )(2du + k − √k2 + 4dku)z−n1−n2−...−nd
(k + √k2 + 4dku)zn1+n2+...+nd − (k − √k2 + 4dku)z−n1−n2−...−nd
].
where z = (cid:2)2du + k + √k2 + 4duk(cid:3) /2du. For fast degradation rate, k ≫ u, we obtain a
much simpler expression,
τ (n1, n2, ..., nd) ≃ n1 + n2 +...+ nd +1
k
,
which for large radial distances, r ≫ 1, can also be written as
r√d
k
.
τ (r) ≃
(25)
(26)
One can see that this expression agrees well with Eq. (11) at large r. In the opposite limit
of the fast diffusion rates (continuum limit), u ≫ k, one can show that the MFPT are equal
to
τ (n1, n2, ..., nd) ≃ n1 + n2 +...+ nd +1
2√kud
.
For r ≫ 1 it modifies into
τ (r) ≃
r
2√uk
,
(27)
(28)
which asymptotically agree with Eq. (12) at large distances. These results again support
the idea that main contribution to the LAT at large distances from the origin are due to the
MFPT, extending the validity of this idea to all dimensions. This is an important observa-
tion since the first-passage analysis is a well developed mathematical tool that was already
successfully employed for analyzing multiple physical, chemical and biological processes.30
10
To support arguments about the importance of the first-passage events in dynamics of
the morphogen gradient development, the ratio of MFPT over LAT is plotted in Fig. 7 for
different sets of parameters. One can see that this ratio is always approaching 1 for large
distances. Larger degradation rates as well as higher dimensions lead to faster converging to
unity, while in the continuum limit (fast diffusion rates) the effect of dimension disappears.
C. Effective Potentials
Analyzing mechanisms of morphogen gradient formation suggested a new idea that degra-
dation can be viewed as an effective potential that drives the signaling molecules away from
the source.29 Thus morphogens are not simply diffusing with equal probability in each di-
rection, but their motion is biased by this effective potential to move further away from the
source. This concept can be extended and applied for the multi-dimensional SDD models of
creating signaling molecules profiles.
The effective potential can be easily calculated from the stationary profile, leading to
Uef f (n1, n2, ..., nd) ≃ kBT ln P (s)(n1, n2, ..., nd)
= kBT ( n1 + n2 +...+ nd ) ln x,
and it can be rewritten as follows (for i = 1, 2, ..., d),
Uef f (n1, n2, ..., nd) ≃
dXi=1
Uef f (ni), Uef f (ni) = kBT ni ln x.
(29)
(30)
This equation has an important physical meaning suggesting that the overall potential is
a sum of potentials along each of the coordinate axes. Consequently, in higher dimensions
the effective potential is stronger than one dimension. The reason for this is that in higher
dimensions morphogens can diffuse in more directions and thus the probability of returning
to the origin decreases as ∼ 1/d. It also suggests that there is a constant force component,
Fi = −
∂Uef f
∂ ni
= −kBT ln x = kBT /λ,
(31)
along each axis that drives signaling molecules away from the source.
The importance of this concept can be seen in explaining most dynamic properties of
morphogen reaction-diffusion systems. The linear dependence of the LAT on distances
from the sources [see Eq.
(9)] is the consequence of the effective potential that changes
11
the unbiased diffusion of morphogen molecules into a driven motion. Similarly, the linear
dependencies of the MFPT on distances have the origin: see Eq.(24). It also provides an
alternative explanation for dependence of the LAT on dimension for sites near the source
(Fig. 6):
increasing d make this potential stronger so it drives particles faster to their
destinations. The same reasoning can be used to understand why the MFPT approximate
the LAT better at higher dimensions or at faster degradations (Fig. 7).
D. Variance of Local Accumulation Times
The advantage of presented theoretical method is that it allows us to calculate all dy-
namic properties of the morphogen gradient formation. To illustrate this, let consider higher
moments of the local accumulation times. The LAT itself is the first moment as indicated
in Eq. (8). The second moment, which is a mean-squared local accumulation time, can be
also calculated from the local relaxation function,
< t2(n1, n2, ..., nd) >= −
∞Z0
t2 dR(n1, n2, ..., nd; t)
dt
dt = −2
deR(n1, n2, ..., nd; s)
ds
s=0
.
(32)
Substituting the explicit expression for eR(n1, n2, ..., nd; s) we obtain,
2( n1 + n2 +...+ nd )2 − 2
< t2(n1, n2, ..., nd) >=
(k2 + 4duk)
+
4(2du + k)( n1 + n2 +...+ nd )
(k2 + 4duk)3/2
+
6(2du + k)2
(k2 + 4duk)2 .
It can be shown that generally the m-th moment of the LAT is given by
< tm(n1, n2, ..., nd) >= (−1)m−1m
dm−1eR(n1, n2, ..., nd; s)
dsm−1
.
s=0
(33)
(34)
The explicit forms for the first and second moments of the LAT allow us to calculate a
variance, which gives a measure of fluctuations in the local accumulation times. The variance
of the local accumulation time is equal to
σ(n1, n2, ..., nd) = √< t2 > − < t >2 =(cid:20)( n1 + n2 +...+ nd )2 − 2
(k2 + 4duk)2(cid:21)1/2
2(2du + k)( n1 + n2 +...+ nd )
(k2 + 4duk)3/2
5(2du + k)2
(k2 + 4duk)
+
+
.
(35)
In terms of the radial distance r the variance can be written as
σ(r) =" dr2 − 2
(k2 + 4duk)
2r√d(2du + k)
(k2 + 4duk)3/2 +
+
5(2du + k)2
(k2 + 4duk)2#1/2
.
12
(36)
In the limit of fast diffusion rates (continuum limit) the expression for the variance is simpler,
√5
2k
+
r
2√5uk
.
σ(r) ≃
(37)
This result implies that the variance becomes independent of the dimension for u ≫ k. For
the case of the fast degradation rates (k ≫ u) the variance behavior is different,
σ(r) ≃ pdr2 + 2r√d + 3
k
.
(38)
In this limit the variance increases with d but becomes independent of the diffusion rate.
The variances normalized with respect to the LAT are presented in Fig. 8. We can
see that at large distance the ration σ(r)/t(r) is always approaching unity. The increase
in the degradation rates lowers the variance, while increasing the diffusion rate make the
system more noisy. At fast degradation rates, increasing the dimension lowers the variance
(Fig. 8a), while for large diffusion rates there is no dependence on d. The importance of
these observations is that they suggest possible ways of how nature might control noise in
morphogen gradient systems.
III. A DISCRETE-STATE STOCHASTIC SDD MODEL WITH EXTENDED
SOURCE REGION
In the model discussed before it was assumed that the source of signaling molecules
is sharply localized at the origin.
In real systems the morphogen production is more
delocalized,1,2,7 and it raises a question of how the size of the source region affects the dynam-
ics of morphogen gradient formation. To answer this question, the original d-dimensional
discrete-state stochastic SDD model is generalized to take into account this effect by assum-
ing that morphogens can be produced from the discrete point sources distributed inside a
hypercubic of volume (2R)d in d-dimensional space. This corresponds to a line of length 2R,
a square of area 4R2 and a cube of volume 8R3 in one, two and three dimensions respectively.
It is assumed that the production rate at each site is equal to Q.
13
It is convenient to introduce a propagator function G(m1, m2, ..., md; t0n1, n2, ..., nd; t)
defined as the conditional probability to find the particle at site ~n = (n1, n2, ..., nd) if it
starts at t0 at site ~m = (m1, m2, ..., md).21 Then the probability P (n1, n2, ..., nd; t) of finding
the particle at site ~n = (n1, n2, ..., nd) at time t can be expressed as a superposition of the
corresponding propagators:
P (n1, n2, ..., nd; t) =
RXm1=−R
RXm2=−R
...
RXmd=−R
tZ0
G(m1, m2, ..., md; t0n1, n2, ..., nd; t)dt0,
(39)
where the summation is performed over the region of space where particle sources are located.
The corresponding master equations for temporal evolution of probabilities can be solved in
the large-time limit (see Appendix A), leading to
P (s)(n1, n2, ..., nd) =
2QΓd
√k2 + 4duk
exp (− n1 − n2 −...− nd
λ
+
dR
λ
),
(40)
with x and λ defined in Eq. (6), while a new function Γ is given by
Γ =
xR+1 + xR − 2
x − 1
.
(41)
It characterizes the extended source region. In the case of R = 0 we get Γ = 1 and the
results of Sec. II are fully recovered. In general, Γ can be a complex function that strongly
depends on geometry and distribution of morphogen sources.
To simplify calculations, here we assumed that the morphogen molecules are produced
in the hypercubic region around the origin, but our analysis can be easily extended to
geometrically more complex source regions.
It is also important to note that our model
differ from the continuum SDD models where morphogens can be produced only at the
surface of the region of size R.27,28 In our case, which is much closer to the real situation,
the sources of the signaling molecules are distributed inside of the production region, and
morphogens can a diffuse in the source area, cross into non-productive region and return
back.
The local accumulation times for the discrete SDD model with the extended source can
be calculated following the same procedure as explained in Sec. II. It yields
t(n1, n2, ..., nd) =
1
√k2 + 4duk(cid:20) 2du + k
√k2 + 4duk
+ n1 + n2 +...+ nd −dR
+d
(R + 1)xR+1 + RxR − xΓ
Γ(x − 1)
(cid:21) ,
(42)
14
which at the distance r (along the vector ~r = (n1, n2, ..., nd) with n1 = n2 = ... = nd )
is modified into
t(r) =
d
√k2 + 4duk(cid:20)
2du + k
d√k2 + 4duk
+
r
√d − R +
(R + 1)xR+1 + RxR − xΓ
Γ(x − 1)
(cid:21) .
In the limit of fast degradation rates, k ≫ u, the resulting expression is simpler,
t(r) ≃
1
k
+
d
k
(
r
√d − R),
(43)
(44)
which for R = 0 reduces, as expected, to Eq. (11). For the special position on the surface
of the producing area at the edge of the hypercube, r = R√d, it predicts even simpler
expression t(R√d) ≃ 1/k. For fast diffusion rates (u ≫ k) one can show that the LAT is
given by
t(r) ≃
1
2k
+
√d
2√uk(cid:20)(
r
√d − R) +
2R2
2R + 1(cid:21) ,
(45)
which for R = 0 reduces to Eq. (12).
The LAT at the surface of the production area for different dimensions are plotted in
Fig. 9. One can see that for large degradation rates the LAT become independent of d
(Fig. 9a), while for larger diffusion rates there is a dependence on the dimensionality. It
can be explained using the following arguments. At very large k, fluxes from other source
sites do not reach this specific location - particles are degraded before they can diffuse to
neighboring sites.
In this case the LAT should not depend on R and on d - only local
dynamics at the given site is important. For faster diffusion (Fig. 9b and 9c) the role of
fluxes from neighboring sites becomes more important so the dependence on R will show up.
But the contribution form the sites that are further away will be much smaller. As a result
there is a saturation behavior at R ≫ 1.
higher stationary concentrations so it takes more time to reach the steady-state conditions,
In addition, increasing the dimension leads to
and this explains why the LAT are the highest for 3D and the lowest for 1D systems.
Our analysis can also be extended for computation of mean first-passage times for the
systems with extended production volume. The explicit formulas for MFPT are quite bulky
and they are fully derived in Appendix B. Here we present simpler limiting expressions. For
slow degradation (conditions close to the continuum limit) we obtain
τ (n1, n2, ..., nd) ≃ ( n1 + n2 +...+ nd +1 − dR)/2√kud,
(46)
15
which for the single-point source (R = 0) reduces to Eq. (27). For the opposite limit of slow
diffusion rates, k ≫ u, the MFPT depends only on the degradation rate,
τ (n1, n2, ..., nd) ≃ ( n1 + n2 +...+ nd +1 − dR)/k,
(47)
which also for the case of R = 0 is identical to Eq. (25). In both limiting cases MFPT
decrease for larger production areas since there are source sites closer to the given position.
Varying the size of the production region modifies also the effective potential that mor-
phogen molecules experience in the system due to degradation. From the stationary densities
we obtain,
Uef f (n1, n2, ..., nd) ≃ kBT Γd( n1 + n2 +...+ nd −dR) ln x,
which can be rewritten as
Uef f (n1, n2, ..., nd) ≃
dXi=1
[Uef f (ni) − Uef f (R)] ,
where we defined (for i = 1, , 2..., d)
Uef f (ni) = kBT Γd ni ln x, Uef f (R) = kBT ΓdR ln x.
(48)
(49)
(50)
The corresponding force along the axis i that effectively pushes morphogens away from the
source is given by
Fi = −
∂Uef f
∂ ni
leading to stronger forces with increasing the size of the producing area.
= −kBT Γd ln x = kBT Γd/λ,
(51)
IV. SUMMARY AND CONCLUDING REMARKS
We developed a multi-dimensional discrete-state stochastic theoretical framework for un-
derstanding reaction-diffusion processes of morphogen gradients formation. The approach
provides a full analytical description of stationary state and dynamic properties of complex
systems where signaling molecules profiles are created. It allowed us to fully analyze the
role of discreteness by comparing with current continuum theoretical models, as well as the
effect of the dimensionality.
It is found that at large times the system will reach stationary exponential density profiles
with the decay length that increases with the dimension, in contrast to the continuum
16
methods which predict the decay length to be independent of d. The differences between two
approaches become larger in analyzing dynamic properties such as the local accumulation
times that describe the relaxation to the stationary-state behavior. Continuum models
predict that the LAT is approaching zero at the source, resulting in multiple time scales
that control dynamics of the system. In contrast, our calculations suggest that the local
accumulation times are always finite and they provide the only time scale to describe the
kinetics of morphogen gradients formation. Thus, it is argued that current continuum models
cannot be used in analyzing these complex reaction-diffusion dynamics at distances closer
to the source, while our discrete approach does not have any problems.
From the presented discrete method an interesting dependence of dynamic properties on
dimensions is observed. It is found that for sites close to the source, when the degradation
is faster than the diffusion, the LAT times decrease with the dimension, while for regions far
away form the source the dependence is reversed. At the same time, for large diffusion rates
no effect is observed at any distance. It is explained by accounting for possible pathways
connecting the source and the given location in the system. We also analyzed another
dynamic property, mean first-passage times. It is shown that at large distances from the
source the MFPT provide an excellent approximation for the LAT, and the approximation
is better for higher dimensions and larger degradation rates, while at the continuum limit
(fast diffusion) there is no dependence on d and the approximation works not as well.
The concept that degradation processes can be viewed as an effective potential that
pushes signaling molecules away from the source has been extended to multi-dimensional
systems. It is found that increasing the dimension makes this potential stronger, and this
simple idea was powerful enough to explain most trends in dynamic properties, such as the
linear dependence of the LAT and MFPT on the distances and the effect of dimensions. In
addition, the method allowed us to compute higher moments of the local accumulation times,
and specific calculations have been made for estimation of variances. Finally, we extended
our method for analyzing reaction-diffusion systems with variable range of production region
by explicitly estimating all dynamic and stationary properties and their dependencies of the
size of the source volume and dimensions.
It was argued that the presented discrete-state stochastic approach allows to capture all
relevant physical-chemical properties of the development of morphogen gradients. The main
success of the method is a full analytical description of all involved processes at all times and
17
distances. Another advantage is that other biochemical and biophysical processes can be
consistently incorporated. For example, it will be crucial to extend the models to take into
account more complex phenomena such as non-uniform production rates, cooperative mech-
anisms of degradation and possible bindings/unbindings of morphogens to other molecules
in the system. It will be also very important to test these theoretical ideas in experimental
studies.
ACKNOWLEDGMENTS
We would like to acknowledge the support from the Welch Foundation (grant C-1559).
APPENDIX A: CALCULATIONS OF STATIONARY DENSITY PROFILES FOR
THE SYSTEM WITH EXTENDED PRODUCTION VOLUME
In Sec.
III we introduced the propagator or Green function, G(m1, m2, ..., md; t0 =
0n1, n2, ..., nd; t), that characterizes the system with extended range of morphogen pro-
duction. Its temporal evolution is governed by the following master equation:
dG(m1, m2, ..., md; t0n1, n2, ..., nd; t)
G(m1, m2, ..., md; t0n1, n2, ..., nd; t)
+(2ud + k)G(m1, m2, ..., md; t0n1, n2, ..., nd; t) = Qδ(n1 − m1)δ(n2 − m2)...δ(nd − md),
dt
− uXnn
(52)
where δ(x) is a Kronecker delta andPnn is the operator that sums over the nearest neigh-
bors,
Xnn
G(m1, m2, ..., md; t0n1, n2, ..., nd; t) = G(m1, m2, ..., md; t0n1 − 1, n2, ..., nd; t)
+G(m1, m2, ..., md; t0n1 + 1, n2, ..., nd; t) + G(m1, m2, ..., md; t0n1, n2 − 1, ..., nd; t) + ...
(53)
Eq. (52) can be rewritten as
LG(m1, m2, ..., md; t0n1, n2, ..., nd; t) = δ(n1 − m1)δ(n2 − m2)...δ(nd − md)
(54)
with L defined as an operator acting on the Green function. Here, the Green function can
be regarded as an auxiliary function which satisfies the appropriate boundary conditions
18
generated by a singularly point source located at ~m = (m1, m2, ..., md). The corresponding
Green function for the steady state reads then
G(s)(m1, m2, ..., md; t0n1, n2, ..., nd; t → ∞) =
2Qxn1−m1+n2−m2+...+nd−md
√k2 + 4duk
.
(55)
Now we can calculate probability function defined in Eq. (39) by summing over the source
region,
P (s)(n1, n2, ..., nd) =
2Qxn1+n2+...+nd
√k2 + 4duk
RXm1=−R
RXm2=−R
...
RXmd=−R
x−m1−m2−...−md.
(56)
The summation over m1, ...md can be performed in the following way,
RXm1=−R
x−m1 = 2
RXm1=0
x−m1 − 1 =
x−R(xR + xR+1 − 2)
x − 1
.
(57)
Substituting this result into Eq.(56) yields
P (s)(n1, n2, ..., nd) =
2Qxn1+n2+...+nd−dR
√k2 + 4duk
(
xR + xR+1 − 2
x − 1
)d =
2QΓdxn1+n2+...+nd−dR
√k2 + 4duk
.
(58)
APPENDIX B: CALCULATIONS OF MEAN FIRST-PASSAGE TIMES FOR THE
SYSTEM WITH EXTENDED PRODUCTION VOLUME
Similarly to the approach explain in Appendix A, we define F (m1, m2, ..., md; t0n1, n2, ..., nd; t)
as a first-passage conditional probability to reach for the first time the site ~n = (n1, n2, ..., nd)
if at t0 the particle starts at ~m = (m1, m2, ..., md).
dF (m1, m2, ..., md; t0n1, n2, ..., nd; t)
dt
= uXnn
F (m1, m2, ..., md; t0n1, n2, ..., nd; t)
−(2ud + k)F (m1, m2, ..., md; t0n1, n2, ..., nd; t).
Here again Pnn is the sum operator explained in Eq.
first-passage probability can be expressed as a sum over these propagators,
(59)
(53). The corresponding mean
f (n1, n2, ..., nd; t) =
RXm1=−R
RXm2=−R
...
RXmd=−R
tZ0
F (m1, m2, ..., md; t0n1, n2, ..., nd; t)dt0.
(60)
It can be shown that
19
1(a − 2du + √a2 − 4d2u2)y2n1+2n2+...+2nd−2dR
2√a2 − 4d2u2yn1+n2+...+nd−dR
− Θd
1
1
Θd
ef (n1, n2, ..., nd; s) =
2(a − 2du − √a2 − 4d2u2)
, (61)
where we defined
Θ1 = (
yR+1
1 + yR
y1 − 1
1 − 2
and
(a + √a2 − 4d2u2)
2du
y1 =
,
y2 =
), Θ2 = (
yR+1
2 + yR
y2 − 1
(a − √a2 − 4d2u2)
,
2du
2 − 2
),
(62)
a = s + 2du + k.
(63)
The MFPT to reach the site ~n = (n1, n2, ..., nd) , can be found from the following expres-
sion,
τ (n1, n2, ..., nd) = −
The final expression is given by
d ef (n1,n2,...,nd;s)
ds
s=0
ef (n1, n2, ..., nd; s)s=0
.
(64)
τ (n1, n2, ..., nd) =
1
√k2 + 4dku
[−
2du + k
√k2 + 4dku
+
+
+
+
Φd
Φd
Φd
2(2du + k − √k2 + 4dku)z−n1−n2−...−nd+dR
2(k − √k2 + 4dku)z−n1−n2−...−nd+dR
1(2du + k + √k2 + 4dku)zn1+n2+...+nd−dR + Φd
Φd
1(k + √k2 + 4dku)zn1+n2+...+nd−dR − Φd
1( n1 + n2 +...+ nd −dR)(2du + k + √k2 + 4dku)zn1+n2+...+nd−dR
2(k − √k2 + 4dku)z−n1−n2−...−nd+dR
1(k + √k2 + 4dku)zn1+n2+...+nd−dR − Φd
2( n1 + n2 +...+ nd −dR)(2du + k − √k2 + 4dku)z−n1−n2−...−nd+dR
2(k − √k2 + 4dku)z−n1−n2−...−nd+dR
1(k + √k2 + 4dku)zn1+n2+...+nd−dR − Φd
Ω1(2du + k + √k2 + 4dku)zn1+n2+...+nd−dR − Ω2(2du + k − √k2 + 4dku)z−n1−n2−...−nd+dR
Φd
1(k + √k2 + 4dku)zn1+n2+...+nd−dR − Φd
2(k − √k2 + 4dku)z−n1−n2−...−nd+dR
Φd
Φd
];
(65)
where
llz1 =
Ω1 = dΦd−1
1
[
(2du + k + √k2 + 4duk)
1 − 2
2du
zR+1
1 + zR
z1 − 1
1 − 2zR+1
1 + 2z1
Φ1 = (
1 − RzR
RzR+2
(z1 − 1)2
,
z2 = (2du+k−√k2+4duk)
2du
;
),
Φ2 = ( zR+1
2 −2
2 +zR
z2−1
);
],Ω2 = dΦd−1
2
[−RzR+2
2 +RzR
2 +2zR+1
2 −2z2
(z2−1)2
(66)
(67)
].
(68)
The auxiliary functions Φ1,Φ2,Ω1 and Ω2 depend on the range of the source production.
In the case of R = 0 we obtain Φ1 = Φ2 = 1 and Ω1 = Ω2 = 0, and Eq.(65) reduces, as
expected, to Eq. (24).
20
1 A. Martinez-Arias and A. Stewart, Molecular Principles of Animal Development (Oxford Uni-
versity Press, New York, 2002).
2 H. Lodish, A. Berk, C.A. Kaiser, M. Krieger, M.P. Scott, A. Bretscher, H. Ploegh, and P.
Matsudaira, Molecular Cell Biology 6-th ed., (W.H. Freeman, New York, 2007).
3 L. Wolpert, J. Theor. Biol. 25, 1 (1969).
4 T. Tabata and Y. Takei, Development 131, 703 (2004).
5 F.H. Crick, Nature 225, 420 (1970).
6 A.M. Turing, Phil. Trans. Roy. Soc. Lond. 237, 37 (1952).
7 A. Porcher and N. Dostatni, Curr. Biol. 20, R249 (2010).
8 T. Gregor, E.F. Wieschaus, A.P. McGregor, W. Bialek and D.W. Tank, Cell 130, 141 (2007).
9 A. Kicheva, P. Pantazis, T. Bollenbach, Y. Kalaidzidis, T. Bittig, F. Julicher and M. Gonzales-
Gaitan, Science 315, 521 (2007).
10 S.R. Yu, M. Burkhardt, M. Nowak, J. Ries, Z. Petrasek, S. Scholpp, P. Schwille and M. Brand,
Nature 461, 533 (2009).
11 M. Kerszberg and L. Wolpert, J. Theor. Biol. 191, 103 (1998).
12 E.V. Entchev, A. Schwabedissen and M. Gonzales-Gaitan, Cell 103, 981 (2000).
13 P. Muller, K. W. Rogers, B. M. Jordan, J. S. Lee, D. Robson, S. Ramanathan, A. F. Schier,
Science 336, 721 (2012).
14 J.A. Drocco, O. Grimm, D.W. Tank and E. Wieschaus, Biophys. J. 101, 1807 (2011).
15 S. C. Little, G. Tkacik, T. B. Kneeland, E. F. Wieschaus, T. Gregor, PLoS Biol. 9 e1000596
(2011)
16 A. Spirov, K. Fahmy, M. Schneider, E. Frei, M. Noll and S. Baumgartner, Development 136
605-614 (2009)
17 S. Zhou, W. C. Lo, J. L. Suhalim, M. A. Digman, E. Grattom, Q. Nie and A. D. Lander,
Current Biology 22 668-675 (2012)
18 T.B. Kornberg, Biophys. J. 103, 2252 (2012).
21
19 J.A. Drocco, E. F. Wieschaus and D. W. Tank, Phys. Biol. 9, 055004 (2012).
20 A.M. Berezhkovskii, C. Sample, and S.Y. Shvartsman, Biophys. J. 99, L59 (2010).
21 A.M. Berezhkovskii, J. Chem. Phys. 135, 07412 (2011).
22 A.M. Berezhkovskii and S.Y. Shvartsman, J. Chem. Phys. 135, 154115 (2011).
23 A.M. Berezhkovskii, C. Sample and S.Y. Shvartsman, Phys. Rev. E 83, 051906 (2011).
24 A. M. Berezhkovskii and A.M. Shvartsman J. Chem. Phys. 138, 244105(2013).
25 A. Mogilner and D. Odde, Trends Cell Biol. 21, 692-700 (2011).
26 P. Muller, K. W. Rogers, S. R. Yu, M. Brand, A. F. Schier, Development 140, 1621 (2013).
27 P. V. Gordon, C. B. Muratov, and S. Y. Shvartsman , J. Chem. Phys. 138, 104121 (2013).
28 A. J. Ellery, M. J. Simpson, and S. W. McCue , J. Chem. Phys. 139, 017101 (2013).
29 A.B. Kolomeisky, J. Phys. Chem. Lett. 2, 1502 (2011).
30 S. Redner,A Guide to First-Passage Processes (Cambridge University Press, New York, 2001).
22
Fig.1. A schematic of the discrete-state SDD model for establishment of morphogen gradi-
ents in d dimensions. A specific case of d = 2 is presented. Signaling molecule are generated
at the origin (shown in red) with a rate Q. Particles can also diffuse along the lattice to the
neighboring sites with a rate u, or they might be degraded with a rate k.
Fig. 2. Local accumulation times in one dimension as a function of distance from the source
r for discrete-state and continuum SDD models. (a) Fast degradation rates, k = 1, u = 0.01;
(b) Comparable diffusion and degradation rates, k = u = 1; and (c) Fast diffusion rates,
k = 1, u = 100. The predictions for the continuum model are taken from Refs.27,28. Insets
show the same plots for larger length scales.
Fig. 3. Local accumulation times in two dimensions as a function of distance from the
source r for discrete-state and continuum SDD models. (a) Fast degradation rates, k = 1,
u = 0.01; (b) Comparable diffusion and degradation rates, k = u = 1; and (c) Fast diffusion
rates, k = 1, u = 100. The predictions for the continuum model are taken from Ref.28.
Insets show the same plots for larger length scales.
Fig. 4. Local accumulation times in three dimensions as a function of distance from the
source r for discrete-state and continuum SDD models. (a) Fast degradation rates, k = 1,
u = 0.01; (b) Comparable diffusion and degradation rates, k = u = 1; and (c) Fast diffusion
rates, k = 1, u = 100. The predictions for the continuum model are taken from Refs.27,28.
Insets show the same plots for larger length scales.
Fig. 5. Local accumulation times at the position r = 10 as a function of spatial dimensions.
Upper curve corresponds to the fast degradation rates, k = 1, u = 0.01. The middle curve
is for comparable diffusion and degradation rates, k = u = 1. The lower curve describes the
fast diffusion regime, k = 1, u = 100.
Fig. 6. Local accumulation times at the origin r = 0 as a function of spatial dimensions.
Upper curve corresponds to the fast degradation rates, k = 1, u = 0.01. The middle curve
is for comparable diffusion and degradation rates, k = u = 1. The lower curve describes the
fast diffusion regime, k = 1, u = 100.
Fig. 7. The ratio of MFPT over LAT as a function of distance from the source for different
23
dimensions for the discrete-state SDD models. (a) Fast degradation rates, k = 1, u = 0.01;
(b) Comparable diffusion and degradation rates, k = u = 1; and (c) Fast diffusion rates,
k = 1, u = 100.
Fig. 8. The ratio of variance over LAT as a function of distance from the source for different
dimensions for the discrete-state SDD models. (a) Fast degradation rates, k = 1, u = 0.01;
(b) Comparable diffusion and degradation rates, k = u = 1; and (c) Fast diffusion rates,
k = 1, u = 100.
Fig. 9. Local accumulation times as a function of the size of the production region at the
source surface, r = R√d. The size of the source region is expressed in units of R√d. (a)
Fast degradation rates, k = 1, u = 0.01; (b) Comparable diffusion and degradation rates,
k = u = 1; and (c) Fast diffusion rates, k = 1, u = 100.
24
Q
u
u
u
u
k
Figure 1. Teimouri and Kolomeisky
25
(a) k=1 , u=0.01
(b) k=1, u=1
0
50 100 150 200
2
1.5
t(r)
1
100
50
0
0 50 100 150 200
1
2
2.5
3
0.5
0
0.5
1
(c) k=1, u=100
2
2.5
3
1.5
r
0.5
1.5
r
continuum SDD
discrete SDD
0.65
t(r)
1000
500
0
15
10
5
0
0
t(r)
0.6
0.55
10
5
0
0
50 100 150 200
0.5
0
0.5
1
2
2.5
3
1.5
r
Figure 2. Teimouri and Kolomeisky
(a) k=1 , u=0.01
(b) k=1 , u=1
26
20
15
t(r)
10
1000
500
0
0
50 100 150 200
5
0
0
0.5
1
continuum SDD
discrete SDD
t(r)
1.5
r
0.6
0.5
0.4
0.3
0.2
0.1
0
2
1.5
t(r)
1
0.5
2
2.5
0
3
0
(c) k=1 , u =100
100
50
0
0 50 100150200
2
2.5
3
0.5
1
1.5
r
10
5
0
0
50 100 150 200
0
0.5
1
1.5
r
2
2.5
3
Figure 3. Teimouri and Kolomeisky
(a) k=1 , u=0.01
(b) k=1 , u=1
27
15
10
5
1000
800
600
400
200
0
t(r)
0 50 100 150 200
0
0
0.5
1
2
1.5
r
continuum SDD
discrete SDD
t(r)
0.6
0.4
0.2
2
1.5
t(r)
1
0.5
0
2.5
3
0
(c) k=1 , u=100
100
80
60
40
20
0
0
50
100 150 200
2
2.5
3
0.5
1
1.5
r
10
5
0
0
50 100 150 200
0
0
0.5
1
2
2.5
3
1.5
r
Figure 4. Teimouri and Kolomeisky
28
k=1; u=0.01
k=1;u=1
k=1; u=100
2
d
3
15
10
)
0
1
=
r
(
t
5
0
1
Figure 5. Teimouri and Kolomeisky
29
)
0
=
r
(
t
1
0.8
0.6
0.4
0.2
0
1
k=1;u=0.01
k=1; u=1
k=1; u=100
2
d
3
Figure 6. Teimouri and Kolomeisky
30
(a) k=1, u=0.01
(b) k=1 , u=1
1
0.8
0.6
0.4
0.2
)
r
(
t
/
)
r
(
τ
1
0.8
0.6
0.4
0.2
)
r
(
t
/
)
r
(
τ
0
25
0
(c) k=1 , u=100
5
10
r
15
20
25
15
20
0
0
5
10
d=1
d=2
d=3
)
r
(
t
/
)
r
(
τ
r
1
0.8
0.6
0.4
0.2
0
0
5
10
r
15
20
25
Figure 7. Teimouri and Kolomeisky
(a) k=1, u=0.01
(b) k=1 , u=1
31
)
r
(
t
/
)
r
(
σ
2.2
2
1.8
1.6
1.4
1.2
1
0
0.5
1
d=1
d=2
d=3
)
r
(
t
/
)
r
(
σ
1.5
r
2.2
2
1.8
1.6
1.4
1.2
1
)
r
(
t
/
)
r
(
σ
2.2
2
1.8
1.6
1.4
1.2
1
2
2.5
3
0
(c) k=1 , u=100
0.5
1
1.5
r
2
2.5
3
0
0.5
1
1.5
r
2
2.5
3
Figure 8. Teimouri and Kolomeisky
(a) k=1, u=0.01
(b) k=1 , u=1
32
2
1.5
1
0.5
0
2
6
4
Rd1/2
2
2
1.5
1
)
2
/
1
d
R
(
t
0.5
8
0
(c) k=1, u=100;
5
10
Rd1/2
15
)
2
/
1
d
R
(
t
d=1
d=2
d=3
)
2
/
1
d
R
(
t
1.5
1
0.5
0
20
80
60
40
Rd1/2
Figure 9. Teimouri and Kolomeisky
|
1803.06493 | 1 | 1803 | 2018-03-17T12:13:53 | Actin filaments growing against an elastic membrane: Effect of membrane tension | [
"physics.bio-ph",
"cond-mat.soft",
"cond-mat.stat-mech"
] | We study the force generation by a set of parallel actin filaments growing against an elastic membrane. The elastic membrane tries to stay flat and any deformation from this flat state, either caused by thermal fluctuations or due to protrusive polymerization force exerted by the filaments, costs energy. We study two lattice models to describe the membrane dynamics. In one case, the energy cost is assumed to be proportional to the absolute magnitude of the height gradient (gradient model) and in the other case it is proportional to the square of the height gradient (Gaussian model). For the gradient model we find that the membrane velocity is a non-monotonic function of the elastic constant $\mu$, and reaches a peak at $\mu=\mu^\ast$. For $\mu < \mu^\ast$ the system fails to reach a steady state and the membrane energy keeps increasing with time. For the Gaussian model, the system always reaches a steady state and the membrane velocity decreases monotonically with the elastic constant $\nu$ for all nonzero values of $\nu$. Multiple filaments give rise to protrusions at different regions of the membrane and the elasticity of the membrane induces an effective attraction between the two protrusions in the Gaussian model which causes the protrusions to merge and a single wide protrusion is present in the system. In both the models, the relative time-scale between the membrane and filament dynamics plays an important role in deciding whether the shape of elasticity-velocity curve is concave or convex. Our numerical simulations agree reasonably well with our analytical calculations. | physics.bio-ph | physics |
September 20, 2018
Actin filaments growing against an elastic membrane: Effect of membrane tension
Raj Kumar Sadhu and Sakuntala Chatterjee
Department of Theoretical Sciences, S. N. Bose National Centre for Basic Sciences,
Block JD, Sector III, Salt Lake, Kolkata 700106, India.
We study the force generation by a set of parallel actin filaments growing against an elastic
membrane. The elastic membrane tries to stay flat and any deformation from this flat state, either
caused by thermal fluctuations or due to protrusive polymerization force exerted by the filaments,
costs energy. We study two lattice models to describe the membrane dynamics. In one case, the
energy cost is assumed to be proportional to the absolute magnitude of the height gradient (gradient
model) and in the other case it is proportional to the square of the height gradient (Gaussian model).
For the gradient model we find that the membrane velocity is a non-monotonic function of the elastic
constant µ, and reaches a peak at µ = µ∗. For µ < µ∗ the system fails to reach a steady state and the
membrane energy keeps increasing with time. For the Gaussian model, the system always reaches a
steady state and the membrane velocity decreases monotonically with the elastic constant ν for all
nonzero values of ν. Multiple filaments give rise to protrusions at different regions of the membrane
and the elasticity of the membrane induces an effective attraction between the two protrusions in the
Gaussian model which causes the protrusions to merge and a single wide protrusion is present in the
system. In both the models, the relative time-scale between the membrane and filament dynamics
plays an important role in deciding whether the shape of elasticity-velocity curve is concave or
convex. Our numerical simulations agree reasonably well with our analytical calculations.
PACS numbers: 05.40.-a, 87.16.aj, 87.16.Ka, 87.15.A-
I.
INTRODUCTION
Cell motility plays an important role in many important biological processes such as morphogenesis, wound repair,
cancer invasion etc.
[1 -- 5]. Actins and microtubules are cytoskeletal proteins whose polymerization and depolymer-
ization generate significant force that induce the cell motility. Inside the cell, these filaments grow against the cell
membrane which acts as a biological barrier. The polymerization force generated by these filaments is measured in
many in vitro cases by applying an opposing external load on the barrier. The velocity of the barrier decreases with
the external load and this dependency, known as the force-velocity curve, is an important characteristic of the force
generation mechanism.
The plasma membrane against which the actin filaments grow has a role of central importance in the motility process
[6 -- 9]. The membrane tension couples different processes that are part of cell migration [10] and hence it is important
to understand how the membrane tension affects actin polymerization. Extension of the plasma membrane that takes
place during cell motility and cell spreading depends strongly on membrane tension. In Ref. [11] it was shown that
when the membrane tension is decreased by adding certain ampiphillic compounds, the lamellopodial extension rate
increases, while in situations where membrane tension is increased using osmotic methods, the extension rate drops.
Although the membrane tension is generally considered as an obstacle to cell movement, it was shown in Ref.
[12]
that the membrane tension also optimizes the motility in Caenorhabditis elegans sperm cells by streamlining actin
polymerization in the direction of movement. In an in vitro reconstitution of lipid bilayer, it has been found that
membrane elasticity causes formation of filament bundles that support membrane protrusions from branched filament
network [13]. A flexible membrane enhances formation of filopodial protrusion and also allows merging of smaller
neighboring protrusions into a larger one [14].
Above studies show that the elastic properties of a flexible membrane have significant influence on actin filament
polymerization and force generation process. This gives rise to a more general question: What happens when a
flexible, elastically deformable obstacle is placed in the path of growing filaments? Some interesting results were
obtained in this direction in Ref. [15], where an elastic membrane was described using a solid-on-solid model [16, 17]
and the leading edge of the actin mesh was modelled by an advancing uncorrelated front. In absence of the advancing
mesh, the equilibrium thermal fluctuations of the membrane can be characterized by Edwards-Wilkinson universality
class [15, 18]. However, when driven by the polymerizing mesh, the growth exponents of the membrane shape
fluctuations show a slow crossover to Kardar-Parisi-Zhang universality class [15, 19]. This similar modeling strategy
was generalized in Ref. [20] to understand certain aspects of crawling motion of a cell on a flat substrate, particularly,
how the asymmetry between the leading and trailing edge of the membrane decides the shape and velocity of the
membrane.
2
In this paper, we study a coupled system consisting of an elastic membrane and a set of growing filaments, which
pushes the membrane from below. Due to elastic energy, the membrane tries to stay flat and any deformation from
this flat state costs energy. A height field describes the membrane configuration such that in a flat state, height is
same everywhere on the membrane and whenever a height gradient develops, it costs energy. We consider two different
lattice models to describe the membrane dynamics. In one case, we use the model introduced in Ref. [17] to describe
protrusions in a membrane and studied later in Ref. [15]. Here, the energy cost of a local membrane deformation is
assumed to be proportional to the absolute magnitude of the local height gradient. Below we refer this model as the
"gradient model". In the second case, the energy cost is proportional to the square of the local height gradient [16].
We call it the "Gaussian model". The filaments are modelled by rigid parallel rods, which are pushing against the
membrane and causing its deformation. We are interested to find out how the membrane elasticity affects the growth
of the filaments and how the polymerization force exerted by the filaments affects the membrane deformations.
We find that for the gradient model, the average membrane velocity shows a peak as a function of the membrane
tension µ. This is somewhat surprising since a larger membrane tension is expected to make the polymerization
process more costly and hence the growth process slower. We show that the peak results from the competition
between the polymerization force of the filaments and the elastic force of the membrane. We also show that for small
µ, the system does not have a steady state and the total contour length of the membrane keeps increasing with time.
For large values of µ, the system reaches a steady state and the membrane velocity then decreases with µ. For the
Gaussian model, however, the membrane velocity decreases monotonically with the membrane elastic constant ν and
steady state is reached for all nonzero values of ν. In both the models, the elasticity-velocity curve plays similar role as
the force-velocity characteristic obtained in earlier studies [21 -- 28]. We show that the relative time-scale between the
membrane dynamics and the filament dynamics is crucial to determine the qualitative shape of the elasticity-velocity
curve. The importance of the relative time scale was also highlighted in our earlier work [21] where instead of an
elastically deformable membrane, we had used a Kardar-Parisi-Zhang surface whose height fluctuations were biased
in the direction opposite to that of polymerization. Even in that case, we had shown that the force-velocity curve can
change its shape depending on the relative time scale between the filament and surface dynamics [21]. These recent
results, along with the current ones reported here, emphasize the importance of taking into account the independent
(thermal) fluctuations of the membrane shape and position, while studying force generation by growing filaments. In
many recent modelling approaches this aspect has been ignored and a rigid barrier with no shape fluctuations has
been considered instead [22, 29, 30].
This paper is organized as follows. In Sec. II, we describe our models. Our results for the gradient model are
presented in Sec. III and the results for the Gaussian model are described in Sec. IV and conclusions are in Sec. V.
II. DESCRIPTION OF THE MODEL
Our system consists of a set of N parallel filaments growing against an elastic membrane as shown in Fig. 1. The
membrane is modelled as a one dimensional lattice of length L and lattice constant d. At each site i of the lattice
a height hi is assigned. For a completely flat membrane, when height of all sites are the same, the elastic energy
is minimum. Presence of a local height gradient stretches the membrane and costs energy. In the absence of any
filament, the membrane undergoes equilibrium thermal fluctuations and the probability to obtain a particular height
profile {hi} follows Boltzmann measure.
We consider two different models to describe the equilibrium membrane dynamics. In the first case, the membrane
is modelled as a solid-on-solid (SOS) surface without diffusion [31] and the Hamiltonian has the form [15 -- 17]
H1 = µ
L
X
i=1
hi − hi+1
(1)
where µ is the tension. The sum in Eq. (1) is related to the total contour length C of the particular height configuration
of the membrane, such that C = PL
i=1 hi − hi+1 + Ld. Note that since the energy is linear in C, the energy cost
for increasing the contour length by an amount δ just depends on δ and is independent of C. We call Eq. (1) the
[17] to study thermally excited protrusions in lipid membranes. In
"gradient model". This model was used in Ref.
Ref.
[15], Eq. (1)
was used to describe a membrane and scaling properties of the height fluctuations were studied when the membrane
was driven by an advancing uncorrelated front representing the actin meshwork. Towards the end of the paper, we
present a discussion on this.
[16], the critical behavior of this model was analysed using functional renormalization. In Ref.
In our second model, the membrane is described as a Gaussian surface with the Hamiltonian [16, 20]
H2 = ν
L
X
i=1
(hi − hi+1)2
3
(2)
where ν denotes the elastic constant. In this case, energy depends on the square of the local height gradients, and
hence the energy cost to extend the membrane by a certain amount also depends on the contour length C. For large C
values, the energy cost also goes up, as expected for an elastic body. We refer this model as 'Gaussian model' below.
To keep our description simple, we have neglected the bending energy of the membrane and have only considered its
elastic energy here. We assume periodic boundary condition on the membrane, hL+1 = h1.
The membrane can undergo independent thermal fluctuations in its local height [see Figs. 1(a) and 1(b)] and tends
to minimize its elastic energy. In our lattice model, we assume that as a result of these fluctuations, the local height
can increase or decrease by a discrete amount δ. The rates of the dynamical moves that changes the energy by an
amount △E are assumed to satisfy local detailed balance
R+
R−
= e−β∆E
(3)
where R+ (R−) is the rate of those processes that increases (decreases) the energy by an amount ∆E [Fig. 1a]. For
∆E = 0, the rate is taken to be unity [Fig. 1b].
The filaments are modelled as rigid polymers, composed of few rodlike monomers of length d. A (de)polymerization
event increases (decreases) the length of a filament by an amount d. Throughout this work, we consider δ, the unit
of barrier height fluctuation, and the monomer size d to be the same. There are two types of filaments we need to
consider: free filaments, which are not in touch with the membrane, and bound filaments, whose tip is in contact
with a membrane site. The point of contact is called the binding site. For a free filament, the polymerization process
happens with a rate U0 and depolymerization happens with a rate W0 [see Fig. 1(d)]. However, for a bound filament,
a polymerization process increases the height of the binding site by an amount d [see Fig. 1(c)] and hence an energy
cost ∆E is involved. Note that the change in energy can be positive or negative, or even zero, depending on the local
height configuration around the binding site. For a positive (negative) energy cost, the bound filament polymerization
rate is taken to be U0R+ (U0R−), while for zero energy cost, the rate is simply U0. The depolymerization rate of
bound filament does not involve any membrane movement and hence is equal to W0.
The elastic interaction of the membrane tends to keep the membrane flat and the bound filaments generally grow
by causing protrusions in the membrane. Although in principle, it is possible that the local height configuration
around the binding site is such that the polymerization of bound filament actually releases some elastic energy, such
configurations are rare and most of the time elastic force acts against polymerization force. The elasticity-velocity
curve therefore plays a similar role as the force-velocity curve measured in many earlier studies [21 -- 28].
Note that the height of the membrane at the binding sites should be such that the membrane always stays above
the filament tips, and this puts some restrictions in height fluctuations at the binding sites. Any height fluctuation
that brings a binding site at a lower height than the filament tip is forbidden. Everywhere else in the membrane, the
height fluctuations will occur in accordance with Eq. (3).
We perform simulations using kinetic Monte Carlo technique. The relative time scale between the filament dynamics
and the membrane dynamics is quantified by a parameter S. For a system consisting of N filaments and L membrane
sites, each Monte Carlo time-step consists of N filament updates and S independent membrane updates.
In case
of a bound filament polymerization, the membrane height is also simultaneously updated at the binding site. More
specifically, for N < S, we first choose a filament at random and perform polymerization or depolymerization move as
described above. Then we choose S/N membrane sites in random sequential order and update them (for noninteger
values of S/N , we replace them by the nearest integer). Repeating this process N times completes one Monte Carlo
step. Similarly, for N > S, we first perform N/S filament updates and then choose one membrane site at random and
update it; repeat this process S times and that defines one Monte Carlo step. Starting from an initially flat membrane
and all filaments of length d (i.e., each filament consists of only one monomer), the system undergoes time-evolution
and after a large number of Monte Carlo steps, we perform our measurements. The value of parameters U0 and W0
are taken from Refs. [1, 32] and the value of d is taken from Refs. [1, 22]. All these simulation parameters are listed
in Table I
III. RESULTS FOR GRADIENT MODEL
In the gradient model, since the elastic energy of the membrane is proportional to its contour length, the change
in local height by an amount d that causes C to change by 2d, brings about a change 2µd in the energy. In our
U0 Free filament polymerization rate 2.784s−1
1.4s−1
W0 Filament depolymerization rate
d
2.7nm
300K
T
Size of an actin monomer
Temperature
TABLE I: Parameters used in our simulation.
(a)
R+
R-
(c)
U0 R+
4
(b)
1
1
(d)
U0
W0
FIG. 1: Schematic representation of our model. The square blocks show actin monomers, which join together to form rod-like
filaments. The thick solid line represents the shape of the elastic membrane. (a): A bulk site of the membrane thermally
fluctuates and changes its height by an amount d, which in turn changes the membrane contour length by 2d. The forward
process increases the energy and occur with rate R+ while the reverse process decreases the energy and happens with rate R−.
(b): A bulk site of the membrane changes its height by an amount d but the energy remains same and thus the movement
happens with rate unity. (c): A bound filament pushes the binding site by an amount d that costs energy and occurs with
rate U0R+. (d): A free filament polymerizes (depolymerizes) with rate U0 (W0).
simulation, we choose R+ = e−βµd and R− = eβµd. The bound filament polymerization that leads to an increase
(decrease) in energy happens with rate U0R+ (U0R−). All other movements of the membrane where energy does not
change, occur with rate unity. We first present the results for a single filament and S/L = 1.
A. Peak in the elasticity-velocity curve for single filament
The polymerization of the bound filament pushes the membrane upward and gives rise to a nonzero membrane
velocity, which is measured as the rate of change of the average membrane height in the long time limit. We measure
the velocity V as a function of µ and present our data in Fig. 2(a), for N = 1 and S/L = 1. We find that V shows
a nonmonotonic dependence on µ: Starting from a nonzero value for µ = 0, it increases with µ for small µ values,
and after reaching a peak at a certain µ∗, the velocity decreases again. Presence of a peak is somewhat surprising,
since increasing µ ought to make it more difficult for the filament to push against the membrane. We also find that
V scales as 1/L. In the remaining part of this sub-section, we explain different aspects of this data in detail.
First we consider µ = 0. In this case, there is no energy cost involved in stretching the membrane, and thus the single
)
c
e
s
/
m
n
(
L
V
5.5
5
4.5
4
V0 L
3.5
3
2.5
2
1.5
1
0.5
0
0
(a)
0.1
)
c
e
s
/
m
n
(
V
0.01
0.001
2
3
4
5
6
7
8
µ (pN)
L=64
L=128
L=256
1
*
µ
2
3
4
5
6
7
8
µ (pN)
)
c
e
s
/
m
n
(
V
4
3
2
1
0
0
5
(b)
Vbind
Vbulk
1
*
µ
2
µ (pN)
3
4
FIG. 2: µ − V curve for a single filament. (a): The average velocity of the membrane as a function of µ shows a peak at µ∗.
While V ∼ 1/L, the peak-position µ∗ does not depend on L. The solid triangle on the y-axis marks V0 = d
L (U0 − W0), which
is the expected value of V at µ = 0. Inset: For large µ, the membrane velocity decreases exponentially with a decay constant
≃ 0.67, which is close to the value of βd. Here, we have used L = 64. (b): The variation of average velocity of the binding
site (Vbind) and the average velocity of a bulk site (Vbulk) with µ, shows that Vbind decreases monotonically with µ while Vbulk
increases with µ for µ < µ∗. For µ > µ∗, these two velocities are equal. Here we have used L = 16. For all the above plots, we
have used S/L = 1. Other simulation parameters are as in Table I.
filament present in the system polymerizes with rate U0, irrespective of whether it is free or bound to the membrane.
The velocity of the barrier in this case turns out to be V0 = d(U0 − W0)/L. We briefly present the calculation here.
Let p0 be the probability that the filament is in the bound state. The height at the binding site can (a) increase due to
bound filament polymerization that happens with an effective rate U0p0, (b) increase due to thermal fluctuation that
happens with rate 1, or (c) decrease due to thermal fluctuation, provided the filament is not in a bound state (since
the membrane always needs to stay above the filament tip) and this process happens with effective rate (1 − p0). It
p0(U0 + 1), where
follows from here that the average velocity of the membrane can be written as V0 = V (µ = 0) =
the pre-factor d/L is the change in average height of the membrane due to d-unit change in the binding site height.
Here, we have assumed S/L = 1. The contact probability p0 can be calculated by noting that the height difference
between the filament tip and the binding site performs a biased random walk, with the restriction that it cannot cross
d
L
U0 − W0
1 + U0
the origin and become negative [21]. Our simple calculation presented in Appendix A yields p0 =
, which
gives the required expression for V0. Comparison with our simulation data in Fig. 2(a) show good agreement.
Next, we explain the presence of the peak in the µ − V curve. Note that for µ = 0 the system has no steady state
and the filament pushes the binding site upward, while other (L − 1) sites of the system which are not coupled (µ
being zero), undergo equilibrium fluctuations and show no net velocity. Consequently, the height difference between
the binding site and the bulk sites and hence the contour length C keeps increasing with time. For nonzero µ there is
an energy cost associated with stretching the membrane. This elastic force tries to reduce C and the bulk sites feel
an upward pull towards the binding site. The strength of this pull increases as µ increases and this explains why the
membrane velocity increases with µ. However, for small values of µ, this elastic force is not strong enough to counter
the polymerization force exerted by the filaments, and the bulk sites are not able to catch up with the binding site,
which still moves at a larger velocity and C keeps increasing with time. Finally, when µ reaches a critical value µ∗,
such that the elastic force exactly balances the polymerization force, the average velocity of the bulk sites becomes
equal to that of the binding site. As µ is increased further, the elastic force becomes stronger than the polymerization
force and it becomes increasingly difficult for the filament to push the binding site. However, as soon as a successful
polymerization takes place, and the binding site height increases, the bulk sites quickly catch up because the elastic
energy associated with a nonflat profile is high for µ values in this range. The whole membrane now moves with the
same velocity and C stabilizes. The membrane velocity in this case is dominated by the polymerization events at the
binding site, the rate of which is U0 exp(−βµd). Thus the membrane velocity decreases exponentially with µ in this
range [see data in Fig. 2(a), inset].
To verify the above mechanism, we measure the velocity of the binding site and the bulk sites separately and plot
the data in Fig. 2(b). As argued above, we find that for small µ, the binding site velocity is higher than the bulk site
and these two velocities become equal for µ ≥ µ∗. The difference between the binding site and the bulk site velocities
can be alternatively measured by <C(t)−C(0)>
, for large t, and our plot in Fig. 3b shows that for µ < µ∗ this quantity
t
6
decreases linearly with µ and becomes zero at µ∗. We also measure the ratio λ =< (hb+1 − hb+ hb − hb−1+ 2d)/C >,
averaged over different configurations. Here, b denotes the binding site. λ gives the fraction of contour length that is
contained between the binding site and its two neighbors. For µ < µ∗, when the binding site moves faster than the
rest of the membrane, for large times, this fraction is very close to 1, since the length of the rest of the membrane
becomes negligible compared to the growing separation between the binding site and its neighbors. Our data in Fig.
3(a) show that λ stays at 1 for small µ and then shows a sharp fall at µ = µ∗. For µ ≫ µ∗, when the polymerization is
almost impossible and the membrane velocity is close to zero, the membrane is flat and λ then becomes 2/L. In this
limit, the membrane behaves like a rigid barrier, which was studied in Refs. [22 -- 24, 29, 30, 33]. We find that after
the sharp jump at µ = µ∗, the ratio of the contour lengths decreases exponentially with µ to the asymptotic value
2/L [Fig. 3(a), inset]. Note that both for small and large µ, the binding site is the only site that is being pushed by
the filament and the dynamics of bulk sites still follow local detailed balance. The overall velocity of the membrane is
thus generated by the drive present at the single binding site and this is the reason V scales inversely with the system
size L.
(a)
2
-
L
λ
10
1
0.1
0.01
2
4
6
8
µ (pN)
t
/
>
)
0
(
C
-
)
t
(
C
<
8
7
6
5
4
3
2
1
0
L=64
L=128
(b)
)
t
(
C
105
104
103
102
x0.5
10-1
101
105
107
103
t
1
∗
µ
2
µ (pN)
3
1
0.9
0.8
0.7
0.6
λ
0.5
0.4
0.3
0.2
0.1
0
0
L=64
L=128
1
∗
µ
2
µ (pN)
3
4
5
0
FIG. 3: The membrane contour length keeps growing for µ < µ∗ and stabilizes for µ > µ∗. (a): The ratio λ = 1 for
µ < µ∗ and decreases sharply at µ∗ to its asymptotic value 2/L for large µ. Inset: The saturation of λ at large µ happens
exponentially. (b): The quantity <C(t)−C(0)>
, for large t, decreases with µ and becomes zero above µ∗. Inset: Variation of
contour length with time at µ∗ = 1.748pN . We see that after large enough time, C(t) grows with time as t0.5 with a diffusion
constant D = 37.41 ± 7.19nm2/s. The analytically calculated value of D = 41.352nm2/s which is close to the numerical value.
Here we use L = 64. For both panels S/L = 1. Other simulation parameters are as in Table I.
t
B. Absence of steady state for µ < µ∗
For µ < µ∗, when the binding site and bulk site velocities are different, and C keeps increasing with time, the system
does not have a steady state. This seems somewhat surprising and in this subsection we discuss this issue in detail.
Let us define z = 2hb − hb+1 − hb−1. Our data in Appendix B [Fig. B-1(a)] show that for µ < µ∗, the sign of z is
always positive and even for µ > µ∗, the probability to find negative z is negligible. This is expected, since for small
µ, the binding site always stays above its neighbors, and for large µ, when the membrane is flat, the binding site is
most of the time at the same level with its neighbors, but does not typically fall below that level. We show below
that z performs a biased random walker with a reflecting boundary condition at the origin.
First note that z can increase either due to increase of hb (from bound filament polymerization or thermal fluctuation
at the binding site) which happens with rate e−βµd(p0U0 + 1), or due to decrease of hb±1, which happens with rate
1. This is because our data in Appendix B (Fig. B-1b) show that the sites (b ± 1) almost always have one neighbor
(site b ± 2) at a lower height, and another neighbor (site b) at a higher height. From Eq. (1) it then follows that the
height at the sites (b ± 1) can increase or decrease without any energy cost and rate of such processes is taken to be
1 [see, for example, Fig. 1(b)]. The value of z can decrease because of decrease in hb which releases elastic energy
and happens with rate eβµd(1 − p0), or because of increase of hb±1, which again happens with rate unity, as explained
above. Let us define P (z, t) as the probability for z at time t. Then the change in P (z, t) in a small time △t can be
written as,
P (z, t + △t) − P (z, t) = △t[(1 − p0)eβµdP (z + 2d, t) + 2P (z + d, t) + 2P (z − d, t) + (1 + U0p0)e−βµdP (z − 2d, t)
−{(1 − p0)eβµd + 4 + (1 + U0p0)e−βµd}P (z, t)] (4)
7
Here, we have considered z > 0 (supported by data in Fig. B-1a) and put a reflecting boundary at the origin
z = 0, such that z can never be negative. In the continuum limit, Eq. (4) becomes a Fokker-Planck equation for a
biased random walker with drift and diffusion given by, respectively, v = 2d{(1 + U0p0)e−βµd − (1 − p0)eβµd} and
D = 2d2{(1 − p0)eβµd + (1 + U0p0)e−βµd + 1}.
It can be easily seen that for small µ, the bias v is positive while for large µ, the bias v is negative. At µ = µ∗,
v becomes zero and the system shows unbiased diffusion with diffusivity D. The value of µ∗ can be obtained by
equating v to zero and using the expression for p0 from Appendix A which gives the resulting equation
W0e3βµ∗d − U0e2βµ∗d − (U 2
0 − U0W0 + U0)eβµ∗d + U0 = 0.
(5)
Numerical solution of the above equation gives only one physical solution for µ∗. We find µ∗ = 1.81pN , which is close
to the value 1.748 ± 0.004pN observed in simulations.
Due to the reflecting boundary condition at the origin, this implies that for µ < µ∗, the system does not reach
a steady state, and hzi increases linearly with time, while for µ > µ∗, there exists a steady state and hzi reaches a
time-independent value. At µ = µ∗, we have an unbiased random walker with reflecting boundary at the origin, for
which there is no steady state either, but hzi in this case grows diffusively with time with a diffusion constant that
matches our analytical expression given above [see Fig. 3(b), inset].
C. Faster membrane dynamics lowers µ∗
So far we have considered the case for S = L membrane updates in one MC step. What happens when the membrane
dynamics is faster or slower than this? Apart from the binding site, all the other (L − 1) bulk sites are being driven
by only the elastic interaction. For large S. when the bulk sites are updated at a higher rate, the effect of the elastic
interaction is felt more strongly. As a result, the point of balance µ∗ where the elastic force and polymerization force
become equal, now shifts towards a smaller value of µ. More specifically, for small but nonzero µ, when the bulk sites
experience an upward bias towards the binding site, their total displacement in the upward direction per MC step
increases, as S increases. Thus the bulk sites are able to catch up with the binding site at a smaller value of µ. In
other words, as S increases, µ∗ decreases. In Fig. 4(a) we plot µ − V data for different S/L values, and find that
the peak of the curve shifts towards left as S/L increases. For S/L >> 1, the value of µ∗ becomes infinitesimally
small and for any finite µ, the curve looks monotonic. In the inset of Fig. 4(a) we show the variation of µ∗ with S/L.
Note, however, that the dependence is not very strong and µ∗ deceases logarithmically slowly with S/L. We see that
over two decades of change of S, the value of µ∗ changes only by a factor of half. However, the membrane velocity
measured at µ∗ shows a more strong dependence on S/L and grows as a power law with an exponent ≃ 0.87 [Fig.
4(b)]. We can generalize our calculation in the last subsection for arbitrary values of S and obtain µ∗ as the only one
physical solution of the equation
S
L (cid:16)W0e3βµ∗d − U0e2βµ∗d − U0eβµ∗d + U0(cid:17) − (U 2
0 − U0W0)eβµ∗d = 0.
(6)
The analytical result shown by the continuous line in the inset of Fig. 4(a) shows good agreement with numerical
data over a wide range of S/L.
Note that our calculation for the contact probability p0 presented in Appendix A shows that for small µ and large
S/L, the contact probability is very small. This means that the filament is unbound most of the time and grows
like a free filament with a net growth rate d(U0 − W0), which is independent of µ. Since we measure the membrane
velocity as the average displacement of the membrane per Monte Carlo step and we define our Monte Carlo step such
that there are S surface updates and one filament update, our data in Fig. 5(a) show that the membrane velocity
also approaches this limit as S/L becomes large. The µ − V curve in this case therefore becomes flat for small µ and
then decreases for large µ, giving rise to a concave curve, as shown in Fig. 5(a). Thus the µ − V curve can change its
shape from convex to concave depending on the choice of the relative time scale S/L [21, 23, 34].
In the limit of very large S/L, when the membrane fluctuations occur much more rapidly than the filament
polymerization, the membrane reaches a thermal equilibrium between two filament movements. The partition function
for the system can be calculated in this case and the average contour length hCi of the membrane can be obtained
from there, which has the form hCi = Ld/(1 − e−βµd). In Fig. 5(b) we compare it with numerics and find reasonably
good agreement for large S/L.
x
a
m
V
V
/
1
0.8
0.6
0.4
0.2
0
0
(a)
*
µ
6
4
2
x0.87
1
(b)
8
10-2
10-1
100
101
S/L
)
∗
µ
(
V
0.1
0.01
0.001
2
4
6
8
10
µ (pN)
0.1
1
S/L
10
FIG. 4: Peak position and peak height of µ − V curve depends on S/L. (a): Scaled V as a function of µ for different S/L
values. The scaling factor Vmax has been used such that the data for different S/L values can be compared. In this plot the
symbol ◦ correspond to S/L = 2−2 (Vmax ≃ 5.84 × 10−2nm/s), the ✷ correspond to S/L = 2−1 (Vmax ≃ 5.84 × 10−2nm/s),
the △ correspond to S/L = 1 (Vmax ≃ 7.89 × 10−2nm/s) and the ∇ correspond to S/L = 26 (Vmax ≃ 2.23nm/s). For smaller
values of S/L, the curve is monotonic as the velocity is mostly determined by the binding site. As S/L increases, the µ − V
curve develops a peak at µ∗. Inset: µ∗ decreases with S/L. The discrete points from simulations match well with continuous
line from Eq. (6). We use L = 64 here. (b): The membrane velocity at µ∗ plotted against S/L shows a power law increase.
The solid line represents a function ∼ x0.87 and goes parallel to our numerical data points. Here we have used L = 64. Other
simulation parameters are as in Table I.
x
a
m
V
V
/
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
(a)
S/L=2-3
S/L=22
S/L=29
S/L=212
2
4
6
8
10
µ (pN)
/
d
L
>
C
<
5
4.5
4
3.5
3
2.5
2
1.5
1
0
(b)
S/L=23
S/L=26
S/L=29
S/L=210
Analytical
0.5
1
1.5
2
2.5
3
3.5
4
µ (pN)
FIG. 5: Variation of membrane velocity and contour length against µ for different values of S/L. (a): The µ−V curve changes
from convex to concave as S/L is increased. For S/L << 1, the curve is convex while for S/L >> 1, it becomes concave.
Here, we have scaled V by Vmax such that we can compare them in the same scale. The values of Vmax are 4.67 × 10−1nm/s,
4.67 × 10−1, nm/s, 3.73nm/s, and 3.73nm/s, respectively, in the increasing order of S/L. Here, we have used L = 8. (b):
Average contour length of the membrane scaled by Ld as a function of µ for different values of S/L. The continuous line is from
analytical calculation in the equilibrium limit, which matches with simulation for very high S/L. Other simulation parameters
are as in Table I.
D. For large µ the membrane behaves as a rigid obstacle
When µ is very large, the membrane remains flat most of the time. Whenever there is a filament polymerization,
the height of the binding site increases, but due to high elastic energy cost associated with such a configuration, the
bulk sites quickly catch up and the membrane is flat again. We find that the membrane behaves like a perfectly rigid
barrier in this case. Let pi be the probability to find a gap of length i between the filament tip and the binding site.
For a rigid barrier, this gap can only increase or decrease due to filament polymerization or depolymerization. For
i > 0 the master equation for a rigid barrier is
dpi
dt
= U0pi+1 + W0pi−1 − (U0 + W0)pi
(7)
and for i = 0,
dp0
dt
= U0p1 − W0p0.
9
(8)
In the steady state, we have pi = (W0/U0)ip0 and along with normalization condition Pi pi = 1, this gives the contact
probability p0 = 1 − W0/U0 ≃ 1/2 (see Table I). In Fig. 6(a), we show the variation of p0 with µ for different S/L
and find that for large µ, the contact probability indeed saturates to 1/2, the rigid barrier limit. Smaller the value
of S/L, faster is the saturation. For small µ, the contact probability can be calculated analytically (see Appendix
A) and in Fig. 6(a) this has been shown by solid lines, which agree well with the simulations. Our calculation
remains valid only for µ < µ∗ and hence the comparison has been done only in this regime. µ∗ becomes too small
for very large S/L, and hence not marked in this plot. In the case for S/L << 1, when the membrane dynamics
is very slow, the thermal fluctuations of the membrane can be neglected and the contact probability is essentially
controlled by the polymerization and depolymerization of the filament. Thus we find that the contact probability
approaches 1/2 even for µ < µ∗ in this case [shown by red triangles in Fig. 6(a)]. However, the system is not in
steady state here and the contour length of the membrane keeps increasing with time. In Fig. 6(b) we plot the ratio
λ =< (hb+1 − hb + hb − hb−1 + 2d)/C > and show this explicitly.
0
p
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
(a)
S/L=2-6
S/L=1
S/L=26
S/L=29
S/L=212
1
2
3
4
5
6
7
8
9
λ
1
0.8
0.6
0.4
0.2
0
0
(b)
S/L=2-6
S/L=2-3
S/L=1
1
2
3
4
5
6
7
8
µ (pN)
µ (pN)
FIG. 6: For µ >> 1, the membrane behaves as a rigid obstacle. (a): The contact probability p0 for different values of S/L.
The continuous lines are from the analytical predictions. For large µ, the value of p0 saturates to 1/2, which is expected for a
rigid barrier. As S/L decreases, the rigid barrier behavior sets in for smaller µ values. The continuous lines show analytical
calculation in µ < µ∗ regime. (b): The variation of λ for different values of S/L. For small µ, the value of λ is unity, which
means that the contour length of the membrane diverges with time. For high enough value of µ, λ saturates to 2/L. We use
L = 64 for both the plots. Other simulation parameters are as in Table I.
E. Results for multiple filaments
So far in this section, we have considered the case of single filament. We end this section with a brief discussion on
multiple filaments. Let us define the filament density as ρ = N/L. We find for ρ ≪ 1, and for uniform distribution
for the binding sites, the qualitative behavior is same as that for single filament. In Fig. 7(a), we show the µ − V
curve for multiple filaments for small ρ. Since the membrane velocity results from the polymerization force exerted
at the binding sites (motion at all other sites follows local detailed balance, as mentioned in Sec. II), our data show
that V scales as ρ. The nonmonotonicity of the µ − V curve means that even for multiple filaments, the system fails
to reach a steady state for small µ and the membrane contour length C keeps increasing with time. We have checked
that (data not shown) at µ = µ∗ the contour length grows diffusively with time and for µ > µ∗ system reaches a
steady state when C has a finite value. Figure 7(b) shows the µ − V curve for different values of S/L. We note that
for S/L << 1, the curve is convex while for S/L >> 1, it becomes concave.
As ρ increases, the elasticity of the membrane induces an effective interaction between the filaments and the single
filament picture does not remain valid any more. Our data in Fig. 8 show that V ∼ ρ scaling is lost and µ∗ now
depends on ρ. For large ρ, most of the sites are binding sites and are driven by the filaments. Due to elastic interaction,
the remaining few bulk sites feel an upward pull and are able to catch up with the binding sites at smaller µ values.
Thus µ∗ decreases with ρ for large ρ. When ρ = 1, the system will reach steady state for all values of µ but for ρ < 1,
there will always be a nonzero µ∗ below which the system does not have a steady state.
)
c
e
s
/
m
n
(
ρ
/
V
5
4
3
2
1
0
(a)
N=4; L=64
N=2; L=128
N=2; L=64
0
1
∗
µ
2
3
4
5
6
7
8
µ (pN)
10
(b)
S/L=2-2
S/L=25
S/L=27
S/L=210
1
0.8
0.6
0.4
0.2
x
a
m
V
V
/
0
0
2
4
6
µ (pN)
8
10
FIG. 7: µ − V curve for multiple filaments with ρ ≪ 1. (a): The velocity scales with ρ similar to the single filament case. It
also shows a peak at µ = µ∗ which is independent of ρ. We have used S/L = 1 here. (b): Scaled velocity as a function of
µ for different time scales which show that the curve changes from convex to concave similar to the single filament case. For
S/L << 1, the curve is convex while for S/L >> 1, the curve becomes concave. The values of Vmax are 0.467nm/s, 3.71nm/s,
3.74nm/s and 3.75nm/s for S/L = 2−2, 25, 27, and 210, respectively. Here we have used N = 4 and L = 32. Other simulation
parameters are as in Table I.
ρ=1/2
ρ=1/3
ρ=1/4
2
1.5
1
∗
µ
0.5
0.001
0.01
ρ
0.1
0.5
5.5
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
)
c
e
s
/
m
n
(
ρ
/
V
0
0
2
4
6
8
10
µ (pN)
FIG. 8: Results for multiple filaments with high filament density. In the main plot, we show µ − V curve for multiple filaments
for ρ ∼ 1. We note that V does not scale with ρ. The peak position µ∗ decreases with ρ. The inset shows the variation of
µ∗ with ρ. We note that for ρ << 1, the value of µ∗ remains constant and then decreases with ρ for high value of ρ. We use
S/L = 1 here. Other simulation parameters are as in Table I.
In the above discussion, we have assumed that the binding sites are distributed uniformly throughout the membrane.
If we consider an inhomogeneous distribution of binding sites, then for small µ the region of the membrane, where the
density of binding sites is high, will have a large local velocity since it is being driven by a large number of filaments.
On the other hand, the part of the membrane where binding sites have a low density, will have a much smaller
local velocity. In order for the whole system to reach a steady state, the velocity should be same everywhere on the
membrane. This means that the polymerization force present in the fastest part of the membrane has to be smaller
than the elastic interaction which pushes the slowest part of the membrane where binding site density is lowest. A
very large value of µ is required to achieve this balance. Thus µ∗ becomes very large in the case of inhomogeneous
distribution of binding sites. We have verified this in our simulations (data not shown here).
11
IV. RESULTS FOR THE GAUSSIAN MODEL
In the Gaussian model, the membrane Hamiltonian is given by Eq. (2). In this case, if the height of the j-th site,
hj changes to hj ± d, then the change in energy, ∆E = 2dν[d ± (2hj − hj+1 − hj−1)], depends on the local height
configuration around the jth site. This is different from the gradient model, where ∆E is constant. Following the
local detailed balance [see Eq. (3)] we use in our simulations R+ = e−β∆E, and R− = 1, while all other movements
of the membrane where the energy of the barrier does not change, occur with rate unity. First we present our results
for the single filament and then we discuss the case of multiple filaments.
A. Membrane velocity decreases with ν
For ν = 0, there is no difference between gradient model and Gaussian model [Eqs. (1) and (2)]. Therefore, as
derived in Sec. III A, the membrane velocity at ν = 0 is given by V0 = d(U0 − W0)/L. The membrane contour length
C keeps increasing with time and the system does not have a steady state. However, as soon as ν is nonzero, any
polymerization event or independent thermal fluctuation that causes an increase in C, has an energy cost which is
higher as C gets larger. Thus the membrane can not stretch indefinitely and for all ν > 0 the system reaches a steady
state. Our data in Fig. 9(a) show that for ν > 0, the membrane velocity V decreases as ν increases. For large ν, the
decay is exponential and we also find V ∼ 1/L scaling, as seen earlier in Fig. 2 for the gradient model.
Note that there is a discontinuity of the ν − V curve at ν = 0 [see Fig. 9(a)]. The data in Fig. 9(a) bottom inset
show that as ν → 0, the velocity saturates to a value, which is different from the value V0 = d(U0 − W0)/L at ν = 0.
This is because at ν = 0, only the binding site of the membrane has nonzero velocity and the bulk sites have zero
velocity. But as soon as ν 6= 0, the system has a steady state and the bulk sites must move with the same velocity as
the binding site. This sharp jump in the bulk sites velocity causes the discontinuity in the ν − V curve.
4.5
4
V0 L
3.5
3
2.5
2
1.5
1
)
c
e
s
/
m
n
(
L
V
0.5
0
(a)
L=64
L=128
L=256
)
c
e
s
/
m
n
(
V
)
c
e
s
/
m
n
(
V
10-2
10-4
0.07
0.06
0.05
1
3
2
4
ν (pN/nm)
10-5
10-3
10-1
ν (pN/nm)
0.2
0.4
0.6
0.8
1
1.2
1.4
d
/
>
C
<
800
600
400
200
64
0
(b)
L
-
d
/
>
C
<
101
10-1
10-3
103
102
L
-
d
/
>
C
<
1
2
3
4
ν (pN/nm)
5
101
10-2
10-1
ν (pN/nm)
100
0.2
0.4
0.6
0.8
1
ν (pN/nm)
ν (pN/nm)
FIG. 9: Variation of velocity and contour length with ν for Gaussian model. (a): The ν − V curve for single filament shows
1/L scaling. The curve is monotonic for all ν 6= 0. Top inset: The velocity falls exponentially with decay constant ≃ 3.13,
which is close to 2βd2. Bottom inset: In the limit of very small ν, the velocity saturates to a value ∼ 0.074nm/sec, that is
distinctly different from V0 = 0.058nm/s at ν = 0. We use L = 64 for the insets and S/L = 1 for all the plots. (b): Average
contour length of the membrane scaled by d as a function of ν. For very high value of ν, < C > becomes equal to the system
size Ld. Top inset: Contour length decreases exponentially for large µ with a decay constant ≃ 3.2 which is again close to
2βd2. Bottom inset: For very small ν, contour length falls off with ν as a power law with an exponent ≃ 0.9. For all the above
plots, we use L = 64 and S/L = 1. Other simulation parameters are as in Table I.
In Fig. 9(b) we plot the steady state average contour length hCi as a function of ν. We find that (top inset) hCi
decreases exponentially for large ν, as found in the gradient model (also see Fig. 3). For very small ν, our data [Fig.
9(b) bottom inset] show that hCi decreases as a power law with an exponent ≃ 0.9. The power law decay can be
explained as follows.
For very small ν, the local height gradient of the membrane decreases sharply as a function of the distance from the
binding site and we have yb > yb+1 > yb+2..., where yi = hi − hi+1 and b denotes the binding site. The bound filament
polymerization and thermal height fluctuations at the binding site may change hb which actually changes C. It is easy
to see that the rate of these processes that increase hb is (1 + U0p0)e−2βνd2(2yb+1) and that decrease hb is (1 − p0).
The height fluctuations of the other sites will not change C. In the steady state, C is time-stationary, the increasing
12
and decreasing rates must be equal, from which it follows that yb ∼ 1/ν. Moreover, in the steady state, the velocity
of all membrane sites are equal. The velocity at the binding site is given by {(1 + U0p0)e−2βνd2(2yb+1) − (1 − p0)},
and that at the neighboring site (b + 1) is {1 − e−2βνd2(yb−yb+1+1)}. Once we equate them, it follows that yb+1 ∼ 1/ν
for small ν. Equating the velocity of the sites (b + 1) and (b + 2), it can be similarly shown that yb+2 ∼ 1/ν, and thus
for any site i, yi ∼ 1/ν holds. Thus the quantity (< C > −Ld = Pi yi) shows an 1/ν dependence. Our numerical
data yields power law exponent 0.9 which is close to this prediction.
B. Faster membrane dynamics yields a concave ν − V curve
The relative time-scale between the membrane and filament dynamics is an important parameter for the Gaussian
model also. The contact probability p0 that is crucial to determine the interaction between the filament and the
membrane depends strongly on whether the membrane dynamics is faster or slower than the filament dynamics. For
very large ν, it is expected that p0 ≃ 1/2 since the membrane behaves like a rigid barrier in this limit (also see our
discussion in Sec.
III D). However, when ν is small, our data in Fig. 10(a) show that the variation of p0 with ν
is qualitatively determined by the value of S/L. For S/L ≫ 1, the thermal fluctuations of the membrane happen
so fast that the contact with the filament tip is lost most of the time and p0 is small. From this small value, p0
increases to 1/2 as ν becomes large. For S/L ≪ 1, the membrane fluctuations become almost negligible and the
contact probability is controlled by the (de)polymerization of the (bound) free filament which does not depend on ν.
Our data for small S/L support this.
0
p
0.6
0.4
0.2
0
0
(a)
S/L=2-6
S/L=1
S/L=26
S/L=212
1
2
3
4
5
x
a
m
V
V
/
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
(b)
S/L=1
S/L=23
S/L=29
S/L=212
0.5
1
1.5
2
2.5
3
3.5
4
ν (pN/nm)
ν (pN/nm)
FIG. 10: Contact probability and velocity as a function of ν in a Gaussian model for different values of S/L. (a): For S/L << 1,
p0 does not show much variation and its value remains close to 1/2. For S/L ∼ 1, in the small ν region, the value of p0 is
higher than 1/2 which then saturates to its rigid barrier limit for high ν. For S/L >> 1, p0 starts with very small value and
then increases with ν, then saturates to 1/2. Higher the value of S/L, slower the tendency to reach 1/2. We use L = 8 here.
(b): The ν − V curve changes from convex to concave as S/L is increased. The values of Vmax are 0.475nm/s, 3.50nm/s,
3.72nm/s, and 3.73nm/s for S/L = 1, 23, 29 and 212 respectively. Other simulation parameters are as in Table I.
The behavior of p0 has a direct influence on the ν − V curve and we show in Fig. 10(b) that the curve becomes
concave from convex as S/L is increased. When S/L ≫ 1, the filament is most of the time unbound for small ν and
hence the membrane velocity remains constant at d(U0 − W0)/L, same as that for a free filament. For large ν, when
p0 starts increasing again, V decreases. The ν − V curve is concave in this case. On the other hand, for very small
S/L, the value of p0 is no longer negligible and it remains nearly 0.5 throughout the region. In this case, due to the
increase in ν, the velocity starts decreasing even in the small ν range, which yields a convex curve.
As done in the gradient model, in the limit of very fast membrane dynamics, we can neglect the filament dynamics
and assuming equilibrium we can calculate the average elastic energy of the membrane which shows good agreement
with our numerics (data not shown) for very large S/L.
The multiple filaments case is qualitatively the same as the single filament case. For low density of filaments
(N/L = ρ ≪ 1), we plot the ν − V curve in Fig. 11. We find that V decreases monotonically with ν for nonzero ν
C. Multiple filaments
and scales as ρ. The decay of V is exponential for large ν with a decay constant which is same as that in the single
filament case. In the limit of large ν, the membrane behaves like a rigid barrier and apart from an overall scaling of
V by a factor ρ, its dependence on ν remains same for single or multiple filaments.
13
N=4; L=64
N=8; L=128
N=8; L=256
10-2
10-4
10-6
)
c
e
s
/
m
n
(
V
4.5
4
3.5
3
2.5
2
1.5
1
0.5
)
c
e
s
/
m
n
(
ρ
/
V
2
3
4
ν (pN/nm)
5
6
0
0
0.5
1
1.5
2
ν (pN/nm)
2.5
3
3.5
4
FIG. 11: Velocity of the barrier as a function of ν for multiple filaments. There is no qualitative difference between the results
of a single filament and the multiple filaments. For large value of ν, the velocity falls exponentially with a decay constant
≃ 3.18, which is close to the single filament case. For the inset we use N = 8; L = 128 and for all the plots, we use S/L = 1.
Other simulation parameters are as in Table I.
One interesting observation that can be made for the multiple filaments case is the merging of membrane protrusions.
In earlier studies [14] of a more detailed modeling of cell membrane, it was shown that actin filament polymerization
gives rise to filopodial protrusions. The protrusion speed depends on elastic properties of the membrane and the
membrane distortion also induces an effective attraction between the filopodia and close by filopodia merge together
to form a larger (wider) protrusions. We study merging of membrane protrusion within our simple model by monitoring
the contour length l of the membrane between two binding sites at a distance k apart. When the protrusions created
at the two binding sites merge with each other, l becomes equal to k. In Fig. 12, we plot (l−k)
as a function of ν for a
fixed k for the barrier with only two filaments. We find that as ν is increased, the ratio decreases and the decrease is
exponential for moderate to large ν values. In Fig. 12 inset, we show how the two protrusions merge as ν increases.
k
k=128
k=256
k=400
4000
3000
2000
1000
>
1
h
-
h
<
i
ν=0.05
ν=0.10
ν=0.50
0
0
250
750 1000
500
i
k
/
)
k
-
l
(
5
1
0.2
0.04
0
0.5
1
1.5
ν (pN/nm)
2
2.5
FIG. 12: Merging of protrusions. We plot the quantity (l−k)
, for the membrane with only two binding sites at a distance k
apart, with l being the contour length between the two binding sites. For a given k, this quantity decreases with ν exponentially
with a decay constant ≃ 1.44, which indicates merging of two protrusions for large ν. Here we use L = 1024 and S/L = 1.
Inset: Two protrusions merge as we increase ν. The average height profile of the membrane is shown for different values of ν.
Note that the height of the protrusions also become shorter as ν increases. The ν values appearing in the legends are in the
unit of pN/nm. These data are for k = 256. We use L = 1024 and S/L = 1. Other simulation parameters are as in Table I.
k
V. CONCLUSIONS
14
In this paper, we study the force generation mechanism by a set of parallel filaments growing against an elastic
membrane. The elastic membrane tends to stay flat and any distortion from this flat state costs energy.
In our
gradient model, where the energy cost is proportional to the absolute magnitude of the local height gradient, we find
that the polymerization force wins over the elastic force of the membrane for low values of the membrane tension and
the system fails to reach a steady state. This gives rise to a nonmonotonic dependence of membrane velocity on its
elasticity. In Gaussian model, where the energy cost of deformation is proportional to the square of the local height
gradient, the membrane velocity monotonically decreases with elasticity and the system always reaches a steady state
for all nonzero elasticity. Our detailed numerical simulations of various different quantities, accompanied by analytical
calculations and simple scaling arguments provide a comprehensive picture of the behavior of the system in the long
time limit.
A similar system was studied in Ref. [15] where the membrane was described by the gradient model and the growing
filaments were represented as a network mesh. The filaments always stay below the membrane and can grow only
when there is some gap between the mesh and the membrane. This is in contrast to our model where the filaments
can push against the membrane and deform it to make space for themselves to grow. This microscopic difference
actually gives rise to qualitatively different scaling behavior for the two systems. We address this important issue in
details elsewhere [35].
We also highlight the importance of the relative time scale between the dynamics of the membrane and the filaments
to determine the qualitative nature of the dependence of membrane velocity on elasticity. While for fast membrane
dynamics, the velocity is a concave function of membrane elasticity, for slow membrane dynamics a convex dependence
is observed. Both these types of dependencies are observed in real experiments. A convex force-velocity curve is
obtained for actin quoted polystyrene beads in Ref.
[25] or for magnetic colloidal particles pushed by unbranched
parallel actin filaments in Refs. [26, 27]. On the other hand, a concave force-velocity curve was reported for branched
actin network in Ref. [28]. Our simple model can yield both these characteristics for different choices of the relative
time scale mentioned above.
Throughout this paper, we have only considered elastic energy of the membrane and neglected its bending rigidity.
We have also assumed the membrane to be homogeneous with same elasticity along the whole membrane. However, in
many physical systems bending rigidity can be important and there can also be tension gradient along the membrane
[36, 37]. Our results for the simple model will pave way for studying more complex models where above mentioned
effects are included.
VI. ACKNOWLEDGEMENTS
We acknowledge useful discussions with M. Barma. SC acknowledges financial support from the Science and
Engineering Research Board, India (Grant No. EMR/2016/001663). The computational facility used in this work
was provided through the Thematic Unit of Excellence on Computational Materials Science, funded by Nanomission,
Department of Science and Technology (India).
Appendix A: Calculation of p0 for the gradient model with single filament for µ < µ∗
In this Appendix, we present an analytical calculation for the contact probability p0 in the gradient model with
µ < µ∗ with arbitrary S/L, in presence of a single filament. Let pi be the probability that there is a gap of size i
between the binding site and the filament tip. Clearly, the contact probability p0 corresponds to i = 0. The master
equation for pi can be written as
dpi
dt
= (U0 +
S
L
eβµd)pi+1 + (W0 +
S
L
e−βµd)pi−1 − (U0 +
S
L
eβµd +
S
L
e−βµd + W0)pi; f or i > 0
and
dp0
dt
= (U0 +
S
L
eβµd)p1 − (
S
L
e−βµd + W0)p0; f or i = 0
where all symbols have their usual meaning. Here, we have used the fact that for µ < µ∗, the binding site moves faster
than the rest of the system and hence has a larger height than all bulk sites. Although the system is not in steady
15
state for µ < µ∗, the probability pi still reaches a stationary value. In steady state one has the recursion relation
pi = (
)ip0; which by applying normalization condition gives the expression for the contact probability,
S
L e−βµd+W0
L eβµd+U0
S
p0 =
U0 − W0 + S
L eβµd − S
L e−βµd
S
L eβµd + U0
; f or µ < µ∗.
(A-1)
Alternatively, the binding site being driven by the growing filament, its velocity must be same as the growth velocity
of the filament. The later quantity is simply U0p0e−βµd + U0(1 − p0) − W0, while the velocity of the binding site has
the form U0p0e−βµd + S
L (1 − p0)eβµd. Equating these two gives the same expression for p0 as in Eq. (A-1).
L e−βµd − S
Appendix B: Probability that z ≥ 0 and that hb±1 ≥ hb±2
In Fig. B-1, we plot that the probabilities for z ≥ 0 and hb±1 ≥ hb±2, where z = 2hb − hb+1 − hb−1. We show that
these probabilities are very close to unity even for µ > µ∗.
)
0
≥
z
(
b
o
r
P
1
0.995
0.99
0.985
0.98
0.975
0.97
0.965
0.96
0.955
0.95
(a)
0
1
∗
µ
2
3
µ
4
5
6
)
2
±
b
h
≥
1
±
b
h
(
b
o
r
P
1
0.995
0.99
0.985
0.98
0.975
0.97
0.965
(b)
1
∗
µ
2
3
µ
4
5
6
FIG. B-1: Height gradient around the binding site. (a): The probability that z ≥ 0 as a function of µ. We see that the
probability is exactly one for µ < µ∗ and even for µ > µ∗, it is very close to one. (b): Here we show the probability that
hb±1 ≥ hb±2 as a function of µ. The red triangles are for prob(hb+1 ≥ hb+2) while the blue circles are for prob(hb−1 ≥ hb−2).
These probabilities are also close to one for entire range of µ. Here, L = 64 and S/L = 1. Other simulation parameters are as
in Table I.
[1] J. Howard, Mechanics of motor proteins and the cytoskeleton, (Sinauer Associates, Sunderland, MA, 2001).
[2] T. D. Pollard and J. A. Cooper, Actin, a central player in cell shape and movement, Science 326, 1208 (2009).
[3] L. Blanchoin, R. B. Paterski, C. Sykes and J. Plastino, Actin dynamics, architecture and mechanics in cell motility, Physiol
Rev 94, 235 (2014).
[4] P. Friedl and D. Gilmour, Collective cell migration in morphogenesis, regeneration and cancer, Nat. Rev. Mol. Cell Biol.
10, 445 (2009).
[5] J. Plastino and C. Sykes, The actin slingshot, Curr. Opin. Cell Biol. 17, 62 (2005).
[6] P. Sens and J. Plastino, Membrane tension and cytoskeleton organization in cell motility, J. Phys.: Condens. Matter 27,
273103 (2015).
[7] A. Diz-Munoz, D. A. Fletcher and O. D. Weiner, Use the force: Membrane tension as an organizer of cell shape and
motility, Trends Cell Biol. 23, 47 (2012).
[8] J. Lemi`ere, F. Valentino, C. Campillo and C. Sykes, How cellular membrane properties are affected by the actin cytoskele-
ton, Biochimie 130, 33 (2016).
[9] K. Keren, Cell motility: The integrating role of the plasma membrane, Eur. Biophys. J. 40, 1013 (2011).
[10] N. C. Gauthier, M. A. Fardin, P. Roca-Cusachs, and M. P. Sheetz, Temporary increase in plasma membrane tension
coordinates the activation of exocytosis and contraction during cell spreading, Proc. Natl. Acad. Sci. U.S.A. 108 14467
(2011).
[11] D. Raucher and M. P. Sheetz, Cell spreading and lamellipodial extension rate is regulated by membrane tension, J. Cell
Biol. 148 127 (2000).
16
[12] E. L. Batchelder, G. Hollopeter, C. Campillo, X. Mezanges, E. M. Jorgensen, P. Nassoy, P. Sens and J. Plastino, Membrane
tension regulates motility by controlling lamellipodium organization, Proc. Natl. Acad. Sci. U.S.A. 108, 11429 (2011).
[13] A. P. Liu, D. L. Richmond, L. Maibaum, S. Pronk, P. L. Geissler and D. A. Fletcher, Membrane-induced budding of actin
filaments, Nat. Phys. 4, 789 (2008).
[14] E. Atilgan, D. Wirtz and S. X. Sun, Mechanics and dynamics of actin-driven thin membrane protrusions, Biophys. J. 90,
65 (2006).
[15] S. L. Narasimhan and A. Baumgaertner, Dynamics of a driven surface, J. Chem. Phys. 133, 034702 (2010).
[16] A. Volmer, U. Seifert and R. Lipowsky, Critical behavior of interacting surfaces with tension, Eur. Phys. J. 5, 811 (1998).
[17] R. Lipowsky and S. Grotehans, Renormalization of hydration forces by collective protrusion modes, Biophys. Chem. 49
27 (1994); R. Lipowsky and S. Grotehans, Hydration vs. protrusion forces between lipid bilayers, Europhys. Lett. 23 599
(1993).
[18] S. F. Edwards and D. R. Wilkinson, The surface statistics of a granular aggregate, Proc. R. Soc. London 381, 17 (1982).
[19] M. Kardar, G. Parisi and Y-C. Zhang, Dynamic scaling of growing interfaces, Phys. Rev. Lett. 56, 889 (1986).
[20] A. Baumgaertner, Crawling of a driven adherent membrane, J. Chem. Phys. 137, 144906 (2012).
[21] R. K. Sadhu and S. Chatterjee, Actin filaments growing against a barrier with fluctuating shape, Phys. Rev. E 93, 062414
(2016).
[22] D. K. Hansda, S. Sen and R. Padinhateeri, Branching influences force-velocity curve and length fluctuations in actin
networks, Phys. Rev. E 90, 062718 (2014).
[23] R. Wang and A. E. Carlsson, Load sharing in the growth of bundled biopolymers, New J. Phys 16, 113047 (2014).
[24] C. S. Peskin, G. M. Odell and G. F. Oster, Cellular motions and thermal fluctuations: The Brownian ratchet, Biophys. J.
65, 316 (1993).
[25] Y. Marcy, J. Prost, M. F. Carlier, and C. Sykes, Forces generated during actin-based propulsion: A direct measurement
by micromanipulation, Proc. Natl. Acad. Sci. U.S.A. 101, 5992 (2004).
[26] C. Brangbour, O. du Roure , E. Helfer, D. D´emoulin, A. Mazurier, M. Fermigier, M. F. Carlier, J. Bibette and J. Baudry,
Force-velocity measurements of a few growing actin filaments, PLoS Biol. 9, e1000613 (2011).
[27] D. D´emoulin, M. F. Carlier, J. Bibette, and J. Baudry, Power transduction of actin filaments ratcheting in vitro against a
load, Proc. Natl. Acad. Sci. U.S.A. 111, 17845 (2014).
[28] S. H. Parekh, O. Chaudhuri, J. A. Theriot and D. A. Fletcher, Loading history determines the velocity of actin-network
growth, Nat. Cell Biol. 7, 1219 (2005).
[29] K. Tsekouras, D. Lacoste, K. Mallick and J. F. Joanny, Condensation of actin filaments pushing against a barrier, New J.
Phys. 13, 103032 (2011).
[30] D. Das, D. Das and R. Padinhateeri, Collective force generated by multiple biofilaments can exceed the sum of forces due
to individual ones, New J. Phys. 16, 063032 (2014).
[31] D. Nelson, T. Piran and S. Weinberg, Statistical Mechanics of Membranes and Surfaces, (World Scientific, Singapore,
2004).
[32] T. D. Pollard, Rate constants for the reactions of ATP-and ADP-actin with the ends of actin filaments, J. Cell. Biol. 103,
2747 (1986).
[33] J. Krawczyk and J. Kierfeld, Stall force of polymerizing microtubules and filament bundles, Europhys. Lett. 93, 28006
(2011).
[34] J. Zhu and A. Mogilner, Mesoscopic model of actin-based propulsion, PLoS Comp. Biol. 8, e1002764 (2012).
[35] R. K. Sadhu and S. Chatterjee (Unpublished).
[36] B. Fogelson and A. Mogilner, Computational estimates of membrane flow and tension gradient in motile cell, PLoS One
9, e84524 (2014).
[37] Y. Schweitzer, A. D. Lieber, K. Keren and M. M. Kozlov, Theoretical analysis of membrane tension in moving cells,
Biophys. J. 106 84 (2014).
|
1304.4025 | 2 | 1304 | 2013-08-27T13:26:42 | The syncytial Drosophila embryo as a mechanically excitable medium | [
"physics.bio-ph",
"cond-mat.soft",
"q-bio.TO"
] | Mitosis in the early syncytial Drosophila embryo is highly correlated in space and time, as manifested in mitotic wavefronts that propagate across the embryo. In this paper we investigate the idea that the embryo can be considered a mechanically-excitable medium, and that mitotic wavefronts can be understood as nonlinear wavefronts that propagate through this medium. We study the wavefronts via both image analysis of confocal microscopy videos and theoretical models. We find that the mitotic waves travel across the embryo at a well-defined speed that decreases with replication cycle. We find two markers of the wavefront in each cycle, corresponding to the onsets of metaphase and anaphase. Each of these onsets is followed by displacements of the nuclei that obey the same wavefront pattern. To understand the mitotic wavefronts theoretically we analyze wavefront propagation in excitable media. We study two classes of models, one with biochemical signaling and one with mechanical signaling. We find that the dependence of wavefront speed on cycle number is most naturally explained by mechanical signaling, and that the entire process suggests a scenario in which biochemical and mechanical signaling are coupled. | physics.bio-ph | physics |
1
The syncytial Drosophila embryo as a mechanically excitable
medium
Timon Idema1,∗, Julien O. Dubuis2, Louis Kang1, M. Lisa Manning3, Philip C. Nelson1, Tom C.
Lubensky1, Andrea J. Liu1,†
1 Department of Physics and Astronomy, University of Pennsylvania, Philadelphia, PA,
USA
2 Department of Physics, Princeton University, Princeton, NJ, USA
3 Department of Physics, Syracuse University, Syracuse, NY, USA
∗ Present address: Department of Bionanoscience, Delft University of Technology, Delft,
The Netherlands; E-mail: [email protected]
† E-mail: [email protected]
Abstract
Mitosis in the early syncytial Drosophila embryo is highly correlated in space and time, as manifested
in mitotic wavefronts that propagate across the embryo.
In this paper we investigate the idea that
the embryo can be considered a mechanically-excitable medium, and that mitotic wavefronts can be
understood as nonlinear wavefronts that propagate through this medium. We study the wavefronts via
both image analysis of confocal microscopy videos and theoretical models. We find that the mitotic
waves travel across the embryo at a well-defined speed that decreases with replication cycle. We find
two markers of the wavefront in each cycle, corresponding to the onsets of metaphase and anaphase.
Each of these onsets is followed by displacements of the nuclei that obey the same wavefront pattern.
To understand the mitotic wavefronts theoretically we analyze wavefront propagation in excitable media.
We study two classes of models, one with biochemical signaling and one with mechanical signaling. We
find that the dependence of wavefront speed on cycle number is most naturally explained by mechanical
signaling, and that the entire process suggests a scenario in which biochemical and mechanical signaling
are coupled.
1
Introduction
The early embryos of many species, including Drosophila [1 -- 3], Xenopus [4 -- 6], Oryzias [7], Fundulus [8],
and zebrafish [9, 10], exhibit metachronous mitosis, in which mitosis progresses as a wavefront through
the embryo. Such wavefronts are reminiscent of biochemical wavefronts that are used to transmit signals
across many cells in other biological systems, such as wavefronts of the molecule cAMP that propagate
in a colony of Dictyostelium when it begins to aggregate to form a fruiting body [11 -- 13]. Propagating
wavefronts, however, need not be purely biochemical in origin. The process of mitosis is a highly mechan-
ical one that involves significant changes in the volume occupied by chromatin [14] as well as separation
of chromosomes [15]. This raises the question of whether mitotic wavefronts are purely biochemical
phenomena or whether they might have a mechanical component as well.
The nuclei of the Drosophila embryo are syncytial (i.e., they share the same cytoplasm and are not
separated into individual cells by plasma membranes) during their first thirteen division cycles. The nuclei
migrate to the egg's surface during the ninth cycle. There they divide five more times, until the fourteenth
cycle, when cell membranes form and gastrulation begins [1]. Mitotic wavefronts are observed in cycles
9 through 13 [1]. In this period, chemical diffusion is unhindered by membrane barriers. For example, it
is known that calcium, a signal carrier that influences many local phenomena including mitosis [16 -- 18],
exhibits spikes of concentration in the syncytial embryo [19 -- 23] that have been resolved into a wavefront
that travels across the embryo at the same speed as the mitotic wavefront [20].
However, mitosis is also a mechanical phenomenon. In the syncytial embryo, nuclei are embedded in
2
an elastic cytoskeleton, which contains both actin and microtubules [24 -- 26]. Actin caps assemble around
each of the nuclei at the end of interphase, and provide anchor points for the mitotic spindles that pull
the two daughter nuclei apart [24 -- 27]. Recent work shows that mechanical interactions are important
for re-organization of the nuclei after mitosis [28], and optical tweezer experiments show that nuclei are
mechanically coupled [29]. Moreover, mechanical deformations of the embryo are known to be able to
induce morphogen expression [30]. However, little is known about how mechanical interactions affect
collective phenomena such as mitotic wavefronts at the level of the entire embryo.
In this paper we report the results of both our image analysis of wavefronts in early Drosophila em-
bryos, and our theoretical studies of models of wavefront propagation. Using novel tracking techniques, we
analyzed confocal microscopy videos taken of Drosophila embryos in which the nuclear DNA/chromosomes
are visualized by labeling their histones with GFP. Our analysis yields the position, shape and dynamics
of the DNA/chromosomes with high temporal and spatial resolution during cycles 9 -- 14. We observe
two distinct markers of the mitotic process in each cycle, one corresponding to the onset of metaphase
(at which point the chromosomes condense in the nuclear midplane, known as the metaphasic plate,
see figure 4 for an illustration of the different stages) and one corresponding to the onset of anaphase.
Both onsets exhibit identical wavefront patterns, indicating that they are indeed two markers of the same
process. Both onsets are also followed by displacements in the positions of the nuclei that also exhibit
the same wavefront patterns. Finally, we find that the wavefront speed slows down from one cycle to the
next.
We treat the embryo theoretically as an excitable medium, consisting of nuclei that can be triggered
into initiating metaphase or anaphase, thereby locally exciting the medium and thus signaling their
neighbors. We not only consider the well-known case of nonlinear wavefront propagation in a chemi-
cally excitable medium [31, 32], but introduce a model for the early embryo as a mechanically excitable
medium [33], through which mitotic wavefronts can propagate via stress diffusion. Comparing the data
with the results of these two models, we find that our observations are difficult to reconcile with a purely
biochemical scenario. In such a scenario, the wavefront speed has a tendency to increase with nuclear
density, and thus with cycle, contrary to our observations. The observations can, however, be explained
quite naturally by a novel scenario in which nuclei not only respond to their mechanical environment,
but also actively use it to signal each other. Our results suggest that mitotic wavefronts in syncytial
Drosophila embryos may constitute one example of a previously unexplored form of mechanical signaling
via nonlinear wavefronts that could also arise in very different biological contexts [33, 34].
2 Results
2.1 Image analysis results
Nuclear cycle and shape
An example image of detected nuclei in a Drosophila embryo is shown in figure 1a.
In each cy-
cle, as the nuclei progress from interphase through metaphase to anaphase, the detected shape of the
DNA/chromosomes changes in a well-defined manner (figure 1b). Newly separated nuclei are small and
spherical, and thus show up in our shape tracking as small circles. During interphase, the nuclear DNA
grows in size over time as it is duplicated. At the onset of metaphase, the chromosomes condense in the
midplane of the nucleus, and appear to elongate into an ellipse. The final step of mitosis, the onset of
anaphase, corresponds to two detectable changes in the shape: a sudden shift of the orientation axis over
a π/2 angle, and a change of aspect ratio. An example plot showing the ratio of the length of the two
axes as a function of time during a cell cycle is given in figure 1c.
3
Wavefront pattern in the onset of metaphase and anaphase
The onsets of metaphase and anaphase, as determined by the axes ratio (figure 1d) are indicated by
dotted blue lines and dashed orange lines, respectively. Evidently the onset of metaphase exhibits a
wavefront pattern, or rather two wavefronts, one propagating from each pole. The two wavefronts do not
necessarily start at the same time. The onset of anaphase exhibits the same wavefront pattern. Mitotic
waves were first observed by Foe and Alberts [1]; with better time resolution, it is evident that these
wavefronts can be resolved into two distinct markers of mitosis, corresponding to the onsets of metaphase
and anaphase. There may well be additional markers that cannot be resolved via histone labeling alone;
for example, the work of Parry et al. [20] indicates that calcium may provide another marker for the
mitotic process, and we find that the nuclear displacements also provide markers (see below).
Effect of shape changes on nuclear positions
The processes of metaphase and anaphase affect not only the shapes of the chromosomes, but also their
positions. After each of the shape changes, the nuclei move collectively through the embryo, almost
exclusively along the long axis (which we designate as the x-axis), resulting in a global 'breathing mode'
of the entire embryo (see SI movie 1 [35]). Remarkably, after an initial transition in which the nuclei
re-organize after anaphase (studied in detail by Kanesaki et al. [28]), the nuclei hardly move with respect
to their nearest neighbors during this collective movement. Figure 1e shows the average displacement ∆x
along the x-axis of a small set of nuclei. Figure 1f shows the same motion for all nuclei. Note that there
are subtle changes in the gray scale that parallel the metaphasic and anaphasic wavefronts but that are
shifted to the right (i.e. occur later in time) with respect to each of those wavefronts. This illustrates that
the nuclear displacements follow the same wavefront pattern as the axes ratio, so that the displacements
also serve as markers for the mitotic wavefront. The existence of such a marker in the displacements as
well as in the axes ratio and in calcium concentration underlines the important interplay of mechanics
and biochemistry in the mitotic process.
The displacement response to the onsets of metaphase and anaphase causes the nuclei to move to
new equilibrium positions (figure 1e). Note that the relaxation time of this response is fairly long, about
half the length of the mitotic phase (∼ 1min) for the onset of metaphase and about half the length of
the following interphase (up to 10min) for the actual divisions. The displacements following the onset
of metaphase therefore occur before the cytoskeletal reconstruction process, which takes place during
anaphase, whereas the displacements following the onset of anaphase happen during the aftermath of the
cytoskeleton reconstruction.
Wavefront speeds
We quantify the wavefront speeds in figure 2 for two sets of movies, where the environmental conditions
(in particular the temperature) were approximately the same for all movies in a given set, but differed
between the two sets (the data of the two sets were taken several months apart). Figure 2a shows an
example of a position vs. time plot of all metaphase (blue diamonds) and anaphase (red pluses) onset
events in a single cycle of a single embryo. The slope, corresponding to the wavefront speed, is clearly
constant across the embryo. Figure 2b shows the ratio of the speeds of the wavefronts as measured by
the onsets of metaphase and anaphase of all embryos, showing that for a given embryo and cycle, these
are identical, confirming that they are two markers of a single process.
From embryo to embryo there are large variations in wavefront speed (figure 2c), but they all show
a consistent reduction in speed from one cycle to the next. This trend is illustrated in figure 2e, where
we plot the same data, normalized by the speed of the first wavefront, on a log-linear scale. Although
our data only span a single decade, this figure suggests that the decrease of wavefront speed with cycle
number is consistent with a decaying exponential.
4
Figure 2d shows that the time interval that separates the onset of metaphase from the onset of
anaphase is the same for all cycles for a given embryo, but is different for the two different sets of data.
By looking at the point at which the nuclear envelope breaks down and reforms, Foe and Alberts [1] also
found that the duration of the mitotic phase is constant through cycles 10, 11 and 12 (3 minutes in their
observations, comparable to our result), but was longer for cycle 13 (5 minutes). The re-formation of
the nuclear envelope membrane may therefore take significantly longer in the last syncytial cycle, even
though the actual mitotic processes continue to follow the pattern of the earlier cycles.
Cycle statistics
The nuclei on the surface are separated by a well-defined distance an, which decreases with cycle number
n. Because the number of nuclei doubles from one cycle to the next, it is not surprising that an decays
exponentially, scaling like an ∼ 2−β(n−n0), with n the cycle number and n0 the number of the first
observed cycle. We consistently found a value of β = 0.46 in our experiments (figure 2f and table 1).
The value of β is slightly less than 1/2, presumably because the curved embryo is being projected onto a
plane. We have also measured the duration of each cycle, tn, and found that, over the observed cycles, it
increases with cycle number n, with a weak exponential growth: tn = t0e0.29·n, where t0 = 33s for set 1
and t0 = 25s for set 2, see table 1 and figure 7.
2.2 Theoretical Analysis
Our observation that the mitotic wavefronts propagate at constant speed across the embryo suggests
that the embryo can be considered as an excitable medium that supports nonlinear front propagation.
Alternatively, the nuclei could all have biological clocks that determine when mitosis starts, which operate
independently; in that case the wavefront would be only a result of a lucky timing of those clocks.
We discuss various timing models and show that they are inconsistent with our observations in the
supplementary material. Here we concentrate on two distinct classes of models for front propagation in
excitable media. In the first model the nuclei communicate by releasing a small chemical species, which
then diffuses to neighboring nuclei, triggering them to initiate mitosis. In the second model we explore
the novel idea that mitotic wavefronts in the early embryo can be described by wavefront propagation
in a medium that is mechanically rather than chemically excitable. In this model, forces exerted at the
onset of the mitotic phase give rise to mechanical stresses that trigger other nuclei to proceed to mitosis
as well.
Biochemical-signaling model
At the end of a cycle, when all nuclei have completed the duplication of their DNA, we assume that
they are in an excitable state, meaning that they can be triggered to initiate mitosis once they receive
an appropriate signal. An obvious candidate for signaling between nuclei is a small protein (e.g. a Cdk,
cyclin or some other activator), which we will denote as A. By definition, nuclei can divide only once
per cycle; therefore, in our model, we introduce a refractory period for each nucleus following anaphase,
equal to the duration of the interphase.
To introduce chemical excitability, we assume that if the local concentration of A exceeds a threshold
α, the nucleus starts its program of mitosis, part of which involves releasing more A. A then diffuses
away, raising the concentration of A at neighboring nuclei, and so on. In our model we allow for a time
delay tdelay between trigger and release, meaning that a nucleus does not release more A until a time tdelay
after its local concentration exceeds α. We model releases of A by the nuclei (or sources) as localized
pulses (Dirac delta functions), and the system is initiated with a single nucleus releasing a quantity Q of
A. The wavefront at any point in time corresponds to the position of all nuclei that release A at that
5
Figure 1. Observation of wavefronts and mechanical response. a) Image of a Drosophila
embryo during mitosis at the end of cycle 11, with the detected chromosomal contours overlaid.
Anaphasic wavefronts (orange dashed curved lines), the long axis (green dashed straight line) and a
typical slice perpendicular to the long axis (green parallel straight lines) are indicated. b) Sketch of the
three main states in image analysis: interphase (circular contours), metaphase (compressed elliptical
contours), and anaphase (highly extended elliptical contours, perpendicular to metaphase contour). See
also figure 4. c) Ratio of the two elliptical axes of the detected shape of the nuclear DNA/chromosomes
vs. time in cycle 11, averaged over an x-slice (as shown in a); error bars indicate variation within the
slice. The transitions between interphase and metaphase, as well as the onset of anaphase, are sharp
and indicated respectively by dotted (blue) and dashed (orange) vertical lines. The slice shown was
taken at x = 200µm. d) Kymograph showing the elliptical axes ratio, a/b (where white indicates values
larger than 1 and black indicates values smaller than 1), as a function of position x and time. The
dotted and dashed lines indicate the onsets of metaphase and anaphase, as in figure c. e) Average
x-displacement ∆x of the nuclei within one slice vs. time. After a nucleus has divided, we use the
average position of its two daughters. The slice shown is identical to the one in figure c. f) Kymograph
showing the collective motion of nuclei in slices taken at different positions along the long axis of the
embryo. White indicates motion in the positive x direction, black in the negative x direction. Dotted
and dashed lines again indicate the onsets of metaphase and anaphase. Note that the displacements
occur sometime after these onsets, but follow the same wavefront pattern.
moment. Details on how to solve the diffusion equation and carry out the other needed calculations are
given the appendices. An example wavefront is shown in figure 3a.
In the case of zero delay time, the speed v of the resulting wavefront is determined by three parameters:
the diffusion constant D, the nuclear spacing a and the concentration threshold α. We obtained the
value of a from direct measurements (figure 2f). Gregor et al. [36] found from diffusion experiments in
Drosophila that the diffusion constant of a molecule with hydrodynamic radius R is well described by a
ceinterphasemetaphaseanaphaseb4006008001000120005101520ΔX (μm)40060080010001200time (s)0.40.60.81.01.21.41.6a/btime (s)abaXfX (μm)400500300200100450600800100012001400time (s)400500300200100450600800100012001400time (s)X (μm)d 6
Figure 2. Wavefront propagation and speeds. a) x-coordinate of nuclei at the onset of metaphase
(blue diamonds) and anaphase (red pluses) vs. time for the wavefront shown in figure 1. Both events
show two clear wavefronts moving in from near the embryo poles (solid lines). b) Ratio of the speeds of
the wavefronts as measured by the onset of anaphase (vap) and metaphase (vmp), for different embryos
and cycles. Each embryo is indicated by a different symbol and color, with the closed and open symbols
representing two different measurement sets. Ratios for a given cycle and different embryos are slightly
separated horizontally. c) Wavefront speed vs. cycle. Two of the embryos contribute two waves per
cycle (coming in from opposite poles, as in figure 1a; blue squares and green diamonds). Although the
actual propagation speeds vary significantly from one embryo to the next, they all follow the same
trend, decreasing with successive cycles. d) Time interval between the onset of metaphase and anaphase
vs. cycle. e) Log-linear plot of wavefront speeds vs. cycle, normalized by the speed of its first observed
wavefront (if the first observed wave front is in cycle 10) or 0.71 times its first observed wavefront (if the
first observed wavefront is in cycle 11). The black open circles connected by a dashed line corresponds
to a scaling of 0.71 per cycle, showing that all embryos follow the same exponentially decaying trend. f)
Average distance between nearest neighbors on a logarithmic plot. The dashed line corresponds to a
dependence 2−βn, where n is the cycle number and β = 0.46. In figures b-f, the same symbol/color
corresponds to the same embryo.
modified Stokes-Einstein relation [37]:
D = kBT /(6πηR) + b,
(1)
where kB is Boltzmann's constant, T the temperature, η = 4.1 ± 0.4cP the effective viscosity of the
syncytial Drosophila embryo, and b = 6.2 ± 1.0µm2/s is an experimentally determined constant. Using
this expression, we estimate that a reasonable value for the diffusion coefficient (from the size of the
activator A), would correspond to a chemical with a radius of approximately 5.0nm and therefore a
diffusion constant of about 10µm2/s.
time (s)X (μm)ca4005006007008000125250375500101112131.02.03.04.05.00.0cyclevwavefront (μm/s)1011121350100150200250300101112130.20.40.60.81.01.21.41011121310203015101112131.000.500.300.70cyclevnormalizedecycleΔtmitosis (s)dcyclevap / vmpbfcycleNN dist. (μm)0.01.63500 Combining the parameters of our model, we define a nondimensional threshold and speed:
¯α = a2α/Q,
¯v =
v
D/a
.
7
(2)
(3)
In a three-dimensional model the power of a in equation (2) is 3. As shown in appendix G.1, for a
steady-state wavefront, we then have ¯v = 1/f (¯α), where f (¯α) increases monotonically with ¯α (figure 9).
Consequently, if both D and α are fixed, the wavefront speed v increases as the nuclear spacing a decreases,
and thus the speed increases with cell cycle, in direct contradiction to our experimental observations.
Thus, the simplest form of the biochemical signaling model cannot describe the data of figure 2c.
We next consider the possibility of a delay tdelay between the time when the local concentration of A
reaches the threshold value α, and the instant when more A is released. In the limit where a2/D (cid:29) tdelay,
the wavefront speed is determined by diffusion as before, v = D/(af (¯α)). In the opposite limit, a2/D (cid:28)
tdelay, we find v = a/tdelay, so v would decrease with cycle number for constant tdelay. We find that for
our system, introducing a small, fixed delay time of 2 − 8s puts us in the crossover regime between these
two types of behavior. Consequently, the model predicts that for the first few cycles, the wavefront speed
should increase, whereas it should level off or slightly decrease in the last cycle. Changing the value of
the threshold α does not qualitatively change this result. Changing the value of the diffusion constant D
simply shifts the position of the crossover.
A key result of our analysis with a fixed time delay is that a physically unrealistic diffusion coefficient
is required in order to reproduce our experimental observations. In order to obtain a strictly decreasing
wavefront speed for the range of interest, a diffusion constant of more than 100µm2/s is required. This
corresponds to a signaling particle that is even smaller than an ion. Thus, a biochemical-signaling model
with a time delay that is independent of cell cycle cannot describe our observations either (figure 3b).
We also investigated the wavefront speed in the case where the delay time is allowed to vary from one
cycle to the next. Naturally, given a value for the diffusion constant and the threshold, for each cycle
we can find a delay time such that the speed predicted by the model matches the observed speed; these
values are listed in table 2. The found values do not show any consistent trend, and differ quite strongly
between the two data sets. There is no obvious explanation for what would set the time delay in each
cycle; the time delay is not proportional to the total duration of the cycle (which increases from one cycle
to the next) or any other obvious time scale. Therefore, this procedure simply shifts the problem from
understanding the trend in the wavefront speed to understanding the trend in the delay time, and does
not provide a satisfactory explanation of our data.
On the basis of these results, we conclude that it is very unlikely that a wavefront that propagates via
diffusion of some chemical species would slow down with cycle number, as observed in our experiments.
We also note that any model in which the biochemical signal is mediated by a method that is faster than
diffusion (such as active transport) suffers from the same problem: the predicted wavespeed would go up
with increasing cycle, because the spacing between the nuclei goes down.
Mechanical-signaling model
The early embryo cannot support ordinary elastic waves because it is heavily damped by the viscosity
of the cytosol. Consequently, displacements do not propagate ballistically as in a wave, but diffusively.
However, just as diffusion of A can lead to nonlinear wavefront propagation in the biochemical signaling
model, diffusion of displacement could lead to wavefront propagation in a mechanical signaling model.
We therefore introduce a model in which the nuclei communicate via stresses or strains that they exert
on the cytoskeleton at the initiation of the mitotic phase. For example, these could be the forces that
cause the chromosomes to condense into sister chromatids in prophase or to align in the nuclear midplane
at the onset of metaphase.
8
In our model, a nucleus starts its program when the largest eigenvalue of the local stress tensor
exceeds a threshold value α. We describe the cytoskeleton as a homogeneous linear elastic medium,
characterized by two elastic parameters, for example its bulk and shear moduli (K and µ, respectively)
or equivalently the Young's modulus E and dimensionless Poisson ratio ν. The viscous fluid in which
the elastic cytoskeleton is immersed exerts a drag force on it, characterized by a damping constant Γ.
Assuming that the nuclei exist in a thin layer near the surface of the embryo, we denote the deformations
i − xi (i = 1, 2), where the deformation maps point (x1, x2) onto
in the plane of the layer by ui = x(cid:48)
point (x(cid:48)
2). In the overdamped limit (zero Reynolds number), the displacement (cid:126)u of a nucleus can be
described by [38]:
1, x(cid:48)
Γ∂tui =
E
2(1 + ν)
∂j∂jui +
E
2(1 − ν)
∂i∂juj.
(4)
The term on the left represents the damping with damping factor Γ, and the two terms on the right are
the elastic force per unit volume. Equation (4) is reminiscent of the diffusion equation: a time derivative
on the left equals second-order space derivatives on the right. This model can therefore be thought of as
describing the diffusion of the vector displacement field ui. The right hand side of equation (4) gives rise
to two quantities with the dimensions of diffusion constants [33]:
D1 =
E
(1 − ν2)Γ
=
1 − ν
2
µ
Γ
and
D2 =
E
2(1 + ν)Γ
=
µ
Γ
.
(5)
In order to introduce mechanical excitability into the model, we assume that if the largest eigenvalue
of the stress tensor at a nucleus at position (cid:126)x0 exceeds a threshold value, α, at time t0, it triggers the
nucleus into action which involves adding additional stress to the system. This stress can be added in the
form of a source term Qij = fixj + xifj, a symmetric tensor of rank 2, corresponding to a force per unit
volume (cid:126)f acting over a distance (cid:126)x. Qij is therefore the symmetric combination of a force and a distance,
with the dimensions of a stress (force per unit area), so it represents a stress source. In two dimensions,
Qij has an isotropic part of the form Q0δij and a traceless anisotropic part of the form Q1(ninj − 1
2 δij)
where n indicates the anisotropy direction.
If n makes an angle θ with the x-axis, we find that the
components of Qij in matrix notation are given by:
(cid:18) 1
0
(cid:19)
0
1
(cid:18) cos 2θ
(cid:19)
Q = Q0
δ((cid:126)x) − Q1
sin 2θ
sin 2θ − cos 2θ
δ((cid:126)x).
(6)
Here δ((cid:126)x) is the two-dimensional Dirac delta function. Similar active force dipoles have previously been
introduced into other tissue-level models, such as those of Bischofs et al. [39] and Ranft et al. [40]. To
include the force due to the added stress at (cid:126)x = (cid:126)x0 and t = t0, we add the term ∂jQijδ((cid:126)x − (cid:126)x0)Θ(t − t0)
to the right hand side of equation (4):
Γ∂tui =
E
2(1 + ν)
∂j∂jui +
E
2(1 − ν)
∂i∂juj + ∂jQijδ((cid:126)x − (cid:126)x0)Θ(t − t0),
(7)
where Θ(t) is the Heaviside step function. Equation (7) essentially describes the diffusion of the vector
displacement field ui due to a tensor source term. It is similar, but not identical, to a scalar reaction-
diffusion equation, which describes the evolution of a scalar concentration field c due to a scalar source
term. It is therefore not surprising that the model described by equation (4) also produces wavefronts,
as can be seen in figure 3c and d.
In order to compare the model results with the data, we need to estimate the values of the elastic
constants and the damping parameter. The speed v now depends on the quantity D = µ/Γ that deter-
mines the dimensional part of both diffusion constants (equation (5)), as well as the nuclear spacing a,
the strengths Q0 and Q1 of the source term, and the threshold value α. It is well known that the values
9
of both the elastic and the viscous modulus of a polymer network depend strongly on filament concen-
tration [42 -- 46, 53], which can differ from one cycle to the next. Because the number of nuclei doubles in
each cycle, the number of actin caps in the network doubles as well (see figures 4 and 5). Thus, the local
concentration of actin and of microtubules should effectively double with cycle number n. We therefore
write c ∼ 2(n−n0), where as before n0 is the number of the first observed cycle. Both the storage and
loss moduli of polymer networks increase with concentration approximately as power laws, but the actual
powers are debated [42, 44 -- 46]. Moreover, in each successive cycle the nuclei get pushed further out into
the plasma membrane encompassing the entire embryo [1], increasing the friction coefficient. Because the
dynamics of our system depend only on the value of the two effective diffusion constants given in equa-
tion (5), we will not be able to distinguish the dependence of the storage and loss moduli independently.
Instead we assume a dependence D = µ/Γ ∼ c−γ ∼ 2−γ(n−n0). We will use γ as a fit parameter.
Because of the mathematical similarity between the mechanical-signaling model (equation (4)) and
the diffusion model for concentration fields, we can use the same type of dimensional analysis as for
the biochemical-signaling model. We again use the dimensionless threshold ¯α and wavefront speed ¯v
defined by equations (2) and (3), where Q is now the typical strength of the source term, and we write
¯v = g(¯α, ν). We determine g(¯α, ν) numerically, and find that it can be well described by the functional
form g(¯α, ν) = −4(c1 + c2 ¯α) log(¯α)/(1 − ν2) + c3, where c1, c2 and c3 are constants that depend on
the choice of source term and boundary conditions [33]. In the analysis that follows, we have adopted
boundary conditions that are free along the long axis and periodic along the short axis to mimic the
elongated shape of the embryo.
Figure 3e shows a fit to a displacement wavefront profile following the first detectable sign of the
mitotic wavefront (onset of metaphase) in the initial (tenth) cycle. We find that in order to fit the profile,
the source term (6) must be nearly isotropic, so that Q1 (cid:28) Q0. We therefore set Q1 = 0 and fit to find
the threshold stress, which gives α = 0.1Q0/a2
10, with a10 the spacing in cycle 10. Thus, the threshold
stress is approximately ten percent of the force exerted per unit area.
Figure 3f shows a fit of the wavefront speed of the two datasets, Q1 = 0 and α = 0.1Q0. Here, the fit
parameter is the exponent γ that governs the change in the displacement diffusion constant from cycle
to cycle. Both datasets are well-described with a value of γ = 1.15. The only difference between the two
datasets is the value of the displacement diffusion constant D = µ/Γ in the 10th cycle, which is about
3µm2/s for set 1 and about 6µm2/s for set 2.
These values for the diffusion constant are comparable to those found in microrheology experiments,
which have measured the frequency-dependent complex shear modulus in a variety of living cells [47 -- 51].
In contrast to pure actin networks, living systems often do not exhibit a low-frequency plateau in the
storage modulus G(cid:48)(ω). Although this makes a precise determination of the shear modulus difficult, we
can still get a decent order-of-magnitude estimate from the experimental data at µ ∼ 5Pa. The damping
constant Γ is given by Γ = cηξ = η/ξ2 [41,45], where c is the filament concentration, η = 4· 10−3Pa· s [36]
is the ambient fluid viscosity, and ξ ∼ 100nm is the mesh size of the actin network. We thus estimate
D ∼ 10µm2/s, in good agreement with our fitting results.
The found value for the exponent γ is also reasonable.
In-vitro experiments on entangled F-actin
solutions indicate that the storage and loss moduli depend on the concentration in the same way [45],
which leads us to expect the shear modulus µ and viscosity η to have similar dependence on c. On the
other hand, for a semidilute solution of rigid rods, the viscosity is expected to rise as c3, where c is the
filament concentration [53]. Because the damping factor Γ scales with the concentration and the mesh
size ξ, which itself depends on the concentration as ξ ∼ c−1/3, we find that γ should be somewhere between
2/3 (for an entangled F-actin solution) and 8/3 (for a semidilute solution). Our value of γ = 1.15 indicates
that our system falls somewhere in between these two regimes, which is reasonable for the Drosophila
embryo, with its hemispherical actin caps enclosing each nucleus (see figure 5).
Figures 3e and f show that we can consistently fit both the wavefront velocity and the displacement
profile of the nuclei as a function of time immediately after the metaphasic wavefront, with the same
10
theory. We note that this is not possible with the chemical signaling model, which cannot provide any
information about the displacement profile. The fact that we can fit both quantities with the same
parameters therefore provides strong evidence in favor of the mechanical signaling model.
In addition, we note that the nuclear displacement profile provides a more discriminating test of the
mechanical signaling model than the wavefront velocity. Although the velocity wavefront speed data alone
can be fitted by either purely isotropic force dipoles or purely anisotropic force dipoles (and presumably
anything in between), the displacement wavefront can only be fit with dipoles with a strong isotropic
component. Moreover, although either the displacement or the velocity data can be fit with different
combinations of the threshold and diffusion constant, the numbers given above are the only ones for
which we can fit both quantities.
In summary, the mechanical signaling model agrees much better with the data than the biochemical
signaling model in two important respects. First, it captures the dependence of the wavefront velocity on
cell cycle number while the biochemical signaling model does not. From dimensional analysis, we have
shown for both models that the wavefront velocity depends mainly on D/a, where D is the diffusion con-
stant and a is the average distance between nuclei. Note that a decreases with cycle number. In the case
of biochemical signaling, the chemical diffusion coefficient D remains constant with cycle number, leading
to a wavefront velocity that tends to increase with cycle number. In the case of mechanical signaling,
however, the displacement diffusion coefficient, D = µ/Γ, decreases quite strongly with cycle number
because the damping coefficient, Γ, should increase more rapidly with filament concentration than the
elastic constant, µ. If we make the reasonable assumption that the filament concentration increases with
cycle number, then this means that the stress diffusion coefficient decreases with cycle number, leading to
a wavefront velocity that decreases with cycle number, in accord with experimental observations. Second,
we have shown that the mechanical signaling model describes not only the wavefront velocity but also the
displacement profile following the metaphasic wavefront. In the biochemical-signaling model, a separate
mechanical description would be necessary in order to describe the nuclear displacements.
Finally, we note that we have assumed that the elastic constants and damping coefficients vary from
cycle to cycle but do not change much during the period that we are focusing on. However, the cytoskele-
ton reconstructs completely during the cell cycle. Our analysis will apply as long as the elastic constants
and damping coefficient do not change appreciably from the time that the original triggering wavefront
is generated to the time that the anaphasic wavefront occurs. Thus, the assumption is that cytoskeletal
reconstruction occurs sometime during anaphase and is finished before the process of mitosis begins in
the next cycle. In particular, this also means that our model should not be able to correctly predict the
much larger displacements following anaphase (see figure 1e), which indeed it cannot.
3 Discussion
During the early cycles of Drosophila development, the cycles of the nuclei are strongly coupled across
the entire embryo by mitotic wavefronts that travel at constant speed across the embryo. We summarize
our observations as follows:
1. There are several markers of the mitotic process in each cycle, corresponding to the onsets of
metaphase and anaphase, which are visible as wavefronts that travel across the embryo (figure 1d).
2. The speed of the mitotic wavefronts slows down in each successive cycle (figures 2c and 2e).
3. The onsets of metaphase and anaphase both trigger a mechanical response of the entire embryo in
the form of displacements of the nuclei that also exhibit a wavefront pattern (figure 1f).
In addition to these observations, we add those of Parry et al. [20]:
4. There is a visible wavefront in calcium release that coincides with the onset of anaphase.
11
Figure 3. Propagation of wavefronts by chemical and mechanical signaling. a) Color plot
showing the chemical wavefront in two dimensions. The wave starts in the center (red dot) with a single
Dirac delta peak release. The color coding indicates when a nucleus releases its chemical to the bulk,
going from red through the different hues of the rainbow to violet. b) Plot showing the best fits (blue
and purple lines) of the diffusion model with time delay to the to the two sets of experimental data
(black and gray dots with error bars). Although the time delay manages to balance the trend that the
wavefront speed increases in the region of interest (but not before), the model fails to describe the
observed data. Here D = 10µm2/s. c) Color plot showing the mechanical wavefront in two dimensions
for totally anisotropic dipoles, including their orientations, which are picked at random, and free
boundary conditions. The color coding is the same as in figure a. d) Color plot showing the mechanical
wavefront in two dimensions for totally isotropic dipoles and semi-periodic boundary conditions
(periodic in vertical direction, free in horizontal direction). Wavefronts are initialized at both free ends
simultaneously and travel to the center, as in the experimental system. e) Plot showing fit (purple) of
the displacements calculated from the model to the experimentally obtained displacements (blue)
following the onset of metaphase. Fit parameters same as in figure e (set 1). Error bars obtained by
averaging over a slice of 40µm, as indicated in figure 1a. f) Plot showing fits (blue and purple lines) of
the mechanical model for isotropic force dipoles and semi-periodic boundary conditions to the two sets
of experimental data (black and gray dots with error bars). Fit parameters: α = 0.1Q/a2
10, where Q is
the dipole strength and a10 the spacing in cycle 10, γ = 1.15, and D = 3µm2/s (blue line/black
datapoints), D = 6µm2/s (purple line/gray datapoints).
101112131.02.03.04.05.06.07.010111213-2.00.02.04.06.0displacement (μm)speed (μm/s)speed (μm/s)500550600650700cyclecycletime (s)cbfe0.01.02.03.04.05.06.07.00.0ad 12
5. The speed of the calcium wavefront slows down in each successive cycle, presumably matching the
speed of the mitotic wavefront.
We have considered two scenarios to assess whether they are consistent with these observations. In
both cases, based on observations (1), (2) and (5), we take the observed metaphase, anaphase and calcium
wavefronts to be different markers of the same mitotic process, and assume that the mitotic wavefront is
triggered by a single event.
Scenario A
Mitosis is triggered by a biochemical signal. Here we assume that a biochemical signal is responsible for
triggering mitosis. The signal is mediated by the release and subsequent diffusion of a small ion, molecule
or protein. The only chemical species that is known to exhibit a wavefront pattern during mitosis is
calcium. However, because the onset of metaphase happens well before the observed calcium wavefront,
which coincides with the onset of anaphase (5), calcium cannot be the signal carrier. Our theoretical
analysis suggests that biochemical signaling is unlikely to be consistent with observation (3), since the
natural tendency of such a model is to produce a wavefront speed that increases with cycle number. The
larger the signaling molecule, the more pronounced this tendency is. Thus, we conclude that Scenario A
is unlikely.
This prediction could be tested by looking for wavefronts in likely signaling species. If the wavefront
propagates biochemically, then wavefronts should be observable in the appropriate signaling molecules
(presumably CDKs or cyclins that are known to govern checkpoints in the cell cycle that precede the
onset of metaphase [56]). If, as our model suggests, the wavefront does not propagate biochemically, then
the original signaling molecule should not exhibit wavefronts.
Scenario B
Mitosis is triggered by a mechanical signal. In this scenario, there is a mechanical wavefront that triggers
mitosis. The signal is transmitted via stress changes in the embryo and amplified by further release
of stress as other nuclei enter the mitotic phase. Because this wavefront propagates mechanically, this
speed slows down with successive cycles (2). Since we observe a metaphasic wavefront whose speed of
propagation slows down with cycle, the metaphasic wavefront itself could be the triggering mechanical
wavefront. It is more likely, however, that the triggering wavefront occurs earlier in the cycle and starts
a clock in each nucleus, which controls the mitotic process. As a result of this clock, there are many
markers of the process that exhibit the same wavefront pattern, including the onsets of metaphase and
anaphase (1), the release of calcium (5), and displacements of the nuclei during metaphase and anaphase
(3). This scenario is consistent with all observations.
Scenario B is consistent as well with independent observations made in Xenopus embryos. These
embryos are not syncytial; instead they are divided into cells from the first cycle. It is unlikely that a
biochemical signal could cross cell membranes to propagate a wavefront. Nevertheless, these embryos do
exhibit metachronous mitosis [4]. They also exhibit calcium oscillations inside each cell, which precede
anaphase [55]. Their behavior is therefore most consistent with Scenario B: an initial mechanical wavefront
triggered by a mechanical process at the onset of metaphase or earlier, is followed by a calcium signal
inside each cell and an anaphasic wavefront.
We emphasize that Scenario B does not imply that the entire process of triggering mitosis is mechani-
cal. Indeed, the mechanism by which additional stress is generated via a force dipole in our model must be
biochemical. First, there must be some sensor components that are activated when the stress exceeds its
threshold value. These components must then activate other biochemical species to eventually generate
additional stress by creating a force dipole. If Scenario B is correct, there should be a way of incorporating
our mechanically signaling model into models of the chemical networks that control the cell cycle, such
13
as those of Tyson and Novak [56]. One question is whether the original triggering mechanical wavefront
serves as a checkpoint in the cycle.
In order to understand how to include mechanical signaling into
such models, it is critical to have new experiments to identify precisely the original triggering wavefront.
Our model would predict that signaling molecules in stages of the cell cycle that follow this triggering
wavefront should exhibit wavefronts that slow down with cycle, while those in stages that precede the
triggering wavefront should not.
In principle, the estimated elastic constants and damping coefficients could be obtained directly from
experiments by measuring the storage and loss moduli of the embryo surface in vivo using two-point
microrheology. Optical tweezer experiments similar to the ones done by Schotz et al. [29] could also be
used to extract the elastic moduli and the drag coefficient we used in our mechanical model. The actin
concentration could be measured at the same time by staining the actin filaments with e.g. rhodamine,
as done by Parry et al. [20] or GFP-moesin, as done by Cao et al. [57].
Even though the process of mitosis is known to require chemical activation, the key assumption in
Scenario B is that the initial wavefront also propagates mechanically. This can be tested by mechanically
poking the embryo at different times within the cell cycle. If the cell is poked just in advance of the original
triggering wavefront, the poking itself should generate a wave that propagates from the poking site with
the same speed as the mitotic wavefront. If the embryo is poked too far in advance of the original triggering
wavefront, however, there should be no response.
If the embryo is poked after the mitotic wavefront
begins, there may be no response because the nuclei have already entered mitosis and can no longer be
triggered. Thus, we would expect that poking could generate a mitotic wavefront only if it is applied in
a certain time window of the cycle that could serve to identify the original triggering wavefront. Note
that experiments by Farge at a slightly later stage of development in Drosophila showed that mechanical
stress applied in the appropriate time window can lead dramatic changes in development [30]; Scenario
B suggests that mechanical stress is important even at the syncytial stages studied here.
Finally, we note that biochemical experiments could also test the mechanical-signaling model. The
most straightforward test would be to to destroy or degrade the filaments that mechanically couple
the nuclei. This should prevent the mechanical wavefronts from propagating and thus the nuclei from
synchronizing their mitosis. This could be done by injecting colcemid or nocodazole to disrupt the
microtubules or latrunculin which affects actin filaments, for example [3]. Other means of disrupting
cytoskeletal filaments, via mutation or laser ablation, should also affect the mechanical wave.
4 Materials and Methods
4.1 Confocal videos
The imaged flies were from a His-GFP stock with a P[w+ ubi-H2A-GFP] insertion on the third chromo-
some. All embryos were collected at 25◦C and dechorionated in 100% bleach for 1 minute. They were
picked using a 70µm nylon strainer (BD Falcon), rinsed in distilled water and laid down on a semiperme-
able membrane (Biofolie). The excess water was absorbed and the embryos were immersed in Halocarbon
oil 27 (Sigma Aldrich) and covered with a 22 × 22µm coverslip (Corning). Embryos were imaged with
a 20× oil immersion objective plan apochromat (Leica, NA=0.7) on a Leica SP5 laser scanning micro-
scope with excitation wavelength of 488nm (argon laser 60mW). 8 bit images were taken every second at
512× 1024 0.45nm pixels and 1.4µs/pixel (734ms/image). An example video is shown in SI movie 1 [35].
4.2 Image analysis
We visualized nuclear DNA/chromosomes by tagging their histones with GFP. To determine the positions,
sizes, aspect ratios and orientations of the DNA/chromosomes from each video frame, we developed a
new image analysis technique, explained in detail in [58]. In brief, we first applied a bandpass filter to
14
eliminate high-frequency noise. We then made a contour plot of the resulting image, found the locally
highest-level contour (i.e., the contours with no other contour inside them), and identified each of them
as a single nucleus. For each nucleus, we fit the contour at half-height with an ellipse to get its position,
shape and orientation. An example of an experimental image with the chromosomal tracking overlaid is
given in figure 1a.
Because the images were taken at high frequency (typically 1 Hz), the nuclei move less than their
own radius from one frame to the next, simplifying tracking. The obvious exception is when nuclei divide
during anaphase, and the observed shape splits in two. Because we detect shapes as well as positions of the
chromosomes in each nucleus, tracking divisions is easy as well: when a nucleus divides, the chromosomes
become highly elongated just before they split, and produce two almost circular daughters close to the
endpoints of the long axis of the mother immediately after it splits, which are easily identified.
Acknowledgments
We thank Thomas Gregor for providing resources for the experiments and for his careful reading of the
manuscript, and Xiaoyang Long for assistance with acquiring the experimental data. We also thank
Gareth Alexander and Michael Lampson for helpful discussions. This work was partially supported
by the Netherlands Organization for Scientific Research through a Rubicon grant (T.I.) and by the
National Science Foundation through grants NSF-DMR-1104637 (A.J.L.) and NSF-DMR-1104707 (L.K.
and T.C.L.).
Appendices
A Embryo layout and replication cycle
The four main stages in the Drosophila embryo replication cycle, which we can detect from our movies,
are illustrated in figure 4. A sketch of a cross-section of the embryo is shown in figure 5, illustrating how
the nuclei are all located at the surface of the embryo for cycles 9-13 [1, 60].
B Experimental data sets
Our image analysis results are for two different sets of experiments, which were carried out at ambient
room temperature several months apart. The ambient temperature was higher for the second set, resulting
in faster embryo development. We only used the data from those embryos which we could track from cycle
10-14 in the first set (Dataset 1, 3 embryos) and cycle 11-14 in the second set (Dataset 2, 4 embryos). SI
movie 1 [35] is the raw data of one of the embryos from set 1. This confocal microscopy imaging movie
shows a developing Drosophila embryo. The chromosomal histones are visualized by labeling with GFP.
The version of the movie shown here shows 1 image per 15s, displayed at 5fps, so sped up 75x. Movies
for data analysis were recorded at 1fps. The dimensions of each frame are 346 × 440µm.
C Additional image analysis results
The average data from the two sets are given in table 1, and their average speeds are plotted on a log-
linear scale in figure 6. The data from set 1 are given as closed symbols (blue, purple and green) in
figure 2 of the main text, the data from set 2 as open symbols (cyan, orange, gold and red). In figure 3c
of the main text and figure 6, the black dots correspond to the mean wavefront speeds of set 1, and the
gray ones to the mean speeds of set 2.
15
Figure 4. Illustration showing the four stages of the Drosophila embryo replication cycle that we can
detect from our movies: interphase (DNA replication), metaphase (condensation of chromosomes in the
nuclear midplane), anaphase (division of the nucleus in two daughter nuclei) and telophase (separation
of daughter nuclei). The plasma membrane is shown in gray, the actin cap (made of actin filaments) in
red, the microtubules in green, the centrosomes in yellow, and the DNA/chromosomes in blue.
interphasemetaphaseanaphasetelophaseinterphaseDrosophila replication cycle 16
Figure 5. Sketch of a cross-section through a Drosophila embryo valid for stages 9-13. Most nuclei are
located at the surface of the embryo. The nuclei are pushed outwards into the plasma membrane (gray),
resulting in the formation of somatic buds. Each nucleus is enclosed in a microtubule basket (green)
and contained in an individual actin cap (red), which gets disassembled after mitosis and re-assembled
during interphase. DNA/chromosomes are shown in blue and centrosomes in yellow. The yolk (light
blue) is a viscoelastic fluid containing water, cytoskeletal elements and necessary building blocks for the
nuclei. The yolk is bounded by an actin cortex over which the nuclei can move. Also shown in this
sketch are the small number of nuclei that reside inside the yolk, and the also small number of somatic
cells that already form in cycle 10 at the posterior end (the pole cells that divide out of sync with the
rest of the embryo). See Foe and Alberts [1] for sketches for each of the first 14 cycles and Schejter and
Wieschaus [60] for a review on the cytoskeletal elements in the early embryo.
In addition to the data shown in figure 2 of the main text, we also measured the duration of each of
the cycles (figure 7a). The numbers we found are consistent with those reported by Foe and Alberts [1]
and Parry et al. [20]. Averaging over the embryos in each set, we find that the cycle duration can be
reasonably approximated by a quadratic dependence on the cycle number (figure 7b).
17
Figure 6. Average speed of each of the two sets of data, on a log-linear plot, fitted by an exponential
v ∼ 2−ε(n−n0), ε = 0.5 ± 0.05. The black dots correspond to the mean wavefront speeds of set 1, and the
gray ones to the mean speeds of set 2.
Data set 1
cycle number
nuclear spacing (µm)
wavefront speed (µm/s)
cycle duration (s)
mitosis duration (s)
10
23.4 ± 0.8
2.9 ± 0.9
600 ± 60
237 ± 8
11
18.2 ± 0.6
2.2 ± 0.9
763 ± 97
231 ± 9
12
13.2 ± 0.3
1.5 ± 0.4
922 ± 96
233 ± 12
13
9.7 ± 0.2
1.0 ± 0.2
1365 ± 100
240 ± 12
cycle number
nuclear spacing (µm)
wavefront speed (µm/s)
cycle duration (s)
mitosis duration (s)
11
Data set 2
18.0 ± 1.2
4.2 ± 0.4
574 ± 139
197 ± 32
12
13.5 ± 0.3
2.9 ± 0.3
757 ± 150
194 ± 26
13
10.0 ± 0.4
2.0 ± 0.4
1077 ± 160
194 ± 24
Table 1. Experimental data averaged over the data sets. Data sets 1 and 2 correspond to two different
sets of measurements, taken on different days. They correspond to respectively the closed and open
symbols in figure 2 of the main text and figure 7a.
101112131.05.02.03.01.5cyclevwavefront (μm/s) 18
Figure 7. Duration of the measured cycles. a) Experimental data. The different symbols and colors
correspond to the ones in figure 2 in the main manuscript. b) Cycle duration averaged over all
experimentally observed embryos (black and gray dots for sets 1 and 2 respectively). The cycle
durations can be fitted reasonably well by a weak exponential tn = t0e0.29·n, where t0 = 33s (set 1) and
t0 = 25s (set 2).
111213cycle200400600800100012001400cycle duration (s)0cycle duration (s)ba10111213200400600800100012001400cycle0 19
D Timing models
Consider the following timing mechanism for generating a wavefront in a row of people. Assume each
person has a (synchronized) watch, and each is told to raise his/her arms at exactly τ seconds, according
to their individual clocks, after an arbitrarily chosen t = 0. If all of the clocks run at the same rate, there
is no wave, but if the clock of each successive person in the line runs more slowly that that of his/her
neighbor to the right, then a wavefront will be generated.
For the timing method to be the cause for the wavefronts observed in our system, each nucleus would
require a clock. That clock could simply be the amount of time it takes to duplicate the DNA, i.e., the
duration of the interphase, which changes from one cycle to the next. However, there is no correlation
between a nucleus' position and the duration of interphase, which means that we would not expect a
mitotic wavefront to emerge in this case. Alternatively, it is well-established that there are several proteins
which exhibit patterning along the anterior-posterior or dorsal-ventral axes of the embryo. A much studied
example is Bicoid, which exhibits an exponential profile along the anterior-posterior axis [52,59], the same
axis along which the wavefront travels. Now if the duration of interphase were affected by the local Bicoid
concentration, that could provide a mechanism for the clocks of the nuclei to get out of sync, and produce
a wavefront in the various markers for mitosis (such as our observed metaphase and anaphase wavefronts).
There are three reasons why the model outlined above cannot explain our data. The first is specific to
Bicoid. As observed by Gregor et al. [52], the total amount of Bicoid steadily increases over time, as more
of the protein is translated in each cycle. In particular, the amount of protein keeps steady pace with
the number of nuclei, such that at the start of each cycle, the actual amount of protein in each nucleus
at a given position in the embryo is always the same. Therefore if Bicoid were responsible for causing
the mitotic wavefronts, the wavefronts would have the same speed in each cycle, which they do not. The
second reason is more general: as we show below, in order to obtain a linear wavefront propagation from
an exponential concentration profile, the actual absolute amount of material does not matter, only the
decay length - which means that once again the predicted wavefront speed would be independent of cycle.
Finally note that to get wavefronts traveling in both directions along the anterior-posterior axis of the
embryo, we would need at least two concentration gradients of different proteins, for which it would be
highly surprising if they produced wavefronts with the same speed. The timing method therefore cannot
describe our data.
We now assume that the nuclei do not signal each other, in any way, that it is time to start mitosis;
instead, they sample their local environment for a given protein, such as Bicoid (Bcd), and have the
length of their cycle depend on the local concentration. Because the concentration of many such proteins
does indeed exhibit a gradient from one of the two poles, this could explain how mitosis starts close to
the poles, and then seems to 'travel' along the embryo, where in reality there is no traveling front at all.
Roughly speaking, there are four kinds of patterns expressed by proteins in the early Drosophila
embryo: an exponential profile along the AP axis starting from one of the poles (of which Bcd is an
example [52, 59]), an exponential profile along the DV axis (such as Dorsal [61]), a terminal morphogen
profile which is high at both poles and both low and flat in between (as for the phosphorelation gradient
of MAPK [61]), and a striped pattern (for e.g. Hunchback, Giant, Paired and Runt [36]). Because
the observed mitotic waves start at the poles and then spread along the AP axis of the embryo, the
striped patterns and the DV-axis gradients cannot be the ones causing them. There are often two mitotic
waves, which start at the two opposite poles, which suggest that the terminal morphogens might be
good candidates, but they hardly show any profile in the mid-60% of the embryo [61]. The most likely
candidates are therefore the proteins that have an exponential profile along the AP axis, although in this
case there must be at least two proteins that can trigger mitosis. This latter observation is a first weak
point of the timing model, but not necessarily cause to rule it out.
During interphase, the nuclei in the syncytial embryo are surrounded by a nuclear membrane (also
known as the nuclear envelope). One of the things the nucleus can do is to concentrate proteins inside
that membrane. This has been observed directly for Bcd by Gregor et al. [52] and confirmed by our own
20
observations. Irrespective of whether the proteins are concentrated in the nuclei during interphase, or
they are distributed throughout the entire cytosol, they always exhibit an exponential decay along the
anterior-posterior axis. As observed by Gregor et al. [52], the total amount of Bcd steadily increases over
time, as more of the protein is translated in each cycle. In particular, the amount of protein keeps steady
pace with the number of nuclei, such that at the start of each cycle, the actual amount of protein in each
nucleus at a given position in the embryo is always the same [52].
Let us denote the position along the long axis by x, and the total length of the axis by L. The local
concentration at x is then given by c(x) = c0e−x/λ, where λ is the characteristic length scale of the
exponential profile. As stated above, the experimental results of Gregor et al. tell us that λ is the same
in all cycles, whereas c0 goes up [52]. Assuming that all nuclei are equally good at collecting material
from their environment, the amount of material collected by a single nucleus in a simple one-dimensional
model of the embryo is given by
C(x, N ) =
c(y) dy = 2c0λ sinh
e−x/λ,
(8)
(cid:90) x+L/2N
x−L/2N
(cid:18) L
(cid:19)
2N λ
where N is the number of nuclei. As stated above, the key assumption of the timing model is that the
duration of the cycle of each nucleus depends somehow on the concentration, or rather, on the amount
of material in the nucleus, so we have
∆tcycle = f (C(x, N )) = f
2c0λ sinh
.
(9)
(cid:18)
(cid:18) L
(cid:19)
2N λ
(cid:19)
e−x/λ
Unfortunately, we do not know what the function f in equation (9) is. The only thing we do know
is that it is monotonously increasing with its argument (the total amount of material in a nucleus). We
will therefore explore two explicit possibilities:
• The simplest possible dependence: f is a linear function.
• The dependence that gives the observed behavior of (effective) wavefronts, i.e., that the result-
ing 'speed' v of mitosis events through the embryo is well-defined and constant throughout, and
∆tcycle = x/v.
For the first option, we write f (α) = t0 − τ α, with t0 some offset time and τ a timescale. We can
then calculate the speed of an observed mitotic wavefront, as a function of the number of nuclei N , by
calculating the time difference between two positions x and y in the embryo:
v(x, y, N ) =
=
y − x
∆tcycle(y, N ) − ∆tcycle(x, N )
2c0λτ sinh(cid:0) L
2N λ
y − x
(cid:1)(cid:0)e−x/λ − e−y/λ(cid:1) .
(10)
Equation (10) shows that v depends on the position, so there is no well-defined wavespeed in this model.
This is of course no big surprise - we just took a random functional dependence for f , so there should
be no reason to expect it would produce a wavespeed that is position-independent. However, this does
illustrate the point that a constant wavespeed is something special: we need to specifically choose f such
that a constant speed comes out.
Note that it may of course be that there is no constant wavespeed, but that it only appears to be
constant within our error bars. Although we can not rule this option out, this also does not come
out naturally. For instance, inserting numbers for Drosophila from Gregor et al. [36] (L = 450µm,
λ = 70µm = 0.15L, N = 50) we find that for x = 0, the measured speed more than doubles as we take
the measuring point y across the embryo, which can certainly not be confused with a constant speed.
21
To get a constant speed, we need to choose a different function f , specifically a logarithm: f (α) =
t0 + τ log α. In this case we find:
v(x, y, N ) =
y − x
(τ /λ)(y − x)
= λ/τ.
(11)
In this case, we do indeed find a constant value of v across the embryo. However, we also find that v is
independent of N . v does depend on the decay length λ, but the value of λ does not change [52]. The
wavefront speed predicted by this model is therefore the same for all cycles, in direct contradiction to our
observations.
E Diffusion model
The process of diffusion is governed by the diffusion equation, here given for a concentration field c((cid:126)x, t):
∂c((cid:126)x, t)
∂t
= D∇2c((cid:126)x, t),
(12)
where D is the diffusion constant. Equation (12) is linear, so we can use the superposition principle:
the sum of any two solutions is itself a solution. The general solution for a system with no boundaries
depends only on the initial condition c((cid:126)x, 0), and is given by
(cid:90)
c((cid:126)x, t) =
G((cid:126)x, (cid:126)y, t)c((cid:126)y, 0) d(cid:126)y,
(13)
where G((cid:126)x, (cid:126)y, t) is the Green's function of the diffusion equation, which for a system in n dimensions is
given by
G((cid:126)x, (cid:126)y, t) =
1
(4πDt)n/2
exp
.
(14)
(cid:18)
−(cid:126)x − (cid:126)y2
4Dt
(cid:19)
The Green's function describes the concentration field at (cid:126)x at time t due to a single delta-function
concentration source at (cid:126)y at time 0.
F Mechanical model
As described in the main text, we can describe the medium in which the nuclei live as an overdamped
elastic medium. Motion in this medium can be described by some displacement vector (cid:126)u from a fixed
reference position. We get force balance by equating the damping forces acting on the nuclei (due to
friction with the cortical actin layer surrounding the yolk or the outer membrane, and drag due to the
viscous fluids the nuclei and their surrounding microtubule baskets are immersed in) to the elastic forces
in the polymer cytoskeleton:
Γ∂tui =
E
2(1 + ν)
∂j∂jui +
E
2(1 − ν)
∂i∂juj.
(15)
As also pointed out in the main text, equation (15) is reminiscent of the diffusion equation: a time
derivative on the left equals second-order space derivatives plus a source term on the right, and it comes as
no surprise that the solution depends on a quantity with the dimension of a diffusion constant D = µ/Γ,
where µ is the material's shear modulus and equals E/(2 + 2ν). Moreover, equation (15) also allows
for a Green's function type solution, but here in the form of a tensor Gijk((cid:126)x, t), relating an arbitrary
input Qijδ((cid:126)x)Θ(t) to a resulting displacement vector uk((cid:126)x, t) [33]. In two dimensions, the input tensor
22
Qij has three independent components, and can be decomposed in a (hydrostatic) expansion/contraction,
and two symmetric traceless parts:
(cid:19)
(cid:18) 1 0
0 1
(cid:18) 1
(cid:19)
0
0 −1
+ Q2
(cid:19)
(cid:18) 0
1
Q = Q0
+ Q1
.
(16)
1
0
(cid:19)
The two symmetric traceless parts can be converted into each other by a rotation over π/4. They can
therefore alternatively be written as a single term with a prefactor and an angle, as done in equation (3)
of the main text:
(cid:19)
(cid:18) cos 2θ
Q = Q0
0
1
− Q1
sin 2θ
sin 2θ − cos 2θ
.
(17)
(cid:18) 1
0
The second term in equation (17) now represents a volume-conserving force dipole that is oriented at
an angle θ with respect to the x-axis. An example displacement field due to a single force dipole at the
origin is given in figure 8.
Figure 8. Example of a displacement field due to a single force dipole located at the origin and having
an angle of π/6 with respect to the x-axis, obtained by taking the t → ∞ limit of Gijk((cid:126)x, t)Qij.
G Wavefronts
Now that we know the solutions to the chemical and mechanical diffusion equations due to a single
source, we can exploit the fact that the chemical and stress diffusion equations (12 and 15) are linear to
compute the behavior of a system with many sources using the superposition principle. For simplicity, we
pre-arrange the nuclei on a triangular grid, with a little noise in the position of each nucleus to prevent
artifacts due to a perfect arrangement. This is consistent with our observation that just before the mitosis
waves the nuclei in an actual embryo have a rather high degree of triangular order, except where there
are defects due to the fact that a nucleus did not divide in an earlier cycle. Alternatively, we can also
!4!2024!4!2024 consider packings with short-range correlations but no long-range order (like the packing of soft repulsive
spheres), which gives the same results [33].
G.1 Wavefronts in the biochemical diffusion model
23
We will describe the release by a nucleus of a biochemical with a Dirac delta function source located at
the position of the nucleus. Because integration is a linear operation, adding two sources, even if they
divide at different times, is trivial - we simply carry out the integration in equation (13) for each nucleus
that has already divided with the time properly offset, then sum over these nuclei. The only problem
is to determine when each nucleus is supposed to divide. To find out, we perform what is essentially a
numerical integration over time. We start with a release of material at the origin, which we model by
having a delta function concentration there at t = 0. By construction, the concentration field is then
given by G((cid:126)x, 0, t) as long as no other sources have released their chemicals into the system. We proceed
in small timesteps ∆t, calculating for each timestep the concentration at the location of each of the
nuclei that have not yet divided, given the total concentration field generated by the nuclei that have
divided so far. Suppose there is a total number of N nuclei, M of which have already released their
chemicals. The ith source is located at (cid:126)x = (cid:126)ai, released its chemicals at t = ti, and the ti's are ordered.
For tM < t < tM +1 we then find by using the superposition principle:
c((cid:126)x, t) =
,
(18)
M(cid:88)
i=1
Q
4πD(t − ti)
exp
−(cid:126)x − (cid:126)ai2
4Dt
(cid:18)
(cid:19)
where we have taken the number of dimensions to be two, and Q is the number of chemicals released
by a single nucleus. From c((cid:126)x, t) we can determine when the next source will release its chemicals, by
solving c((cid:126)aj, t) = α for M + 1 ≤ j ≤ N . We thus check all nuclei that have not yet released anything,
and determine which one will be the next source by finding the one with the smallest value of t, which
sets tM +1. If there is a nonzero delay time between activation and release, we simply add it to the found
value of tM +1. Given the positions of the sources, the system thus has three parameters: the diffusion
constant D, the release concentration α, and the delay time tdelay. The 'release-wave' is then the time
at which a given source releases its chemicals versus its distance to the origin (i.e., the first source). An
example is given in figure 3a of the main text, which reveals a clear wavefront with a well-defined wave
speed.
G.1.1 Algorithm for finding wavefronts
In summary, we use the following algorithm to numerically find the wavefronts within the biochemical
diffusion model (and, with the proper adaptations, for the stress diffusion model as well):
• Generate a grid of hexagonally arranged nuclei with some small positional noise, centered at the
origin.
• Start with a delta function concentration at the origin at t = 0.
• Increase time in steps of ∆t. For each timestep, calculate the local concentration at each of the
sites of the nuclei, due to the nuclei that have released chemical so far. If one of these exceeds the
critical concentration α, add a delta function concentration peak at this location and time.
• Stop after either all nuclei have divided or a predefined time has been reached.
G.1.2 Analysis of the steady-state wavefront
24
In the case without delay time, it is fairly easy to determine the speed of the wavefront for the regime
in which the wavefront is well-established, i.e., when its curvature is small. Suppose the lattice spacing
is a, the time it takes to get from one row to the next is t, and the amount of material released by each
nucleus is Q. The speed, in a triangular lattice, is then v = a · 1
3/t, because the spacing between the
two rows is a· 1
3. To find the time, it turns out to be sufficient to only consider the 2 nearest neighbors
in the previous row. We choose coordinates such that the neighbors are located at (± a
√
2 , 0), and our next
nucleus is at (0, a
3). Then the time at which this next nucleus is activated is given by the solution of
2
√
√
2
2
(cid:20)
(cid:21)
α =
2Q
4πDt
exp
− a2
4Dt
.
(19)
The results obtained using equation (19) are almost identical to those obtained by numerical solution of
the full equations. If necessary, corrections could of course be made by including additional sources.
To analyze equation (19) further, we introduce dimensionless variables ¯α = a2α/Q and τ = Dt/a2.
The equation then becomes
¯α =
e−1/4τ ,
1
2πτ
(20)
which can of course not be solved analytically, but is easy to solve numerically. We note that the right
hand side of (20) has a maximum value of ¯αmax = 2/πe at τ = 1/4, giving the condition that there can
only be a wave if α < 2Q/πea2. If we write the inverse of (20) as τ = f (¯α), we can write for the wavefront
speed
(21)
so v ∝ D if both a and α are fixed, but unfortunately the scaling of the speed with a and α is hidden in
f (¯α). Based on the numerical determination of f (¯α) we can capture its features fairly accurately with
the following function
f (¯α)
v =
3
,
(cid:19)1/m(cid:35)
f (¯α) ≈ b1 ¯α1/n
− b2
1
4
− ¯α
,
(22)
√
D
a
1
2
(cid:34)
1
(cid:18) 8
e
where fitting gives b1 = 1.38, b2 = 0.28, n = 6.3 and m = 3.6 (see figure 9). Our numerical solutions
of the full equations show that the data do indeed collapse onto the curve described by equations (21)
and (22) for different diffusion coefficients and nuclear spacings (figure 10). In particular, we find that
the wavefront speed v always increases as the nuclear spacing a decreases, so v increases with increasing
cycle number.
G.2 Wavefronts in the biochemical diffusion model with delay time
As indicated in Section G.1, we can include a delay time tdelay into our diffusion model. Now a nucleus
divides (i.e., releases its chemicals) a time tdelay after the local concentration first reaches the threshold
value α. To first approximation, the total time between the activation of a nucleus in a given row and one
in the next row is simply the sum of the delay time and the travel time, which we found in the previous
section. We therefore find:
√
Da
3
v =
2Dtdelay + 2a2f (a2α/Q)
.
(23)
Equation 23 breaks down for long delay times, or small thresholds, as in those cases the effects of earlier
releases become important. However, for our system these effects are small. Figure 11 shows the wavefront
speed v as a function of cycle number for four different values of the delay time, and for both the cases
that α and ¯α are constant. Note that for constant α there is a lower limit to the cycle number below which
25
Figure 9. Plot of f (¯α), defined in equation (21), determined numerically (black dots) and fitted with
equation (22) (red line).
the model predicts no wavefronts (as the activation threshold is not reached). Note also that although
the introduction of the delay time causes the increasing trend in the wavefront speed to chance in the
later cycles, it will still increase in earlier cycles.
G.2.1 Fitting the data with time delay
Naturally, the more parameters we have, the easier it is to fit any set of experimental data points. In
the diffusion model with delay time we have four parameters: the diffusion constant D, the grid size a,
the threshold value α and the delay time tdelay. The grid size (i.e. the spacing between the nuclei) is
measured independently, leaving us with three parameters which we can vary. Reasonable values for
the diffusion constant for a small chemical in the early Drosophila embryo are D = 5 − 100 µm2/s, as
measured by Gregor et al. [36]. As indicated in the main text, we cannot fit the experimental trend (a
wavefront speed that decreases exponentially with cycle number) with fixed values for tdelay and α for
any value of D within this range (see figure 3b of the main document). The only way we can thus fit the
experimental data within this model is if either (or both) of α and tdelay change with the cycle number.
We have systematically investigated a number of options, changing α or tdelay with cycle number.
Some results are given in table 2. We did not find any result that fit the data in which the numbers
change in a well-defined way (e.g.
the delay time increasing linearly or exponentially with the cycle
number). Moreover, in the case of variable delay time, we find that in the first cycle (cycle 10), the
interaction is between nearest-neighbors as in the model without time delay, but the interaction range
goes up every cycle, up to 5 rows apart in cycle 13. In the case of variable concentration threshold, we
need a change of at least an order of magnitude in each cycle to fit the experimental data. Even in
the case where we allow both variables to change, we keep finding at least one of these two problems.
Even though we cannot strictly rule out the diffusion model with variable time delay and concentration
threshold, these results make it very unlikely that this model is actually correct.
f (α)0.050.100.150.200.250.050.100.150.200.25α 26
Figure 10. Dimensionless wavefront speed ¯v = v/(D/a) as a function of the dimensionless
threshold ¯α = a2α/Q in the biochemical signaling model. The figure shows ¯v for several choices of the
parameters D and a (different symbols). The data all collapse on a single curve, as described by
equation (21). The dashed line is the fit to that curve of equation (22).
G.3 Wavefronts in the mechanical model
The analysis leading to wavefront propagation in the mechanical model is described in Ref. [33]. As in
the case of diffusion, we must set a threshold to determine when a source (a nucleus) is activated. The
simplest option is to look at the eigenvalues of the stress tensor: if the largest of those (taking absolute
values) exceeds a certain threshold α, the nucleus is activated. An activated nucleus adds an additional
force dipole term to the stress field in the system, which in turn of course affects the displacement field,
as described by the Green tensor solution of equation (15). Note that we assume linear elasticity, so that
the strain is linear in the displacement, and the stress is linear in the strain. The superposition principle
therefore holds not just for the displacements but for the strain and stress fields as well.
The implementation of the stress-mediated signaling model follows the same pattern as that of the
chemical-diffusion-mediated signaling model, with the concentration c replaced by the stress tensor σij.
An example implementation on a grid of 21 × 21 nuclei is shown in figure 12. The figure shows a clear
wavefront which has a well-defined speed.
As in the diffusion model, we analyze our mechanical model in terms of dimensionless parameters.
There is only one quantity in our model that has the dimensions of a speed, namely µ/aΓ, which means
that the resulting wavefront speed has to scale linearly with this factor, as indeed it does. We again
define a dimensionless wavefront speed ¯v = aΓ
D/a and a dimensionless stress threshold ¯α = a2α/Q,
where Q is now the strength of the force dipole. We can then write
µ v = v
v =
µ
aΓ
g(¯α, ν)
(24)
We determine the function g(¯α, ν) by numerically solving the model. As detailed in [33], we find that
it can be well described by the following functional form, derived using a similar argument we used to
arrive at equation (20) for the diffusion model:
g(¯α, ν) = −4
(c1 + c2 ¯α) log(¯α)
1 − ν2
+ c3.
(25)
0.00.20.40.60.8510152025αv 27
Figure 11. Log-linear plots of the wavefront speed as a function of cycle number in the diffusion model
with time delay. Top: fixed threshold α/Q = 0.001µm−2, D = 10µm2/s, and four values of tdelay: 0s
(blue), 2s (red), 5s (gold), and 8s (green). Bottom: same graph for fixed rescaled threshold ¯α = 0.06.
Note that the form in equation (25) differs slightly from that in [33] because of the use of the shear
modulus µ instead of the Young's modulus E in the definition of ¯v, and the extra constant c3, which
is due to the introduction of semi-periodic boundary conditions. We use equations (24) and (25) with
c1 = 4.5, c2 = 1.5 and c3 = 5.4 to fit the experimental data in figure 3f of the main text.
78910111213n1.010.05.02.020.03.01.515.07.0v78910111213n10.05.02.03.01.57.0v 28
References
1. Foe VE, Alberts BM (1983) Studies of nuclear and cytoplasmic behaviour during the five mitotic
cycles that precede gastrulation in Drosophila embryogenesis. J Cell Sci 61: 31-70.
2. Foe VE (1989) Mitotic domains reveal early commitment of cells in Drosophila embryos. Develop-
ment 107: 1-22.
3. Foe VE, Odell G, Edgar BA (1993) Mitosis and morphogenesis in the Drosophila embryo: Point and
Counterpoint. In: Bate M, Martinez-Arias A, editors, Development of Drosophila melanogaster .
Cold Spring Harbor: Cold Spring Harbor Press, pp. 149-300.
4. Satoh N (1977) Metachronous cleavage and initiation of gastrulation in amphibian embryos. De-
velop, Growth and Differ 19: 111 -- 117.
5. Newport J, Kirschner M (1982) A major developmental transition in early Xenopus embryos: I.
Characterization and timing of cellular changes at the midblastula stage. Cell 30: 675-686.
6. Saka Y, Smith JC (2001) Spatial and temporal patterns of cell division during early Xenopus
embryogenesis. Dev Biol 229: 307-318.
7. Kageyama T (1986) Mitotic wave in the yolk syncytial layer of embryos of Oryzias latipes originates
in the amplification of mitotic desynchrony in early blastomeres. Zool Sci 3: 1046.
8. Trinkaus JP (1992) The midblastula transition, the YSL transition and the onset of gastrulation
in Fundulus. Development 116: 75-80.
9. Kane DA, Warga RM, Kimmel CB (1991) Mitotic domains in the early embryo of the zebrafish.
Nature 360: 735-737.
10. Kane DA, Kimmel CB (1993) The zebrafish midblastula transition. Development 119: 447-456.
11. Devreotes P (1989) Dictyostelium discoideum: a model system for cell-cell interactions in develop-
ment. Science 245: 1054-1058.
12. Levine H, Reynolds W (1991) Streaming instability of aggregating slime mold amoebae. Phys Rev
Lett 66: 2400 -- 2403.
13. Lee KJ, Cox EC, Goldstein RE (1996) Competing patterns of signaling activity in Dictyostelium
Discoideum. Phys Rev Lett 76: 1174 -- 1177.
14. Kleckner N, Zickler D, Jones GH, Dekker J, Padmore R, et al. (2004) A mechanical basis for
chromosome function. Proc Natl Acad Sci USA 101: 12592-12597.
15. Alberts B, Johnson A, Lewis J, Raff M, Roberts K, et al. (2008) Molecular biology of the cell. New
York, NY, U.S.A.: Garland Science, 5th edition.
16. Silver RB (1989) Nuclear envelope breakdown and mitosis in sand dollar embryos is inhibited by
microinjection of calcium buffers in a calcium-reversible fashion, and by antagonists of intracellular
Ca2+ channels. Dev Biol 131: 11-26.
17. Silver RB (1990) Calcium and cellular clocks orchestrate cell division. Ann N Y Acad Sci 582:
207-221.
18. Thomas AP, Bird GSJ, Hajn´oczky G, Robb-Gaspers LD, Putney JW (1996) Spatial and temporal
aspects of cellular calcium signaling. FASEB J 10: 1505-1517.
29
19. Allbritton NL, Meyer T (1993) Localized calcium spikes and propagating calcium waves. Cell
Calcium 14: 691-697.
20. Parry H, McDougall A, Whitaker M (2005) Microdomains bounded by endoplasmic reticulum
segregate cell cycle calcium transients in syncytial Drosophila embryos. J Cell Biol 171: 47-59.
21. Parry H, McDougall A, Whitaker M (2006) Endoplasmic reticulum generates calcium signalling
microdomains around the nucleus and spindle in syncytial Drosophila embryos. Biochem Soc Trans
34: 385-388.
22. Jaffe LF (2008) Calcium waves. Phil Trans R Soc B 363: 1311-1316.
23. Whitaker M (2008) Calcium signalling in early embryos. Phil Trans R Soc B 363: 1401-1418.
24. Warn RM, Magrath R, Webb S (1984) Distribution of f-actin during cleavage of the Drosophila
syncytial blastoderm. J Cell Biol 98: 156-162.
25. Karr TL, Alberts BM (1986) Organization of the cytoskeleton in early Drosophila embryos. J Cell
Biol 102: 1494-1509.
26. Sullivan W, Theurkauf WE (1995) The cytoskeleton and morphogenesis of the early Drosophila
embryo. Curr Opin Cell Biol 7: 18-22.
27. Stevenson VA, Kramer J, Kuhn J, Theurkauf WE (2001) Centrosomes and the scrambled protein
coordinate microtubule-independent actin reorganization. Nat Cell Biol 3: 68-75.
28. Kanesaki T, Edwards CM, Schwarz US, Grosshans J (2011) Dynamic ordering of nuclei in syncytial
embryos: a quantitative analysis of the role of cytoskeletal networks. Integr Biol 3: 1112-1119.
29. Schotz EM (2004) In vivo manipulation of Drosophila syncytial blastoderm embryos using optical
tweezers. Diploma thesis, Universitat Konstanz, Konstanz, Germany.
30. Farge E (2003) Mechanical induction of Twist in the Drosophila foregut/stomordeal primordium.
Curr Biol 13:1365-1377.
31. Winfree AT (1972) Spiral waves of chemical activity. Science 175: 634-636.
32. Agladze KI, Krinsky VI (1982) Multi-armed vortices in an active chemical medium. Nature 296:
424-426.
33. Idema T, Liu AJ (2013) Mechanical signaling via nonlinear wavefront propagation in a
mechanically-excitable medium. Phys Rev Lett (under review) / arXiv:1304.3657.
34. Majkut S, Idema T, Swift J, Krieger C, Liu AJ, Discher DE (2013) Heart-specific stiffening in
early embryos parallels matrix and myosin expression to dynamically match cell function. Curr
Biol (under review).
35. Supplementary confocal microscopy imaging movie of a developing Drosophila embryo available at
http://idemalab.tudelft.nl/research/drosophilamovie.html.
36. Gregor T, Bialek W, de Ruyter van Steveninck RR, Tank DW, Wieschaus EF (2005) Diffusion and
scaling during early embryonic pattern formation. Proc Natl Acad Sci USA 102: 18403-18407.
37. Lang I, Scholz M, Peters R (1986) Molecular mobility and nucleocytoplasmic flux in hepatoma
cells. J Cell Biol 102: 1183-1190.
30
38. Chaikin PM, Lubensky TC (1995) Principles of condensed matter physics. Cambridge, U.K.:
Cambridge University Press.
39. Bischofs IB, Safran SA, Schwarz, US (2004) Elastic interactions of active cells with soft materials.
Phys Rev E 69: 021911.
40. Ranft J, Basan M, Elgeti J, Joanny JF, Prost J, Julicher F (2010) Fluidization of tissues by cell
division and apoptosis. Proc Natl Acad Sci USA 107: 20863-20868.
41. Schmidt CF, Barmann M, Isenberg G, Sackmann E (1989) Chain dynamics, mesh size, and diffusive
transport in networks of polymerized actin: a quasielastic light scattering and microfluorescence
study. Macromolecules 22: 3638-3649.
42. Palmer A, Mason TG, Xu J, Kuo SC, Wirtz D (1999) Diffusing wave spectroscopy microrheology
of actin filament networks. Biophys J 76: 1063-1071.
43. Muller O, Gaub HE, Barmann M, Sackmann E (1991) Viscoelastic moduli of sterically and chem-
ically cross-linked actin networks in the dilute to semidilute regime: measurements by oscillating
disk rheometer. Macromolecules 24: 3111-3120.
44. Janmey PA, Hvidt S, Kas J, Lerche D, Maggs A, et al. (1994) The mechanical properties of actin
gels. J Biol Chem 269: 32503-32513.
45. Gardel ML, Valentine MT, Crocker JC, Bausch AR, Weitz DA (2003) Microrheology of entangled
F-actin solutions. Phys Rev Lett 91: 158302.
46. Gardel ML, Shin JH, MacKintosh FC, Mahadevan L, Matsudaira PA, et al. (2004) Scaling of
F-actin network rheology to probe single filament elasticity and dynamics. Phys Rev Lett 93:
188102.
47. Yamada S, Wirtz D, Kuo SC (2000) Mechanics of living cells measured by laser tracking microrhe-
ology. Biophys J 78: 17361747.
48. Fabry B, Maksym G, Butler J, Glogauer M, Navajas D, Fredberg J (2001) Scaling the microrheology
of living cells. Phys Rev Lett 87: 148102.
49. Guigas G, Kalla C, Weiss M (2007). Probing the nanoscale viscoelasticity of intracellular fluids in
living cells. Biophys J 93: 316323.
50. Wilhelm C (2008). Out-of-equilibrium microrheology inside living cells. Phys Rev Lett 101: 028101.
51. Wirtz D (2009). Particle-tracking microrheology of living cells: principles and applications. Annu
Rev Biophys 38: 301326.
52. Gregor T, Wieschaus EF, McGregor AP, Bialek W, Tank DW (2007) Stability and nuclear dynamics
of the bicoid morphogen gradient. Cell 130: 141-152.
53. Doi M, Edwards SF (1986) The theory of polymer dynamics. Oxford, UK: Oxford University
Press.
54. MacKintosh FC, Kas J, Janmey PA (1995) Elasticity of semiflexible biopolymer networks. Phys
Rev Lett 75: 4425 -- 4428.
55. Keating TJ, Cork RJ, Robinson KR (1994) Intracellular free calcium oscillations in normal and
cleavage-blocked embryos and artificially activated eggs of Xenopus laevis. J Cell Sci 107: 2229-
2237.
31
56. Tyson JJ, Novak B (2011) Cell cycle: who turns the crank? Curr Biol 21: R185-R187.
57. Cao J, Crest J, Fasulo B, Sullivan W (2010) Cortical actin dynamics facilitate early-stage centro-
some separation. Curr Biol 20: 770 - 776.
58. Idema T (2013) A new way of tracking motion, shape and divisions. Eur Biophys J 42: 647-654.
59. Driever W, Nusselein-Volhard C (1988) A gradient of bicoid protein in Drosophila embryos. Cell
54: 83-93.
60. Schejter ED, Wieschaus E (1993) Functional elements of the cytoskeleton in the early Drosophila
embryo. Annu Rev Cell Biol 9: 67-99.
61. Sample C, Shvartsman SY (2010) Multiscale modeling of diffusion in the early Drosophila embryo.
Proc Natl Acad Sci USA 107: 10092-10096.
32
cycle
nuclear spacing wavefront speed
number
10
11
12
13
(µm)
23.4 ± 0.8
18.2 ± 0.6
13.2 ± 0.3
9.7 ± 0.2
(µm/s)
2.9 ± 0.9
2.2 ± 0.9
1.5 ± 0.4
1.0 ± 0.2
tdelay (s) α/Q (10−4/µm2)
(α fixed)
(tdelay fixed)
(tdelay, α/Q)
(s, 10−4/µm2)
1.5
5.7
16.0
36.1
0.0044
0.29
6.9
31
(5, 0.254)
(10, 0.300)
(20, 0.685)
(40, 1.057)
Table 2. Experimentally determined values of the nuclear spacing and wavefront speed (set 1), and
numerically determined values of the required delay time tdelay and threshold concentration α to fit the
experimental data. In column four, D = 15 µm2/s and α/Q = 5 · 10−4 µm−2; in column 5,
D = 10 µm2/s and tdelay = 10 s; in column 6, D = 10 µm2/s and tdelay is assumed to double in each
cycle. None of the columns show a systematic dependency of the parameters on the cycle number,
making it impossible to assign predictive power to the numbers found, or to find a model to explain the
dependencies. Note also that for the case of fixed delay time (column 5), we need to assume that the
threshold value goes up by at least an order of magnitude in each cycle. In the case of variable delay
time (columns 4 and 6), the interactions become very long-ranged in the later cycles, with nuclei up to
5 rows apart triggering each other in cycle 13, even though in cycle 10 the interaction only involves
nearest neighbors.
33
Figure 12. Wavefront from a simulation with 21 × 21 nuclei. The nuclei are arranged on a hexagonal
grid with random small offsets. The wave starts at the center point which generates a stress dipole of
unit strength along the x-axis at t = 0. Whenever the absolute value of the largest eigenvalue of the
stress tensor at another nucleus exceeds the threshold value α, it also divides, adding a unit stress
dipole in a random direction to the total stress field. a) Distance of the dividing nuclei to the center vs.
their activation time, with a linear fit. b) Graphical representation of the 2 dimensional field, with the
dipoles indicated in the orientation in which they divide, and color-coded according to the time that
they divide, on a hue scale (red-yellow-green-blue-violet).
0.51.01.52.02.53.03.5510152025rtimeab |
1905.02474 | 1 | 1905 | 2019-05-07T11:14:07 | Applying allometric scaling to predator-prey systems | [
"physics.bio-ph",
"q-bio.PE"
] | In population dynamics, mathematical models often contain too many parameters to be easily testable. A way to reliably estimate parameters for a broad range of systems would help us obtain clearer predictions from theory. In this paper, we examine how the allometric scaling of a number of biological quantities with animal mass may be useful to parameterise population dynamical models. Using this allometric scaling, we make predictions about the ratio of prey to predators in real ecosystems, and we attempt to estimate the length of animal population cycles as a function of mass. Our analytical and numerical results turn out to compare reasonably to data from a number of ecosystems. This paves the way for a wider usage of allometric scaling to simplify mathematical models in population dynamics and make testable predictions. | physics.bio-ph | physics |
Applying allometric scaling to predator-prey systems
Niels Bohr Institute, University of Copenhagen, Blegdamsvej 17, 2100 København Ø, Denmark.
Andreas Eilersen∗
Kim Sneppen†
Niels Bohr Institute
(Dated: May 8, 2019)
In population dynamics, mathematical models often contain too many parameters to be easily
testable. A way to reliably estimate parameters for a broad range of systems would help us obtain
clearer predictions from theory. In this paper, we examine how the allometric scaling of a number of
biological quantities with animal mass may be useful to parameterise population dynamical models.
Using this allometric scaling, we make predictions about the ratio of prey to predators in real
ecosystems, and we attempt to estimate the length of animal population cycles as a function of
mass. Our analytical and numerical results turn out to compare reasonably to data from a number
of ecosystems. This paves the way for a wider usage of allometric scaling to simplify mathematical
models in population dynamics and make testable predictions.
PACS numbers: 87.23.Cc, 87.10.Ca
Keywords: allometry, population dynamics, predator, prey, Lotka-Volterra
I.
INTRODUCTION
When modeling the dynamics of ecological communi-
ties, a recurring problem is the difficulty of estimating
model parameters. If we desire to develop a model that
can describe real ecosystems, a common approach is to
add terms and parameters to account for as many real-
world complications as possible. The result is unfortu-
nately that many of the models end up being too com-
plicated to actually make any definitive predictions due
to uncertainties about the often large number of parame-
ters. A model that requires precise measurements of pa-
rameters for every individual system one wishes to study
will of course be interesting for the isolated case, but it
will be difficult to derive more general principles from
it. We believe that a simplified model that makes ap-
proximate but clear predictions might be a more useful
approach. In this paper, we will argue that, by using allo-
metric mass scaling, it is possible to estimate the parame-
ters of the classic Lotka-Volterra predator-prey equations
in such a way that this highly idealised model can be used
to predict the behaviour of actual populations. It is our
hope that we will be able to rewrite all parameters of the
equations in terms of only two quantities: Prey mass and
predator mass. We will also look at the implications of
body size for the period of animal population cycles. By
doing so, we wish to conclusively demonstrate the use-
fulness of allometric mass scaling relations in population
modeling.
The fact that many ecological variables scale allomet-
rically with animal body mass has attracted increasing
attention in recent years. Ginzburg & Colyvan [1] go as
far as to call the allometries fundamental laws of ecology,
∗ [email protected]
† [email protected]
comparing them to Kepler's laws in physics. Peters [2]
compiled a list of variables exhibiting allometric scaling,
which we will make use of in this paper. For example,
generation time and metabolic rate correlate with mass
to powers of (approximately) 1/4 and 3/4, respectively.
It is these relationships that we will exploit to write the
Lotka-Volterra equations in terms of animal body mass.
For a compelling attempt at finding a theoretical foun-
dation for these quarter-power scaling laws, see the work
of West et al. [3].
On the larger ecosystem scale, there are also examples
of allometric scaling. In particular, many animals - most
prominently rodents such as lemmings - exhibit a regu-
lar population cycle. The time elapsed between peaks in
abundance of such animals tends to scale with the aver-
age mass of the animal. Empirically, the scaling relation
is found to be T ∝ m0.26 [4]. We wish to argue for a
theoretical basis of this relationship.
Yodzis & Innes [5] use the mass to parameterise a sys-
tem of equations similar to generalised Lotka-Volterra
equations, with consumer and resource (whether plant
or animal) substituted for predator and prey. Their
model assumes that the predator reproduction will satu-
rate with increasing prey population, giving the predator
a Holling type II or III functional response. Also, they
argue that the strength of the predator-prey interaction
should scale with the ratio of prey mass to predator mass
to some power, and that it should be possible to deter-
mine the coefficients of this scaling law from measurable
biological quantities. With the model in place, they anal-
yse the linear stability of the dynamical system and find
that, for certain predator-prey mass ratios, it will have a
limit cycle with a period T ∝ m1/8
R , where mC is the
consumer (predator) mass and mR is the resource (prey)
mass.
C m1/8
We will here proceed down a similar path, though our
model will be notably simplified and our approach to the
predator-prey functional response will be different. The
original Lotka-Volterra equations on which we will be
basing our model assume that the predation and preda-
tor reproduction rates increase in proportion with prey
population density, a so-called Holling type I functional
response. We here assume that prey population is always
far from the carrying capacity of the ecosystem, resulting
in a prey reproduction rate that is also proportional to
prey population. Instead of trying to determine a biolog-
ically reasonable coefficient for the scaling of interaction
strength with the predator-prey mass ratio, we will let
the coefficient remain unknown. We will determine the
equilibrium populations in terms of this unknown coeffi-
cient. Luckily, it turns out that when we look at the ratio
of the populations, this coefficient cancels out. Thus, our
method still yields useful information.
C m1/8
Finally, we will look at the period of population cy-
cles. The simplest version of the Lotka-Volterra equa-
tions has a non-trivial equilibrium which is a center,
rather than a limit cycle. Here, we likewise find a pe-
riod of T ∝ m1/8
R as mentioned above. In order to
obtain the empirically determined m1/4 relationship with
population cycle length, Yodzis & Innes point out that
one can assume a direct proportionality between preda-
tor (consumer) size and prey (resource) size. While this
relationship may hold in many systems (see e.g.
[6]), it
certainly does not in such cases as the wolf-moose sys-
tem studied by Peterson et al.
[4], and the relationship
is hardly well-defined in systems where the resource is a
plant. Ginzburg & Colyvan [1] even present a critique
of the whole idea of using only the linearisation of the
Lotka-Volterra equations to predict the length of popu-
lation cycles. It would therefore be preferable if we could
derive a relationship between prey mass and cycle period
that is independent of predator mass. This is what we
will attempt to do in the following.
The model put forward here is thus an application
of the basic idea of Yodzis & Innes to a heavily sim-
plified system of equations, without making attempts
at determining the exact interaction strength between
predator and prey directly. It is our hypothesis that even
such a simplified model will still give reasonable order-
of-magnitude predictions about real ecosystems.
2
[7]. Here, x denotes prey, y predator, α is the per-
capita reproduction rate of prey, and δ is the per capita
death rate of predators in the absence of prey. The in-
teraction strengths β and γ are slightly harder to define.
β denotes the risk of each prey being eaten per predator,
and γ represents the increase in predator reproduction
rate per prey. These latter two parameters are of course
more difficult to estimate than the first two, and we will
therefore need to find a way around this obstacle.
As opposed to Yodzis & Innes, we choose to work with
animal abundances rather than biomass densities. We do
this because it is conceptually easier and data are more
readily available for abundances than for biomass densi-
ties for the systems that we wish to study. A complica-
tion arising from this is that when working with abun-
dances, there is a distinction between somatic growth (in-
dividuals growing larger) and reproductive growth, which
would be unimportant if we were to work with biomass
densities. We shall therefore ignore the finer details of an-
imal reproduction and growth, and simply assume that
all growth results in the production of new individu-
als. Furthermore, we assume that the populations are
large enough and reproductive events evenly distributed
enough in time that population growth can be modelled
as continuous rather than discrete.
According to Peters [2] we then have the following em-
pirical relation for reproduction rate:
α =
1
400
m−1/4
x
[day−1]
(3)
In the cited mass scaling relations, all masses are in
kilograms. As the predator-prey pairs we will be examin-
ing here are all mammals, we shall be using the mass scal-
ing relations that apply to mammals. For cold-blooded
animals such as reptiles the relations will be different,
though not radically so.
It should be possible to calculate the death rate of
predators in the absence of prey from the so-called
turnover time. This is defined as the average time it will
take an animal to metabolise its entire energy reserves.
In turn, this can be calculated from the metabolic effect.
Again from [2]
II. PARAMETERISING THE
LOTKA-VOLTERRA EQUATIONS
This implies
tturnover = 19m1/4
y
[day]
(4)
The original Lotka-Volterra predator-prey equations
read as follows:
δ = t−1
turnover =
1
19
m−1/4
y
[day−1]
(5)
dx
dt
dy
dt
= αx − βxy
= γxy − δy
(1)
(2)
The coupling coefficient β we assume to be propor-
tional to predator ingestion rate. We believe this to be
justified, since the more a given predator consumes, the
higher the per-capita risk of being eaten by it should be
for the prey. The predator ingestion rate in terms of
energy scales with mass as
I ∝ m3/4
y
[J · (day · predator)−1]
(6)
[2]. The number of individual prey that a predator
needs to eat to satisfy this energetic demand is inversely
proportional to prey mass, and we therefore write β as
3
will discuss below suggests that the wolves have a much
lower efficiency than we would expect based on the above
(η ≈ 2 %) [10].
In laboratory experiments, a figure of
about η = 10 % is observed [10], and for lack of a bet-
ter estimate, we shall use this so-called ten percent law
in our calculations. Given that we are not going for an
exact description of any one particular interaction, we
believe that it is justified to use this rough estimate.
The relation between mass of consumed prey (mx,c)
β = k ·
m3/4
y
mx
[(day · predator)−1]
(7)
and mass of produced predator (my,p) is now
where k is an unknown proportionality constant.
Knowing the equilibrium population of prey or predator
should make it possible to determine k if this is desired.
Our parameterisation thus deviates notably from
that of Yodzis & Innes, since they assume that the
predator death rate and the interaction strength scale
with the ratio of prey mass to predator mass to the
power of 3/4 (here converted to abundance rather
than biomass, as was originally used). Strictly speak-
ing, the ingestion rate of y predators reflects some
kind of average prey consumption rate at average prey
abundance. What we really need here is the slope
of predator kill rate as a function of prey abun-
dance. Furthermore, the units of the ingestion rate is
[J · (predator · day)−1] ∝ [prey · (predator · day)−1] and
not [(predator · day)−1] as we need it to be for our units
to match. Despite all this, we still believe that the al-
lometric scaling of the ingestion rate is a reasonable ap-
proximate measure of the predator's ability to consume
and therefore of the dependence of consumption rate on
prey abundance. We now only need to find a way around
not knowing the exact proportionality.
The slope of predator kill rate with prey abundance
that really constitutes β depends on a number of factors
(temperature, prey population density, predator satia-
tion etc.
[2]), and it is probably not possible to make
a universal estimate of it. Instead, we let k embody all
these complications and tune it to fit the systems that
we will study. As mentioned above, it fortunately can-
cels out in the final calculation of the prey to predator
population ratio anyway.
The relation between the number of prey eaten and
the number of predators produced can be derived ap-
proximately if we know the ecological efficiency η of the
predator-prey interaction. The ecological efficiency here
refers to the percentage of prey biomass that is converted
into predator biomass. Ecological efficiencies vary con-
siderably depending on the nature of the interaction [8],
and it is therefore difficult to find an estimate that is
both precise and general. For systems with a low preda-
tor to prey mass ratio and positive correlation of biomass
density with body mass, ecological efficiency should be
high (η ≈ 35 %) according to a review by Trebilco et
al.[9], which, however, deals with aquatic ecosystems.
Lindeman's original paper similarly shows an efficiency
that rises with trophic level [8]. On the other hand, a
case study of the Isle Royale wolf-moose system that we
my,p = η · mx,c
(8)
assuming that prey and predator have similar energy
content per unit mass. Rewriting this in terms of num-
bers of individual predators produced (Ny,p) and prey
consumed (Nx,c) we get
Ny,p = my,p/my =
mx,c
my
η =
mxNx,c
my
η
(9)
In the Lotka-Volterra equations, the number of preda-
tors produced per unit time is given by the term
Ny,p = γxy
and the number of prey consumed by the term
Nx,c = βxy
(10)
(11)
Thus, we get the following relation between β and γ :
γ =
mx
my
η · β = k · m−1/4
y
· η [(day · prey)−1]
(12)
We have now written all the parameters of the equa-
tions in terms of the animal body masses alone, with k
from eq. (7) being the only parameter that remains to
be determined. However, we can get around this by fo-
cusing our attention on the equilibrium predator-to-prey
population ratio instead of the absolute populations.
The Lotka-Volterra equations have the non-trivial
equilibrium
(x, y) = (cid:18) δ
γ
,
α
β(cid:19)
(13)
which is neutrally stable. The equilibrium ratio be-
tween prey and predator populations is therefore
x/y =
βδ
αγ
=
21
η (cid:18) my
mx(cid:19)3/4
(14)
This number depends only on the masses. We see that
due to the factor 1/γ this ratio is inversely proportional
to ecological efficiency, so that if our estimated 10 % ef-
ficiency is a factor 2 too great, we will estimate a ratio
that is half the "correct" value.
III. THE PERIOD OF POPULATION CYCLES
y
x m1/8
The Jacobian matrix of the Lotka-Volterra equa-
tions at the non-trivial steady state has the eigenval-
, contrary to the observed T ∝ m1/4
ues (i√αδ, −i√αδ), meaning that for small perturbations
away from equilibrium, the system will oscillate over time
with a period of T = 2π√αδ
. This leads to the afore-
mentioned scaling of population cycle period with mass
T ∝ m1/8
x . A
problem with using linearisation in this case is that the
period thus obtained only applies when oscillations are
relatively small. Population cycles in actual predator-
prey pairs, such as the vole-weasel pair in northern Scan-
dinavia, can involve fluctuations over two orders of mag-
nitude [11]. When solving the equations numerically, we
see that much of the time, the population of prey will be
in a state of slow, exponential recovery, while the preda-
tor population slowly approaches zero. When the prey
population recovers, the predator population quickly ex-
plodes, initiating a swift collapse of the prey population.
The collapse phase observed in real rodent cycles does in-
deed appear to be notably shorter than the growth and
peak phases, and the corresponding predator cycles are
similarly observed to be very sharply peaked [11, 12].
We therefore believe that the dynamics can be realisti-
cally modelled as consisting of a slow exponential growth
phase and a fast collapse phase. Using this two-timescale
assumption, we will try to derive an expression for the
period T of population cycles. Splitting more complex
predator-prey models into slow and fast phases has pre-
viously been done by Rinaldi & Muratori [13].
In the
following, we shall use a similar basic idea, but a differ-
ent mathematical approach and solve for the period T ,
rather than maximal abundance as they did. An illus-
tration of the of the cycle and its fast and slow segments
can be seen in fig. (1). For our derivation, we will use the
maximum and minimum prey density of a cycle, which
should be easily obtainable from observations and avail-
able in the literature.
The slow approach to and subsequent drifting away
from the saddle point at (0, 0) is what takes up the ma-
jority of the orbital period of the system. For this reason,
we will here attempt to derive an approximate relation for
the cycle length by looking at the behaviour around the
saddle point at (0, 0) instead of the center at (cid:16) δ
Although the period of the cycle is mainly determined
by the hyperbolic approach to the saddle point, the os-
cillation still happens around the center equilibrium at
β(cid:17).
γ , α
(cid:16) δ
γ , α
β(cid:17). As can be seen in fig. 1, the time average pop-
ulations are very close to the equilibrium populations at
the center. We therefore do not believe that there is a
contradiction between using the saddle point linearisa-
tion to determine the oscillation period, but determining
population ratios based on center equilibrium values.
Using the linearisation around the center equilibrium,
we obtain a period that is independent of initial con-
4
Trajectory
Mean populations
LV Equilibrium
Timestep
103
]
2
-
m
k
[
e
r
a
H
102
101
100
10-10
10-8
10-6
10-4
10-2
100
102
Lynx [km -2]
FIG. 1. (Colour online) An illustration of the dynamics of the
predator-prey system. This numerical solution is based on pa-
rameters appropriate for the lynx-hare system discussed be-
low. The line shows the trajectory of the system in predator-
prey space, and the circles are all spaced evenly in time, at
a separation of 50 days. The distinction between a fast and
a slow segment of the trajectory can be clearly seen from the
spacing of the circles. Note also that equilibrium abundances
are practically identical to mean abundances, meaning that
we can use the two interchangeably.
ditions, but which does not match observations, as the
assumption that initial conditions are close to the equi-
librium breaks down in the real systems studied here. In-
stead, we assume that the initial conditions are far from
the center equilibrium. For this asymptotic approxima-
tion, the period will depend on initial conditions and the
calculated period matches observations better.
Starting from a population xmin, the prey population
should grow as follows:
x(t) = xmineαt
(15)
When predator population is low, prey population
grows unobstructed. After one period of length T , we
should have the maximal population density
xmax = x(T ) = xmineαT
(16)
The time it will take the population to recover to a
density of xmax now becomes
T =
1
α
ln(cid:18) xmax
xmin(cid:19) = 400·ln(cid:18) xmax
xmin(cid:19)·m1/4
x
[day] (17)
We thus get the m1/4
x -relation found empirically. The
above expression should be valid when the amplitude
of oscillations is very large, so that the period of the
predator-prey cycle is dominated by the slow growth
phase which in the Lotka-Volterra model occurs at low
Hare
Lynx
Hare
Lynx
]
2
-
m
k
[
n
o
i
t
a
u
p
o
p
l
200
100
0
0
)
g
o
l
(
n
o
i
t
a
u
p
o
p
l
1010
100
10-10
0
2000
4000
6000
8000
10000
time [days]
2000
4000
6000
8000
10000
time
5
]
2
-
m
k
[
n
o
i
t
a
u
p
o
p
l
15000
10000
5000
0
0
)
g
o
l
(
n
o
i
t
a
u
p
o
p
l
1020
100
10-20
0
Vole
Weasel
2000
4000
6000
8000
10000
time [days]
Vole
Weasel
2000
4000
6000
8000
10000
time
(a)
(b)
FIG. 2. (Colour online) (a) A numerical simulation of the Lotka-Volterra equations for lynx and hare. xmax ≈ 180 km−2, x0 =
xmin ≈ 8km−2, y0 = 0.3 and k = 1.05 · 10−2. The period is just under 2000 days, or 5.5 years, and the average hare density
is 51 km−2. Average lynx density is 0.059 km−2. The ratio of the averages is 860 hares per lynx. As can be seen from the
logarithmic plot, the predicted predator oscillations are too violent, with extinction of lynx at the cycle minimum. When this
does not actually happen, it may be due to the fact that lynx can survive partially on other prey when hare population is low
[14]. (b) The solution obtained using the masses of voles and weasels. xmax ≈ 104 km−2, x0 = xmin ≈ 102 km−2, y0 = 20, and
k = 2 · 10−4. We still see a cycle somewhat shorter than the observed, with an estimated T ≈ 2.3 years. Again, the predator
oscillation is unrealistically violent. Average vole density is 2100 km−2 and weasel density is 4.6 km−2, giving 460 voles/weasel.
predator abundances. Note, however, that we at no point
have assumed that the population crash should be due
to the influence of a predator. We just assumed that
the crash was fast and did not extend the period length
or influence the exponential growth phase significantly.
The derivation here should therefore be equally valid if a
population crash is caused by e.g. a shortage of food or
an epidemic. Given that in the case of many rodents it
is unclear if it is actually predation that drives the cycle
[15], this is a significant advantage.
Another interesting feature of this expression is the
logarithmic scaling with population maximum-minimum
ratio. Hanski has already hinted at such a scaling rela-
tion for the vole-weasel system [16]. In his 1991 paper, he
shows that ln(cid:16) xmax
xmin(cid:17) correlates with latitude, and that
oscillation period also correlates with latitude. Oscilla-
tion period thus also correlates with the logarithm of the
maximum-minimum ratio.
It is possible that we have
found a theoretical explanation for this correlation.
In the next section, we will demonstrate that eq. (17)
roughly fits the pattern seen in oscillating populations in
nature, although there is a significant deviation between
predicted and observed numbers. For the prey-predator
ratios on the other hand, the parameters derived above
mostly give realistic results.
IV. COMPARING THEORY WITH DATA
The classic example of a system described well by the
Lotka-Volterra equations is the interaction between the
canadian lynx (lynx canadensis) and the snowshoe hare
(lepus americanus). Although there has been some doubt
as to whether the hare population cycle is driven primar-
ily by predation or other factors, there seems to be evi-
dence that changes in hare mortality are mainly due to
predation [23]. The population density of hares oscillates
from around 8 to just under 200 per square kilometer over
the 8-10 years long cycle [24]. The average density of lynx
ranges from 0.03 to 0.3 km−2 [25].
To see how well our model fits with observations, we
plug the average masses of lynx - on average roughly 11
kg [20] and hares - roughly 1.6 kg [26] into the equations
and solve them numerically. We choose initial conditions
corresponding to the density per square kilometer when
hare abundance is lowest (x0 ≈ xmin = 8 and y0 = 0.3 -
due to the phase difference between lynx and hare popu-
lation oscillations, we let lynx population start out high
and hare population start out low). We then tune the pa-
rameter k to obtain the correct ratio between cycle highs
and lows. Initial lynx abundance is taken to be slightly
above minimum, as the predator cycle will lag behind
the prey. The result can be seen in fig. 2 (a). Our
simulation predicts an average prey to average predator
population ratio that is quite close to the observed val-
ues. The period is off by about a third, which, given
the simplifications of the model is not a bad estimate.
The fact that the population collapse takes such a short
time in our simulation contributes to our underestimat-
ing the period. In reality, the collapse takes about 1-2
years [24]. The spiky appearance of the graph is also not
very naturalistic. However, taking increasing predation
from other predators, increasing susceptibility to disease
and other complicating factors that increase with popu-
lation density into account would most likely lead to a
System
Lynx-hare
Vole-weasel
Wolf-moose
Lemming osc.
my [kg]
11 ± 1
mx
1.6 ± 0.1
xmax
180 ± 80
xmin Observed x/y ratio Theoretical x/y Obs. T [days] Theoretical T
8 ± 4
1400 ± 300
0.08 ± 0.01 0.025 ± 0.002 (10 ± 2) · 104 100 ± 50
33 ± 1
350 ± 10
-
-
-
0.064 ± 0.003 1000 ± 200
14 ± 5
40 ± 20
-
42 ± 2
-
1460 ± 0
860 ± 80
600 ± 400
200 ± 200
850 ± 70
510 ± 60
3000 ± 200
1600 ± 200
-
730 ± 90
-
6
TABLE I. Table of the data used and the values calculated, including uncertainties. Numbers are rounded to the highest
uncertain digit [16 -- 22].
more rounded shape of the peaks, similar to the one seen
in actual observations. It turns out that the time average
abundances are fairly close to the predicted equilibrium
abundances in all of our numerical solutions. We shall
therefore use mean abundances and equilibrium abun-
dances interchangeably when validating our results.
We also plug the masses into eqs. (14) and (17). The
theoretical estimates obtained this way and their uncer-
tainties can be seen in table I. For this particular system,
we estimate a period of T ≈ 1400 days. Compared to the
observed period of around 3000 days, the error is about
50 %. As far as order-of-magnitude estimates go, this is
still reasonable. Neglecting the duration of the collapse
phase is probably part of the reason for this error. For
comparison, the cycle period obtained from linearisation
(mxmy)1/8 = 770 days, which is far too
gives us
short. This again underlines the usefulness of approx-
imating the cycle as a series of instantaneous collapse
phases followed by exponential growth phases.
2π√αδ
=
550
Another case where the basic Lotka-Volterra equa-
tion might be useful is the interaction between the vole
(microtus agrestis) and least weasel (mustela nivalis) in
northern Scandinavia, as mentioned above. Although
there still is some doubt about the role of predation in the
cycle here as well, there is evidence that predation plays
at least a significant part. Vole density ranges from 102
to 104
km−2 over a cycle, while weasel density ranges
from 1 to 20 km−2 and is strongly correlated with vole
density at northern latitudes [16]. The cycle is observed
to be about 4 years long in the areas we are interested
in [11]. A numerical solution of the Lotka-Volterra equa-
tions for these parameter values can be seen in fig. 2 (b).
This numerical solution gives us an estimate of the pe-
riod T ≈ 830 days = 2.3 years, and of the prey-predator
ratio of 460 voles per weasel. A comparison between the-
oretical results calculated using the derived expressions
and observations can again be seen in table I.
Large population oscillations are observed in some ro-
dent species even when there is no single obvious predator
feeding on the rodent. One example of this is the north-
ern collared lemming of Greenland (dicrostonyx groen-
landicus) [17]. We of course cannot use such an example
to test our hypothesis about prey-predator population
ratios, but we may still use it to examine the accuracy of
the derived period. The results of our examination can
be seen in the table, and both the estimated period and
the error are similar to those of the vole.
As a final example, we will consider the wolves (canis
lupus) and moose (alces alces americanus) of Isle Royale
in Lake Superior, Michigan. On this island, wolves and
moose coexist in isolation, with very little interference
from other animals. Due to the small size of the island,
animal populations are so small that random events (such
as the introduction of parvovirus to the wolf population
in 1980) will have a large influence on the population,
which seems to fluctuate almost erratically [27]. There-
fore, we cannot determine an observational population
cycle length for this system. However, the average popu-
lations should reflect an equilibrium ratio that should be
predictable from wolf and moose mass. As can be seen
in table I, we obtain an accurate estimate of this ratio.
Based on these cases we may conclude that our ide-
alised model works as an order-of-magnitude estimate of
the behaviour of ecosystems. There is a discrepancy be-
tween the derived period of population oscillations and
what is observed, and this discrepancy cannot be ex-
plained entirely by experimental uncertainty. However,
our results reproduce two patterns observed empirically,
which have not yet been theoretically explained. One
is the apparent scaling of oscillation periods with mass
to the quarter power. Another is the scaling of period
with ln(cid:16) xmax
xmin(cid:17). We will therefore argue that the derived
expression is of interest despite the discrepancy.
V. DISCUSSION
As predicted in the introduction, we have been able to
parameterise the Lotka-Volterra equations using animal
body mass in such a way that they provide fairly accu-
rate predictions of the equilibrium predator-prey popu-
lation ratio. When we also know the amplitude of the
fluctuations of prey population, we obtain analytical es-
timates of the oscillation periods that reproduce the pat-
terns found in nature, albeit with a discrepancy. Notably,
our approximate expression for the cycle period exhibits
the same allometric mass scaling as the one found em-
pirically. Furthermore, it shows a logarithmic scaling of
period with the ratio of maximum to minimum popula-
tions, which is also found in data. The ratios of average
prey population to average predator population found in
our simulations fit relatively well with real-world data.
For the population ratios, the uncertainty of population
counts and animal weights explain the errors in two of
three cases. Our prediction of the amplitude of predator
2000
1800
1600
1400
1200
y
/
x
1000
800
600
400
200
0
0
(a)
Observations
Theory
(21/ )*(my/mx)3/4
= 10 %
= 5 %
= 20 %
1
2
3
4
my/mx
5
6
7
8
(b)
3500
3000
2500
2000
1500
1000
500
]
s
y
a
d
[
T
0
10-2
7
Observations
Theory
1/4
400*ln(xmax/xmin)*mx
2
/
( * ) ; my = 0.08 kg
( * ) ; my = 11 kg
2
/
10-1
mx [kg]
100
FIG. 3. (Colour online) (a) shows the observed and theoretically calculated prey-predator population ratios for the wolf-moose,
vole-weasel, and hare-lynx systems. Here, the full line shows the predicted power law. The dotted line shows the theoretically
calculated prey-predator ratio for half the ecological efficiency used in this paper (5 %), while the dashed line shows the
prey-predator ratio calculated for twice the used ecological efficiency (20 %). (b) shows the observed and calculated periods
of population oscillations for voles, lemmings, and hares. The black line shows the corresponding mass power law where we
xmin (cid:17) = 4. This number is close to the values for the vole and lemming oscillation. The dotted and dashed lines
have set ln(cid:16) xmax
show the (mxmy)1/8 scaling law predicted from linearisation, where my is that of lynx and weasel respectively. We could not
calculate an oscillation period for the moose of Isle Royale, and no single predator is known to cause the lemming oscillation,
so they each only occur in one of the plots. Data points and error bars show the numbers without any rounding.
oscillations, however, is unreasonable in comparison with
observations, possibly because of the assumption that the
predator is entirely dependent on one prey species.
Of course, even though our model was only meant as
a crude estimate, we need to address why we see the dis-
crepancy that we do between theory and observations. In
the case of the prey-predator population ratios, the un-
certain estimate of ecological efficiency is a likely source
of error. The range of ecological efficiencies observed in
the real world is so large that it poses a challenge to this
kind of population dynamical modelling. If our estimated
efficiency is a bit too low, it will explain the discrepancy,
as can be seen in fig. 3.
The period is off by a larger percentage, and it is less
clear what might cause the error. One drastic assumption
that we have made is that it takes no time for animals to
grow to adult size. We have considered whether this de-
lay might explain some of the error. To take the time re-
quired to reach full size into account, we have attempted
a numerical solution of the equations while including a
time delay in predator and prey reproduction. Unfortu-
nately, this does not significantly change the oscillation
period. Another possible source of error is the assump-
tion that collapse is instantaneous.
In reality, it does
take some time, though not as long as the exponential
growth phase. If the duration of the collapse phase also
scales with animal mass, it would help explain why we
consistently underestimate the period by about 50 %.
Finally, our model is a mean-field theory, whereas in
reality, geographical separation does play a role. Maybe
the fact that real predators have to seek out the prey,
and that prey may survive for longer in some locations
than in others may serve to slow the dynamics of real,
geographically extended ecosystems. This, however, is a
subject that we will leave for future studies.
Despite these discrepancies, our work demonstrates
that, by using the many available allometric mass scaling
laws, it is possible to obtain reasonable predictions from
even very simple population dynamical models. This fact
should have wide applications in population dynamics.
Another area where this could be applicable is in epidemi-
ology. The incubation and recovery times of a variety
of diseases with multiple host species have already been
shown to scale with host mass [28], and Dobson [29] has
studied a multi-host disease model parameterised using
mass. A possible further use of the model described here
could be to construct an epidemiological model taking
predation into account. Models of epidemics in predator-
prey systems have been proposed before [30], but they
often contain so many unknown parameters that an ex-
amination of parameter space becomes difficult. Here,
a parameterisation using mass could significantly reduce
the number of free parameters.
Already in 1992, Yodzis & Innes pointed out that the
application of mass scaling relations to population dy-
namics can potentially make it a lot easier to make re-
alistic estimates of the parameters involved. Still, to
our knowledge it is not until now that the predictions of
a mass-parameterised population dynamical model have
been tested against real-world data. The scaling of re-
production rate with animal mass has also provided us
with a possible explanation for the relationship between
population cycle length and mass, at least in systems
where the amplitude is large. This is for example very
much the case for several rodent and lagomorph species.
In conclusion, we find that the allometric mass scaling
laws that apply to a variety of biological quantities could
[1] Lev R. Ginzburg and Mark Colyvan, Ecological orbits:
how planets move and populations grow (Oxford Univer-
sity Press, 2004).
[2] Robert Henry Peters, The ecological implications of body
size (Cambridge University Press, Cambridge, 1983).
[3] Geoffrey B. West, James H. Brown, and Brian J. En-
quist, "A general model for the origin of allometric scal-
ing laws in biology," Science 276, 122 -- 126 (1997).
[4] R.O. Peterson, R.E. Page, and K.M. Dodge, "Wolves,
moose, and the allometery of population cycles," Science
224, 1350 -- 1352 (1984).
[5] P. Yodzis and S. Innes, "Body size and consumer-resource
dynamics," The American Naturalist 139, 1151 -- 1175
(1992).
[6] Ulrich Brose, "Body-mass constraints on foraging be-
haviour determine population and food-web dynamics,"
Functional Ecology 24, 28 -- 34 (2010).
[7] Alfred J. Lotka, "Analytical note on certain rhythmic
relations in organic systems," PNAS 6, 410 -- 415 (1920).
[8] Raymond Lindeman, "The trophic-dynamic aspect of
ecology," Ecology 23 (1942).
[9] R. Trebilco, J. K. Baum, A. K. Salomon, and N. K.
Dulvy, "Ecosystem ecology: size-based constraints on the
pyramids of life," Trends In Ecology & Evolution 28,
423 -- 431 (2013).
[10] P. A. Colinvaux and B. D. Barnett, "Lindeman and the
ecological efficiency of wolves," The American Naturalist
114, 707 -- 718 (1979).
[11] Ilkka Hanski, Heikki Henttonen, Erkki Korpimaki, Lauri
and Peter Turchin, "Smallrodent dynamics
Oksanen,
and predation," Ecology 82, 1505 -- 1520 (2001).
[12] C. J. Krebs, S. Boutin, R. Boonstra, A. R. Sinclair, J. N.
Smith, M. R. Dale, K. Martin, and R. Turkington, "Im-
pact of food and predation on the snowshoe hare cycle,"
Science (New York, N.Y.) 269 (1995).
[13] S. Rinaldi and S. Muratori, "Slow-fast limit cycles in
predator-prey models," Ecological Modelling 61, 287 -- 308
(1992).
[14] Melvin E Sunquist, Wild cats of the world (University of
Chicago Press, Chicago, Ill, 2002).
[15] P. Turchin, L. Oksanen, P. Ekerholm, T. Oksanen, and
H. Henttonen, "Are lemmings prey or predators?" Nature
405 (2000).
[16] I. Hanski, L. Hansson, and H. Henttonen, "Specialist
predators, generalist predators, and the microtine rodent
cycle," The Journal of Animal Ecology 60 (1991).
[17] Olivier Gilg, Ilkka Hanski, and Benoıt Sittler, "Cyclic
dynamics in a simple vertebrate predator-prey commu-
nity," Science (New York, N.Y.) 302 (2003).
[18] Heikki Henttonen, Tarja Oksanen, Aarre Jortikka, and
Voitto Haukisalmi, "How much do weasels shape micro-
8
potentially prove highly useful in population dynamics.
ACKNOWLEDGMENTS
The authors wish to thank Dr. David Vasseur for his
detailed and useful comments on an earlier draft of this
article.
tine cycles in the northern fennoscandian taiga?" Oikos
50, 353 -- 365 (1987).
[19] Michael Carlsen, Jens Lodal, Herwig Leirs,
and
Thomas Secher Jensen, "The effect of predation risk on
body weight in the field vole, microtus agrestis," Oikos
87, 277 -- 285 (1999).
[20] Ron Moen, James M. Rasmussen, Christopher L. Bur-
dett, and Katharine M. Pelican, "Hematology, serum
chemistry, and body mass of free-ranging and captive
canada lynx in minnesota," Journal of Wildlife Diseases
46, 13 -- 22 (2010).
[21] Bruce J. Gillingham, "Meal size and feeding rate in the
least weasel (mustela nivalis)," Journal of Mammalogy
65, 517 -- 519 (1984).
[22] U. S. Seal and L. D. Mech, "Blood indicators of seasonal
metabolic patterns in captive adult gray wolves," The
Journal of Wildlife Management 47, 704 -- 715 (1983).
[23] Charles J. Krebs, Rudy Boonstra, Stan Boutin,
and
A.R.E. Sinclair, "What drives the 10-year cycle of snow-
shoe hares?" BioScience 51, 25 -- 35 (2001).
[24] Charles J. Krebs, "Of lemmings and snowshoe hares: the
ecology of northern canada," Proceedings of the Royal
Society B 278, 481 -- 489 (2011).
[25] Garth Mowat, Mark O ' Donoghue,
and Kim Poole,
"Ecology of lynx in northern canada and alaska," in Ecol-
ogy and conservation of lynx in the United States, edited
by L.F. Ruggiero, K.B. Aubry, S.W. Buskirk, G. M.
Koehler, C. J. Krebs, K.S. McKelvey, and J.R. Squires
(Boulder: University of Colorado Press, 2000) Chap. 9,
pp. 265 -- 306.
[26] Ronald L. Smith, Dennis J. Hubartt,
and Russell L.
Shoemaker, "Seasonal changes in weight, cecal length,
and pancreatic function of snowshoe hares," The Journal
of Wildlife Management 44, 719 -- 724 (1980).
[27] John A. Vucetich and Rolf O. Peterson, "The influence of
prey consumption and demographic stochasticity on pop-
ulation growth rate of isle royale wolves (canis lupus),"
Oikos 107, 309 -- 320 (2004).
[28] Jessica M. Cable, Brian J. Enquist,
and Melanie E.
Moses, "The allometry of host-pathogen interactions (al-
lometry and disease)," PLoS ONE 2 (2007).
[29] A. Dobson, "Population dynamics of pathogens with mul-
tiple host species," The American Naturalist 164, 64 -- 78
(2004).
[30] Ying-Hen Hsieh and Chin-Kuei Hsiao, "Predator-prey
model with disease infection in both populations," Math-
ematical Medicine and Biology: A Journal of the IMA
25, 247 -- 266 (2008).
|
1105.5061 | 1 | 1105 | 2011-05-25T15:18:21 | Broadband Dielectric Spectroscopy on Human Blood | [
"physics.bio-ph",
"cond-mat.soft",
"q-bio.TO"
] | Dielectric spectra of human blood reveal a rich variety of dynamic processes. Achieving a better characterization and understanding of these processes not only is of academic interest but also of high relevance for medical applications as, e.g., the determination of absorption rates of electromagnetic radiation by the human body. The dielectric properties of human blood are studied using broadband dielectric spectroscopy, systematically investigating the dependence on temperature and hematocrit value. By covering a frequency range from 1 Hz to 40 GHz, information on all the typical dispersion regions of biological matter is obtained. We find no evidence for a low-frequency relaxation (alpha-relaxation) caused, e.g., by counterion diffusion effects as reported for some types of biological matter. The analysis of a strong Maxwell-Wagner relaxation arising from the polarization of the cell membranes in the 1-100 MHz region (beta-relaxation) allows for the test of model predictions and the determination of various intrinsic cell properties. In the microwave region beyond 1 GHz, the reorientational motion of water molecules in the blood plasma leads to another relaxation feature (gamma-relaxation). Between beta- and gamma-relaxation, significant dispersion is observed, which, however, can be explained by a superposition of these relaxation processes and is not due to an additional delta-relaxation often found in biological matter. Our measurements provide dielectric data on human blood of so far unsurpassed precision for a broad parameter range. All data are provided in electronic form to serve as basis for the calculation of the absorption rate of electromagnetic radiation and other medical purposes. Moreover, by investigating an exceptionally broad frequency range, valuable new information on the dynamic processes in blood is obtained. | physics.bio-ph | physics | Broadband Dielectric Spectroscopy on Human Blood
M. Wolfa, R. Gulicha, P. Lunkenheimera,∗, A. Loidla
aExperimental Physics V, Center for Electronic Correlations and Magnetism, University of Augsburg, 86135 Augsburg, Germany
1
1
0
2
y
a
M
5
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
1
6
0
5
.
5
0
1
1
:
v
i
X
r
a
Abstract
Background
Dielectric spectra of human blood reveal a rich variety of dynamic processes. Achieving a better characterization and understand-
ing of these processes not only is of academic interest but also of high relevance for medical applications as, e.g., the determination
of absorption rates of electromagnetic radiation by the human body.
Methods
The dielectric properties of human blood are studied using broadband dielectric spectroscopy, systematically investigating the
dependence on temperature and hematocrit value. By covering a frequency range from 1 Hz to 40 GHz, information on all the
typical dispersion regions of biological matter is obtained.
Results and conclusions
We find no evidence for a low-frequency relaxation (”α-relaxation”) caused, e.g., by counterion diffusion effects as reported
for some types of biological matter. The analysis of a strong Maxwell-Wagner relaxation arising from the polarization of the cell
membranes in the 1-100 MHz region (”β-relaxation”) allows for the test of model predictions and the determination of various
intrinsic cell properties.
In the microwave region beyond 1 GHz, the reorientational motion of water molecules in the blood
plasma leads to another relaxation feature (”γ-relaxation”). Between β- and γ-relaxation, significant dispersion is observed, which,
however, can be explained by a superposition of these relaxation processes and is not due to an additional ”δ-relaxation” often
found in biological matter.
General significance
Our measurements provide dielectric data on human blood of so far unsurpassed precision for a broad parameter range. All data
are provided in electronic form to serve as basis for the calculation of the absorption rate of electromagnetic radiation and other
medical purposes. Moreover, by investigating an exceptionally broad frequency range, valuable new information on the dynamic
processes in blood is obtained.
Keywords: dielectric spectroscopy, blood, relaxation, specific absorption rate, dielectric loss, dielectric constant
1. Introduction
Blood is a highly functional body fluid, it delivers oxygen
to the vital parts, it transports nutrients, vitamins, and metabo-
lites and it also is a fundamental part of the immune system.
Therefore the precise knowledge of its constituents, its physi-
cal, biological, and chemical properties and its dynamics is of
great importance. Especially its dielectric parameters are of rel-
evance for various medical applications [1], like cell separation
(e.g., cancer cells from normal blood cells [2]), checking the
deterioration of preserved blood [3], and dielectric coagulom-
etry [4]. In addition, the precise knowledge of the dielectric
properties of blood is prerequisite for fixing limiting values for
∗Corresponding author. Tel.: + 49 821 5983603 FAX + 49 821 5983649
Email address: [email protected] (P.
Lunkenheimer)
electromagnetic pollution (via the conductivity in the specific
absorption rate (SAR))[5, 6, 7, 8].
Early measurements of the electrical properties of blood con-
tributed significantly to unravel the constitution of red blood
cells (RBC). For example, the results by Hober [9] provided
the first indications of a dispersion (i.e. frequency dependence),
caused by the membrane of RBCs, in the radio frequency (RF)
spectrum of the dielectric properties of blood. This relaxation
process is nowadays identified as being of Maxwell-Wagner
type [10, 11] and termed β-relaxation in biophysical literature
[12, 13].
Various early works [14, 15, 16, 17, 18, 19, 20, 21] were fol-
lowed by measurements at very high frequencies [22, 23, 24,
25]. Some of them revealed an additional dispersion with a re-
laxation rate near 18 GHz, which can be assigned to the reori-
entation of water molecules and which is named γ-dispersion.
Furthermore, a third relaxation, termed α-relaxation and lo-
Preprint submitted to BBA - General Subjects
March 12, 2015
cated in the low-frequency regime, ν < 100 kHz was de-
tected in some biological materials [26, 27]. However, inter-
estingly, an α-relaxation seems to be absent in whole blood
[28] and only is found in hemolyzed blood cells [27]. This
was speculated to be due to a higher ion permeability of the
membranes in the latter case, shifting the relaxation spectrum
into the experimental frequency window [27]. The origin of
the α-relaxation is a matter of controversy; most commonly it
is assumed to arise from counterion diffusion effects [13, 29].
Finally, a dispersion with low dipolar strength located in the
frequency regime between the β- and γ-dispersion was identi-
fied by Schwan [12]. The origin of the δ-dispersion and the
possible role of bound water in its generation is controversially
discussed [30, 31, 32, 33, 34, 35, 36, 37, 38].
Taking together all these results, it is clear that there are
three main dispersion regions in the dielectric frequency spec-
trum of blood between some Hz and 50 GHz, termed β, γ, and
δ [12, 13]. This nomenclature should not be confused with
that used in the investigation of glassy matter like supercooled
liquids or polymers. Within the glass-physics community the
terms α-, β-relaxation, etc. are commonly applied to com-
pletely different phenomena than those considered above (see,
e.g., refs 39 and 40). In the present work we follow the bio-
physical nomenclature.
A lot of additional research has been done until the early
1980s [1, 41, 42, 43, 44] and a detailed review was given in
1983 by Schwan [13]. Later on, in the course of the upcoming
debates about electromagnetic pollution, dielectric properties of
body tissues and fluids received renewed interest as they deter-
mine the SAR, a measure for the absorption of electromagnetic
fields by biological tissue [8, 45, 46, 47, 48]. But also various
other important medical questions can be addressed by dielec-
tric spectroscopy [1, 2, 3, 4, 49].
In the last two decades, a
number of papers on dielectric spectroscopy on blood and ery-
throcyte solutions were published [35, 50, 51, 52, 53, 54, 55,
56, 57], most of them treating special aspects only.
On the theoretical side, a number of models for the descrip-
tion of cell suspensions have been proposed. Most models fo-
cus on the β-relaxation [12, 14, 15, 16, 17, 20, 58, 59, 60, 61,
62, 63, 64], including the often employed Pauly-Schwan model
[12, 58], discussed below. Some of them also account for the
non-spherical shape of cells [14, 15, 16, 17, 61]. It seems clear
that diluted solutions and whole blood with a hematocrit value
of 86% have to be treated differently. The Bruggeman-Hanai
model [20, 59] was specially developed for highly concentrated
suspensions. A recent summary of various models can be found
in ref. [61].
Concluding this introduction, it has to be stressed that, af-
ter more than one century of research, many aspects of the
dielectric properties of blood (e.g., the presence and origins
of α- and δ-dispersion) are still unclear.
It should be noted
that, in addition to the three main dispersion effects, from a
theoretical viewpoint a number of further relaxation features
may show up in blood. For example, it is well known that
RBC’s are far from being of spherical shape and in principle
for shelled ellipsoidal particles up to six relaxations can be ex-
pected [61]. Furthermore, the hemoglobin molecules within the
RBC’s should show all the typical complex dynamics as found
in other proteins. Based on the available literature data (e.g.,
[13]), it seems that most, if not all, of these additional pro-
cesses do not or only weakly contribute to the experimentally
observed spectra. However, one should carefully check for pos-
sible deviations from the simple three-relaxation scenario men-
tioned above, which may well be ascribed to these additional
processes.
Maybe the best and most cited broadband spectra of blood
covering several dispersion regions are those by Gabriel et al
[46], taken at 37 C, which are commonly used for SAR cal-
culations and for medical purposes. However, even these data
are hampered by considerable scatter and they are composed
from data collected by different groups on different samples.
Clearly, high-quality spectra covering a broad frequency range
measured on identical samples are missing. A systematic in-
vestigation of the hematocrit and temperature dependence is
essential to achieve a better understanding of the different dis-
persion contributions of blood. The present work provides the
dielectric constant ε′, the loss ε′′, and the conductivity σ′ of
human blood in a broad frequency range (1 Hz to 40 GHz), by
using a combination of different techniques of dielectric spec-
troscopy applied to identical samples. In addition, the tempera-
ture (280 K - 330 K) and hematocrit value (0 - 86%) dependence
is thoroughly investigated.
2. Models and Data Analysis
Dielectric spectroscopy is sensitive to dynamical processes
that involve the reorientation of dipolar entities or displace-
ment of charged entities, which can cause a dispersive behav-
ior of the dielectric constant and loss. However, also non-
intrinsic Maxwell-Wagner effects caused by interfacial polar-
ization in heterogeneous samples can lead to considerable dis-
persion [10, 11, 65]. As mentioned above, biological mat-
ter shows various dispersions in the frequency regime 1 Hz to
40 GHz, which have different microscopic and mesoscopic ori-
gins and therefore have to be described differently.
2.1. Intrinsic Relaxations
Intrinsic processes like, e.g., the cooperative reorientation of
dipolar molecules are often described by the Debye formula
[66]:
ε∗(ν) = ε∞ +
∆ε
1 + i2πντ
(1)
The relaxation strength is given by ∆ε = εs − ε∞. εs and ε∞
are the limiting values of the real part of the dielectric constant
for frequencies well below and above the relaxation frequency
νrelax = 1/(2πτ), respectively. This frequency is characterized
by an inflection point in the frequency dependence of the di-
electric constant and a peak in the dielectric loss. If taking into
account an additional dc-conductivity contribution, the conduc-
tivity shows a steplike increase close to νrelax. The Debye theory
assumes that all entities do relax with the same relaxation time
τ. In reality, a distribution of relaxation times often leads to
2
a considerable smearing out of the spectral features [67, 68].
Those can be described, e.g., by the Havriliak-Negami formula,
which is an empirical extension of the Debye formula by the
additional parameters α and β [69, 70]:
(2)
ε∗(ν) = ε∞ +
∆ε
(cid:2)1 + (i2πντ)1−α(cid:3)β
Special cases of this formula are the Cole-Cole formula [71]
with 0 ≤ α < 1 and β = 1 and the Cole-Davidson formula
[72, 73] with α = 0 and 0 < β ≤ 1.
In most materials with intrinsic relaxations, the inevitable dc
conductivity σdc arising from ionic or electronic charge trans-
port cannot be neglected. Usually it leads to a 1/ν divergence in
the loss at frequencies below the loss peak and can be taken into
account by including a further additive term ε′′
dc = σdc/(ε0ω) in
Eq. 2 (ε0 denotes the permittivity of free space, ω is the circular
frequency).
2.2. Maxwell-Wagner Relaxations
The β-dispersion in biological matter is commonly accepted
to be of Maxwell-Wagner type [12, 13, 44, 47, 48]. As shown
by Maxwell and Wagner [10, 11], strong dispersive effects
mimicking those of intrinsic dipolar relaxations can arise in
samples composed of two or more regions with different elec-
trical properties (e.g., plasma, cytoplasma, and cell membranes
in the case of RBCs). It should be noted that this dispersion
can be completely understood from the heterogeneity of the in-
vestigated samples without invoking any frequency-dependent
microscopic processes within the involved dielectric materials.
If one of the regions in the sample is of interfacial type and rel-
atively insulating, e.g., an insulating surface layer [65] or the
membranes of biological cells [12, 13], very high apparent val-
ues of the dielectric constant are detected at low frequencies.
A straightforward approach for understanding the dielectric be-
havior of heterogeneous systems is an equivalent-circuit anal-
ysis. Here any interfacial layer can be modeled by a parallel
RC element with the resistance R and capacitance C of the in-
terfacial element much higher than the corresponding bulk val-
ues [65]. This leads to a relaxation spectrum where the low-
frequency capacitance and conductance are dominated by the
interface. For the calculation of ε′(ν) from the measured ca-
pacitance, usually the overall geometry of the sample is used
instead of that of the thin interfacial layer (i.e., the assumed C0
is much smaller than that of the layer). Thus, an artificially high
dielectric constant is detected at low frequencies. At high fre-
quencies, the interface capacitor becomes shorted and the bulk
properties are detected. This leads to the steplike decrease of
ε′(ν) and increase of σ′(ν) with increasing frequency, typical
for relaxational behavior. The increase of σ′(ν) arises from the
fact that the bulk conductance usually is much higher than the
interface conductance (the cell membrane in the case of RBCs),
the latter being shorted by the interface capacitor at high fre-
quencies.
Instead of an equivalent-circuit analysis [65], for biological
matter it is common practice to treat the β-relaxation analogous
to an intrinsic relaxation process, i.e., to fit it with the Debye
3
equation or its extensions (Eqs. 1 and 2). A variety of models
have been developed to connect the obtained fitting parameters
with the intrinsic dielectric properties of the different regions
of the samples (see section 2.5) and a lot of modeling work of
experimental data was performed [12, 14, 15, 16, 17, 20, 57,
58, 59, 60, 61, 62, 63, 64, 74, 75, 76, 77, 78, 79].
2.3. Electrode Polarization
Blood exhibits strong ionic conductivity. At low frequencies
the ions arrive at the metallic electrodes and accumulate in thin
layers immediately below the sample surface [80, 81, 82, 83].
The frequency, below which this effect sets in, critically de-
pends on electrode distance and ionic mobility and in biologi-
cal matter typically is located in the kHz - MHz region. These
insulating layers represent large capacitors leading to an appar-
ent increase of ε′(ν) and decrease of σ′(ν) at low frequencies,
quite similar to the Maxwell-Wagner effects discussed above.
These non-intrinsic contributions can hamper the unequivocal
detection of the parameters of the β-relaxation. Various exper-
imental techniques have been applied to avoid the influence of
electrode polarization (see, e.g., refs. 28, 50, 81). An alterna-
tive way is the exact modeling of these non-intrinsic contribu-
tions. The most common models are a parallel RC circuit or
a so-called constant-phase element, both connected in series to
the bulk sample [81, 82, 83].
A parallel RC circuit corresponds to an additional impedance
RC(ν) =
Z ∗
RRC
1 + i2πνRRCCRC
,
(3)
which has to be added to the bulk impedance. RRC and CRC are
the resistance and capacitance of the insulating layers, respec-
tively. From the resulting total impedance, the total capacitance
and conductance (and thus ε′(ν) and σ′(ν)) can be calculated
resulting in a behavior equivalent to a Debye-relaxation (this
scenario corresponds to a conventional Maxwell-Wagner relax-
ation). Alternatively a ”constant phase element”, which is an
empirical impedance function, given by ZCPE = A(iω)−α (refs
82, 83), can be used. When defining τRC = RRCCRC, Eq. 3
formally has the same mathematical structure as Eq. 1. Thus,
in analogy to Eq. 2 a distributed RC-circuit can be introduced
by writing
RC(ν) =
Z ∗
.
(4)
RRC
(cid:2)1 + (i2πντRC)1−α(cid:3)β
We want to emphasize, that in contrast to Eq. 3, which leads to
a frequency dependence identical to that of a Debye relaxation,
an equivalent-circuit evaluation using Eq. 4 does not lead to
fit curves identical to those of a Havriliak-Negami relaxation:
For the latter case, the relaxation time τ in Eq. 1 is assumed to
be distributed. In the equivalent-circuit case, the corresponding
quantity, determining, e.g., the loss peak position, is τ = RbCRC
with Rb the bulk resistance [65]. However, the distributed quan-
tity in the equivalent-circuit case is τRC = RRCCRC, thus leading
to different curve shapes.
2.4. Temperature Dependence
and
The fitting of relaxation spectra directly provides the relax-
ation times (τ), the width parameters (α and β), the relaxation
strengths (∆ε), the dielectric constant for ν −→ ∞ (ε∞), and
the dc conductivity (σdc) (see section 2.1). For the temperature
dependence of τ and σdc, thermally activated behavior
τ = τ0 exp Eτ
kBT!
and
σdc =
σ0
T
exp −
Eσ
kBT! ,
(5)
(6)
can be assumed. σ0 is a prefactor. Eτ and Eσ denote the hinder-
ing barriers for the relaxational process and the diffusion of the
charge carriers (i.e., dissolved ions of the plasma in the present
case), respectively. τ0 is an inverse attempt frequency, often as-
sumed to be of the order of a typical phonon frequency. Equa-
tion 6, with the extra 1/T term, is derived by considering the
difference of the forward and backward hopping probabilities
of an ion between two sites in a potential that becomes asym-
metric due to the external field [84]. Here the field drives the
ionic motion. In contrast, for dielectric relaxations, within the
framework of the fluctuation-dissipation theorem it is assumed
that the dielectric measurement is sensitive to reorientational
fluctuations, which are present even without field. Thus, τ in
Eq. 5 is proportional to the inverse of the reorientation proba-
bility of a molecule experiencing a hindering barrier, which is
just given by the exponential term.
The temperature dependence of the dielectric strength of
dipolar relaxation mechanisms often can be characterized by
the Curie law [85, 86]:
∆ε =
C
T
(7)
Deviations from the Curie law are usually thought to signify
dipole-dipole interactions.
For most dielectric materials, the broadening of the loss peak
diminishes with increasing temperature, i.e., α → 0 and β → 1
for high temperatures. This finding can be ascribed to the fast
thermal fluctuations, which cause every relaxing entity ”seeing”
the same environment [39], thus leading to an identical relax-
ation time for each entity, which implies Debye-like behavior.
2.5. Cell Models
Two commonly employed models for the description of the
β-dispersion of cell suspensions and tissue will be introduced
now and applied to the experimental data in section 4.3.2. Both
models are claimed to be applicable to high cell concentrations
and thus are especially suited for the samples of the present
work. Based on the Maxwell-Wagner model [10, 11], the Pauly
and Schwan takes into account the membranes of cells [58, 48].
Using appropriate approximations (e.g., a negligible membrane
conductance) some simple relations are derived:
∆εβ =
9prCm
4ε0 · (1 + p/2)2
(8)
4
σdcβ =
1 − p
1 + p/2
· σa
(9)
Equation 8 allows the calculation of the membrane capaci-
tance per area unit, Cm, from the relaxation strength of the β-
dispersion, ∆εβ, and the volume fraction p and radius r of the
suspended particles. Via Eq. 9, the conductivity of the sus-
pending medium (plasma in the case of blood), σa, can be de-
termined from the volume fraction and the measured dc con-
ductivity of the suspension, i.e. the limiting low frequency con-
ductivity of the β-dispersion, σdcβ. Moreover, resolving the ex-
pression for the β-relaxation time τβ [58],
τβ = rCm
2σa + σi − p(σi − σa)
σiσa(2 + p)
,
(10)
one can approximate the conductivity of the cell interior (cyto-
plasma), σi by
σi =
σarCm(2 + p)
σaτβ(2 + p) − (1 − p)rCm
.
(11)
An alternative access to σi is provided by the following relation
[58]:
σi =
2σ2
a(1 − p) − σaσ∞β(2 + p)
σ∞β(1 − p) − σa(1 + 2p)
(12)
Here σ∞β denotes the high-frequency plateau of the step in
σ′(ν). Also the dielectric constant of the cytoplasma, εi, can
be determined [58]:
εi =
2ε2
a(1 − p) − εaε∞β(2 + p)
ε∞β(1 − p) − εa(1 + 2p)
(13)
ε∞β is the limiting high-frequency dielectric constant of the β-
relaxation (cf. ε∞ in Eq. 1) and εa is the dielectric constant of
the suspending medium (plasma).
The Pauly-Schwan model includes correction factors for high
concentrations (e.g., the 1 + p/2 factor in eq. 8) and is claimed
to be valid for all values of p (see, e.g., ref. [48]). Another
model, especially developed for highly concentrated suspen-
sions, is the one by Bruggemann [20] and Hanai [59, 62, 63]
taking into account the polarization of particles in the presence
of neighboring ones. The model leads to the equations
εp =
εa(1 − p) − ε∞βk
1 − p − k
,
k = εa
ε∞β!1/3
and
σp =
σa(1 − p) − σdcβk
1 − p − k
,
k = σa
σdcβ!1/3
(14)
(15)
for the dielectric properties εp and σp of the particles. It should
be mentioned that this model assumes homogeneous particles
(i.e. without shell) and Eqs. 14 and 15 can be considered pro-
viding average values of the whole cell only.
One should note that the exact solution of the dielectric the-
ory of suspensions of ellipsoidal particles leads to the prediction
of six separate relaxation processes, namely two per ellipsoid
axis, arising from the Maxwell-Wagner relaxation of the shell
and of the particle interior [61]. For spheroids, i.e. ellipsoids
with two equal semi-diameters, which may be a good approxi-
mation of RBCs, still four relaxations are expected. Usually in
the application of the Maxwell-Wagner model to cell suspen-
sions, including the above treated models, various reasonable
approximations are made (e.g., that the membrane thickness is
much smaller than the cell radius) that lead to the prediction of
a single relaxation only.
3. Materials and methods
To determine the complex dielectric permittivity and con-
ductivity in a broad frequency range (from 1 Hz to 40 GHz),
different measurement techniques were combined [87]. In the
frequency range 1 Hz - 10 MHz, high precision measurements
were performed by means of a Novocontrol Alpha-A Analyzer.
This frequency response analyzer directly measures the sam-
ple voltage and the sample current by the use of lock-in tech-
nique. The ac voltage is applied to a parallel-plate capaci-
tor made of platinum containing the sample material (diame-
ter 5 mm, plate distance 0.6 mm). In our earlier measurements
of various materials, platinum was found to minimize contri-
butions from dissolved ionic impurities arising from the elec-
trode material. The capacitor is mounted into a N2-gas cryo-
stat (Novocontrol Quatro) for temperature-dependent measure-
ments. For the measurements in the frequency range 1 MHz -
3 GHz a coaxial reflection method was used employing the Ag-
ilent Impedance/Material Analyzer E4991A. Here the sample,
again placed in the same parallel-plate capacitor, is connected
to the end of a specially designed coaxial line, thereby bridg-
ing inner and outer conductor [88]. For additional measure-
ments between 40 Hz-110 MHz, the autobalance bridge Agi-
lent 4294A was used. Its measurement range overlaps with that
of the other devices. In all the measurements described above,
the applied ac voltage was 0.1 V. The Agilent ”Dielectric Probe
Kit” 85070E using the so called ”performance probe” with an
E8363B PNA Series Network Analyzer covered the high fre-
quency range from 100 MHz to 40 GHz. It uses a so-called
open-end coaxial reflection technique, where the end of a coax-
ial line is immersed into the sample liquid. The applied ac volt-
age was 32 mV. Calibration was performed with the standards
Open, Short, and Water. As any contributions from parasitic
elements are excluded by this technique, the obtained absolute
values were used to correct the results obtained with the low-
frequency techniques, discussed above, for contributions from
stray capacitance.
Blood samples from a healthy person were taken at the hos-
pital ”Klinikum Augsburg”. All blood samples were taken from
the same person and various vital parameters of the blood sam-
ples were checked. All samples were taken before a meal, at
the same time of the day. We did not find any significant dif-
ference in the measurement results obtained on samples taken
at different days. To avoid clotting, the samples were prepared
with EDTA (ethylenediaminetetraacetic acid). The influence of
different coagulation inhibitors on the dielectric properties were
tested and found to be insignificant. Besides the whole blood,
which was measured as taken from the body, four other samples
with different hematocrit values (Hct) were prepared. Hct is
given by the ratio of volume fraction of the corpuscles (erythro-
cytes, leukocytes, and thrombocytes) of the blood and the total
volume. The whole blood used in the present work was found to
have Hct = 0.39. After centrifugation of the whole blood, the
corpuscular parts could be separated from the plasma by pipet-
ting. By remixing with the plasma obtained in this way, four
additional samples were prepared: plasma (Hct < 5) and blood
with Hct = 0.23, 0.57, and 0.86. The exact Hct values were de-
termined by taking hemograms with a Beckman-Coulter Hema-
tology Analyzer at the hospital. For each measurement run with
the different devices fresh blood samples were used.
One may suspect that sedimentation of RBCs and other cells
during the temperature-dependent measurement runs could in-
fluence the measurements. Covering the whole investigated
temperature range with the different measurement devices did
take about 2-3 hour only.
In the high-frequency measure-
ments with the ”open-end” coaxial technique, where relatively
large amounts of material (about 25 ml) held in transparent test
tubes were used, visible inspection revealed no indications of
sedimentation even after much longer time. Nevertheless the
sample material was thoroughly stirred before each frequency
sweep at the different temperatures. The results did well match
those at lower frequencies, where small sample amounts of
about 0.01 ml contained in a platinum capacitor were used.
With this capacitor, up to three separate temperature-dependent
measurement runs were performed, using different devices and
partly using different thermal histories. All the results did
match very well. Finally, the measurements usually were done
by first cooling the sample from room temperature and subse-
quently heating it up to the highest temperatures. The results
from the cooling and heating runs, which were done at differ-
ently aged samples, did always agree within experimental reso-
lution.
The integrity of the erythrocytes was retained during most of
the dielectric measurements, which was checked by a compari-
son of the room temperature results before and after cooling or
heating. However, as expected this was no longer the case when
the samples were subjected to the highest temperatures inves-
tigated, extending up to 330 K. Therefore these measurements
were performed at the end of each temperature-dependent mea-
surement run.
If not otherwise indicated,
the experimental data points
shown in the plots of the present work have errors that are not
exceeding the size of the symbols.
4. Results and Discussion
4.1. Broadband Spectra
Figure 1 shows the broadband spectra of the different sam-
ples at body temperature (≈ 310 K), extending from 1 Hz to
40 GHz.
5
electrode
polarization
310K
plasma
Hct=0.23
Hct=0.39
Hct=0.57
Hct=0.86
'
'
'
108
106
104
102
106
104
102
100
)
1
-
m
c
1
-
10-2
(
'
10-4
0
2
4
6
log10[ (Hz)]
(a)
(b)
(c)
10
8
Figure 1: (a) Dielectric constant, (b) dielectric loss, and (c) real part of the
conductivity of whole blood, blood plasma (Hct = 0.39), and blood samples
with different hematocrit values as function of frequency, measured at body
temperature (310 K). The lines are fits assuming a distributed RC equivalent
circuit to account for the electrodes and, for samples with Hct ≥ 0.23, two
Cole-Cole functions for the β- and γ-relaxation. For the plasma data, a single
Cole-Cole function was used instead.
The dielectric constant of the blood plasma (Fig. 1(a), or-
ange circles) reveals a low-frequency plateau between 1 and
100 Hz, followed by a steplike decrease towards higher fre-
quencies that passes into another plateau between about 1 MHz
and 10 GHz. The behavior below about 1 MHz can be as-
cribed to electrode polarization (see section 4.2 for a detailed
discussion). At frequencies beyond about 1 GHz a further de-
crease of ε′(ν) indicates the beginning γ-dispersion arising from
the reorientational motion of the water molecules (see section
4.4). ε′′(ν) shows a plateau at low frequencies, followed by a
strong decrease above about 300 Hz and the γ-relaxation peak
at ca. 20 GHz. Accordingly, the conductivity σ′(ν) exhibits a
strong increase at low frequencies, followed by a plateau be-
tween 1 kHz and 1 GHz. At ν > 1 GHz another strong increase
at the end of the measured spectrum shows up, again corre-
sponding to the γ-relaxation.
Just as the plasma, the dielectric spectra of the other samples
also show a γ-relaxation and a electrode-polarization contribu-
tion, the latter leading to a strong increase of ε′(ν) below about
10 - 100 kHz and decrease of σ′(ν) below about 1 - 10 kHz.
However, between these two features, a further process shows
6
up, the well-known β-dispersion, located at about 1 - 100 MHz
in ε′(ν). It is caused by the Maxwell-Wagner relaxation aris-
ing from the heterogeneity of the solute/cell system (see section
2.2). It is evidenced by a steplike decrease in the dielectric con-
stant at about 1 - 100 MHz, an s-shaped bend in the decrease of
the dielectric loss around 1 MHz, and a steplike increase in the
conductivity around 1 MHz. Comparing the different samples,
it becomes evident that the absolute values of ε′(ν), ε′′(ν), and
σ′(ν) decrease with increasing hematocrit value over almost the
whole frequency range. A detailed analysis of the β-dispersion
will be provided in section 4.3. The δ-dispersion, which is sup-
posed to be located in the frequency range between the β- and
γ-relaxation cannot be detected on this scale and will be treated
in section 4.5 below.
In the present work the complete broadband spectra are
fitted by combining several relaxational dispersions and the
electrode-polarization contribution. In addition, the dc conduc-
tivity has to be included in the fitting routine, since the blood
plasma contains free ions that contribute to the conductivity.
The lines shown in Fig. 1 are fits with the sum of two relax-
ational dispersions described by Eq. 2 (with β = 1) and the
dc-conductivity, which are assumed to be connected in series
to the electrode impedance given by Eq. 4 (with β = 1). The
fits were simultaneously performed for real and imaginary part
of the permittivity. A qualitative inspection of Fig. 1 reveals
that an excellent match of the experimental spectra could be
achieved in this way, which also is the case for the other tem-
peratures investigated.
In the following sections, we discuss
the different contributions to the spectra and the resulting rele-
vant fit parameters in detail. To serve for SAR calculations and
medical purposes, the fit curves for all temperatures and hema-
tocrit values investigated in the present work are provided in
electronic form.
4.2. Electrode polarization and α-Dispersion
To examine the contributions from electrode polarization to
the spectra in more detail, as an example in Fig. 2 the spectra
of whole blood at 310 K are shown in the low frequency range,
1 Hz - 200 kHz. As demonstrated by the dash-dotted magenta
line, the observed relaxationlike feature cannot be satisfactorily
fitted by assuming a simple RC equivalent circuit (Eq. 3) in se-
ries to the bulk sample, which leads to a symmetric loss peak
just like an intrinsic Debye relaxation (see section 2.3). The de-
viations are strongest in ε′′ below the peak frequency, where the
measured ε′′(ν) obviously varies far too weakly with frequency
to be describable by the strongly increasing fit curve. Using a
Debye relaxation function (Eq. 1) with an additional conductiv-
ity contribution σ′
dc, which leads to a minimum in ε′′(ν) below
the peak frequency, also cannot account for the experimental
data (dotted blue line). Replacing the Debye by a Cole-Cole
function (Eq. 2 with β = 1; green line) only leads to a marginal
improvement of the fits. Instead, a distributed RC equivalent
circuit as described by Eq. 4 with β = 1 but α , 0 (Cole-Cole
case) was found to provide very accurate fits of the measured
spectra (solid red line). The same can be said for the other
blood samples with different hematocrit values investigated in
the present work. A constant phase element in series with the
bulk also very well accounts for the experimental data (black
dashed line). However, in our further analysis we decided to use
the distributed RC circuit instead because its parameters seem
to have more physical signification than those of the constant
phase element.
whole blood
310 K
4
5
'
'
'
107
106
105
104
107
106
105
)
1
-
m
c
1
-
(
'
10-3
10-4
10-5
(a)
(b)
(c)
0
1
2
3
log10[ (Hz)]
Figure 2: ε′(ν) (a), ε′′(ν) (b), and σ′(ν) (c) of whole blood (Hct = 0.39) at
310 K and low frequencies (circles). The lines are fits using different functions:
RC equivalent circuit (Eq. 3; dash dotted magenta line), distributed RC equiv-
alent circuit (Eq. 4; red line), Debye function with additional dc-conductivity
contribution (dotted blue line), Cole-Cole function, also with dc contribution
(green line), and constant phase element (black dashed line).
Good fits with this approach can also be achieved for the re-
sults obtained at different temperatures. As an example, Fig. 3
shows the dielectric quantities of whole blood in the frequency
range, dominated by electrode polarization, for selected tem-
peratures. The lines represent the fits of the complete broad-
band spectra (cf. Fig. 1) where a distributed RC circuit, Eq. 4,
was used for the description of the low-frequency data. In all
cases the agreement of fits and experimental curves are excel-
lent. The onset of the electrode effects, i.e. the increase of ε′
and the decrease of σ′ when lowering the frequency, is found
to shift to lower frequencies with decreasing temperature. This
can be ascribed to the reduced mobility of the ionic charge car-
riers at low temperatures, which thus arrive at the electrodes for
smaller frequencies only. As revealed by Fig. 1, for increas-
ing hematocrit value the onset of the electrode effects shifts to
lower frequencies, too. This is in accord with the well-known
increase of the viscosity of blood with increasing hematocrit
7
value, corresponding to a reduction of ion mobility.
'
'
'
107
106
105
104
107
106
105
)
1
-
m
c
1
-
(
'
10-3
10-4
10-5
(a)
(b)
(c)
0
1
whole blood
electrode
polarization
280K
300K
320K
330K
4
5
2
3
log10[ (Hz)]
Figure 3: ε′(ν) (a), ε′′(ν) (b), and σ′(ν) (c) of whole blood (Hct = 0.39) in
the low-frequency regime dominated by electrode effects shown for selected
temperatures (symbols). The lines represent fit curves as in Fig. 1 using a dis-
tributed RC equivalent circuit for the description of the electrode polarization.
The fits reveal a width parameter α close to 0.15 and nearly
independent of temperature and hematocrit value (not shown),
except for Hct = 0.86, where α ≈ 0.20 is found, slightly in-
creasing with temperature. α characterizes the distribution of
relaxation times of the RC equivalent circuit that describes the
electrode polarization (see section 2.3). The deviations from
Debye behavior may be explained, e.g., by the surface rough-
ness of the electrodes [89, 90, 91]. The fits reveal electrode
capacitances CRC of the order of 10 µF. CRC is found to be only
weakly temperature dependent and it shows a tendency to de-
crease with increasing hematocrit value. This can be ascribed to
the mentioned reduction of the ionic mobility leading to a less
effective formation of the insulating electrode layers. The fits
do not reveal reliable information on the electrode resistance as
no clear low-frequency plateau in σ′(ν) is seen (cf. Fig. 1 and
Fig. 3).
From the presented results, it is clear that the electrode po-
larization is the dominant effect in the low frequency spectrum
of blood. The equivalent-circuit description in terms of a dis-
tributed RC circuit provides nearly perfect fits of the experi-
mental data. The typical deviations between fit and measured
data are around 10% or less, which is negligible compared to
the many decades the dielectric quantities vary with frequency.
Thus the presence of an additional contribution from a possible
α-relaxation seems unlikely but due to the mentioned devia-
tions, a weak α-relaxation cannot be fully excluded. However,
it should be noted that in earlier investigations also no indica-
tions for an α-relaxation in blood were found [28].
4.3. β-Dispersion
4.3.1. Phenomenological Evaluation
Figure 4 shows the spectra of whole blood in the frequency
range of the β-dispersion (10 kHz to 200 MHz) at different tem-
peratures. Except for the plasma, not containing any RBC’s
that would cause a β-process, all samples show a similar relax-
ational behavior and temperature dependence in this frequency
regime (see Fig. 1). Just as for intrinsic relaxations, the per-
mittivity curves in the β-dispersion regime shift to higher fre-
quencies with increasing temperature. The lines in Fig. 4 again
represent the fits of the complete broadband spectra (cf. Fig.
1) using a Cole-Cole function (Eq. 2 with β = 1) for the β-
relaxation. As discussed in detail in section 2.2, the β-relaxation
is generated by the heterogeneity of the sample material, which
is composed of plasma and RBC’s. ε′(ν) exhibits the typical
steplike decrease with increasing frequency (Fig. 4(a)).
Its
additional increase towards the lowest frequencies observed in
Fig. 4(a) corresponds to the onset of the electrode-polarization
effects (cf. Fig. 3), well taken into account in the fits by the
distributed RC equivalent circuit (see section 4.2). In ε′′(ν) a
loss peak is expected but only its high-frequency flank can be
seen (Fig. 4(b)).
Its low-frequency part is superimposed by
the strong ionic dc conductivity, which leads to a contribution
dc = σdc/(ε0ω). Thus, instead of a loss peak, only a slight
ε′′
shoulder at about 3 MHz is revealed.
The dc conductivity also leads to the low-frequency plateau
in σ′(ν) (Fig. 4(c)) while the shoulders observed around 3 MHz
arise from the relaxation. Via the relation σ′ = ε0ε′′ω,
the nearly Debye-like behavior of the β-relaxation (implying
ε′′(ν) ∼ ν−1 on the high frequency side of the peaks) leads to
the nearly frequency independent σ′(ν) at ν > 10 MHz. The
low- and high-frequency plateaus of σ′ are labeled as σdcβ and
σ∞β, respectively. The steplike increase of σ′(ν) from σdcβ to
σ∞β can be qualitatively understood assuming a shorting of the
cell membrane capacitances at high frequencies. Thus, at high
frequencies the RBC’s no longer obstruct the current path and
an enhanced conductivity is detected. Therefore σ∞β can be re-
garded as good approximation of the intrinsic conductivity of
the plasma, denoted as σa in section 2.5 (in fact it is a mixture
of plasma and cytoplasma conductivity, which we here assume
to be of not too different magnitude). This is nicely corrobo-
rated by the approximate agreement of this plateau value with
the conductivity of the pure plasma as seen in Fig. 1(c) for all
Hct values (the small deviations for higher Hct values are due
to the larger volume fraction of cytoplasma having somewhat
lower conductivity; see section 4.3.2). The absolute values of
both conductivity plateaus revealed in Fig. 4(c) increase with
increasing temperature, mirroring the thermally activated ionic
charge transport in the plasma.
As mentioned above,
the best fitting results of the β-
relaxation were achieved by using a Cole-Cole function. Only
104
103
'
102
(a)
105
103
(b)
(c)
2x10-2
10-2
2x10-3
4
'
'
)
1
-
m
c
1
-
(
'
whole blood
-relaxation
280K
300K
320K
330K
5
6
7
log10[ (Hz)]
8
Figure 4: ε′(ν) (a), ε′′(ν) (b), and σ′(ν) (c) of whole blood (Hct = 0.39) in the
β-dispersion region for selected temperatures. The lines represent fit curves as
in Fig. 1 using the Cole-Cole function for the description of the β-relaxation.
for the highest hematocrit values, significant deviations of fits
and experimental data were observed, which will be treated in
section 4.5. The temperature dependence of the width param-
eter αβ, the relaxation strength ∆εβ, and the relaxation time τβ
obtained from the fits are shown in Fig. 5. The present fitting
of the complete broadband spectra, including the contributions
from electrode polarization and the γ-relaxation, minimizes the
influence of any additional processes on the obtained fit param-
eters.
The width parameter α (Fig. 5(a)), which for intrinsic re-
laxations usually is assumed to arise from a distribution of re-
laxation times [67, 68], is almost temperature independent. As
α assumes rather small values between 0.07 and 0.11, the β
process shows nearly Debye-like behavior. The width parame-
ter increases with increasing RBC content, i.e., the deviations
from the Debye case become stronger. According to the Pauly-
Schwan model (see section 2.5, Eq. 10) the relaxation time τβ
depends on the conductivity outside (σa) and inside of the cell
(σi) and on the membrane capacitance (Cm). It is unlikely that
the cell parameters σi or Cm should be influenced by the hema-
tocrit value and thus a distribution of the outer plasma conduc-
tivity seems the most likely cause of the non-Debye behavior.
But also an alternative explanation seems possible: α , 0 im-
plies a shallower high-frequency flank of the β-peak. This flank
is essentially determined by the intrinsic plasma conductivity
8
T (K)
300
320
280
0.12
0.10
0.08
(a)
0.06
15 (b)
-relaxation
0
0
0
1
/
)
s
n
(
10
5
0
300
200
100
50
Hct=0.23
Hct=0.39
Hct=0.57
Hct=0.86
(c)
3.0
3.2
3.4
1000 / [T (K)]
3.6
Figure 5: Temperature dependence of width parameter (a), relaxation strength
(b), and relaxation time (c) as obtained from fits assuming a Cole-Cole-function
for the description of the β-relaxation (cf Fig. 4). The lines in the Arrhenius
plot of τβ (c) are linear fits corresponding to thermally activated behavior, Eq.
5.
and corresponds to the high-frequency plateaus seen in Fig.
4(c). Thus α , 0 implies an increase of σ′(ν) with frequency,
which is typical for hopping conductivity as commonly found
for ionic charge transport [92, 93, 94, 95]. Finally, it has to be
mentioned that additional relaxations arising from the other cell
types in blood could also influence the observed β-relaxation.
As RBCs are by far the dominating cell species (e.g., volume
fraction about 45% vs. ∼ 1% of white blood cells), these con-
tributions can be expected to be small. Nevertheless, it cannot
be excluded that they may contribute to the observed deviations
from Debye behavior.
As revealed by Fig. 5(b), the relaxation strength of the β-
relaxation is nearly temperature independent. Thus, according
to Eq. 8, the membrane capacity also can be assumed to be tem-
perature independent. The strong drop of ∆εβ at T > 320 K, ob-
served for all samples except for Hct = 0.86, is most likely due
to the hemolysis of the RBC’s at high temperatures. This as-
sumption is supported by the fact that no such deviations can be
found for the γ-relaxation (see section 4.4), which is indepen-
dent of the RBC’s. However, it is not clear why the sample with
9
the highest RBC content (black triangles in Fig. 5(b)) remains
unaffected. Possibly, cell-cell interactions prevent hemolysis at
high Hct.
Figure 5(b) also reveals a decrease of the relaxation strength
with increasing content of erythrocytes. This is a rather surpris-
ing result and cannot be explained within the proposed theo-
ries. Especially, it contradicts the increase of ∆εβ with p pre-
dicted by Eq. 8. In literature, suspensions of erythrocytes and
other cells using non-plasma solvents, like phosphate buffered
saline, quite generally show a continuous increase of ∆εβ with
Hct [50, 57, 96, 97, 98]. However, blood samples (i.e. with the
suspending medium being plasma) can show more complex be-
havior, especially for higher Hct values [56, 99] and as shown in
ref 54, substitution of plasma by some other solute can strongly
influence ∆εβ. In the framework of a simple equivalent-circuit
picture, ∆εβ is determined by a complex superposition of the
membrane capacitances of all RBCs. Increasing the number of
RBCs could lead to an increase or decrease of ∆εβ depending on
whether parallel or series connections of the cell capacitances
(relative to the field direction) are prevailing. The latter seems
to be the case in our blood samples. The reason is unclear but
cell aggregation as, e.g., rouleaux formation may play a role
here [54, 99, 100, 101].
According to Eq. 10, the β-relaxation time should depend on
σi, σa, and Cm. It was shown above that the membrane capaci-
tance is nearly temperature independent. Thus, the conductivi-
ties should dominate the temperature dependence of τβ. Indeed
the Arrhenius representation of Fig. 5(c) reveals that τβ(T ) is
in accord with the expected thermally activated behavior, typ-
ical for ionic conductivity. The hindering barriers Eτ, calcu-
lated from the slopes in Fig. 5(c) (cf. Eq. 5), seem to slightly
increase with growing Hct and barriers varying between 0.11
and 0.15 eV were obtained. However, due to the rather small
temperature region that could be covered in these biological
samples (compared, e.g., to supercooled liquids [39, 40]), the
significance of these values should not be overemphasized. In
addition, Fig. 5(c) reveals a decrease of the relaxation times
with increasing hematocrit value. In principle, such a behavior
seems to be consistent with Eq. 10 but the observed variation by
about a factor of three is stronger than expected. However, one
should be aware that the relaxation time is directly proportional
to Cm (Eq. 10) while Cm itself is proportional to ∆εβ (Eq. 8).
Thus it is clear that the observed rather strong Hct-dependent
variation of τβ is directly connected to that of ∆εβ revealed by
Fig. 5(b).
To compare the results on the β-relaxation parameters pre-
sented above with earlier publications only partly is possible,
because (to the best knowledge of the authors) no such sys-
tematic (temperature and hematocrit dependent) and broadband
research on human blood has been done before. Moreover,
the available literature values deviate quite strongly from each
other. For various erythrocyte suspensions, the reported values
of the relaxation time τβ are, for example, 254 ns (Hct = 0.07,
room temperature) [50], 29 ns (Hct = 0.30, T = 298 K) [102],
or 230 ns (Hct = 0.47, T = 310 K) [52, 103]. In blood the
following values were found: τβ = 133 ns (sheep blood, T =
310 K) [46], τβ = 89 − 65 ns (human blood, T = 288 − 308 K,
Hct = 0.43) [25], and τβ = 53.1 ns (bovine blood, room tem-
perature, Hct = 0.50) [13]. The values in the present work
vary between 35.9 ns (Hct = 0.86, T = 330 K) and 274.1 ns
(Hct = 0.23, T = 280 K). The literature results for the re-
laxation strength also show rather strong variation. In ref 52,
literature values between 1100 and 5000 were reported. Fricke
found, dependent on Hct, εs ≈ ∆ε = 900 −4000 for dog, rabbit,
and sheep blood [96]. For human blood, Pfutzner published val-
ues between approximately 2000 and 6000 (Hct = 0.10 − 0.90)
[99]. In the present work we have obtained ∆ε ≈ 3300 − 13800
(for Hct = 0.86, T = 330 K and Hct = 0.23, T = 320 K, re-
spectively). Even less data are available for the width parameter
αβ, because often other fitting functions were used. But mostly
they are around 0.1 [46, 52, 104], similar to the values in the
present work.
4.3.2. Application of Cell Models
As mentioned in section 2.5, by using appropriate models it
should be possible to determine intrinsic cell parameters as the
membrane capacitance or the conductivity of the cytoplasma
from the parameters of the β-relaxation. Using the fitting pa-
rameter ∆εβ and Eq. 8, the membrane capacitance Cm can be
calculated. As discussed in the previous section, the experi-
mentally determined dielectric strength decreases with increas-
ing Hct, in contrast to the increase predicted by Eq. 8. The
use of Eq. 8 therefore would imply a strongly Hct-dependent
membrane capacitance (see Table 1 for the results at 310 K).
This can hardly be interpreted in a physical way. Literature
values vary between 0.17 µF/cm2 (ref 102) and 3 µF/cm2 (ref
54). However, most authors assume a membrane capacity of
about 1 µF/cm2 [52, 96, 97, 105, 106], whereas some report
Hct-dependent membrane capacities [51, 56]. Possible reasons
for the unexpected behavior of ∆εβ(Hct) and thus of Cm(Hct)
were discussed in the previous section. It seems that Eq. 8 is
not able to account for the observations in ”real” blood samples,
in contrast to suspensions of erythrocytes in common solvents.
Hct
0.23
0.39
0.57
0.86
Cm ( µF
cm2 )
11 (±3)
4.9 (±1.4)
3.5 (±1.1)
1.2 (±0.4)
σa ( 10−2
Ωcm )
0.79 (±0.15)
0.63 (±0.15)
0.74 (±0.15)
0.78 (±0.15)
σi ( 10−2
Ωcm )
0.75 (±0.08)
0.50 (±0.08)
0.78 (±0.09)
0.83 (±0.11)
εi
36.9 (±3.0)
41.5 (±2.5)
42.0 (±2.5)
44.7 (±2.0)
Via Eq. 11, the conductivity of the cell interior σi can be
calculated from σa, Cm, and the measured β-relaxation times.
However, as Cm determined from Eq. 8 shows an unphysi-
cal Hct dependence (Table 1), also an unreasonable Hct de-
pendence of σa would result. An alternative determination of
σi is provided by Eq. 12. Using the experimentally deter-
mined σa and σ∞β at 310 K we arrive at the values for σi
listed in Table 1. Literature results are distributed around 0.6 ×
10−2(±0.1) Ω−1cm−1 [41, 42, 54, 56, 57, 98, 105, 106, 107].
Deviating results were reported by Cook [24, 25] (ca. 1.0 ×
10−2 Ω−1cm−1) and Asami [108] (ca. 0.32 × 10−2 Ω−1cm−1).
Our values for σi are by about a factor 2-3 smaller than the
measured conductivity of the plasma, σa ≈ 1.7 ×10−2 Ω−1cm−1.
It seems reasonable that the conductivity of the cytoplasma
should be lowered by the presence of the large hemoglobin
molecules and their bound water shells within the RBCs (about
37% volume fraction [41]), if compared to the conductivity of
the outer plasma (see following discussion of Fig. 6 for a quan-
titative treatment). Indeed such behavior was found previously
[42, 54, 56]. The reported ratios between about 1.5 and 2.7
are consistent with our findings. Obviously, Eq. 12 is able to
provide reasonable estimates for σi. It is based on the deter-
mination of σ∞β, which is read off at high frequencies, where
the cell membranes are shorted and thus cell aggregation has no
effect on the results.
)
K
1
-
m
c
1
-
(
T
i
10
plasma
(a)
Hct
0.23
0.39
1
3.0
0.57
0.86
3.2
3.4
1000 / [T (K)]
50
40
30
i
(b)
Table 1: Membrane capacitance Cm, plasma conductivity σa, conductivity of
the cell interior σi, and dielectric constant of the cell interior εi at 310 K as
determined from Eqs. 8, 9, 12, and 13 respectively.
280
300
T (K)
320
Using Eq. 9 should allow for the calculation of the conduc-
tivity of the suspending medium σa from the dc conductivity
σdcβ of the blood samples. σa also can be directly determined
from the dc conductivity of the plasma, measured in the present
work (1.7 × 10−2 Ω−1cm−1 at 310 K). The calculated values are
provided in Table 1. While being nearly Hct-independent as
expected, they differ from the directly measured value by a fac-
tor of about two. Again cell aggregation causing a lowering of
the experimentally observed σdcβ may explain this finding.
Figure 6: Temperature dependent conductivity (Arrhenius plot) (a) and dielec-
tric constant (b) of the cell interior calculated from Eqs. 12 and 13, respec-
tively. For comparison, in (a) also data for pure plasma are provided. The solid
line in (a) indicates approximately linear behavior implying thermally activated
charge transport (Eq. 6). The dashed line shows an approximate description of
the blood data (except for Hct = 0.39) using the same energy barrier as for the
plasma.
In Fig. 6(a) the temperature dependence of σi is shown in
10
the Arrhenius type of presentation (log(σiT ) vs 1000/T ) com-
monly used for ionic conductivity (cf Eq. 6). For comparison,
also the dc conductivity determined from fits of the spectra of
pure blood plasma (cf. Fig. 1) is included. In the determi-
nation of σi(T ) via Eq. 12, for σa the plasma dc conductivity
was used and σ∞β was calculated from the fit parameters of the
β-relaxation shown in Fig. 5. As the β-relaxation shows slight
deviations from Debye behavior, σ′(ν) exhibits a slight increase
in the region of its high-frequency plateau (see, e.g., Fig. 4(c)).
As an estimate of σ∞β, we used the value of σ′(ν) at a frequency
1.5 decades above the peak frequency. Except for whole blood,
the temperature dependence of σi is nearly independent of Hct
and seems to follow thermally activated behavior (dashed line
in Fig. 6(a)). The deviations at the two highest temperatures
are directly related to similar problems in the β-relaxation pa-
rameters (Fig. 5). As discussed in section 4.3, this may arise
from an onset of hemolysis of the RBCs. Interestingly, the en-
ergy barrier of 0.17 eV, deduced from the slope of the linear fit
curve of the plasma data (solid line) is in good accord with the
results on the blood samples (dashed line). Nevertheless, the
absolute values of the conductivity of the cytoplasma are about
a factor of 2-3 lower than those of the plasma. As mentioned in
the previous paragraph, this can be explained by the presence
of the hemoglobin molecules within the cell. For a quantitative
estimate one can use Maxwell’s mixture equation for the effec-
tive conductivity σeff of a suspension of particles with volume
concentration p [10]:
σeff − σs
σeff + 2σs
= p
σp − σs
σp + 2σs
(16)
Here σs and σp are the conductivity of the solute and the par-
ticle, respectively. If we regard the hemoglobin molecules as
insulating particles (i.e., σp = 0) suspended in plasma with the
same conductivity as the extracellular plasma, we arrive at the
following ratio of solute conductivity (σs = σa) and effective
conductivity (σeff = σi) [41]:
σa
σi
=
1 + p/2
1 − p
(17)
Using p = 0.37 (ref 41), a ratio of about two is obtained, which
is in quite reasonable agreement with the findings of Fig. 6(a).
The energy barrier of 0.17 eV for charge transport within the
intra- and extracellular plasma is of the same order of magni-
tude as the one deduced from τβ(T ) (0.11 - 0.15 eV, see previous
section). This seems reasonable as the temperature dependence
of τβ should be mainly governed by the conductivity of inner
and outer plasma (Eq. 10). In any case one should bear in mind
that the absolute values of the energy barriers have rather high
uncertainty due to the restricted temperature range.
The dielectric constant of the cell interior εi was calculated
using Eq. 13. The results for 310 K are listed in Table 1 and the
temperature dependence is shown in Fig. 6(b). One should be
aware that εi is the dielectric constant at frequencies below the
onset of the γ-relaxation. As expected, the obtained values are
smaller than the dielectric constant of the suspending medium
(εa ≈ 73 - 67 for T = 290 - 310 K, respectively), deduced from
the fits of the spectra of pure plasma (see Fig. 1 for 310 K). This
difference is reasonable because the main contribution to the ε′
of the plasma arises from the highly dipolar water molecules
(εs of water ≈ 74 at 310 K (ref 109)) and the additional con-
stituents of the cytoplasma (mainly hemoglobin) should lower
its permittivity. In literature, εi values ranging around 40-70
were reported [41, 106, 107, 108]. The calculation of εi by Eq.
13 should provide the same results for each hematocrit value.
However, as revealed by Table 1 and Fig. 6(b), the calculated
εi increases by about 20% with increasing Hct, pointing out the
limits of the model. Similar behavior was also reported in ref
107. The decrease of εi with increasing temperature revealed
by Fig. 6(b) is consistent with Curie behavior (Eq. 7) expected
for the dielectric strength (and thus approximately also for the
static dielectric constant) of dipolar materials. Here one should
be aware that εi represents the static dielectric constant of a γ-
like relaxation of the cell interior that will take place at higher
frequencies and in fact this relaxation contributes to the actu-
ally observed γ-relaxation of blood (see section 4.4). The un-
certainty of the data in Fig. 6(b) is too large to allow for a quan-
titative evaluation in terms of Eq. 7. Overall, Eq. 13 seems to
lead to reasonable values of the dielectric constant of the cyto-
plasma. It is based on the determination of the plasma dielectric
constant (experimentally determined) and ε∞β, which is not in-
fluenced by possible cell aggregation effects.
An alternative determination of the dielectric properties of
the RBCs is provided by the Hanai-Bruggemann model, Eqs.
14 and 15, which was especially proposed for highly concen-
trated suspensions. From Eq. 15 the cell conductivity was
calculated, leading, however, to negative values. Using Eq.
14, the dielectric constant εp of the cell is found to vary be-
tween 34 (Hct = 0.57) and 42 (Hct = 0.39) at room tempera-
ture. Those values are slightly smaller than the ones calculated
from the Pauly-Schwan theory. This is a reasonable result since
the Hanai-Bruggemann model does not account for the shelled
structure of the cells and thus the obtained values represent the
average of cell membrane and interior. The dielectric constant
of the membrane can be expected to be much lower than that of
the cytoplasma (in contrast to its capacitance, which is high due
to its small thickness), which leads to the reduced values of the
total dielectric constant.
4.4. γ-Dispersion
The dielectric spectra of bulk water exhibit a strong relax-
ation feature near 18 GHz (at room temperature) [110], which
is also observed in electrolytic solutions [111]. It is commonly
ascribed to the reorientational dynamics of the dipolar water
molecules (but also alternative scenarios are discussed; see,
e.g., ref 112) and denoted as α-relaxation within the nomencla-
ture of dipolar liquids and glass formers. The same relaxational
process also arises from the free water molecules in blood sam-
ples. Figure 7 shows real and imaginary part of the permittivity
(a, b) and the real part of the conductivity (c) of whole blood in
the frequency range 1 to 40 GHz at different temperatures. The
lines represent fits of the broadband spectra as shown in Fig. 1,
using a Cole-Cole function for the γ-relaxation. In ε′(ν) (Fig.
7(a)), the onset of the typical relaxation steps is seen but their
11
whole blood
-relaxation
(a)
(b)
(c)
75
50
25
40
20
0.5
0.1
'
'
'
)
1
-
m
c
1
-
(
'
0.01
9.0
9.5
10.0
log10[ (Hz)]
280K
300K
320K
330K
10.5
T (K)
300
320
280
0.10
0.05
(a)
0.00
-relaxation
70
60
50
40
15
10
5
)
s
p
(
(b)
(c)
plasma
Hct=0.23
Hct=0.39
Hct=0.57
Hct=0.86
3.0
3.2
3.4
1000 / [T (K)]
3.6
Figure 7: ε′(ν) (a), ε′′(ν) (b), and σ′(ν) (c) of whole blood (Hct = 0.39) in the
γ-dispersion region for selected temperatures. The lines represent fit curves as
in Fig. 1 using the Cole-Cole function for the description of the γ-relaxation
.
high frequency plateaus, ε∞γ, are located beyond the investi-
gated frequency range. Thus, exact values for ε∞γ could not be
determined and in the fitting procedure a lower limit of 2.5 was
used leading to values between 2.5 and 6. The low-frequency
plateau εsγ of the γ-dispersion decreases with increasing tem-
perature. The relaxation steps and loss peaks (Fig. 7(b)) show a
strong temperature-dependent frequency shift due to the slow-
ing down of the molecular dynamics with decreasing tempera-
ture. We find the Cole-Cole formula (Eq. 2 with β = 1) to pro-
vide the best fits of the γ-relaxation. Figure 7(c) shows the con-
ductivity spectra with the corresponding rise and the onset of
the high frequency plateau. The low-frequency plateau of σ′(ν)
corresponds to the combined conductivity of plasma and cyto-
plasma as discussed in section 4.3.2. The γ-dispersion shows
similar behavior for the other investigated samples.
The fitting parameters of the γ-relaxation, αγ, ∆εγ, and τγ,
are provided in Fig. 8. The width parameter (a) shows a ten-
dency to increase with increasing Hct. This seems reasonable
as a higher number of RBC’s should lead to stronger disorder
in the system and therefore the distribution of relaxation times
should broaden. The observed decline of αγ with increasing
temperature, corresponding to an approach of Debye behavior,
is a common phenomenon in dipolar liquids [39, 113]. It can
Figure 8: Temperature dependence of width parameter (a), relaxation strength
(b), and relaxation time (c) as obtained from fits assuming a Cole-Cole func-
tion for the description of the γ-relaxation (cf Fig. 7). The dashed line in
(b) shows literature data for pure water (using the I.U.P.A.C. standard data for
εs(T )) [109]. The solid lines in (b) are fits with a Curie-law, Eq. 7. The line
in the Arrhenius plot of τγ (c) is a linear fit of the curve for Hct = 0.39 (whole
blood) corresponding to thermally activated behavior, Eq. 5. The dashed line
shows the curve for pure water [110].
be explained by the growing thermal fluctuations of the envi-
ronment of the water dipoles. At very high temperatures, each
dipole ”sees” the time average of the quickly fluctuating envi-
ronment, which is the same for every dipole, leading to Debye
behavior [39].
The hematocrit dependence of the relaxation strength (Fig.
8(b)) shows the expected tendency: Increasing Hct values cause
a decrease of the volume fraction of plasma and thus of wa-
ter in the sample, causing the reduction of the γ-relaxation
strength. The temperature dependence of ∆εγ can be well pa-
rameterized by a Curie-law, Eq. 7, (solid lines) with some de-
viations for Hct = 0.23 only. The obtained Curie constant, C,
increases smoothly from 13400 to 19800 with decreasing Hct.
The dashed line in Fig. 8(b) corresponds to ∆ε(T ) of pure wa-
ter, calculated from the I.U.P.A.C. (International Union of Pure
and Applied Chemistry) standard values for the static permit-
tivity εs of bulk water [109] (see also [114]) via ∆ε = εs − ε∞
assuming ε∞ = 4 [115, 116]. The relaxation strength of water
matches the general trend revealed by the other curves in Fig.
12
8(b). However, obviously it shows a somewhat stronger tem-
perature dependence. This may indicate weaker interactions
between the water molecules in blood than in pure water, which
can be rationalized by the presence of the other constituents of
blood (e.g., proteins or salt ions).
Figure 8(c) shows the temperature dependence of the relax-
ation times τγ in an Arrhenius plot. The observed linear in-
crease is in accord with thermally activated behavior, Eq. 5.
As an example, a linear fit of the data at Hct = 0.39 is shown
(solid line). From its slope an energy barrier of 0.19 eV is ob-
tained. There seems to be a slight increase of energy barriers
with growing Hct value (from 0.18 to 0.20 eV). However, this
variation is too small to be considered significant, especially
when taking into account the rather small temperature range
that could be investigated in the present experiments due to
the restricted robustness of blood to stronger temperature varia-
tions. The present results agree reasonably with those reported
by Cook [24] who found values for τγ of whole blood varying
between 11.9 and 7.0 ps at three temperatures between 298 and
308 K. Gabriel et al.[46] reported 8.4 ps at 310 K, about 30%
higher than our result of 6.5 ps. The dashed line in Fig. 8(c)
represents τ(T ) of pure water as measured by Kaatze [110].
Obviously the γ-relaxations in the investigated blood samples
exhibit nearly identical dynamics as the main relaxation of pure
water. Water shows some small deviations from Arrhenius be-
havior, which seem to be absent in blood but these differences
are of limited significance. However, one may speculate that
the non-Arrhenius behavior of water arises from increasing co-
operativity of the molecular motions at low temperatures as of-
ten invoked to explain corresponding findings in glass forming
liquids [117, 118]. In blood, its other constituents can be ex-
pected to lead to a reduction of the direct interactions between
the water molecules and thus to less cooperativity. In addition,
there are speculations of a first-order phase transition in super-
cooled water [119], which may lead to critical power-law be-
havior even in the normal liquid state, thus also explaining the
deviation from Arrhenius behavior in Fig. 8(c). Both scenar-
ios are consistent with the different temperature dependence of
∆ε(T ) of water and blood discussed in the previous paragraph.
4.5. Further Dispersions
In section 4.3 it was shown that fits using the Cole-Cole func-
tion provide a reasonable description of the β-relaxation region.
However, there are some minor deviations of fits and experi-
mental data especially for the higher hematocrit values. This is
demonstrated in Fig. 9(a) where dielectric-constant data for the
blood samples with the highest and lowest hematocrit values are
shown. In contrast to the 23% sample, the fit of the spectrum
of the highly concentrated sample clearly is of inferior quality.
Similar deviations were previously also observed in Cole-Cole
fits of data on disc-shaped rabbit-erythrocyte suspensions [57].
A close inspection of the spectrum at Hct = 0.86 (Fig. 9(b))
seems to indicate that it may be composed of two separate re-
laxation steps. However, one could suspect an experimental
artifact because in the β-dispersion region the spectrum is com-
posed of results from two different experimental methods with
the transition close to 10 MHz (see section 3). To exclude this,
'
'
104
103
102
103
102
(a)
(b)
5
60
40
8
7
6
8
log10[ (Hz)]
310 K
Hct=0.23
Hct=0.86
9
9
10
10
Figure 9: (a) Comparison of ε′(ν) of the blood samples with the lowest (Hct =
0.23) and highest hematocrit value (Hct = 0.86) in the frequency range of the
β- and δ-dispersion. The lines represent fit curves as in Fig. 1 and Fig. 4
using the Cole-Cole function for the description of the β-relaxation. (b) ε′(ν)
for Hct = 0.86 (triangles: same data as in (a), crosses: measurement with
different experimental setup). The solid line in (b) shows an alternative fit with
two Cole-Cole functions for the β-relaxation. The two separate relaxation steps
are indicated by the dashed lines. The inset shows a magnified view of the
high-frequency region.
in Fig. 9(b) additional data extending from 100 kHz to 50 MHz
obtained with a different apparatus (autobalance bridge Agilent
4294A) are shown, which exactly reproduce the two other data
sets. Thus, a fit using the sum of two separate relaxation contri-
butions was performed (solid line in Fig. 9(b)). It provides an
excellent description of the spectrum revealing relaxation times
of 16 ns and 146 ns. A very similar fit using two Cole-Cole
functions was shown by Asami and Yamaguchi [74] to provide
a good description of data on human erythrocyte suspensions.
In blood there are various possibilities for additional relax-
ational processes, in addition to those considered for the ex-
planation of the α-, β-, and γ-relaxation:
(i) the reorienta-
tion of protein-bound water molecules, (ii) the hemoglobin β-
relaxation (i.e., the tumbling motion of the protein molecules),
(iii) the motion of polar protein subgroups, (iv) the Maxwell-
Wagner relaxation of the cell interior, or (v) the additional
Maxwell-Wagner relaxations due to the non-spherical cell
shape,
to name just the most likely ones. Most of them
can be simply excluded based on the very large amplitude of
∆εs ≈ 1000 of the additional relaxation suggested by the fit
shown in Fig. 9(b): (i) Bound water cannot have a larger ∆ε
than free water.
(ii) The hemoglobin β-relaxation in aque-
ous solution was found to have a ∆ε of the order of 100
[12, 13, 33].
It seems unreasonable that it should be higher
in the hemoglobin/cytoplasma solution of the cell interior. (iii)
Polar protein subgroups can be expected to have smaller relax-
ation strength that the main tumbling relaxation. (iv) In princi-
13
ple, the capacitance of the cell interior should also be shorted
at high frequencies and, indeed, the Maxwell-Wagner model
of shelled particles predicts a corresponding relaxation [61].
However, for any reasonable choice of parameters this capac-
itance is too small to lead to any considerable contribution to ε′
and this additional relaxation usually is considered negligible
[12, 13, 33]. Thus, the non-spherical shape of the RBCs seems
to be the most likely cause of the additional relaxation observed
in Fig. 9(b). Already in ref 12 deviations from simple relaxation
behavior of erythrocyte suspensions were ascribed to the non-
spherical form of RBCs and also Asami and Yamaguchi [74]
explained their results in this way. As mentioned in section 2.5,
the Maxwell-Wagner model for suspensions of spheroid parti-
cles predicts up to four relaxations [61] (two of them arise from
the cell interior and can be neglected). However, the found
relaxation-time ratio of the order of 10 is too high to be ex-
plainable by this model, at least if assuming a reasonable ratio
of the two semi-diameters of the spheroids [61, 120]. The spec-
tra on rabbit erythrocytes, mentioned above, also could not be
described by the Maxwell-Wagner model for spheroid particles
[57]. However, one should be aware that RBCs only roughly
can be approximated by spheroids and most likely their bicon-
cave shape plays a role in the found discrepancies.
An additional δ-dispersion between β- and γ-relaxation is
often invoked to explain a slow continuous decrease of ε′(T ),
observed in the region from several 10 MHz to about 3 GHz
including protein solutions
in various biological materials,
[12, 30, 33, 121, 122] and blood [12].
It has been ascribed
to various mechanisms as the dynamics of protein-bound water
molecules or polar subgroups of proteins. Indeed such a disper-
sion is also found in our present results on blood and becomes
most pronounced for the high hematocrit values (see inset of
Fig. 9 for an example). However, it can be completely de-
scribed by the broadband fits promoted in the previous sections
(line in inset), especially if including a second relaxation in the
β-relaxation region as shown in Fig. 9(b). Thus in our blood
samples the apparent dispersion in this region arises from the
superposition of β- and γ-relaxation and we find no evidence for
a δ-relaxation. However, the presence of a small δ-relaxation is
not completely excluded by this finding.
5. Summary and Conclusions
In the present work, we have provided dielectric spectra of
human blood for an exceptionally broad frequency range and at
different temperatures and hematocrit values. A combination of
models for the different dispersion regions enabled nearly per-
fect fits of the broadband spectra. The obtained fit curves rep-
resent an excellent estimate of the dielectric properties of blood
for a wide range of parameters. They are provided for electronic
download in the supporting information and can be employed
for SAR calculations and other application purposes. The dif-
ferent dispersion regions have been analyzed in detail. The ob-
served electrode-polarization effects are accounted for by an
equivalent circuit model assuming a distribution of relaxation
times. While our analysis of the low-frequency region cannot
completely rule out the presence of an α-dispersion in blood,
we can satisfactorily describe our low-frequency data without
invoking such a relaxation. This finding agrees with earlier re-
sults stating the absence of an α-relaxation in blood [28]. The
analysis of the β-relaxation using standard cell models partly
leads to unreasonable results for the intrinsic dielectric proper-
ties. This most likely can be ascribed to cell aggregation play-
ing an important role in ”real” blood samples, in contrast to
suspensions of erythrocytes prepared by standard solutes, of-
ten reported in literature. Cell aggregation seems to be impor-
tant especially for the dielectric behavior at the low-frequency
side of the β-relaxation. In contrast, using only parameters de-
termined at frequencies beyond the β-peak frequency, leads to
reasonable estimates of the conductivity and dielectric constant
of the cell interior. In addition, we find strong hints that the
β-relaxation is in fact composed of two separate relaxation pro-
cesses, which we ascribe to the marked deviations of the RBCs
from spherical geometry.
We observe clear dispersion effects in the region between the
β- and γ-relaxation, which often is ascribed to a so-called δ-
relaxation. However, our analysis of the broadband spectra in-
cluding electrode polarization, β-dispersion, and γ-relaxation
leads to excellent fits in this region, which thus is revealed to
be a superposition of different contributions and not due to a
separate relaxation process. Thus, while there clearly is dis-
persion in blood between the β- and γ-relaxation, there is no
evidence for a δ-relaxation. Finally, detailed information on the
γ-relaxation in blood is provided. Its properties closely resem-
ble those of the relaxation caused by reorientational molecular
motions in pure water. However, some minor differences arise,
which seem to indicate less cooperative motions of the water
molecules in blood samples.
Overall, the dielectric spectra of blood are of astonishing
simplicity if considering the complexity of blood, being com-
posed of a variety of different constituents.
In fact we have
described our broadband spectra without assuming any intrin-
sic frequency dependence in the complete range from 1 Hz up
to about 1 GHz and only the γ-dispersion arising from the tum-
bling motion of the water molecules is of intrinsic nature. Of
course, there is the strong β-relaxation, which may be regarded
as quasi-intrinsic but as it is of Maxwell-Wagner type, in a nar-
rower sense it should be considered as artificial. However, of
course for many applications (e.g., the calculation of SAR val-
ues) the overall dielectric properties and not only the intrinsic
ones are of essential importance and we hope our work will
serve for these purposes in the future.
6. Acknowledgements
We gratefully acknowledge the help of Dr. W. Behr, Dr. K.
Doukas, and Prof. Dr. W. Ehret at the ”Klinikum Augsburg” in
the taking and preparing of the blood samples.
Appendix A. Supplementary data
Fit curves of the broadband spectra and fit parameters for all
samples and temperatures investigated in the present work are
available for electronic download.
14
References
[1] E. H. Grant, R. J. Sheppard, G. P. South, Dielectric behaviour of biolog-
ical molecules in solution, Clarendon Press, Oxford, 1978.
[2] F. F. Becker, X. B. Wang, Y. Huang, R. Pethig, J. Vykoukal, P. R. C.
Gascoyne, Separation of human breast-cancer cells from blood by dif-
ferential dielectric affinity, Proc. Natl. Acad. Sci. USA 92 (1995) 860–
864.
[3] Y. Hayashi, Y. Katsumoto, I. Oshige, S. Omori, A. Yasuda, K. Asami,
Dielectric inspection of erythrocytes, J. Non-Cryst. Solids 356 (2010)
757–762.
[4] Y. Hayashi, Y. Katsumoto, S. Omori, A. Yasuda, K. Asami, M. Kaibara,
I. Uchimura, Dielectric coagulometry: A new approach to estimate ve-
nous thrombosis risk, Anal. Chem. 82 (2010) 9769–9774.
[5] A. Ahlbom, U. Bergqvist, J. H. Bernhardt, J. P. Cesarini, L. A. Court,
M. Grandolfo, M. Hietanen, A. F. McKinlay, M. H. Repacholi, D. H.
Sliney, J. A. J. Stolwijk, M. L. Swicord, L. D. Szabo, M. Taki, T. S.
Tenforde, H. P. Jammet, R. Matthes, Guidelines for limiting exposure
to time-varying electric, magnetic, and electromagnetic fields (up to 300
GHz), Health Physics 74 (1998) 494–522.
[6] H. Bassen, Ieee recommended practice for measurements and compu-
tations of radio frequency electromagnetic fields with respect to human
exposure to such fields, 100 kHz-300 GHz (ieee std c95.3-2002), The
Institute of Electrical and Electronics Engineers, New York, 2003.
[7] C.-K. Chou, Ieee standard for safety levels with respect to human expo-
sure to radio frequency electromagnetic fields, 3 kHz to 300 GHz (ieee
std c95.1-2005), The Institute of Electrical and Electronics Engineers,
Inc., New York, 2006.
[8] A. Loidl, P. Lunkenheimer, R. Gulich, A. Wixforth, M. Schnei-
der, P. Hanggi, G. Schmid, Untersuchungen zu der Fragestellung, ob
makroskopische dielektrische Gewebeeigenschaften auch auf Zellebene
bzw. im subzellularen Bereich uneingeschrankte Gultigkeit besitzen.
Abschlussbericht
http://www.emf-forschungsprogramm.de/forschung/dosimetrie/dosi-
metrieabges/dosi075ab.pdf, 2008.
[9] R. Hober, Messungen der inneren Leitfahigkeit von Zellen III, Arch.
Ges. Physiol. 150 (1913) 15–45.
[10] J. C. Maxwell, A treatise of electricity and magnetism, Part 2, Chapter
10, Oxford University Press, London, 1873.
[11] K. W. Wagner, Erklarung der dielektrischen Nachwirkungsvorgange auf
Grund Maxwellscher Vorstellungen, Arch. Elektrotech. 2 (1914) 371–
387.
[12] H. P. Schwan, Electrical properties of tissue and cell suspensions, Adv.
Biol. Med. Phys. 5 (1957) 147–209.
[13] H. P. Schwan, Electrical properties of blood and its constituents - alter-
nating current spectroscopy, Blut 46 (1983) 185–197.
[14] H. Fricke, The electric capacity of cell suspensions, Phys. Rev. 21
(1923) 708–709.
[15] H. Fricke, A mathematical treatment of the electric conductivity and
capacity of disperse systems, I. The electric conductivity of a suspension
of homogenous spheroids, Phys. Rev. 24 (1924) 575–587.
[16] H. Fricke, A mathematical treatment of the electric conductivity and ca-
pacity of disperse systems, II. The capacity of a suspension of conduct-
ing spheroids surrounded by a non-conducting membrane for a current
of low frequency, Phys. Rev. 26 (1925) 678–681.
[17] H. Fricke, The electrical capacity of suspensions with special reference
to blood, J. Gen. Physiol. 9 (1925) 137–152.
[18] K. S. Cole, Electric impedance of suspensions of spheres, J. Gen. Phys-
iol. 12 (1928) 29–36.
[19] H. Danzer,
Uber das Verhalten biologischer Korper bei Hochfrequenz
2, Ann. Phys. 413 (1934) 783–790.
[20] D. A. G. Bruggemann, Berechnung verschiedener physikalischer Kon-
stanten von heterogenen Substanzen. I. Dielektrizitatskonstanten und
Leitfahigkeiten der Mischkorper aus isotropen Substanzen, Ann. der
Phys. 24 (1935) 636–664.
[21] B. Rajewsky, H. Schwan, Die Dielektrizitatskonstante und Leitfahigkeit
des Blutes bei ultrahohen Frequenzen, Naturwissenschaften 35 (1948)
315–316.
[22] H. Schwan, Electrical properties of blood at ultrahigh frequencies, Am.
J. of Phys. Med. 32 (1953) 144–152.
[23] T. S. England, Dielectric properties of the human body for wave-lengths
in the 1-10 cm range, Nature 166 (1950) 480–481.
15
[24] H. F. Cook, Dielectric behaviour of human blood at microwave frequen-
cies, Nature 168 (1951) 247–248.
[25] H. F. Cook, A comparison of the dielectric behaviour of pure water
and human blood at microwave frequencies, Br. J. Appl. Phys. 3 (1952)
249–255.
[26] H. P. Schwan, Die elektrischen Eigenschaften von Muskelgewebe bei
Niederfrequenz, Z. Naturforsch. 9b (1954) 245–251.
[27] H. P. Schwan, E. L. Carstensen, Dielectric properties of the membrane
of lysed erythrocytes, Science 125 (1957) 985–986.
[28] T. P. Bothwell, H. P. Schwan, Electrical properties of the plasma mem-
brane of erythrocytes at low frequencies, Nature 178 (1956) 265–266.
[29] H. P. Schwan, G. Schwarz, J. Maczuk, H. Pauly, On the low-frequency
dielectric dispersion of colloidal particles in electrolyte solution, J. Phys.
Chem. 66 (1962) 2626–2635.
[30] H. P. Schwan, Electrical properties of bound water, Ann. N.Y. Acad.
Sci. 125 (1965) 344–354.
[31] E. H. Grant, Structure of water neighboring proteins peptides and amino
acids as deduced from dielectric measurements, Ann. N.Y. Acad. Sci.
125 (1965) 418–427.
[32] E. H. Grant, S. E. Keefe, S. Takashima, Dielectric behavior of aque-
ous solutions of bovine serum albumin from radiowave to microwave
frequencies, J. Phys. Chem. 72 (1968) 4373–4380.
[33] B. E. Pennock, H. P. Schwan, Further observations on electrical proper-
ties of hemoglobin-bound water, J. Phys. Chem. 73 (1969) 2600–2610.
[34] N. Nandi, K. Bhattacharyya, B. Bagchi, Dielectric relaxation and sol-
vation dynamics of water in complex chemical and biological systems,
Chem. Rev. 100 (2000) 2013–2045.
[35] Y. Feldman, I. Ermolina, Y. Hayashi, Time domain dielectric spec-
troscopy study of biological systems, IEEE Trans. Dielectr. Electr. Insul.
10 (2003) 728–753.
[36] A. Knocks, H. Weingartner, The dielectric spectrum of ubiquitin in
aqueous solution, J. Phys. Chem. B 105 (2001) 3635–3638.
[37] G. E. Thomas, S. Bone, G. Drago, Determination of protein denatura-
tion and glass transition temperatures using high-frequency time domain
reflectometry, J. Phys. Chem. B 112 (2008) 15903–15906.
[38] P. Sasisanker, H. Weingartner, Hydration Dynamics of Water near an
Amphiphilic Model Peptide at Low Hydration Levels: A Dielectric Re-
laxation Study, Chemphyschem 9 (2008) 2802–2808.
[39] P. Lunkenheimer, U. Schneider, R. Brand, A. Loidl, Glassy dynamics,
Contemp. Phys. 41 (2000) 15–36.
[40] F. Kremer, A. Schonhals (Eds.), Broadband dielectric spectroscopy,
Springer, Berlin, 2002.
[41] H. Pauly, H. P. Schwan, Dielectric properties and ion mobility in ery-
throcytes, Biophys. J. 6 (1966) 621–639.
[42] J. Krupa, B. Kwiatkow, J. Terlecki, Method of calculating conductivity
of human erythrocytes interior based on measurement of electric magni-
tudes of suspension, Biophysik 8 (1972) 227–236.
[43] O. Schanne, P. Ruiz, E. Ceretti, Impedance measurements in biological
cells, John Wiley & Sons, 1978.
[44] R. Pethig, Dielectric and electronic properties of biological materials,
John Wiley & Sons, Chichester, 1979.
[45] S. Gabriel, R. W. Lau, C. Gabriel, The dielectric properties of biological
tissues. 2. Measurements in the frequency range 10 Hz to 20 GHz, Phys.
Med. Biol. 41 (1996) 2251–2269.
[46] S. Gabriel, R. W. Lau, C. Gabriel, The dielectric properties of biological
tissues. 3. Parametric models for the dielectric spectrum of tissues, Phys.
Med. Biol. 41 (1996) 2271–2293.
[47] K. R. Foster, H. P. Schwan, Dielectric properties of tissues and biological
materials: a critical review, Crit. Rev. Biomed. Eng. 17 (1989) 25–104.
[48] K. R. Foster, H. P. Schwan, Dielectric properties of tissues, in: C. Polk,
E. Postow (Eds.), Handbook of Biological Effects of Electromagnetic
Fields, CRC Press, Boca Raton, 2. edition, 1995.
[49] G. H. Markx, C. L. Davey, The dielectric properties of biological cells at
radiofrequencies: Applications in biotechnology, Enzyme and Microbial
Technology 25 (1999) 161–171.
[50] R. Lisin, B. Z. Ginzburg, M. Schlesinger, Y. Feldman, Time domain
dielectric spectroscopy study of human cells.1. Erythrocytes and ghosts,
Biochim. Biophys. Acta 1280 (1996) 34–40.
[51] T. X. Zhao, Electrical impedance and hematocrit of human blood with
various anticoagulants, Physiol. Meas. 14 (1993) 299–307.
[52] F. Bordi, C. Cametti, R. Misasi, R. Depersio, G. Zimatore, Conduc-
tometric properties of human erythrocyte membranes: dependence on
haematocrit and alkali metal ions of the suspending medium, Eur. Bio-
phys. J. 26 (1997) 215–225.
[53] F. Jaspard, M. Nadi, A. Rouane, Dielectric properties of blood: an inves-
tigation of haematocrit dependence, Physiol. Meas. 24 (2003) 137–147.
[54] H. Beving, L. E. G. Eriksson, C. L. Davey, D. B. Kell, Dielectric-
properties of human blood and erythrocytes at radio frequencies (0.2-10
MHz) - dependence on cell-volume fraction and medium composition,
Eur. Biophys. J. 23 (1994) 207–215.
[55] J. Z. Bao, C. C. Davis, M. L. Swicord, Microwave dielectric measure-
ments of erythrocyte suspensions, Biophys. J . 66 (1994) 2173–2180.
[81] H. P. Schwan, Electrode polarization impedance and measurements in
biological materials, Ann. N. Y. Acad. Sci. 148 (1968) 191–209.
[82] J. R. MacDonald, I. D. Raistrick, D. R. Franceschetti, The electri-
in: Impedance Spec-
cal analogs of physical and chemical processes,
troscopy, John Wiley & Sons, New York, 1. edition, 1987.
[83] F. Bordi, C. Cametti, T. Gili, Reduction of the contribution of electrode
polarization effects in the radiowave dielectric measurements of highly
conductive biological cell suspensions, Bioelectrochem. 54 (2001) 53–
61.
[84] J. Zarzycki, Glasses and the vitreous state, Cambridge University Press,
Cambridge, 1. edition, 1991.
[56] T. Chelidze, Dielectric spectroscopy of blood, J. Non-Cryst. Solids 305
[85] P. Debye, Einige resultate einer kinetischen theorie der isolatoren,
(2002) 285–294.
[57] Y. Hayashi, I. Oshige, Y. Katsumoto, S. Omori, A. Yasuda, K. Asami,
Dielectric inspection of erythrocyte morphology, Phys. Med. Biol. 53
(2008) 2553–2564.
[58] H. Pauly, H. P. Schwan,
Uber die Impedanz einer Suspension von
kugelformigen Teilchen mit einer Schale, Z. Naturforsch. 14B (1959)
125–131.
[59] T. Hanai, Theory of the dielectric dispersion due to the interfacial polar-
ization and its application to emulsions, Kolloid Z. 171 (1960) 23–31.
[60] H. Looyenga, Dielectric constants of heterogeneous mixtures, Physica
31 (1965) 401–406.
[61] K. Asami, Characterization of heterogeneous systems by dielectric spec-
troscopy, Prog. Polymer Science 27 (2002) 1617–1659.
Phys. Z. 3 (1912) 97–100.
[86] L. Onsager, Electric moments of molecules in liquids, J. Am. Chem.
Soc. 58 (8) (1936) 1486–1493.
[87] U. Schneider, P. Lunkenheimer, A. Pimenov, R. Brand, A. Loidl, Wide
range dielectric spectroscopy on glass-forming materials: an experimen-
tal overview, Ferroelectrics 249 (2001) 89–98.
[88] R. Bohmer, M. Maglione, P. Lunkenheimer, A. Loidl, Radio-frequency
dielectric measurements at temperatures from 10 K to 450 K, J. Appl.
Phys. 65 (1989) 901–904.
[89] L. Nyikos, T. Pajkossy,
Fractal dimension and fractional power
frequency-dependent impedance of blocking electrodes, Electrochim.
Acta 30 (1985) 1533–1540.
[90] S. H. Liu, Fractal model for the ac response of a rough interface, Phys.
[62] T. Hanai, A remark on theory of dielectric dispersion due to interfacial
Rev. Lett. 55 (1985) 529–532.
polarization, Kolloid Z. 175 (1961) 61–62.
[91] T. Pajkossy, L. Nyikos,
Impedance of fractal blocking electrodes,
J.
[63] T. Hanai, Electrical properties of emulsions,
in: Emulsion Science,
Electrochem. Soc. 133 (1986) 2061–2064.
Academic Press, London, 1968.
[92] A. K. Jonscher, Dielectric relaxations in solids, Chelsea Dielectrics
[64] Y. Katsumoto, Y. Hayashi, I. Oshige, S. Omori, N. Kishii, A. Yasuda,
K. Asami, Dielectric cytometry with three-dimensional cellular model-
ing, Biophys. J. 95 (2008) 3043–3047.
[65] P. Lunkenheimer, V. Bobnar, A. Pronin, A. Ritus, A. Volkov, A. Loidl,
Origin of apparent colossal dielectric constants, Phys. Rev. B 22 (2002)
052105.
[66] P. Debye, Polar molecules, Dover Publications, Inc., New York, 1929.
[67] H. Sillescu, Heterogeneity at the glass transition: a review,
J. Non.
Cryst. Solids 243 (1999) 81–108.
Press, London, 1983.
[93] S. R. Elliott, Ac conduction in amorphous-chalcogenide and pnictide
semiconductors, Advances In Physics 36 (1987) 135–218.
[94] S. R. Elliott, A. P. Owens, The diffusion-controlled relaxation model for
ionic transport in glasses, Philos. Mag. B 60 (1989) 777–792.
[95] K. Funke,
Ion dynamics and correlations - translational and localized
ionic hopping motion in solid electrolytes, Philos. Mag. A 68 (1993)
711–724.
[96] H. Fricke, Relation of the permittivity of biological cell suspensions to
[68] M. D. Ediger, Spatially heterogeneous dynamics in supercooled liquids,
fractional cell volume, Nature 172 (1953) 731–732.
Annu. Rev. Phys. Chem. 51 (2000) 99–128.
[69] S. Havriliak, S. Negami, A complex plane analysis of alpha-dispersions
in some polymers, J. Polym. Sci. C 14 (1966) 99–117.
[70] S. Havriliak, S. Negami, A complex plane analysis of dielectric and me-
chanical relaxation process in some polymers, Polymer 8 (1967) 161–
210.
[71] K. S. Cole, R. H. Cole, Dispersion and absorption in dielectrics,
J.
Chem. Phys. 9 (1941) 341–351.
[97] C. L. Davey, H. M. Davey, D. B. Kell, On the dielectric properties of
cell suspensions at high volume fractions, Bioelectrochem. Bioenerg. 28
(1992) 319–340.
[98] F. Bordi, C. Cametti, T. Gili, Dielectric spectroscopy of erythrocyte cell
suspensions. a comparison between Looyenga and Maxwell-Wagner-
Hanai effective medium theory formulations, J. Non-Cryst. Solids 305
(2002) 278–284.
[99] H. Pfutzner, Dielectric analysis of blood by means of a raster-electrode
[72] D. W. Davidson, R. H. Cole, Dielectric relaxation in glycerine, J. Chem.
technique, Med. Biol. Eng. Comput. 22 (1984) 142–146.
Phys. 18 (1950) 1417–1417.
[73] R. H. Cole, D. W. Davidson, High frequency dispersion in n - propanol,
J. Chem. Phys. 20 (1952) 1389 –1391.
[74] K. Asami, T. Yamaguchi, Electrical and morphological changes of hu-
man erythrocytes under high hydrostatic pressure followed by dielectric
spectroscopy, Annals Biomed. Eng. 27 (1999) 427–435.
[75] E. Gheorghiu, On the limits of ellipsoidal models when analyzing di-
electric behavior of living cells - emphasis on red blood cells, Ann. N.Y.
Acad. Sci. 873 (1999) 262–268.
[76] A. Di Biasio, C. Cametti, Effect of the shape of human erythrocytes on
the evaluation of the passive electrical properties of the cell membrane,
Bioelectrochemistry 65 (2005) 163–169.
[77] C. Merla, M. Liberti, F. Apollonio, C. Nervi, G. D’Inzeo, Dielectric
spectroscopy of blood cells suspensions: study on geometrical structure
of biological cells., Conf Proc IEEE Eng Med Biol Soc 1 (2006) 3194–7.
[78] J. Sudsiri, D. Wachner, J. Gimsa, On the temperature dependence of the
dielectric membrane properties of human red blood cells, Bioelectro-
chemistry 70 (2007) 134–140.
[79] Y. Hayashi, Y. Katsumoto, I. Oshige, S. Omori, A. Yasuda, K. Asami,
The effects of erythrocyte deformability upon hematocrit assessed by the
conductance method, Phys. Med. Biol. 54 (2009) 2395–2405.
[80] J. L. Oncley, The investigation of proteins by dielectric measurements,
Chem. Rev. 30 (1942) 433–450.
[100] A. Pribush, H. J. Meiselman, D. Meyerstein, N. Meyerstein, Dielectric
approach to the investigation of erythrocyte aggregation: I. experimental
basis of the method, Biorheology 36 (1999) 411–423.
[101] A. Pribush, H. J. Meiselman, D. Meyerstein, N. Meyerstein, Dielec-
tric approach to investigation of erythrocyte aggregation. ii. kinetics of
erythrocyte aggregation-disaggregation in quiescent and flowing blood,
Biorheology 37 (2000) 429–441.
[102] F. Bordi, C. Cametti, A. Dibiasio, Determination of cell membrane pas-
sive electrical properties using frequency domain dielectric spectroscopy
technique - a new approach, Biochim. Biophys. Acta 1028 (1990) 201–
204.
[103] C. L. Davey, D. Kell, Electrical field phenomena in biological systems,
volume 21, Institute of Physics, London, 1989.
[104] C. Gabriel, S. Gabriel, Dielectric properties of body tissues - appendix
b: part 1: literature survey http://niremf.ifac.cnr.it/tissprop/#over, 1997-
2007.
[105] C. Ballario, A. Bonincontro, C. Cametti, A. Rosi, L. Sportelli, Con-
ductivity of normal and pathological human erythrocytes (homozygous
beta-thalassemia) at radiowave frequencies, Z. Naturforsch., C: Biosci.
39 (1984) 160–166.
[106] K. Asami, Y. Takahashi, S. Takashima, Dielectric properties of mouse
lymphocytes and erythrocytes, Biochim. Biophys. Acta 1010 (1989)
49–55.
16
[107] B. Bianco, G. P. Drago, M. Marchesi, C. Martini, G. S. Mela, S. Ridella,
Measurements of complex dielectric constant of human sera and ery-
throcytes, IEEE Trans. Instrum. Meas 28 (1979) 290–295.
[108] K. Asami, T. Hanai, N. Koizumi, Dielectric approach to suspensions
of ellipsoidal particles covered with a shell in particular reference to
biological cells, Jpn. J. Appl. Phys. 19 (1980) 359–365.
[109] H. Kienitz, K. N. Marsh, Recommended reference materials for real-
ization of physicochemical properties section - permittivity, Pure Appl.
Chem. 53 (1981) 1847–1862.
[110] U. Kaatze, Complex permittivity of water as a function of frequency and
temperature, J. Chem. Eng. Data 34 (1989) 371–374.
[111] R. Gulich, M. Kohler, P. Lunkenheimer, A. Loidl, Dielectric spec-
troscopy on aqueous electrolytic solutions, Radiat. Environ. Biophys.
48 (2009) 107–114.
[112] U. Kaatze, R. Behrends, R. Pottel, Hydrogen network fluctuations and
dielectric spectrometry of liquids, J. Non-Cryst. Solids 305 (2002) 19–
28.
[113] A. Schonhals, F. Kremer, A. Hofmann, E. W. Fischer, E. Schlosser,
Anomalies in the scaling of the dielectric alpha-relaxation, Phys. Rev.
Lett. 70 (1993) 3459–3462.
[114] W. J. Ellison, Permittivity of pure water, at standard atmospheric pres-
sure, over the frequency range 0-25 thz and the temperature range 0-100
degrees c, J. Phys. Chem. Ref. Data 36 (2007) 1–18.
[115] J. B. Hasted, Aqueous dielectrics, Chapman and Hall, London, 1973.
[116] R. Buchner, J. Barthel, J. Stauber, The dielectric relaxation of water
between 0 degrees C and 35 degrees C, Chem. Phys. Lett. 306 (1999)
57–63.
[117] M. D. Ediger, C. A. Angell, S. R. Nagel, Supercooled liquids and
glasses, J. Phys. Chem. 100 (1996) 13200–13212.
[118] K. L. Ngai, Dynamic and thermodynamic properties of glass-forming
substances, J. Non-Cryst. Solids 275 (2000) 7–51.
[119] C. A. Angell,
Insights into phases of liquid water from study of its
unusual glass-forming properties, Science 319 (2008) 582–587.
[120] K. Asami, Characterization of biological cells by dielectric spec-
troscopy, J. Non-Cryst. Solids 305 (2002) 268–277.
[121] E. H. Grant, Dielectric dispersion in bowine serum albumen, J. Mol.
Biol. 19 (1966) 133–139.
[122] E. H. Grant, B. G. R. Mitton, G. P. South, R. J. Sheppard, Investigation
by dielectric methods of hydration in myoglobin solutions, Biochem. J.
139 (1974) 375–380.
17
|
1210.2562 | 1 | 1210 | 2012-10-09T11:11:02 | Modelling bacterial flagellar growth | [
"physics.bio-ph",
"cond-mat.stat-mech"
] | The growth of bacterial flagellar filaments is a self-assembly process where flagellin molecules are transported through the narrow core of the flagellum and are added at the distal end. To model this situation, we generalize a growth process based on the TASEP model by allowing particles to move both forward and backward on the lattice. The bias in the forward and backward jump rates determines the lattice tip speed, which we analyze and also compare to simulations. For positive bias, the system is in a non-equilibrium steady state and exhibits boundary-induced phase transitions. The tip speed is constant. In the no-bias case we find that the length of the lattice grows as $N(t)\propto\sqrt{t}$, whereas for negative drift $N(t)\propto\ln{t}$. The latter result agrees with experimental data of bacterial flagellar growth. | physics.bio-ph | physics | epl draft
Modelling bacterial flagellar growth
M. Schmitt and H. Stark
Institut fur Theoretische Physik, Technische Universitat Berlin - Hardenbergstrasse 36, 10623 Berlin, Germany, EU
2
1
0
2
t
c
O
9
]
h
p
-
o
i
b
PACS 87.10.Hk -- Lattice models in biological physics
PACS 87.16.Ka -- Filaments in subcellular structure and processes
PACS 87.16.Qp -- Flagella
Abstract -- The growth of bacterial flagellar filaments is a self-assembly process where flagellin
molecules are transported through the narrow core of the flagellum and are added at the distal
end. To model this situation, we generalize a growth process based on the TASEP model by
allowing particles to move both forward and backward on the lattice. The bias in the forward
and backward jump rates determines the lattice tip speed, which we analyze and also compare
to simulations. For positive bias, the system is in a non-equilibrium steady state and exhibits
boundary-induced phase transitions. The tip speed is constant. In the no-bias case we find that
the length of the lattice grows as N (t) ∝
result agrees with experimental data of bacterial flagellar growth.
√t, whereas for negative drift N (t) ∝ ln t. The latter
.
s
c
i
s
y
h
p
[
1
v
2
6
5
2
.
0
1
2
1
:
v
i
X
r
a
Bacterial flagella act as motility devices that allow bac-
teria such as Escherichia coli and Salmonella to swim and
to respond to chemical stimuli by performing chemotaxis
[1]. A flagellum mainly consists of a long helical hollow
structure with a length of up to 20µm and a typical outer
diameter of 0.02µm. The growth mechanism of the flagel-
lar filament is a self-assembly process. Flagellin molecules
are transported through the hollow core of the filament
and are added one by one at the tip of the filament. An
important quantity of such a growth process is the time
dependence of the filament length N (t). A recent model
that treats flagellin molecules as diffusing particles in a
single-file process established a growth function N (t) ∝ √t
[2]. However, the only available experimental data to our
knowledge show a logarithmic growth N (t) ∝ ln t [3]. To
model the growth process and to try to verify the experi-
mental data, we make use of a variant of the well known to-
tally asymmetric simple exclusion process (TASEP). This
model takes into account the important fact that flagellar
growth happens in a single-file process meaning flagellin
molecules cannot pass each other.
The TASEP itself has developed into a paradigmatic
model of non-equilibrium physics [4 -- 9].
In particular,
TASEP models with open boundaries showed to be very
fruitful for the study of non-equilibrium steady states,
i.e., states that are characterized by non-vanishing cur-
rents. These systems have been used to model the mo-
tion of ribosome along mRNA [10], molecular motors along
microtubule filaments [11], or that of cars on a highway
p
q
p
q p
γ
1
i
α
N
Fig. 1: Growing asymmetric simple exclusion process with open
boundary conditions. The labels indicate rates at which par-
ticle hops can occur. At site 1 particles transform into a new
lattice site with rate γ, which then becomes the new site 1.
[12,13], just to name a few applications. All of these mod-
els consider lattices with a fixed number of lattice sites
N . However, there are recent approaches that general-
ize the TASEP to dynamically extending exclusion pro-
cesses. One approach considers the TASEP with a fluc-
tuating boundary that can be pushed away by particles
on the lattice [14], another one takes into account a death
probability of the leading particle [15, 16]. In a different
variant of the TASEP, particles that reach the end of the
lattice can transform into new lattice sites resulting in a
dynamically extending lattice [17]. Initially, the extend-
ing TASEP was formulated to model fungal hyphal growth
[18]. In the reminder of this paper we will generalize this
model to a partially asymmetric model where particles can
move both forward and backward on the lattice. In par-
ticular, we calculate the growth function N (t) and discuss
implications for the bacterial flagellar growth.
The model is defined in the following way (see fig. 1 for
p-1
M. Schmitt et al.
a schematic). Particles enter the lattice with rate α at the
boundary on the right. They either move to the left with
rate q or to the right with rate p so that q + p = 1. Finally,
at the tip of the lattice at site 1 they transform into a
new lattice site with rate γ. A crucial feature of TASEP
models is that particles can only move if the target site is
empty. This hard-core interaction results in a nonlinear
relationship between current J and density ρ, which itself
is the origin of the many interesting features of the TASEP.
In a first step we study systems where particle hops
possess a positive drift q > p. To begin the analysis, we
write down the averaged rate equations for occupancy ni ∈
{0, 1} of site i:
hn1i = qh(1 − n1)n2i − ph(1 − n2)n1i − γhn1i ,
hn2i = qh(1 − n2)n3i − qh(1 − n1)n2i − ph(1 − n3)n2i
+ph(1 − n2)n1i − γhn1n2i ,
hnii = qh(1 − ni)ni+1i − qh(1 − ni−1)nii
−ph(1 − ni+1)nii + ph(1 − ni)ni−1i
+γhn1(ni−1 − ni)i ,
i ≥ 3 .
The term γhn1i in the first equation describes the growth
of the lattice. Further terms proportional to γ occur since
after each growth event the lattice sites are renumbered.
This corresponds to working in the frame of reference of
the tip so that the new site at the left boundary becomes
site 1. To tackle this system of coupled equations, we
employ a mean-field approximation by setting hninji =
hniihnji = ρiρj, where ρi is the density of site i. We
now identify the particle current Ji,j from site i to its
neighboring site j as
J1,0 = γρ1 ,
J2,1 = q(1 − ρ1)ρ2 − p(1 − ρ2)ρ1 ,
(1)
Ji+1,i = q(1 − ρi)ρi+1 − p(1 − ρi+1)ρi − γρ1ρi , i ≥ 2 .
In a steady state, the current J is constant throughout
the lattice. Moreover, since particles at the boundary on
the left are used solely for the growth of the lattice, the
current must exactly match the tip speed J = v = γρ1.
This allows us to solve the set of equations (1) for the
density throughout the lattice
ρ1 = v
γ ,
ρ2 =
γ )
v(1+ p
γ (q−p) ,
q− v
(2)
ρi+1 = v+(v+p)ρi
q−(q−p)ρi
,
i ≥ 2 .
The last equation shows two fixed points (see fig. 2) upon
setting ρi = ρi+1:
ρ± =
q − p − v
2(q − p) ±s(cid:18) q − p − v
2(q − p) (cid:19)2
−
v
q − p
,
(3)
where ρ− is stable and ρ+ is unstable as indicated in fig.
2(a). Iterating eq. (2) results in four possible density pro-
files [see fig. 2(b)]. Whereas the low-density profiles relax
1
+
ρ
i
ρ+
ρ−
ρi
1
i
N
Fig. 2: Sketch of possible solutions to recurrence relation (2)
on the left and corresponding density profiles on the right. ρ−
is a stable fixed point whereas ρ+ is unstable.
1
towards the stable fixed point ρ−, the high-density pro-
files start in close vicinity of the unstable fixed point ρ+
and deviate from it when the other end of the lattice is
reached. Upon changing bias q and boundary conditions
α, γ, it is now possible to identify three different phases.
For low input rate α at lattice site N , the system is in the
low density (LD) phase, whereas low growth or release rate
γ at lattice site 1 leads to the high density (HD) phase.
The crossover between low density and high density phase
is governed by a first-order transition, where both phases
coexist. The locations of the phases as determined by
the mean-field theory are similar to the refined α, γ phase
diagram in fig. 3, which we will discuss below. When
both rates α and γ are high, the fixed points of eq. (2) do
no longer exist and the system is in the maximal current
(MC) phase. The density profile is then no longer con-
trolled by the boundary conditions α and γ. Instead, the
bottleneck in the system is given by the bulk dynamics,
namely due to the implicit hard-core interaction, which
restricts the current or tip speed to its maximum value
JMC = vMC = (3 − 2√2)(q − p). It is determined from
eq. (3) by setting the root to zero, i.e., when the two fixed
points merge. Decreasing the hopping rate bias q below
1 slows down the bulk dynamics which results in an ex-
panding maximal current phase in the phase diagram.
In order to determine the phases just discussed within
the formulated mean-field theory, we extended and gener-
alized the procedure of ref. [17] to q < 1. We shortly sum-
marize the reasoning. Setting ρN −1 = ρN in the mean-
field rate equation for the boundary on the right,
ρN = p(1− ρN )ρN −1 − q(1 − ρN −1)ρN + α(1 − ρN ) , (4)
In the
yields the boundary condition ρN = α/(q − p).
LD phase, ρN equals the stable fixed point ρ− of eq. (3)
which gives an expression for the LD current or tip speed
JLD = vLD. However, the iteration of eq. (2) starting at
ρ2 only converges to the stable fixed point ρ− if it starts
below the unstable fixed point ρ+, i.e., if ρ2 < ρ+. Using
the expression for JLD = vLD, this condition leads to a
phase boundary against the HD phase. In the HD phase
the tip velocity vHD and the high density value ρ+ follow
by setting ρ2 = ρ+. From ρ2 = ρ+ the density then has
to decay towards ρN = α/(q − p) at the boundary on the
p-2
right. This sets up a stationary shock profile at the phase
transition between LD and HD phase. Close to this phase
transition in the HD phase, the shock moves away from
the tip but never reaches the other boundary of the lattice.
Only further away from the phase transition, does the high
density expand throughout the lattice except in a narrow
region close to the boundary on the right.
Monte Carlo simulations of our model were done using
a random sequential update where sites are picked ran-
domly one after another and updated according to rates
α, γ, q. One time step consists of updating all N lattice
sites. The tip velocity v is then determined as a func-
tion of either α or γ for a given value of q. We localize the
HD-MC or LD-MC phase transition when v reaches a con-
stant maximum value which indicates the MC phase. As
already indicated, the resulting phase diagram obtained
by the mean-field approach does not coincide very well
with results from Monte-Carlo simulations. An improved
mean-field calculation that takes into account correlations
between site 1 and 2 by setting hn1n2nii = hn1n2ini, pre-
dicts a more precise phase diagram. This was already ob-
served for q = 1, i.e., in the totally asymmetric case [17].
In contrast to the mean-field approach discussed above,
iteration of eq. (2) then starts at site 3 instead of site
2. In particular, to eliminate hn1n2i in the equation for
ρ3, one needs to evaluate the rate equation for the state
hn2(1 − n1)i. The resulting phase diagram is shown in
fig. 3 and compared to Monte-Carlo simulations. The im-
proved mean-field calculations correctly predict the phase
boundaries between LD and MC phases for various q val-
ues whereas the boundaries between HD and MC phases
still show some deviations from simulations. This was al-
ready observed in [17]. Since determining the HD-MC
boundary involves the starting value of iteration (2), the
result depends strongly on the number of site correlations
that are taken into account at the tip. Figures 4(a) and
(b) show density profiles at constant α, γ that convert
from the respective LD or HD phase at q = 1 to the MC
phase at q = 0.8 and 0.6.
For our purposes the main conclusion is that the system
reaches a non-equilibrium steady state characterized by a
constant current or tip speed v. Hence, the lattice grows
linearly in time: N (t) = vt. This is valid for all three
phases and the actual value of v depends on the parame-
ters α, γ and q. With vMC = (3 − 2√2)(q − p) in the MC
phase, we obtain
N (t) = (3 − 2√2)(q − p)t ,
q > p ,
(5)
which agrees with simulations [see fig. 6(a)]. Note that
Eq. (5) predicts N (t) → 0 for decreasing bias q → 0.5.
However, Monte-Carlo simulations show a non-vanishing
growth of the lattice for q = p. So, we have to study this
case separately.
When q = p = 0.5, the bulk of the lattice is not driven
anymore, i.e., the condition of detailed balance is fulfilled
in the bulk but not at the boundaries. The system is still
Modelling bacterial flagellar growth
γ
1
0.8
0.6
0.4
0.2
0
0
LD
LD
LD
LD
LD
LD
LD
LD
LD
LD
.
.
.
.
.
.
.
.
.
.
8
8
8
8
8
8
8
8
8
8
0
0
0
0
0
0
0
0
0
0
=
=
=
=
=
=
=
=
=
=
q
q
q
q
q
q
q
q
q
q
.
.
.
.
.
.
.
.
.
.
6
6
6
6
6
6
6
6
6
6
0
0
0
0
0
0
0
0
0
0
=
=
=
=
=
=
=
=
=
=
q
q
q
q
q
q
q
q
q
q
.
.
.
.
.
.
.
.
.
.
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
=
=
=
=
=
=
=
=
=
=
q
q
q
q
q
q
q
q
q
q
(a)
(a)
(a)
(a)
(a)
(a)
(a)
(a)
(a)
(a)
MC
MC
MC
MC
MC
MC
MC
MC
MC
MC
(b)
(b)
(b)
(b)
(b)
(b)
(b)
(b)
(b)
(b)
D
D
D
D
D
D
D
D
D
D
H
H
H
H
H
H
H
H
H
H
0.2
0.4
α
0.6
0.8
1
Fig. 3: Phase diagram of extending TASEP. Monte Carlo sim-
ulation (symbols) and improved mean-field calculations (lines)
are shown for q = 1.0, 0.8, and 0.6. Horizontal lines are boun-
daries between HD and MC phases. Vertical lines separate LD
from MC phases. Bullets: parameters of density plots in figs.
4(a), (b).
kept out of equilibrium. For q = p, eqs. (2) reduce to
ρ1 = v/γ,
ρ2 = 2v + v/γ,
ρi+1 = ρi(1 + 2v) + 2v ,
i ≥ 2 .
(6)
lim
i→∞
The recurrence relation does not have a fixed point with
ρ > 0 and the density grows to infinity,
ρi = ∞. To
bypass this behavior, we impose that the density reaches
the physical maximum at the boundary on the right, ρN =
1. For N ≫ 1 and with ρN = 1 one readily shows that
ρ1 = v/γ ≪ 1 and therefore negligible in eqs. (6). Then,
the recurrence relation of (6) is solved by
ρi = (1 + 2v)i−1 − 1 .
(7)
The condition ρN = 1 gives 2 = (1 + 2v)N (t), which in-
dicates that v is time-dependent now. By identifying v(t)
as dN (t)/dt, we can solve for N (t) and arrive at
N (t) = √ln 2√t
for
t ≫ 1 .
(8)
This result is confirmed by simulations as indicated in fig.
6(a). It is independent of α and γ, i.e., the MC phase fills
the whole phase diagram. It also shows that the current
in the lattice is ohmic, J = v = dN (t)/dt ∝ 1/N (t).
We note that the growth function of eq. (8) is consis-
tent with a continuum approach for this particular system.
The non-extending TASEP with fixed N , open bound-
aries, and q = p reduces to the diffusion equation in the
continuum limit. Hence, to obtain a continuous model of
the extending TASEP with q = p, one has to formulate
the diffusion equation on a growing domain. Similar to
writing a reaction-diffusion equation on a growing domain
[19], we find [20]
∂ρ
∂t
=
1
2N (t)2
∂2ρ
∂x2 −
1
dN (t)
N (t)
dt
ρ,
x ∈ [0, 1] ,
(9)
p-3
M. Schmitt et al.
ρ
ρ
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
1
(a)
α=0.3
γ=0.7
q=1.0
q=0.8
q=0.6
(c)
α=1γ=1
100
site
200
q=0.5
eq. (11)
100
site
200
ρ
ρ
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
1
100
site
200
(d) α=1γ=1
q=0.45
q=0.40
q=0.30
25
site
50
Fig. 4: Monte Carlo simulations of density profiles for different
phases: (a) LD phase with crossover to MC phase (q = 0.8, 0.6),
(b) HD phase with crossover to MC phase (q = 0.8, 0.6) [Bullets
in fig. 3 give the location in the phase diagram], (c) q = p, and
(d) q < p.
where we replaced ρi(t) by ρ(x, t) and the growing domain
is mapped on the interval [0, 1]. This equation only has a
stationary solution in the case of diffusive growth N (t) ∝
√t. For ∂ρ/∂t = 0 and using growth function (8), we
arrive at the time-independent equation
1
ln 2
d2ρ
dx2 − ρ = 0 .
(10)
Under the assumption of Dirichlet boundary conditions
ρ(0) = 0 and ρ(1) = 1, where the tip of the lattice is
located at x = 0, the stationary solution reads
ρ(x) =
sinh(√ln 2 x)
sinh√ln 2
.
(11)
Monte Carlo simulations confirm this density profile (see
fig. 4(c)). The deviation from a linear density profile, i.e.,
the solution to the non-extending TASEP with q = p, is
rather small.
When q < p, the particles in the bulk prefer to hop to
the right, even though the boundary conditions enforce a
current from right to left. As a result, the growth func-
tion (5) becomes negative indicating that the mean-field
calculation is not applicable in this case. Studies of the
non-extending TASEP with q < p exist [21, 22] and exact
solutions revealed a reverse bias phase where the parti-
cle current decreases exponentially with system size N .
In our case, we expect that the lattice grows very slowly
for q < p. Furthermore, most of the particles will stay
in a domain at the right boundary, whereas near the left
boundary most of the sites are vacant. Monte Carlo sim-
ulations confirm that there is a domain interface between
an empty domain and a "jammed" domain. As fig. 4(d)
shows, this domain wall is roughly located in the middle
of the lattice. Now, whenever the left-most particle in the
(b)
α=0.7
γ=0.3
q=1.0
q=0.8
q=0.6
x
d1
τ1
d2
τ2
t
Fig. 5: Trajectory of a biased random walker with fixed wall
at the origin. The random walker starts at x = 0. The height
of an excursion is called di, the length is called τi.
lattice reaches the tip, the lattice will extend by one site.
After such a growth event, the next particle in the lattice
becomes the left-most particle. Therefore, the growth of
the lattice is equivalent to finding the probability that a
biased random walker reaches a maximum distance from
a bounding wall. In this picture, the wall is given by the
domain interface between the empty domain on the left
and the "jammed" domain on the right and the maximum
distance is the distance of the lattice tip from the domain
interface. Figure 5 depicts a trajectory of a biased random
walker with fixed wall at the origin. The random walker
is on excursions that last for a timespan of τi and have a
maximum distance di from the origin. Ref. [23] determines
the probability for di to be smaller than some threshold c
in the long-time limit as
P (di < c) = 1 − e−c/κ ,
(12)
where κ characterizes the step-length distribution W (Xi)
of the random walker by the condition
heXi/κi = 1 .
(13)
In our case, the respective probabilities for hops to the
left or right, i.e., away from or towards the origin, are
W (Xi = +1) = q and W (Xi = −1) = p. The condition
(13) gives
qe1/κ + pe−1/κ = 1 ,
which results in
κ =(cid:18)ln
p
q(cid:19)−1
.
(14)
At time t the random walker made t/hτii excursions and
the maximum Mt is given by
Mt = max(d1, d2, . . . , dt/hτii) .
(15)
The probability that Mt < c becomes
P (Mt < c) = P (d1 < c) · P (d2 < c) . . . P (dt/hτii < c)
= (cid:16)1 − e−c/κ(cid:17)t/hτii
.
Replacing c by c + κ ln t yields
P (Mt < c + κ ln t) =(cid:18)1 −
e−c/κ
t (cid:19)
t/hτii
(16)
p-4
106
106
104
104
102
102
N
N
(a) α=1γ=1
(a) α=1γ=1
q=1.0
q=0.8
q=0.6
q=0.5
eq. (5)
eq. (8)
102
102
104
104
106
106
t
t
150
150
100
100
N
N
50
50
0
0
(b) α=1γ=1
(b) α=1γ=1
q=0.45
q=0.40
q=0.30
eq. (19)
c=-48
c=-48
7
7
= - 1
= - 1
c
c
c = - 4 . 5
c = - 4 . 5
102
102
104
104
106
106
t
t
0
10-2
10-3
v
10-4
10-5
10-6
0
Modelling bacterial flagellar growth
N
10000
20000
30000
10
1
/
)
h
m
µ
(
v
0.1
0
L (µm)
15
5
L (µm)
10
15
Fig. 6: Lattice growth determined from Monte Carlo simula-
tions. (a) Upper 3 curves: positive drift with q > p, bottom
curve: no drift with q = p = 0.5. (b) Negative drift with q < p.
and in the limit of large t/hτii, one obtains the Gumbel
distribution [24]
P (Mt < c + κ ln t) = exp(−e−c/κ/hτii) .
(17)
This means that the probability for a maximum excursion
stays constant in time, when in the long-time limit the
maximum grows as
Mt = κ ln t + c ,
(18)
where c is an undetermined constant. Since the domain
interface is situated approximately in the middle of the
lattice, the growth function of the whole lattice is then
N (t) = 2(cid:18)ln
p
q(cid:19)−1
with tip speed
ln t + c ,
q < p ,
(19)
v =
dN (t)
dt ∝ exp(−
N
2κ
) .
(20)
Interestingly, the dependence of v on N agrees quali-
tatively with the dependence of the current on N in
the non-extending TASEP with negative drift [21, 22].
Simulations of an extending lattice for different q < 0.5
agree very well with this theory by reproducing the log-
arithmic growth law (19) with its prefactor 2κ [see fig.
6(b)]. This is even true for q approaching 0.5 when the
domain interface in fig. 4(d) becomes broader and, strictly
speaking, the bounding wall of our model does not exhibit
a hard-core repulsion for the random walker. Just as in
the q = p case, simulations of an extending lattice are
independent of α and γ. Note that the analogy with a
biased random walker bounded by a wall also gives the
growth laws N (t) ∝ t for q > p and N (t) ∝ √t for q = p.
We checked this by Monte-Carlo simulations.
Summarizing our results, we found three distinct
regimes for the bias q in the extending TASEP with qual-
itatively different growth functions. The experimental ob-
servation of logarithmic growth for flagellar filaments [3]
Fig. 7: Lattice growth speed from simulations for small neg-
ative bias (q = 0.4999). The inset shows flagellar elongation
data of three samples of Salmonella and fits, taken from [3].
mentioned in the introduction is in qualitative agreement
with our model when the drift is negative (q < 0.5). In
order to compare our result with the experiment in more
detail, we carried out simulations with very small negative
drift. By setting q = 0.4999, it was possible to achieve
a lattice length of the order of N = 20000 in reason-
able simulation time while still reaching the logarithmic
growth regime. If each growth event mimics one flagellin
molecule, N = 20000 corresponds to a filament of about
10µm. For our estimate, we used the length 55 A for a
flagellin molecule and the fact that 11 flagellin molecules
make up the circular cross section of the hollow flagellar
filament. The speed of elongation of flagellar filaments in
the experiment was found to decrease exponentially with
increasing length for filament lengths in the range from
4 to 14µm (see inset of fig. 7) [3]. This is in agreement
with our simulations as illustrated in fig. 7. However,
it remains to establish an explanation why the flagellin
in the hollow filament should exhibit a negative drift to-
wards the cell body.
It has been suggested that proton
motive force (PMF) is biasing the Brownian motion of
the flagellin translocation process [25]. The details of this
mechanism are, however, very unclear at this stage and
should be subject to further research.
REFERENCES
[1] Berg H. C., E. coli in Motion (Springer, New York) 2004.
[2] Keener J. P., Bull. Math. Biol., 68 (2006) 1761.
[3] Iino T., J. Supramol. Struct., 2 (1974) 372.
[4] Derrida B., Domany E. and Mukamel D., J. Stat. Phys,
69 (1992) 667.
[5] Derrida B., Evans M. R., Hakim V. and Pasquier V.,
J. Phys. A: Math. Gen., 26 (1993) 1493.
[6] Schutz G. and Domany E., J. Stat. Phys., 72 (1993) 277.
[7] Krug J., Phys. Rev. Lett., 67 (1991) 1882.
[8] Parmeggiani A., Franosch T. and Frey E., Phys. Rev.
Lett., 90 (2003) 086601.
p-5
M. Schmitt et al.
[9] Mukamel D., Soft and Fragile Matter: Nonequilibrium
Dynamics, Metastability and Flow, edited by Cates M. E.
and Evans M. R. (Taylor & Francis) 2000, pp. 237-258.
[10] MacDonald C. T., Gibbs J. H. and Pipkin A. C.,
Biopolymers, 6 (1968) 1.
[11] Frey E., Parmeggiani A. and Franosch T., Genome
Informatics, 15 (2004) 46.
[12] Barlovic R., Santen L., Schadschneider A. and
Schreckenberg M., Eur. Phys. J. B, 5 (1998) 793.
[13] Chowdhury D., Santen L. and Schadschneider A.,
Phys. Rep, 329 (2000) 199.
[14] Nowak S. A., Fok P. W. and Chou T., Phys. Rev. E,
76 (2007) 031135.
[15] Dorosz S., Mukherjee S. and Platini T., Phys. Rev.
E, 81 (2010) 042101.
[16] Kwon H. W. and Choi M. Y., Phys. Rev. E, 82 (2010)
013101.
[17] Sugden K. E. P. and Evans M. R., J. Stat. Mech,
(2007) P11013.
[18] Sugden K. E. P., Evans M. R., Poon W. C. K. and
Read N. D., Phys. Rev. E, 75 (2007) 031909.
[19] Crampin E. J., Gaffney E. A. and Maini P. K., Bull.
Math. Biol., 61 (1999) 1093.
[20] Schmitt M., A growth model
for bacterial flagella
(diploma thesis, Technische Universitat Berlin) 2010.
[21] Blythe R. A., Evans M. R., Colaiori F. and Essler
F. H. L., J. Phys. A: Math. Gen., 33 (2000) 2313.
[22] De Gier J., Finn C. and Sorrell M., arXiv,
(2011)
1107.2744v1 [cond-mat.stat-mech].
[23] Iglehart D. L., Annals Math. Stat., 43 (1972) 627.
[24] Gumbel E. J., Statistical theory of extreme values and
some practical applications, Vol. 33 (U.S. National Bu-
reau of Standards Applied Mathematics Series, Washing-
ton, DC) 1954.
[25] Minamino T., Imada K. and Namba K., Mol. BioSyst.,
4 (2008) 1105.
p-6
|
1106.2038 | 1 | 1106 | 2011-06-10T12:38:14 | Folding and unfolding of a triple-branch DNA molecule with four conformational states | [
"physics.bio-ph",
"cond-mat.stat-mech"
] | Single-molecule experiments provide new insights into biological processes hitherto not accessible by measurements performed on bulk systems. We report on a study of the kinetics of a triple-branch DNA molecule with four conformational states by pulling experiments with optical tweezers and theoretical modelling. Three distinct force rips associated with different transitions between the conformational states are observed in the folding and unfolding trajectories. By applying transition rate theory to a free energy model of the molecule, probability distributions for the first rupture forces of the different transitions are calculated. Good agreement of the theoretical predictions with the experimental findings is achieved. Furthermore, due to our specific design of the molecule, we found a useful method to identify permanently frayed molecules by estimating the number of opened basepairs from the measured force jump values. | physics.bio-ph | physics |
Folding and unfolding of a triple-branch DNA
molecule with four conformational states
Sandra Engela∗, Anna Alemanyb,c, Nuria Fornsb,c,
Philipp Maassa†, and Felix Ritortb,c ‡
aFachbereich Physik, Universitat Osnabruck,
Barbarastr. 7, 49076 Osnabruck, Germany
bDepartament de F´ısica Fonamental, Facultat de F´ısica,
Universitat de Barcelona, Diagonal 647, 08028 Barcelona, Spain
cCIBER-BBN Networking center on Bioengineering,
Biomaterials and Nanomedicine, Spain
Abstract
Single-molecule experiments provide new insights into biological processes
hitherto not accessible by measurements performed on bulk systems. We report
on a study of the kinetics of a triple-branch DNA molecule with four conforma-
tional states by pulling experiments with optical tweezers and theoretical mod-
elling. Three distinct force rips associated with different transitions between the
conformational states are observed in the folding and unfolding trajectories. By
applying transition rate theory to a free energy model of the molecule, probability
distributions for the first rupture forces of the different transitions are calculated.
Good agreement of the theoretical predictions with the experimental findings is
achieved. Furthermore, due to our specific design of the molecule, we found a
useful method to identify permanently frayed molecules by estimating the number
of opened basepairs from the measured force jump values.
Keywords: Nonequilibrium systems, single-molecule experiments,
optical tweezers, DNA
PACS: 82.37.-j, 05.70.Ln, 82.39.Pj, 87.80.Nj
1 Introduction
In recent years, single-molecule experiments became of great importance in biophysi-
cal research since progress in nano- and microscale manufacturing technologies facili-
tated the design of scientific instruments with sufficient sensitivity and precision to en-
able the controlled manipulation of individual molecules (for reviews, see, for example,
[22, 10]). In contrast to the traditionally used bulk assays, where individual biomolec-
ular dynamics can get masked, single-molecule experiments provide new insights into
∗[email protected]
†[email protected],
phone:
0049-541-969-3460,
http://www.statphys.uni-osnabrueck.de
‡[email protected],
phone:
0034-934035869,
http://www.ffn.ub.es/ritort/index.html
fax:
fax:
0049-541-969-2351,
0034-934021149,
1
Folding and unfolding of a triple-branch DNA molecule with four conformational states
2
the thermodynamics and kinetics of biophysical and biochemical processes hitherto not
accessible. They complement standard spectroscopy and microscopy methods used in
molecular biology and biochemistry and hence have to be regarded as an important
source of additional information helping in the interpretation of biomolecular pro-
cesses. Furthermore, single-molecule experiments permit the measurement of small
energies and the detection of large fluctuations.
The manipulation of single molecules offers a powerful new tool in molecular and
cellular biophysics allowing for the exploration of processes occuring inside the cell at
an unprecedented level. To instance just a few of the recently investigated biochem-
ical processes: the transport of matter through pores or channels [19, 20, 1], interac-
tions between DNA and proteins [13] or DNA and RNA [36], the motion of single-
molecular motors [2, 4, 32], DNA transcription and replication [34, 15], virus infection
[26, 8], DNA condensation [23] and ATP generation [35]. In addition, the structure
of biological networks [30] and the viscoelastic and rheological properties of the DNA
[27, 31, 6, 29] have been studied.
An important class of single-molecule experiments are performed with optical tweez-
ers. By means of an optical trap generated by a focused laser beam, this useful tech-
nique renders it possible to exert forces on micron sized objects, achieving sub-piconewton
and sub-nanometer resolution in force and extension, respectively. Accordingly, one
can study force-induced folding-unfolding dynamics and in this way get insight into
corresponding processes in the cell and typical bond forces. Of particular interest is
the unfolding of DNA molecules, where the hydrogen bonds between the complemen-
tary base pairs (bps) are disrupted. This so-called unzipping is connected to the DNA
replication mechanism.
An interesting field of biophysical studies is the investigation of junctions in molecules
since they present manifold ways to interact with other substances, for instance cations.
Three-way junctions are especially interesting because metal ions such as magnesium
can bind to them and alter the tertiary structure. Here a first step of such a study is
presented where we investigate a molecule with a three-way junction alone, without
cation binding.
Many of the single-molecule experiments so far focused on molecules with a rel-
atively simple free energy lanscape (FEL) exhibiting just two states, a folded and an
unfolded one, or including an additional misfolding state, leading to different kinetic
pathways. In this work we will consider a richer situation, where metastable states as
intermediates occur during the folding-unfolding route. In this context, we will address
the following key questions:
(i) Can a corresponding molecule with such intermediate states be designed on the
basis of a suitable model for a FEL?
(ii) Is it possible to observe the intermediate states by perfoming pulling experiments
with optical tweezers?
(iii) Can phenomenological Bell-Evans kinetic models be applied to describe the
folding-unfolding processes including intermediate states? In particular, when
validating the kinetic theory against the experimental results, how do the first
rupture force distributions compare with the ones predicted by the theory?
A further important aspect that we looked at in some detail is the heterogenity
of molecular folding-unfolding behaviour that we observed in the experiments. By
measuring force-distance curves (FDCs) of several molecules we classified them into
Folding and unfolding of a triple-branch DNA molecule with four conformational states
3
different reproducible patterns. This leads to a useful method to identify irreversible
molecular fraying, a phenomenon which is often observed in single-molecule studies.
2 Description of the experiments
Based on Mfold folding predictions [25, 37] and taking FEL considerations into ac-
count (see sec. 5), we designed and synthesised a DNA molecule which is composed
of three parts and hence referred to as triple-branch molecule. It consists of a stem
as introduced in [21] with 21 bps and two nearly identical hairpin branches which
are formed of 16 bps and a loop with four bases, thus comprising a total number of
114 bases, see fig. 1. To avoid misfolding, the second hairpin branch differs at two
positions from the first one.
hairpin branch 1 hairpin branch 2
16bps + 4bs loop
16bps + 4bs loop
stem
21bps
Total: 114bs
Figure 1: Structure of the triple-branch DNA molecule.
The triple-branch molecule is inserted between two identical short double-stranded
DNA (dsDNA) handles of 29 bps each [9], leading to a total number of 172 bases,
corresponding to a total contour length in the unfolded state of about 100 nm. Each of
these polymer spacers is chemically linked to a bead. One of the beads is retained with
the help of a pipette via air suction, the other one is optically trapped in a laser focus
[28]. On the 5′ end of the DNA molecule, biotin is attached to enable a connection with
Folding and unfolding of a triple-branch DNA molecule with four conformational states
4
a streptavidin-coated bead (SA bead) whose diameter is 1.8 µm. Biotin is a vitamin
which establishes a strong linkage to the proteins avidin and streptavidin. The 3′ end is
modified with the antigen digoxigenin, able to interact with an antidigoxigenin-coated
bead (AD bead). The latter has a diameter of 3.0 µm. Figure 2 shows the different
components of the molecular construct, that is to say the triple-branch DNA molecule,
the handles and the beads, captured in optical trap and micropipette, respectively. Note
that it is not a true-to-scale representation.
The pulling experiments are carried out with a miniaturised dual-beam laser optical
tweezers apparatus [11] at room temperature (≃ 25 °C) and at salt concentration of 1
M NaCl aqueous buffer with neutral pH (7.5) stabilised by Tris HCl and 1 M EDTA.
The dual-beam optical tweezers collect data at 4 kHz and can operate with a feedback
rate of 1 kHz. Spatial resolution constitutes 0.5 nm with a maximal distance range of
≃ 10 µm. Forces up to 100 pN can be achieved, whereas the force resolution is 0.05 pN.
DNA triple-branch molecule
Loop
Hairpin
branches
Loop
Optical trap
Stem
AD bead
Handle 1
Handle 2
SA
bead
Micropipette
molecular extension
absolute distance
Figure 2: Sketch of the experimental setup. In our experiments we measure the relative
distance X rather than its absolute value.
Optical tweezer pulling experiments permit the measurement of the force f as well
as the total distance X between the centre of the optical trap and the tip of the mi-
cropipette, see fig. 2. In the experiment we vary the trap-pipette distance X(t) with
a constant speed v = dX/dt in the range of 45 nm/s to 200 nm/s, which corresponds
to a constant average loading rate r of 3.0 pN/s and 13.4 pN/s in between rip events,
respectively. The experiment consists of loading cycles which in turn are divided into
an unfolding part (during "pulling") and a folding part (during "pushing"). The load-
ing cycles are repeated as long as the tether connection is unbroken. Otherwise a new
connection has to be established, possibly a new molecule must be searched and linked
to a new bead. In sec. 4 we will discuss different patterns found in the measured curves
and present a detailed analysis of two representative molecules. We chose them among
seven molecules exhibiting the first and among five molecules featuring the second
pattern. For each molecule we recorded, on average, approximately 50 cycles. The
pulling speeds, ranging from 45 to 200 nm/s, influence the experimental results only
weekly due to a logarithmic dependence of the first rupture force with the speed. We
found compatible data for sets of similar molecules. In the theoretical analysis of the
data in sec. 5, we concentrate on the largest set of 82 loading cycles for the molecule
Folding and unfolding of a triple-branch DNA molecule with four conformational states
5
at a speed of 200 nm/s. The study of the above mentioned second molecule comprises
55 cycles.
3 Analysis of unfolding and folding trajectories
Based on the design of the triple-branch molecule, we have to distinguish between four
conformational states (see fig. 3):
1. a completely folded molecule.
2. a completely unfolded stem with the hairpin branches still folded.
3. stem and either hairpin branch 1 or 2 are completely unfolded.
4. a completely unfolded molecule.
1
2
d′
0
4
3
d0
d0
d0
x(n, f ) = d0
n = 0
x(n, f ) = ul(f ) + d′
0
x(n, f ) = ul(f )
n = 21
l = 2nd
x(n, f ) = ul(f ) + d0
n = 53
n = 37
l = 2nd + nloopd
l = 2nd + 2nloopd
Figure 3: The four stable or metastable states of the triple-branch molecule: 1 - folded
molecule, 2 - unfolded stem, 3 - stem and one hairpin branch unfolded, 4 - unfolded
molecule. The values of x(n, f ) refer to the end-to-end distance given in eq. (2) with
the number n of opened bps corresponding to the molecular construct shown in fig. 1
and the contour length l according to eq. (3).
Figure 4(a) displays a typical unfolding route and fig. 4(b) a typical unfolding and
folding trajectory in form of a FDC, which in fact records the evolution of the force as a
function of time and relative trap position X1. With rising X, the force f first increases
almost linearly according to an elastic response of the DNA handles, which consist of
dsDNA and are stable over the whole range of forces where the unfolding / folding
of the triple-branch molecule takes place. The overstretching transition of the linkers,
typically at 65 pN, lies much above the forces we explore. At a first rupture force f1
a sudden decrease ("jump") ∆ f1 occurs, which is caused by the unfolding of the stem.
This unfolding goes along with an abrupt change in the length of the molecule when
single-stranded DNA (ssDNA) is released. As a consequence, the bead in the optical
1 In our experiments we measure the relative distance between trap and pipette, X, rather than the absolute
value. The force f exerted on the molecular construct leads to a displacement f /kb of the bead in the optical
trap, where kb = 0.08 pN/nm is the rigidity of the trap. Hence, the relative distance X is related to the relative
molecular extension xm (see fig. 2) by xm = X − f /kb.
Folding and unfolding of a triple-branch DNA molecule with four conformational states
6
]
N
p
[
e
c
r
o
f
18
17
16
15
14
13
12
11
20
18
16
14
150
200
250
100
150
200
250
relative distance X [nm]
(a)
19
18
17
16
15
14
13
12
11
]
N
p
[
e
c
r
o
f
unfolding
f1
f2
f3
folding
f1,rf
f2,rf
f3,rf
100
150
200
250
relative distance X [nm]
(b)
Figure 4: Force as a function of change in the trap-pipette distance (a) for one typical
unfolding trajectory for the first investigated molecule and (b) for one typical unfolding
and folding trajectory for the second investigated triple-branch molecule. Indicated are
the three first rupture forces belonging to the transitions 1 ( f1), 2 ( f2) and 3 ( f3) and the
respective refolding forces, labelled fi,rf. The inset in (a) shows four further unfolding
curves for the first investigated molecule.
trap moves towards the centre of the trap, visible as the force drops. Following the
jump ∆ f1, there is again a linear increase up to the next force rip at a first rupture force
f2, where one of the hairpin branches unfolds, which in turn leads to the force jump
∆ f2. Eventually, the second hairpin branch unzips at a first rupture force f3 with a jump
∆ f3. The linear regime following this last force rip corresponds to the stretching of the
whole molecular construct including handles and the already unfolded triple-branch
molecule.
Upon decreasing X from the completely unfolded state 4, the force first follows
closely the corresponding unfolding part of the trajectory. However, a backward tran-
sition does not occur at ( f3 − ∆ f3) ≃ 16.2 pN, but at a considerably lower value of
(14.0 ± 0.8) pN, cp. fig. 4(b), hence manifesting a hysteresis effect. Moreover, an in-
vestigation of a larger number of folding trajectories reveals that the bases of the hairpin
branches do not always pair conjointly in well-defined events during a short time inter-
Folding and unfolding of a triple-branch DNA molecule with four conformational states
7
val. In contrast, the corresponding force rips during unfolding indicate a cooperative
behaviour of the biomolecule, where the breakage of all hydrogen bonds stabilising the
DNA structure happens almost simultaneously. The second transition takes place at
(13.3 ± 0.8) pN instead of ( f2 − ∆ f2) ≃ 15.5 pN. After both hairpin branches refolded,
one can identify another sharp transition to the folded state 1 around (11.2 ± 0.8) pN,
which is again considerably lower than ( f1 − ∆ f1) ≃ 13.7 pN.
During unfolding, one can assume that the breakage of hydrogen bonds follows the
sequence of the molecular construct. This makes it useful to introduce the number n
of broken bonds as state variable and to calculate a FEL as function of this variable
(see sec. 5). The refolding of the hairpin branches in the folding trajectories exhibit
less sharp transitions, see fig. 4(b). During folding, in particular at the beginning in the
unfolded state, a huge number of secondary structures can be found which implies that
the kinetic pathways are less predefined and accordingly, the transitions get smeared
out. With respect to a theoretical treatment, moreover, a description in terms of the
simple state variable n becomes unlikely to be sufficient. In a refined analysis, many
more configurations should have to be included as relevant states in a coarse-grained
description [17]. Such refined analysis, however, goes beyond the scope of this work
amd we therefore concentrate on the unfolding process in the following.
There are plenty of possible ways to analyse the FDCs in order to find out the first
rupture forces and the force jumps. In our procedure we arranged the normalised data,
i.e. the relative distance X and the force f , in windows of a certain size of data points.
For all consecutive windows we then calculated the slope of the considered data points,
the span, i.e. the maximum distance between the lowest and the highest force value,
and the mean of the relative distance X j as a moving average. Transitions between
the conformational states take place where the slope is minimal and the span maximal
under the condition that an appropriate number of contiguous windows is connected.
Having found the X j of the three force rips, slope and axis intercept are calculated by
linear regression for each conformational state. One can now easily calculate the first
rupture forces as the intersection points with the four fitted lines and extract the force
jump values, as exemplified in fig. 5 for both molecules whose unfolding trajectories
were depicted in figs. 4(a) and 4(b), respectively. Since the data acquisition rate is
]
N
p
[
e
c
r
o
f
20
19
18
17
16
15
14
13
12
]
N
p
[
e
c
r
o
f
19
18
17
16
15
14
13
12
11
150
200
250
100
150
200
250
relative distance X [nm]
(a)
relative distance X [nm]
(b)
Figure 5: The first rupture forces and force jump values of the unfolding trajectories
shown in fig. 4 are extracted as indicated here from the intersection of the fitted grey
lines and the transitions (vertical lines).
In part (a) the procedure is shown for the
molecule used in the theoretical analysis in sec. 5 and in (b) for the molecule with
permanent fraying behaviour.
Folding and unfolding of a triple-branch DNA molecule with four conformational states
8
constant, at higher pulling speeds less data points are collected and therefore the values
of the analysis are broader distributed. The three first rupture forces fi as well as
the jumps ∆ fi are subject to stochastic fluctuations, as can be seen in fig. 4(a) where
we show four unfolding trajectories belonging to different pulling cycles of the same
molecule in the inset. An analysis revealed that the fluctuations of the force jumps ∆ fi
are about ten times smaller than the fluctuations of the fi. Accordingly, in sec. 5, we
will disregard the fluctuations in the ∆ fi and use only their averages ∆ fi that will be
discussed in more detail in the following section.
4 Different unfolding patterns related to the number of
opened bps
Applying the above mentioned procedure to analyse the experimental data of several
molecules, we found two predominating patterns in the unfolding trajectories which
are reflected in the distributions of the first rupture forces fi as follows. In the first
pattern, see fig. 6(a), these distributions have a similar shape for all three force rips. The
histograms indicate the existence of one maximum slightly below 17 pN. The mean
values for the three first rupture forces are f1 = (17.0 ± 0.9) pN, f2 = (16.4 ± 0.6) pN
and f3 = (16.8 ± 0.7) pN. In contrast, in a second pattern we detected a strikingly
lower value for the first rupture force of the first rip f1 = (14.6 ± 0.8) pN, as depicted in
fig. 6(b), whereas the other two rip forces lie basically in the same range of 16 to 18 pN.
As in the former case, the second rip tends to have a slightly smaller first rupture force,
f2 = (16.5 ± 0.8) pN, than the third rip, f3 = (17.2 ± 0.5) pN.
The small f1 observed in the second molecule suggests the occurrence of permanent
molecular fraying. Obviously less force is needed to unfold the stem than typically, cp.
fig. 4(a) with 4(b), since some bps at its basis are partly or completely melted. In other
words, during folding, this molecule does not reach an entirely folded state but some
bps of the stem next to the handles remain irreversibly and permanently open. This
phenomenon has been observed previously in several pulling experiments [11, 33]. A
possible reason for irreversible fraying is the formation of reactive oxidative species
due to the impact of the laser light of the optical trap leading to a degradation of the
DNA bases [12]. The so generated singlet oxygens are known to oxidise certain nucleic
acids, such as guanine and thymine, irreversibly. This could explain our observation
that once a molecule shows fraying it does not change back again to normal behaviour.
Due to the fact that we work with polystyrene microspheres which are more prone to
photodamage than the DNA bases themselves, their wide ranging interaction with the
bases might be reduced replacing polystyrene by silica beads, which exhibit consid-
erably minor irreversible oxidative damage. It would be very interesting to carry out
such experiments.
The insets in figs. 6(a) and 6(b) depict the corresponding force jump distributions
of both molecules. While the first molecule possesses a large first force jump value
of ∆ f1 = (1.3 ± 0.2) pN and two smaller force jumps at the second and third rip of
∆ f2 = (0.9 ± 0.2) pN and ∆ f3 = (1.0 ± 0.2) pN, the frayed molecule features three
force jumps of approximately the same value, i.e. ∆ f1 = (0.93 ± 0.05) pN, ∆ f2 =
(0.95 ± 0.06) pN and ∆ f3 = (0.97 ± 0.04) pN, respectively. This illustrates clearly the
influence of irreversible fraying in the latter case since, roughly estimated, the same
number of bps is expected to open in all three rips which should not be the case in an
entirely folded molecule.
Folding and unfolding of a triple-branch DNA molecule with four conformational states
9
0.3
0.2
0.1
0
0.3
0.2
0.1
0
y
c
n
e
u
q
e
r
f
e
v
i
t
a
l
e
r
y
c
n
e
u
q
e
r
f
e
v
i
t
a
l
e
r
0.2
0.1
0
0.3
0.2
0.1
0
0.5
1
1.5
force jump [pN]
first rip
second rip
third rip
14
15
16
17
18
19
first rupture force [pN]
(a)
0.8
0.9
1
force jump [pN]
first rip
second rip
third rip
12 13 14 15 16 17 18
first rupture force [pN]
(b)
Figure 6: Histograms of the three first rupture forces during unfolding for two repre-
sentative triple-branch molecules: (a) the one used in the theoretical analysis in sec. 5
and (b) the one exhibiting permanent fraying behaviour. Note that the average force
value of the first rip has decreased in (b) as compared to (a). The insets show the
corresponding histograms for the force jump values.
During a force rip, the relative distance X is constant and thus ∆X = 0. Therefore
the change in the relative molecular extension,
∆xm = ∆X − ∆ f /keff = ∆x(n, f ),
(1)
is only related to the force jump ∆ f and the combined stiffness of bead and handles
keff, given by 1/keff = 1/kb + 1/kh, where kb is the trap stiffness and kh the rigidity
of the handles, respectively. With the help of a linear least squares fit, the effective
stiffness keff of the molecular construct is extracted from the average slope of the FDCs
and amounts to (0.067 ± 0.003) pN/nm. Note that, as expected, keff is smaller than kb
(≃ 0.08 pN/nm).
For a certain force value f and assuming an elastic model for the released ssDNA,
the number n of opened bps is related univocally to the equilibrium end-to-end distance
of the DNA molecule x(n, f ), whereas its change ∆x(n, f ), in turn, equals ∆xm. Using
this relation, it is now possible to estimate the change in the number ∆ni of bps which
Folding and unfolding of a triple-branch DNA molecule with four conformational states 10
are opened sequentially during each force rip. Considering only the configurations of
the four conformational states, the equilibrium end-to-end distance x(n, f ) can be de-
composed into two parts, cp. fig. 3. The first part, the elongation ul( f ) of the mean
end-to-end distance of the ssDNA along the force direction, accounts for the ideal elas-
tic response of the ssDNA, where l is the contour length. The second part contains the
contribution of the diameter of stem and hairpin branches, respectively. Acccordingly,
x(n, f ) = ul( f ) +
0, n =21 + 16 + 16 bps
d0,
d′
0,
n = 0 and 21 + 16 bps
n =21 bps
,
(2)
where d0 ≃ 2 nm, in accordance with the diameter of the B-DNA helix. The exact value
of the diameter contribution of both hairpin branches d′
0, when the stem is unfolded,
depends on the orientation of the branches. We set d′
0 = 2d0 as a working value2.
Regarding the contour length l, which depends, amongst others (cp. sec. 5), on the
number n of opened bps, and considering again solely the configurations of the four
conformational states, one gets
0,
nloopd,
2nloopd,
n = 0 and 21 bps
n =21 + 16 bps
n =21 + 16 + 16 bps
,
(3)
l = 2nd +
where the interphosphate distance d is taken to be 0.59 nm/base and nloop = 4 is the
number of bases per loop. For every force rip fi the corresponding number of opened
bps ni is calculated separately by considering the differences of l between the states.
Different types of models can be used to calculate ul( f ). Prominent examples bor-
rowed from polymer physics are the freely jointed chain (FJC) and the worm-like chain
(WLC) model. According to ref. [27], the FJC model includes an extra term, leading
to the expression
ul( f ) = l 1 +
f
Y!"coth b f
kBT! −
kBT
b f # .
(4)
Here Y denotes the Young modulus, b is the Kuhn length, kB the Boltzmann constant
and T the temperature. Typical values of the Kuhn length and the Young modulus under
working conditions of T ≃ 25 °C and 1 M NaCl concentration are b = 1.42 nm and
Y = 812 pN [27] or, as published recently, b = 1.15 nm and Y = ∞ [11], respectively.
In the WLC model [3], the force f (ul), due to an elongation ul, is given by
f (ul) =
kBT
P "
1
4 (1 − ul/l)2 −
1
4
+
ul
l # ,
(5)
and to obtain ul( f ), this equation has to be inverted. Based on the WLC model, we
tested the influence of the persistence length P in a typical range of 1.0 to 1.5 nm [6].
From the data shown in table 1 it is apparent that, depending on the model and
parameters, the results for the estimated change in the number of opened bps vary in an
acceptable range when the errors are taken into account. In addition, one can see that it
is not evident which model and parameters should be considered as the best ones. Good
results are found for the FJC model using recent values of [11] and for the WLC model
2Due to this simplification, the ∆n1 of the first rip is likely to be slightly underestimated and the second
rip's ∆n2 overestimated. However, it will not affect the change in the total number ∆ntot of opened bps since
we consider the change of x(n, f ), and the d′
0 contributions will cancel each other out.
Folding and unfolding of a triple-branch DNA molecule with four conformational states 11
molecule 1
molecule 2
model
FJC
WLC
parameter
b & P [nm], Y [pN]
b = 1.42, Y = 812 [27]
b = 1.15, Y = ∞ [11]
P = 1.0
P = 1.3 [24]
P = 1.5
expected values
∆n1
18 (2)
19 (3)
20 (3)
20 (3)
19 (3)
21
∆n2
14 (2)
15 (2)
16 (3)
15 (2)
15 (2)
16
change in no. of opened bps [bps]
∆n2
∆n3
15 (1)
15 (2)
16 (2)
16 (1)
17 (1)
17 (3)
16 (1)
16 (2)
16 (2)
16 (1)
∆ntot
46 (3)
50 (3)
54 (3)
51 (3)
50 (3)
∆n1
13 (1)
14 (1)
14 (1)
14 (1)
13 (1)
16
53
21
16
∆n3
15 (1)
16 (1)
17 (1)
16 (1)
16 (1)
16
∆ntot
42 (2)
46 (2)
49 (2)
46 (2)
45 (2)
53
Table 1: Overview over the change in the number of opened bps for different models
and parameters for the two representative molecules. The numbers in brackets are the
standard deviations.
with P = 1.3 nm [24]. We chose to work with the FJC model with the parameters of
[11] in sec. 5.
The second molecule indeed reveals a considerably smaller ∆n1, depending on the
model around 13 or 14 bps, so that 7 or 8 bps are not closed after the folding process
is completed. Performing single-molecule experiments without knowing the exact in-
fluence of permanently frayed bps can lead to misinterpreted results. Checking the
appropriate parameters for the polymer models with the help of the change in the num-
ber of opened bps of the hairpin branches, one can estimate the number of irreversibly
frayed bps at the basis of the stem3.
5 Theory for the unfolding kinetics
The kinetics of the unfolding process can be described on a coarse-grained level based
on a Gibbs free energy G(n, f ) as a function of the number n of sequentially opened
bps for an applied force f . For small forces, including f = 0, the FEL is expected
to have a shape as displayed in fig. 7(a). In general, G(n, 0) increases monotonously
with n. However, local minima occur at the metastable states 2, 3 and 4 because there
is an increase of entropy associated with the release of additional degrees of freedom
when the stem-hairpin-junction and end-loops of the hairpin branches are opened. With
rising force the FEL is expected to get tilted, so that the energies of the metastable states
are lowered. With a knowledge of G(n, f ) we can apply standard transition rate theory
and write for the transition rate from state i to i + 1
Γi,i+1( f ) = γ
0
i γi,i+1( f ) ,
(6)
0
i is an attempt rate and γi,i+1( f ) is the Boltzmann factor corresponding to the
where γ
activation barrier ∆Gi,i+1( f ) that has to be surmounted,
γi,i+1( f ) = exp −
∆Gi,i+1( f )
kBT
! .
(7)
When considering only sequential configurations in the evaluation of the FEL, dif-
ferent structures compatible with a given n can occur once the stem is completely un-
folded. These refer to different possibilities of breaking the bps in the hairpin branches
3We like to note that the checking of the change in the number of opened bps can be, in principle, also
applied to non-permanent, reversible molecular fraying.
Folding and unfolding of a triple-branch DNA molecule with four conformational states 12
G(n, f )
1
FEL at f=0
FEL at f=16.21 pN
γ2,3(f )
3
4
)
f
,
n
(
G
2
∆G2,3(f )
140
120
100
80
60
40
20
0
0
n
10
20
30
40
50
no. of opened bps n [bps]
(a)
(b)
Figure 7: (a) Sketch of the expected FEL for the triple-branch molecule at zero force
as a function of the number n of opened bps. (b) FEL calculated from eqs. (8) to (11)
at two different forces in units of the thermal energy kBT .
1 and 2. In order to point out this "degeneration", a new parameter α is introduced,
leading to the FEL G(n, α, f ). At a given n, we must calculate the (restricted) partition
sum over the configurations α to get G(n, f ). In order to find G(n, α, f ), we consider
the following decomposition,
G(n, α, f ) = Gform(n, α) + Gss
str(n, α, f ) − f ∆xss
l ,
(8)
where Gform(n, α) is the free energy of formation of the configuration (n, α), Gss
is the strain energy of the unfolded ssDNA and f ∆xss
nition of ∆xss
str(n, α, f )
l is a Legendre term (for the defi-
l see eq. (10) below).
The free energy of formation Gform(n, α) is written as
Gform(n, α) = Xall bps
gµ,µ+1 + Gjunc(n) + G(1)
loop(n) + G(2)
loop(n) .
(9)
The first term refers to the nearest neighbour model developed in [5, 7], which specifies
the interaction gµ,µ+1 between a base pair µ and the directly adjacent one µ + 1. It was
shown to provide reasonable agreement with experiments [16, 21, 33]. For example,
applying this model onto a sequence 5′-TCCAG. . . -3′ and its complementary part 3′-
AGGTC. . . -5′, the stack energy reads Gstack = gTC/AG + gCC/GG + gCA/GT + gAG/TC + . . .
The most recent values of gµ,µ+1 lie in the range of -2.37 to −0.84 kcal/mol at 25°C
[11]. The terms Gjunc(n), G(1)
loop(n) in eq. (9) describe the free energy re-
duction due to the release of the stem-hairpin-junction and end-loops and are estimated
from [25, 37] as G(1)
loop = 1.58 kcal/mol and Gjunc = 4.90 kcal/mol.
loop(n) and G(2)
The strain energy Gss
str(n, α, f ) of the unfolded single-stranded part [21] with contour
length l = l(n, α)4 can be calculated from the work needed to stretch the unpaired
bases. We like to remind the reader that we denote the elongation of the mean end-
to-end distance of the ssDNA in force direction by ul( f ). In what follows we chose
loop = G(2)
4In previous publications of some of the authors this contour length was denoted by ln,α to emphasize the
dependence on n (and, in addition, α here). This dependence is caused by the change of the contour length
in the transitions. For easier reading we suppress to give it explicitely in the following. Further details about
the contour length were already discussed in sec. 4.
Folding and unfolding of a triple-branch DNA molecule with four conformational states 13
to work with the FJC model, see eq. (4), with parameters of [11] instead of using the
WLC model, despite the fact that both approaches give similar good results (cp. table
1). Since ul( f ) is monotonously increasing with f , it has an inverse fl(u) = u−1
l ( f ),
which is the force that is exerted by a ssDNA chain with contour length l, if its mean
end-to-end distance is elongated by u. Accordingly, setting ∆xss
l = ul( f ), we can write
Gss
str(n, α, f ) =
∆xss
l
Z
0
du′ fl(u′) = f ∆xss
l −
f
Z
0
d f ′ ul( f ′) .
(10)
Finally, we computed G(n, f ) by
G(n, f ) = −kBT lnXα
exp −
G(n, α, f )
kBT
! .
(11)
In fig. 7(b) the FEL is depicted for f = 0 and f = 16.21 pN.
It exhibits the
behaviour anticipated in fig. 7(a): for zero force it has minima at the stable/metastable
states and it becomes tilted with rising force. When approaching the force regime
where the rips occur in fig. 4(a), the levels of the minima become comparable. We want
to point out that the DNA sequences shown in fig. 1 have been designed deliberately to
yield the multiple-state structure seen in fig. 7(b). This gives us some confidence in the
model underlying the construction of the G(n, α, f ) in eq. (8).
Based on the FEL we can easily calculate the transition probability W( fi fi−1) for
the first rupture force fi in the ith transition if the first rupture force was fi−1 in the
(i − 1)th transition. The result is
W( fi fi−1) =
0
i
γ
r γi,i+1( f ) exph−
γ
0
i
r Z f
f ∗
i−1
d f ′
γi,i+1( f ′)i ,
(12)
i = fi − ∆ fi (and f0 = f ∗
0 = 0); r was the loading rate, see sec. 2, and γ
where f ∗
i and
γi,i+1( f ) were defined in eq. (6). The activation energy ∆Gi,i+1( f ) appearing in eq. (7)
was calculated, as indicated in fig. 7(a), from the G(n, f ) by determining the energy
Gmin
i,i+1 ( f )
of the ith transition between the ith and (i +1)th state, ∆Gi,i+1( f ) = Gsaddle
( f ).
i
The attempt rate γ
( f ) of the local minimum belonging to state i and the saddle point energy Gsaddle
i,i+1 ( f ) −Gmin
i was used as the only fitting parameter.
i
0
0
For the joint probability density of the three first rupture forces we then obtain
Ψ3( f1, f2, f3) = W( f10) W( f2 f1) W( f3 f2) ,
(13)
which allows us to calculate the distributions shown in fig. 6(a).
Figure 8 displays the histograms for the three rips from fig. 6(a) in comparison
with the distributions calculated from our theory. In view of the available statistics (82
cycles, see sec. 2), the agreement is quite satisfactory.
6 Conclusions
An important class of biophysical studies is the investigation of junctions in molecules
since they present manifold ways to interact with other substances, for instance cations.
Three-way junctions are especially interesting because metal ions such as magnesium
[14, 18] can bind to them and alter the tertiary structure. Here a first step of such a
Folding and unfolding of a triple-branch DNA molecule with four conformational states 14
experiment
experiment
experiment
theory
theory
theory
y
c
n
e
u
q
e
r
f
e
v
i
t
a
l
e
r
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
14 16 18
16 17
14 16 18
Figure 8: Comparison of the first rupture force distributions from fig. 6(a) with the
theory for the first ( f1), the second ( f2) and the third ( f3) transition, see left, middle and
1 = 9 · 106 Hz,
right panel, respectively. The best fit was obtained for attempt rates γ
2 = 1.1 · 106 Hz and γ
γ
3 = 3.5 · 105 Hz.
0
0
0
study is presented where we investigate a molecule with a three-way junction alone,
without in vivo relevant substances.
One of our aims of this work was to study whether the construction and kinetics
of more complex DNA molecules with richer folding-unfolding behaviour can be de-
scribed by proper extensions of theories developed successfully for two-state systems
so far. Our results show that this is indeed possible, at least for the unfolding trajec-
tories. A triple-branch molecule has been specifically designed to produce a four-state
system based on a model for the free energy landscape. This design was successful
and we were able to prove the existence of these states by the emergence of associated
force rips in pulling experiments. The first rupture forces have been systematically
recorded in these pulling experiments and their distributions have been calculated. A
transition rate theory based on the free energy landscape was successful in describing
these distributions.
Two patterns have been found in the measured unfolding trajectories, one indicat-
ing the anticipated unfolding behaviour and the other one pointing to the occurrence
of irreversible molecular fraying. This characterisation was possible by connecting the
extracted force jump values to the change in the number of opened bps at each transi-
tion. For this estimation we tested the validity of two polymer models for the elastic
response of ssDNA (FJC and WLC) and different sets of parameters in order to find the
best agreement with the expected values. This analysis is useful to compare the elastic
properties measured in DNA unzipping experiments with those obtained by stretching
ssDNA polymers [11].
One class of molecules required a smaller force than anticipated to unfold the stem
since some bps at its basis are partly or completely melted due to photodamaging.
Permanent molecular fraying is an usually undesired, but frequent effect in single-
Folding and unfolding of a triple-branch DNA molecule with four conformational states 15
molecule studies and deserves special attention in order to reduce its distorting influ-
ence on experimental results. It is important to find means to avoid irreversible fraying
since not fully closed molecules change the measured unfolding-folding trajectories so
that an average over all molecules, including permanently frayed ones, can lead to de-
viations of the real values and to misinterpreted results. To improve the statistics of the
results, it is therefore necessary to identify irreversibly frayed molecules and remove
them from the analysis. Within our analysis, we found a useful method to identify
permanent molecular fraying. With the appropriate parameters for the polymer models
one can estimate the number of irreversibly frayed bps at the basis of the stem. Further
experiments performed with silica beads instead of polystyrene microspheres could
clarify under which conditions permanent molecular fraying can be decreased.
All this knowledge paves the way to further interesting studies such as the refold-
ing problem [17] or two topics which require further experimental research. Firstly,
the binding of metal ions to the three-way junctions could be examined in order to
find out about structural changes due to the formation of tertiary contacts and there-
fore altered kinetics of the unfolding process [18]. Secondly, the translocation motion
of helicases that unwind dsDNA could be addressed as another interesting subject, in-
cluding the investigation of how they move along bifurcation points. Eventually, the
kinetic approach for the prediction of the probability distributions of the first rupture
forces could be a well suited starting point for future in-depth modelling in this domain
on a more microscopic basis.
Acknowledgements
S. E. thanks the Deutscher Akademischer Austauschdienst (DAAD) for providing fi-
nancial support (FREE MOVER and PROMOS) for stays at the Small Biosystems Lab
in Barcelona where the experiments have been performed. A. A. is supported by grant
AP2007-00995. F. R. is supported by the grants FIS2007-3454, Icrea Academia 2008
and HFSP (RGP55-2008).
References
[1] Ammenti, A., F. Cecconi, U. M. B. Marconi, and A. Vulpiani. 2009. J Phys
Chem B 113 (30):10348–10356.
[2] Block, S. M., C. L. Asbury, J. W. Shaevitz, and M. J. Lang. 2003. Proc Natl
Acad Sci U S A 100 (5):2351–2356.
[3] Bustamante, C., J. F. Marko, E. D. Siggia, and S. Smith. 1994. Science 265
(5178):1599–1600.
[4] Carter, N. J., and R. A. Cross. 2006. Curr Opin Cell Biol 18 (1):61–67.
[5] Crothers, D. M., and B. H. Zimm. 1964. J Mol Biol 9:1–9.
[6] Dessinges, M.-N., B. Maier, Y. Zhang, M. Peliti, D. Bensimon, and V. Cro-
quette. 2002. Phys Rev Lett 89 (24):248102.
[7] Devoe, H., and I. J. Tinoco. 1962. J Mol Biol 4:500–517.
Folding and unfolding of a triple-branch DNA molecule with four conformational states 16
[8] Dumont, S., W. Cheng, V. Serebrov, R. K. Beran, I. Tinoco, A. M. Pyle, and
C. Bustamante. 2006. Nature 439 (7072):105–108.
[9] Forns, N., S. de Lorenzo, M. Manosas, K. Hayashi, J. M. Huguet, and F. Ritort.
2011. Biophys J 100 (7):1765–1774.
[10] Hormeno, S., and J. R. Arias-Gonzalez. 2006. Biol Cell 98 (12):679–695.
[11] Huguet, J. M., C. V. Bizarro, N. Forns, S. B. Smith, C. Bustamante, and
F. Ritort. 2010. Proc Natl Acad Sci U S A 107 (35):15431–15436.
[12] Landry, M. P., P. M. McCall, Z. Qi, and Y. R. Chemla. 2009. Biophys J 97
(8):2128–2136.
[13] Leger, J. F., J. Robert, L. Bourdieu, D. Chatenay, and J. F. Marko. 1998. Proc
Natl Acad Sci U S A 95 (21):12295–12299.
[14] Liphardt, J., B. Onoa, S. B. Smith, I. Tinoco, and C. Bustamante. 2001. Sci-
ence 292 (5517):733–737.
[15] Maier, B., D. Bensimon, and V. Croquette. 2000. Proc Natl Acad Sci U S A 97
(22):12002–12007.
[16] Manosas, M., D. Collin, and F. Ritort. 2006. Phys Rev Lett 96 (21):218301.
[17] Manosas, M., I. Junier, and F. Ritort. 2008. Phys Rev E Stat Nonlin Soft Matter
Phys 78 (6 Pt 1):061925.
[18] Manosas, M., and F. Ritort. 2005. Biophys J 88 (5):3224–3242.
[19] Melchionna, S., M. Fyta, E. Kaxiras, and S. Succi. 2007. Int. J. Mod. Phys. C
18:685.
[20] Meller, A., L. Nivon, E. Brandin, J. Golovchenko, and D. Branton. 2000. Proc
Natl Acad Sci U S A 97 (3):1079–1084.
[21] Mossa, A., M. Manosas, N. Forns, J. M. Huguet, and F. Ritort. 2009. J. Stat.
Mech. P02060.
[22] Ritort, F. 2006. J. Phys. Condens. Matter 18:R531–R583.
[23] Ritort, F., S. Mihardja, S. B. Smith, and C. Bustamante. 2006. Phys Rev Lett
96 (11):118301.
[24] Rivetti, C., C. Walker, and C. Bustamante. 1998. J Mol Biol 280 (1):41–59.
[25] SantaLucia, J. J. 1998. Proc. Natl. Acad. Sci. USA 95:1460–1465.
[26] Smith, D. E., S. J. Tans, S. B. Smith, S. Grimes, D. L. Anderson, and C. Bus-
tamante. 2001. Nature 413 (6857):748–752.
[27] Smith, S. B., Y. Cui, and C. Bustamante. 1996. Science 271 (5250):795–799.
[28] Smith, S. B., Y. Cui, and C. Bustamante. 2003. Methods in Enzymology
361:134–162.
Folding and unfolding of a triple-branch DNA molecule with four conformational states 17
[29] van Mameren, J., P. Gross, G. Farge, P. Hooijman, M. Modesti, M. Falken-
berg, G. J. L. Wuite, and E. J. G. Peterman. 2009. Proc Natl Acad Sci U S A 106
(43):18231–18236.
[30] Wagner, B., R. Tharmann, I. Haase, M. Fischer, and A. R. Bausch. 2006. Proc
Natl Acad Sci U S A 103 (38):13974–13978.
[31] Wang, M. D., H. Yin, R. Landick, J. Gelles, and S. M. Block. 1997. Biophys J
72 (3):1335–1346.
[32] Wen, J.-D., L. Lancaster, C. Hodges, A.-C. Zeri, S. H. Yoshimura, H. F.
Noller, C. Bustamante, and I. Tinoco. 2008. Nature 452 (7187):598–603.
[33] Woodside, M. T., W. M. Behnke-Parks, K. Larizadeh, K. Travers, D. Her-
schlag, and S. M. Block. 2006. Proc Natl Acad Sci U S A 103 (16):6190–6195.
[34] Wuite, G. J., S. B. Smith, M. Young, D. Keller, and C. Bustamante. 2000.
Nature 404 (6773):103–106.
[35] Yasuda, R., H. Noji, M. Yoshida, K. Kinosita, and H. Itoh. 2001. Nature 410
(6831):898–904.
[36] Yin, H., M. D. Wang, K. Svoboda, R. Landick, S. M. Block, and J. Gelles.
1995. Science 270 (5242):1653–1657.
[37] Zuker, M. 2003.
Nucleic Acids Res. 31 (13):3406–3415.
See also
http://mfold.rna.albany.edu/.
|
1706.04752 | 1 | 1706 | 2017-06-15T06:53:05 | Coarse-Grained model of the demixing of DNA and non-binding globular macromolecules | [
"physics.bio-ph",
"q-bio.BM"
] | The volume occupied by the unconstrained genomic DNA of prokaryotes in saline solutions is thousand times larger than the cell. Moreover, it is not separated from the rest of the cell by a membrane. Nevertheless, it occupies only a small fraction of the cell called the nucleoid. The mechanisms leading to such compaction are the matter of ongoing debates. The present work aims at exploring a newly proposed mechanism, according to which the formation of the nucleoid would result from the demixing of the DNA and non-binding globular macromolecules of the cytoplasm, like ribosomes. To this end, a coarse-grained model of prokaryotic cells was developed and demixing was analyzed as a function of the size and number of crowders. The model suggests that compaction of the DNA is actually governed by the volume occupancy ratio of the crowders and remains weak almost up to the jamming critical density. Strong compaction is however observed just before jamming, suggesting that crowding and electrostatic repulsion work synergetically in this limit. Finally, simulations performed with crowders with different sizes indicate that the DNA and the largest crowders demix preferentially. Together with the recent observation of the gradual compaction of long DNA molecules upon increase of the concentration of BSA proteins and silica nanoparticles, this work supports the demixing mechanism as a key player for the formation of the nucleoid. | physics.bio-ph | physics | Coarse-Grained Model of the Demixing of DNA and
Non-Binding Globular Macromolecules
Marc Joyeux*
LIPHY, Université Grenoble Alpes and CNRS, Grenoble, France
(*) email : [email protected]
1
ABSTRACT: The volume occupied by the unconstrained genomic DNA of prokaryotes in
saline solutions is thousand times larger than the cell. Moreover, it is not separated from the
rest of the cell by a membrane. Nevertheless, it occupies only a small fraction of the cell
called the nucleoid. The mechanisms leading to such compaction are the matter of ongoing
debates. The present work aims at exploring a newly proposed mechanism, according to
which the formation of the nucleoid would result from the demixing of the DNA and non-
binding globular macromolecules of the cytoplasm, like ribosomes. To this end, a coarse-
grained model of prokaryotic cells was developed and demixing was analyzed as a function of
the size and number of crowders. The model suggests that compaction of the DNA is actually
governed by the volume occupancy ratio of the crowders and remains weak almost up to the
jamming critical density. Strong compaction is however observed just before jamming,
suggesting that crowding and electrostatic repulsion work synergetically in this limit. Finally,
simulations performed with crowders with different sizes indicate that the DNA and the
largest crowders demix preferentially. Together with the recent observation of the gradual
compaction of long DNA molecules upon increase of the concentration of BSA proteins and
silica nanoparticles, this work supports the demixing mechanism as a key player for the
formation of the nucleoid.
2
INTRODUCTION
The genomic DNA of eukaryotes is separated from the cellular cytoplasm by a nuclear
envelope and displays several levels of compaction ranging from the initial wrapping of the
DNA helix around histone proteins to the final X-shape of chromosomes. In contrast, the
genomic DNA of prokaryotes lacks such a detailed organization and is not separated from the
rest of the cell by any membrane. As has been known for decades, it nevertheless occupies
only a fraction of the cell called the nucleoid, which is rather surprising because the volume
occupied by the unconstrained molecule in saline solutions is several thousands of times
larger than the volume of the cell. The mechanism leading to the compaction of the bacterial
genomic DNA has puzzled the scientists for decades1 and is still the matter of ongoing
debates.2 The crucial point is that none of the mechanisms proposed until recently provide a
convincing explanation for the formation of the nucleoid. More precisely, supercoiling
provokes only mild compaction2,3 and the number of nucleoid associated proteins capable of
bridging two DNA duplexes is too small to induce significant global compaction.2,4 In
contrast, the conjunction of DNA charge neutralization by small polycations5 and the action
of fluctuation correlation forces6 is able to compact the DNA significantly, but this is
essentially an all-or-none mechanism, with the DNA molecule being either in the coil state or
in a globular state much denser than the bacterial nucleoid.7,8 Such an abrupt transition from
the coil state to a too dense globule is also observed upon addition to the buffer of long
neutral polymers and salt9 (to compensate for the weakness of depletion forces10) or long
anionic polymers.11,12 Finally, long cationic polymers are able in vitro to compact
progressively the DNA molecule to shrunken coil structures that resemble those of the DNA
inside the nucleoid,8 but this associative phase separation mechanism cannot play a role in the
3
compaction of the bacterial DNA in vivo, because prokaryotes cells do not contain sufficient
amounts of long polycations or proteins with large positive charges.
However, it has been shown recently that negatively charged globular macromolecules
may also play a role in the formation of the nucleoid. Indeed, long DNA molecules could be
compacted gradually to densities comparable to that of the nucleoid by adding 5 to 10% (w/v)
of bovine serum albumin (BSA) to the solution13,14. It can be argued that this result is not
completely unambiguous, because the surface of BSA proteins displays small positively
charged patches despite its total charge of approximately -18e and the formation of weak
BSA-DNA coacervates has been reported.15 Fortunately, it has been checked even more
recently that a few percents of negatively charged silica nanoparticles with diameters ranging
from 20 to 135 nm are also able to compact gradually the DNA,16 thereby confirming the
results obtained with BSA. These results suggest that the compaction of the bacterial nucleoid
may result from the demixing of DNA and other non-binding globular macromolecules
contained in the cytoplasm,17 this hypothesis being all the more sensible as approximately
30% of the dry mass of cells is composed of ribosomes, which are almost spherical and highly
negatively charged complexes with diameter 20-25 nm. The formation of the nucleoid would
consequently result from a segregative phase separation18 leading to a phase rich in DNA (the
nucleoid) and another phase rich in the other macromolecules (the rest of the cytoplasm). This
hypothesis has received little attention up to now, although it has been evoked on theoretical
grounds almost 20 years ago.19,20
The purpose of the present work is to elaborate further on this putative mechanism for
nucleoid compaction by addressing two important questions. The first one deals with the
determination of the volume occupancy ratio at which segregation takes place. It is indeed not
easy to estimate the volume effectively occupied by proteins or nanoparticles in a given
experiment, because the Debye length cannot be determined precisely. While calculations
4
suggest that silica nanoparticles occupy approximately 15-20% of the volume at the point of
DNA collapse,16 the authors nevertheless suspect that the actual Debye length was actually
larger than the value they used in their calculations and speculate that 'the observed critical
concentrations of nanoparticles approach the overlapping concentration, which is ca. 74%
for most densely packed spheres'16. The second question relates to size dispersion of
macromolecules in the cytoplasm. Indeed, experiments are usually performed with
macromolecules and/or particles that are all of the same size, while the size of
macromolecules in the cytoplasm varies over a large range. It is consequently important to get
an idea of how size dispersion affects demixing in real conditions. It should be stressed that
these two questions are rather difficult to handle from a purely theoretical point of view,
because the outcome of statistical physics calculations depends critically on the quality of the
description of the interactions between crowders and/or between crowders and the DNA
chain.21,22 In the present work, we therefore sought for answers through simulations based on
a coarse-grained model that will now be described.
COMPUTATIONAL DETAILS
As illustrated in Fig. 1, the coarse-grained model used in this work consists of a
circular chain of
=n
1440
beads of radius
78.1=a
nm separated at equilibrium by a distance
=l
0
0.5
nm (the genomic DNA) enclosed in a large confining sphere of radius
=R
0
120
nm
(the cell), together with N spheres of radius b (the crowding macromolecules). As in previous
work,23-28 each bead represents 15 consecutive DNA base pairs, so that the model may be
thought of as a 1:200 reduction of a typical E. coli cell with a nucleotide concentration around
10mM, close to the physiological value. The potential energy of the system,
potE , is the sum
of four terms
5
E
pot
=
V
DNA
+
V
DNA/C
+
V
C/C
+
V
wall
,
(1)
which describe the internal energy of the DNA molecule, the DNA-crowder interactions, the
crowder-crowder interactions, and the repulsive potentials that maintain the DNA and the
crowders inside the confining sphere, respectively. The internal energy of the DNA molecule,
DNAV
, is further written as the sum of 3 contributions23-28
n
∑
=
1
k
(
l
k
+
2
l
0
)
g
2
V
DNA
=
h
2
where
q
2
k
+
2
e
DNA
n
∑
=
1
k
n
2
n
∑ ∑
=
1
k
+=
kK
(
r
k
H
2
r
K
)
,
)(
rH
=
1
pe
4
-
exp
r
r
Dr
,
(2)
(3)
which describe the stretching, bending, and electrostatic energy of the DNA chain,
respectively.
kr denotes the position of DNA bead k, kl the distance between two successive
beads, and
q the angle formed by three successive beads. The stretching energy is a
k
computational device without biological meaning, which is aimed at avoiding a rigid rod
description. h was set to
1000
/
lTk
B
2
0
, in order for
lk
0l
to remain on average as small as
002.0
l despite the forces exerted by the crowders (in this work, all energies are expressed in
units of TkB
, where
=T
298
K). In contrast, the bending rigidity constant,
g
82.9=
Tk
B
, was
chosen so as to provide the correct persistence length for DNA,
x
=
gl
0
/( B
Tk
)
49
nm,
equivalent to 10 beads.29 Finally, the electrostatic energy is expressed as a sum of Debye-
Hückel potentials30, where
-=
e
DNA
15.12
e
(with e the absolute charge of the electron)
denotes the value of the charge placed at the centre of each DNA bead,23
e =
80 e
the
0
dielectric constant of the medium, and
07.1=
Dr
nm the Debye length inside the medium.
Since each bead represents 15 consecutive DNA base pairs, the total charge carried by the
corresponding phosphate groups is
e30
, but the smaller value
-=
e
DNA
15.12
e
was used in
6
-
-
-
-
»
-
the Debye-Hückel potential to take counter-ion condensation into account.30 Moreover, the
small value of the Debye length,
07.1=
Dr
nm, corresponds to a concentration of monovalent
salts close to 100 mM and reflects the order of magnitude of the Debye length expected in
bacterial cells.16 Admittedly, the equilibrium separation of two DNA beads,
=l
0
0.5
nm, is too
large compared to the value of the Debye length to warrant that different parts of the DNA
chain will never cross. However, such crossings are very infrequent and appear to affect the
geometry of the DNA chain only in a limited fashion. Therefore, the term of the potential
energy that would fully prevent chain crossing (Eq. (A.5) of Ref. 2), which is quite expensive
from the computational point of view, was discarded from the lengthy simulations reported
here. Finally, electrostatic interactions between nearest-neighbours are not included in Eq. (2),
because it is considered that they are already accounted for in the stretching and bending
terms.
In the same spirit, the DNA-crowder and crowder-crowder interactions are expressed
as sums of Debye-Hückel potentials with hard cores
V
DNA/C
=
e
DNA
e
C
n
N
∑∑
=
1
k
=
1
j
H
(
Rr
k
j
b
)
(4)
V
C/C
=
2
e
C
N
1
N
∑ ∑
H
+=
1
j
=
1
j
J
(
RR
j
J
,)2
b
where
)(rH is the function described in Eq. (3),
Ce is an electrostatic charge that
characterizes the spherical crowders and
jR denotes the position of crowding sphere j. The
2
(
rHe
C
)2
b
potential in Eq. (4) differs from the usual DLVO potential31,32
DLVO
W
=
)(
r
2
e
M
pe
4
+
r
1(
2
)
b
r
D
exp(
r
2
b
r
D
)
,
(5)
7
-
-
-
-
-
-
-
-
where Me
is the total charge of the sphere, essentially by the fact that it diverges for a
separation r between the centers of the spheres equal to 2b, while the DLVO potential
diverges for
0=r
. However, for
e
M
=
+
e
C
1(
)
b
r
D
r
0
r
0
2
b
(6)
the
2
(
rHe
C
)2
b
and
W
DLVO r
)(
potentials remain close to each other in a broad range of
separation values around 0r . For example, Fig. 2 shows that the
2
(
rHe
C
)2
b
potential with
=
-=
e
DNA
e
C
15.12
e
and
5.6=b
nm remains close to the DLVO potential with
-=
e
M
210
e
for all interaction values ranging from 0 to
TkB8
. In this work, we choose to use the
2
(
rHe
C
)2
b
potential rather than the
W
DLVO r
)(
potential, because the increase in the
effective width of the crowder due to electrostatic interactions does not depend on the radius b
of the spheres for the former one, while it decreases with increasing b for the later one. As
will be emphasized in the Results and Discussion Section, this feature makes the
interpretation of
the results obtained with
the
2
(
rHe
C
)2
b
potential particularly
straightforward. The same argument holds of course for the
e
DNA
brHe
C
(
)
potential that
describes DNA-crowder interactions. The three electrostatic repulsion potentials,
2
e
DNA
)(
rH
,
e
DNA
brHe
C
(
)
, and
2
(
rHe
C
)2
b
, are plotted in Fig. 3 for
e =
C
e
DNA
and
5.6=b
nm. Most
simulations discussed in the present paper were performed with
e =
C
e
DNA
, but a series of
simulations were performed with
e =
C
8.0 e
DNA
for the sake of comparison. Finally, wallV is
written in the form
V
wall
=
z
(
n
∑
=
1
k
f
(
r
k
+
)
N
∑
=
1
j
f
(
R
j
))
,
(7)
where the repulsive force constant z was set to
1000
TkB
and the function
)(rf
is defined
according to
8
-
-
-
-
-
-
-
-
if
0Rr £
:
0)( =rf
if
0Rr >
:
)(
rf
=
r
R
0
6
-
1
.
(8)
The dynamics of the system was investigated by integrating numerically overdamped
Langevin equations. Practically, the updated positions at time step n+1 were computed from
the positions at time step n according to
+
)1
n
=
(
r
k
)(
n
r
k
+
)1
R
(
n
j
=
R
)(
n
j
t
D+
ph
6
D+
ph
6
a
t
+
f
)(
n
k
2
tTk
B
ph
6
a
x
)(
n
k
+
F
)(
n
j
b
2
tTk
B
ph
6
b
X
)(
n
j
,
(9)
where
20=
t
ps is the integration time step,
)(n
kf
and
)(n
jF are vectors of inter-particle forces
arising from the potential energy potE ,
=T
298
K is the temperature of the system,
)(n
kx
and
)(n
jX are vectors of random numbers extracted from a Gaussian distribution of mean 0 and
variance 1, and
.0=h
00089
Pa s is the viscosity of the buffer at 298 K. After each integration
step, the position of the centre of the sphere was slightly adjusted so as to coincide with the
centre of mass of the DNA molecule.
RESULTS AND DISCUSSION
Simulations were performed by first letting the DNA chain equilibrate inside the
confining sphere. Since the unconstrained chain forms a coil with a radius of gyration larger
than
xnl
0
6/
242
nm (this estimation based on the Worm-Like Chain model neglects
electrostatic repulsion between DNA segments),33 which is substantially larger than the radius
=R
0
120
nm of the confining sphere, the relaxed chain occupies the whole space inside the
sphere and the repulsive potential wallV acts directly on some of the DNA beads to repel them
9
D
D
D
»
inside the confining sphere, thereby preventing further expansion of the chain. A typical
conformation of the equilibrated chain is shown in the bottom left vignette of Fig. 1.
Simulations indicate that the mean value of the radius of gyration of the equilibrated confined
chain is close to 82 nm for a Debye length
07.1=
Dr
nm. A number N of crowding spheres
with radius b are then introduced at random (but non-overlapping) positions inside the
confining sphere, as shown in the top vignette of Fig. 1, and the system is allowed to
equilibrate again. After equilibration times ranging from a few ms (at low volume occupancy)
to more than 10 ms (close to jamming), the system reaches a new steady state, which is
characterized by a (usually) lower value of the radius of gyration of the DNA chain, as shown
in the bottom right vignette of Fig. 1. Practically, simulations were performed with four
different values of N (N=500, 1000, 2000, or 3000 crowding spheres) and 8 to 15 different
values of b for each value of N. The results are shown in Fig. 4, where the mean value of the
radius of gyration of the DNA chain after equilibration with the crowding spheres,
gR , is
plotted as a function of the radius b of the crowders. It is observed in this figure that the four
curves corresponding to the different values of N display the same evolution as a function of
b, which can be divided into three different regimes. Indeed, for the lowest values of b,
gR
decreases regularly down to a threshold value close to
gR
62
nm. Close examination of the
simulations reveals that, over the whole range of corresponding values of b, the repulsive
potential
wallV still contributes to repelling directly some of the DNA beads inside the
confining sphere. In contrast, for the largest values of b, the system is jammed and neither the
DNA chain nor the crowders are able to move significantly. As a result,
gR remains close to
its initial value of 82 nm. Finally, in a rather narrow interval of values of b located between
these two regimes,
gR drops sharply to values as low as about 50nm. Strikingly, in this
regime no part of the DNA chain is any longer in contact with the confining sphere, so that
10
»
wallV contributes uniquely to maintaining the crowders inside the confining sphere, while
compaction of the DNA chain results exclusively from interactions among DNA beads and
crowding spheres. It is certainly not by chance that, for the four values of N, maximum
compaction of the DNA chain occurs just before jamming. The obvious interpretation of this
observation is that crowding conditions close to the jamming transition constrain DNA and
the crowders to remain at short distances from one another and feel electrostatic interactions
despite the shortness of the Debye length, which ultimately favors demixing of DNA and the
crowders and compaction of the DNA chain.
Examination of Fig. 4 further indicates that the optimum value of b leading to
maximum DNA compaction decreases with increasing N, being close to 11.5, 8.5, 6.5, and 5.5
nm, for N=500, 1000, 2000, and 3000, respectively. Estimation of the volume occupancy ratio
according to
0RbN
/(
3)
suggests that the optimum ratio also decreases with increasing N,
being close to 0.44, 0.36, 0.32, and 0.29, for N=500, 1000, 2000, and 3000, respectively.
However, this estimation neglects the electrostatic repulsion between crowders. A natural way
of taking such repulsion into account consists in considering that the effective radius
b D+
b
of a crowding sphere is equal to half the separation at which the interaction energy of two
crowders is equal to the thermal energy TkB
. The interesting point in using the
2
(
rHe
C
)2
b
potential of Eq. (4) instead of the DLVO potential of Eq. (5) to describe the interaction
between two crowders is precisely that
bD
does not depend on b for the
2
(
rHe
C
)2
b
potential and a particular value of Ce , while it does depend on b for the DLVO potential and a
particular value of Me . As illustrated in Fig. 3, bD
is close to 1.8 nm for
e =
C
e
DNA
and
07.1=
Dr
nm. It is quite noteworthy that estimation of the effective volume occupancy ratio
according to
11
-
-
=r
b
N
(
D+
0R
b
3)
(10)
leads to nearly identical values of r at maximum DNA compaction for all values of N, namely
68.0»r
, 0.63, 0.66, and 0.68, for N=500, 1000, 2000, and 3000, respectively. Even more
striking is the fact that the four plots for
gR collapse on a single master curve when r is
chosen as the abscissa axis instead of b, as can be seen in Fig. 5. This plot indicates that,
within the validity of the model, the compaction dynamics of the DNA chain is driven
exclusively by the effective volume occupancy ratio of the crowders. More precisely, the
radius of gyration of the DNA chain decreases linearly with increasing values of r up to
6.0»r
, while significantly stronger DNA compaction, which no longer requires the direct
action of the confining wall on the DNA beads, takes place in the range
6.0
£ r
7.0
.
Finally, the system becomes jammed around
75.0»r
, meaning that jamming occurs for the
soft spheres considered here for values of r only slightly larger than for hard spheres
66.0»r
(
).34
The leading role of r is further confirmed by the plot of the mean value of the
interaction energy between the confining wall and the DNA chain as a function of
r-75.0
,
the gap to the volume occupancy ratio at jamming, which is shown in Fig. 6. Again, the
curves corresponding to the four different values of N collapse on a single master curve.
Moreover, the use of log-log axes highlights the fact that the average work exerted by the wall
to maintain the DNA chain inside the confining sphere decreases as the cube of
r-75.0
in
the whole range of values of r extending from weak crowding to the threshold where strong
compaction takes place. While this power law is probably model-dependent, this graph
nevertheless emphasizes the fact that r controls not only the profile of DNA concentration
12
£
inside the confining sphere but also the strength of the interactions between the DNA chain
and the confining wall.
Finally, let us mention that the results presented above do not depend on the exact
value of Ce , the parameter that characterizes the electrostatic charge carried by the crowding
sphere. To ascertain this point, a set of simulations were run with N=2000 crowders and
e =
C
8.0 e
DNA
, that is for crowder/crowder interactions reduced by a factor 0.64 compared to
simulations run with
e =
C
e
DNA
. Still, it proved to be sufficient to plug the appropriate values
of bD
in Eq. (10), that is
D b
6.1=
nm for
e =
C
8.0 e
DNA
against
D b
8.1=
nm for
e =
C
e
DNA
, for
the plots obtained with the two different values of Ce to superpose (results not shown).
The simulations reported above are therefore consistent with the progressive
compaction of long DNA molecules, which is observed upon increase of the concentration of
anionic silica nanoparticles added to the solution containing the DNA molecule.16 They
furthermore support the conjecture of the authors that maximum compaction is observed
when the 'concentrations of nanoparticles approach the overlapping concentration',16 which
is the first point this work aimed to ascertain. The simulations moreover indicate clearly that,
close to jamming, crowding and electrostatic repulsion forces work synergetically in favor of
particularly efficient DNA-crowders demixing and strong compaction of the DNA chain.
The second point addressed in this work deals with the effect of crowders size
dispersion, and more precisely how macromolecules size dispersion may affect demixing in
real cells. To investigate this point, several sets of simulations were run with crowders with
two different radii enclosed simultaneously in the confining sphere and compared with
simulations involving only mono-disperse crowders. All sets of simulations with bi-disperse
crowders actually led to similar results and the discussion below focuses on the comparison of
results obtained with 1400 spheres with radius b=7.2 nm and 600 spheres with radius b=3.5
nm, on one side, and 2000 crowders with radius b=6.5 nm, on the other side. In both cases,
13
the effective volume occupancy ratio is close to the optimum value for maximal DNA
compaction (
65.0»r
). Practically, we will compare the density distributions
)(rpX
, such
that the mean number of particles of species X, which centers are located in a distance interval
rr +
,[
dr
[
from the center of the confining sphere, is
2p
4
)(
rpnr
X
X
dr
, where
Xn is the total
number of particles of type X (
== n
nX
1440
for DNA,
nX
= N
=
2000
for 2000 crowders
with radius b=6.5 nm, etc). Fig. 7A shows the distribution obtained after equilibration of 2000
crowders with radius b=6.5 nm in the absence of DNA. The regularly spaced peaks, which are
separated by about 14.5 nm, denote the quasi-crystalline order that prevails inside the
confining sphere just below the jamming critical density. The distribution becomes
progressively flatter with decreasing values of r. Similarly, the distributions obtained after
inserting the 2000 crowders with radius b=6.5 nm at random positions in the confining sphere
containing the pre-equilibrated DNA chain and allowing the system to equilibrate again, as
for the results in Figs. 4-6, are shown in Fig. 7B. As discussed above, the DNA beads and the
crowding spheres demix, with the effect that the concentration of DNA beads increases
steeply close to the center of the sphere but vanishes beyond 90 nm, while the concentration
of crowders decreases regularly towards the center of the confining sphere compared to the
system without DNA (Fig. 7A). These distributions can be contrasted with those shown in
Figs. 7C and 7D, which were obtained along the same lines, but with 1400 spheres with radius
b=7.2 nm and 600 spheres with radius b=3.5 nm instead of 2000 spheres with radius b=6.5
nm. Quite interestingly, it may be observed that, in the absence of DNA (Fig. 7C), the outer
shell is composed almost exclusively of spheres with the largest radius, while the two
distributions remain roughly equal for all the inner shells. Note also that the crystalline order
is progressively lost when moving away from the edge of the sphere. Finally, Fig. 7D shows
the distributions obtained with both the DNA chain and bi-disperse crowders. It is clearly seen
in this figure that the distribution of smaller crowders varies little compared to the system
14
without DNA (Fig. 7C), while the distribution of larger crowders decreases regularly towards
the center of the confining sphere, as for mono-disperse crowders (Fig. 7B). Stated in other
words, the total volume fraction occupied by crowders of any size is the crucial quantity for
strong DNA compaction, while segregation out of the nucleoid affects primarily the crowders
with largest size. All in all, these results suggest that, in real prokaryotic cells, demixing
involves the DNA molecule and the largest globular macromolecules present in the
cytoplasm, that is essentially the ribosomes, although their concentration is smaller than the
critical crowder concentration leading to strong compaction.
CONCLUSIONS
Triggered by the experimental observation of the gradual compaction of long DNA
molecules upon increase of the concentration of BSA proteins13,14 and silica nanoparticles16,
the main purpose of the present work was to elaborate further on the conjecture that the
formation of the bacterial nucleoid may be driven by the demixing of DNA and non-binding
globular macromolecules present in the cytoplasm. For this purpose, several sets of
simulations were performed by varying the number and size of spherical crowders in a coarse-
grained model of prokaryotic cells. Demixing of DNA and the crowders and compaction of
the DNA chain were observed in all simulations starting from homogenously distributed
particles. The simulations moreover highlight the fact that the gradual compaction of the
DNA chain is governed by the volume occupancy ratio of the crowders. DNA compaction
increases with this ratio but remains weak almost up to the jamming critical density. Much
stronger DNA compaction is however observed just before jamming, suggesting that
crowding and electrostatic repulsive interactions work synergetically in favor of particularly
15
efficient demixing. Finally, simulations performed with crowders of different sizes indicate
that the DNA chain and the largest crowders demix preferentially.
It should furthermore be stressed that, in the simulations discussed above, the volume
fraction occupied by naked crowders at the critical concentration for strong DNA compaction
ranges from 0.29 to 0.44, depending on the number N of crowders, so that the volume fraction
left for virtual water ranges from 0.56 to 0.71, which is in qualitative agreement with the 50-
70% water content that is usually reported for prokaryotic cells. Moreover, the translational
diffusion coefficient of macromolecules is known to be much smaller in prokaryotic cells than
in water and in eukaryotic cells,35 which indicates that the cytoplasm of prokaryotic cells is
indeed close to jamming. Put together, these two facts confirm that the mechanism described
here may indeed play an important role in vivo.
This work therefore adds weight to the hypothesis that the formation of the nucleoid in
bacteria may actually result from the demixing of DNA and non-binding globular
macromolecules present in the cytoplasm, that is, in other words, from a segregative phase
separation18 leading to a phase rich in DNA (the nucleoid) and another phase rich in the other
macromolecules (the rest of the cytoplasm).13,14,16 Since about 30% of the dry mass of
prokaryotic cells is composed of ribosomes, which are highly and almost uniformly charged
anionic complexes with diameter 20-25 nm, and are excluded from the nucleoid in their
functional form,36,37 it is furthermore tempting to argue that ribosomes (and complexes made
of ribosomes) actually play the role of the larger crowders in the coarse-grained model, while
most other non-binding macromolecules play the role of the smaller crowders, which are only
weakly segregated or not segregated at all.
To conclude, let me argue that it is probably confusing to describe the demixing
mechanism discussed above as resulting from the action of 'depletion forces', in spite of the
fact that it is governed by the volume occupancy ratio of the crowders. Indeed, the term
16
'depletion
force'
traditionally describes
the effective attraction
force between
macromolecules, which results from the preferential exclusion of smaller cosolutes from the
vicinity of these macromolecules. Such depletion forces can be dominated either by entropy10,
as is probably the case for the condensation of DNA by neutral polymers9, or by enthalpy,38,39
as may be the case for the condensation of DNA by anionic polymers.11,12 In both cases,
depletion forces are very short-ranged and compact the DNA molecule very abruptly to
almost crystalline densities above a certain polymer concentration threshold. When the
diameter of the crowders is much larger than the diameter of the DNA duplex, as is the case
here, it is instead more natural to compare the strength of the pair interaction between DNA
and the crowder, on one side, with the average strength of the pair interactions between two
DNA segments and between two crowders, on the other side, to get an insight whether the
system will remain globally homogenous or will demix. This difference in the strength of pair
interactions is precisely the c parameter of Flory-Huggins polymer solution theory,18 which
sign determines whether the two species demix (
0>c
) or not (
0<c
), and which absolute
value determines the extent of demixing, which is consequently gradual instead of being an
all-or-none process, as the condensation provoked by depletion forces.
ACKNOWLEDGEMENT
This work was supported by the Centre National de la Recherche Scientifique and
Université Grenoble Alpes.
17
REFERENCE LIST
(1) Zimmerman, S. B. Shape and Compaction of Escherichia coli Nucleoids. J. Struct. Biol.
2006, 156, 255-261.
(2) Joyeux, M. Compaction of Bacterial Genomic DNA: Clarifying the Concepts. J. Phys.:
Condens. Matter 2015, 27, 383001.
(3) Cunha, S.; Woldringh, C. L.; Odijk, T. Polymer-Mediated Compaction and Internal
Dynamics of Isolated Escherichia coli Nucleoids. J. Struct. Biol. 2001, 136, 53-66.
(4) Rouvière-Yaniv, J.; Yaniv, M. E. coli DNA Binding Protein HU Forms Nucleosome-Like
Structure With Circular Double-Stranded DNA. Cell 1979, 17, 265-274.
(5) Manning, G. S. Limiting Laws and Counterion Condensation in Polyelectrolyte Solutions
I. Colligative Properties. J. Chem. Phys. 1969, 51, 924-933.
(6) Oosawa, F. Interaction Between Parallel Rodlike Macroions. Biopolymers 1968, 6, 1633-
1647.
(7) Gosule, L. C.; Schellman, J. A. Compact Form of DNA Induced by Spermidine. Nature
1976, 259, 333-335.
(8) Akitaya, T.; Seno, A.; Nakai, T.; Hazemoto, N.; Murata, S.; Yoshikawa, K. Weak
Interaction Induces an ON/OFF Switch, Whereas Strong Interaction Causes Gradual Change:
Folding Transition of a Long Duplex DNA Chain by Poly-L-lysine. Biomacromolecules
2007, 8, 273-278.
(9) Lerman, L. S. A Transition to a Compact Form of DNA in Polymer Solutions. Proc. Natl.
Acad. Sci. U.S.A. 1971, 68, 1886-1890.
(10) Asakura, S.; Oosawa, F. On Interaction Between Two Bodies Immersed in a Solution of
Macromolecules. J. Chem. Phys. 1954, 22, 1255-1256.
18
(11) Jordan, C. F.; Lerman, L. S.; Venable, J. H. Structure and Circular Dichroism of DNA in
Concentrated Polymer Solutions. Nature New Biol. 1972, 236, 67-70.
(12) Ichiba, Y.; Yoshikawa, K. Single Chain Observation on Collapse Transition in Giant
DNA Induced by Negatively-Charged Polymer. Biochim. Biophys. Res. Commun. 1998, 242,
441-445.
(13) Krotova, M. K.; Vasilevskaya, V. V.; Makita, N.; Yoshikawa, K.; Khokhlov, A. R. DNA
Compaction in a Crowded Environment With Negatively Charged Proteins. Phys. Rev. Lett.
2010, 105, 128302.
(14) Yoshikawa, K.; Hirota, S.; Makita, N.; Yoshikawa, Y. Compaction of DNA Induced by
Like-Charge Protein: Opposite Salt-Effect Against the Polymer-Salt-Induced Condensation
With Neutral Polymer. Phys. Chem. Lett. 2010, 1, 1763-1766.
(15) Wagh, J.; Patel, K. J.; Soni, P.; Desai, K.; Upadhyay, P.; Soni, H. P. Transfecting pDNA
to E. coli DH5α Using Bovine Serum Albumin Nanoparticles as a Delivery Vehicle.
Luminescence 2015, 30, 583-591.
(16) Zinchenko, A.; Tsumoto, K.; Murata, S.; Yoshikawa, K. Crowding by Anionic
Nanoparticles Causes DNA Double-Strand Instability and Compaction. J. Phys. Chem. B
2014, 118, 1256-1262.
(17) Joyeux, M. In Vivo Compaction Dynamics of Bacterial DNA: A Fingerprint of
DNA/RNA Demixing ? Curr. Opin. Colloid Interface Sci. 2016, 26, 17-27.
(18) Hsu, C. C.; Prausnitz, J. M. Thermodynamics of Polymer Compatibility in Ternary
Systems. Macromolecules 1974, 7, 320-324.
(19) Sear, R. P. Coil-Globule Transition of a Semiflexible Polymer Driven by the Addition of
Spherical Particles. Phys. Rev. E 1998, 58, 724-728.
(20) van der Schoot, P. Protein-Induced Collapse of Polymer Chains. Macromolecules 1998,
31, 4635-4638.
19
(21) Castelnovo, M.; Gelbart, W. M. Semiflexible Chain Condensation by Neutral Depleting
Agents: Role of Correlations Between Depletants. Macromolecules 2004, 37, 3510-3517.
(22) de Vries, R. Depletion-Induced Instability in Protein-DNA Mixtures: Influence of Protein
Charge and Size. J. Chem. Phys. 2006, 125, 014905.
(23) Jian, H.; Vologodskii, A. V.; Schlick, T. A Combined Wormlike-Chain and Bead Model
for Dynamic Simulations of Long Linear DNA. J. Comp. Phys. 1997, 136, 168-179.
(24) Florescu, A.-M.; Joyeux, M. Description of Non-Specific DNA-Protein Interaction and
Facilitated Diffusion With a Dynamical Model. J. Chem. Phys. 2009, 130, 015103.
(25) Florescu, A.-M.; Joyeux, M. Dynamical Model of DNA Protein Interaction: Effect of
Protein Charge Distribution and Mechanical Properties. J. Chem. Phys. 2009, 131, 105102.
(26) Florescu, A.-M.; Joyeux, M. Comparison of Kinetic and Dynamical Models of DNA-
Protein Interaction and Facilitated Diffusion. J. Phys. Chem. A 2010, 114, 9662-9672.
(27) Joyeux, M.; Vreede, J. A Model of H-NS Mediated Compaction of Bacterial DNA.
Biophys. J. 2013, 104, 1615-1622.
(28) Joyeux, M. Equilibration of Complexes of DNA and H-NS Proteins on Charged
Surfaces: A Coarse-Grained Model Point of View. J. Chem. Phys. 2014, 141, 115102.
(29) Frank-Kamenetskii, M. D.; Lukashin, A. V.; Anshelevich, V. V. Torsional and Bending
Rigidity of the Double Helix From Data on Small DNA Rings. J. Biomol. Struct. Dynam.
1985, 2, 1005-1012.
(30) Vologodskii, A. V.; Cozzarelli, N. R. Modeling of Long-Range Electrostatic Interactions
in DNA. Biopolymers 1995, 35, 289-296.
(31) Derjaguin, B.; Landau, L. Theory of the Stability of Strongly Charged Lyophobic Sols
and of the Adhesion of Strongly Charged Particles in Solution of Electrolytes. Acta
Physicochim. URSS 1941, 14, 633-662.
20
(32) Verwey, E. J. W.; Overbeek, J. T. G. Theory of the Stability of Lyophobic Colloids;
Elsevier: Amsterdam, Netherlands, 1948.
(33) Teraoka, I. Polymer Solutions: An Introduction to Physical Properties; Wiley: New
York, U.S.A., 2002.
(34) Chaudhuri, P.; Berthier, L.; Sastry, S. Jamming Transitions in Amorphous Packings of
Frictionless Spheres Occur Over a Continuous Range of Volume Fractions. Phys. Rev. Lett.
2010, 104, 165701.
(35) Mika, J. T.; Poolman, B. Macromolecule Diffusion and Confinement in Prokaryotic
Cells. Curr. Opin. Biotech. 2011, 22, 117-126.
(36) Sanamrad, A.; Persson, F.; Lundius, E. G.; Fange, D.; Gynna, A. H.; Elf, J. Single-
Particle Tracking Reveals That Free Ribosomal Subunits are not Excluded From the
Escherichia coli Nucleoid. Proc. Natl. Acad. Sci. U.S.A. 2014, 111, 11413-11418.
(37) Bakshi, S.; Siryaporn, A.; Goulian, M.; Weisshaar, J. C. Superresolution Imaging of
Ribosomes and RNA Polymerase in Live Escherichia coli Cells. Mol. Microbiol. 2012, 85,
21-38.
(38) Sapir, L.; Harries, D. Origin of Enthalpic Depletion Forces. J. Phys. Chem. Lett. 2014, 5,
1061–1065.
(39) Sapir, L.; Harries, D. Is the Depletion Force Entropic? Molecular Crowding Beyond
Steric Interactions. Curr. Opin. Coll. Int. Sci. 2015, 20, 3–10.
21
Figure 1 : The top vignette shows a typical conformation of the equilibrated DNA chain with
1440 beads (red) and the 2000 crowding spheres with radius
5.6=b
nm (cyan), which have
just been added randomly. The bottom left vignette shows the same snapshot with the
crowders having been removed and ¼ of the confining sphere being shown. The bottom right
vignette shows a typical conformation of the DNA after equilibration with the crowding
spheres.
22
Figure 2 : Plot of
2
(
rHe
C
)2
b
, which describes crowder-crowder repulsion (red solid line),
and the alternative DLVO potential of Eq. (5) (cyan dashed line), as a function of r, for
e =
. The vertical dash-dotted line is located at
5.6=b
nm, and
210
e
2=
b
.
r
e
DNA
,
C
-=
e
M
23
-
Figure 3 : Plot of
2
e
DNA
)(
rH
, which describes DNA-DNA repulsion (blue long-dashed line),
e
DNA
brHe
C
(
)
, which describes DNA-crowder repulsion (green short-dashed line), and
)2
b
e
DNA
, which describes crowder-crowder repulsion (red solid line), as a function of r,
r = and
nm. The two vertical dash-dotted lines are located at
5.6=b
and
b
2
(
rHe
C
e =
for
2=
b
r
C
. The horizontal dash-dotted line located at
TkB1
indicates that thermal energy
corresponds to the repulsion of two spheres, which centers are separated by 16.6 nm (see
text).
24
-
-
Figure 4 : Plot of
gR , the mean radius of gyration of the equilibrated DNA chain, as a
function of b, the radius of the crowding spheres, for N=500 (circles), 1000 (triangles), 2000
(squares) and 3000 (lozenges) crowding spheres. Both
gR and b are expressed in nm.
Simulations were performed with
e =
C
e
DNA
.
25
Figure 5 : Plot of
gR , the mean radius of gyration of the equilibrated DNA chain, as a
function of r, the effective volume occupancy ratio of the crowding spheres, for N=500
gR is
(circles), 1000 (triangles), 2000 (squares) and 3000 (lozenges) crowding spheres.
expressed in nm. Simulations were performed with
e =
C
e
DNA
. The black dash-dotted line is
just a guide for the eyes.
26
Figure 6 : Log-log plot of the mean wall/DNA interaction energy per DNA bead as a function
of
r-75.0
, the gap to the volume occupancy ratio at jamming, for N=500 (circles), 1000
(triangles), 2000 (squares) and 3000 (lozenges) crowding spheres. The mean interaction
energy is expressed in units of TkB
. Simulations were performed with
e =
C
e
DNA
. The black
dash-dotted line is a guide for the eyes that helps visualize cubic evolution.
27
Figure 7 : Plot of the mean density distributions of DNA beads and crowding spheres,
, for equilibrated systems without DNA (vignettes A and C) or with DNA (vignettes B
)(rpX
and D), and with 2000 crowding spheres with radius b=6.5 nm (vignettes A and B) or 1400
spheres with radius b=7.2 nm and 600 spheres with radius b=3.5 nm (vignettes C and D).
28
|
1601.07126 | 4 | 1601 | 2017-05-17T11:42:21 | Population Density Equations for Stochastic Processes with Memory Kernels | [
"physics.bio-ph",
"q-bio.NC"
] | We present a novel method for solving population density equations (PDEs), where the populations can be subject to non-Markov noise for arbitrary distributions of jump sizes. The method combines recent developments in two different disciplines that traditionally have had limited interaction: computational neuroscience and the theory of random networks. The method uses a geometric binning scheme, based on the method of characteristics, to capture the deterministic neurodynamics of the population, separating the deterministic and stochastic process cleanly. We can independently vary the choice of the deterministic model and the model for the stochastic process, leading to a highly modular numerical solution strategy. We demonstrate this by replacing the Master equation implicit in many formulations of the PDE formalism, by a generalization called the generalized Montroll-Weiss equation - a recent result from random network theory - describing a random walker subject to transitions realized by a non-Markovian process. We demonstrate the method for leaky- (LIF) and quadratic-integrate and fire (QIF) neurons subject to spike trains with Poisson and gamma distributed spike intervals. We are able to model jump responses for both models accurately to both excitatory and inhibitory input under the assumption that all inputs are generated by one renewal process. | physics.bio-ph | physics |
Population Density Equations for Stochastic Processes with Memory Kernels
Yi Ming Lai and Marc de Kamps
Institute for Artificial and Biological Computation
School of Computing
University of Leeds
LS2 9JT Leeds
United Kingdom
(Dated: August 7, 2018)
We present a novel method for solving population density equations (PDEs) - a mean field tech-
nique describing homogeneous populations of uncoupled neurons - where the populations can be
subject to non-Markov noise for arbitrary distributions of jump sizes. The method combines recent
developments in two different disciplines that traditionally have had limited interaction: compu-
tational neuroscience and the theory of random networks. The method uses a geometric binning
scheme, based on the method of characteristics, to capture the deterministic neurodynamics of the
population, separating the deterministic and stochastic process cleanly. We can independently vary
the choice of the deterministic model and the model for the stochastic process, leading to a highly
modular numerical solution strategy. We demonstrate this by replacing the Master equation implicit
in many formulations of the PDE formalism, by a generalization called the generalized Montroll-
Weiss equation - a recent result from random network theory - describing a random walker subject
to transitions realized by a non-Markovian process. We demonstrate the method for leaky- (LIF)
and quadratic-integrate and fire (QIF) neurons subject to spike trains with Poisson and gamma
distributed interspike intervals. We are able to model jump responses for both models accurately
to both excitatory and inhibitory input under the assumption that all inputs are generated by one
renewal process.
I.
INTRODUCTION
Population density techniques are widely used in
physics, biology, chemistry, finance and other areas of
science, often in the form of stochastic differential equa-
tion equations, or more generally in the form of the dif-
ferential Chapman-Kolmogorov (dCK) equation [1]. The
basic idea is always the same: the state of individuals in
the population is described by a combination of deter-
ministic laws that are known, and a noise process which
is statistically similar for all individuals, causing irregular
random state changes.
Population density techniques have a long standing his-
tory in computational neuroscience starting with [2 -- 5].
In particular, the last twenty years have seen an explosion
of interest in this area [6], as it now becomes clear that
although brain-sized simulations are technically possible
[7], the resulting models are unwieldy, in terms of the
number of parameters involved and the amount of data
generated. Increasingly, the population level is seen as
an appropriate mesoscopic description level for modeling
complex neural systems. For example, recently a cor-
tical column has been simulated with population-based
approaches, e.g.
[8, 9]. The development of techniques
that relate the mesoscopic population level to that of in-
dividual neurons is therefore vital to the brain sciences.
In the past, many applications have used stochas-
tic differential equations or alternatively Fokker-Planck
approaches:
initially often for leaky-integrate-and-fire
(LIF) neurons e.g. [10, 11], but later also for other models
such as quadratic- (QIF) or exponential-integrate-and-
fire neurons [12, 13], or even more complex ones such
as the conductance-based model in [14]. Many studies
have assumed weak synaptic effects, allowing the intro-
duction of a diffusion approximation and thereby the use
of Fokker-Planck or Langevin equations. However, it has
been argued that shot noise rather than Gaussian white
noise is required for realistic simulations, for example
by Richardson and Swarbrick [15] in the context of neo-
cortical populations. Furthermore, post-synaptic effects
are not necessarily small. Implicit in the formulation by
Omurtag et al. [16] is the possibility that synaptic jumps
are subject to Poisson statistics and may be large. Us-
ing a similar framework Nykamp and Tranchina [17] used
a smoothness approximation for the population density
that allows large synaptic inputs to be incorporated in a
numerical approach. de Kamps [18] and Iyer et al. [19]
have demonstrated that by using the method of charac-
teristics, a numerical scheme can be found for arbitrarily
large jumps without relying on a smoothness assumption.
By constructing a geometric binning scheme from the
characteristics, we are able to model the deterministic
neurodynamics by a shift of probability mass through
the bins, thereby avoiding the numerical difficulties in-
troduced by the drift term of the dCK equation. In this
non-equidistant binning scheme, the full dCK equation
is now reduced to a Poisson Master equation.
This means that the system is represented by a combi-
nation of probability shifts and a Master equation which
describes transitions due to a point process. In this pa-
per, we relax the assumption that the noise is Poisson
in nature. We can model the stochastic process as a
continuous-time random walk (CTRW) on a network of
states, and follow the approach of Hoffmann et al. [20] to
derive a generalized Montroll-Weiss equation; this leads
to an equation analogous to the Poisson master equation,
but with a convolution with a memory kernel based on
the inter-arrival distribution of the point process.
The importance of non-Poisson statistics has been
pointed out by Cateau and Reyes [21], who demonstrated
that some experimental data is better described by a
gamma distribution, and in a theoretical study showed
that the dynamics of a synfire chain is substantially af-
fected by the statistics of spike trains. Using a renewal-
based approach, Ly and Tranchina [22] were able to study
non-Poisson inter-spike intervals by constructing a two-
dimensional population density where one of the vari-
ables is the membrane voltage, and the other the time
since the last spike. In this paper, we present a different
approach, allowing us to create a general scheme suit-
able for different one-dimensional neuronal models for
arbitrary transition matrices, thereby treating excitation
and inhibition on the same footing, under the assumption
that all transitions are generated by the same renewal
process. Instead of a full two-dimensional treatment, we
start with a one-dimensional method and find that the
non-Markovian characteristics can be accounted for by
a convolution with the recent history of the probability
density of the population. Other studies on the effect of
non-Poisson noise, not directly related to the approach
here, consider various forms of colored noise injected into
individual neuron models and studied the output statis-
tics, for example [23, 24].
Figure 1: Dynamics of LIF (left) and QIF neurons (right).
Time t is shown as a function of V0 on an interval [Vmin, Vmax].
II. METHOD
We consider a population of neurons in the mean field
approximation:
in a homogeneous population neurons
are uncoupled, identical, but individually see different
realizations of the same statistical process. Larger inho-
mogeneous networks must be described as homogeneous
subpopulations and require assumptions on how the net-
work connectivity transforms output of one population
into the input of others, e.g. see [25]. Alternatively, we
consider a single neuron subjected to a large number of
repetitions of the same process. Under these assump-
tions a population of individuals behaving according to
V = F (V ), with V the membrane potential of a neuron,
can be described by the dCK equation (we will refer to
2
Figure 2: A geometric grid for LIF neurons. (The bin [0, V ∗]
is a fiducial bin used to avoid having exponentially many bins
near V = 0.)
the potential in lower case as an argument in the density
and in upper case when discussing individual neurons for
legibility):
(cid:90)
∂ρ
∂t
+
∂
∂v
(F ρ) =
dw {W (v w)ρ(w, t) − W (w v)ρ(v, t)} .
M
(1)
Here ρ(v, t) is the population density defined on an in-
terval M : ρ(v)dv is the fraction of neurons with potential
in [v, v + dv). W (v w) describes the transition density:
the probability per unit time that a neuron moves from
state v to state w. F (v) defines the neuron model. For
example, the LIF neuron is defined by F (V ) = −V /τ
where τ is the neuron time constant. Other models in-
clude the QIF:
F (V ) = (V 2 + I)/τ,
(2)
with I often interpreted as a control parameter, or the
exponential-integrate-and-fire model that we will not
discuss here. The method here applies to any one-
dimensional neural model. More complex neuronal mod-
els require more than one dimension; elsewhere we show
that it is possible to apply the geometric binning scheme
to two dimensional neural models [26]. In this paper we
will focus on one dimensional neural models as it allows
a simpler exposition of the method.
A. Geometric Binning
Our objective is to describe the evolution of the density
In the absence of synaptic input this
function ρ(v, t).
is described by the advective part of the dCK equation,
which could be solved numerically. However, geometrical
considerations give a particularly simple method. The
non-dimensionalised LIF neuron is usually stated as:
τ
dV
dt
= −V,
(3)
where V is the membrane potential, τ the membrane
time constant of the neuron, so the membrane voltage
decays exponentially in the absence of stochastic input.
Explicitly, t = τ ln V0
V is the time it takes for the neuron
to decay to a voltage V from an initial voltage V0.
We can use this to construct a set of characteristics
(Fig. 1 left).
If we consider a starting distribution of
neurons in a population ρ(V, t = 0), it is clear that neu-
rons between curves will remain between those curves as
V1-0.8-0.6-0.4-0.2-00.20.40.60.81t (s)00.0050.010.0150.020.0250.030.0350.040.0450.05LIFV10-8-6-4-2-0246810t (s)00.0050.010.0150.020.0250.030.0350.040.0450.05QIFVVmaxVmaxe(−∆tτ)Vmaxe(−2∆tτ)V∗V=0 time progresses, in the absence of input. We can dis-
cretize state space using equidistant steps in time, rather
than potential: starting at Vmax, we evolve Eq. 3 during
a time ∆t, and use the new value of V as a bin bound-
ary. Repeating the process, we approach the equilibrium
point V = 0. This is shown diagrammatically in Fig. 2.
The bins get exponentially smaller closer to V = 0, and
therefore we define a small constant V ∗ close to V = 0.
We define a fiducial bin [0, V ∗] where probability mass re-
mains stationary. Similarly, we approach the equilibrium
from the left hand side by starting at Vmin, and calcu-
lating the potential decay in steps of ∆t, which yields a
series of bin boundaries and break off in a similar way.
In practice we are free to pick Vmin, and usually pick a
value that yields the same bin boundaries left and right
of the equilibrium.
Starting from an arbitrary distribution of probability
over the grid, its evolution can be done essentially with-
out computation: each time step ∆t, the mass in each
bin - the fraction of the population present in that bin -
moves to the next bin in the direction of the equilibrium.
Mass that enters the equilibrium bin remains there. This
simple observation suggests that problem can be solved
by interleaving two steps: a shift of mass through a geo-
metric grid, followed by a numerical solution of the Mas-
ter equation which implements transport of mass from
bin to bin. In the following section we prove this.
B. The Master Equation in a Moving Coordinate
System
A key observation [19, 22, 27, 28] is that the method
of characteristics can be used to transform Eq. 1 into a
simpler one.
Consider the ordinary differential equation:
3
Figure 3: A shift of probability mass through the geometric
grid is sufficient to capture the evolution of the density profile
due to the deterministic neuronal dynamics. For LIF neurons
(top) mass moves in the direction of the equilibrium point.
Mass that enters the fiducial bin surrounding the equilibrium
point remains there. For QIF neurons with I < 0 (bottom),
there are two equilibrium points: one stable, one unstable.
Movement towards the stable equilibrium is similar to the
LIF case. Movement away from the unstable equilibrium to-
wards the threshold is upwards. Probability mass reaching
threshold must be removed from the system and reinserted at
a potential Vreset, possibly after observing a refractive period.
total derivative of the density i.e. along curves V (t) that
are solution to Eq. (4) we have
dV
dt
= F (V ),
(4)
dρ(V (t), t)
dt
=
∂ρ(V (t), t)
∂t
+
∂ρ(V (t), t)
∂V
dV
dt
,
and let V (t, v(0)) be a solution of Eq. (4) with v = v0
for t = 0 . It is possible to interpret this as a coordinate
transformation:
v(cid:48) = V (t, v)
t(cid:48) = t
(5)
In the new coordinate system Eq. (1) assumes a sim-
(cid:82) t
0
ρ(cid:48)(V (t), t) ≡ e
∂F (V (ξ))
∂v
dξρ(V (t), t)
(7)
This simpler form is explained by the observation that
along integral curves of the system, one can calculate the
(cid:90)
pler form:
dρ(cid:48)(v(cid:48), t)
dt
with
=
M
dw {W (v(cid:48) w)ρ(w) − W (w v(cid:48))ρ(v(cid:48))} ,
(6)
using the chain rule. The definition of Eq. (7) directly
leads to Eq. (6).
Equation (6) is just the Master equation of the noise
Intu-
process, albeit in a moving coordinate system.
itively, this makes sense:
in a coordinate system that
co-moves with the neuronal dynamics, all change must
come from the stochastic process. As an example, con-
sider Poisson distributed spike trains. For shot noise:
W (w v) = νδ(w − v − h) + (1 − ν)δ(w − v),
(8)
where h is the synaptic efficacy and ν the rate of the
Poisson process, and w, v arbitrary potential values. This
indicates that the only possibility for a jump is from a
potential v to v + h as the transition probability is 0
for all other transitions. The transition probability ex-
presses that an input spike causes an instantaneous jump
in membrane potential. For a Markov process ν and h
can be time dependent.
VmaxVmaxe(−∆tτ)Vmaxe(−2∆tτ)V=0VminVmine(−∆tτ)LIFVmaxV=√−IV=−√−ItospikingVminQIF Consider the case of a LIF neuron, F (V ) = −V /τ .
v(cid:48) = ve− t
t(cid:48) = t
τ
(9)
with ρ(cid:48)(v(cid:48), t) = e− t
τ ρ(v(cid:48), t), and Eq. (1) reduces to:
∂ρ
∂t −
1
τ
∂
∂v
(ρv) = ν(ρ(v − h) − ρ(v)),
(10)
4
Of course a single synaptic efficacy is unrealistic and in
practice one uses [17]:
(cid:90)
∂ρ
∂t −
1
τ
∂
∂v
(ρv) =
dhp(h)ν(ρ(v − h) − ρ(v)),
(11)
As we will argue below, this does not fundamentally
change the method.
After the coordinate transformation this becomes:
dρ(cid:48)(v(cid:48), t)
dt
= ν(ρ(cid:48)(v(cid:48) − he
t
τ ) − ρ(cid:48)(v(cid:48))),
(12)
where we have taken into account that Eq. (8) must now
be represented in v(cid:48)-space. This constitutes a consider-
able simplification: instead of solving partial differential
equation, one is faced with a system of ordinary differ-
ential equations. This comes at a price: one is forced to
represent the density not in a fixed interval in potential
space, but in a frame that moves with respect to that
interval. Moreover, as Eq. (12) shows, in that frame the
jumps are time dependent, even if they are constant in
the original frame. This precludes the analytic solution
for constant h given in [29].
The geometric binning scheme provides a method for
representing density in v(cid:48) coordinates. The entire method
now becomes a two step process. The first step consists of
a shift of probability mass, as explained in Fig. 3, which
represents the movement of neurons under the influence
of deterministic dynamics during a time step ∆t. The
second is the solution of the Master equation over a time
∆t
step ∆t, small enough for he
τ and ν(t) to be constant.
In the following section we will describe this process in
detail.
Although shown for LIF neurons, the method gener-
alizes in an obvious way. It is always the case that the
characteristics of Eq. 1 are given by the solutions of the
system τ dV
dt = F (V ), and therefore a geometric grid can
always be constructed by integrating this equation re-
gardless of whether analytic solutions are available, like
for QIF neurons, or a numerical solution is required. The
only subtlety that needs to be observed is there may be
multiple equilibria present; mass movement may be in op-
posite directions at either side of the equilibrium point.
This is illustrated in Fig. 3: LIF neurons have a sin-
gle stable equilibrium point, QIF neurons a stable and
unstable one (see Fig. 1).
It is best to think of a po-
tential interval bounded by two equilibrium points (or
the minimum or maximum potential) as an independent
strip, and capture probability mass movement in each
Figure 4: The shift of probability mass can largely be re-
placed by an index update. The black bars at the top indicate
a geometric grid. The array V stores the values of the grid
boundaries; the relationship between the contents of V and
the grid boundaries are indicated symbolically by red lines.
This relationship is immutable; it remains constant through-
out simulation. P is an array representing the probability
mass at time t. The top figure represents the situation at
t = 0: each element of P contains a numerical value repre-
senting the amount of probability mass. For a given element,
the blue arrow indicates the potential interval containing this;
together the V and P arrays represent a discretized density
profile. The evolution of the density profile is realized by up-
dating the relationship between the P and the V array, as
indicated in the bottom panel representing the density profile
at t = ∆t, by the change of the blue arrows. The contents
of the V array remain unchanged, as do the contents of P,
with the exception of two bins. The situation depicted in
this Figure shows decay towards a steady state that is rep-
resented by the third and fourth potential bin from the left.
Mass represented by the two outward pointing arrows on the
extreme left and right represents mass that has cycled from
the equilibrium bins to potential values at the extreme end of
the potential interval. If the stationary point is stable, this
is undesirable and this mass should be removed and added
to the elements of the mass array that currently point to the
equilibrium bins; this is indicated by the dashed arrows. This
Figure shows a neural model that has a single stable equilib-
rium point, such as the LIF model. For a model with more
than one stationary point, such as the QIF neuron for I < 0,
the V and P arrays must be separated into isolated strips,
each with their own relationship between the P and V array.
By updating the relationship between elements of P and V,
the shift of mass is captured almost entirely without moving
data around.
strip independently. The full potential interval is then
represented by a collection of these strips.
Finally, a point that is implementationally important.
Rather than shifting the data around as described, which
is computationally expensive, it makes more sense to keep
track of the position of each portion of probability mass
in the geometric grid. This reflects the observation that
the density profile is constant in v(cid:48) space. The process is
shown in Fig. 4.
Algorithmically, this introduces a considerable amount
VPVPt=0t=Δt of bookkeeping, which is described in some detail in [18],
but which we will ignore in the remainder of the paper
as it is not conceptually different from the method as
described above.
C. The Master Equation in a Geometric Grid
For simplicity, we will describe the solution to the Mas-
ter equation of the Poisson process. Extension to the
gMW equation will be straightforward. First, consider
the Master equation at t = k∆t. We need to formulate
the Master equation in a non-equidistant grid. Consider
the probability mass in bin i. This bin corresponds to
a potential interval [Vk(i), Vk(i + 1)]. Neurons that are
present in this mass bin will, when they receive an input
spike, move to a different potential and will be in the
interval [Vk(i) + h, Vk(i + 1) + h]. It is therefore a matter
of finding out which potential intervals are covered by
this interval, and by what proportion. This is a straight-
forward geometrical problem which is illustrated in Fig.
5.
Denote the set of mass bins covered by [Vk(i)+h, Vk(i+
1) + h] by Vk,i(h) and for bin j ∈ Vk,i(h) let mij denote
what proportion of bin j is covered by [Vk(i) + h, Vk(i +
1) + h] - then the Master equation becomes:
(cid:88)
j∈Vk,i(h)
dP[i]
dt
= ν{−P[i] +
or in vector-matrix notation:
mijP[j]},
(13)
dP
dt
= ν(−I + M )P,
(14)
where the elements of M are mij. The vector P is the
probability mass in our non-equidistant bins and corre-
sponds to a discrete version of the quantity ρ(cid:48)(V )dV from
the previous section. The process is identical for excita-
tory (h > 0) or inhibitory (h < 0) input.
Once we have computed this transition matrix M , we
know the probability of an event causing a transition
from any state to another. Therefore the full implemen-
tation of our method consists of two interleaved steps:
a probability shift between bins of the probability array,
corresponding to the deterministic neuronal dynamics,
while the effect of the stochastic input is captured by
solving the Master equation (14) on the corresponding
non-equidistant binning scheme.
The matrices M will be band matrices, i.e. sparse (see
Fig. 10 for an explicit example). A row typically re-
flects a position in the interval from where the neurons
leave and a position where the neurons arrive, with the
intermediate positions filled with zeros. Synaptic smear-
ing broadens the band of the arrivals, but as long as the
width of the synaptic distribution is small compared to
the simulation interval, the overall matrix M will still be
sparse. In practice, one samples the synaptic distribution
5
Figure 5: Coefficients for the Poisson Master equation are
purely determined by synaptic efficacy h and the bin bound-
aries. Two grids are shown above, with the bottom displaced
by h. The transition matrix is determined by what fraction
the displaced grid covers each bin of the original grid. For
example, the bin labelled "bin k" has length Lk. After dis-
placement by h, it overlaps bin k by some interval lk,k and bin
k + 1 by an interval lk,k+1. Therefore the transition matrix
would have entries mk,k = lk,k
, and so on.
Lk
At the same time, part of bin k + 1 is pushed above threshold,
which corresponds to spiking. The probability mass pushed
above threshold is placed in the bin containing Vreset at the
next time step (or after a delay if a refractory period is de-
sired). This procedure for bin k allows us to calculate the
k-th row of the transition matrix, so repeating the procedure
for all bins gives us the full matrix.
, mk,k+1 = lk,k+1
Lk
with a few well chosen synaptic efficacies, yielding a num-
ber of matrices - the overall matrix M is then a weighted
sum of these matrices. As this is done before simulation
starts, there is only a small effect on simulation time.
III. GENERALIZED MONTROLL-WEISS
EQUATION: BEYOND MARKOV
The master equation, when standing on its own, de-
scribes the behaviour of a random walker on a network,
where each interval in v-space is a node. The walker is
locked on a node, unless a connection to another node
appears at which point the walker must move instanta-
neously after which the connection vanishes. In the con-
text of computational neuroscience, the appearance of a
connection is the arrival of an input spike, which allows
the receiving neuron to move from its current membrane
potential to a different one. The probability of a con-
nection appearing is given by the previously-calculated
transition matrix.
Having used this abstraction, we now are able to ex-
tend our master equation method to other renewal pro-
cesses by using a generalized Montroll-Weiss (gMW)
equation for this network. The Montroll-Weiss equation
was originally used to model anomalous diffusion on reg-
ular lattices, and was recently generalized to networks by
Hoffmann et al. [20]. We now briefly restate the deriva-
tion of generalized Montroll-Weiss equation for our ex-
ample, following the approaches of [20, 30, 31]. We start
with a random walker in our state space with a waiting
time distribution (WTD) f (t).
We
are
interested
in
determining P(t)
=
VmaxV=0hspikinglk,klk,k+1Lkbink the
that
probability
(Hence qi(t) = (cid:80)∞k=0 qk
{P0(t), P1(t), . . . , PN (t)}
the
walker will be at any given state at a time t. We define
qi(t) to be the probability that a walker arrives at a
state i at exactly t, and qk
i (t) to be the probability that
a walker arrives at state i at time t having taken exactly
k steps.
If we know the
qk
i (t)s, we know that the probability of a walker being
at state j after k + 1 steps is the sum of the qk
i (t)s,
weighted by the probability of making a step from j to i
at the required time. Hence we construct the recursion
relation:
i (t).)
q(k+1)
j
(t) =
mijf (t − τ )qk
i (τ )dτ ,
(15)
(cid:90) t
(cid:88)
0
∀i
where the mijs are the coefficients of the transition ma-
trix induced by the synaptic efficacy h, as calculated in
Fig. 5. In Laplace space:
q(k+1)
j
(s) =
mij f (s)qk
i (s) .
(16)
∞(cid:88)
Summing over all k and adding q0
j (s) to both sides gives:
q0
j (s) +
q(k+1)
j
(s) =
mij f (s)qk
i (s) + q0
j (s) ,
k=0
k=0
∀i
which in matrix-vector notation is
q(s) = M f (s)q(s) + q0(s) .
so that
q(s) =
(cid:16)
(cid:17)−1
I − M f (s)
P(0) .
(17)
(18)
(19)
Now we know that the probability of a being at state i at
a time t must be equal to the probability of arriving at
state i at some τ < t and an event not occurring between
τ and t:
(cid:88)
∀i
∞(cid:88)
(cid:88)
(cid:90) t
0
Pi(t) =
(cid:82) t
Gi(t − τ )q(τ )dτ ,
(20)
where Gi(t) = 1−
0 f (t)dt is the probability of an event
Pi(s) =
not occurring in time t.
f (s)qi(s), allowing us to use our expression for q to give
us the gMW equation for our network:
In Laplace space:
(cid:16)
(cid:17)−1
P(s) =
1 − f (s)
s
I − M f (s)
P (0) .
(21)
Next, we use the identity L{dP/dt} = s P(s) − P (0) and
(1 − M f (s))P(s) from Eq.
substitute in P (0) =
s
1− f (s)
(21). Some rearrangement yields:
6
where the function K is the memory kernel and is defined
so that
K(s) :=
s f (s)
1 − f (s)
.
(24)
We note that this looks similar to our previous Master
equation - in the case where we have a Poisson process
of rate ν, K = νδ(t) and we obtain our previous Master
equation.
In general K does not have a closed-form solution and
has to be evaluated numerically. For a gamma distri-
bution of shape α and rate ν, K is the Laplace inverse
of
K =
sνα
(s + ν)α − να .
In the interests of simplicity, in the subsequent examples
we will consider α = 2 and 3, giving
K(t) α=2 = ν2 exp(−2νt)
ν2 exp(−
K(t) α=3 =
2√3
3
3
2
νt) sin(
√3
2
νt)
(25)
respectively. This is motivated by experimental data
showing gamma-distributed inter-spike intervals [21, 32].
However, we stress that our method also works on other
distributions for which the memory kernel has to be nu-
merically evaluated. Efficient computation of the Laplace
transform of other ubiquitous probability distributions
of inter-event statistics of renewal processes, such as the
Weibull or Pareto distributions, is an open area of re-
search, for example [33].
define Knorm := K/(cid:82) ∞
our examples, the normalisation constant (cid:82) ∞
We will compare the results of simulations between
Poisson input and gamma input. In order to do this, we
0 K(t)dt, so that Knorm∗ P is also
a probability distribution. This simplifies the equation
in some cases, for example, if the population reaches a
steady-state distribution Ps, then Knorm ∗ Ps = Ps. In
0 K(t)dt is
equal to the expectation value of the input spike train
(which is ν/α for a Γ(α, ν) distribution). Hence we can
cast the integro-differential gMW equation for gamma
input in the form:
dP
dt
=
ν
α
(M − I) (Knorm(t) ∗ P(t)) ,
(26)
allowing us to compare gamma distributions with differ-
ent shapes by varying the rate ν accordingly. For other
distributions, this comparison cannot always be done; for
example, the method is also suitable for evaluating inputs
with a power-law distribution which does not necessarily
have finite moments.
L{
dP
dt } = (M − I) K P(s)
dP
dt
= (M − I) (K(t) ∗ P(t)) ,
(22)
(23)
IV. RESULTS
We consider our population of LIF neurons to have a
membrane time constant of τ = 0.05 s, and begin with
a single Poisson input of 800Hz with synaptic efficacy of
0.03, which has been used as a benchmark in earlier stud-
ies [16]. We then take the natural extension to gamma
distributed inputs, and verify our method against Monte
Carlo simulations. In the figures here, the initial condi-
tion is that all the neurons in the population are at their
equilibrium V = 0, we have normalized the threshold
potential Vth so that Vth = 1 and dimensionless.
7
Figure 7: (a): Firing rates of the LIF neuron with inputs
from a Γ(α, ν) distribution. Lines are calculated using our
method, while markers are from Monte Carlo simulations of
10000 neurons. h = 0.1 in all cases. Solid line and crosses:
α = 1, ν = 150, i.e. a Poisson process with rate 150. Dashed
line and circles: α = 2, ν = 300. Dotted line and triangles:
α = 3, ν = 450. (ν is varied such that the expectation of the
input process remains the same across all cases.)
(b): the
steady state density profiles for the different shape factors.
reach its steady state firing rate.
In Fig. 7, we show
that changing the shape factor of the input distribution
can even change the steady-state firing rate in low firing
rate regimes. The density profiles are also significantly
affected by the shape factor.
We contrast this with a system without threshold,
where we see that decreased shape factor results in
a broader steady-state density distribution around the
same mean - see Fig. 8. The system we consider is a
generalization of the Ornstein-Uhlenbeck (OU) process.
The OU process is one of the most fundamental exam-
ples of a stochastic process and is often used as a canon-
ical example when developing techniques in the study of
stochastic differential equations (SDEs) in various fields
([34 -- 36]). It is often written as: dxt = θ(µ−xt)dt+σdWt,
where Wt is the Wiener process. Here we replace the
Wiener process jump process with an arbitrary proba-
Figure 6: Firing rates of the LIF neuron with inputs from a
Γ(α, ν) distribution. Lines are calculated using our method,
while markers are from Monte Carlo simulations of 10000 neu-
rons. h = 0.03, Vth = 1 in all cases. Solid line and crosses:
α = 1, ν = 800, i.e. a Poisson process with rate 800. Dashed
line and circles: α = 2, ν = 1600. Dotted line and triangles:
α = 3, ν = 2400. (ν is varied such that the expectation of the
input process remains the same across all cases.)
In Fig. 6 we observe good agreement with Monte Carlo
simulations in the firing rate. We note that our method
works much faster, as the computational load scales ap-
proximately linearly with the number of bins in our dis-
cretized characteristic space (which does not depend on
the system size), while the Monte Carlo simulations scale
with the number of neurons. In this paper we use on the
order of tens of thousands of neurons in our Monte Carlo
simulations.
We also see that for higher shape factors, the popula-
tion experiences stronger transients and takes longer to
0.000.050.100.150.200.250.30Time (s)051015202530Firing rate (Hz)0.000.050.100.150.200.250.30Time (s)0.00.51.01.52.02.53.03.54.0Firing rate (Hz)(a)0.00.20.40.60.81.0V0.00.51.01.52.02.53.03.54.0Probability density(b) bility density function for the time between jumps.
In
the absence of noise, the variable xt relaxes to µ with
a time constant θ. We consider a dimensionless version
where dx/dt = −x between jumps. For the stochastic
part, we consider the variable x to have jumps of size h
with the interval between jumps distributed according to
the gamma distribution with shape α and rate ν.
8
Figure 9: Gain curves for h = 0.15. Solid line: shape =
1 (Poisson process). Dashed line: shape = 2. Dotted line:
shape = 3. Input firing rate is expected input ν/α.
Figure 8: Steady-state density of the generalised OU process.
Lines are calculated using our method, while markers are from
Monte Carlo simulations of 20000 neurons. h = 0.1 in all
cases. Solid line and crosses: α = 1, ν = 10, i.e. a Poisson
process with rate 10. Dashed line and circles: α = 2, ν = 20.
Dotted line and diamonds: α = 3, ν = 30.
Returning to our study of the LIF neuron, by consid-
ering a 'gain curve' (Fig. 9) of steady-state output firing
rate against input firing rate for different inter-spike dis-
tributions, we can identify regions of parameter space
where one would expect to see significant differences in-
duced by different shape factors. As we can see from
the gain curves, for the same expected input and efficacy
an increased shape factor decreases the firing rate. While
this effect is only slight at high input firing rates, it is sig-
nificant at lower firing rates, and we can see it changes
the threshold input required for firing.
By using the assumption by Cateau and Reyes [21]
that a neuron experiences a superposition of many spike
trains with little connectivity between them, so that the
conglomerate spike train can be modeled as a single re-
newal process, we can study the balance of excitation and
inhibition (Fig. 11). In the vein of previous studies such
as [25], we consider a 4:1 ratio of excitatory to inhibitory
input, and a corresponding 1:4 ratio of synaptic efficacy.
We generate a spike train with gamma distributed in-
terspike intervals for different shape factors with a given
input rate ν as a marked point process. We perform a
Bernoulli trial on each spike to determine whether it is
excitatory (pe = 0.8), or inhibitory (pi = 1 − pe). An ex-
citatory spike will contribute an instantaneous jump of
magnitude he = 0.05 to the membrane potential while
an inhibitory spike contributes a jump of magnitude
hi = −4he. The resulting Monte Carlo simulations are
given in Fig. 11. To construct a population density ver-
sion of this process, we generated two matrices Me and
Mi by the process outlined in Fig. 5 and add them to
obtain a single matrix M ≡ peMe + piMi, whose struc-
ture we represent visually in Fig. 10. Using discretized
versions of the kernel Eq. (25) then allows us to solve
Eq. (26) numerically for this case. Again, we interleave
solutions over a time ∆t with the mass shift procedure
to obtain the results of Fig. 11 (solid curves). There
is good agreement with the Monte Carlo process, and
both methods predict a small output firing rate, which
is variability driven, given that the expectation value of
the input contribution is 0. Surprisingly, we see no dis-
cernible dependency on the shape factor here.
Finally, we can easily obtain results from other neu-
ronal models as well. In Fig. 12 we show the steady state
density profile of a population of QIF neurons (I = −1),
as well as the transient firing rate as response to a jump
in input. Whilst the rate responses look qualitatively
similar, the density profile looks different: neurons tend
to cluster in the ghost of the attractor (the stable fixed
point at V = −1). However, both the LIF (Fig. 7 (bot-
tom)) and QIF cases display a shift in the peak of the
probability density due to the shape factor.
We attribute this shift in the peak to neurons return-
ing to a lower potential value after having been pushed
through threshold. Lower shape factors imply larger vari-
ability, and therefore more neurons being pushed across
threshold. These neurons will reappear at the reset po-
tential and move upwards in V , contributing to the den-
sity below the expectation value and a leftwards shift in
the peak. We note that in the absence of a threshold no
such shift is observed, as seen in Fig. 8.
0.00.51.01.52.0x−0.50.00.51.01.52.02.53.03.5Probability density6080100120140160180200220240Input firing rate (Hz)0510152025Output firing rate (Hz) 9
Figure 10: The transition matrix M as an operator for moving
probability mass from bin i to bin j. Synaptic noise removes
neurons from their current position, resulting in a loss term
along the diagonal. Neurons undergoing an excitatory jump
move up in potential and thereby end up in the mass array
with a higher bin number. Neurons undergoing inhibition end
up at a lower bin number, at a larger distance from the di-
agonal, reflecting hi = −4he. The complex shape of the two
bands is a result of using a geometric grid: near the reversal
potential the same jump in potential covers more bins. The
reversal bin is larger than the neighboring geometric bins, so
upon translation covers a large number of them. This repre-
sents the straight part of the bands.
Multiple renewal processes
A key assumption in our analysis so far is that our in-
puts, whether excitatory or inhibitory, can be assumed
to be from a single conglomerate renewal process. Re-
laxing this assumption is difficult since superpositions of
renewal processes are not themselves renewal processes
(except the case where the component processes are Pois-
son [37, 38]). Hoffmann et al. [20] derive a generalized
Montroll-Weiss equation for a random walker on a net-
work where transitions between nodes can be from dif-
ferent renewal processes. They do this by assuming that
after a move is made by the random walker, the clocks
of all renewal processes are reset. As such, only the joint
probability distribution of the first event has to be used
in the derivation, as opposed to a full description of a
superposition of processes.
However, in the context of a neuronal population re-
ceiving inputs from external sources or other populations,
we cannot usually rely on this assumption. We briefly ex-
amine what occurs if we naively use the approach from
[20] to model a population receiving excitatory and in-
hibitory inputs, each of which is a process with inter-
arrival times given by Γ(2, 2ν).
In the case of a single
conglomerate gamma process, this would give us a nor-
0 K(t) = ν, and for two inputs,
one would expect a combined value of 2ν. However, when
we compute the memory kernel for two inputs, we instead
malization constant (cid:82) ∞
Figure 11: Firing rates of the LIF neuron with inputs from a
Γ(α, ν) distribution. Lines are calculated using our method,
while markers are from Monte Carlo simulations of 10000 neu-
rons.
In all cases, an input spike has an 0.8 probability of
being excitatory (h = 0.05) and an 0.2 probability of being in-
hibitory (h = −0.2). Solid line and crosses: α = 1, ν = 2000,
i.e. a Poisson process with rate 2000. Dashed line and circles:
α = 2, ν = 4000. Dotted line and triangles: α = 3, ν = 6000.
(ν is varied such that the expectation, ν/α, of the input pro-
cess remains the same across all cases.)
obtain(cid:82) ∞
0 K(t) = 8ν/5, i.e. a suppression to 80% of the
value that one would expect based on the individual pro-
cesses.
(cid:82) ∞
In Fig. 13 we examine the accuracy of this assump-
tion. We see that in the Fig. 13 (a), when both pro-
cesses make a comparable contribution, that there is rea-
sonable agreement between the method and Monte Carlo
simulations. However, where he (cid:29) hi, one would expect
a convergence to the single channel result (i.e. where
0 K(t) = 2ν).
In the Fig. 13 (b), we extend our
regime to higher values of he, and indeed we see that
Monte Carlo simulations approach the single channel re-
sult (which we label as "theoretical correction"), while
the gMW equation keeps predicting a reduction. This
is because despite the relative insignificance of the in-
hibitory input spikes, they reset the clock for the excita-
tory process, leading to erroneous predictions.
0.000.050.100.150.200.250.30Time (ms)0123456Firing rate (Hz) 10
Figure 12: (a): Firing rates of the QIF neuron with inputs
from a Γ(α, ν) distribution. Lines are calculated using our
method, while markers are from Monte Carlo simulations of
10000 neurons. The population has a time constant τ = 0.01
and a constant current I = −1, and the stochastic input has
a synaptic efficacy h = 0.2 in all cases. Solid line and crosses:
α = 1, ν = 500, i.e. a Poisson process with rate 500. Dashed
line and circles: α = 2, ν = 1000. Dotted line and triangles:
α = 3, ν = 1500. (ν is varied such that the expectation of
the input process remains the same across all cases.) (b): the
steady state density profiles for the different shape factors.
The membrane potential has been renormalized so the Vth =
1.0 and is dimensionless.
V. CONCLUSION AND FUTURE WORK
We have demonstrated a method for numerically solv-
ing population density equations that separates the de-
terministic and stochastic processes. The dynamics of
the deterministic process are reflected in the choice of
grid for the probability mass. Deterministic motion can
be accounted for by shifting the mass through the grid.
This just leaves the problem of solving the equation de-
termining the mass transfer due to the stochastic pro-
cess. For a Poisson process this is an extremely simple
Figure 13: The firing rate of a LIF population with two in-
put processes, as a function of excitatory synaptic efficacy.
The solid lines are the simulation of Poisson noise with our
method, while the dash-dotted line is the solution of the
Montroll-Weiss equation for the gamma processes. The Monte
Carlo simulations are the triangles and crosses with error bars.
τ = 0.05, ν = 500. The inhibitory synaptic efficacy hI is fixed
at 0.15.
system of ordinary differential equations, with a resulting
method that is manifestly insensitive to the gradient of
the density profile.
The separation between deterministic and stochastic is
general and makes no assumptions about the nature of
the stochastic process. Therefore other methods for de-
scribing stochastic processes than Master equations can
be incorporated. We demonstrated this explicitly by
adopting a recent result from random network theory:
the generalized Montroll-Weiss equations. This leads im-
mediately to a formulation of population density equa-
tions for stochastic processes with a memory kernel.
Arbitrary synaptic distributions can be specified by
choosing the appropriate transition matrices, and inhibi-
tion does not have to be considered as a separate special
case - we can study systems with balanced excitation
and inhibition.
In general, one can model a synaptic
distribution by using a superposition of transition matri-
(a)−10−50510V0.00.20.40.60.81.01.2Probability density(b)0.100.120.140.160.180.200.22hE051015202530354045Firing rate (Hz)Calculated firing rates, Gamma processCalculated firing rates, Poisson processMonte Carlo simulation, Gamma processMonte Carlo simulation, Poisson process(a)0.100.120.140.160.180.200.22hE051015202530354045Firing rate (Hz)Calculated firing rates, Gamma processCalculated firing rates, Poisson processMonte Carlo simulation, Gamma processMonte Carlo simulation, Poisson process(b) ces. Therefore, modeling learning can be easily accom-
modated for as it amounts to a reweighting of these ma-
trices in the computation of the final transition matrix.
As this final matrix is sparse, that can be done efficiently.
To summarize, we first construct a geometric binning
using the method of characteristics, such that the deter-
ministic dynamics of a neuron model can be captured
by a probability shift from one bin to the next at each
time step ∆t. Between these steps we solve Eq.
(26)
numerically using the forward Euler method. This re-
quires sampling the history at each time step; neverthe-
less, the resulting algorithm is still more efficient than
Monte Carlo simulation. Furthermore, in most cases the
memory kernel is of finite width, allowing further com-
putational savings on the convolution and less memory
of the process stored.
We note that Eq. (26) is of a simpler form than that in
Hoffmann et al. [20]. There the authors construct a gen-
eral method for a random walker on a network with ar-
bitrary WTDs between nodes, whereas we consider that
our input process has a single common WTD. This is due
to the difference in the underlying assumptions - in their
case they assume that the clocks of all WTDs are reset
when the walker makes a move. On the other hand, from
a neuroscience perspective, a neuron receiving an input
spike should not affect the clocks of the neurons emitting
said spikes. A similar assumption was used in [22]. In
this framework, we are able to consider distributions of
synaptic efficacies, as well as mixed excitation and inhi-
bition, as long as the conglomeration of spike trains can
still be modeled as a renewal process.
Dealing with the superposition of renewal processes in
general is still an open problem in mathematics, as the
superposition is in general no longer a renewal process:
it is a renewal process if and only if both processes are
Poisson [37, 38].
The method is not necessarily restricted to one dimen-
sional neuronal models - as would be required for e.g.
neurons displaying a limit cycle. We have successfully
implemented a method for two dimensional neural mod-
els [26], and anticipate that there may be some additional
computational overhead due to the need to retain a his-
tory of densities, but expect an otherwise straightforward
generalization.
11
where Eα,β(z) :=(cid:80)∞n=0 zn/Γ(β + αn) is the generalized
convolution K ∗ P = (cid:82) t
We have briefly mentioned power-law distributions.
One probability distribution that behaves as a power
law asymptotically (for 0 < ν ≤ 1) is the Mittag-Leffler
(t) = tβ−1Eβ,β(−tβ),
distribution [39], which has f M L
Mittag-Leffler function. This has the nice property that
f M L
(s) = (1 + sβ)−1, so the memory kernel in our case
β
would be K = L−1[s1−β] = tβ−2/Γ(β − 1). Hence the
(t−τ )β−2
Γ(β−1) P(τ )dτ , which is sim-
ply the Caputo fractional derivative [40] D1−β
t P(t). Our
gMW equation therefore becomes (by taking a fractional
integral on both sides):
β
0
Dβ
t P(t) = (M − I)P(t) .
(27)
It would be interesting to explore the implications of this
fractional differential equation for the population den-
sity. For example, there may be a connection with the
model for adaptation posed by Teka et al. [41], where a
fractional derivative is introduced in the LIF model itself.
As our method can be applied to any dynamical sys-
tem with jump noise, we hope that our method is use-
ful beyond computational neuroscience. An obvious
application area is queuing theory, where the class of
G/D/k queues handles events that arrive stochastically,
but where the queues themselves operate deterministi-
cally.
The main limitations of our method are in studying su-
perpositions of processes which cannot be approximated
by conglomerate renewal process; and renewal processes
with time-varying parameters, which is an important out-
standing problem.
VI. ACKNOWLEDGMENT
This project received funding from the European
Union's Horizon 2020 research and innovation pro-
gramme under grant agreement No.
720270 (Human
Brain Project).
[1] C. W. Gardiner, Handbook of Stochastic Methods
(Springer, 1997), 2 ed.
[2] R. B. Stein, Biophysical Journal 5(2), 173 (1965).
[3] P. I. M. Johannesma, Stochastic neural activity: A theo-
[7] S. Kunkel, T. C. Potjans, J. M. Eppler, H. E. Plesser,
A. Morrison, and M. Diesmann, Frontiers in Neuroinfor-
matics 5 (2011).
[8] N. Cain, R. Iyer, C. Koch, and S. Mihalas, PLoS Comput
retical investigation (1966).
Biol 12(9), e1005045 (2016).
[4] B. W. Knight, The Journal of General Physiology 59(6),
[9] T. Schwalger, M. Deger, and W. Gerstner, PLOS Com-
767 (1972).
[5] W. Gerstner and J. L. van Hemmen, Network: Compu-
tation in Neural Systems 3(2), 139 (1992).
[6] B. W. Knight, D. Manin, and L. Sirovich, Dynamical
models of interacting neuron populations in visual cortex
(1996).
putational Biology 13(4), e1005507 (2017).
[10] L. M. Ricciardi, Diffusion Processes and Related Topics
in Biology (Lecture Notes in Biomathematics) (Springer-
Verlag, 1977).
[11] N. Brunel and S. Sergi, Journal of Theoretical Biology
195(1), 87 (1998).
12
[12] N. Brunel and P. E. Latham, Neural Computation
15(10), 2281 (2003).
[13] N. Fourcaud-Trocm´e, D. Hansel, C. van Vreeswijk, and
N. Brunel, The Journal of Neuroscience 23(37), 11628
(2003).
[14] M. J. Richardson, Phys. Rev. E (2004).
[15] M. J. E. Richardson and R. Swarbrick, Phys. Rev. Lett.
105(17) (2010).
[16] A. Omurtag, B. W. Knight, and L. Sirovich, Journal of
[26] M. de Kamps and Y. M. Lai (In Preparation).
[27] M. de Kamps, Neural Computation 15(9), 2129 (2003).
[28] T. Schwalger and B. Lindner, Physical Review E 92(6),
062703 (2015).
[29] An analytic solution of the reentrant Poisson master
equation and its application in the simulation of large
groups of spiking neurons (2006).
[30] E. W. Montroll and G. H. Weiss, J. Math. Phys. (1965).
[31] V. M. Kenkre, E. W. Montroll, and M. F. Schlesinger, J.
Computational Neuroscience 8(1), 51 (2000).
Stat. Phys. (1973).
[17] D. Q. Nykamp and D. Tranchina, Journal of Computa-
tional Neuroscience 8(1), 19 (2000).
[18] M. de Kamps, ArXiv e-prints (2013), 1309.1654.
[19] R. Iyer, V. Menon, M. Buice, C. Koch, and S. Mihalas,
PLoS Comput Biol 9(10), e1003248+ (2013).
[32] J. Troy and J. Robson, Visual Neuroscience 9, 535 (1992).
[33] A. G. Rossberg 45(2), 531 (2008).
[34] J. L. Doob, Annals of Mathematics 43(2), pp. 351 (1942).
[35] K. C. Chan, G. A. Karolyi, F. A. Longstaff, and A. B.
Sanders, The Journal of Finance 47(3), 1209 (1992).
[20] T. Hoffmann, M. A. Porter, and R. Lambiotte, Phys.
[36] E. Bibbona, G. Panfilo, and P. Tavella, Metrologia 45(6),
Rev. E 86, 046102 (2012).
S117 (2008).
[21] H. Cateau and A. D. Reyes, Phys. Rev. Lett. 96(5)
[37] S. M. Samuels, Journal of Applied Probability pp. 72 -- 85
(2006).
(1974).
[22] C. Ly and D. Tranchina, Neural Computation 21(2), 360
[38] J. A. Ferreira, Stochastic processes and their applications
(2009).
86(2), 217 (2000).
[23] F. Muller-Hansen, F. Droste., and B. Lindner, Phys. Rev.
[39] Mittag-Leffler Functions, Related Topics and Applica-
E (2015).
[24] L. Shiau, T. Schwalger, and B. Lindner, J. Comput. Neu-
rosci. (2015).
tions (Springer, 2014).
[40] M. Caputo, Geophys. J. R. Astron. Soc. (1967).
[41] W. Teka, T. M. Marinov, and F. Santamaria, PLoS Com-
[25] D. J. Amit and N. Brunel, Network: Comput. Neural
put. Biol. (2014).
Syst. (1997).
|
1509.00418 | 1 | 1509 | 2015-09-01T17:59:08 | The mean shape of transition and first-passage paths | [
"physics.bio-ph",
"cond-mat.soft",
"cond-mat.stat-mech"
] | We calculate the mean shape of transition paths and first-passage paths based on the one-dimensional Fokker-Planck equation in an arbitrary free energy landscape including a general inhomogeneous diffusivity profile. The transition path ensemble is the collection of all paths that do not revisit the start position $x_A$ and that terminate when first reaching the final position $x_B$. In contrast, a first-passage path can revisit but not cross its start position $x_A$ before it terminates at $x_B$. Our theoretical framework employs the forward and backward Fokker-Planck equations as well as first-passage, passage, last-passage and transition-path time distributions, for which we derive the defining integral equations. We show that the mean time at which the transition path ensemble visits an intermediate position $x$ is equivalent to the mean first-passage time of reaching the starting position $x_A$ from $x$ without ever visiting $x_B$. The mean shape of first-passage paths is related to the mean shape of transition paths by a constant time shift. Since for large barrier height $U$ the mean first-passage time scales exponentially in $U$ while the mean transition path time scales linearly inversely in $U$, the time shift between first-passage and transition path shapes is substantial. We present explicit examples of transition path shapes for linear and harmonic potentials and illustrate our findings by trajectories generated from Brownian dynamics simulations. | physics.bio-ph | physics | The mean shape of transition and first-passage paths
Won Kyu Kim∗ and Roland R. Netz†
Department of Physics, Freie Universitat Berlin,
Arnimallee 14, 14195 Berlin, Germany
We calculate the mean shape of transition paths and first-passage paths based on the one-
dimensional Fokker-Planck equation in an arbitrary free energy landscape including a general in-
homogeneous diffusivity profile. The transition path ensemble is the collection of all paths that do
not revisit the start position xA and that terminate when first reaching the final position xB. In
contrast, a first-passage path can revisit but not cross its start position xA before it terminates at
xB. Our theoretical framework employs the forward and backward Fokker-Planck equations as well
as first-passage, passage, last-passage and transition-path time distributions, for which we derive
the defining integral equations. We show that the mean time at which the transition path ensemble
visits an intermediate position x is equivalent to the mean first-passage time of reaching the starting
position xA from x without ever visiting xB. The mean shape of first-passage paths is related to
the mean shape of transition paths by a constant time shift. Since for large barrier height U the
mean first-passage time scales exponentially in U while the mean transition path time scales linearly
inversely in U , the time shift between first-passage and transition path shapes is substantial. We
present explicit examples of transition path shapes for linear and harmonic potentials and illustrate
our findings by trajectories generated from Brownian dynamics simulations.
I.
INTRODUCTION
For a reaction involving a free energetic barrier, the
ensemble of transition paths is the collection of all paths
that lead from the reactant to the product ensemble
without recrossing the boundaries between the transi-
tion domain and the reactant domains [1 -- 3]. For con-
tinuous paths described by the Fokker-Planck equation,
transition paths can be generated by imposing absorb-
ing boundary conditions on the boundaries between the
reactant, transition and product domains [4]. The mean
transition path time τ T P is the first moment of the transi-
tion path time distribution. Based on an explicit formula
derived by A. Szabo for the one-dimensional case [4], τ T P
is for a large free-energetic barrier U much shorter than
Kramers' mean first-passage time τ KF P . Note that a
first-passage path is allowed to revisit its origin many
times and in the Fokker-Planck description is obtained
by imposing a reflecting boundary condition at the start
position. In fact, while Kramers' mean first-passage time
grows exponentially with the energy barrier height U , the
mean transition path time decreases linearly inversely in
U for a fixed separation between the start and final po-
sition along the one-dimensional reaction coordinate [5].
This means that in a reaction involving a large energetic
barrier, the system spends an exponential amount of time
revisiting the reactant state, while the actual transition
occurs very quickly [6, 7].
Although transition paths are crucial for the under-
standing of rare events, they are in typical experiments
that measure reaction rates not directly accessible. This
∗Electronic address: [email protected]
†Electronic address: [email protected]
situation dramatically changed with the advent of high
resolution single molecule experiments that allow to ac-
tually observe the folding and unfolding transition paths
of proteins [6 -- 10] as well as nucleic acid molecules [11 --
13]. Note that in these experiments, reaction paths are
typically obtained from the FRET efficiency between
fluorophores connected to molecular positions that al-
low to separate folded from unfolded state. As such,
these experiments project the complex molecular dynam-
ics onto a one-dimensional reaction coordinate that cor-
responds to an intramolecular distance, which motivated
extensive theoretical work using models restricted to one-
dimensional diffusion (though it is clear that a projec-
tion into one dimension does not necessarily mean that a
Markovian description is valid). Indeed, in these experi-
ments it was found that the mean transition path time is
significantly smaller than the folding or unfolding time.
In fact, the transition typically occurs so quickly that
only upper estimates can experimentally be obtained,
which from early estimates of about τ T P < 200 µs for
proteins as well as RNA [6, 7, 11], has come down to
τ T P < 10 µs with improved experimental time resolu-
tion [8 -- 10, 12].
The experimental advances created theoretical inter-
est in transition paths and led to intense simulation ac-
tivities [14 -- 16] as well as the development of analytic
approaches [5, 17, 18]. In this work, we present a theo-
retical framework for transition paths involving a combi-
nation of the backward Fokker-Planck equation, the for-
ward Fokker-Planck equation, and the renewal equation
approach, and use it to derive the mean shape of transi-
tion paths. We use the same framework to also calculate
the mean shape of Kramers' first-passage paths. Interest-
ingly, first-passage and transition path shapes are identi-
cal modulo a shift by constant time which correspond to
the residence time at the start position and is given by
5
1
0
2
p
e
S
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
8
1
4
0
0
.
9
0
5
1
:
v
i
X
r
a
the difference of Kramers' mean first-passage time and
the mean transition path time. We present explicit re-
sults for transition path shapes for constant, linear and
harmonic potentials and illustrate our findings with tran-
sition and first-passage paths generated using Brownian
dynamics simulations.
II. DERIVATION OF TRANSITION PATH
TIMES AND SHAPES
The Fokker-Planck (FP) operator is defined as [19 -- 21]
L(x) = ∂xD(x)e−F (x)∂xeF (x),
(1)
where F (x) is the free energy in units of the thermal en-
ergy kBT and D(x) is the position-dependent diffusivity.
In our previous analysis of protein folding trajectories
from molecular dynamics trajectories we found that the
diffusivity profile has a pronounced spatial dependence,
together with the free energy profile it allows to predict
kinetics that is rather insensitive on the precise definition
of the reaction coordinate [22]. But even for the much
simpler system of two water molecules diffusing relative
to each other the diffusivity profile is not constant and
therefore is important to take into account [23]. The
Green's function can be formally written as
G(x, tx0) = etL(x)δ(x − x0).
(2)
It fulfills the initial condition
G(x, 0x0) = δ(x − x0),
and solves the forward FP equation
∂tG(x, tx0) = L(x)G(x, tx0).
The adjoint FP operator [19 -- 21]
L†(x0) = eF (x0)∂x0D(x0)e−F (x0)∂x0 ,
solves the backward FP equation
∂tG(x, tx0) = L†(x0)G(x, tx0).
(3)
(4)
(5)
(6)
We will in the following section first use the backward FP
equation, as it allows to derive transition path times and
first-passage times in a most transparent and direct fash-
ion. We will then use the forward FP approach, which
requires careful normalization of expectation values but
allows to calculate mean passage times and from that var-
ious relations between mean transition path, first-passage
and passage times. Finally, we use the renewal equation
approach to derive constitutive relations between transi-
tion path time, first-passage time and last-passage time
distributions. Here we will be able to present a clear in-
terpretation of the expression derived for the mean shape
of transition and first-passage paths.
2
A. Backward Fokker-Planck approach
1. First-passage time distributions
(cid:90) xB
The derivation in this section uses concepts and tech-
niques presented previously in [20, 21]. By assuming ab-
sorbing boundary conditions at positions xA and xB we
calculate first-passage times for paths that start at x0
with xA < x0 < xB and reach the boundaries for the
first time. For this we define the survival probability
S(x0, t) =
dxG(x, tx0),
(7)
xA
that the paths have not reached yet an absorbing bound-
ary with the obvious properties S(x0, 0) = 1 and, for
regular free energies, S(x0,∞) = 0. The first-passage
distribution for reaching either one of the boundaries is
defined as
K(xA ∨ xB, tx0) = −∂tS(x0, t),
and by using Eq. (4) can be rewritten as
K(xA ∨ xB, tx0) = −
dxL(x)G(x, tx0)
(cid:90) xB
(cid:90) xB
xA
(8)
(9)
(10)
dx ∂xj(x, tx0)
=
= j(xB, tx0) − j(xA, tx0),
xA
where we used the flux at position x defined as
j(x, tx0) = −D(x)e−F (x)∂xeF (x)G(x, tx0).
This shows that the total first-passage distribution
can be decomposed into the two first-passage distribu-
tions K(xA, tx0) = −j(xA, tx0) and K(xB, tx0) =
j(xB, tx0) corresponding to the respective boundary
fluxes according to
K(xA ∨ xB, tx0) = K(xA, tx0) + K(xB, tx0).
(11)
(10) on
By applying the flux operator defined in Eq.
both sides of the backward FP equation Eq. (6) we ob-
tain explicit equations for the first-passage distributions
K(xA, tx0) and K(xB, tx0) as
∂tK(xA/B, tx0) = L†(x0)K(xA/B, tx0).
(12)
(cid:90) ∞
Defining the n-th moments of the first-passage distribu-
tions as
K (n)(xA/Bx0) =
dt tnK(xA/B, tx0),
(13)
0
we obtain from Eq. (12) the set of equations
− nK (n−1)(xA/Bx0) = L†(x0)K (n)(xA/Bx0),
(14)
where in the derivation we used the boundary condition
that K(xA/B, tx0) = 0 for t = 0 and t = ∞. Thus
all moments can be calculated recursively by straight-
forward integration of Eq. (14). The zeroth moment of
the first-passage distribution is nothing but the splitting
probability,
not detailed. We write Eq. (16) explicitly for φB(x0),
eF (x0)∂x0D(x0)e−F (x0)∂x0φB(x0) = 0.
(21)
φA/B(x0) = K (0)(xA/Bx0),
(15)
Integrating once we obtain
which gives the probability that a path starting at x0
reaches the boundary at xA or xB. From Eq. (14) we
obtain for n = 0
eF (x)
D(x)
= C∂xφB(x),
(22)
3
L†(x0)φA/B(x0) = 0.
and S(x0,∞) = 0 we conclude that(cid:82) ∞
From Eq. (8) and the boundary conditions S(x0, 0) = 1
0 dt [K(xA, tx0) +
K(xB, tx0)] = 1, in other words, the sum of the split-
ting probabilities is unity, eventually the path reaches a
boundary,
(16)
φA(x0) + φB(x0) = 1.
(17)
For n = 1 we obtain from Eq. (14)
L†(x0)K (1)(xA/Bx0) = −φA/B(x0).
(18)
Since the first-passage distributions K (1)(xAx0) and
K (1)(xBx0) are not normalized, reflected by the fact
that the splitting probabilities φA/B(x0) are smaller than
unity, the mean first-passage times are after normaliza-
tion given by
τ F P (xA/Bx0) =
K (1)(xA/Bx0)
φA/B(x0)
.
(19)
As a side remark, the mean first-passage time to reach
either the boundary xA or xB is given by the sum of
the first moments τ F P (xA ∨ xBx0) = K (1)(xAx0) +
K (1)(xBx0). Adding the two equations for K (1)(xAx0)
and K (1)(xBx0) in Eq. (18) and using that φA(x0) +
φB(x0) = 1 we arrive at the familiar equation [20, 21]
L†(x0)τ F P (xA ∨ xBx0) = −1.
(20)
2. Splitting probabilities
We explicitly show the calculation of the splitting prob-
abilities, all further calculations proceed similarly and are
(cid:90) x0
xA
where C is an integration constant that will be deter-
mined later. Another integration yields
C−1
dx
eF (x)
D(x)
= φB(x)x0
xA
= φB(x0),
(23)
where we used that φB(xA) = 0, i.e., a path that starts
at the absorbing boundary at xA will be immediately
absorbed and the probability to reach xB vanishes. Con-
versely, φB(xB) = 1 and thus
(cid:90) xB
C =
dx
eF (x)
D(x)
.
(24)
xA
For φA(x0) we obtain
φA(x0) = 1 − φB(x0) =
(cid:90) xB
x0
1
C
dx
eF (x)
D(x)
.
(25)
3. Mean first-passage times
From Eq.
(18) and using the results for φA(x0)
and φB(x0) in Eqs.
(23) and (25) we can straightfor-
wardly calculate the first moments of the first-passage
distributions. The boundary conditions require some
thought: The mean first-passage time to reach either
absorbing boundary, τ F P (xA ∨ xBx0) = K (1)(xAx0) +
K (1)(xBx0), vanishes at the boundaries, i.e., τ F P (xA ∨
xBxA) = τ F P (xA∨xBxB) = 0. It follows that both first
moments K (1)(xAx0) and K (1)(xBx0) must individually
vanish at the absorbing boundaries, i.e. K (1)(xAxA) =
K (1)(xAxB) = 0 and K (1)(xBxA) = K (1)(xBxB) = 0.
With these boundary conditions we obtain
K (1)(xAx0) = CφB(x0)
K (1)(xBx0) = CφA(x0)
(cid:90) xB
x0
(cid:90) x0
xA
dx e−F (x)φ2
A(x) + CφA(x0)
dx e−F (x)φ2
B(x) + CφB(x0)
(cid:90) x0
xA
(cid:90) xB
x0
dx e−F (x)φA(x)φB(x),
dx e−F (x)φA(x)φB(x).
(26)
(27)
and
From Eq. (19) the mean first-passage time to reach boundary A when starting from x0 reads
dx e−F (x)φA(x)φB(x),
τ F P (xAx0) = C
dx e−F (x)φ2
A(x) + C
φB(x0)
φA(x0)
while the mean first-passage time to reach boundary B when starting from x0 reads
τ F P (xBx0) = C
φA(x0)
φB(x0)
dx e−F (x)φ2
B(x) + C
dx e−F (x)φA(x)φB(x).
(cid:90) x0
(cid:90) xB
xA
x0
(cid:90) xB
(cid:90) x0
x0
xA
4
(28)
(29)
As we will show explicitly below, because of reversibility,
the mean first-passage time τ F P (xAx0) in fact equals the
mean-time a transition path that starts at the boundary
xA and ends at the boundary xB needs in order to reach
the intermediate position x0, it thus determines the mean
shape of the transition path,
The same result is obtained from Eq. (28) by the limiting
procedure τ T P (xAxB) = τ F P (xAx0 → xB), reflecting
that transition paths are reversible, i.e. τ T P (xBxA) =
τ T P (xAxB).
shape(x0xA) = τ F P (xAx0),
τ T P
(30)
B.
Forward Fokker-Planck approach
parameterized in terms of the mean time as a function
of the position. Likewise, the mean first-passage time
τ F P (xBx0) corresponds to the mean time a transition
path that starts at boundary xB and ends at boundary
xA needs in order to reach the intermediate position x0,
shape(x0xB) = τ F P (xBx0).
τ T P
(31)
Note that the paths that contribute to the shape
shape(x0xA/B) revisit the position x0 multiple times,
τ T P
as will be illustrated later on when we present explicit
Brownian dynamics paths.
4. Transition path times
The transition path time denotes the mean time a path
takes to reach from the absorbing boundary xA to the
other absorbing boundary at xB. It is thus defined by
τ T P (xBxA) = τ F P (xBx0 → xA).
(32)
In the limit x0 → xA the first term in Eq. (29) vanishes
and we obtain in agreement with Szabo's result [4]
τ T P (xBxA) = C
dx e−F (x)φA(x)φB(x).
(33)
(cid:90) xB
xA
(cid:90) ∞
It is instructive to describe transition paths also us-
ing the forward FP equation [4] as this allows to define
passage and residence times and to derive various use-
ful relations between transition path times, first-passage
times, and passage times.
Defining moments of the Green's function as
G(n)(xx0) =
dt tnG(x, tx0),
(34)
0
we obtain from the forward FP Eq. (4) for n > 0 the
recursive relations
− nG(n−1)(xx0) = L(x)G(n)(xx0).
(35)
For n = 0 we obtain
− δ(x − x0) = L(x)G(0)(xx0).
(36)
We again impose absorbing boundary conditions at xA
and xB, i.e. G(xA, tx0) = G(xB, tx0) = 0, which means
that all moments satisfy G(n)(xAx0) = G(n)(xBx0) = 0.
Equations (35) and (36) are solved straightforwardly by
integration, yielding
G(0)(xx0) = Ce−F (x) {φA(x0)φB(x) − θ(x − x0)[φA(x0) − φA(x)]} ,
and
G(1)(xx0) = Ce−F (x)
(cid:26)
(cid:90) x
φA(x)
dx(cid:48) G(0)(x(cid:48)x0)φB(x(cid:48)) + φB(x)
dx(cid:48) G(0)(x(cid:48)x0)φA(x(cid:48))
xA
x
where θ(x− x0) denotes the Heavyside function with the
properties θ(x − x0) = 1 for x > x0 and zero otherwise.
(cid:90) xB
(cid:27)
(37)
,
(38)
Note that we assume the start and end positions x0 and
x of the paths to be inside the absorbing boundary con-
ditions, i.e., xA < x < xB and xA < x0 < xB. The mean
time to reach the position x when starting out from po-
sition x0 follows from proper normalization as
τ P (xx0) =
G(1)(xx0)
G(0)(xx0)
,
(39)
we call this time the mean passage time and it is always
larger than the mean first-passage time unless the target
position is an absorbing boundary. The mean passage
time is the mean time to reach the target at position x,
while allowing for multiple recrossing events. We obtain
for x0 < x the result
τ P (xx0) = C
dx(cid:48) e−F (x(cid:48))φ2
(cid:90) x0
B(x(cid:48))
dx(cid:48) e−F (x(cid:48))φA(x(cid:48))φB(x(cid:48))
dx(cid:48) e−F (x(cid:48))φ2
A(x(cid:48)),
φA(x0)
φB(x0)
(cid:90) x
+C
+C
x0
φB(x)
φA(x)
xA
(cid:90) xB
x
(cid:90) x
(cid:90) xB
xA
x0
φA(x)
φB(x)
(cid:90) x0
+C
+C
x
φB(x0)
φA(x0)
(40)
(41)
while for x < x0 we obtain
τ P (xx0) = C
dx(cid:48) e−F (x(cid:48))φ2
B(x(cid:48))
dx(cid:48) e−F (x(cid:48))φA(x(cid:48))φB(x(cid:48))
dx(cid:48) e−F (x(cid:48))φ2
A(x(cid:48)).
Obviously, the two expressions are connected by the sym-
metry τ P (xx0) = τ P (x0x) that reflects the reversibility
of the underlying processes described by the FP equation.
We note that this symmetry also holds when x0 and/or
x are located on the absorbing boundaries xA and xB
. This symmetry also holds when we shift the absorb-
for xA → −∞
ing boundary conditions to infinity, i.e.
and/or xB → ∞, that is in the absence of absorbing
boundary conditions.
The mean first-passage times in Eqs.
(28) and (29)
follow from the passage times by the limiting procedures
τ F P (xAx0) = τ P (x → xAx0),
and
τ F P (xBx0) = τ P (x → xBx0).
(42)
(43)
The expression
τ P (x0x0) = C
φA(x0)
φB(x0)
dx(cid:48) e−F (x(cid:48))φ2
B(x(cid:48))
(cid:90) x0
(cid:90) xB
xA
+C
φB(x0)
φA(x0)
x0
dx(cid:48) e−F (x(cid:48))φ2
A(x(cid:48)),
(44)
5
measures the mean time a path stays at the starting po-
sition x0, we call this time the residence time. By explicit
consideration of the results in Eqs. (28), (29), (33), (44)
it turns out that the transition path time τ T P (xBxA) in
Eq. (33) is related to the first-passage times of reaching
the absorbing boundaries at xA and xB from an interme-
diate position x0 by subtracting the residence time,
τ T P (xBxA) = τ F P (xAx0) + τ F P (xBx0) − τ P (x0x0).
(45)
This shows that a transition path time can be constructed
by adding the mean first-passage times of two paths start-
ing at an arbitrary position x0 that reach the boundaries
xA and xB. Since each path recrosses the starting posi-
tion, the residence time τ P (x0x0) has to be subtracted
in order not to overcount these recrossing events. By
a tedious but straightforward calculation one can show
that
τ F P (xBx0) − τ P (x0x0) = τ T P (xBx0) = τ T P (x0xB),
(46)
holds for the transition path time of going from x0 to xB
or from xB to x0. Combining this with Eq. (45) we thus
find
τ T P (xAxB) = τ T P (xAx0) + τ T P (xBx0) + τ P (x0x0)
(47)
= τ F P (xAx0) + τ T P (x0xB).
Equation (47) demonstrates that the transition path time
from xA to xB can be decomposed into the first-passage
time starting from an intermediate position x0 and the
transition path time continuing to the other boundary.
Together with our definition for the shape of a transition
path in Eq. (30), we conclude
shape(x0xA) = τ T P (xBxA) − τ T P (xBx0)
τ T P
= τ T P (x0xA) + τ P (x0x0),
(48)
i.e., the mean shape of a transition path from xA to x0
is the transition path time from xA to xB minus the
transition path time from x0 to xB, or, alternatively, the
transition path from xA to x0 plus the residence time at
x0.
Finally, and as mentioned before, the symmetry of
mean passage times τ P (xx0) = τ P (x0x) also holds when
we move the point x onto the absorbing boundary xA,
this turns the mean passage time τ P (xAx0) into the
mean first-passage time and we obtain τ F P (xAx0) =
τ P (x0xA). Combining this with the definition Eq. (30)
we find
shape(x0xA) = τ F P (xAx0) = τ P (x0xA),
τ T P
(49)
and we see that the transition path shape corresponds
to the mean passage time of paths that start from the
absorbing boundary xA. Note that the formulas Eqs.
(45)-(49) have been explicitly derived in the presence of
absorbing boundaries at positions xA and xB, we will
show in the next section that similar relation can be de-
rived from integral equations for the distribution of pas-
sage times.
C. Renewal equation approach
or
6
1. First-passage time distribution
Explicit expressions for the transition path time can
also be derived within the renewal equation approach
without referral to an explicit underlying diffusive model.
The relations derived in this section are thus more general
than the previous derivations which were based on the
one-dimensional FP equation. Also, the present deriva-
tion allows to understand more deeply in what sense the
first-passage time τ F P (xAx0) in the presence of an ab-
sorbing boundary at xB can be interpreted as the shape
shape(x0xA).
of a transition path starting from xA, τ T P
In this section we do not impose absorbing boundaries
unless explicitly mentioned. Although we use a one-
dimensional reaction coordinate, our results can be read-
ily generalized to higher dimensions.
We start with the renewal equation [24, 25]
0
G(x, tx0) = Gx(cid:48)(x, tx0)+
dt(cid:48) G(x, t−t(cid:48)x(cid:48))K(x(cid:48), t(cid:48)x0),
(50)
which can be viewed as a general definition of the
first-passage time distribution K(x(cid:48), tx0) and where
Gx(cid:48)(x, tx0) denotes the Green's function in the presence
of an absorbing boundary condition at x(cid:48). It is an alter-
native more general definition than the one presented in
Eq. (12). The renewal equation states that the ensemble
of all paths starting at time zero at x0 and that are at
position x at time t can be decomposed into paths that
never reach the absorbing boundary condition at x(cid:48) and
paths that hit the boundary x(cid:48) for the first time at time
t(cid:48) and from there on diffuse freely to x. By letting the
position of the absorbing boundary x(cid:48) coincide with x we
obtain the special case
(cid:90) t
G(x, tx0) =
dt(cid:48) G(x, t − t(cid:48)x)K(x, t(cid:48)x0).
0
(cid:82) ∞
0 dtG(x, tx0)e−ωt Eq. (51) becomes
In terms of
the Laplace transform G(x, ωx0) =
(51)
G(x, ωx0) = G(x, ωx) K(x, ωx0).
(52)
(cid:90) ∞
Using that moments can be calculated from the Laplace
transform by
G(n)(xx0) ≡
dt tnG(x, tx0) = (−∂ω)n G(x, ωx0)ω=0,
(53)
0
the normalized first moments are related by
(cid:90) t
τ F P (xx0) = τ P (xx0) − τ P (xx).
(55)
In other words, the mean first-passage time τ F P (xx0) of
going from x0 to x in the absence of any additional ab-
sorbing or reflecting boundaries can be constructed from
the mean passage time τ P (xx0) of going from x0 to x by
subtracting the residence time τ P (xx) of staying at x.
By symmetry of the passage time (derived in the previous
section) we can write
τ F P (xx0) = τ P (x0x) − τ P (xx).
(56)
This relation holds also in the presence of an absorbing
boundary condition at x0 (note that an absorbing bound-
ary condition can be simply imposed by creating a po-
tential well of infinite depth in the region x < x0, which
turns x0 into an absorbing boundary for all paths that
come from x > x0). This turns τ F P (xx0) into the tran-
sition path time τ T P (xx0), the passage time τ P (x0x)
into the first-passage time τ F P (x0x), and the residence
time τ P (xx) without specified boundary conditions into
the residence time at x in the presence of an absorbing
(xx). We thus
boundary at x0, which we denote by τ P
x0
obtain from Eq. (56)
τ T P (xx0) = τ T P (x0x) = τ F P (x0x) − τ P
(xx),
(57)
x0
which is equivalent to Eq. (46) (note that Eq. (46) by
way of derivation holds in the presence of two absorb-
ing boundary conditions at xA and xB, so to make the
equivalence perfect we can either shift the boundary xA
in Eq. (46) to infinity or impose an additional absorbing
boundary condition in Eq. (57)).
In order to derive Eq. (47) we need a convolution equa-
tion for first-passage times. For this we choose in the re-
newal equation (50) the absorbing boundary condition x(cid:48)
at a relative position x0 < x(cid:48) < x and in this case obtain
G(x, tx0) =
dt(cid:48) G(x, t − t(cid:48)x(cid:48))K(x(cid:48), t(cid:48)x0).
(58)
We now impose an absorbing boundary condition at x,
which turns both Green's functions into first-passage
time distributions so that we obtain
K(x, tx0) =
dt(cid:48) K(x, t − t(cid:48)x(cid:48))K(x(cid:48), t(cid:48)x0),
(59)
0
valid for arbitrary positions x(cid:48) with x0 < x(cid:48) < x. By
using Laplace transformation, similarly as the calculation
leading to Eq. (55), this yields
τ F P (xx0) = τ F P (x(cid:48)x0) + τ F P (xx(cid:48)).
(60)
(cid:90) t
0
(cid:90) t
− ∂ω ln K(x, ωx0)ω=0 =
=
K (1)(xx0)
K (0)(xx0)
G(1)(xx0)
G(0)(xx0)
− G(1)(xx)
G(0)(xx)
,
(54)
Imposing an additional absorbing boundary condition at
x0 turns this into
τ T P (xx0) = τ T P (x(cid:48)x0) + τ F P
x0
(xx(cid:48)),
(61)
where the subindex x0 in the last term indicates that an
absorbing boundary is present at x0. This is identical
to Eq. (47), remembering that Eq. (47) was derived in
the presence of an absorbing boundary at xB. We next
combine Eqs. (57) and (61) and obtain
τ T P (xx0) = τ F P
(xx(cid:48)) + τ F P (x0x(cid:48)) − τ P
(x(cid:48)x(cid:48)), (62)
x0
x0
which is equivalent Eq. (45) if we impose an additional
absorbing boundary condition at x.
2. Transition path time distribution
We now impose an absorbing boundary condition at
position x0 in the convolution equation (59), this turns
the two first-passage time distributions starting at x0 into
transition path time distributions and we obtain
(cid:90) t
0
boundaries at x0 and x, and finally a transition path from
x(cid:48) to the final destination x.
7
We now use the renewal equation (51) and impose an
absorbing boundary condition at x0 and replace the vari-
able x by x(cid:48) to yield
H(x(cid:48), tx0) =
dt(cid:48) Gx0 (x(cid:48), t − t(cid:48)x(cid:48))T (x(cid:48), t(cid:48)x0),
(68)
which is an explicit integral equation for the last-passage
time distribution. We now impose an additional absorb-
ing boundary condition at x with the ordering x0 < x(cid:48) <
x and obtain
Hx(x(cid:48), tx0) =
dt(cid:48) Gx0,x(x(cid:48), t − t(cid:48)x(cid:48))T (x(cid:48), t(cid:48)x0). (69)
By comparison with Eq. (67) we obtain
(cid:90) t
0
(cid:90) t
(cid:90) t
0
0
T (x, tx0) =
dt(cid:48) Kx0(x, t − t(cid:48)x(cid:48))T (x(cid:48), t(cid:48)x0),
(63)
T (x, tx0) =
dt(cid:48) T (x, t − t(cid:48)x(cid:48))Hx(x(cid:48), t(cid:48)x0).
(70)
(cid:90) t
where Kx0(x, t − t(cid:48)x(cid:48)) is the first-passage time distribu-
tion with an additional absorbing boundary condition at
x0 with x0 < x(cid:48) < x. Note that Eq. (61) follows directly
from this integral equation via Laplace transformation.
It means that a transition path can be decomposed into
a transition path to an intermediate position x(cid:48) followed
by a first-passage path from x(cid:48) that does not revisit x0.
To go on with our derivation we define the last-passage
distribution via the integral equation
G(x, tx0) = Gx(cid:48)(x, tx0)+
dt(cid:48) H(x, t−t(cid:48)x(cid:48))G(x(cid:48), t(cid:48)x0).
(64)
In essence, the last-passage distribution H(x, t(cid:48)x(cid:48)) com-
prises all paths that go from x(cid:48) to x without revisiting
the starting point at x(cid:48). By moving the starting position
x0 to the absorbing boundary at x(cid:48) we obtain
0
(cid:90) t
G(x, tx(cid:48)) =
dt(cid:48) H(x, t − t(cid:48)x(cid:48))G(x(cid:48), t(cid:48)x(cid:48)).
(65)
0
We now impose two absorbing boundary conditions, one
at x and the other at x0 with the condition x0 < x(cid:48) < x,
and obtain
Kx0(x, tx(cid:48)) =
dt(cid:48) T (x, t − t(cid:48)x(cid:48))Gx0,x(x(cid:48), t(cid:48)x(cid:48)). (66)
(cid:90) t
0
By inserting this integral equation into Eq. (63) we ob-
tain
T (x, tx0) =(cid:82) t
0 dt(cid:48)(cid:82) t−t(cid:48)
0
dt(cid:48)(cid:48) T (x, t − t(cid:48) − t(cid:48)(cid:48)x(cid:48))
Gx0,x(x(cid:48), t(cid:48)(cid:48)x(cid:48))T (x(cid:48), t(cid:48)x0),
(67)
which has a nice intuitive interpretation: a transition
path from x0 to x can be decomposed into a transition
path from x0 to an arbitrary mid-point position x(cid:48), a path
that starts from x(cid:48) and returns to x(cid:48) without reaching the
Also this expression, from which we will derive the tran-
sition path shape, has an intuitive interpretation: a tran-
sition path from x0 to x can be decomposed into a last-
passage path from x0 to an arbitrary mid-point position
x(cid:48) followed by a transition path from x(cid:48) to the final des-
tination x. Note that the last-passage paths from x0 to
x(cid:48) do not visit the absorbing boundary condition x which
is indicated by the subscript.
By construction, the integrand in Eq. (70) is the joint
probability that a transition path starting from x0 and
ending at x has a duration of t and is at time t(cid:48) at the
position x(cid:48). This is so because paths for times later than
t(cid:48) proceed on transition paths to x and do not visit back
to x(cid:48) and therefore do not contribute to the probability of
being at x(cid:48). The average shape of a transition path thus is
obtained by averaging theT (x, t− t(cid:48)x(cid:48))Hx(x(cid:48), t(cid:48)x0) both
over the intermediate time t(cid:48) and transition path dura-
tion t. We thus obtain for the shape of a transition path
from xA to xB
0 dt(cid:48) t(cid:48)T (xB, t − t(cid:48)x)HxB (x, t(cid:48)xA)
0 dt(cid:48) T (xB, t − t(cid:48)x)HxB (x, t(cid:48)xA)
.
(71)
(cid:82) ∞
0 dt(cid:82) t
(cid:82) ∞
0 dt(cid:82) t
(cid:82) ∞
(cid:82) ∞
0 dt(cid:48) t(cid:48)HxB (x, t(cid:48)xA)
0 dt(cid:48) HxB (x, t(cid:48)xA)
xA,xB
= τ P
shape(xxA) =
τ T P
(xxA),
(72)
and thus have derived the important result that the shape
of a transition path is given by the passage time from an
absorbing boundary at xA to a midpoint x in the pres-
ence of a second absorbing boundary at xB, as presented
in Eq. (49). We remind the reader of the relation Eq.
(49) which shows that because of the symmetry of pas-
sage times, instead of averaging over paths that come
from the absorbing boundary xA, one can equally well
average over first-passage paths that start from x and
By slightly rearranging we obtain
shape(xxA) =
τ T P
that end at the boundary xA, the latter ensemble is for
simulations much more easy to implement and we will
explicitly demonstrate the equivalence of both ensembles
in our simulations.
III. THE SHAPE OF KRAMERS'
FIRST-PASSAGE PATHS
Here we consider the mean shape of the Kramers' first-
passage paths defined as paths that start from a reflecting
boundary and reach an absorbing boundary. We basi-
cally repeat the derivation steps from the previous sec-
tion but replace the absorbing boundary condition at xA
by a reflecting one. If we impose a reflecting boundary
at position xA in the convolution relation for the first-
passage distribution Eq. (59) we obtain
KxA(xB, txA) =
dt(cid:48) KxA(xB, t − t(cid:48)x)KxA (x, t(cid:48)xA),
(73)
where we denote a reflecting boundary condition by a
subscript with a tilde and an adsorbing boundary condi-
tion by a subscript without a tilde.
(cid:90) t
0
(cid:90) t
We next impose an absorbing boundary condition at
xB and a reflecting boundary condition at xA in the in-
tegral relation for the last-passage distribution Eq. (65)
and obtain
KxA(xB, tx(cid:48)) =
dt(cid:48) T (xB, t − t(cid:48)x)GxA,xB (x, t(cid:48)x).
(74)
By inserting this integral equation into Eq. (73) we ob-
tain
KxA (xB, txA) =(cid:82) t
0 dt(cid:48)(cid:82) t−t(cid:48)
dt(cid:48)(cid:48) T (xB, t − t(cid:48) − t(cid:48)(cid:48)x)
0
0
GxA,xB (x, t(cid:48)(cid:48)x)KxA(x, t(cid:48)xA),
(75)
which has a similar interpretation as the corresponding
result for an absorbing boundary condition at the origin
in Eq. (67): a Kramers' first-passage path from xA to
xB can be decomposed into a first-passage path from xA
to an arbitrary mid-point position x, a path that starts
from x and returns to x without reaching the absorb-
ing boundary at xB and without crossing the reflecting
boundary at xA, and finally a transition path from x to
the final destination xB.
We next impose an absorbing boundary condition at
xB and a reflecting boundary condition at x0 = xA on
the definition of the first-passage distribution Eq. (51),
from which we obtain
GxA,xB (x, txA) =
dt(cid:48) GxA,xB (x, t− t(cid:48)x)KxA(x, t(cid:48)xA).
(76)
Comparison with Eq. (75) gives the integral equation
dt(cid:48) T (xB, t − t(cid:48)x)GxA,xB (x, txA).
KxA (xB, txA) =
(77)
0
(cid:90) t
(cid:90) t
0
8
We now use similar arguments leading to our expression
for the transition path shape in Eq. (72): The integrand
in Eq. (77) is the joint probability that a first-passage
path starting from xA and ending at xB has a duration
of t and is at time t(cid:48) at position x. The average shape of
a first-passage path is obtained by averaging over both
intermediate time t(cid:48) and the first-passage path duration
t, we thus obtain for the mean shape of a Kramers' first-
passage path from xA to xB
shape(xxA) =
τ KF P
(xxA).
(78)
The only difference to the result for the transition path
shape Eq. (72) is that the absorbing boundary condition
at xA is replaced by a reflecting boundary condition.
(cid:82) ∞
(cid:82) ∞
0 dt(cid:48) t(cid:48)GxA,xB (x, t(cid:48)xA)
0 dt(cid:48) GxA,xB (x, t(cid:48)xA)
= τ P
xA,xB
By Laplace transformation of Eq. (77) we obtain (sim-
ilarly as when we derived Eq. (55) from Eq. (51))
(xxA),
τ KF P (xBxA) = τ T P (xBx) + τ P
xA,xB
(79)
where we defined the Kramers' mean first-passage time
shape(xBxA) and which is explicitly
as τ KF P (xBxA) = τ KF P
(cid:90) xB
given by [26]
(cid:90) x
dx(cid:48)e−F (x(cid:48)).
(80)
τ KF P (xBxA) =
dx
eF (x)
D(x)
xA
xA
By combining Eq. (48), Eq. (78) and Eq. (79) we find
shape(xxA) = τ T P
τ KF P
shape(xxA) + τ P
xA,xB
(xAxA),
(81)
showing that the mean shape of Kramers' first-passage
shape(xxA) and the mean shape of transition
paths τ KF P
shape(xxA) are identical and shifted by a con-
paths τ T P
(xAxA). This shift corresponds
stant given by τ P
to the passage time at the reflecting boundary xA and
(xAxA) =
is according to Eq.
τ KF P (xBxA) − τ T P (xBx).
(79) given by τ P
xA,xB
xA,xB
IV. RESULTS FOR EXPLICIT POTENTIALS
We next present exemplaric transition path shapes for
a few different simple potential shapes shown in Fig. 1-
(a)-(c). We consider a reaction coordinate x in the range
of 0 ≤ x ≤ L, where L is the transition length scale, and
restrict ourselves from now on to a homogeneous diffusion
constant D
A. Brownian dynamics simulations and trajectory
analysis
We also present trajectories obtained from one di-
mensional overdamped Brownian dynamics (BD) simula-
tions. The simulations are based on the Langevin equa-
tion
dx(t)
dt
= −D
dF (x)
dx
+
ζ(t)
γ
,
(82)
9
FIG. 1: Illustrations of the used rescaled potentials F (x)/U as a function of the rescaled length x/L, where L is the transition
length scale and U is the barrier height: (a) linear potential F (x) = U x/L, (b) full harmonic potential F (x) = 4U (1− x/L)x/L
and (c) harmonic ramp F (x) = U (2− x/L)x/L. (d) A typical transition path trajectory xT P (t) for the force-free case, obtained
(x00) when the transition path crosses the position x0 are indicated
from Brownian dynamics (BD) simulations. The times tT P
(e) A typical Kramers' first-passage path trajectory xKF P (t) for the force-free case, obtained from BD
by vertical lines.
simulations, in the presence of a reflecting boundary condition at x = 0 and an absorbing boundary condition at x = L. The
transition path is the last part of the trajectory indicated by the gray region.
i
where γ = kBT /D is the friction constant and ζ(t) is
a Gaussian random force which fulfills (cid:104)ζ(t)(cid:105) = 0 and
(cid:104)ζ(t)ζ(t(cid:48))(cid:105) = 2γkBT δ(t−t(cid:48)). The discretized and rescaled
Langevin equation reads
2dt r(t),
(83)
(cid:112)
x(t + dt) = x(t) − dF
dx
dt +
with zero mean and unit standard deviation. We iterate
Eq. (83) with a typical time step dt = 10−4.
To obtain mean first-passage times τ F P (0x0) and
τ F P (Lx0) we vary the initial position from x0 = 0 to
x0 = L and measure the time needed to reach one of the
two absorbing boundaries xA = 0 or xB = L for the first
time, we typically average over 105 first-passage times.
where x = x/L is the rescaled position, t = tD/L2 is
the rescaled time, and r(t) is a Gaussian random number
We also generate transition path trajectories within
BD simulations. In practice we initiate a trajectory at a
00.5100.51xLFxUa00.5100.51xLFxUb00.5100.51xLFxUc0.000.020.040.060.080.100.120.140.00.20.40.60.81.0tDL2xtLdtiTPx00x00.00.10.20.30.40.50.60.70.00.20.40.60.81.0tDL2xtLeTransitionpath reflecting boundary at x = 0 and record until it reaches
the absorbing boundary at x = L, the transition path
trajectory is the last portion of the trajectory after it
has last returned to the reflecting boundary at x = 0, as
shown in Fig. 1-(e). The mean transition path shape is
obtained by averaging the time transition paths take to
reach a certain position x0
shape(x00) =
τ T P
tT P
i
(x00)
N
,
(84)
N(cid:88)
i=1
i
(x00) denotes the time at which a transition
where tT P
path trajectory that starts out at x = 0 crosses the po-
sition x0, as illustrated in Fig. 1-(d). Note that a single
transition path crosses the position x0 multiple times,
the averaging in Eq. (84) is done over the entire transi-
tion path ensemble and over all crossing events, N thus
counts the total number of crossing events in the entire
transition path ensemble. For our final results we typi-
cally generate 104 transition paths.
In a similar manner, we analyze Kramers' first-passage
trajectories, which start from a reflecting boundary at
x = 0 and eventually reach the absorbing boundary at
x = L, an example of which is shown in Fig. 1-(e). To
obtain the mean shape of Kramers' first-passage trajecto-
shape(x00), we average the mean time
ries, denoted by τ KF P
it takes such a path to reach a certain position x0
shape(x00) =
τ KF P
tKF P
i
(x00)
N
,
(85)
N(cid:88)
i=1
(x00) denotes the time at which a path that
where tKF P
starts from x = 0 crosses x = x0.
i
B. Force-free case
We first consider the force-free case, F = 0. The split-
ting probabilities read φA(x) = 1−x/L and φB(x) = x/L
with C = 1, and the transition path time according to
Eq. (33) reads
10
FIG. 2:
(a) The normalized distribution functions for the
transition path time τ T P (L0) (circles) and the Kramers' first-
passage time τ KF P (L0) (squares) in the force-free case, ob-
tained from BD simulations. (b) Three typical transition path
trajectories x(t). (c) Three typical Kramers' first-passage tra-
jectories x(t).
time distribution is more sharply peaked compared with
the Kramers' first-passage time distribution. The trajec-
tories shown in Fig. 2-(b) and (c) reflect this difference
of the two distributions.
The mean transition path shape is, according to Eqs.
(28) and (30), given as
(cid:16)
(cid:17)
τ T P (L0) =
L2
6D
,
(86)
shape(x00) = τ F P (0x0) =
τ T P
Lx0
6D
2 − x0
L
,
(88)
which is three times smaller than Kramers' mean first-
passage time
τ KF P (L0) =
L2
2D
,
(87)
according to Eq. (80). This decrease is due to the sub-
traction of the part of the Kramers' first-passage tra-
jectories that contains multiple returns to the origin, as
illustrated in Fig. 1-(e).
The normalized distribution functions for the transi-
tion path time τ T P (L0) (circles) and the Kramers' first-
passage time τ KF P (L0) (squares) are shown in Fig. 2-
(a), obtained from BD simulations. The transition path
and is depicted in Fig. 3-(a) by a solid line. Note that the
transition path shape is a quadratic function, transition
paths start out with finite velocity at the origin and reach
the final destination with infinite velocity. This asym-
metry, which is a universal property of mean transition
path shapes for all potentials, can be easily understood
by considering Eq. (48) and realizing that a mean transi-
tion path time scales quadratic with the diffusion length
scale in the limit of small diffusion length scale. The filled
symbols in Fig. 3-(a) show the BD simulation results for
the first-passage time τ F P (0x0) while the open square
shape(x00) obtained
symbols show the BD results for τ T P
via Eq. (84), both simulation results agree well with the
(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:230)ΤTPL0(cid:224)ΤKFPL00.00.20.40.60.81.00123456Pt0.00.20.40.60.81.00.00.20.40.60.81.0xtL0.00.20.40.60.81.00.00.20.40.60.81.0xtLabctDL2 11
The solid curve in Fig. 3-(b) shows the Kramers' mean
shape(x00), calculated from Eqs.
first-passage shape τ KF P
(81) and (88). The Kramers' mean first-passage shape
shape(x00) is, according to Eqs.
τ KF P
(79) and (81), iden-
tical to the transition path shape τ T P (x00) shifted by
xA=0,xB =L(x) = τ KF P (L0) − τ T P (L0) =
the amount τ P
L2/(3D). The symbols in Fig. 3-(b) show the BD results
using Eq. (85), again, the agreement is very good.
C. Transition path in linear potential
For a linear potential F = U x/L we find for the tran-
sition path time
τ T P (L0) =
L2
D
U coth(cid:0) U
(cid:1) − 2
2
U 2
,
(89)
which is an even function of U . This means that the
transition path time is the same irrespective of whether
the transition paths go up the linear potential or whether
they go down. This of course follows directly from the
general symmetry of passage times in Eqs. (40) and (41)
but is worthwhile pointing out again at this point. To
leading order in U the asymptotic behavior reads
(cid:40) 1
6 − U 2
1/U
360
τ T P D/L2 ≈
, U (cid:28) 1
, U (cid:29) 1.
(90)
The red solid curve in Fig. 4-(a) shows τ T P (L0) in Eq.
(89) while the asymptotic expressions in Eq.
(90) are
depicted by broken curves. Note that the Kramers' mean
first-passage time τ KF P (L0) = L2(eU − 1 − U )/(DU 2),
shown by a solid blue curve in Fig. 4-(a), shows very
different behavior and in particular is a monotonically
increasing function of U . For large potential strength
U (cid:29) 1 we find an exponential increase to leading order,
τ KF P (L0) ∼ eU /U 2. The symbols in Fig. 4-(a) show
BD simulation results for the transition path time, which
agree well with the theory.
A further noteworthy fact is that the mean transi-
tion path time τ T P is for non-zero values of U strictly
smaller than the force-free result τ T P = L2/(6D) cor-
responding to the maximum value obtained for U = 0.
This means that transition paths in a linear potential
are faster than force-free transition paths, regardless of
whether the slope is positive or negative.
(a) Mean shape of transition paths τ T P
shape(x00) in
FIG. 3:
the force-free case. The solid line shows the analytic result
Eq. (88). Filled circles show BD simulation results for the
mean first-passage time τ F P (0x0) while open squares show
the mean shape from the analysis of transition paths accord-
ing to Eq. (84). (b) Mean shape of Kramers' first-passage
shape(x00) in the force-free case. Symbols show BD
paths τ KF P
simulation results while the solid line shows analytic results
according to Eqs. (81) and (88). Note that the two curves in
(a) and (b) are identical except a vertical shift by a constant
time.
theoretical result Eq. (88).
The transition path shapes read
csch(cid:0) U
2
(cid:1) csch(cid:0) U
2 − U x0
2L
(cid:1)(cid:2)(x0/L − 2) sinh(cid:0) U x0
2L
(cid:1) + x0
L sinh(cid:0)U − U x0
2L
(cid:1)(cid:3)
coth(U/2) − x0
L coth( U x0
2L )
U
,
2U
shape(x00) =
τ T P
shape(x0L) =
τ T P
L2
D
L2
D
,
(91)
(92)
(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:230)ΤFP0x0(cid:225)ΤshapeTPx000.00.20.40.60.81.00.000.050.100.150.20ΤshapeTPx00DL2(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)ΤshapeKFPx000.00.20.40.60.81.00.350.400.450.50x0LΤshapeKFPx00DL2ab 12
FIG. 4: Results for a linear potential F = U x/L. (a) The solid red curve shows the mean transition path time τ T P (0L) from
Eq. (89) as a function of U on a log-linear scale. The broken curves depict the asymptotic expressions from Eq. (90). For
comparison, the solid blue curve shows Kramers' mean first-passage time τ KF P (L0) which monotonically increases with U .
shape(x00) starting
The symbols denote BD simulation results. (b) Mean shapes of transition paths τ T P
shape(x0L) starting from the right boundary from Eq. (92).
from the left boundary from Eq. (91), while black curves depict τ T P
Symbols denote BD simulation results for U = −5 and U = −10 while broken red curves depict the asymptotic expressions for
U = ±5 from Eq. (93).
shape. Blue curves depict τ T P
where τ T P
shape(x00) has the asymptotic limits
shape(x00)D/L2 ≈
τ T P
(cid:40) U−sinh U
U (1−cosh U )
τ T P (L0)D/L2 − 1
x0
L
L − 1)2
6 ( x0
, x0 (cid:28) L
, x0 ≈ L.
(93)
Figure 4-(b) shows the transition path shapes τ T P
shape as
function of the position x0, where the blue curves depict
shape(x00) in Eq.
τ T P
(91), and the black curves depict
shape(x0L) in Eq. (92). Symbols denote BD simulation
τ T P
results for U = −5 and U = −10. The broken red curves
(93) for U = ±5.
depict the asymptotic limits in Eq.
Due to the symmetry of passage times, the shapes τ T P
shape
are symmetric with respect to an exchange of starting
positions.
D. Harmonic potential
For a harmonic potential F = 4U x(1− x/L)/L we find
for the transition path time
τ T P (L0) =
F2,2(−U )
L2
4D
−
√
L2
√
πU erf(
2D
(cid:90) √
U
U )
0
dy y2e−y2
F2,2(−y2),
(94)
where F2,2(x) = F2,2({1, 1};{3/2, 2}; x) is the general-
ized hypergeometric function. For the small barrier limit
U (cid:28) 1 we find to leading order
τ T P (L0) ≈ L2
D
− 2
45
U
,
(95)
(cid:21)
(cid:20) 1
6
which decreases from the force-free transition path time
τ T P = L2/(6D). For the large barrier limit U → ∞ we
recover the known asymptotic result [5, 7]
τ T P (L0) ≈ L2 ln(2eγU )
8DU
,
(96)
√
0 dy y2e−y2
where γ ≈ 0.577 is the Euler gamma constant, and
U ) ≈ 1, F2,2(−U ) ≈ ln(4eγU )/(2U )
we used erf(
π/4) ln(2) for large
U . We note that the denominator 8U in Eq.
(96)
can be reinterpreted as the rescaled curvature ω2 =
(cid:12)(cid:12)(cid:12) at the barrier top of the harmonic
and (cid:82) ∞
L2(cid:12)(cid:12)(cid:12)(cid:0)d2F/dx2(cid:1)
F2,2(−y2) = (
x=L/2
√
potential, yielding the previously published form [7]
τ T P (L0) ≈ L2 ln(2eγU )
Dω2
.
(97)
(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:45)30(cid:45)20(cid:45)1001020300.050.100.200.501.00UΤTP0LDL2(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)0.00.20.40.60.81.00.000.050.100.150.200.000.050.100.150.20x0LΤshapeTPx0LDL2ΤshapeTPx00DL2U(cid:61)0U(cid:61)(cid:177)5U(cid:61)(cid:177)10U(cid:61)0U(cid:61)(cid:177)5U(cid:61)(cid:177)10ab For fixed potential curvature and varying potential
height, Eq. (97) shows that the transition path time in-
creases logarithmically with increasing potential height
U , while for fixed diffusion L, Eq. (96) shows that the
transition path time decreases inversely linearly with in-
creasing potential height U [5].
In Fig. 5 we present τ T P (L0) as a function of the
barrier height U . In Fig. 5-(a) we show τ T P (L0)D/L2
from Eq. (94) on a log-log scale (solid red curve), which is
seen to decrease from the force-free case τ T P D/L2 = 1/6
as U increases. We also show the asymptotic expressions
Eqs. (95) and (96) by dashed curves. In Fig. 5-(b) we
show τ T P (L0) from Eq. (94) on a log-linear scale (solid
red curve), here we also compare with BD simulation
results obtained via Eq. (84). The solid blue curves in
Fig. 5 depict the Kramers' mean first-passage time given
by
(cid:16)√
(cid:17)
(cid:16)√
(cid:17)
τ KF P (L0) =
U
erfi
U
πerf
L2
(cid:82) x
D
0 e−t2
dt is the imaginary error function.
dt is the error function, and
(98)
8U
,
(cid:82) x
where erf (x) = 2√
π
erfi (x) = 2√
0 et2
π
The leading order result for large U reads
√
√
τ KF P (L0)D/L2 =
eU
U3/2
π
8
=
π
ω2
eU(cid:112)U . (99)
In Fig. 5 we see that the transition path time τ T P (L0)
is a monotonically decreasing function of the barrier
height U , while the Kramers' time τ KF P (L0) is a sym-
metric function and has a minimum of τ KF P = L2/(2D)
13
at U = 0. In fact, transition paths over a harmonic bar-
rier with U > 0 are faster, while transition paths over a
harmonic well characterized by U < 0 are slower com-
pared to the force-free case with U = 0. This can be
rationalized by Eq. (47), since the transition path time
for reaching from the boundaries to the center of the
harmonic potential are rather insensitive on whether U is
positive or negative (as will be shown in the next section),
but the residence time at the center of the harmonic po-
tential is much larger for the case of a harmonic well with
U < 0 than for a harmonic barrier with U > 0. The sym-
metric behavior of the Kramers' mean first-passage time
can be understood based on Eq. (60) since first-passage
time are transitive: the first-passage time for traversing
a harmonic potential is the sum of the first-passage times
from the boundary to the middle and from the middle to
the other boundary. We reiterate that mean first-passage
times are transitive, as shown in Eq. (60), while transi-
tion path times are not, as shown in Eq. (47).
In Fig. 6 we show the normalized distribution func-
tions for the transition path time (circles) and for the
Kramers' first-passage time (squares) for U = 3, obtained
from BD simulations. The transition path time distri-
bution shows a pronounced peak around τ D/L2 = 0.1,
close to the mean transition path time τ T P (L0)(U =
3)D/L2 ≈ 0.1, as seen in Fig.
In contrast, the
Kramers' first-passage time distribution is quite broad,
the first moment is given by τ KF P (L0)(U = 3)D/L2 ≈ 1
and thus is 10 times larger than the mean transition path
time.
5.
For the transition path shape we find
shape(x00) = τ T P (L0) − L2
τ T P
2DU
where D+(x) = e−x2(cid:82) x
0 dtet2
reduces to −τ T P (L0) given in Eq. (94) for x0 = 0.
(cid:90) √
U (2x0/L−1)
√
U
(cid:32)
dy
√
erf(y) − erf(
U (2x0/L − 1)) − erf(
U )
√
√
erf(
− 1
2
U )
(cid:33)
D+(y),
(100)
is the Dawson integral function. The second term in Eq. (100) vanishes for x0 = L and
Figure 7 depicts the mean transition path shapes
shape(x00) in Eq. (100) for different values of the bar-
τ T P
rier height U . Transition paths are faster for positive
values of U , i.e.
for paths that have to go over a har-
monic barrier top, while the slow down for negative val-
ues of U , i.e.
for paths that have to traverse a har-
monic well. Again, we observe a pronounced asymmetry
of the mean shape of transition paths, paths start out
quickly and reach the boundary at x = L with vanish-
ing slope. Filled symbols show BD simulation results for
τ F P (0x0) while open symbols show BD simulation re-
shape(x00), both for U = 3. We observe good
sults for τ T P
agreement between the two different ways of extracting
transition path shapes, as expected based on our ana-
lytical results, as well as with our analytically derived
shape.
E. Harmonic ramp
Here we consider the harmonic potential F (x) =
U x(2 − x/L)/L which has a barrier top F = U at the
final position x = L.
14
FIG. 5: Results for the harmonic potential F = 4U x(1 − x/L)/L as a function of the barrier height U . (a) Mean transition
path time τ T P (L0) from Eq. (94) (solid red curve) on a log-log scale, compared with the asymptotic expressions Eqs. (95) and
(96) (dashed lines). (b) Mean transition path time τ T P (L0) (solid red curve) on a log-linear compared with BD simulation
data (symbols). Solid blue curves depict the Kramers' mean first-passage time τ KF P (L0) from Eq. (98). The horizontal
dashed line depicts the force-free transition path time τ T P = L2/(6D).
FIG. 6: Normalized distribution functions for the transition
path time τ T P (L0) (circles) and the Kramers' first-passage
time τ KF P (L0) (squares) in a harmonic potential at U = 3,
obtained from BD simulations.
The transition path time reads
L2(cid:82) √
0
U
D
τ T P (L0) =
dy y2e−y2
√
√
πU erf(
F2,2(y2)
U )
.
(101)
For small U we find the asymptotic expression
− 2U 2
945
τ T P (L0) D/L2 ≈ 1
6
− U
90
,
(102)
shape(x00) from Eq.
FIG. 7: Mean transition path shape τ T P
(100), for different values of the barrier height U of the har-
monic potential F = 4U x(1−x/L)/L. Symbols show BD sim-
shape(x00)
ulation results for τ F P (0x0) (filled circles) and τ T P
(open squares) for U = 3. The horizontal dashed line depicts
the force-free transition path time τ T P = L2/(6D).
while for large U we find
τ T P (L0) D/L2 ≈ ln U
4U
(103)
Figure 8 depicts τ T P (L0) as function of the barrier
In Fig. 8-(a) we show, on double logarith-
height U .
mic scales, the numerically integrated τ T P (L0) from Eq.
.
tTPt0TPt¥TPtKFP10-210-110010110210310-210-1100101UtTPL0DL2(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)-10-5051010-1100101UtTPL0DL2abln2egU8U(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:224)(cid:230)tTPL0(cid:224)tKFPL00.00.20.40.60.81.01.21.4024681012tDL2Pt(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)(cid:225)U(cid:61)(cid:45)2U(cid:61)0U(cid:61)3U(cid:61)50.00.20.40.60.81.00.00.10.20.30.4x0LΤshapeTPx00 15
FIG. 8: Results for the harmonic ramp F (x) = U x(2− x/L)/L. (a) Mean transition path time τ T P (L0) from Eq. (101) (solid
red curve) on log-log scales, the asymptotic expressions Eqs. (102) and (103) are shown by dashed black lines. The blue line
shows Kramers' mean first-passage time τ KF P (L0)D/L2 from Eq. (104). (b) Same curves shown on log-linear scales, compared
with BD simulation data for transition paths starting from the left, τ T P (L0), (circles) and for transition paths starting from
the right, τ T P (0L), (triangles). The horizontal dashed line depicts the force-free transition path time τ T P (L0)D/L2 = 1/6.
(101) by the solid red curve and compare with the asymp-
totic expressions Eqs. (102) and (103) (dashed curves).
In Fig. 8-(b) we show τ T P (L0) from Eq.
(101) on a
log-linear scale, the symbols show BD simulation results.
The solid blue curves in Fig. 8 depict the Kramers' mean
first-passage time, which is given by
(cid:16)√
(cid:17)
(cid:16)√
(cid:17)
U
erfi
U
πerf
− F2,2(−U )
2
,
(104)
τ KF P (L0) =
L2
D
for large U .
The transition path time τ T P (L0)D/L2 is nonmono-
tonic and is maximal for finite U around U ≈ −21/8,
implying that transition paths that move down a weak
harmonic ramp are slower than in the force-free case.
For large U, τ T P (L0) decreases, similar to the linear
In contrast, the
potential case shown in Fig.
Kramers' mean first-passage time τ KF P (L0) exponen-
tially increases as U increases.
4-(a).
and has the leading order expression
4U
√
τ KF P (L0) =
πL2
4D
eU
U 3/2
,
(105)
For the transition path shapes we find
shape (x00) = τ T P (L0) −
τ T P
shape (x0L) = τ T P (L0) −
τ T P
√
πL2
2DU
√
πL2
2DU
(cid:90) y0
(cid:90) √
0
y0
(cid:20)
(cid:21)
dyey2
erf(y)
(cid:104)
1 +
√
erf(
erf(y)
erf(y0)
U ) − erf(y)
(cid:105)
√
erf(y)
− 2
erf(
[erf(y0) − erf(y)]
U )
,
U
dyey2
erf(y0) − erf(
U )
√
(106)
,
(107)
where y0 ≡ √
U (1 − x0/L).
Figure 9 depicts the transition path shapes starting
shape(x00) (solid curves) from Eq. (106),
shape(x0L) (broken curves)
from the left, τ T P
and starting from the right, τ T P
from Eq. (107), for different values of the barrier height
U . The symbols show the corresponding results from
BD simulations. Note that the transition path shapes
shape(x0L) at constant U are asym-
shape(x00) and τ T P
τ T P
ΤTPΤ0TPΤ(cid:165)TPΤKFP10(cid:45)210010210410(cid:45)510(cid:45)410(cid:45)310(cid:45)210(cid:45)1100101102UΤTPL0DL2(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:243)(cid:45)20(cid:45)10010200.100.501.00UΤTPL0DL2ablnU4U 16
paths for arbitrary free energy and diffusivity landscapes.
We use a combination of the backward and forward
Fokker Planck approaches to derive explicit expressions
for transition and first-passage path shapes. To clarify
the interpretation of our results, we also present convo-
lution expressions for the distribution functions of tran-
sition path and passage times. We show that the mean
shape of Kramers' first-passage paths is identical to the
shape of transition paths shifted by a constant. Based
on our analytic theory, we present mean shapes for sev-
eral simple model potentials. We illustrate our results by
trajectories generated from Brownian dynamics simula-
tions. Interestingly, transition path shapes are intrinsi-
cally asymmetric, they start out with finite velocity and
reach the target position with infinite velocity, which is
easily understood from our sum rules for transition path
and passage times.
The transition path shapes we predict can be compared
straightforwardly with simulations for proteins that un-
dergo folding and unfolding events and will allow for a
crucial test of the assumptions underlying the projection
onto a one-dimensional reaction coordinate. With fur-
ther developments of experimental single-molecule tech-
niques, our results for the transition path shapes can also
be compared with experimental results in the future. For
such a comparison, note that a reflecting boundary con-
dition at x = xA, as used in our calculations, is typically
not present in molecular dynamics simulations nor in ex-
periments. To apply our formulas, one can easily shift
the reflecting boundary conditions to a position where
the trajectory never visits. Alternatively, one can cut
out all trajectory sections that visit the region behind
the reflecting boundary condition and merge the remain-
ing trajectory parts with a continuous concatenated time,
which is valid in the limit of vanishing memory and ef-
fective mass.
FIG. 9: Mean transition path shapes starting from the left,
shape(x00) (solid curves) from Eq. (106), and mean transi-
τ T P
shape(x0L) (bro-
tion path shapes starting from the right, τ T P
ken curves) from Eq. (107), for different values of the barrier
height U of the harmonic ramp F (x) = U x(2 − x/L)/L. The
symbols show the corresponding BD simulation results.
metric with respect to the exchange of starting and end
positions, due to the asymmetry of the barrier poten-
tial (this becomes clear by comparing the mean shapes
for U = 0 (grey line) and for U = −5 (red line) start-
ing from the left boundary and starting from the right
boundary).
V. CONCLUSION
Acknowledgements
Based on the one-dimensional Fokker-Planck equation,
we develop the theoretical formalism to calculate mean
shapes of transition paths and of Kramers' first-passage
The authors thank Bill Eaton for stimulating discus-
sions. Financial support from the DFG (SFB 1078) is
acknowledged.
[1] P. G. Bolhuis, D. Chandler, C. Dellago, and P. L.
Acad. Sci. U.S.A. 106, 11837 (2009),
Geissler, Annu. Rev. Phys. Chem. 53, 291 (2002),
[8] H. S. Chung, K. McHale, J. M. Louis, and W. A. Eaton,
[2] R. B. Best and G. Hummer, Proc. Natl. Acad. Sci. U.S.A.
Science 335, 981 (2012),
102, 6732 (2005),
[3] P. Metzner, C. Schutte, and E. Vanden-Eijnden, J.
Chem. Phys. 125, 084110 (2006),
[9] H. Yu, A. N. Gupta, X. Liu, K. Neupane, A. M. Brigley,
I. Sosova, and M. T. Woodside, Proc. Natl. Acad. Sci.
U.S.A. 109, 14452 (2012),
[4] G. Hummer, J. Chem. Phys. 120, 516 (2004),
[5] S. Chaudhury and D. E. Makarov, J. Chem. Phys. 133,
[10] H. S. Chung and W. A. Eaton, Nature 502, 685 (2013).
[11] T.-H. Lee, L. J. Lapidus, W. Zhao, K. J. Travers, D. Her-
034118 (2010),
[6] E. Rhoades, M. Cohen, B. Schuler, and G. Haran, J. Am.
Chem. Soc. 126, 14686 (2004).
[7] H. S. Chung, J. M. Louis, and W. A. Eaton, Proc. Natl.
schlag, and S. Chu, Biophys. J. 92, 3275 (2007).
[12] K. Neupane, D. B. Ritchie, H. Yu, D. A. N. Foster,
F. Wang, and M. T. Woodside, Phys. Rev. Lett. 109,
068102 (2012),
(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:231)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)(cid:230)U=-5U=0U=2U=50.00.20.40.60.81.00.000.050.100.150.000.050.100.15x0LtshapeTPx0LDL2tshapeTPx00DL2 17
[13] K. Truex, H. S. Chung, J. M. Louis, and W. A. Eaton,
72, 4350 (1980),
Phys. Rev. Lett. 115, 018101 (2015),
[21] R. Zwanzig, Nonequilibrium Statistical Mechanics (Ox-
[14] D. E. Shaw, P. Maragakis, K. Lindorff-Larsen, S. Piana,
R. O. Dror, M. P. Eastwood, J. A. Bank, J. M. Jumper,
J. K. Salmon, Y. Shan, W. Wriggers, Science 330, 341
(2010).
ford University Press, USA, 2001),
[22] M. Hinczewski, Y. von Hansen, J. Dzubiella, and R. R.
Netz, J. Chem. Phys. 132, 245103 (2010),
[23] Y. von Hansen, F. Sedlmeier, M. Hinczewski, and R. R.
[15] Z. Zhang and H. S. Chan, Proc. Natl. Acad. Sci. U.S.A.
Netz, Phys. Rev. E 84, 051501 (2011).
109, 20919 (2012),
[24] D. R. Cox, Renewal theory, vol. 4 (Methuen London,
[16] R. Frederickx, T. in't Veld, E. Carlon, Phys. Rev. Lett.
1962).
112, 198102 (2014).
[25] N. G. Van Kampen, Stochastic processes in physics and
[17] M. Sega, P. Faccioli, F. Pederiva, G. Garberoglio, and
chemistry, vol. 1 (Elsevier, 1992).
H. Orland, Phys. Rev. Lett. 99, 118102 (2007),
[18] H. Orland, J. Chem. Phys. 134, 174114 (2011),
[19] G. H. Weiss, Adv. Chem. Phys. 13, 1 (1967).
[20] A. Szabo, K. Schulten, and Z. Schulten, J. Chem. Phys.
[26] C. W. Gardiner, Handbook of stochastic methods, vol. 4
(Springer Berlin, 1985).
|
1611.03637 | 2 | 1611 | 2017-05-08T15:49:21 | The impact of anticipation in dynamical systems | [
"physics.bio-ph",
"nlin.AO"
] | Collective motion in biology is often modelled as a dynamical system, in which individuals are represented as particles whose interactions are determined by the current state of the system. Many animals, however, including humans, have predictive capabilities, and presumably base their behavioural decisions---at least partially---upon an anticipated state of their environment. We explore a minimal version of this idea in the context of particles that interact according to a pairwise potential. Anticipation enters the picture by calculating the interparticle forces from linear extrapolations of the particle positions some time $\tau$ into the future. Simulations show that for intermediate values of $\tau$, compared to a transient time scale defined by the potential and the initial conditions, the particles form rotating clusters in which the particles are arranged in a hexagonal pattern. Analysis of the system shows that anticipation induces energy dissipation and we show that the kinetic energy asymptotically decays as $1/t$. Furthermore, we show that the angular momentum is not necessarily conserved for $\tau >0$, and that asymmetries in the initial condition therefore can cause rotational movement. These results suggest that anticipation could play an important role in collective behaviour, since it induces pattern formation and stabilises the dynamics of the system. | physics.bio-ph | physics | The impact of anticipation in dynamical systems
P. Gerlee1, K. Tunstrøm2, T. Lundh1 and B. Wennberg1
1Mathematical Sciences, Chalmers University of Technology
and University of Gothenburg, 412 96 Goteborg and
2Department of Physics, Chalmers University of Technology, 412 96 Goteborg
(Dated: February 5, 2018)
Abstract
Collective motion in biology is often modelled as a dynamical system, in which individuals are
represented as particles whose interactions are determined by the current state of the system.
Many animals, however, including humans, have predictive capabilities, and presumably base their
behavioural decisions-at least partially-upon an anticipated state of their environment. We
explore a minimal version of this idea in the context of particles that interact according to a
pairwise potential. Anticipation enters the picture by calculating the interparticle forces from
linear extrapolations of the particle positions some time τ into the future. Simulations show that
for intermediate values of τ , compared to a transient time scale defined by the potential and the
initial conditions, the particles form rotating clusters in which the particles are arranged in a
hexagonal pattern. Analysis of the system shows that anticipation induces energy dissipation and
we show that the kinetic energy asymptotically decays as 1/t. Furthermore, we show that the
angular momentum is not necessarily conserved for τ > 0, and that asymmetries in the initial
condition therefore can cause rotational movement. These results suggest that anticipation could
play an important role in collective behaviour, since it induces pattern formation and stabilises the
dynamics of the system.
PACS numbers: 89.75.Kd, 87.23.-n, 05.45.-a
7
1
0
2
y
a
M
8
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
7
3
6
3
0
.
1
1
6
1
:
v
i
X
r
a
1
INTRODUCTION
Countless examples of collective motion are found in biological systems, spanning from
swarming bacteria to human crowds [1]. The dynamical patterns exhibited by groups of ani-
mals have fascinated humans across millennia, but with modern technology, this fascination
has been channeled into an active research area, where methodologies across disciplines-
from biology to physics and engineering sciences-are essential to achieve progress [2–4].
From a modeling perspective, much research derives from the idea that simple rules
of interaction between animals, e.g. attraction, repulsion and alignment [5], can explain
observed swarming patterns [6]. Typically, swarming behaviour is modelled as a collection
of particles that represent the organisms in question. To each particle one assigns a position
and velocity. The velocity determines the evolution of the position, and the velocity is in
turn influenced by the position and velocity of neighbouring particles. Taking inspiration
from physics, the interactions between individuals are often assumed to result from pairwise
forces that only depend on the distance between the individuals. Typically the force is
defined via a potential that contains a repulsive and attractive region such that at some
intermediate distance the interaction energy is minimised. In terms of animals this would
correspond to some preferred distance between neighbouring individuals [7]. In addition,
most physics inspired models rely on self-propulsion to drive pattern formation [8] and often
include a noise term as well [9].
An underlying assumption in most current collective motion models is that individuals
react and update their velocity according to the current state of other individuals. Con-
trary to this, we know that many animals including humans have the ability to anticipate
movement, and act on predicted states. Humans anticipate the movement of visual cues,
such as other people in a moving crowd, by extrapolating their positions in time [10]. This
process occurs within the retina itself, and is in fact necessary if we are to respond to rapid
visual cues, since the delay induced by phototransduction is on the order of 100 ms [11]. In
addition to performing extrapolation the retina also detects when its prediction fails and
signals this downstream to the visual cortex [12]. Anticipation is not restricted to humans,
but has also been detected among insects [13], amphibians [14] and fish [15]. Therefore
anticipation most likely plays an important role in the behaviour of animals that engage in
flocking and swarming.
2
A recent study of human interactions in crowds quantified the effect of anticipation by
showing that the strength of physical interaction does not depend on distance, but on the
time to collision [16]. This suggests that humans act in an anticipatory fashion and use cur-
rent positions and velocities to extrapolate possible future collisions and update their current
velocity to avoid such collisions. In terms of the above mentioned interaction potential one
might then assume that individuals interact according to anticipated future positions and
adjust their current velocities according to the predicted state of the system. This idea has
been investigated by Morin et al. [17], who considered the continuous-time Viscek model,
that assumes self-propelled particles that interact with neighbouring particles within a cer-
tain interaction radius. The classical model was modified so that the difference in angle
between particle i and j, θi − θj, was altered to θi − (θj + ασj), where σj is the sign of the
angular velocity of particle j and α is some positive parameter. Depending on the value of
α and the magnitude of the noise that model can exhibit isotropic behaviour, spinning and
flocking, which shows that anticipation indeed can have important consequences. The effect
of anticipation has also been investigated in a lattice-based model inspired by the swarm-
ing of soldier crabs [18]. They showed that mutual anticipation leads to dense collective
motion with a high degree of polarisation, and that the turning response depends on the
distance between two individuals rather than the relative heading, which is in agreement
with empirical data.
In this paper we also explore the idea of anticipation, but in the context of an interacting
particle system where the interaction forces are calculated not from current positions, but
from positions extrapolated some time τ into the future. For the sake of simplicity we assume
that the future positions are given by a linear extrapolation from the current velocities,
although more elaborate means of extrapolation are conceivable (see Discussion). In order
to get a thorough understanding of the dynamics that anticipation induces we disregard self-
propulsion, alignment and other processes commonly found in models of collective behaviour
and focus on the effects in a simple of model where individuals are represented as interacting
particles.
3
A MODEL OF ANTICIPATION IN DYNAMICAL SYSTEMS
To introduce the concept of anticipation let us begin with a simple example. Consider
two particles of mass m connected by a spring with rest length zero and spring constant
k/2, where k > 0. The distance x between the two particles obeys the equation mx = −kx,
which is simply that of a harmonic oscillator. Now let us assume that the particles anticipate
the movement of one another, and hence that the forces acting on the particles are not given
by their instantaneous positions, but by the anticipated positions some time τ in the future,
i.e. we calculate the forces based on predicted positions xp
i (t + τ ) = xi(t) + τ vi(t) for particle
i = 1 and 2. In this case the equation of motion of the interparticle distance x is given by
mx = −kx − kτ x. This is the equation for a damped harmonic oscillator, and hence we
conclude that in this simple system anticipation has a damping effect on the dynamics. In
contrast to the undamped system there is dissipation of energy and for any initial condition
the system will reach a stationary state with x = 0 as t → ∞.
We now extend the idea of anticipation to N identical particles that interact via some
potential U (r) and hence obey the following equations of motion:
(cid:126)xi(t) = (cid:126)vi(t),
m (cid:126)vi(t) = −(cid:88)
j(cid:54)=i
(cid:126)∇U(cid:0)(cid:126)xp
j (t + τ )(cid:1) ,
i (t + τ ) − (cid:126)xp
(1)
where m = 1 is the mass and (cid:126)xi, (cid:126)vi ∈ R2, i.e. the system is defined in the plane. Again
we assume that the individuals make use of a linear prediction of future positions such that
(cid:126)xp
i (t + τ ) = (cid:126)xi(t) + τ(cid:126)vi(t). We note that for τ = 0 the system reduces to a standard system
of interacting particles.
We will use a generalized Morse potential,
U (x) = Cre−x/lr − Cae−x/la ,
(2)
to define the interaction between pairs of particles, where Cr and Ca represent the ampli-
tude of the repulsive and attractive component of the potential, and lr and la denote their
respective ranges. In the simulations presented here we will use Cr = 15, Ca = 2.5 and
la = 0.1, lr = 0.05. We note that the force between particle i and j is in the direction of the
vector
i,j = (cid:126)xi(t) − (cid:126)xj(t) + τ ((cid:126)vi(t) − (cid:126)vj(t)).
rτ
(3)
4
This implies that for τ > 0 the direction of the force is generally not parallel to the vector
(cid:126)xi(t) − (cid:126)xj(t), pointing from particle i to j, but is influenced by the velocity of the two
particles.
In the following we will consider particles moving in two dimensions with open boundaries.
We typically initialise the system with the particles at random positions within a disk, and
zero initial velocities. All numerical solutions have been carried out using a Runge-Kutta
method of order 8 from the GNU Scientific Library [19].
RESULTS
Before investigating the impact of anticipation on the dynamics we need to set the an-
ticipation time τ into relation with other time and length scales present in the system. The
model has one natural spatial scale, given by the minimum of the Morse potential,
¯rp = argmin U (r) ,
which with the Morse potential and parameters used here is given by
¯rp =
lalr
la − lr
log
Cr/Ca
lr/la
≈ 0.25.
(4)
(5)
The anticipation time, τ , gives one natural time scale, but because the system is defined
in the plane without boundaries it is difficult to define time scales or spatial scales valid
asymptotically for long time intervals, to compare ¯rp and τ with. However, it is possible to
define a relevant transient time scale expressed in terms of the initial data. We let
¯fT = E[∇U ((cid:126)xj − (cid:126)xk)] ,
(6)
be the expected interparticle force, where the expectation is taken over a randomly chosen
pair of initial positions of the particles. From this we define a transient time scale
(cid:113)
¯tT =
2¯rp/ ¯fT ,
(7)
which is the time needed for a particle with initial velocity zero to move a distance ¯rp when
subject to a constant force of strength ¯fT .
By solving the equations of motion (1) numerically with initial positions uniform random
within a disk and initial velocities equal to zero, we observe for τ /¯tT ≈ 1, independent
5
of particle number, the rapid formation of milling structures, where the particles organise
into rotating patterns. For smaller τ /¯tT (cid:28) 0.1 the particles behave much like the classical
system where nearby particles tend to form clusters, whereas others disperse. For τ /¯tT (cid:29) 10
particles also disperse, but do so in smaller clusters of aligned particles. The following section
are devoted to understanding this behaviour for both small and large particle numbers.
Small particle numbers
To get a better understanding of the effect of anticipation we start by looking at the
dynamics of small systems with N ∈ [2, 11] and set τ = 1. The particles are placed at
random positions in the unit circle and the initial velocities are zero for all particles. This
implies that the initial density and therefore the transient time scale will vary with N , but
for small N we have that τ and ¯tT are of same order of magnitude, and hence we expect
collective behaviour to emerge.
For N = 2 the system behaves much like the two-particle system mentioned in the
introduction. The particles attract one another and rapidly converge to a distance given
by the equilibrium distance of the pairwise potential. This suggests that energy is being
dissipated, a fact we will return to later.
For N = 3, the particles start rotating on the same orbit around their common centre of
gravity. A similar behaviour is seen for larger systems, but the number of unique orbits and
their radii depend on the number of particles N (see fig. 1). We note that the configurations
closely resemble a partial hexagonal lattice built from equilateral triangles. These patterns
are rigid rotations of a locally crystalline configuration, much like the dynamics reported in
region VI of [8]. The main difference being that in our case there is no self-propulsion, and
the particles are only influenced by a pairwise potential with anticipation.
Analysis of milling patterns
The reason behind the observed rotation for N > 2 can be understood by considering
the forces acting on each particle (see fig. 2). For simplicity we consider the case N = 3.
We assume that the three particles are symmetrically distributed, i.e. with angle 2π
3 between
them, on the inner dashed circle of radius r(t) and with the same speed (cid:126)v(t) in the clockwise
6
FIG. 1. Rotational configurations observed for N ∈ [3, 11]. Particles are shown as circles and red
lines correspond to the unit velocity vector of each particle. The lines show the trajectories of the
particles where the first 10 time steps have been discarded.
direction. For generality we assume that the particles move with velocity (cid:126)v(t) which has
both a tangential and radial component. The non-zero velocity gives rise to anticipated
positions that lie on a circle with radius R(t). A third radius is also of importance, R0,
which is the radius at which the interparticle distance equals the equilibrium distance of the
potential. For the case N = 3, we have that the equilibrium (or resting) radius is
R0 =
lalr log( Crla
)
√
Calr
3(la − lr)
.
A symmetric configuration on R0 implies that the force from the potential (cid:126)F (R0) = 0,
where (cid:126)F (R) is the net force acting on each particle when at a distance R from the centre.
If R(t) > R0 the net force will be radially inwardly directed from the point (cid:126)R(t). Now,
7
-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4-0.400.4 this force, (cid:126)F (R(t)) is not applied on the point (cid:126)R(t) which is the anticipated position on the
outer circle, but instead at the current positions at (cid:126)r(t). By decomposing this force into
its radial and tangential components we get a triangle which is similar to the right-angled
triangle which has (cid:126)R(t) as the hypotenuse (see fig. 2). The net force can be decomposed into
a radial component, Fr, that is generating a central force, and a tangential force Fθ which
is retarding the rotation. Due to the fact that the anticipated position is outside R0, the
particles will experience attractive forces, although they in fact are located in the repulsive
region of the potential. However, note that this is only true if (cid:126)v(t)τ is large enough so that
the anticipated position lies outside of R0.
To obtain a quantitative understanding of the three particle system we formulated an
ODE-system for the special case of N particles located on a circle (as in figure 2). We describe
the dynamics of a single particle, but since the configuration is symmetric the positions and
velocities of the other particles can be obtained by shifting the solution appropriately. The
equations of motion are given by: (cid:126)r(t) = (cid:126)v(t)
(cid:126)v(t) = −F (R(t))
where (cid:126)R(t) = (cid:126)r(t) + τ(cid:126)v(t) and R(t) = (cid:126)R(t). The force is given by
(cid:126)R(t)
R(t)
(cid:19)
F (R(t)) = 2
where
(cid:18) Ca
2(cid:88)
n−1
la
√
k=1
αk =
e−αk/la − Cr
lr
e−αk/lr
cos(
π − k 2π
n
2
)
(cid:114)
2R(t)
1 − cos(k
2π
n
)
(8)
(9)
(10)
is the interparticle distance between the focal particle and particle k.
We solve this system numerically for the the case N = 3 (see fig. 3), and observe two
distinct phases: (i) an accelerating phase, where the particles move onto a circle with radius
slightly smaller than the equilibrium distance of the potential, and (ii) a milling phase in
which the particles in unison slowly spiral outwards towards the equilibrium circle with
radius R0.
From the equation system (8) and figure 2, we can obtain an analytic expression of how
the speed decays in the milling phase in the following way. The second equation in (8) gives
that
vθ = −Fθ
8
FIG. 2. A schematic picture of a symmetric three particle system that is rotating on a circle
with a radius r(t) which is smaller than the equilibrium radius R0. The small red circles show
the current positions and the green circles denote the anticipated positions obtained by adding
the vector (cid:126)v(t)τ to the current position, i.e. (cid:126)R(t) = (cid:126)r(t) + (cid:126)v(t)τ . Since the anticipated position
is located outside the equilibrium circle, there will be an attracting force applied at the actual
position (cid:126)r(t). By decomposing this force, we get a central acceleration and a tangential retardation
that eventually makes the speed of the particles go to zero.
where vθ is the tangential velocity. By the similarity of the two right angled triangles in
9
R0r(t)vϑτFϑFr FIG. 3. The solution of the ODE-system (8) with the focal particle (blue circle) at initial position
(cid:126)r(0) = (0, 0.1) with (cid:126)v(0) = (0, 0.35). The particles spiral out towards the equilibrium radius R0.
figure 2 we have that
Fθ =
vθτ.
F (R(t))
R(t)
Note that both F (R(t)) and vθ tend to zero as t → ∞. We are interested in studying
how quickly that happens, and therefore assume that the system has reached the milling
phase. This means that the particles are orbiting very close to the equilibrium circle and
that F (R(t)) ≈ Fr and that R(t) ≈ R0. By making these assumptions, we can simplify the
10
-0.15-0.10-0.050.050.100.15-0.15-0.10-0.050.050.100.15 differential equation in the following way:
vθ = −Fθ = −F (R(t))
R(t)
vθτ ≈ − Fr
R0
vθτ = − v2
θ
R2
0
vθτ,
where the last equality holds because the particles are following a circular movement with
radius R0 for which we have Fr = v2
θ /R0. Thus, we end up with the following asymptotic
separable differential equation
with solution
vθ(t) =
vθ = −v3
θ τ
R2
0
1(cid:113) 2τ t
R2
0
.
(11)
+ 1
vθ(0)2
From this expression it is clear that v ∼ t−1/2 as t → ∞, and from that we conclude
that the kinetic energy should go as K(t) = mv2/2 ∼ t−1. This result is in agreement
with the long-term kinetic energy calculated for a three particle system with random initial
conditions (see fig. 4). This indeed shows that the milling phase is transient, but also that
the rate of energy dissipation is low and the kinetic energy scales as 1/t.
Large particle numbers
We now consider systems with N (cid:29) 10 particles and look for large scale patterns in the
dynamics. Figure 5 contains snapshots from a simulation with N = 100 particles which
show that the particles first aggregate into local clusters (t ≈ 12) that eventually coalesce
and form a single milling structure (t ≈ 90). Similar dynamics are seen with N = 1000
particles initialised at the same density (see fig. 6), with the difference that not all clusters
merge within the duration of the simulation. Although not visible due to the large system
size, the particles are organised in a hexagonal lattice.
In the case of N = 1000 particles initialised at a 10-fold higher density (see fig. 7) we
observe an initial repulsive phase (at t ≈ 4.5), which leads to an expansion of the cluster
(at t ≈ 15). However, the cluster remains coherent and contracts into an approximately
In conclusion, we observe
hexagonal lattice, which rotates around the centre of gravity.
similar dynamics independent of system size. The system which is initially disordered self-
organises into an approximately hexagonal lattice which rotates as a rigid body around its
centre of gravity.
11
FIG. 4. The kinetic energy K(t) as a function of time for N = 3. The dashed line corresponds to
1/t. After an initial acceleration the kinetic energy decays approximately as 1/t in agreement with
(11).
It appears as if the inclusion of anticipation allows the system to dissipate energy and
reach a state which minimises the total potential energy. In order to investigate this we
compared the configuration in figure 5 at t = 250 with a configuration of stationary par-
ticles that minimises the total potential energy (calculated using Mathematica's function
Minimize). The comparison is shown in figure 8 and reveals almost perfect agreement be-
tween the milling configuration obtained with anticipation (black dots) and the stationary
configuration that minimises the total potential energy, which consists of a partial hexagonal
lattice.
We now move on to the question of energy dissipation for large N , which seems to play
a crucial role in the dynamics of the system.
12
10-210-110010110210310-410-310-210-11001/tN=3log tlog K(t) FIG. 5. Positions of 100 particles at different times, starting from a random initial configuration
with zero velocities. The initial distribution was uniform in a disk with radius 2.4. The anticipation
time τ = 0.25, and the transient time scale equals 1. The particles are coloured by the cluster they
belong to at t = 12.
Energy dissipation
13
FIG. 6. Positions of 1000 particles different times, starting from a random initial configuration with
zero velocities. The initial distribution was uniform in a disk with radius 10.0. The anticipation
time τ = 0.25, and the transient time scale equals 0.8. The particles are coloured by the cluster
they belong to at t = 2000.
The example with two spring-coupled particles, and the special case of a symmetric
configuration of three particles shows that anticipation may result in energy dissipation.
The following calculation shows that the total energy, E(t), computed as the sum of kinetic
14
FIG. 7. Positions of 1000 particles different times, starting from a random initial configuration with
zero velocities. The initial distribution was uniform in a disk with radius 2.5. The anticipation
time τ = 0.05, and the transient time scale equals 0.06.
energy and the anticipated potential energy,
N(cid:88)
E(t) =
(cid:126)vi2 +
U ((cid:126)xi − (cid:126)xj + τ ((cid:126)vi − (cid:126)vj)) ,
(12)
N(cid:88)
i=1
i,j=1
i(cid:54)=j
15
FIG. 8. The black dots show the final point configuration, and the coloured dots a local energy
minimum. The colours represent the stiffness of each particle position, calculated from the largest
eigenvalue of the linearised system as a function of only one particle, given that the others are
fixed.
is strictly decreasing along the orbits, when not all velocities are constant. Here we consider
a more general anticipating dynamical system, of which the particle system studied above
is but a special case:
(cid:126)x = (cid:126)v
(cid:126)v = −∇U ((cid:126)x + τ(cid:126)v) ,
16
(13)
where (cid:126)x, (cid:126)v ∈ Rn, and U : Rn → R is a smooth and bounded potential. Then
d
dt
(E(t)) =
2
(cid:17)
d
dt
+ U ((cid:126)x + τ(cid:126)v)
(cid:18)(cid:126)v2
(cid:19)
= (cid:126)v · (cid:126)v + ∇U ((cid:126)x + τ(cid:126)v) ·(cid:16) (cid:126)x + τ (cid:126)v
(cid:17)
= (cid:126)v ·(cid:16) (cid:126)v + ∇U ((cid:126)x + τ(cid:126)v)
(cid:125)
(cid:123)(cid:122)
(cid:124)
−τ (cid:126)v2 .
=0
(14)
(15)
Hence the sum of kinetic energy and the anticipated potential energy are strictly decreasing
until all velocities are constant. This behaviour can be seen in figure 9A, which shows the
how the total energy evolves for the simulation presented in figure 5 with N = 100 particles.
The anticipated total energy is strictly decreasing and the jumps correspond to collision and
merger of clusters. Note that the classical total energy (calculated with τ = 0) is not strictly
decreasing but increases transiently during collisions.
The rate of dissipation
For the three particle system we could show that the energy decays as 1/t. We now
analyse the rate of energy dissipation for a general potential U (x) and arbitrary system size
N .
With a potential given by a positive definite quadratic form, U (x) = (cid:126)xT A(cid:126)x, the system
is again a damped oscillator, whose solutions converge exponentially to zero, but for a
semidefinit form, say U = 1
n, the system never comes to rest, because the
potential is constant along the degenerate x1-direction. The anticipation does not change
2λnx2
2λ2x2
2 + ··· + 1
this, and seen from the point of view of simulations, the same will be observed if the quadratic
form has an almost degenerate subspace, in the sense that some of the eigenvalues of A are
several magnitudes smaller than the the dominant eigenvalues.
However, if the potential U is constant, or nearly constant, on a curved manifold, e.g. a
curve, the situation is different. The three particles on a circle is an example of this more
general situation. Here the potential attains its minimum on a one-dimensional curved
submanifold, the circle, with (cid:126)x(t) very close to the circle, and with (cid:126)v(t) almost tangential to
the manifold. A Taylor expansion of the potential U then gives
(cid:126)v(cid:107)(t) = O((cid:126)v(cid:107)2)
d
dt
17
FIG. 9. (A) The evolution of energy for the simulation represented in fig 5. The jumps in the two
curves correspond to the merger of clusters. (B) The evolution of the kinetic energy for N = 3, 10, 20
and 50. The initial condition was chosen with positions according to a global minimum of the
potential, and velocities corresponding to a rigid rotation of the point configuration. The slope of
the dashed lines is −1.
at least if τ(cid:126)v(cid:107) is small compared to the curvature of the manifold. This calculation can
be carried out in a more general setting, for example when the minimum of U is attained
18
timekinetic energyanticipatedclassicalAlog tlog K(t)B along a curve in Rn: Assume that Γ = {(cid:126)φ(x) = (φ1(s), ..., φn(s)) s ∈ I ⊂ R} is a segment
of a curve parametrised by s, and assume that U (x) is a quadratic function of (cid:126)x− (cid:126)x∗, where
(cid:126)x∗ is the point on Γ closest to x. Though not completely general, this would cover many
generic situations. We may choose a coordinate system so that
φk(0) = 0 for k = 1, ..., n ,
φ1(s) = s,
φ(cid:48)
k(0) = 0 for k = 2, ..., n ,
φ(cid:48)(cid:48)
k(0) = µk for k = 1, ..., n ,
so that the curve is tangent to the x1-axis of the coordinate system, and that it may be
parameterised with x1. The point (cid:126)x∗ = (s, φ2(s), ..., φn(s)) is found by minimising the
function(cid:80)(xk − φk(s))2, and at a minimum, this must satisfy
s − x1 +
(φk(s) − xk)φ(cid:48)
k(s) = 0 ,
(16)
which determines s as a function of (cid:126)x, at least near the origin. Hence
k=2
U (x1, ..., xn) =
=
1
2
1
2
((cid:126)x − (cid:126)φ(s))T A((cid:126)x − (cid:126)φ(s))
ν1(x1 − s)2 +
1
2
νk(φk(s) − xk)2 .
n(cid:88)
k=2
where for simplicity we have assumed that A is diagonal with diagonal elements νk. Re-
stricting Equation (16) to points on the x1-axis, we find that
n(cid:88)
so that
k=2
n(cid:88)
s(x1) = x1 − n(cid:88)
x1(s) = s +
k=2
1
2
U (x1, 0, ..., 0) = O(s6) +
=
1
2
1
2
n(cid:88)
n(cid:88)
k=2
j=2
19
µks3 + O(s4)
or
1
2
µkx3
1 + O(x4
1) ,
n(cid:88)
(cid:18)1
νk
µks2 + O(s3)
2
1 + O(x5
1)
k=2
µjµkνkx4
(cid:19)2
Therefore, with (cid:126)v = (v1, 0, ..., 0) and (cid:126)x = (cid:126)0,
(cid:126)v = −∇U ((cid:126)x + τ(cid:126)v) = −∇U (τ v1, 0, ..., 0) ,
d
dt
and finally
where γ = 2(cid:80)n
k=2
(cid:80)n
(cid:126)v2
2
d
dt
= −(cid:126)v · ∇U (τ v1, 0, ..., 0)
= −γv1τ 3v3
= −γτ 3v4
1 + O(v5
1)
1 + O(v5
1) ,
dt(cid:126)v2 ∼ −(cid:126)v4, which implies that
(17)
(18)
(19)
j=2 µjµkνk, and therefore d
(cid:126)v(t)2 ∼ 1
t
when t is large, and the exact expression depends on the constants µk, which determine
the curvature of the curve. Of course this is not a full proof, but rather motivation for
the observed behaviour. A complete proof would require an additional calculation showing
that the the component of (cid:126)F normal to Γ is sufficiently strong to confine (cid:126)x(t) to remain
sufficiently close to Γ, and (cid:126)v(t) to rest almost tangent to Γ, that the estimates above are
still valid.
For the N -particle systems with pair interactions that is the main theme of this paper,
the potential energy is invariant under rotations and translations, and hence the potential
miniumum is attained on a three-dimensional curved submanifold of the 2n-dimensional
configuration space. In simulations with a large number of particles it is difficult to observe
the 1/t behaviour, however, because the dynamics is complex and one may need to wait a
very long time before this asymptotic behaviour can be seen. To circumvent this problem we
have computed numerical solutions with initial data such that the positions (cid:126)xi(0) correspond
to a minimum for the potential energy, and with initial velocities corresponding to a rigidly
rotating body, and find excellent agreement with the predicted 1/t-decay of energy (see fig.
9B).
Situations where the mimimum manifold has higher dimension appear in simulations
with a large number of particles, where we often observe how the system breaks up into
smaller clusters that essentially do not interact with the other clusters, and hence represent
a case with a many dimensional subspace with almost constant potential energy, defined by
translations and rotations of each cluster.
20
Linear and angular momenta
So far we have shown that anticipation induces dissipation and generally leads to a 1/t-
decay in kinetic energy. This leads to a gradual reduction in the rotation of the cluster, but
we have not explained how this rotation comes about in the first place. To do this we need
to analyse the time evolution of the linear and angular momenta, (cid:126)M (t) and L(t), which we
define as
N(cid:88)
i=1
(cid:126)M (t) =
(cid:126)vi
and
L(t) =
N(cid:88)
i=1
(cid:126)vi × (cid:126)xi ,
(20)
where (cid:126)vi × (cid:126)xi denotes the cross product of (cid:126)v and (cid:126)x, which is a scalar in this two dimensional
setting. First,
N(cid:88)
(cid:88)
i=1
i,j
i(cid:54)=j
d
dt
(cid:126)M (t) =
=
d
dt
(cid:126)vi
Fi,j =
(cid:88)
i,j
i(cid:54)=j
1
2
(Fi,j + Fj,i) = 0 .
(21)
(cid:80)
1
N
By a change of variables we may assume that M (t) = (cid:126)0, and hence that the center of mass,
i (cid:126)xi is constant, and may therefore be set equal to zero. The angular momentum satisfies
(22)
(23)
(24)
d
dt
L(t) =
=
(cid:88)
(cid:88)
i
i,j
i(cid:54)=j
(cid:17)
(cid:16) (cid:126)vi × (cid:126)xi + (cid:126)vi × (cid:126)xi
(cid:88)
Fi,j × (cid:126)xi =
1
2
i,j
i(cid:54)=j
Fi,j × ((cid:126)xi − (cid:126)xj) .
i,j
(cid:12)(cid:12)(cid:1)
(cid:12)(cid:12))rτ
Fi,j = −U(cid:48)(cid:0)(cid:12)(cid:12)rτ
i,j ≡ −V ((cid:12)(cid:12)rτ
(cid:12)(cid:12)rτ
(cid:12)(cid:12)
(cid:88)
(cid:12)(cid:12))((cid:126)vi − (cid:126)vj) × ((cid:126)xi − (cid:126)xj)
V ((cid:12)(cid:12)rτ
i,j,
rτ
i,j
i,j
i,j
d
dt
L(t) = −1
2
τ
i,j
i(cid:54)=j
Note that the force Fi,j may be written
where rτ
i,j = (cid:126)xi − (cid:126)xj + τ ((cid:126)vi − (cid:126)vj), and we define V (r) = U(cid:48)(r)/r. Therefore
So with τ = 0, the angular momentum is constant as it should for a classical system. All
simulations presented here start with all initial velocities equal to zero, and therefore also
21
L(0) = 0, but the second derivative need not be. Computing L(t) gives
d2
dt2 L(t) =
(cid:32)
(cid:88)
i,j
i(cid:54)=j
V (cid:48)((cid:12)(cid:12)rτ
V ((cid:12)(cid:12)rτ
i,j
i,j
(cid:12)(cid:12))
(cid:12)(cid:12))
i,j
i,j · rτ
rτ
(cid:12)(cid:12)rτ
(cid:12)(cid:12)
(cid:16) (cid:126)vi − (cid:126)vj
i,j
((cid:126)vi − (cid:126)vj) × ((cid:126)xi − (cid:126)xj) +
(cid:17) × ((cid:126)xi − (cid:126)xj)
(cid:33)
.
(25)
At t = 0, when the velocities are zero, the first terms in the sum disappear, and using the
symmetries in the sum this yields
d2L
dt2 (0) =
=
(cid:88)
τ
2
i,j,k
i(cid:54)=j(cid:54)=k
V ((cid:126)xi − (cid:126)xj)V ((cid:126)xi − (cid:126)xk)((cid:126)xi − (cid:126)xj) × ((cid:126)xi − (cid:126)xk)
(26)
(27)
which typically will be non-zero for a random initial configuration. Therefore, contrary to
conservative systems, asymmetries in the initial configuration of a particle system initially
at rest, may result in a non-zero angular momentum, which is then observed as a milling
structure.
The time evolution of the angular momentum of a typical simulation is exhibited in
figure 10A which shows L(t) for the simulation presented in figure 5. The jumps in the
curve correspond to the merger of clusters, and we note that collisions typically increase
the absolute value of the angular momentum, while it decreases in between collisions due
to the 1/t dissipation of kinetic energy discussed above. Since L(t) is proportional to the
t, we expect the angular momentum to asymptotically scale
√
speed v(t), which scales as 1/
√
as 1/
t, which is precisely what is observed in simulations (see fig. 10B).
DISCUSSION
In this paper we have shown that internal prediction or anticipation leads to milling
behaviour in a system of particles that interact via a pairwise potential. The behaviour of
the system depends on the relation between the anticipation time τ and the transient time
scale ¯tT (7), such that when τ is of the same order of magnitude as ¯tT we observe the rapid
formation of rotating clusters that merge upon collision. These clusters have a crystalline
structure made up of a hexagonal arrangement of particles and rotate around the centre
of gravity as a rigid body. We have shown analytically that rotation emerges due to an
22
FIG. 10. (A) The evolution of angular momentum for the simulation represented in fig 5. The jumps
in the curve correspond to the merger of clusters. (B) The evolution of the angular momentum
for N = 3, 10, 20 and 50. The initial condition was chosen with positions according to a global
minimum of the potential, and velocities corresponding to a rigid rotation of the point configuration.
The slope of the dashed lines is −1/2.
increase in the angular momentum, which only occurs for τ > 0. Further we have shown
23
timeangular momentumAlog tlog L(t)B that anticipation leads to dissipative behaviour where the kinetic energy typically decreases
as 1/t.
These results show that motion anticipation – the capability to extrapolate the future
location of their peers – which is known to be present in humans and other animals, has a
large impact on the dynamics of interacting particle models that are often used for describing
collective behaviour. It acts to stabilise the system, which leads to an ordered crystalline
structure, and also induces milling which is a ubiquitous feature of flocking – even in the
absence of self-propulsion.
Using models that incorporate anticipation it would be possible to analyse existing data
on e.g. pedestrian to infer the anticipation time τ that is used by humans [16]. This could
then be compared with anticipation time obtained from other species and other systems,
and could give novel insight into how anticipation influences collective behaviour at different
spatial and temporal scales.
In relation to experimental data we note that the fixed value of τ most likely is a sim-
plification, and that most animals (including humans) make use of a dynamical value of τ
which depends on the distance to the object whose trajectory one is trying to predict. For
larger distance a larger value of τ is beneficial, while for smaller distances one runs the risk
of over-shooting unless τ is chosen sufficiently small. In terms of the interacting particle
systems we have studied here we conjecture that including a dynamic anticipation time will
lead to a faster equilibration of the system dynamics.
A possible extension to the current model is to take into account the fact that an individ-
ual might have knowledge about the predictions made by other individuals. That knowledge
could be captured in an internalised model that would give individual predictions that go
beyond the simple linear extrapolation used in the current model. This would lead to a kind
of second order prediction that might lead to other dynamical behaviours.
It is also possible to consider a model in which individuals have perfect knowledge about
the future state of the system. This implies that x(t+τ ) = x(t)+τ v(t), which when inserted
into eq. (1) (instead of xp(t + τ )) would give rise to a reversed delay-equation. Although
such a system would contradict the principle of causality it could still be interesting from a
theoretical point of view.
The observation that anticipation introduces dissipation and makes it possible for the
system to reach steady states might have implications for our understanding of more com-
24
plicated dynamical systems. Many biological and sociological systems contain an element of
anticipation and models that explicitly take this into account might give novel insights into
the dynamics of these systems.
[1] D. Sumpter, Philosophical Transactions of the Royal Society B: Biological Sciences 361, 5
(2006).
[2] I. Giardina, HFSP J. 2, 205 (2008).
[3] T. Vicsek and A. Zafeiris, Physics Reports 517, 71 (2012).
[4] U. Lopez, J. Gautrais, and I. D. Couzin, Interface . . .
(2012).
[5] I. D. Couzin, J. Krause, R. James, G. D. Ruxton, and N. R. Franks, J. Theor. Biol. 218, 1
(2002).
[6] K. Tunstrøm, Y. Katz, C. C. Ioannou, C. Huepe, M. J. Lutz,
and I. D. Couzin, PLoS
Computational Biology 9, e1002915 (2013).
[7] Y. Katz, K. Tunstrøm, C. C. Ioannou, C. Huepe, and I. D. Couzin, Proc. Natl. Acad. Sci.
108, 18720 (2011).
[8] M. R. D' Orsogna, Y. L. Chuang, A. L. Bertozzi, and L. S. Chayes, Physical Review Letters
96, 104302 (2006).
[9] T. Vicsek, A. Czir´ok, E. Ben-Jacob, I. Cohen, and O. Shochet, Physical review letters 75,
1226 (1995).
[10] R. Nijhawan, Nature 370, 256 (1994).
[11] M. J. Berry, I. H. Brivanlou, T. A. Jordan, and M. Meister, Nature 398, 334 (1999).
[12] G. Schwartz, S. Taylor, C. Fisher, R. Harris, and M. J. Berry II, Neuron 55, 958 (2007).
[13] T. S. Collett and M. F. Land, Journal of comparative physiology 125, 191 (1978).
[14] B. G. Borghuis and A. Leonardo, The Journal of neuroscience : the official journal of the
Society for Neuroscience 35, 15430 (2015).
[15] S. Rossel, J. Corlija, and S. Schuster, Journal of experimental biology 205, 3321 (2002).
[16] I. Karamouzas, B. Skinner, and S. J. Guy, Physical Review Letters 113, 238701 (2014).
[17] A. Morin, J.-B. Caussin, C. Eloy, and D. Bartolo, Physical Review E 91, 012134 (2015).
[18] H. Murakami, T. Niizato, and Y.-P. Gunji, Scientific Reports 7 (2017).
[19] B. Gough, GNU scientific library reference manual (Network Theory Ltd., 2009).
25
|
1607.04190 | 2 | 1607 | 2017-02-23T08:12:11 | Trade-offs between spatial and temporal resolutions in stochastic super-resolution microscopy techniques | [
"physics.bio-ph",
"cond-mat.soft",
"physics.data-an",
"physics.optics"
] | Widefield stochastic microscopy techniques such as PALM or STORM rely on the progressive accumulation of a large number of frames, each containing a scarce number of super-resolved point images. We justify that the redundancy in the localization of detected events imposes a specific limit on the temporal resolution. Based on a theoretical model, we derive analytical predictions for the minimal time required to obtain a reliable image at a given spatial resolution, called image completion time. In contrast to standard assumptions, we find that the image completion time scales logarithmically with the ratio of the image size by the spatial resolution volume. We justify that this non-linear relation is the hallmark of a random coverage problem. We propose a method to estimate the risk that the image reconstruction is not complete, which we apply to an experimental data set. Our results provide a theoretical framework to quantify the pattern detection efficiency and to optimize the trade-off between image coverage and acquisition time, with applications to $1$, $2$ or $3$ dimension structural imaging. | physics.bio-ph | physics | Trade-offs between spatial and temporal resolutions in stochastic super-resolution
microscopy techniques
Jean-Fran¸cois Rupprecht,1 Ariadna Martinez-Marrades,2 Rishita Changede,1 and Gilles Tessier2
1Mechanobiology Institute, National University of Singapore, 5A Engineering Drive 1, 117411 (Singapore).∗
2Sorbonne Universit´es, UPMC Univ Paris 05, Paris (France).
(Dated: September 17, 2018)
Widefield stochastic microscopy techniques, such as PALM or STORM, rely on the progressive
accumulation of a large number of frames, each containing a scarce number of super-resolved point
images. We justify that the redundancy in the localization of detected events imposes a specific
limit on the temporal resolution. Based on a theoretical model, we derive analytical predictions
for the minimal time required to obtain a reliable image at a given spatial resolution, called image
completion time.
In contrast to standard assumptions, we find that the image completion time
scales logarithmically with the image size to spatial resolution volume ratio, which is the hallmark
of a random coverage problem. We discuss how an increased background noise can affect the image
completion time and impact the latter scaling. We finally propose a method to estimate the risk that
the image reconstruction is not complete. Our results provide a theoretical framework to quantify
in real-time the pattern detection efficiency with applications to structural imaging in 1, 2 or 3
dimension.
I.
INTRODUCTION
Optical microscopy is a convenient tool to study biolog-
ical processes, but its resolution is fundamentally limited
by Abbe's diffraction. The image of a point source is
a pattern whose size is comparable to the optical wave-
length (∼ 250 nm), hence source points separated by a
distance smaller than a wavelength are hardly distin-
guishable [1]. Electron microscopy provides a higher spa-
tial resolution (∼ 1 nm) but at the cost of a more complex
sample preparation which is incompatible with in vivo-
imaging [2]. The recently developed super-resolution
imaging techniques aim at combining the best of these
two worlds. Using these techniques, spatial resolution as
low as 10 nm have been achieved for imaging biological
cell structures. However, their applicability to the study
of dynamical biological processes is limited by their long
acquisition times [3, 4].
Though relying on different optical probes, the super-
resolution techniques known as PALM (Photoactivation
Localization Microscopy) or STORM (Stochastic Opti-
cal Reconstruction Microscopy) rely on a common prin-
ciple: sources that lie within the same diffraction-limited
volume are separated by a sequential activation process,
which introduces a temporal separation between source
points [5]. Within each frame, a small and random
fraction of probes is activated by illumination. This
sparse subset of randomly activated probes is imaged
to produce a frame. Then, finding the centroid of each
diffraction patterns leads to a set of coordinates, hav-
ing a nanometer-level precision [1, 6]. Merging all the
single-molecule positions obtained on successive frames
produces the final image. Since only a small fraction of
probes is imaged per frame, a certain number of frames
∗Electronic address: [email protected]
is required in order to obtain a reliable reconstructed im-
age. Multiplying this number by the typical acquisition
time of frames (typically in the 10 − 100ms range), we
obtain the minimal time, denoted T , to obtain an im-
age at a nanometer-scale resolution. A typical reported
value is T ∼ 30 min for a whole cell imaging at a 10 nm
resolution [4]. This value is too large to study many dy-
namical processes that occur in living cells, such as the
contraction of acto-myosin units [7], reorganization of fo-
cal adhesion complexes [8] or protein cluster formation
within the plasma membrane [9, 10].
In addition, stochastic microscopy is prone to local-
ization errors. These errors may either originate from
overlapping spread functions or from emission outside of
the region of interest. However, given a set of localized
observations, one can generally assume that spurious de-
tections corresponds to regions with low count-density,
e.g. the density-based spatial clustering of applications
with noise (DBSCAN) algorithm operates noise filtering
by eliminating observations whose nearest neighbors are
further than a prescribed threshold distance [11]. Here,
we consider the reconstruction criteria in which a region
of space is assumed to belong to the region of interest if
and only if it has collected at least a number r of obser-
vations during the duration of the experiments T . If the
number of observations is insufficient, the reconstructed
structural image displays voids within the region of in-
terest (ROI) -- parts of the ROI are assigned to the back-
ground noise. We refer to these voids as stochastic aber-
ration. This leads to the following two formulations of
the main question of the present paper: What should be
the minimal acquisition time in order to reliably discrim-
inate between the region of interest from the rest of the
field of view? How can we reliably discriminate whether
a hole in the reconstructed super-resolved image is a gen-
uine gap in the structure rather than an aberration due
to a lack of observations?
It is generally thought that in widefield stochastic tech-
7
1
0
2
b
e
F
3
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
0
9
1
4
0
.
7
0
6
1
:
v
i
X
r
a
niques, such as PALM, the imaging time T is solely
controlled by the density (denoted ρ) of activated flu-
orophores per frame and by the spatial resolution (σ)
following the relation T ∼ 1/(ρσ). The latter relation
does not depend on the total size of the field of view
(S), which represents a subtantial advantage of stochas-
tic techniques over deterministic one.
In raster-scan-
based techniques, e.g. STimulated Emission Depletion
(STED), the resolution is exempt of stochastic aberra-
tion but the acquisition time t increases linearly with the
size of the field of view [14].
In this paper, we argue that the imaging completion
time T should be expected to depend on the size of the
field of view, due to the random localization process that
results in an uneven spatial distribution of events. Based
on a stochastic model, we derive the relation:
T ∼
(cid:19)(cid:21)(cid:27)
(cid:18) S
(cid:18) S
+ (r + γD) ln
ln
(cid:26)
(cid:19)
1
σρ
σ/S(cid:28)1
ln
σθ
(cid:20)
,
(1)
σ
which means that the trade-off between the spatial (σ)
and temporal (T ) resolutions depends (i) on the ratio of
the size of the field of view to the desired spatial resolu-
tion (ii) on the necessity to separate the ROI from the
noisy background, via a minimal number of redundant
observations r that is an increasing function of the back-
ground noise intensity, (iii) on the risk of an incomplete
coverage of the ROI (i.e. of stochastic aberrations) via
the 5% -- centile parameter θ = 0.95 and finally (iv) on the
dimensionality D ∈ {1, 2, 3} of the ROI via the constant
γD. The prefactor ln(S/σ) in Eq.
(1) can be signifi-
cantly larger than 1, eg. a cell of extension S = 103 µm2
contains 107 squares of area σ = 10−4 µm2 (ie. a typi-
cal size for an Abel diffraction pattern), which leads to
ln(S/σ) = 16.
The result Eq. (1) applies to experimental situations in
which a high reconstruction fidelity is needed. Obtain-
ing a complete image reconstruction can be of critical
interest in structural reconstructions, e.g. when evaluat-
ing the integrity of a DNA segment [15] or the tenseg-
rity of the actin network within a cell [16].
Indeed, a
broken actin filament cannot support tension, similarly
to a nano-wire which cannot conduct current when it is
cut in two. Mind that the logarithmic scaling with the
image size stems also holds for a near complete image
reconstruction (see Sec. III B). However, in some other
experimental contexts in which a high fraction of missed
pixels in the reconstructed image is tolerable, we show
that the image time should be expected to scale linearly
with size of the sample (see Sec. III G).
We tested the applicability of Eq. (1) on experimental
localization sequences. We find that Eq. (1) is no longer
valid at high level of background noise. This is expected
since Eq. (1) does not hold when the minimal number of
redundant observations r is larger than a critical value
In the regime r (cid:29) rc, we find that the
rc = ln(S/σ).
image time behaves as T ∼ r/(ρσ).
The paper is organized as follows. We first present the
two experimental setups: (i) a PALM setup and (ii) a
2
Total Internal Reflection Microscopy (TIRM) experiment
in which we measure the scattered light from Brownian
nano-particles at the surface of a two-dimensional sam-
ple. We then define two image rendering schemes, called
patch and box-filling methods (see 1). We then prove
the relation Eq. (1) and we show its connexion to the
coupon-collector problem [32 -- 34]. Therefore, we refer to
the result of Eq. (1) as the coupon-collector scaling.
We then consider the robustness of the coupon-
collector scaling for several image completion require-
ments, and in particular the effect of correlations between
successive frames. This case is particularly motivated by
the Brownian setup, in which the escapes and returns of
the Brownian particles within the detection zone leads to
temporally correlated scattering events between succes-
sive frames. We point out that there is a close analogy
between the gold nanoparticle experiments and PALM
techniques relying on organic dyes whose blinking statis-
tics exhibit time-correlations [31]. We recall that bleach-
ing refers to an irreversible transition of a probe to an
inactive state [5]. The analogy holds both on correlated
blinking events -- which corresponds to the correlated re-
turns of the Brownian particles to the illuminated region
-- and on bleaching events -- which corresponds to the es-
cape of the Brownian particle far from the illuminated
region.
We conclude our article by presenting a procedure by
which, in real-time during the acquisition, we can esti-
mate the risk that the image is prone to stochastic aber-
rations.
II. METHODS
A. PALM experiments
We analyzed the sequence of localization events from
two sets of samples: (i) silane sample with quasi-uniform
sampling in fluorophores, and (ii) fibroblast cell with
tagged actin structure.
In particular, we illustrate our
noise removal procedure in Fig. 2, by requiring a min-
imal number of r = 30 events. Details on the PALM
experiments are provided in the SI [47].
B. TIRM experiments
In a recent work [30], we presented a new stochastic
imaging technique to map an electromagnetic field with
a nano-scale resolution using light-scattering Brownian
particles as local probes of the field intensity. The Brow-
nian motion of the scatterers eventually lead to a full
coverage of the imaged field. Following [30], we consider
the imaging problem of an evanescent wave created by a
Total Internal Reflection Microscopy setup. In this setup,
we consider that the optical intensity of the electromag-
3
Figure 1: (Color online) Structural reconstruction by stochas-
tic super-localization microscopy. Probes (colored dots) are
bound to a structure of interest (green line). (a -- b) Circular
patches representation: (a) Upper left inset: Abel diffraction
pattern observed in a CCD camera. The super-resolution al-
gorithm yields a set of coordinate corresponding to the center
of the pattern (black cross). In the patch method represen-
tation, each point coordinate is represented by a disk with a
radius σ that is proportional to the uncertainty of the super-
localization procedure (blue disk).
(b) Patches accumulate
with the acquisition time, eventually covering the whole struc-
ture of interest (patches are represented by different colors for
separate time frames). (c -- f) Box-filling representation, which
leads to a density map in terms of a number of accumulated
events per pixel. (c) The field of view is divided into N = 9
pixels among which F = 5 pixels contain probes. (d) A se-
quence of frames (blue circle: size of the Abel pattern). (e)
Target image. (f) Map of the cumulative number of observa-
tions M (t)
, for each pixel j and for each frame t. Complete
image completion (with r ≥ 1) is obtained after t = 4 frames.
At t = 50, all pixels have been observed at least r = 10 times.
j
netic field can be modelled as
I(x, y, z) = I0(x, y) exp(−z/β(x, y)),
(2)
where β is the penetration length of the field, and I0 is
proportional to the optical intensity of the field at the
surface -- with a proportionality constant related to the
scattering cross section of the particles.
In principle, in most situations of interest, both quan-
tities β and I0 can vary with the location (x, y) on the
surface. In this context, the term image acquisition refers
to the determination of the field intensity I0 and β. How-
ever, as a first test of the method, the experimental data
set from [30] corresponds to a situation where both I0
and β are homogeneous within the whole field of view.
We detail a procedure that leads to the determination of
I0 and β in the SI [47].
Figure 2: (Color online) PALM imaging of the actin mesh
within a fibroblast cell. (a) Image obtained by representing
detection events as circles of radius 5nm (i.e. patch method).
(b) Density plot, in which the field-of-view is divided into
100 × 100 pixels of width x nm. (i.e. box-filling method).
(c-d) Zoom on the cell boundary plot, which reveals three re-
gions defined according to the density of observations:
low
(above red line), intermediate (below black line) and high
(between the two lines). (d) Low density filtering: only pixels
which have collected more than r = 30 observations are repre-
sented, to isolate the region of interest (i.e. actin fibers at the
boundary). Mind that r represents the number of redundant
observation required to separate a region of interest from a
noisy background.
C. Two image rendering methods
Super-resolution techniques rely on the localization of
the center of diffraction spots, which provides a set of
points. However, a spatial extention needs to be at-
tributed to each point to obtain an image that is readable
to the human eye. In the following, we will be interested
in the two following image rendering methods: (i) the
box-filling method (BFM), which is adapted to a den-
sity image representation [24] and (ii) the patch method
(PM), which is associated to a pointillist representation
[25].
The BFM considers the structure of interest as tessel-
lated into F square pixels of equal area, which can there-
fore be expressed as the ratio of the total volume by the
resolution volume: F = S/σ. Each new event falls within
a specific pixel, thereby increasing by one the cumulative
number of observations of this pixel. This method is
naturally adapted to measure the densities. Though we
employ the term pixel in the following, our method also
applies to 3D imaging problems in which F refers to the
number of voxels within the structure of interest [26 -- 29].
σt= 1t= 20ce1000000101100010102200011014120001110120......dfΔx220001110t = 1t = 2t = 3t = 4t = 500r= 1r= 10Patch method Box filling methodabacr= 30outside celloutside fiberscountcount02004006008001000db In the other hand, the PM associates to each event
a surrounding extension, characterized by the quantity
σ, which is either a length (1D), an area (2D) or a vol-
ume (3D). Generally, the spatial extension is chosen to
correspond to the spatial uncertainty associated to the lo-
calization procedure (e.g. a few nanometers, [25]). The
image completion time is related to the minimal number
of patches required to cover the structure of interest.
D. Statistics of events
We assume that fluorescent events are distributed ac-
cording to a homogeneous Poisson process, such that
the probability density dP that an event occurs in an
infinitesimal space of volume ds reads dP = ρds [17].
We now consider a regular domain of volume S within
a D dimensional space, in which we assume a constant
density of fluorophore d. Furthermore, we assume that
at each frame, only a fraction f of fluorophores are de-
tected. The number of detected fluorescence events after
one frame, denoted N (1), is a Poisson process of den-
sity ρ = f d; hence P(cid:2)N (1) = n(cid:3) = exp(−ρS)(ρS)n/n!.
If S = A refers to the volume of the Abel diffrac-
tion pattern, the mean number of fluorescence events
per frame ρA should be lower than 1 in order to limit
the risk of overlapping point spread functions. Typi-
cally A = (102 nm)D [5], hence ρ < 10−2Dnm−D.
In
the case of membrane (D = 2) with fluorophore den-
sity d = 104 µm−2, the corresponding maximal fraction
of activated fluorophores should be f < 10−3. After a
number T of frames, the total number of collected events
is distributed according to a Poisson distribution, with
P(cid:2)N (T ) = n(cid:3) = exp(−ρST )(ρST )n/n!.
We finally assume that the density of events ρ is time-
independent, hence neglecting the progressive bleaching
of fluorophores [31]. Our time-independent assumption
corresponds to two situations, in which either (i) the to-
tal number fluorophores per elementary resolution vol-
ume remains large compared to the number bleached flu-
orophores, or (ii) if the activation laser is increased as a
function of time in order to balance the effect of bleach-
ing.
E. Estimation of the structure size
In many situations, the volume of the structure is ei-
ther completely unknown a priori, or can only be par-
tially inferred - e.g. by assuming randomly oriented lin-
ear order. Within the BFM, we show (see SI, Sec. B 1
[47]) that the maximum likelihood estimator of the num-
ber of relevant pixels F corresponds to the quantity:
4
j
where M (t)
is the cumulative number of measures of the
pixel j, e.g. M (t)
j = 0 if the pixel j has never collected
j ≥ 1 if the pixel has been
any event up to time t and M (t)
observed at least once up to time t (see Fig. 1). Simi-
larly, within the patch-method framework, the maximum
likelihood estimator of the structure volume consists in
the covered volume at the time t. These two estimators
are biased, as they tend to underestimate the structure
volume.
F. Mathematical definition of the image
completion time
We call image completion time the minimal number of
frames required to obtain a complete image of the region
of interest. The term complete refers to the condition
that every pixel or point (among those that should be
observed) has been covered at least a certain number of
times, denoted r ≥ 1. More precisely, the image com-
pletion time T is the random variable (called stopping
time) that corresponds to the minimal time t such that
= r; where j ∈ [1, . . . F ] in the BFM frame-
minj
work, or j refers to any point within the volume of inter-
est in the PM framework. We will be mainly interested
in the centile of T , denoted tθ and defined as:
M (t)
(cid:16)
(cid:17)
j
(cid:20)
(cid:16)
(cid:21)
(cid:17) ≥ r
P [T ≤ tθ] = P
min
j
M (tθ)
j
= 1 − θ,
(4)
where θ is the tolerated risk. To summarize, the quantity
t0.05 refers to the minimal number of frames that guar-
antees, with 95% probability, that there is no stochastic
aberration within the reconstructed ROI image.
G. Simulations
Both in the BFM and PM frameworks, the volume of
the region of interest is tessellated into a grid of elemen-
tary squares.
In the BFM, each event covers a single
elementary square; while in the PM, each patch σ covers
a square matrix of elementary squares. In both frame-
works, we generate a large sample of coverage events and
we analyse the resulting distribution of coverage times
using Matlab's prctile function.
III. RESULTS AND DISCUSSION
A. The image completion time follows a
coupon-collector scaling
(cid:98)F (t) =
F(cid:88)
j=1
(cid:16)
(cid:17)
min
M (t)
j
, 1
,
(3)
We first derive the main result of Eq. (1) in the case
of the BFM with no time-correlation between frames.
Here, we assume that the value of the total number of
pixels F is known. Under the assumption that detection
events occurring in separate pixels are independent, the
probability that exactly M pixels have been observed at
least once (r = 1) reads:
(cid:104)(cid:98)F (t) = M
(cid:105)
P
=
(cid:18) F
(cid:19)
M
(cid:0)1 − pt
0
(cid:1)M
p(M−F )t
0
,
(5)
where p0 = 1 − p1, and p1 = ρσ is the probability that
an event occurs in a given pixel and at a given frame.
In particular, the probability that the estimator (cid:98)F (t) is
equal to its target value F reads:
(cid:104)(cid:98)F (t) = F
P
.
(6)
(cid:105)
=(cid:0)1 − (1 − p1)t(cid:1)F
(cid:104)(cid:98)F (tθ) = F
(cid:19)
(cid:18) F
(cid:18) S
(cid:105)
(cid:19)
tion time, defined according to Eq.
lution of the equation P
We now determine the centile of the image comple-
(4) as the so-
= 1 − θ, hence
tθ = ln(cid:0)1 − (1 − θ)1/F(cid:1) / ln (1 − p1). In the limit p1 (cid:28) 1
and for sufficiently high centiles (θ < 0.1), we find that
the centile of the imaging time reads:
tθ ∼
1(cid:28)F
F
µ
ln
θ
=
ln
1
ρσ
θσ
,
(7)
where µ = F p1 is the mean number of observations
per frame. The latter expression corresponds to the an-
nounced Eq. (1) with r = 1.
A key feature shared by Eqs. 1 and 7 is the non-linear
dependence of the imaging time in terms of the number
F of pixels that characterize the structure. This scal-
ing is related to the classical coupon-collector problem
[32 -- 34]. The problem consists in buying a minimal num-
ber of the boxes (each containing a random coupon) in
order to gather a complete collection of coupons, with
a sufficiently high probability. Here, we focus on the
case where each box contains, at random, either 0 (with
probability p0) or 1 coupon -- in which case the mean
number of coupons per box is equal to µ = 1 − p0. A
straightforward proof leads to the following exact ex-
pression for the mean number of bought boxes t (i.e.
frames) required to collect all coupons (i.e. all pixels)
is E [T ] = F (1 + 1/2 + . . . + 1/F )/µ. If the number of
coupons F is large, the latter expression takes the asymp-
totic form E [T ] = F ln(F )/µ. Adapting the identity (2)
of Ref.
[32], one shows that the centile of the stopping
time reads tθ = (F/µ)×ln(F/θ) in the same limit F (cid:29) 1,
which corresponds to Eq. (6) after identification of the
mean number of coupons per box to the mean number of
events per frame.
Mind that Eq. (7) weakly depends on the risk level θ,
which is another characteristic property of the coupon-
collector problem [32 -- 34].
Furthermore, the expression for the image completion
time in Eq. (7) only depends on the number of pixels but
not their spatial organization, e.g. on the 1D, 2D or 3D
nature of the structure. This is expected since pixels are
considered to be independent.
5
B. The coupon-collector scaling holds for a near
complete coverage
A simple argument shows that Eq. (1) holds even in
the near total coverage, ie. when the final image should
contain a significant fraction (e.g. 90%) of the total num-
ber of pixels within the ROI. Consider that a single pixel
i is missing after t frames. The additional number of
frames ∆t that is required to find the missing pixel i is of
the order of the total number of pixels, ie. ∆t ∝ F . This
increment is small compared to the total completion time
T = Tnear−complete + ∆t ≈ F ln(F ). Provided that the
missing fraction of pixels is small, the near-completion
time Tnear−complete is approximatively equal to the com-
pletion time T .
C. The coupon-collector scaling holds when
redundant observations per pixel are required
To distinguish relevant observations from spurious
ones, we consider that a pixel should collect a min-
imal number of observations denoted r to be consid-
ered as being part of the region of interest. Assum-
ing that all pixels within the ROI are equivalent, the
probability that all pixels have collected at least r ob-
servations can be expressed in terms of the probability
that the pixel 1 have collected at least r observations
(cid:104)(cid:98)F (t) = F
(cid:105)
(cid:16)
(cid:17)F
1 ≥ r)
as P
=
P(M (tθ)
. We show in the
SI B 2 a [47] that, in this case, the centile of the image
completion time reads
tθ ∼
1(cid:28)F
F
µ
{ln (F/θ) + (r − 1) ln (ln F )} .
(8)
The latter relation corresponds to the centile of the
coupon collector's problem when r copies of each coupon
need to be collected (see [32, 37]).
We emphasize that Eq. (8) requires that the required
number of coverage r is sufficiently small, ie. that r (cid:28)
ln(F ). We also generalize the result of Eq. (8) to a multi-
color imaging problem (see SI B 3, [47]).
D. The coupon-collector scaling holds in the
presence of spatial inhomogeneities
In this section, we discuss the case of a non-
homogeneous rate of activation, which is particularly
important in PALM (see Fig 2). We model the
non-homogeneity of detection events by assuming that,
among pixels, the probability p0 is distributed according
to a probability distribution ψ(q). Under this assump-
tion, the probability that there has been more than r
observation in a particular pixel i reads:
P(M (t)
i ≥ r) =
dq P(Mi ≥ rp0,i = q)ψ(q).
(9)
(cid:90) 1
0
where P(M1 ≥ rp0,i = q) is given in 5 (see also SI, Eq.
(B3), [47]). From the expression in Eq. (9), we numer-
ically solve the relation Eq. (4) to obtain the imaging
time tθ.
In the Experimental Comparison section IV, we con-
sider a model in which the probability ψ is Gaussian dis-
tributed. We find that the coupon-collector logarithmic
scaling still holds in that case. However, the precise dis-
tribution of spatial hetereogeneities is needed in order to
obtain a quantitative fit to the experimental data (see
Fig. 5).
in the structure? We identify P () as being equal to the
empty-space distribution defined in [39], hence we find
that
P () = 1 − exp(−ρt D/Ω),
(10)
where Ω = πD/2/Γ [1 + D/2] is the volume of a sphere of
radius 1. We expect Eq. (10) to hold within the BFM
framework, hence providing the probability that a con-
nected set of N = /σ missing pixels corresponds to a
genuine hole.
6
E. The coupon-collector scaling holds with the
patch image-rendering method
We now consider that the image results form the ac-
cumulation of circular patches, whose radius σ corre-
sponds to the spatial resolution. The patch centers are
distributed according to a homogeneous Poisson distri-
bution within the region of interest, of volume S.
The study of coverage problem has a long history
[38, 39]. However, analytical results concerning coverage
problems in two dimensions are rather recent [40, 41].
These studies were motivated by the study of the wifi
coverage resulting from randomly located routers. We
will make use of results concerning the expression of the
centile nθ of the number of patches required to cover a
circle [38] or a square [40] by circular patches.
Here, we seek an expression of the centile time tθ, i.e.
a time expressed in terms of a number of frames t, rather
than the centile time expressed in terms of the number
of patches n. We expect that tθ = nθ/µ where µ is
the number of events per frame.
Indeed, in the small
patch limit σ/S (cid:28) 1, full coverage events occur when
the number of events is large (n (cid:29) 1) in which case the
number of events is simply proportional to the number
of frames t. This approximation is further justified in
the SI [47]. Therefore, following Refs.
[38] and [40], we
find that Eq.
(1) corresponds to the time required to
obtain a r-fold coverage of a D-dimensional ROI of total
volume S by circular patches of volume σ. In particular,
we obtained that γ1 = 0 in 1D (following [38]) and that
γ2 = 2 in 2D (following [40]) and finally that γ3 = 3.
Remarkably, Eq.
(1) takes a similar form as the
coupon-collector problem from Eq.
(8). This similar-
ity suggests that in the limit σ/S (cid:28) 1, regularly spaced
patches of size σ/S behave as if they were independent.
Mind, however, that the expression from Eq.
(8) cor-
responds to a value γD = −1 for any space dimension.
The origin of this discrepancy at second order in the ra-
tio σ/S (cid:28) 1 is discussed in Ref. [38]. In conclusion, we
have shown that both the PM and BFM lead to similar
expressions for the image completion time.
Another interest of the PM representation is that we
can answer the following question: after a time t has
elapsed, what is the probability P () that a hole of size
in reconstructed image corresponds to a genuine gap
F. The coupon-collector scaling holds in the
presence of correlations between frames
In the Brownian scatterers experiments, the gold par-
ticle may enter, escape or return within the field of
view, leading to correlated observations between succes-
sive frames. In contrast to the discussion leading to Eq.
(6), these temporal correlations invalidate the indepen-
dence hypothesis that allows to factorize the final time
probability distribution. We encompass these correlated
observations through the following box-filling model, in
which the number of events per pixel and per frame is
assumed to be a random variable K with a general prob-
ability law pk = P(K = k) for all k ≥ 0. The statis-
tics of K encompass the effect of time-correlated obser-
vations by neglecting the time between successive corre-
lated events. Comparison of this model to experiments
is satisfactory, as visible in Figs. 3(e) and (f), in which
we represent the experimental data from Ref.
[30] and
simulated evolutions of the cumulative number of events
M (t)
.
observation per pixel per frame as ν = (cid:80)∞
σ2 =(cid:80)∞
j
We define the mean and variance of the number of
k=1 kpk and
k=1 k2pk − ν2, respectively. We assume that the
set of probabilities pk, k ≥ 0 is identical for each of the
pixels of the structure to be imaged. In SI B 2 a [47], we
(cid:19)(cid:21)(cid:27)
show that the imaging completion time reads
(cid:26)
(cid:19)
(cid:20)
tθ ∼
σ(cid:28)S
(1 − p0)−1
ln
(cid:18) S
σθ
(cid:18) S
σ
+ (r − 1) ln
ln
,
(11)
provided that p1 (cid:54)= 0. Mind that Eq. (11) differs from
Eq. (7) due to the prefactor 1 − p0, which is determined
by the precise statistics of K and may significantly differ
form the value of µ/S. In particular, at a constant total
mean number of events per frame µ, an increase in the
mean number of correlated events ν also increases the
imaging time.
We conclude that, temporal correlations can signifi-
cantly affect the value of the image completion time, yet
without affecting the coupon-collector scaling of the im-
age completion time.
7
different pixels have been acquired. We find that the
probability defined in Eq. (5) is maximal after a number
of frames topt(M ) ∼ M/(νF ) in the limit µ/F (cid:28) 1 and
M/F (cid:28) 1. Hence, the image completion time is pro-
portional to M , with no logarithmic dependence on the
parameters.
Secondly, when the ratio of structure signal to the
background noise is weak, a large number r (cid:29) F of re-
dundant observations per pixel is required; we show that
the coupon collector scalings from Eqs. 8 and 1 does not
hold in this limit. Indeed, due to the central limit the-
orem, the number of observations collected in the pixel
j eventually converges with t towards a Gaussian dis-
tribution: M (t)
previous section, µ and Σ2 are the mean and variance
of the number of observation per frame within the total
field of view (which, in principle, can be different from
ν and σ2 in the presence of temporally correlated noise).
Under the Gaussian assumption, we find that the proba-
bility distribution of the image completion time T reads:
j ∼ N(cid:0)tµ/F, tΣ2/F(cid:1) , where, as in the
(cid:40)
1 − erf
(cid:33)(cid:41)F
(cid:32)
r − µt/F
(cid:112)2Σ2t/F
P [T ≤ t] = 2−1/F
where erf(x) = (cid:82) x
√
−∞dt exp(−t2)/
π is the error func-
tion [44]. In the limit of a large number of observations
r (cid:29) ln(F ), the expansion of the error function around 0
provides the following approximate expression:
,
(12)
(cid:115)
(cid:19)
(cid:18) F
√
2
2πθ
tθ ∼ F r
µ
+
2rΣ2
µ
log
,
(13)
in the limit r (cid:29) ln(F/µ). The key feature from Eq.
(13) is that the image completion time tθ does not follow
the coupon-collector scaling. Similarly to a deterministic
imaging techniques, the image time scales linearly with
F - which should be expected since the effects of the
localization randomness are all the more averaged out
that the required redundancy per pixel is large.
To conclude, we have obtained analytical results for
the image completion time problem in the limits r (cid:28)
ln(F/µ) and r (cid:29) ln(F/µ) while we resorted to numerical
simulation to describe the intermediate regime.
IV. COMPARISON TO EXPERIMENTS
A. Comparison to PALM experiments
We analyzed the sequence of localization events from
a silicon wafer with quasi-uniform coating. The non-
uniformity in the fluorophore density leads to hetero-
geneities in the value of p0, i.e. the probability per frame
that a pixel does not collect any observation (see Fig.
4). We fit the distribution p0 by a Gaussian distribution
ψ(p0) = N exp((p0 − νp)2/(2σ2
p)), where N corresponds
to the normalization over p0 ∈ [0, 1].
3
Figure 3:
(Color online) Schematic view of the stochastic
Brownian TIRM technique and its modeling. (a) Global view
of the total field of observation, divided into F pixels (BFM
framework).
(b) Zoom on pixel 3, including a particle la-
belled by the frame instant. After t ≥ 6, the probe is not
detected again. (c-d) Scheme of the evolution of the cumu-
lative number of observations M (t)
occuring in pixel 3 (c) as
seen experimentally (d) as represented in our model, where all
correlated observations are collapsed into one single instanta-
neous events.
(e-f) Evolution of the cumulative number of
observations obtained, either (e) from the experimental data
set of Ref [30], or (f) from Monte-Carlo simulations, with the
fitted jump distributed. The different colors correspond to
different pixels (4 among F = 100).
(g-h) Distribution of
jump length pk ∝ exp(−k/kc) (in log-scale): (g) from experi-
ments (blue circles), where the maximum likelihood estimator
of the exponential model [46] provides the value kc = 2.9±0.1
(black line); (h) from simulations, with kc = 2.9 and the same
number of 2792 simulated events.
G. Situations in which the coupon-collector scaling
does not hold
First, when a small subset of observation is sufficient
to reconstruct the image, the coupon-collector scaling
should not be expected. This may include situations in
which the structure can be inferred, e.g. by assuming
randomly oriented linear shapes [16]. Consider that the
image is considered to be complete as soon as M (cid:28) F
1234573experiments061256112312567abcd01000200002040600100020000204060modelexperimentsmodelefttttM(t)jM(t)3M(t)3M(t)jj= 2j= 4j= 2j= 4gh 8
Figure 5: (Color online) Centile tθ of the image completion
time as a function of the required number of events per pixel
in the regime r ≥ ln(F ). (a) Simulations with at most one
observation per pixel, F = 15 and p1 = 0.1: (red cross) centile
from stochastic simulations; (solid green line) approximate so-
lution from Eq. (13); (blue circles) exact centile time obtained
by numerical inversion of Eq. (12). (b) TIRM experiments
from Ref. [30], with F = 4: (red bars) centile estimation from
experiments, where the error is estimated by bootstrapping
[43]; (black dots) theoretical prediction from Eq. (13).
The experimental histogram is fitted by the distribution
pk = Akc exp(−k/kc), where Akc = 1/(1 − exp(−1/kc)),
and kc = 2.9 (see Fig. 3.g.). This leads to a mean
jump size ν = 1/(1 − exp(−1/kc)) = 3.4 and a vari-
ance σ2 = 1/(cosh(1/kc) − 1) = 17. We check that our
results weakly depend on the specific value attributed to
the separation time ∆.
As described in Figs. 4a and 5b, we show that our the-
oretical expressions from Eqs. (11) and (13) both fit to
the experimental estimation of the centile time in their
respective validity range. We point out that a straight-
forward implementation of Eq. (8), which would neglect
temporal correlations, leads to a value that is an order of
magnitude lower than what is experimentally observed.
In the SI [47], Sec. D, we justify that a large number
of redundant observations r ≈ 4 · 103 is required in order
to obtain a reliable measure of the penetration length β.
V. REAL-TIME ESTIMATION OF THE RISK
OF STOCHASTIC ABERRATION
Experimentally, the two quantities F and µ are un-
known a priori. These quantities are indeed associated
to the structure to be imaged, whose properties are un-
known prior to imaging. Here, we propose a real-time
procedure to determine whether we can safely consider
that the image is complete. We emphasize that this pro-
cedure is not specific to a choice of image representation
method, nor on the required number of redundant obser-
vations r.
based on the two estimators (cid:98)F (t) and (cid:98)µ(t) of number of
We evaluate the probability that the image is complete
pixels and of the mean number of events per frame, re-
spectively. For example, in the BFM with r = 1, the
estimator of the image completion probability reads:
.
(14)
(cid:104)(cid:98)F (t) = F
(cid:105)
(cid:98)P
(cid:16)
1 − (1 −(cid:98)µ(t)/(cid:98)F (t))t(cid:17)(cid:98)F (t)
=
Figure 4: (Color online) Centile t0.05 of the image comple-
tion time as a function of the required number of redundant
observations per pixel r. (a) Simulations with at most one ob-
servation per pixel and per frame, F = 15 and p1 = 0.1. The
analytical expression from Eq. (8) (solid blue line) provides
a better fit of simulations (red error bars, obtained by boot-
strapping [43]) than the approximate solution from Eq. (13)
(black circle) . (b) Comparison of our theoretical expression
to the TIRM experiments from [30], with F = 15: (red er-
ror bars) centile estimation from the experiments; (blue solid
line) theoretical prediction from Eq. (8) with a jump prob-
ability distribution pk ∝ exp(−k/kc) with kc = 2.9; (green
crosses) stochastic simulations. (c-d) PALM imaging of a sili-
con wafer with quasi-uniform coating in fluorophores (the ROI
corresponds to the whole field of view, with N = F = 100).
(c) Non-uniform fluorophore density leads to spatial hetero-
geneities in p0, i.e.
the probability per frame that a pixel
does not collect any observation: the experimental distribu-
tion (blue boxes) is fitted by a Gaussian (dashed red line).
Inset: field of view in terms of the number of collected ob-
servations per pixel. (d) Centile t0.05 for (blue and red) two
distinct samples with identical concentration of fluorophores:
(solid lines) analytical expression and (error bars) experimen-
tal result.
Based on the estimation of probability distribution
from Eq.
(9), we find our predicted centile time is in
quantitative agreement with the analysis of two experi-
mental data sets. We point out that taking into account
the spatial heterogeneities in p0 is required to obtained
the quantitative fit represented Fig. 4d.
B. Comparison to TIRM experiments
We represent the TIRM experiments data from Ref.
[30] within the BFM framework and we include tempo-
ral correlations between frames. First, the mean num-
ber of particles per frame and over the whole field of
view reads µ = 0.70. Secondly, the jump distribution is
estimated as follow: two successive events are assumed
to correspond to the return of the same particle if (i)
they occur within the same pixel and (ii) they are sep-
arated by a time interval of less than ∆ = 5 frames.
abSimulationsTheoryGaussian Th.TheoryExperimentscd0.810.9CountTheoryExperiments0106010p0ExactSimulationsApproximateab 9
Fig. 6b,d. After t = 300 frames, we estimate that about
420 additional frames are required, which is consistent
with the theoretical value of the centile time tθ = 760.
Based on the estimators(cid:98)P[(cid:98)F (t) = F ] and (cid:98)tθ
, we pro-
pose the following procedure to analyse an imaging ex-
periment in which t frames have been collected:
(t)
1. Compute the estimators of the number of pixels
((cid:98)F (t)) and of the mean number of events per frame
((cid:98)µ(t)).
2. Compute the estimator of the probability that the
image is complete. If this estimator is higher than
a desired confidence threshold, the imaging process
can be stopped.
tion time (cid:98)tθ
and return to step 1 with the substitution t ← (cid:98)tθ
3. Otherwise, compute the estimated image comple-
(t) − t additional frames
.
. Perform (cid:98)tθ
(t)
(t)
The above procedure is not specific to any particular
criteria for the image completion. For example, if a large
redundancy is required (r (cid:29) ln(F )), one should use the
expressions of Eq. (12) for the probability that the im-
age is complete and Eq. (13) for the image completion
time. In the SI C 3 [47], we provide an expression for the
probability that the image is complete within the PM
framework.
Conclusion Our theoretical model provides a unified
framework to describe the temporal resolution of sev-
eral types of stochastic microscopy techniques. These
include PALM, in which a large number of fluorescent
probes are attached to the sample and are stochastically
activated, or techniques in which a smaller number of
scattering probes stochastically explore the imaged re-
gion. We derive analytical expressions for the centile of
the imaging time for several types of image completion
criteria. When a sufficiently low number of accumulated
events per pixel are required, the temporal resolution is
shown to be logarithmically coupled to the spatial resolu-
tion (pixel size), due to the spatial redundancy of detec-
tion events. However, the temporal resolution becomes
linearly coupled to the spatial resolution when a large
spatial redundancy of events is needed, as the effects of
the localization randomness are averaged out. Our re-
sults on the imaging time are readily applicable to esti-
mate the minimal time required to reliably characterize
spatial patterns by stochastic imaging, with applications
ranging from the detection of protein clusters by PALM
[9] to the detection of the electromagnetic field around
nano-antennas by Brownian particles [30].
Supplementary material Electronic supplementary
material is available at XXX or via YYY.
(t)
after a number t = 50 of observations (light orange) and t =
300 (dark blue); (red vertical line) the limit value is F = 100.
Figure 6: (Color online) Numerical simulation of the real-time
imaging method (Sec. V), where the total number of pixels
is F = 100, the probability of an event observation per frame
and per pixel is p1 = 10−2, averaging is performed over 104
after a
number t = 50 of observations (light orange) and t = 300; (red
vertical line) the limit value is t0.95 = 760. (c) Evolution of the
samples. (a) Histogram of the values of the estimator (cid:98)F (t)
(b) Histogram of the values of the estimator (cid:100)t0.05
mean value of the estimator (cid:98)F (t) (solid blue curve), together.
We also represent the probability P ((cid:98)F (t) = F ) for the image
experiment) is sufficient to obtain a good estimate of (cid:98)F (t) and
P ((cid:98)F (t) = F ). (d) Probability distribution for the estimator
(cid:100)t0.05
(red vertical line) as t increases. The value (cid:100)t0.95
completion (solid magenta curve). In both cases, error bars
indicate the standard deviation estimated from the random
sampling. Hence a single random realization (i.e. a single
. The distribution converges to the centile t0.95 = 760
= 264 is
significantly larger than the current number of frame t = 50:
this is consistent with the conclusion that more observations
are required.
(50)
(t)
We represent the evolution of the estimated probability
corresponding to Eq. (14) in Fig. 6c. We set the values
to F = 100 and p1 = 10−2. At t = 300 the image
completion probability is lower than 4.10−3: hence more
frames are needed. The question is now to determine
how many additional number of frames are required.
Our analytical expression of the image completion
time can then be used to infer the required additional
=
number of frames. For example, the quantity (cid:98)tθ
((cid:98)F (t)/(cid:98)µ(t)) ln((cid:98)F (t)/θ) is an estimator of the image com-
pletion time, where we consider the BFM with r = 1 for
simplicity. We represent the convergence of the latter
estimator to the expected value of the centile time tθ in
(t)
abcdFFF Authors contributions A. M. M. and G. T. carried
out the gold nano-particles experiments and localization
analysis, and instigated the theoretical problem. R. C.
carried out the PALM experiments and localization anal-
ysis. J.-F. R. performed the theoretical calculations, sim-
ulations, centile time analysis of the experiments and
wrote the manuscript. We have no competing interests.
Acknowledgements We thank Xu Xiaochun (MBI Mi-
croscopy core) for designing the localization code of the
PALM setup, V. Studer, M. Coppey and B. Hajj for en-
lightening discussion on the PALM technique, and S. Tlili
for comments on the manuscript.
10
[1] C. W. McCutchen, Journal of the Optical Society of
[25] A. Triller and D. Choquet, Trends in Neurosciences 28,
America 57, 1190 (1967).
133 (2005).
[2] A. J. Koster and J. Klumperman, Nature reviews. Molec-
ular cell biology Suppl, SS6 (2003).
[3] E. Betzig et al.Science 313, 1642 (2006).
[4] Z. Liu, L. D. Lavis, and E. Betzig, Molecular cell 58,
644 (2015).
[5] L. Schermelleh, R. Heintzmann, and H. Leonhardt, The
Journal of Cell Biology 190, 165 (2010).
[6] E. Betzig, Optics Letters 20, 237 (1995).
[7] H. Wolfenson, G. Meacci, S. Liu, M. R. Stachowiak,
T. Iskratsch, S. Ghassemi, P. Roca-Cusachs, B. OS-
haughnessy, J. Hone,
and M. P. Sheetz, Nature Cell
Biology (2015), 10.1038/ncb3277.
[8] C. Bertocchi, W. I. Goh, Z. Zhang, and P. Kanchana-
wong, Critical Reviews in Biomedical Engineering 41,
281 (2013).
[26] M. F. Juette, T. J. Gould, M. D. Lessard, M. J.
Mlodzianoski, B. S. Nagpure, B. T. Bennett, S. T. Hess,
and J. Bewersdorf, Nature Methods 5, 527 (2008).
[27] G. Shtengel, et al., Proceedings of the National Academy
of Sciences of the United States of America 106, 3125
(2009) .
[28] B. Hajj, J. Wisniewski, M. El Beheiry, J. J. Chen,
A. Revyakin, C. Wu, and M. Dahan, Proc Natl Acad
Sci U S A 111, 17480 (2014).
[29] R. Galland, G. Grenci, A. Aravind, V. Viasnoff,
V. Studer, and J.-B. Sibarita, Nature methods 12, 641
(2015).
[30] A. Martinez-Marrades, J.-F. Rupprecht, M. Gross, and
G. Tessier, Optics express 22, 29191 (2014).
[31] P. Annibale, S. Vanni, M. Scarselli, U. Rothlisberger,
[9] R. Changede, X. Xu, F. Margadant, and M. P. Sheetz,
and A. Radenovic, Nature Methods 8, 527 (2011).
Developmental Cell , 1 (2015).
[10] K. H. Biswas et al.Proceedings of the National Academy
of Sciences of the United States of America 112, 10932
(2015).
[11] M. Ester, H.-P. Kriegel, J. Sander, X. Xu, et al., in Pro-
ceedings of the 2nd International Conference on Knowl-
edge Discovery and Data mining , 226 (1996).
[12] G. Shtengel, Y. Wang, Z. Zhang, W. I. Goh, H. F. Hess,
and P. Kanchanawong, Methods in Cell Biology, 1st ed.,
Vol. 123 (Elsevier Inc., 2014) pp. 273 -- 294.
[13] I. Schoen, J. Ries, E. Klotzsch, H. Ewers, and V. Vogel,
Nano letters 11, 4008 (2011).
[14] K. I. Willig, R. R. Kellner, R. Medda, B. Hein, S. Jakobs,
and S. W. Hell, Nature Methods 3, 721 (2006).
[15] M. Alexeyev, I. Shokolenko, G. Wilson, and S. Ledoux,
Cold Spring Harb Perspect Biol. , 5(5): a012641 (2013).
[16] Z. Zhang, Y. Nishimura, and P. Kanchanawong, Molec-
ular biology of the cell , E16 (2016).
[17] P. Hall, Introduction to the theory of coverage processes
[32] P. Erdos and A. R´enyi, Magyar Tudom´anyos Akad´emia
Matematikai Kutat´o Int´ezet´enek Kozlem´enyei 6, 215
(1961).
[33] W. Feller, Wiley Series, Vol. 2 (1968) p. 509.
[34] R. P. Stanley and H. S. Wilf, The American Mathematical
Monthly, Vol. 97 (1990) p. 864.
[35] A. Orlitsky, N. P. Santhanam, and J. Zhang, Science
(New York, N.Y.) 302, 427 (2003).
[36] S. Cox et al.Nature Methods 9, 195 (2011).
[37] D. J. Newman, Science (New York, N.Y.) 98, 104 (1943).
[38] L. Flatto, Israel J. Math. 15, 167 (1973).
[39] H. Solomon, Geometric Probability (SIAM, 1978) p. 174.
[40] H. Zhang and J. C. Hou, The Proceedings of the 5th
ACM international symposium on Mobile ad hoc net-
working and computing , 121 (2004).
[41] G. L. Lan, Z. M. Ma, and S. S. Sun, "Discrete Geome-
try, Combinatorics and Graph Theory: 7th China-Japan
Conference," (Springer Berlin Heidelberg, Berlin, Hei-
delberg, 2007) Chap. Coverage P, pp. 88 -- 100.
(John Wiley & Sons Australia, Limited, 1988) p. 408.
[42] H. Cang, A. Labno, C. Lu, X. Yin, M. Liu, C. Gladden,
[18] C. E. Shannon, Proceedings of the IEEE 86, 447 (1998).
[19] H. Shroff, C. G. Galbraith, J. a. Galbraith, and E. Bet-
zig, Nature methods 5, 417 (2008).
[20] S. van de Linde, A. Loschberger, T. Klein, M. Heidbreder,
S. Wolter, M. Heilemann, and M. Sauer, Nature proto-
cols 6, 991 (2011).
[21] I. Izeddin, C. G. Specht, M. Lelek, X. Darzacq, A. Triller,
C. Zimmer, and M. Dahan, PLoS ONE 6, e15611 (2011).
[22] U. Endesfelder, S. van de Linde, S. Wolter, M. Sauer,
Y. Liu, and X. Zhang, Nature 469, 385 (2011).
[43] B. Efron, The Annals of Statistics 7, 1 (1979), .
[44] I. M. Ryzhik and I. S. Gradstein, Tables of Series, Prod-
ucts and Integrals (1957) p. 438.
[45] H. Shroff, C. G. Galbraith, J. A. Galbraith, H. White,
J. Gillette, S. Olenych, M. W. Davidson, and E. Betzig,
Proceedings of the National Academy of Sciences of the
United States of America 104, 20308 (2007).
[46] V. Rivoirard and G. Stoltz, Statistique math´ematique en
and M. Heilemann, ChemPhysChem 11, 836 (2010).
action (Vuibert, 2012) p. 448.
[23] B.-C. Chen et al. Science 346, 1257998 (2014) .
[24] S. Cox, E. Rosten, J. Monypenny, T. Jovanovic-
Talisman, D. T. Burnette, J. Lippincott-Schwartz, G. E.
Jones,
and R. Heintzmann, Nature Methods 9, 195
(2011).
[47] See Supplemental Material [url]. .
Appendix A: Description of the PALM experiments
1. Materials and methods
Silane
functionalization Biotinylated
silane
(Methoxy Silane PEG biotin) was mixed with Methoxy
Silane in concentration equivalent to have a final concen-
tration of 103 molecules/µm2 of Biotin on the surface.
Clean cover slips were coated with this silane mixture
using vaporization under vacuum. Silane functionalized
cover slips were then washed with PBS and incubated
with Dylite650 Neutravidin for 1 hour followed by a
subsequent wash before PALM imaging.
Cell Culture and sample preparation Mouse embry-
onic fibroblasts were grown in DMEM media containing
1mM sodium pyruvate and 10% fetal bovine serum at
37◦ C with 5% of carbon dioxyde. Cell were spread for
4 to 6 hours on a fibronectin coated glass dish. Spread
fibroblasts were then were fixed with 4% formaldehyde
at 37◦ C for 10 minutes followed by mild detergent
permeablization with 0.5% Triton X 100 for 15minutes.
Alexa647- Phalloidin was used to label actin (125nM for
1 hour immediately prior to imaging). Multiflurophore
beads of 0.14 µm diameter (Spherotech, Cat no. FP0257-
2) were added as fiducial markers for PALM imaging.
2. Photoactivated light microscopy (PALM)
Fresh imaging buffer was made for every sample. Imag-
ing buffer contained we oxygen scavenging imaging buffer
constituting of the following solutions with a volume ra-
tio of 90:10:1.
(1) 50 mM Tris-HCl (pH 8.0), 10 mM
NaCl, 10(2) 1M mercaptoethylamine with pH adjusted
to 8.5 using HCl (3) Anti-bleaching oxygen scavenger
system containing 14 mg Glucose Oxidase, 50 l Cata-
lase (17 mg/ml) in 200 µL 10mM Tris-HCl (pH 8.0), and
50 mM NaCl DPBS solution. The samples were sealed
with parafilm to prevent air exchange.
PALM imaging was performed using Ziess Elyra. 100X
objective (Alpha Plan Apochromat 100X oil NA 1.46,)
with 1.6 magnification to have a final pixel size of 100 nm
by 100 nm was used. The camera on the system is Andor
iXon DU897 512x512 electron multiplier CCD camera.
A total of 2 · 104 images were collected with continuous
streaming at 50 ms per frame for each sample. PALM im-
ages were reconstructed using a custom-made maximum
likelihood software [9].
Appendix B: Image completion time within the
box-filling framework
1. Maximum likelihood estimator of the number of
pixel in the PALM
We consider the result of a particular simulation or ex-
periment in which the cumulative number of observation
P
=
pkj
(cid:17)
(cid:19)
M (t) = k
F(cid:89)
(cid:18) t
11
j = kj for all j ≤ F . The
within the pixel j reads: M (t)
likelihood of such outcome is defined as the probability:
(cid:16)
1 (1 − p1)t−kj 1(cid:98)F (t)≤F , (B1)
for all k ≥ 0, and where 1(cid:98)F (t)≤F is the indicative function,
equal to 1 if (cid:98)F (t) ≤ F and 0 otherwise. The product in
global minimum of Eq. (B1) is achieved for F = (cid:98)F (t) --
therefore (cid:98)F (t) is called the maximum likelihood estimator
Eq. (B1) spans from j = 1 to j = F as M (t)
j = 0 with
probability 1 for all pixels which do not correspond to the
structure of interest. Due to the indicative function, the
j=1
kj
of F .
2. Proofs for the coupon-collector scaling in the
regime 1 < r < ln(F )
a. One observation per pixel per frame
Here we consider the case where the number of
observations of a pixel at each frame is either 0 or 1 (i.e.
pk = 0 for all k > 1).
We define the probability q(t)
j
that the pixel m has been
j = P(M (t)
observed a number j times at the time t: q(t)
m =
j). Successive observations are considered as indepen-
1 ≤ j ≤
dent in time, hence q(t+1)
r − 1 for all 1 ≤ j < r. As we are interested in the time
required to reach the state j = r, we consider the state
j = r to be an absorbing state q(t+1)
r−1. As
soon as j ≤ t, the probability to have reached j ≤ r − 1
observations of the pixel is:
j + p1 q(t)
r + p1 q(t)
= p0q(t)
= q(t)
j−1,
r
j
P(M (t)
m = j) =
t!
(t − j)! j!
1 (1 − p1)t−j,
pj
j ≤ r − 1,
(B2)
from which we deduce the probability that the pixel has
been observed at least r ≥ 2 times is: q(t)
j=0 q(t)
.
In the long-time limit 1 (cid:28) t, t!/(t − j)! ∼ tj and the
absorption probability q(t)
r
tends to 1 as
j
r = 1−(cid:80)r−1
(cid:19)r−1
for
(cid:18) p1
p0
1 (cid:28) t.
(B3)
P(M (t)
m = r) = 1 − tr−1
pt
0
(r − 1)!
The probability that all pixels have been observed r
r )F . We are interested in the
=
times at the time t is (q(t)
centile time tθ given by the condition: P
(q(tθ)
In order to obtain a simple explicit
expression for tθ, we approximate the probability q(t)
r by
its long-time behavior from Eq. (B3) (which is valid for
)F = 1 − θ.
(cid:16){F δjr}j
(cid:17)
r
θ is sufficiently small or for F sufficiently large) to obtain
that:
(cid:18)
(cid:19)r−1
1 − (1 − θ)1/F =
ptθ
0
(r − 1)!
tθ
p1
p0
.
(B4)
Given that 1 − (1 − θ)1/F ∼ θ/F in the limit θ (cid:28) 1, we
obtain from Eq. (B4):
ln(p0)tθ + (r − 1) ln
tθ
= ln
(cid:19)
(cid:18) p1
p0
(cid:18) θ
(cid:19)
F
+ ln [(r − 1)!] ,
(B5)
which, in the limit r (cid:28) F , leads to:
(cid:110)
ln(cid:0) F
θ
(cid:1) + (r − 1) ln
(cid:104)
p0(− ln(p0)) ln(cid:0) F
p1
θ
tθ = −
ln(p0)
(cid:1)(cid:105)
+ C1
(cid:111)
,
(B6)
where C1 is a constant of F . In the regime of rare hits
(1−p0 (cid:28) 1), then ln(p0) = ln(1−(1−p0)) = −(1−p0) =
−p1 = −µ/F , where µ is the mean number of hits per
frame, Eq. (B6) then reads
tθ ∼
1(cid:28)F
F
µ
{ln(F/θ) + (r − 1) ln(ln(F/θ)) + C1/F} ,
(B7)
in the limit r (cid:28) F . This proves the relation of Eq. (8).
b. Random number of observations per frame
In this section, we consider that, at each frame, the
number of observations of a given pixel is random variable
equal to (i) 0 with probability p0 and (ii) to a value k ∈
[0, r] with a probability law pk.
Following the method of the previous paragraph B 2 a,
we consider the coverage dynamic for a single pixel. The
probability that the single pixel has been observed j-
times, with 1 ≤ j ≤ r − 1, during a sequence of t frames
is:
(cid:18) p1
(cid:19)j1
(cid:88)
that guarantee the condition that (cid:80)r
(ii) ju = (cid:80)r
(t − ju)! . . . jr!
where (i) the sum holds over the sets of indices (j1, . . . jr)
m=1 mjm = j, and
m=1 jm is the total number of adsorption
(cid:18) pr
(cid:19)jr
q(t)
j =
,
(B8)
pt
0
. . .
p0
p0
t!
events.
r
r = 1 −(cid:80)r−1
)F = 1 − θ, in agreement with Eq.
The imaging time tθ is defined by the equation:
(q(tθ)
In or-
der to obtain a more explicit expression for tθ, we focus
on two simple cases where the asymptotic behavior of
for t (cid:29) 1 can be analytically stud-
q(t)
ied.
We first review the case where steps are all of equal
height: pk = psδ(s−k) (e.g. p1 = 0) and r is a multiple of
j=0 q(t)
(4).
j
12
s i.e. there exists q such that r = qs. Then the situations
amounts to the case considered in the section B 2 a, with
the substitution r ← q and p1 ← ps.
The second case relies on the hypothesis that p1 > 0.
The set of indexes that maximizes the exponent ju in
m=1 mjm = j is
(j, 0, . . . 0). Moreover, j = r − 1 maximizes the exponent
ju = j1 = j. At the leading order in t, Eq. (B8) reads
Eq. (B8) under the constraint that (cid:80)r
(cid:18) p1
(cid:19)r−1
P(M (t)
m ≥ r) = 1 − tr−1
,
(B9)
pt
0
(r − 1)!
p0
which is identical to Eq. (B3), and leads to the scaling
Eq. (B7), and which therefore proves Eq. (11).
3. Multi-colored images
Our results are readily adaptable to the case of a col-
ored image, i.e. resulting from the combination of several
channels of light emission produced by different imag-
ing probes. This technique is frequently used in cell bi-
ology to image simultaneously actin, myosin and other
proteins [45]. The number of distinct types of imaging
probes is denoted C ; the number of pixels that contain
the j -- type probe is denoted Fj ; the probability (per
pixel and per frame) to detect an imaging probe is de-
are defined similarly to
Eq. (3). The imaging time is now defined by the rela-
= 1 − θ. As-
tion P
suming that the imaging probes act independently, the
imaging time tθ (with C colors) is given by the relation:
noted p1,j. The estimators (cid:98)F (t)
(cid:111) ∩ . . . ∩(cid:110)(cid:98)F (t)
(cid:0)1 − (1 − p1,j)tθ(cid:1)Fj = 1 − θ.
(cid:16)(cid:110)(cid:98)F (t)
C(cid:89)
(cid:111)(cid:17)
C = FC
1 = F1
(B10)
j
j=1
The imaging time defined in Eq. (B10) exhibit a coupon
collector scaling in the following two situations:
(i) if the emission probabilities are identical for all
probes (i.e. p1,j = p1), the expression from Eq.
(7)
holds after the substitution of (a) F by F1 + . . . + FC,
and (b) µ by µ = µ1 + . . . + µC, which corresponds to
the total number of events per frame. Therefore, if we
further assume that Fj = F for all j, we show that the
imaging time exhibits a coupon-collector type behavior
(CF ln(CF )) in terms of total number of pixels CF .
(ii) if one channel is characterized by a weak blinking
probability compared to all the other probes (e.g. p1,1 (cid:28)
p1,j for all other j, then it will likely be the limiting factor
in the imaging process, in which case Eq. (7) holds after
the substitution of F and p1 by F1 and p1,1.
4.
Inhomogeneous field of view
We first consider the case in which at most one obser-
vation per pixel and per frame can occur. We suppose
that the probability for p0 in each pixels follows a Gaus-
sian distribution, with:
2. Evolution of the probability distribution as a
function of the number of frames
13
P(p0,i = q) =
.
(B11)
patches cover the whole circle as
Following [39], we express the exact probability that n
We express the conditional relation that
PS [0 gapn events] =
(cid:18)n
(cid:19)
k(cid:88)
j
j=0
(−1)j(cid:16)
1 − j
σ
S
(cid:17)n−1
,
(C2)
where k is the greatest integer smaller than S/σ, and PS
denotes the probability measure with a structure volume
equal to S.
ter
(cid:80)∞
The
that
a
n=k+1 P [0 gapsn events] P
probability
time
then
t
reads:
N (t)
the set of coefficients that
exp(−µt)
∞(cid:88)
a(t)
j =
(cid:104)
(cid:16)(cid:0)n
j
(cid:105)
no
gap
remains
P [0 gaps]
af-
=
. We define
e = n
(cid:1)((µt)(1 − jσ/S))n(cid:17)
n=k+1
= (µt)je−j(µt)σ/S(1 − jσ/S)jγj,
n!
where
γj =
Γ(−j + k + 1) − Γ(−j + k + 1, (µt)(1 − jσ/S))
Γ(j + 1)Γ(−j + k + 1)
.
With these definitions, we conclude using Fubini's theo-
rem that the probability reads
P [T ≤ t] = P [0 gap] =
(−1)j
a(t)
j
1 − jσ/S
.
(C3)
k(cid:88)
j=0
We mention that, from Ref. [39], the probability that a
number i ≤ k of gaps remains after n events, denoted
P [i gapsn events], reads:
(cid:19) k−i(cid:88)
(cid:18)n
(cid:18)n − i
(cid:19)
(−1)j(cid:16)
i
j=0
j
(cid:17)t−1
1 − (i + j)
σ
S
,
(C4)
which is the continuous analogue of Eq. (5) in the box-
filling model.
3. The area estimation problem
Following the method presented in the final section
in the main text, we propose an estimator of the risk
that the image is not complete, based on the collected
information after t frames. In particular, we provide an
estimate of the additional number of frames that should
be taken to obtain a complete image at a given confidence
θ ). This estimator rely on the current covered area
((cid:98)t(t)
of the experimental realization (cid:98)S(t), which corresponds
to a maximum-likelihood estimator of the volume of the
exp(cid:0)−(q − p0)/(2σ2)(cid:1)
(cid:82) 1
0 du exp (−(u − p0)/(2σ2))
(cid:90) 1
P(M (t)
i ≥ r) =
dq P(Mi ≥ rp0,i = q)P(p0,i = q).
0
(B12)
We notice that P(M1 ≥ rp0,i = q) is given by Eq.
(B3). The imaging time is now defined by the relation
= 1 − θ. Assuming
P
that each pixels are independent and using Eq. (9), the
imaging time tθ is defined by the relation
(cid:111) ∩ . . . ∩(cid:110)
F ≥ r
M (t)
1 ≥ r
M (t)
(cid:111)(cid:17)
(cid:19)F
1 ≥ rp0,i = q)P(p0,1 = q)
dq P(M (t)
(cid:16)(cid:110)
(cid:18)(cid:90) 1
0
= 1 − θ.
(B13)
From the experimental data set, we estimate the local
value of p0,i for each pixel i. For a number of pixel F =
2500, we find that the values of p0,i can be qualitatively
considered as Gaussian distributed (see Fig. ). The mean
probability is E(p0) = 0.95 and the standard deviation
is σs = 0.010. We point out that the latter value of σs
is significantly larger than the standard deviation σu <
10−5 associated to the uncertainty in the estimation of
p0,i in each pixels.
Appendix C: Patch-method
In the following sections Sec. C 1 to Sec. C 2, we focus
on the 1D -- coverage problem of a circle by circular arcs.
(cid:104)(cid:98)Y 1 events
(cid:105)
1. Evolution of the mean coverage
We denote by (cid:98)Y (t) the fraction of points which are still
left uncovered after t frames. For a single patch, F = 1,
= 1 − σ/S.
the mean uncovered area is E
Since patches occur at independent positions within S,
we can factorize the mean covered area after n number
= (1− s/S)n. After averaging
of patches: E
over the distribution of the number of events up to the
time t, we find that the mean covered volume after t
frames reads [39]
(cid:105)
(cid:104)(cid:98)Y n events
(cid:105)
(cid:104)(cid:98)S(t)/S
E
= 1 − exp(−µtσ/S),
(C1)
which holds for arbitrary values of σ/S and t.
14
Figure 7: (Color online) (a) Probability of the full coverage of
a circle of circumference S by a number of t of patches (arcs
segments) of size σ/S = 10−2: (blue solid line) mean covered
area (cid:98)Y (t) after t-frames (with 1 event per frame) ; (orange
at time t, as constructed from the mean covered area (cid:98)Y (t),
line) exact probability of a full coverage at time t from Eq.
(C2) (black circles) estimated probability for a full coverage
(C5).
from Eq.
(b) Evolution of the centile of the image
time as a function of the ratio of the volume covered at each
localization event (σ) by the total volume of the field of view
(S).
Figure 8: (Color online) Analysis of the experiments from
Ref. [30] using the Gaussian model defined in Eq. (D1). (a)
Density representation of the experimental data: measured
scattered intensity Ii as a function of the measure height Zi:
(colormap) experimental number of observations per boxes
(red line) linear regression for ln(I) in terms of z. (b) Residual
analysis and validation the linear Gaussian model: we verify
that residual Ri =
line) and maintains a constant variance (red box) with the
intensity Iobs.
Ii ln(Ii/(cid:98)I0n)) is centered (red
IiZi +
√
√
structure to be imaged. We assume that the covered
length (cid:98)F (t) is well described by the mean covered area at
time t, hence that
(cid:98)P(t)(cid:16)(cid:98)S(t) = S
(cid:17) ≈ PS(1−e−σt/S) [0 gap] ,
(C5)
where PS(1−e−µσt/S) refers to the probability defined
(C2), that the total area to be covered is
in Eq.
S(cid:0)1 − e−µσt/S(cid:1). We test the method on Fig. SI. 7a,
in a case where the ratio S/σ is known, and we find that
the approximation of Eq. (C5) is very satisfactory. Fol-
lowing the procedure defined in Sec. V, we estimate the
additional number of frames required to obtained a com-
plete image, which can be deduced from Eq. (1), through
the substitution of S and µ by their corresponding esti-
mators.
Appendix D: Calculation of the minimum number r
of observations per pixel
In this section, we justify that a large number of
observations is required per pixel to obtain a reliable
measurements of the electromagnetic field considered in
Ref. [30]. We recall that the experimental data consists
in 2792 measurements of heights and intensities (Zi, Ii)
(i ≤ 2792). We model the noise on the height measure-
ment through the following linear Gaussian model:
IiZi = −β
Ii ln (Ii/I0) + σηi,
(D1)
(cid:112)
where η is a standard Gaussian white noise process.
We define the vector of unknown parameters B =
β(−1, ln I0), where β refers to the penetration length and
IiZi
I0 to the intensity at the surface, as well as Y =
√
Figure 9: (Color online) Convergence of estimators and the
imaging process of a patterned surface (14 × 14 pixels) in
terms of the penetration length field β and maximal intensity
I0. (a) The penetration length is β = β(2) by default and
β = β(1) = 1.50β(2) within specific pixels (MBI pattern).
(b) Convergence of the estimator (cid:98)βn, defined in Eq. (D2).
estimator (cid:98)I0n defined in Eq. (D3).
(c) The maximal intensity I0 = I (1)
0 = 0.75I (1)
within specific pixels (NUS pattern). (d) Convergence of the
or I0 = I (2)
0
0
√
√
Ii ln Ii,
Ii) (ex-
(vector of observations) and X = (
plicative matrix). In terms of these quantities, the model
defined in Eq. (D1) reads Y = XB + ση, which corre-
sponds to the well-known linear Gaussian model. The
estimator of the vector B that maximizes the likelihood
function is (cid:98)Bn = (X T X)−1X T Y [46] ; the developed ex-
SSSS/ab-101-5-4-3-2-1024681012141618ab-0.4-0.200.2-6-4-20051015 15
We now apply our results to the detection of a pattern
in the electromagnetic field (see Fig. SI. 9), in a simulated
experiment. We consider a local electromagnetic field
that takes the form of Eq. (2), and in which β and I0 may
take either of two values. In Fig. SI. 9, the error of the
estimators β is lower than 10% ; the image indeed results
from the superposition of t = 106 frames, a number in
agreement with the predicted threshold from Eq. (13):
F × r ≈ 106, where F = 14 × 14 and r = 4000.
(cid:18) 1
pression corresponds to the two estimators
(cid:98)βn = −C(cid:8)I · ZI ln I − I ln I · ZI(cid:9) ,
(cid:98)I0n = exp
where I =(cid:80)n
(cid:8)I · ZI ln I − I ln I · ZI(cid:9)(cid:19)
2(cid:17)−1
I · I ln2 I − I ln I
(cid:16)
βn
,
(D2)
(D3)
i=1
√
(D4)
i=1 Ii and C =
n = (
√
i )/(n − 2),
R2
. The
variance of the noise is also an unknown variable that can
be evaluated by the estimator:
where Ri =
For a Gaussian distribution of noise, we expect to have:
n(cid:88)
(cid:98)σ2
Ii ln(Ii/(cid:98)I0n)) is called the residual.
n − 2(cid:1) in the limit n (cid:29) 1.
n/σ2 ∼ F(cid:0)1, 1/
(cid:98)σ2
within the confidence interval C((cid:98)βn):
(cid:104)(cid:98)βn ±(cid:98)σnt(n−2)
We now define confidence intervals for the estimators
defined in Eqs. (D2 -- D3). We consider a risk level α =
0.05: with a 1 − α = 0.95 probability, the quantity β lies
β ∈ C((cid:98)βn) =
IiZi +
√
(cid:112)
1−α/2
(D5)
(cid:105)
IC
,
where t(n−2)
0.975 is the one-sided quantile of the Student dis-
tribution, with t(n−2)
0.975 = 1.96 . . . in the limit n (cid:29) 1. As
the quantity C is inversely proportional to the number of
√
observations n, the confidence interval Eq. (D5) narrows
on β with a 1/
n speed as n increases.
From Eq. (D5), we obtain an estimate of r, i.e. the
minimal number n = r of observations required so that
the estimator of β has an error lesser than 10%, with
probability 95% probability. From experimental values,
we find that the confidence interval for β is 60, 1± 7.3 nm
for n = 2792.
Therefore, the estimate for the minimal observation
n )
n )
(cid:105)
(cid:105)
= (P
per pixel r should around 4000.
(cid:104)
β ∈ CI((cid:98)β(1)
(cid:105)
probability that all the estimators (cid:98)β(j)
(cid:104)∀j, β ∈ CI((cid:98)β(j)
(cid:104)∀j, β ∈ CI((cid:98)β(j)
(cid:105)
We now consider a system of F identical pixels. The
n , where 1 ≤ j ≤ F
is the pixel label, are precise at 10% to the exact value
)F . We set
β is P
n )
= 1 − where is the accepted risk
P
level (in the following = 0.05). Therefore, compared to
the case of single pixel, the risk level on a single pixel is to
= 1−0.05/F .
be divided by a factor F : P
The confidence interval from Eq.
(D5) holds provided
that α = 0.05/F , which corresponds to a narrower con-
fidence interval compared to the single pixel case. How-
ever the quantity t(n−2)
1−/(2F ) increases weakly with the
t(∞)
1−0.05/(2×10) ≈ 2.80 and
number of pixels F (e.g.
t(∞)
1−0.05/(2×100) ≈ 3.40).
(cid:104)
β ∈ CI((cid:98)β(1)
n )
In conclusion, increasing the number of pixels F does
not significantly increase the required number of obser-
vations per pixel r.
|
1906.11944 | 1 | 1906 | 2019-06-24T07:53:03 | Mechanical characterization of cells and microspheres sorted by acoustophoresis with in-line resistive pulse sensing | [
"physics.bio-ph",
"physics.app-ph"
] | Resistive Pulse Sensing (RPS) is a key label-free technology to measure particles and single-cell size distribution. As a growing corpus of evidence supports that cancer cells exhibit distinct mechanical phenotypes from healthy cells, expanding the method from size to mechanical sensing could represent a pertinent and innovative tool for cancer research. In this paper, we infer the cells compressibility by using acoustic radiation pressure to deflect flowing cells in a microchannel, and use RPS to sense the subpopulations of cells and particles at each acoustic power level. We develop and validate a linear model to analyze experimental data from a large number of particles. This high-precision linear model is complemented by a more robust (yet less detailed) statistical model to analyze datasets with fewer particles. Compared to current acoustic cell phenotyping apparatus based on video cameras, the proposed approach is not limited by the optical diffraction, frame rate, data storage or processing speed, and may ultimately constitute a step forward towards point-of-care acousto-electrical phenotyping and acoustic phenotyping of nanoscale objects such as exosomes and viruses. | physics.bio-ph | physics |
Mechanical characterization of cells and
microspheres sorted by acoustophoresis with
in-line resistive pulse sensing
Antoine Riaud,∗,†,‡ Anh L. P. Thai,‡,¶ Wei Wang,† and Valerie Taly∗,‡
†ASIC and System State Key Laboratory, School of Microelectronics, Fudan University,
Shanghai 200433, China
‡INSERM UMR-S1147, CNRS SNC5014, Paris Descartes University, Equipe labellis´ee
Ligue Nationale contre le cancer, Paris, France
¶The Physics of Living Matter Group, Department of Physics and Materials Sciences,
University of Luxembourg
E-mail: antoine [email protected]; [email protected]
Abstract
Resistive Pulse Sensing (RPS) is a key label-free technology to measure particles
and single-cell size distribution. As a growing corpus of evidence supports that cancer
cells exhibit distinct mechanical phenotypes from healthy cells, expanding the method
from size to mechanical sensing could represent a pertinent and innovative tool for
cancer research.
In this paper, we infer the cells compressibility by using acoustic
radiation pressure to deflect flowing cells in a microchannel, and use RPS to sense
the subpopulations of cells and particles at each acoustic power level. We develop
and validate a linear model to analyze experimental data from a large number of
particles. This high-precision linear model is complemented by a more robust (yet
less detailed) statistical model to analyze datasets with fewer particles. Compared
1
to current acoustic cell phenotyping apparatus based on video cameras, the proposed
approach is not limited by the optical diffraction, frame rate, data storage or processing
speed, and may ultimately constitute a step forward towards point-of-care acousto-
electrical phenotyping and acoustic phenotyping of nanoscale objects such as exosomes
and viruses.
Introduction
Resistive pulse sensing (RPS) is a key method for the label-free analysis of cells and nanopar-
ticles. 1 It measures the size of particles by monitoring the electrical resistance of a channel
filled with electrolyte such as Phosphate Buffer Saline (PBS). When an insulating particle
travels through this channel (or pore), the resistance changes by an amount proportional to
the particle volume. This method is faster than video analysis and is not limited by optical
diffraction. This makes it pertinent to analyze viruses 2,3 colloids, 4,5 macromolecules 6 and es-
pecially DNA. 7 For all these advantages, RPS is a mainstream method with well-established
standards adapted for medicine and diagnostics based on liquid biopsies. 8,9
In recent years, the accumulation of evidences showing that cancer cells exhibit distinct
mechanical phenotypes from healthy cells 10,11 has prompted a major research effort to add
mechanosensing capabilities to the RPS framework. The most common strategy 12 -- 15 is to
use a pair of constrictions of different widths. The widest constriction is slightly larger than
the cell diameter and measures the cell size while the smallest one is slightly narrower than
the cell. The time needed for the cell to deform and travel through the second constriction
may then be fed into biomechanical models to estimate the cell deformability. 16,17 Two ma-
jor requirements of this squeezing method are (i) that the constriction size must be very
close to the cell size, which requires specific devices for each different cell size and make it
unsuitable for complex mixtures of different cells (such as whole blood and heterogeneous
populations 18), and (ii) that the objects traveling through the channel must all be highly
deformable, which excludes the study of solid microparticles and most nanoparticles (includ-
2
ing viruses and exosomes). Indeed, the shear due to cytoplasm flow in deforming cells is
proportional to 1~a2, 17 with a the cell radius, and by analogy it is inferred that viral and
exosome components could not deform enough to cross a narrow constriction and provide
meaningful elasticity data.
While deformation is a quasi-static measurement of cell mechanical properties, acoustic
characteristics such as sound speed and compressibility complement this picture with a
high-frequency viewpoint. Such measurement is commonly achieved by acoustophoresis,
ie. by tracking the migration of objects due to acoustic forces. 19 Compared to deformation
measurements, acoustophoresis characterization is applicable to both solid and soft particles
and is contactless, which minimizes cross-contamination risks. A major shortcoming of
this characterization is that the migration speed depends strongly on the particle diameter,
which has to be evaluated externally, for instance using RPS. 19 A promising alternative is the
isoacoustic method 20 but it requires an elaborated optical setup and modified media. Both
methods also rely on a microscopy setting for video analysis which restricts the measurement
throughput, is limited by optical diffraction and precludes point-of-care applications.
In this paper, we combine the well-accepted RPS method for the measurement of particle
size to acoustophoresis for the measurement of particle compressibility. We first introduce
a theoretical model that relates the particle deviation to the acoustic field intensity and the
particle size, density and compressibility. Hence, depending on the acoustic power level,
different populations of particles can be sorted. The size of each of these populations is then
measured by an in-line RPS chip. Eventually, the signal is numerically analyzed to recover
the particle compressibility depending on its density.
After introducing the necessary theoretical background for each component of the exper-
iment (acoustophoresis chip, resistive pulse sensing chip and signal analysis) and detailing
the experimental protocol, we coupled an acoustophoresis and an RPS chips for a proof-of-
concept experiment involving polystyrene microspheres and Jurkat cells (a common model
for blood cancer already studied with the RPS technology alone). Our experiments not only
3
yield to the compressibility and the size distribution of the polystyrene microspheres and
Jurkat cells but also reveal that the polystyrene microspheres formed doublets that could be
distinguished from Jurkat cells based on their compressibility.
System principle and theory
The proposed system combines acoustophoresis to measure particle compressibility and re-
sistive pulse sensing to obtain particle size (Fig. 1). By modulating the acoustic power level,
the acoustophoresis chip sorts various populations of particles. The relation between the
acoustic power and the deviation of particles is established at the beginning of this section.
It is shown that the particle compressibility can only be computed if the size of the particles
is known. To obtain the particle size, the sorted particles are guided towards another chip
equipped with an RPS sensor. In principle, knowing the acoustic power and the particle size
should yield the acoustic contrast of the particle and thus its compressibility in a straight-
forward fashion. However, the two chips were designed as stand-alone and are connected
by a short yet non-negligible capillary tubing that generates a lag between the sorting and
detection events. The signal processing when accounting for this delay is non-trivial and is
discussed at the end of this section.
Tilted-angle standing SAW acoustophoresis
At high power, acoustic waves generate a steady stress called acoustic radiation pressure.
The resulting force is used in acoustophoresis experiments to displace particles. 21,22 Since the
migration speed depends on the particles size, density and compressibility, 23 acoustophoresis
is routinely used for sorting small objects in microfluidic channels. 20,24 -- 27
A widespread technology to generate the acoustic field is using surface acoustic waves
(SAW) as shown in Fig. 2(a). At the center of the picture, cells suspended in their cul-
ture medium flow through a PolyDiMethylSiloxane (PDMS) channel. This channel is placed
4
Figure 1: Experimental setup. The experiment combines an acoustostophoresis chip (top-
center) and a RPS chip (bottom-center). On the bottom-right, two pressurized vials con-
taining a buffer fluid (PBS) and the sample to analyze (cells and/or microspheres) supply
the acoustophoresis section of the experiment (top). This section comprises a microfluidic
chip (top-right), a pair of interdigitated transducers and a power-modulated ultrasonic power
supply (top-left). Depending on the acoustic power, the particles are directed to a default
outlet (bottom-right) or to the resistance pulse sensing chip (bottom-center). The flow rates
in the system are controlled by three precision flowmeters.
Figure 2: Sorting chip schematic. (a) cross-section: the IDTs are located on each side of
the disposable PDMS chip. The generated SAWs (wavy arrows) travel freely along the solid
surface and radiate once they reach the PDMS. The x direction in the drawing corresponds
to the crystallographic X direction. The generated bulk waves are then transmitted into the
liquid (straight arrows). (b) top view: the particles carried by the fluid are deviated by the
acoustic radiation force. The blue and red colors indicate the (oscillating) electric potential
while the green to yellow color gradient refers to the (oscillating) acoustic pressure field.
(c) Three scenarios of particle deviation depending acoustic radiation to drag force ratio M
(Eq. (8)): when M→ 0, the particles follow the flow (ψ= θC), whereas when M > sin θC
the particles are locked along the acoustic wavefronts (ψ= 0). At intermediate values, the
particles follow a striated path forming an angle ψ with the wavefronts given by Eq. (7).
5
SAW80 um500 um(a)xzSAWout1out2(b)yx C out1out2yx! M 00< M < sin C M > sin C (c)FradFdrag" C between a pair of interdigitated transducers (in blue and red) that generate two counter-
propagative surface acoustic waves (oscillating arrows). These waves propagate along the
solid surface until they reach the microchannel base. At this stage, the SAW interfere to form
a standing surface acoustic wave and the vertical component of the surface wave vibration
radiates into the PDMS and then in the channel as a bulk acoustic wave (BAW). A consid-
erable advantage of using SAW is the possibility to use cheap disposable microchannels that
can be detached from the ultrasonics transducer in order to minimize cross-contamination
risks. 28 -- 30
Under the action of the acoustic wave, the cells are attracted towards the pressure nodes.
At best, this allows a separation distance of a quarter-wavelength.
In order to overcome
this limitation, Collins et al. 31 proposed using tilted-angle SAW to deflect the particles as
shown in Fig. 2(b). In this tilted configuration, the trapped particles will travel along the
acoustic nodes while the drifting particles will follow the flow more closely (cf. analytical
model thereafter). This tilted configuration has later been improved by Ding et al. 32 to sort
circulating tumor cells from peripheral blood mononuclear cells (PBMC) 33 and exosomes
from whole blood 34 and is adopted here.
The acoustic force Frad due to a standing SAW reads:
Frad = −4πa3
Φ = f1− 3
E = 1
pRM S = ρ0ω2uRM S
κ0pRM S
2
2
kXEΦ sin(2kX⋅ r),
f2 cos(2θR),
3
2
kL cos θR
(1)
(2)
(3)
(4)
with a the particle radius, kX the SAW wave-vector, kL the BAW wavenumber (in the liquid),
ω the SAW angular frequency,E the acoustic energy density, pRM S the pressure fluctuation
root-mean-square (RMS) of the BAW, uRM S the RMS of the SAW vertical oscillations, Φ
the acoustic contrast between the particle and the fluid, and r the position vector of the
6
scattering coefficients of the particle. κp and κ0 stand for the particle and fluid
particle. The acoustic contrast factor, given by Eq. (2), depends on the propagation angle
of the radiated SAW θRࣃ 22o (Rayleigh angle) and the monopolar f1= 1− κp
f2= 2(ρp−ρ0)
2ρp+ρ0
factor− cos(2θR)ࣃ−0.67, that is approximately the opposite of the unitary value needed to
recover the usual contrast factor Φ= f1+ 3
compressibility respectively, and ρp and ρ0 the particle and fluid density respectively. As
pointed out by Simon et al., 35 the SAW acoustic contrast differs from its BAW value by a
and dipolar
κ0
The acoustic radiation force is balanced by the drag force:
2f2 obtained when θR→ π
Frad+ Fdrag= 0.
2 .
(5)
(7)
(8)
Neglecting particle acceleration, and for particles far away from channel walls, the drag force
reads:
Fdrag= 6πηa(vF− vp),
(6)
where vF and vp stand for the flow and particle speed respectively, and η is the dynamic
viscosity.
In previous studies of tilted-angle standing SAW acoustophoresis, no solutions to Eq. (5)
were available and it had to be integrated numerically. However, the analog optical problem
was previously solved by Pelton et al. 36 by a clever change of coordinates. They showed that
the particle travels with an angle ψ relatively to the wavefronts (see Fig. 2(c)):
tan ψ = 0,
sin2 θC−M 2
M = 2a2ΦkXE
cos θC
.
9ηvF
sin θC < M
sin θC > M
,
Interestingly, Pelton et al. also derived a relatively simple analog of Eq. (7) for nanoparticles
that accounts for Brownian diffusion.
7
In the deterministic (non-Bronwnian) case, this deviation angle only depends on the tilt
angle θC and the dimensionless migration speed M given by Eq.(8). When sin θC < M ,
the particles follow the acoustic wavefronts (locked mode) whereas for smaller values of M
the particles travel more tangentially to the flow (ie. tan ψ→ tan θC). Remarkably, tan ψ is
independent of the sign of M so that particles with positive and negative acoustic contrast
follow the same trajectory. Since these theoretical results were previously unknown to the
field of acoustics, we first confirmed them against previously published data in Fig. 3. The
perfect match between analytical and numerical results validates the calculations, while the
good agreement with experimental results supports the validity of this opto-acoustic analogy.
Figure 3: Comparison of previously published deviation angle of polystyrene microspheres of
various radii with the numerical integration of equation (5) and its analytical solution given
by Eq. (7). a: Adapted from Collins et al. (Fig. 7B), 31 b: Adapted from Ding et al. (Fig.
S1). 32 In both cases, the SAW magnitude was assumed proportional to the actuation voltage
and this conversion coefficient was the only fitting parameter (one conversion coefficient was
regressed for each figure). The angle φ= θC− ψ was deemed closer to experimental concerns
and thus more convenient for comparisons.
According to Eq. (7), the particles trajectory depends only on the parameter M . Since
M depends on the particle radius and acoustic contrast, previous studies 19 were unable
to obtain the compressibility directly and had to assume a given particle radius or use an
average value obtained by an independent measurement instead. Hence, one experiment had
8
to be performed for each particle size, and the heterogeneity in size was difficult to take into
account. Here, the particle size is directly measured after sorting using the RPS chip.
Resistive pulse sensing
Figure 4: Electrical schematic of the sorting chip. The cells and microspheres flow through
the 200 µm long constriction where they are sensed by the electrode triplet. This channel
forms half of Wheatstone bridge completed by a variable resistor. The 50 kHz AC excitation
allows to bypass the capacitive nature of the electrical double layer.
Resistive pulse sensing works by monitoring the electrical resistance of a channel contain-
ing a conductive solution (Fig. 4). In the absence of particles, a voltage applied between the
blue and red electrodes generates a baseline electrical current. When an insulating particle
travels through this channel, it blocks some of the electrical current (ie. the channel electrical
resistance RCH increases). The method is also able to determine biological cell radii thanks
to the cells membrane that blocks the electrical current. For a channel of hydraulic diameter
DH and length L, 37 the resistance increase ∆RCH= Rliq− Rcell reads:
F(8a3~DH
1
1+ DH
L 2
where F is a correction factor close to unity when a< 0.25DH
is given by DH=
4wh
π with w the channel width and h its height.
DH
2L2 +
∆RCH
RCH
2
= 8a3
LDH
2
3),
(9)
38 and the hydraulic diameter
9
CR1varR2varV+V-Vexc20 um100 um A major problem when working with cells and smaller particles is that miniaturized elec-
trodes become a current bottleneck due to the insulating electrical double layer (depicted as
capacitors CDL in Fig. 4). Hence, measurements of continuous current are flawed due to this
parasitic capacitance. This limitation is mitigated by working with an alternative current.
Another advantage of alternating current is to protect the electrodes from electrolysis.
In order to improve the signal-to-noise ratio, the channel electrical resistance is monitored
by comparing two symmetric electrodes (shown in red), which form an electrical half-bridge.
Measurement errors are further minimized by completing a 4-resistors Wheatstone bridge to
obtain the differential signal directly:
V+− V−ࣃ ∆RCH
RCH+ R1
Vexc,
(10)
This signal is subsequently amplified by lock-in amplifier and digitized for numerical
processing.
Signal analysis
Depending on the acoustic power level, different populations of particles are deflected towards
the RPS chip where the particle size is evaluated. In principle, knowing the dimensionless
migration speed M and the particle size should yield the acoustic contrast of the particle
and thus its compressibility. In practice, the two chips were designed as stand-alone and are
connected by a short yet non-negligible capillary tubing. Due to this additional distance,
there is a lag of several minutes between the application of the acoustic power and the
detection of particles in the RPS chip. This situation is further complicated by the Taylor
dispersion so that particle arrival time itself spreads over tens of seconds. Furthermore, due
to experimental constraints detailed thereafter, the acoustic power levels had to be changed
randomly over relatively short timescales. In order to process the time-delayed RPS signal
and obtain the particles compressibility, we propose two approaches. The first strategy is
10
closely related to a physical model of the particle transport in the capillary model (linear
model) while the second one is purely statistical and focuses on the acoustic power levels
that preceded the detection of a particle.
Linear method
In the physical approach, we consider the probability g to observe the particle i exiting the
tubing at a specific time ti. In physical terms, g is the impulse response of the tubing for
particle transport. A fundamental assumption of our model is that particles do not interact
with each other, such that the arrival time distribution of the particles only depends on the
particle (its size) and whether it was sorted or not (acoustic contrast and acoustic energy
density). Another assumption is that the acoustic contrast distribution of each kind K of
particle do not overlap. Here, kind is purposefully vague as it may refer to the material (for
polystyrene microspheres), or to the cell type, strain, etc.
The experiments were conducted with a finite set of p+ 1 acoustic power levels (Ek ∈
{E0..Ep}). The time was also discretized into nt periods so that the delay τ between sorting
and observation belongs to {τ1...τnt}. Assuming that the energy density was always zero
except between ti− τj and ti− τj+1 where it reachedEk, we define g
as the probability of
(K)
jk
observing a particle of kind K. Thus, the total probability of observing a particle of kind K
at time ti knowing all the sequence of power levels reads:
(11)
(12)
pQ
k=0
(K)
jk δk
ij,
(K) = ntQ
j=0
ij = 1 E(ti− τj)=Ek,
otherwise.
δk
Bi
g
0
In Eq. (12), δk
ij encodes reconstruction of the history of the acoustic energy density that
preceded the detection of the particle i.
Since the power levels were randomly sampled from a uniform distribution, the total
11
probability to observe a particle at time ti knowing all the sequence of power levels reads:
pQ
k=0
Bi = ntQ
j=0
gjk = Q
K
gjkδk
ij,
(K)
jk ,
xKg
(13)
(14)
where xK the fraction of particles of kind K. Compared to (11), this equation indicates
that the function gjk should exhibit one peak for each kind of particle even if they share the
same acoustic contrast. Introducing the linear index α= j+ ntk, the unknown g-function is
obtained by comparing the model predictions B=DG to the ground truth B for N particle
This comparison is done by minimizing the Euclidian distance B− B2 with the constraint
gα≥ 0:
detection events and N controls (randomly sampled signals when no particles were detected).
,
B1⋮
B2N
δ1
δ2N
D =
1 δ1
nt(p+1)
⋮
⋮
1 δ2N
nt(p+1)
, B=
G =
g1⋮
gnt(p+1)
Bi = 1 particle,
control.
0
,
(15)
(16)
(17)
which allows determining the impulse response for each type of particle at each power level.
This detailed information can then be used to determine the minimum acoustic power to
deflect them, and thus the acoustic contrast and particle compressibility.
12
Statistical method
When the number of particles is too low, the linear model is badly conditioned and yields
inaccurate results. If the sorting threshold is the only valuable measurement to extract from
the RPS dataset, the statistical method detailed thereafter tends to fare better. Compared
to the previous model, an additional assumption is that there exists a sharp thresholdEmin
that a discrete time-seriesE(ti− τj), j ∈{1..nt} of random power levels yielded a particle
i, we know that at least one of those power levels exceededEmin. Reversely, if no particle
below which no particles are deviated and above which all particles are sorted. Assuming
was sorted we may assume that either no particle was present in the sorting section or
that the threshold power has not exceeded at the critical time when the particle was in the
sorting section. The former hypothesis becomes overwhelmingly more likely if the number
of particles is small, which allows lifting the ambiguity. The problem then becomes analog
to rolling an s-sided dice nt− 1 times and then rolling once a different dice that only has
markings above a thresholdEmin. The average result of the rolls will be higher than if the
unbiased dice had been kept for all the rolls. Similarly the mean of the power series preceding
particles detection should deviate from the mean of the power applied to the microchannel:
{Ei}=(nt− 1)µ(E)+ µ(E≥Emin)
nt
,
(18)
N∑
i=1
N
xi denotes the mean of quantity x over all the particles,x=
x(τj)
where{x}= 1
is the time-average of the quantity x, µ(E) denotes the expected value of the s-sided dice
and µ(E>Emin) the expected value of the biased dice:
Ek,
pQ
p+ 1
k=0
1+ p− pmin
µ(E) =
µ(E≥Emin) =
pQ
k=pmin
τnt−τ1
1
(19)
(20)
1
1
Ek,
nt∑
j=1
13
In our experiments, the acoustic power levels are regularly spaced (Ek= kE0). Combining
Eq. (18) to (20) and after some algebraic manipulation, we get the shift between the average
acoustic power levels that preceded the detection of a particle and those that do not:
{Ei}− µ(E)=Emin
2nt
,
(21)
Eq. (21) clearly illustrates the trade-offs of this statistical method: the acoustic power history
must be long enough to capture the power level that triggered the sorting and subsequent
detection of the particle, but it should not be too long as this tends to dilute the information.
Materials and methods
Materials
An aqueous suspension of 7.32 µm diameter PS microspheres (FS06F/9559) was purchased
from Bangs Laboratories. According to the manufacturer, the PS beads density is 1062
kg~m3.
cells is approximately 11.5± 1.5µm and their density was assumed similar to lymphoblasts 40
(1075 kg~m3).
Jurkat cells were prepared as described by Fernandez et al. 39 The diameter of Jurkat
Cells and PS microspheres suspensions were mixed together to a final number density of
0.5 million microspheres and 0.5 million cells/mL. This number density was chosen so that
at most one cell or one particle was in the RPS sensor at any given time.
Acoustophoresis chip and transducers
The SAW transducer was a two-side polished 3" diameter Y-128o cut of LiNbO3 crystal
equipped with a pair of interdigitated transducers. The electrode width and gap were set
to 25 µm in order to become resonant at the 40 MHz excitation frequency. The transducers
14
were positioned to generate an X-propagating SAW (velocity cSAW= 3990 m/s). Disposable
microfluidic chips made of PDMS were prepared by soft-lithography. The chips bottom were
closed by a 100 µm thick membrane. According to profilometer measurements, the sorting
section was 4 mm long, 500 µm wide and 80 µm high. It made a 3.4o angle with the IDT.
The fabrication process is detailed in the supplementary information.
The cells and microspheres were introduced at the channel center at a flow rate of 4
µL/min, while a buffer flow with a flow rate of 9 µL/min was added symmetrically to focus
the particles before sorting. Such flow rate was chosen as an acceptable compromise between
slower flow rates that yield an easier sorting and high flow rates less prone to sedimentation
issues. The flow was supplied by a pressure-based microfluidic flow controller (MFCS-EZ,
Fluigent) and the flow rates were monitored with three microfluidic flow sensors (FRP,
Fluigent). The pressure was then regulated by a control loop to maintain a constant desired
flow rate.
The default outlet of the acoustophoresis chip was discarded in a pressurised container,
while the sorted outlet was connected to the RPS chip via a 4.5 mm long PTFE tubing (0.3
mm inner diameter). The flow rate towards the RPS chip was regulated to 3 µL/min.
Acoustic power modulation
During operation, the SAW transducers were powered with a sinusoidal voltage of amplitude
500 mVpp generated by an arbitrary waveform generator (AWG) gated by an Arduino module
and then amplified by a 30 dB PARF310004 power amplifier (ETSA) (see Fig. 1). According
to a calibration procedure using acoustic streaming in micro-droplets 41 described in the
supplementary information, the standing SAW displacement at the center of the channel
was approximately 0.52 nmRM S. Literature data indicate that the leaky SAW decays by
approximately 0.4 dB/wavelength, that is 2 dB over the channel width. 42 -- 44
In order to adjust the average magnitude of the acoustic radiation force, we used the
duty cycle of the Arduino Power Modulation (PWM). Provided that the gating frequency
15
(490 Hz) is much slower than the SAW frequency, equation (1) remains valid. Meanwhile, as
long as the modulation period is much shorter than the time particles take to travel across
the sorting section, the particles only experience the average acoustic power.
Besides time-dependence constraints, the choice of power levels during the acoustophore-
sis was further guided by two aspects: (i) the power modulation frequency of the Arduino
chip is close to the lock-in amplifier, which adds noise to the measured signal from the sensor
chip. In order to increase the signal-to-noise ratio (SNR), we alternated periods of 5 s on
and 5 s off. This 5 s duration was chosen much smaller than the characteristic duration
of the impulse response (approximately 50 s, see the results section) and much longer than
the residence time of the particles in the microchannel. All the RPS measurements were
conducted during the 5 s off, and the 5 s on samples were discarded. (ii) PDMS is a strongly
attenuating material that absorbs quickly the SAW power which drives significant temper-
ature increase in the vicinity of the SAW, 45,46 hence the power must remain low enough
not to perturb significantly the experiment, In preliminary experiments, the PDMS showed
evidence of thermal damage when the time-averaged acoustic energy density exceeded 22.8
J/m3 (ie when the duty ratio was above 50%). Therefore, this duty ratio was selected as the
upper bound for subsequent experiments.
The main experiment time scales are illustrated in Fig. 5.
In order to measure the
compressibility of the particles, the acoustic power was selected randomly within 11 regularly
spaced values every 10 s cycle. The exact sequence of power and the detection events were
recorded for further analysis.
Resistive Pulse Sensing
The RPS chip was composed of a series of filters to prevent clogging (smallest cross-section
20 µm) followed by a LRP S= 100 µm-long 20× 20 µm wide sensing section (see Fig. 4). In
order to generate a quasi-uniform electric field required by equation (9), we set the distance
between the electrodes to 100 µm. The uniformity of the field was verified using Comsol
16
Figure 5: Main experiment timescales. a: on and off deviation periods (5 s each) to improve
the SNR. b: 490 Hz power modulation controlled by the Arduino chip. The duty cycle
controls the average acoustic power experienced by the particles as they travel through the
sorting chip. c: 40 MHz sinusoidal wave to generate the SAW.
(data in SI). The electrodes themselves were a symmetric assembly of an a active electrode
and two sensing electrodes. Each of the electrodes was 20 µm wide. The fabrication process
is detailed in the supplementary information.
We probed the channel resistance by powering the Wheatstone bridge with a 5 Vpp AC-
excitation at 50 kHz followed by lock-in demodulation with a gain of 1,000 (SCITEC 441).
Results and discussion
In order to clearly compare the current method with earlier approaches based on video
analysis, we first recorded the deflection of particles at the outlet of the sorting section over
a range of power levels. Video analysis indicates that sorting occurs if the particle position
in the channel exceeds 382± 40 µm. Hence, according to Fig. 6 the polystyrene microspheres
are sorted for power levels exceedingE= 11.4 J/m3, whereas the cells have no clear threshold
even though some deflection occurs as early asE= 9.1 J/m3.
Next, we compare this video analysis to the RPS chip output. A typical signal trace
17
μ(a)(b)(c) Figure 6: Video analysis of particles deflection for a range of power levels. The x axis
represents the microspheres (a) or cells (b) position just before exiting the sorting channel,
while the y axis indicates the relative power levelE. For each bin delimiting a power level
(±4.5%) and position ((±8.3 µm), the number of particles or cells leaving the channel over a
1s interval is indicated by the color scale.
18
(a)(b) (after amplification) is shown in Fig. 7. Each half peak lasts approximately 2.3 ms. In the
experiment, we used the 7.32 µm diameter microspheres (V = 205 µm3) as a calibration
standard to establish the voltage-volume relation coefficient, and obtained 170 µm3/V. The
experimental standard deviation of the particle size distribution is inferior to the data from
manufacturer (see SI), which supports that our design does not introduce additional bias
due to the vertical position of the particles in the channel. 27
Figure 7: Waveform trace recorded from the RPS chip after amplification. The smaller peak
on the left is a single PS particle while the peak on the right cannot be identified without
knowledge of the acoustic power level.
Over the course of 4 hours, we recorded the arrival rate of particles. Except for very few
outliers, most particles volume ranged from 150 µm3 up to 1,600 µm3. Fig. 8 is a composed
histogram showing the number of particles with a volume smaller than 350 µm3 arriving over
time intervals (75 s) much larger than the power fluctuation period (10 s). This attenuates
the stochastic nature of the power sequence and demonstrates that the particles arrive in the
detector at a constant rate (top histogram). This population of particles is further subdivided
into smaller groups of identical volumes which yields the center two-dimensional histogram.
This graphic shows a downward trend that indicates a slight decay in the average detected
particle volume over time (larger particles sediment faster). Finally, the histogram on the
right indicates the total number of particles detected over 4 hours and is representative of
the size distribution of the PS microspheres.
Unlike these small PS beads, the arrival rate of larger objects decays much faster as
shown in Fig. 9. We believe this decay is due to the sedimentation of the particles despite a
19
Figure 8: Arrival rate of particles with a volume ranging from 150 µm3 up to 350 µm3. The
heatmap at the center indicates the number of particles counted in the time interval t± 324 s
with a volume V± 37 µm3. The bar graph at the top represents the arrival time distribution
of the entire population of particles, while the bar graph on the right represents the volume
distribution of the entire population of particles regardless of their arrival time.
20
continuous stirring of the liquid reservoir. 47 According to the literature, Jurkat cells diameter
(volume) ranges between 10.5 and 12.5 µm (600 up to 1000 µm3). Hence, the objects with
a volume below 600 µm3 are suspected to be PS microspheres doublet. This hypothesis will
be verified thereafter using compressibility data.
Figure 9: Arrival rate of particles with a volume ranging from 350 µm3 up to 1,600 µm3.
The heatmap at the center indicates the number of particles counted in the time interval
t± 324 s with a volume V ± 32 µm3. The histogram at the top represents the arrival time
distribution of the entire population of particles, while the histogram on the right represents
the volume distribution of the entire population of particles regardless of their arrival time.
Calculation of the deviation thresholds
In the previous section, we have exposed a stream of microspheres and cells to increasing
levels acoustic power until a deviation thresholdEmin was reached, at which point the mi-
yields exactly M = sin θC . However, knowing M is not enough to immediately deduce the
crospheres and then the cells started to be deflected towards the analysis chip. According to
Eq. 7, the deviation only depends on the number M (Eq. (8)), therefore the onset of sorting
particle acoustic properties. Besides acoustic contrast, M also depends on external factors
21
such as the flow velocity, fluid viscosity and acoustic energy density, but also on the particle
radius. The latter is directly evaluated with the RPS sensor (assuming a spherical geometry
for the particles). Hence, the acoustic contrast is obtained from:
Φexp= 9ηvF sinθC
2a2kXEmin
(22)
Once the acoustic contrast is known, recovering f1 and f2 (knowing the particle density) is
straightforward.
In the following, we first use the linear method developed above to recover the deviation
threshold for the polystyrene microspheres, then apply the statistical methods to analyze less
abundant particles. Eventually, the acoustic contrast of each type of particle is recovered
and the particle compressibility is computed.
Linear method
The systemDG− B2 (Eqs. (15-17)) is minimized with the nonzero least square solver of
Octave. The reshaped version of G is presented in Fig. 10.
According to Fig. 10, the PS microspheres take between 70 s and 120 s to flow along the
4.5 mm tubing between the sorting and the sensing sections. Almost no particles are sorted
unless the relative acoustic energyE exceeds 13.7 J/m3 which is comparable to the results
from the video analysis (11.4 J/m3).
Statistical method
When the number of particles is too low, the linear model yields inaccurate results, so the
statistical model should be used instead. We recall that this model studies the difference of
average acoustic energy density in the moments that precede the detection of a particle (as
compared to when no particle was detected). The analytical formula from Eq. 18 is first
validated against numerical simulations, then we assess the robustness of the method against
22
Figure 10: PS particles (7.32 µm diameter) sorting impulse response estimated from the
linear regression solution G. The heatmap at the center shows the impulse response as a
function of the power level and the delay after application of the pulse. The histogram on the
top represents the cumulated impulse response across all power levels while the histogram
on the right indicates the likelihood of deviation at a given power level.
23
statistical noise, and then use this method to analyze experimental results.
In order to validate the statistical method, we simulated the sorting and detection of
particles with a hard threshold below which no particles are sorted and above which all
the particles are sorted. Since the PS microspheres were detected within 150 s (15 random
power levels), we also used nt= 15 in these simulations. The results of 200,000 simulations
of sorting events and 200,000 negative controls are depicted in Fig. 11 by the blue and red
dots respectively. The linear trend of these dots compares well to the analytical formula
from Eq. 18 (solid lines).
Due to the statistical nature of the model, we then wanted to evaluate its reliability:
the 200,000 simulations were grouped into 1,000 sets of 200 particles, yielding 1,000 possible
outcomes. The standard deviation between these outcomes is also presented in Fig. 11 in
a series of shades. Each shade represents a standard deviation between the 1,000 outcomes.
This process was repeated for each discrete level of acoustic power used in our experiments,
which yields the bands shown in Fig. 11. According to the simulations, differences in sorting
threshold above 5 J/m3 should exceed a standard deviation.
The experimental results of the PS microspheres, cells and PS microsphere doublets are
also reported on Fig. 11 with the associated error bars based on the standard deviation
from the simulations (adjusted based on the number of observed particles in each case). The
three types particles have a deviation threshold that differs by less than a standard deviation.
Nonetheless the size of the particles has not yet been accounted for. These thresholds can
also be compared to the video analysis and the physical model used earlier to analyze the
deviation of PS microspheres. This statistical approach yields a deviation threshold of 9.8
J/m3, which is reasonably close to the video analysis (11.4 J/m3) and lower than the physical
model (13.4 J/m3). Yet, all three methods agree within 20%.
24
Figure 11: Statistical inference of the particle deviation threshold. We simulated 1,000
independent experiments in which N = 200 particles are observed, and for each particle
the power is averaged over nt = 15 discrete time steps. Of these 15 time steps, only 1 is
expected to determinate the sorting result of the particle. The solid blue (red) line represent
the analytical result given by equation (18), and the dots show the population and time-
averaged power levels for the particles (control) averaged over 1,000 simulations. Each degree
of shaded blue (red) areas indicate a standard deviation between simulated experiments.
The crosses represent the experimental results from a single experiment involving 200 cells
or particles doublet and more than 3,000 PS particles. The width of the cross is the standard
deviation obtained from the simulations using similar population sizes.
Calculation of scattering coefficients
Once the particle size and deviation threshold are known, we use Eq. (22) to recover the
particles acoustic contrast. f1 and f2 are then immediately obtained which allows computing,
the particles compressibility. The results are presented in Table 1 and Fig. 12.
The polystyrene microspheres compressibility (obtained from the statistical method) is
very close to the tabulated value 19 of 2.2× 10−10 Pa−1. The microspheres doublets com-
pressibility is estimated to be 2.20× 10−10 Pa−1, which is almost the same as for the PS
microspheres. We also note that the f2 coefficient (that depends on the density ratio) is gen-
erally much smaller than the compressibility-related f1 coefficient. Hence, the exact value
of the particle density is not critical for the results accuracy. Even though the cells and the
doublets had a similar deviation threshold, the larger size of the cells yields a very different
compressibility (3.28× 10−10 Pa−1). Furthermore, despite the small number of cells detected
during the experiment, the estimated compressibility of Jurkat cells is consistent with earlier
25
iPS x2cellsPS studies and intermediate between red blood cells (κp = 3.18× 10−10 Pa−1) and MCF-12A
(κp= 3.54× 10−10 Pa−1). 19
Table 1: Analysis of cells and particles deviation. (a) acoustic energy density estimated from
the 0.52 nmRM S displacement.
particle
PS (video)
PS (linear)
PS (statistical)
cells (video)
cells (statistical)
PS×2 (statistical)
ρp
(kg/m3)
1062
1062
1062
1075
1075
1062
min
(J/m3)
E(a)
9.8± 1.1
10.2± 0.9
8.5± 3.6
6.2± 3.6
11.4
13.4
Vp
(µm3)
210
210
210
800
800
400
Φexp
f1
f2
0.398
0.339
0.462
0.175
0.213
0.472
0.441
0.390
0.505
0.226
0.264
0.515
0.0397
0.0397
0.0397
0.0476
0.0476
0.0397
(×1010, Pa−1)
κp
2.49
2.72
2.21
3.45
3.28
2.20
Our final results are synthesized in Fig. 12. The three methods (video, linear and statis-
tical) yield slightly different results for the compressibility of PS microspheres. Nonetheless,
PS microspheres are clearly distinct from other kinds of particles in terms of size and can
be identified with certainty. The cells and PS doublets show a slight overlap in size and
compressibility, but the combination of both parameters lifts the ambiguity and indicates
more clearly that these two populations do refer to two different types of particles. This
result could not have been obtained from any of these two methods alone.
Figure 12: Experimental particle size and acoustic contrast. The analysis method is indicated
by the symbol color, with "stats" standing for "statistical".
26
6789101112131415particle diameter (μm)0.00.51.01.52.02.53.03.54.0particle compressibility (Pa−1)1e−10statisticalvideolinear Perspectives
This work marks a first step towards the integration of RPS sensing and acoustophoresis on
a single chip. Since the proposed method does not require high-speed camera or microscopes,
a natural continuation would be to integrate the acoustophoresis and RPS in a single chip
for point-of-care diagnostic. From a more fundamental point of view, acoustophoresis and
RPS are not restricted by the diffraction limit, hence combining these two technologies may
allow probing the mechanical properties of nano-objects such as nanoparticles, exosomes
and viruses. Nonetheless, the current system still faces several challenges that need to be
addressed before the technology reaches its full potential for point-of-care applications and
nanoparticle analysis.
In our opinion, the two major limitations of the current device are (i) that it is not yet
truly single-cell and (ii) that the density of the particles has to be calibrated in a different
experiment. The single-cell limitation stems from the delay between sorting and detection,
which is mostly due to the tubing interconnect between the sorting and RPS chips. This
tubing generates a Taylor diffusion such that individual detection events cannot be directly
linked to the acoustic energy density. This issue should disappear once the sorting and
sensing functions are integrated on a single chip (thereby eliminating the tubing and thus
the delay and need for probabilistic models). Since both chips are fabricated using the
same process on similar substrates, such integration may be within reach. Regarding the
need to know the particle density, we anticipate two approaches. According to Eq. (2),
choosing θR= 45o makes Φ independent of the particle density. Such Rayleigh angle can be
achieved by lowering the SAW velocity, for instance by switching material or using thinner
substrates. An alternative approach would be measuring the speed of sedimentation in a
configuration similar to Grenvall et al. 27 In this work, the authors showed that the vertical
position of the particles can be sensed by an RPS system with neighboring electrodes, thus
a pair of such electrodes can measure the sedimentation speed of the particles and thus their
density. Besides density, compressibility and size, higher-end lock-in amplifier may also allow
27
to record the cell electrical impedance at various frequencies with a similar setup.
Conclusion
Resistive pulse sensing has long been limited for the measurement of cells and particles me-
chanical properties. In this work, we used acoustophoresis to provide the mechanical insight
to the RPS. This required several theoretical and technological advances, including study-
ing the deviation of particles in a tilted-angle acoustic field, designing a modular two-chips
experiment and accounting for the time delay between particle sorting and detection when
analyzing the data. Our results were scrutinized by three different methods, which approxi-
mately agreed on the particle compressibility. Furthermore, in contrast to constriction-based
methods, mechanical phenotyping can be performed over a much broader range of particle
size and elasticity, including cells and solid particles. With further integration, this strategy
could yield point-of-care mechanical phenotyping devices and allow analyzing nanoparticles,
exsosomes and viruses.
Acknowledgement
The authors gratefully acknowledge Gabriele Pitingolo, Shufang Renault and Leonard Jagot
Lagoussiere for their useful discussion, Aloysa Guerra and Catherine Dode for the generous
gift of Jurkat cells, Philippe Nizard for his help with cell manipulation and Michael Baudoin
for his fruitful discussions.This research was supported by the Fondation pour la Recherche
Medicale (SPF20160936257)
Supporting Information Available
The following files are available free of charge.
28
Fabrication of the microchannels, fabrication of the interdigitated transducers, calibration
of the SAW power and evaluation of the electric field homogeneity in the sensing channel.
References
(1) Guo, J.; Huang, X.; Ai, Y. On-demand lensless single cell imaging activated by differ-
ential resistive pulse sensing. Analytical chemistry 2015, 87, 6516 -- 6519.
(2) Harms, Z. D.; Mogensen, K. B.; Nunes, P. S.; Zhou, K.; Hildenbrand, B. W.; Mitra, I.;
Tan, Z.; Zlotnick, A.; Kutter, J. P.; Jacobson, S. C. Nanofluidic devices with two pores
in series for resistive-pulse sensing of single virus capsids. Analytical chemistry 2011,
83, 9573 -- 9578.
(3) Zhou, J.; Kondylis, P.; Haywood, D. G.; Harms, Z. D.; Lee, L. S.; Zlotnick, A.; Ja-
cobson, S. C. Characterization of Virus Capsids and Their Assembly Intermediates
by Multicycle Resistive-Pulse Sensing with Four Pores in Series. Analytical chemistry
2018, 90, 7267 -- 7274.
(4) Weatherall, E.; Hauer, P.; Vogel, R.; Willmott, G. R. Pulse size distributions in tunable
resistive pulse sensing. Analytical Chemistry 2016, 88, 8648 -- 8656.
(5) Willmott, G. R. Tunable Resistive Pulse Sensing: Better Size and Charge Measurements
for Submicrometer Colloids. Analytical chemistry 2018, 90, 2987 -- 2995.
(6) Billinge, E. R.; Broom, M.; Platt, M. Monitoring Aptamer -- Protein Interactions Using
Tunable Resistive Pulse Sensing. Analytical chemistry 2013, 86, 1030 -- 1037.
(7) Clarke, J.; Wu, H.-C.; Jayasinghe, L.; Patel, A.; Reid, S.; Bayley, H. Continuous base
identification for single-molecule nanopore DNA sequencing. Nature nanotechnology
2009, 4, 265.
29
(8) Braylan, R. C.; Fowlkes, B. J.; Jaffe, E. S.; Sanders, S. K.; Berard, C. W.; Herman, C. J.
Cell volumes and DNA distributions of normal and neoplastic human lymphoid cells.
Cancer 1978, 41, 201 -- 209.
(9) Chapman, E. H.; Kurec, A. S.; Davey, F. Cell volumes of normal and malignant mononu-
clear cells. Journal of clinical pathology 1981, 34, 1083 -- 1090.
(10) Suresh, S. Biomechanics and biophysics of cancer cells. Acta Materialia 2007, 55, 3989 --
4014.
(11) Darling, E. M.; Di Carlo, D. High-throughput assessment of cellular mechanical prop-
erties. Annual review of biomedical engineering 2015, 17, 35 -- 62.
(12) Koutsouris, D.; Guillet, R.; Lelievre, J.; Guillemin, M.; Bertholom, P.; Beuzard, Y.;
Boynard, M. Determination of erythrocyte transit times through micropores. I-Basic
operational principles. Biorheology 1988, 25, 763 -- 772.
(13) Zheng, Y.; Shojaei-Baghini, E.; Azad, A.; Wang, C.; Sun, Y. High-throughput biophys-
ical measurement of human red blood cells. Lab on a Chip 2012, 12, 2560 -- 2567.
(14) Zheng, Y.; Wen, J.; Nguyen, J.; Cachia, M. A.; Wang, C.; Sun, Y. Decreased deforma-
bility of lymphocytes in chronic lymphocytic leukemia. Scientific reports 2015, 5, 7613.
(15) Zhou, Y.; Yang, D.; Zhou, Y.; Khoo, B. L.; Han, J.; Ai, Y. Characterizing Deformability
and Electrical Impedance of Cancer Cells in a Microfluidic Device. Analytical chemistry
2017, 90, 912 -- 919.
(16) Raj, A.; Dixit, M.; Doble, M.; Sen, A. A combined experimental and theoretical ap-
proach towards mechanophenotyping of biological cells using a constricted microchan-
nel. Lab on a Chip 2017, 17, 3704 -- 3716.
(17) Ye, T.; Shi, H.; Phan-Thien, N.; Lim, C. T.; Li, Y. Relationship between transit time
30
and mechanical properties of a cell through a stenosed microchannel. Soft matter 2018,
14, 533 -- 545.
(18) Ruban, G.; Goncharova, N.; Marinitch, D.; Loiko, V. Mononuclears Size-Distribution
as Marker of Acute Leukemia. International Journal of Advance in Medical Science
2015, 3, 1 -- 12.
(19) Hartono, D.; Liu, Y.; Tan, P. L.; Then, X. Y. S.; Yung, L.-Y. L.; Lim, K.-M. On-chip
measurements of cell compressibility via acoustic radiation. Lab on a Chip 2011, 11,
4072 -- 4080.
(20) Augustsson, P.; Karlsen, J. T.; Su, H.-W.; Bruus, H.; Voldman, J. Iso-acoustic focusing
of cells for size-insensitive acousto-mechanical phenotyping. Nature communications
2016, 7, 11556.
(21) Riaud, A.; Baudoin, M.; Bou Matar, O.; Becerra, L.; Thomas, J.-L. Selective Manipula-
tion of Microscopic Particles with Precursor Swirling Rayleigh Waves. Physical Review
Applied 2017, 7, 024007.
(22) Baudoin, M.; Gerbedoen, J.-C.; Riaud, A.; Matar, O. B.; Smagin, N.; Thomas, J.-L.
Folding a focalized acoustical vortex on a flat holographic transducer: Miniaturized
selective acoustical tweezers. Science Advances 2019, 5.
(23) Williams, P. S.; Martin, M.; Hoyos, M. Acoustophoretic Mobility and Its Role in Op-
timizing Acoustofluidic Separations. Analytical chemistry 2017, 89, 6543 -- 6550.
(24) Petersson, F.; Aberg, L.; Sward-Nilsson, A.-M.; Laurell, T. Free flow acoustophoresis:
microfluidic-based mode of particle and cell separation. Analytical chemistry 2007, 79,
5117 -- 5123.
(25) Augustsson, P.; Magnusson, C.; Nordin, M.; Lilja, H.; Laurell, T. Microfluidic, label-
31
free enrichment of prostate cancer cells in blood based on acoustophoresis. Analytical
chemistry 2012, 84, 7954 -- 7962.
(26) Muller, P. B.; Rossi, M.; Marin, A.; Barnkob, R.; Augustsson, P.; Laurell, T.;
Kaehler, C. J.; Bruus, H. Ultrasound-induced acoustophoretic motion of microparti-
cles in three dimensions. Physical Review E 2013, 88, 023006.
(27) Grenvall, C.; Antfolk, C.; Bisgaard, C. Z.; Laurell, T. Two-dimensional acoustic particle
focusing enables sheathless chip Coulter counter with planar electrode configuration.
Lab on a Chip 2014, 14, 4629 -- 4637.
(28) Kishor, R.; Seah, Y. P.; Zheng, Y. J.; Xia, H.; Wang, Z.; Lu, H. J.; Lim, T. T. Character-
ization of an acoustically coupled multilayered microfluidic platform on SAW substrate
using mixing phenomena. Sensors and Actuators A: Physical 2015, 233, 360 -- 367.
(29) Guo, F.; Xie, Y.; Li, S.; Lata, J.; Ren, L.; Mao, Z.; Ren, B.; Wu, M.; Ozcelik, A.;
Huang, T. J. Reusable acoustic tweezers for disposable devices. Lab on a Chip 2015,
15, 4517 -- 4523.
(30) Ma, Z.; Collins, D. J.; Ai, Y. Detachable Acoustofluidic System for Particle Separation
via a Traveling Surface Acoustic Wave. Analytical Chemistry 2016, 88, 5316 -- 5323,
PMID: 27086552.
(31) Collins, D. J.; Alan, T.; Neild, A. Particle separation using virtual deterministic lateral
displacement (vDLD). Lab on a Chip 2014, 14, 1595 -- 1603.
(32) Ding, X.; Peng, Z.; Lin, S.-C. S.; Geri, M.; Li, S.; Li, P.; Chen, Y.; Dao, M.;
Suresh, S.; Huang, T. J. Cell separation using tilted-angle standing surface acoustic
waves. Proceedings of the National Academy of Sciences 2014, 111, 12992 -- 12997.
(33) Li, P.; Mao, Z.; Peng, Z.; Zhou, L.; Chen, Y.; Huang, P.-H.; Truica, C. I.; Drabick, J. J.;
32
El-Deiry, W. S.; Dao, M.; Suresh, S.; Huang, T. J. Acoustic separation of circulating
tumor cells. Proceedings of the National Academy of Sciences 2015, 112, 4970 -- 4975.
(34) Wu, M.; Ouyang, Y.; Wang, Z.; Zhang, R.; Huang, P.-H.; Chen, C.; Li, H.; Li, P.;
Quinn, D.; Dao, M. Isolation of exosomes from whole blood by integrating acoustics
and microfluidics. Proceedings of the National Academy of Sciences 2017, 114, 10584 --
10589.
(35) Simon, G.; Andrade, M. A.; Reboud, J.; Marques-Hueso, J.; Desmulliez, M. P.;
Cooper, J. M.; Riehle, M. O.; Bernassau, A. L. Particle separation by phase modu-
lated surface acoustic waves. Biomicrofluidics 2017, 11, 054115.
(36) Pelton, M.; Ladavac, K.; Grier, D. G. Transport and fractionation in periodic potential-
energy landscapes. Phys. Rev. E 2004, 70, 031108.
(37) Coulter, W. H. Means for counting particles suspended in a fluid. 1953; US Patent
2,656,508.
(38) DeBlois, R.; Bean, C. Counting and sizing of submicron particles by the resistive pulse
technique. Review of Scientific Instruments 1970, 41, 909 -- 916.
(39) Fern´andez-Ramos, A. A.; Marchetti-Laurent, C.; Poindessous, V.; Antonio, S.; Laurent-
Puig, P.; Bortoli, S.; Loriot, M.-A.; Pallet, N. 6-mercaptopurine promotes energetic
failure in proliferating T cells. Oncotarget 2017, 8, 43048.
(40) Zipursky, A.; Bow, E.; Seshadri, R. S.; Brown, E. J. Leukocyte density and volume in
normal subjects and in patients with acute lymphoblastic leukemia. Blood 1976, 48,
361 -- 371.
(41) Riaud, A.; Baudoin, M.; Matar, O. B.; Thomas, J.-L.; Brunet, P. On the influence of
viscosity and caustics on acoustic streaming in sessile droplets: an experimental and a
33
numerical study with a cost-effective method. Journal of Fluid Mechanics 2017, 821,
384 -- 420.
(42) Royer, D.; Dieulesaint, E. Elastic waves in solids, vol. 1; Springer-Verlag Berlin Heidel-
berg, 2000; Chapter 5, pp 337 -- 339.
(43) Toru, S. R´ealisation d'une pince acoustofluidique pour la manipulation de bioparticules.
Ph.D. thesis, Ecully, Ecole centrale de Lyon, 2014.
(44) Jo, M. C.; Guldiken, R. Effects of polydimethylsiloxane (PDMS) microchannels on
surface acoustic wave-based microfluidic devices. Microelectronic Engineering 2014,
113, 98 -- 104.
(45) Ha, B. H.; Lee, K. S.; Destgeer, G.; Park, J.; Choung, J. S.; Jung, J. H.; Shin, J. H.;
Sung, H. J. Acoustothermal heating of polydimethylsiloxane microfluidic system.
Scientific reports 2015, 5, 11851.
(46) Ha, B. H.; Park, J.; Destgeer, G.; Jung, J. H.; Sung, H. J. Generation of dynamic
free-form temperature gradients in a disposable microchip. Analytical chemistry 2015,
87, 11568 -- 11574.
(47) Baret, J. A remote syringe for cells, beads and particle injection in microfluidic channels.
Lab on a Chip. Chips & Tips. http://www. rsc. org/Publishing/Journals/lc/Chips
and Tips/remote syringe. asp 2009,
34
|
1105.2992 | 1 | 1105 | 2011-05-16T02:18:52 | Shock of three-state model for intracellular transport of kinesin KIF1A | [
"physics.bio-ph",
"q-bio.SC"
] | Recently, a three-state model is presented to describe the intracellular traffic of unconventional (single-headed) kinesin KIF1A [Phys. Rev. Lett. {\bf 95}, 118101 (2005)], in which each motor can bind strongly or weakly to its microtubule track, and each binding site of the track might be empty or occupied by one motor. As the usual two-state model, i.e. the totally asymmetric simple exclusion process (TASEP) with motor detachment and attachment, in steady state of the system, this three-state model also exhibits shock (or domain wall separating the high-density and low density phases) and boundary layers. In this study, using mean-field analysis, the conditions of existence of shock and boundary layers are obtained theoretically. Combined with numerical calculations, the properties of shock are also studied. This study will be helpful to understand the biophysical properties of the collective transport of kinesin KIF1A. | physics.bio-ph | physics | Shock of three-state model for intracellular transport of
kinesin KIF1A
Yunxin Zhang∗
Shanghai Key Laboratory for Contemporary Applied Mathematics,
Centre for Computational System Biology,
School of Mathematical Sciences, Fudan University, Shanghai 200433, China.
Recently, a three-state model is presented to describe the intracellular traffic of
unconventional (single-headed) kinesin KIF1A [Phys. Rev. Lett. 95, 118101 (2005)],
in which each motor can bind strongly or weakly to its microtubule track, and each
binding site of the track might be empty or occupied by one motor. As the usual
two-state model, i.e. the totally asymmetric simple exclusion process (TASEP) with
motor detachment and attachment, in steady state of the system, this three-state
model also exhibits shock (or domain wall separating the high-density and low density
phases) and boundary layers. In this study, using mean-field analysis, the conditions
of existence of shock and boundary layers are obtained theoretically. Combined with
numerical calculations, the properties of shock are also studied. This study will be
helpful to understand the biophysical properties of the collective transport of kinesin
KIF1A.
PACS numbers: 87.16.Nn, 87.16.A-, 05.60.-k, 05.70.Ln
Keywords: TASEP; molecular motor; shock; domain wall; KIF1A
∗Email: [email protected]
1
1
0
2
y
a
M
6
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
2
9
9
2
.
5
0
1
1
:
v
i
X
r
a
I.
INTRODUCTION
2
Molecular motors are biogenic force generators acting in the nanometer range.
They are responsible for intracellular transport of wide varieties of cargo from one
location to another in eukaryotic cells [1 -- 6]. Linear motors produce sliding move-
ments along filamentous structures called protein tracks; for example, myosin slides
along actin filament [7 -- 10], kinesin [11 -- 15] and dynein [16 -- 20] along microtubule.
Microtubule and filamentary protein actin are protein filaments which form part of
a dual-purpose scaffolding called cytoskeleton, act like struts or girders for the cel-
lular architecture and, at the same time, also serve as tracks for the intracellular
transportation networks.
However, experiments found, one single filamentary track is usually traveled along
by multiple motors. Fundamental understanding of these collective physical phenom-
ena may expose the causes of motor-related diseases (e.g., Alzheimer's disease). In
literatures, these phenomena are usually described by the totally asymmetric simple
exclusion process (TASEP), which is originally proposed in [21], consists of particles
hopping unidirectionally with hard-core exclusion along a 1D lattice. TASEP is one
of many examples for driven systems with stationary nonequilibrium states, which
cannot be described in terms of Boltzmann weights. To model the attachment and
detachment of motors to and from tracks, Parmeggiani et al [22, 23] discussed a class
of driven lattice gas obtained by coupling 1D TASEP to Langmuir kinetics, in which
the attachment and detachment of motors is modeled as particle creation and anni-
hilation respectively. Furthermore, Lipowsky et al [24, 25] suggested a more general
model, in which the diffusion of motors in the cell is considered explicitly. However,
in reality, a motor protein is not a mere particle, but an enzyme whose mechanical
movement is coupled with its biochemical cycle. Therefore, recently, a three-state
model is presented by Nishinari et al to describe the intracellular transport of single-
headed kinesin KIF1A [26, 27]. In which, the microtubule (MT) binding motor might
be in two states: strongly MT binding state and weakly MT binding state, denoted
by S, W respectively. Biochemically, the strongly binding state corresponds to bare
3
motor or ATP binding motor state, and the weakly binding state corresponds to ADP
binding state.
One of the important feature of the collective motion of motors along one single
track is the possible appearance of shock or domain wall, which is defined as the
interface between the low-density and high density regions. For the usual two-state
TASEP, using mean field method, the existence and properties of the shock have
been discussed recently [28]. In this study, similar analysis to the Nishinari's three-
state model will be presented, the efficient and necessary conditions of the existence
of shock will be given theoretically, and with the aid of numerical calculations the
properties of the shock will also be discussed. The method used in this study can be
regarded as a generalization of the one presented in [28].
In the next section, the three-state model and its mean field approximation will
be briefly introduced, and then in Sec. III the conditions of the existence of shock
will be presented. The existence of boundary layers and the properties of shock will
be discussed in Sec. IV and V. Finally, this study will be shortly summarized in Sec.
VI.
II. TASEP WITH THREE INTERNAL STATES
The three-state TASEP given in [27] can be mathematically desccribed as follows.
Let Si and Wi denote the probabilities of finding a molecular motor in the states 1
and 2 at the lattice site i at time t respectively (states 1 and 2 correspond to strongly
bound and weakly bound states of molecular motors). Then Si, Wi are governed by
the following master equations
dSi
dt
=ωa(1 − Si − Wi) − ωhSi − ωdSi + ωsWi
+ ωf Wi−1(1 − Si − Wi) + (1 − c)ωf Wi(Si+1 + Wi+1),
(1)
dWi
dt
= − (ωs + ωf )Wi(1 − Si+1 − Wi+1) + ωhSi
− [ωs + (1 − c)ωf ]Wi(Si+1 + Wi+1)
− ωbWi(2 − Si+1 − Wi+1 − Si−1 − Wi−1)
+ ωb(Wi−1 + Wi+1)(1 − Si − Wi),
4
(2)
where ωa is the rate of a molecular motor binding to the empty lattice site i, i.e. the
transition rate of state 0 to state 1, ωh is the transition rate of state 1 to state 2, i.e.
the rate of ATP hydrolysis, ωd is the transition rate of state 1 to state 0, i.e. the
rate of detachment, ωb is the rate of random Brownian motion. After the release of
ADP, the motor steps forward to the next binding site in front with rate ωf , stays at
the current location with rate ωs. c is an interpolating parameter (0 ≤ c ≤ 1). The
corresponding equations for the left boundary (i = 1) are given by
dS1
dt
=α(1 − S1 − W1) − ωhS1 − γ1S1 + ωsW1 + (1 − c)ωf W1(S2 + W2),
dW1
dt
= − (ωs + ωf )W1 + ωhS1 + cωf W1(S2 + W2)
(3)
(4)
+ ωbW2(1 − S1 − W1) − ωbW1(1 − S2 − W2) − γ2W1,
where α is the rate of attachment of the motors at the left boundary (i.e. the lattice
site i = 1), γ1, γ2 are the rates of detachment of motors in state 1 and state 2 at the
left boundary respectively. The equations for the right boundary (i = N) are give by
dSN
dt
=δ(1 − SN − WN ) + ωf WN −1(1 − SN − WN )
+ ωsWN − ωhSN − β1SN ,
dWN
dt
=ωhSN − ωsWN − β2WN + ωbWN −1(1 − SN − WL)
− ωbWN (1 − SN −1 − WN −1),
(5)
(6)
where δ is the rate of attachment of the motors at the right boundary (i.e. the lattice
site i = N), β1 β2 are the rates of detachment of motors in state 1 and state 2 at the
right boundary respectively.
It should be pointed out that the exclusion process described above is different from
the one discussed in [29], where multiple occupancy of sites is allowed if particles are
5
in different internal states. Here, the multiple occupancy is unallowed. However,
the particles bounding to the lattice site might be in two different states 1 and 2,
corresponding to the strongly bound and weakly bound states. As in [22], attachment
and detachment of a motor are modeled as, effectively, creation and annihilation of
motors on the lattice. Moreover, the transition between states 1 and 2 is described
by the rates ωh, ωs, ωf , the Brownian ratched mechanism is described by rate ωb.
In the large N limit, we can make the continuum
N −1 and x = (i − 1)△x.
Mean Field Approximation:
mean field approximation to Eqs. (1) and (2). Let ∆x = 1
Obviously, 0 ≤ x ≤ 1, since 1 ≤ i ≤ N. Using the Taylor expansion
Xk=1
S(x±∆x) = S(x)+
, W (x±∆x) = W (x)+
(±∆x)k
Xk=1
∂kS(x)
∂xk
+∞
+∞
k!
(±∆x)k
k!
∂kW (x)
∂xk
.
(7)
(8)
(9)
The continuum limits of Eqs. (1) and (2) are then
∂S(x, t)
∂t
∂W (x, t)
=ωa(1 − S − W ) + ωsW − (ωh + ωd)S
+ ωf(cid:18)W − ∆x
+ (1 − c)ωf W (cid:18)S + W + ∆x
∂x (cid:19) (1 − S − W )
∂S(x, t)
∂x
+ ∆x
∂W (x, t)
∂x (cid:19)
+ O(∆x2),
∂W (x, t)
∂t
∂S(x, t)
=cωf W (cid:18)S + W + ∆x
− (ωs + ωf )W + ωhS + O(∆x2).
∂x
+ ∆x
∂W (x, t)
∂x (cid:19)
Thus, the probability density ρ(x, t) = S(x, t) + W (x, t) of finding a molecular motor
at lattice site x at time t satisfies [summing Eqs. (8) and (9)]
∂ρ(x, t)
∂t
=ωa(1 − ρ) − ωdS + ωf
∂W (ρ − 1)
∂x
∆x + O(∆x2).
(10)
As the discussion in [30], in the thermodynamic limit N → ∞, there are three
regimes to be distinguished. If ωa and ωd are of order [1/(N − 1)]α with α < 1, then
at the steady state, the system, Eqs. (9) and (10), reduces to
ωa(1 − W − S) − ωdS = 0,
cωf W (S + W ) − (ωs + ωf )W + ωhS = 0.
(11)
6
So the probability S satisfies
ck(k + 1)S2 + [(k + 1)(l + 1) + n − c(2k + 1)]S + (c − l − 1) = 0,
(12)
and W = 1− (k + 1)S, ρ = S + W , where l = ωs/ωf , n = ωh/ωf , k = ωd/ωa (because
of the particle-hole symmetry, we restrict the discussion to the case ωa ≥ ωd, i.e.
0 ≤ k ≤ 1).
For the cases that ωa and ωd are of order [1/(N − 1)]α but with α > 1, the local
kinetics is negligible and the system will be
W (W + S − 1) = C,
cωf W (S + W ) − (ωs + ωf )W + ωhS = 0
,
(13)
where the constant C is determined by the left or right boundary conditions [31, 32].
The case of the local rates ωa and ωd being of the order 1/(N − 1) is the most
interesting one, and will be investigated further in this study. In the following, we
always assume that the local rates ωa and ωd are of the order 1/(N − 1).
III. THE EXISTENCE OF SHOCK
Let
Ωa =
ωa
∆x
,
Ωd =
ωd
∆x
,
then at steady state, the leading terms of ∆x of Eqs. (9) (10) are
ωf
∂W (ρ − 1)
∂x
+ Ωa(1 − ρ) − ΩdS = 0
cωf W (S + W ) − (ωs + ωf )W + ωhS = 0
ωf
∂W (ρ − 1)
∂x
+ Ωa(1 − ρ) − Ωd(ρ − W ) = 0,
or
cωf W ρ − (ωs + ωf )W + ωh(ρ − W ) = 0.
The second equation implies
W =
nρ
n + l + 1 − cρ
.
(14)
(15)
(16)
(17)
The steady state flux is then proportional to
J = W (1 − ρ) =
nρ(1 − ρ)
n + l + 1 − cρ
.
7
(18)
The same as in [27], at the steady state, we can obtain the left boundary conditions
S(0) =
α − [cα(α − ωs)/ωf ]
cα + ωh
,
W (0) =
α
ωf
,
and the right boundary conditions
S(1) =
ωs + β
ωh
(cid:20)
ωh
ωh + ωs + β −
β
ωf(cid:21) ,
W (1) =
ωh
ωh + ωs + β −
β
ωf
.
(19)
(20)
One can find that Eqs.
(16) involves only the first-order derivatives of ρ and
W with respect to x whereas there are two sets of boundary conditions (19) and
(20). Therefore, if we integrate the equations (16) with the left boundary conditions
(19), the solution (denoted by ρl, Wl, Sl respectively) may not, in general, match
smoothly with the solution (denoted by ρr, Wr, Sr respectively) obtained for the
same equations but with the right boundary conditions (20). This discontinuity
corresponds to a shock or domain wall. However, at any position x, the continuity
condition of motor flux, or equivalently Jl(x) = Jr(x), should be satisfied, where
Jl(x) = W (x−)[1− ρ(x−)] and Jr(x) = W (x+)[1− ρ(x+)]. At the shock position xs,
W (xs−) = Wl(xs),
W (xs+) = Wr(xs),
ρ(xs−) = ρl(xs),
ρ(xs+) = ρr(xs).
So the continuity condition Jl(xs) = Jr(xs) implies
nρl(xs)(1 − ρl(xs))
n + l + 1 − cρl(xs)
= J(xs) =
nρr(xs)(1 − ρr(xs))
n + l + 1 − cρr(xs)
,
(21)
(22)
or
ρl(xs) + ρr(xs) = 1 +
c
n
J(xs) = 1 + γρl(xs)ρr(xs),
(23)
where γ = c
n+l+1 < 1.
From Eqs. (16), one can show the probability ρ satisfies
(γρ2 − 2ρ + 1)ρx = Ωah(1 − γρ)(cid:20)(k + 1)cρ2 −(cid:18) c
γ
+ c + k(l + 1)(cid:19) ρ +
c
γ(cid:21) ,
(24)
8
with Ωah = Ωa/ωh. It can be proved that, for 0 ≤ c ≤ 1 and l ≥ 0, the discriminant
△ :=(cid:20) c
γ
+ c + k(l + 1)(cid:21)2
4(k + 1)c2
γ
−
≥ 0.
So the equation (24) can be reformulated as
(ρ − ρ3)(ρ − ρ4)ρx = −Ωah(k + 1)c(ρ − ρ0)(ρ − ρ1)(ρ − ρ2),
(25)
.
(26)
(27)
(28)
where
ρ0 =
1
γ
,
ρ3,4 =
1 ∓ √1 − γ
γ + c + k(l + 1)i ∓rh c
γ
,
ρ1,2 = h c
γ + c + k(l + 1)i2
One can easily show that, for 0 ≤ c ≤ 1, l ≥ 0 and k ≥ 0,
1
2 ≤ ρ3 ≤ 1.
ρ4 ≥ 1,
ρ2 ≥ 1,
ρ0 ≥ 1,
ρ1 ≤ 1
2(k + 1)c
− 4(k+1)c2
γ
Particularly,
2c
γ
ρ1 =
γ + c + k(l + 1)i +rh c
h c
→
l+n+1+k(l+1) ≤ 1
l+n+1
1
γ + c + k(l + 1)i2
as k → 0,
as c → 0.
− 4(k+1)c2
γ
Moreover, one can easily show that the function f (x) = x − √x2 − x decreases with
x ≥ 1 monotonously. So, for c ≤ (1 − k)(l + 1) + n,
ρ1 =h c
γ + c + k(l + 1)i −rh c
γ
2(k + 1)c
γ + c + k(l + 1)i2
− 4(k+1)c2
−s(cid:18)(k + 1)(l + 1) + n + c
−s(cid:18)(k + 1)(l + 1) + n + c
(cid:19)2
2(k + 1)c
2(k + 1)c
l + n + 1
−
c
(cid:19)2
(cid:19)2
(k + 1)(l + 1) + n + c
2(k + 1)c
(k + 1)(l + 1) + n + c
2(k + 1)c
l + n + 1
c
−s(cid:18)l + n + 1
c
=
≥
≥
=ρ3.
l + n + 1
(k + 1)c
(k + 1)(l + 1) + n + c
2(k + 1)c
−
−
(29)
9
Therefore, in the following, we always assume that ρ0, ρ2, ρ4 ≥ 1, and 1
2 ≤ ρ3 ≤ ρ1 ≤ 1.
The general solutions of Eq. (25) are
F (ρ) = x + C,
(30)
where C is an arbitrary constant and
F (ρ) = −
1
Ωah(k + 1)c
[A lnρ − ρ0 + B lnρ − ρ1 + D lnρ − ρ2] ,
(31)
with
A =
ρ0(ρ0 − 1)
(ρ0 − ρ1)(ρ2 − ρ0)
, B = −
ρ2
1 − 2ρ0ρ1 + ρ0
(ρ0 − ρ1)(ρ1 − ρ2)
, D =
ρ2
2 − 2ρ0ρ2 + ρ0
(ρ0 − ρ2)(ρ1 − ρ2)
.
(32)
So the solution of Eq. (25), which satisfies the left boundary condition ρ(0) = S(0) +
W (0), see Eq. (19), is
F (ρl) = x + F0,
or ρl(x) = F −1(x + F0),
(33)
where F0 = F [ρ(0)]. Similarly, the solution of Eq. (25), which satisfies the right
boundary condition ρ(1) = S(1) + W (1), see Eq. (20), is
F (ρr) = x + F1 − 1,
or ρr(x) = F −1(x + F1 − 1),
(34)
where F1 = F (ρ(1)). In the following, we assume ρl 6≡ ρr (otherwise, there would be
no shock and boundary layers).
Combining (21) (23) (33) (34), one sees that, at the shock position xs
F (ρl(xs)) − F (ρr(xs)) + F1 − F0 − 1 = 0,
ρl(xs) + ρr(xs) = 1 + γρl(xs)ρr(xs).
(35)
These are the efficient and necessary condition of the existence of shock at the position
xs. In other words, at the shock position xs, H(xs, Ωa, k, ωf , l, n, c) :=
F (ρl(xs)) − F (cid:18) 1 − ρl(xs)
1 − γρl(xs)(cid:19) + F1 − F0 − 1 = 0.
(36)
If there exists 0 < xs < 1, such that H(xs, Ωa, k, ωf , l, n, c) = 0, the shock will appear
at xs, and the height of the shock is
εs = ρr(xs) − ρl(xs) =(cid:12)(cid:12)(cid:12)(cid:12)
1 − 2ρl(xs) + γρ2
1 − γρl(xs)
l (xs)
.
(cid:12)(cid:12)(cid:12)(cid:12)
(37)
Let
H(ρ) := F (ρ) − F (cid:18) 1 − ρ
1 − γρ(cid:19) + F1 − F0 − 1 = 0,
then from (31), one can obtain
10
(38)
dH(ρ)
dρ
=
×
=
dF (ρ)
dρ −
d
dρ
F (cid:18) 1 − ρ
1 − γρ(cid:19) =
γρ2 − 2ρ + 1
Ωah(1 − γρ)
[(1 − ρ) − ρ1(1 − γρ)][(1 − ρ) − ρ2(1 − γρ)] + (γ − 1)(ρ − ρ1)(ρ − ρ2)
(ρ − ρ1)(ρ − ρ2)[(1 − ρ) − ρ1(1 − γρ)][(1 − ρ) − ρ2(1 − γρ)]
γ(1 − ρ)2(1 − γρ1)(ρ − ρ3)(ρ − ρ4)(cid:16) 1−ρ1
Ωah(1 − γρ)3(ρ − ρ1)(ρ − ρ2)(cid:16) 1−ρ
1−γρ − ρ1(cid:17)(cid:16) 1−ρ
1−γρ1 − ρ2(cid:17)
1−γρ − ρ2(cid:17)
> 0
if
0 ≤ ρ < 1−ρ1
1−γρ1
,
< 0
> 0
if
if
1−ρ1
1−γρ1
< ρ < ρ1,
ρ1 < ρ ≤ 1.
(39)
Using this property of the function H(ρ), we can obtain the following results:
(I) For 0 ≤ ρ(0) < ρ3 and ρ3 < ρ(1) ≤ ρ1, the conditions of existence of shock in
interval (0, 1) is
F (cid:18) 1 − ρ(1)
1 − γρ(1)(cid:19) < F0 + 1 and F (cid:18) 1 − ρ(0)
1 − γρ(0)(cid:19) < F1 − 1,
(40)
see Fig. 1 (a). From (25) (31), one can find the function F (ρ) increases with ρ for
0 ≤ ρ < ρ3 and ρ1 < ρ ≤ 1, and decreases with ρ for ρ3 < ρ < ρ1. Thus
l (cid:18) 1 − ρ(1)
1 − γρ(1)(cid:19) < 1,
1 − ρ(0)
1 − γρ(0)
1 − γρ(1)(cid:19) < F0 + 1 ⇐⇒ ρ−1
1 − γρ(0)(cid:19) < F1 − 1 ⇐⇒ ρr(0) <
F (cid:18) 1 − ρ(1)
F (cid:18) 1 − ρ(0)
(41)
.
Therefore, the conditions presented in (40) are generalizations of the ones obtained
in [28] for the usual TASEP with motor detachment and attachment:
ρ−1
l (1 − ρ(1)) < 1,
and ρr(0) < 1 − ρ(0).
In fact, if the parameter c = 0 (i.e. γ = 0), the Eq. (24) reduces to
(1 − 2ρ)ρx = Ωah[(l + n + 1) − [(l + n + 1) + k(l + 1)]ρ],
(42)
(43)
F (cid:18) 1 − ρ(1)
F (cid:18) 1 − ρ(0)
1 − γρ(1)(cid:19) < F0 + 1 ⇐⇒ ρ−1
1 − γρ(0)(cid:19) > F1 − 1 ⇐⇒ ρ−1
l (cid:18) 1 − ρ(1)
r (cid:18) 1 − ρ(0)
1 − γρ(1)(cid:19) < 1,
1 − γρ(0)(cid:19) > 0.
Therefore, conditions (44) are also generalizations of the ones for the usual TASEP
[28]:
ρ−1
l (1 − ρ(1)) < 1,
and ρ−1
r (1 − ρ(0)) > 0.
(III) For 0 ≤ ρ(0), ρ(1) ≤ ρ3, the condition of the existence of shock in (0, 1) is
(44)
(45)
(46)
(47)
which is similar as the model discussed in [22] for the usual TASEP. For such reduced
cases, ρ3 = 0.5, and the conditions (40) of existence of shock is reduced to (42).
(II) For 0 ≤ ρ(0) < ρ3 and ρ1 ≤ ρ(1) ≤ 1, the conditions of existence of shock in
interval (0, 1) is
11
F (cid:18) 1 − ρ(1)
1 − γρ(1)(cid:19) < F0 + 1,
see Fig. 1 (b). Similar as in (I),
and F (cid:18) 1 − ρ(0)
1 − γρ(0)(cid:19) > F1 − 1,
ρ−1
l (ρ3) < 1
and
ρr(0) <
1 − ρ(0)
1 − γρ(0)
where ρr(x) is one of the solutions of differential equation (24), which satisfies ρr(1) =
ρ3 and ρr(0) > ρ3. See Fig. 1 (c).
(IV) For ρ(0), ρ(1) > ρ3, there exists no shock in (0, 1). It can be readily verified that
the function f (ρl, ρr) = ρl + ρr−γρlρr−1 increases monotonously with 0 ≤ ρl, ρr ≤ 1,
and f (ρ3, ρ3) = 0. For ρ(0), ρ(1) > ρ3, one knows that ρ3 ≤ min(ρl, ρr) < max(ρl, ρr)
(note: we always assume ρl 6≡ ρr). Thus f (ρl, ρr) = ρl + ρr − γρlρr − 1 > 0. It is to
say that there exists no shock [see (35)].
(V) For 0 ≤ ρ(1) < ρ3 and ρl(0) > ρ3, there exists no shock in (0, 1).
In conclusion, ρ(0) < ρ3 is one necessary condition of the existence of shock in
(0, 1).
IV. THE EXISTENCE OF BOUNDARY LAYER
Generally speaking, if ρl 6≡ ρr and there is no shock in (0, 1), boundary layer will
appear at least at one of the boundaries 0 and 1. Similar as in [28], we have the
following results:
(I) For 0 ≤ ρ(0) < ρ3 and ρ3 < ρ(1) ≤ ρ1: if
F (cid:18) 1 − ρ(1)
1 − γρ(1)(cid:19) > F0 + 1 and F (cid:18) 1 − ρ(0)
1 − γρ(0)(cid:19) < F1 − 1,
there exists boundary layer at the right boundary x = 1, see Fig. 1 (d); if
F (cid:18) 1 − ρ(1)
1 − γρ(1)(cid:19) < F0 + 1 and F (cid:18) 1 − ρ(0)
1 − γρ(0)(cid:19) > F1 − 1,
12
(48)
(49)
there exists boundary layer at the left boundary x = 0 (in these cases, the shock
position xs < 0), see Fig. 2 (b).
In view of the property (39) of function H(ρ), the conditions (48) can be simplified
as
and the conditions (49) can be simplified as
F (cid:18) 1 − ρ(1)
1 − γρ(1)(cid:19) > F0 + 1,
1 − γρ(0)(cid:19) > F1 − 1.
F (cid:18) 1 − ρ(0)
(II) For 0 ≤ ρ(0) < ρ3 and ρ1 ≤ ρ(1) ≤ 1: if
F (cid:18) 1 − ρ(0)
1 − γρ(0)(cid:19) < F1 − 1,
(50)
(51)
(52)
there exists boundary layer at the left boundary x = 0 (i.e. the shock position xs < 0),
see Fig. 2 (c); if
there exists boundary layer at the right boundary x = 1 (i.e.
F (cid:18) 1 − ρ(1)
1 − γρ(1)(cid:19) > F0 + 1,
(53)
the shock position
xs > 1), see Fig. 2 (a).
(III) For 0 ≤ ρ(0), ρ(1) ≤ ρ3 and ρl(1) 6= ρ(1), there exists boundary layer at x = 1.
See Fig. 1 (c), Fig. 2 (d) and Fig. 3 (a). If
1 − ρ(0)
1 − γρ(0)
there is also the boundary layer at the left boundary x = 0.
ρ−1
l (ρ3) < 1
and ρr(0) >
,
(54)
13
(IV) For ρ3 ≤ ρ(0) < 1, 0 ≤ ρ(1) ≤ ρ3 and ρ(0) 6= ρr(0), there exist boundary layers
at both x = 0 and x = 1, see Fig. 3 (b) and Fig. 3 (c). For these cases, ρ(x) = ρr(x)
for 0 < x < 1.
(V) For ρ3 ≤ ρ(0), ρ(1) ≤ 1 and ρ(0) 6= ρr(0), there exists boundary layer at x = 0.
For these cases, ρ(x) = ρr(x) for 0 < x ≤ 1, see Fig. 3 (d) and Fig. 4.
V. THE PROPERTIES OF SHOCK
Finally, we discuss the properties of shock briefly. From the discussion in Sec. III,
one knows that ρ(0) < ρ3 is necessary for the existence of shock. So, at the shock
position xs, ρl(xs) < ρr(xs), see Eq. (23), and the height of the shock is [see Eq. (37)]
εs = ρr(xs) − ρl(xs) =
1 − 2ρl(xs) + γρ2
1 − γρl(xs)
l (xs)
.
The derivative of the height εs with respect to ρl(xs) is
∂εs
∂ρl(xs)
=
γ − 1
(1 − γρl(xs))2 − 1 < 0.
At the same time, ρl(0) = ρ(0) < ρ3 means ∂ρl(xs)
∂xs
> 0 [see (25)]. So
∂εs
∂xs
=
∂εs
∂ρl(xs)
∂ρl(xs)
∂xs
< 0,
(55)
(56)
(57)
which means, the shock height εs decreases with the shock position xs. Therefore, we
only need to give the relations between shock position xs and the model parameters
Ωa, Ωd, ωf , ωs, ωh, ωb, c.
Because of the complexity of the function F [see (31)], it is difficult to get theoreti-
cal results as in [28]. From numerical calculations, we find that, the shock position xs
decreases with parameters Ωa, α, ωb, ωs, but increases with parameters Ωd, β, ωf , ωh, c,
see Fig. 5. In the calculations, xs is obtained by (36) and (33).
VI. CONCLUDING REMARKS
In this study, the three-state process, which is presented in [26, 27] to model
the intracellular transport of single-headed kinesin KIF1A , is theoretically analyzed
14
using mean field approximation. By similar methods as for the usual TASEP [28], the
conditions of the existence of shock or domain wall, which is defined as the interface
of low-density and high-density phases, are obtained. With the aid of numerical
calculations, the parameters dependent properties of the shock are also discussed.
The results obtained in this study will be helpful to understand the real biophysical
properties of motor traffic in eukaryotic cells.
Acknowledgments
This study is funded by the Natural Science Foundation of Shanghai (under Grant
No. 11ZR1403700).
[1] J. Howard. Mechanics of Motor Proteins and the Cytoskeleton. Sinauer Associates and
Sunderland, MA, 2001.
[2] D. Bray. Cell movements: from molecules to motility, 2nd Edn. Garland, New York,
2001.
[3] G. M. Cooper. The Cell: A Molecular Approach, 2nd Edn. Sinauer Associates, Inc.,
Sunderland, Mass., 2000.
[4] M. Schliwa. Molecular Motors. Wiley-Vch, Weinheim, 2003.
[5] A. O. Sperry. Molecular Motors: Methods and Protocols (Methods in Molecular Biology
Vol 392). Humana Press Inc., Totowa, New Jersey, 2007.
[6] R. D. Vale. The molecular motor toolbox for intracellular transport. Cell, 112:467 -- 480,
2003.
[7] A. D. Mehta, R. S. Rock, M. Rief, J. A. Spudich, M. S. Mooseker, and R. E. Cheney.
Myosin-V is a processive actin-based motor. Nature, 400:590 -- 593, 1999.
[8] A. M. Hooft, E. J. Maki, K. K. Cox, and J. E. Baker. An accelerated state of myosin-
based actin motility. Biochemistry, 46:3513 -- 3520, 2007.
15
[9] J. C. M. Gebhardt, A. E.-M. Clemen, J. Jaud, and M. Rief. Myosin-V is a mechanical
ratchet. Proc. Natl. Acad. Sci. USA, 103:8680 -- 8685, 2006.
[10] K. Shiroguchi and Jr. K. Kinosita. Myosin V walks by lever brownian motion. Science,
316:1208 -- 1212, 2007.
[11] S. M. Block, L. S. B. Goldstein, and B. J. Schnapp. Bead movement by single kinesin
molecules studied with optical tweezers. Nature, 348:348 -- 352, 1990.
[12] R. D. Vale, T. Funatsu, D. W. Pierce, L. Romberg, Y. Harada, and T. Yanagida.
Direct observation of single kinesin molecules moving along microtubules. Nature,
380:451 -- 453, 1996.
[13] A. Yildiz, M. Tomishige, R. D. Vale, and P. R. Selvin. Kinesin walks hand-over-hand.
Science, 303:676 -- 678, 2004.
[14] N. J. Carter and R. A. Cross. Mechanics of the kinesin step. Nature, 435:308 -- 312,
2005.
[15] N. R. Guydosh and S. M. Block. Direct observation of the binding state of the kinesin
head to the microtubule. Nature, 08259, 2009.
[16] R. Mallik, B. C. Carter, S. A. Lex, S. J. King, and S. P. Gross. Cytoplasmic dynein
functions as a gear in response to load. Nature, 427:649 -- 652, 2004.
[17] S. L. Reck-Peterson, A. Yildiz, A. P. Carter, A. Gennerich, N. Zhang, and R. D. Vale.
Single-molecule analysis of dynein processivity and stepping behavior. Cell, 126:335 --
348, 2006.
[18] S. Toba, T. M. Watanabe, L. Yamaguchi-Okimoto, Y. Y. Toyoshima, and H. Higuchi.
Overlapping hand-over-hand mechanism of single molecular motility of cytoplasmic
dynein. Proc. Natl. Acad. Sci. USA, 103:5741 -- 5745, 2006.
[19] A. Gennerich, A. P. Carter, S. L. Reck-Peterson, and R. D. Vale. Force-induced bidi-
rectional stepping of cytoplasmic dynein. Cell, 131:952 -- 965, 2007.
[20] A. Houdusse and A. P. Carter. Dynein swings into action. Cell, 136:395 -- 396, 2009.
16
[21] C. MacDonald, J. Gibbs, and A. Pipkin. Kinetics of biopolymerization on nucleic acid
templates. Biopolymers, 6:1 -- 25, 1968.
[22] A. Parmeggiani, T. Franosch, and E. Frey. Phase coexistence in driven one-dimensional
transport. Physical Review Letters, 90:086601, 2003.
[23] A. Parmeggiani, T. Franosch, and E. Frey. Totally asymmetric simple exclusion process
with langmuir kinetics. Phys. Rev. E, 70:046101, 2004.
[24] R. Lipowsky, S. Klumpp, and T.M. Nieuwenhuizen. Random walks of cytoskeletal
motors in open and closed compartments. Physical Review Letters, 87:108101, 2001.
[25] R. Lipowsky, Y. Chai, S. Klumpp, S. Liepelt, and M. J. I. Muller. Molecular motor
traffic: From biological nanomachines to macroscopic transport. Physica A, 372:34 -- 51,
2006.
[26] K. Nishinari, Y. Okada, A. Schadschneider, and D. Chowdhury. Intracellular transport
of single-headed molecular motors kif1a. Phys.Rev. Lett., 95:118101, 2005.
[27] P. Greulich, A. Garai, K. Nishinari, A. Schadschneider, and D. Chowdhury. Intracellu-
lar transport by single-headed kinesin kif1a: Effects of single-motor mechanochemistry
and steric interactions. Phys. Rev. E, 75:041905, 2007.
[28] Y. Zhang. Domain wall of the totally asymmetric exclusion process without particle
number conservation. Chin. J. Phys., 48:607 -- 618, 2010.
[29] T. Reichenbach, T. Franosch, and E. Frey. Exclusion processes with internal states.
Phys. Rev. Lett. 74,, 97:050603, 2006.
[30] V. Popkov, A. Rakos, R. D. Willmann, A. B. Kolomeisky, and G. M. Schutz. Localiza-
tion of shocks in driven diffusive systems without particle number conservation. Phys.
Rev. E, 67:066117, 2003.
[31] B. Derrida, M. R. Evans, V. Hakim, and V. Pasquier. Exact solution of a 1d asymmetric
exclusion model using a matrix formulation. J. Phys. A: Math. Gen., 26:1493 -- 1517,
1993.
[32] G. Schutz and E. Domany. Phase transitions in an exactly soluble one-dimensional
exclusion process. J. Stat. Phys., 72:277 -- 304, 1993.
17
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
(1−ρ(0))/(1−γ*ρ(0))
(a)
ρ
1
ρ
3
ρ
(x)
r
ρ
(x)
l
x
s
(1−ρ(1))/(1−γ*ρ(1))
ρ
−1[(1−ρ(1))/(1−γ*ρ(1))]
l
Ω
=25; Ω
=10;
d
a
α=10; β=25;
b=0; c=1;
ω
=145; ω
=155;
s
f
ω
=200 N=101
h
0.2
0.4
0.6
0.8
1
(1−ρ(0))/(1−γ*ρ(0))
(c)
ρ
1
ρ
3
ρ(0)
ρ
(x)
r
x
s
ρ
(x)
l
ρ(1)
Ω
=15; Ω
=10;
d
a
α=15; β=50;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=1001
h
0.2
0.4
0.6
0.8
1
18
(b)
Ω
=15; Ω
=10;
d
a
α=3; β=8;
b=0.5; c=0.5;
=145; ω
ω
f
s
ω
=200 N=101
h
=155;
ρ
(x)
l
x
s
(1−ρ(1))/(1−γ*ρ(1))
(1−ρ(0))/(1−γ*ρ(0))
ρ
(x)
r
ρ
1
ρ
3
ρ
−1((1−ρ(1))/(1−γ*ρ(1)))
l
0.2
0.4
0.6
0.8
1
Ω
=10; Ω
=2;
d
a
α=1; β=15;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=101
h
(d)
ρ(1)
(1−ρ(1))/(1−γ*ρ(1))
ρ
1
ρ
3
0.2
0.4
0.6
0.8
1
ρ
(x)
l
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1−γρ(0)(cid:17) < F1−1 (a); Or 0 ≤ ρ(0) < ρ3, ρ1 ≤ ρ(1) ≤ 1 and F (cid:16) 1−ρ(1)
1−γρ(1)(cid:17) <
1−γρ(1)(cid:17) < F0 +1,
1−γρ(0)(cid:17) > F1 − 1 (b). There exists shock in (0, 1) and boundary layer at x = 1 if
FIG. 1: There exists shock in (0, 1) if 0 ≤ ρ(0) < ρ3, ρ3 < ρ(1) ≤ ρ1 and F (cid:16) 1−ρ(1)
F0 +1, F (cid:16) 1−ρ(0)
F (cid:16) 1−ρ(0)
0 ≤ ρ(0), ρ(1) ≤ ρ3 and ρ−1
x = 1 if 0 ≤ ρ(0) ≤ ρ3, ρ3 ≤ ρ(0) ≤ ρ1 and F (cid:16) 1−ρ(1)
1−γρ(1)(cid:17) > F0 + 1 (d).
(ρ3) < 1, ρr(0) < 1−ρ(0)
l
1−γρ(0) (c). There exists boundary layer at
19
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0.45
0
ρ
1
ρ
3
Ω
=5; Ω
=2;
d
a
α=0.1; β=5.5;
b=0.5; c=0.5;
ω
=145; ω
=155;
s
f
ω
=200 N=101
h
ρ(1)
(1−ρ(1))/(1−γ*ρ(1))
ρ
(x)
l
0.2
0.4
0.6
0.8
(a)
1
ρ
(x)
r
(1−ρ(0))/(1−γ*ρ(0))
ρ
3
ρ(0)
ρ
1
Ω
=20; Ω
=5;
d
a
α=10; β=5;
b=0.5; c=0.5;
=145; ω
ω
=155;
f
s
ω
=200 N=101
h
0.2
0.4
0.6
0.8
(c)
1
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
0
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
ρ
1
ρ
3
ρ
1
ρ
(x)
r
(1−ρ(0))/(1−γ*ρ(0))
ρ
3
ρ(0)
Ω
=20; Ω
=10;
d
a
α=20; β=20;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=101
h
0.2
0.4
0.6
0.8
(b)
1
Ω
=16; Ω
=10;
d
a
α=5; β=50;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=201
h
ρ
(x)
l
ρ(1)
0.2
0.4
0.6
0.8
(d)
1
FIG. 2: There exists boundary layers at x = 1 if 0 ≤ ρ(0) ≤ ρ3, ρ1 ≤ ρ(1) ≤ 1 and
F (cid:16) 1−ρ(1)
1−γρ(1)(cid:17) > F0 + 1 (a); Or 0 ≤ ρ(0), ρ(1) ≤ ρ3 and ρl(1) > ρ(1) (d). There exists
boundary layers at x = 0 if 0 ≤ ρ(0) < ρ3, ρ3 < ρ(1) ≤ ρ1 and F (cid:16) 1−ρ(0)
1−γρ(0)(cid:17) > F1 − 1 (b);
Or 0 ≤ ρ(0) ≤ ρ3, ρ1 ≤ ρ(1) ≤ 1 and F (cid:16) 1−ρ(0)
1−γρ(0)(cid:17) < F1 − 1 (c).
0.2
0.4
0.6
0.8
ρ
1
ρ
3
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
ρ(0)
ρ
3
Ω
=16; Ω
=10;
d
a
α=60; β=50;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=801
h
0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
ρ
1
ρ(0)
ρ
3
Ω
=16; Ω
=10;
d
a
α=30; β=50;
b=0.5; c=0.5;
ω
=145; ω
=155;
s
f
ω
=200 N=1001
h
20
(b)
ρ
(x)
r
ρ(x)
ρ(1)
0
0.2
0.4
0.6
0.8
1
0.75
(d)
ρ
(x)
r
ρ
1
0.7
0.65
0.6
0.55
0.5
0
ρ(x)
ρ(1)
Ω
=16; Ω
=10;
d
a
α=30; β=25;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=201
h
ρ(0)
ρ
3
0.2
0.4
0.6
0.8
1
Ω
=16; Ω
=10;
d
a
α=1; β=30;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=201
h
ρ(1)
ρ
(x)
l
(a)
1
(c)
ρ
(x)
r
ρ
1
ρ(x)
ρ(1)
0
0.2
0.4
0.6
0.8
1
FIG. 3: There exists boundary layers at x = 1 if 0 ≤ ρ(0), ρ(1) ≤ ρ3 and ρl(1) < ρ(1) (a).
There exist boundary layers at both x = 0 and x = 1 if ρ3 ≤ ρ(0) < 1, 0 ≤ ρ(1) ≤ ρ3
and ρ(0) < ρr(0) (b); Or ρ3 ≤ ρ(0) < ρ1, 0 ≤ ρ(1) ≤ ρ3 and ρ(0) > ρr(0) (c). There are
boundary layers at x = 0 if ρ3 ≤ ρ(0), ρ(1) ≤ ρ1 and ρ(0) < ρr(0) (d).
ρ(0)
0.85
(a)
0.8
0.75
0.7
0.65
ρ
(x)
r
ρ(x)
Ω
=16; Ω
=10;
d
a
α=60; β=25;
b=0.5; c=0.5;
ω
=145; ω
=155;
s
f
ω
=200 N=201
h
0.6
0.55
0.5
0
ρ(0)
0.85
(b)
0.8
0.75
0.7
ρ
(x)
r
ρ(x)
ρ
1
ρ(1)
ρ
3
21
ρ(1)
ρ
1
Ω
=6; Ω
=5;
d
a
α=60; β=25;
b=0.5; c=0.5;
ω
=145; ω
=155;
s
f
ω
=200 N=201
h
0.2
0.4
0.6
0.8
1
0.2
0.4
0.6
0.8
1
0.85
(c)
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0
ρ
(x)
r
ρ(x)
ρ(0)
ρ
3
ρ(1)
ρ
1
Ω
=6; Ω
=5;
d
a
α=30; β=15;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=201
h
0.2
0.4
0.6
0.8
1
ρ(x)
ρ(0)
ρ
3
ρ(1)
ρ
1
Ω
=6; Ω
=5;
d
a
α=40; β=10;
b=0.5; c=0.5;
=145; ω
ω
=155;
s
f
ω
=200 N=201
h
0.2
0.4
0.6
0.8
1
ρ
3
(d)ρ
(x)
r
0.65
0.6
0.55
0.5
0
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0
FIG. 4: There are boundary layers at x = 0 if ρ3 ≤ ρ(1) ≤ ρ1, ρ1 ≤ ρ(0) ≤ 1 (a); Or
ρ3 ≤ ρ(0), ρ(1) ≤ 1 and ρ(0) > ρr(0) (b); Or ρ3 ≤ ρ(0) ≤ ρ1, ρ1 ≤ ρ(1) ≤ 1 (c); Or
ρ3 ≤ ρ(0), ρ(1) ≤ 1 and ρ(0) < ρr(0) (d).
1
0.9
0.8
0.7
0.6
s
x
0.5
0.4
0.3
0.2
0.1
0
10
0.5
0.45
0.4
0.35
s
x
0.3
0.25
0.2
0.15
0.1
0
0.7
0.6
0.5
0.4
0.3
0.2
0.1
s
x
d=10;
Ω
α=10; β=25;
b=0; c=1;
ω
f=145; ω
s=155;
ω
h=200 N=101
15
20
25
Ω
a
30
35
40
45
d=10;
Ω
a=25; Ω
α=10; β=25;
b=0; c=1;
ω
f=145;
ω
h=200 N=101
50
100
150
ω
s
200
250
300
d=10;
Ω
a=25; Ω
α=10; β=25;
b=0; c=1;
ω
s=155;
ω
h=200 N=101
0
50
100
150
200
250
300
ω
f
0.8
0.7
0.6
0.5
s
x
0.4
0.3
0.2
0.1
0
0
0.6
0.55
0.5
0.45
0.4
s
x
0.35
0.3
0.25
0.2
0.15
0.1
0
0.4
0.35
0.3
0.25
s
x
0.2
0.15
0.1
0.05
0
50
d=10;
Ω
a=25; Ω
β=25;
b=0; c=1;
ω
f=145; ω
s=155;
ω
h=200 N=101
2
4
6
8
α
10
12
14
16
a=25;
Ω
α=10; β=25;
b=0; c=1;
ω
f=145; ω
s=155;
ω
h=200 N=101
5
10
Ω
d
15
20
25
d=10;
Ω
a=25; Ω
α=10; β=25;
b=0; c=1;
ω
f=145; ω
N=101
s=155;
100
150
ω
h
200
250
300
22
d=10;
Ω
a=25; Ω
α=10; β=25;
c=1;
ω
f=145; ω
s=155;
ω
h=200 N=101
0.2
0.4
ω
b
0.6
0.8
1
d=10;
Ω
a=25; Ω
α=2;
b=0; c=1;
ω
f=145; ω
s=155;
ω
h=200 N=101
5
10
15
β
20
25
30
d=10;
Ω
a=25; Ω
α=10; β=25;
b=0;
ω
f=145; ω
s=155;
ω
h=200 N=101
0.2
0.4
c
0.6
0.8
1
0.6683
0.6682
0.6682
s
x
0.6682
0.6682
0.6682
0.6682
0
0.8
0.7
0.6
0.5
s
x
0.4
0.3
0.2
0.1
0
0
0.68
0.66
0.64
0.62
0.6
0.58
0.56
0.54
0.52
0.5
0
s
x
FIG. 5: The shock position xs decreases with parameters Ωa, α, ωb, ωs, but increases with
parameters Ωd, β, ωf , ωh, c. In the calculations, xs is obtained by Eqs. (36) and (33).
|
1303.2669 | 1 | 1303 | 2013-03-11T20:16:36 | Crawling scallop: Friction-based locomotion with one degree of freedom | [
"physics.bio-ph"
] | Fluid-based locomotion at low Reynolds number is subject to the constraints of the scallop theorem, which dictate that body kinematics identical under a time-reversal symmetry (in particular, those with a single degree of freedom) cannot display locomotion on average. The implications of the theorem naturally compel one to ask whether similar symmetry constraints exist for locomotion in different environments. In this work we consider locomotion along a surface where forces are described by isotropic Coulomb friction. To address whether motions with a single degree of freedom can lead to transport, we analyze a model system consisting of two bodies whose separation distance undergoes periodic time variations. The behavior of the two-body system is entirely determined by the kinematic specification of their separation, the friction forces, and the mass of each body. We show that the constraints of the scallop theorem can be escaped in frictional media if two asymmetry conditions are met at the same time: the frictional forces of each body against the surface must be distinct and the time-variation of the body-body separation must vary asymmetrically in time (so quick-slow or slow-quick in the extension-contraction phases). Our results are demonstrated numerically and interpreted using asymptotic expansions. | physics.bio-ph | physics | Crawling scallop: Friction-based locomotion with one degree of freedom
Department of Mechanical and Aerospace Engineering, University of California San Diego, 9500 Gilman Drive, La Jolla,
CA 92093-0411, USA.
Gregory L. Wagner∗, Eric Lauga∗∗
3
1
0
2
r
a
M
1
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
9
6
6
2
.
3
0
3
1
:
v
i
X
r
a
Abstract
Fluid-based locomotion at low Reynolds number is subject to the constraints of the scallop theorem, which
dictate that body kinematics identical under a time-reversal symmetry (in particular, those with a single
degree of freedom) cannot display locomotion on average. The implications of the theorem naturally compel
one to ask whether similar symmetry constraints exist for locomotion in different environments. In this work
we consider locomotion along a surface where forces are described by isotropic Coulomb friction. To address
whether motions with a single degree of freedom can lead to transport, we analyze a model system consisting
of two bodies whose separation distance undergoes periodic time variations. The behavior of the two-body
system is entirely determined by the kinematic specification of their separation, the friction forces, and the
mass of each body. We show that the constraints of the scallop theorem can be escaped in frictional media
if two asymmetry conditions are met at the same time: the frictional forces of each body against the surface
must be distinct and the time-variation of the body-body separation must vary asymmetrically in time (so
quick-slow or slow-quick in the extension-contraction phases). Our results are demonstrated numerically
and interpreted using asymptotic expansions.
1. Introduction
The capacity for locomotion is essential for the survival of much of life on Earth and is manifested in
strategies as diverse as the organisms which depend on it.
In fluids such as water or air and on scales
ranging from microns to tens of meters, creatures swim and fly by beating flagella, tails, wings, undulating
their bodies, or actuating pumps (Vogel, 1994). On land, animals crawl, walk, run, hop, climb, and slither
using friction between their bodies and the ground, or undulate and appear to "swim" through sand or soil
(Alexander, 2003).
The physics and scale of an environment determine the scope of successful propulsive strategies. The
strategy employed by a scallop, for example, which is to quickly open its shell, displacing a large amount
∗Corresponding author. Telephone: (781) 710-0871. Fax: (858) 822-3107.
∗∗Telephone: (858) 822-7925
Email addresses: [email protected] (Gregory L. Wagner), [email protected] (Eric Lauga)
URL: http://maeresearch.ucsd.edu/lauga/ (Eric Lauga)
Preprint submitted to Elsevier
September 18, 2018
of fluid, and then close it slowly, displacing a small amount of fluid, is ineffective if attempted in a fluid
environment where inertial forces are overwhelmed by viscous forces. This notion forms the basis for the
"scallop theorem", which holds that locomotion at low Reynolds number is not possible if the kinematics of
the body are identical under a time-reversal symmetry -- which is always true if the deformation is controlled
by a single degree of freedom (Purcell, 1977). The theorem relies on the linearity and time-independence of
the equations of motion for the fluid (the Stokes equation) and states that in order to achieve self-propulsion,
a body at low Reynolds number must deform in a manner indicating a clear direction of time, for example
in a waving motion (Lauga and Powers, 2009).
The beauty and important implications of this theorem, which greatly restrict the viable propulsive
strategies available to organisms and machines at small Reynolds numbers, naturally compel one to seek
similar fundamental results for other types of locomotion. For example, consider terrestrial locomotion
for which organisms make use of friction forces between their bodies and a surface. Are strategies in this
physical environment similarly constrained?
In general, strategies for terrestrial locomotion may be divided into two categories: those relying on the
movement of limbs (walking, hopping, climbing, and running) and those relying on movements of the body
(crawling and slithering). A number of studies have explored slithering locomotion, including its mechanics
(Home, 1812; Mosauer, 1932; Gray, 1946; Hu and et al., 2009), evolutionary advantages (Gans, 1975), ener-
getic efficiency (Walton and et al., 1991), optimization (Jing and Alben, 2012) as well as its application to
robotic propulsion (Hirose and Morishima, 1990; Transeth et al., 2009). On crawling-types of locomotion,
which is perhaps the simplest of all forms of terrestrial locomotion and is the subject of this work, investiga-
tions have examined the locomotion of maggots (Berrigan and Pepin, 1995) and earthworms (Quillin, 1999),
as well as hypothetical discrete-mass systems including systems consisting of two masses (Chernous'ko, 2002;
Zimmerman and et al., 2004) or a chain of three or more (Figurina, 2004; Zimmerman and et al., 2004, 2007,
2009; Bolotnik and et al., 2011) connected by springs or rigid mechanisms and actuated by external forces
or kinematic constraints.
A feature common to studies of the discrete-mass crawling strategies is that they consider systems
with either many degrees of freedom (Zimmerman and et al., 2004, 2007, 2009; Bolotnik and et al., 2011),
anisotropic friction coefficients (Zimmerman and et al., 2004), or analyze only locomotion in the case where
the masses alternately stick and slide (Chernous'ko, 2002). Here we consider the most basic situation
where deformation with one degree of freedom (and therefore time-reversible in its sequence of shapes)
actuates a system with isotropic Coulomb friction coefficients and which always interacts with the surface
through sliding friction forces, and propose to quantify the minimal requirements necessary to achieve
locomotion. Is locomotion even possible with isotropic friction using time-reversible deformations? What
are the associated minimum necessary mechanical or kinematic properties? How do they determine the
direction and magnitude of motion?
2
Figure 1: Two-body system (1 and 2) with prescribed relative distance, l(t), translating along the x direction.
To answer these questions, we consider a simple system consisting of two bodies which rest on a flat
surface and are joined by a mechanism that enforces the time-variation of their separation distance. The
variation of in the distance between the bodies gives rise to friction forces which determine their motion.
Friction forces are assumed to be isotropic (independent of the direction of body velocity) as it is evident
that friction anisotropy will trivially lead to locomotion. We demonstrate that locomotion is possible in a
system with only one degree of freedom provided that there is an asymmetry both in the friction properties
of the system as well as the deformation kinematics (meaning the time-periodic variations in length display
a quick-slow or slow-quick sequence of extension-contraction strokes). Finally, we develop a qualitative
physical explanation for the mechanics of strategy of locomotion through an asymptotic analysis of the
equations of motion (Bender and Orszag, 1999) and present a brief exploration of the space of physical
parameters available to the system.
2. Mathematical description of the two-body system
Consider a system consisting of two bodies with position xi (i = 1, 2), velocity xi, acceleration xi, and
mass mi which are supported by a flat surface and are constrained to translate along a horizontal line (the
x direction). The bodies are connected by a mechanism which prescribes their relative distance and is the
single degree of freedom available to the system to form the basis for its locomotion. The prescribed relative
position of the bodies is denoted l(t) = x2 − x1. A schematic representation of this system is shown in Fig. 1.
The variation in time of l(t) gives rise to a driving force, FG,i, which is exerted on each body by the
linkage and is opposed by a friction force, Ff,i, exerted on each body by the solid surface. The motion of
the bodies is described by Newton's second law,
xi =
1
mi
(FG,i + Ff,i) .
(1)
The friction force may be described by a variety of different models. Here, we restrict our analysis to
isotropic Coulomb friction forces which are independent of the sliding velocity. We use a standard model
which introduces a "static" friction force that requires that the force opposing it be greater in order for the
3
given body to translate, which we take to be the same as the sliding friction force. Mathematically we may
express this by writing the friction law
where Fi is then the magnitude of the sliding friction force exerted on body i by the surface. Here we choose
−Fi sgn( xi)
if FG,i > Fi,
−FG,i
if FG,i ≤ Fi,
(2)
Ff,i =
to represent the physical properties of the bodies and the surface as a friction force as opposed to a friction
coefficient in order to simplify the analysis; friction coefficients µi can be found by computing µi = Fi/mig.
Without loss of generality we assume that F1/m1 > F2/m2 (or µ1 > µ2) to ensure that only body 1 sticks
to the surface if the driving force is too weak to overcome the friction of both bodies.
For the prescription of the relative distance between the bodies, we assume that l(t) is a periodic function
which consists of an extension phase in which the bodies are pushed apart and a contraction phase in which
the bodies are pulled together. Again without loss of generality, we specify the kinematics such that the
extension phase is followed by the contraction phase and the bodies begin and end each period of the
kinematic specification at their minimum separation.
If the extension phase and contraction phase are
opposite and equal in magnitude, we say that the kinematic specification is "symmetric"; otherwise the
kinematic specification is "asymmetric".
In order to determine the driving force between the bodies arising from the variation of their relative
distance, we subtract the acceleration of each body to obtain
x1 − x2 = l(t) =
1
m2
(Ff,2 + FG,2) −
1
m1
(Ff,1 + FG,1) .
(3)
The relative position is enforced by a rigid mechanism which implies that FG,1 = −FG,2, and so we have
Inserting this relation into the equations of motion for each body, we find
FG2 =
m1m2
m1 + m2 (cid:18)l −
Ff,2
m2
+
Ff,1
m1 (cid:19) .
x1 = −
m2
m1 + m2
m1
l +
Ff2 + Ff1
m1 + m2
,
l +
Ff2 + Ff1
m1 + m2
,
x2 =
m1 + m2
(4)
(5)
which completes our mathematical description of the body motion. For this equation of motion to hold
for the blocks, the driving force must be large enough (and the kinematic acceleration large enough) such
that both blocks slide. When the kinematic acceleration is too small for both blocks to slide, only the
block with the smaller friction force will move. Also, since the kinematics are specified externally, the
driving force is necessarily always large enough to force motion in one of the blocks. We therefore find
that the Froude number, Fr, which is the ratio between driving forces and friction forces and is defined as
Fr = (m1 + m2)L/T 2(F1 + F2), must be relatively large. In contrast, the locomotion of snakes is usually
4
associated with Fr < 1 (Hu and et al., 2009), and while insufficient data is available to calculate Fr for
organisms which use closely related mechanisms (such as maggots, see Berrigan and Pepin (1995)), they are
almost certainly associated with low Froude numbers as well. It seems likely that the reason for this is tied
to the energy costs associated with maintaining high Froude numbers in frictional media.
The most striking feature of the equations of motion is the discontinuous dependence of the friction force
on the body velocity. This discontinuity is the critical nonlinearity which enables locomotion. However,
while the equations are globally nonlinear, it is apparent that they are linear on intervals for which the
direction of body motion does not change and that within these the position and velocity of the bodies can
be found analytically. The full solution can therefore be determined by pasting together the exact solutions
for adjacent intervals. Another technique to find steady solution computes the position and velocity of the
system for a single period and equates initial and final conditions.
3. Locomotion with one degree of freedom is possible
With the mathematical description of body motion established, we ask whether sliding locomotion with
a single degree of freedom is even possible. For our simple two-body system we observe two distinct modes
of asymptotic motion, as illustrated in Fig. 2: stationary oscillation around a fixed point (no locomotion, see
Fig. 2 A and B) and net translation (locomotion, see Fig. 2 C and D). In both examples from Fig. 2, bodies
1 and 2 have the same mass (m1 = m2) but are subject to different friction forces on the surface (F2 6= F1).
The two examples only differ in their kinematic specification: in A and B, the bodies are actuated with a
symmetric kinematic specification, resulting in stationary oscillation, whereas in C and D an asymmetric
kinematic specification leads to locomotion. In both cases the bodies exhibit transient behavior at the onset
of motion which progresses into an asymptotic, stable mode of motion after long times.
In general, three different regimes exist in the choice of relative friction force and kinematic specification:
(a) equal friction forces and arbitrary kinematics, (b) unequal friction forces and symmetric kinematics, and
(c) unequal friction forces and asymmetric kinematics. Regimes (a) and (b) lead to stationary oscillation,
whereas parameter choices falling into regime (c) lead to translation. Locomotion with a single degree of
freedom is therefore possible provided that there is an asymmetry in both the kinematics (actuation occurs
at a different rate one way than the other) and the friction forces exerted by each block on the surface.
4. Physics of locomotion with one degree of freedom: double symmetry-breaking
The locomotion of the system depends on the frictional symmetry between the bodies as well as the
symmetry between the extension and contraction phases of the kinematic specification. Physically, the
asymmetry in friction forces is required because if the friction forces are symmetric, a solution exists in which
the velocities of the bodies always oppose one another, and accordingly their friction forces cancel each other
5
Figure 2: Example of asymptotic stationary oscillation for equal masses (m1 = m2 = 1) but unequal friction forces (F1 = 4,
F2 = 1) and a symmetric kinematic specification. (A) Positions of both bodies (body 1, thick blue line; body 2, dashed red
line) and geometric center (dash-dotted black line). (B) Time-variation of the normalized separation distance between the two
bodies, (x2 − x1)/L, showing symmetric kinematics. (C) and (D): same as in (A) and (B) but with asymmetric kinematics
leading to locomotion.
such that their resulting equations of motion are merely proportional to the kinematic specification. The
asymmetry in the kinematic specification is required because if the extension and contraction phases are
opposite and equal in magnitude, a solution exists for which the net motion of the bodies over each phase
is equal and opposite, which cumulatively result in no net motion over the kinematic period. The breaking
of each of these symmetries divides the space of possible system properties into the three aforementioned
fundamental regimes; we now examine each regime in detail.
4.1. Symmetric friction forces
When the friction forces exerted by each bodies are identical, the system does not translate. If the bodies
are pushed apart in opposing directions after an initial time t = t0, the velocity and position of body 1 are
6
given by
x1 = −
x1 = −
m2
m1 + m2
m2
m1 + m2
l,
[l(t) − l(t0)] + x1(t0),
(6)
(7)
where we have used x1(0) = l(0) = 0. The position of body 2 can be found similarly, and if for simplicity
we use xC (0) = 0, we find the position of the center of mass to be
xC (t) =
1
2(cid:18) m1
m2
− 1(cid:19) [x1(t) − x1(t0)] .
(8)
Since x1(t) is periodic in time (via Eq. 6), we find that the position of the geometric center is periodic as
well and the system exhibits stationary oscillation. When m1 = m2, the center of the system is stationary
for all times.
4.2. Asymmetric friction forces, symmetric kinematics
Further, even if the friction forces exerted by the bodies are asymmetric, the system will not translate if
the kinematic specification is symmetric. This is best described mathematically by stating that l(t) = l(T −t)
for all t, where T is the period of oscillation of l(t). In this case, we are able to find a solution in which the
system does not exhibit locomotion. To simplify the argument we consider the extension and contraction
phases separately, which we denote A and B respectively. Each phase begins at a time t0 and ends at time
tf , and the position of the geometric center during each phase is xA and xB.
When the friction forces exerted by the bodies are unequal and the system is oriented as in Fig. 1, we
observe that the system velocity increases over the course of an extension phase. It is then possible to find
a small negative initial system velocity at the beginning of the extension phase such that − xA(t0) = xA(tf ).
Then, by examination of the equations of motion (1), it is evident that if the relative acceleration imposed
by the contraction phase is opposite that imposed by the extension phase, then xB(t0) = xA(tf ) implies
xB(t0) = − xB(tf ) and xA(t0) = xB(tf ), and the considered initial system velocity constitutes an equilibrium
initial system velocity in a periodic solution for the system motion. Further, the same logic applies to the
system position: if the kinematic specification is opposite in sign and equal in magnitude, and if the body
velocities are opposite in sign, then the motion of the geometric center over the extension and contraction
phases, respectively, will be opposite in sign and equal, and the system will experience no net motion
cumulatively over both phases. Locomotion is therefore not possible when the extension and contraction
phases of the kinematic specification are symmetric, even under asymmetric friction. The statements of §4.1
and §4.2 constitute thus a form of scallop theorem for locomotion under isotropic friction.
4.3. Asymmetric friction forces, asymmetric kinematics
We have demonstrated that asymmetric friction forces and asymmetric kinematics can result in loco-
motion, but the question remains: why? Some insight can be gained by considering the system in the
7
Figure 3: Illustration of body position (A) and velocity (B) in an extension phase when both bodies start from rest; (A):
normalized body positions, xi/L; (B): normalized body velocities,
xi × (T /L). There are two intervals distinguished by the
signs of the body velocities; in the first, the bodies have opposite velocity and are pushed apart; in the second, both bodies
translate forward.
asymptotic limit of zero kinematic period, or fast kinematics and strong driving forces. We make two key
observations. First, over the duration of an extension phase, the body with a smaller friction force (body
2) will move farther forward (x > 0) than the body with a smaller friction force (body 1) when the bodies
are pushed part. This implies trivially that the geometric center of the system will move forward during an
extension phase and backwards during a contraction phase. Second, if the extension phase is longer (and the
kinematic velocity faster) than the contraction phase, the net change in system position over both intervals
will be positive. This can be understood by examining the first terms of an asymptotic expansion in the
extension phase duration for the change in system velocity and position over the phase. Due to the piecewise
nature of the equations, we must consider separately the case where the bodies begin the extension phase
at rest from the case where they begin with some initial velocity.
4.3.1. A single extension phase when bodies are initially at rest
Consider the two-block system when it is initially at rest at time t0 = 0. In this case, the extension
phase is divided into two intervals distinguished by the sign of the velocity of body 1, as illustrated in Fig. 3
where we show both the body positions (A) and velocities (B). In the first interval, which constitutes the
greater part of the extension phase, the bodies are pushed apart from their stationary position so that body
2 moves forward and body 1 backwards, and l(t) > 0. This interval ends when the velocity of body 1, which
is lesser in magnitude than body 2 due to its larger friction force, reaches zero and changes direction. In the
second interval, both bodies move forward until the end of the interval is reached.
The motion of each body is described by Eq. (5).
In the first interval we have that Ff1 = F1 and
8
Ff2 = −F2, which yields
x1 = −
m2
m1 + m2
l(t) +
F1 − F2
m1 + m2
·
Integrating from t = 0 to t, and noting that the body is initially at rest, we find the body velocity,
x1(t) = −
m2
m1 + m2
l(t) +
F1 − F2
m1 + m2
t,
and integrating again and setting x1(0) = 0, we obtain the body position,
x1(t) = −
m2
m1 + m2
l(t) +
F1 − F2
m1 + m2
t2
2
·
(9)
(10)
(11)
The first interval ends when the velocity of body 1 reaches zero. We expect this to occur very close to the
end of the interval at a time t = T − ∆T , where ∆T ≪ T (see Fig. 3). We therefore expand l(t) near t = T
to find
which implies
x1(T − ∆T ) = 0 = ∆T hm2l(T ) − F1 + F2i + (F1 − F2) T + O(∆T 2),
∆T = −
F1 − F2
m2l(T ) − F1 + F2
T.
(12)
(13)
To emphasize the dependence of quantities on the phase duration T , we note that if the length of extension
remains constant we expect the kinematic acceleration to scale as l ∼ 1/T 2. We therefore introduce the
notation l(T ) = lT /T 2, where lT is some constant which is proportional to the maximum separation between
the bodies l(T ), and write
∆T = −
F1 − F2
m2lT
T 3 + O(T 5).
The position of body 1 at the end of the interval is
x1(T − ∆T ) = −
m2
m1 + m2
l(T ) +
F1 − F2
m1 + m2
T 2
2
+ O(T 3).
(14)
(15)
We can now find the final body velocities and positions. In the second interval we have that both body
velocities are positive implying that Ff1 = −F1 and Ff2 = −F2. This yields the equation of motion
x1 = −
m2
m1 + m2
l(t) −
F1 + F2
m1 + m2
·
(16)
The velocity of body 1 is found by integration, where by specification the initial velocity of the body in this
interval is zero,
x1(t) = −
and at time t = T we find
m2
m1 + m2 hl(t) − l(T − ∆T )i −
F1 + F2
m1 + m2
[t − (T − ∆T )] ,
x1(T ) =
F1 − F2
m1 + m2
T + O(T 3).
9
(17)
(18)
It is not necessary to solve for the velocity of body 2, because by definition of the kinematics it will equal
the velocity of body 1 at time t = T . The position of body 1 is found by integrating the body velocity,
x1(T ) = −
m2
m1 + m2
l(T ) +
F1 − F2
m1 + m2
T 2
2
+ O(T 4).
Since x2(T ) = x1(T ) + l(T ), we find that
x2(T ) =
m1
m1 + m2
l(T ) +
1
2
F1 − F2
m1 + m2
T 2 + O(T 4),
and the geometric center xC = (x1 + x2)/2 is
xC (T ) =
1
2
m1 − m2
m1 + m2
l(T ) +
1
2
F1 − F2
m1 + m2
T 2 + O(T 4).
(19)
(20)
(21)
The first term in this expression is the change in position due to the difference in body inertia and is
independent of the phase duration. The second term, however, is proportional to the difference in friction
forces between the blocks and therefore contains the key to the direction of body translation. Its origin is in
the short interval of time that both bodies are moving forward: because an increase in phase duration causes
the the length of this interval to increase, the change in final system position is correspondingly increased
as well.
4.3.2. A single extension phase when bodies have an initial positive velocity
The situation detailed in the previous section is modified somewhat by the existence of an initial system
velocity at the beginning of the kinematic period. A similar asymptotic analysis (the details of which are
given in Appendix A) with an initial positive velocity x0 yields the change in system velocity and position,
respectively,
and
xC (T ) − x0 =
F1 − F2
m1 + m2
T − x0
2F1
m2 (cid:18) 1
l0
−
1
lT(cid:19) T 2 + O(T 3),
xC (T ) =
1
2
m1 − m2
m1 + m2
l(T ) + x0T +
1
2
F1 − F2
m1 + m2
T 2 − x0
2F1
m2l0
T 3 + O(T 4),
(22)
(23)
where lT and l0 are lengths which characterize the acceleration term in the kinematic specification. The
effect of the small initial velocity is to increase the total change in system position and to decrease the
change in system velocity. The first is intuitive and easy to understand. The second is also intuitive and
explains why the system exhibits stable dynamics, since period-wise increases in the initial velocity decay
until the initial velocity reaches some steady-state.
4.3.3. A full kinematic period when bodies translate in steady-state
We can then address an extension phase followed by a contraction phase when the bodies are translating
in steady-state. This requires us to first obtain the steady-state velocity for particular system parameters,
and then insert this into our equation for the final body position after the full kinematic period. To derive
10
the steady-state velocity, we consider the change in velocity after an extension phase and a contraction phase
and set the final and initial velocities equal to one another.
For the contraction phase, the final velocity and position of the body when starting with some small
initial positive velocity are simply the opposite of that for the extension phase,
xC (T ) − x0 = −
F1 − F2
m1 + m2
T + x0
2F2
m1 (cid:18) 1
l0,c
−
1
lT,c(cid:19) T 2 + O(T 3),
and
xC (T ) = −
1
2
m1 − m2
m1 + m2
l(T ) + x0T −
1
2
F1 − F2
m1 + m2
T 2 + x0
2F2
m1l0,c
T 3 + O(T 4),
(24)
(25)
where we have used the subscript c to differentiate the kinematic constants in the contraction phase (c) from
the extension phase (e). The details of this calculation are found in Appendix B. For a contraction phase
which is functionally equivalent and opposite to the extension phase, we have l0,e = −l0,c, and lT,e = −lT,c,
even if the phases have different durations. We then find the steady-state velocity by combining an extension
phase with a contraction phase and and setting the final velocity equal to the initial velocity. By writing
the extension phase duration as αT and the contraction phase duration as (1 − α)T with kinematic period
T , we obtain
x0 =
(F1 − F2) (2α − 1)
2T (1/l0 − 1/lT ) [α2F1m1 + (1 − α)2F2m2](cid:18) m1m2
m1 + m2(cid:19) + O(T ).
(26)
We can simplify this expression by defining a reduced system mass m∗ = m1m2/(m1 + m2) and lumping the
kinematic-dependent parameters into a single parameter with dimensions of length Lkin = l0lT /[2(lT − l0)];
if L is the difference between the greatest and least separation of the bodies, then Lkin = L for a piecewise
quadratic form for the kinematic specification and π2L/8 for a sinusoidal form. The relation becomes
x0 = (F1 − F2) (2α − 1)
Lkinm∗
α2F1m1 + (1 − α)2F2m2
1
T
+ O(T ).
(27)
Accordingly we find that the steady-state initial velocity is inversely proportional to the total phase duration
(or proportional to the prescribed kinematic velocity). Using this initial velocity to compute the final position
of the bodies, we find that
xC = (F1 − F2) (2α − 1)
Lkinm∗
α2F1m1 + (1 − α)2F2m2
+ O(T 4).
(28)
This expression encapsulates the main result of this paper and explicitly establishes the two conditions
necessary for locomotion. The first condition is that α 6= 1/2; or there must be an asymmetry between the
extension and contraction phases of the kinematics. The second condition is that F1 6= F2; or there must be
an asymmetry between the friction force exerted by each body on the supporting surface. It is important
to note that condition requires an asymmetry in the friction force specifically, regardless of how this force
depends on material parameters (such as mass or friction coefficient). Furthermore, we see that the change
in position of the bodies therefore goes to a constant as the kinematic period vanishes, and the effective
velocity of the system scales as xC /T ∼ 1/T , or with the velocity of the kinematics.
11
Figure 4: Change in system position (A) and velocity (B) over a single extension phase versus phase duration; (A): change in
normalized system position, ∆xC /L; (B) change in normalized system velocity, ∆ xC × T /L. The full calculation is plotted
with solid lines while the asymptotic expansion presented in the text for small phase durations is shown with dashed lines.
The parameters are l(T ) = 10, F2 = 1, m1 = m2 = 1 and the plots are shown for three different friction asymmetries
F1/F2 = {2, 10, 40}.
4.4. A physical description for all parameter regimes
With an explanation in hand for when the driving force is large compared to the friction force, we now
examine the relationship between phase interval and the change in system velocity and position for the full
range of extension phase durations. In Fig. 4 we plot the change in system position (A) and velocity (B)
against the length of the extension phase for all physically valid extension phase durations and a kinematic
specification which is piecewise quadratic and continuous in position and velocity.
The dependence on the change in system position on phase duration (Fig. 4A) provides an indication
of the direction of system translation. We see that this dependence is monotonic, which implies that the
system will always translate forward when the extension phase is longer than the contraction phase, and
vice versa.
The dependence of system velocity on phase duration (Fig. 4B) provides an indication of the nature of
the system dynamics. Here we find that the change in system velocity is not monotonic with phase duration.
When the phase duration is short (large driving force, high Froude number) the change in system velocity
increases as the phase duration increases. When the phase duration is long (smaller driving force, relatively
smaller Froude number) this dependence is reversed.
The combination of these two effects implies that there are two distinct regimes of translation for sliding
locomotion of this system. When the change in system velocity is increasing with increasing duration,
period-wise initial velocity of the system will be positive when the bodies are translating in the forward
direction. However when the system velocity is decreases with increasing phase duration, we find that the
12
Figure 5: (A): Dependence of the period-averaged velocity, hU i, on the kinematic asymmetry, α, for F2 = 1 and three values
of F1; (B) Dependence of hU i on the friction asymmetry F1/F2 for α = 3/4 and three values of F2. In both cases we have
m1 = m2 = 1, l(T ) = 10.
period-wise initial velocity of the system is antagonistic to the overall direction of translation. Finally, the
maximum period-wise displacement of the system and therefore maximum average system velocity for a
given kinematic specification occurs at the boundary between these two regimes. This can be explained by
observing that a zero initial velocity at the beginning of the extension phase implies that the initial velocity
for the contraction phase is at a maximum, which minimizes the decrease in system position over this phase.
4.5. Parameter studies
Once it is known that both an asymmetry in the specified kinematics and an asymmetry in the critical
friction forces are necessary for the system to achieve net locomotion, we may ask how the motion of the
system is dependent on the extent of the asymmetries. One metric which characterizes the capacity of the
system for locomotion is its velocity, and so to probe the effect of physical asymmetry on system behavior
we define a "period-averaged velocity" as
hU i =
xC (t0 + T ) − xC (t0)
T
·
(29)
For the kinematic specification, we choose a simple piecewise quadratic form and define a parameter α
to characterize the asymmetry of the kinematics. The length of the extension phase is then αT and the
contraction phase is (1 − α)T , and α = 1/2 implies a symmetric kinematic specification.
In Fig. 5, we plot the period-averaged velocity against both the kinematic asymmetry (left) and the
asymmetry in friction forces (right). For the kinematic asymmetry, we observe that the period-averaged
velocity of the system increases as the specified kinematics become more asymmetric.
13
For the frictional asymmetry, we observe that the period-averaged velocity increases to a maximum and
then decays to zero as the asymmetry increases, which is expected since no translation occurs either for
F1/F2 = 1 or as F1/F2 → ∞ (when the friction force is too large for the kinematic specification to force
the system into a sliding mode of locomotion). Further, as explained in the previous section, we observe
the two distinct regimes of translation in which the period-wise initial system velocity is either in the same
direction as translation (small F1/F2) or in the opposite direction (large F1/F2).
5. Discussion
In this work, we attempt to understand the behavior of the conceptually simplest crawler possible: a
one-dimensional, one-degree of freedom system consisting of two mechanically connected point masses. We
find that such a system is able to achieve time-averaged translation even when its frictional interactions
with the surface are isotropic and are mediated only through sliding contact, in addition to the more evident
cases of intermittent static contact and anisotropic friction. We use computations and physical reasoning to
show which and in what manner symmetries must to be broken to obtain net locomotion. In doing so, we
demonstrate that:
• Friction-based locomotion with one degree of freedom is possible because of the non-linear dependence
of the friction force on body velocity;
• Two symmetries must be broken in order for the system to achieve locomotion: the two components
of the body must exert a different friction force on the supporting surface, and the body kinematics
must be asymmetric on a single period of actuation,
• The physical mechanism of locomotion results from an interval within a single kinematic period in
which the body with a smaller friction force (faster) carries the other body (slower) forward
• For each chosen set of physical parameters and kinematic function the system always eventually
achieves a steady-state of locomotion;
• Two fundamental regimes of translation exist which correspond to the relative strength of the driving
force to the friction force (or Froude number): when the driving force is large (higher Froude numbers)
the system velocity at the beginning of each period is in the direction of translation; when the driving
force is smaller (relatively lower Froude numbers) this initial system velocity opposes the direction of
translation; and the maximum velocity occurs on the boundaries of these two regimes.
Further, we observe that locomotion in this system occurs necessarily at relatively high Froude number.
While experimental data on the forces exerted by crawling organisms is difficult to obtain, it seems likely
that most crawling organisms locomote low Froude numbers. This may have something to do with the
14
energetic cost of maintaining high velocities (and therefore high Froude numbers) in frictional media; in our
system, much mechanical energy is wasted in applying sufficiently large forces to the blocks to ensure that
they always slide.
The established criterion for the asymmetries of a two mass, friction-based system is fundamental knowl-
edge in the physics of terrestrial locomotion. Additionally we hope that this study will help guide the
analysis and design of simple crawlers for use in exploration and medicine, as well aid in the development
of intuition and understanding of more advanced and complex modes of terrestrial locomotion.
6. Acknowledgments
This work was supported by the National Science Foundation by Grant No. CBET-0746285 to E.L. and
a Focht-Powell Fellowship to G.L.W.
Appendix A. Analysis of a single extension phase when bodies start with some small initial
velocity
We examine here the effect some small initial positive velocity has on the final position and velocity of
the system after a single extension phase. When both bodies begin with a positive velocity the equation of
motion for body 1 is
Integration yields the body velocity,
x1 = −
m2
m1 + m2
l(t) −
F1 + F2
m1 + m2
·
and the body position,
x1(t) = −
m2
m1 + m2
l(t) −
F1 + F2
m1 + m2
t + x0,
x1(t) = −
m2
m1 + m2
l(t) −
F1 + F2
m1 + m2
t2
2
+ x0t,
(Appendix A.1)
(Appendix A.2)
(Appendix A.3)
where we set the initial position of body 1 to x1(0) = 0. The first interval ends at time t = t1 when the
velocity of body 1 decreases to 0. In general this can only be solved if l(t) is specified, but to start let us
assume that t1 is small. We can then expand l(t1) to obtain
x1(t1) = −
m2
m1 + m2
t1l(0) −
F1 + F2
m1 + m2
t1 + x0 + O(t2
1),
We then find that x1(t1) = 0 when
t1 = x0
m1 + m2
m2l(0) + F1 + F2
·
(Appendix A.4)
(Appendix A.5)
15
When T → 0 and using the notation l(0) = l0/T 2 we have that
t1 =
=
x0(m1 + m2)
m2l0
x0(m1 + m2)
m2l0
= x0
(m1 + m2)
m2l0
1
1 + T 2 F1+F2
T 2
m2l0 ! ,
T 2"1 − T 2 F1 + F2
T 2(cid:18)1 − T 2 F1 + F2
m2l0
m2l0 (cid:19)2
+ T 4(cid:18) F1 + F2
+ ...# ,
(Appendix A.6)
m2l0 (cid:19) + O(T 6),
which confirms our assumption that t1 is small when T → 0. The position of body 1 at the end of the first
interval is
x1(t1) = −
m2
m1 + m2
= t1 x0 −
t1
2
l(0) −
t2
F1 + F2
1
2
m1 + m2
m2l(0) + F1 + F2
m1 + m2
t2
1
2
! ,
+ x0t1,
=
x0t1
2
= O(T 2).
(Appendix A.7)
In the second interval, body 1 is moving backwards and body 2 is moving forwards. The equation of motion
is
x1 = −
m2
m1 + m2
l(t) +
F1 − F2
m1 + m2
·
(Appendix A.8)
Integration yields the body velocity
x1(t) = −
m2
m1 + m2 h l(t) − l(t1)i +
F1 − F2
m1 + m2
and integrating again yields the body position
(t − t1),
(Appendix A.9)
x1(t) = −
m2
m1 + m2
[l(t) − l(t1)] +
m2
m1 + m2
l(t1)(t − t1)
+
1
2
F1 − F2
m1 + m2
(t − t1)2 + x1(t1).
(Appendix A.10)
This interval ends at time t = t2 when the velocity of body 1 goes to zero again and both bodies are
translating forwards. We use the ansatz that t2 is very close to the end of the interval T and, as in the case
where the bodies start from rest, define ∆T = T − t2. We then have that
l(T − ∆T ) = l(T ) − ∆T l(T ) + O(∆T 2),
= −∆T l(T ) + O(∆T 2),
(Appendix A.11)
and
l(T − ∆T ) = l(T ) − ∆T l(T ) +
∆T 2
2
l(T ) + O(∆T 3).
=
∆T 2
2
16
l(T ) + O(∆T 3),
(Appendix A.12)
The equation for body velocity becomes
0 = −
m2
m1 + m2 h−∆T l(T ) − t1l(0)i +
F1 − F2
m1 + m2
[(T − ∆T ) − t1] ,
= ∆T hm2l(T ) − (F1 − F2)i + t1hm2l(0) − (F1 − F2)i
+ (F1 − F2)T,
(Appendix A.13)
= ∆T hm2l(T ) − (F1 − F2)i + x0(m1 + m2)
+ (F1 − F2)T,
m2l(0) − F1 + F2
m2l(0) + F1 + F2
This implies that
∆T = −
1
m2l(T ) − F1 + F2" x0(m1 + m2)
· · · + (F1 − F2)T#.
m2l(0) − F1 + F2
m2l(0) + F1 + F2
· · ·
(Appendix A.14)
Using the notation l(0) = l0/T 2 and l(T ) = lT /T 2, we find
∆T = −
m2lT !" x0(m1 + m2)
1 − T 2 F1−F2
m2l0
1 + T 2 F1+F2
m2l0
1 − T 2 F1−F2
1
T 2
m2lT
+ (F1 − F2)T#,
(Appendix A.15)
(Appendix A.16)
= −T 3 F1 − F2
m2lT (cid:18)1 + T 2 F1 − F2
m2lT
− T 2 x0(m1 + m2)
· · ·(cid:18)1 − T 2 F1 + F2
m2l0 (cid:19) + O(T 6).
m2lT (cid:19) + O(T 7)...
(cid:18)1 + T 2 F1 − F2
m2lT (cid:19)(cid:18)1 − T 2 F1 − F2
m2l0 (cid:19) · · ·
Rearranging, we find
∆T = −(cid:20) x0(m1 + m2)
−(cid:20) x0(m1 + m2)
(cid:21) T 2 −(cid:18) F1 − F2
(cid:18) F1 − F2
m2lT
m2lT
m2lT
m2lT (cid:19) T 3
−
2F1
m2l0(cid:19)(cid:21) T 4 + O(T 5).
17
The position of body 1 at the end of the second interval is
x1(T − ∆T ) = −
= −
= −
+
+
+
+
m2
m1 + m2
m2
m1 + m2
1
2
m2
F1 − F2
m1 + m2
l(T ) −
m1 + m2
m2
m1 + m2
1
2
m2
F1 − F2
m1 + m2
[l(T − ∆T ) − l(t1)]
l(t1)(T − ∆T − t1)
(T − ∆T − t1)2 + x1(t1),
1
2
m2
m1 + m2 h∆T 2l(T ) − t2
1
l(0)i
t1l(0)(T − ∆T − t1)
(T − ∆T − t1)2 +
x0t1
2
+ O(T 4),
2
+
1
2
1
2
+ t2
F1 − F2
F1 − F2
m1 + m2
m2l(T )
m1 + m2
l(T ) + T 2(cid:18) 1
m1 + m2(cid:19)
m1 + m2#
m1 + m2#
m1 + m2(cid:19) + t1T" m2l(0)
m1 + m2# +
+ ∆T 2"−
1"−
− T ∆T(cid:18) F1 − F2
+ t1∆T "−
m2l(0)
m1 + m2
m2l(0)
m1 + m2
F1 − F2
F1 − F2
1
2
1
2
+
+
m1 + m2
−
F1 − F2
m1 + m2#
x0t1
2
+ O(T 4).
After much algebra, we obtain
x1(T − ∆T ) = −
m2
m1 + m2
l(T ) + x0T
+
+
1
m1 + m2
2(cid:20) F1 − F2
m2 (cid:18) F1 − F2
x0
lT
+ ( x0)2 m1 + m2
m2lT
(cid:21) T 2
l0 (cid:19) T 3 + O(T 4).
2F1
−
(Appendix A.17)
(Appendix A.18)
In the final interval both bodies have a positive velocity. The equation of motion for body 1 is
x1 = −
m2
m1 + m2
l(t) −
F1 + F2
m1 + m2
,
(Appendix A.19)
so, noting that body 1 is at rest at the beginning of the final interval, the velocity of body 1 is
and its position,
x1 = −
m2
m1 + m2 h l(t) − l(t2)i −
F1 + F2
m1 + m2
(t − t2),
x1 = −
m2
m1 + m2
[l(t) − l(t2)] +
m2
m1 + m2
l(t2) (t − t2)
−
1
2
F1 + F2
m1 + m2
(t − t2)2 + x1(t2).
18
(Appendix A.20)
(Appendix A.21)
At time t = T we find for the body velocity
x1(T ) = −
m2
m1 + m2 h l(T ) − l(T − ∆T )i −
F1 + F2
m1 + m2
∆T,
m1 + m2
= −∆T (cid:20) m2
= −∆T (cid:18) m2
m1 + m2
= x0 +
T − x0
F1 − F2
m1 + m2
l(T ) +
lT T −2 +
F1 + F2
F1 + F2
m1 + m2(cid:21) + O(T 4),
m1 + m2(cid:19) + O(T 4),
lT(cid:19) T 2 + O(T 3),
m2 (cid:18) 1
−
l0
1
2F1
and for the position of body 1,
x1 = −
m2
m1 + m2
[l(T ) − l(T − ∆T )] +
m2
m1 + m2
l(T − ∆T )∆T
(Appendix A.22)
∆T 2 + x1(T − ∆T ),
∆T 2l(T ) −
1
2
F1 + F2
m1 + m2
∆T 2 + x1(T − ∆T ) + O(T 6),
(Appendix A.23)
−
1
2
F1 + F2
m1 + m2
m2
1
2
1
2
= −
= −
= −
m1 + m2
m2
m1 + m2
m2
m1 + m2
lT ∆T 2
T 2 + x1(T − ∆T ) + O(T 4),
m1 + m2(cid:19) T 2 + x0
2(cid:18) F1 − F2
l(T ) + x0T +
1
2F1
m2l0
T 3 + O(T 4).
At the end of the extension phase, the velocity of body 2 is equal to that of body 1, and the position of
body 2 is x2(T ) = x1(T ) + l(T ). The velocity of the geometric center xC = ( x1 + x2)/2 is therefore
xC (T ) = x0 +
F1 − F2
m1 + m2
T − x0
2F1
m2 (cid:18) 1
l0
−
1
lT(cid:19) T 2 + O(T 3),
(Appendix A.24)
and the position of the geometric center xC = (x1 + x2)/2 is
xC (T ) =
1
2
m1 − m2
m1 + m2
l(T ) + x0T +
1
2(cid:18) F1 − F2
m1 + m2(cid:19) T 2 + x0
2F1
m2l0
T 3 + O(T 4).
(Appendix A.25)
Appendix B. Analysis of a single contraction phase when bodies start with some small initial
velocity
We examine here the change in velocity of the two body system after a contraction phase when the
bodies begin and end with some small initial positive velocity. This velocity is presumed small enough such
that the contraction kinematics induce a reversal in the velocity of body 2, but large enough such that the
body still has a positive velocity at the end of the phase. This analysis is very similar to the extension phase
analysis, except that it is the velocity of body 2 which is now of interest. The equation of motion for body
2 is
x2 =
m1
m1 + m2
l(t) −
F1 + F2
m1 + m2
·
(Appendix B.1)
19
This equation of motion holds until the velocity of body 2 changes sign. Using the same steps as for the
extension phase and letting T → 0, we find the time t1 at which x2(t1) = 0,
t1 = x0
m1 + m2
−m1l(0) + F1 + F2
,
= − x0
m1 + m2
m1l0
T 2(cid:18)1 + T 2 F1 + F2
m1l0 (cid:19) + O(T 6).
The position of body 2 at t = t1 is
x2(t1) = x2(0) +
m1
m1 + m2
(l(t1) − l(0)) −
1
2
F1 − F2
m1 + m2
t2
1 + x0t1,
= x2(0) + t1" x0 −
t1
2
F1 + F2 − l(0)m1
m1 + m2
# ,
= x2(0) +
x0t1
2
+ O(T 2).
(Appendix B.2)
(Appendix B.3)
In the second interval, body 1 is moving forwards and body 2 is moving backwards, and the equation of
motion for body 2 is
x2 =
m1
m1 + m2
l(t) −
F1 − F2
m1 + m2
·
(Appendix B.4)
As in the extension phase with body 1, this interval ends when the velocity of body 2 changes sign. It is
convenient to again define ∆T = T − t2, and we find
∆T =
1
m1l(T ) − F1 + F2" x0 (m1 + m2)
− (F1 − F2)T#.
m1l(0) − F1 + F2
m1l(0) − F1 − F2
(Appendix B.5)
With the notation l(0) = l0/T 2 and l(T ) = lT /T 2, and expanding in powers of T , this becomes
F1 − F2
∆T = −
T 3(cid:18)1 + T 2 F1 − F2
m1lT (cid:19) + O(T 7)
(cid:18)1 + T 2 F1 − F2
m1lT
+ T 2 x0(m1 + m2)
· · ·(cid:18)1 + T 2 F1 + F2
m1lT
m1l0 (cid:19) + O(T 6),
(cid:21) T 2 −(cid:18) F1 − F2
(cid:18) F1 − F2
m1lT
m1lT (cid:19) T 3
+
=(cid:20) x0(m1 + m2)
m1lT
+(cid:20) x0(m1 + m2)
m1lT
2F2
m1l0(cid:19)(cid:21) T 4 + O(T 5).
m1lT (cid:19)(cid:18)1 − T 2 F1 − F2
m1l0 (cid:19) · · ·
(Appendix B.6)
20
Expanding the kinematic specification in the same way as in the extension phase calculation, and noting
that in this case l(T ) = 0, we find the position of body 2 at the end of the second interval,
x2(T − ∆T ) =
m1
m1 + m2
[l(T − ∆T ) − l(t1)]
l(t1) (T − ∆T − t1)
(T − ∆T − t1)2 + x2(t1),
m1 + m2
m1
l(0) +
1
2
m1
m1 + m2 h∆T 2l(T ) − t2
1
l(0)i
t1l(0) (T − ∆T − t1)
m1
m1 + m2
1
F1 − F2
2
m1 + m2
m1
m1 + m2
1
F1 − F2
2
m1 + m2
m1
−
−
−
−
= −
= −
(T − ∆T − t1)2 + x2(0) +
x0t1
2
+ O(T 4),
F1 − F2
m1 + m2(cid:19)
m1 + m2# + T ∆T(cid:18) F1 − F2
m1 + m2(cid:19)
F1 − F2
m1 + m2
l(0) − T 2(cid:18) 1
2
−
m1l(T )
m1 + m2
2
m1 + m2
− t1T " m1l(0)
+ ∆T 2" 1
1 m1l(0)
+ t1∆T " m1l(0)
m1 + m2
+ t2
m1 + m2
−
1
2
−
1
2
F1 − F2
m1 + m2#
F1 − F2
m1 + m2!
m1 + m2#
F1 − F2
−
+ x2(0) +
x0t1
2
+ O(T 4),
which simplifies to
x2(T − ∆T ) = x2(0) −
m1
m1 + m2
l(0) + x0T
−
+
1
m1 + m2
2(cid:20) F1 − F2
m1 (cid:18) F1 − F2
x0
lT
+ ( x0)2 m1 + m2
m1lT
(cid:21) T 2
l0 (cid:19) T 3 + O(T 4)·
2F2
+
(Appendix B.7)
(Appendix B.8)
In the third interval, both bodies are translating forward. The equation of motion for body 2 is then
x2 =
m1
m1 + m2
l(t) −
F1 + F2
m1 + m2
·
(Appendix B.9)
21
By integrating from T − ∆T to T and employing the fact that x2(T − ∆T ) = 0, we find the velocity of body
2 at the end of the contraction phase,
x2(T ) =
=
m1 + m2 hl(T ) − l(T − ∆T )i −
m1
m1
∆T l(T ) −
∆T,
m1 + m2
F1 + F2
m1 + m2
2F2
= x0 −(cid:18) F1 − F2
m1 + m2(cid:19) T + x0
m1 (cid:18) 1
l0
F1 + F2
m1 + m2
∆T,
(Appendix B.10)
−
1
lT(cid:19) + O(T 3).
The position of body 2 is
x2(T ) =
m1
m1 + m2
[l(T ) − l(T − ∆T )] +
m1
m1 + m2
l(T − ∆T )
∆T 2l(T ) + x2(T − ∆T ) + O(T 4),
(Appendix B.11)
∆T 2 + x2(T − ∆T ),
−
F1 + F2
m1 + m2
1
2
m1
m1 + m2
=
1
2
= x2(0) −
m1
m1 + m2
2F2
m1l0
+ x0
T 3 + O(T 4).
l(0) + x0T −
1
2
F1 − F2
m1 + m2
T 2
Finally, we find the velocity of the geometric center by observing that at the end of the contraction phase we
have x1(T ) = x2(T ) = xC (T ), and we find the position of the geometric center by observing that because
l(T ) = 0 then we must have xC (T ) = x2(T ). Further, if we presume that, as in the extension interval
calculation, that the initial position of the geometric center is xC (0) = 0, we find that x2(0) = l(0)/2. The
velocity of the geometric center at the end of the contraction phase is therefore
xC (T ) = x0 −(cid:18) F1 − F2
m1 + m2(cid:19) T + x0
and the position of the geometric center is
2F2
m1 (cid:18) 1
l0
−
1
lT(cid:19) + O(T 3),
(Appendix B.12)
xC (T ) = −
1
2
m1 − m2
m1 + m2
l(0) + x0T −
1
2
F1 − F2
m1 + m2
T 2 + x0
2F2
m1l0
T 3 + O(T 4).
(Appendix B.13)
22
References
Alexander, R., 2003. Principles of animal locomotion. Princeton University Press.
Bender, C., Orszag, S., 1999. Advanced Mathematical Methods for Scientists and Engineers I: Asymptotic Methods and
Perturbation Theory. Springer-Verlag.
Berrigan, D., Pepin, D.J., 1995. How maggots move: Allometry and kinematics of crawling in larval diptera. J. Insect Physiol.
41, 329 -- 337.
Bolotnik, N., et al., 2011. The undulatory motion of a chain of particles in a resistive medium. Z. Angew. Math. Mech. 91.
Chernous'ko, F.L., 2002. The optimum rectilinear motion of a two-mass system. J. Appl. Math. Mech-USS 66.
Figurina, T.Y., 2004. Controlled quasistatic motions of a two-link robot on a horizontal plane. J. Comput. Sys. Sci. 43.
Gans, C., 1975. Tetrapod limblessness: Evolution and functional corollaries. Am. Zool. 15, 455 -- 467.
Gray, J., 1946. The mechanism of locomotion in snakes. J. Exp. Biol. 23.
Hirose, S., Morishima, A., 1990. Design and control of a mobile robot with an articulated body. Int. J. Robot. Res. 9.
Home, E., 1812. Observations intended to show that the progressive motion of snakes is partly performed by means of the ribs.
Philos. Trans 163.
Hu, D.L., et al., 2009. The mechanics of slithering locomotion. P. Natl. Acad. Sci. USA 106.
Jing, F., Alben, S., 2012. Optimization of two- and three-link snake-like locomotion. ArXiv 1212.0062v1.
Lauga, E., Powers, T.R., 2009. The hydrodynamics of swimming microorganisms. Rep. Prog. Phys. 72, 096601.
Mosauer, W., 1932. On the locomotion of snakes. Science 76.
Purcell, E.M., 1977. Life at low Reynolds number. Am. J. Phys. 45.
Quillin, K.J., 1999. Kinematic scaling of locomotion by hydrostatic animals: ontogeny of peristaltic crawling by the earthworm
lumbricus terrestris. J. Exp. Biol. 202, 661 -- 674.
Transeth, A.A., Pettersen, K.Y., Liljeback, P., 2009. A survey on snake robot modeling and locomotion. Robotica 27, 999 -- 1015.
Vogel, S., 1994. Life in moving fluids. Princeton University Press.
Walton, M., et al., 1991. The energetic cost of limbless locomotion. Science 249.
Zimmerman, K., et al., 2004. An approach to worm-like motion. XXI ICTAM, 15-21 August, Warsaw, Poland .
Zimmerman, K., et al., 2007. Forced nonlinear oscillator with nonsymmetric dry friction. Arch. Appl. Mech. 77.
Zimmerman, K., et al., 2009. Motion of a chain of three point masses on a rough plane under kinematical constraints. Modeling,
Simulation and Control of Nonlinear Engineering Dynamical Systems , 61 -- 70.
23
|
1506.01934 | 2 | 1506 | 2015-08-03T13:57:23 | Rheology of fractal networks | [
"physics.bio-ph",
"cond-mat.soft"
] | We model the cytoskeleton as a fractal network by identifying each segment with a simple Kelvin-Voigt element, with a well defined equilibrium length. The final structure retains the elastic characteristics of a solid or a gel, which may support stress, without relaxing. By considering a very simple regular self-similar structure of segments in series and in parallel, in 1, 2 or 3 dimensions, we are able to express the viscoelasticity of the network as an effective generalised Kelvin-Voigt model with a power law spectrum of retardation times, $\cal L\sim\tau^{\alpha}$. We relate the parameter $\alpha$ with the fractal dimension of the gel. In some regimes ($0<\alpha<1$), we recover the weak power law behaviours of the elastic and viscous moduli with the angular frequencies, $G'\sim G''\sim w^\alpha$, that occur in a variety of soft materials, including living cells. In other regimes, we find different power laws for $G'$ and $G''$. | physics.bio-ph | physics |
Rheology of fractal networks
P. Patr´ıcio,1, 2, ∗ C. R. Leal,1, 3 J. Duarte,1, 4 and C. Janu´ario1
1ISEL - Instituto Superior de Engenharia de Lisboa,
Instituto Polit´ecnico de Lisboa, 1959-007 Lisboa, Portugal.
2CEDOC, Faculdade de Ciencias M´edicas, Universidade Nova de Lisboa, 1169-056 Lisboa, Portugal.
3Centro de Investiga¸cao em Agronomia, Alimentos, Ambiente e Paisagem, LEAF,
Instituto Superior de Agronomia, Universidade de Lisboa, 1349-017 Lisboa, Portugal.
4CAMGSD, Instituto Superior T´ecnico, Universidade de Lisboa, 1049-001 Lisboa, Portugal
(Dated: July 20, 2015)
We model the cytoskeleton as a fractal network by identifying each segment with a simple Kelvin-
Voigt element, with a well defined equilibrium length. The final structure retains the elastic char-
acteristics of a solid or a gel, which may support stress, without relaxing. By considering a very
simple regular self-similar structure of segments in series and in parallel, in 1, 2 or 3 dimensions, we
are able to express the viscoelasticity of the network as an effective generalised Kelvin-Voigt model
with a power law spectrum of retardation times, L ∼ τ α. We relate the parameter α with the fractal
dimension of the gel. In some regimes (0 < α < 1), we recover the weak power law behaviours of
the elastic and viscous moduli with the angular frequencies, G(cid:48) ∼ G(cid:48)(cid:48) ∼ wα, that occur in a variety
of soft materials, including living cells. In other regimes, we find different power laws for G(cid:48) and
G(cid:48)(cid:48).
Microrheology measurements on the cytoskeleton of
the cell revealed interesting weak power law behaviours
[1] (see [2, 3] for recent reviews), which are frequently
associated to the phenomenological "Soft Glassy Materi-
als" (SGM) model [4, 5]. Based on the idea of structural
disorder and metastability, common to all SGMs, this
model relates the power law exponent of the elastic and
viscous moduli, G(cid:48)(w) ∼ G(cid:48)(cid:48)(w) ∼ wα, with 0 < α < 1,
to a mean-field noise temperature x = α + 1, with a glass
transition occurring at x = 1.
However, rather than a generic fluidic system above
the glass transition, the cytoskeleton could be more easily
associated with a polymer network near the sol-gel tran-
sition. Using the ideas of percolation and self-similarity,
the power law exponent α has been previously related
to the fractal dimension df of a flexible polymer cluster.
This relationship is not unique and depends on the un-
derlying assumptions of the proposed microscopic models
(monodispersity vs polydispersity, unscreening vs screen-
ing of excluded volume, etc. -- see [6] for a review). In
particular, if a polydisperse polymeric fractal (prescribed
by bond percolation theory), following Rouse chain dy-
namics (for flexible polymer chains) with fully screened
hydrodynamic interactions is considered [7], the expo-
nent α can take values between 0 and 1 for df ranging
from 2.5 to 1.25, respectively.
More recently, the weak power law behaviour of the
cytoskeleton has been associated with the "Glassy Worm
Like Chain" (GWLC) model [8] (so-called from its anal-
ogy to SGM model). This model defines an average sep-
aration between the crosslinks along the filaments. If the
relaxation modes have a wavelength shorter than this
separation, its relaxation time follows the "Worm Like
Chain" (WLC) model for semiflexible polymers, which
largely compose the cytoskeleton. Otherwise, the relax-
ation spectrum is stretched through an effective Boltz-
mann factor with a characteristic energy that must be
overcome to induce a conformational change of the net-
work. This model predicts a high frequency regime with
a power law exponent α = 3/4 (corresponding to WLC
model) and a low frequency regime with a second power
law exponent that depends on the phenomenological pa-
rameter . Extensions of this model deal with the possi-
bility of transient crosslinking between filaments [9].
It is known that the cytoskeleton, or a cellular tissue,
even in equilibrium, supports a certain amount of stress,
which is imposed by a substrate or other neighbouring
cells. The structural cytoskeleton filaments must retain
a solid character, without a full relaxation. In this article,
we will consider a very simple model of a solid gel, com-
posed of a regular self-similar network of segments with
well defined lengths and rigid bonds. By identifying each
segment with a Kelvin-Voigt viscoelastic element (an hy-
pothesis suggested in [10, 11]), we are able to express
the viscoelasticity of the network as an effective gener-
alised Kelvin-Voigt model with a power law spectrum of
retardation times, L ∼ τ α, where α is related to the net-
work power law distribution of lengths (and eventually to
its fractal dimension), and to the Kelvin-Voigt particular
characteristics of each segment. This relation is not di-
rect, since in 2 or 3 dimensions we have a large collection
of Kelvin-Voigt elements in series and in parallel. When
0 < α < 1, we recover the weak power law behaviours
G(cid:48) ∼ G(cid:48)(cid:48) ∼ wα. In other regimes, for α < 0 and α > 1,
we obtain, first analytically (with some approximations),
and then numerically, different and interesting power law
behaviours for G(cid:48) and G(cid:48)(cid:48).
Let us initially consider a simple 1D self-similar struc-
ture that is defined by first dividing the system size L by
a number NL. We will get NL segments of size L/NL.
Now, we pick only KL < NL segments and repeat the
procedure n times (see Fig. 1, top, for NL = 3 and
Thus, the total creep compliance becomes
(cid:90) l1
ln
J(t) ∼
l−ξ−1Jl(t)dl
2
(5)
where lmin = ln and lmax = l1.
Let us now suppose, in very general terms, that Jl ∼ lδ
and τ = τl ∼ lγ, with γ > 0. In these conditions, we have
(cid:90) τ1
(cid:90) τ1
τn
τn
(cid:16)
τ (δ−ξ)/γ(cid:16)
l−ξ−1Jl
1 − e−t/τl
(cid:17) dl
1 − e−t/τ(cid:17) dτ
dτ
τ
J(t) ∼
∼
dτ
(6)
where τmin = τn and τmax = τ1.
The function defined by L ∼ τ (δ−ξ)/γ corresponds to
the generalised Kelvin-Voigt retardation spectrum [12].
If we make the association
δ − ξ
γ
α1D = α =
(7)
we may write L ∼ τ α. The rheological response of the
structure depends essentially on this parameter α, and
the minimum and maximum retardation times.
To calculate the elastic and viscous moduli, we deter-
mine first the Laplace transform of the creep compliance:
(cid:90) ∞
(cid:90) τ1
J(s) =
J(t)e−stds ∼
0
τn
τ α−1
1
s(1 + sτ )
dτ
(8)
Then, the complex creep compliance, which is given by
the relation J∗(w) = (iw) J(iw). The elastic and viscous
compliances, defined through J∗ = J(cid:48) − iJ(cid:48)(cid:48), are respec-
tively given by
(cid:90) τ1
(cid:90) τ1
τn
τn
J(cid:48)(w) ∼
J(cid:48)(cid:48)(w) ∼
τ α−1
τ α−1
1
1 + (wτ )2 dτ
wτ
1 + (wτ )2 dτ
(9)
(10)
Finally, the complex modulus is related to the complex
compliance through G∗J∗ = 1. The elastic and viscous
moduli, defined by G∗ = G(cid:48) + iG(cid:48)(cid:48), may be determined
from the relations:
J(cid:48)
J(cid:48)(cid:48)
G(cid:48) =
J(cid:48)2 + J(cid:48)(cid:48)2 , G(cid:48)(cid:48) =
J(cid:48)2 + J(cid:48)(cid:48)2
(11)
For w (cid:28) 1/τ1 (wτ (cid:28) 1 for any τ < τmax = τ1), the
integrals of Eq. 9 and 10 are much simplified, and we
obtain the scalings J(cid:48) ∼ w0 and J(cid:48)(cid:48) ∼ w1, with J(cid:48) (cid:29) J(cid:48)(cid:48).
The elastic and viscous moduli scale as
FIG. 1: Top: simple 1D self similar structure of segments.
Bottom: 2D network of segments generated from the 1D self
similar structure represented at the top.
KL = 2). We obtain (NL − KL)K i−1
li = L/N i
of size ln = L/N n
L .
L (with 1 < i < n − 1), and NLK n−1
L
segments of size
segments
L
We may associate each segment, of size l, with a
Kelvin-Voigt element, composed of a spring of stiffness
Gl in parallel with a dashpot of viscosity ηl. The (un-
deformed) spring ensures the length l. The element be-
haves elastically on long times scales and its dynamics
comes from the viscous element.
Its creep compliance
is Jl(t) = Jl(1 − e−t/τl ) [12], where Jl = 1/Gl and
τl = ηl/Gl is the element retardation time. Because the
elements are in series, the total creep compliance is just
the sum of all the elements' creep compliances [12]. Tak-
ing into account the repetition of segment sizes, which
are a consequence of the self similar construction, we may
write
J(t) = (NL − KL)
K i−1
L Jli(t) + NLK n−1
L Jln (t) (1)
n−1(cid:88)
i=1
This sum may be approximated by an integral by mul-
tiplying it by di = 1. If di is considered to be small, we
may write
K i
J(t) ∼
(2)
Ldi/dl corresponds to the density of seg-
The product K i
ments of size l (per unit size). If we invert the relation
l = li = L/N i
LJl(t)
di
dl
dl
l1
(cid:90) ln
L, we obtain
L
l
ln NL
ln
1
i =
and we may write
K i
L = ei ln KL =
di
dl
(cid:19)ξ
= − 1
ln NL
1
l
,
ξ =
ln KL
ln NL
(3)
(4)
G(cid:48) ≈ 1
J(cid:48) ∼ w0, G(cid:48)(cid:48) ≈ J(cid:48)(cid:48)
J(cid:48)2 ∼ w1
(12)
For w (cid:29) 1/τn (wτ (cid:29) 1 for any τ > τmin = τn), the
integrals of Eq. 9 and 10 are again simplified, yielding
the scalings J(cid:48) ∼ w−2 and J(cid:48)(cid:48) ∼ w−1, with J(cid:48) (cid:28) J(cid:48)(cid:48). We
,
(cid:18) L
l
1D 2D obtain then the same scalings for the elastic and viscous
moduli:
3
G(cid:48) ≈ J(cid:48)
J(cid:48)(cid:48) ∼ w1
J(cid:48)(cid:48)2 ∼ w0, G(cid:48)(cid:48) ≈ 1
(13)
The coefficients of G(cid:48) and G(cid:48)(cid:48) are dependent of α but
their scalings are not. In fact, the exponents of w coincide
with the scalings of a simple Kelvin-Voigt model.
For 1/τ1 (cid:28) w (cid:28) 1/τn, we may in some cases extend
the limits of the integrals of Eq. 9 and 10 to τmin → 0
and τmax → ∞, which allow us to obtain the results:
J(cid:48)(w) ∼ π
2
J(cid:48)(cid:48)(w) ∼ π
2
w−αcsc
w−αsec
πα
2
πα
2
(if 0 < α < 2)
(if − 1 < α < 1)
(14)
(15)
When 0 < α < 1, both integrals are well defined and we
recover the weak power law behaviours:
G(cid:48)(w) ∼ G(cid:48)(cid:48)(w) ∼ wα,
G(cid:48)(cid:48)
G(cid:48) = tan
πα
2
(16)
If α < 0, the integral for the elastic compliance (Eq.
9) diverges as τmin = τn → 0. The elastic compliance is
then dominated by the smallest retardation times. In this
case, we have wτn (cid:28) 1, and J(cid:48) ∼ w0. On the contrary,
if α > 2, then this integral is dominated by the largest
retardation times, τmax = τ1. We have wτ1 (cid:29) 1, and
the power law behaviour J(cid:48) ∼ w−2. By the same line
of reasoning, we may determine from Eq. 10 the power
law behaviours J(cid:48)(cid:48) ∼ w1 for α < −1 and J(cid:48)(cid:48) ∼ w−1
for α > 1. Applying these results to each interval of α,
and using the approximations J(cid:48) (cid:29) J(cid:48)(cid:48) for α < 0 and
J(cid:48) (cid:28) J(cid:48)(cid:48) for α > 1 (which we may infer from Eq. 14 and
15), we obtain the power law behaviours:
(α < −1)
(−1 < α < 0)
(0 < α < 1)
(1 < α < 2)
(17)
(18)
(19)
(20)
G(cid:48) ∼ w0
G(cid:48) ∼ w0
G(cid:48) ∼ wα
G(cid:48) ∼ w1−α
G(cid:48) ∼ w0
G(cid:48)(cid:48) ∼ w1
G(cid:48)(cid:48) ∼ w−α
G(cid:48)(cid:48) ∼ wα
G(cid:48)(cid:48) ∼ w1
G(cid:48)(cid:48) ∼ w1
(2 < α)
(21)
We note that for α < −1 or α > 2, we have a single power
law behaviour for G(cid:48) ∼ w0 and G(cid:48)(cid:48) ∼ w1, for all values
of the angular frequency w. Indeed, for these ranges of
α, the whole structure is entirely dominated by only one
Kelvin-Voigt element, corresponding respectively to the
minimum (α < −1) or the maximum (α > 2) retardation
times.
The creep compliance of the 1D self similar struc-
ture (Eq. 1) corresponds to an effective discrete gen-
eralised Kelvin-Voigt model. The accurate values for the
exponents of the power law behaviours, G(cid:48) ∼ wx and
G(cid:48)(cid:48) ∼ wy, for any value of α, are shown in Fig. 2, for
τmax/τmin = 106. Representative plots of G(cid:48)(w) and
G(cid:48)(cid:48)(w) are shown in Fig. 3 (see [11] for calculation de-
tails).
FIG. 2: Exponents x and y vs α of the power laws G(cid:48) ∼ wx
and G(cid:48)(cid:48) ∼ wy, for a generalised Kelvin-Voigt model with
a power law spectrum of retardation times L ∼ τ α,
for
τmax/τmin = 106. The highlighted region refers to the weak
power law behaviour, in which x ≈ y ≈ α (0 < α < 1).
FIG. 3: Elastic and viscous moduli, G(cid:48)(w) (solid line) and
G(cid:48)(cid:48)(w) (dashed line) for a generalised Kelvin-Voigt model with
a power law spectrum of retardation times L ∼ τ α, with α =
−0.5, 0.25, 0.75, 1.5, τmin = 10−3, τmax = 103 (in arbitrary
units).
To create a 2D or a 3D network structure of segments,
we may follow the same self similar construction in the
other dimensions (see 2D network in Fig. 1, bottom).
The boundaries between the segments are identified with
the network crosslinks, which allow us to extend the seg-
ments into the interior part of the system, of size L2 or
L3.
The total creep compliance reflects the deformation of
the 2D or 3D network along the direction of the applied
force, or stress, which we take to be the direction of the
1D structure considered before. But now we have an in-
tricate combination of segments in series and in parallel.
Let us consider first the 2D network structure depicted
in Fig. 1, bottom. In the first iteration of our construc-
tion, we have the boundary segments of the square of
size L divided in NL = 3 parts. At the boundaries be-
tween the segments, we extend new segments into the
interior of the system, creating a matrix of NL × NL ad-
−2−1012300.51αxy10−410−210010210410−410−2100α=−0.5wG'G''10−410−210010210410−410−310−210−1100101α=0.25wG'G''10−410−210010210410−610−410−2100α=0.75wG'G''10−410−210010210410−1010−810−610−410−2α=1.5wG'G'' jacent squares of size L/NL. We have thus generated
ML = NL + 1 lines in the direction of the force. These
lines are in parallel. Each line is composed of NL seg-
ments of size L/NL, in series. This regular structure of
ML lines in parallel with NL segments in series, each of
one with equal creep compliance, gives the matrix creep
compliance:
J mat
l1
(t) =
NL
ML
Jl1(t)
(22)
At the nth iteration, after considering all the different
elements in series and in parallel, we get a surprisingly
simple result. In fact, the number of segments in series
cancels with the number of segments in parallel, yielding:
J 2D(t) =
NL − KL
NL + 1
Jli(t) +
NL
NL + 1
Jln (t)
(23)
In the case of the 3D network structure, due to the
extra dimension, the number of elements in parallel cor-
responds to the square of the number of the elements in
series. After counting all the contributions at the nth
iteration, we have:
J 3D(t) =
NL − KL
(NL + 1)2
Jli(t)
K i−1
L
+
NL
(NL + 1)2
Jln (t)
K n−1
L
(24)
As for the 1D case, these sums may also be approxi-
mated by the integrals:
n−1(cid:88)
i=1
n−1(cid:88)
i=1
(cid:90) ln
l1
(cid:90) ln
l1
J 2D(t) ∼
Jl(t)
di
dl
dl ∼
J 3D(t) ∼
Jl(t)
K i
L
di
dl
dl ∼
(cid:90)
τ δ/γ−1(cid:16)
τ (δ+ξ)/γ−1(cid:16)
1 − e−t/τ(cid:17)
1 − e−t/τ(cid:17)
dτ
(25)
dτ
(cid:90)
The weak power law behaviours exponents become
α2D =
δ
γ
,
α3D =
δ + ξ
γ
(26)
(27)
The 1D self-similar structure analysed here is simply
a line of connected segments. However, the 2D and
3D self-similar structures correspond to networks with
non-integer fractal dimensions. We may cover the whole
structures (except a number of lines) with (K i
L)D boxes
(D = 2, 3 for the 2D or the 3D network, respectively)
of size L/N i
L, yielding a fractal dimension df = Dξ (if
df > 1).
The parameters δ and γ depend on the choice of our
particular model. Several possibilities may be consid-
ered. It is reasonable to assume, as in Stokes's law, that
ηl is proportional to the viscosity of the solvent and to
the size of the element, ηl ∼ l. Then, the retardation
time τl = Jlηl ∼ lδ+1, leading to γ = δ + 1. If we take
4
Jl = 1/Gl constant (this choice was done for the rheo-
logical stiffnesses of the SGM model, or the Rouse model
[13]; in the latter case, it followed from the equipartition
theorem), δ = 0, and we may obtain the interesting result
α3D =
df
3
,
(δ = 0, γ = δ + 1 = 1)
(28)
There are other reasonable scaling laws. We may for
instance invoke the idea of springs in series to justify
Jl ∼ l. In this case, τl ∼ l2, and α3D = (1 + df /3)/2.
In this article, we have presented a new paradigm for
soft solid or gelled materials. We have shown that a
self similar network, with a power law distribution of
segment lengths, may lead to a generalised Kelvin-Voigt
model with a power law spectrum of retardation times,
L ∼ τ α, where α is related to the fractal dimension of
the network. We recover the weak power law behaviours
G(cid:48) ∼ G(cid:48)(cid:48) ∼ wα for 0 < α < 1, often observed in the
cell cytoskeleton, and other soft materials. For α < 0
or α > 1, we also obtain other interesting power law
behaviours, which are characteristic of this effective gen-
eralised Kelvin-Voigt model.
The system here presented compares with the "Soft
Glassy Material"(SGM) model which, in what regards its
linear viscoelastic regime, may be associated with a gen-
eralised Maxwell model with a power law spectrum of re-
laxation times H ∼ 1/τ α (see [11] for a detailed compar-
ison). With this model, we also recover the weak power
law behaviours G(cid:48) ∼ G(cid:48)(cid:48) ∼ wα for 0 < α < 1, but we
have other, substantially different power law behaviours
for α < 0 or α > 1 [5, 11]. The generalised Maxwell
model (and the SGM model) reflects a more fluidic sys-
tem, which contrasts with the solid like or gelled gener-
alised Kelvin-Voigt model (obtained from a self-similar
network) described in this article.
Soft materials, including the cell cytoskeleton, are usu-
ally very complex materials, in which we probably have
a mixture of microscopic Kelvin-Voigt elements associ-
ated with Maxwell relaxation structures, both in series
and in parallel, leading to different scaling behaviours in
different ranges of angular frequencies. Furthermore, the
cytoskeleton is an active structure from which we may
expect novel behaviours, at least for particular ranges of
characteristic times. We hope, however, that this new
solid like or gelled paradigm model may bring improved
understanding of these rheological weak power law be-
haviours, that appear so often in so many complex soft
materials.
∗ Electronic address: [email protected]
[1] B. Fabry, G. N. Maksym, J. P. Butler, M. Glogauer,
D. Navajas, and J. J. Fredberg, Physical review letters
87, 148102 (2001).
5
[2] P. Kollmannsberger and B. Fabry, Annual Review of Ma-
[9] L. Wolff, P. Fernandez, and K. Kroy, New Journal of
terials Research 41, 75 (2011).
Physics 12, 053024 (2010).
[3] R. H. Pritchard, Y. Y. S. Huang, and E. M. Terentjev,
Soft matter 10, 1864 (2014).
[4] P. Sollich, F. Lequeux, P. H´ebraud, and M. E. Cates,
Physical review letters 78, 2020 (1997).
[5] P. Sollich, Physical Review E 58, 738 (1998).
[6] H. H. Winter and M. Mours, in Neutron spin echo spec-
troscopy viscoelasticity rheology (Springer, 1997), pp.
165 -- 234.
[7] M. Muthukumar, Macromolecules 22, 4656 (1989).
[8] K. Kroy and J. Glaser, New Journal of Physics 9, 416
(2007).
[10] M. Balland, N. Desprat, D. Icard, S. F´er´eol, A. Asnacios,
J. Browaeys, S. H´enon, and F. Gallet, Physical Review
E 74, 021911 (2006).
[11] P. Patricio
and C. R. Leal,
arXiv
preprint
arXiv:1506.01927 (2015).
[12] N. Phan-Thien, Understanding Viscoelasticity: An Intro-
duction to Rheology (Springer Science & Business
Media, 2012).
[13] M. Rubinstein and R. Colby, Polymers Physics (Oxford,
2003).
|
1401.0036 | 1 | 1401 | 2013-12-26T07:04:18 | Cyclic and Coherent States in Flocks with Topological Distance | [
"physics.bio-ph"
] | A simple model of the two dimensional collective motion of a group of mobile agents have been studied. Like birds, these agents travel in open free space where each of them interacts with the first $n$ neighbors determined by the topological distance with a free boundary condition. Using the same prescription for interactions used in the Vicsek model with scalar noise it has been observed that the flock, in absence of the noise, arrives at a number of interesting stationary states. In the `single sink state' the entire flock maintains perfect cohesion and coherence. In the `cyclic state' every agent executes a uniform circular motion, and the entire flock executes a pulsating dynamics i.e., expands and contracts periodically between a minimum and a maximum size of the flock. When refreshing rate of the interaction zone is the fastest, the entire flock gets fragmented into smaller clusters of different sizes. On introduction of scalar noise a crossover is observed when the agents cross over from a ballistic motion to a diffusive motion. Expectedly the crossover time is dependent on the strength of the noise $\eta$ and diverges as $\eta \to 0$. In simpler version the translational degrees of freedom of the agents are suppressed but their angular motion are retained. Here agents are the spins, placed at the sites of a square lattice with periodic boundary condition. Every spin interacts with its $n$ = 2, 3 or 4 nearest neighbors. In the stationary state the entire spin pattern moves as a whole when interactions are anisotropic with $n$ = 2 and 3; but it is completely frozen when the interaction is isotropic with $n=4$. These spin configurations have vortex-antivortex pairs whose density increases as the noise $\eta$ increases and follows an excellent finite-size scaling analysis. | physics.bio-ph | physics |
Cyclic and Coherent States in Flocks with Topological Distance
Biplab Bhattacherjee1, K. Bhattacharya2 and S. S. Manna1
1Satyendra Nath Bose National Centre for Basic Sciences, Block-JD, Sector-III, Salt Lake, Kolkata-700098, India
2Department of Physics, Birla Institute of Technology and Science, Pilani - 333 031, Rajasthan, India
A simple model of the two dimensional collective motion of a group of mobile agents have been
studied. Like birds, these agents travel in open free space where each of them interacts with the
first n neighbors determined by the topological distance with a free boundary condition. Using the
same prescription for interactions used in the Vicsek model with scalar noise it has been observed
that the flock, in absence of the noise, arrives at a number of interesting stationary states. One
of the two most prominent states is the 'single sink state' where the entire flock travels along the
same direction maintaining perfect cohesion and coherence. The other state is the 'cyclic state'
where every individual agent executes a uniform circular motion, and the correlation among the
agents guarantees that the entire flock executes a pulsating dynamics i.e., expands and contracts
periodically between a minimum and a maximum size of the flock. We have studied another limiting
situation when refreshing rate of the interaction zone is the fastest. In this case the entire flock gets
fragmented into smaller clusters of different sizes.
On introduction of scalar noise a crossover is observed when the agents cross over from a ballistic
motion to a diffusive motion. Expectedly the crossover time is dependent on the strength of the
noise η and diverges as η → 0.
An even more simpler version of this model has been studied by suppressing the translational
degrees of freedom of the agents but retaining their angular motion. Here agents are the spins,
placed at the sites of a square lattice with periodic boundary condition. Every spin interacts with
its n = 2, 3 or 4 nearest neighbors.
In the stationary state the entire spin pattern moves as a
whole when interactions are anisotropic with n = 2 and 3; but it is completely frozen when the
interaction is isotropic with n = 4. These spin configurations have vortex-antivortex pairs whose
density increases as the noise η increases and follows an excellent finite-size scaling analysis.
PACS numbers: 64.60.ah 64.60.De 64.60.aq 89.75.Hc
1. INTRODUCTION
Flights of a flock of birds or a swarm of bees are
well known examples of living systems exhibiting collec-
tive behavior which are often modeled by groups of self-
propelled mobile agents [1 -- 4]. These groups are called
'cohesive' since each agent maintains a characteristic dis-
tance from other agents and at the same time they are
'coherent' since all agents move along a common direc-
tion.
In addition, the internal structure and organiza-
tion of a flock is found to be complex and far from being
static. Examples include leader-follower relationships in
bird flocks [5], fission-fusion events in fish schools [6] and
social groups in human crowd [7]. Such flocks often travel
over long duration covering distances much larger than
the size of the flocks along arbitrary directions and there-
fore effectively an infinite amount of space is available to
the flock for it's motion. It is also known that while in-
dividual agents often change their directions of flight the
whole flock maintains motion along the same direction in
a stable fashion.
The generic feature of collective motion is, agents are
short-sighted. While an individual agent's behavior is
influenced by the behavior of a small group of local agents
around it, the whole group behaves in unison. In other
words a short range interaction among the agents may
lead to a unique global behavior of the entire group which
implies the existence of a long range correlation among
the agents. Therefore the question is, given a random
distribution of positions and velocities what kind of short
range dynamics can lead to global correlation reflected in
cohesion and coherence among the agents.
This question was first explored in an assembly of self-
propelled particles, known as the Vicsek model [8]. Here
particles (agents) are released at random locations within
a unit square box on the x−y plane with periodic bound-
ary condition and with random velocities. However, in
the deterministic motion the direction of velocity of each
agent i is oriented along the resultant velocity vRi of all
agents within an interaction zone (IZ) of range R around
i.
In reality each agent may make an error in judging
the resultant direction of motion and this has been in-
troduced in the stochastic version of the model where
noise is introduced by topping the orientational angle of
vRi by a random amount ∆θ. Each individual agent is
then moved along the updated velocity direction. A co-
herent phase is observed in the noise-free case with high
agent densities. Moreover a continuous phase transition
is observed on increasing the strength of noise where the
mean flock speed continuously decreases to zero. How-
ever, facets like high density traveling bands occurring at
low noise were revealed in later studies [9, 10] and argu-
ments were put in favor of a discontinuous transition.
In a recent field study by the Starflag group on flocks
it has been shown that the interac-
of Starlings [11],
tions among the birds of a flock do not depend on the
2
(a)
(b)
FIG. 1: (Color online) (a) A flock of N = 16 agents with n = 5 neighbors in the interaction zone and at any arbitrary time
t. The central agent is denoted by the subscript 0 and the velocity vectors of this agent and its 5 neighbors in the interaction
zone are shown using red arrows. (b) In the next time step the velocity of agent 0 is calculated by Eqn. (1) which is along the
resultant of all n + 1 velocity vectors and has the magnitude v.
metric distance but on the topological distance. More
quantitatively they found that each bird interacts with
a fixed number of neighbors, about six or seven in num-
ber, rather than all neighbors within a fixed radial dis-
tance. Observing flocks of Starlings the angular density
distribution of neighboring birds have been found to be
anisotropic e.g., a bird is more likely to keep its near-
est neighbor at its two sides rather than on the front
and back. Fishes [12] have also been found to interact
with neighbors determined by topological rules. Theo-
retical investigations [13, 14] revealed that the behavior
of topology based models are very different from metric
based models [15].
The concept of graph theory based topology was, how-
ever, used [16] to analyze the Vicsek model itself from the
perspective of control theory. The metric distance based
interactions were modeled using graphs with "switching
topology". Such studies also derived the conditions for
the formation of coherent flocks for agents with fixed
topologies [17]. The relevance of underlying graphs or
networks on the nature of collective motion has also been
studied [18, 19].
The observations of the StarFlag group prompted us
to study the collective motion of flocking phenomena in
two dimensions using the interactions depending on the
topological distance. Most crucially we have obtained
very interesting stationary states which have not been
observed before, mainly the cyclic states. At the same
time an increasingly large number of states are found
to be completely cohesive and coherent. The paper is
organized as follows. In section 2 we describe our topo-
logical distance dependent model for collective motion.
The connectivity among such a collection of agents has
been studied as the Random Geometric Graph in section
3. The stationary states of such flocks have been stud-
ied in section 4, the two most prominent states being the
Single Sink State and the Cyclic State. The effect of the
noise on the dynamics and the critical point of transition
have been studied in section 5. Study of the dynamics of
the flock with the fastest refreshing rate of the Interac-
tion Zone has been done in section 6. A simpler version
of the model with its vortex-antivortex states have been
studied on the square lattice in section 7. Finally we
summarize and discuss in section 8.
2. MODEL
In our model the interaction zone has been defined in
the following way. During the flight, each agent i inter-
acts with a short list of n other selected agents that con-
stitute the IZ. It determines it's own velocity using the
Eqn. (1) given below following a synchronous dynamics.
In general, the agent often refreshes the group of agents
in IZ. For example, at the early stage, when the flock is
relaxing to arrive at the stationary state and also during
some stationary states, the inter-agent distances change
with time. Every time the IZ is refreshed, we assume the
criterion of selecting n agents is that they are the first
n nearest neighbors of i. We introduce at this point a
"refreshing rate" which controls how frequently an agent
updates it's IZ. In this paper we study two limiting sit-
uations when these rates are slowest and fastest. In the
slowest rate, the agents do not change at all the the list
of other n agents in their IZs. The IZ for each agent,
constructed at the initial stage, remains the same ever
after, even if n initial neighbors of an agent no longer
remain nearest neighbors as time proceeds. The other
v0(t)v1(t)v2(t)v3(t)v4(t)v5(t)v0(t)v1(t)v2(t)v3(t)v4(t)v5(t)Resultantv0(t+1) 3
FIG. 2: The undirected RGG with N = 1000 nodes distributed randomly within a square box with a free boundary condition.
Each node is linked to its n nearest neighbors; n = 1, 2, 3 and 4 increasing from left to the right. For small n there are many
components of the graph which merge with one another as n increases. The largest component has sizes 9, 150, 988, 1000.
limiting case is when the refreshing rate is the fastest,
the IZ is refreshed for every agent at each time step. The
slowest case has been discussed in sections 4 and 5. The
fastest refreshing rates have been discussed in section 6.
For the spins on the square lattice discussed in section 7,
these two cases actually mean the same since the spins
are firmly fixed at their lattice positions.
The number n of agents in IZ is considered as an in-
teger parameter of the model. As in Vicsek model [8]
the system is updated using a discrete time dynamics.
While the speeds v of all agents are always maintained
to be the same, the orientational angles θi of their ve-
locities are updated by the direction of the resultant of
velocity vectors of all n agents in the interaction zone
and the agent i itself (Fig. 1),
θi(t + 1) = tan−1[Σj sin θj(t)/Σj cos θj(t)]
(1)
where the summation index j runs over all (n + 1) agents
in IZ. The whole flock moves in the infinite space. Fol-
lowing this dynamics, the flock reaches the stationary
state after a certain period of relaxation time. It is ob-
served that the stationary state depends on the initial
positions, initial velocities of the agents, as well as the
neighbor number n. Though a number of different sta-
tionary states have been observed, often the state is a
fixed point or a cycle. We have studied the statistical
properties of these fixed points and cycles and observed
that cohesion and / or coherence are indeed present in
different stationary states.
There are two crucial differences of our model and the
Vicsek's model which we summarize as follows: a) We
have used the same prescription of interaction that had
been claimed in the StarFlag experiment. Accordingly,
an agent here interacts with a fixed number n of nearest
neighbors and not on all neighbors up to a fixed radial
distance. It may be noted that, these two mechanisms
are essentially serving the same purpose. Both prescrip-
tions keep the mutual interactions active among the local
agents only. Therefore our choice of neighbourhood and
Vicsek model's range of interaction are actually on the
same footing.(b) No periodic boundary condition is im-
posed in our model. Therefore agents move out in open
space, yet they often form flocks that exhibit consider-
able amount of cohesion and coherence. Compared to the
Vicsek model, use of the free boundary condition makes
our model less restrictive.
In the limiting case, when the IZ is refreshed at the
slowest rate, it may appear that for an agent, any of
the n neighbours can be at an arbitrarily large distance.
Certainly this is not the case for a cohesive and coherent
flock by definition. Moreover, we will see in the following
that for most of the stationary states, the neighboring
agents indeed remain within the close proximity of an
agent.
3. RANDOM GEOMETRIC GRAPHS
At the initial stage N agents are uniformly distributed
at random locations within a unit square box on the x−y
plane without periodic boundary condition. A random
geometric graph (RGG) [20] is constructed whose vertices
are the agents. At the same time for any arbitrary pair
of vertices i and j, j is defined as a neighbor of i if it
is among the n vertices nearest to i. Then an edge is
assumed to exist from i to j. This implies that the edges
are in general 'directed' since if j is the neighbor of i then
i may or may not be the neighbor of j. Therefore the
resulting graph is inherently a directed graph. However
one can also define a simplified version of the graph by
ignoring the edge directions and consider the graph as
an undirected graph. In the following we refer such an
undirected graph as the RGG.
In Fig. 2 we exhibit the pictorial representation of an
undirected RGG for N = 1000 as n is increased step by
step. For small values of n the graph has many differ-
ent components. As n is increased the components grow
gradually in size, merge into one another and finally the
RGG becomes a single component connected graph cov-
ering all vertices for a certain value of n. Here we have
4
using the detailed knowledge of the set {(cid:126)vi(t)}. Implica-
tion of this is, the positions and velocities are completely
decoupled during the time evolution since the actual po-
sitions of agents do not play any role to determine the
velocities. Therefore the topological connectivity of RGG
remains invariant and is a constant of motion.
4. STATIONARY STATES
4.1. Single Sink States
Initially the N agents are randomly distributed with
uniform probabilities within the unit square box on the
x− y plane. If the corresponding RGG is fully connected
then all agents are assigned the same speed v but along
different directions. The angles θi of the velocity vec-
tors with respect to +x axis are assigned by drawing
them randomly from a uniform probability distribution
between 0 and 2π. As time proceeds the agents soon
come out of the initial unit square box and spread out in
the open two dimensional space. After some initial relax-
ation time the flock arrives at the stationary state. One
of the most common stationary state is the one where
the flock is completely coherent and cohesive. The entire
flock moves along the same direction without changing
the flock's spatial cohesive shape and therefore θi(t) = C
for all i and are independent of time. We call these states
as the 'Single Sink States' (SSS). Therefore this station-
ary state is a fixed point of the dynamical process. A
picture of such a flock has been shown in Fig. 4(a).
4.2. Cyclic States
In cyclic states the velocity directions θi of all agents
change at a constant rate in absence of noise. Therefore
the angular velocity θi(t) = D for all agents is a constant
of motion. Each agent moves in a circular orbit of its
own depending on its initial position, but their radii and
time periods are the same. Consequently the magnitude
of the resultant velocity of the whole flock in the CS has
also a constant value but its direction changes with the
same angular velocity. In addition typically the shape of
the flock is another circle though with some irregularities
and interestingly its radius changes periodically with the
same period of individual agents. Therefore the whole
circular flock pulsates, i.e., periodically expands and con-
tracts where each agent moves on its own fixed circular
trajectory. We explain this motion in Fig. 4(b) by plot-
ting the flock at different instants of time and also show
two individual agent trajectories. We call these states as
the 'Cyclic States' (CS).
For an arbitrary CS, let the probability that the radius
of individual agent's circular trajectory between R and
R + dR be P (R)dR. Given that the uniform speed of the
FIG. 3: (Color online) The fractions g(N, n) of single compo-
nent connected graphs, in a sample of 1000 RGGs, are plotted
against n. For N agents this fraction grows as the number
of neighbors n is gradually increased. System sizes N are
256 (black), 512 (red), 1024 (green), 2048 (blue) and 4096
(magenta), increased from left to right. The number of inde-
pendent configurations used for each value of n is 1000.
shown four figures for n =1, 2, 3 and 4. The randomly
selected positions of all vertices are exactly the same in
these figures. The size of a component is measured by
the number of vertices in that component. In this figure
the RGG becomes fully connected for n = 4.
The structure and connectivity of RGG depend on the
initial positions of N vertices. Therefore we have first
studied how the fraction g(N, n) of connected graphs
grows with n when the flock size N is increased. For
a particular RGG the connectivity is checked using the
'Burning Algorithm' [21] where the fire, initiated at an
arbitrary vertex, propagates along the edges and finally
burns all vertices if and only if the RGG is a single com-
ponent connected graph.
In Fig. 3 we show the plots of g(N, n) against n for
different values of N . To find out if a minimum value
of the neighbor number n exists, one can artificially pre-
pare a linear initial configuration of agents where each
agent has it's right neighbor as the nearest one. This
corresponds to n = 1 but occurrence of such a config-
uration by random selection of positions of the agents
is extremely improbable. Numerically we find that for
small n the g(N, n) takes vanishingly small values. How-
ever, on increasing n, g(N, n) increases very rapidly and
when n is around 7, g(N, n) ≈ 1 i.e., nearly all configu-
rations become connected. With increasing flock size N
the curves slowly shifts to higher values of the neighbor
number n.
Only those flocks whose RGGs are single component
connected graphs are considered for their dynamical evo-
lution. The initial neighbor list is maintained for the en-
tire dynamical evolution of the flock and is never updated
even if all n initial neighbors of an agent no longer re-
main nearest neighbors as time evolves. This means that
the set of agents' velocities {(cid:126)vi(t + 1)} is fully determined
01234567n0.00.20.40.60.81.0g(N,n) 5
(a)
(b)
FIG. 4: (Color online) Flocks of size N = 512 and n = 10, moving with the speed of v =0.03 without noise. (a) Single sink
state: The fully cohesive and coherent motion of the flock is exhibited by its position at three different instants: 10000 (blue),
11000 (green) and 12000 (magenta). Three straight line trajectories of individual agents are also shown. The frame size is
90 × 90 units. (b) Cyclic state: The stationary state pulsating flock has been shown at different time instants: 173,000 (black),
176000 (red), 180000 (green), 188000 (blue), 192000 (brown), 197000 (violet) and 200000 (magenta). The time period is 28835.
Two individual agents' circular trajectories with radius ≈ 137.67 are also shown. The frame size is 600 × 600 units.
period remaining the same. Therefore it appears that
even in the continuous limit of v → 0, the characteristic
features of the flocks reported here remain same.
Starting from the initial state, when the random posi-
tions and velocities are assigned to all agents, the frac-
tions of stationary states that exhibit the SSS and CS are
estimated and are denoted by gSSS(N, n) and gCS(N, n)
respectively. In Fig. 6 these two quantities are plotted
against the neighbor number n for different flock sizes N .
For a certain N , gSSS(N, n) gradually increases with in-
creasing n (Fig. 6(a)). For a given flock size N , however
large, if the neighbor number n is increased to N − 1,
then on using the dynamics mentioned in Eqn. (1) the
stationary state flock must be both cohesive and per-
fectly coherent i.e., gSSS(N, N − 1) = 1. No stationary
state other than SSS can exist in this limiting situation.
On the other hand when n < N − 1 but n is increased,
then gSSS(N, n) also gradually increases and approaches
the value of unity for any arbitrary value of N . At the
same time, gCS(N, n) decreases with n for a fixed N but
increases with N for a fixed n (Fig. 6(b)). Finally in
Fig. 6(c) we plot the sum gSSS(N, n) + gCS(N, n) which
is less than unity for small n, but on increasing n, this
sum gradually increases and reached ≈ 1 for n = 8 for all
N . It is therefore concluded that if the neighbor num-
ber is increased all other states gradually disappear and
only SSS and CS states mostly dominate but ultimately
for even larger value of n it is the SSS state that only
survives.
FIG. 5: The probability distribution of the radii of the indi-
vidual agent's circular orbits in the cyclic states. The power
law has an exponent of τ = 1.99(2).
agents is v and their angular velocities is θ, the radius
of the circular trajectory is R = v/ θ. We have studied
a large number of such cyclic states and measured the
radii of the agents' orbits. In Fig. 5 we show the prob-
ability distribution of these radii which follows a power
law distribution P (R) ∼ R−τ with τ = 1.99(2).
Throughout this paper we have used only one value of
the agent speed, i.e., v = 0.03. If the speed is reduced by
a certain factor a CS state remains CS but all the char-
acteristic lengths are reduced by the same factor. The
radius of the circular orbit of every agent and also the
size of the flock are reduced by the same factor, the time
100101102103104R10-510-410-310-210-1100P(R) 6
FIG. 6: (Color online) The occurrence of two most prominent stationary states when the neighbor number n has been varied
over a range from 5 to 10 and for for different flock sizes N = 64 (black), 256 (red), 512 (green) and 1024 (blue). (a) The
fraction gSSS(N, n) of SSS has been plotted against n. (b) The fraction gCS(N, n) of CS has been plotted against n. (c) The
sum of gSSS(N, n) + gCS(N, n) has been plotted and it is seen that beyond n ≈ 8 the sum is approximately unity.
(a)
(b)
(c)
FIG. 7: (Color online) Flocks of size N = 512 and n = 5, moving with the speed of v =0.03 without noise. The positions of
the agents are marked by black dots and three individual agent's trajectories are shown in each case by red, blue and magenta
colors. (a) Distributed sink state: Every agent moves along a fixed direction θi = Ci of its own which is different in general
from the directions of motion of other agents. (b) Cycloid state: In the stationary state the trajectory of each agent is a cycloid.
(c) Space-filling state: The trajectory of an agent never repeats itself but gradually fills up the space between two concentric
circles. The frame sizes are 40000, 15000 and 350 units respectively.
4.3. Other states
In addition there are a number of other stationary
states, few of them are described below, but the list may
not be exhaustive.
(i) Each agent has a constant velocity but their direc-
tions are different for different agents. For example the
i-th agent has its direction of velocity θi = Ci. In this
case the agents, after some relaxation time, moves out-
ward radially. The shape of the flock is approximately
circular, again with some irregularities, and the radius of
the flock increases at a uniform rate. We call these states
as the 'Distributed Sink States'(DSS). In Fig. 7(a) an ex-
ample of the DSS has been shown. The position of the
flock is shown at t = 500000 and three agents' trajectories
have been shown using different colors.
(ii) In another type of stationary state the trajectories
of the individual agents are very similar to cycloids. Each
agent moves radially outward in a nearly cycloidal motion
(Fig. 7(b)). A considerable number of agents form a
flock of circular shape but others are scattered around
this circular flock. We call these as 'Cycloid States'.
(iii) Thirdly there can be rosette type stationary states.
The trajectory of each agent is like a rosette which never
closes and lies between two concentric circles. Conse-
quently in the long time limit the trajectories fill the
space between the two circles. This means that the mean
separation between consecutive intersections of the agent
trajectory with a radial section gradually vanishes as the
trajectory evolves for a longer time. We call these as
'Space-Filling States'. Three such rosette trajectories
and the position of the flock have been shown in Fig.
7(c).
Few points may be mentioned here about the charac-
teristics of the different stationary states. For example a
possible anisotropic effect on the stationary states may
exist due to the choice of the unit square box for releas-
ing the agents. Initially the positions of the agents are
selected randomly within unit square box on the two di-
mensional plane. We have compared that if the agents'
5678910n0.00.20.40.60.81.0gSSS(N,n)(a)5678910n0.00.20.40.60.81.0gCS(N,n)(b)5678910n0.850.900.951.00gSSS(N,n)+gCS(N,n)(c) locations are selected randomly within a circle of radius
1/2, no appreciable change have been observed in the
fractions of different stationary states.
The center of mass of the entire flock has different kinds
of trajectories in different stationary states. In SSS, the
center of mass moves in a straight line exactly similar to
all other agents. In CS, the trajectory of the center of
mass is also a circle, but it's radius is not the same as the
radius of the orbit of the individual agents, but it is some
what larger. In DSS the dynamics of the center of mass is
similar to that of the agents. Although, the shape of the
flock is approximately circular, but the fact that, at any
instant the positioning of the agents on the circumference
is not uniform, makes the center of mass move radially
outwards in a straight line. Motion is indeed unbounded.
In cycloid states, the trajectory of the center of mass is
also a cycloid and radially outwards. The motion here
is also unbounded. In space filling states, the trajectory
of the center of mass is rosette type.
In this case the
trajectory is bounded.
How sensitive are the final stationary states on the
choice of the random initial values of {xi, yi} and {θi}
for the N agents? To study this point, we tried with
a flock that evolves from a certain initial configuration
that evolves to a CS. Now we again evolve the same flock,
but this time we slightly change the initial configuration
randomly by xi = xi + a.10−4.r and yi = yi + a.10−4.r
where r is a random number. The directions {θi} of
the velocity vectors are maintained the same. We then
tune a, and found that with 0 < a < 1.70 the stationary
state is still a CS, but with different values of the orbit
radius. When a ≥ 1.75 the stationary state becomes SSS.
We conclude that with some amount of perturbation the
character of the stationary state remains same, but with
even stronger perturbation the stationary state changes.
A preliminary calculation with our model in three di-
mensions shows the following features.
In general, to
obtain connected graphs, the value of n needs to be large
compared to what is required in two dimensions. Almost
always the dynamics leads to a SSS in the steady state.
We did not find any other state starting from random
initial conditions.
5. DYNAMICS IN PRESENCE OF NOISE
Studying the role of noise on the dynamics of the flock
is very crucial. It is assumed that every agent makes a
certain amount of error in judging the angle of its velocity
vector at each time step. More precisely given the angles
θ(t) of velocity vectors of all n + 1 agents within the
interaction zone at time t, it first calculates the resultant
of these vectors using the Eqn.
It then tops up
this angle by a random amount ζ(η) which is uniformly
distributed within {−η/2, η/2}. Therefore the modified
(1).
7
Eqn. (1) reads as:
θi(t + 1) = tan−1[Σj sin θj(t)/Σj cos θj(t)] + ζ(η).
(2)
The role of the noise is to randomize the determinis-
tic dynamics and quite expectedly the stationary state
structures of the flocks exhibited in the SSS and CS pat-
terns are gradually lost. We have studied the effect of
noise on both these states by gradually increasing the
strength of the noise η.
In both cases we use a flock
of N = 512 agents, each of them interacts with n = 10
nearest neighbors and travel with speed v = 0.03. Ini-
tially all of them are released within the square box of
size unity. We first run the dynamics without any noise
i.e., η = 0 and ensure that the stationary state pattern
is indeed a single sink state. As the dynamics proceeds
we calculate the maximal distance Rm(t) and the aver-
age distance Ra(t) of an agent from the center of mass
(xc(t), yc(t)) of the flock. In Fig. 8(a) we plot these two
quantities against time and they are exactly horizontal
curves which are the signatures of the SSS state. These
simulations are then repeated for η > 0 and the varia-
tions of Rm(t) and Ra(t) have been shown in Figs. 8(b),
8(c) and 8(d) for η = 0.2, 0.5 and 1 respectively. In all
four cases the flock starts with the same positions and
velocities of the agents. It is seen that on increasing the
strength of noise the stochasticity gradually sets in and
the variations of Rm(t) and Ra(t) gradually become ran-
dom. A similar plot has been exhibited in Fig. 9 for
the CS but for only a single flock. The Fig. 9(a) shows
the zero noise case and the curves are periodic. How-
ever when the noise level is increased (Figs. 9(b)-(c)) it
distorts the periodicity. With small values of η the vari-
ations are slightly distorted from the periodic variations,
however with larger strength of noise the distortion is
much more. Finally for η = 1 the fluctuations look ran-
dom and similar to those in the SSS (Fig. 9(d)).
Next we calculated the mean square displacement
(cid:104)r2(t, η)(cid:105) from the origin as time passes. The averag-
ing has been done for a single agent within a flock and
over many such independent flock samples. The noise
strength has been varied over a wide range of values.
In Fig. 10 we have displayed the variation of (cid:104)r2(t, η)(cid:105)
against the time t using a log − log scale for six different
values of η. Here again we consider flocks with fully con-
nected RGGs. On the other hand with zero noise these
configurations may lead to any of the possible stationary
states. Simulating up to a maximal time of T = 108 we
observed a cross-over behavior in the mean square dis-
placement. When η is very small (cid:104)r2(t, η)(cid:105) ∼ t2κ with
κ ≈ 1 which implies that the flock maintains a ballis-
tic motion at the early stage, i.e., the coherence is still
maintained during this period. On the other hand after
a long time one gets κ = 1/2 which indicates the diffu-
sive behavior. This implies that even if a little noise is
applied for a long time, the effect of the noise becomes
so strong that the flock can no longer maintain a cohe-
sive and coherent structure any more and agents diffuse
away in space. Therefore for any value of η there is a
cross-over from the ballistic to diffusive behavior. Con-
sequently a cross-over time tc(η) can be defined such that
for short times t << tc(η) the dynamics is ballistic with
κ = 1 and for t >> tc(η) the dynamics is diffusive with
κ = 1/2. In Fig. 10 we show this behavior and observe
that the crossover time depends explicitly on the value
of η and diverges as η → 0. The value of tc(η) has been
estimated by the time coordinate of the point of inter-
section of fitted straight lines in the two regimes of Fig.
10: for t << tc(η) and for t >> tc(η). The value of tc(η)
so estimated diverges as tc(η) ∼ η−2.52.
This transition is more explicitly demonstrated using
a plot of the Order Parameter (OP) M (η) against η (Fig.
11). The OP is defined as the time averaged magnitude
of the resultant of all agents' velocity vectors, scaled by
its maximum value
M (η) = (cid:104)ΣN
j=1vj(cid:105)/(N v)
(3)
where (cid:104)...(cid:105) denotes the time average over a long period
of time in the stationary state. In Fig. 11 (a) we have
plotted M (η) against η at an interval of ∆η = 0.1 for
different system sizes from N = 64 to 1024.
In these
simulations the initial conditions are chosen to be com-
pletely coherent so that the velocity vectors of all agents
are in the same direction. In the absence of noise this
situation is maintained and M (0) = 1 at all times for all
system sizes. However for η > 0 noise sets in, but in the
stationary state one still gets a non-zero OP. On further
increasing η the OP decreases monotonically and ulti-
mately vanishes. Therefore there exists a critical value
ηc(N ) of the noise parameter where the transition from
the ordered state to disordered state takes place. It is ob-
served in Fig. 11(a) that as the system size N is enlarged
the transition becomes more and more sharp and shifts
to the regime of small η. Further we have calculated the
Binder cumulant G(η) = 1 − (cid:104)M 4(η)(cid:105)/3(cid:104)M 2(η)(cid:105)2 and
plotted against η in Fig. 11(b) for the same system sizes
[22]. The value of G(η) drops from a constant value of
around 2/3 at the small η regime to about 1/3 for large
values of η.
The transition point ηc can be estimated in the follow-
ing two ways. For each curve in Fig. 11(a) we calculate
the value of η1/2(N ) for which M (η) = 1/2. We define
η1/2(N ) is the characteristic noise level where the tran-
sition takes place. By interpolation of the plots in Fig.
11(a) of the points around M (η) = 1/2 we have esti-
mated η1/2(N ). These estimates are then extrapolated
in Fig. 11(c) as:
η1/2(N ) = η1/2(∞) + AN−1/ν.
(4)
On tuning the trial values of ν very slowly we found that
for ν ≈ 2.86 the error in the least square fit of the above
finite-size correction formula is minimum. Therefore the
8
extrapolated η1/2(∞) ≈ 1.70 is the critical noise strength
ηc according to our estimate. A similar calculation has
also been done using the Binder cumulant. From this
calculation we estimated ηc = 1.82 and ν = 3.50. The
differences between the two estimates are considered as
the error in the measured values which are 0.12 and 0.64
for ηc and ν respectively.
Since the agents are uniformly distributed initially,
the edges of RGG also are homogeneously distributed
as shown in Fig. 2(d). However with time evolution
these edges change their positions but their connectivity
do not change, i.e., the end nodes of every edge are al-
ways fixed since the neighbor list does not change. How
they look in the stationary state has been exhibited in
Fig. 12. This is the picture of the circle shaped flock
in the cyclic state. The blue dots represent the agents
and the red lines represent the edges. What is interest-
ing to note is that the system self-organizes itself so that
not only agents but also the edges are constrained to be
within a very limited region of the space. Very few edges
criss-cross the flock from one side to the opposite side.
Initially each agents had its n neighbors at its closest dis-
tances. After passing through the relaxation stage and
arriving at the stationary state, when the shape of the
flock is completely different from its initial shape, most of
the agents maintain their connections with other agents
in their local neighborhood only.
6. THE FASTEST REFRESHING RATE OF THE
INTERACTION ZONE
Here we consider the case corresponding to the fastest
refreshing rate of the interaction zone, i.e., when every
agent updates its n nearest neighbors at every time step.
Consequently, the RGG is no longer a constant of motion
in this case and is updated at each time step. While per-
forming simulations of this version, we first notice that
in the long time stationary state the entire flock becomes
fragmented with probability one into different clusters.
For a flock of N agents with neighbor number n, the
minimum number of agents in a cluster is n + 1. An
agent of a particular cluster has all n neighbors which
are members of that cluster only. In the stationary state
all agents of a cluster has exactly the same direction of
velocity and the entire cluster moves along this direction
with uniform speed v. Therefore, the velocity direction
θ of a particular cluster can be looked upon as the iden-
tification label of that cluster. The shape of the flock
is approximately circular since each cluster travels out-
ward with same speed (Fig. 13(a)). Moreover, within a
cluster if the agent i is a neighbor of the agent j then j
is also a neighbor of i. Therefore the sub-graph of the
entire RGG specific to a cluster is completely undirected
and the corresponding part of the adjacency matrix is
symmetric (Fig. 13(b)). This immediately implies that
9
FIG. 8: (Color online) Effect of noise is exhibited on a single flock (N = 512 agents, each having n = 10 neighbors) which goes
to a single sink state without any noise. Variation of the maximal radius Rm (blue) and the average radius Ra (red) have been
shown with different strengths of the noise parameter: (a) η = 0, (b) 0.2, (c) 0.5 and (d) 1. The initial positions and velocities
are same in all four cases.
FIG. 9: (Color online) Effect of noise is exhibited on a single flock (N = 512 agents, each having n = 10 neighbors) which
goes to a cyclic state without any noise. Variation of the maximal radius Rm (blue) and the average radius Ra (red) have been
shown with different strengths of the noise parameter: (a) η = 0, (b) 0.2, (c) 0.5 and (d) 1. The initial positions and velocities
are same in all four cases.
the N × N adjacency matrix of the entire flock can be
written in a block-diagonal form by assigning suitable
identification labels of different agents.
A natural question would be how the probability dis-
tribution D(s) of different cluster sizes depends on the
cluster size s. To answer this question a large number
of independent flocks have been simulated and each of
them was evolved to its stationary state. In the station-
ary state the sizes of the individual clusters are measured
using the burning method. In Fig. 14(a) a plot of D(s)
050000100000150000200000t01234567Ra(t), Rm(t)(a)050000100000150000200000t010203040Ra(t), Rm(t)(b)050000100000150000200000t020406080Ra(t), Rm(t)(c)050000100000150000200000t050100Ra(t), Rm(t)(d)050000100000150000200000t0100200300400500Ra(t), Rm(t)(a)050000100000150000200000t0100200300400500600Ra(t), Rm(t)(b)050000100000150000200000t0500100015002000Ra(t), Rm(t)(c)050000100000150000200000t050100150200250300Ra(t), Rm(t)(d) 10
the sites of a regular lattice, but the directions θi(t) of
the spins are the only dynamical variables that evolve
with time following Eqn.
(1). More specifically, spins
are placed on a square lattice with different choices for
the first n neighbors and we study the spatio-temporal
patterns that emerge during the time evolution of the
angular variables {θi(t)}. The arrangement of the spins
allows to draw a parallel with the dynamics of planar
spins in the two dimensional XY model.
The connections between the Vicsek model [8] and the
2d XY model [23] have been explored since long [3, 19,
24]. It is well known that in the limit of speed v → 0
the dynamics of the Vicsek model would exactly map on
to the finite temperature Monte Carlo dynamics of the
2d XY model. However, in the latter model any long-
range ordered phase is absent. Instead a quasi-long-range
ordered phase appears at the low temperatures and the
transition to the disordered phase is associated with the
simultaneous unbinding and increase of vortex-antivortex
(VAV) pairs. In the low temperature phase VAV pairs
are to be found in tightly bound states. We find that
our model defined on the square lattice also gives rise to
VAV pairs and we determine the density of such pairs, as
a function of the noise amplitude η.
We define the IZ of an agent with respect to its n near-
est neighbors on the square lattice of size L× L with the
periodic boundary condition. For n = 2, the IZ includes
the top and the right nearest neighbors. For n = 3, the
left nearest neighbor is also included and in the case of
n = 4, all the four nearest neighbors are included. We
notice that for the n = 4 case, the Vicsek model with
spins similarly placed on the square lattice and interact
with a range R = 1 and our model are same. As be-
fore we study the dynamics of the spin system with and
without noise.
In the absence of noise, beginning from arbitrary initial
conditions for θi's, the dynamics results in the formation
of VAV pairs. For n = 2 and 3, the interactions are
anisotropic. Consequently the entire spin pattern in the
stationary state as well as all VAV pairs are mobile and
in general all the spin orientations θi's change with time.
In comparison, for n = 4, all the θi's remain frozen in
time which also implies that all the VAV pairs are an-
chored. We find that the choice of the IZ, in addition
to the periodic boundary condition, fixes the direction of
motion of the VAV pairs. In the Fig. 15(a) an instan-
taneous configuration of the spins is plotted for a lattice
with L = 64 and n = 2. In general for the n = 2 case
the entire spin pattern moves on the average along the
diagonal direction from top-right to bottom-left. For the
case, n = 3, the vortices travel from the bottom to the
top.
Let the V (η, L) be the number of VAV pairs observed
at noise η in lattice of size L. The number of VAV pairs
observed at zero noise, V (0, L) is plotted against the
value of n in the Fig. 15(b) for different lattice sizes. We
FIG. 10:
(Color online) The mean square displacement
(cid:104)r2(t, η)(cid:105) of an agent from the origin has been plotted against
time for different values of the noise parameter η = 0.2
(black), 0.5 (red), 1.0 (green), 2.0 (blue), 3.0 (magenta) and
4.0 (brown). The flock size N = 512 and the neighbor num-
ber n = 10. Two short straight lines are the guides to the eye
whose slopes are κ = 1/2 and 1.
vs. s for s > n has been shown on a semi-log scale for N
= 512 and n = 5, the data being collected using a sample
size of 20000 independent flocks. Apart from some noise
at the tail end and a maximum around the smallest value
of s the plot fits well to a straight line, implying an ex-
ponentially decaying form of the probability distribution.
We conclude D(s) ∼ exp(−s/sc) where sc ≈ 7.2(1).
Next, the average number of clusters (cid:104)ns(N, n)(cid:105) has
been calculated and plotted in the inset of Fig. 14(b)
for N = 128, 256 and 512 and n = 5 on a log − log scale.
Again, apart from the tail end, the plots fit very nicely to
parallel straight lines, the slopes of which are estimated
to be 1.168(5). In the main part of Fig. 14(b) a scal-
ing has been shown which exhibits a nice data collapse,
corresponding to the following form
(cid:104)ns(N, n)(cid:105)N−1.1 ∼ n−1.168.
(5)
This implies that as the neighbor number n increases,
there would be fewer clusters in the stationary state. On
the other hand, for a specific value of n, the average clus-
ter number grows with the flock size as N 1.1. Assuming
that the above scaling relation holds good for the en-
tire range of n, for a given N one can define a cut-off
value of n = nc such that (cid:104)ns(N, nc)(cid:105) = 1 which leads
to nc(N ) ∼ N 1.1/1.168 = N 0.94. However our simulations
suggest that due to the presence of an upward bending
at the tail end, the above scaling relation does not work
at this end and nc(N ) is actually of the order of N .
7. VORTICES ON THE SQUARE LATTICE
In this section we studied a simpler version of our
model where every agent is a spin vector. They are no
more mobile, their positions are completely quenched at
100101102103104105106107108t1001021041061081010<r2(t,η)>κ=1/2κ=1 11
FIG. 11: (Color online) (a) The stationary state order parameter M (η) and (b) the Binder cumulant G(η) have been plotted
for the system sizes N = 64 (black), 128 (red), 256 (green), 512 (blue) and 1024 (magenta) where all agents start with their
initial velocities in the same direction. System size increases from right to left. (c) Extrapolation of ηc(N ) values determined
from (a) and (b). We obtained ηc(N ) = 1.70 + N−1/2.86 and ηc(N ) = 1.82 + N−1/3.50 respectively.
basic frequencies f0, f1 and f2 all of which are rational
multiples of 1/L (Fig. 16(b)). The time evolution can be
explained as a periodic oscillation with f1 and f2 (since
f2 = 2f1) riding on a very slow mode. These features
carry over to L = 128 as well. We find that in the case
of n = 3, the spectrum is similar but the frequencies are
not simple multiples of 1/L.
It is known that metastable vortices are produced at
low temperatures in the 2d XY model when equilibrium
is achieved beginning from arbitrary initial conditions.
However, these vortices are not responsible for the VAV
unbinding transition [25]. Therefore, we study the effect
of noise by "cooling down" [26] the system to zero noise
level starting from a high value of noise. At each noise
level the system initially passes through 105 time steps;
after this relaxation it passes through an additional 104
time steps and then moves to the next lower level of noise.
We begin around the value of noise given by η = 3.7 and
decrease η by an amount 0.07 in each step. This method
suppresses the generation of the VAV pairs at low noise.
At zero noise the vortices are absent in contrast to the
statistics discussed in the previous paragraphs where the
cooling down method was not employed.
At zero noise we find the spin system reaches the glob-
ally ordered state where all the spin vectors are oriented
in the same direction. We believe that this is due to the
finite size effect of the lattice. The spin configuration at
low noise of η = 0.15 has been shown in the Fig. 17(a)
for L = 64 and n = 4 where not only vortices are ab-
sent but the long-range order is also not present. At the
higher noise levels VAV pairs start appearing and there
is a rapid increase in the number of pairs with further
increase in noise. All the VAV pairs appearing initially
are tightly bound i.e., lattice spacing is small between the
members in a pair as in Fig. 17(b)) but at higher noise
members in a pair are seen to unbound (Fig. 17(c)).
The order parameter M (η) tends to unity as η → 0 as
shown in the Fig. 18(a). In Fig.18(b) we plot the vortex-
pair density, defined as ρ = V /L2. The natural collapse
of the plots belonging to different system sizes indicates
a functional dependence of ρ on n and η independent of
FIG. 12: (Color online) The position of a flock (N = 512, n
= 10) in CS by blue dots and links by red lines. The frame
size is 150 units.
observe during the time evolution for a given initial con-
dition that the number of VAV pairs initially decays and
then becomes stationary. However, different initial con-
ditions leads to different values at the stationary states.
The bars indicate the dispersion that is observed for dif-
ferent initial conditions which lead to non-zero number
of VAV pairs. For the lattice size L = 128 we wait for 105
steps before calculating number of VAV pairs. The inset
to the Fig. 15(b) shows the percentage of initial condi-
tions that lead to non-zero number of VAV pairs. We
find as lattice size increases, arbitrary initial conditions,
almost always, lead to states with VAV pairs.
The nature of the time variation of the spin angle θi's
for the cases n = 2 and n = 3 are found to be quite
complex. In the Fig. 16(a) we plot the time series corre-
sponding to the oscillation of the x-component of a typi-
cal spin vector for L = 64 and n = 2. The corresponding
power spectrum apparently reveals the presence of three
012345η0.00.20.40.60.81.0M(η)(a)012345η0.30.40.50.60.7G(η)(b)00.10.20.3N-1/ν234ηc(N)(c) 12
(a)
(b)
FIG. 13: (Color online) (a) Positions of a flock of N = 512 agents with n = 5 neighbors in the interaction zone and at four
different time instants: 100 (black), 200 (red), 300 (green) and 400 (magenta). The RGG has been updated at every time step.
Three individual agents' straight line trajectories have also been shown. There are a total of 29 clusters. (b) The sub-graph
of the RGG corresponding to a specific cluster of 16 agents have been shown. An arrow has been drawn from the agent i to
the agent j if j is one of the n neighbors of i. There are a total of 16 × 5 = 80 distinct links and each link is directed in both
directions. It may also be noted that whole set of links are restricted to the nodes of the cluster only.
(a)
(b)
FIG. 14: (Color online) (a) The probability distribution D(s) of the cluster sizes s seems likely to have an exponentially
decaying form exp(−s/sc) where sc ≈ 7.2. (b) The inset shows the plot of the average cluster size (cid:104)s(N, n)(cid:105) for N = 512 and
n = 5 with neighbor number n on a log − log scale: N = 128 (blue), 256 (green) and 512 (red). In the main plot the vertical
axis has been scaled by N 1.1 which leads to a data collapse.
L. To understand this behavior we obtain the collapse
of the logarithm of 1/ρ in Fig.18(c) for different values
of L and n. The plot reveals that in the region where
VAV pairs start proliferating ρ ∼ exp(− Anβ
ηα ), where A,
β and α are constants. We estimate α by averaging slope
of individual curves which yields α = 2.68 ± 0.13. This
gives β = γα = 0.75. This result is in contrast to the
relation ρ ∼ exp(− a
T ), where a is VAV pair energy and
T is the temperature for the 2d XY model [26].
8. SUMMARY AND DISCUSSION
To summarize, we have studied a simple model of the
collective behavior of N interacting mobile agents which
travel in the open free space.
In this model we have
incorporated the observation of the StarFlag experiment
which advocates the necessity of using the topological dis-
tance instead of the metric distance. Each agent interacts
with a group of n selected agents around it who are po-
sitioned within an interaction zone. Every agent freshly
selects at a certain rate the group of agents within the IZ.
-2525-2525 12.88 13.06 27.64 27.82020406080100s100101102103104105D(s)56810142030404050n0.0030.0060.010.020.040.07<ns(N,n)> N-1.15681014203040n1248163264<ns(N,n)> 13
(a)
(b)
FIG. 15: (Color online) (a) Vortex-antivortex pairs of the spin systems in the stationary state in a square lattice with size
L = 64 and n = 2 with zero noise. The orientations of the spins change with time in such a way so that the entire spin pattern
with vortices and the antivortices move with a uniform speed from the top-right corner to the bottom-left corner. Vortices
and antivortices are marked by filled red and blue circles. (b) The number V (0, L) of vortex-antivortex pairs at zero noise for
different values of n in lattices of three different sizes: L = 32 (circles), 64 (squares) and 128 (triangles). The bars indicate
the standard deviation in the values obtained from around 100 configurations in each case.
In the inset the percentage of
configurations which lead to steady states with vortex pairs is plotted against n.
FIG. 16: (Color online) (a) The x-component of a typical spin vector is plotted against time for lattice with L = 64 and
n = 2 at zero noise. (b) The corresponding power spectrum indicates the presence of three basic frequencies f0 = 1/(4096L),
f1 = 1/(3L) and f2 = 2/(3L). The higher frequencies can be expressed as combinations of the f1 and f2 and essentially become
harmonics of f1 since f2 = 2f1.
The selection criterion is to choose the first n neighbors.
In this paper we have studied two limiting situations, i.e.,
when the refreshing rates are fastest and the slowest. We
first studied when the refreshing rate is slowest i.e., when
the IZ for each agent is determined at the beginning and
is never updated. All the agents follow the interaction
rule in Vicsek model. It has been observed that in ab-
sence of noise, starting from a small localized region of
space the agents gradually spread as time passes, so that
after some relaxation time the flock arrives at a station-
ary state. The most prominent stationary states are the
single sink state and the cyclic state. Using numerical
methods we claim that the frequencies of occurrence of
other stationary states like the distributed sink states,
cycloid states and the space-filling states goes to zero as
the neighbor number increases to about 8. Beyond n = 8
only the SSS and CS states dominate. Finally as n ap-
proaches the flock size N , only the SSS states dominate.
Interestingly, it has also been observed that the actual
metric distances of the n topological neighbors do not
become arbitrarily large in all these stationary states in
absence of noise. For the SSS states it is true exactly.
For the Cyclic states also the topological neighbors re-
main within a finite distance from an agent (see Fig. 12).
234n05101520V(0,L)234n050100Vortex bearing states (%)02004006008001000t-0.6-0.4-0.200.20.40.6Cos(θ(t))(a)10-610-510-410-310-210-1100f00.10.20.30.40.50.6P(f)(b)f0f1f2 14
(a)
(b)
(c)
FIG. 17: (Color online) The figure shows proliferation of vortices as noise is increased in a square lattice of length L = 64 with
n = 4. (a) For η = 0.15 there are no vortices, (b) for η = 0.35 vortices begin to appear, there are only two vortex-antivortex
pairs and (c) for η = 0.40 twenty one vortex-antivortex pairs can be seen. Filled circles of two different colors, red for vortices
and blue for antivortices, have been drawn around the vortex centers.
(a)
(b)
(c)
FIG. 18: (Color online) The behavior of different quantities against noise for lattice sizes L = 32 (black), 64 (red) and 128
(green) with values of n = 2 (circle), 3 (square) and 4 (triangle). (a) The variation of the order parameter M (η, L) against
noise η are shown for L = 128 and for n = 2, 3 and 4. (b) The vortex-pair density ρ(η, L) against η. (c) Scaling collapse of
ln(1/ρ(η, L)) against nγ/η. For the collapse we use γ = 0.28. The dashed straight line, having slope α = 2.68, is a guide for
the eye and indicates the power-law nature in the low noise regime.
Moreover, when the refreshing rate is the fastest, every
fragmented cluster travels in a single sink state. There-
fore, it is this close proximity of the neighbors that gives
rise to the cohesiveness present in our model.
Further, on the application of noise a crossover takes
place from the ballistic motion to the diffusive motion
and the crossover time depends on the strength of the
noise η, which diverges as η → 0. Further the calculation
of the Order Parameter M (η) and the Binder cumulant
G(η) lead us to estimate the critical noise ηc required for
the continuous transition from the ordered to the disor-
dered phase. Secondly, when the refreshing rate is fastest,
each agent freshly determines its neighbors in the interac-
tion zone at every time step. In the stationary state the
flock gets fragmented into a number of smaller clusters of
different sizes. The agents in a cluster move completely
coherently, different cluster has different direction of mo-
tion.
A simpler version of the model has also been studied in
the limit of the speed v → 0 when the positions of spins
are completely frozen at the sites of a square lattice, but
their orientational angles θi(t) evolve with time again by
the Vicsek interaction. Here for n = 4 the spin configura-
tion is completely static. On the other hand for n = 2 and
3, the entire spin configuration moves along the diagonal
and parallel to the asymmetry axis respectively. Further
01234η00.20.40.60.81M(η,L)01234η00.050.10.15ρ(η, L)0.30.40.50.60.70.80.91.0nγ / η23451020ln (1/ρ(η,L)) we have observed that the density of vortex-antivortex
pairs increases with the strength of the noise and fits to
a nice finite-size scaling behavior.
Overall, our findings suggest that complex spatio-
temporal patterns may emerge in the interplay between
an underlying network structure and collective motion.
We believe that our study would also be relevant in the
general problem of consensus development in networked
agents [27] and as such issues like undesired synchroniza-
tion observed in real-world networks [28]. We observed
that multiple frequencies develop during oscillations of
different dynamical variables. Whether there is a pos-
sibility that a cascade of frequencies develop eventually
leading to chaotic behavior remains an open question.
ACKNOWLEDGMENT
Useful discussion with Shraddha Mishra is thankfully
acknowledged. K.B. acknowledges the support from the
BITS Research Initiation Grant Fund.
[1] Reynolds, C. W. (1987), Flocks, herds and schools: A dis-
tributed behavioral model, Computer Graphics, 21, 25-
34.
[2] Vicsek, T. and Zafeiris, A. (2012), Collective motion,
Phys. Rep., 517, 71-140.
[3] Toner, J. and Tu, Y. (1995), Long-Range Order in a Two-
Dimensional Dynamical XY Model: How Birds Fly To-
gether, Phys. Rev. Lett., 75, 4326-4329.
[4] Buhl, J., Sumpter, D. J. T., Couzin, I. D., Hale, J. J.,
Despland, E., Miller, E. R., and Simpson, S. J. (2006),
From disorder to order in marching locusts, Science, 312,
1402-1406.
[5] Nagy, M., ´Akos, Z., Biro, D., and Vicsek, T. (2010), Hi-
erarchical group dynamics in pigeon flocks, Nature, 464,
890-893.
[6] Croft, D. P., Arrowsmith, B. J., Bielby, Skinner, K.,
White, E., Couzin, I. D., Magurran, A. E., Ramnarine,
I., and Krause, J. (2003), Mechanisms underlying shoal
composition in the Trinidadian guppy, Poecilia reticu-
lata, Okios, 100, 429.
[7] Moussaıd, M., Perozo, N., Garnier, S., Helbing, D., and
Theraulaz, G. (2010), The Walking Behaviour of Pedes-
trian Social Groups and Its Impact on Crowd Dynamics,
PLoS ONE, 5, e10047.
[8] Vicsek, T., Czir´ok, A., Ben-Jacob, E., Cohen, I., and
Shochet, O. (1995), Novel Type of Phase Transition in
a System of Self-Driven Particles, Phys. Rev. Lett., 75,
1226-1229.
[9] Gr´egoire, G. and Chat´e, H. (2004), Onset of Collec-
tive and Cohesive Motion, Phys. Rev. Lett., 92, 025702-
025705.
[10] Chat´e, H., Ginelli, F., Gr´egoire, G., and Raynaud, F.
(2008), Collective motion of self-propelled particles in-
teracting without cohesion, Phys. Rev. E., 77, 046113-
046127.
[11] Ballerini, M., Cabibbo, N., Candelier, R., Cavagna, A.,
Cisbani, E., Giardina, I., Lecomte, V., Orlandi, A.,
15
Parisi, G., Procaccini, A., Viale, M., and Zdravkovic,
V. (2008), Interaction ruling animal collective behavior
depends on topological rather than metric distance: Ev-
idence from a field study, Proc. Natl. Acad. Sci., 105,
1232-1237.
[12] Gautrais, J., Ginelli, F., Fournier, R., Blanco, S., Soria,
M., Chat´e, H., Theraulaz, G. (2012), Deciphering Inter-
actions in Moving Animal Groups, PLoS Comput. Biol.,
8, e1002678.
[13] Ginelli, F. and Chate´e, H. (2010), Relevance of Metric-
Free Interactions in Flocking Phenomena, Phys. Rev.
Lett., 105, 168103-168106.
[14] Peshkov, A., Ngo, S., Bertin, E., Chat´e, H., and Ginelli,
F. (2012), Continuous Theory of Active Matter Sys-
tems with Metric-Free Interactions Phys. Rev. Lett., 109,
098101-098106.
[15] Heupe, C. and Aldana, M. (2008), New tools for char-
acterizing swarming systems: A comparison of minimal
models, Physica A, 387, 2809-2822.
[16] Jadbabaie, A., Lin, J., and Morse, A. S. (2003), Coor-
dination of groups of mobile autonomous agents using
nearest neighbor rules, Automatic Control, IEEE Trans-
actions, 48, 988-1001.
[17] Tanner, H. G., Jadbabaie, A., and Pappas, G. J. (2003),
Stable flocking of mobile agents, part I: fixed topology,
Proceedings of the 42nd IEEE Conference on Decision
and Control, 2, 2010-2015.
[18] Bode, N. W. F., Wood, A. J., and Franks, D. W. (2011),
The impact of social networks on animal collective mo-
tion, Anim. Behav., 82, 29-38.
[19] Chat´e, H., Ginelli, F., and Montagne, R. (2006), Sim-
ple Model for Active Nematics: Quasi-Long-Range Order
and Giant Fluctuations, Phys. Rev. Lett., 96, 180602-
180605.
[20] Dall, J. and Christensen, M. (2002), Random geometric
graphs, Phys. Rev. E, 66, 016121-016129.
[21] Herrmann, H.J., Hong, D.C., and Stanley, H.E. (1984),
Backbone and elastic backbone of percolation clusters
obtained by the new method of "burning", J. Phys. A,
17, L261-L266.
[22] Nagy M., Daruka I. and Vicsek T. (2007), New aspects of
the continuous phase transition in the scalar noise model
(SNM) of collective motion, Physica A, 373, 445-454.
[23] Kosterlitz, J. M. and Thouless, D. J. (1973), Ordering,
metastability and phase transitions in two-dimensional
systems, J. Phys. C, 60, 1181-1203.
[24] Toner, J., Tu, Y., Ramaswamy, S., (2005), Hydrodynam-
ics and phases of flocks, Annals of Physics, 318, 170244.
[25] Miyashita, S., Nishimori, H., Kuroda, A., and Suzuki,
M. (1978), Monte Carlo Simulation and Static and Dy-
namic Critical Behavior of the Plane Rotator Model,
Prog. Theor. Phys., 60, 1669-1685.
[26] Tobochnik, J. and Chester, G. V. (1979), Monte Carlo
study of the planar spin model, Phys. Rev. B, 20, 3761-
3769.
[27] Olfati-Saber, R. and Murray, R. M. (2007), Consen-
sus and Cooperation in Networked Multi-Agent Systems,
Proceedings of the IEEE, 95, 215-233.
[28] Floyd, S. (1994), The Synchronization of Periodic Rout-
ing Messages, IEE/ACM TRANSACTIONS ON NET-
WORKING, 2, 122-136.
|
1908.01054 | 1 | 1908 | 2019-08-02T21:02:38 | Cross-validation tests for cryo-EM maps using an independent particle set | [
"physics.bio-ph",
"q-bio.BM"
] | Cryo-electron microscopy is a revolutionary technique that can provide 3D density maps at near-atomic resolution. However, map validation is still an open issue in the field. Despite several efforts from the community, it is possible to overfit the reconstructions to noisy data. Here, inspired by modern statistics, we develop a novel methodology that uses a small independent particle set to validate the 3D maps. The main idea is to monitor how the map probability evolves over the control set during the refinement. The method is complementary to the gold-standard procedure, which generates two reconstructions at each iteration. We low-pass filter the two reconstructions for different frequency cutoffs, and we calculate the probability of each filtered map given the control set. For high-quality maps, the probability should increase as a function of the frequency cutoff and of the refinement iteration. We also compute the similarity between the probability distributions of the two reconstructions. As higher frequencies are added to the maps, more dissimilar are the distributions. We optimized the BioEM software package to perform these calculations, and tested the method on several systems, some which were overfitted. Our results show that our method is able to discriminate the overfitted sets from the non-overfitted ones. We conclude that having a control particle set, not used for the refinement, is essential for cross-validating cryo-EM maps. | physics.bio-ph | physics | Cross-validation tests for cryo-EM maps using
an independent particle set
Sebastian Ortiz1, Luka Stanisic2, Boris A Rodriguez3, Markus
Rampp2, Gerhard Hummer4,5, and Pilar Cossio1,4,*
1 Biophysics of Tropical Diseases, Max Planck Tandem Group, Universidad
de Antioquia UdeA, Calle 70 No. 52-21, Medell´ın, Colombia.
2Max Planck Computing and Data Facility, 85748 Garching, Germany.
3 Grupo de F´ısica At´omica y Molecular, Instituto de F´ısica, Facultad de
Ciencias Exactas y Naturales, Universidad de Antioquia UdeA, Calle 70
No. 52-21, Medell´ın, Colombia.
4Department of Theoretical Biophysics, Max Planck Institute of Biophysics,
5Institute of Biophysics, Goethe University, 60438 Frankfurt am Main,
60438 Frankfurt am Main, Germany.
*email: [email protected]; [email protected]
Germany.
arXiv:1908.01054v1 [physics.bio-ph] 2 Aug 2019
1
Abstract
Cryo-electron microscopy is a revolutionary technique that can
provide 3D density maps at near-atomic resolution. However, map
validation is still an open issue in the field. Despite several efforts
from the community, it is possible to overfit the reconstructions to
noisy data. Here, inspired by modern statistics, we develop a novel
methodology that uses a small independent particle set to validate
the 3D maps. The main idea is to monitor how the map probability
evolves over the control set during the refinement. The method is
complementary to the gold-standard procedure, which generates two
reconstructions at each iteration. We low-pass filter the two recon-
structions for different frequency cutoffs, and we calculate the prob-
ability of each filtered map given the control set. For high-quality
maps, the probability should increase as a function of the frequency
cutoff and of the refinement iteration. We also compute the similar-
ity between the probability distributions of the two reconstructions.
As higher frequencies are added to the maps, more dissimilar are the
distributions. We optimized the BioEM software package to perform
these calculations, and tested the method on several systems, some
which were overfitted. Our results show that our method is able to
discriminate the overfitted sets from the non-overfitted ones. We con-
clude that having a control particle set, not used for the refinement,
is essential for cross-validating cryo-EM maps.
2
Cryo-electron microscopy (cryo-EM) has revolutionized structural biol-
ogy by providing electron density maps of biomolecules that were difficult
to resolve with X-ray crystallography or nuclear magnetic resonance [1 -- 3].
The introduction of direct electron detection cameras [4, 5] and novel com-
putational algorithms [6, 7] has enabled the reconstruction of density maps
with near-atomic details. To date, thousands of maps, and their correspond-
ing atomic models, have been deposited in the electron microscopy [8] and
protein data banks [9] (EMDB and PDB, respectively).
Typically, cryo-EM maps are reconstructed using the gold-standard pro-
cedure [10, 11]. The particle images are divided into two sets, and two
independent reconstructions are generated. The reconstructions are refined
iteratively using maximum-likelihood [12, 13] or Bayesian techniques [14, 15].
At each iteration the Fourier Shell Correlation (FSC) [16, 17] between the
two independent reconstructions is computed. Fixed FSC threshold criteria
at 0.143 [18] or 0.5 [17] are used to determine the resolution of the reconstruc-
tions (i.e., the size of the smallest reliable detail). The refinement process
is halted when the resolution of the reconstructions stops improving. In the
end, the maps are masked and a final resolution is determined.
However, in spite of several efforts from the cryo-EM community, map
validation is still problematic.
In the recent Map Challenge it has been
shown that there is no absolute 'gold standard' [19]. The protocols are user-
dependent and there can be biases due to processing workflows. For instance,
in the FSC calculation, the resolution estimate is dependent on the radius of
the shell in Fourier space, and on the point symmetry of the molecule [20, 21].
The use of a fixed threshold for the FSC is restricted by the assumption that
the noise and the signal are orthogonal [20]. In addition, the mask can be
a source for overestimating the resolution [18, 22, 23]. Therefore, the best
criteria to estimate the map resolution are still debated in the cryo-EM com-
munity [20, 21]. These issues can lead to overfitted cryo-EM reconstructions.
For example, the reported values of the resolution in the model (from the
PDB) and in the map (from the EMDB) are different for about 30% of the
deposited data [24]. Moreover, it has been found that more than 70% of the
maps in the EMDB have moderate to low agreement with the model, mostly
because of the limited resolvable features of the maps [25]. In extreme cases,
maps can be reconstructed from pure-noise images [26, 27].
Therefore, methods that validate the quality of the maps and models are
fundamental for cryo-EM. Randomization of the phases beyond a frequency
threshold can give signatures of overfitting in the FSC curve [11, 28]. Bet-
3
ter resolution estimates are obtained with reference-free pipelines using the
1/2 bit non-fixed FSC threshold [20, 29]. The local resolution in a map
can be evaluated using the background noise of the reconstruction [30] or
by masking different regions with the FSC [23, 31]. Predictability of the
particle alignment provides quality indicators of the reconstruction [32, 33].
Moreover, several metrics that monitor cross-correlations in real or Fourier
space between the maps and models indicate the reliability of the resolu-
tion [24, 25, 34]. Recently, deep learning algorithms have been introduced
to automatically classify maps into high, medium, and low resolution [35].
However, all these methods have the limitation that they do not use the raw
data, which ultimately comes from the individual particles, but they only
use the maps or models that are product of processing and averaging. For
instance, in cryo-EM there is no cross-validation method, such as the R-free
in X-ray crystallography [36], which uses an independent control set from the
pure experimental data.
Inspired by modern statistical methods, we here propose an unbiased
strategy that validates cryo-EM reconstructions using a small control set of
particle images that are omitted from the refinement process. We do not fo-
cus on determining a specific value for the resolution but we develop a simple
cross-validation technique that monitors how the quality of the reconstruc-
tions evolves during the refinement procedure. We first calculate the BioEM
[37, 38] probability of the maps, given the control set, as a function of a
low-pass frequency cutoff of the reconstructions. High-quality maps should
increase in probability for higher frequency cutoffs and higher refinement
iterations. We then show that the similarity between the probability distri-
butions of the two reconstructions from the gold-standard procedure is an
additional quality indicator. Finally, we test the method on different systems
and asses its effectiveness to discriminate overfitted maps.
Results
Cross-validation protocol.
We propose a statistical framework for the cross-validation of cryo-EM recon-
structions. First, and foremost, the validation analysis is done over a small
control set of particle images not used in the refinement process. Analogously
to the R-free in X-ray crystallography [36], this independent set should give
4
an unbiased estimate of the quality of the reconstructions.
Figure 1: Cross-validation protocol for unbiased map validation in cryo-EM.
(left) Gold-standard refinement procedure in cryo-EM. Two particle sets are
used to generate two independent reconstructions. These reconstructions are
compared using the Fourier shell correlation (FSC). A fixed FSC threshold is
used to extract the resolution of the reconstructions. The process is iterated
until the resolution stops improving. (right) Novel cross-validation protocol
using a small control particle set. At each iteration of the refinement, the
reconstructions are low-pass filtered to different frequency cutoffs kc. The
BioEM probabilities [37, 38], over the independent control set, are calculated
as a function of kc. Two tests validate the quality of the reconstructions:
1) the cumulative log-posterior and 2) the statistical similarity between the
probability distributions (measured with a normalized Jensen-Shannon di-
vergence). The results from both tests should increase as a function of the
frequency cutoff. The maps represented correspond to the RAG1-RAG2
complex (see the Methods).
Fig. 1 shows the work-flow of the methodology. The refinement is done
following the gold-standard procedure (Fig. 1 -- left), where two reconstruc-
5
tions are generated at each iteration step. These two reconstructions are
validated using the control particle set (Fig. 1 -- right). At each iteration, the
two maps are low-pass filtered to different frequency cutoffs, kc (see the Meth-
ods). The BioEM [37] probability, Piω(kc), for each set i = 1, 2 is calculated
over the control set, ω ∈ Ω, with Nω particles. As a first cross-validation
tion of kc for each set i. This cumulative evidence should increase or remain
constant as higher frequencies are added to the maps. Failing this test is a
prime indicative that there is a problem in the refinement process.
test, we monitor the cumulative log-posterior,Pω ln(Piω(kc))/Nω, as a func-
The second cross-validation test consists on measuring the similarity be-
tween the probability distributions of the two reconstructions, also as a func-
tion of the frequency cutoff. For this purpose, we calculate a normalized
Jensen-Shannon divergence (NJSD) (see the Methods). The NJSD is a pos-
itive, symmetric and bound metric that measures how distinguishable are
the probability distributions from the reconstructions sets 1 and 2. We ex-
pect that as more frequencies are added to the reconstructions, more noise
is added, and the probability distributions are more uncorrelated (i.e., less
similar).
In the following, we describe in detail the two cross-validation tests.
Map evidence from the cumulative log-posterior.
We tested the methodology over several cryo-EM datasets:
the synaptic
RAG1-RAG2 complex (RAG1-RAG2) [39], the human HCN1 channel (HCN1)
[40], and the TRPV1 ion channel (TRPV1) [41]. These systems represent
a diverse set of biomolecular families, with membrane proteins and protein-
nucleicacids complexes. The reconstruction refinement was performed using
the gold-standard procedure in RELION [14]. The final resolution of these
systems ranges from approximately 3 to 6 A(see the Methods). To analyze
the impact of overfitting, we studied two additional systems: cryo-EM recon-
structions from the HIV-1 envelop trimer (HIV-ET) [42] and a set of synthetic
pure-noise images that act as a 'false' control set with the RAG1-RAG2 re-
constructions (see the Methods). This was motivated by the fact that some
reconstructions might have been generated from pure-noise particles, and
their resolution might have been over-estimated [26, 27, 43].
In Fig. 2, we examine the improvement of the maps by monitoring the
cumulative log-posterior relative to noise, Pω ln(Piω(kc))/Nω − ln(PNoise),
over the control set with Nω = 5000, as a function of kc for the reconstructions
6
Figure 2: The cumulative log-posterior relative to noise Pω ln(Piω)/Nω −
ln(PNoise), over the control set with Nω images, as a function of the frequency
cutoff for reconstructions from set i = 1 and 2 (solid and dashed lines, re-
spectively). The results are shown for different refinement iteration steps
with a gradient color code: the first iteration is maroon and the last itera-
tion is green. On the top row, we show the results for the standard cryo-EM
systems: HCN1, TRPV1 and RAG1-RAG2 for Nω = 5000. Systems that
exhibit signs of overfitting, i.e. a noise-particle control set with Nω = 1000
and HIV-ET with Nω = 5000, are shown in the bottom row, highlighted with
a red box.
from sets i = 1, 2. The results are shown for different refinement iterations
with a gradient color scheme (first iteration: maroon; last iteration: green).
These results measure how probable each filtered map is relative to PNoise (see
the Methods). For the RAG1-RAG2, HCN1 and TRPV1 systems, we find
an increase of the map evidence (given by the cumulative log-posterior) as a
function of the frequency cutoff. For very high frequencies, the cumulative
evidence plateaus. We only observe minor differences between the results
from set i = 1 and 2 (solid and dashed lines, respectively, in Fig. 2). This
is an indication of the similarity between the reconstructions generated from
the two sets.
Importantly, the results highlight the ability of the BioEM
7
posterior to correctly rank maps of different resolutions. The reconstructions
from the last iterations (i.e., the most refined) are the most probable. This
is in agreement with what one expects from the 3D-refinement algorithms
[7].
In contrast, for the HIV-ET and noise-particle set, we find a different be-
havior of the map evidence. We find that the cumulative log-posterior does
not increase as a function of the frequency cutoff but decreases or remains
constant. For the noise-particle set, the map evidence relative to PNoise is
small, and the differences between iterations are almost two orders of mag-
nitude smaller than for the non-overfitted sets. Moreover, for this case, as
the refinement iterations increase, the maps are slightly less probable. This
analysis monitors overfitting in cryo-EM: if the map evidence does not in-
crease as a function of the frequency cutoff or the refinement iteration, then
there are signs of overfitting in the data.
Similarity between the probability distributions.
As a second validation test, we compare the distributions of the posterior
probabilities generated by the reconstructions from sets i = 1, 2 over the
control set. In the Supplementary Information, we show an example of the
probability distributions for the HCN1 system for two frequency cutoffs at
a given iteration (Supplementary Fig. 1-top). We find that the probability
distributions, over the independent set, are quite similar for both reconstruc-
tions. However, there are small differences between them, and the higher-
frequency maps present larger fluctuations (Supplementary Fig. 1-bottom).
These differences can be quantified using a normalized Jensen-Shannon di-
vergence (NJSD; see the Methods).
In Fig. 3, we plot the NJSD as a function of the frequency cutoff kc.
Interestingly, for the RAG1-RAG2, HCN1 and TRPV1 systems, we observe
that as the filtered maps contain higher frequencies, the larger the value of
the NJSD. This implies that the probability distributions between maps with
higher frequencies are less similar, possibly because they are more uncorre-
lated due to the high-frequency noise. For these standard systems, we also
find that as the iteration increases the NJSD reaches at higher frequencies a
plateau value. This behavior can be fit with an inverse exponential function
−Ae−kc/γ + B (see below and solid lines in Fig. 3). On the contrary, for
the HIV-ET and noise-particle set, we find that the NJSD remains constant
or has random behavior, suggesting that distributions do not consistently
8
Figure 3: Normalized Jensen-Shannon divergence (NJSD) as a function of
the frequency cutoff. This metric calculates the similarity between the dis-
tributions of the BioEM probabilities computed for the two reconstructions
from sets 1 and 2. We use a gradient color code for the refinement iteration
steps: the first iteration is maroon and the last iteration is green. On the top
row, we show the results for the standard cryo-EM systems: HCN1, TRPV1
and RAG1-RAG2. For these systems, we fit the data points to an inverse
exponential function −Ae−kc/γ + B (solid lines). Systems that present signs
of overfitting, a noise-particle control set and HIV-ET, are shown in the bot-
tom row with dashed lines as a guide. The red box highlights the overfitted
systems. The number of images in the control sets are the same as for the
data in Fig. 2.
change when higher frequencies are added to the maps.
Cross-validation tests versus resolution.
We explored how the cross-validation results depend on the map resolution.
For the HCN1, TRPV1 and RAG1-RAG2 systems, we find that the NJSD
curves can be fitted to an inverse exponential function, −Ae−kc/γ + B (solid
Intuitively, the frequency γ indicates where the
lines shown in Fig. 3).
9
plateau of the NJSD is reached. In Fig. 3, we can qualitatively see that γ is
larger for higher refinement iterations. In Fig. 4, we plot the frequency γ as
a function of the inverse of the resolution (calculated using the FSC at the
threshold 0.143). Interestingly, we find that the frequency γ is highly corre-
lated to the inverse of the resolution with correlation coefficient r2 = 0.93,
0.91, and 0.85, for HCN1, TRPV1 and RAG1-RAG2, respectively. These
results show that even from a small independent control set, it is possible
to extract unbiased information about the map resolution. We note that for
the HIV-ET and noise-particle sets it is not possible to fit the NJSD data to
an inverse exponential function. Therefore, we can only estimate the corre-
lation between γ and the inverse of the resolution for the standard cryo-EM
systems.
Figure 4: Frequency γ versus the inverse of the resolution for the standard
cryo-EM systems: HCN1, TRPV1 and RAG1-RAG2. The NJSD curves
for these systems were fitted to an inverse exponential function −Ae−kc/γ +
B. We find large correlations between γ and the inverse of the resolution
(calculated using the 0.143 criteria). The correlation coefficients are r2 =
0.93, 0.91, and 0.85, for HCN1, TRPV1 and RAG1-RAG2, respectively. Solid
lines show the linear fits.
10
0
0.08
0.12
0.16
0.2
0.24
0.28
1/Resolution [1/Å]
HCN1 r2=0.93
TRPV1 r2=0.91
RAG1−RAG2 r2=0.85
0.1
0.08
0.06
0.04
0.02
Frequency γ [1/Å]
Convergence over a small cross-validation set.
We assessed how the results depend on the number of particles in the control
set. In Supplementary Fig. 2, we show an example of the cumulative log-
posterior and NJSD as a function of the number of images in the control set.
We find that after approximately 1000 particles these observables converge,
suggesting that only a small set is needed to perform the cross-validation
analysis. This is confirmed in Supplementary Fig. 3, where we plot the cu-
mulative log-posterior and NJSD as a function of the frequency cutoff for a
validation set of 1000 images. For the same set, in Supplementary Fig. 4, we
plot the frequency γ as a function of the inverse of the map resolution, show-
ing high correlations for the standard cryo-EM systems. These results are
very similar to those obtained for the cross-validation set with 5000 particles.
Discussion
In this work, we have developed a novel methodology for cross-validating
cryo-EM reconstructions. Importantly, the procedure is performed over an
independent particle set that is not used to generate the reconstructions.
Two cross-validation tests are proposed. The first consists of monitoring
the cumulative log-posterior of the maps as a function of a low-pass filter
frequency cutoff. The posterior should increase as a function of the frequency
cutoff and the refinement iteration. In the second test, we assess the similarity
between the probability distributions generated from the two reconstructions
from the gold-standard procedure. The distributions should become less
similar as higher frequencies are added to the reconstructions.
We performed the cross-validation tests over several systems: three stan-
dard cryo-EM reconstruction sets, and two datasets with noise particles that
mimic overfitting. The results show substantial differences. While for the
standard cryo-EM sets the results are as expected, the overfitted sets present
almost no increment (even sometimes decrease) of the cumulative posterior
or the NJSD. Thus, signatures of overfitting can be monitored with the pro-
posed cross-validation tests.
Our methodology is general and robust. The mathematical framework is
not only valid for the BioEM posterior but also for any posterior probability
that measures the likelihood of a 3D density given a particle set. The tests
converge over a small particle set, typically only 1000 particles. Moreover, the
11
methodology has the potential to be applicable for directly refining atomic
models (instead of 3D maps) using an independent control set.
Determining an unbiased estimate of the reconstruction resolution re-
mains an open issue. However, our procedure could shed light on how to
tackle this problem with a different perspective. For example, the resolution
could be defined as a multiple of γ that determines the frequency at which
the information between the probability distributions is governed by noise.
All-in-all, our work provides a novel way to monitor overfitting in cryo-
EM. We conclude that having a control particle set which is not used to
generate the reconstructions should become a standard for any cryo-EM ap-
plication.
Methods
Benchmark systems.
We used the following benchmarks that represent diverse biomolecular families and cryo-
EM systems:
The human hyperpolarization-activated cyclic nucleotide-gated channel (HCN1) is a
voltage-dependent ion channel, which was resolved to high resolution using cryo-EM [40].
The system was resolved in two conformational states, an apo state and a cAMP-bound
state, to ∼ 3.5 A using RELION 3D-refinement [14]. 55870 particles images belonging to
the apo state together with the defocus information of each particle are available in the
Electron Microscopy Public Image Archive (EMPIAR) [44] with code 10081. Their pixel
and image size are also available in that archive.
The recombination-activating genes RAG1-RAG2 form a complex (RAG1-RAG2) that
plays an essential role in the generation of antibodies and antigen-receptor genes in a
process called V(D)J recombination. Two main structures of the RAG1-RAG2 complex
can be distinguished during the V(D)J recombination, a synaptic paired complex and
the signal end complex (SEC). These states were resolved to 3.7 and 3.4 A, respectively,
using cryo-EM [39]. 81946 processed picked particles from the SEC state are deposited in
the EMPIAR data bank with code 10049. The defocus information is available for these
particles.
The mammalian transient receptor potential TRPV1 ion channel (TRPV1) is the
Its structure was determined to 3.4 A using cryo-EM [41]. A
receptor for capsaicin.
set of 35645 processed particles for this system are found in the EMPIAR data bank with
code 10005. The defocus information is also available for these particles.
The human immunodeficiency virus type 1 envelope glycoprotein trimer (HIV-ET) is
a membrane-fusing machine which mediates virus entry into host cells. The structure of
the apo HIV-1 envelope glycoprotein in the trimer-conformation was determined to 6 A
using the 0.5 FSC threshold with cryo-EM [42]. A set of 124478 particles used in the
12
refinement process is available in EMPIAR with code 10008. The defocus information is
also available for these particles.
For all of the above cases, a subset of 5000 particles was randomly selected to be
used as the cross-validation set. Specifically, these particles are not used in the refinement
processes.
Pure-noise images: we generated a set of synthetic 1000 pure-noise particles. Each
particle contains random intensities following a Gaussian distribution with zero mean and
unit variance (for details see the Supplementary Information). These images were used as
a "false" control set to assess the RAG1-RAG2 reconstructions.
3D refinement.
System
HCN1
RAG1-RAG2
TRPV1
HIV-ET
#Particles Symmetry #iterations Final resolution*
50870
79946
30645
119478
C4
C2
C4
C3
17
26
24
10
*using the 0.143 FSC threshold
4.2A
3.8A
5.3A
9.9A
Table 1: Summary of the results from the 3D-refinement using RELION [14]
for the cryo-EM systems.
The RELION [14] software was used to reconstruct the cryo-EM maps. For all systems,
we assume that the deposited particles correspond to the same state. Therefore, the
preprocessing steps of 2D or 3D classification are not performed. As the initial reference
map for the 3D refinement, we use the final map reported by the authors low-pass filtered to
60 A. This was done to minimize the risk of overfitting [11]. The 3D-refinement procedure
implements the gold-standard approach by splitting the data into two random halves (sets
i = 1, 2) and performing two independent reconstructions. We note that the number
of particles used for these reconstructions was slightly less than those of the original
works because the particles from the control set were taken out.
In all cases, we used
the RELION default parameters, and point-group symmetries reported by the authors.
Table 1 summarizes the results obtained from the 3D refinement. The resolutions are in
accordance with the reported ones, taking into account that the post-processing steps were
not performed, and that the control set of particles was excluded from the refinement.
Low-pass filter.
Consider a map m generated from an iteration of the 3D refinement. Let Fm(k) be its
3D-Fourier transform, where k is the reciprocal vector. We perform a low-pass filter on
the map, F kc
m (k), up to a frequency cutoff kc. The resulting filtered map is
13
m (k) =(cid:26)Fm(k)
F kc
0
k ≤ kc
otherwise.
(1)
We use the code lowpassmap fftw available from the Rubinstein lab webpage [45] to per-
form this calculation. We then convert the map into real space by applying the inverse
Fourier transform of F kc
m (k). The real-space filtered map is masked and then used as input
for the BioEM computation (see below).
BioEM posterior probabilities.
The BioEM method [37] uses a Bayesian framework to quantify the consistency between an
experimental image ω and a given map m (or model) by calculating a posterior probability
Pmω. BioEM takes into account the relevant physical parameters (Θ) for the image forma-
tion: center displacement, normalization, offset, noise, orientation and CTF parameters
(defocus, amplitude, and B-factor). Pmω is calculated by integrating-out all parameters
Pmω ∝Z L(ωΘ, m)p(Θ)p(m)dΘ ,
(2)
where p(m) and p(Θ) are the prior probabilities of the map and parameters, respectively,
and L(ωΘ, m) is the likelihood function. We considered the prior probabilities of maps
and parameters uniform over the integration intervals. In Eq. 2, the integrals over the
offset, noise and normalization are performed analytically [37], and that over the center
displacement is described in ref. [38]. The integral over the orientations and CTF defocus
is done using a double-round algorithm, which is described in the following subsection.
Similarly as in ref. [37], we define a noise model PNoise = (2πλ2e)−Npix/2 where Npix
is the number of pixels and λ is the image variance (by default λ = 1). PNoise is used as a
reference to compare the posterior probabilities.
BioEM algorithm.
To optimize the computations, we divided the BioEM posterior calculation into two rounds.
The objective of the first round is to obtain the best orientations for each particle. In this
round, an all-orientations to all-particles algorithm is performed [38]. As the BioEM input
map, we used the final reconstruction from the refinement with a broad mask and without
low-pass filtering. To sample the orientations, we used 36864 quaternions that sample
uniformly orientation space [46]. The particles were grouped into sets with similar exper-
imental defocus with 0.4µm range, and an independent orientation search was performed
for each group.
In this round, the best 10 orientations for each particle are obtained.
An example of the BioEM input for the first round is presented in the Supplementary
Information.
In the second round, a zoom around the best 10 orientations from the first round
and experimental defocus is performed for each low-pass filtered reconstruction from the
different refinement iterations. The zoom around each best orientation is done using 125
quaternions with approximately 0.01 grid spacing, resulting in 1250 zoomed-orientations
14
for each particle. This procedure is described in detail in ref.
[47]. The defocus of
each particle is fixed to its experimental value. We used 8 filtering-frequencies for each
reconstruction; these were distributed uniformly from 1/(pspNpix) to 1/(3ps) where ps is
the pixel size. All reconstructions were masked using the same broad mask as for round
1. An example of the BioEM input file for round 2 is presented in the Supplementary
Information.
BioEM code.
The BioEM code has been extended with several optimizations, which drastically increase
performance for the second round of calculations. Most importantly, the main data struc-
tures and algorithm were modified to allow for a parallel comparison of multiple orien-
tations to a single particle image. Initial reading of the input files has been parallelized,
and the overall memory consumption decreased. These code changes lead to more efficient
utilization of the computing resources, and hence to a faster calculation of posterior prob-
abilities, especially for the workloads specific to the second round. For more information,
we refer the reader to the BioEM user manual: https://readthedocs.org/projects/bioem/.
Normalized Jensen-Shannon divergence.
Measuring a distance among probability distributions is a common task in statistics. Most
distance measures include concepts from information theory, such as the Kullback-Leibler
divergence [48, 49] or the Shannon entropy [50]. In this work, we measure the statistical
similarity between the probability distributions from reconstructions from set 1 and set 2
calculated over the control set. We define a metric that is the Jensen-Shannon divergence
[49, 50] normalized by the individual Shannon entropies
(3)
NJSD = Pω[P1ω ln(P1ω/Mω) + P2ω ln(P2ω/Mω)]
2(Pω P1ω ln(P1ω)Pω P2ω ln(P2ω))1/2 ,
where P1ω and P2ω are the probabilities of the reconstructions from set 1 and 2, respec-
tively, over image ω, and Mω = (P1ω +P2ω)/2. For simplicity of notation, we have omitted
the dependency of the probabilities on the frequency cutoff kc. To calculate Eq. 3, we
normalize the posterior probabilities such that P1ω + P2ω = 1 for each image ω, frequency
cutoff and iteration.
In Eq. 3, the numerator measures the correlation between the probability distribu-
tions, and the Shannon entropies in the denominator play the role of a normalization
factor. Some important properties of the NJSD metric are that it is positive, symmetric
and its lower bound is 0 if and only if P1ω = P2ω for all particles ω.
Data availability
The BioEM code is available at https://github.com/bio-phys/BioEM. A tu-
torial to perform the cross-validation protocol is available at:
15
https://github.com/bio-phys/BioEM-tutorials.
Acknowledgements
The authors thank Dr. Alessandro Laio for insightful discussions, Dr. Jose
Maria Carazo for information about the overfitted systems, and Dr. Frank
Avila for proof reading. S.O. and P.C. were supported by Colciencias, Uni-
versity of Antioquia and Ruta N, Colombia. G.H., and P.C. acknowledge the
support of the Max Planck Society. Some computations were performed on a
local server with an NVDIA Titan X GPU. PC gratefully acknowledges the
support of NVIDIA Corporation for the donation of this GPU. Other com-
putations were performed at the Max Planck Computing and Data Facility.
Author contributions
G.H. and P.C. conceived the presented idea. S.O., G.H, and P.C. developed
the theory. S.O. and P.C. performed the computations. L.S. and M.R. co-
developed and optimized the code. All authors discussed the results and
contributed to the final manuscript.
References
[1] Kuhlbrandt, W. Cryo-EM enters a new era. eLife 3, e03678 (2014).
[2] Cheng, Y. Single-Particle Cryo-EM at Crystallographic Resolution. Cell
161, 450 -- 457 (2015).
[3] Murata, K. & Wolf, M. Cryo-electron microscopy for structural analysis
of dynamic biological macromolecules. Biochimica et Biophysica Acta
1862, 324 -- 334 (2018).
[4] Wu, S., Armache, J.-P. & Cheng, Y. Single-particle cryo-EM data acqui-
sition by using direct electron detection camera. Microscopy 65, 35 -- 41
(2016).
[5] McMullan, G., Faruqi, A. & Henderson, R. Direct Electron Detectors.
Methods in Enzymology 579, 1 -- 17 (2016).
16
[6] Kervrann, C., Sanchez Sorzano, C. O., Acton, S. T., Olivo-Marin, J.-C.
& Unser, M. A Guided Tour of Selected Image Processing and Analysis
Methods for Fluorescence and Electron Microscopy. IEEE Journal of
Selected Topics in Signal Processing 10, 6 -- 30 (2016).
[7] Cossio, P. & Hummer, G. Likelihood-based structural analysis of elec-
tron microscopy images. Current Opinion in Structural Biology 49,
162 -- 168 (2018).
[8] Lawson, C. L. et al. EMDataBank.org: unified data resource for Cry-
oEM. Nucleic Acids Res 39, D456 -- D464 (2011).
[9] Berman, H. et al. The Protein Data Bank. Nucleic Acids Res 28,
235 -- 242 (2000).
[10] Henderson, R. et al. Outcome of the First Electron Microscopy Valida-
tion Task Force Meeting. Structure 20, 205 -- 214 (2012).
[11] Scheres, S. H. W. & Chen, S. Prevention of overfitting in cryo-EM
structure determination. Nature methods 9, 853 (2012).
[12] Sorzano, C. et al. XMIPP: a new generation of an open-source image
processing package for electron microscopy. Journal of Structural Biol-
ogy 148, 194 -- 204 (2004).
[13] Tang, G. et al. EMAN2: An extensible image processing suite for elec-
tron microscopy. Journal of Structural Biology 157, 38 -- 46 (2007).
[14] Scheres, S. H. W. RELION: implementation of a Bayesian approach
to cryo-EM structure determination. Journal of structural biology 180,
519 -- 530 (2012).
[15] Punjani, A., Rubinstein, J. L., Fleet, D. J. & Brubaker, M. A.
cryoSPARC: algorithms for rapid unsupervised cryo-EM structure de-
termination. Nature Methods 14, 290 -- 296 (2017).
[16] Saxton, W. O. & Baumeister, W. The correlation averaging of a reg-
ularly arranged bacterial cell envelope protein. Journal of Microscopy
127, 127 -- 138 (1982).
[17] Harauz, G. & van Heel, M. Exact Filters for General Geometry Three
Dimensional Reconstruction. Optik 78, 6 -- 30 (1986).
17
[18] Rosenthal, P. B. & Rubinstein, J. L. Validating maps from single particle
electron cryomicroscopy. Current Opinion in Structural Biology 34,
135 -- 144 (2015).
[19] Heymann, J. B. et al. The first single particle analysis Map Challenge: A
summary of the assessments. Journal of structural biology 204, 291 -- 300
(2018).
[20] Van Heel, M. & Schatz, M. Fourier shell correlation threshold criteria.
Journal of structural biology 151, 250 -- 262 (2005).
[21] Sorzano, C. O. S. et al. A review of resolution measures and related
aspects in 3D Electron Microscopy. Progress in biophysics and molecular
biology 124, 1 -- 30 (2017).
[22] Penczek, P. A. Resolution Measures in Molecular Electron Microscopy.
Methods in Enzymology 482, 73 -- 100 (2010).
[23] Pintilie, G., Chen, D.-H., Haase-Pettingell, C., King, J. & Chiu, W.
Resolution and Probabilistic Models of Components in CryoEM Maps
of Mature P22 Bacteriophage. Biophysical Journal 110, 827 -- 839 (2016).
[24] Afonine, P. V. et al. New tools for the analysis and validation of cryo-EM
maps and atomic models. Acta Crystallographica Section D Structural
Biology 74, 814 -- 840 (2018).
[25] Neumann, P., Dickmanns, A. & Ficner, R. Validating Resolution Rev-
olution. Structure 26, 785 -- 795.e4 (2018).
[26] Henderson, R. Avoiding the pitfalls of single particle cryo-electron mi-
croscopy: Einstein from noise. Proceedings of the National Academy of
Sciences 110, 18037 -- 18041 (2013).
[27] Shatsky, M., Hall, R. J., Brenner, S. E. & Glaeser, R. M. A method
for the alignment of heterogeneous macromolecules from electron mi-
croscopy. Journal of structural biology 166, 67 -- 78 (2009).
[28] Chen, S. et al. High-resolution noise substitution to measure overfitting
and validate resolution in 3D structure determination by single particle
electron cryomicroscopy. Ultramicroscopy 135, 24 -- 35 (2013).
18
[29] Afanasyev, P. et al. Single-particle cryo-EM using alignment by classifi-
cation (ABC): the structure of Lumbricus terrestris haemoglobin. IUCrJ
4, 678 -- 694 (2017).
[30] Kucukelbir, A., Sigworth, F. J. & Tagare, H. D. Quantifying the local
resolution of cryo-EM density maps. Nature methods 11, 63 -- 5 (2014).
[31] Cardone, G., Heymann, J. B. & Steven, A. C. One number does not fit
all: mapping local variations in resolution in cryo-EM reconstructions.
Journal of structural biology 184, 226 -- 36 (2013).
[32] Vargas, J., Melero, R., G´omez-Blanco, J., Carazo, J. M. & Sorzano,
C. O. S. Quantitative analysis of 3D alignment quality: its impact on
soft-validation, particle pruning and homogeneity analysis. Scientific
Reports 7, 6307 (2017).
[33] Vargas, J., Ot´on, J., Marabini, R., Carazo, J. M. & Sorzano, C. O. S.
Particle alignment reliability in single particle electron cryomicroscopy:
a general approach. Scientific Reports 6, 21626 (2016).
[34] Brown, A. et al. Tools for macromolecular model building and re-
finement into electron cryo-microscopy reconstructions. Acta Crystal-
lographica Section D Biological Crystallography 71, 136 -- 153 (2015).
[35] Avramov, T. K. et al. Deep Learning for Validating and Estimating
Resolution of Cryo-Electron Microscopy Density Maps. Molecules 24
(2019).
[36] Brunger, A. T. Free R value: a novel statistical quantity for assessing
the accuracy of crystal structures. Nature 355, 472 -- 475 (1992).
[37] Cossio, P. & Hummer, G. Bayesian analysis of individual electron mi-
croscopy images: Towards structures of dynamic and heterogeneous
biomolecular assemblies. Journal of structural biology 184, 427 -- 437
(2013).
[38] Cossio, P. et al. BioEM: GPU-accelerated computing of Bayesian infer-
ence of electron microscopy images. Computer Physics Communications
210, 163 -- 171 (2017).
19
[39] Ru, H. & Others. Molecular Mechanism of V(D)J Recombination
from Synaptic RAG1-RAG2 Complex Structures. Cell 163, 1138 -- 1152
(2015).
[40] Lee, C.-H. & MacKinnon, R.
Structures of the human HCN1
hyperpolarization-activated channel. Cell 168, 111 -- 120 (2017).
[41] Liao Maofu, Cao Erhu, Julius David & Cheng Yifan. Structure of the
TRPV1 ion channel determined by electron cryo-microscopy. Nature
504, 107 (2013).
[42] Mao, Y. et al. Molecular architecture of the uncleaved HIV-1 envelope
glycoprotein trimer. Proceedings of the National Academy of Sciences
110, 12438 -- 12443 (2013).
[43] Subramaniam, S. Structure of trimeric HIV-1 envelope glycoproteins.
Proceedings of the National Academy of Sciences 110, E4172 -- E4174
(2013).
[44] Iudin, A., Korir, P. K., Salavert-Torres, J., Kleywegt, G. J. & Patward-
han, A. EMPIAR: a public archive for raw electron microscopy image
data. Nature Methods 13, 387 -- 388 (2016).
[45] Rubinstein
Lab.
URL
https://sites.google.com/site/
rubinsteingroup/home.
[46] Yershova, A., Jain, S., LaValle, S. M. & Mitchell, J. C. Generating
uniform incremental grids on SO(3) using the Hopf fibration. Int. J.
Robot. Res. 29, 801 -- 812 (2010).
[47] Cossio, P. et al. Bayesian inference of rotor ring stoichiometry from
electron microscopy images of archaeal ATP synthase. Microscopy 67,
266 -- 273 (2018).
[48] Kullback, S.
Information theory and statistics. (Dover Publications,
1968).
[49] Lin, J. Divergence measures based on the Shannon entropy.
IEEE
Transactions on Information Theory 37, 145 -- 151 (1991).
[50] Cover, T. M. & Thomas, J. A. Elements of information theory 2nd
edition (2006).
20
Supplementary Information: Cross-validation
tests for cryo-EM maps using an independent
particle set
Sebastian Ortiz1, Luka Stanisic2, Boris A Rodriguez3, Markus
Rampp2, Gerhard Hummer4,5, and Pilar Cossio1,4,*
1 Biophysics of Tropical Diseases, Max Planck Tandem Group, University
of Antioquia UdeA, Calle 70 No. 52-21, Medell´ın, Colombia.
2Max Planck Computing and Data Facility, 85748 Garching, Germany.
3 Grupo de F´ısica At´omica y Molecular, Instituto de F´ısica, Facultad de
Ciencias Exactas y Naturales, Universidad de Antioquia UdeA, Calle 70
No. 52-21, Medell´ın, Colombia.
4Department of Theoretical Biophysics, Max Planck Institute of Biophysics,
5Institute of Biophysics, Goethe University, 60438 Frankfurt am Main,
60438 Frankfurt am Main, Germany.
*email: [email protected]; [email protected]
Germany.
arXiv:1908.01054v1 [physics.bio-ph] 2 Aug 2019
1
Supplementary Figures
Supplementary Figure 1: Differences in the log-posterior distributions. (top)
Examples of the distributions of the log-posterior relative to noise over the
independent particle set. The distributions are calculated for the reconstruc-
tions from set 1 and set 2 at two cutoff frequencies kc = 0.05 and 0.25 A−1 for
the fifth iteration of refinement of the HCN1 system. The vertical lines are
the averages of the distributions. (bottom) Absolute value of the difference
between the probability distributions from set 1 and set 2 for kc = 0.05 and
0.25 A−1. The distributions calculated for the maps with higher frequencies
are less similar.
2
Supplementary Figure 2: Convergence of the observables. (top) The cumu-
lative log-posterior relative to noisePω ln(Piω)/Nω − ln(PNoise) for set i = 1
and 2 (solid and dashed lines, respectively), and (bottom) the normalized
Jensen-Shannon divergence as a function of the number of particles in the
control set. The results are shown for the TRPV1 system for iteration 12
and cutoff frequency kc = 0.21 A−1. The observables converge if more than
approximately 1000 particles are used.
3
Supplementary Figure 3: Cumulative log-posterior and NJSD for a control
set with 1000 particles. (top) The cumulative log-posterior relative to noise
and (bottom) the normalized Jensen-Shannon divergence as a function of
the frequency cutoff. We use a gradient color code for the refinement iteration
steps: the first iteration is maroon and the last iteration is green. The results
are shown for the standard cryo-EM systems: HCN1, TRPV1 and RAG1-
RAG2. The cumulative log-posterior is shown for the reconstructions from
set 1 as solid lines and set 2 as dashed lines. NSJD data is fit to an inverse
exponential function −Ae−kc/γ + B (solid lines; bottom).
4
Supplementary Figure 4: Frequency (γ) versus the inverse of the resolution
for a control set with 1000 particles. The results are shown for the standard
cryo-EM systems: HCN1, TRPV1 and RAG1-RAG2. The correlation coeffi-
cients are r2 = 0.95, 0.93, and 0.78, respectively. Solid lines show the linear
fits.
5
0
0.08
0.12
0.16
0.2
0.24
0.28
1/Resolution [1/Å]
Control set with 1000 particles
HCN1 r2=0.95
TRPV1 r2=0.93
RAG1−RAG2 r2=0.78
0.1
0.08
0.06
0.04
0.02
Frequency γ [1/Å]
Supplementary Text
BioEM input file examples
Round 1: Example of the BioEM input file for the TRPV1 system for round
1. The best orientations for each particle are obtained using the final map
from the refinement. The following input file is for a subset of particles that
have experimental defocus between 1.3 and 1.7 µm. The best 10 orientations
for each particle are selected.
PIXEL SIZE 1.22
NUMBER PIXELS 256
USE QUATERNIONS
CTF DEFOCUS 1.3 1.7 10
CTF B ENV 0 10 2
CTF AMPLITUDE 0.1 0.1 1
PRIOR DEFOCUS CENTER 1.5
SIGMA PRIOR DEFOCUS 0.8
SIGMA PRIOR B CTF 1
DISPLACE CENTER 30 1
WRITE PROB ANGLES 10
Round 2: Example of the BioEM input file for the TRPV1 system for
round 2. The input file is for a single particle that has an experimental de-
focus of 1.9 µm.
PIXEL SIZE 1.22
NUMBER PIXELS 256
USE QUATERNIONS
CTF DEFOCUS 1.9 1.9 1
CTF B ENV 0 10 2
CTF AMPLITUDE 0.1 0.1 1
PRIOR DEFOCUS CENTER 1.9
SIGMA PRIOR DEFOCUS 0.3
SIGMA PRIOR B CTF 1
DISPLACE CENTER 30 1
6
Pure-noise particles
We generated a set of 1000 synthetic pure-noise particles. Each particle has
an image size of 180 × 180 and a pixel size of 1.23 A. The particles contain
random intensities following a Gaussian distribution with zero mean and unit
variance. Because there is no experimental defocus, the BioEM probabilities
are computed by performing round 1 with defocus range between 0.5 and 4.5
µm and using 4608 quaternions uniformly distributed in orientation space.
This analysis was performed for each of the refined maps of the RAG1-RAG2
system.
7
|
1807.07332 | 1 | 1807 | 2018-07-19T10:30:39 | An Open-Source OpenSim Oculomotor Model for Kinematics and Dynamics Simulation | [
"physics.bio-ph",
"q-bio.TO"
] | Physics-based modeling and dynamic simulation of human eye movements has significant implications for improving our understanding of the oculomotor system and treating various visuomotor disorders. We introduce an open-source biomechanical model of the human eye that can be used for kinematics and dynamics analysis. This model is based on the passive pulley hypothesis, constructed based on the data reported in literature regarding physiological measurements of the human eye and made publicly available. The model is implemented in OpenSim, which is an open-source framework for modeling and simulation of musculoskeletal systems. The model incorporates an eye globe, orbital suspension tissues and six extraocular muscles. The excitation and activation patterns for a variety of targets can be calculated using the proposed closed-loop fixation controller that drives the model to perform saccadic movements in a forward dynamics manner. The controller minimizes the error between the desired saccadic trajectory and the predicted movement. Consequently, this model enables the investigation muscle activation patterns during static fixation and analyze the dynamics of eye movements. | physics.bio-ph | physics |
An Open-Source OpenSim Oculomotor Model for Kinematics
and Dynamics Simulation
Konstantinos Filip, Dimitar Stanev1, and Konstantinos Moustakas
July 20, 2018
1Electrical and Computer Engineering Department, University of Patras, Greece; Corresponding author:
[email protected]
Abstract
Physics-based modeling and dynamic simulation of human eye movements has significant implications
for improving our understanding of the oculomotor system and treating various visuomotor disorders.
We introduce an open-source biomechanical model of the human eye that can be used for kinematics
and dynamics analysis. This model is based on the passive pulley hypothesis, constructed based on the
data reported in literature regarding physiological measurements of the human eye and made publicly
available1. The model is implemented in OpenSim, which is an open-source framework for modeling
and simulation of musculoskeletal systems. The model incorporates an eye globe, orbital suspension
tissues and six extraocular muscles. The excitation and activation patterns for a variety of targets can be
calculated using the proposed closed-loop fixation controller that drives the model to perform saccadic
movements in a forward dynamics manner. The controller minimizes the error between the desired
saccadic trajectory and the predicted movement. Consequently, this model enables the investigation
muscle activation patterns during static fixation and analyze the dynamics of eye movements.
1SimTK project: https://simtk.org/projects/eye
Introduction
Rapid and accurate eye movements are of great importance for natural vision and thus studying human
eye movement can improve our understanding of the oculomotor system and treating various visuo-
motor disorders Lee and Terzopoulos (2006); Wei et al. (2010). Over the past decades, biomechanics
simulation has provided the means to analyze different human movements Delp et al. (2007). The same
principles can be used to analyze visual tasks by modeling the musculoskeletal properties of the ocu-
lomotor system. Consequently, this model can be used to investigate muscle activation patterns during
static fixation, analyze the dynamics of various eye movements, calculate metabolic costs and simu-
late eye disorders, such as different forms of strabismus Wong (2004). Furthermore, it can be easily
integrated with available full body models in order to analyze the relation between the vestibular and
oculomotor systems.
Eye movements are a generated from the coordinated activation of the six Extraocular Muscles
(EOMs). Clinical trials have provided a profound knowledge on the properties of the EOMs and their
line of action on the eye globe Robinson et al. (1969) and the resistive tension of the surrounding tis-
sues Collins et al. (1981); Iskander et al. (2018). Various computational models of the extraocular
muscles and orbital mechanics have been proposed, which provide insight for oculomotor biomechan-
ics, control of eye movement Bach-y Rita et al. (1971) and binocular misalignment. These models focus
on the realism of muscle behavior and they were based on the viscoelastic properties and physiological
data EOMs.
The first 3D biomechanical model was developed by Robinson (1964); Robinson and Fuchs (1969),
who simplified the formulation by only considering the elasticity of the EOMs ignoring their dynamics.
The model incorporates anatomically realistic muscle paths and empirical innervation-length-tension
relationships. In order to study the neural control of rapid saccadic movements, models using anatom-
ical and mechanical properties of EOMs have been developed by accounting for the nonlinear muscle
dynamics Thelen et al. (2003); Millard et al. (2013). Such models, having the advantage of supporting
dynamics simulation, are used in conjunction with brain level controllers Angelaki and Hess (2004);
James et al. (2018).
Methods
Eye Modeling
The eye model consists of the eye globe, three pairs of EOMs and the connective passive tissues. The size
of an emmetropic human adult eye is approximately 0.0242m (transverse, horizontal), 0.0237m (sagittal,
vertical), 0.022 − 0.0248m (axial, anteroposterior) with no significant difference between sexes and age
groups. In the transverse diameter, the eye may vary from 0.021m to 0.027m, thus it can be approximated
by a solid sphere of radius r = 0.012m. The weight of an average human eye is m = 0.0075kg and the
moment of inertia can be calculated assuming a spherical homogeneous and isotropic model I = 2/5mr2.
The eye has three rotational Degrees of Freedom (DoFs), namely incyclotosion-excyclotosion (x-axis),
adduction-abduction (y-axis) and supraduction-infraduction (z-axis).
Muscle Modeling
The six EOMs, including four rectus muscles and two oblique muscles, are controlled by the cranial
nerves so as to track a visual target and to stabilize the image of the object of interest on the retina.
The Lateral Rectus (LR) and Medial Rectus (MR) muscles form an agonist/antagonist pair that produce
horizontal eye movements. The Superior Rectus (SR) and Inferior Rectus (IR) muscles form the vertical
agonist/antagonist pair, which mainly controls vertical eye movement and also affects rotation about the
1
line of sight (secondary action) and the horizontal plane (tertiary action). The Superior Oblique (SO)
muscle passes through the cartilaginous trochlea attached to the orbital wall, which reflects the SO path
by 51◦. The Inferior Oblique (IO) muscle originates from the orbital wall anteroinferior to the globe
center and inserts on the sclera posterior to the globe equator. The primary actions of SO and IO cause
rotation of the globe around the visual axis, but also affect vertical (secondary action) and horizontal
(tertiary action) movements.
The model relies on the passive pulley assumption, which states that the pulleys have fixed to the
orbit pulley points Clark et al. (1977); Miller (2007). Table 1 shows the positions of the origin, insertion
and pulleys for the EOMs, defined in the local body coordinates of the eye globe. The data are based
on physiological measurements Iskander et al. (2018), with some minor modification so as to prevent
unrealistic muscle-surface penetration. Since no position was documented for the origin of the SO, a
point close to the origins of the rectus muscles was chosen to match the fiber length in the primary
position of the SO muscle.
Table 1: Muscle path points for the six EOMs defined in the local frame of the eye globe (dimensions
are given in meters).
Muscle
LR
MR
SR
IR
SO
IO
Origin
Oy
0.0006
0.0006
0.0036
-0.0024
0.0122
-0.0154
Ox
-0.034
-0.030
-0.0317
-0.0317
0.0082
0.0113
Oz
-0.013
-0.017
-0.016
-0.016
-0.0152
-0.0111
Px
-0.0102
-0.0053
-0.0092
-0.0042
-0.030834
-0.00718
Pulley
Py
0.0003
0.00014
0.012
-0.0128
0.001145
-0.0135
Insertion
Iy
0
0
0.0104
-0.0102
0.011
0
Ix
0.0065
0.0088
0.0076
0.00805
0.0044
-0.008
Iz
0.0101
-0.0096
0
0
0.0029
0.009
Pz
0.012
-0.0146
-0.002
-0.0042
-0.01644
0
The Millard muscle model Millard et al. (2013) has been adopted for the modeling of the EOMs,
permitting parameterization of the characteristic curves according to the experimental measured data.
The muscles were modeled using the rigid tendon assumption that ignores the elasticity of the tendon.
This means that the series element of the muscle model is not included (the tendon length lT is equal
to the tendon slack length lT
s ). EOMs are considered parallel-fibered muscles, so the pennation angle
is zero (α = 0). The values for the maximum isometric force f M
o and tendon
length lT are presented in Table 2.
o , optimal fiber length lM
Table 2: Millard muscle parameters for the EOMs.
Muscle Maximum Isometric
Force (N)
Optimal Fiber
Length (m)
Tendon Slack
Length (m)
Maximum Contraction
Velocity (m / s)
LR
MR
SR
IR
SO
IO
1.4710
1.5740
1.1768
1.4269
0.6031
0.5590
0.04898
0.04084
0.04487
0.04549
0.03956
0.04110
0.0084
0.0038
0.0054
0.0048
0.0265
0.0015
3.8483
4.6155
4.2009
4.1437
4.7648
4.5863
The active Force-Length (F-L) and Passive-Force-Length (F-PE) characteristic curves of the EOMs
differ significantly from those of a skeletal muscle. As shown in Figure 1, we can fine-tune the curve
parameters so as to fit the experimental data available for the LR muscle. The values for the active F-L
2
Table 3: Parameters of the active F-L charac-
teristic curve for the EOMs.
Table 4: Parameters of the F-PE characteristic
curve for the EOMs.
Parameter
min norm active fiber length
transition norm fiver length
max norm active fiver length
shallow ascending slope
minimum value
Value
0.55
0.7
1.8
2.4
0.0
Parameter
strain at zero force
strain at one norm force
Value
-0.18
0.4
and F-PE characteristic curves are summarized in Tables 3 and 4, respectively. We safely assume that
the parameters of the characteristic curves for the other EOMs are the same.
(a) active F-L curve
(b) F-L curve
Figure 1: The active F-L and F-PE curve definition for the Millard muscle model as implemented in
OpenSim.
EOMs have a higher fraction of fast twitch fibers and thus different Force-Velocity (F-V) behavior,
due to different structures compared to skeletal muscles. Despite that, the default Millard F-V curve
was used for the six EOMs, since the behavior of the selected muscle model depends heavily on the
maximum contraction velocity vmax. The maximum muscle contraction velocity is tuned so as to match
the peak velocity of saccadic eye movement ωmax = 15.7rad/s (900◦
/s). Following this definition, the
maximum muscle contraction velocity is given in optimal fiber length per seconds and it is thus different
for each EOMs, as their optimal fiber length is different (vmax = ωmaxr/lM
o ). Furthermore, since the optic
nerve is much shorter that the average muscle nerve, activation and deactivation delays (τa = τd = 5ms)
are smaller. Finally, two separate wrapping spheres for the rectus muscles and the oblique muscles were
created, to avoid abnormal changes on the F-L curve as the eye rotates.
Passive Connective Tissues
The passive connective tissues of the orbit apply a restoring force, which brings the eye back to the
central position when the net force from the EOMs is zero. These tissues include all non-muscular
suspensory tissues, such as Tenon's capsule, the optic nerve, the fat pad and the conjunctiva. The force-
displacement elasticity force can be represented as
3
ft = −kpq − kcq3 − kd q
(1)
where, ft is the passive tissue forces, kp = 2.225 · 10−3Nm/rad, kc = 34.53 · 10−3Nm/rad3 and kv =
2 · 10−3Nm/(rad/s) the physiological constants Robinson et al. (1969); Collins et al. (1981); Priamikov
et al. (2016) and q, q ∈ R3 the rotational coordinates and velocities of the mode. These forces are
modeled using OpenSim's expression based coordinate force.
Results
Fixation Controller
A fixation controller that calculates the EOMs excitations required to track a desired saccade was im-
plemented. The controller actuates the model in a closed-loop Forward Dynamics (FD) manner. The
parameters of the controller are the desired horizontal and vertical fixation angles, the saccade onset and
velocity, and the gains of Proportional Derivative (PD) tracking controller. A sigmoid function is used
for generating smooth saccade trajectories in the horizontal and vertical direction, while the torsional
component is maintained close to zero. More formally,
(cid:0) tanh(b(t − t0)) + 1(cid:1)
(cid:0)1 − tanh2(b(t − t0))(cid:1)
a
2
ab
2
θd(t) =
θd(t) =
(2)
where θd(t) and θd(t) represent the desired orientation and velocity at time t, a the magnitude of the
trajectory, b the slope and t0 a time shift constant. Provided a fixation goal θg, a desired saccade velocity
θg and a saccade onset tg, the parameters of the sigmoid function are defined as a = θg, b = 2θg/θg and
t0 = ts. The output of the PD tracking controller has the following form
u(t) = kp(θd(t) − θ(t)) + kd(θd(t) − θ(t))
(3)
where kp, kd are the tracking gains, and θ(t), θ(t) the simulated response of the model.
The sign and magnitude of u(t), representing the deviation from the fixation target for each axis of
rotation respectively, are used to calculate the muscle excitation levels, by assuming that each individual
muscle rotates the eye globe in a particular direction. Figure 2 presents an instance of the model during
simulation with the corresponding muscles activated. Figure 3 depicts the simulated coordinates, an-
gular velocities and estimated EOMs excitation levels that reproduce the desired saccade trajectory for
different model parameters. Finally, Figure 4 shows alternations in the saccadic movements both in the
horizontal and vertical direction so as to examine the activation and deactivation patterns of the EOMs.
Conclusion
A realistic oculomotor model representing the motility of a normal human eye was presented and made
publicly available. The parameters of the model were calibrated using available experimental measured
data. The model can be used for kinematics and dynamics analysis or as a tool for obtaining the muscle
activations that generate a desired saccade, using a closed-loop fixation controller in a FD manner. There
is of course space for further improvement, which will enhance the accuracy and the predictability of
the proposed computational model. In this study, we didn't attempt to model the muscle pulleys Kono
et al. (2002), where the position of the pulleys vary as a function of the model coordinates. Therefore,
the users should consider performing further validation of the eye model based on the requirements of
the targeted utility and the variables of interests.
4
Figure 2: Model with a fixation target at θH = −15◦, θV = 0◦ during simulation. Blue denotes low and
red high muscle activation levels.
(a) θH = −15◦, θV = 0◦, b = 600◦
/s and kv = 0.002Nm/(rad/s)
(b) θH = −15◦, θV = 0◦, b = 600◦
/s and kv = 0Nm/(rad/s)
Figure 3: Simulated saccade response for different model parameters. Left subplot represents the simu-
lated generalized coordinates, middle the generalized velocities and right the estimated EOMs excitation
levels.
5
00.20.40.60.81Time (s)-20-15-10-505Coordinates (deg)SimulatedReference00.20.40.60.81Time (s)-600-500-400-300-200-1000100Velocity (deg/s)SimulatedReference00.20.40.60.81Time (s)00.20.40.60.81Activation LevelLRMRSRIRSOIO00.20.40.60.81Time (s)-20-15-10-505Coordinates (deg)SimulatedReference00.20.40.60.81Time (s)-600-500-400-300-200-1000100Velocity (deg/s)SimulatedReference00.20.40.60.81Time (s)00.20.40.60.81Activation LevelLRMRSRIRSOIO Figure 4: Simulated saccade response performing abduction/adduction (0 − 2s) and supraduction/infra-
duction (2 − 4s) with kv = 0.002Nm/(rad/s).
6
012345Time (s)-15-10-5051015Coordinates (deg)Ad/AbductionSup/InfraductionIn/Extorsion012345Time (s)-300-200-1000100200300Velocity (deg/s)Ad/AbductionSup/InfraductionIn/Extorsion012345Time (s)00.20.40.60.81Activation LevelLRMRSRIRSOIO Bibliography
S.-h. Lee and D. Terzopoulos, "Heads Up ! Biomechanical Modeling and Neuromuscular Control of the
Neck," ACM Transactions on Graphics, vol. 1, no. 212, pp. 1188–1198, 2006.
S. L. Delp, F. C. Anderson, A. S. Arnold, P. L. Loan, A. Habib, C. T. John, E. Guendelman, and
D. G. Thelen, "OpenSim : Open-Source Software to Create and Analyze Dynamic Simulations of
Movement," IEEE Transactions on Biomedical Engineering, vol. 54, no. 11, pp. 1940–1950, 2007.
D. A. Robinson, D. M. O'meara, A. B. Scott, and C. C. Collins, "Mechanical components of human eye
movements," Journal of Applied Physiology, vol. 26, no. 5, pp. 548–553, 1969.
C. C. Collins, M. R. Carlson, a. B. Scott, and a. Jampolsky, "Extraocular muscle forces in normal human
subjects." Investigative Ophthalmology & Visual Science, vol. 20, no. 5, pp. 652–664, 1981.
J. Iskander, M. Hossny, S. Nahavandi, L. del Porto, and L. Porto, "An ocular biomechanic model for
dynamic simulation of different eye movements," Journal of Biomechanics, 2018.
D. A. Robinson, "The mechanics of human saccadic eye movement," The Journal of Physiology, pp.
245–264, 1964.
D. A. Robinson and A. F. Fuchs, "Eye movements evoked by stimulation of frontal eye fields." Journal
of Neurophysiology, vol. 32, no. 5, pp. 637–648, 1969.
D. G. Thelen, F. C. Anderson, and S. L. Delp, "Generating dynamic simulations of movement using
computed muscle control," Journal of Biomechanics, vol. 36, no. 3, pp. 321–328, mar 2003.
M. Millard, T. Uchida, A. Seth, and S. L. Delp, "Flexing computational muscle: modeling and simula-
tion of musculotendon dynamics," Journal of Biomechanical Engineering, vol. 135, no. 2, pp. 1–12,
mar 2013.
S. S. James, C. Papapavlou, A. Blenkinsop, A. J. Cope, S. R. Anderson, K. Moustakas, and K. N. Gur-
ney, "Integrating brain and biomechanical models-A new paradigm for understanding neuro-muscular
control," Frontiers in Neuroscience, vol. 12, no. FEB, 2018.
A. Priamikov, M. Fronius, B. Shi, and J. Triesch, "OpenEyeSim: A biomechanical model for simulation
of closed-loop visual perception," Journal of Vision, vol. 16, no. 15, p. 25, dec 2016.
R. Kono, R. A. Clark, and J. L. Demer, "Active pulleys: Magnetic resonance imaging of rectus muscle
paths in tertiary gazes," Investigative Ophthalmology and Visual Science, vol. 43, no. 7, pp. 2179–
2188, 2002.
R. A. Clark, J. M. Miller, and J. L. Demer, "Three-dimensional Location of Human Rectus Pulleys by
Path Inflections in Secondary Gaze Positions," Investigative Ophthalmology & Visual Science, vol. 41,
no. 12, pp. 3787–3797, nov 1977.
7
J. M. Miller, "Understanding and misunderstanding extraocular muscle pulleys," Journal of Vision,
vol. 7, no. 11, p. 10, 2007.
Q. Wei, S. Sueda, and D. K. Pai, "Physically-based modeling and simulation of extraocular muscles,"
Progress in Biophysics and Molecular Biology, vol. 103, no. 2-3, pp. 273–283, dec 2010.
P. Bach-y Rita, C. C. Collins, Smith-Kettlewell Institute of Visual Sciences., and University of the
Pacific. Department of Visual Sciences., The control of eye movements. Academic Press, 1971.
D. E. Angelaki and B. J. M. Hess, "Control of eye orientation : where does the brain ' s role end and the
muscle ' s begin ?" Neuroscience, vol. 19, 2004.
A. M. Wong, "Listing's law: clinical significance and implications for neural control," Survey of Oph-
thalmology, vol. 49, no. 6, pp. 563–575, nov 2004.
8
|
1909.06528 | 1 | 1909 | 2019-09-14T04:32:52 | Effect of geometry on the positioning of a single spot in reaction-diffusion systems | [
"physics.bio-ph"
] | We consider the formation of a single spot (localized solution) in reaction-diffusion (RD) equation on a curved manifold. Specifically, we study the direction (alignment) of the normal to interface between maxima and minima of concentration in the steady-state on a prolate and on an oblate ellipsoid. We further analyse the effect of shape asymmetry on l = 1 eigenmode of the sphere by assuming a small deformation from the spherical geometry. Our analysis shows that the eigenfunction corresponding to highest eigenvalue align along the symmetry axis for a prolate ellipsoid, and perpendicular to the symmetry axis for an oblate ellipsoid. Finally, we compare the direction of variation of the most unstable mode (eigenfunction with highest growth rate) in the system obtained by assuming a small deformation from the sphere and the alignment of interface normal obtain from the numerical simulations. | physics.bio-ph | physics | Effect of geometry on the positioning of a single spot in reaction-diffusion systems
Sankaran Nampoothiri∗
International Centre for Theoretical Sciences, Tata Institute of Fundamental Research,
Survey no 151, Shivakote, Hesaraghatta Hobli, Bengaluru 560089, India
(Dated: September 17, 2019)
We consider the formation of a single spot (localized solution) in reaction-diffusion (RD) equation
on a curved manifold. Specifically, we study the direction (alignment) of the normal to interface
between maxima and minima of concentration in the steady-state on a prolate and on an oblate
ellipsoid. We further analyse the effect of shape asymmetry on l = 1 eigenmode of the sphere by
assuming a small deformation from the spherical geometry. Our analysis shows that the eigenfunc-
tion corresponding to highest eigenvalue align along the symmetry axis for a prolate ellipsoid, and
perpendicular to the symmetry axis for an oblate ellipsoid. Finally, we compare the direction of
variation of the most unstable mode (eigenfunction with highest growth rate) in the system obtained
by assuming a small deformation from the sphere and the alignment of interface normal obtain from
the numerical simulations.
PACS numbers: 87.10.-e, 82.40.Ck, 82.20.-w, 02.40.-k
I.
INTRODUCTION
The emergence of pattern in space and time is an
ubiquitous phenomenon in nature. Hence its understand-
ing and modeling is of fundamental importance in many
fields. After the seminal work of Turing [1], reaction-
diffusion systems play a central role in the mathematical
modeling of spatial pattern formation [2 -- 5].The appli-
cability of RD equations ranges through many different
fields. Recently RD equations have also found its role
in understanding the spatial organization of molecules in
biological membranes
[6]. For example, oscillations of
Min protein system in E.coli cell are modelled using RD
equations [7 -- 9].
In most of the previous studies RD equations have
been proposed and analyzed on flat surfaces to under-
stand the formation of patterns. It is important to note
that the features of surface geometry has been under-
appreciated in all these studies. But, the role of surface
shape has been highlighted in many of the recent studies.
For example, the spatial patterning of protein molecules
are known to be sensitive to cell shape [10]. Geometry
can also play an important role in the formation of com-
plex patterns observed on animal surfaces [2]. Thus the
shape of the surface can be crucial in the formation of
wide variety of patterns.
Owing to the importance of understanding RD on
complex geometries, some of the previous studies have
analyzed the effect of geometry in RD systems [11 -- 20].
For example, the parr-marks formation on fish skins are
studied in the work [12] where shape of the skin is mod-
eled as growing elliptic cylinder which indicates the im-
portance of surface shape in understanding the patterns
observed in nature. Some of the rececnt works about the
study on nucleation of RD waves on curved surfaces [13],
∗ email:[email protected]
spiral waves on curved surfaces [21] and the effect of ge-
ometry on Min-protein dynamics [22, 23] again suggest
the importance of surface shape. These studies suggest
that the shape of the surface strongly influence the for-
mation of spatial patterns. It is also interesting to note
that the importance of the spectrum of the Laplace op-
erator on a curved surface in RD equation is highlighted
in the work [11].
The localized state (single spot) holds a significant po-
sition in RD like systems [24]. The localized solutions can
have important applications in morphogenesis and tech-
nologies. One interesting appearance of localized struc-
ture is in the RD models of blood coating [25]. Another
important application of localized solution can come in
variuos cellular processes. For example, single spot in
RD like equation can play a significant role in cell polar-
ization [26].
In the light of above studies it would be imperative
to analyze the interplay between geometry and the po-
sitioning of a single spot in RD. In the current work,
we have analyzed the role of shape asymmetry of the
surface on the positioning of a single spot in RD sys-
tems. To analyze the role of shape asymmetry in RD, we
have numerically evolved RD equations of Schnakenberg
model on both prolate and oblate ellipsoid. Specifically,
we have studied the positions of single spot on both cases
by varying one of the parameters in the system. Our
analysis suggests that the geometry can act as a cue for
the positioning of a single spot in RD systems.
In Sec.
The paper is organized as follows.
II, we
outline the general model of RD equation. Then we in-
troduce the Schnakenberg model of RD and its linear
stability analysis on a sphere. We then consider the for-
mation of a single spot on a sphere using Schnakenberg
model.
III, we analyse the formation of spot
on a prolate and on an oblate ellipsoid by solving the
RD equations numerically. In Sec. IV, we carry out a
perturbative analysis to understand the effect of defor-
mation on l = 1 mode by assuming a small deformation
In Sec.
9
1
0
2
p
e
S
4
1
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
8
2
5
6
0
.
9
0
9
1
:
v
i
X
r
a
from the spherical geometry. In Sec V, we compare the
conclusions from perturbative analysis and the numerical
observations. We summarize our results in Sec. VI.
which satisfies the linearized equation
∂ (δW )
∂t
= LδW,
II. MODEL
where
In general, the dynamics of RD system on a given
curved surface can be modelled by the following set of
equations
= F1(A, B) + DA (cid:52)LB A,
= F2(A, B) + DB (cid:52)LB B,
∂A
∂t
∂B
∂t
(1a)
(1b)
and
2
(5)
(6)
(7)
L = γC + D∇2,
(cid:18) 1 0
0 d
D =
(cid:19)
, C =
∂f
∂U
∂g
∂U
U0,V0
,
∂f
∂V
∂g
∂V
f (U, V ) = γ(a0 − U + U 2 V ),
g(U, V ) = γ(b0 − U 2 V ) .
The solution to the Eq. (5) can be written as
∞(cid:88)
l(cid:88)
m=−l
l=0
δW (θ, φ, t) =
C m
l eλ(l)tP m
l (cos θ)eimφ,
where the constants C m
l
conditions. The eigenvalues λ(l) satisfy
can be determined from initial
λ2 +λ[(
l(l + 1)
a2
)(1+d)−γ(fu +gv)]+h(l(l+1)) = 0. (8)
Hence the growth rate λ(l) corresponding to a partic-
ular mode l can be written as
−( l(l+1)
(cid:113)
( l(l+1)
a2
±
(1 + d) − γ(fu + gv))
a2
λ± =
(1 + d) − γ(fu + gv))2 − 4h(l(l + 1))
2
2
, (9)
where fu = ∂f
can be given as
∂U U0,V0, gv = ∂g
∂V U0,V0. The h(l(l + 1))
h(l(l + 1)) = d(l(l + 1)/a2)2 − γ(d fu + gv) l(l + 1)/a2+
γ2(fugv − fvgu).
Note that the set of modes having the postive growth
rate can lead to spatial inhomogeneity in concentration
in the steady-state.
It is also important to note that
the eigenvalues of Laplce-Beltrami operator is crucial in
determining the growth rate for RD systems.
First we illustrate the formation of single spot on a
sphere. Note that the parameters (a0, b0, d, γ) can con-
trol the growth rate of each modes. Hence, these pa-
rameters can play a significant role in determining the
number of spots.
In this case, we chose the values of
parameters in such a way that l = 1 mode is unstable
and all other modes are stable. We have then numeri-
cally solved Eq.(4a,4b) on the surface of a sphere using
FEniCS [28]. The single spot obtained on a sphere is
shown in the Fig. 1.
the
represent
where A, B are the concentrations of chemicals,
F1(A, B), F2(A, B)
reaction kinetics,
DA, DB are diffusion coefficients of the chemicals A and
B respectively, and (cid:52)LB is the Laplace-Beltrami opera-
tor on curved surface. The reaction terms in the equation
control the degredation and production of chemicals on
the surface, and in general, independent of the surface
shape. Hence the RD system senses the presence of ge-
ometry through the Laplace-Beltrami operator of that
surface.
A. Schnakenberg model on a sphere
to
the well-studied
restrict
[27] as an example for our studies.
Schnakenberg
We
model
The
model has the following advantages a) The model has
simplest kinetics b) The space of parameters where the
model can exhibit Turing instability is large and robust.
The reaction kinetics of the model is written as [27]
F1(A, B) = k1 − k2A + k3A2B,
F2(A, B) = k4 − k3A2B,
(2)
(3)
where the reaction kinetics F1 and F2 controls the pro-
duction and depletion of chemicals A and B. To proceed
further, we now write the non-dimensional version of the
equation as
∂U
∂τ
∂V
∂τ
= (cid:52)LBU + γf (U, V ),
= d (cid:52)LBV + γg(U, V ),
(4a)
(4b)
DA
a2 , γ = a2k2
, d = DB
DA
where τ = DAt
, U = A(k3/k2)1/2
and V = B(k3/k2)1/2 and the (cid:52)LB Laplace-Beltrami
operator on a sphere in scaled variable. The reaction
kinectics is given by f (U, V ) = (a0−U +U 2V ), g(U, V ) =
(b0 − U 2V ) where a0 = k1
)1/2 and b0 = k4
k2
The linear stability analysis about the homogeneous
(a0+b0)2 ) follows. A small
steady state (U0, V0) = (a0+b0,
variation in the homogeneous steady state can be denoted
as
)1/2.
( k3
k2
( k3
k2
k2
b0
(cid:18) δU − U0
δV − V0
(cid:19)
,
δW =
FIG. 1: Spot on a sphere. The parameters are γ = 8,
d = 10, a0 = 0.1, b0 = 0.9. The red represents the
maxima of the concentration of chemical U and blue
represents the minimum.
In the following section we analyse the positioning
of a single spot on a prolate and on an oblate ellipsoid
by keeping (a0, b0, d) same as in the case of a sphere and
vary the value of parameter γ. The specific importance
of parameter γ and its role in pattern selection is thor-
oughly discussed [2]. Note from Eq. (9) that the changes
in the value of γ can affect the growth rate of modes. The
growth rate of higher modes (modes with lower eigenval-
ues) can increase as a result of increasing the value of γ.
In other words, the maximum of growth rate can shift
towards higher modes as we increase the value of γ as
shown in the Fig. 2.
3
Beltrami operator is connected to the geometry of the
surface. Thus, by controlling the spectrum of Lapalce-
Beltrami operator, the geometry can influence the nature
of the steady-state solutions of RD. Hence, in this sec-
tion, we have explored the effect of deformation from the
spherical geometry on the localized solution (single spot)
in RD systems.
In order to understand the role of geometry, we have
numerically solved the Schnakenberg model on both pro-
late and oblate ellipsoid. Specifically, we have studied the
positioning of a single spot on both ellipsoids as we vary
the parameter γ. Our numerical simulation shows that
the parameter γ can play a significant role in determining
the positioning of a single spot .
A. Schnakenberg model on ellipsoid
In this section we consider the formation of single
spot on prolate and oblate ellipsoid using Schnakenberg
model. We now briefly mention the geometrical charac-
terestics of ellipsoids. The equation of an ellipsoid is
x2 + y2
a2 +
z2
b2 = 1,
(10)
where the case with a > b is called oblate ellipsoid, while
the case with a < b is the prolate ellipsoid. The ellipsoid
can be parametrized as
a sin θ cos φ
a sin θ sin φ
,
X(θ, φ) =
(11)
b cos θ
where θ and φ are the coordinates on the surface. Note
that on an ellipsoid both curvatures are θ dependent.
The Gaussian curvature of an ellipsoid is positive (see
appendix) where the curvature varies from b2/a4 (at θ =
0) to 1/b2 (at θ = π/2). Note that the Gauss curvature
is maximum at θ = 0 and minimum at θ = π/2 for a
prolate ellipsoid. In the case of an oblate ellipsoid, the
maximum of Gauss curvature occurs at θ = π/2 and
minimum occurs at θ = 0.
We have numerically solved the RD equation for
Schnakenberg model on both ellipsoids using FEniCS.
Initially we have considered a homogeneous distribution
of chemicals on both surfaces. The initial condition is
then provided by adding random perturbation to the ho-
mogeneous steady-state.
First, we have solved RD equations on prolate el-
lipsoid by considereing different values of γ. The spot
obtained in each cases are presnted in the Fig. 3 and
Fig. 4. To begin with, we have chosen γ = 8 and obtained
a single spot in the steady-state where the concentration
contains one maxima and one minima as shown in the
Fig. 3. In this case, the normal to interface is perpendic-
ular to the axis of symmetry. In other words, concentra-
tion is varying perpendicular to the axis of symmetry for
these values of γ. Here the maxima of concentration is
FIG. 2: The figure illustrates the effect of γ on growth
rate λ(l). Note that the curve shift towards the left
(right) as we decrease (increase) the γ.
III. ROLE OF GEOMETRY ON THE
POSITIONING OF A SINGLE SPOT
In the previous section we have seen that the role of
eigenvalues of Laplace-Beltrami operator in determining
the growth rate (λ±) of different modes. It is obvious that
spectrum (eigenvalues and eigenfunctions) of Laplace-
4
in the Fig 6. The interface normal is aligning perpen-
dicular to symmetry axis for both values of γ. Here the
concentration is high around the positions of maximum
Gauss curvature. In the case of an oblate ellipsoid also,
as similar to prolate ellipsoid, the concentration can vary
along and perpendicular to symmetry axis.
peaked near to the points of minimum Gauss curvature.
We have then considered γ = 5.6 and obtained the same
positioning of spot as in the previous case as shown in
the Fig. 3.
We have then carried out the simulation using γ = 5.5
and γ = 4 as shown in the Fig. 4. In both cases, the nor-
mal to interface is aligning along the symmetry axis. In
other words, the variation of concentration occurs along
the symmetry axis in this case. Here the maxima of
concentration is peaked near to the regions of maximum
Gauss curvature . Our simulation shows that concentra-
tion can vary along and perpendicular to symmetry axis
depending on the parameter γ.
FIG. 5: Left is the single spot on an oblate ellipsoid with
γ = 8 and right is the spot obtained for γ = 5.5. Note
that concentration is varying along the symmetry axis.
Here b = 1, a = 1.1.
FIG. 3: Left is the single spot on a prolate ellipsoid with
γ = 8 and right is the spot obtained for γ = 5.6. The
white line represents the interface. The other parameter
values are a0 = 0.1, b0 = 0.9, d = 10. We have chosen
semi-major axis b = 1.1 and semi-minor axis a = 1. Note
that concentration is varying perpendicular to symmetry
axis.
FIG. 6: Left is the single spot on an oblate ellipsoid with
γ = 4 and right is the spot obtained for γ = 5.3. Note
that concentration is varying perpendicular to symmetry
axis.
IV. EFFECT OF DEFORMATION: PROLATE
AND OBLATE ELLIPSOID
In this section, we have analyzed the role of shape
asymmetry on l = 1 mode by assuming a small deforma-
tion from the spherical geometry. The deformation can
remove the degeneracy of l = 1 mode. In other words,
the eigenvalues of different modes l = 1; m = 1,−1, 0
can be different due to shape asymmetry. This result into
different growth rates for these modes. Note from Fig. 2
that the modes with lower (higher) eigenvalues can have
larger growth rate for high (low) values of γ.
In the case of both prolate and oblate ellipsoid, we
have computed perturbatively the correction to eigen-
value of modes l = 1; m = 1,−1, 0 to understand which
mode has got highest/lowest eigenvalue and also the cor-
responding eigenfunction to zeroth order. To summa-
rize, the most unstable mode (eigenfunction with high-
FIG. 4: Left is the single spot on a prolate ellipsoid with
γ = 4 and right is the spot obtained for γ = 5.5. Note
that the interface normal is aligning along symmetry axis.
We have then solved the system of RD equations on an
oblate ellipsoid and the spot observed for different values
of γ is shown in the Fig. 5 and Fig. 6. In this case, the
paramater γ = 8 and γ = 5.5 result into a patterned state
where the normal to interface is aligning along symmetry
axis as shown in the Fig. 5. Note that the peak of the
maxima is formed around the points of minimum Gauss
curvature.
We have then considered parameter values γ = 5.3 and
γ = 4 in our simulation and obtained a spot as shown
est growth rate) and its direction of variation (along or
perpendicular to symmetry axis) on the surface can be
obtained from the perturbative analysis.
A.
prolate ellipsoid
First we consider the case of a prolate ellipsoid where
we calculate the correction to eigenvalue and eigenfunc-
tion corresponding to l = 1; m = 1,−1, 0 modes.
In
oder to obtain the eigenvalues and eigenfunctions we need
to compute the form of Laplace-Beltrami operator. The
form of Lapalce-Beltrami operator on any curved surface
is given by 1√
ggij∂j where gij is the inverse of the
metric gij and g is the determinant of the metric. This
can be given as
g ∂i
√
(cid:79)2 =
1
a2(cos2 θ + b2
a2 sin2 θ)
(1 − b2
a2 ) sin 2θ
2a2(cos2 θ + b2
a2 sin2 θ)2
∂2
∂θ2 +{
} ∂
∂θ
+
+
cot θ
a2(cos2 θ + b2
a2 sin2 θ
1
a2 sin2 θ
∂2
∂φ2 .
(12)
The above form of (cid:53)2 for a small deformation from the
spherical geometry can be written as
(cid:53)2 = (cid:53)2
sphere + A,
where A is given as
A = − 2
a3 sin2 θ
∂2
∂θ2 − 2
a3 sin 2θ
(13)
(14)
∂
∂θ
.
Next, for l = 1 case where we have a 3-fold degeracy,
we compute the 3×3 matrix of the perturbation A which
is given by
1 (cid:105)
1 A Y 1
(cid:104)Y 1
(cid:104)Y 0
1 A Y 1
1 (cid:105)
1 (cid:105) (cid:104)Y −1
A Y 1
1 (cid:105)
1 A Y 0
(cid:104)Y 1
(cid:104)Y 0
1 A Y 0
1 (cid:105)
1 (cid:105) (cid:104)Y −1
A Y 0
(cid:105)
1 A Y −1
(cid:105)
1 A Y −1
A Y −1
(cid:105)
1
1
1
1
1
(15)
We consider the following form of spherical harmonics for
calculating the above matrix elements.
(cid:104)Y 1
(cid:104)Y 0
(cid:104)Y −1
1
.
(cid:112)3/2π sin θeiφ,
(cid:112)3/π cos θ,
(cid:112)3/2π sin θe−iφ.
−1
2
1
2
1
2
Y 1
1 (θ, φ) =
Y 0
1 =
Y −1
1 =
(cid:104)Y 0
1 A Y 0
1 (cid:105) =
5
12
5a3 .
The perturbation matrix can now be explicitly written
as
4/5a3
0
0
.
0
12/5a3
0
0
0
4/5a3
(16)
We can now write the correction to eigenvalue for l = 1
mode due to a small deformation from spherical geometry
as
1 = −2/a2 + 4/5a3,
α1
1 = −2/a2 + 12/5a3,
α0
1 = −2/a2 + 4/5a3,
α−1
(17)
(18)
(19)
1
where α1,0,−1
are the new eigenvalues calculated upto
O(). Note from above expression that the eigenvalue α0
1
is higher compared to other eigenvalues. The three-fold
degenaracy of l = 1 mode is lifted to two-fold degenaracy
due to deformation. Now we need to calculate the eigen
functions corresponding to these eigenvalues.
The eigenvectors of the perturbation matrix are
given by
I(cid:105) =
1
, II(cid:105) =
0
, III(cid:105) =
0
0
1
0
0
.
0
1
(20)
Now one can write the eigenvectors to zeroth order cor-
responding to α1
1, α0
1
1 and α−1
as
ψ1 1(cid:105) = Y 1
1 ,
ψ1 2(cid:105) = Y 0
1 ,
ψ1 3(cid:105) = Y −1
1
(21)
(22)
(23)
.
1 is given by Y 0
Note that the eigenfunction correponding to highest eigen
value α0
1 which is varying along the axis
of symmetry for a prolate ellipsoid. The eigenfunction
(Y −1
1 ) corresponding to lowest eigenvalue is varying
perpendicular to symmetry axis.
, Y 1
1
Note that the growth rate of modes are different as a
result of removing the degeneracy in the eigenvalues due
to deformation. We can write the growth rate λ± corre-
sponding to the different modes Y 0
1 by following
the Eq. (9) as
1 and Y 1
2
−(−α0
(cid:112)(−α0
(cid:112)(−α1
1(1 + d) − γ(fu + gv))
λ0
1± =
1(1 + d) − γ(fu + gv))2 − 4h(α0
1)
,
2
1(1 + d) − γ(fu + gv))
λ1
1± =
1(1 + d) − γ(fu + gv))2 − 4h(α1
1)
−(−α1
2
,
±
2
(24)
(25)
Because of the orthogonality relation we need to calculate
only the integrals in the diagonal terms of the perturba-
tion matrix. These are evaluated to be
±
(cid:104)Y 1
(cid:104)Y −1
1 A Y 1
A Y −1
1 (cid:105) =
(cid:105) =
1
1
4
5a3 ,
4
5a3 ,
1 and Y 1
1± and λ1
where λ0
1± are the growth rates corresponding
to modes Y 0
1 . Note from Fig. 2 that the higher
modes (modes with lower eigenvalues) can have largest
growth rate for high values of γ. The lower modes can
become more unstable for low values of γ.
1 . Hence the growth rate λ1
In the case of a prolate ellipsoid, we have seen from
the Eq. (17, 18, 19) that the eigenvalue of mode Y 1
1 is
lower compared to Y 0
1+ of
the mode Y 1
1 with lower eigenvalue can be larger com-
pared to the growth rate of Y 0
1 for high values of γ. Note
that Y 1
1 is varying perpendicular to the axis of symmetry.
The mode Y 0
1 with higher eigenvalue can become more
unstable (λ0
1+ > λ1
1+) for lower values of γ. The mode
Y 0
1 is varying along the symmetry axis. The schematic
illustration of the effect of γ on the growth rate of modes
is shown in the Fig. 7.
1 = −2/a2 − 4/5a3.
α−1
6
(30)
1 or α1
. The eigenvalue of the mode Y 0
In the case of an oblate ellipsoid, the eigenfunction corre-
ponding to higher eigenvalue α1
1 is given by Y 1
1 or
Y −1
1 is lower compared
to eigenvalue of the mode Y 1
1 in the case of an oblate
ellipsoid.
1
In the case of an oblate ellipsoid, the growth rate of
Y 0
1 with lower eigenvalue can be higher compared to the
growth rate of Y 1
1 with
higher eigenvalue can become more unstable for low val-
ues of γ. The schematic illustration of the effect of γ on
the growth rate of modes is shown in the Fig 8.
1 for high γ values. The mode Y 1
FIG. 7: Schematic illustration of the effect of γ on
growth rate of the modes Y 0
1 . Note that the
1 and Y 1
mode Y 0
γ from γ2 to γ1. Y 0
1 can become more unstable as we decrease the
1 vary along symmetry axis and Y 1
1
vary perpendicular to symmetry axis.
B.
oblate ellipsoid
Here, similar to the analysis carried out in the case of
a prolate ellipsoid, we calculate the growth rate corre-
sponding to different modes. We can write the form of
(cid:53)2 for a small deformation from the spherical geometry
as
(cid:53)2 = (cid:53)2
sphere + A,
where A is given by
A =
2
a3 sin2 θ
∂2
∂θ2 +
2
a3 sin 2θ
∂
∂θ
.
(26)
(27)
We can now calculate the correction eigenvalues for l = 1
mode using the perturbation theory and can be given as
1 = −2/a2 − 4/5a3,
α−1
1 = −2/a2 − 12/5a3,
α0
(28)
(29)
FIG. 8: Schematic illustration of the effect of γ on
growth rate of the modes Y 0
1 . Note that the
1 and Y 1
γ from γ2 to γ1.
mode Y 1
1 can become more unstable as we decrease the
V. DIRECTION OF VARIATION OF MOST
UNSTABLE MODE AND NUMERICAL
OBSERVATIONS: A COMPARISON
In this section, we compare the direction of variation
of the most unstable mode obtained by the perturbative
analysis and the direction of interface normal observed
in our numerical simulations. The normal to interface
is aligning perpendicular to symmetry axis for γ = 8
and γ = 5.6 and parallel to symmetry axis for γ = 5.5
and γ = 4 in the case of a prolate ellipsoid. Thus our
numerical simulations shows that concentration can vary
perpendicular (parallel) to symmetry axis for high (low)
values of γ .
The perturbative analysis shows that the mode Y 1
1
(Y 0
1 ) can be more unstable for high (low) values of γ
as schematically shown in the Fig. 7 in the case of a
prolate ellipsoid. Hence, our analysis suggests that the
mode varying perpendicular (parallel) to symmetry axis
can be more unstable for the high (low) values of γ. Note
the similarity between the directions of variation of con-
centration observed in the simulations and the directions
of variation of most unstable mode obtained by the per-
turbative analysis for high (low) values of γ.
The interface normal is aligning along the symmetry
axis for γ = 8 and γ = 5.5 and perpendicular to symme-
try axis for γ = 5.3 and γ = 4 in the case of an oblate
ellipsoid. The perturbative analysis shows that the mode
varying parallel (pependicular) can be more unstable for
high (low) values of γ as schematically illustrated in the
Fig. 8. The analysis again indicates the similarity be-
tween the directions of interface normal observed in the
simulations and the directions of most unstable mode.
VI. SUMMARY
To sum up, we have studied the role of shape asym-
metry on the positioning of a single spot using Schnaken-
berg model on both prolate and oblate ellipsoid.
In
the case of a prolate ellipsoid, the normal to interface
is aligning perpendicular to symmetry axis for γ = 8 and
γ = 5.6. For values of γ = 5.5 and γ = 4, the interface
normal is aligning along the symmetry axis. In the case
of an oblate ellipsoid, for γ = 8 and γ = 5.5, the normal
to interface is aligning along the symmetry axis. The
normal to interface is aligning pependicular to symmetry
axis for γ = 5.3 and γ = 4. In both prolate and oblate
ellipsoid, the concentration can vary along and perpen-
dicular to the symmetry axis depending on the parameter
value γ.
We have analysed the effect of shape asymmetry on
l = 1 mode by assuming a small deformation from the
spherical geometry. Our analysis shows that the mode Y 1
1
(Y 0
1 ) can become more unstable for high (low) values of γ
for a prolate ellipsoid. In the case of an oblate ellipsoid,
the mode Y 0
1 (Y 1
1 ) can become more unstable for high
(low) values of γ.
We have then compared the direction of variation of
concentration obtained in the numerical simulation with
the direction of variation of the most unstable mode ob-
tained by the perturbative analysis. The concentration
can vary perpendicular (parallel) to symmetry axis as we
move along high (low) values of γ for a prolate ellipsoid.
The mode Y 1
1 ) can be more unstable for high (low)
values of γ as schematically illustrated in Fig. 7 in the
case of a prolate ellipsoid.
1 (Y 0
In the case of an oblate ellipsoid, concentration can
vary parallel (perpendicular) to symmetry axis as we
move along high (low) values of γ. In this case, the mode
1 (Y 1
Y 0
1 ) can become more unstable for high (low) val-
ues of γ as schematically shown in the Fig. 8. Thus,
in the case of both prolate and oblate ellipsoid, we have
observed a similarity between the directions of interface
normal observed in the simulations and the directions
of most unstable mode obtained from the perturbative
analysis.
The analysis presented in the work can be extended
to understand the role of geometry in any RD models
like BVAM model [14] and other models [2, 3]. Another
important application of this work can come in under-
standing the role of geometry in various cellular process.
7
Many important processes in cell and developmental bi-
ology are controlled by the spatial distribution of pro-
teins [29] where the effect of geometry can be signifi-
cant [30]. Note that RD like equations play a crucial
role [7 -- 9, 31, 32] in understanding these spatial distribu-
tion of proteins. Hence the analysis presented here can
be incorporated into above studies which may provide
useful insights about the role of cell geometry in various
cellular processes.
The localized solution (single spot) of RD systems can
play a significant role in determining the positioning of
plane of division in cell division processes [26]. The cur-
rent study hints that identifying the possible directions
of variation of eigenfunctions can lead to a model inde-
pendent (neglect the details of reaction kinetics) under-
standing of the positioning of a single spot on arbitarly
shaped surfaces. Thus the analysis presented here may
be useful to give insights about the possible planes of
division without knowing the details of reaction kinetics.
ACKNOWLEDGEMENTS
I thank Vijaykumar Krishnamurthy for many useful
discussions about the FEniCS. I also thank Vinayak Ja-
gadish for careful reading of the manuscript and valuable
suggestions.
Appendix A: Laplace operator and perturbative
calculations
Here we calculate the correction to eigenvalues and
eigenfunction of l = 1 mode due to deformation from
spherical geometry. First we consider the case of a pro-
late ellipsoid.
The ellipsoid can be parametrized as
x = a sin θ cos φ,
y = a sin θ sin φ,
z = b cos θ,
(A1)
(A2)
(A3)
where the range of φ and θ is given by 0 (cid:54) φ (cid:54) 2π and
0 (cid:54) θ (cid:54) π. The vector (cid:126)X is given by
(cid:126)X = a sin θ cos φi + a sin θ sin φj + b cos θk.
(A4)
The metric gθθ and gφφ is given by
b2
a2 sin2 θ),
gθθ = a2(cos2 θ +
gφφ = a2 sin2 θ,
gθφ = gφθ = 0.
The form of Lapalce-Beltrami operator is given by
1√
ggij∂j. This can be explicitily calculated as
√
g ∂i
(cid:79)2 =
1
a2(cos2 θ + b2
a2 sin2 θ)
∂2
∂θ2 +{
cot θ
a2(cos2 θ + b2
a2 sin2 θ
+
(1 − b2
a2 ) sin 2θ
2a2(cos2 θ + b2
a2 sin2 θ)2
} ∂
∂θ
+
1
a2 sin2 θ
∂2
∂φ2 .
Note that when b = a the form of the Laplace operator
reduces to sphere as expected. Now consider a small
deformation of the form b = a + . We can write the
term (cos2 θ + b2
a sin2 θ) by neglecting
O(2) term. Now we can write (cid:53)2 to O() as
∂2
a2 sin2 θ) as (1 + 2
1
(cid:53)2 =
1
a2
cot θ
a2
∂2
∂
∂θ2 +
∂θ
2
a3 sin2 θ
∂θ2 − 2
The above form of (cid:53)2 can be written as
∂2
a3 sin 2θ
∂
∂θ
.
+
a2 sin2 θ
∂φ2 −
(cid:53)2 = (cid:53)2
sphere + A,
(A5)
(A6)
∂
∂θ
.
The elements in the peturbation matrix A is given by
where A is given as
A = − 2
a3 sin2 θ
(cid:104)Y 1
1 A Y 1
(cid:90) π
∂2((cid:112)3/2π sin θ)
1 (cid:105) = 2π
0
−1
2
{ 2
a3 sin2 θ
∂2
a3 sin 2θ
∂θ2 − 2
(cid:112)3/2π sin θ
∂((cid:112)3/2π sin θ)
∂θ
1
(cid:104)Y −1
1 A Y 0
(cid:104)Y 0
∂θ2
+
2
a3 sin 2θ
4
5a3 .
A Y −1
=
1
4
5a3 .
1 (cid:105) = 2π
(cid:105) =
−1
2
(cid:90) π
∂2((cid:112)3/π cos θ)
∂((cid:112)3/π cos θ)
0
{ 2
a3 sin2 θ
(cid:112)3/2π cos θ
8
Similar procedure can be done for an oblate ellipsoid also.
Appendix B: Ellipsoid
The ellipsoid can represented as
x2 + y2
a2 +
z2
b2 = 1,
(B1)
where the case with a > b is called oblate spheroid, while
the case with a < b is prolate spheroid. The ellipsoid can
be parametrized as
a sin θ cos φ
a sin θ sin φ
,
b cos θ
X(θ, φ) =
(B2)
where θ and φ are the coordinates on the surface. Using
the above parametrization we read intrinsic and extrinsic
quantities related to curvature as
gθθ = a2(cos2 θ +
b2
a2 sin2 θ), gφφ = a2 sin2 θ, gθφ = gφθ = 0,
κθθ =
} sin θ dθ
b
(cos2 θ + b2
a2 sin2 θ)1/2
, κφφ =
b sin2 θ
(cos2 θ + b2
a2 sin2 θ)1/2
,
κθφ = κφθ = 0,
and then Gauss and mean curvature on an ellipsoid is
given by
b2
K =
a4(cos2 θ + b2
1 + (cos2 θ + b2
2a2(cos2 θ + b2
a2 sin2 θ)2
a2 sin2 θ)
a2 sin2 θ)3/2
,
.
+
2
a3 sin 2θ
∂θ
∂θ2
} sin θ dθ =
H = b
12
5a3 .
[1] A. M. Turing, Philosophical Transactions of the Royal
Society of London B: Biological Sciences 237, 37 (1952).
[2] J. D. Murray, Mathematical Biology. II Spatial Models
and Biomedical Applications {Interdisciplinary Applied
Mathematics V. 18} (Springer-Verlag New York Incor-
porated, 2001).
[7] M. Howard, A. D. Rutenberg, and S. de Vet, Physical
review letters 87, 278102 (2001).
[8] G. Meacci and K. Kruse, Physical biology 2, 89 (2005).
[9] H. Meinhardt and P. A. de Boer, Proceedings of the Na-
tional Academy of Sciences 98, 14202 (2001).
[10] D. Thalmeier, J. Halatek, and E. Frey, Proceedings of
[3] A. Koch and H. Meinhardt, Reviews of modern physics
the National Academy of Sciences 113, 548 (2016).
66, 1481 (1994).
[11] J. R. Frank, J. Guven, M. Kardar, and H. Shackleton,
[4] F. Borgogno, P. D'Odorico, F. Laio, and L. Ridolfi, Re-
arXiv preprint arXiv:1710.00103 (2017).
views of Geophysics 47 (2009).
[5] S. Getzin, H. Yizhaq, B. Bell, T. E. Erickson, A. C. Pos-
tle, I. Katra, O. Tzuk, Y. R. Zelnik, K. Wiegand, T. Wie-
gand, et al., Proceedings of the National Academy of Sci-
ences 113, 3551 (2016).
[12] C. Venkataraman, T. Sekimura, E. A. Gaffney, P. K.
and A. Madzvamuse, Physical Review E 84,
Maini,
041923 (2011).
[13] F. Kneer, E. Scholl, and M. A. Dahlem, New Journal of
Physics 16, 053010 (2014).
[6] M. Loose, K. Kruse, and P. Schwille, Annual review of
[14] C. Varea, J. Aragon, and R. Barrio, Physical Review E
biophysics 40, 315 (2011).
60, 4588 (1999).
9
[15] V. Zykov, A. Mikhailov, and S. Muller, Physical review
letters 78, 3398 (1997).
[16] J. Gomatam and F. Amdjadi, Physical Review E 56,
3913 (1997).
[24] V. K. Vanag and I. R. Epstein, Chaos: An Interdisci-
plinary Journal of Nonlinear Science 17, 037110 (2007).
[25] E. Lobanova, E. Shnol, and F. Ataullakhanov, Physical
review E 70, 032903 (2004).
[17] S. Liaw, C.-C. Yang, R. Liu, and J. Hong, Physical Re-
[26] R. Diegmiller, H. Montanelli, C. B. Muratov, and S. Y.
view E 64, 041909 (2001).
[18] R. G. Plaza, F. Sanchez-Garduno, P. Padilla, R. A. Bar-
rio, and P. K. Maini, Journal of Dynamics and Differen-
tial Equations 16, 1093 (2004).
[19] A. L. Krause, M. A. Ellis, and R. A. Van Gorder, Bulletin
of mathematical biology 81, 759 (2019).
[20] S. Nampoothiri
and A. Medhi,
arXiv preprint
arXiv:1705.02119 (2017).
[21] H. Dierckx, E. Brisard, H. Verschelde, and A. V. Pan-
filov, Physical Review E 88, 012908 (2013).
[22] J. C. Walsh, C. N. Angstmann, A. V. McGann, B. I.
Henry, I. G. Duggin, and P. M. Curmi, AIMS Biophysics
(2016).
[23] L. Wettmann, M. Bonny, and K. Kruse, PloS one 13,
e0203050 (2018).
Shvartsman, Biophysical journal 115, 26 (2018).
[27] J. Schnakenberg, Journal of theoretical biology 81, 389
(1979).
[28] H. P. Langtangen, A. Logg, and A. Tveito, Solving PDEs
in Python: The FEniCS Tutorial I (Springer Interna-
tional Publishing, 2016).
[29] J. Halatek, F. Brauns, and E. Frey, Philosophical Trans-
actions of the Royal Society B: Biological Sciences 373,
20170107 (2018).
[30] R. Parthasarathy and J. T. Groves, Soft Matter 3, 24
(2006).
[31] A. B. Goryachev and A. V. Pokhilko, FEBS letters 582,
1437 (2008).
[32] M. Otsuji, S. Ishihara, K. Kaibuchi, A. Mochizuki,
S. Kuroda, et al., PLoS computational biology 3, e108
(2007).
|
1307.3530 | 1 | 1307 | 2013-06-27T20:54:01 | Vibrations, Quanta and Biology | [
"physics.bio-ph",
"physics.chem-ph",
"quant-ph"
] | Quantum biology is an emerging field of research that concerns itself with the experimental and theoretical exploration of non-trivial quantum phenomena in biological systems. In this tutorial overview we aim to bring out fundamental assumptions and questions in the field, identify basic design principles and develop a key underlying theme -- the dynamics of quantum dynamical networks in the presence of an environment and the fruitful interplay that the two may enter. At the hand of three biological phenomena whose understanding is held to require quantum mechanical processes, namely excitation and charge transfer in photosynthetic complexes, magneto-reception in birds and the olfactory sense, we demonstrate that this underlying theme encompasses them all, thus suggesting its wider relevance as an archetypical framework for quantum biology. | physics.bio-ph | physics | Vibrations, Quanta and Biology
S. F. Huelga and M. B. Plenioa,b ∗
aInstitut fur Theoretische Physik, Universitat Ulm, Albert-Einstein-Allee 11, 89073 Ulm, Germany
bCenter for Integrated Quantum Science and Technologies, Albert-Einstein-Allee 11, 89073 Ulm, Germany
Quantum biology is an emerging field of research that concerns itself with the experimental and
theoretical exploration of non-trivial quantum phenomena in biological systems.
In this tutorial
overview we aim to bring out fundamental assumptions and questions in the field, identify basic
design principles and develop a key underlying theme – the dynamics of quantum dynamical
networks in the presence of an environment and the fruitful interplay that the two may enter. At
the hand of three biological phenomena whose understanding is held to require quantum mechanical
processes, namely excitation and charge transfer in photosynthetic complexes, magneto-reception
in birds and the olfactory sense, we demonstrate that this underlying theme encompasses them all,
thus suggesting its wider relevance as an archetypical framework for quantum biology.
Keywords: Biology, Quantum dynamics, Environments, Vibrations, Excitons, Electrons, Protons,
Transport, Coherence
INTRODUCTION
Following early speculations concerning the potential
role of quantum physics in biology [1], recent progress in
science and technology has led to the rapid emergence
of a new direction of research whose aim is the experi-
mental and theoretical exploration of quantum effects in
biology (see e.g. [2, 3]) which are taking place on length
and timescales that allow quantum dynamics and envi-
ronmental fluctuations to enter an intricate and fruitful
interplay.
Before we enter into more detailed discussions, let us
first make some points as to why the existence of quan-
tum effects in biology may be considered surprising and
why there is rapidly growing excitement for developing
what is being called quantum biology.
To begin with, biological systems are, almost by def-
inition, open systems, as they need to be continuously
supplied with energy to maintain the out of equilibrium
state that life represents. Open systems, however, espe-
cially warm, wet and noisy biological systems, are sub-
ject to environmental fluctuations that are usually ex-
pected to result in fast decoherence and, as a result, the
suppression of well controlled quantum dynamics. Thus
quantum phenomena may at first sight seem unlikely to
play a significant role in biology. There are arguments
however to counter this pessimistic view. At the level
of molecular complexes and proteins, processes that are
of fundamental importance for biological function can be
very fast (taking place within picoseconds) and well lo-
calised (extending across a few nanometers, the size of
proteins) and may therefore exhibit quantum phenom-
ena before the environment has had an opportunity to
destroy them. Furthermore, early work in quantum in-
formation science, for example, has shown that thermal
noise in stationary non-equilibrium systems may in fact
support the existence of quantum coherence and entan-
glement [4, 5]. Hence the possible existence of significant
quantum dynamics is not only a question of sufficiently
short length and time scales but may also depend on a
constructive interplay between a quantum dynamical sys-
tem and its environment such that quantum correlations
are not simply washed out or suppressed but may in fact
be enhanced or regenerated by the interaction with the
environment.
These arguments suggest that quantum effects in bi-
ology are possible at the right length- and time scales.
Indeed, quantum phenomena such as electron tunneling
[6, 7] have been observed in biological systems and there
is some evidence for proton tunneling in enzymes (see e.g.
[8]). As such, tunneling phenomena are not intimately
related with biology. Electron tunneling for example is
a well-known and important phenomenon in solid state
physics. The question thus remains whether, on the one
hand, biological systems will exhibit more complex quan-
tum dynamical phenomena that may either involve sev-
eral interacting particles or multiple interacting compo-
nents of a network or, on the other hand, whether the
specifics of the biological systems and their environments
will play a crucial role in allowing or supporting certain
quantum dynamical phenomena in biology. Only then
would we call these ”non-trivial” quantum effects in bio-
logical systems.
We do expect however that in the course of evolution
Nature will have learnt to make use of quantum phe-
nomena only if these enable or make more efficient a
useful biological function that provides an evolutionary
advantage.
It is indeed well established from a quan-
tum information perspective that pure quantum dynam-
ics of multi-component systems can provide qualitative
performance improvements over classical systems for ex-
ample where transport is concerned [9]. This provides
3
1
0
2
n
u
J
7
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
0
3
5
3
.
7
0
3
1
:
v
i
X
r
a
further support for the expectation that nature has de-
veloped non-trivial quantum phenomena in the dynam-
ics of biological systems, possibly supported by their en-
vironment whose presence and influence is unavoidable.
These quantum phenomena are not merely a by-product
of the underlying quantum nature of chemical bonds but
are actually exploited by biological systems to enhance
performance and achieve novel functionalities. The clear
demonstration that Nature makes use of quantum effects
would bring about the necessity for a significant change of
thinking for biologists as they would be required to grasp
quantum concepts in order to understand some funda-
mental biological processes. The very same fact would
however also present the opportunity to learn from biol-
ogy by unraveling the mechanisms by which quantum
dynamics and its interplay with environments lead to
enhanced performance. The resulting design principles
have the potential to lead to the development of new ap-
plications at the bio-nano scale [10].
FIG. 1. Biological systems are organized in hierarchical struc-
tures. The continuous refinement of experimental tools per-
mits the investigation of ever finer detail giving rise to the
discovery of novel phenomena. At a certain level we expect
quantum physical properties to become relevant. Whether na-
ture has evolved to enhance them to take benefit from them
(quantum enhanced efficiency) or to suppress them to avoid
their detrimental effects (quantum noise) represents one of
interesting open question at the heart of quantum biology.
(Figure courtesy of Alipasha Vaziri.)
Last but not least, the recent acceleration of the de-
velopment of quantum biology also has a very practical,
technological reason. Indeed, it is worthwhile noting that
quantum biology is not a new field but does go back a
long time, perhaps to Jordan’s book ”Die Physik und
das Geheimnis des Lebens” [1] in which he posed the
question ”Sind die Gesetze der Atomphysik und Quan-
tenphysik fur die Lebensvorgange von wesentlicher Be-
deutung?” (Are the laws of atomic and quantum physics
2
of essential importance for life?) and coined the term
Quanten-Biologie (quantum biology). So, why this re-
newed and rapidly growing interest in the field? To un-
derstand this, it is helpful to remember the development
of quantum information science, whose theoretical foun-
dations had been studied for some time by the 1990’s.
This research had revealed that concepts of quantum in-
formation have the potential to provide real performance
advantages over classical systems. Crucially, however,
it also became clear that quantum technologies had ad-
vanced sufficiently to turn these theoretical ideas into
reality. This led to the emergence of a rapid develop-
ment of both theory and experiment which is continuing
to this day.
In recent years, quantum biology too has been bene-
fitting considerably from the refinement of experimental
tools that are beginning to provide direct access to the
observation of quantum dynamics in biological systems
[11–14] thanks to their increasing sensitivity to quantum
phenomena at short length and time scales (see fig. 1
for a suggestive illustration of this point). These newly
found technological capabilities have helped to elevate
the study of quantum biology from a largely theoretical
endeavor to a field in which theoretical questions, con-
cepts and hypotheses may be tested experimentally and
thus be subjected to experimental verification or falsifi-
cation.
Indeed we should stress here that experiments
are essential (even more so than in the well-controlled
present day systems of quantum information science) to
verify theoretical models because biological systems un-
der investigation have a complexity and structural variety
that prevents us from knowing and controlling all their
aspects in detail. Results obtained by these refined ex-
perimental techniques lead to new theoretical challenges
and thus stimulate the development of novel theoreti-
cal approaches.
It is this mutually beneficial interplay
between experiment and theory that promises an accel-
erated development of the field.
In this tutorial overview we will explore the type of
phenomena that currently define the field of quantum bi-
ology and aim to bring out what are key questions at
the present stage of development. We will then focus our
discussions on a principle that is rapidly gaining recog-
nition as being of central importance in this field – the
crucial role of the interplay between quantum dynamics
of multi-site systems, a.k.a. networks, on the one hand
and of complex, structured environments on the other.
We will illustrate the generality of this principle by using
it to analyse and interpret three key examples of quantum
effects in biological systems, environment assisted exci-
tation energy transport in photosynthesis [11, 15, 16],
magneto-reception of birds [17–19] and the mechanism
underlying olfaction [20, 21], for which there is some
compelling theoretical and/or experimental evidence that
quantum phenomena are essential for their understand-
ing. These three examples will hopefully stimulate the
Typical spatial scale [m]10 -1110 -1010 -910 -810 -610 -510 -410 -210 -210 -310 -410 -610 -810 -1210 -1410 -1Typical time scale [s]Typical spatial scale [m]10 -1110 -1010 -910 -810 -610 -510 -410 -210 -210 -310 -410 -610 -810 -1210 -1410 -1Typical time scale [s]Function ? Tools Quantum Classical discovery of many more biological phenomena for which
non-trivial quantum effects are of fundamental impor-
tance and thus come to be seen as the seeds from which
a rich phenomenology of quantum effects in biology may
grow.
FIG. 2. Cartoon illustrating the broad line of argument in
this article: Following the identification of general questions,
we will argue that biological systems use protein structure to
adjust the properties of transport or sensory networks and, at
the same time, those of the environment of these networks.
Mutual tuning of these structures through evolutionary adap-
tation may achieve optimal performance which in turn can be
explained from generalizable design principles. These design
principles can lead to the formulation of novel structures and
experiments to amplify and verify quantum effects. The ac-
curate description of the interplay of structured environments
and quantum dynamics especially in the non-perturbative
regime requires the development of novel theoretical meth-
ods. These concepts will be discussed and shown to apply
to the current three examples of quantum effects in biology,
photosynthesis, avian magneto-reception and olfaction. It is
the hope that these examples will be joined by many others
and lead to the emergence of a new research branch, quantum
biology.
GENERAL QUESTIONS AND PRINCIPLES
In the following we would like to draw attention to a
number of general points related to the study of quantum
effects in biology and identify broad questions that might
be worth exploring.
Biology, dynamics, transport — Although perhaps ob-
vious, we would like to stress that biology is not merely
about static structures but that the dynamics that is en-
abled by these highly organized structures plays a key
role. Indeed, this dynamics is essential as biological sys-
tems need to be supplied with energy and drained of
entropy to maintain the out-of-equilibrium state that life
represents. This requires a wide variety of transport pro-
cesses including excitation energy transfer but extend-
3
ing to charge transport including electrons, protons and
ions as well as the transport of larger molecules, peptides
and proteins. But the role of transport processes extends
much further as they are also essential ingredients in pro-
cesses such as signal recognition and transduction which
requires the transfer of excitations or real physical parti-
cles. Hence transport and more general dynamical phe-
nomena play a fundamental role in biology and therefore
in the following discussion of quantum effects in biology.
Quantum traits & advantages — Quantum physics of-
fers a wide variety of features that distinguish it from its
classical counterpart and which may in some cases allow
for enhanced performance and functionalities. This in-
cludes coherence and interference, that is, wave-like fea-
tures of particles which may lead to faster propagation
for example in quantum random walks [9]. More sophis-
ticated multi-particle coherence phenomena distributed
across different components of a system such as entan-
glement hold the potential to achieve higher sensitivity
to external signals [22, 23]. Another quantum trait, the
quantization of energy, leads to well-defined energy levels
that can be excited in discrete portions, quanta of energy,
only. In the exchange of energy between two quantum
systems these features will then enforce highest efficiency
if the respective energy levels, such as vibrational eigen-
frequencies, are well matched. This in turn permits the
unambiguous identification of frequencies and thus can
facilitate the construction of sensing devices [20].
Therefore elucidating to what extent and under what
conditions quantum traits may be realised in biological
systems and how they are exploited for enhanced perfor-
mance is of considerable interest.
Open quantum systems — Biological structures in gen-
eral are not isolated as they are open systems that are in
permanent contact with their environments. While many
combinations of system and environment are conceivable,
in this text the system will tend to be formed of electronic
degrees of freedom (excitons, electrons, ...) while the en-
vironment will either be of vibrational nature or com-
posed of electron and nuclear spin degrees of freedom.
While thermal fluctuations imparted by the environment
may be beneficial for example to overcome potential bar-
riers between distinct classical configurations and there-
fore facilitating processes such as protein folding or chem-
ical reactions, the benefit of system-environment inter-
actions is less evident in the quantum world. The un-
controlled fluctuations of these environments will lead to
changes in the local structures in which the electronic
degrees of freedom are embedded and may therefore lead
to decoherence. The advantages of quantum coherence
in random walks [9] may for example be destroyed by
decoherence [24] and therefore the normal expectation
in quantum technologies tended to be that almost com-
plete isolation of the quantum system from its environ-
ment is required to reap the benefits of quantum effects.
However, it was realized in that field that the interac-
Protein Molecules Control of electronic and vibrational structures to create & optimize function Refined experiments Classical Biology Quantum Biology New numerical methods for modelling Photosynthesis Olfaction Avian Magnetoreception tion between system and environment may lead to the
creation of quantum properties such as coherence and
entanglement [4] and investigations in quantum biology
found that they may enhance transport for example in
photosynthetic quantum networks [15, 16].
The elucidation of mechanisms by which the quan-
tum dynamics on networks may enter a fruitful inter-
play with their environment to achieve enhanced perfor-
mance and long-lived quantum coherence in transport,
signalling and sensing is thus of fundamental importance.
Environments are structured — As we shall discover,
the role of the environment in biological systems is not
a passive one. Biological environments are not feature-
less sources of white noise nor do they represent merely
weak perturbations. The environmental spectral densi-
ties [25] describing their interaction with the system tend
to display two principal structures, a broad smooth back-
ground which has a short memory time and interacts with
the system mainly through its fluctuations, i.e. causing
noise, and well defined narrow features corresponding, for
example, to long-lived vibrational motion that can lead to
quasi-coherent dynamical non-equilibrium exchanges be-
tween system and environment. As we will see later the
smooth background is due mostly to the protein environ-
ment as well as noise processes originating from solvents
while the sharp features tend to originate from long-lived
vibrations that belong to molecules that are held within
the protein scaffold [26, 27].
The rich structure of biological environments leads to
non-Markovian dynamics which is also non-perturbative
being neither weak nor strong compared to the intra-
system dynamics. These features deviate from environ-
ments that are typically studied in quantum technologies
and uncovering non-trivial consequences of these struc-
tures that may have been exploited by nature and are
therefore worthy of careful study.
Out of equilibrium & back-action — Many processes of
interest to quantum biology are initiated by the sudden
generation of an initial excitation, such as a quantum of
energy to be transported, which drives the system away
from equilibrium. As a consequence of this sudden ex-
citation neither the system nor its environment will re-
main stationary. The ensuing out-of-equilibrium dynam-
ics then leads to perturbation of the environment by the
system and a back-action of the very same environment
on the system initiating as a result a dynamical exchange
between system and environment. If the perturbation of
the environment concerns the broad smooth background
it will relax back very rapidly to its equilibrium value due
to its short memory (correlation) time. If, on the other
hand, a narrow spectral feature, relating to a long-lived
vibrational mode, is excited then this will lead to long
lasting quasi-coherent motion of the environment with
the possibility of triggering quasi-coherent exchange with
the system. This and other non-equilibrium exchanges
between system and environment can have considerable
4
influence on the system dynamics [28–35], influence the
direction of energy transfer [28] and may even lead to the
generation of long-lived quantum coherence in the system
[28, 36–38].
The functional role of this non-equilibrium exchange
and in particular the interaction of electronic system de-
grees of freedom with long-lived vibrational modes rep-
resents a key question in quantum biology. Thus the in-
vestigation of the non-equilibrium, non-perturbative and
non-Markovian character of the system-environment in-
teraction is of considerable interest. Besides the concep-
tual impact that it may have it also calls for the devel-
opment of novel methods for the accurate numerical de-
scription of such system-environment interaction which
must be capable of going well beyond the usual pertur-
bative treatments in the description of most quantum
technologies [87, 154, 163, 168, 178].
Control of Structure — The discussions so far sug-
gest that quantum dynamics, structured environments
and their mutual interplay may provide increased effi-
ciencies and potentially even novel functionalities.
In
order to optimize this interplay between system and en-
vironment, biological systems cannot simply be random
conglomerates but must possess structure that they can
control. Examples are proteins which are essentially one-
dimensional amino-acid sequences, representing the pri-
mary structure, that have the ability to fold into spe-
cific three-dimensional arrangements whose short range,
secondary, structure defines the local arrangement via
the relative orientation of the amino acid residues, while
the global three-dimensional arrangement, denoted as the
tertiary structure, provides the long-range structure. It
will become crucial that far from being passive scaffolds
whose mere purpose is to hold other molecules in their
place, the primary, secondary and especially the tertiary
structure of proteins play a much broader and active role
for the emergence of quantum effects in biology. In fact,
their purpose is to control and tune the properties of the
actual molecular network, the structure of the environ-
ment in which this network is embedded and, crucially,
the interaction and interplay between the two. The pro-
tein can achieve this by tuning properties such as local
energies, by adjusting the local environment mainly via
its secondary structure, and the coupling rates within the
transport network, by adjusting the distance and relative
orientation of constituents as determined mainly through
its tertiary structure, and at the same time the properties
of the environmental fluctuations that this transport net-
works are subjected to are determined and tuned mainly
via the secondary protein structure. The latter may not
only be achieved by structuring the protein itself but also
by placing within it molecules that provide desirable vi-
brational features. This optimization that is required
here will have taken place on evolutionary time scales.
It is noteworthy though that, distinct from evolution-
ary optimization, via the control of conformation within
an existing structure nature also possesses the means to
switch between different states of operation.
How this control is realised in detail and which config-
urations exploit optimally quantum properties as well as
system-environment interaction is of considerable inter-
est for the exploration of quantum effects in biology.
Technologies & Experiments: Hardware and Software
– The questions and principles stated so far will guide our
thinking and allow theory to progress. It must be stressed
however that the study of quantum effects in biology is an
endeavour in which theory alone cannot succeed and has
to proceed hand-in-hand with experiment. So much is
unknown about biological systems that assumptions that
are entering theoretical models as well as the predictions
of these models must be tested against experiment. In-
deed, as we already stated in the introduction, the quest
for quantum effects in biology has gained momentum re-
cently in no small part due to the technological advances
that are increasingly permitting the direct observation
of quantum effects in biological systems and the prob-
ing of these systems on previously unknown length- and
time-scales.
The future development of the field will depend criti-
cally on the identification and development of technolo-
gies, classical or quantum physical, that can lead to
deeper insights into the workings of biological systems
at the length and time scales at which quantum effects
can act. New technologies and new experimental set-ups
as well as protocols need to be developed (see e.g. [39–42]
for a few examples). Furthermore, the complexity of the
systems under investigation suggests that methods from
fields such as signal processing [43–45] may prove fruitful
to optimize experiments and subsequent analysis.
It will be exciting to see how new experimental phe-
nomena and theoretical principles will stimulate the de-
velopment of novel experimental tools and clever proto-
cols. This, we believe, will be essential for the devel-
opment of the field as only irrefutable direct proof of
quantum phenomena and their functional role will force
biologists to adopt quantum theoretical concepts to un-
derstand foundational aspects of biology.
QUANTUM DYNAMICS IN BIOLOGICAL
ENVIRONMENTS
Before we start to discuss a set of general and gener-
alizable principles that govern the quantum dynamics of
biological systems we would like to discuss here an exam-
ple that presents us with a ”smoking gun” which suggests
that the study of the interplay between quantum dynam-
ics in the presence of structured environments may lead
to interesting insights.
5
FIG. 3. A schematic picture of the transport network such
as the one realized in the Fenna-Matthews-Olson complex.
Molecules, such as Bacteriochlorophyll a (BChla), are ar-
ranged in space giving rise to specific distances and relative
orientations between individual BChla thereby adjusting the
strength of the dipolar interaction (indicated by black arrows)
between excitations on different BChla molecules. The sur-
rounding protein (not shown) is also able to control the local
environment of the BChla molecules and thus adjust their ex-
citation energies. The specific arrangement and the nature of
the environment determines the transport efficiency from the
left upper site (site 1 in the FMO) that accepts excitations
from the antenna complex and transfers them via the lower
right site (site 3 in the FMO complex) to the reaction center
where charge separation is initiated to begin the irreversibly
process of binding the exciton energy in chemical form.
Transport & Environment: An illustrative example
Let us consider a transport network that is composed
of a set of highly absorptive entities such bacteriochloro-
phyll molecules each of which may support an electronic
excitation (an Frenkel exciton). These molecules will be
arranged in space by a protein scaffold and together they
form a pigment-protein complex. In our concrete exam-
ple this will be the Fenna-Matthews-Olson (FMO) com-
plex [46, 47] which forms an integral part of the photosyn-
thetic light harvesting complexes of green sulphur bacte-
ria [48]. Such complexes serve to transport electronic
excitations, excitons [49], in the presence of a vibrational
environment.
We will contend that nature can construct and opti-
mize pigment-protein complexes, for example the FMO
complex, to create transport networks that exploit both
the quantum dynamics of its electronic degrees of free-
dom and their interaction with a structured vibrational
environment. The detailed mechanisms that nature has
used to achieve this and the design principles it is follow-
ing will be discussed in the next section and then used
as the basis to understand other seemingly unrelated bi-
ological quantum phenomena.
Reaction center Antennae Let us first provide evidence, by means of a numeri-
cal example, that such tuning and optimization may be
taking place and that the vibrational environment may
indeed have a significant and indeed beneficial impact on
the transport dynamics of the FMO complex. Then we
will move on to derive generalizable conclusion from these
findings. The full Hamiltonian describing the exciton-
vibrational interaction as well as the exciton-exciton in-
teraction is given by H = Hex + HI + HB where
Enn(cid:105)(cid:104)n +
[Jmnm(cid:105)(cid:104)n + h.c.],
(cid:88)
m(cid:54)=n
1
2
(1)
(2)
N(cid:88)
(cid:88)
(cid:88)
n=1
i,k
1
2
Hex =
HB =
HI =
†
ωka
ikaik,
(cid:88)
(cid:112)
[
n
k
†
nk)n(cid:105)(cid:104)n + h.c.].(3)
Snkωk(ank + a
Here n(cid:105) describes an excitation on site n, the Jmn de-
scribe the dipolar interaction between excitation on sites
†
m and n and the operators ank, a
nk denote bosonic de-
struction and creation operators for the kth independent
vibrational mode coupled to site n [46]. The exciton-
mode interaction is determined by the strength of their
Huang-Rhys factors Snk [50]. Note, that HI describes a
purely dephasing interaction because the vibrational de-
grees of freedom have energies that are at least 10 − 100
times smaller than the excitation energies of the excitons
thus suppressing direct exciton-phonon interconversion.
The dephasing can be understood to originate from the
fact that vibrations will change the local environment of
each site (e.g. moving charges) and thus affect the excita-
tion energy of the relevant site [51]. We assume that the
dynamics is dominated by contributions to the spectral
density where each site interacts with its own indepen-
dent environment an assumption that is corroborated by
first-principles numerical studies of photosynthetic com-
plexes [52, 54–57] and normal-mode analysis combined
with quantum chemical methods [58, 59]. The interaction
between site n and its environment is characterized by
kδ(ω − ωk) which
is a joint property of the environment and the system
combining the strength of interaction of modes with the
mode density.
the spectral density Jn(ω) =(cid:80)
k Snkω2
The full dynamical equation also need to include two
further contributions, one describing the spontaneous an-
nihilation of an exciton and the concomitant loss of the
energy into the general environment at a rate γloss, a
process that nature would like to avoid. Secondly, the
transfer of the excitation from a specific site, in the FMO
complex, that is the site labeled 3, into the reaction cen-
ter which is again described by an irreversible decay at
rate γRC motivated by the fact that in the reaction cen-
ter charge separation is achieved to irreversibly stabilize
the excitonic energy. Both contributions enter the equa-
tions of motion via typical Lindblad terms [60] so that
we find the global time evolution of the density operator
describing system and environment to be governed by
6
dρS,E
= −i[H, ρSE]
N(cid:88)
n=1
dt
−γloss
−γRC(3(cid:105)(cid:104)3ρSE + ρSE3(cid:105)(cid:104)3 − 20(cid:105)(cid:104)3ρSE3(cid:105)(cid:104)0)
(n(cid:105)(cid:104)nρSE + ρSEn(cid:105)(cid:104)n − 20(cid:105)(cid:104)nρSEn(cid:105)(cid:104)0)
(4)
where 0(cid:105) denotes the electronic ground state of the
pigment-protein complex. Needless to say, these dynam-
ical equations describing the full state of system and vi-
brational environment are generally too complex to be
solved exactly. Crucially however we will see shortly that
the structure of the environment does play an important
role and thus the development of methods that can cap-
ture these features will be of considerable importance for
quantitative studies of such systems. In order to bring
out the main points clearly, we would like to delay the dis-
cussion of these numerical challenges to final part of this
article. In order to obtain our first observation we follow
the approximate treatments presented in [15, 16, 61, 62]
and will improve on these later on.
To this end we will model the system-environment in-
teraction perturbatively and derive a master equation for
the system evolution only to model the transport through
the FMO complex under low light conditions, i.e.
the
rate γin at which excitations enter the network via site
1 is much smaller than the rate γRC at which excitation
leave the network from site 3 into the reaction center.
As a consequence, the mean population in the transport
network is much smaller than unity at any time. The
typical but at first sight perhaps surprising result of such
an analysis is presented in Fig. 4. Contrary to what one
might have expected, the conductivity of the electronic
transport network in the FMO complex (quantified as the
rate at which the reaction center is populated in steady
state divided by the rate at which excitations enter the
transport network) exhibits a maximum at a finite de-
phasing rate, that is, dephasing noise can actually assist
the electronic transport [15, 16].
This immediately raises the question as to whether the
regime that results in optimal transport performance is
found to be essentially classical in the sense that the dy-
namics is well represented by a rate equation model or
whether, despite the dephasing noise, it remains firmly
in the quantum mechanical regime in which quantum co-
herent dynamics is only weakly perturbed by dephasing
noise. Questions of this type can be answered both in
a qualitative and a more quantitative manner. First, an
examination of the parameters that are typically enter-
ing the dynamical equations when describing photosyn-
thetic complexes reveals that the strength of the intra-
system coupling, e.g.
the dipolar interaction, is com-
parable in strength to the system-environment coupling.
Hence one already expects that the dynamics is taking
7
pendently from the others at a fixed frequency and phase.
If the phase for each pendulum is chosen at random then
the global signal appears incoherent while clearly each
pendulum is coherent). While natural conditions do not
resemble laser light, laser spectroscopic experiments on
individual specimens provide the sharpest tools for the
identification of the dynamical equations that govern the
system evolution and thus have a crucial role to play in
the determination of quantum effects in biological sys-
tems.
These observations raise the questions as to why opti-
mal performance is achieved in this intermediate regime.
Answering this question will lead us to identify the dy-
namical and structural principles that are underlying op-
timal performance of quantum transport networks.
It
will drive us towards uncovering a rich interplay between
electronic degrees of freedom and their vibrational envi-
ronment and point towards the possibility that nature
has optimized both electronic networks and vibrational
environment in an evolutionary process. This will be the
subject of the next section.
DESIGN PRINCIPLES
Here we will elucidate basic principles that have been
found to underlie the fruitful interplay between vibra-
tional environments and coherent quantum dynamics
[15, 16, 28, 61, 62, 66, 67, 71–78]. Identifying and un-
derstanding these principles at a deep,
intuitive level
and seeing how nature may have used them to opti-
mize performance does provide additional value even if
individual processes in specific circumstances have been
known in different physical situations. The principles
that we present here will also allow us to bring under one
umbrella several seemingly unrelated biological phenom-
ena, namely excitation, electron and proton transfer pro-
cesses, the chemical compass model of magneto-reception
of birds and the mechanisms underlying olfaction, each
of which is suspected to be governed in an essential way
by quantum phenomena.
Controlling resonances – The phonon antenna
We will begin by elucidating a first principle which will
provide an understanding why optimal transport perfor-
mance in the FMO complex can be achieved at inter-
mediate noise levels. More importantly, this principle
is sufficiently general to provide a mechanism that can
support the surprisingly long-lasting oscillatory features
observed in recent ultra-fast laser spectroscopy experi-
ments [11–14] and explain key aspects of the dynamics
that may underlie the process olfaction.
We will approach this topic by means of a simple but
instructive question concerning the optimization of a sim-
FIG. 4. Plot of the conductivity of the FMO complex where
excitations enter the FMO complex at site 1 and exit at site
3 versus the overall strength of the dephasing noise due to
the interaction between electronic and vibrational degrees of
freedom. The key observation is that optimal performance is
found for an intermediate level of dephasing noise. No noise or
very strong noise are counterproductive for the performance
of the transport network. See [62] for the Hamiltonain param-
eters and the spontaneous emission rate, the noise is modeled
as local dephasing Lindblad master equation in the site basis.
place in a regime in which neither dephasing noise nor
quantum coherent dynamics clearly dominate. This is
further corroborated by the examination of the coherence
and entanglement [63, 64] properties of states [62, 65]
and, more importantly, the dynamics of the system [66]
which demonstrate quite clearly that on shorter length-
and timescales quantum coherence is present in the sys-
tems while for longer distances and times classical prop-
erties dominate. This suggests that indeed, the optimal
operating regime in this setting is found to be ”halfway”
between the classical and the quantum world.
In the light of recent discussions concerning the rel-
evance of coherence in photosynthetic systems, we feel
that an important remark is in order here. The coherence
properties of the states and dynamics of a system that
are observable in an experiment may depend very much
on the specific experimental set-up and the specific exci-
tation regime that the system is subjected to. Excitation
by incoherent sunlight as well as ensemble averages may
suppress the observed coherence and have tempted some
researchers to reach the conclusion that coherence may
be of no relevance in these systems under natural condi-
tions. This however is not necessarily correct as the cru-
cial point are the coherence properties of the underlying
dynamics and not of signatures of coherence in an exper-
imental signal. It is the equations of motion that deter-
mine the performance of each individual system and are
unaffected by the nature of the initial preparation or en-
semble averages that may obscure the observed coherence
(consider a set of pendula each of which oscillates inde-
−2−101234560.650.70.750.80.850.90.951Log(Dephasing rate)Conduction Efficiency 8
spectral density has a single maximum and thus takes
roughly the shape depicted in figure 6. A numerical op-
timization employing Redfield equations to take account
of the spectral structure of the environment (see [34, 61]
for theoretical and numerical details) now finds that the
optimal position of site 2 is close to site 1 such that it
exhibits a strong coherent dipolar interaction and close
in excitation energy. Having found this numerical results
we now would like to rationalize its origin and thereby
arrive at a very useful design principle – the phonon an-
tenna.
ple transport network (see fig. 5 for a schematic repre-
sentation of the following). Consider a network made
up of only three sites, namely site 1, which accepts exci-
tations from the antenna and site 3 which is connected
to the reaction center. Both, site 1 and site 3 are fixed
in their properties (position, orientation and excitation
energy). The system is completed by site 2 whose excita-
tion energy, position and orientation, and hence dipolar
interaction strength with sites 1 and 3 we are free to
choose. We assume that site 3 provides the zero of ex-
citation energy, site 1 has an excitation energy that is
300 wavenumbers higher (for readers unaccustomed to
wavenumbers, note that 1 wavenumber, also denoted as
1cm−1, corresponds to ω = 1.88· 1011s−1). The question
that we would like to answer concerns the optimal choice
of excitation energy, position and orientation of site 2 or
in other words the optimal choice of the excitation energy
of site 2 and its dipolar coupling strengths to sites 1 and
3? As such, this question cannot be answered unambigu-
FIG. 5. A network made up of only three sites, site 1 which
accepts excitation from the antenna, site 3 which is connected
to the reaction center both of which are fixed in their prop-
erties (position, orientation and excitation energy) as well as
another site 2 whose excitation energy, position and orienta-
tion, and hence dipolar interaction strength with sites 1 and
2 we are free to chose. What is the optimal choice of the ex-
citation energy of site 2 and its dipolar coupling strengths to
sites 1 and 3?
ously as we are missing a crucial piece of information,
namely that of the structure of the environmental fluc-
tuations. This structure is characterized by the spectral
density of the environment which is a combination of the
density of environmental modes and their individual cou-
pling strength to the system. Typical spectral densities
in pigment-protein complexes possess considerable struc-
ture with sharp peaks originating from long-lived vibra-
tional modes as well as a broad background whose maxi-
mum tends to be in the range of around 200cm−1 which
we will now assume for the subsequent optimization. For
the sake of clarity, let us assume that the environmental
FIG. 6. In the upper figure, two closely spaced energy levels
are separated from a third level to which excitations should
be delivered. They are subject to dephasing noise from an en-
vironment with a finite bandwidth that exhibits a maximum.
A coherent interaction between the upper two energy levels
leads to dressed states ±(cid:105) with an energy splitting which,
if matched to the maximum of the environment spectral den-
sity, will optimize transport from the upper to the lower level.
Hence the dressed states act as an antenna to harvest envi-
ronmental fluctuations to enhance transport.
Indeed, the strong coherent dipolar interaction be-
tween sites 1 and 2 suggests that we move to a new
basis made up of the eigenstates of the coherent part
of the dynamics of these two sites, that is the exci-
tonic states of that system or, for quantum opticians,
the dressed state picture. This change of picture leads
us to rewrite the Hamiltonian eq. (3) that describes the
system-environment interaction in the excitonic basis of
nen(cid:105),
eigenstates {en(cid:105)} of eq. (1), so that i(cid:105) = (cid:80)
n C i
and the coupling terms
where
HI =
Qnm =
1
2
(cid:88)
(cid:88)
(cid:112)
n,m
ik
(Qnmen(cid:105)(cid:104)em + h.c.),
SkωkC i
nC i
†
m(aik + a
ik).
(5)
(6)
? ? Antennae Reaction center J2)(SJ0 J21233Dressed levels 12 This leads us to two insights. Firstly, in the exciton
(dressed state) basis the action of the dephasing noise
now leads to transitions between excitons, that is ampli-
tude noise, which facilitates transport towards the lower
of the two exciton states. Secondly, the two excitons
(dressed states) are separated by an energy difference
that is related to the coherent dipolar coupling strength
and the energy difference of sites 1 and 2. The dom-
inant contribution to the transition between these ex-
citons (dressed states) arises from those environmental
modes whose frequency closely matches the energy dif-
ference between dressed states. Indeed, the optimal so-
lution is such that the energy separation of the dressed
states matches the maximum of the environmental spec-
tral density, i.e. where the environmental fluctuations
are strongest, so that the environment may bring about
transitions between the dressed states most effectively.
In this sense, we can argue that the two eigenstates of
the coupled Hamiltonian are tuned to harvest environ-
mental fluctuations to achieve optimal excitation energy
transport through the formation of a tunable ”phonon
antenna”.
Observing these two points alone, that is inducing
strong coherent coupling to move to a dressed basis and
tuning the coupling such that it matches the maximum
of the spectral density of the environmental fluctuations,
one already comes close to the numerically obtained so-
lution of the above optimization problem and can thus
optimize excitation energy transport. It should be noted
that the phonon antenna principle is also capable of
making predictions about more complex transport net-
works such as that of the FMO complex. Indeed, it was
found that the physically important relaxation pathway
between sites 1 and 3 is mediated by pigments which
are spectrally and spatially positioned by the protein to
efficiently sample the spectral function of the proteins
fluctuations [34, 61]. Whether this optimality is a de-
terminant in the emergence of this structure in nature is
another matter altogether, but it is striking how well the
phonon antenna concept can be used to rationalise the
site energies and couplings of the pigments participating
in this pathway. The role of vibrational modes has also
been studied for other light harvesting complexes such as
the cryptophyte antenna protein phycoerythrin 545 [35].
It now becomes transparent why the optimal operat-
ing regime for excitation energy transport may actually
be found to be where coherent dipolar interactions and
system-environment interactions, i.e. dephasing noise,
are of broadly comparable strength.
Indeed, if the en-
vironmental noise is too weak, then the formation of
dressed states will present little benefit for transport. On
the other hand if dephasing noise is very strong then it
will suppress the formation of the dressed states and thus
of the phonon antenna effect in the first place. As a con-
sequence, an intermediate regime where dephasing noise
of intermediate strength is present naturally appears as
9
the optimal operating regime according to the phonon
antenna. We will later see that there are a number of
additional mechanisms in which noise and coherent dy-
namics coexist to lead to similar conclusions [62, 79].
Before we move on to the discussion of these effects, it
is instructive to highlight a connection that the phonon
antenna concepts shares with the concept of Hartmann-
Hahn resonances [80] in nuclear magnetic resonance and
spin sensing. Here one is faced with the challenge that the
sensor, e.g., an electron spin, may not be resonant with
an external system generating a signal (e.g., a different
electron or nuclear spin). This problem may be overcome
by continuously driving the electron spin in the sensor at
a strength that leads to dressed states whose splitting
matches the frequency of the external signal and thus
permits a strong coherent response of the sensor (transi-
tions between upper and lower dressed states) [80–82].
With this in mind we now move on to bring out two
other conclusions and connections that we can draw from
the phonon antenna concept in the case when the spec-
tral density of the environment possesses very sharply
peaked features, that is well-defined long lived vibrational
modes. Indeed, this will provide a both a mechanism ex-
plaining the origin of long-lived coherences observed in
recent ultrafast spectroscopy experiments and a connec-
tion to biological sensors.
Long-lived coherences as a non-equilibrium process
Experimental observations employing ultrafast 2-D
spectroscopy on various photosynthetic complexes exhib-
ited long-lived oscillatory features which were interpreted
as evidence for long-lived electronic coherence in the sys-
tems under investigation [11, 13, 14]. Under this hy-
pothesis electronic coherence appear to exhibit lifetimes
that can reach the picosecond range thus exceeding ex-
pectations from condensed matter systems at least ten-
fold. This interesting observation gave rise to a variety
of attempts for explanations of the long-lived coherences
including (i) overall reduction of dephasing [57, 83] which
are however not compatible with the observed very short
lifetimes of optical exciton coherences in the system, (ii)
correlations in the noise sources between different sites
[52, 53, 85] which are however not supported by first
principles calculations of spectral densities [54–57] and
normal-mode analysis combined with quantum chemi-
cal methods [58, 59] and (iii) variations of the electronic
structure of the FMO complex [86] which are not suffi-
cient however to explain the observed durations. In the
following we will show that the inclusion of significant
coupling of electronic motion to long-lived vibrational
modes [28, 34] are capable of explaining the observations
[36, 37] and even more so to give support to the idea
that vibrational motion plays and important role for elec-
tronic transport, quantum or classical – a principle which
we will show here to be of broader importance in biology.
Fig. 7d). Secondly, we note that the sudden displace-
10
To gain an insight into possible mechanisms that sup-
port long-lived electronic coherences in biological sys-
tems, we will now take the phonon antenna principle to
its extreme by applying it to a system-environment inter-
action in which the broad features of the spectral density
are supplemented by sharp features due to the presence of
some long-lived and well-defined vibrational mode as pro-
posed in [28]. As we have learnt in the previous section,
tuning the dipolar interaction in the electronic system to
the energy difference of the excitonic states that matches
the maximum of the spectral density will maximize the
rate at which transitions between these states occur. For
a broad and smooth spectral density these transitions will
be dominantly incoherent (following essentially a Fermi’s
golden rule argument). In the presence of a long-lived vi-
brational mode the phonon antenna principle remains the
same but, crucially, the nature of the interaction changes
as the interaction between a single vibrational mode and
the electronic degrees of freedom is coherent, at least for
as long as the mode itself remains coherent.
Let us examine this mechanism in more detail by con-
sidering an exciton transport network in which excitons
enter in one site and exit in another. Initially the net-
work is in its ground state and no excitons are present.
This situation is depicted schematically in Fig. 7a where
a pendulum represents the long-lived vibrational mode
which we assume to be in a thermal state with small ex-
citation number and where the black dot represents the
population of the ground state of the electronic system.
The higher lying electronic levels, representing the vari-
ous exciton eigenstates of the electronic system, are not
excited even at room temperature as the excitation en-
ergy is in the range of eV. The initial (fast) injection of
an exciton, either coherently or incoherently, populates
one of the exciton states of the system (raised black dot
in Fig. 7b) and creates a sudden force on the electrons
and nuclei and thus change their equilibrium positions
(compressed spring in Fig. 7b). Now the environment
will start to react to these forces which initiates transient
oscillations of the modes at approximately their natural
frequency ωk. The continuous background of the spectral
density will relax very rapidly into the new equilibrium
state as it contains a broad range of frequencies and thus
possesses a very short correlation time. The well-defined
long-lived vibrational mode will oscillate for a consider-
able time (which can be up to several picoseconds) and
will interact with the electronic system (in Fig. 7c we
see that the spring connecting the vibrational mode to
the electronic motion is periodically expanded and com-
pressed). This in turn leads to oscillations between dif-
ferent exciton states. We make two observations: Firstly,
these oscillations will have the largest amplitude between
those exciton states whose energy difference is nearly res-
onant with the frequency of the vibrational mode (see
FIG. 7. Simplified mechanical illustration of a possible princi-
ple behind long-lived coherence in biological systems. Details
are presented in the main text.
ment of this mode implies that it is now found to be in a
displaced thermal state which will be close to a coherent
state if the frequency of the mode is such that its thermal
occupation number is low. As is well-known in quantum
optics a mode in a coherent state acts on a two-level
system essentially like a time-dependent classical driving
field resulting in coherent Rabi-like oscillations following
approximately the Hamiltonian
((cid:104)Qnm(cid:105)(t)en(cid:105)(cid:104)em + h.c.),
(7)
(cid:88)
where (cid:104)Qnm(cid:105)(t) ∝ (cid:80)
Hdriving ≈ 1
2
n(cid:54)=m
√
ik
nC i
SkωkC i
m sin(ωkt) in the
above-mentioned approximation of the initial, transient
and coherent response of the modes to exciton injection.
As a result we will observe coherent transitions between
dissipative excitonic states and thus coherences that will
last for the coherence time of the vibrational mode –
an expectation that can be corroborated by more so-
phisticated numerical treatments employing methods de-
scribed in [28, 87–89]. This electron-vibrational coupling
regenerates electronic coherences in the system to replace
those that are continuously damped out by the fluctua-
tions of the smooth background environment and may
even lead to stationary entanglement [90]. We note that
in the language of chemical physics, this phenomenon
arises through coherent non-adiabatic coupling (via dis-
crete mode motion) which induces oscillatory crossing of
potential surfaces. The physical picture presented here
illustrates a key point: electronic coherence may emerge
from transiently exciting robust, weakly dephasing vibra-
tional coherences which then transfer back coherence to
those exciton transitions that are well matched to the
mode [28, 36, 37].
The importance of vibrational modes for interpret-
ing experimental observations in multidimensional spec-
troscopy has only recently begun to be appreciated
[28, 36–38, 87] and from the discussion above is seen to
be intimately related to the phonon antenna concept and
in fact quantum sensing.
Phonon-assisted electron tunneling & olfaction
In the following we will explain how the coupling of
electronic and vibrational motion may also underlie the
function of biological sensors and exemplify these ideas
at the hand of a mechanism suggested to be an important
contribution to the function of olfactory sensors.
Despite considerable progress concerning the under-
standing of the structure of olfactory receptors that in-
volved in the early stages of the olfactory process, the
detailed mechanisms by which we are able to discrim-
inate between the vast number of odorants are not yet
fully understood [91]. This is emphasized by the fact that
for nearly 100 years researchers have striven, with limited
success, to identify principles that allow for the predic-
tion of smell. The principal reason for this failure can
be traced back to the lack of a detailed understanding
of what is actually happening during and shortly after
the process of binding of odorants to the bindings site of
receptors. There are currently two mechanisms proposed
that are based on quite different mechanisms and which,
crucially, lead to different experimental predictions. It is
possible that both contribute in a critical way to olfac-
tion.
On the one hand there is the idea of the lock-and-
key principle. The olfactory stimulus is provided my
molecules, odorants, that arrive at the receptor via dif-
fusion through the air. In the absence of the odorant the
binding pocket and the receptor exhibits small thermal
fluctuations about some equilibrium conformation. Only
certain types of odorants will be capable of attaching
themselves to the binding pocket, a choice determined
by chemical affinity, shape etc (this is the ”key in the
lock” part of the principle). Once attached, the interac-
tion between the odorant and the receptor results in a
change of the average conformation of the receptor. This
conformational change is then posited to induce further
processes and initiates a signalling chain (this part con-
stitutes the ”turning of the key” part of the principle).
For many receptors, especially those binding only a very
specific molecule, this appears to be a useful and valid
principle. Therefore it seems natural to adopt the very
same principle also for olfactory receptors. There are at
least 100000 odorants but far fewer olfactory receptors –
several hundreds in humans, so that there is not a specific
receptor for each single odorant. The ability to differen-
tiate such a large number of odorants would thus require
that each odorants may bind to a variety of receptors.
This would give rise to a vast number of distinct binding
patterns and the subsequent sensation of smell.
11
Despite its attractiveness and applicability in certain
cases the lock-and-key principle has to be subjected to ex-
perimental test. This is where it has recently experienced
some significant challenges. Firstly, it is not straightfor-
ward to explain by means of the lock-and-key mechanism
alone why outwardly very similar molecules may smell
completely different while molecules with rather different
shape may smell similar. Even more remarkable is the ob-
servation in recent experiments that Drosophila flies are
capable of discriminating between molecules in their hy-
drogenated and their deuterated form and, importantly,
are able to generalize from deuterated molecules to other
molecules that exhibit a vibrational modes similar in fre-
quency to the Carbon-Deuterium stretch mode [21] (see
[92] for recent experiments with similar outcomes on hu-
mans).
Indeed, these observations provide some support for an
alternative theory that is based on physical properties of
molecules rather than chemical or shape-based ideas that
are underlying the lock-and-key principle. Remarkably,
it was well before the advent of the lock-and-key princi-
ple that Dyson in 1938 proposed [93] that the smell of
a molecule may be determined by its vibrational spec-
trum and that hence the olfactory system effectively op-
erates as a vibrational spectrometer (and idea that has
also been pursued later, starting in the 1950’s, by Wright
[94]). This is an attractive idea, especially to a physicist,
but at the time it suffered from a quite severe drawback –
the lack of a plausible mechanism by which the olfactory
system would in fact be able to identify specific vibra-
tional modes. It was Luca Turin [20] who promoted the
idea that inelastic electron tunneling (IET) may play a
role in olfaction which, if true, would have the attrac-
tive feature that it would help to predict the smell of
a molecule from the analysis of its vibrational spectrum
[95]. IET is in fact a well established physical method for
determining the vibrational spectrum of molecules [96].
It was discovered in solid state physics by Jaklevik and
Lambe in 1966 when they identified anomalous behaviour
in the conductivity through tunnel junctions in the pres-
ence of organic adsorbates [97, 98]. They realised that
these anomalies arose for certain voltages such that the
energy difference of the electrons on both sides of the
tunneling barrier would match a vibrational quantum of
energy of the adsorbed molecules and went on to demon-
strate that this mechanism, inelastic electron tunneling
spectroscopy, allows them to identify vibrational spectra
of molecules (apparently, they recognized its potential for
the theory of smell but were dissuaded by less far-sighted
colleagues [99]).
The proposed phonon assisted electron tunneling
mechanism for olfaction makes use of two phenomena
which, together, make it a proper quantum biological
phenomenon. There is on the one hand the tunneling
process of a massive particle, here the electron, and on
the other hand the fact that a vibrational mode, that is a
quantized harmonic oscillator, can only take up energy in
discrete quanta proportional to the relevant vibrational
frequency ωodor. It is the second aspect that makes it pos-
sible for this process to discriminate effectively between
different vibrational modes and thus between the vibra-
tional fingerprints of different molecules. A schematic
picture of the process is presented in Fig. 8. The bind-
ing pocket of an olfactory receptor, which begins empty
(upper part of Fig. 8), is assumed to be situated close to
an electron donor and an electron acceptor (these may
either be oriented alongside the binding site or separated
by the binding pocket itself, a difference that is impor-
tant in practice but not relevant for the following basic
argument). Crucially, it is assumed that electrons in the
donor have an energy that is ∆E higher than for electrons
in the acceptor and that the energy distribution in donor
and acceptor is much narrower than ∆E. In that case
tunneling is suppressed in the absence of the odorant be-
cause of the impossibility to satisfy energy conservation.
If the energy difference ∆E is well matched to the ex-
12
reside there for a certain time long enough for electron
transfer to take place. The presence of the odorant with
a vibrational mode that is well-matched to ∆E (see lower
part of Fig. 8) then allows for energy conservation dur-
ing a tunneling process – an electron can now tunnel from
donor to acceptor while using its excess energy to excite
a vibrational quantum of the odorant. This process will
be unidirectional for sufficiently low temperatures such
that kT (cid:46) ωodor. The concomitant charge redistribu-
tion then triggers a further sequence of events that leads
to a signalling cascade.
Needless to say very detailed calculations are required
to ascertain that the described mechanism is possible in
principle for reasonable choices of parameters [100, 101].
A wide variety of considerations need to be taken account
in these studies which include the determination of spec-
tral densities describing the electron-vibration coupling,
the effect that orientation of the odorant in the bind-
ing pocket may have on it [102], temperature effects etc
(see [103] for a more detailed analysis and [104] for an
overview).
What we would like to stress here though is the very
close resemblance of the above mechanism to the phonon
antenna principle and thus the importance of electron-
vibrational coupling. In both cases an electronic degree
of freedom senses the presence of a vibrational mode
due to the tuning into resonance of the energy differ-
ence of two states (dressed states in the phonon antenna
and donor/acceptor in olfaction) to the energy of single
quanta of the vibrational motion. This resonance condi-
tion ensures increased transport efficiency and thus leads
to a detectable effect at the physiological level (deliv-
ery of an electronic excitation to the reaction center or
electron transfer which stimulates a subsequent response
of a receptor). Both examples demonstrate the potential
usefulness of controlled resonances for biological systems.
Noise assisted processes
FIG. 8. A schematic representation of the principle behind
phonon assisted electron tunneling as a basis for olfaction. A
binding site is close to a electron donor and acceptor that are
at different energies which suppresses tunneling. If a molecule
with a vibrational mode whose energy closely matches the en-
ergy difference between donor and acceptor enters the binding
site then electron tunneling can become energetically allowed.
Then a charge redistribution becomes possible and a subse-
quent signalling chain may be triggered.
citation energy of a specific vibrational mode ωodor of
the odorant, then its presence may facilitate a detectable
and very specific change in the dynamics of the receptor
by the following process: The odorant enters the binding
pocket and structural changes may take place which give
space for the lock-and-key principle to play a certain role.
Once the odorant has entered the binding pocket it will
Needless to say, not all biological systems will possess
long-lived vibrational modes that are strongly coupled
to electronic degrees of freedom and can therefore play a
role in the dynamics of the system. What is present in
all biological systems however are thermal fluctuations
of the molecular and protein structures as well as the
surrounding solvents, water etc. These may lead to a
broad noisy background mainly resulting in dephasing
noise on the electronic degrees of freedom.
Given the omnipresence of these noise sources one may
ask as to whether nature has evolved to make use of in
conjunction with quantum coherent dynamics to support
processes that are of relevance to life. That this may
indeed be the case we will explain in the following by first
presenting several mechanisms by which dephasing noise
may in fact support transport phenomena. This will be
D A E D A E followed by a discussion of two phenomena in which noise
can be shown to assist fundamental biological processes.
In the first example we will briefly revisit transport in
photosynthetic complexes while in the other example we
will show that the mechanism that is proposed to underlie
the magneto-reception of birds depends crucially on the
presence of an environment.
for electron transfer reaction is given by
k = Ae
− −∆G
kB T ,
with ∆G given by
∆G =
(cid:18)
λ
4
1 +
∆G0
λ
(cid:19)2
13
(9)
(10)
which achieves a maximum for ∆G0 = λ. Here A is a
system dependent term, ∆G0 is the standard free energy
of the electron transfer reaction and λ is the reorganiza-
tion energy which quantifies the strength of interaction
between the electron and its environment composed of
vibrational modes and the solvent [105].
FIG. 9. Local dephasing, for example due to random fluctua-
tions of the energy levels generated from random vibrational
motion, leads to line-broadening and hence increased overlap
between sites. Viewing these fluctuations dynamically, as il-
lustrated by the double arrows, one finds that the energy gap
between levels varies in time. The resulting nonlinear depen-
dence of the transfer rate on the energy gap may therefore
lead to an enhancement of the average transfer rate in the
presence of dephasing noise.
It is worth noting that while the application of an ex-
cessive amount of dephasing noise suppresses transport
this may in itself serve a useful function if for example a
transport network contains sites, for example for struc-
tural reasons, that may lead to leakage of excitations into
domains from where further transport may be slow. In
such a case dephasing may be useful to reduce the effec-
tive transition rate to such sites and thus block unfavor-
able transfer paths from being followed [74, 106].
Bridging energy gaps & blocking paths
Pigment-protein complexes consist of a number of
sites, schematically depicted in Fig. 3, whose excitation
energies generally exhibit a certain degree of static dis-
order, that is, their on-site energies will differ from site
to site and also from one pigment-protein complexes to
another. If the energy difference between sites that ex-
change excitation is larger than the intersite coupling ma-
trix element in the relevant Hamiltonian, then transitions
will be severely reduced because of energy conservation
unless of course there is are quasi-resonant vibrational
modes present that, as was explained in the phonon-
antenna mechanism, can take up the energy difference.
The presence of a well-matched mode is not necessary
though. Broadband dephasing noise alone may already
come to the rescue in a manner that can be understood
from two different viewpoints. On the one hand one no-
tices that dephasing noise will lead to a broadening of the
excitation energy of each site and thus to an increased
overlap between the two energy levels while it does not
cause the loss of excitations from the system (see Fig.
9). Alternatively, one may take a dynamical viewpoint of
the same phenomenon by realising that dephasing noise
arises from the random fluctuations of the excitation en-
ergies of each site. As a consequence, these fluctuating
energy levels will occasionally come sufficiently close in
energy to allow for excitation energy transfer between the
sites as the energy difference has been reduced to a value
smaller than the direct coupling matrix element (see Fig.
9). Again, we observe that this mechanism will lead to an
optimal operating regime at intermediate levels of envi-
ronmental noise. Indeed, a low level of fluctuations will
not bring the site energies sufficiently close and trans-
port remains suppressed, while excessive fluctuations of
the site energies will reduce the time intervals in which
the sites are energetically sufficiently close to allow for
efficient energy transfer. This can be estimated easily by
computing the overlap between Lorentzian lines of width
γ that are displaced by an amount ω0 in which case we
find
(cid:90) ∞
1
π2
γ
γ
γ2 + (ω − ω0)2 dω =
2γ
π(4γ2 + ω2
0)
. (8)
Destructive interference, symmetry and noise
−∞
γ2 + ω2
which takes on a maximum at γ = ω/2.
An analogous behaviour arises already in the Marcus
theory of electron transport where the rate constant k
While linear networks can already exhibit interesting
noise assisted transport phenomena [16, 107–111], multi-
site networks may exhibit more complex behaviour which
dephSite i Site i Site j Site j deph arises due to the interplay between a wealth of construc-
tive and destructive interference effects in a quantum dy-
namical system on the one hand and environmental noise
on the other hand [16, 62, 67–70, 77, 112].
14
A basic example that exhibits the essential nature of
this type of effect consists of a simple three-site network
depicted in Fig.10. Here two sites 1 and 2 are coupled to a
third site which in turn leaks excitations irreversibly into
a reaction center. The coherent interaction is described
by a Hamiltonian
3(cid:88)
2(cid:88)
H =
Eii(cid:105)(cid:104)i +
Jk3(k(cid:105)(cid:104)3 + h.c),
(11)
k=1
k=1
where i(cid:105) corresponds to an excitation in site i and we
assume J13 = J23. Let us begin by considering an exci-
tation initially prepared in the antisymmetric state
ψ(cid:105) =
1√
2
(1(cid:105) − 2(cid:105))
(12)
which forms an eigenstate of this Hamiltonian whose
overlap with the site 3, which we assume to be coupled
dissipatively to a reaction center, vanishes. As a con-
sequence this excitation will remain localized and does
not propagate through the system. Eventually the finite
lifetime of the excitation implies that it will be lost to
the general environment due to a spontaneous annihi-
lation process, an event that is not in the interest of a
transport network. Hence coherence effects may lead to
a strongly reduced or even vanishing transport rate.
One may argue however, that under natural conditions
a pigment-protein complex is not excited in such an an-
tisymmetric state, but will tend to receive a single exci-
tation locally, for example on site 1 (this is for example
the case of the FMO complex [46]). Nevertheless it is
easy to see that the subsequent dynamics has a propen-
sity for leaving the system in an anti-symmetric state or
more generally a state that propagates slowly through
the network due to quantum interference [62]. To this
end note that we can write the initial state localised on
site 1 as an equally weighted coherent superposition of
the symmetric and the anti-symmetric states, i.e.
(cid:18)1(cid:105) − 2(cid:105)√
2
1(cid:105) =
1√
2
(cid:19)
1(cid:105) + 2(cid:105))√
2
+
.
(13)
Thanks to constructive interference a symmetric super-
position experiences a coherently enhanced coupling to
the site 3 to which it will then propagate rapidly, and
from there into the reaction center, while the antisym-
metric part will not evolve at all. Hence, in 50% of the
cases the system will remain in the anti-symmetric state
while in the other 50% of the cases the excitation reaches
the reaction center. Therefore the transfer efficiency is
limited to 50% in this setting.
Now it becomes evident that dephasing noise whose
strength is correctly tuned can have a beneficial effect
FIG. 10. A three-site network in which two sites 1 and 2
are each coupled to a third site 3 via an exchange interaction
of the same strength. Site three is irreversibly connected to
a sink. In (a) the excitation is delocalized over two sites (red
and green) with equal probability of being found at either site
but with a wave function that is antisymmetric with respect
to the interchange of red and green. This state will not evolve
due to destructive interference and hence no excitation will
ever reach the reaction center. In (b) pure dephasing causes
the loss of phase coherence and the two tunneling amplitudes
no longer cancel, eventually leading to a complete excitation
transfer to the sink.
in such situations. Indeed, uncorrelated dephasing noise
acting locally on each site will randomly flip the rela-
tive phase between 1(cid:105) and 2(cid:105) and thus leads to tran-
sitions between the symmetric and the antisymmetric
state. Hence, the presence of dephasing noise inhibits
both constructive and destructive interference and there-
fore slows the propagation of an excitation in a system
initiated in the symmetric state and accelerates the prop-
agation of an excitation in a system initiated in the anti-
symmetric state. As a consequence we expect again that
an intermediate noise level will be optimal, too low it will
not suppress destructive interference efficiently and too
high it will suppress all transport. Noise will be benefi-
cial if the overall propagation under the noisy environ-
ment is still sufficiently rapid to be completed within the
natural lifetime of the excitation. Similar considerations
have also been conjectured independently to play a role
in biological electron transport in photosynthetic reac-
tion centers [113] (note however that recent experimental
and theoretical work suggests that long-lived nuclear vi-
brations are present in the reaction center dynamics and
may have a role to play here too [114–120]).
This is the simplest example for the more general phe-
nomenon that complex quantum networks may possess
subspaces whose members do not propagate and do not
have overlap with the site connected to the reaction cen-
ter. Systems in which such a subspace will be populated
during the time evolution, e.g. because it includes the
site that receives energy, will benefit from environmental
noise which can drive the systems out of the trapping
subspace [62].
It should also be noted that an energy
mismatch and the resulting time-evolution of the rela-
(a) −12ψSite 3 +A+−−=AA(b) 12ρSite 3 Site 2 Site 1 tive phase between the two sites 1 and 2 also leads to
transitions between symmetric and anti-symmetric state
and can thus assists transport [16, 62, 121]. Which of the
two processes if any plays a role in the actual dynamics
of the the FMO complex needs to be determined by care-
fully designed experiments. In both cases the key point
is that some process breaks the symmetry of the system
and thus inhibits the system from getting trapped in an
unfavorable state. It is this principle of breaking sym-
metries that will also play a key role in another sensory
process in nature, namely the magnetic sense of birds as
we consider shortly.
Robustness of excitation energy transport in noisy quantum
networks
As we have seen earlier in Fig. 4, noise does indeed sup-
port transport through realistic transport networks such
as that of the FMO complex. We have identified a vari-
ety of mechanisms for white as well as coloured noise to
lead to these observations. Here we wish to highlight an-
other aspect of noise assisted transport originating from
the broad part of the spectrum, namely that it confers a
certain stability on transport dynamics. In this case, nei-
ther variations of the fine details of the spectral density
nor variations of the electronic structure of the transport
network have a significant effect. As shown in [62, 122]
transport networks that exhibit efficient white noise as-
sisted transport will suffer relatively small variations of
the transport efficiency, in the sub-percent range, even if
the electronic parameters of onsite and coupling energies
vary by up to 20%. Such a stability can be an attrac-
tive feature to ensure robustness in an organism that is
subject to changes in its environment.
Magneto-reception in birds
The effects of weak magnetic field on the growth of
plants as well as the remarkable orientation and navi-
gation abilities of birds, mammals, reptiles, amphibians,
fish, crustaceans and insects are well documented [123].
The mechanism by which this magnetic field sense is
achieved, however, is less well understood and at least
two alternative ideas are being considered, one that ex-
ploits the forces exerted on ferrimagnetic iron oxide par-
ticles embedded in the body, essentially a classical effect,
and another that is based on magnetically sensitive free
radical reactions [17, 18, 124].
Behavioral experiments with birds such as the Euro-
pean robin to study avian magneto-reception [125] have
led to the observation that the process of avian magneto-
reception depends on the wavelength of the ambient light
[126] and can be disrupted by very weak external oscil-
lating magnetic fields [19, 127]. This together with the
15
experimental demonstration of magnetic field effects on a
radical pair reaction at typical earth magnetic fields [128]
provide pieces of evidence to support the idea that the
chemical compass mechanism may be involved in avian
magneto-reception. The cryptochromes in the retina of
migratory birds provide a potential physiological imple-
mentation of such a mechanism [17] which has motivated
the recently growth in attention from quantum physicists
[124, 130–138].
Here we briefly outline the idea of the chemical com-
pass based on the radical pair mechanism and stress that
it represents an example of the crucial role of the in-
terplay between electronic spin quantum dynamics and
the nuclear spin environment of that electron spin (see
Fig. 11 for a schematic representation of the following).
A donor-acceptor pair is initially in its electronic ground
state characterized by a paired electron in a singlet state.
Absorption of a photon induces an electron transfer of
a single electron from the donor to the acceptor thus
creating a radical pair, that is, two molecules with an
unpaired electron each. For simplicity we assume that
the electronic spin state remains unaffected in this step
so that the electrons remain in a singlet state. At this
stage no magnetic field sensitivity can be expected as the
spin singlet state is rotationally symmetric and hence
insensitive to the orientation and magnitude of the ex-
ternal magnetic field. Hence, a further ingredient is re-
FIG. 11. A schematic description of the chemical compass
of avian magneto-reception on the basis of the radical-pair
mechanism. A photon excites two electrons from the ground
state to form a radical pair distributed across two molecules.
The local hyperfine interaction leads to transitions between
singlet and triplet space whose rate is affected by the orien-
tation of the external magnetic field. Reaction products will
depend on the the electrons being in the singlet or triplet
space.
quired to break this symmetry. This ingredient is pro-
vided by the nuclear spin environment of the donor and
Electron transfer D + A D A A D Singlet Triplet Incident Light Magnetic Interaction Singlet products Triplet products + + acceptor molecules. Due to the distance dependence of
the dipolar interaction between electron and nuclear spin
the unpaired electrons on donor and acceptor see domi-
nantly uncorrelated interactions which induce symmetry-
breaking transitions from the singlet to the triplet man-
ifold. The Hamiltonian of the radical-pair plus nuclear
environment is given by
(cid:88)
(cid:88)
H =
siTijIij − gµBB(sA + sD)
(14)
i=D,A
j
where Iij and si are the nuclear and electron spin oper-
ators respectively while Tij denotes the hyperfine cou-
pling tensor. B denotes the external magnetic field, µ0
the Bohr magneton and g the gyromagnetic ratio. These
transitions and the resulting time varying population of
the singlet and the triplet manifold will now be dependent
of the relative orientation of the molecules with respect
to the external magnetic field via the anisotropy of the
hyperfine interaction as well as the strength of the mag-
netic field as it dictates the energy splitting in the triplet
manifold.
Relative population differences between singlet and
triplet manifolds result in different rates of occurrence of
chemical products. A donor acceptor pair in the singlet
state may either recombine to reach the groundstate or
lead to a reaction product at rate kS. Spins in the triplet
state cannot recombine to reach the ground state but may
lead to different reaction products at rate kT . Hence the
amount and nature of reaction products will depend on
the relative populations of the singlet and triplet man-
ifolds whose time evolution is usually described by a
simple phenomenological quantum master equation (the
Haberkorn approach [129])
d
dt
ρ = − i
[H, ρ]−kS(QSρ+ρQS)−kT (QT ρ+ρQT ) (15)
(cid:82) tr[ρ(t)QS]dt.
where QS (QT ) are the projectors onto the singlet
(triplet) subspace. The singlet yield is then determined
as kS
The fact that local interactions between electron spins
and their nuclear environment lead to a breaking of the
symmetry of the electronic spin state gives a first indica-
tion of the importance of the presence of local environ-
ments. It does not however reveal the full depth of the
role of the system-environment interaction. Indeed, con-
trary to the assumption of early studies, it is not merely
the coherence properties of the electron spin state that
determine the sensitivity of the chemical compass. To
gain a full understanding of the chemical compass and
the role of coherence one must study the entire system
composed of electron spins and nuclear spins [138].
It
turns out that it is the coherence of the initial state (elec-
tronic singlet state and unpolarized nuclear spins) in the
basis made up of the eigenstates of the full system in the
absence of a magnetic field that has predictive power of
16
the sensitivity of the chemical compass. The principal
reason is down to the fact that the coherences in sys-
tem and environment that are present initially will accu-
mulate phase factors due to the presence of an external
magnetic field which in turn will be seen, when observing
the electron spin only, as oscillations between singlet and
triplet states. Indeed, a large amount of initial coherence
in this picture correlates strongly with high sensitivity of
the chemical compass and the impact of different types
of decoherence can be predicted from it. Hence it is only
when studying the properties of system and environment
that we can gain and understanding of the chemical com-
pass and an opportunity to predict optimal designs [138].
This example brings out quite clearly that it is the co-
operation of electron spin quantum dynamics on the one
hand and the interaction with the nuclear spin environ-
ment of these electrons on the other hand that lead to
the magnetic field sensitivity of the radical pair mecha-
nism. Without the environment no magnetic field sensi-
tivity can be expected. Therefore a more refined picture
emerges when the structure of the nuclear spin environ-
ment is taken account of by including it in the system
dynamics. From this viewpoint the chemical compass
can be seen as an environment assisted quantum inter-
ferometer thus affirming avian magneto-reception as an
example of quantum biology that benefits from an inter-
play of system-environment interaction [138].
Structure adaption
In the preceding section we have presented mechanisms
by which the quantum dynamics of electronic systems can
enter a beneficial interplay with the dynamics originating
from its interaction with an environment. The principles
expounded here are more general though and apply to
any quantum network whose task may include for exam-
ple transport or sensing and that is in contact with an
environment of charges, spins or vibrations. We have also
explained that it is likely that nature has optimized both
the properties of its quantum networks and the structure
of the environment to achieve optimal performance.
Here we would like to summarize briefly some of the
means that nature has available for achieving this struc-
tural adaptation and tuning for optimality. There are
two principal aspects that can be controlled, the elec-
tronic network and its environment. While in many re-
spects it appears easier to achieve considerable changes
in the electronic structure as compared to the vibrational
environment it is likely that both will have been subject
to evolutionary adaptation.
Position & Orientation:
Interaction strength – The
protein can arrange molecules, such as chromophores,
controlling their respective distances as well as the rela-
tive orientation of their dipole-moments. Thanks to the
1/r3 distance dependence of the dipolar interaction and
its angular dependence 3(Di · rij)(Dj · rij) − (Di · Dj),
where Di is the optical dipole moment and rij the unit
vector connecting the two molecules, significant changes
of the interaction strength and even its sign can be
achieved.
Static properties of Local Environments: Onsite exci-
tation energies – In vacuum two realizations of the same
molecule, such as chlorophyll are identical in all its as-
pects. In a protein scaffold however their properties can
vary considerably. These variations are being controlled
by the local environment such as the presence and dis-
tance of partial charges or even free charges that will
affect the local excitation energy of a molecule through
their electrostatic interaction. These changes can be sig-
nificant and can reach the range of 100(cid:48)s of wavenumbers
or more even for very small distance changes in the en-
vironment. For larger structural changes of the protein
scaffold, e.g. due to variations in pH-value or other more
drastic changes the shift in wavelength can even reach
dozens of nanometers. Thus the arrangement of the lo-
cal environment is of crucial importance for the definition
of the electronic properties of a biological quantum net-
work.
Dynamics of Local Environments: Dephasing noise –
The structure of the local environment plays another cru-
cial role as the motion of partial charges or even free
charges will lead to fluctuations in the excitation energy
of the nearby sites. It is this process for example that
couples the vibrational motion of proteins and molecules
to their electronic energies and is thus the cause of both
dephasing and the interaction of long-lived vibrational
modes with electronic degrees of freedom.
Vibrational and spin environments – The spectral den-
sity describing the interaction of electronic degrees of
freedom and their environment is a combination of two
aspects, both of which may be tuned. Firstly, the cou-
pling strength of individual environmental modes to the
system which in turn depends on the amplitude of the
motion of charges due to the vibration relative to the
site of interest and the magnitude and number of these
charges.
Secondly, the density of vibrational modes
which is determined by the quasi-continuous mode den-
sity of the protein as well as the discrete modes provided
by the specific molecules that are realizing the quantum
network. For spin environments, e.g. the nuclear spin
environment in magneto-reception, the structure of the
relevant molecule and the form of the electronic wave-
function determines the strength and anisotropy of the
hyperfine interaction.
PC card principle – Besides the adjustment of the
broad vibrational spectrum of the protein a key principle
is the ability to affect the discrete part of the vibrational
spectra by the choice of the molecule that realizes the
quantum network. Each molecule will be characterized
by a set of vibrational modes of varying lifetime. De-
pending on the fit of the molecule in the protein scaffold
17
the lifetime and coupling of these modes and thus their
contribution to the overall spectral density will be af-
fected.
In this way, it is conceivable that similar to a
PC card the insertion of a specific molecule may allow
for certain functionality to be achieved. Likewise the
phonon assisted tunneling process in olfaction suggests
that this process may be used to recognise molecules and
to initiate charge transfer processes.
Enhancing or decreasing noise – The presence of en-
vironmental noise may or may not be of advantage for a
biological system. Certainly, strong featureless noise can
affect the ability to act and react specifically to external
input. Hence, in various circumstances it may be essen-
tial to reduce the level of noise for example by protecting
a molecule inside a protein structure or a hydrophobic
binding pocket. On the other hand it may on occasion
be beneficial to increase the level of noise by arranging
easily movable charges close to a site which will respond
strongly to vibrations and thus create significant changes
in the excitation energy of the site.
These are some of the principal means by which bio-
logical systems may affect both their quantum dynam-
ical networks as well as the environment around them
to achieve optimal performance. This suggests that in-
deed nature has at its hand a wide variety of possibilities
for tuning quantum networks and their environments to
achieve robust and efficient devices.
NUMERICAL DESCRIPTION OF QUANTUM
DYNAMICS IN STRUCTURED
ENVIRONMENTS
One of the key messages of the preceding sections is the
special role of the interplay between the quantum dynam-
ics of a system and its environment, in particular when
this environment does not merely represent white noise
but possesses structure. Furthermore,
it has become
clear that the optimal operating regime for quantum bio
dynamics tends to favour parameter ranges in which the
interaction between system components is comparable to
the interaction of these system components and their en-
vironment.
Both features are not well modeled by the traditional
perturbative treatments that lead to master equations
of Lindblad [139, 140], Redfield [141] or modified Red-
field type [46, 142, 143] (see [60] for a review of master
equations). Indeed the essential importance of the non-
Markovian nature of the system-environment interaction
calls for the development of non-perturbative methods
that can accurately, certifiably and efficiently model the
resulting dynamics.
In recent years a wide variety of
methods has been developed with the aim of address-
ing some or all of these issues. In the following we will
present the common theoretical setting in which these
methods are formulated and then briefly outline key as-
pects of these methods. Details can be intricate and will
be left for further reading.
Standard model of open-quantum system
We begin by defining more formally a standard model
of an open quantum system as it is relevant to many of
the biological settings that we have discussed so far. In
this model the quantum system interacts with a macro-
scopic number of environmental degrees of freedom and
the total system and environment state evolves under a
purely unitary dynamics. This model is of considerable
importance for the description of a wide variety of biolog-
ical environments. Dissipation and decoherence appear
when the system is observed without any knowledge of
the state of its environment, leading to a non-unitary
effective dynamics for the sub-system’s reduced density
matrix. The total Hamiltonian of system and environ-
ment can be written as
H = HS + HE + HI ,
(16)
where HS is the Hamiltonian that refers to the degrees
of freedom of the open quantum system of interest, HE
is the Hamiltonian describing the free evolution of the
environment while HI describes the interaction between
the open quantum system and its environment.
The environment is assumed to be well described by
a continuum of harmonic oscillator degrees of freedom
labeled by some real number. The internal dynamics of
these harmonic oscillators is described by bosonic modes
and is thus given by the Hamiltonian
(cid:90) xmax
HE =
0
dx g(x)a†
xax.
(17)
In a physical context x would usually represent some con-
tinuous real variable such as the frequency or momentum
of each mode while xmax is the maximum value that this
variable may assume. The creation and annihilation op-
erators satisfy the continuum bosonic commutation rules
y] = δ(x− y) with the Dirac delta function δ(x− y).
[ax, a†
We have adopted a continuum description of the envi-
ronment but discrete environments can be treated in the
same way by replacing the integral by a summation.
The internal dynamics of the system is described by a
completely general Hamilton operator HS. For the in-
teraction between the system and its environment we
assume a linear coupling between arbitrary system op-
erators and the position operator of the environment
dxh(x) A(ax + a†
x),
(18)
(cid:90) xmax
HI =
0
where A is an arbitrary operator of the open quantum
system and the (real) coupling function h(x) describes
18
(cid:90) xmax
(cid:90) xmax
the coupling strength with each mode. It is convenient,
though not essential, to consider g(x) = x [144]. In that
case the total Hamiltonian for system plus environment
is given by
0
dx xa†
xax +
0
H = HS +
dxh(x) A(ax + a†
x).
(19)
For Gaussian initial states of the environment [145], the
dynamics induced in the quantum system S by its inter-
action with the environment is determined by the spec-
tral density J(ω) [146, 147]. For the continuum model
of the reservoir that we are considering, i.e. choosing
g(x) = x, this function is given by,
J(ω) = πh2(ω).
(20)
The spectral function thus describes the overall strength
of the interaction of the system with the reservoir modes
of frequency ω.
Despite the relative simplicity of this model its dy-
namics is not exactly solvable and one has to resort to
numerical methods. This is particularly so, when the
environmental spectral density does exhibit considerable
structure or when the coupling strength between system
and environment lies in the non-perturbative regime. In
the following we discuss a variety of methods that have
been developed in recent years to treat this situation.
Desiderata
There is a wide variety of simulations methods and in
order to evaluate them it will be helpful to first establish
some properties that one may wish a method to possess.
Such desiderata for the method include (i) that it is ef-
ficient in the system size, (ii) it is able to take account
of arbitrary spectral densities without losing efficiency or
having to redesign the protocol, (iii) it has the ability to
deal with both high, intermediate and low temperatures,
and (iv) it should be certifiable, that is it possesses a
known and controllable error.
Time adaptive renormalization group methods
We begin by presenting an approach to the system-
environment interaction that preserves the full informa-
tion about the environment state, is able to treat ar-
bitrary spectral densities and coupling strengths within
a single unified framework with known and controllable
error and is computationally efficient in the size of the
environment.
The key idea of this method which was developed in
[87–89] is simple but effective. It begins by mapping the
spin-boson model exactly onto a 1D system thus permit-
ting the deployment of the time-adaptive density matrix
renormalization group (t-DMRG) technique to integrate
the time evolution of the full system-environment dynam-
ics efficiently.
We begin by demonstrating that a system linearly cou-
pled to a reservoir characterized with a spectral density
J(ω) is unitarily equivalent to a semi-infinite chain with
only nearest-neighbors interactions, where the system it-
self only couples to the first site in the chain (see figure
12) [148].
In other words, there exists a unitary oper-
ator Un(x) such that the countably infinite set of new
operators
dx Un(x)a†
x, a†
x =
Un(x)b†
n
(21)
(cid:90) xmax
b†
n =
0
(cid:88)
n
∞(cid:88)
satisfies the bosonic commutation relations [bn, b†
δnm and leads to the transformed Hamiltonian
m] =
H(cid:48) = HS+c0 A(b0+b
†
0)+
ωnb†
†
n+1bn+b†
nbn+tn(b
nbn+1),
where c0, tn, ωn are real constants. The proof of this
n=0
FIG. 12. Illustration of the effect of the transformation Un(x).
On the left side a central system is interacting with a reservoir
with continuous bosonic or fermionic modes, parametrized by
x, after Un(x) the system is the first place of a discrete semi-
infinite chain parametrized by n.
statement becomes quite direct after one key observation.
Since J(ω) is positive, we can always define the measure
dµ(x) = h2(x)dx and write the unitary
Un(x) = h(x)pn(x)
(22)
where pn(x) are some set of real orthonormal polynomi-
als with respect to the measure dµ(x) = h2(x)dx with
support on [0, xmax] [149]. Employing p0(x) = 1 we im-
mediately find
HI = A(b0 + b
†
0).
HE in the new basis is obtained by employing the fact
that orthogonal polynomials satisfy the recursion relation
19
pk+1(x) = (Ckx− Ak)pk(x)− Bk pk−1(x) for k = 0, 1, 2...
and p−1(x) ≡ 0. We find
b†
nbn+1 +
[
1
Cn
b†
nbn +
An
Cn
Bn+1
Cn+1
†
n+1bn]
b
(cid:88)
HE =
n
with [88]
ωn = An/Cn,
tn = Bn+1/Cn+1 = 1/Cn.
This approach provides us with a way to construct an ex-
act mapping, one has just to look for a family of orthog-
onal polynomials with respect to the measure dµ(x) =
h2(x)dx. This can be done analytically in some impor-
tant cases (see [88] for examples), however even if the
weight h2(x) is a complicated function, families of or-
thogonal polynomials can be found by using very sta-
ble numerical algorithms such as the ORTHPOL package
[150].
The reformulation of the system-environment interac-
tion in the chain picture allows us to apply t-DMRG
techniques quite directly. As it can be proven for a
wide variety of spectral densities, in the large n limit
the chain parameters ωn to tn converge and their ra-
tio approaches ωn/tn → 2. A translationally invariant
harmonic chain with this ratio ωn/tn has a gapless dis-
persion and excitations can escape down the chain with-
out being scattered back towards the system (see Fig.
13 for illustration). Thus the buildup of excitations in
any region of the chain is limited over time, and corre-
lations can be expected to be bounded. This in turn
suggests that t-DMRG will provide an accurate descrip-
tion for long times [151]. The universal asymptotics
of environments revealed by this analysis are discussed
in [88]. Practical experience, though not mathematical
proof, suggests that the method itself appears to satisfy
all the desiderata listed above.
Indeed it has been es-
sential in the accurate numerical modeling dynamics of
complex environments which explain the emergence of
long-lived quantum coherence in photosynthetic complex
[28] as explained earlier. It should be noted however that,
in contrast to some other methods, it becomes computa-
tionally less efficient with increasing temperature. On the
other hand it easily incorporates time-dependent pertur-
bations of the system, e.g. via laser pulses and it does
preserve all the information about the environment and
is thus capable to following the dynamics after the exci-
tation of complex wave-packets in the environment.
Hierarchy equations of motion
Recently, the numerical hierarchy technique [152–154],
which has a longstanding history [155–157], has received
renewed attention in the context of excitation energy
transfer across pigment-protein complexes. This ap-
proach is non-perturbative and is capable of interpolat-
ing, for example, between the Bloch-Redfield and the
)(xUnbosonSpinnnt 20
result does not change significantly any more. Neverthe-
less, it should be noted, that the hierarchy method has
been applied successfully to a dimer [152, 154] as well as
the seven-site FMO complex [154, 159, 160] with a Brow-
nian harmonic oscillator spectral density of the environ-
ment. Note also recent numerically intensive numerical
calculations of 2-D spectra [161] using this method and
recent algorithmic improvements [162].
Path integral techniques
A variety of other methods to study transport pro-
cesses and in consequence also transport in noisy environ-
ments have been developed in condensed matter physics.
These methods represent various approaches for finding
numerically the formal path integral solution of the time
evolution. These include the quasi-adiabatic path inte-
gral approach [163–167] and the iterative summation of
real-time path integrals [168] to name just two. These
procedures are expected to give good results in the high
temperature limit and hence short correlation time of
the environments. With decreasing temperatures and
thus increasing correlation time the computational ef-
fort grows rapidly. Temperatures not too far below the
typical system frequencies appear to be accessible [169].
For highly structured environments in which for exam-
ple both narrow and broad features are combined these
methods find challenges.
In this case, sharply peaked
modes can be added to the system and their damp-
ing is treated in the bath [170] but such an approach
will be challenging for the treatment of quantum net-
works when addition of modes make the network itself
too high-dimensional for numerical treatment. Path in-
tegral methods have been applied mostly to dimers (see
e.g.[166]) and their scaling to larger systems, just as for
the transformation approach to be discussed below, re-
mains to be demonstrated. A comparison of the hierar-
chy and the path integral methods has been carried out
recently for a Brownian harmonic oscillator environment
for which both methods are expected to be applicable
[171].
Other numerical methods
Many more methods exist which include methods
based on a combination of the polaron transformation
and a variational ansatz [172–174], schemes based on lin-
earization of the environment dynamics [175–178], semi-
classical methods in which the interaction between the
system and its vibrational environment are replaced by
mean-field type terms thus obviating the need for ten-
sor product structures [28], schemes based on time-
convolutionless master equations [179, 180], methods
based on quantum state diffusion [181, 182] . For a recent
FIG. 13. Illustrative sketch of open-system dynamics in the
chain representation. (a) Subsystem initially injects excita-
tions (shown as wave packets) into inhomogenous region of
the chain. Scattering from inhomogeneity causes back ac-
tion of excitations on the system at later times and leads to
memory effects and non-Markovian subsystem dynamics. (b)
At long times, after multiple scattering, excitations penetrate
into the homogenous region and propagate away from the
system without backscattering. This leads to irreversible and
Markovian excitation absorption by the environment.
Forster regimes [152, 154]. It derives a hierarchy of equa-
tions in which the reduced density operator of the system
couples to auxiliary operators which in principle allow for
the simulation of complex environments. The depth of
this hierarchy and the structure of its coefficients depend
on the correlation time of the bath and its spectral den-
sity. This approach appears sufficiently flexible to take
account of spatial correlations in the noise as well. For
specific choices of the bath spectral density, namely the
Brownian harmonic oscillator and/or high temperatures
the temporal bath correlations decay exponentially so
that the hierarchy can be terminated early with small but
uncontrolled error and one obtains a manageable struc-
ture of the hierarchy. An estimate suggests that in this
case the set of operators in the hierarchy will scale at
least proportional τ k where τ is the correlation time and
k is the number of sites in the system to be studied (see
[158] for a more detailed discussion). More complex spec-
tral densities will lead to considerably more demanding
evaluations of the coefficients of the hierarchy. A non-
exponential decay of the temporal bath correlations also
leads one to estimate an exponential growth of the num-
ber of operators in the hierarchy. It is possible that spe-
cific numerical simulations turn out to be more efficient
than those estimates suggest but there is no certificate
that provides error bounds. In fact, it is a challenge to
determine the errors introduced by the various approxi-
mation steps that are involved in the numerical hierarchy
technique. Hence one has to test for convergence empir-
ically by increasing the depth of the hierarchy until the
(a) (b) Contemporary Physics
review see of simulation methods see [183].
SUMMARY & CONCLUSIONS
The preceding sections have identified and discussed
In this con-
questions of interest to quantum biology.
text the unavoidable presence of partially uncontrolled
environments in biology, often perceived as noise, has led
us to recognize a theme of fundamental importance that
is central to the study of quantum effects in biology –
the interplay between quantum coherent dynamics of a
system on the one hand and the interaction of this very
system with its environment on the other.
It is this unavoidable lack of isolation that provides the
boundary conditions under which natural evolution had
to operate and therefore it may, in hindsight, not be sur-
prising that nature has found solutions in which optimal
biological quantum dynamics tends to be achieved in a
regime where the interaction within the quantum system
are of the order of its interaction with the environment so
that both contributions do not merely coexist but enter
a fruitful interplay.
That this is so, is not an accident. It can be understood
from simple and generalizable principles that we have
identified in our discussions to explain that too much
or too little coherence can be detrimental and that it is
in fact natural to expect that there is an intermediate
regime that is optimal.
Moreover, we argued, biological systems have available
a variety of tools that enable them to achieve these op-
timal regimes in a controlled fashion. Indeed, by the use
of protein structure to arrange molecules and their local
environment they are capable of tuning the properties
of transport or sensory networks and, at the same time,
adjust the environment of these networks by providing
isolation or for example by inserting specific molecules
with desirable vibrational or spin properties to fashion
an environment. It is this toolbox that allows for mutual
tuning through evolutionary adaptation which can then
be used to achieve optimal performance whose origin we
understand from generalizable design principles.
The importance of these design principles goes be-
yond merely understanding what has been created al-
ready but also paves the way by which these principles,
when spelled out and made quantitative, can allow for
the rational design of optimal structures as well as the
execution of optimized experiments by which to amplify
and verify quantum effects in biology.
We have followed this path in the preceding sections
and have determined and discussed such design princi-
ples and have then applied them to bring under one um-
brella the phenomena of photosynthesis, olfaction and
avian magneto-reception. We hope that guided by the
considerations and principles presented here we will be
able to add to photosynthesis, avian magneto-reception
21
and olfaction, the three clouds at the otherwise blue sky
of classical biology, and discover many more biological
phenomena for which quantum effects are of fundamen-
tal importance and thus come to be seen as the seeds
from which a rich phenomenology of quantum effects in
biology may grow.
ACKNOWLEDGEMENTS
Over the course of the last few years we have benefitted
from conversations and collaboration on topics of quan-
tum biology with a wide variety of researchers, in partic-
ular J. Almeida, A. Aspuru-Guzik, A. Bayat, J.M. Cai,
T. Calarco, F. Caruso, F. Caycedo-Soler, A. W. Chin, A.
Datta, M. del Rey, G. S. Engel, F. Jelezko, R. Ghosh, S.
Montangero, A. Olaya-Castro, H. Plenio, J. Prior, Th.
Renger, E. Romero, E. Solano, L. Turin, R. van Gron-
delle and A. Vaziri. We gratefully acknowledge support
by the EU STREP project PAPETS, the ERC Synergy
grant BioQ and an Alexander von Humboldt Professor-
ship.
[1] P. Jordan, Die Physik und das Geheimnis des organis-
chen Lebens, Friedrich Vieweg & Sohn, Braunschweig
1943.
[2] M. Mohseni, Y. Omar, G.S. Engel, and M.B. Plenio
(eds.), Quantum effects in biology, Cambridge Univer-
sity Press, Cambridge 2013.
[3] Proceedings of the 22nd Solvay Conference in Chem-
istry on ”Quantum Effects in Chemistry and Biology,
Procedia Chemistry 3 (2011).
[4] M.B. Plenio and S.F. Huelga, Entangled light from white
noise, Phys. Rev. Lett. 88 (2002), 197901.
[5] L. Hartmann, W. Dur and H.J. Briegel, Steady state en-
tanglement in open and noisy quantum systems at high
temperature, Phys. Rev. A 74 (2006), 052304.
[6] D. DeVault and B. Chance, Studies of photosynthesis
using a pulsed laser I. Temperature dependence of cy-
tochrome oxidation rate in Chromatium. Evidence for
tunneling, Biophys J 6 (1966), pp. 825-847.
[7] D. DeVault, Quantum mechanical tunneling in biological
systems, Quart. Rev. Biophys. 13 (1980), pp. 387 - 564.
[8] Y. Cha, C.J. Murray and J.P. Klinman, Hydrogen tun-
neling in enzyme-reactions, Science 243 (1989), pp. 1325
- 1330.
[9] J. Kempe, Quantum random walks - an introductory re-
view, Contemp. Phys. 44 (2003), pp. 307 - 327.
[10] G.D. Scholes, G.R. Fleming, A. Olaya-Castro, and R.
van Grondelle, Lessons from nature about solar light
harvesting, Nature Chem. 3 (2011), pp. 763 - 774.
[11] G.S. Engel, T.R. Calhoun, E.L. Read, T.K. Ahn, T.
Mancal, Y.C. Cheng, R.E. Blankenship, and G.R. Flem-
ing, Evidence for wavelike energy transfer through quan-
tum coherence in photosynthetic systems, Nature 446
(2007), pp. 782 - 786.
Contemporary Physics
22
[12] I.P. Mercer, Y.C. El-Taha, N. Kajumba, J.P. Maran-
gos, J.W.G. Tisch, M. Gabrielsen, R.J. Cogdell, E.
Springate, and E. Turcu, Instantaneous mapping of co-
herently coupled electronic transitions and energy trans-
fers in a photosynthetic complex using angle-resolved co-
herent optical wave-mixing, Phys. Rev. Lett. 102 (2009),
057402.
[13] G. Panitchayangkoon, D. Hayes, K.A. Fransted, J.R.
Caram, E. Harel, J.Z. Wen, R.E. Blankenship, and
G.S. Engel, Long-lived quantum coherence in photosyn-
thetic complexes at physiological temperature, Proc. Nat.
Acad. Sci. Am. 107 (2010), pp. 12766 - 12770.
[14] E. Collini, C. Wong, K. Wilk, P. Curmi, P. Brumer, and
G. Scholes, Coherently wired light-harvesting in photo-
synthetic marine algae at ambient temperature, Nature
463 (2010), pp. 644 - 649.
[15] M. Mohseni, P. Rebentrost, S. Lloyd, and A. Aspuru-
Guzik, Environment-Assisted Quantum Walks in Pho-
tosynthetic Energy Transfer, J. Chem. Phys. 129 (2008),
174106.
[16] M.B. Plenio and S.F. Huelga, Dephasing assisted trans-
port: Quantum networks and biomolecules, New J.
Phys. 10, (2008), 113019.
[17] T. Ritz, S. Adem, and K. Schulten, A model
for
photoreceptor-based magnetoreception in birds, Biophys.
J. 78 (2000), pp. 707 - 718.
[18] K. Schulten, C. E. Swenberg, and A. Weller, A biomag-
netic sensory mechanism based on magnetic field mod-
ulated coherent electron spin motion, Z. Phys. Chem.
NF111 (1978), pp. 1 - 5.
[19] T. Ritz, P. Thalau, J.B. Phillips, R. Wilschko, and
W. Wilschko, Resonance effects indicate a radical-pair
mechanism for avian magnetic compass, Nature 429
(2004), pp. 177 - 180.
[20] L. Turin, A spectroscopic mechanism for primary olfac-
tory reception, Chem. Sens. 21 (1996), pp. 773 - 791.
[21] M.I. Franco, L. Turin, A. Mershin, and E.M.C.
Skoulakis, Molecular vibration-sensing component in
Drosophila melanogaster olfaction, Proc. Natl. Acad.
Sci. Am. 108 (2011), pp. 3797 - 3802.
[22] D.J. Wineland, J.J. Bollinger, W.M. Itano, F.L. Moore
and D.J. Heinzen, Spin squeezing and reduced quantum
noise in spectroscopy, Phys. Rev. A 46 (1992), pp. R6797
- R6800.
[23] S.F. Huelga, C. Macchiavello, T. Pellizzari, A.K. Ekert,
M.B. Plenio and J.I. Cirac, Improvement of frequency
standards with quantum entanglement, Phys. Rev. Lett.
79 (1997), pp. 3865 - 3868.
[24] J.P. Keating, N. Linden, J.C.F. Matthews, and A. Win-
ter, Localization and its consequences for quantum walk
algorithms and quantum communication, Phys. Rev. A
76 (2007), 012315.
[25] The environmental spectral density is a combination of
mode density in the environment and the strength of
interaction between the system and each of these modes
[146].
[26] F. Caycedo-Soler, A.W. Chin, J. Almeida, S.F. Huelga,
and M.B. Plenio, The nature of the low energy band
of the Fenna-Matthews-Olson complex: vibronic signa-
tures, J. Chem. Phys. 136 (2012), 155102.
[27] M. Ratsep, Z.-L. Cai, J.R. Reimers and A. Freiberg,
Demonstration and interpretation of significant asym-
metry in the lowresolution and high-resolution Qy fluo-
rescence and absorption spectra of bacteriochlorophyll a,
J. Chem. Phys. 134 (2011), 024506.
[28] A. W. Chin, J. Prior, R. Rosenbach, F. Caycedo-Soler,
S. F. Huelga, and M. B. Plenio, Vibrational structures
and long-lasting electronic coherence, Nature Phys. 9
(2013), pp. 113 - 118.
[29] J.D. Biggs and J.A. Cina, Studies of Impulsive Vi-
brational Influence on Ultrafast Electronic Excitation
Transfer, J. Phys. Chem. A 116 (2012), pp. 1683 - 1693.
[30] G.H. Richards, K.E. Wilk, P.M.G. Curmi, H.M. Quiney,
and J.A. Davis, Excited state coherent dynamics in light-
harvesting complexes from photosynthetic marine algae,
J. Phys. B 45 (2012), 154015.
[31] J.M. Womick and A.M. Moran, Exciton Coherence and
Energy Transport in the Light-Harvesting Dimers of Al-
lophycocyanin, J. Phys. Chem. B 113 (2009), pp 15747
- 15759.
[32] J.M. Womick and A.M. Moran, Vibronic Enhancement
of Exciton Sizes and Energy Transport in Photosyn-
thetic Complexes, J. Phys. Chem. B 115 (2011), pp. 1347
- 1356.
[33] J.M. Womick, B.A. West, N.F. Scherer and A.M.
Moran, Vibronic effects in the spectroscopy and dynam-
ics of C-phycocyanin, J. Phys. B. 45 (2012), 154016.
[34] A.W. Chin, S.F. Huelga, and M.B. Plenio. Coherence
and Decoherence in Biological System: Principles of
Noise Assisted Transport and the origin of Long-lived
Coherences. Phil. Trans. Roy. Soc. A 370 (2012), pp.
3638 - 3657.
[35] A. Kolli, E.J. O’Reilly, G.D. Scholes, A. Olaya-Castro,
The fundamental role of quantized vibrations in coherent
light harvesting by cryptophyte algae, J. Chem. Phys.
137 (2012), 174109.
[36] V. Tiwari, W.K. Peters and D.M. Jonas, Electronic
resonance with anticorrelated pigment vibrations drives
photosynthetic energy transfer outside the adiabatic
framework, PNAS 110 (2013), 1203 - 1208.
[37] J. Almeida, S.F. Huelga, and M.B. Plenio, Long-
lived oscillatory signals in electronic 2-D spectra from
exciton-vibrational interaction, E-print arXiv:1307.xxxx
[38] N. Christensson, H.F. Kauffmann, T. Pullerits, and T.
Mancal, Origin of long-lived coherences in light harvest-
ing complexes, J. Phys. Chem B 116 (2012), pp. 7449 -
7454.
[39] E. Harel, A.F. Fidler and G.S. Engel, Real-time mapping
of electronic structure with single-shot two-dimensional
electronic spectroscopy, Proc. Nat. Acad. Sci. 107
(2010), pp. 16444 - 16447.
[40] F. Caruso, S.K. Saikin, E. Solano, S.F. Huelga, A.
Aspuru-Guzik, and M.B. Plenio, Probing biological
light-harvesting phenomena by optical cavities, Phys.
Rev. B 85 (2012), 125424.
[41] F. Caruso, S. Montangero, T. Calarco, S.F. Huelga, and
M.B. Plenio, Coherent open-loop optimal control of light
harvesting dynamics, Phys. Rev. A 85 (2012), 042331.
[42] S. Hoyer, F. Caruso, S. Montangero, Mohan Sarovar,
T. Calarco, M.B. Plenio, and K.B. Whaley, Coher-
ently controlled preparation and verification of excitonic
states in a light harvesting complex by ultrafast spec-
troscopy with shaped pulses, arXiv:1306.xxxx
[43] E.J. Candes and M.B. Wakin, An Introduction To Com-
pressive Sampling, IEEE Sig. Proc. Mag. 25 (2008), pp.
21 - 30.
[44] J. Almeida, J. Prior and M.B. Plenio, Computation of
2-D spectra assisted by compressed sampling, J. Phys.
Contemporary Physics
23
Chem. Lett. 3 (2012) 2692 - 2696.
Chem. Lett. 4 (2013), pp. 903 - 907.
[45] J. N. Sanders, S. Mostame, S. K. Saikin, X. Andrade,
J. R. Widom, A. H. Marcus, A. Aspuru-Guzik, Com-
pressed sensing for multidimensional electronic spec-
troscopy experiments, J. Phys. Chem. Lett. 3 (2012)
2697 - 2702.
[46] J. Adolphs and T. Renger, How proteins trigger excita-
tion energy transfer in the FMO complex of green sulfur
bacteria, Biophys. J. 91 (2006), pp. 2778 - 2797.
[47] M.S.A. Busch, F. Muh, M.E. Madjet, and T. Renger,
The Eighth Bacteriochlorophyll Completes the Excita-
tion Energy Funnel in the FMO Protein, J. Phys. Chem.
Lett. 2 (2011), pp. 93 - 98.
[48] H. van Amerongen, L. Valkunas and R. van Grondelle,
Photosynthetic Excitons, World Scientific 2000.
[49] A.S. Davydov, The Theory of Molecular Excitons, Sov.
Phys. Usp. 7 (1964), pp. 145 - 178.
[50] M. Wendling, T. Pullerits, M.A. Przyjalgowski, S.I.E.
Vulto, T.J. Aartsma, R. van Grondelle and H. van
Amerongen, Electron-vibrational coupling in the Fenna-
Matthews-Olson complex of Prosthecochloris aestuarii
determined by temperature-dependent absorption and
fluorescence line-narrowing measurements, J. Phys.
Chem. B 104 (2000), pp. 5825 - 5831.
[51] V. May and O. Kuhn, Charge and Energy Transfer Dy-
namics in Molecular Systems, Wiley-VCH Verlag 2004
[52] J. Strumpfer and K. Schulten, The effect of correlated
bath fluctuations on exciton transfer, J. Chem. Phys.
134 (2011), pp. 095102.
[53] J. Lim, M. Tame, K.H. Yee, J.-S. Lee and J. Lee,
Phonon-induced dynamic resonance energy transfer, E-
print arXiv:1304.3967.
[54] C. Olbrich, J. Strumpfer, K. Schulten, and U.
Kleinekathofer, Theory and Simulation of the Environ-
mental Effects on FMO Electronic Transitions, J. Phys.
Chem. Lett. 2 (2011), pp. 1771 - 1776.
[55] C. Olbrich, T.L.C. Jansen, J. Liebers, M. Aghtar,
J. Strumpfer, K. Schulten, J. Knoester, and U.
Kleinekathofer, From Atomistic Modeling to Excita-
tion Transfer and Two-Dimensional Spectra of
the
FMO Light-Harvesting Complex, J. Phys. Chem. B 115
(2011), pp. 8609 - 8621.
[56] C. Olbrich, J. Strumpfer, K. Schulten, and U.
Kleinekathofer, Quest for Spatially Correlated Fluctu-
ations in the FMO Light-Harvesting Complex, J. Phys.
Chem. B 115 (2011), pp. 758 - 764.
[57] S. Shim, P. Rebentrost, S. Valleau, and A. Aspuru-
Guzik, Atomistic study of the long-lived quantum coher-
ences in the Fenna-Matthews-Olson complex, Biophys.
J. 102 (2012), pp. 649-660.
[58] T. Renger, A. Klinger, F. Steinecker, M. Schmidt am
Busch, J. Numata and F. Muh, Normal Mode Analy-
sis of the Spectral Density of the Fenna-Matthews-Olson
Light-Harvesting Protein: How the Protein Dissipates
the Excess Energy of Excitons, J. Phys. Chem. B 116
(2012), pp. 14565 – 14580.
[59] T. Renger and F. Muh, Understanding photosynthetic
light-harvesting: a bottom up theoretical approach, Phys.
Chem. Chem. Phys. 15 (2013), pp. 3348 - 3371.
[60] A. Rivas and S.F. Huelga, Open Quantum Systems – An
Introduction, Springer Briefs in Physics 2012
[61] M. del Rey, A.W. Chin, S.F. Huelga, and M.B. Ple-
nio, Exploiting structured environments for efficient en-
ergy transfer: The phonon antenna mechanism, J. Phys.
[62] F. Caruso, A.W. Chin, A. Datta, S.F. Huelga, and
M.B. Plenio, Highly efficient energy excitation transfer
in light-harvesting complexes: The fundamental role of
noise-assisted transport , J. Chem. Phys. 131 (2009),
105106.
[63] M.B. Plenio and V. Vedral, Teleportation, Entangle-
ment and Thermodynamics in the Quantum World,
Contemp. Phys. 39 (1998), pp. 431 - 446.
[64] M.B. Plenio and S. Virmani, An introduction to entan-
glement measures, Quant. Inf. Comp. 7 (2007), pp. 1 -
71.
[65] F. Fassioli and A. Olaya-Castro, Distribution of entan-
glement in light-harvesting complexes and their quantum
efficiency, New. J. Phys. 12 (2010), 085006.
[66] F. Caruso, A.W. Chin, A. Datta, S.F. Huelga, and M.B.
Plenio, Entanglement and entangling power of the dy-
namics in light-harvesting complexes, Phys. Rev. A 81
(2010), 062346.
[67] J.S. Cao and R.J. Silbey, Optimization of Exciton Trap-
ping in Energy Transfer Processes, J. Phys. Chem. A
113 (2009), pp. 13825-13838.
[68] F. Caruso, S.F. Huelga and M.B. Plenio, Noise-
Enhanced Classical and Quantum Capacities in Com-
munication Networks, Phys. Rev. Lett. 105 (2010),
190501.
[69] P. Giorda, S. Garnerone, P. Zanardi and S. Lloyd, Inter-
play between coherence and decoherence in LHCII pho-
tosynthetic complexes, E-print arXiv:1106.1986.
[70] J. Wu, R.J. Silbey and J. Cao, Generic Mechanism of
Optimal Energy Transfer Efficiency: A Scaling Theory
of the Mean First-Passage Time in Exciton Systems,
Phys. Rev. Lett. 110 (2013), 200402.
[71] L. Campos Venuti and P. Zanardi, Excitation transfer
through open quantum networks: a few basic mecha-
nisms, Phys. Rev. B 84 (2011), 134206.
[72] J.L Wu, F. Liu, Y. Shen, J.S. Cao and R.J. Silbey,
Efficient energy transfer in light-harvesting systems, I:
optimal temperature, reorganization energy and spatial-
temporal correlations, New J. Phys. 12 (2010) 105012.
[73] S. Hoyer, M. Sarovar, and K.B. Whaley, Limits of quan-
tum speedup in photosynthetic light harvesting, New J.
Phys. 12 (2010), 065041.
[74] A.W. Chin, A. Datta, F. Caruso, S.F. Huelga, and M.B.
Plenio, Noise-assisted energy transfer in quantum net-
works and light-harvesting complexes, New J. Phys. 12
(2010), 065002.
[75] P. Rebentrost, M. Mohseni, I. Kassal, S. Lloyd, and A.
Aspuru-Guzik, Environment-Assisted Quantum Trans-
port, New J. Phys. 11 (2009), 033003.
[76] J. Moix, J.L. Wu, P.F. Huo, D. Coker and J.S. Cao,
Efficient Energy Transfer in Light-Harvesting Systems,
III: The Influence of the Eighth Bacteriochlorophyll on
the Dynamics and Efficiency in FMO, J. Phys. Chem.
Lett. 2 (2011), pp. 3045 - 3052.
[77] A. Olaya-Castro, C.F. Lee, F. Fassioli Olsen, and
N.F. Johnson, Efficiency of energy transfer in a light-
harvesting system under quantum coherence, Phys. Rev.
B 78 (2008), 085115.
[78] M.B. Plenio and S.F. Huelga, Quantum dynamics of bio-
molecular systems in noisy environments, Proceedings
of the 22nd Solvay Conference in Chemistry on ”Quan-
tum Effects in Chemistry and Biology, Procedia Chem-
istry 3 (2011), pp. 248 - 255.
Contemporary Physics
24
[79] S. Lloyd, M. Mohseni, A. Shabani and H. Rab-
itz, The quantum Goldilocks effect: on the conver-
gence of
timescales in quantum transport, E-print
arXiv:1111.4982
[80] S.R. Hartmann, and E.L. Hahn, Nuclear Double Res-
onance in the Rotating Frame, Phys. Rev. 128 (1962),
pp. 2042.
[81] J.-M. Cai, A. Retzker, F. Jelezko, and M.B. Plenio, A
Large-Scale Quantum Simulator on a Diamond Surface
at Room Temperature, Nat. Phys. 9 (2013), pp. 168-173.
[82] J.-M. Cai, F. Jelezko, M.B. Plenio, and A. Retzker,
Diamond-Based Single-Molecule Magnetic Resonance
Spectroscopy, New J. Phys. 15 (2013), pp. 013020.
[83] L.A. Pachon and P. Brumer, Physical Basis for Long-
Lived Electronic Coherence in Photosynthetic Light-
Harvesting Systems, J. Phys. Chem. Lett. 2 (2011), pp.
2728 - 2732.
[84] Ch. Kreisbeck and T. Kramer, Long-Lived Electronic
Coherence in Dissipative Exciton-Dynamics of Light-
Harvesting Complexes, J. Phys. Chem. Lett. 3 (2012),
pp. 2828 - 2833.
[85] F. Fassioli, A. Nazir, and A. Olaya-Castro, Quantum
State Tuning of Energy Transfer in a Correlated En-
vironment, J. Phys. Chem. Lett. 1 (2010), pp. 2139 -
2143.
[86] G. Ritschel, J. Roden, W. T. Strunz, A. Aspuru-Guzik,
A. Eisfeld, Absence of quantum oscillations and depen-
dence on site energies in electronic excitation transfer
in the Fenna-Matthews-Olson trimer, J. Phys. Chem.
Lett. 2 (2011), pp. 2912 - 2917.
[87] J. Prior, A.W. Chin, S.F. Huelga, and M.B. Plenio, Ef-
ficient simulation of strong system-environment inter-
actions, Phys. Rev. Lett. 105 (2010), 050404.
[88] A.W. Chin, A. Rivas, S.F. Huelga, and M.B. Plenio, Ex-
act mapping between system-reservoir quantum models
and semi-infinite discrete chains using orthogonal poly-
nomials, J. Math. Phys. 51 (2010), 092109.
[89] A.W. Chin, S.F. Huelga, and M.B. Plenio, Chain rep-
resentations of open quantum systems and their nu-
merical simulation with time-adapative density matrix
renormalisation group methods, in Semiconductors and
Semimetals eds. Uli Wurfel, Michael Thorwart, Eicke
R. Weber and Chennupati Jagadish, vol 85, Burlington:
Academic Press, 2011, pp. 115-144.
[90] S.F. Huelga, A. Rivas, and M.B. Plenio, Non-
Markovianity assisted Steady State Entanglement, Phys.
Rev. Lett. 108 (2012), 160402.
[91] M. Zarzo, The sense of smell: molecular basis of odorant
recognition, Biol. Rev. 82 (2007), pp. 455 - 479.
[92] S. Gane, D. Georganakis, K. Maniati, M. Vamvakias,
N. Ragoussis, E.M.C. Skoulakis, and L. Turin, Molecu-
lar Vibration-Sensing Component in Human Olfaction,
PLOS ONE 8 (2013), e55780.
[93] G.M. Dyson, The scientific basis of odour, Chem. Ind.
57 (1938), pp. 647 - 651.
[94] R. Wright, Odor and molecular vibration: neural coding
of olfactory information, J. Theor. Biol. 64 (1977), pp.
473 - 502.
[95] L. Turin, A method for the calculation of odor character
from molecular structure, J. Theor. Biol. 21 (2002), pp.
367 - 385.
[96] J. Hihath and N. Tao, Electron-phonon interactions in
atomic and molecular devices, Prog. Surf. Sci. 87 (2012),
pp. 189 - 208.
[97] R.C. Jaklevic and J. Lambe, Molecular vibration spectra
by electron tunneling, Phys. Rev. Lett. 17 (1966), pp.
1139 - 1140.
[98] J. Lambe and R.C. Jaklevic, Molecular vibration spectra
by inelastic electron tunneling, Phys. Rev. 165 (1968),
pp. 821 - 832.
[99] L. Turin, The Secret of Scent. Adventures in Perfume
and the Science of Smell, Faber and Faber Ltd. London
2006.
[100] J.C. Brookes, F. Hartoutsiou, A.P. Horsfield, and A.M.
Stoneham, Could humans recognize odor by phonon as-
sisted tunneling?, Phys. Rev. Lett. 98 (2007), 038101.
[101] I.A. Solovyov, P.-Y. Chang, and K. Schulten, Vibra-
tionally assisted electron transfer mechanism of olfac-
tion: myth or reality?, Phys. Chem. Chem. Phys. 14
(2012), pp. 13861 - 13871.
[102] E.R. Bittner, A. Madalan, A. Czader, and G. Roman,
Quantum origins of molecular recognition and olfaction
in drosophila, J. Chem. Phys. 137 (2012), 22A551.
[103] J.C. Brookes, A.P. Horsfield, and A.M. Stoneham, The
Swipe Card Model of Odorant Recognition, Sensor 12
(2012), pp. 15709 - 15749.
[104] J.C. Brookes, Olfaction: the physics of how smell works,
Contemp. Phys. 52 (2011), pp. 385 - 402.
[105] R.A. Marcus, Electron transfer reactions in chemistry.
Theory and experiment, Rev. Mod. Phys. 65 (1993), pp
599 - 610.
[106] G.-Y. Chen, N. Lambert, C.-M. Li, Y.-N. Chen and
F. Nori, Rerouting Excitation Transfer in the Fenna-
Matthews-Olson Complex, E-print arXiv:1304.2613.
[107] K.M. Gaab and Ch. J. Bardeen, The effects of con-
nectivity, coherence, and trapping on energy transfer in
simple light-harvesting systems studied using the Haken-
Strobl model with diagonal disorder, J. Chem. Phys. 121
(2004), pp. 7813 - 7820.
[108] F.L. Semiao, K. Furuya and G.J. Milburn, Vibration-
enhanced quantum transport, New J. Phys. 12 (2010),
083033.
[109] A. Asadian, M. Tiersch, G.G. Guerreschi, J.M. Cai, S.
Popescu, and H.J. Briegel, Motional effects on the ef-
ficiency of excitation transfer, New J. Phys. 12 (2010),
075019.
[110] A. Vaziri and M.B. Plenio, Quantum coherence in ion
channels: resonances, transport and verification, New
J. Phys. 12 (2010), 085001.
[111] I. Kassal and A. Aspuru-Guzik, Environment-assisted
quantum transport in ordered systems, New J. Phys. 14
(2012), 053041.
[112] M. Mohseni, A. Shabani, S. Lloyd, Y. Omar and H. Ra-
bitz, Geometrical effects on energy transfer in disordered
open quantum systems, E-print arXiv:1212.6804.
[113] I.A. Balabin and J.N. Onuchic, Dynamically controlled
protein tunneling paths in photosynthetic reaction cen-
ters, Science 290 (2000), pp. 114 - 117.
[114] M.H. Vos, F. Rappaport, J.C. Lambry, J. Breton, J.L.
Martin, Visualization of coherent nuclear motion in a
membrane-protein by femtosecond spectroscopy, Nature
363 (1993), pp. 320 – 325.
[115] M.H. Vos, M.R. Jones, C.N. Hunter, J. Breton, J.L.
Martin, Proc. Natl. Acad. Sci. 91 (1994), pp. 12701 –
12705.
[116] A.M. Streltsov, A.G. Yakovlev, A.Y. Shkuropatov,
V.A. Shuvalov, Femtosecond kinetics of electron trans-
fer in the bacteriochlorophyll(M)-modified reaction cen-
Contemporary Physics
25
ters from Rhodobacter sphaeroides (R-26), Febs Letters,
383 (1996), pp. 129 - 132.
[117] V.I. Novoderezhkin, A.G. Yakovlev, R. van Grondelle
and V.A. Shuvalov, Coherent Nuclear and Electronic
Dynamics in Primary Charge Separation in Photosyn-
thetic Reaction Centers: A Redfield Theory Approach,
J. Phys. Chem. B 108 (2004), pp. 7445 - 7457.
[118] S. Westenhoff, D. Palecek, P. Edlund, P. Smith and D.
Zigmantas, Coherent picosecond exciton dynamics in a
photosynthetic reaction center, J. Am. Chem. Soc. 134
(2012), pp. 16484-16487. (2012).
[119] E. Romero, R. Augulis, V.I. Novoderezhkin, M. Ferretti,
J. Thieme, D. Zigmantas and R. van Grondelle, private
communication.
[120] C.A. Rozzi, S.M. Falke, N. Spallanzani, A. Rubio, E.
Molinari, D. Brida, M. Maiuri, G. Cerullo, H. Schramm,
J. Christoffers and Ch. Lienau, Quantum coherence con-
trols the charge separation in a prototypical artificial
light-harvesting system, Nature. Comm. 4 (2013), 1602.
[121] J. Lim, J. Ryu, C. Lee, S. Yoo, H. Jeong, and J. Lee,
Role of Energy-Level Mismatches in a Multi-Pathway
Complex of Photosynthesis, New J. Phys. 13 (2011),
103002.
[122] J.L. Wu, F. Liu, J. Ma, R.J. Silbey and J.S. Cao,
Efficient energy transfer in light-harvesting systems:
Quantum-classical comparison, flux network, and ro-
bustness analysis, J. Chem. Phys. 137 (2012) 174111.
[123] S. Johnsen and K. J. Lohmann, The physics and neu-
robiology of magnetoreception. Nature Rev. Neurosci. 6
(2005), 703712.
[124] T. Ritz, Quantum effects in biology: Bird navigation,
Procedia Chemistry 3 (2011), pp. 262 - 275.
[125] W. Wiltschko, R. Wiltschko, and T. Ritz, The mecha-
nism of the avian magnetic compass, Procedia Chem-
istry 3 (2011), pp. 276 - 284.
[126] J. B Phillips and S. C. Borland, Behavioural evidence for
use of a light-dependent magnetoreception mechanism by
a vertebrate, Nature 359 (1992), pp. 142 - 144.
[127] P. Thalau, T. Ritz, K. Stapput, R. Wiltschko, and W.
Wiltschko, Magnetic compass orientation of migratory
birds in the presence of a 1.315 MHz oscillating field,
Naturwissenschaften 92 (2005), pp. 86 - 90.
[128] K. Maeda, K. B. Henbest, F. Cintolesi, I. Kuprov, C. T.
Rodgers, P. A. Liddell, D. Gust, C. R. Timmel, and P.
J. Hore, Chemical compass model of avian magnetore-
ception, Nature 453 (2008), pp. 387 - 390.
[129] U.E. Steiner, T. Ulrich, Magnetic Field Effects in Chem-
ical Kinetics and Related Phenomena, Chem. Rev. 89
(1989), pp. 51
[130] I.K. Kominis, Quantum Zeno effect explains magnetic-
sensitive radical-ion-pair reactions, Phys. Rev. E 80
(2009), 056115.
[131] J.M. Cai, G.G. Guerreschi, and H.J. Briegel, Quantum
control and entanglement in a chemical compass, Phys.
Rev. Lett. 104 (2010), 220502.
[132] E.M. Gauger, E. Rieper, J.J.L. Morton, S.C. Benjamin,
and V. Vedral, Sustained Quantum Coherence and En-
tanglement in the Avian Compass, Phys. Rev. Lett. 106
(2011), 040503.
[133] J.M. Cai, Quantum Probe and Design for a Chemi-
cal Compass with Magnetic Nanostructures, Phys. Rev.
Lett. 106 (2011), 100501.
[134] J.M. Cai, F. Caruso, and M.B. Plenio, Quantum limit
for avian magnetoreception: How sensitive can a chem-
ical compass be?, Phys. Rev. A 85, (2012), 040304(R).
[135] J. N. Bandyopadhyay, T. Paterek, and D. Kaszlikowski,
Quantum Coherence and Sensitivity of Avian Magne-
toreception, Phys. Rev. Lett. 109 (2012), 110502.
[136] E. M. Gauger, S. C. Benjamin, Comment on Quantum
Coherence and Sensitivity of Avian Magnetoreception,
E-Print arXiv:1303.4539.
[137] H. J. Hogben, T. Biskup, and P. J. Hore, Entangle-
ment and Sources of Magnetic Anisotropy in Radical
Pair-Based Avian Magnetoreceptors, Phys. Rev. Lett.
109 (2012), 220501.
[138] J.-M. Cai and M.B. Plenio, Chemical compass for
avian magnetoreception as a quantum device, E-Print
arXiv:1304.4143.
[139] G. Lindblad, Generators of Quantum Dynamical Semi-
groups, Comm. Math. Phys. 48 (1976), pp. 119 - 130.
[140] V. Gorini, A. Kossakowski, and E.C.G. Sudarshan,
Completely Positive Dynamical Semigroups of N-Level
System, J. Math. Phys. 17 (1976), pp. 821 - 825.
[141] A. Redfield, On the theory of relaxation processes, IBM
J. Res. Dev. 1 (1957), pp. 19 - 31.
[142] W. M. Zhang, T. Meier, V. Chernyak and S. Mukamel,
Exciton-migration and three-pulse femtosecond optical
spectroscopies of photosynthetic antenna complexes, J.
Chem. Phys. 108 (1998), pp. 7763 7774.
[143] T. Renger and R. A. Marcus, Variable-range hopping
electron transfer through disordered bridge states: Ap-
plication to DNA, J. Chem. Phys. 107 (2003), pp. 8404
8419.
[144] In fact a redefinition of modes can always be made to
achieve this choice with a suitable redefinition of h(x).
[145] These include thermal states and states displaced in
phase space which covers a wide class of reasonable ini-
tial states in biological systems. It will be highly chal-
lenging to create states that differ significantly from
these.
[146] A.J. Leggett, S. Chakravarty, A.T. Dorsey, M.P.A.
Fisher, A. Garg, and W. Zwerger, Dynamics of the dis-
sipative two-level system, Rev. Mod. Phys. 59 (1987),
pp. 1 - 85.
[147] U. Weiss, Quantum Dissipative Systems, World Scien-
tific, Singapore (2001).
[148] R.S. Burkey and C.D. Cantrell, Discretization in the
quasi-continuum, J. Opt. Soc. Am. B 1 (1984), pp. 169
- 175.
[149] P. Borwein and T. Erd´elyi, Polynomials and Polynomial
Inequalities, Spinger 1995
[150] W. Gautschi, ORTHPOL-a package of routines for gen-
erating orthogonal polynomials and Gauss-type quadra-
ture rules, ACM Trans. Math. Soft. 20 (1994), pp. 21-62.
[151] J. Eisert, M. Cramer, and M.B. Plenio, Area laws for
the entanglement entropy - a review, Rev. Mod. Phys.
82 (2010), pp. 277 - 306.
[152] A. Ishizaki and G.R. Fleming, Unified treatment of
quantum coherent and incoherent hopping dynamics in
electronic energy transfer: Reduced hierarchy equation
approach, J. Chem. Phys. 130 (2009), 234111.
[153] A. Ishizaki and Y. Tanimura, Quantum dynamics of sys-
tem strongly coupled to low-temperature colored noise
bath: Reduced hierarchy equations approach, J. Phys.
Soc. Jpn. 74 (2005), pp. 3131 - 3134.
[154] A. Ishizaki and G.R. Fleming, Theoretical examina-
tion of quantum coherence in a photosynthetic system
at physiological temperature, Proc. Nat. Acad. Sci. Am.
Contemporary Physics
26
106 (2009), pp. 17255 - 17260.
[155] T. Takagahara, E. Hanamura, and R. Kubo, Stochastic
models of intermediate state interaction in 2nd order
optical processes - stationary response .1, J. Phys. Soc.
Jpn. 43 (1977), pp. 802 - 810.
[156] T. Takagahara, E. Hanamura, and R. Kubo, Stochastic
models of intermediate state interaction in 2nd order
optical processes - stationary response .2, J. Phys. Soc.
Jpn. 43 (1977), pp. 811 - 816.
[157] Y. Tanimura and R. Kubo, Time evolution of a quantum
system in contact with a nearly Gaussian-Markoffian
noise bath, J. Phys. Soc. Jap. 58 (1989), pp. 101 - 114.
[158] A. Ishizaki, T.R. Calhoun, G.S. Schlau-Cohen, and G.R.
Fleming, Quantum coherence and its interplay with pro-
tein environments in photosynthetic electronic energy
transfer, Phys. Chem. Chem. Phys. 12 (2010), pp. 7319
- 7337.
[159] A.G. Dykstra and Y. Tanimura, The role of the envi-
ronment time scale in light-harvesting efficiency and co-
herent oscillations, New J. Phys. 14 (2012), 073027.
[160] Ch. Kreisbeck, T. Kramer, M. Rodriguez, and B. Hein,
High-performance solution of hierarchical equations of
motions for studying energy-transfer in light-harvesting
complexes, J. Chem. Theo. Comp. 7 (2011), pp. 2166 -
2174.
[161] B. Hein, C. Kreisbeck, T. Kramer, and M. Rodr´ıguez,
Modelling of Oscillations in Two-Dimensional Echo-
Spectra of the Fenna-Matthews-Olson Complex, New J.
Phys. 14 (2012), 023018.
[162] J. Zhu, S. Kais, P. Rebentrost, and A. Aspuru-Guzik,
Modified Scaled Hierarchical Equation of Motion Ap-
proach for the Study of Quantum Coherence in Photo-
synthetic Complexes, J. Phys. Chem. B 115 (2011), pp.
1531 - 1537.
[163] D. E. Makarov and N. Makri, Path -integrals for dissipa-
tive systems by tensor multiplication - Condensed-phase
quantum dynamics for arbitrarily long times, Chem.
Phys. Lett. 221 (1994), pp. 482 - 491.
[164] N. Makri and D. E. Makarov, Tensor propagator for
iterative quantum time evolution of reduced density-
matrices .1. Theory, J. Chem. Phys. 102 (1995), pp.
4600 - 4610.
[165] N. Makri and D. E. Makarov, Tensor propagator for
iterative quantum time evolution of reduced density-
matrices .2. Numerical methodology, J. Chem. Phys. 102
(1995), pp. 4611 - 4618.
[166] M. Thorwart, J. Eckel, J.H. Reina, P. Nalbach, and
S. Weiss, Enhanced quantum entanglement in the non-
Markovian dynamics of biomolecular excitons, Chem.
Phys. Lett. 478 (2009), pp. 234 - 237.
[167] P. Nalbach and M. Thorwart, Quantum Coherence and
Entanglement in Photosynthetic Light-Harvesting Com-
plexes, in Quantum efficiency in complex systems, Pt
I: Biomolecular systems, Book Series: Semiconductors
and Semimetals, Vol 83 (2010), pp. 39 - 75.
[168] S. Weiss, J. Eckel, M. Thorwart, and R. Egger, Itera-
tive real-time path integral approach to nonequilibrium
quantum transport, Phys. Rev. B 77 (2008), 195316.
[169] M. Thorwart, P. Reimann and P. Hanggi, Iterative al-
gorithm versus analytic solutions of the parametrically
driven dissipative quantum harmonic oscillator, Phys.
Rev. E 62 (2000) pp. 5808 - 5817.
[170] M. Thorwart, E. Paladino, and M. Grifoni, Dynamics
of the spin-boson model with a structured environment,
Chem. Phys. 296 (2004), pp. 333 - 344.
[171] P. Nalbach, A. Ishizaki, G.R. Fleming, and M. Thor-
wart, Iterative path-integral algorithm versus cumulant
time-nonlocal master equation approach for dissipative
biomolecular exciton transport, New J. Phys. 13 (2011),
063040.
[172] D.P.S. McCutcheon and A. Nazir, Coherent and inco-
herent dynamics in excitonic energy transfer: Corre-
lated fluctuations and off-resonance effects, Phys. Rev.
B 83 (2011), 165101.
[173] D.P.S. McCutcheon and A. Nazir, Consistent treatment
of coherent and incoherent energy transfer dynamics us-
ing a variational master equation, J. Chem. Phys. 135
(2011), 114501.
[174] A. Kolli, A. Nazir, and A. Olaya-Castro, Electronic
excitation dynamics in multichromophoric systems de-
scribed via a polaron-representation master equation, J.
Chem. Phys. 135 (2011), 154112.
[175] G. Stock and M. Thoss, Mapping approach to the semi-
classical description of nonadiabatic quantum dynamics,
Phys. Rev. A 59 (1999), pp. 564 - 579.
[176] M. Thoss and G. Stock, Semiclassical description of
nonadiabatic quantum dynamics, Phys. Rev. Lett. 78
(1997), pp. 578 - 581.
[177] E.R. Dunkel, S. Bonella, and D.F. Coker, Iterative lin-
earized approach to nonadiabatic dynamics, J. Chem.
Phys. 129 (2008), 114106.
[178] P. Huo and D.F. Coker, Iterative linearized density ma-
trix propagation for modeling coherent excitation energy
transfer in photosynthetic light harvesting, J. Chem.
Phys. 133 (2010), 184108.
[179] A. Shabani, M. Mohseni, H. Rabitz, and S. Lloyd, Ef-
ficient estimation of energy transfer efficiency in light-
harvesting complexes, Phys. Rev. E 86 (2012), 011915.
[180] H.-P. Breuer and F. Petruccione, The Theory of Open
Quantum Systems, Oxford University Press, New York,
2002.
[181] J. Roden, A. Eisfeld, W. Wolff, and W.T. Strunz, In-
fluence of Complex Exciton-Phonon Coupling on Opti-
cal Absorption and Energy Transfer of Quantum Aggre-
gates, Phys. Rev. Lett. 103 (2009), 058301.
[182] G. Ritschel, J. Roden, W.T. Strunz, and A. Eis-
feld, An efficient method to calculate excitation en-
ergy transfer in light-harvesting systems: application to
the Fenna-Matthews-Olson complex, New J. Phys. 13
(2011), 113034.
[183] L.A. Pachon and P. Brumer, Computational methodolo-
gies and physical insights into electronic energy trans-
fer in photosynthetic light-harvesting complexes, Phys.
Chem. Chem. Phys. 14 (2012), pp. 10094 - 10108.
|
1306.2065 | 1 | 1306 | 2013-06-09T22:06:25 | Emergence of the advancing neuromechanical phase in a resistive force dominated medium | [
"physics.bio-ph"
] | Undulatory locomotion, a gait in which thrust is produced in the opposite direction of a traveling wave of body bending, is a common mode of propulsion used by animals in fluids, on land, and even within sand. As such it has been an excellent system for discovery of neuromechanical principles of movement. In nearly all animals studied, the wave of muscle activation progresses faster than the wave of body bending, leading to an advancing phase of activation relative to the curvature towards the tail. This is referred to as "neuromechanical phase lags" (NPL). Several multi-parameter neuromechanical models have reproduced this phenomenon, but due to model complexity the origin of the NPL has proved difficult to identify. Here we use perhaps the simplest model of undulatory swimming to accurately predict the NPL during sand-swimming by the sandfish lizard, with no fitting parameters. The sinusoidal wave used in sandfish locomotion, the friction-dominated and non-inertial granular resistive force environment, and the simplicity of the model allow detailed analysis, and reveal the fundamental mechanism responsible for the phenomenon: the combination of synchronized torques from distant points on the body and local traveling torques. This general mechanism should help explain the NPL in organisms in other environments; we therefore propose that sand-swimming could be an excellent system to quantitatively generate and test other neuromechanical models of movement. Such a system can also provide guidance for the design and control of robotic undulatory locomotors in complex environments. | physics.bio-ph | physics |
Emergence of the advancing neuromechanical phase in a
resistive force dominated medium
Yang Ding1
,
∗, Sarah S. Sharpe2, Kurt Wiesenfeld1 and Daniel I. Goldman1
,
2
,
∗∗
1. School of Physics,
2. Interdisciplinary Bioengineering Program,
Georgia Institute of Technology, Atlanta, United States
* (current) Department of Aerospace and Mechanical Engineering,
University of Southern California, Los Angeles, USA
** To whom correspondence should be addressed.
837 State Street NW, Atlanta, GA 30332-0430,
E-mail: [email protected]
July 30, 2018
Abstract
Undulatory locomotion, a gait in which thrust is produced in the opposite direction of a trav-
eling wave of body bending, is a common mode of propulsion used by animals in fluids, on land,
and even within sand. As such it has been an excellent system for discovery of neuromechanical
principles of movement. In nearly all animals studied, the wave of muscle activation progresses
faster than the wave of body bending, leading to an advancing phase of activation relative to the
curvature towards the tail. This is referred to as "neuromechanical phase lags" (NPL). Several
multi-parameter neuromechanical models have reproduced this phenomenon, but due to model
complexity the origin of the NPL has proved difficult to identify. Here we use perhaps the sim-
plest model of undulatory swimming to accurately predict the NPL during sand-swimming by
the sandfish lizard, with no fitting parameters. The sinusoidal wave used in sandfish locomotion,
the friction-dominated and non-inertial granular resistive force environment, and the simplicity
of the model allow detailed analysis, and reveal the fundamental mechanism responsible for the
phenomenon: the combination of synchronized torques from distant points on the body and local
traveling torques. This general mechanism should help explain the NPL in organisms in other envi-
ronments; we therefore propose that sand-swimming could be an excellent system to quantitatively
generate and test other neuromechanical models of movement. Such a system can also provide
guidance for the design and control of robotic undulatory locomotors in complex environments.
Animal movement emerges from the complex interplay of nervous and musculoskeletal systems
with the environment. Much progress has been made for understanding the neural control patterns
and motor systems responsible for effective locomotion ([1, 2, 3, 4, 5, 6, 7]). While the environment's
influence on neural control is increasingly recognized [8, 9], challenges remain in understanding how
environments shape the control strategy of locomotion. Particular behaviors, gaits and environments
have revealed themselves to be amenable to detailed comparison of experiment and theory to elucidate
neuromechanical principles of control [4, 10, 11, 12]. A form of locomotion where there has been
much progress is undulatory locomotion, a movement strategy employed by numerous, phylogenetically
1
diverse animals such as fish, snakes, worms and sandfish lizards (Scincus scincus) (see Fig. 1) [13, 14,
15, 16, 17, 18] to traverse fluids, solids, and even sand.
In undulatory locomotion, a traveling wave of muscle activation (and curvature) propagates from
head to tail resulting in forward movement. The forces produced on different "segments" of the
body can be decomposed into thrust and drag, and integrating these over the body at any instant
in time determines the propulsion of the animal. Many robots have also been built that use such
a gait [19, 20, 21]. A feature of undulatory locomotion that is observed across a range of animal
sizes and environments is that the wave of muscle activation travels faster than the wave of curvature
[22, 23, 24, 25, 26]. Consequently, the relative phase of the muscle activation to the curvature advances
along the body. Physically, this means that more posterior muscles begin activating earlier in the
muscle strain cycle (i.e. while the muscle is lengthening) and produce more negative work than
anterior muscles. The phenomenon of the advancing neuromechanical phase is often referred to as the
"neuromechanical phase lags" (Fig. 1c & d), or "NPL" for short.
Two complementary modeling approaches are used to understand movement principles. The
"bottom-up" approach (referred to as "anchoring" in [27]) integrates realistic models of multiple bio-
components and the complex interactions among them, as well as the with models of the environment.
For example, a model (see [28]) might incorporate tens to hundreds of muscles, hundreds to thousands
of neurons, chemical kinetics, and the nonlinear couplings among them. Further complexity could be
added by coupling these models to fluids which are governed by complex partial differential equations.
In contrast, the "top-down" approach (referred to as "templates" in [27]) identifies coordinated compo-
nents as one single element to generate reduced models and seeks general principles of system behavior.
Using the first approach, many multi-parameter neuromechanical models [29, 30, 10, 31, 28, 32, 33]
have been proposed to model undulatory locomotion. While such models qualitatively reproduce the
NPL in undulatory swimming, due to uncertainties about the passive body properties and the hydro-
dynamical forces, as well as the model complexity and number of parameters, it remains a challenge
to explain the origin of the phenomenon.
In this paper we show that what might seem to be a specialized and complex system, a lizard "swim-
ming" in sand using an undulatory gait, facilitates quantitative comparison of experiment and theory,
and helps explain the fundamental origin of the NPL in undulatory locomotion in other environments.
We base the present work on our previous biological muscle activity measurements [26] which revealed
that the sandfish displays neuromechanical phase lags when targeting a particular behavior -- escape.
Using a template approach -- inputting kinematics of the lizard which confer swimming speed and en-
ergetic benefits [34] into a previously developed granular resistive force model of sand-swimming and
abstracting the nervous system and musculoskeletal system as a "black box" -- we are able to repro-
duce internal torque timing patterns (i.e. from muscle contractions) with no fitting parameters. The
simple kinematics combined with the relatively simple rheological features of organism-fluidized sand
allow us to analyze the model and thus make statements about general principles of neuromechanics
in swimming, applicable to organisms and robots in other environments.
1 Model
1.1 Resistive force theory model
Previously we developed a granular resistive force theory (RFT) model and a numerical simulation that
explained the swimming performance of the sandfish [18, 35, 34]. The models showed that the lizards
swim within a self-generated "frictional fluid" where frictional forces between the granular particles
dominate over both the body inertia and inertial forces from the environment. As before, we prescribe
body kinematics (in the frame of the animal) based on the experimental observation that the body
position of a sandfish in the body frame is approximately a single-period sinusoidal wave traveling
2
a
b
Swimming direction
Traveling wave direction
−θ
1 cm
+θ
50
c
0.9 SVL
5
50
d
0.3 SVL
2
)
ο
(
θ
0
−50
Muscle Lengthening
Muscle Shortening
0.25
Time (s)
0
-5
−50
0.5
0
Time (s)
-2
0.25
E
M
G
(
m
V
)
Figure 1: Neuromechanical phase lags of the sandfish during sand-swimming. (a) A sandfish lizard
resting on 0.3 mm diameter glass particles.
(b) A trace of an x-ray image of the sandfish during
subsurface sand-swimming at time, t = 0.21 s in panels c and d. Opaque markers (black circles) are
attached to the exterior midline to facilitate tracking. Electrodes are implanted in epaxial musculature
on the right side of the body at 0.3 (magenta), 0.5 (green), 0.7 (blue), 0.9 (yellow) snout-vent length
(SVL) locations (where the vent is just posterior to the pelvic girdle and the SVL is approximately
0.75 of the total body length). (c) & (d) EMG recordings at 0.9 and 0.3 SVL, respectively, during
sand-swimming. Gray regions indicate time duration over which the rectified filter EMG is above a
threshold (equal to the mean of the rectified-filtered signal) indicating muscle activation (see [26] for
more details). The blue line shows the measured angle between consecutive markers (see panel b). The
red circles show the maximum or minimum of the best 2nd order polynomial fit to the angle vs. time
series for each half cycle. Arrows indicate the difference in time between the onset of muscle activation
and maximal convexity. Note the different scales for EMG due to different electrode constructions.
3
v
t=0
Swimming direction
Traveling wave direction
F
κ(+)
.
κ(+)
τ(−)
ψ
Figure 2: Diagram of the model. Magenta arrows represent velocity and green arrows represent forces
from the medium. Inset shows the signs of the torque (τ ), the curvature (κ), and the rate of change
of curvature ( κ) at approximately 0.6 body length on the body. Negative τ corresponds to no muscle
activation on the right side of the body (red thick line). ψ indicates the angle between the segment
axis and its velocity.
posteriorly (Fig. 2):
yb = A sin[2π(
xb
λ
+
t
T
)],
(1)
λ = 2π
where yb is the lateral displacement from the midline of a straight animal, A is the amplitude, T is the
period of undulation, λ is the wavelength, t is the time, and xb is the distance along a line parallel to
the direction of the traveling wave measured from the tail tip. Here, we normalize both the wavelength
and period to 2π such that 2π
T = 1. Since the trunk of the sandfish is quite uniform (with
body width variations less than about 5% from 0.1 snout to vent length (SVL) to 1.0 SVL) and the
diameter of the body decreases significantly after about 1.2 SVL, we used a uniform body shape and
took the total arc length (L) in the model to be 1.2 times the average SVL (8.9 ± 0.3 cm) of the
animal. Dissection revealed after 1.2 SVL the tail is composed of mostly adipose tissue and a small
amount of muscle; therefore, both the external and internal torques on the tail should be minimal for
the tail beyond 1.2 SVL. We neglected the variation of the horizontal position xb of a segment within
a cycle, so the normalized position on the animal body s 2π
L corresponded to the horizontal position
xb in the model, where s is the arc length from the tail end. When a smaller amplitude was used, the
wavelength was kept as 2π.
For swimming in sand, the granular force ~F on any infinitesimal segment of the swimmer is in-
dependent of the segment speed (and thus undulation frequency), proportional to its depth, and is a
function of the angle (ψ in Fig. 2) between the segment axis and its velocity direction. See Figure S1
and Equation S1 in SI for the empirically determined granular force ~F (ψ). The depth of a segment is
calculated assuming the model sandfish swims with its center 3.5 cm below the horizontal plane and
at an entry angle of 22 degrees (an average value for the sandfish [18]). The entry angle is the angle
between the horizontal plane and the plane in which the animal moves [26].
Since the estimated inertial force is negligible, the swimmer moves in a way such that net external
force and torque are approximately zero. In this study, we consider all three degrees of freedom in
4
κ
1
0
a
−1
0
Time
2
π
c
x=0.71π
x=1.28π
b
2π
0
Time
M
4
τ
0
−4
2π
4
a
x
κ
Offset
M
a
x
κ
e
m
T
i
π
M
i
n
κ
M
i
n
κ
Onset
0
0
Tail
π
x
2
π
Head
τ
(
N
.
c
m
)
0
-4
Figure 3: Neuromechanical phase lags in the model. (a) & (b) The curvature (blue lines), torque
(green lines), and the predicted muscle activation (gray shaded region) from the RFT model at two
representative points indicated by black dots in Figure 2 and gray vertical lines in (c). (c) Torque
as a function of time and position along the body. Gray vertical bars indicate the predicted muscle
activation durations at two representative points. Solid and dashed black lines represent the time when
the maximal curvature and minimal curvature are reached, respectively.
a plane, namely the forward (the only degree of freedom in our previous RFT models), lateral and
yaw motion ("recoil"), and determine the velocities of the three degrees of freedom by solving the
force/torque balance equations at every instant of time. Since the motion of a point on the body is the
~R, θ)
superposition of the prescribed and center of mass (CoM) motions, the net external force ~Fnet(
~R, θ) = (0, 0, τnet) are functions of the center of mass
and net external torque about the CoM ~τnet(
velocity
~R and rotation rate about the CoM θ. For the CoM movement, Newton's laws give
~Fnet(
~τnet(
~R, θ) = M
~L
~R, θ) =
~R
(2)
where M is total mass and L is angular momentum. By setting the inertial terms on the right sides
of these equations to zero, the center of mass velocities (
~R and θ) can be numerically determined.
1.2 Torque calculation in RFT
Because inertia is negligible, the net torque due to the granular force on a portion (e.g.
[xb, 2π]) of
the sandfish body about any point of interest xb is also approximately zero. From this we calculate
5
the internal torque (i.e the torque generated by muscle) at xb:
~τmuscle + Z 2π
xb
~r × ~f ds =
~L ≈ 0
− τmuscle = τ (xb, t)
= Z 2π
xb
{(z − xb)fy(z, t) − [y(z, t) − y(xb, t)]fx(z, t)}q1 + y ′2
b dz.
(3)
where ~f is the granular force per unit length. We assume the muscle must only overcome torque
from resistive forces τ and thus internal passive body forces are small compared to external resistive
forces. This assumption was tested by performing in vivo bending tests on an anaesthetized animal (see
Methods and SI)and measuring stiffness and damping coefficients at varying rotation rates; we estimate
that the maximal torques from elastic (0.094 ± 0.027 N·cm) and damping (0.055 ± 0.034 N·cm) forces
are over an order of magnitude smaller than the maximal torque from resistive forces (4.1 N·cm). We
also assume the time lag between neural activation and muscle force development is small compared to
the sandfish undulation period (≈ 0.5 s). We thus assume activation timing approximately corresponds
to "muscle" torque timing. Therefore, we use the sign of τ to predict muscle activation (Figs. 2 & 3):
positive τ (or negative τmuscle) corresponds to muscle activation on the right side of the body.
2 Results and Discussion
We find that phase lags between internal torque and curvature in the model can explain the NPL
between electromyogram (EMG) and curvature seen in experiments.
τ displays a traveling wave
pattern and positive τ occurs in a range close to that of measured EMG activation (Figs. 3 & 4).
Without corrections from body passive forces or consideration of muscle physiology or body structure,
the average phase difference between the beginning and ending of positive τ in the model compared
to EMG onset and offset in experiments is less than 5%, where 2π is the range of possible phase
lags (Video S1 in SI). A large portion of the positive torque region overlaps with the region where the
curvature decreases (negative κ), but the positive τ region lags the negative κ region near the head and
leads it near the tail. The agreement between experiment and theory is striking, particularly because
our model has no fitting parameters; we posit this is largely a consequence of the simple movement
and the relatively simple but strong environmental interaction.
To gain more insight into how the phase lags arise due to torque contributions from different parts
of the body, we consider a simplified case where amplitude is small, forward motion is negligible, and
the resistive force is viscous. This makes analytical calculation of torque straightforward but does not
change the results qualitatively. In this simpler case, the torque from the fore-aft forces is negligible,
and only the lateral force (per unit length) fy(x, t) = −c y(x, t) = −cA cos(x + t) need be considered.
For simplicity and to separate the effects, we first neglect yaw motion. The torque can be calculated
analytically from Equation 3: τ (x0, t) = (2π − x0) sin(t) − cos(x0 + t) + cos(t). For example, if we take
two points x1 = π and x2 = π − ∆ near the middle of the body, we obtain τ1 = 3.7Ac sin(t + φ) and
τ2 = (3.7 + 1.7∆)Ac sin(t + φ − 0.29∆), where φ = 0.57. The NPL is still captured since the phase
difference between τ2 and τ1 is a fraction (0.29) of ∆, the phase difference between κ1 and κ2. The
torque contribution can be approximately divided into three parts, as follows:
6
2
π
e
m
T
i
π
.
κ(+)
.
κ(−)
0
Tail
0
Onset
π
x
Model
Experiment
Offset
.
κ(−)
2
π
Head
Figure 4: The predicted onset and offset of muscle activation from the model (green lines) compared
to EMG measurements from the sandfish experiment (black error bars indicates standard deviation,
adapted from [26]). Gray areas indicate the periods of negative κ. A/λ = 0.22 and the model sandfish
body is oriented at a downward entry angle of 22 degrees relative to the horizontal. The corresponding
positions of the electrodes in the model are approximated based on the curvature phases.
7
}
(4)
τ2 =Z 2π
x2
}
{z
head
}
{z
head
τ1 =Z 2π
x1
fy(z, t)(z − x1) dz
≈ δfy(x1, t)δ/2
+Z 2π−δ
x1+δ
fy(z, t)(z − x1) dz
local
{z
+ δfy(2π, t)(2π − x1)
fy(z, t)(z − x2) dz
≈ δfy(x2, t)δ/2
+Z 2π−δ
x2+δ
fy(z, t)(z − x2) dz
local
{z
+ δfy(2π, t)(2π − x2)
}
where δ is a small length. The phase difference between the torque contributions from local forces
for the two points is ∆, which is the same as the phase difference of other local variables (e.g. κ)
on the traveling wave (Fig. 5). In contrast, the phase of the torque transmitted from a distant point
on the body (e.g. the head) is the same for both points (even though the magnitude differs). This
synchronized torque contribution can be thought of as either a standing wave or a traveling wave
with infinite speed. Because of the combination of the torques from local and distant forces and the
continuous force distribution, the net phase difference between τ2 and τ1 is less than ∆ and the torque
wave speed is greater than the curvature wave speed. A similar analysis can be performed if the
integration is done on the posterior side of the body (toward the tail).
The balance of torque on the body leads to an overall yaw motion, whose phase is the same along
the body. Superposition of yaw motion and lateral motion of the body results in variation of both the
magnitude and phase of the lateral motion along the body in the lab frame (see Fig. S2 and derivation
in SI). However, the overall speed of the lateral displacement in the lab frame is the same as the
prescribed lateral displacement (sinusoidal wave) in the body frame. Therefore, the yaw motion only
changes the relative phase between the curvature wave and the apparent displacement (or force) wave
locally. Since the only requirement for this mechanism is a traveling wave pattern of force, it predicts
the NPL are general for torques from distributed forces. As shown in Figure 6a, the localized elastic
and damping forces by definition have constant phase differences with the curvature. In accord with
previous studies [30, 28], our calculations show that the relative phase between the torque from inertial
forces and curvature advances in the posterior direction. However, the overall phase of the inertial
torque advances by about 0.4π compared to the sandfish EMG phase. The phase lags persist if the
granular resistive forces in the model are replaced with viscous resistive forces, which low Reynolds
number swimmers like nematodes experience [13].
Although passive body forces are not responsible for the NPL, they can still influence the observed
pattern. For example, we find that the inclusion of viscous forces in the body shifts the phase of the
torque in granular media toward the phase pattern produced from only viscous forces (dash-dotted red
lines in Fig. 6a). That is, the phase difference between the torque and κ is smaller and the torque wave
speed is smaller, in accord with previous studies in fluids [29, 10]. This suggests that the small internal
viscous forces within the body may partially account for the phase differences we observe between the
torque from resistive forces and EMG. For swimming in a high Reynolds number fluid, the muscle
activation duration is in general smaller than those observed for the sandfish (≈ 0.5) [23]. Previous
studies (e.g. [29]) suggest that the torque from external forces may be overcome by passive elements
8
F2
F1
Traveling wave
X2
X
1
Head
Fh
Figure 5: Local and distant forces contribute to the torque on two points near the middle of the body.
Green arrows represent the forces on the body. The red and orange arrows indicate the force adjacent
to x2 and x1, respectively. The blue arrow indicates the force at the head (Fh). Note that this is
an analysis for a small amplitude case, and the lateral displacement is exaggerated in the figure for
visibility.
of the body. The nearly 0.5 duty factor of the muscle is evidence that resistive forces dominate in a
granular environment and the slight decrease of the duty factor (a relatively larger decrease is typical
during swimming in fluids [23]) implies passive forces play a small role for swimming of the sandfish.
Variations of locomotor kinematics also affect the timing of the torque (Fig. 6b). For example,
a downward entry angle (observed in the animal experiments [26]) advances the phase of the torque
compared to the horizontal swimming case. This is because when the body is oriented downward,
the head, which has a more advanced phase, contributes more to the overall torque due to the head's
greater depth and correspondingly larger resistive forces. Also, a larger number of periods (longer
body and smaller ratio of wave length to body length) both delays the phase of torque and reduces
the torque wave speed. The phase shift is due to the contribution of the extra tail length, where the
phase of the force lags that at anterior positions. The effect of period (body length) can be used to
estimate the error in timing that may occur due to neglecting the tail after 1.2 SVL: The error should
be a small fraction of the difference between the 1.2-period case and the control case. Further, we
found that a smaller undulation amplitude reduced the variation in torque wave speed.
The time delay between EMG activation and force production ([36]) might affect the phase lag
timing of EMG activation, but we argue that this delay is small compared to the typical period of
undulation for the sandfish (≈2 ms latency compared to ≈500 ms undulation period).
If the time
delay was significant and approximately constant, the EMG-curvature phase relation would change for
different frequencies.
3 Conclusions
We developed a theory to explain the basic control signals needed to generate a particular undulatory
movement pattern in a sand-swimming lizard. We abstracted the nervous/musculoskeletal system
by assuming that passive body forces are small and that internal torque is synchronized with neural
activation timing; this abstraction revealed that the NPL is intrinsic to undulatory locomotion provided
that distributed forces, such as resistive or inertial forces, play major roles. For undulatory locomotion
in other environments, the principle of the simultaneous response to distant torques should also apply,
9
2
π
-
a
-
Damping
Elastic
Inertial
Granular
Low Re
e
m
T
i
π
b
1.2 period
Horizontal
Angle
Small A
0
0
Tail x Head
π
π 0
2
Tail x Head
π
2
π
Figure 6: NPL for varying model parameters. (a) The starting (slightly thicker lines) and ending of
positive torque generated by granular force (solid green lines), viscous fluid force (solid black lines),
inertial force (solid magenta lines), damping force (dash-dotted red lines), and elastic force from the
body (dotted blue lines). (b) The beginning and ending of positive torque when the model sandfish
swims in a horizontal plane (solid black line, control case), at an entry angle of 22 degrees (solid green
line), at a small amplitude A/λ = 0.05 (dotted blue line) , and with a 1.2 periods of wave on its body
(dash-dotted red line). We aligned the head for the 1.2 period case and only the anterior portion (1
period of the wave) is shown in the figure. Gray areas indicate negative κ.
10
though quantities such as the phase of the force will differ from the sandfish case. Building on this
principle could help future studies explain other variations of the NPL.
Because we now have a system in which experiment and theory are in quantitative agreement, we
can begin to develop more detailed models (i.e. anchors [27])which answer specific questions about
nervous system control, muscle configuration, morphology, etc. For example, it has been established
that the intersegmental coordination of neural oscillators along the body of swimmers is influenced
by sensory feedback ([37]). Detailed models of central pattern generators (CPGs), sensory neurons
and muscles can be used to understand how external torque and neural activation interact so that the
intersegmental phase lags produce single period sinusoidal motion. As such, a hierarchy of anchors can
be used to generate testable hypotheses and understand actuation timing for animals in a variety of
environments.
More broadly, we have demonstrated that the seemingly specific and peculiar sand-swimming be-
havior could be an excellent system in which to develop quantitative models of neuromechanics. Due
to relatively simple but dominant environmental interactions, the neuromechanical control pattern is
greatly constrained by the environment. In addition, the granular RFT provides an excellent model
for interaction with the substrate; this is in contrast to locomotion in true fluids in which more com-
plex theories [38] are needed to quantitatively compare experiment and model. We hypothesize that
by studying subarenaceous animals within dry and saturated granular substrates (like those on the
bottom of the ocean floor), animal models with potentially fewer parameters can be analyzed in detail.
This in turn can help provide guidance for the design and control of artificial undulatory locomotors
in complex environments [39, 40, 41]. Better physical models can also improve our understanding of
the biological systems.
4 Materials and Methods
4.1 EMG Recordings
Previous work [26] using a micro-CT scan of a single sandfish revealed 26 vertebrae in the trunk
and more than 13 anterior caudal vertebrae in the tail. The iliocostalis musculature was targeted for
implantation and is located on the dorso-lateral portion of the trunk. Dissection revealed qualitatively
similar muscle morphology to that described for Iguana iguana [42, 43]; where iliocostalis musculature
spanned approximately 1 vertebrae.
Electrodes were implanted in one side of the body at 0.3 (magenta), 0.5 (green), 0.7 (blue), 0.9
(yellow) snout-vent length (SVL)(Fig. 1b) where the average SVL was 8.9 cm (N = 5 animals). EMG
data used in this paper were taken from n=37 sandfish swimming trials. The EMG signal was filtered
with a second-order Chebyshev filter and rectified in order to facilitate EMG burst detection. A burst
threshold was set equal to the mean of this rectified-filtered EMG trace. Burst onset was defined as
the time when the filtered EMG signal exceeded the threshold and afterwards remained above it for
a minimum of 0.04s. EMG burst offset was defined as when the filtered EMG signal became lower
than the threshold and remained below for at least 0.08 s [44]. This burst detection was necessary to
exclude small voltage changes that did not constitute an EMG burst, such as noise due to movement
artifact. See [26] for more details on the EMG recording and analysis technique.
4.2 Dynamic Bending Tests
Three anesthetized sandfish (mass = 15, 16 and 25 g) were gently clamped at approximately 0.5 SVL
and at 0.6 SVL (Fig. 7a) with adjustable grips. The grips were attached to a rigid platform and to a
rotating platform, respectively. A motor rotated the anterior region of the sandfish through ± 15◦ for
3 cycles at angular velocities of 1, 10 and 20◦/s. The first and last half cycle were excluded from the
11
analysis due to varying rotation velocities. The anterior end of the sandfish was clamped to a platform
with two strain gages (Omega, KFG-3-120-C1-11L1M2R) used to record resulting torques. Signals were
amplified (INA125P; Digi-Key) by 5000 before data acquisition and analyzed using custom software
(LabVIEW, NI, Austin, TX, USA). Black points were marked on the animal midline at increments of
0.1 SVL. The best fit line through the markers circled in red were used to calculate the angle θ.
Body stiffness, K, was estimated using the slope of the best fit line through the torque-angular
displacement curve (Fig. 7b) for a single cycle (n=8 trials each). Using a viscoelastic model (or Voigt
model), the viscous damping coefficient, c, was approximated by quantifying the viscous torque (τv)
at zero angular displacement during steady-state rotation and dividing by the angular speed ( θ), (i.e.
c = τv(θ = 0)/ θ).
For the hysteretic damping model, the structural damping coefficient h was proportional to angular
displacement, θ and π/2 out phase. The loss factor η = h/K. h was estimated by finding the torque
at zero displacement during steady state rotation and dividing by the maximum angular displacement:
h = τh(θ = 0)/θmax. The area contained within the work loop (Eloss) was determined using polynomial
fits to the torque vs angle curves for increasing and decreasing angle.
To interrogate stiffness and damping coefficients at higher speeds, we repeated the experiment with
one of the sandfish (animal 2, Fig. S3) using angular speeds of 1, 10, 20, 50 and 100◦/s and compared
with previous results.
We substituted the values during sand-swimming (angular excursion of 30◦ and angular velocity
of 240◦/s) and the average calculated K and c at 20◦/s into our viscous damping model to estimate
torques during sand-swimming (Figs. 7c & S4). For hysteretic damping, we estimated the damping
torque at 240◦/s by extending the trend line between calculated torque from hysteretic damping and
angular speed between 20 and 100◦/s.
4.3 Pendulum swing tests
In the second technique, sandfish were modeled as a physical pendulum to estimate K, c and η at
higher angular frequencies. The same sandfish were used as in the previous experiment. Animals were
oriented vertically and clamped at approximately 0.5 SVL. The tail of the sandfish was bent upward
and released, allowing the body to swing freely. The sandfish body was modeled as a rigid cylinder
and the tail as a cone with uniform density. The angular motion, θ was fit to a damped harmonic
oscillator:
I θ + c θ + Kθ + mgdCoM sin(θ) = 0,
(5)
where dCoM is the distance from the point of rotation to the center of mass, m is the mass of the
unclamped portion of the sandfish, I is the moment of inertia, and θ is the angular acceleration.
Angular motion during the first half cycle after the tail was released was neglected due to large angles
and body bending. θ was measured between the 0.5 and 0.8 SVL body positions.
We also fit the motion using a hysteretic damping model:
I θ + (1 + iη)Kθ + mgdCoM sin(θ) = 0.
(6)
For both models, we used the small angle approximation sin(θ) ≈ θ. Best fit parameters were de-
termined using minimization techniques (Matlab, Mathworks, Natick, MA, USA). Both viscous and
hysteretic models fit the angular displacement trajectory well (r2 < 0.9). See SI for experimental setup
diagrams and detailed results.
12
Foam
Animal 1
a
Axis of
Rotation
Rotating Platform
Grips
Rotate
0.08
b
)
m
c
.
N
(
τ
0.04
0
-0.04
-0.08
-10
c
)
m
c
.
N
τ
(
3
2
1
τ rms(0.62 SVL)
τ rms(0.75 SVL)
τ rms(0.25 SVL)
0
θ (deg.)
10
0.15
0.1
0.05
0
40
80
120
0
0
100
.
θ (deg./ s)
200
(15◦) , decreases to ≈ −0.26 rad.
Figure 7: Experimental measurements of sandfish body elasticity and damping. (a) Top view of the
setup. (b) Representative work loop for Animal 1 with angular velocities of 1◦/s (red), 10◦/s (blue) and
20◦/s (green). The direction is indicated by the arrows, where the angle is initially zero and increases
to ≈ 0.26 rad.
(−15◦), then returns to zero. Work loops are
shown for constant angular velocity (i.e. the system has reached a steady state). Dotted trajectories
represent the experimentally recorded force; solid curves are the best polynomial fits. (c) Estimated
average torque using calculated elastic (circles) and damping (triangles) coefficients compared to the
maximum τrms due to external forces calculated from discrete element simulation [34] (0.62 body
length (BL))(dashed horizontal black line). Average τrms for 0.25 BL(light gray dashed line) and 0.75
BL (dark gray dashed line) are also shown. For animal 1 (red), 2 (blue) and 3 (green) coefficients were
measured for 1, 10 and 20◦/s. In animal 4 (magenta) coefficients were measured for 1, 10, 20, 50 and
100 ◦/s. The average angular speed the sandfish operates at is 240 ◦/s. Inset shows zoomed in region
of data in the main figure (units are the same). 13
5 Acknowledgments
We thank Paul B. Umbanhowar, Silas Alben, George Lauder, Tom Daniel, and Robert J. Full for
helpful discussion and Humaira Taz for assistance with construction of experimental apparatus. We
thank Elizabeth A. Gozal for providing the EMG analysis code (SpinalMOD). This work was supported
by NSF Physics of Living Systems (PoLS) grants No. PHY-0749991 and No. PHY-1150760, the ARL
MAST CTA W911NF-11-1-0514, and the Burroughs Wellcome Fund.
References
[1] Holmes P, Full RJ, Koditschek D, Guckenheimer J (2006) The dynamics of legged locomotion:
Models, analyses, and challenges. SIAM Review 48:207 -- 304.
[2] Ijspeert A (2008) 2008 special issue: Central pattern generators for locomotion control in animals
and robots: A review. Neural Networks 21:642 -- 653.
[3] Alexander RM (2003) Principles of Animal Locomotion (Princeton University Press, Princeton,
USA).
[4] Nishikawa K, et al. (2007) Neuromechanics: an integrative approach for understanding motor
control. Integrative and Comparative Biology 47:16 -- 54.
[5] Orlovsky GN, Deliagina T, Grillner S, Orlovskii G, Grillner S (1999) Neuronal control of locomo-
tion: from mollusc to man (Oxford University Press, New York, USA).
[6] Josephson R (1985) The mechanical power output of a tettigoniid wing muscle during singing and
flight. Journal of Experimental Biology 117:357 -- 368.
[7] Hof AL (1984) EMG and muscle force: an introduction. Human Movement Science 3:119 -- 153.
[8] Tytell E, Holmes P, Cohen A (2011) Spikes alone do not behavior make: why neuroscience needs
biomechanics. Current Opinion in Neurobiology 21:816 -- 822.
[9] Chiel H, Ting L, Ekeberg O, Hartmann M (2009) The brain in its body: motor control and sensing
in a biomechanical context. The Journal of Neuroscience 29:12807 -- 12814.
[10] Tytell E, Hsu C, Williams T, Cohen A, Fauci L (2010) Interactions between internal forces, body
stiffness, and fluid environment in a neuromechanical model of lamprey swimming. Proceedings of
the National Academy of Sciences 107:19832 -- 19837.
[11] Dickinson MH, et al. (2000) How animals move: An integrative view. Science 288:100.
[12] Schmitt J, Holmes P (2000) Mechanical models for insect locomotion: dynamics and stability in
the horizontal plane - ii. application. Biological Cybernetics 83:517 -- 527.
[13] Cohen N, Boyle J (2010) Swimming at low reynolds number: a beginners guide to undulatory
locomotion. Contemporary Physics 51:103 -- 123.
[14] Gray J, Hancock G (1955) The propulsion of sea-urchin spermatozoa. Journal of Experimental
Biology 32:802 -- 814.
[15] Hu D, Nirody J, Scott T, Shelley M (2009) The mechanics of slithering locomotion. Proceedings
of the National Academy of Sciences 106:10081 -- 10085.
14
[16] Jayne B, Lauder G (1995) Speed effects on midline kinematics during steady undulatory swimming
of largemouth bass, micropterus salmoides. Journal of Experimental Biology 198:585 -- 602.
[17] Sfakiotakis M, Lane D, Davies J (1999) Review of fish swimming modes for aquatic locomotion.
IEEE Journal of Oceanic Engineering 24:237 -- 252.
[18] Maladen RD, Ding Y, Li C, Goldman DI (2009) Undulatory Swimming in Sand: subsurface
Locomotion of the sandfish lizard. Science 325:314 -- 318.
[19] Wright C, et al. (2007) Design of a modular snake robot. In Proceedings of the IEEE International
Conference on Intelligent Robots and Systems (IEEE), pp. 2609 -- 2614.
[20] Crespi A, Ijspeert A (2008) Online optimization of swimming and crawling in an amphibious
snake robot. IEEE Transactions on Robotics 24:75 -- 87.
[21] Choset H, et al. (2000) Design and motion planning for serpentine robots. In Proceedings of SPIE,
vol. 3990, p. 148.
[22] Jayne B, Lauder G (1995) Red muscle motor patterns during steady swimming in largemouth
bass: effects of speed and correlations with axial kinematics. Journal of Experimental Biology
198:1575 -- 1587.
[23] Wardle C, Videler J, Altringham J (1995) Tuning in to fish swimming waves: body form, swimming
mode and muscle function. Journal of Experimental Biology 198:1629 -- 1636.
[24] Gillis G (1998) Environmental effects on undulatory locomotion in the American eel Anguilla
rostrata: kinematics in water and on land. Journal of Experimental Biology 201:949.
[25] Williams L, et al. (1989) Locomotion in lamprey and trout: the relative timing of activation and
movement. Journal of Experimental Biology 143:559 -- 566.
[26] Sharpe S, Ding Y, Goldman D (2013) Environmental interaction influences muscle activation strat-
egy during sand-swimming in the sandfish lizard Scincus scincus. The Journal of Experimental
Biology 216:260 -- 274.
[27] Full RJ, Koditschek DE (1999) Templates and anchors: Neuromechanical hypotheses of legged
locomotion on land. Journal of experimental biology 202:3325 -- 3332.
[28] McMillen T, Williams T, Holmes P (2008) Nonlinear muscles, passive viscoelasticity and body ta-
per conspire to create neuromechanical phase lags in anguilliform swimmers. PLoS Computational
Biology 4:e1000157.
[29] Cheng J, Pedley T, Altringham J (1998) A continuous dynamic beam model for swimming fish.
Philosophical Transactions of the Royal Society of London Series B: Biological Sciences 353:981 --
997.
[30] Bowtell G, Williams T (1991) Anguilliform body dynamics: modelling the interaction between
muscle activation and body curvature. Philosophical Transactions of the Royal Society of London
Series B: Biological Sciences 334:385 -- 390.
[31] Chen J, Friesen W, Iwasaki T (2011) Mechanisms underlying rhythmic locomotion: body -- fluid
interaction in undulatory swimming. Journal of Experimental Biology 214:561 -- 574.
[32] Pedley T, Hill S (1999) Large-amplitude undulatory fish swimming: fluid mechanics coupled to
internal mechanics. Journal of Experimental Biology 202:3431 -- 3438.
15
[33] Fang-Yen C, et al. (2010) Biomechanical analysis of gait adaptation in the nematode Caenorhab-
ditis elegans. Proceedings of the National Academy of Sciences 107:20323 -- 20328.
[34] Ding Y, Sharpe SS, Masse A, Goldman DI (2012) Mechanics of undulatory swimming in a frictional
fluid. PLoS Computational Biology 8:e1002810.
[35] Maladen RD, Ding Y, Umbanhowar PB, Kamor A, Goldman DI (2011) Mechanical models of
sandfish locomotion reveal principles of high performance subsurface sand-swimming. Journal of
the Royal Society Interface 8:1332 -- 1345.
[36] Daley MA, Biewener AA (2003) Muscle force-length dynamics during level versus incline loco-
motion: a comparison of in vivo performance of two guinea fowl ankle extensors. Journal of
Experimental Biology 206:2941 -- 2958.
[37] Marder E, Bucher D, Schulz D, Taylor A (2005) Invertebrate central pattern generation moves
along. Current Biology 15:R685 -- R699.
[38] Rodenborn B, Chen CH, Swinney HL, Liu B, Zhang H (2013) Propulsion of microorganisms by a
helical flagellum. Proceedings of the National Academy of Sciences 110:E338 -- E347.
[39] Hirose S, Biologically inspired robots: Snake-like locomotors and manipulators,(1993).
[40] Roper D, Sharma S, Sutton R, Culverhouse P (2011) A review of developments towards biologically
inspired propulsion systems for autonomous underwater vehicles. Proceedings of the Institution of
Mechanical Engineers, Part M: Journal of Engineering for the Maritime Environment 225:77 -- 96.
[41] Colgate J, Lynch K (2004) Mechanics and control of swimming: a review. Oceanic Engineering,
IEEE Journal of 29:660 -- 673.
[42] Carrier D (1990) Activity of the hypaxial muscles during walking in the lizard Iguana iguana. J
Exp Biol 152:453 -- 470.
[43] Ritter D (1996) Axial muscle function during lizard locomotion. Journal of experimental biology
199:2499 -- 2510.
[44] Hochman S, et al. (2012) Enabling techniques for in vitro studies on mammalian spinal locomotor
mechanisms. Frontiers in Bioscience: a Journal and Virtual Library 17:2158.
16
|
1802.08608 | 2 | 1802 | 2018-07-26T05:14:55 | Tubulation pattern of membrane vesicles coated with bio filaments | [
"physics.bio-ph",
"cond-mat.soft"
] | Narrow membrane tubes are commonly pulled out from the surface of phospholipid vesicles using forces applied either through laser or magnetic tweezers or through the action of processive motor proteins. Recent examples have emerged where such tubes spontaneously grow from vesicles coated with bioactive cytoskeletal filaments (e.g. FtsZ, microtubule) in the presence GTP. We show how a soft vesicle deforms due to the interplay between its topology, local curvature and the forces due to the active filaments. We present results from Dynamically Triangulated Monte Carlo simulations of a spherical continuum membrane coated with a nematic field and show how the intrinsic curvature of the filaments and their ordering interactions drive membrane tubulation. We predict interesting patterns of nematic defects, on curved 2D membrane surfaces, which promote tube formation. Implication of our model for more dynamic cases where vesicles coated with an active mixture of microtubule and myosin show shape oscillation, are also discussed. All these cases point to a common theme that defect locations on 2D membrane surfaces are hot spots of membrane deformation activity. | physics.bio-ph | physics |
Spontaneous tubulation of membrane vesicles coated with bio-active filaments.
1 Gaurav Kumar, 2 N. Ramakrishnan, and 1 Anirban Sain∗
1 Physics Department, Indian Institute of Technology-Bombay,
Powai, Mumbai, 400076, India.
2 Department of Bioengineering,
University of Pennsylvania, Philadelphia, PA, 19104, USA
Narrow membrane tubes are commonly pulled out from the surface of phospholipid vesicles using
forces applied either through laser or magnetic tweezers or through the action of processive motor
proteins. Recent examples have emerged where array of such tubes spontaneously grow from vesicles
coated with bioactive cytoskeletal filaments (e.g. FtsZ, microtubule) in the presence GTP/ATP. We
show how a soft vesicle deforms as a result of the interplay between its topology, local curvature and
the forces due to filament bundles. We present results from Dynamically Triangulated Monte Carlo
simulations of a spherical continuum membrane coated with a nematic field (the filaments) and
show how the intrinsic curvature of the filaments and their bundling interactions drive membrane
tubulation. We predict interesting patterns consisting of large number of nematic defects which
accompany tubulation. A common theme emerges that defect locations on vesicle surfaces are
hot spots of membrane deformation activity, which could be useful for vesicle origami. Although
our equilibrium model is not applicable to the nonequlibrium shape dynamics exhibited by active
microtubule coated vesicles, we show that some the features like size dependent vesicle shape can
still be understood from our equilibrium model. E-mail: [email protected]
PACS numbers: PACS : 87.16.-b, 87.15.Aa, 81.40.Jj, 87.15.Rn
INTRODUCTION
Narrow membrane tubes are ubiquitous in eukaryotic
cells and are essential for a number of cell functions in-
cluding signalling and trafficking. Examples of tubular
structures are the axons and dendrites of nerve cells,
tubular networks in the Golgi body and the Endoplas-
mic reticulum. The curvature energy of a tubular pro-
trusion E ∝ κ(l/r), where l is the length of the tube,
r the radius and κ the membrane bending rigidity. As
a result the formation of narrow tubes (characterised by
l >> r) requires extremely high energies and hence they
are not expected to be stable unless stabilised by external
forces [1]. Such forces can be applied by laser tweezers
[2], suction pressure [3] or processive molecular motors.
[4]. Forces internal to the cell or membrane vesicles, for
example, growing bundles of actin [5, 6] or microtubule
(MT) filaments [7, 8] can also push out membrane pro-
trusions which can grow into long tubes upon further
polymerization of the filaments. Tubulation can also be
promoted by spiral shaped protein filaments like dynamin
[9? ] and ESCRT-III [11], or proteins with curved do-
mains like Bar [12], ENTH [13], Exo70 [14], etc that wrap
around the tube. Such induced membrane curvatures are
interpreted as local spontaneous curvature [15]. Sponta-
neous membrane curvature can also naturally result from
lipid heterogenity [16] or lipid tilt [17] in a phospholipid
membrane.
Here we address experiments [18] where an array of
membrane tubes emerge spontaneously due to surface
active filamentous proteins FtsZ, attached to the outer
surface of artificial membrane vesicles (liposomes).
In-
terestingly, here the FtsZ filaments align with the tube
axis, instead of wrapping around the tube. We attribute
420 nm
FIG. 1: (color online) Tubulation of a vesicle, (a) to (d), due
to FtsZ filaments (fluorescent regions), in real time (minutes).
(e) shows loss of spherical shape due to excessive tubulation.
(f) and (g) show high resolution images focused at the equa-
torial plane and the surface, respectively, of the tubulated
vesicle. Nearly regular pattern of tubulation sites are visible.
(h) shows reverse tubulation i.e., convex bulges on the vesicle
surface accompanied by membrane invaginations protruding
inward. (i) shows two different ways of anchoring FtsZ onto
the membrane, causing opposite membrane curvatures.
(j)
shows parallel arrangement of FtsZ filaments along the tubu-
lar axis. Scale bars are 10µm in the upper and middle rows,
and 420nm in the lower row. Adapted from Fig.1,2,3 and 6
of Ref[18], with permission.
this to the attractive interaction among parallel FtsZ fil-
aments, in the presence of GTP, which gives rise to FtsZ
bundles [19]. In our model, this bundling interaction will
be the main driving force towards multiple tube forma-
tion. Secondly, the regular arrangement of many tubes
on the curved vesicle surface is rather striking and has
not been modelled before. As we will see later that the
formation of such tubes cannot be reliably modelled by
considering just a single axisymmetric tube because col-
lective effects turn out to be important. For example,
in the present case of tubular array the boundary con-
dition at the base of a single tube is not axisymmetric
but instead has five or six-fold discrete rotational sym-
metry depending on the number of neighbouring tubes
surrounding it.
In this experiment Osawa et. al. [18] attached mem-
brane targeted FtsZ filaments onto the outer surface
of large Giant Unilamellar Vesicles (GUVs) of diame-
ter ∼ 10µm and observed tubes of diameter 50 − 200nm
grow spontaneously. The filaments caused either con-
cave depressions or convex bulges on the membrane, as
illustrated in Fig. 1(i), depending on whether the mem-
brane anchors (FtsA) were attached to the C terminal
or the N terminal of the filaments, respectively. Depend-
ing on the type of attachment, membrane tubes either
grew outward (see time series in Fig. 1(a-d)) or inward
(see Fig. 1h). While elongation of tubes was observed
only when GTP was in abundance, tubes shrunk when
medium was depleted of GTP, implying the essential role
played by GTP. But unlike in active systems here GTP
hydrolysis neither generate any active forces nor move
the FtsZ filaments. In fact, here the FtsZ filaments are
anchored to the membrane surface and they can at the
most diffuse slowly.
In order to understand the FtsZ assisted membrane
tubulation it is necessary to review the well known phys-
ical properties of FtsZ filaments. FtsZ filaments have in-
trinsic curvature of order (100-200 nm)−1 and they play
important role during cytokinesis of rod shaped bacte-
ria like E. coli and B. subtilis [20, 21]. In the presence of
GTP, FtsZ filaments also condense into bundles via weak,
lateral, inter-filament attraction [22]. These bundles can
also locally bend membrane and generate constriction
forces (few pN) [23] on relatively wide membrane tubes
of diameter 1 − 2µm. Given these properties it has been
unclear how these tubes form.
In this article, treating
the filaments as a nematic field and using a generalised
Canham-Helfrich [24 -- 28] model for the lipid membrane,
we suggest a mechanism for vesicle tubulation and pre-
dict the pattern of arrangement of FtsZ filaments on the
vesicle.
According to the Hairy-Ball theorem [29] filaments, ap-
proximated as nematics here, cannot be arranged on a
closed surface (tangentially) without forming topological
defects and the topological charge of these defects must
add up to two. Arrangement of these defects is also a the-
oretically well studied problem [29 -- 31]. Keber et al [8]
has recently studied active dynamics of such defects on
spherical as well as deformed vesicles. However the influ-
ence of the nematics (filaments) on the elastic membrane
is also very interesting because it leads to nontrivial de-
formation of the vesicle shape [8, 18], which is the focus
2
of this article. In fact, in the case of FtsZ shape defor-
mation of the vesicle is accompanied by large number of
high energy defects which was not considered before.
Qualitatively, Fig. 1b-d,f suggest that concave depres-
sions and tubes go hand in hand and we can guess that
tubes may form at the junctions where such concave
patches meet. The question then arises, how will these
concave patches arrange themselves on a spherical sur-
face. Fig. 1g further shows a regular arrangement of
bright patches, with coordination number five or six. Can
we then approximate the patches as the pentagons and
hexagons that cover the surface of a soccer ball ? The
tubes would then emerge from the vertices. We will find
out later that a variant of this picture is correct.
MODEL
In our coarse-grained approach, we model the FtsZ
coated membrane as a nematic field adhering to a de-
formable fluid membrane surface. This model was devel-
oped and many of its properties were studied by one of
the authors [26 -- 28]. We will later highlight the results
that are new here.
In this model the local orientation
of the nematic field is denoted by the unit vector n(~r)
which lies in the local tangent plane of the membrane
and is free to rotate in this plane. Filament-membrane
interactions are modelled as anisotropic spontaneous cur-
vatures of the membrane, in the vicinity of the filament,
while filament-filament interactions are modelled by the
splay and bend terms of the Frank's free energy for ne-
matic liquid crystals. The total energy is
E = Z dAh κ
+ Z dAh K1
2
2
κk
2
(Hk − ck)2 +
( ∇.t)2i.
K3
2
κ⊥
2
(H⊥ − c⊥)2i
(1)
(2H)2 +
( ∇.n)2 +
Here the first term is the Canham-Helfrich elastic en-
ergy for membranes [24, 25] with bending rigidity κ and
membrane mean curvature H = (c1 +c2)/2. Here, c1 and
c2 are the local principal curvatures on the membrane
surface along orthogonal tangent vectors t1 and t2. κk
and κ⊥ are the induced membrane bending rigidities and
ck and c⊥ are the induced intrinsic curvatures along n,
the orientation of the filament in the local tangent plane,
and t, its perpendicular direction, respectively. Origin of
a nonzero c⊥, an induced curvature perpendicular to the
filament, is not obvious. In fact it will turn out to be the
driving force towards filament bundling which accompa-
nies membrane tubulation. Tubulation can also occur
due to the ck term alone, however it does not promote
formation of long straight filament bundles. The filament
orientations on the tubes are different in these two cases.
The membrane curvature along n and t are given by
Hk = c1 cos2 φ + c2 sin2 φ and H⊥ = c1 sin2 φ + c2 cos2 φ,
where φ denotes the angle between filament orientation
n and principal direction t1. K1 and K3 are the splay
and bend elastic constants for the in plane nematic in-
teractions and ∇ is the covariant derivative on the curved
surface [32].
As in Ref [26], we use a discrete form of this energy
functional to perform Monte-Carlo simulations on a tri-
angulated membrane, to study the equilibrium shapes.
Each vertex i hosts a orientation vector ni. In particu-
lar, we use the standard Lebwohl-Lasher model Enn =
2 ) , [33] to mimic the in plane ne-
matic interaction terms. Here ǫLL is strength of the ne-
matic interaction, in one constant approximation( K1 =
−ǫLLPi>j( 3
2 (ni.nj)2 − 1
K3 ), and the sum Pi>j is over all the nearest neigh-
bour (i, j) vertices on the triangulated grid, promoting
alignment among the neighbouring orientation vectors.
Models with anisotropic membrane curvatures have
been developed and used by various authors [34? , 35].
But all of these efforts focussed on membrane structures
that are axis-symmeric in nature and mainly sought to
study formation of single tubes. Previous work by one of
the authors [26 -- 28] focussed on effect of nonzero κk and
ck. This amounts to introducing one preferred length
scale, namely the radius of the tube.
In contrast, the
present work focusses on the effects of non-zero κ⊥ and
c⊥.
In fact, using these parameters we account for a
new physical effect, namely bundling of filaments due to
inter-filament attraction.
Nonzero values of κ⊥ and c⊥ induce parallel alignment
of filaments, on the outer surface of a membrane tube, all
pointing along the tube axis (z). Higher the c⊥ smaller is
the tube radius; this is equivalent to filament bundling.
In our simulation, with fixed number of vertices, the den-
sity of vertices goes up relatively at the high curvature re-
gions of the triangulated surface. This makes the number
density of vertices relatively higher at the tube surfaces,
leading to higher filament density on the tubes, consis-
tent with filament bundle picture. Since our model does
not distinguish between the outer and the inner surfaces
of a tube, this effect can also account for MT bundles
which are responsible for pushing out membrane tubes
from a vesicle. Note that, the nematic interaction term
in our model is also minimized when the filaments align
with the tube axis (z). This is also true for continuum
Frank's free-energy where both splay the bend terms are
minimized. But the nematic interaction term does not
control the tube radius because it is not dependent on
the curvature of the underlying membrane as long as all
the nematics on the tube are aligned along the common
z axis of the straight tube.
We will later show that bundling assisted membrane
tubulation is possible even in the absence of intrinsic fil-
ament curvature (i.e., ck = 0). This will be relevant for
MT filaments which does not have any intrinsic curva-
ture. But in the case of FtsZ, since intrinsic curvature
and bundling both are known to be involved, we will use
nonzero values for both ck and c⊥ for modeling FtsZ.
3
RESULTS
Monte-Carlo simulations of our model (Eqn. (1)) show
that, for ck < 0, c⊥ > 0, with ck << c⊥, regularly
spaced narrow tubes emerge (see Fig. 2) from an initially
spherical vesicle. We checked that c⊥ mainly controls
the tube radii and ck controls the curvature of the valleys.
However the total number of tubes is a joint effect of both
c⊥ and ck. Reversing the signs of the intrinsic curvatures,
i.e., ck > 0, c⊥ < 0 tubes grow into the vesicle (see Fig. 2
a,d,e). Since FtsZ coated membrane is expected to be
stiffer than the bare bilayer membrane (for which κ ∼
20kBT , we set κk = 35kBT and κ⊥ = 25kBT . We found
that setting κ = 0 or κ = 20kBT only makes minor
qualitative differences in the tubular structures; tubes
are thicker and little less in number with κ = 20kBT .
Furthermore, nonzero κ offers an initial energy barrier
for tube nucleation making tubulation slow, which could
be bypassed by raising the temperature temporarily in
our Monte-Carlo simulation.
FIG. 2:
(color online) Growth of tubes, outward (a,b,c)
or inward (a,d,c), depend on the signs of ck and c⊥. We
use ck = −0.05 and c⊥ = 1 for outward growing tubes, and
reverse the signs of ck and c⊥ to get inward growing tubes.
Other parameters are κk = 35, κ⊥ = 25 and ǫLL = 3, in
units of kBT . Volume has been held fixed. In (e) the vesicle
is sectioned to make the inner tubes visible The results from
constant area ensemble are nearly identical at same parameter
values.
In Fig.3 we show few shapes, some of them quite un-
expected ones, when the bundling effect is switched off
i.e., κ⊥ = 0. Yet, tubes or inverted tubes can form (see
Fig.3a and b, respectively) provided ck ≫ 1/R, the cur-
vature of the original spherical vesicle. However, Fig.3c
and d, are examples where tubes do not form due to vari-
ation of other parameters. The criteria for emergence of
a tube and dependence of its radius (r) on various pa-
rameters of the model can be understood by the stability
analysis of a single tube given in Ref[27]. However the
specific arrangement of many tubes on the vesicle surface
4
Fig.4 shows that, even without any intrinsic curvature
(i.e., ck = 0) of the adhering filaments, tubes may still
emerge due to bundling interaction among filaments. As
we discuss later this may be the driving force for tubu-
lation for vesicles coated with MT, which does not have
any intrinsic curvature like FtsZ.
FIG. 3:
(color online) Shapes with fixed c⊥ = 0, κk = 25 and
ǫLL = 5.0, with other parameters varying, a) κ = 0, κ⊥ =
25, ck = 1.0 b) same as (a) except ck = −1.0, c) κ = 10, κ⊥ =
0, ck = 1.0 and d) same as (c) except ck = −1.0.
cannot be inferred analytically. Minimization of the free
energy of a straight, uniform cylindrical tube [27] yields
equilibrium tube radius
r2 =
κ
2 (κk + κ⊥) + κkκ⊥
κkκ⊥(ck + c⊥)2
(2)
and inclination (φ) of the nematic [27] on the surface of
the tube,
cos2 φ =
(κkck − κ⊥c⊥)r + κ⊥
κk + κ⊥
(3)
Note that, if any of κk or κ⊥ is zero, while κ 6= 0, the
radius becomes infinite, siganalling that tubes will not
form, which is the case for Fig.3-c,d. On the other hand if
κ = 0 then the tube radius is (ck +c⊥)−1. This is the case
for Fig.3-a,b, where only ck is nonzero and same for Fig.2
where c⊥ dominates over ck. In this respect we note that
in Ref[26], Fig.11 despite κ⊥ = 0, and κ, κk 6= 0, tubes
still emerged. But these tubes were bent and also had
nonuniform cross-section which violate the assumption in
the stability analysis. Furthermore, Eq.3 suggests that
for physically acceptable solutions the right hand side
(r.h.s.) must be between zero and one. Indeed φ = π/2
for Fig.2, while φ = 0, for Fig.3a, consistent with the
longitudinal and azimuthal orientation of filaments on
the tube, respectively. But when the r.h.s.
is less than
zero a boundary minima occurs for the free energy at
φ = π/2 and similarly for the r.h.s larger than one the free
energy at φ = 0 is the physically acceptable minimum
value. Correspondingly, the formula for equilibrium r
also changes (see Ref[27]).
FIG. 4:
(color online) Tubulation even without intrin-
sic filament curvature, i.e., ck = 0. Other parameters are
κ = 0, κk = κ⊥ = 25, c⊥ = 1 and ǫLL = 5.0
In previous work on this model [26 -- 28] neither volume
nor area of the deformed vesicle was conserved during
Monte-Carlo simulation.
In this work we ran separate
simulations for fixed volume and fixed area ensembles.
We noticed no significant change in the qualitative results
between these two different ensembles, except that the
simulation with constant area was relatively faster. Ex-
perimentally, total fluorescence of the membrane bound
FtsZ was found to increase during tubulation [18]. This
could be due to addition of higher density of filaments
to the tubes from the solution or could be due to release
of entropic membrane folds leading to increase in vesicle
area, attracting more FtsZ from the solution.
In Osawa's experiments [18] tubes grew indefinitely
(see Fig 1e) as long as GTP was supplied and they shrunk
when GTP was depleted.
In our simulation too tube
growth did not stop when the vesicle volume was kept
fixed but the area was unconstrained. GTP depletion is
known to cause two things: a) it switches off the attrac-
tion between FtsZ filaments and as a result they unbun-
dle [19], and b) it raises the intrinsic curvature of the FtsZ
filaments [18]. We implement GTP depletion by turning
off κ⊥ to zero which causes the tubes to shrink. We also
raise the value of ck simultaneously, which marginally
increases undulations on the nearly spherical vesicle.
It can be argued that in a Monte-Carlo simulation only
the final equilibrium state has physical relevance, while
the intermediate states may not follow the actual system
kinetics. Therefore we considered ensembles where both
the volume and the area of the vesicle were fixed. This is
a typical recipe for simulating, for example, red blood cell
shapes [37]. We fixed the area to be about 10% excess
over that of the corresponding sphere at a given volume.
This recipe is physically meaningful for our FtsZ case
also because there is a limit to the maximum amount
of area that a vesicle can reserve in the form of mem-
brane folds, beyond which area stretching elasticity be-
comes important. For this ensemble, our system took the
same pathway as before (when area was not fixed) but
the tubulation was not indefinite and the system reached
equilibrium, as in Fig 1d,f or g. This ensemble becomes
particularly important for the MT induced tubulation of
deflated membrane vesicles in the experiments of Keber
et al [8], which we discuss now.
Keber et. al. [8] attempted to mimic the active cell
cortex by assembling arrays of microtubules on the inner
surface of a spherical membrane vesicle, along with high
concentration of kinesin motors. This rendered activity
to the MT layer. MT filaments showed incessant growth,
shrinkage, bundling and sliding motion at the spherical
surface. Four topological nematic defects of charge +1/2
formed at the surface and showed interesting spatio-
temporal dynamics. Upon partial deflation these vesicles
deformed into ellipsoidal shape and four narrow tubes
emerged (Fig.6c), with MT bundles inside. The tubes
kept changing their positions. Furthermore, for smaller
vesicles only two tubes persisted (Fig.6d) along with
shape changing dynamics. Few groups [38, 39], including
Keber et al [8], modelled the active MT-membrane sys-
tem as nematics on a spherical surface and studied the
effective dynamics of the nematic defects. But so far no
study has included vesicle deformation due to the active
MT-motor dynamics.
Although the dynamics of this active system is be-
yond the scope of our equilibrium Monte-Carlo simula-
tion, some aspects of this system can still be understood
from equilibrium physics of our model studied at phys-
ically relevant parameters. Towards this we set ck = 0
in our model, since MT does not have intrinsic curvature
like FtsZ, and retain nonzero values of kk, k⊥, c⊥ and ǫLL.
It is well known that polymerization of actin and MT
bundles, even in the absence of activity can give rise to
tubular protrusions [5, 7]. That implies that membrane
tubulation can occur at equilibrium also, however the dy-
namics of the tube and the vesicle has a purely nonequi-
librium origin (motor acivity). The location of the tubes
in Keber et. al's experiment [8] coincides with the lo-
cations of +1/2 nematic defects, where elastic strain in
the nematic field is the highest. However, tubulation
also requires extra area. So when the vesicle is tout (i.e.,
membrane tension high), the MT bundles cannot over-
come the membrane tension to push out tubes, instead
the bundles buckle. This effect was seen in Ref[7]. For
such an undeformed spherical vesicle the equilibrium po-
sitions of four +1/2 defects are the vertices of a symmet-
ric tetrahedron, since +1/2 defects repel each other. Due
to activity in Ref[8], such an equilibrium state turns out
to be unstable and the defects oscillate between a tetra-
5
hedral and a planar configuration [8]. Although, we can-
not model this instability, our Monte-Carlo simulation at
high bending rigidity (κ) and at high temperature shows
that the defect positions indeed fluctuate between tetra-
hedron and planar configurations (see Fig.5) indicating
closeness of these two states in the equilibrium energy
landscape.
FIG. 5:
(color online) The simplest vesicle shape, at fixed
volume, with only nonzero isotropic bending modulus κ = 20
and weak nematic interaction ǫLL = 1.5. The four s = +1/2
defects (blue solid circles) fluctuate, at finite temperature,
between (a) tetrahedral and (b) planar arrangements. The
lines are the connectors between the centroid and the defects.
The numerous small dots (red) show the vertices of the tri-
angulated mesh outlining the surface of the vesicle. Here the
nematic interaction was chosen weak because at stronger in-
teraction the free energy barrier between (a) and (b) states
will be higher which cannot be overcome only by thermal fluc-
tuations. In the active case [8] active fluctuations make these
states unstable.
Furthermore, when Keber et al [8] deflate the vesicle,
by applying hypertonic stress, it amounts to generating
excess area (as the volume reduces) and as a result the
vesicle deforms. To mimic this system we allow about
10% excess area for the vesicle as before, but switch
to relatively stronger nematic and bundling interactions:
ǫLL = 9 and c⊥ = 1.2. In addition, we set ck = 0 as MT
filaments do not have intrinsic curvature. When started
from a sphere, the vesicle grows into an ellipsoid utilising
the extra area. The four +1/2 defects arrange themselves
into two pairs, one pair each migrating approximately to
the opposite ends of the major axis. The strong nematic
interaction ensures that more defects are not nucleated
as defects possess high elastic energy. Subsequently four
tubes emerge from the four defects (Fig.6a). However as
our system do not have any active dynamics the tubes do
not change their positions. Interestingly, at these same
parameter values, when we reduce the volume of the vesi-
cle further to one third of its value (with corresponding
reduction in area) only two tubes emerged (see Fig.6b).
This occurred via merging of the defect pair into one +1
defect, at each end of the major axis. The corresponding
size difference in the experimental figures are indicated
in Fig.6c and d.
In our
the filaments
cannot
slide/translate, unlike in the active case of Keber et al
equilibrium model
6
[8], however the defects can move due to rearrangement
of the nematics. As mentioned earlier, the areal density
of nematics can be non-uniform in our model, because al-
though there is one filament per vertex, the areal density
of vertices is higher on the tubes than in the valleys. This
could very well be the case in the experiment. For exam-
ple, parallel arrangement of filaments on tubes, parallel
to the tube axis, will produce denser filament coverage
and than that in the valleys. This is consistent with
higher fluorescent intensity from the tubes, reported in
Osawa's experiments [18]
FIG. 7:
(color online) Arrangement of nematic defects on a
tubulating vesicle. The red circles denote s = −1/2 defects
which encircle the aster like s = +1 defects. The parameters
are same as in Fig. 2.
of s = +1/2, in nearly 2D epithelial tissue. It turns out
that energetically it is favourable to co-localize nematic
defects and high membrane curvature. However in-plane
arrangement of nematic defects on a closed surfaces must
obey Poincare's index theorem (popularly known as the
hairy ball theorem) which states that the total topologi-
cal charge (s) of all defects on a sphere must add up to 2
(more generally to 2(1 − g) on a surface of genus g) [29].
This constrains the number and location of the tubes on
the vesicle surface. What precedes a tube is a defect of
charge s = 1 with an aster like arrangement of nematics.
The membrane deforms into a pointed structure (vertex)
around the aster and produce a tube. The total positive
charge increases with the number of vertices but this is
efficiently nullified by the local arrangement of s = −1/2
defects around each vertex. Each vertex is surrounded by
typically five or six, and seldom seven, other vertices. Ac-
cordingly, the central vertex forms five, six or seven trian-
gles, with the surrounding vertices. Each triangle forms a
concave valley and hosts a s = −1/2 defect (red circles in
Fig.3). But each s = −1/2 defect is shared by three s = 1
defects. To compute total defect charge on the vesicle,
we consider contributions from polygons formed by the
−1/2 defects (red circles in Fig. 7). The net charge of a
pentagon is 1 + 5 × (−1/2) × (1/3) = 1/6, while that for
a hexagon is 1 + 6 × (−1/2) × (1/3) = 0, and heptagon is
−1/6. One realisation of this pentagon-hexagon arrange-
ment is the minimal soccer ball structure which has 12
pentagons, 20 hexagons and equivalently, thirty-two +1
and sixty −1/2 defects, with total charge adding up to
2. The simple picture that emerges is that each polygon,
made of s = −1/2 defects, hosts one s = +1 defect (tube)
at its centre. So in the framework of soccer ball struc-
ture the polygonal units are formed by the −1/2 (valley)
defects at the vertices. In our initial guess we considered
the hexagonal lattice dual to this where tubes were at
FIG. 6:
(color online) Effect of excess area: at fixed volume
the area is fixed at 10% excess over the corresponding sphere.
(a) and (b) model partially deflated vesicles (c and d) of Ref[8]
(with permission) where initially spherical vesicles deformed
into ellipsoidal shapes due to excess available area. (d) had
lesser volume than (c), which resulted into half the number
of tubes in (d). Model parameters are the same for (a) and
(b) : κ0 = 0, κk = 25, κ⊥ = 20, ck = 0, c⊥ = 1.2, ǫLL = 9,
except that volume of (b) is 1/3-rd of (a). Despite presence
of bundling effect c⊥, strong nematic interaction allowed only
finite number of defects (and hence tubes) as defects cause
high elastic energy.
Nematic defects turn out to be hot spots of activity
on the membrane.
In our simulation tubes grow from
sites of nematic defects with a positive defect charge.
The charge s of a nematic defect is defined as the to-
tal amount of rotation a nematic director undergoes as a
closed loop is traversed around a defect. In Keber et. al.
[8] defect sites on ellipsoid shaped vesicle constitute weak
spots that encourage growth of MT bundles which in turn
induce tubulation. Ladoux's group [40] has shown cell
death and extrusion occurring predominantly at the sites
the vertices of the polygons. However, smaller and larger
vesicles have different number of defects, for example,
seventy-six −1/2 and forty +1 defects, still adding up
to 2 (in fact Fig. 7 has this structure). The regular ar-
rangement of valleys and tubes now can be matched with
Figs. 1f and 1g, respectively. Note from Fig. 7, that each
tube (+1 defect) typically has five to six nearest neigh-
bours while each valley (-1/2 defect) has three nearest
neighbours. Charge cancellation among nematic defects,
on deformed axis-symmetric vesicles, have been discussed
in Ref[41].
In summary, we showed how filament bundle induced
tubulation, can be modeled by introducing an induced
anisotropic membrane curvature perpendicular to the fil-
ament's alignment. We showed that, within our model,
either of ck and c⊥ is individually capable of causing
tubulation. But the corresponding inclination of the ne-
matic field on the tube surface is different in the two
cases. Furthermore, when nematic interaction is weak (in
case of FtsZ) the vesicle allows formation of many defects,
and subsequently many tubes where bundling interaction
leads to maximum energy gain. On the other hand for
MT with strong nematic interaction only minimal num-
ber of defects and tubes form. Althouh GTP/ATP hy-
drolysis is common to both cases (FtsZ/MT), we empha-
size that the FtsZ case is not active in the conventional
sense, because no physical filament movement is gener-
ated (in the absence of motors) unlike the case of kinesin
driven MT case. Although our model cannot address the
active MT dynamics, our simulations indicates that some
of the features observed in the active MT induced vesi-
cle shapes may have equilibrium origin. This includes
emergence of the ellipsoidal shape upon reduction of the
vesicle volume and the influence of the vesicle volume in
determining the number of tubes (four versus two). The
novel filament arrangement on the deformed vesicle sur-
face is an interesting outcome of our numerical investiga-
tion and is difficult to predict a priori. This link between
filament arrangement and vesicle shape may be useful for
vesicle origami.
Gaurav kumar would like to acknowledge financial sup-
port from CSIR (India).
∗ Electronic address: [email protected]
[1] I. Derenyi, F. Julicher, and J. Prost . Physical review
letters 88, 238101 (2002)
7
biology 12 ,e1004982 (2016).
[6] Mesarec, L., W. G o z d z, S. Kralj, M. Fosnaric, S.
Penic, V. Kralj-Iglic, and A. Iglic,. European Biophysics
Journal 46 , 705-718 (2017).
[7] Elbaum, M., D. K. Fygenson, and A. Libchaber,. Phys-
ical Review Letters 76 , 4078(1996).
[8] F. C. Keber, E. Loiseau, T. Sanchez, S. J. DeCamp, L.
Giomi, M. J. Bowick, M. C. Marchetti, Z. Dogic, and A.
R. Bausch. Science 345 ,1135 (2014).
[9] Low, H. H., C. Sachse, L. A. Amos, and J. L owe. Cell
139, 13421352 (2009).
[10] Hinshaw, J.. Annual review of cell and developmental
biology 16, 483519 (2000).
[11] Adell, A. Y., D. Teis. FEBS letters 585, 31913196
(2011).
[12] Peter, B. J., H. M. Kent, I. G. Mills, Y. Vallis, P. J. G.
Butler, P. R. Evans, and H. T. McMahon,. Science 303
, 495499 (2004).
[13] Ford, M. G., I. G. Mills, B. J. Peter, Y. Vallis, G. J.
Prae- fcke, P. R. Evans, and H. T. McMahon, . Nature
, 419:361 (2002).
[14] Zhao, Y., J. Liu, C. Yang, B. R. Capraro, T. Baumgart,
R. P. Bradley, N. Ramakrishnan, X. Xu, R. Radhakrish-
nan,T. Svitkina,. Developmental cell 26 ,:266278 (2013).
[15] Zimmerberg, J., and M. M. Kozlov . Nature reviews
Molecular cell biology , 7:9 (2006).
[16] Y. Li, R. Lipowsky, and R. Dimova. Proceedings of the
National Academy of Sciences 108 ,4731 (2011).
[17] Schnur, J. M. . Science 262 , :16691676 (1993).
[18] M. Osawa, D. E. Anderson, and H. P. Erickson. The
EMBO journal 28,3476 (2009).
[19] Erickson, H. P., D. W. Taylor, K. A. Taylor, and D.
Bramhill. Proceedings of the National Academy of Sci-
ences 93 , 519523 (1996).
[20] X. Yang, Z. Lyu, A. Miguel, R. McQuillen, K. C. Huang
and J. Xiao. Science 355 ,744 (2017).
[21] A. W. Bisson-Filho, Y.-P. Hsu, G. R. Squyres, E. Kuru,
F. Wu, C. Jukes, Y. Sun, C. Dekker, S. Holden, M. S.
VanNieuwenhze, et al.. Science 355,739 (2017).
[22] B. Ghosh and A. Sain.
Physical review letters
101,178101 (2008).
[23] M. Osawa, D. E. Anderson, and H. P. Erickson. Science
320 ,792 (2008).
[24] Canham, P. B.. Journal of theoretical biology 26 ,6181
(1970).
[25] Helfrich, W . Z. Naturforsch. C 81 , :041922 (1973).
[26] N. Ramakrishnan, P. S. Kumar, and J. H. Ipsen. Physical
Review E 81,041922 (2010).
[27] N. Ramakrishnan, P. S. Kumar, and J. H. Ipsen. Bio-
physical journal 104 ,1018 (2013).
[28] N. Ramakrishnan, P. Kumar, and J. H. Ipsen. Macro-
molecular Theory and Simulations 20 ,446 (2011).
[29] Lubensky, T., and J. Prost . Journal de Physique II 2,
:371382 ( 1992 ).
[30] Nelson, D. R . Nano Letters 2 , :11251129 ( 2002).
[31] Shin, H., M. J. Bowick, and X. Xing. Physical review
[2] F. Hochmuth, J.-Y. Shao, J. Dai, and M. P. Sheetz .
letters 101 , :037802 ( 2008 ).
Biophysical journal 70, 358 (1996).
[32] Chaikin, P. M., and T. C. Lubensky,. Cambridge uni-
[3] R. Hochmuth, H. Wiles, E. Evans, and J. McCown .
versity press , : ( 2000 ).
Biophysical journal 39,83 (1982).
[4] A. Roux, G. Cappello, J. Cartaud, J. Prost, B. Goud
and P. Bassereau . Proceedings of the National Academy
of Sciences 99,5394 (2002).
[5] Weichsel, J., and P. L. Geissler. PLoS computational
[33] P. Lebwohl. Phys. Rev. A 6 ,426 (1972).
[34] Iglic, A., B. Babnik, U. Gimsa, and V. Kralj-Iglic. Jour-
nal of Physics A: Mathematical and General 38 , :8527
( 2005 ).
[35] G omez-Llobregat, J., F. El as-Wolff, and M. Lind en .
Biophysical journal 110 ,197204 (2016 ).
[39] F. Alaimo, C. Kohler, and A. Voigt. arXiv preprintarXiv:
[36] Mesarec, L., A. Iglic, and S. Kralj, . Adv Biomed Res
1703.03707 (2017).
Innov ,4:2. ( 2018).
[37] HW, G. L., M. Wortis, and R. Mukhopadhyay,. Proceed-
ings of the National Academy of Sciences 99 ,1676616769
(2002 ).
[40] T. B. Saw, A. Doostmohammadi, V. Nier, L. Kocgo-
zlu, S. Thampi, Y. Toyama, P. Marcq, C. T. Lim, J. M.
Yeomans, and B. Ladoux. Nature 544 ,212 (2017).
[41] Mesarec, L., W. G o z d z, A. Iglic, and S. Kralj.
[38] Zhang, R., Y. Zhou, M. Rahimi, and J. J. De Pablo .
Scientific reports 6 ,27117. (2016 ).
Nature communications 13 ,13483 (2016 ).
8
|
1503.03853 | 1 | 1503 | 2015-03-12T19:38:03 | Gyrification from constrained cortical expansion | [
"physics.bio-ph",
"cond-mat.soft",
"nlin.PS",
"q-bio.TO"
] | The exterior of the mammalian brain - the cerebral cortex - has a conserved layered structure whose thickness varies little across species. However, selection pressures over evolutionary time scales have led to cortices that have a large surface area to volume ratio in some organisms, with the result that the brain is strongly convoluted into sulci and gyri. Here we show that the gyrification can arise as a nonlinear consequence of a simple mechanical instability driven by tangential expansion of the gray matter constrained by the white matter. A physical mimic of the process using a layered swelling gel captures the essence of the mechanism, and numerical simulations of the brain treated as a soft solid lead to the formation of cusped sulci and smooth gyri similar to those in the brain. The resulting gyrification patterns are a function of relative cortical expansion and relative thickness (compared with brain size), and are consistent with observations of a wide range of brains, ranging from smooth to highly convoluted. Furthermore, this dependence on two simple geometric parameters that characterize the brain also allows us to qualitatively explain how variations in these parameters lead to anatomical anomalies in such situations as polymicrogyria, pachygyria, and lissencephalia. | physics.bio-ph | physics | Gyrification from constrained cortical expansion
Tuomas Tallinena, Jun Young Chungb, John S. Bigginsc, and L. Mahadevand,e,f,1
aDepartment of Physics and Nanoscience Center, University of Jyväskylä, FI-40014 Jyväskylä, Finland; bSchool of Engineering and Applied Sciences,
Harvard University, Cambridge, MA 02138; cCavendish Laboratory, Cambridge University, Cambridge CB3 0HE, United Kingdom; dWyss Institute for
Biologically Inspired Engineering, eKavli Institute for Bionano Science and Technology, and fSchool of Engineering and Applied Sciences, Department
of Organismic and Evolutionary Biology, and Department of Physics, Harvard University, Cambridge, MA 02138
Edited* by John W. Hutchinson, Harvard University, Cambridge, MA, and approved July 15, 2014 (received for review April 1, 2014)
The exterior of the mammalian brain—the cerebral cortex—has
a conserved layered structure whose thickness varies little across
species. However, selection pressures over evolutionary time scales
have led to cortices that have a large surface area to volume ratio in
some organisms, with the result that the brain is strongly convo-
luted into sulci and gyri. Here we show that the gyrification can
arise as a nonlinear consequence of a simple mechanical instability
driven by tangential expansion of the gray matter constrained by
the white matter. A physical mimic of the process using a layered
swelling gel captures the essence of the mechanism, and numerical
simulations of the brain treated as a soft solid lead to the formation
of cusped sulci and smooth gyri similar to those in the brain. The
resulting gyrification patterns are a function of relative cortical ex-
pansion and relative thickness (compared with brain size), and are
consistent with observations of a wide range of brains, ranging
from smooth to highly convoluted. Furthermore, this dependence
on two simple geometric parameters that characterize the brain
also allows us to qualitatively explain how variations in these
parameters lead to anatomical anomalies in such situations as poly-
microgyria, pachygyria, and lissencephalia.
brain morphogenesis elastic instability
The mammalian brain is functionally and anatomically com-
plex. Over the years, accumulating evidence (1, 2) shows
that there are strong anatomical correlates of its information-
processing ability; indeed the iconic convoluted shape of the
human brain is itself used as a symbol of its functional com-
plexity. This convoluted (gyrified) shape is associated with the
rapid expansion of the cerebral cortex. Understanding the evo-
lutionary and developmental origins of the cortical expansion
(1–6) and their mechanistic role in gyrification is thus an im-
portant question that needs to be answered to decipher the
functional complexity of the brain.
Historically there have been three broad hypotheses about the
origin of sulci and gyri. The first is that gyri rise above sulci by
growing more (7), requiring the pattern of sulci and gyri to be
laid down before the cortex folds, presumably by a chemical
morphogen. There is no evidence for this mechanism. The sec-
ond hypothesis considers that the outer gray matter consists of
neurons, and the inner white matter is largely long thin axons
that connect the neurons to each other and to other parts of the
nervous system and proposes that these axons pull mechanically,
drawing together highly interconnected regions of gray matter to
form gyri (8–10). However, recent experimental evidence (11)
shows that axonal tension when present is weak and arises deep
in the white matter and is thus insufficient to explain the strongly
deformed gyri and sulci. The third hypothesis is that the gray
matter simply grows more than the white matter, an experi-
mentally confirmed fact, leading to a mechanical buckling that
shapes the cortex (11–14). Evidence for this hypothesis has re-
cently been provided by observations of mechanical stresses in
developing ferret brains (11), which were found to be in patterns
irreconcilable with the axonal tension hypothesis. In addition,
experiments show that sulci and gyri can be induced in usually
smooth-brained mice by genetic manipulations that promote
cortical expansion (15, 16), suggesting that gyrification results
from an unregulated and unpatterned growth of the cortex rel-
ative to sublayers.
Nevertheless, there is as yet no explicit biologically and physi-
cally plausible model that can convincingly reproduce individual
sulci and gyri, let alone the complex patterns of sulci and gyri
found in the brain. Early attempts to mechanically model brain
folding (13) were rooted in the physics of wrinkling and assumed
a thin stiff layer of gray matter that grows relative to a thick soft
substrate of white matter. This model falls short in two ways. First,
the gray matter is neither thin nor stiff relative to the white matter
(17, 18). Second, this model predicts smooth sinusoidal wrinkling
patterns, sketched in Fig. 1A, whereas even lightly folded brains
have smooth gyri but cusped sulci. More complicated mechanical
models including, e.g., elasto-plasticity and stress-related growth
(14, 19, 20), lead to varying morphologies, but all produced simple
smooth convolutions rather than cusped sulci.
A fundamentally different mechanical instability that occurs
on the surface of a uniformly compressed soft solid (21, 22) has
recently been exposed and clarified, theoretically, computation-
ally, and experimentally (23–26). This sulcification instability
arises under sufficient compression leading to the folding of the
soft surface to form cusped sulci via a strongly subcritical tran-
sition. In Fig. 1B, we show a geometry dual to that associated
with wrinkling: A soft layer of gray matter grows on a stiff white-
matter substrate. Unlike wrinkling, this instability can produce
the cusped centers of sulci, but the flat bottom of the gray matter
is not seen in the brain. This is a consequence of the assumption
that the gray matter is much softer than the white matter—in
reality the two have very similar stiffnesses (17, 18). We are thus
led to the final simple alternate, sketched in Fig. 1C, where the
stiffnesses of the gray and white matter are assumed to be
identical. Such a system is subject to a cusp-forming sulcification
Significance
The convolutions of the human brain are a symbol of its
functional complexity and correlated with its information-
processing capacity. Conversely, loss of folds is correlated with
loss of function. But how did the outer surface of the brain, the
layered cortex of neuronal gray matter, get its folds? Guided
by prior experimental observations of the growth of the cortex
relative to the underlying white matter, we argue that these
folds arise due to a mechanical instability of a soft tissue that
grows nonuniformly. Numerical simulations and physical
mimics of the constrained growth of the cortex show how
compressive mechanical forces sculpt it to form characteristic
sulci and gyri, consistent with observations across species in
both normal and pathological situations.
Author contributions: T.T., J.Y.C., J.S.B., and L.M. designed research; T.T., J.Y.C., J.S.B., and
L.M. performed research; T.T., J.Y.C., J.S.B., and L.M. analyzed data; T.T. and J.Y.C. con-
tributed new reagents/analytic tools; and T.T., J.Y.C., J.S.B., and L.M. wrote the paper.
The authors declare no conflict of interest.
*This Direct Submission article had a prearranged editor.
1To whom correspondence should be addressed. Email: [email protected].
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1406015111/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1406015111
PNAS September 2, 2014 vol. 111 no. 35 12667–12672
L
A
C
I
S
Y
H
P
D
E
I
L
P
P
A
S
E
C
N
E
I
C
S
D
N
A
S
C
I
S
Y
H
P
O
I
B
Y
G
O
L
O
I
B
L
A
N
O
I
T
A
T
U
P
M
O
C
A
D
B
E
C
F
the z direction, although folding can only occur in the x – y plane;
we find that when transversely isotropic tangential expansion
exceeds g = gx = gz ≈ 1.29 sulcification of the gray matter becomes
energetically favorable over a smooth surface, and the gray matter
forms cusped folds largely internal to the gray matter and remi-
niscent of the folds in lightly sulcified brains such as the porcupine
(Fig. 2A). As gx is increased further (for simplicity gz = 1.29 was
fixed) the gray matter folds down into the white matter forming
a big cusped sulcus and smooth gyrus, reminiscent of the sulci and
gyri found in more folded brains such as a cat (Fig. 2B). Our plots
also indicate regions of compressive and tensile stress, which agree
with observations in developing ferret brains (11).
We plot the geometric characteristics of the sulcus, such as
depth and width, as a function of gx in Fig. 2C, which allow us to
establish several nontrivial similarities between our geometry
and actual brains (Fig. S2). After the transition from smooth to
A
B
C
Fig. 1. Wrinkling and sulcification in a layered material subject to differ-
ential growth. (A) If the growing gray matter is much stiffer than the white
matter it will wrinkle in a smooth sinusoidal way. (B) If the gray matter is
much softer than the white matter its surface will invaginate to form cusped
folds. (C) If the two layers have similar moduli the gray matter will both
wrinkle and cusp giving gyri and sulci. Physical realizations of A, B, and C,
based on differential swelling of a bilayer gel (Materials and Methods),
confirm this picture and are shown in D, E, and F, respectively.
instability discussed earlier, and can lead to an emergent pattern
very reminiscent of sulci and gyri in the brain.
Results
A physical experiment to mimic these patterns can be easily created
using a hemispherical polydimethylsiloxane (PDMS) gel coated
with a layer of PDMS that can swell by absorbing a solvent such
as hexanes (Materials and Methods). By varying cross-linking den-
sities we can prepare samples with different ratios of the moduli of
the two layers and capture both the wrinkled morphology shown in
Fig. 1D (when the outer layer is stiffer) and the sulcified mor-
phology shown in Fig. 1E (when the outer layer is softer). In par-
ticular, we see the appearance of brainlike morphologies with deep
sulci when the modulus ratio is close to unity (Fig. 1F).
To study gyrification quantitatively, we first construct a nu-
merical model in two dimensions. We start with a rectangular
domain consisting of a layer of gray matter on top of a deep layer
of white matter, both having the same uniform shear modulus μ.
The material is assumed to be neo-Hookean with volumetric
strain energy density
h
Tr
i
J−2=3 − 3
ðJ − 1Þ2;
[1]
FFT
W =
μ
2
+ K
2
where F is the deformation gradient, J = det(F), and the bulk
modulus K = 103μ makes the tissues almost incompressible. To
model growth of the gray matter relative to the white matter, we
apply a tangential growth profile,
gðyÞ = 1 +
α
1 + e10ðy=T−1Þ;
[2]
so that g = 1 in the white matter and g = 1+ α in the gray matter,
with a smoothed step at the interface (Fig. S1). Here, y is dis-
tance from the top surface in material coordinates, T is the un-
deformed thickness of the gray matter, and α controls the
magnitude of expansion. Later on we denote g ≡ 1 + α ≈ g(0).
We use a custom finite element method to minimize the elastic
energy (details Please select the in Materials and Methods). Our
2D plane-strain calculations also include constrained expansion in
Fig. 2. Formation of a minimal sulcus. The 2D sulci with tangential expansion
ratio of (A) g = 1.30 and (B) g = 2.25 of the gray matter (Eq. 2 and Fig. S1).
Coloring shows radial and circumferential tensile stress in the left and right sulci,
respectively. The stress is compressive in the noncolored areas. Grid lines corre-
spond to every 20 rows or columns of the numerical discretization with nodes.
The width W, depth D, and thickness of the gray matter in the sulcus (Ts) and
gyrus (Tg) are indicated in B. For comparison with observations of brains, we also
show sections of porcupine and cat brains, taken from www.brainmuseum.org.
(C) Scaled dimensions of the simulated sulcus (solid lines) as a function of g
compared with those in porcupine (triangles), cat (dots), and human (squares)
show that our model can capture the basic observed geometry. Width and depth
are given relative to the undeformed thickness T of the gray matter (for details
of the measurements and error bars, see Fig. S2).
12668 www.pnas.org/cgi/doi/10.1073/pnas.1406015111
Tallinen et al.
sulcified, the sulcal depth increases continuously although the
width between sulci is always finite, agreeing with observations in
weakly convoluted brains. In the high g regime, the optimal spacing
is about 4T whereas depth continues to increase, in agreement with
observations in highly folded brains. Finally, the deformed thick-
ness of the gray matter varies such that, at the gyrus, it is nearly
twice that at the base of the sulcus; the same pattern as seen in all
real brains. Our 2D model thus captures the essential features of
individual sulci and gyri and the intersulcal spacing.
Although sulci are fundamentally different from wrinkles,
a qualitative understanding of our results follows by using the
classical formula λ = 2πt[μ/(3μs)]1/3 for the wrinkling wavelength
of a compressed stiff film (modulus μ, thickness t) on a soft
substrate (modulus μs) (27). Extrapolating this to the case here
(μs = μ) yields λ ≈ 4.36t in rough agreement with the simulated
sulcal spacing. A rigorous analytical treatment of gyrification is,
however, presently out of reach due to the subcritical nature of
the instability that is accompanied by finite strains and cusplike
features. Although the underlying mechanical principle is that
the gray matter folds to relax its compressive stress and that is
balanced by deforming the white matter, we emphasize that the
details are quite different from wrinkling and buckling, because
sulcification is a scale-free nonlinear subcritical instability (24).
We now explore the patterns of sulci and gyri in 3D by mod-
eling the brain as a thick spherical shell, with outer radius R and
inner radius Ri = R/2, including both the gray and white matter;
we note that for such geometries the resulting gyrification pat-
terns are independent of the (presence or absence of a) core. As
in the 2D model above, the domain is assumed to be of uniform
elastic material described by Eq. 1, but for numerical conve-
nience we now adopt modest compressibility with K = 5μ, cor-
responding to Poisson’s ratio ν ≈ 0.4. Brain tissues actually show
time-dependent compressibility owing to poroelasticity (28), but
this is irrelevant over the long times associated with morpho-
genesis, when we may safely limit ourselves to considering just
elastic effects. We assume that tangential expansion, given by Eq. 2,
is transversely isotropic so that the area expansion is given by g2.
We model small brains as complete thick spherical shells and
patches of large brains as patches of thick spherical shells with
periodic boundary conditions along the edges, and discretize
them using tetrahedral elements (Materials and Methods).
G
Our 3D model of the brain has three geometrical parameters,
brain radius (size) R, cortical thickness T and the tangential ex-
pansion g2 with experimentally observable analogs. Indeed,
in
mammals the cortical thickness T, gray-matter volume VG and
white-matter volume VW are linked by robust scaling laws that relate
brains varying over a millionfold range in weight (29, 30). These laws
can be written, in dimensionless units, as T ∼ V 0:1
G and VW ∼ V 1:23
(29). Using the spherical geometry of our model, we can relate these
quantities to our model parameters as VG ∼ g2[R3 − (R − T)3] and
VW ∼ (R − T)3 − (R/2)3. These empirical scaling laws, together with
an estimate that g2 = 5 for R/T = 20, eliminate two degrees of
freedom from the model, leaving us with a single parameter
family of models describing brains of different sizes. This is shown
in Fig. 3 where we plot g2 against relative brain size R/T and
display images of real brains as well as numerically simulated
brain shapes for a range of representative relative sizes.
Our model correctly predicts that brains with R/T K 5 (cor-
responding to physical size of R ≈ 5 mm) should be smooth as g is
insufficient to cause buckling. Intermediate-size brains are cor-
rectly predicted to have isolated sulci that are largely localized
within the gray matter. Larger brains become increasingly folded,
with sulci penetrating the white matter and the brain surface
displaying complicated patterns of branched sulci, similar to
those in large real brains (Fig. 3). The degree of folding is con-
ventionally quantified by the gyrification index (GI), the ratio of
the surface area to the area of the convex hull. For the largest
brain that we simulate (R/T = 20, g2 = 5, corresponding to
L
A
C
I
S
Y
H
P
D
E
I
L
P
P
A
S
E
C
N
E
I
C
S
D
N
A
S
C
I
S
Y
H
P
O
I
B
Y
G
O
L
O
I
B
L
A
N
O
I
T
A
T
U
P
M
O
C
Fig. 3. Known empirical scaling laws for gray-matter volume and thickness
are mapped on a g2 vs. R/T diagram. Corresponding simulations for spherical
brain configurations, with images shown at a few points, show that the
surface remains smooth for the smallest brains, but becomes increasingly
folded as the brain size increases. We also show patterns for ellipsoidal
configurations (major axis = 1.5 × minor axes) that lead to anisotropic gyr-
ification. Images of rat, lemur, wolf, and human brains illustrate the in-
creasingly prominent folding with increasing size in real brains. Also shown
are images of our physical mimic of the brain using a swelling bilayer gel of
PDMS immersed in hexanes. The smooth initial state gives rise to gyrified
states for different relative sizes of the brain R/T = 10, 15 (see also Fig. 5). All
of the brain images are from www.brainmuseum.org.
physical size R ≈ 36 mm) we find GI ≈ 2.8, which can be compared
with the modestly larger human brain that has GI ≈ 3 in regions
that exclude the sylvian fissure (31). Most actual brain shapes de-
viate from spherical so that sulci, especially in small- and medium-
sized brains, tend to align with the direction of least curvature. By
repeating our calculations on an ellipsoidal geometry (Fig. 3) we
capture this qualitative trend. The numerical brain shapes are
complemented by experimental realizations in Fig. 3. Our bilayer
gel brain models (Materials and Methods) capture the realistic
sulcal spacing of about 4T and the qualitative trends in variation of
sulcal patterns with R/T up to modest-size brains.
Because many brain atlases show different sections of the brain
to highlight the anatomical complexity of the folds, we show
sections of our simulated patterns in Fig. 4A. For comparison, we
show sections from a raccoon brain, which has a similar size to the
simulated brain, and see that the two appear very similar. An
important observation that becomes apparent is that cutting
through gyri and sulci with various alignments with respect to the
section plane gives the impression of rather complex gyrification
and exaggerates depths of sulci, especially when the section plane
is off the center of curvature (plane 2; Fig. 4A). Simulated cross
sections display features such as buried gyri and regions with
disproportionally thick cortices seen in sections of real brains, but
they are really just geometric artifacts of sectioning.
As can be observed, the calculated gyri are rounded rather than
flattened as in, e.g., brain samples that have been fixed before re-
moval. Experimental evidence suggests,
in agreement with our
model, that the compressive constraint of the skull or meninges is
not required for gyrification (30), but it could affect the appearance
by flattening the gyral crowns. Our simulation of a brain confined by
a rigid shell that mimics the skull confirms this (Fig. 4B).
Tallinen et al.
PNAS September 2, 2014 vol. 111 no. 35 12669
A
B
C
D
Fig. 4.
(A) Sections of a simulated brain (section planes indicated at right)
are compared with coronal sections of a raccoon brain (from www.brain-
museum.org). Cuts through the center of the brain (Upper) and the off the
center (Lower) show that we can capture the hierarchical folds but em-
phasize how misleading sections can be in characterizing the sulcal archi-
tecture. (B) Confining our simulations with a uniform pressure of 0.7μ to
mimic the meninges and skull leads to a familiar flattened sulcal morphol-
ogy. (C) Changing the gray matter thickness in a small patch of the growing
cortex leads to morphologies similar to polymicrogyria in our simulations.
Here g2 = 5 and R/T = 20 except in the densely folded region where R/T = 40.
(D) A simulated brain of same physical size as that in C but with a thickened
cortex (R/T = 12) and reduced tangential expansion (g2 = 2) displays wide
gyri and shallow sulci reminiscent of pachygyric brains.
The empirical scaling laws we used to restrict the parameter
space of our model describe normal brains. Deviations from the
scaling laws relating gyrification index to the relative cortical
thickness are known to lead to some pathological morphologies in
brains. Our model can reproduce these variations by considering
deviations in the parameters R/T and g. For example, poly-
microgyria is typically associated with the gray matter having only
four layers rather than six (32), and hence being thinner. In the
context of our model, reducing T while maintaining g for a fixed
brain size results in more and smaller sulci and gyri, as shown in Fig.
4C, consistent with observations. Pachygyria and lissencephalia are
associated with reduced neuronal migration to the cortex and thus
less surface expansion leading to smaller values of g (33). In the
context of our model, smaller surface expansion g or a thicker cortex
T leads to fewer sulci and wider gyri, as shown in the simulations
in Fig. 4D, consistent with observations.
Discussion
Our mechanical model for gyrification based on biological obser-
vations and physical constraints uses realistic but simplified geom-
etries, growth profiles, and mechanical properties. Real brain tissues
are known to change their growth rates in response to stress (20),
and other mechanisms lead to relaxation of elastic stresses over
time. Nevertheless, our simple elastic description suffices to explain
the essentials of gyrification while more complicated inelastic
material behavior has a minor role. Our theory naturally leads to
cusped sulci and smooth gyri, and shows that gray-matter thickness
should vary by about a factor of two between gyral crowns and sulcal
fundi, and that sulcal depths should vary substantially relative to
gray-matter thickness whereas gyral spacings should be always about
4 times the gray-matter thickness; these are consistent with nu-
merous experimental observations (12, 30). Furthermore, the model
captures the variations in the patterns of sulci as a function of brain
size showing that small brains should be smooth and large brains
should have complex branched sulcal networks, in line with obser-
vations. Finally, our model predicts that gray matter is under com-
pression and white matter is under radial tension beneath gyri and
circumferential tension beneath sulci, also consistent with recent
observations (11). From a broader perspective, our study comple-
ments other recent studies that highlight the role of mechanical
forces on morphogenesis in such instances as gut patterning (34)
where patterns similar to gyri and sulci yield villi in the gut lumen.
The shape of the mammalian brain has a number of prominent
large-scale features: the longitudinal fissure, the sylvian fissure, and
many specific gyri and sulci with important functional correlates that
we cannot capture with our minimal model. Thus, our approach
should be seen as complementing the detailed variations in pro-
liferation patterns, cell migration, shape change, etc., that give rise to
cortical expansion and in turn are regulated by it. Indeed, the
growing insight into regulatory mechanisms of neuronal proliferation
and migration (3, 4, 5, 6) implicates genetic mutations responsible for
brain size (e.g., ASPM, CENPJ for microcephaly), cortex thickness
(e.g., GPR56 for polymicrogyria), and the gyrification index (e.g.,
RELN for lissencephaly), and thus links naturally to the parameters
in our model for brain morphology. Simultaneously, comparative
phylogenies that look at expression profiles for these genes (6) may
also provide an evolutionary window into brain gyrification, wherein
proximal genetic causes lead to pathways that allowed for the
growth of the brain and the expansion of the cortex, turning prim-
itive smooth brains to current gyrified brains, leaving open the
possibility that the process is now significantly regulated. Finally, our
geometric mechanics approach to gyrification might also serve as
a template for brain morphometrics in terms of experimentally
measurable geometrical parameters that vary across individuals,
complementing current geometric statistics approaches (35) by ac-
counting for the underlying physical processes.
Materials and Methods
Physical Models. To create a simulacrum of the brain, we used a poly(dime-
thylsiloxane) (PDMS) elastomer bilayer consisting of an outer layer coated on
the surface of a hemispherical core. Differential swelling of the top layer when
exposed to a solvent (hexanes) mimics constrained cortical expansion and leads
to a variety of surface morphologies. The hemispherical core is fabricated with
PDMS elastomer (Sylgard 184; Dow Chemical Co., Midland, MI) using replica
molding, using a 30:1 mass ratio of base monomer to cross linker that is mixed
and poured over a concave lens-shaped silicone mold with r ≈ 11 mm. The
mixture is left at room temperature to allow trapped air bubbles to escape
and then cured at 75 °C for about 30 min to produce a partially cured contact
surface. After cooling, the PDMS core is carefully peeled off from the mold.
The partially cured PDMS surface is sticky and thus is used to enhance the
interfacial bonding between the top layer and the core for the preparation of
a bilayer specimen; μ0c ≈ 100 kPa in the fully cured state (36). The specimen
geometry and the experimental setup are illustrated in Fig. 5.
The top layers are prepared with different compositions of the PDMS
mixture by a drop coating technique, in which layer thickness is controlled by
adjusting the amount and viscosity of uncured PDMS mixture deposited on
the hemispherical core in a preheated oven. The selected elastomer com-
position is first thoroughly mixed and degassed to eliminate air bubbles.
Then, drops of a mixture are deposited and uniformly spread onto the surface
of a preprepared PDMS core using a pipette. Finally, the specimen is com-
pletely cured at 75 °C overnight. The thickness of the prepared top layers
T0 is determined by comparing the optical images before and after layer
coating using an image analysis program.
The prepared bilayer specimen is fully immersed in a Petri dish filled with
hexanes (Sigma-Aldrich) at room temperature for a period t typically less than
12670 www.pnas.org/cgi/doi/10.1073/pnas.1406015111
Tallinen et al.
topmost rows of elements are included in the gray matter), to minimize the
effect of substrate thickness on the sulcus and gyrus. At each simulated value
of tangential expansion the sulcus is initiated by applying a downward force
to the surface node of a lateral boundary, and self-contact of the sulcus is
accounted for by preventing the surface crossing the vertical line that defines
the boundary. The initiating force is then removed, the system allowed to
relax, and the aspect ratio of the domain is adjusted quasistatically (by
sweeping the height of the domain over a finite interval enclosing the energy
minimum) to find the energetically optimal relative width of the gyrus.
The 3D simulations are based either on irregular tetrahedral meshes (small
brain simulations that implement a full spherical or ellipsoidal thick shell) or
a curved cubical mesh where each cube is divided to five tetrahedrons (large
brain simulations that implement a patch of a spherical thick shell), see Fig. 6.
Inner surfaces of the thick shells are clamped, but the use of free boundary
conditions or simulations with full solid spheres would yield similar results. The
irregular 3D meshes consist of ≈ 2 × 106 nodes (about 107 elements) and the
regular meshes consist of 320 × 320 × 80 nodes (about 4 × 107 elements). In
each case the mesh density is such that the gray-matter layer contains at least
eight layers of elements through its thickness. The spherical patch spans an
angle of π/2 about the x and z axes, with periodic boundary conditions so that
if a copy of the domain is rotated by an angle π/2 about the x or z axis, it would
connect seamlessly to the original domain. The simulations based on regular
meshes have small random spatial variations in growth to break the otherwise
perfect rotational symmetry. These random perturbations do not affect the
shape or size of sulci and gyri or qualitative features of the sulcal pattern, but
different perturbation fields produce different patterns because they are the
only mechanism breaking the symmetry in the system. In the simulations based
on irregular meshes, the mesh provides sufficient randomness.
h
A stress-free initial configuration of a tetrahedron is defined by the matrix
^A =
, where ^x1, ^x2 and ^x3 are vectors describing the tetrahe-
dron, assuming a Cartesian coordinate system. The deformed configuration
of the tetrahedron, including growth and elastic deformation, is defined by
^x2
^x3
^x1
i
A =½ x1 x2 x3 = FG ^A ,
[3]
where x1, x2, and x3 are the deformed basis vectors and F is the elastic de-
formation gradient. The growth tensor
G = gI +ð1 − gÞ^ns ⊗ ^ns
[4]
describes tangential expansion perpendicular to the surface normal ^ns, with
g given by Eq. 2. At each time step we obtain F from Eq. 3 by using
F = AðG ^AÞ−1. The Cauchy stress, i.e., the force per unit area in the deformed
configuration, is derived from the strain energy density WðFÞ (Eq. 1) by
σ = 1
J
∂W
∂F
FT,
[5]
where J = det(F). Surface traction of each deformed face (i = 1, 2, 3, 4) of the
tetrahedron is given by si = −σni, where ni are normals with lengths pro-
portional to the deformed areas of the faces. Nodal forces are obtained by
distributing the traction of each face equally for its three vertices.
Self-avoidance of the surface is implemented by preventing nodes pen-
etrating element faces at the surface. If a separation d between a node and
face is less than the contact offset h (we use h = a/3, where a is the mesh
spacing in the initial configuration) it is considered a contact and penalized
by an energy 4Ka2
is in-
terpolated to the vertex nodes of the face and an opposite force is given to
the node in contact with the face.
2. The contact force from this potential
h − d
h
Fig. 6. Cross-section views of 3D simulation geometries for small and large
brains in their initial undeformed states. The gray-matter thickness T, brain
radius R, and boundary conditions are indicated. A detailed image of the
regular mesh structure of the large brain domain shows the reflection
symmetry between every pair of elementary cubes that share a face.
L
A
C
I
S
Y
H
P
D
E
I
L
P
P
A
S
E
C
N
E
I
C
S
D
N
A
S
C
I
S
Y
H
P
O
I
B
Y
G
O
L
O
I
B
L
A
N
O
I
T
A
T
U
P
M
O
C
Fig. 5. A physical model of brainlike instability. To mimic the growth of the
gray matter in the brain, a hemispherical elastomer (radius r, shear modulus
μ0c) is coated with a top elastomer layer (thickness T0, shear modulus μ0t) that
swells by absorbing solvent over time t. Representative images of a bilayer
specimen in the initial (dried) state and swollen state (modulus ratio μt/μc ≈ 1)
are shown at right.
10 min. This exposure time is short enough to prevent the solvent from pene-
trating into the core, but is sufficient to swell the top layer. Immediately after
solvent exposure, the bilayer specimen is withdrawn from the bath and sub-
sequently imaged with a digital camera (Nikon D80) equipped with a zoom lens
(Sigma 105 mm f/2.8 EX DG Macro). Less polar solvents such as chloroform and
toluene (both having a similar high PDMS swelling coefficient; ref. 37) yield
similar results, but hexanes has a relatively low evaporation rate, which makes
the deswelling process slow enough to capture the surface patterns.
The shear modulus of the outer layer in the swollen state μt is lower than that
in the initial dry state (μ0t), but the modulus of the core remains unchanged over
the short period of solvent exposure time (μc ≈ μ0c). We consider three classes of
models, sketched in Fig. 1 A–C, where μt/μc > 1, μt/μc < 1, and μt/μc ≈ 1. If μ0t
μ0c, even the swollen modulus μt of the top layer remains considerably greater
than μc, so that μt/μc > 1. If however μ0t ≈ μ0c, after swelling of the top layer
μt/μc < 1. An intermediate situation is when the initial μ0t is slightly greater than
μ0c. If this bilayer is immersed in a solvent, then the modulus of the top layer in
the swollen state is comparable to that of the core, i.e., μt/μc ≈ 1 as in the brain.
The above three situations are experimentally tested by varying the modulus of
the outer layer (μ0t ≈ 1 MPa, ≈ 500 kPa, and ≈ 100 kPa, monomer to cross-linker
ratios 5:1, 10:1, and 30:1, respectively; ref. 36) while keeping the core modulus
constant (μ0c ≈ 100 kPa). The resulting instability patterns are shown in Fig. 1 D–F.
We also experimentally examine the effects of varying the thickness of the top
layer T0 while keeping all other properties fixed (μ0t ≈ 500 kPa, μ0c ≈ 100 kPa,
and r ≈ 11 mm) by using a set of nine bilayer specimens with top layers of three
different thicknesses, namely ≈300 μm, ≈800 μm, and ≈1.2 mm (hence, the
relative initial radius R/T0 ≈ 38, 15, and 10, respectively, where R = r + T0). Each
specimen is immersed in hexanes for a specific period t, and the resulting surface
ffiffiffiffiffiffiffiffiffi
morphologies induced by differential growth and mechanical buckling are im-
aged (Fig. S3). The absorption of solvent is a diffusive process, with T =
4Dt
,
where T is the penetration depth (or thickness) of solvent into a top layer and D
is the diffusion coefficient. Assuming D ≈ 6 × 10−10 m2/s for hexanes in PDMS (38,
39), the estimated T (and R/T) for t = 1 min, t = 4 min, and t = 9 min is found to
be ≈380 μm (30), 760 μm (15), and 1.1 mm (10), respectively. Highlighted images
along a diagonal line in Fig. S3 show the case where the initial radius R/T0 closely
matches the estimated relative radius R/T. Two of these images are compared
with the corresponding images of real and simulated brain shapes in Fig. 3.
p
Numerical Simulations. Our finite element model for brain folding is based on
constant strain triangle (2D) or tetrahedron (3D) elements and an explicit
solver for quasistatic equilibration of the system and allows simulation of the
large strains and highly nonlinear mechanics involved in gyrification, but
necessitates the use of high density meshes.
In 2D simulations the domain is discretized into a rectangular lattice of 150 ×
600 nodes (width × depth) and filled with plane-strain triangle elements
forming a mesh of crossed triangles. The top surface of the gray matter is free
and we apply symmetric boundary conditions along the lateral sides of the
domain. Owing to symmetry, the simulation domain contains only one half of
the sulcus/gyrus. The domain is 10 times the gray-matter thickness (i.e., 60
Tallinen et al.
PNAS September 2, 2014 vol. 111 no. 35 12671
The energy of the system is minimized by damped second-order dynamics,
using an explicit scheme,
a node, respectively. The 2D simulations are implemented similarly but with
triangular elements instead of tetrahedra.
fðtÞ − γvðtÞ
vðt + ΔtÞ = vðtÞ +
Δt,
xðt + ΔtÞ = xðtÞ + vðt + ΔtÞΔt:
m
[7]
Here Δt = 0:05a=
is the time step, m = a3 mass of a node, and γ = 10 m
viscous damping. Vectors f, v, and x are force, velocity, and position of
ffiffiffiffi
p
K
[6]
ACKNOWLEDGMENTS. We thank Teijo Kuopio for discussions. Computations
were run at CSC–IT Center for Science, Finland. The brain images were
obtained from the collections of the University of Wisconsin, Michigan State,
and the National Museum of Health and Medicine, funded by the National
Science Foundation and the National Institutes of Health. We thank the
Academy of Finland (T.T.), the Wyss Institute for Biologically Inspired Engi-
neering (J.Y.C. and L.M.), the Royal Society (J.S.B.), the Singleton Foundation
(L.M.), and the MacArthur Foundation (L.M.) for partial financial support.
1. Striedter GF (2005) Principles of Brain Evolution (Sinauer Associates, Sunderland, MA).
2. Lui JH, Hansen DV, Kriegstein AR (2011) Development and evolution of the human
neocortex. Cell 146(1):18–36.
3. Gabi M, et al. (2010) Cellular scaling rules for the brains of an extended number of
primate species. Brain Behav Evol 76(1):32–44.
20. Bayly PV, Okamoto RJ, Xu G, Shi Y, Taber LA (2013) A cortical folding model in-
corporating stress-dependent growth explains gyral wavelengths and stress patterns
in the developing brain. Phys Biol 10(1):016005.
21. Biot MA (1965) Mechanics of Incremental Deformations (Wiley, New York).
22. Southern E, Thomas AG (1965) Effect of constraints on the equilibrium swelling of
4. Neal J, et al. (2007) Insights into the gyrification of developing ferret brain by mag-
rubber vulcanizates. J Polym Sci A 3(2):641–646.
netic resonance imaging. J Anat 210(1):66–77.
5. Lancaster MA, et al. (2013) Cerebral organoids model human brain development and
microcephaly. Nature 501(7467):373–379.
6. Bae BI, et al. (2014) Evolutionarily dynamic alternative splicing of GPR56 regulates
regional cerebral cortical patterning. Science 343(6172):764–768.
7. Lefèvre J, Jean-François M (2010) A reaction-diffusion model of human brain de-
velopment. PLoS Comp Biol 6:e1000749.
8. Van Essen DC (1997) A tension-based theory of morphogenesis and compact wiring in
the central nervous system. Nature 385(6614):313–318.
9. Hilgetag CC, Barbas H (2006) Role of mechanical factors in the morphology of the
primate cerebral cortex. PLoS Comp Biol 2:146–159.
10. Herculano-Houzel S, Mota B, Wong P, Kaas JH (2010) Connectivity-driven white
matter scaling and folding in primate cerebral cortex. Proc Natl Acad Sci USA 107(44):
19008–19013.
11. Xu G, et al. (2010) Axons pull on the brain, but tension does not drive cortical folding.
J Biomech Eng 132(7):071013.
12. Bok ST (1959) Curvature of the cerebral cortex. Histonomy of the Cerebral Cortex
(Elsevier, Amsterdam).
13. Richman DP, Stewart RM, Hutchinson JW, Caviness VS, Jr (1975) Mechanical model of
brain convolutional development. Science 189(4196):18–21.
14. Toro R, Burnod Y (2005) A morphogenetic model for the development of cortical
convolutions. Cereb Cortex 15(12):1900–1913.
15. Chenn A, Walsh CA (2002) Regulation of cerebral cortical size by control of cell cycle
exit in neural precursors. Science 297(5580):365–369.
16. Stahl R, et al. (2013) Trnp1 regulates expansion and folding of the mammalian ce-
rebral cortex by control of radial glial fate. Cell 153(3):535–549.
17. Prange MT, Margulies SS (2002) Regional, directional, and age-dependent properties
of the brain undergoing large deformation. J Biomech Eng 124(2):244–252.
18. Kaster T, Sack I, Samani A (2011) Measurement of the hyperelastic properties of ex
vivo brain tissue slices. J Biomech 44(6):1158–1163.
19. Nie J, et al. (2010) A computational model of cerebral cortex folding. J Theor Biol
264(2):467–478.
23. Trujillo V, Kim J, Hayward RC (2008) Creasing instability of surface attached hydro-
gels. Soft Matter 4:564–569.
24. Hohlfeld E, Mahadevan L (2011) Unfolding the sulcus. Phys Rev Lett 106(10):105702.
25. Hohlfeld E, Mahadevan L (2012) Scale and nature of sulcification patterns. Phys Rev
Lett 109(2):025701.
26. Tallinen T, Biggins JS, Mahadevan L (2013) Surface sulci in squeezed soft solids. Phys
Rev Lett 110(2):024302.
27. Chen X, Hutchinson JW (2004) Herringbone buckling patterns of compressed thin
films on compliant substrates. J Appl Mech 71(5):597–603.
28. Cheng S, Bilston LE (2007) Unconfined compression of white matter. J Biomech 40(1):
117–124.
29. Zhang K, Sejnowski TJ (2000) A universal scaling law between gray matter and white
matter of cerebral cortex. Proc Natl Acad Sci USA 97(10):5621–5626.
30. Welker W (1990) Why does cerebral cortex fissure and fold: a review of determinants
of gyri and sulci. Cereb Cortex 8B:3–136.
31. Ronan L, et al. (2014) Differential tangential expansion as a mechanism for cortical
gyrification. Cereb Cortex 24(8):2219–2228.
32. Nieuwenhuijse P (1913) Zur Kenntnis der Mikrogyrie. Psychiat Neurol B1(17):9–53.
33. Chang BS, et al. (2007) The role of RELN in lissencephaly and neuropsychiatric disease.
Am J Med Genet B Neuropsychiatr Genet 144B(1):58–63.
34. Shyer AE, et al. (2013) Villification: How the gut gets its villi. Science 342(6155):
212–218.
35. Mumford D, Desolneux A (2010) Pattern Theory (A.K. Peters, Natick, MA).
36. Nase J, Lindner A, Creton C (2008) Pattern formation during deformation of a con-
fined viscoelastic layer: From a viscous liquid to a soft elastic solid. Phys Rev Lett
101(7):074503.
37. Lee JN, Park C, Whitesides GM (2003) Solvent compatibility of poly(dimethylsiloxane)-
based microfluidic devices. Anal Chem 75(23):6544–6554.
38. Favre E, Schaetzel P, Nguygen QT, Clement R, Neel J (1994) Sorption, diffusion and
vapor permeation of various penetrants through dense poly(dimethylsiloxane)
membranes: A transport analysis. J Membr Sci 92:169–184.
39. Guo CJ, De Kee D, Harrison B (1992) Effect of molecular structure on diffusion of
organic solvents in rubbers. Chem Eng Sci 47:1525–1532.
12672 www.pnas.org/cgi/doi/10.1073/pnas.1406015111
Tallinen et al.
Supporting Information
Tallinen et al. 10.1073/pnas.1406015111
Fig. S1. Tangential growth profiles for α = 0.3 and α = 1.25 applied in simulations of Fig. 2 A and B, respectively.
Fig. S2. Geometric parameters from brain sections of a porcupine, cat, and human. Brain radius R is indicated by the red arcs. Gyral widths in the porcupine
and cat are determined as the length of the red arc over each gyrus. In the human the sulcal geometry is more complicated and some gyri are inclined with
respect to the sectioning plane. Therefore, in the human gyral widths are determined more selectively as indicated by magenta line segments. Sulcal depths are
indicated by blue line segments (the sylvian fissure and sulci that are clearly inclined with respect to the sectioning plane are excluded in the human). The
thickness of the gray matter at the gyri is indicated by the yellow line segments, and thickness of the gray matter at the sulci by the green line segments (not
shown for the human). The undeformed thickness of the gray matter is approximated by T = Tg/1.5 using the mean thickness Tg of the gray matter at the gyri.
Tangential expansion g is estimated by dividing the length of the surface contour by the length of the red arc (excluding the sylvian fissure in the human). The
data shown for W/T, D/T, and Tg/Ts are given as the mean ± SD. All images are cell-stained (porcupine and cat) or fiber-stained (human) coronal sections from
www.brainmuseum.org.
Tallinen et al. www.pnas.org/cgi/content/short/1406015111
1 of 2
Fig. S3.
For the swelling gel model that mimics the growing brain, we show the sulcal patterns as a function of the scaled initial radius R/T0 and solvent
exposure time t. The sulcal spacing of the patterns increases with decreasing R/T0 and increasing t in qualitative agreement with our theoretical predictions.
The images highlighted with a blue border represent states where the solvent penetration depth is comparable with the upper layer thickness.
Tallinen et al. www.pnas.org/cgi/content/short/1406015111
2 of 2
|
1705.10653 | 1 | 1705 | 2017-05-30T14:10:15 | Michaelis-Menten at 100 and allosterism at 50: driving molecular motors in a hailstorm with noisy ATPase engines and allosteric transmission | [
"physics.bio-ph",
"physics.chem-ph",
"q-bio.SC"
] | Cytoskeletal motor proteins move on filamentous tracks by converting input chemical energy that they derive by catalyzing the hydrolysis of ATP. The ATPase site is the analog of an engine and hydrolysis of ATP is the analog of burning of chemical fuel. Moreover, the functional role of a segment of the motor is analogous to that of the transmission system of an automobile that consists of shaft, gear, clutch, etc. The operation of the engine is intrinsically "noisy" and the motor faces a molecular "hailstorm" in the aqueous medium. In this commemorative article, we celebrate the centenary of Michaelis and Menten's landmark paper of 1913 and the golden jubilee of Monod et al.'s classic paper of 1963 by highlighting their relevance in explaining the operational mechanisms of the engine and the transmission system, respectively, of cytoskeletal motors. | physics.bio-ph | physics |
Michaelis-Menten at 100 and allosterism at 50:
driving molecular motors in a hailstorm with
noisy ATPase engines and allosteric transmission
Debashish Chowdhury
Department of Physics, Indian Institute of Technology,
Kanpur 208016, India
January 3, 2019
Abstract
Cytoskeletal motor proteins move on filamentous tracks by converting input
chemical energy that they derive by catalyzing the hydrolysis of ATP. The
ATPase site is the analog of an engine and hydrolysis of ATP is the analog of
burning of chemical fuel. Moreover, the functional role of a segment of the motor
is analogous to that of the transmission system of an automobile that consists
of shaft, gear, clutch, etc. The operation of the engine is intrinsically "noisy"
and the motor faces a molecular "hailstorm" in the aqueous medium. In this
commemorative article, we celebrate the centenary of Michaelis and Menten's
landmark paper of 1913 and the golden jubilee of Monod et al.'s classic paper of
1963 by highlighting their relevance in explaining the operational mechanisms
of the engine and the transmission system, respectively, of cytoskeletal motors.
1
Introduction
In 1913 Leonor Michaelis and Maud Menten published a paper [1] (see ref.[2] for
english translation) in which they derived an analytical expression (from now
onwards referred to as the MM equation) for the rate of an extremely simple
model of enzymatic reactions [3]. Even after enriching several overlapping sci-
entific disciplines for 100 years, MM kinetics still continues to raise new exciting
fundamental questions on the conditions for its own validity as well as on its
applicability to hitherto unexplored situations. One of the systems where such
issues are being debated is molecular motors [4, 5, 6, 7, 8, 9, 10].
These chemo-mechanical machines are either single proteins or macromolec-
ular complexes; their stepping on the respective filamentous tracks are random,
albeit biased, and these run on "noisy engines". For the sake of concreteness
in this article we'll consider mostly cytoskeletal motors whose engines are AT-
Pases (more precisely, the ATP-hydrolyzing sites in the head domains of the
motors). These motors drive many processes that are involved in sub cellular
and cellular motility as well as contractility, including mitosis and cytokinesis
[9, 10, 11]. The "transmission system" of the motor amplifies the sub-nanometer
movements, powered by ATP hydrolysis, in the engine to several nanometer long
step size along its track. Moreover, the ATPase cycle of the engine also regu-
lates the track-binding affinity of the motor whose track-binding site is located
a few nano-meters away from the engine. Long-distance intra-machine commu-
nications required for track-binding regulation and for the transmission system
operation are based on allosteric mechanisms. The basic principle of allostery
was laid down 50 years ago in a classic paper by Monod, Changeux and Jacob
[12].
The aim of this commemorative note is to celebrate the centenary and golden
jubilee of the pathbreaking papers of Michaelis and Menten [1] and Monod et
al.[12], respectively, by (a) highlighting some important features of the fluctu-
ating ATPase activity of the engine and the allosteric mechanisms of the trans-
mission system of a molecular motor, and (b) exploring the dependence of the
velocity of the motor on the rate of "fuel consumption" by its engine. We also
point out some interesting formal analogies between the enzymatic turnover
and the mechanical stepping of a molecular motor. We emphasize only the
main conceptual developments over the last two decades, but make no attempt
to cover all aspects of molecular motors; interested readers are referred to two
very recent review articles [9, 10].
2 Key mechanical properties of molecular mo-
tors and core concepts in their operational
mechanisms
The tracks for molecular motors are polar, that is the two ends of these filaments
are not identical. On a given track, the members of the same family of motors
1
tend to move in the same direction. The more processive a motor is the longer
is the distance it covers in a single run in between its attachment to the track
and the next complete detachment from it. Thus, each family of motors, that
move on a given filamentous track, is characterized by a specific directionality
and processivity. The fraction of the cycle time that each head remains attached
to the track, on the average, is called its duty ratio.
An external force on a motor against the direction of its natural stepping is
called a load force. The average velocity of a motor decreases with increasing
magnitude of the load force and eventually vanishes at a value that is called the
stall force. The force-velocity relation is one of the most fundamental charac-
teristics of a molecular motor. Since the stepping of molecular motors is noisy,
two different motors with identical mean velocities may exhibit widely different
variances. Therefore, the stochastic stepping kinetics of a motor is character-
ized in more detail in terms of the distributions of the times of its dwell at the
successive spatial positions on its track.
Cytoskeletal motor proteins, which are listed in table 1, function as intra-
cellular "porters" [13]. These motors carry intracellular cargoes (e.g., vesicles,
organelles, etc.) over long distances by "walking" along their respective tracks
hydrolyzing ATP which is the most widely used "fuel" for molecular machines.
In contrast, the motors listed in the table 2 slide one cytoskeletal polar fila-
ment with respect to another. The sliding occurs when a slider motor, that
crosslinks the two filaments, tends to walk on both simultaneously hydrolyzing
ATP. Some sliders work in groups and each detaches from the filament after
every single stroke. These are often referred to as rowers because of the obvious
analogy with rowing where the oars remain in contact with water for a very
brief period in each stroke [13]. Sliders and rowers drive contractility at cellular
and subcellular levels. It is worth pointing out that occasionally a molecular
motor can undergo a mechanical transition that requires more work than the
free energy supplied by the hydrolysis of one molecule of ATP. The molecular
motor accomplishes such a rare feat, that no macroscopic motor can perform,
because of fluctuations arising from its interactions with the thermal reservoir.
Polynucleotides (DNA and RNA) and polypeptides (proteins) are linear
polymers. The sequence of the monomeric subunits of each of these polymers is
dictated by that of the corresponding template. The sequence of the subunits
of a template directs the correct sequence of the successive monomers that are
to be selected, and added step-by-step, by the machine thereby elongating the
product polymer. Specific machine to be used for a template-directed polymer-
ization depends on the nature of the template and product polymers; these are
listed in the table 3. The template also serves as the track for the polymerizer
machine and its movement along the track is powered by input chemical energy.
Therefore, these machines are also regarded as motors[14, 15].
MT and F-actin serve not only as tracks for cytoskeletal motors, but can also
generate pushing and pulling forces during their polymerization and depolymer-
ization, respectively. For example, a polymerizing MT can exert a pushing force
against a membrane thereby mimicing a nano-piston [10]. Similarly, during its
depolymerization a MT can pull a molecular ring by inserting its hook-like out-
2
wardly curled depolymerizing tip into the ring [10]. In addition to the motors
listed above, cell uses also several other types of motors for various specialized
functions; a more exhaustive list is available in ref.[9].
3 Average rate of a MM reaction: implications
for the average velocity of a motor
The MM reaction is normally represented as
E + S k1(cid:42)(cid:41)
k−1
ES k2→E + P
(1)
where E,S,P denote the enzyme, substrate and product, respectively. The most
commonly used expression for V , the rate of the reaction (1) in bulk, was derived,
more than a decade after Michaelis and Menten's original derivation, under a
less restrictive approximation. In that work [16], Briggs and Haldane assumed
that the concentration of the intermediate complex [ES] remains steady (or,
stationary), i.e., independent of time, during the progress of the reaction in
bulk. One of the standard forms in which the corresponding V is expressed is
V =
Vmax[S]
Km + [S]
(2)
which, in addition to the substrate concentration [S], involves two important
parameters KM and Vmax. The parameter Vmax = k2[E]0 is proportional to the
total amount of enzyme [E]0 and it can be interpreted as the maximum rate of
the reaction attainable in the limit [S] → ∞. The other parameter
KM =
k−1 + k2
k1
,
(3)
called the Michaelis constant, can be interpreted as the particular concentration
of substrate for which the rate of the reaction is exactly half of its maximum
possible value.
The importance of Michaelis and Menten's classic work [1] can be appre-
ciated in its historical context by revisiting the current state of that research
field in 1913 (see, for example, the review by Barendrecht [17]). Michaelis and
Menten's achievement also demonstrated how, inspired by the correct insight,
even a minimal theoretical model and an extremely simplifying assumption can
still account for empirical data obtained for varieties of real systems under wide
range of conditions [18]. Whether or not real enzymes follow the equation (2)
has been scrutinised in the last 100 years by analysing the biochemical data for
an enormously large number of enzymes (see refs. [19, 20]). The history of the
kinetics of enzymatic reactions over the last 100 years [21, 22, 23] is testimony
to the success of MM reaction and its various extensions.
Since a cytoskeletal motor runs on an ATPase engine, one would like to know
the dependence of the average velocity v of the motor on the ATP concentration
3
in the surrounding medium. the enzymatic cycle can be expressed symbolically
as (see fig.1)
E + AT P (cid:42)(cid:41) E.AT P → E.ADP.Pi → E.ADP → E
(4)
where E.AT P is a complex of the enzyme E and an ATP molecule whereas
E.ADP.Pi and E.ADP are macromolecular complexes of the enzyme with the
products of hydrolysis. In this case, in each cycle one molecule of ATP is hy-
drolyzed to produce one molecule of ADP and one inorganic phosphate molecule.
If no futile reaction takes place, i.e., every round of chemical reaction drives
mechanical stepping, the chemical and mechanical processes are said to be
tightly coupled. For a tightly coupled molecular motor, v should be given by
v = V L where L is the step size of the motor and V is the average rate of the
ATPase reaction given by equation (2) with [S] = [AT P ]. Therefore, v should
saturate with increasing concentration of ATP. In reality, for many motors the
chemical and mechanical processes are coupled loosely [24, 25, 26, 27]. In case
of loosely coupled motors, it has been observed that v = κV L where κ < 1 is
the strength of the mechano-chemical coupling. What is even more surprising
is that even when the natural forward movement of the motor is opposed by a
load force F , a MM-like form
v(F ) = κ(F )L
Vmax(F )[AT P ]
Km(F ) + [AT P ]
(5)
seems to capture the dependence of v on the ATP concentration. This is, per-
haps, not the appropriate forum for detailed discussion [5] to find out which, if
any, of the parameters κ(F ), Vmax(F ) and Km(F ) are independent of the load
force F . Nevertheless, if an individual rate constant for a specific mechano-
chemical step, indeed, varies with the load force F what is the form of that
dependence? In the literature on molecular motors, the force-dependence of the
individual rate constants associated with the forward and reverse transitions,
at a given ATP concentration, are assumed to have the form [8, 9]
and
kf (F ) = kf (0)e−θF (∆x)/(kB T )
kr(f ) = kr(0)e(1−θ)F (∆x)/(kB T ),
(6)
(7)
respectively, where kB is the Boltzmann constant, T is the absolute temperature
and θ is a fraction of the step size ∆x. Note that θ determines how the work F ∆x
performed by the external load is shared by the forward and reverse transitions
[8].
The average velocity of a motor is, in general, a combination of the individ-
ual rate constants and, therefore, expected to depend on the load-force through
the F -dependence of the rate constants. It has been established experimentally
that, at a given concentration of ATP, v(F ) decreases with increasing magni-
tude of the load force F , and eventually vanishes at the stall force Fs. Both
the load-free velocity v(0) and the stall force Fs depend on the details of the
4
mechano-chemical kinetics of the motor. The experimentally observed force-
velocity relation is often expressed in the simple form [28]
v(F ) = v(0)[1 − (F/Fs)α]
(8)
in terms of the three parameters v(0), Fs and α. The force-velocity relation
is linear in the special case α = 1 whereas it is sub-linear for all α < 1 and
super-linear for α > 1. Note that the parameter α determines the curvature
(i.e., convex-up or concave-up shape) of the force-velocity relation (see Fig.2).
Motors with superlinear force-velocity relation are strong in the sense that their
velocity is practically unaffected by the load force unless the magnitude of the
force is sufficiently large. On the other hand, if the force-velocity relation of
a motor is sub-linear, even a small load force drastically reduces its velocity
thereby exposing its weakness.
It is worth pointing out that load force can affect the mechano-chemistry of
the motor in many different ways. For example, a load force, even if weaker
than the stall force, can increase the frequencies of back-stepping. It can also
tilt the free energy landscape to such an extent that the motor steps backward
while binding with ADP and inorganic phosphate, and it subsequently releases
the reformed ATP.
What happens to the motor if the strength of the load force opposing it
exceeds Fs? In principle, one can envisage three possible scenarios: (i) the
motor may detach from the track, (ii) the motor may start moving backward
hydrolysing one molecule of ATP in each backward step, or (iii) the motor start
moving backward without hydrolyzing ATP as it gets pushed by the load force
[29]. Recal that v(F ) is the average velocity of the motor. Because of the
stochastic nature of its mechano-chemistry, the three possible scenarios painted
here can occur, with increasing frequency as the load forces increases, even if it
remains weaker than Fs. The actual outcome depends on the family to which
the motor belongs.
4 Distribution of turnover times of an enzyme:
implications for the noisy ATPase engine of a
motor
Next, we consider multiple turnovers by a single-enzyme. Obviously, there are
nonzero waiting periods between the successive turnovers of the enzyme [30, 31,
32]. The time interval τ in between the successive turnovers of an enzyme is
called the turnover time [33, 34]. Suppose in each round of the reaction catalyzed
by the enzyme n product molecules are produced. Thus, the population of the
product molecules increases by n in each discrete jump (see Fig.3).
In the
specific case of the ATPase engine of a molecular motor, the time interval in
between the arrival of the successive ATP molecules and their binding to the
engine is random. Moreover, the actual time taken for conversion of the ATP
molecule to the product(s) of hydrolysis also fluctuates from one round of the
5
reaction to the next. Furthermore, the conformation of a motor, like that of
any other enzyme, itself fluctuates (the so-called dynamic disorder) giving rise
to further sources of noise in the turnover time (see Fig.3).
In general, the
distribution of the turnover times characterizes the stochastic kinetics of the
enzymatic reaction; the mean turnover time < τ > happens to be just the first
moment of this distribution.
In the last decade sophisticated techniques of single-molecule enzymology has
made it possible to monitor the enzymatic turnover of a single enzyme [35, 36].
For technical reasons, turnover statistics has been studied experimentally using
some non-motor enzymes although the conclusions drawn from these experiment
are believed to be valid also for the ATPase activities of the engines of the
motors.
Interestingly, barring a few exceptional situations, the average rate
1/ < τ > of the enzymatic reactions obey the MM equation (2) [31, 32, 33, 34].
Thus, in spite of the fluctuations, the average rate of fuel burning by the noisy
ATPase engine of a cytoskeletal motor is still given by the MM equation (2).
5 Analogy between turnover time of enzyme and
dwell time of motor: implications for the dis-
tributions
Next we discuss a formal analogy between the turnover times of the engine
and the times of dwell of the motor at its successive positions on the track.
Single-motor experiments have demonstrated that the movement of each motor
consists of an alternating sequence of pause and translocation. In other words, a
motor dwells at each discrete position on its track for a duration in between its
arrival at that position and the next departure from it (see Fig.4). The nature
of the walk of a given motor is characterised by the distribution of the dwell
times. In fact, both these distributions also have a practical utility; the quantity
nmin =< τ >2 /(< τ 2 > − < τ >2)
(9)
provides a lower bound on the number of kinetic states [37, 36]. The similarities
between the trace of successive turnovers in fig.3 and the trace of successive steps
of the motor in fig.4 indicates the possibility of using the same mathematical
formalism for extracting their statistical properties.
Next, we explicitly mention the nature of the dwell time distribution for
motors with simple mechano-chemical kinetics and point out some of its key
features. For the operation of a molecular motor driven by an ATPase engine,
the hydrolysis of ATP generates a mechanical force. However, this force may
be generated at different stages of the enzymatic reaction for different motors.
Assuming that the force is generated in only one single step of the enzymatic
cycle, four distinct possible stages of force generation are listed below where the
subscripts j and j + 1 label the two successive positions of the motor M on its
track [4].
6
Mechano-chemical binding:
(M + AT P )j (cid:42)(cid:41) (M.AT P )j+1 (cid:42)(cid:41) (M.ADP.Pi)j+1 (cid:42)(cid:41) (M.ADP )j+1 (cid:42)(cid:41) (M )j+1
Mechano-chemical reaction:
(M + AT P )j (cid:42)(cid:41) (M.AT P )j (cid:42)(cid:41) (M.ADP.Pi)j+1 (cid:42)(cid:41) (M.ADP )j+1 (cid:42)(cid:41) (M )j+1
Mechano-chemical release:
(M + AT P )j (cid:42)(cid:41) (M.AT P )j (cid:42)(cid:41) (M.ADP.Pi)j (cid:42)(cid:41) (M.ADP )j+1 (cid:42)(cid:41) (M )j+1
Mechano-chemical trigger:
(M + AT P )j (cid:42)(cid:41) (M.AT P )j (cid:42)(cid:41) (M.ADP.Pi)j (cid:42)(cid:41) (M.ADP )j (cid:42)(cid:41) (M )j+1
The transition from position j to position j + 1 is the force-generating step
of the cycle. Each of these four schemes is based on the assumption of tight
mechano-chemical coupling although these can be easily extended to model
loose-coupling motors. In all the four cases the average velocity of the motor
is given by the same expression which, however, differs from the MM equation
(2); the only difference between the four cases is that different rate constants
are force-dependent in the different models [4]. Moreover, under appropriate
circumstances, by short-circuiting relatively fast transitions (i.e., by combining
more than one step into a single effective step) the number of intermediate
complexes can be reduced to only one. Furthermore, under some conditions
one or more of the transitions can also become practically irreversible. For
example, if the products of ATP hydrolysis are removed immediately after their
formation, their rebinding with the motor can be ruled out. The typical example
shown below
(M + AT P )j (cid:42)(cid:41) I 1
j → (M )j+1,
where I 1
j is the intermediate complex and the last transition has been assumed
to be irreversible, mimics exactly the MM scheme (1). This scheme has been
depicted slightly differently in Fig.5, where the discrete allowed positions corre-
spond to the motor-binding sites on the track.
For the generic kinetics shown in Fig.5, the probability density of the dwell
(cid:18) k1k2
(cid:19)(cid:26)
2B
(cid:27)
e−k−t − e−k+t
(10)
(11)
(12)
times is given by
P (t) =
where
with
k± = A ± B
A =
B =
(k1 + k−1 + k2)
(cid:18) (k1 + k−1 + k2)2
2
4
7
(cid:19)1/2
.
− k1k2
Note that the distribution P (t) for the two-step process shown in Fig.5 is a
sum of two exponentials where the rates k± in these two exponentials are linear
combinations of the rate constants k1, k−1, k2. In the special limiting case k−1 =
0, the expression (10) simples to
P (t) =
e−k2t − e−k1t
(13)
(cid:19)(cid:26)
(cid:18) k1k2
k1 − k2
(cid:27)
where the two rates in the two exponentials are the two non-zero rate constants
k1, k−1. Moreover, in this special limit the mean dwell time becomes k−1
1 + k−1
2 ,
i.e., sum of the average times spent sequentially in the two states.
Even among the molecular motors that never step backward and do not
have branched pathways, very few have mechano-chemical kinetics as simple
as that shown in Fig.5. One such example is the ribosome, one of the largest
and most complex molecular motors, that moves on a messenger RNA (mRNA)
track while synthesising a polypeptide using its mRNA track as a template. Its
mechano-chemical cycle, during the stage of elongation of the polypeptide, has a
few more intermediate states after I1. However, in spite of the existence of more
than one intermediate states, the average rate of elongation of the polypeptide
(which is identical to the average velocity of the ribosome on its track) is gov-
erned by a MM-like equation [38] when kinetic proofreading and translational
error are not captured explicitly by the model. This is not surprising because
the average rate of a MM reaction with more than one intermediate complex is
also known to obey a MM-like equation. However, what is non-trivial is that
even after inclusion of the pathways (a) for tRNA rejection by kinetic proofread-
ing and (b) for incorporation of wrong amino acid into the growing polypeptide,
the mean rate of elongation of the polypeptide still satisfies a MM-like equation
[39]. But, there are few kinetic models of molecular motors, including one that
we [40] developed for the single-headed kinesin KIF1A, whose average velocity
deviate from the MM-like behaviour. We'll not discuss the nature and plausible
causes of these deviations here because the details are available in ref.[36].
During protein synthesis in-vivo usually a large number of ribosomes simul-
taneously move on the same mRNA track, each polymerizing one copy of the
same protein whose amino acid sequence is directed by the mRNA that serves
also as a template. This phenomenon is often referred to as the ribosome traffic
[9] because of its superficial similarity with vehicular traffic on highways. A
ribosome poised for forward translocation to the next codon has to wait at its
present location, if the target codon is already covered by another ribosome
in front of it. The following ribosome can move forward only after its target
codon is vacated by the leading codon. Thus, increasing crowding lengthens the
dwell time of a ribosome at a codon. The effect of the crowding on the overall
distribution of the dwell times of the ribosomes has been calculated analyticaly
[39].
8
6 Forward, backward and reverse processes: con-
ditional dwell times
In each cycle, the position of the motor can change by ±L where L is the fixed
step size of the motor. Starting from any arbitrary initial condition with given
position of the motor and given number of ATP molecules, all the eight pos-
sible transitions can be depicted schematically as shown in Fig.6. The passive
transition M and the corresponding reverse transition Mr are purely mechan-
ical processes which do not change the number of ATP molecules. Similarly,
the transitions C and Cr are purely chemical processes which do not change
the position of the motor. But, both the active forward step (F) or the ac-
tive backward step (B) are driven by hydrolyzing one molecule of ATP (see
fig.6). Does this motor synthesize, rather than hydrolyze, a molecule of ATP in
the corresponding "reverse processes" (Fr and Br), respectively [9] (see fig.6)?
Since addressing such fundamental questions on thermodynamics and kinetics
of molecular motors [7] is beyond the scope of this article, interested readers are
referred to the discussion in ref.[41].
Thus, the stepping pattern of a molecular motor can be more complex than
the progress of an enzymatic reaction. For example, a motor that can step
both forward and backward, a forward step may coincide with the formation of
the product of an enzymatic reaction catayzed by it whereas its backward step
may coincide with the reformation of the substrate [37]. Thus, the complete
mechano-chemical state attained by a forward step, for example, depends on
the direction in which the preceding step was taken. For such motors dwell
times corresponding to different kinetic pathways need to be sorted out for
appropriate analysis of their distributions. The stochastic stepping of a motor
that can step both forward and backward is characterised most appropriately
in terms of the distributions of the four conditional dwell times τ±± which are
defined as follows (see Fig.7):
τ++ = dwell time between a + step followed by a + step
τ+− = dwell time between a + step followed by a − step
τ−+ = dwell time between a − step followed by a + step
τ−− = dwell time between a − step followed by a − step
(14)
where + and − refer to the forward and backward steps, respectively. Moreover,
at least for some motors, step size is not a constant; the distribution of step sizes
itself is an interesting characteristic of the motor. Furthermore, because of the
unavoidable noise in the recordings, extracting the dwell time requires careful
processing of the data. However, new ideas are needed to gainfully exploit these
conditional dwell times to extract useful information on the architecture of the
network of the mechano-chemical states of the motor and kinetic pathways on
that network.
9
7 Allosteric mechanisms of intra-machine com-
munication and the transmission system
Allostery is a mechanism for regulation of the structure, dynamics and function
of an enzyme by the binding of another molecule, called effector, which can
be a small molecule (a ligand) or another macromolecule [42, 43, 44, 45]. The
three defining characteristics of allostery are [46]: (i) the effector is chemically
distinct from the substrate, (ii) the binding site for the effector is spatially well
separated from that of the substrate, and (iii) binding of an effector molecule
affects at least one of the functional properties of the enzyme; the functional
property could be either (a) the binding affinity for its specific substrate or (b)
the rate of the reaction it catalyzes.
The affinity of the motor for its track varies depending on the absence or
presence (and nature) of the ligand (that is, ATP or ADP, etc.) in the "en-
gine" of the motor. The ATPase cycle of the engine, and the associated cyclic
conformational change, is coordinated with the cyclic alteration of the motor's
track-binding affinity. In general, on a given motor, the "engine" (where ATP
gets hydrolyzed) and the site that binds to the track do not overlap and com-
municate with each other by allosteric mechanism [47, 48]. The hydrolysis of
ATP causes only sub-nanometer movements in the ATPase engine. On the other
hand, the step size of the cytoskeletal motors are at least an order of magnitude
longer. The designs of the cytoskeletal motors are such that a "mechanical ele-
ment", that extends from the engine, amplifies the subnanometer movements in
the ATPase engine site up to step size of the motor [49, 50]. This mechanism is
analogous to the transmission system of an automobile. The component devices
of the transmission system include, for example, the piston, shaft, clutch and
gears, etc. and it transmits the engine-generated power to the wheels.
Both myosins and kinesins use a "sensor" to detect the presence or absence
of a single phosphate group. The sensor consists of loops called switch I and
switch II. To serve as the specific sensor for the phosphate group, the switch II
loop swings in and out, respectively,in the presence and absence of the phos-
phate group. When ATP is hydrolyzed to ADP the small movements of the
sensor attached to the engine of the motor is transmitted to the track-binding
site by a helix that is quite long and rigid. In both kinesins and myosins the
conformational changes of the helix have, at least, a superficial similarities with
the movement of a piston. The inward motion of the switch II towards the phos-
phate causes the upstroke of the piston whereas phosphate release initiates the
downstroke. The mechanical elements of kinesin and myosin, called neck-linker
and lever arm, respectively, executes movements driven by the corresponding
piston-like helices described above. In myosin, the motion of the helix drives the
angular motion of the lever arm that, in turn, gives a push to an actin filament
. In kinesin movement of the the piston-like helix causes docking of the neck
linker which, in turn, is responsible for the coordinated movement of the two
heads of the kinesin motor in a directed manner along a MT track.
Investigations on the pathways (amino acid sequence) for intra-machine al-
10
losteric communication began only in the recent past. Although normal mode
analysis of coarse-grained theoretical models of the motors [51] is a very popu-
lar theoretical technique for identifying the allosteric communication pathways,
other alternative methods have also been used (see, for example, ref.[52] for
treatment of myosin)
Establishing the mechanism of communication between the engine and track-
binding site of a dynein is more challenging than that for most of the other
molecular motors. Out of the six domains that form a ring-like structure, only
two seem to operate as engines by hydrolysing ATP. But, the track-binding
site is not located anywhere on this ring.
Instead, the track-binding site is
located at the end of a 15 nm long stalk that emerges from this ring. One of the
biggest mysteries in the designs of cytoskeletal motors is how the sub-nanometer
conformational change in the engine of dynein is allosterically communicated to
the tip of the stalk [53, 54]. Experimental data for dynein have been interpreted
to suggest that it also has gears [55] and clutch [56]; allosteric interactions are
believed to be responsible for the gear and clutch mechanisms of dynein.
8 Summary and conclusion
In this commemorative article we have discussed some topics of current research
interest in molecular motors from the perspective of biochemistry. In particular,
we have examined the roles of Michaelis-Menten scheme of enzymatic reactions
and allosteric mechanism of intra-molecular communication, respectively, in the
operation of the engine and transmission system of the cytoskeletal motors.
A class of motors, that includes ribosome, use their nucleic acid tracks also
as a template for polymerizing other macromolecules. Replication of DNA,
transcription and translation are examples of such template-directed polymeri-
sation. The machines that drive such phenomena have the daunting task of cat-
alyzing the polymerization reaction with much higher fidelity than what might
be demanded by the laws of thermodynamics. The amplification of substrate
specificity of these enzymes is achieved through kinetic processes like kinetic
proofreading, energy relay, etc. How these machines optimise the opposing de-
mands for speed and fidelity is one of the interesting questions at the interface
of enzymatic kinetics and biophysics of molecular motors [57].
The open questions on the mechanism of operation of the transmission sys-
tem of a motor are closely related to the fundamental general questions on the
physical modes of long-distance communication in macromolecular systems [58].
For example, does nature select the path of shortest distance or shortest time for
allosteric communication? Better methods of characterization of the pathways
of allosteric interactions may reveal generic principles that nature might have
exploited in designing the transmission systems through evolutionary tinkering.
Although equipped with a noisy engine and a shaky transmission system,
a molecular motor moves, on the average, in a directed manner in a molecular
hailstorm- a remarkable feat, indeed. Explaining the physical origin of the
MM-like expression for its average velocity and understanding the allosteric
11
mechanism of the transmission system are some important items in the overall
agenda of research on molecular motors [9, 10, 11]. Michaelis-Menten at 100
is as exciting, if not more, as it was in 1913. The concept of allostery at 50
is, perhaps, finding applications in more diverse systems than might have been
anticipated in 1963. Happy birthday allostery! Long live Michaelis-Menten!
Acknowledgements:
I thank Ajeet K. Sharma for a critical reading of the manuscript. This work
is supported by Dr. Jag Mohan Chair professorship, by J.C. Bose national
fellowship, and by a research grant from DBT, government of India. I also thank
the visitors program of the Max-Planck Institute for the Physics of Complex
Systems for hospitality in Dresden where part of this article was written.
References
[1] L. Michaelis and Miss Maud L. Menten, Die kinetik der invertinwirkung,
Biochem. Z. 49, 333-369 (1913). (The kinetics of invertase action, Biochem.
50 (2011).
[2] K.A. Johnson and R.S. Goody, The original Michaelis constant: translation
of the 1913 Michaels-menten paper, Biochem. 50, 8264-8269 (2011).
[3] A. Cornish-Bowden, Fundamentals of enzyme kinetics, (Wiley, 2012).
[4] D. Keller and C. Bustamante, The mechanochemistry of molecular motors,
Biophys. J. 78, 541-556 (2000).
[5] J. Howard, Mechanics of motor proteins and the cytoskeleton, (Sinauer
Associates, Sunderland, 2001).
[6] A. Mogilner, T.C. Elston, H. Wang and G. Oster, Molecular motors: theory,
Comp. Cell Biol. 20, 320-353 (2002).
[7] S.M. Block, Kinesin motor mechanics: binding, stepping, tracking, gating
and limping, Biophys. J. 92, 2986-2995 (2007).
[8] A.B. Kolomeisky and M.E. Fisher, Molecular motors: a theorist's perspec-
tive, Annu. Rev. Phys. Chem. 58, 675-695 (2007).
[9] D. Chowdhury, Stochastic mechano-chemical kinetics of molecular motors:
A multidisciplinary enterprise from a physicist's perspective, Phys. Rep.
529, 1-197 (2013).
[10] D. Chowdhury, Modeling stochastic kinetics of molecular machines at mul-
tiple levels: from molecules to modules, Biophys. J. 104, 2331-2341 (2013).
[11] T. D. Pollard, The Cytoskeleton, cellular motility and the reductionist
agenda, Nature 422, 741-745 (2003).
12
[12] J. Monod, J.P. Changeux and F. Jacob, Allosteric proteins and cellular
control systems, J. Mol. Biol. 6, 306-329 (1963).
[13] S. Leibler and D.A. Huse, Porters versus rowers: a unified stochastic model
of motor proteins, J. Cell Biol. 121, 1357-1368 (1993).
[14] N. Cozzarelli, G.J. Cost, M. Nollmann, T. Viard and J.E. Stray , Giant
proteins that move DNA: bullies of the genomic playground, Nat. Rev. Mol.
Cell Biol. 7, 580-588 (2006).
[15] D. Dulin, J. Lipfert, M.C. Moolman and N.H. Dekker, Studyin enomic
processes at the single-molecule level: introducing the tools and applications,
Nat. Rev. Genet. 14, 9-22 (2013)
[16] G.E. Briggs and J.B.S. Haldane, A note on the kinetics of enzyme action,
Biochem. J. 19, 338-339 (1925).
[17] H.P. Barendrecht, Enzyme-action, facts and theory, Biochemical J. 7, 549-
561 (1913).
[18] J. Gunawardena, Some lessons about models from Michaelis and Menten,
Mol. Biol. Cell 23, 517-519 (2012).
[19] C.M. Hill, R.D. Waight and W.G. Bardsley, Does any enzyme follow the
Michaelis-Menten equation?, Mol. Cell. Biochem. 15, 173-178 (1977).
[20] S. Schnell and P.K. Maini, A century of enzyme kinetics: reliability of the
KM and vmax estimates, Comments on Theor. Biol. 8, 169-187 (2003).
[21] H. Gutfreund, Kinetics:
(Suppl.) E13-E19 (1976).
the grammar of enzymology, FEBS Lett. 62
[22] K.J. Laidler, A brief history of enzyme kinetics, in: New beer in an old
bottle: Eduard Buchner and the growth of biochemical knowledge, ed. A.
Cornish-Bowden 127-133 (Unversitat de Valencia, Valencia, Spain, 1997).
[23] A. Baici, Enzyme kinetics: the velocity of reactions, Biochem. J. (2006).
[24] A.E. Knight and J.E. Moloy, Nat. Cell Biol. 1, E87 (1999).
[25] F. Oosawa, The loose coupling mechanism in molecular machines of living
cells, Genes to Cells 5, 9-16 (2000).
[26] A. Mehta, Myosin learns to walk, J. Cell Sci. 114, 1981-1998 (2001).
[27] B.E. Clancy, W.M. Behnke-Parks, J.O.L. Andreasson, S.S. Rosenfeld and
S.M. Block, Nat. Struct. Mol. Biol. 18, 1020 (2011).
[28] A. Kunwar and A. Mogilner, Robust transport by multiple motors with non-
linear force-velocity relations and stochastic load sharing, Phys. Biol. 7,
016012 (2010).
13
[29] R.A. Cross, Myosin's mechanical ratchet, Proc. Natl. Acad. Sci. USA 103,
8911-8912 (2006).
[30] W. Min, B.P. English, G. Luo, B.J. Cherayil, S.C. Kou and X.S. Xie,
Fuctuating enzymes: lessons from single-molecule studies, Acc. Chem. Res.
38, 923-931 (2005).
[31] S.C. Kou, B.J. Cherayil, W. Min, B.P. English and X.S. Xie, Single molecule
Michaelis-Menten equations, J. Phys. Chem. B 109, 19068-19081 (2005).
[32] W. Min, I. V. Gopich, B.P. English, S.C. Kou, X.S. Xie and A. Szabo,
When does the Michaelis-Menten equation hold for fluctuating enzymes?,
J. Phys. Chem. B 110, 20093-20097 (2006).
[33] S.C. Kou, Stochastic networks in nanoscale biophysics: modeing enzymatic
reaction of a single protein, J. Am. Stat. Assoc. 103, 961-975 (2008).
[34] S. Yang, J. Cao, R.J. Silbey and J. Sung, Quantitative interpretation of
the randomness in single enzyme turnover times, Biophys. J. 101, 519-524
(2011).
[35] X.S. Xie, Enzymology and life at the single molecule level,
in: Single
molecule spectroscopy in chemistry, physics and biology Nobel symposium,
ed. A. Graslund et al., 435-448 (Springer, 2010).
[36] J. R. Moffitt and C. Bustamante, this special issue of FEBS J. (2013).
[37] J. R. Moffitt, Y.R. Chemla and C. Bustamante, Methods in statistical ki-
netics, Methods in Enzymology 475, 221-257 (2010).
[38] A. Garai, D. Chowdhury, D. Chowdhury and T.V. Ramakrishnan, Stochas-
tic kinetics of ribosomes: single motor properties and collective behavior,
Phys. Rev. E 80, 011908 (2009).
[39] A.K. Sharma and D. Chowdhury, Distribution of dwell times of a ribosome:
effects of infidelity, kinetic proofreading and ribosome crowding, Phys. Biol.
8, 026005 (2011).
[40] A. Garai and D. Chowdhury, Stochastic kinetics of a single-headed motor
protein: dwell time distribution of KIF1A, EPL 93, 58004 (2011).
[41] R.D. Astumian, Thermodynamics and kinetics of molecular motors, Bio-
phys. J. 98, 2401-2409 (2010).
[42] J. Monod, J. Wyman and J.P. Changeux, On the nature of allosteric tran-
sitions: a plausible model, J. Mol. Biol. 12, 88-118 (1965).
[43] D.E. Koshland Jr., G. Nemethy and D. Filmer, Comparison of experi-
mental binding data and theoretical models in proteins containing subunits,
Biochemistry 5, 365-385 (1966).
14
[44] J.P. Changeux, Allostery and the Monod-Wyman-Changeux model after 50
years, Annu. Rev. Biophys. 41, 103-133 (2012).
[45] D.E. Koshland, Jr. and K. hamadani, Proteomics and models for enzyme
cooperativity, J. Biol. Chem. 277, 46841-46844 (2002).
[46] A.W. Fenton, Allostery: an illustrated definition of the 'second secret of
life', Trends in Biochem. Sci. 33, 420-425 (2008).
[47] E. Goldsmith, Allosteric enzymes as models for chemomechanical energy
transducing assemblies, FASEB J. 10, 702-708 (1996).
[48] A. Vologodskii, Energy transformation in biological molecular motors,
Phys. of Life Rev. 3, 119-132 (2006).
[49] R.D. Vale and R.A. Milligan, The way things move: looking under the hood
of molecular motor proteins, Science 288, 88-95 (2000).
[50] P. Llinas, O. Pylypenko, T. Isabet, M. Mukherjea, H.L. Sweeney and A.M.
Houdusse, How myosin motors power cellular functions- an exciting jour-
ney from structure to function, FEBS J. 279, 551-562 (2012).
[51] W. Zheng, B.R. Brooks and D. Thirumalai, Allosteric transitions in bi-
ological nano machines are described by robust normal modes of elastic
networks, Current Protein and Peptide Science 10, 128-132 (2009).
[52] S. Tang, J.C. Liao, A.R. Dunn, R.B. Altman, J.A. Spudich and J.P.
Schmidt, Predicting allosteric communication in myosin via a pathway of
conserved residues, J. Mol. Biol. 373, 1361-1373 (2007).
[53] J.A. Spudich, Molecular motors, beauty in complexity, Science 331, 1143-
1144 (2011).
[54] C. Cho and R.D. vale, The mechanism of dynein motility:
insight from
crystal structures of the motor domain, Biochim. Biophys. Acta 1823, 182-
191 (2012).
[55] R. Mallik, B.C. Carter, S.A. Lex, S.J. King and S.P. Gross, Cytoplasmic
dynein functions as a gear in response to load, Nature 427, 649-652 (2004).
[56] J. Huang, A.J. Roberts, A.E. Leschiner and S. L. Reck-Peterson, Lis1 acts
as a "clutch" between the ATPase and microtubule-binding domains of the
dynein motor, Cell 150, 975-986 (2012).
[57] A.K. Sharma and D. Chowdhury, Quality control by a mobile molecular
workshop: quality versus quantity, Phys. Rev. E 82, 031912 (2010).
[58] R. Nussinov, How do dynamic cellular signals travel long distances? Mol.
BioSyst. 8, 22-26 (2012).
15
Figure 1: The main pathway of ATP hydrolysis by an ATPase (ATP-hydrolyzing
enzyme). See the text for explanation.
16
E E.ADP.Pi E.ATP E.ADP Figure 2: Typical force-velocity relations of a molecular motor obtained by
graphically plotting (8) for v(0) = 150s−1, Fs = 10 pN and a four different
values of the parameter α. The unit "number per second" refers to the number
of subunits of its track.
17
Figure 3: A hypothetical noise-free trace of the population of the product
molecules of an enzymatic reaction during multiple turnovers catalyzed by a
single enzyme. The slope of the dashed line indicates the average rate of in-
crease of the reaction product with the passage of time.
18
Time
Product
popula0on
Turnover
Time
Average
behavior:
Slope=av.
rate
n
Figure 4: A hypothetical noise-free trace of the positions of a molecular motor
with the progress of time. The length of the vertical segments are the step sizes
while that of the horizontal segments are the dwell times. No backward step
is observed on this trace. The slope of the dashed line indicates the average
velocity of the motor.
19
Time
Motor
posi,on
Dwell
Time
Step
Size
Average
behavior:
Slope=av.speed
Figure 5: A schematic representation of a simple mechano-chemical kinetics of a
motor that walks on a linear track. The discrete allowed positions of the motor
on the track are labelled by an integer j. The motor can move only forward
by one step; backward steps are not allowed. I1 is an intermediate "chemical"
(or "internal") state. The vertical arrows indicate purely "chemical" transitions
in which the mechanical position of the motor does not change whereas the
transition represented by the tilted arrow is a mechano-chemical transition in
which both the chemical state and mechanical position of the motor change
simultaneously.
20
j+2
j+3
j+1
j--‐1
j--‐2
j--‐3
j
k1
k--‐1
k2
I1
Figure 6: Possible changes in the position and the number of ATP molecules in
a cycle (adapted from ref.[9]; see text for details).
21
No. of fuel molecules Position on the track 0 -L,-1 -L,0 -L,1 0,-1 0,0 0,1 L,-1 L,0 L,1 F B Br Fr C Cr M Mr 0 -1 1 -L L Figure 7: A schematic depiction of the typical trace of the positions of a molecu-
lar motor with the progress of time. Both forward (+) and backward (-) stepping
take place although the latter are quite rare.
22
Time
Motor
posi,on
+
followed
by
--‐
--‐
followed
by
+
+
followed
by
+
--‐
followed
by
--‐
Step
Size
Motor family
Myosin-V & VI
Example of function
Melanosome transport
Kinesin-1 & cytoplasmic dynein Microtubule Mitochondria transport
Track
F-actin
Step size
36 nm
8 nm
Table 1: Few examples of ATP-driven porters are listed along with their corre-
sponding tracks. The actual step size can be an integral multiple of the minimum
value quoted in this table (see ref.[9] for references to original papers).
Motor family
Myosin-II
Filament Example of function
Muscle contraction
F-actin
Kinesin-5, Kinesin-14
Axonemal Dynein
MT
MT
Mitosis
Flagellar beating
Table 2: Few example of rowers and sliders (see ref.[9] for references to original
papers).
Motor
DdDP
DdRP
RdDP
RdRP
Ribosome
Template Product
Function
Step-size
DNA
DNA
RNA
RNA
mRNA
DNA
RNA
DNA
RNA
DNA replication
Transcription
Reverse transcription
RNA replication
Protein
Translation
0.34 nm (1 nt)
0.34 nm (1 nt)
0.34 nm (1 nt)
0.34 nm (1 nt)
1.02 nm (3 nt)
Table 3: Types of polymerizing machines, the templates that direct the sequence
of monomers of the polymeric products. The abbreviations DdDP, DdRP, RdDP
and RdRP refer to DNA-dependent DNA polymerase, DNA-dependent RNA
polymerase, RNA-dependent DNA polymerase and RNA-dependent RNA poly-
merase, respectively (see ref.[9] for references to original papers).
23
|
1309.3585 | 1 | 1309 | 2013-09-13T21:17:25 | Stored Luminescence Computed Tomography | [
"physics.bio-ph",
"physics.med-ph"
] | The phosphor nanoparticles made of doped semiconductors, pre-excited by well-collimated X-ray radiation, were recently reported for their light emission upon NIR light stimulation. The characteristics of X-ray energy storage and NIR stimulated emission is highly desirable to design targeting probes and improve molecular and cellular imaging. Here we propose stored luminescence computed tomography (SLCT), perform realistic numerical simulation, and demonstrate a much-improved spatial resolution in a preclinical research context. The future opportunities are also discussed along this direction. | physics.bio-ph | physics | Stored Luminescence Computed Tomography
Wenxiang Cong, Chao Wang, and Ge Wang*
Biomedical Imaging Center/Cluster
Center for Biotechnology and Interdisciplinary Studies
Department of Biomedical Engineering
Rensselaer Polytechnic Institute, Troy, NY 12180, USA
[email protected], [email protected], [email protected]
Abstract
The phosphor nanoparticles made of doped semiconductors, pre-excited by well-collimated X-ray
radiation, were recently reported for their light emission in the range 650–770 nm upon NIR light
stimulation. The characteristic of X-ray energy storage and NIR stimulated emission is highly desirable
to design targeting probes and improve molecular and cellular imaging. Here we propose “stored
luminescence computed tomography” (SLCT), perform realistic numerical simulation, and demonstrate
a much-improved spatial resolution in a preclinical research context. The future opportunities are also
discussed along this direction.
1. Introduction
X-ray luminescence CT (XLCT) is a hot topic for preclinical imaging [1-3], which combines X-ray and
optical imaging to improve image resolution relative to purely optical imaging modalities such as
fluorescence tomography [4, 5] and bioluminescence tomography [6, 7]. XLCT utilizes X-ray
luminescent nanophosphors (NPs) as imaging probes. NPs can be excited with a pencil, fan or cone
beam of X-rays, and the NP luminescence can be readily generated and efficiently collected using a
sensitive light detection system. XLCT is an analog to fluorescence diffuse optical tomography (FDOT)
but the former has an advantage over the latter in terms of penetration depth and photo-stability.
However, this type of luminescent light signals comes from all excited NPs and is highly diffusive,
limiting the accuracy of image reconstruction.
Recently, MgGa2O4:Cr3+ nanophosphors, as well as others, were reported to be capable of
restraining luminescence emission for ten hours or longer after one-time X-ray excitation [8, 9]. The
stored energy is released upon NIR light stimulation at subsequent time points leads to effectively
controlled luminescence emission in the range of 650–770 nm. The mechanism for the stored
luminescence in such nanophosphors lies in the modified energy diagram of the doped semiconductor
1
MgGa2O4:Cr3+ of nearly 44% cationic site inversion due to nominal Mg deficiency. The dopant Cr3+
ions occupy octahedral sites which are associated with the stored luminescence emission spectrum
around 707 nm (corresponding to the dopant associated 2E(2G) → 4A2(4F) transition); please see Figure
A-1 in the appendix for more details. The advent of such nanophosphors opens a door to significant
protocol flexibility and performance gain for molecular and cellular imaging.
Here we propose “stored luminescence computed tomography” (SLCT), perform realistic
numerical simulation, and demonstrate a much-improved spatial resolution in a preclinical research
context. SLCT is an analog to XLCT, but the former outperforms the latter because stored luminescence
can be readout in a much more sophisticated way than instantaneous emitted luminescence on which
XLCT relies. As shown in this paper, the freedom for selective data acquisition implies a major
improvement in image resolution, holding a great potential for sensitive and specific small animal
imaging in general.
2. Imaging Principles
MgGa2O4:Cr3+ nanoparticles can be functionalized to target specific cells and then introduced into an
object such as a living mouse [8, 10]. X-rays from an x-ray tube are collimated into a narrow beam such
as a pencil or fan beam for excitation of regions of interest (ROI) in the mouse. Part of the x-ray energy
is then deposited in the nanophosphors. When stimulated by NIR laser light in various patterns, the pre-
excited nanophosphors will emit luminescence photons in the range 650–770 nm. The retrospectively
stimulated luminescence emission data are collected on the surface of the mouse for tomographic image
reconstruction.
SLCT imaging is to localize and quantify a distribution of energy-storing nanophosphors such as
MgGa2O4:Cr3+ nanoparticles in a 3D object such as a mouse. Since the measurement of NIR light
signals can be spatially and temporally resolved by stimulating energy-storing nanophosphors at any
location on the surface of the object and any time instant as long as the energy is kept in the
nanoparticles, the resultant dataset will carry more tomographic information than the counterpart in a
corresponding XLCT experiment. Generally speaking, many problems of optical tomography are
underdetermined with possible false solutions due to the inherent non-uniqueness and data noise [7, 11].
Interestingly, this ill-posedness can be effectively overcome with informative combinations of well-
defined x-ray beam shapes and NIR stimulation patterns.
2
It is underlined that two powerful imaging features are unique to SLCT, which are impossible
with other optical imaging modalities such as XLCT. First, a field of view (FOV) for a SLCT study can
be clearly defined with X-rays. This means not only a direct energy distribution from a single X-ray
beam but also a synthetic energy distribution from multiple X-ray beams. The latter scheme is an analog
to tomotherapy for radiation oncology. Second, an energy distribution carried by energy storing
nanoparticles can be optically read out in a multiplexing fashion. NIR light patterns can be projected
externally anywhere around the FOV in a time sequence, which is more effective than measuring all
luminescent signals simultaneously from the distribution as a whole. The NIR stimulation patterns can
be coded in different frequencies as well. The light penetration depth depends on its spectrum and
intensity. For example, SLCT imaging can be performed in an onion peeling fashion. In other words,
external shells can be initially stimulated and reconstructed, depleting X-ray energies in these shells.
Then, the subsequent signals must come from internal shells. With such an onion peeling strategy, the
FOV can be step-by-step shrunk for more accurate and more reliable image reconstruction, which is
impossible with XLCT, fluorescence tomography and bioluminescence tomography.
2.1. X-ray Excitation
Incident X-rays can be easily collimated into a narrow beam to excite energy-storing nanoparticles such
as MgGa2O4:Cr3+ in a living mouse. The X-ray intensity distribution
in the animal can be
described by the Lambert-Beer law:
, (1)
where
is a source position,
the incident x-ray intensity, and
the linear attenuation coefficient
[mm-1] which can be computed with X-ray computed tomography (CT). Energy stored in the
nanoparticles is determined by local X-ray flux intensities. Therefore, we need to compute a stored X-
ray energy density distribution inside the animal according to Eq. (1). As mentioned above, a more
desirable X-ray energy distribution can be synthesized using a tomotherapy approach [12]. The
literature on tomotherapy is extensive. An established radiation therapeutic planning technique, such as
[13], can be directly applied to deposit a pre-specified X-ray energy distribution such as targeting an
ROI.
2.2. NIR Stimulation
3
rIdttXX0rr00rrrrrr000exp0r0X NIR light is moderately absorbed and strongly scattered in biological tissues. The diffusion
approximation (DA) model is commonly used to describe the NIR light propagation in this scenario [6,
14]:
, (3)
where
is a position vector,
an NIR fluence rate [Watts/mm2],
an NIR source [Watts/mm3],
the absorption coefficient [mm-1],
the diffusion coefficient defined by
,
the
reduced scattering coefficient [mm-1], and
. If no NIR photon travels across the boundary
into the tissue domain Ω, the DA is constrained by the Robin boundary condition
, (4)
where
is the outward unit normal vector on
, and
the boundary mismatch factor. The boundary
mismatch factor between the tissue of a refractive index
and the air can be approximated by
with
[15]. Thus, the measurable exiting
photon flux on the surface of the animal is expressed as
. (5)
The intensity of the NIR luminescence emission depends on the density of energy-storing
nanoparticles
, the X-ray intensity
, the laser intensity
, and the stored luminescence photon
yield
of the nanoparticles, which can be defined as the quantum yield per a unit nanoparticle density.
Although this dependency is nonlinear, it is assumed in this feasibility study that the intensity of the
stimulated light emission,
, is linearly proportional to each of the involved densities:
where the stimulating laser intensity distribution in the animal can be calculated by Eq. (3) with a known
laser source projected on the surface of the animal. When all the involved quantities are small, the linear
, (6)
system model should work well.
2.3. Discretization
Eqs. (3)-(4) can be discretized into a matrix equation linking the nanoparticle distribution
and
the NIR photon fluence rate
at a node
using the finite element analysis [14]:
4
rrrrrr,SDarrrSaD13asDs3R20,Drrrrn11211.43990.70990.66810.0636nnnrrνrr,DmrρrXrLrSrrrrρLXSρrr where the elements of the matrix A are
and the elements of the matrix F are
, (7)
, (8)
. (9)
where
we have
are the element shape functions. Since the matrix A in Eq. (7) is positive definite,
. (10)
2.4.
Image Reconstruction
In the compressive sensing (CS) framework, an image can be reconstructed from far less samples than
what the Nyquist sampling theorem requires [16]. Based on the image characteristics encountered in the
biomedical imaging applications, targeting nanoparticles are often attached to cells of a preferred type
and accumulated locally, forming a sparse and/or smooth distribution of nanoparticles. Using a CS
technique [17], we can reconstruct a nanophosphor concentration distribution by solving the following
optimization problem
(5)
The l1 term is to induce the solution sparsity. An interior-point method can be applied to solve a large-
scale l1-regularized optimization problem Eq. (5), aided by the preconditioned conjugate gradient
direction. An important property of the l1-regularized optimization is that a bound can be computed on
the sub-optimality of
.
3. Numerical simulation
3.1. Imaging System
5
ρFArrrrrrrrrrrd2ddjijiajiijDarrrrrdLXjiijf,,2,1iiρFAΦ10 1ρ ΦρFAρ1tosubjectminimum ρ The simulated imaging system consists of a high sensitive EMCCD camera, a mirror imaging device, an
X-ray tube, an X-ray collimator, a laser diode, and a sample stage. The EMCCD camera (iXon3 897,
Andor Technology) has a 512×512 resolution, 16μm×16μm pixel size, and >90% QE. As shown in Fig.
1, the camera faces the two mirrors, and the sample stage mounted on a motorized linear stage for focal
plane adjustment. The two-mirror position is adjusted by another linear stage. A mouse with a lie
prostrate position is caged in a transparent box positioned between the two mirrors, which is easy to
implement for the in vivo experiments. The two-mirror configuration is incorporated to expand the field
of view of the camera and acquire two views of the mouse simultaneously. The entire imaging system is
housed within an optically opaque box.
(a)
(b)
Fig. 1. Rendering of a stored luminescence CT (SLCT) system for small animal imaging. (a) The optical components, (b) a n
overview of the system (Dr. Fenglin Liu in our group made the system illustration).
3.2. Numerical Simulation
We performed representative numerical tests to evaluate the proposed SLCT
approach with a digital mouse phantom [cite ref for the phantom]. As shown in
Fig. 2, the mouse phantom was established from the CT slices of a mouse using
Amira (Amira 4.0, Mercury Computer Systems, Inc. Chelmsford, MA, USA).
The phantom was discretized into 306,773 tetrahedral elements with 58,244
nodes. The stored nanophosphor luminescence re-emission peak was chosen as
700 nm, which was based on the emission characteristics of MgGa2O4 : Cr3+ [8,
9]. The luminescence photon yield
was assumed to be 0.15cm3/mg for
Fig.2. Mouse phantom.
50keV [1]. Appropriate optical parameters were assigned to the mouse model accordingly. The reduced
scattering coefficient
relies on the stored luminescence emission wavelength
and is
approximated by an empirical function:
(14)
6
snmbsa10 where a and b are the constants depending on the tissue type. The organ-specific values for a and b can
be found in [18]. The tissue absorption depends on the local oxy-hemoglobin (HbO2), deoxy-
hemoglobin (Hb) and water (W) concentrations. The absorption coefficient
can be approximated
as the weighted sum of the three absorption coefficients
,
and
, which were
calculated from the corresponding absorbance spectra reported in [18]:
(15)
where
is the ratio between oxy-hemoglobin and total hemoglobin concentration,
and
are scaling factors. The nanophosphor concentration distribution was set in the mouse lung
region from 5µg/mL to 25µg/mL. Figures 3 (a) or 4 (a) show the nanophosphor clusters at the cross
section of z=10mm in the phantom.
X-ray luminescence computed tomography (XLCT): X-ray luminescence CT (XLCT) is a synergistic
imaging modality defining a permissible source region with X-rays to help reconstruct a nanophosphors
(NPs) distribution. These conventional NPs are instantaneously excited by common medical X--rays,
and the luminescent data are efficiently collected using a highly sensitive light camera. In the first
generation CT scanning mode, XCLT uses an x-ray pencil beam excitation, the emitted light can be
measured as a line integral. The classic filtered backprojection method can be used to reconstruct an
image, with image resolution being decided by the x-ray pencil beam aperture [1, 2]. This scanning
mode needs long data acquisition time and is not practical for most preclinical applications. To shorten
the scanning time, a cone beam x-ray luminescence computed tomography strategy was proposed [19].
In the cone beam stimulation mode, the X-rays illuminate the whole sample to stimulate all the
nanophosphors, and a CCD camera acquires luminescent photons for tomographic imaging. Clearly,
this cone beam scanning mode does not sufficiently utilize the primary benefit of XLCT in terms of a
reduced permissible source region.
Here, we present a fan-beam stimulation mode for XLCT, which uses a fan-beam of X-rays to irradiate
an object such as a mouse, and the nanoparticles on a cross-section of the mouse emit NIR light. The
measured NIR light signal (2D) on the external surface of the object is used to reconstruct a nanoparticle
distribution (2D) on the excited cross-section. Theoretically, the dimensionality of measure information
matches that of the unknown image. The fan-beam scanning mode is the optimal balance between the
7
a2aHbOaHbaWaWWaHbOaHbBaSxxS21HbHbOHbOx22BSWS pencil beam mode and the cone-beam mode for XLCT in terms of imaging efficiency and image quality,
and will be focused on in this project to show the advantages of SLCT.
An X-ray tube was operated at 50
KeV and 30 mA and collimated into a fan
beam with a 1mm thickness. A cross-
section was excited at the transverse
position of 10mm of the mouse phantom
described above. The NIR light emission
from
the excited nanoparticles was
recorded on the surface of the phantom by
the CCD camera. The collected NIR data
were then corrupted by 5% Gaussian
noise to simulate practical conditions. A
reconstruction method comparable
to
what has been proposed in Section 2 was
employed
to
reconstruct
the
nanophosphor distribution from the NIR
true nanophosphor
the
Fig. 3. Comparison between
distribution and the XLCT reconstruction (Unit: pico
Watts/mm3) (a) The true nanophosphor concentration
distribution
in
the mouse phantom; and
(b)
the
reconstructed nanophosphor concentration distribution
using XLCT in the fan-beam mode.
data. The reconstructed image revealed
the accumulation of the nanophosphors,
and generated a 37% relative error, which
was
defined
as
, where
and
are the true and reconstructed densities on the k-th mesh
element respectively, bg was assigned as the background value, and
the number of
elements in the set
. Fig. 3 presents a comparison between the true and reconstructed
nanophosphor distributions.
Stored Luminescence Computed Tomography (SLCT): The same mouse phantom and parameters
settings were used to evaluate the performance of the proposed SLCT approach. The SLCT experiment
8
bgtkktkrktktkbgNummax1tkrkbgNumtkbgktk: was performed in an onion peeling fashion. The phantom was stimulated with laser radiation shell by
shell. That is, peripheral regions were first stimulated for “cleaning-up”, and the data were collected for
recursive image reconstruction. The NIR emission signals were gradually collected synchronized to
stepwise NIR light stimulation (but still within a very short time window), instead of being collected
once for all as for XLCT. This divide-and-conquer data acquisition procedure ensures that the number of
unknowns can be significantly reduced to improve spatial resolution and stability of the image
reconstruction. To implement this procedure, two laser beams of 650nm stimulate the phantom. One
laser beam has the same illumination direction as that of the stimulating X-ray beam, while the other
laser beam came from the other side of the phantom in the opposite direction of the first laser beam to
stimulate the phantom simultaneously. At the first step, the NIR laser light stimulation mainly excited
the peripheral region of the phantom to read-out the nanophosphors near the surface of the phantom with
a highly sensitive CCD camera. When the NIR signal acquisition reached a sufficient signal-to-noise
ratio, the camera was recorded the first set of NIR light signals for the first shell tomographic imaging.
Then, the second shell was similarly approached, so on and so forth. The simulated NIR data on the
phantom surface were also corrupted by 5% Gaussian noise. The reconstruction method proposed in
Section 2 was step-wise/shell-wise employed to reconstruct the nanophosphor distribution from the NIR
data. The algorithm gave an excellent performance in terms of convergence and stability. The
reconstructed images are in a close agreement with the truth, as shown in Fig. 4., and the averaged
relative error of the reconstructed nanoparticle distribution was less than 22%.
9
Fig.4. Comparison between
the SLCT reconstruction and
true nanophosphor
concentration distributions (Unit: pico Watts/mm3: (a) True nanophosphor concentration
distribution in the numerical mouse phantom; (b) the reconstructed nanophosphor
concentration distributions in the peripheral region in the object at the first stage of NIR
stimulation; (c) the reconstructed nanophosphor concentration distributions in the
central region in the object at the second stage of NIR stimulation.
It can be observed in Figs. 3 and 4 that SLCT is more accurate and more stable than XLCT, as shown in
Table 1.
10
Table 1: Performance comparison between SLCT and XLCT (SLCT performance can be further improved with
an optimized NIR excitation scheme).
Modality
X-ray Stimulation
NIR Excitation Probe
Stability
Accuracy Error
SLCT
Fan beam
Applied
Nanophosphors
Strong
22%
XLCT
Fan beam
No
Nanophosphors Week
37%
4. Discussions and Conclusion
We have proposed the stored luminescence tomographic imaging modality which is based on
image reconstruction from the shrunk ROIs that physically results from stepwise peripheral clean-ups
through stored luminescence re-emissions upon NIR laser stimulations. Owing to deep penetration of
well-collimated X-ray excitation as well as NIR laser stimulations, the peripheral clean-ups and shrunk
ROIs can be thoroughly and flexibly achieved, making the imaging methodology applicable to high
resolution image reconstruction. It is advisable to consider various X-ray excitation patterns and NIR
light stimulation approaches for specific purposes in the image reconstruction. For convenience of raw
data acquisition and image reconstruction, this modality can achieve stepwise shrunk regions of interest
(ROIs) in a 3D object through one-time X-ray excitation followed by successive NIR light stimulations.
This stepwise treatment could satisfy accurate and reliable image reconstruction of nanophosphor
distribution. Numerical simulations demonstrated the feasibility of the proposed approaches. The stored
luminescence tomographic imaging modality may find its pre-clinical applications in monitoring drug
delivery and assessing cancer therapy.
References
Pratx G, Carpenter C, Sun C, Xing L: X-ray luminescence computed tomography via selective excitation:
a feasibility study. IEEE Trans Med Imag 2010, 29:1992-1999
Pratx G, Carpenter CM, Sun C, Rao RP, Xing L: Tomographic molecular imaging of x-ray-excitable
nanoparticles. Opt Lett 2010, 35(20):3345-3347.
Pratx G, Carpenter CM, Sun C, Xing L: X-Ray Luminescence Computed Tomography via Selective
Excitation: A Feasibility Study. Ieee T Med Imaging 2010, 29(12):1992-1999.
Ntziachristos V, Weissleder R: Experimental three-dimensional fluorescence reconstruction of diffuse
media using a normalized Born approximation. Opt Lett 2001, 26(12):893-895.
11
1.
2.
3.
4.
9.
11.
5.
6.
7.
8.
10.
Milstein AB, Oh S, Webb KJ, Bouman CA, Zhang Q, Boas DA, Millane RP: Fluorescence optical
diffusion tomography. Appl Optics 2003, 42(16):3081-3094.
Wang G, Cong W, Durairaj K, Qian X, Shen H, Sinn P, Hoffman E, McLennan G, Henry M: In vivo
mouse studies with bioluminescence tomography. Opt Express 2006, 14(17):7801-7809.
Wang G, Li Y, Jiang M: Uniqueness theorems in bioluminescence tomography. Med Phys 2004,
31(8):2289-2299.
Basavaraju N, Sharma S, Bessi`ere A, Viana B, Gourier D, Priolkar KR: Red persistent luminescence in
MgGa2O4 :Cr3+; a new phosphor for in vivo imaging. J Phys D: Appl Phys 2013, 46:375401.
Liu F, Yan W, Chuang YJ, Zhen Z, Xie J, Pan Z: Photostimulated near-infrared persistent luminescence
as a new optical read-out from Cr(3)(+)-doped LiGa(5)O(8). Scientific reports 2013, 3:1554.
Brannon-Peppas L, Blanchette JO: Nanoparticle and targeted systems for cancer therapy. Advanced Drug
Delivery Reviews 2004, 56(11):1649-1659.
Cong A, Cong W, Lu Y, Santago P, Chatziioannou A, Wang G: Differential evolution approach for
regularized bioluminescence tomography. IEEE transactions on bio-medical engineering 2010,
57(9):2229-2238.
12. Ma CM, Coffey CW, DeWerd LA, Liu C, Nath R, Seltzer SM, Seuntjens JP, American Association of
Physicists in M: AAPM protocol for 40-300 kV x-ray beam dosimetry in radiotherapy and radiobiology.
Med Phys 2001, 28(6):868-893.
13. Weersink RA, Qiu J, Hope AJ, Daly MJ, Cho BC, Dacosta RS, Sharpe MB, Breen SL, Chan H, Jaffray
DA: Improving superficial target delineation in radiation therapy with endoscopic tracking and
registration. Med Phys 2011, 38(12):6458-6468.
Cong W, Wang G, Kumar D, Liu Y, Jiang M, Wang LV, Hoffman EA, McLennan G, McCray PB,
Zabner J et al: Practical reconstruction method for bioluminescence tomography. Opt Express 2005,
13(18):6756-6771.
Schweiger M, Arridge SR, Hiraoka M, Delpy DT: The Finite-Element Method for the Propagation of
Light in Scattering Media - Boundary and Source Conditions. Med Phys 1995, 22(11):1779-1792.
Donoho DL: Compressed sensing. IEEE Transactions on Information Theory 2006, 52(4):1289-1306.
Candes EJ, Romberg J, Tao T: Robust uncertainty principles: Exact signal reconstruction from highly
incomplete frequency information. IEEE Transactions on Information Theory 2006, 52(2):489-509.
Alexandrakis G, Rannou FR, Chatziioannou AF: Tomographic bioluminescence imaging by use of a
combined optical-PET (OPET) system: a computer simulation feasibility study. Phys Med Biol 2005,
50(17):4225-4241.
Chen D, Zhu S, Yi H, Zhang X, Chen D, Liang J, Tian J: Cone beam x-ray luminescence computed
tomography: a feasibility study. Med Phys 2013, 40(3):031111.
16.
17.
18.
19.
14.
15.
12
Appendix: Energy diagram of doped semiconductors for stored nanophosphor luminescence.
Figure A-1: Energy diagram of doped semiconductors for stored nanophosphor luminescence.
13
|
1006.4374 | 1 | 1006 | 2010-06-22T21:00:11 | Surface residues dynamically organize water bridges to enhance electron transfer between proteins | [
"physics.bio-ph"
] | Cellular energy production depends on electron transfer (ET) between proteins. In this theoretical study, we investigate the impact of structural and conformational variations on the electronic coupling between the redox proteins methylamine dehydrogenase and amicyanin from Paracoccus denitrificans. We used molecular dynamics simulations to generate configurations over a duration of 40ns (sampled at 100fs intervals) in conjunction with an ET pathway analysis to estimate the ET coupling strength of each configuration. In the wild type complex, we find that the most frequently occurring molecular configurations afford superior electronic coupling due to the consistent presence of a water molecule hydrogen-bonded between the donor and acceptor sites. We attribute the persistence of this water bridge to a "molecular breakwater" composed of several hydrophobic residues surrounding the acceptor site. The breakwater supports the function of nearby solvent-organizing residues by limiting the exchange of water molecules between the sterically constrained ET region and the more turbulent surrounding bulk. When the breakwater is affected by a mutation, bulk solvent molecules disrupt the water bridge, resulting in reduced electronic coupling that is consistent with recent experimental findings. Our analysis suggests that, in addition to enabling the association and docking of the proteins, surface residues stabilize and control interprotein solvent dynamics in a concerted way. | physics.bio-ph | physics | Surface residues dynamically organize water bridges to enhance
electron transfer between proteins
Aurélien de la Lande,1,2,‡ Nathan S. Babcock,3,4 Jan Řezáč,1,2,∋
Barry C. Sanders,3,4 Dennis R. Salahub1,2,5,§
1 Department of Chemistry
2 Institute for Biocomplexity and Informatics
3 Department of Physics and Astronomy
4 Institute for Quantum Information Science
5 Institute for Sustainable Energy, Environment and Economy
University of Calgary, 2500 University Drive N. W., Calgary, Alberta, Canada, T2N 1N4
‡ Present address: Laboratoire de Chimie Physique—Centre National de la RechercheScientifique -
Unité Mixte de Recherche 8000, Université Paris-Sud 11, Bâtiment 349, Campus d’Orsay. 15, Rue
Georges Clémenceau, 91405 Orsay Cedex, France.
∋ Present address: Institute of Organic Chemistry and Biochemistry, Academy of Sciences of the Czech
Republic and Center for Biomolecules and Complex Molecular Systems, Flemingovo náměstí 2, 166 10
Prague 6, Czech Republic
§ To whom correspondence should be addressed: [email protected]
Relevant Topics: respiratory chain, Marcus theory, pathway model, dynamic docking, blue copper
proteins
This article contains supporting information online at
www.pnas.org/lookup/suppl/doi:10.1073/pnas.0914457107/-/DCSupplemental.
1
Abstract
Cellular energy production depends on electron transfer (ET) between proteins. In this
theoretical study, we investigate the impact of structural and conformational variations on the
electronic coupling between the redox proteins methylamine dehydrogenase and amicyanin from
Paracoccus denitrificans. We used molecular dynamics simulations to generate configurations
over a duration of 40ns (sampled at 100fs intervals) in conjunction with an ET pathway analysis
to estimate the ET coupling strength of each configuration. In the wild type complex, we find that
the most frequently occurring molecular configurations afford superior electronic coupling due
to the consistent presence of a water molecule hydrogen-bonded between the donor and acceptor
sites. We attribute the persistence of this water bridge to a “molecular breakwater” composed of
several hydrophobic residues surrounding the acceptor site. The breakwater supports the
function of nearby solvent-organizing residues by limiting the exchange of water molecules
between the sterically constrained ET region and the more turbulent surrounding bulk. When
the breakwater is affected by a mutation, bulk solvent molecules disrupt the water bridge,
resulting in reduced electronic coupling that is consistent with recent experimental findings. Our
analysis suggests that, in addition to enabling the association and docking of the proteins, surface
residues stabilize and control interprotein solvent dynamics in a concerted way.
2
Introduction
The electron transport chain is the cornerstone of biological energy transduction. All known
life-forms use membrane-bound chains of redox proteins to convert energy from food or sunlight into
chemical energy stored in adenosine triphosphate (1 ). Biological electron transfer (ET) often occurs
over long distances (>1nm) between protein-encapsulated redox cofactors separated by intervening
protein or solvent molecules. Over the last two decades, there has been increasing interest in water as
an “active constituent in cellular biology” (2 ). Today there is a growing body of evidence suggesting
that water plays an important role mediating long-range ET and that conformational fluctuations are
critical to protein-solvent interactions at ET interfaces (3 ,4 ,5 ,6 ,7 ,8 ,9 ,10 ). Notably, previous authors
have suggested that “ordered water molecules in the protein-protein interface may considerably
influence electronic coupling between redox centers” (4) and that “water may be a particularly strong
tunneling mediator when it occupies a sterically constrained space between redox cofactors with strong
organizing forces that favor constructively interfering coupling pathways” (7).
In general, however, the degree of sophistication of solvent-organizing effects at aqueous ET
interfaces remains unknown. In this study, we show that a pair of solvent-organizing residues in direct
contact with a bridging water molecule may be aided by surrounding residues that help stabilize local
solvent dynamics by mediating contact with the bulk. We predict that surface residues at protein-
protein interfaces can act collectively to organize and stabilize solvent structures and dynamics during
long-range ET.
The timely passage of electrons from protein to protein is crucial for proper metabolic
regulation (11 ) and relies on all the physical and chemical phenomena (i.e., diffusion, protein docking,
and ET reaction steps) that participate in overall electron transmission. Here we investigate the redox
reaction between methylamine dehydrogenase (MADH) and amicyanin taken from Paracoccus
denitrificans. This redox pair is representative of a broad class of interprotein ET reactions involving
blue copper proteins (12 ). Under methylotrophic growth conditions, MADH supplies electrons to
amicyanin, a blue copper protein that in turn shuttles electrons to various c-type cytochromes (13 ). In
vitro, the transfer occurs between the reduced tryptophan tryptophylquinone (TTQ) group on the
MADH β subunit and a cupric complex buried just under the amicyanin surface (Fig. 1). The oxidation
of MADH by amicyanin is a “true” ET reaction, limited by the ET reaction rate kET ( 14 ). Ma et al.
recently reported a series of site-directed mutations performed on the amicyanin methionine 51 residue
3
by alanine, lysine and leucine (Fig. 1) (15 ). Kinetic measurements revealed nearly a tenfold decrease in
kET without significant changes to the proteins’ overall structural, binding, or redox potentials. Ma et al.
concluded that “surface residues of redox proteins may not only dictate specificity for their redox
protein partners but also be critical to optimize the orientations of the redox centers and intervening
media within the protein complex for the ET event” (15).
To analyze an interprotein ET reaction like the reduction of amicyanin by MADH, it is
necessary to consider large ensembles of protein-protein complex configurations that contribute to the
average ET rate (16 ). For a “true” ET reaction, the ET rate kET can be estimated by nonadiabatic
Marcus-Hush-Levich theory (17 ,18 ),
k
ET
=
2
π
h
1
Tλk
π
B
4
T
DA
2
exp
‡
G
⎛ Δ−
⎜⎜
Tk
⎝
B
⎞
=⎟⎟
⎠
2
π
h
1
Tλ
k4
π
B
T
DA
2
exp
2
)
(
°
G
Δ−
k4
π
λ+
T
B
⎛
⎜⎜
⎝
,
⎞
⎟⎟
⎠
(1)
B is Boltzmann’s constant, T is the temperature, ΔG‡ is the
where
is the reduced Planck’s constant, k
h
Gibbs free energy of activation, ΔG° is the Gibbs free energy of reaction, λ is the reorganization
energy, and TDA is the superexchange matrix element which couples the donor and acceptor electronic
states quantum mechanically.
The experimental trends reported by Ma et al. strongly suggest the existence of a correlation
between protein motion and ET activity in the MADH−amicyanin complex. While most numerical
studies have addressed nanosecond timescales or shorter, we investigated longer timescales (40ns) that
are less well understood, thereby obtaining statistics for vibrational modes spanning several temporal
orders of magnitude. In addition to the wild type complex, we consider four amicyanin mutants: M51A,
M51L, M51K and M51C. The first three mutations correspond to those reported by Ma et al. (15),
whereas M51C was added to investigate the impact of a thiol replacing the original thioether. We
employed Molecular Dynamics (MD) simulations to generate the configurations within the transient
ET complex along with an ET pathway analysis to characterize each configuration’s intrinsic ET
activity (see Methods section). This computationally intensive investigation has allowed us to identify
important solvent-stabilizing functions of interprotein surface residues. Because of this solvent-
stabilizing effect, the most ET-active wild type conformations are also the most statistically favored
ones, an effect that is lost in the mutant complexes.
4
Results
We obtained 400 000 molecular configurations from 40ns of Molecular Dynamics simulations
of the wild type complex and each mutant. We employed the semi-empirical pathway model originally
developed by Beratan, Onuchic and Hopfield (19 ) to determine the tunneling pathway with the largest
electronic coupling matrix element TDA for each configuration. Although this model does not provide
an absolute value for the superexchange coupling matrix element TDA and does not account for
interferences between multiple pathways, it can be used to estimate the total electronic coupling decay
factor εtot, where TDA ≈ HDA⋅εtot and HDA is the theoretical “close contact” coupling matrix element
(20 ,21 ). Despite its simplicity, the pathway model has previously demonstrated excellent predictive
power when comparing different molecular configurations (8,22 ,23 ). We recorded the pathway with
the largest decay factor εtot for each configuration and labeled it by the surface residue through which
the electron exits the MADH. We found that in the vast majority of pathways (>99%), the electron
leaves the MADH through one of Ser β 56, Trp β 57, Val β 58, or Trp β 108 before tunneling through
one or more water molecules and entering the amicyanin through His 95. Adopting the terminology
already in use (21), it is convenient to define four distinct collections of similar pathways, or “pathway
tubes,” labeled by the letters A (Ser β 56), B (Trp β 57), C (Val β 58), and D (Trp β 108) (Figs. 2 and
S1). All the remaining excess pathways (labeled E) afford comparatively weak electronic coupling
(Table S1).
Pathway tube A is of particular interest due to its large coupling strength and high frequency of
occurrence (~60%, Table 1). For this reason, we further divided it into three subcategories: A1
represents a single pathway with a hydrogen-bonded bridge from the Ser β 56 O carbonyl oxygen
through a single water molecule to the His 95 HE2 proton; A2 represents all the remaining completely
hydrogen-bonded water bridges between Ser β 56 and His 95; and A3 represents all the partially broken
(i.e., van der Waals coupled) pathways from Ser β 56 to His 95. In a previous pathway analysis
performed on the crystal structure of the MADH−amicyanin dimer (24 ), Brook et al. found that the
strongest pathway required a through-vacuum jump from MADH Trp β 108 to amicyanin Pro 94 (25 ).
In contrast, in our solvated system we find that pathways involving a direct jump from Trp 108 to Pro
94 make up less than 0.01% of our data and are on average one tenth as strong as the completely
hydrogen-bonded A1 pathway. In another analysis performed on the MADH−amicyanin−cytochrome
ternary crystal structure, Chen et al. concluded that the strongest pathway involved a trapped interfacial
water molecule, even though it’s efficiency depended “critically on the presence of the water molecule
5
which may not always be occupied” (26 ). Chen et al.’s water-mediated pathway is identifiable as our
A1 pathway, assuming a hydrogen-bonded arrangement for the hydrogen atoms that were not resolved.
Thus, our computational study strongly supports Chen et al.’s water-bridge hypothesis and moreover
stresses the importance of the dynamical behavior of the ET pathways; very frequent switches between
the pathways are obtained in the course of the simulations (Fig. S2) but pathway A1 was favored only in
the wild type complex.
Configurations associated with pathway tube A1 depend critically on the probability Phb of a
water molecule forming two simultaneous hydrogen bonds with atoms Ser β 56 O and His 95 HE2. The
wild type complex’s statistical affinity for pathway A1 depends on this high probability (Phb > 50%)
compared with those of the mutants (Phb < 20%) (Table 1). In turn, the presence of this water bridge is
linked to the discrepancy in the average number of water molecules found at the interface between the
proteins. The consistent presence of the water molecule joining Ser β 56 and His 95 in the wild type, its
corresponding absence in the mutants, and the resulting reduction in the mutant coupling strength
indicate that solvent organization is vital to this reaction. Any destabilization of the A1 water bridge
results in a statistical shift towards less efficient pathways (Table 1).
The A1 pathway is disrupted when other water molecules jostle or compete with the bridging
water molecule (Fig. 2 A1). To determine the impact of surrounding water molecules on the A1 bridge,
we computed the number of water molecules within the “ET region”, which we define to be a sphere of
radius R centered between the MADH Ser β 56 O and amicyanin His 95 HE2 atoms. Even for large
radii (R = 5Å), an average of only 2.5 water molecules are present within the wild type ET region,
compared with mutant averages ranging from 4.7 to 6.3 (Fig. 3). Comparatively few water molecules
are exchanged between the wild type ET region and the surrounding bath. The mutant ET regions are
much more turbulent, as evidenced by the larger number of water molecules present, the lower
probability of hydrogen-bonded bridge formation Phb, and the higher rate of turnover τ between the
individual water molecules involved in forming the A1 bridge (Table 1). These results indicate that the
dynamical organization of the intervening solvent is crucial to the formation of the most efficient ET
configurations.
the calculated decay
factors
type
ratios of
A comparison of mutant-to-wild
mut
wt
r
mut
k
=
k
ET
mut
wt
k
ET
provides insight about the
6
2
ε
tot
2
ε
tot
and the experimental rates
mutr
ε =
impact of the Met 51 mutation on experimental rates through modifications of the electronic coupling
term. Our results are in overall qualitative agreement with experiment, as the decreases in the mutant
decay factors are of comparable size to the experimentally observed decreases in the ET rate constants
(Table 1). There is, however, a discrepancy between the ordering of the experimentally determined
factors
decay
pathway model
the
and
)
(
rates
mutant
1
r
r
r
L51M
K51M
A51M
k
k
k
>
>
>
(
1
>
>
>
27 ) produces
), discussed below. We find that the packing density model (
r
r
r
A51M
K51M
L51M
ε
ε
ε
decay factors similar to those of the pathway model for the M51L and M51K mutants, but predicts the
M51A and M51C mutants to be at least as kinetically competent as the wild type.
Discussion
It is remarkable that a single strongly-coupled pathway should dominate the thermal statistics of
the wild type complex alone (>50%, Fig. 3a). There is no a priori correlation between the binding
affinity of the complex and ET activity: Maxwell-Boltzmann statistics govern the probability of
occurrence of a given configuration, whereas the non-adiabatic Marcus expression (eq. 1)
independently determines the ET rate constant kET for that configuration. It is apparent that the protein
structure at the wild type interface is specifically suited to favor conformations most amenable to ET.
On the other hand, configurations associated with lower efficiency pathways become more statistically
prominent in the mutant distributions, thereby decreasing the average decay factor 〈εtot〉 (Table 1). This
statistical change also carries implications for the overall reaction kinetics.
In order to relate solvent dynamics directly to the mutation, it is necessary to consider
interactions between the amicyanin 51 residue and other protein or solvent molecules. Previously, when
performing a P52G mutation upon amicyanin, Ma et al. attributed the resulting reduction in kET to a
loss of interactions between the amicyanin Met 51 residue and the MADH Val 58 (28 ). This conclusion
is compatible with our simulation results, which show that conformational variations in the amicyanin
M51K and M51L residues allow an increased number of water molecules to “sneak” into the ET region
(Fig. S3). Intuitively, one expects the replacement of Met 51 by a smaller (alanine) or hydrophilic
(lysine) residue to allow more water molecules to enter the ET region. The cases of cysteine and
leucine are less straightforward to analyze, as both residues are hydrophobic and similar in size to
methionine. Many complex interactions influence the dynamics of these residues, and a future analysis
7
will have to examine a variety of chemical effects (e.g., methionine is a Lewis base whereas leucine is
more acidic). In this regard, our study highlights the importance of subtle interprotein surface dynamics
to the formation of efficient ET pathways.
Ma et al. used the “true, gated, and coupled ET” (17) framework to rationalize the decrease in
the ET rates in terms of protein motion at the interface. Based on large increases in the inferred values
of TDA and λ, as well as observed changes to the rate-limiting reaction kinetics for the N-quinol TTQ
form of MADH (15,28), Ma et al. inferred that the rate constants kET measured for the M51A and
M51K mutants were not those of the “true” ET reaction, as is the case for the wild type system. Rather,
they proposed that the M51A and M51K reactions were “gated” by an unidentified, separate, slower
pre-ET step x that imposed its rate kx over that of the actual ET event (15). On the other hand,
experimental data for the M51L mutant is similar to that of the true wild type reaction and is not
consistent with conformationally gated ET for which kET > kx. For the M51L mutant, Ma et al.
concluded that either kx ~ kET or that the ET reaction is kinetically “coupled” to a rapid but unfavorable
conformational rearrangement with equilibrium constant Kx, so that the observed rate is actually kET ×
Kx (15). This kinetically coupled picture (29 ) is compatible with the “dynamic docking” framework in
which “a large ensemble of weakly bound protein-protein configurations contribute to binding, but only
a few are reactive” (30 ). It is not clear why the mutation of the MADH Met 51 residue would lead to
gated ET (kET > kx) in the M51A and M51K mutants, but coupled ET (kET < kx) in the M51L mutant.
Our numerical analysis is consistent with the viewpoint that ET is modulated by rapid
interconversion within an ensemble of configurations of varying ET reactivity. The configurations
produced by our simulations exhibit a continuum of ET affinities, whereas the kinetically coupled and
dynamic docking models assume a simple active/inactive model of ET activity (29,30). This
active/inactive dichotomy fails to capture the variation in intermediate coupling strengths revealed by
our pathway analysis (Fig. 3). ET rate reductions comparable to the experimental ones are obtained by
summing the contributions to TDA arising from the various accessible molecular configurations within
the transient ET complex, without assuming a distinct pre-organization step. Because each
configuration is associated with an intrinsic ET coupling strength, it is enough to modulate the ET rate
simply by modifying each configuration’s respective statistical weight. We propose the hypothesis that
the increased amount of water at the ET interface dynamically modulates the ET rate in the mutants,
akin to kinetically coupled ET as described above.
8
Further work will be required to reproduce the exact experimental trend in the mutant rate
constants (Table 1). Variations in relevant parameters like ΔG° and λ contribute to the experimental rate
kET, but given that the wild type ET rate is kET = 10s−1, 40ns of MD simulation may not fully account
for these parameters. Furthermore, although the use of Langevin dynamics improves the sampling of
the configuration space, the artificial noise inherent to this method can also become a source of error.
The pathway model itself is limited by its inability to account for complex-valued interferences
between tunneling pathways. It successfully estimates the electronic coupling for molecular
configurations with a single dominant tunneling-pathway or a few constructively-interfering pathways,
but its accuracy is limited for configurations with multiple destructively-interfering pathways (7).
Because pathway “tube” A1 represents only one strongly-coupled pathway and because very few water
molecules are present at the wild type interface, the pathway model is expected to provide a good
coupling estimate for the statistically-favored wild type A1 configurations. The increased number of
water molecules at the mutant interfaces makes inter- and intra-tube destructive interference more
likely in the mutant complexes, and the pathway model may over-estimate the coupling strengths for
these configurations. The question of multiple interferences accentuates the potential importance of
solvent control to create one dominant strongly-coupled pathway at the protein interface.
Conclusions
Earlier studies on both inter- and intraprotein ET revealed the possibility of water-mediated ET
pathways in biological ET (6,31 ), as well as the specific role of protein residues stabilizing well-
defined ET pathways (8,32 ). Our study, however, provides the first evidence that several protein
surface residues can act together in concert to organize bridging water molecules, enhancing electron
transfer between proteins. Our numerical simulations indicate that MADH Ser β 56 and amicyanin His
95 work together to form a solvent-linked bridge between donor and acceptor, while the surrounding
hydrophobic residues act as a “molecular breakwater” to support the stability of this bridge (Fig. 1b).
Comparisons of the solvent organization in the wild type and mutant complexes show that the
amicyanin Met 51 residue plays an essential role, repelling bulk water molecules from the ET region
(Fig. 4). Any modification of the steric or electrostatic interactions at the Met 51 site—by either
replacement (15) or repositioning (28)—may disrupt this solvent-repelling mechanism. In this respect,
9
we believe that site-directed mutagenesis studies of the nearby amicyanin Met 28 and Met 71 residues
(Fig. 1) would also be of great interest. If Met 28 and Met 71 function in the same manner as Met 51,
mutations to these residues will produce reductions in kET similar to those found for Met 51.
More generally, our proposed solvent repelling mechanism depends on a patch of hydrophobic
surface residues surrounding the acceptor site, a characteristic shared by other blue copper proteins
(12). So far, this surface characteristic was believed to ensure a weak binding affinity of the redox
partners, allowing fast association and dissociation processes. Our study reveals another possible role
for a blue copper protein’s hydrophobic surface (Fig. 1b). It may enhance the ET activity of the redox
complex, controlling solvent dynamics to significantly improve the strength and stability of water-
mediated ET pathways.
Methods
Molecular Dynamics Simulations. We carried out Molecular Mechanics computations using the
CHARMM 33a package (33 ). We selected the ternary MADH−amicyanin−cytochrome-c551i complex
resolved to 1.9 Å by X-ray crystallography (Protein Data Bank entry 2GC4). This is a reasonable
starting structure for simulations since crystalline MADH has been demonstrated to be catalytically
competent to transfer electrons to amicyanin (34 ,35 ). After deleting the cytochrome-c551i from the
ternary complex, hydrogen atoms were added with the HBUILD routine (as implemented in
CHARMM) and the proteins were solvated in a TIP3P (36 ) water box of dimensions 115×80×80 Å3.
Approximately 40 Na+ ions were added to ensure electrical neutrality (depending on the mutant).
Histidine residues, including His 53 and His 95 were mono-protonated consistent with the experimental
pH of 7.5 (15). The mutant complexes were generated from this structure in silico using Molden (37 ).
The amicyanin cupric center was treated using the Force Field parameters developed by P. Comba et al.
for blue copper proteins (38 ). The Lennard-Jones parameters for the copper ion were ε = 0.05 kJ/mol
and σ = 2.13 Å (39 ). The wild type and mutant structures were first geometrically optimized by 500
steps of steepest descent algorithm and subsequent 1500 steps of Adopted Basis Newton-Raphson
optimizer. This was followed by 1ns of Langevin dynamics to ensure equilibration and a further 40ns
from which configurations were sampled every 100fs. The Shake algorithm was employed to constrain
hydrogenated bonds at their equilibrium bond lengths. A friction coefficient of 15 ps−1 and a bath
10
temperature of 298K were used to propagate the equations of motion within the Langevin approach.
Periodic boundary conditions were applied to simulate a continuous medium. Finally, a shift function
was used to compute electrostatic interactions between distant pairs of atoms, with a 12 Å cut-off. A
switch function was applied for van der Waals interactions (starting from 10 Å and set to zero at 12 Å).
This is the recommended (default) scheme in CHARMM to compute non-bonded terms.
Choice of Donor and Acceptor. We defined the donor based on the Density Functional Theory (DFT)
Highest Occupied Molecular Orbital (HOMO) of the tryptophan tryptophylquinol (TTQ) cofactor. For
these computations we used the deMon2k code (40 ) with the Perdew–Burke–Ernzerhof functional (41 )
and the DZVP-GGA basis sets. The catecholate ring represents 63% of the donor molecular orbital,
whereas the full MADH Trp β 57 aromatic ring represents almost 73%. There is very little orbital
delocalization onto the MADH Trp β 108 residue (less than 15% spread over its aromatic ring) and as
such it cannot be considered part of the donor group. We therefore restrict our definition of the donor to
the Trp β 57 catecholate ring, assigning a decay factor of 1 between the Trp β 57 atoms. Consequently,
the best pathway for a given configuration does not depend on the choice of the starting atom within
the MADH Trp β 57 catecholate ring. We note that our DFT-based definition of the donor orbital is
different from the one chosen in a previous pathway analysis where the electron density was assumed
to be delocalized across both cycles of the TTQ group and the Trp β 108 residue was therefore taken as
part of the donor group (25). The copper atom was taken as the acceptor since the beta Lowest
Unoccupied Molecular Orbital (LUMO) essentially consists of the copper dxy orbital (some
contributions are found on the Cys 92 residue but do not extend further than the sulphur atom 3p
orbital).
ET analysis. We chose the empirical pathway model originally developed by Beratan et al. (20) to
estimate εtot for the huge number of sampled molecular configurations. The pathway model allowed us
to classify the configurations in terms of distinct geometric motifs, directly relating conformational
fluctuations to variations in the coupling strength. The pathway model assumes that the electron can
tunnel from atom to atom along a given pathway, each interatomic step i contributing a coupling decay
factor εi. Individual covalently-bonded, hydrogen-bonded, and through-vacuum decay factors (denoted
εc, εhb, εv, respectively) were calculated based on semi-empirical formulae (eqs. 3-6). TDA is the product
of the first order close-contact matrix coupling element HDA and the total semi-empirical decay factor
εtot (eq. 3), where εtot is the product of N individual decay factors (N = Nc + Nhb + Nv, respectively). To
11
improve the accuracy of εtot, we used refined parameters derived recently from constrained Density
Functional Theory (DFT) (42 ,43 ) for the εhb term which depends on the hydrogen-bond angle φ and the
atom-to-atom distance R .We employed Dijkstra's algorithm (44 ) to find the pathway with the largest
coupling for each configuration. To make the search tractable, each protein complex was pruned to
about 300 atoms at the ET interface belonging to the following residues: amicyanin Met 28, Met 51,
Pro 52, His 53, Met 71, Cys 92, Pro 94, His 95, Met 98, Cu(II); MADH Ala β 55, Ser β 56, Trp β 57,
Val β 58, Pro β 100, Glu β 101, Trp β 108; and all water molecules within 7 Å of the amicyanin 95
HE2 or MADH 108 CD2 atoms.
DA H
T
=
ε⋅
tot
DA
=ε
tot
N
N
N
c
hb
v
∏∏∏
i
j
⋅ε
⋅
ε
c
hb
i
j
k
1
1
1
=
=
=
ε
k
v
c =ε
6.0
[pathway model]
=ε
hb
36.0
⋅
e
)01.2R(64.0
−
−
23.2
−
(cos
)1
+φ
⋅
e
⋅
e
01.2R(83.1
)(cos
−
−
)1
+φ
(2)
(3)
(4)
(5)
=ε
v
6.0
e
)4.1R(7.1
−⋅
−
(6)
For comparison with the pathway model, the packing density (27) approach was also tested. In
this case, εtot is written as a product of two exponential decay factors that involve the fraction of filled
space (i.e., space within the atoms’ van der Waals radii), the complementary fraction of vacuum space,
and the donor-acceptor separation RDA (eq. 8). The associated decay factor parameters βfill (0.45 Å−1)
and βvac (1.4 Å−1) were taken from Page et al. (27). To evaluate the fraction of filled space ffill we
defined 200 points regularly spaced along the donor-acceptor axis and determined for each of them the
presence of any surrounding atom within their van der Waals radius.
=ε
tot
e
−
f
β
fill
fill
R
DA
⋅
e
1(
−−
f
)
β
R
vac
fill
DA
[packing density model]
(7)
Figures. Molecular graphics were prepared with VMD (version 1.8.6) (45 ).
12
Acknowledgements
The authors wish to thank Sergei Noskov (Univ. Calgary, Canada) and Pascal Permot (Univ. Paris-Sud
11 – CNRS, France) for helpful discussions. This work was supported by the Natural Sciences and
Engineering Research Council of Canada (NSERC), the Canadian Institute for Advanced Research
(CIFAR), the Informatics Circle of Research Excellence (iCORE), and the Alberta Ingenuity Fund
(AIF). Computational resources were provided by WestGrid. B.C.S. is a CIFAR Fellow.
Figure 1: a) Solvated amicyanin (blue) in contact with MADH subunit β (red). Residues of interest to
ET are represented as liquorice. Redox cofactors are shown in purple and surface methionine residues
in orange. The dominant pathway A1 is shown as a transparent blue tube connecting the redox
cofactors. b) Direct (“head on”) view of interfacial residues represented by their chemical nature:
hydrophobic (white), hydrophilic (yellow), positively charged acidic (blue) and negatively charged
basic (red). The dotted green circles indicate the ET region on the surface of each protein. The
“molecular breakwater” is visible as a white ring of hydrophobic residues surrounding the amicyanin
His 95.
13
Figure 2: Representative pathways for each pathway tube in colors corresponding to those in Fig. 3.
Hydrogen bonds are represented by dotted lines and through-space jumps by solid ovals. The arrows on
A1 illustrate perturbations to the hydrogen bond network caused by another nearby water molecule.
14
Figure 3: Normalized distributions of sampled pathway tubes (a-e) and the average number of water
molecules (f) within the ET region defined by a sphere of radius R (sphere shown in Fig. 4).
15
Figure 4: Representative snapshots of the wild type (left) and M51A (right) interfaces, revealing the
breach in the molecular breakwater due to the mutation (other mutants shown in Fig. S3). The grey
sphere represents the ET region for radius R = 5.5 Å. Water molecules are shown in the van der Waals
representation.
16
kr
mut
〈εtot〉 × 103
rε
mut
(experiment)
wild type
1.0
M51L
0.68
M51K
0.49
M51A
0.13
M51C
—
(pathway analysis)
0.90 ± 0.03 0.47 ± 0.03 0.61 ± 0.02 0.65 ± 0.02 0.73 ± 0.02
(pathway analysis)
1.0
0.36 ± 0.04 0.52 ± 0.04 0.57 ± 0.04 0.76 ± 0.05
〈εtot〉 × 103
(packing density)
0.70 ± 0.03 0.42 ± 0.04 0.51 ± 0.03 0.62 ± 0.05 1.03 ± 0.05
rε
mut
(packing density)
1.0
0.56 ± 0.09 0.76 ± 0.07 0.89 ± 0.15 2.29 ± 0.26
Phb
τ (ns−1)
Table 1: Expectation values for 〈εtot〉 and the ratios
0.53
0.40
0.15
0.57
0.19
0.60
0.18
0.65
0.16
0.68
mutr
ε =
2
ε
tot
mut
wt
2
ε
tot
and
r
mut
k
=
k
ET
mut
wt
k
ET
obtained from packing density and pathway analysis of the configurations gathered along the molecular
simulations. The uncertainties account for the sampling errors of the computational protocol (see
Supporting Information). Experimental rates kET were obtained from k3 (at 30°C) in Table 3 of ref. (15)
(M51C was not reported). Phb is the unit-normalized likelihood that a water molecule is simultaneously
hydrogen bonded to both the MADH Ser β 56 O and amicyanin His 95 HE2 atoms during our
simulations. The turnover τ of the bridging water molecule is defined as the number of different water
molecules that participate in pathway A1 divided by the length of the simulation in nanoseconds.
17
References
1. Lane N (2006) Power, sex, suicide: mitochondria and the meaning of life (Oxford University Press, New York).
2. Ball P (2008) Water as an active constituent in cell biology. Chem. Rev. 108:74-108.
3. Tezcan FA, Crane BR, Winkler JR, Gray HB (2001) Electron tunneling in protein crystals. Proc. Natl. Acad. Sci.
98:5002-5006.
4. van Amsterdam IMC et al (2002) Dramatic modulation of electron transfer in protein complexes by crosslinking. Nat.
Struct. Biol. 8:48-52.
5. Francisco WA, Wille G, Smith AJ, Merkler DJ, Klinman JP (2004) Investigation of the pathway for inter-copper
electron transfer in peptidyclycine α-amidating monooxygenase. J. Am. Chem. Soc. 126:13168-13169.
6. Miyashita O, Okamura MY, Onuchic JN (2005) Interprotein electron transfer from cytochrome c2 to photosynthetic
reaction center: Tunneling across an aqueous interface. Proc. Natl. Acad. Sci. 102:3558-3563.
7. Lin J, Balabin IA, Beratan DN (2005) The nature of aqueous tunneling pathways between electron-transfer proteins.
Science 310:1311-1313.
8. de la Lande A, Martí S, Parisel O, Moliner V. (2007) Long distance electron-transfer mechanism in peptidylglycine α-
hydroxylating monooxygenase: A perfect fitting for a water bridge. J. Am. Chem. Soc. 129:11700-11707.
9. Balabin IA, Beratan DN, Skourtis SS (2008) Persistence of structure over fluctuations in biological electron-transfer
reactions. Phys. Rev. Lett. 101:158102.
10. Beratan DN, et al. (2009) Steering electrons on moving pathways. Acc. Chem. Res. 42:1669-1678.
11. de Rienzo F, Gadboulline RR, Wade RC, Sola M, Menziani MC (2004) Computational approaches to structural and
functional analysis of plastocyanin and other blue copper proteins. Cell. Mol. Life. Sci. 61:1123-1142.
12. de Rienzo F, Gabdoulline RR, Menziani MC, Wade RC (2000) Blue copper proteins : a comparative analysis of their
molecular interaction properties. Protein Sci. 9:1439-1454.
13. Harms N, van Spanning RJM (1991) C1 Metabolism in Paracoccus denitrificans: Genetics of Paracoccus
denitrificans. J. Bioenerg. Biomembr. 23:187-210.
14. Brooks HB, Davidson VL (1994) Free energy dependence of the electron transfer reaction between methylamine
dehydrogenase and amicyanin. J. Am. Chem. Soc. 116:11201-11202.
15. Ma JK, Wang Y, Carrell CJ, Mathews FS, Davidson VL (2007) A single methionine residue dictates the kinetic
mechanism of interprotein electron transfer from methylamine dehydrogenase to amicyanin. Biochem. 46:11137-11146.
16. Leys D, Basran J, Talfournier F, Sutcliffe MJ, Scrutton NS (2003) Extensive conformational sampling in a ternary
electron transfer complex. Nat. Struct. Biol. 10:219-225.
17. Davidson VL (2008) Protein control of true, gated, and coupled electron transfer reactions. Acc. Chem. Phys. 41:730-
738.
18. Marcus RA, Sutin N (1985) Electron transfers in chemistry and biology. Biochim. Biophys. Acta 811:265-322.
19. Beratan DN, Onuchic JN, Hopfield JJ (1987) Electronic coupling through covalent and noncovalent pathways in
proteins. J. Chem. Phys. 86:4488-4498.
20. Beratan DN, Betts JN, Onuchic JN (1991) Protein electron transfer rates set by the bridging secondary and tertiary
structure. Science, 252:1285-1288.
21. Curry WB. et al. (1995) Pathways, pathway tubes, pathway docking, and propagators in electron transfer proteins. J.
Bioener. Biomembr. 27:285-293.
22. Gray HB, Winkler JR (2003) Electron tunnelling through proteins. Q. Rev. Biophys. 36:341-372.
23. Jasaitis A, Johansson MP, Wikström M, Vos MH, Verkhovsky M (2007) Nanosecond electron tunneling between the
hemes in cytochrome bo3, Proc. Nat. Acad. Soc. 104:20811-20814.
24 Chen L et al (1992) Crystal structure of an electron-transfer complex between methylamine dehydrogenase and
amicyanin. Biochem. 31: 4959-4964.
25 Brooks HB, Davidson VL (1994) Kinetic and thermodynamic analysis of a physiologic intermolecular electron-transfer
reaction between methylamine dehydrogenase and amicyanid. Biochem. 33:5696-5701.
26. Chen, L., Durley, R.C.E., Mathews, F. S. & Davidson, V. L. (1994) Structure of an electron transfer complex:
methylamine dehydrogenase, amicyanin, and cytochrome c551i. Science, 264, 86-89.
27. Page CC, Moser MC, Chen X, Dutton PL (1999) Natural engineering principles of electron tunnelling in biological
oxidation-reduction. Nature, 402:47-52.
28. Ma JK Carrell CJ, Mathews FS, Davidson VL (2006) Site-directed mutagenesis of proline 52 to glycine in amicyanin
converts a true electron transfer reaction into one that is conformationally gated. Biochem. 45:8284-8293.
29. Davidson VL (2000) Effects of kinetic coupling on experimentally determined electron transfer parameters. Biochem.
45:4924-4928
30. Liang Z-X et al. (2004) Dynamic docking and electron-transfer between cytochrome b5 and a suite of myoglobin
surface-charge mutants. Introduction of a functional-docking algorithm for protein-protein complexes J. Am. Chem.
18
Soc. 126:2785-2798.
31. Bizzarri AR, Brunori E, Bonanni B, Cannistraro S (2007) Docking and molecular dynamics simulation of the azurin–
cytochrome c551 electron transfer complex. J. Mol. Recognit. 20:122-131.
32. Prytkova TR, Kurnikov IV, Beratan DN (2007) Sensitivity in protein electron transfer coupling coherence distinguishes
structure. Science, 315:622-625.
33. Brooks BR et al. (1983) CHARMM: A program for macromolecular energy, minimization, and dynamics calculations.
J. Comput. Chem. 4:187-217.
34. Merli A et al. (1996) Enzymatic and electron transfer activities in crystalline protein complexes. J. Biol. Chem.
271:9177-9180.
35. Ferrari D et al. (2004) Electron transfer in crystals of the binary and ternary complexes of methylamine dehydrogenase
with amicyanin and cytochrome c551i as detected by EPR spectroscopy. J. Biol. Inorg. Chem. 9:231-237.
36. Jorgensen WL, Chandrasekhar J, Madura JD, Impey RW, Klein ML (1983) Comparison of simple potential functions
for simulating liquid water. J. Chem. Phys. 79:926-935.
37. Schaftenaar G, Noordik JH, (2000) Molden: a pre- and post-processing program for molecular and electronic structures.
J. Comput. Aided Mol. Design, 14:123-124.
38. Comba P, Remeny R (2002) A new molecular mechanics force field for the oxidized form of blue copper proteins. J.
Comp. Chem. 23:697-705.
39. Buning C, Comba P. (2000) protonation of the copper(I) form of the blue copper proteins plastocyanin and amicyanin -
A molecular dynamics study. Eur. J. Inorg. Chem. 2000:1267-1273.
40. Köster AM et al deMon developers, (2006). http://www.demon-software.com/public_html/index.html
41. Perdew JP, Burke K, Ernzerhof M (1996) Generalized gradient approximation made simple. Phys. Rev. Lett. 77:3865-
3868.
42. Wu Q, van Voorhis T (2006) Extracting electron transfer coupling elements from constrained density functional theory.
J. Chem. Phys. 125:164105-8.
43. de la Lande A, Salahub DR (2010) Derivation of interpretative models for long range electron transfer from
constrained density functional theory. J. Mol. Struct. THEOCHEM 943:115-120
44. Dijkstra EW (1959) A note on two problems with connections with graphs Num. Math. 1, 269-271
45. Humphrey W, Dalke A, Schulten K. (1996) VMD: Visual Molecular Dynamics. J. Mol. Graphics 14, 33-38.
19
|
1712.01758 | 1 | 1712 | 2017-12-05T16:59:03 | Multiphasic profiles for the biologically important ion activities of NaCl and KCl | [
"physics.bio-ph"
] | When plotted in linear transformations of the Michaelis-Menten equation, ion uptake in plants has been shown to be multiphasic, i.e. to be represented by a series of straight lines separated by discontinuous transitions (Nissen 1971, 1974, 1991, 1996). Reanalysis of data for other transport, binding and enzyme systems has also given such kinetics (Nissen and Mart\'in-Nieto 1998). Recently, such profiles have been found for a variety of biological as well as non-biological processes and phenomena, including activation of ion channels, binding, pH, folding/unfolding, effects of various interactions and chain length (Nissen 2015a,b, Nissen 2016a-d). These biological and nonbiological processes have little in common beyond the ions involved. Here I show that the multiphasic property seen in so many systems is present in the fundamental physical properties of the ion themselves.
The characteristics of multiphasic profiles have now also been found in profiles for activities of ions in simple inorganic solutions, indicating that they are somehow causing the multiphasic profiles in more complex systems. A set of extensive and precise data (Lee et al. 2002) for the activities of the biologically very important ions Na+, K+ and Cl- (in NaCl and KCl) at four different temperatures (15, 25, 35 and 45oC) will be reanalyzed in the present paper. The activities of NaBr and KBr will be plotted in a forthcoming paper. Extensive data for these and other ion activities Wilczek-Vera et al. (2004) can also be precisely represented by multiphasic profiles (in preparation). | physics.bio-ph | physics | Multiphasic profiles for the biologically important ion activities of
NaCl and KCl
Per Nissen
Norwegian University of Life Sciences
Department of Ecology and Natural Resource Management
P. O. Box 5003, NO-1432 Ås, Norway
[email protected]
2
Introduction
When plotted in linear transformations of the Michaelis-Menten equation, ion uptake in plants has been
shown to be multiphasic, i.e. to be represented by a series of straight lines separated by discontinuous
transitions (Nissen 1971, 1974, 1991, 1996). Reanalysis of data for other transport, binding and enzyme
systems has also given such kinetics (Nissen and Martín-Nieto 1998). Recently, such profiles have been
found for a variety of biological as well as non-biological processes and phenomena, including activation of
ion channels, binding, pH, folding/unfolding, effects of various interactions and chain length (Nissen
2015a,b, Nissen 2016a-d). These biological and nonbiological processes have little in common beyond the
ions involved. Here I show that the multiphasic property seen in so many systems is present in the
fundamental physical properties of the ion themselves.
The characteristics of multiphasic profiles have now also been found in profiles for activities of ions in
simple inorganic solutions, indicating that they are somehow causing the multiphasic profiles in more
complex systems. A set of extensive and precise data (Lee et al. 2002) for the activities of the biologically
very important ions Na+, K+ and Cl- (in NaCl and KCl) at four different temperatures (15, 25, 35 and 45oC)
will be reanalyzed in the present paper. The activities of NaBr and KBr will be plotted in a forthcoming
paper. Extensive data for these and other ion activities Wilczek-Vera et al. (2004) can also be precisely
represented by multiphasic profiles (in preparation).
Reanalysis
Data for the activity coefficients of Na+ and Cl- in aqueous solutions of various concentrations at 288.15,
298.15, 308.15 and 318.15 K (Tables 1 and 6 in Lee et al. 2002) have been plotted against each other (Figs
1-8). Data for the activity coefficients of K+ and Cl- (Tables 2 and 5 in Lee et al. 2002) have also been
plotted (Figs 9-16).
As shown by the very high absolute r values for the straight lines (47 of the 58 r values in Figs 1-16 are
0.999 or higher), the data can, without exception, be well represented by straight lines. When lines intersect
in a common point, the transition is obviously discontinuous (intersection between lines VII and VIII in Fig.
3, between lines I and II and between lines II and III in Fig. 5, between lines V and VI in Fig. 6, between
lines I and II in Fig. 7, between lines III and IV in Fig. 9). When lines intersect close to a point, the
transition is probably also discontinuous.
Adjacent lines are quite often parallel or nearly so and are then necessarily noncontiguous, i.e. the transition
is in the form of a jump (lines IV and V in Fig. 3, lines V and VI in Fig. 4, lines III, IV and VI in Fig. 5,
lines IV and V in Fig. 6, lines III and IV and lines V and VI in Fig. 8, lines II and III and lines V and VI in
Fig. 10, lines III and IV and lines V and VI in Fig. 11, lines IV and V in Fig. 13, lines V and VI in Fig. 14,
lines V and VI in Fig. 15, lines IV and V in Fig. 16). Many of the jumps are tiny, but the often high r values
of the adjacent lines indicate that the assignments are correct.
3
Fig. 1. Eight phases. Transitions at 0.148, 0.519, 0.670,
0.828 and 0.957, between 1.095 and 1.225, and between
1.225 and 1.342. Insufficiently detailed data for resolution
of phase VII. Lines IV, VI and VIII are parallel, precisely
so for VI and VIII. Very and exceedingly high absolute r
values for lines I and II.
Fig. 2. Nine phases. Transitions at 0.143 and 0.388,
between 0.548 and 0.632 (jump), at 0.732, between 0.837
and 0.894 (jump), at 0.967, between 1.095 and 1.225, and
between 1.225 and 1.342. Insufficiently detailed data for
resolution of phase VIII. Lines III and IV are parallel, as
are approximately also lines VII and IX. Very high abso-
lute r value for line I. Lines IV and V are probably not a
single line (r = 0.685.
Fig. 3. Eight phases. Transitions at 0.140 and 0.377,
between 0.548 and 0.632 (jump), between 0.707 and
0.775 (jump), and at 0.852, 0.970 and 1.225. Lines III-V
are roughly parallel. Very high absolute r values for lines
I, VII and VIII.
Fig. 4. Eight phases. Transitions at 0.152, 0.540, 0.660
and 0.822, between 0.949 and 1.000 (small jump),
between 1.095 and 1.225, and between 1.225 and 1.342.
Insufficiently detailed data for resolution of phase VII.
Lines V, VI and VIII are parallel or approximately so.
Very high absolute r values for lines I and II.
4
Fig. 5. Eight phases. Transitions at 0.316 and 0.548,
between 0.707 and 0.775 (small jump), between 0.837
and 0.894, between 0.894 and 0.949, between 1.000 and
1.095, and between 1.095 and 1.225. Insufficiently detailed
data for resolution of phases V and VII. Lines III, IV and VI
are about parallel. Very high absolute r values for lines I-III.
Fig. 6. Seven phases. Transitions at 0.369, between 0.548
and 0.632, between 0.632 and 0.707, between 9.775 and
0.837 (tiny jump), at 0.949, and between 1.095 and 1.225
(jump). Insufficiently detailed data for resolution of phase III.
Lines IV and V are parallel. Very high absolute r values for
Lines I, V, VI and VII.
Fig. 7. Six phases. Transitions at 0.447, between 0.837 and
0.894, between 0.894 and 0.949, at 1.157, and between 1.342
and 1.414. Insufficiently detailed data for resolution of phases
III and VI. High to very high absolute r values for lines I, II
and IV. There may possibly be two phases in the range of line I
(r for the first 6 points = -0.9998).
Fig. 8. Seven phases. Transitions between 0.224 and 0.316
(jump), between 0.447 and 0.548 (tiny jump), between 0.707
and 0.775 (tiny jump), between 0.949 and 1.000 (tiny jump),
between 1.095 and 1.225 (tiny jump), and between 1.342 and
1.414. Insufficiently detailed data for resolution of phase VII.
Lines IV and V are parallel. Very to exceedingly high absolute
r values for lines I, III and IV.
5
Fig. 9. Eight phases. Transitions at 0.116, 0.324 and
0.632, between 0.775 and 0.837 (tiny jump), between
1.000 and 1.095, between 1.095 and 1.183, and at 1.303.
Insufficiently detailed data for resolution of phase VI.
Very high absolute r values for lines I, III, IV and V. A
single line in the range of phases VI-VIII seems
unlikely (r = -0.990).
Fig. 10. Six phases. Transitions at 0.278, between 0.548
and 0.632 (tiny jump), between 0.775 and 0.837 (tiny
jump), between 1.000 and 1.095 (tiny jump), and between
1.22 and 1265 (jump, not visible). Lines II-IV are roughly
parallel, but the very high absolute r values show that the
lines are separate.
Fig. 11. Six phases. Transitions at 0.259 and 0.499,
between 0.707 and 0.775 (small jump), between 1.000 and
1.095 (jump), and between 1.225 and 1.265 (jump, not visible).
Lines III and IV are precisely parallel, as are lines V and
VI. Very high absolute r values for lines I, IV and VI.
Fig. 12. Six phases. Transitions between 0.100 and
0.224 (jump), at 0.510, between 0.707 and 0.775 (jump),
between 1.000 and 1.095, and between 1.183 and 1.225
(tiny jump). Lines I and II are about parallel, as are lines
V and VI. Very high absolute r values for lines I-IV.
6
Fig. 13. Eight phases. Transitions at 0.160, between 0.316
and 0.447 (jump), at 0.658, between 0.837 and 0.894
(tiny jump), between 1.000 and 1.095, between 1.095 and
1.183, and at 1.239. Insufficiently detailed data for
resolution of phase VI. Lines IV and V are parallel, but
probably not a single line (r = -0.999 vs. r = -0.9996
for line IV.) Very high absolute r values for lines I, III, IV
and VIII.
Fig. 14. Nine phases. Transitions at 0.190, 0.398 and
0.606, between 0.707 and 0.775 (tiny jump), between 0.894
and 0.949 (tiny jump), between 1.000 and 1.095, between
1.095 and 1.183, and at 1.324. Insufficiently detailed data
for resolution of phase VII. Lines IV-VI are parallel
or approximately so, but are probably not a single line
(r = -0.995). Very high absolute r values for lines I, V
and VIII.
Fig. 15. Nine phases. Transitions at 0.186, between 0.316
and 0.447, at 0.685, between 0.775 and 0.837 (tiny jump),
between 0.894 and 0.949 (tiny jump), between 1.000 and
1.095, between 1.095 and 1.183, and at 1.287. Insuffi-
ciently detailed data for resolution of phase VII. Lines
V and VI are parallel, as is approximately also line IV.
Lines IV-VI are probably not a single line (r = -0.992).
Very high absolute r values for lines I and III.
Fig. 16. Ten phases. Transitions at 0.185, between 0.316
and 0.447, between 0.447 and 0.548, between 0.707 and
0.775 (tiny jump), at 0.860, between 0.949 and 1.000 (tiny
jump), between 1.095 and 1.183 (jump), at 1.243, and
between 1.342 and 1.414. Insufficiently detailed data for
resolution of phases III and X. Lines IV and V are about
parallel, as are lines VI and VII (line VI is horizontal).
Lines VI and VII are probably not a single line (r = -0.944).
Very high absolute r values for lines I and IV.
7
As concluded by the authors, the ion activity coefficients decreased with increasing temperature. However,
in contrast to their representation (Figs 2 and 3 in Lee et al. 2002) of the profiles for ion activities as
curvilinear, the profiles are precisely multiphasic. The profiles for Na+ (from NaCl) can be represented by 8
or 9 phases (Figs 1-4). The profiles has a minimum at 0.632-0.837 and increase again at higher
concentrations. The profiles for Cl- (from NaCl) can be represented by 6-8 phases (Figs 5-8) and decrease
with increasing ion concentration. The profiles for K+ (from KCl) can be represented by 6 phases for the
three highest temperatures (Figs 10-12), but by 8 phases for the lowest temperature (Fig. 9). In marked
contrast to the profiles for Na+, the ion activities decreased over the entire concentration range. The profiles
for Cl- (from KCl) can be represented by 8-10 phases (Figs 13-16). Again, the activities decreased with
increasing concentrations, maybe except for a slight increase at a few of the highest concentrations. The
profiles for any one ion are very similar at the various temperatures, and it seems that any differences in the
patterns may be accidental.
Conclusions and Questions
As shown, the present data are very if not exceedingly precise. It seems that they allow some far-reaching
conclusions, but they do also present difficult questions. In many experimental studies, in non-biological as
well as in biological systems, the systems have been taken to be continuous as a function of the parameter
being varied, without any disruptions. However, it has now been shown (see Introduction) that many
systems are in fact discontinuous rather than continuous. There has been little or no awareness of these
discontinuities which seem to originate from discontinuities in ion activities in the system being studied.
Clearly, such studies should now be carried out with sufficient detail and precision for any discontinuities to
be recognized.
The molecular basis of the discontinuities remains unclear. Why should there be clear and reproducible
discontinuities in simple salt solutions? Why should the lines be straight, apparently perfectly so? Why are
adjacent lines quite often parallel or nearly so? Why do the profiles for Na+ and K+ differ so markedly? To
what extent does the finding of multiphasic profiles invalidate conclusions from less thorough studies?
Acknowledgment – I am very grateful to Bob Eisenberg for his continued interest and encouragement.
References
8
Lee L-S, Chen T-M, Tsai K-M, Lin C-L (2002) Individual ion and mean activity coefficients in NaCl, NaBr,
KCl and KBr aqueous solutions. J Chin Inst Chem Engrs 33: 267-281.
Nissen P (1971) Uptake of sulfate by roots and leaf slices of barley. Physiol Plant 24: 315-324.
Nissen P (1974) Uptake mechanisms. Inorganic and organic. Annu Rev Plant Physiol 25: 53-79.
Nissen P (!991) Multiphasic uptake mechanisms in plants. Int Rev Cytol 129: 89-134.
Nissen P (1996) Uptake mechanisms. Pp 511-527 in: Waisel Y, Eshel A, Kafkafi U (eds). Plant Roots. The
Hidden Half (2. ed). Marcel Dekker, Inc, New York.
Nissen P (2015a) Discontinuous transitions: Multiphasic profiles for channels, binding, pH, folding and
chain length. Posted on arXiv.org with Paper ID arXiv:1511.06601.
Nissen P (2015b) Multiphasic pH profiles for the reaction of tris-(hydroxymethyl)-aminomethane with
phenyl esters. Posted on arXiv.org with Paper ID arXiv:1512.02561.
Nissen P (2016a) Profiles for voltage-activated currents are multiphasic, not curvilinear. Posted on arXiv.org
with Paper ID arXiv:1603.05144.
Nissen P (2016b) Multiphasic profiles for voltage-dependent K+ channels: Reanalysis of data of MaKinnon
and coworkers. Posted on arXiv.org with Paper ID arXiv:1606.02977.
Nissen P (2016c) 'Perfectly' curvilinear profiles for binding as determined by ITC may in fact be
multiphasic. Posted on arXiv.org with Paper ID arXiv:1606.09133.
Nissen P (2016d) Multiphasic interactions between nucleotides and target proteins. Posted on arXiv.org with
Paper ID at arXiv:1608.07459.
Nissen P, Martín-Nieto J (1998) 'Multimodal' kinetics: Cyanobacteria nitrate reductase and other enzyme,
transport and binding systems. Physiol Plant 104: 503-511.
Wilczek-Vera G, Rodil E, Vera JH (2004) On the activity of ions and the junction potential: Revised values
for all data. AIChE Journal 50: 445-462.
|
1709.04698 | 2 | 1709 | 2017-10-24T10:27:11 | Controlling Anomalous Diffusion in Lipid Membranes | [
"physics.bio-ph"
] | Diffusion in cell membranes is not just simple two-dimensional Brownian motion, but typically depends on the timescale of the observation. The physical origins of this anomalous sub-diffusion are unresolved, and model systems capable of quantitative and reproducible control of membrane diffusion have been recognised as a key experimental bottleneck. Here we control anomalous dif- fusion using supported lipids bilayers containing lipids derivatized with polyethylene glycol (PEG) headgroups. Bilayers with specific excluded area fractions are formed by control of PEG-lipid mole fraction. These bilayers exhibit a switch in diffusive behaviour, becoming anomalous as bilayer continuity is disrupted. Diffusion in these bilayers is well-described by a power-law dependence of the mean-square displacement with observation time. The parameters describing this diffusion can be tailored by simply controlling the mole fraction of PEG-lipid, producing bilayers that exhibit diffusive behaviour with similar characteristics to those observed in biological membranes. | physics.bio-ph | physics | Controlling Anomalous Diffusion in Lipid Membranes
H. L. E. Coker* and M. R. Cheetham*
*Contributed equally to this work.
Department of Chemistry, Britannia House, King's College London, London SE1 1DB and
Chemistry Research Laboratory, 12 Mansfield Road, University of Oxford, Oxford OX1 3TA
D. R. Kattnig
Living Systems Institute & Department of Physics,
University of Exeter, Stocker Road, Exeter EX4 4QD
Department of Physics, Strand Building, King's College London, London WC2R 2LS
Y. J. Wang and S. Garcia-Manyes
Department of Chemistry, Britannia House, King's College London, London SE1 1DB
(Dated: October 25, 2017)
M. I. Wallace
Diffusion in cell membranes is not just simple two-dimensional Brownian motion, but typically
depends on the timescale of the observation. The physical origins of this anomalous sub-diffusion
are unresolved, and model systems capable of quantitative and reproducible control of membrane
diffusion have been recognized as a key experimental bottleneck. Here we control anomalous dif-
fusion using supported lipids bilayers containing lipids derivatized with polyethylene glycol (PEG)
headgroups. Bilayers with specific excluded area fractions are formed by control of PEG-lipid mole
fraction. These bilayers exhibit a switch in diffusive behavior, becoming anomalous as bilayer con-
tinuity is disrupted. Diffusion in these bilayers is well-described by a power-law dependence of
the mean square displacement with observation time. The parameters describing this diffusion can
be tailored by simply controlling the mole fraction of PEG-lipid, producing bilayers that exhibit
anomalous behavior similar to biological membranes.
I.
INTRODUCTION
Diffusion is an essential transport mechanism in
membrane biology, vital for a wide range of biological
function including protein organization [1], signalling
[2, 3] and cell survival [4]. Interestingly, such living
systems do not in general display the Brownian mo-
tion predicted by a simple random walk model, and
instead exhibit 'anomalous' diffusion [5] where the dif-
fusivity is dependent on the timescale of observation.
This phenomenon has been reported both for three-
dimensional diffusion in the cell cytosol [6] and two-
dimensional diffusion in the plasma membrane [7–9].
Here we focus on membrane diffusion.
Why and how anomalous diffusion exists in the
plasma membrane has been the subject of consider-
able investigation (reviewed in [10]). The common
underlying mechanism is thought to be the crowded
environment found in the cell membrane [11], and the
presence of slower-moving obstacles [12, 13], pinning
sites, and compartmentalization [11, 14, 15] have all
been suggested as potential contributors to anoma-
leity in membrane diffusion. Confinement in cellu-
lar membranes is observed on the order of tens to
hundreds of nanometers, with anomalous diffusion re-
ported in a large number of cell types [14, 15]. Overall
this work has led to the adoption of a compartmen-
talized 'picket fence' model of the cell membrane as
a proposed improvement to the 'fluid-mosaic' model
[11].
Artificial lipid bilayers have played a key role in
improving our understanding of anomalous diffusion
[16–21], where both phase separation [17] and pro-
tein binding [18] in supported lipid bilayers (SLBs)
have been used to generate anomalous diffusion. Sim-
ulations have also been vital in advancing our under-
standing, with much pioneering [5, 22, 23] and recent
[24–29] work in this area. In particular, simulations
have helped elucidate the role of mobile and immobile
obstacles in causing anomalous behavior [12, 13]. Rel-
evant to our work, simulations have also been used to
better interpret single particle tracking data [30] and
provide methods to discriminate between classes of
anomalous diffusion [31].
Despite these advances,
the specific molecular
mechanisms that give rise to anomalous diffusion in
vivo remain elusive. This is most clearly highlighted
by Saxton et al., who published a call for 'a positive
control for anomalous diffusion' as a solution to this
problem [10]. This positive control would be a sim-
ple and reproducible experimental model exhibiting
'readily tuneable' anomalous diffusion spanning sev-
eral orders of magnitude in timescale. Here we seek to
address this call by engineering a simple experimental
model in which it is possible to select the anomalous
behavior. We take advantage of previous work on the
disruption of SLB formation by PEG-DPPE [32] to
control nanoscale obstacle formation in a bilayer (Fig.
1A). By varying the PEG-DPPE composition in a bi-
layer, we expect that an increased fraction of polymer
in the brush regime will result in the formation of
specific defects in the bilayer, similar to interfacial or
grain boundary defects caused by phases separating
mixtures [33]. Similar defects have also been reported
using Atomic Force Microscopy (AFM) of incomplete
SLB formation from small unilamellar vesicles (SUVs)
[34] as well as in SLBs formed in the presence of mem-
brane active peptides [35].
arXiv:1709.04698v2 [physics.bio-ph] 24 Oct 2017
2
τ can be interpreted in terms of a length scale (λ)
associated with the anomalous behavior in 2D (λ =
√4Dτ ).
III. RESULTS
A. Supported Lipid Bilayers
We produced SUVs from DOPC doped with PEG-
DPPE (0 - 10 mol% PEG-DPPE; 1,2, & 5 kDa PEG-).
Fusion of these SUVs onto a glass coverslip created a
SLB. We confirmed the physical nature of the bilay-
ers using AFM (Fig. 2A): As PEG-DPPE content
increases, small defects appear in the bilayer. Further
increase in the concentration of PEG-DPPE results
in the extension of the interfacial defects, until the
system crosses the percolation threshold. This leads
to confined bilayer patches.
Image binarization and
autocorrelation were used to calculate the excluded
area fraction and the length scale associated with the
defects.
creased from 0 to 6 mol%,
log(cid:0)h∆r2i/4∆t(cid:1) vs.
The diffusive properties of these bilayers were as-
sessed using single-particle tracking:
smTIRF mi-
croscopy was used to follow Texas Red-labelled lipids
(∼ 10−6 mol% TR-DHPE) at 200 Hz;
iSCAT
tracked 40 nm antibiotin-conjugated gold nanoparti-
cles (AuNPs) tethered to biotinylated lipids at 5 kHz.
As the concentration of PEG(2K)-DPPE was in-
the gradient of the
log(∆t) plot deviates from zero
(Fig. 2B). Figure 2C shows the equivalent PEG(2K)-
DPPE dataset arising from iSCAT. Similar plots were
produced for PEG(1K)- and PEG(5K)-DPPE (Fig.
S1). Gradients extracted from these plots (α − 1) al-
low calculation of the anomalous exponent, while the
y-axis intercept reports the transport coefficient. The
values of α for all three PEG molecular weights are
collated in Figure 2D; α transitions from 1 (free dif-
fusion) to 0 (confined diffusion). These data were fit
empirically by a simple sigmoid. The midpoint of each
sigmoid is a measure of the transition between contin-
uous and discontinuous diffusion. Both smTIRF and
iSCAT measurements give rise to the same trend (Ta-
ble S1). Figure 2F shows the equivalent variation of
Γ with mol% PEG-DPPE. Again, in agreement with
Figure 2D, the apparent diffusion coefficient slows as
the particles become confined. It is worth emphasiz-
ing that as Γ scales with α, only points with the same
α values can be compared directly. Limiting values
for Γ differ as expected between smTIRF and iSCAT
experiments due to the size difference of fluorescently
labelled lipids and AuNPs [40].
Using our AFM calibration (Fig. S2), we are able
to convert mol% to excluded area fraction (Fig. 2E).
For both α and Γ, the sigmoids for the three differ-
ent PEG molecular weights now overlap, showing the
same trend with excluded area fraction. For α vs. ex-
cluded area fraction a single sigmoid fit yields a mid-
point at α = 0.232 ± 0.001.
FIG. 1. PEG bilayer model.
(A) Schematic of sup-
ported lipid bilayers. As the mole fraction of PEGylated
lipids increases (left to right), defects form in the bilayer
that that act as obstacles, generating anomalous diffusion.
Representative single-particle tracking of smTIRF (B) and
iSCAT (C) images (scale 10 µm and 1 µm, respectively).
The complex nanometer-scale confinement reported
in cell membranes gives rise to anomalous behav-
ior that spans from microseconds to seconds. Thus
to properly characterize anomalous diffusion,
it is
important to apply techniques capable of studying
these timescales. Here we exploit a combination
of single-molecule Total Internal Reflection Fluores-
cence (smTIRF) [36] and Interferometric Scattering
(iSCAT) microscopies [37, 38] (Figs. 1B and C, Sup-
plementary Methods) to characterize diffusion using
single-particle tracking that spans over four orders of
magnitude in time.
II. THEORY
Anomalous diffusion describes random molecular
motion that does not display a linear scaling of the
second moment with time. The most common model
for anomalous diffusion is to allow the second moment
to scale as a power of time [5, 39],
(cid:10)∆r2(cid:11) = 4Γ∆tα,
(1)
where α is the anomalous exponent and D is replaced
by Γ, the anomalous transport coefficient. Anomalous
sub- and super-diffusion are defined by α < 1 and α >
1. Given the form of equation 1, α can be determined
from the gradient of a logarithmic plot of(cid:10)∆r2(cid:11) /4∆t
vs. ∆t.
The transport coefficient Γ is somewhat more diffi-
cult to interpret as it has dimension of [L]2/[T]α, thus
its dimensions are changing for different degrees of
anomalous behavior. This apparent problem can be
overcome by de-dimensionalizing the observation time
[5] using a 'jump time', τ :
(cid:10)∆r2(cid:11) = 4D∆t(cid:18) ∆t
τ (cid:19)α−1
.
(2)
A
B
C
3
FIG. 2. Anomalous diffusion in PEG bilayers. (A) AFM shows an increase in defect area fraction with increasing
mol% PEG-DPPE (scale 500 nm). (B) Anomalous sub-diffusion increases as amount of PEG-DPPE increases from 0
(black) to 10 (yellow) mol%, here for PEG(2K)-DPPE. (C) Equivalent iSCAT data for 0 to 2.6 mol% PEG(2K)-DPPE.
Variation of α and Γ with mol% PEG-DPPE (D&F), and excluded area fractions (E&G). PEG(1K)- (blue triangles),
PEG(2K)- (red circles) and PEG(5K)-DPPE (green squares). Error bars (grey) throughout represent standard errors.
B. Monte Carlo Simulations
IV. DISCUSSION
To help improve our understanding of these exper-
imental results we constructed a simple Monte-Carlo
simulation of anomalous diffusion: A periodic square
lattice of circular, immobile obstacles of radius R was
simulated using a unit cell with side-length s (Fig.
4A). A discrete-time random walk was subject to the
constraint that the walk cannot enter the circular ob-
stacle. As expected, anomalous behavior arises in the
simulation as the excluded area fraction was increased
(Fig. 4B), and again a plateau of normal diffusion at
short times was observed. When values are collated a
sigmoidal trend was present with α = 0 being reached
near to the percolation threshold for circular obsta-
cles on a square lattice (0.785). Γ was best fit by a
double-sigmoid (Fig. 4D, Table S2).
Combining equations 1 and 2, a linear variation of
log10 (Γ/D) with α is expected; and reproduced by our
data. The characteristic length scales (λ) calculated
from both simulation (Fig. 4E) and experiment (Fig.
4F) are summarized in Table S3.
The presence of PEG-DPPE disrupts SLB forma-
tion leaving a network of defects whose area fraction
is dependent on the concentration. We have exploited
this defect formation to create predictable and tune-
able anomalous behavior. It is bilayer continuity (not
the presence of PEG as an obstacle) that causes the
anomalous behavior. Our controls (Fig. S4) confirm
that normal diffusive behavior can be rescued by fill-
ing in bilayer defects.
The variation of α and Γ with excluded area fraction
that we observe shows a sigmoidal transition between
free and confined diffusion. This relationship can be
used to tune anomalous behavior. It has been shown
using simulations [24] and in cell membranes [15, 16,
46–48] that α values of 0.5 to 0.7 are most biologically
relevant. Using our model, we can make specific and
controlled changes to obstacle extent that match this
range; providing an opportunity to use this simple
model to help predict and study biological systems
that exhibit complex membrane diffusion in vivo.
We must also address the limitations of this model
A
10
0.0 2.0 2.6 3.0 5.0 mol%
0.0
2.0
2.6
3.0
5.0 mol%
D
E
F
G
0
Height / nm
B
C
4
FIG. 3. Monte Carlo simulations of anomalous diffusion. (A) Schematic of the unit cell. (B) Diffusion analysis
of the resultant tracks showed similar behavior to experiment (R = 500 nm, D = 3 µm2 s−1). (C&D) A similar trend
to experiment was also present for α and Γ for R = 150 nm, 500 nm and 1 µm (grey triangles, black circles, teal squares
respectively). (E) Plots of log10 (Γ/D) vs. α show the expected linear relation, dependent on obstacle size. (F) Similar
plots for our experimental data show an essentially static linear relationship for different PEG lengths. PEG(1K)- (blue
triangles), PEG(2K)- (red circles) and PEG(5K)- (green squares).
system. Confinement in cell membranes is not created
directly by membrane defects, but is likely due to the
excluded area created by membrane proteins and their
interactions with lipids. Despite these fundamental
differences, both result in a similar restriction to free
diffusion in the bilayer, and a parallel can be drawn
between the excluded area controlled in this simple
model and that inaccessible to diffusing species in cell
membranes.
Theory predicts that in a system with finite hier-
archy, the diffusion will return to normal behavior
(with a reduced diffusion coefficient) at sufficiently
long observation times [44]. Figure 2B shows that
over the time scales observed in these experiments,
the diffusion here remains anomalous. As normal be-
havior returns at around 100 ms for similarly sized
compartments in cells [15], our model must not pos-
sess the restricted range of compartment size that are
presumably present in cell membranes. However, we
predict that additional control of bilayer defect for-
mation would enable a return to normal diffusion at
these timescales, for example by nano-patterning of
the substrate before SLB formation [45].
By using two single-particle microscopy techniques
we have sampled the anomalous behavior of this model
over four orders of magnitude of time. Alone, fluo-
rescence microscopy cannot access the divergence of
the diffusivity at short times and the onset of anoma-
lous behavior. However iSCAT is not without its own
limitations:
iSCAT image analysis requires efficient
background subtraction [38, 49] which fails if parti-
cles do not move sufficiently e.g. within the confined
regime. The high frame rate of iSCAT also presents
its own challenges in data management, preventing us
from probing all the relevant timescales with a single
technique.
Our simulations helped us to understand the rela-
tionship between α and excluded area fraction.
In
the square lattice model, an excluded area fraction of
0.785 represents the point at which the obstacle di-
ameter is equal to the size of the unit cell, s. This is
therefore the point at which confinement occurs. The
mid point of the sigmoidal fit occurs at an excluded
area fraction of 0.773 and the fit effectively reaches α
= 0 at 0.808. If we consider α to sample the probabil-
ity of being confined, we can say that the mid-point
is indicative of the percolation threshold. Using this
to interpret the experimental data we find that the
percolation threshold of our model falls around an ex-
cluded area fraction of 0.232 ± 0.054.
Perhaps of greatest interest is the difference be-
tween plots of log10(D/Γ) vs. α between our simula-
tions and experiment (Table S3): Only our simulation
shows a variation in length scale with obstacle size. In
contrast, our experimental data shows a similar gra-
dient for all three molecular weights of PEG-DPPE.
This suggests that in the experiment the size (but not
number or extent) of defects produced are of a simi-
lar scale ( 150 nm) and are independent of the PEG
molecular weight. The values extracted from AFM
FWHMs are smaller by around a factor of 3. There
is also a modest negative correlation between FWHM
and excluded area fraction (see Fig. S4). The dif-
ferences between experiment and simulation are most
A
B
C
E
s
D
F
5
likely due to the different topologies for defects in the
two cases: To preserve simplicity, defects in the sim-
ulation were simple circles. This is in contrast to our
experimental AFM images (Fig. 2A) that show not
circular, but rather interfacial defects for intermedi-
ate PEG-lipid concentrations. We therefore interpret
this characteristic length scale as a descriptor of the
barrier to diffusion.
In our model, that is the scale
associated with bilayer defects, and in cells it is likely
to be associated with the cytoskeleton.
imental model, readily tuneable in anomaleity over
the length scales observed in vivo. This study also
opens the way to further experiments that exclude
membrane area using more complex methods than the
simple inclusion of PEG-lipids presented here. Future
work must be directed to expand our understanding of
cell membranes - to recreate biological pathways con-
trolled by diffusion, and to enable the rational design
of devices with tailored bilayer properties.
V. CONCLUSION
By controlling SLB formation using PEGylated
lipids we are able to produce bilayers with defined
anomalous diffusive properties, dependent on the ex-
cluded area fraction. Thus, we hope our work in part
answers the call for a simple and reproducible exper-
VI. ACKNOWLEDGEMENTS
We thank the European Research Council for pro-
viding funding for this work (ERC-2012-StG-106913
CoSMiC).
VII. REFERENCES
[1] E. D. Sheets, G. M. Lee, R. Simson, and K. Jacobson,
[18] M. R. Horton, F. Hofling, J. O. Radler, and T. Fra-
Biochemistry 36, 12449 (1997).
nosch, Soft Matter 6, 2648 (2010).
[2] D. Choquet and A. Triller, Nature Reviews Neuro-
science 4, 251 (2003).
[3] B. N. Kholodenko, Nature reviews. Molecular cell bi-
ology 7, 165 (2006).
[4] U. Cheema, Z. Rong, O. Kirresh, A. J. MacRobert,
P. Vadgama, and R. A. Brown, Journal of Tissue
Engineering and Regenerative Medicine 6, 77 (2012).
[5] M. J. Saxton, Biophysical journal 66, 394 (1994).
[6] B. M. Regner, D. Vucini´c, C. Domnisoru, T. M. Bar-
tol, M. W. Hetzer, D. M. Tartakovsky, and T. J.
Sejnowski, Biophysical Journal 104, 1652 (2013).
[7] F. Hofling and T. Franosch, Reports on Progress in
Physics 76, 046602 (2013).
[8] T. K. Fujiwara, K.
Iwasawa, Z. Kalay, T. A.
Tsunoyama, Y. Watanabe, Y. M. Umemura, H. Mu-
rakoshi, K. G. N. Suzuki, Y. L. Nemoto, N. Morone,
and A. Kusumi, Molecular biology of the cell 27, 1101
(2016).
[9] Y. Golan and E. Sherman, Nature Communications
8, 15851 (2017).
[10] M. J. Saxton, A. Philipse, and A. Philipse, Biophys-
ical journal 103, 2411 (2012).
[11] A. Kusumi, C. Nakada, K. Ritchie, K. Murase,
K. Suzuki, H. Murakoshi, R. S. Kasai, J. Kondo,
and T. Fujiwara, Annual review of biophysics and
biomolecular structure 34, 351 (2005).
[12] M. J. Saxton, Biophysical journal 52, 989 (1987).
[13] H. Berry and H. Chat´e, Physical Review E 89, 022708
(2014), arXiv:1103.2206.
[14] T. Fujiwara, K. Ritchie, H. Murakoshi, K. Jacobson,
and A. Kusumi, The Journal of cell biology 157, 1071
(2002).
[15] K. Murase, T. Fujiwara, Y. Umemura, K. Suzuki,
R. Iino, H. Yamashita, M. Saito, H. Murakoshi,
K. Ritchie, and A. Kusumi, Biophysical journal 86,
4075 (2004).
[16] G. Schutz, H. Schindler, and T. Schmidt, Biophysical
[19] K. M. Spillane, J. Ortega-Arroyo, G. de Wit,
C. Eggeling, H. Ewers, M. I. Wallace, and P. Kukura,
Nano letters 14, 5390 (2014).
[20] H.-M. Wu, Y.-H. Lin, T.-C. Yen, and C.-L. Hsieh,
Scientific reports 6, 20542 (2016).
[21] M. Rose, N. Hirmiz, J. Moran-Mirabal,
and
C. Fradin, Membranes 5, 702 (2015).
[22] M. J. Saxton, Biophysical journal 56, 615 (1989).
[23] M. J. Saxton, Biophysical journal 81, 2226 (2001).
[24] S. Stachura and G. R. Kneller, Molecular Simulation
40, 245 (2014).
[25] Y. Mardoukhi, J.-H. Jeon,
and R. Metzler, Phys.
Chem. Chem. Phys. Phys. Chem. Chem. Phys 30134,
30134 (2015).
[26] H. Koldsø, T. Reddy, P. W. Fowler, A. L. Duncan,
and M. S. P. Sansom, J. Phys. Chem. B 120, 8873
(2016).
[27] J.-H. Jeon, M. Javanainen, H. Martinez-Seara,
R. Metzler, and I. Vattulainen, Physical Review X
6, 021006 (2016).
[28] E. Bakalis, S. Hofinger, A. Venturini, and F. Zer-
betto, The Journal of Chemical Physics 142, 215102
(2015).
[29] M. Javanainen, H. Hammaren, L. Monticelli, J.-H.
Jeon, M. S. Miettinen, H. Martinez-Seara, R. Metzler,
I. Vattulainen, S. W. Hell, C. Eggeling, and S. W.
Hell, Faraday Discuss. 161, 397 (2013).
[30] E. Kepten, A. Weron, G. Sikora, K. Burnecki, and
Y. Garini, PLoS ONE 10, e0117722 (2015).
[31] R. Metzler, J.-H. Jeon, A. G. Cherstvy, E. Barkai,
I. Reich, S. Cova, L. Xun, X. S. Xie,
and N. F.
Scherer, Phys. Chem. Chem. Phys. 16, 24128 (2014).
[32] S. Kaufmann, G. Papastavrou, K. Kumar, M. Textor,
[33] D. Keller, N. B. Larsen,
and E. Reimhult, Soft Matter 5, 2804 (2009).
I. M. Møller,
and
O. G. Mouritsen, Physical Review Letters 94, 025701
(2005).
Journal 73, 1073 (1997).
[34] R. P. Richter and A. R. Brisson, Biophysical Journal
[17] T. V. Ratto and M. L. Longo, Langmuir 19, 1788
88, 3422 (2005).
(2003).
[35] V. Oliynyk, U. Kaatze, and T. Heimburg, Biochimica
et Biophysica Acta (BBA) - Biomembranes 1768, 236
6
[55] J. Y. Tinevez, N. Perry, J. Schindelin, G. M. Hoopes,
G. D. Reynolds, E. Laplantine, S. Y. Bednarek, S. L.
Shorte, and K. W. Eliceiri, Methods, Methods 115,
80 (2016).
(2007).
[36] D. Axelrod, T. P. Burghardt, and N. L. Thompson,
Annu Rev Biophys Bioeng 13, 247 (1984).
[37] K. Lindfors, T. Kalkbrenner, P. Stoller, and V. San-
doghdar, Physical Review Letters 93, 037401 (2004).
[38] J. Ortega-Arroyo and P. Kukura, Physical chemistry
chemical physics : PCCP 14, 15625 (2012).
[39] S. Havlin and D. Ben-avraham, Advances in Physics
36, 695 (1987).
[40] P. Mascalchi, E. Haanappel, K. Carayon, S. Maz`eres,
L. Salom´e, Y. Sako, S. Wieser, G. J. Schutz, F. Pin-
aud, S. Clarke, A. Sittner, M. Dahan, A. Triller,
D. Choquet, R. Mach´an, M. Hof, L. Groc, M. J.
Murcia, D. E. Minner, G.-M. Mustata, K. Ritchie,
C. A. Naumann, G. M. Lee, A. Ishihara, K. A. Ja-
cobson, M. Fein, X. Michalet, A. Lopez, L. Dupou,
A. Altibelli, J. Trotard, J. F. Tocanne, D. M.
Soumpasis, A. Serg´e, N. Bertaux, H. Rigneault,
D. Marguet, N. Meilhac, L. L. Guyader, L. Salom´e,
N. Destainville, C. Yoshina-Ishii, D. H. Murray, L. K.
Tamm, V. Kiessling, Y. Min, N. Pesika, J. Zasadzin-
ski, J. Israelachvili, L. Salom´e, J. L. Cazeils, A. Lopez,
J. F. Tocanne, L. K. Tamm, H. M. McConnell, M. L.
Wagner, L. K. Tamm, K. J. Seu, L. R. Cambrea,
R. M. Everly, J. S. Hovis, K. J. Seu, L. Guo, C. Scom-
parin, S. Lecuyer, M. Ferreira, T. Charitat, B. Tin-
land, M. J. Saxton, K. Jacobson, F. Pinaud, S. Huet,
C. Bouzigues, M. Dahan, A. Sonnleitner, G. Schutz,
T. Schmidt, H. Qian, M. P. Sheetz, E. L. Elson, S. L.
Goodman, G. M. Hodges, L. K. Trejdosiewicz, and
D. C. Livingston, Soft Matter 8, 4462 (2012).
[41] H. Bayley, B. Cronin, A. Heron, M. a. Holden, W. L.
Hwang, R. Syeda, J. Thompson, and M. Wallace,
Molecular bioSystems 4, 1191 (2008).
[42] S. Leptihn, O. K. Castell, B. Cronin, E.-H. Lee,
L. C. M. Gross, D. P. Marshall, J. R. Thompson,
M. Holden, and M. I. Wallace, Nature protocols 8,
1048 (2013).
[43] J. T. Sengel and M. I. Wallace, Proceedings of the
National Academy of Sciences 113, 5281 (2016).
[44] M. J. Saxton, Biophysical journal 92, 1178 (2007).
[45] J. Tsai, E. Sun, Y. Gao, J. C. Hone, and L. C. Kam,
Nano Letters 8, 425 (2008).
[46] P. Schwille, J. Korlach, and W. W. Webb, Cytometry
36, 176 (1999).
[47] T. J. Feder, I. Brust-Mascher, J. P. Slattery, B. Baird,
journal 70, 2767
and W. W. Webb, Biophysical
(1996).
[48] P. R. Smith,
I. E. Morrison, K. M. Wilson,
N. Fern´andez, and R. J. Cherry, Biophysical Jour-
nal 76, 3331 (1999).
[49] P. Kukura, H. Ewers, C. Muller, A. Renn, A. He-
lenius, and V. Sandoghdar, Nature methods 6, 923
(2009).
[50] A. A. Brian and H. M. McConnell, Proceedings of the
National Academy of Sciences of the United States of
America 81, 6159 (1984).
[51] J. Schindelin,
I. Arganda-Carreras, E. Frise,
V. Kaynig, M. Longair, T. Pietzsch, S. Preibisch,
C. Rueden, S. Saalfeld, B. Schmid, J.-Y. Tinevez,
D. J. White, V. Hartenstein, K. Eliceiri, P. Tomancak,
and A. Cardona, Nature Methods 9, 676 (2012).
[52] D. Axelrod, D. E. Koppel, J. Schlessinger, E. El-
son, and W. W. Webb, Biophysical journal 16, 1055
(1976).
[53] D. Soumpasis, Biophysical Journal 41, 95 (1983).
[54] J. Ortega Arroyo, D. Cole, and P. Kukura, Nature
protocols 11, 617 (2016).
Controlling Anomalous Diffusion in Lipid Membranes
Supplementary Information
H. L. E. Coker* and M. R. Cheetham*
*Contributed equally to this work.
Department of Chemistry, Britannia House,
King's College London, London SE1 1DB and
Chemistry Research Laboratory, 12 Mansfield Road,
University of Oxford, Oxford OX1 3TA
D. R. Kattnig
Living Systems Institute & Department of Physics,
University of Exeter, Stocker Road, Exeter EX4 4QD
Y. J. Wang and S. Garcia-Manyes
Department of Physics, Strand Building,
King's College London, London WC2R 2LS
M. I. Wallace
Department of Chemistry, Britannia House,
King's College London, London SE1 1DB
(Dated: October 25, 2017)
arXiv:1709.04698v2 [physics.bio-ph] 24 Oct 2017
1
I. SUPPLEMENTARY FIGURES
FIG. S1. (cid:10)∆r2(cid:11) /4∆t vs. ∆t Plots. PEG(1K)-DPPE; (A) fluorescence and (B) iSCAT and PEG(5K)-
DPPE; (C) fluorescence and (D) iSCAT to accompany the PEG(2K)-DPPE data in Fig. 2B/C. The
same behaviour is seen for all three PEG molecular weights. Along with Fig. 2B/C the data from
these plots are used to extract Γ) and α as outlined in the Results section of the main text.
2
A
C
B
D
FIG. S2. Conversion to excluded area fraction. Image auto-thresholding (see Supplementary
Methods) was used to get an excluded area fraction from the AFM images for several concentrations
of PEG-DPPE. Plotted against mol% PEG-DPPE and fitted with sigmoidal functions, we get a
calibration curve by which to covert mol% PEG-DPPE to excluded area fraction.
FIG. S3. Extended plot of (A) excluded area fraction vs. α and (B) excluded area fraction
vs. Γ. Fig. 2E & 2G show the same data for 0 6 α 6 0.6 to highlight the overlap of sigmoids for the
three datasets. Here, all data is shown (0 6 α 6 1). PEG(1K)-, PEG(2K)- and PEG(5K)-DPPE are
represented in blue triangles, red circles and green squares respectively.
3
A
B
FIG. S4. Anomalous behaviour is a result of confinement in bilayer patches. (A) single-
particle tracking of anomalous diffusion was repeated on a SLB exhibiting anomalous behaviour
(DOPC; 2.6% PEG(2K)-DPPE). This bilayer was then incubated with DOPC SUVs. The additional
vesicles ruptured within the defects, repaired the bilayer and returned normal diffusive behaviour.
(B) FRAP shows similar behaviour, with 2.6% PEG(2K)-DPPE bilayers exhibiting a large slow and
immobile fraction. Upon addition of DOPC SUVs, diffusion again recovers (D = 1.75 µm2 s−1, mo-
bile fraction 0.99) to values comparable to a pure DOPC bilayer (D = 1.81 µm2 s−1, mobile fraction
1.00).(C) Droplet interface bilayers (DIBs) [? ? ] (Fig. S5) were used to create unsupported lipid
bilayers and confirm PEG-lipids alone do not result in anomalous behaviour. In DIBs, defects would
result in conductance across the bilayer, unstable droplets and ultimately bilayer rupture [? ]. Fol-
lowing our previous methods [? ], 'lipid-in' DIBs (2.6 or 5.0 mol% PEG(2K)-DPPE; ∼ 10−6 mol%
TR-DHPE) were formed. Single-particle tracking using TIRF microscopy showed normal diffusion
even for 5.0 mol% PEG-DPPE, well above the threshold value for anomalous behaviour in SLBs.
Bilayers containing 0 (black), 2.6 (red) and 5.0 (yellow) mol% PEG(2K)-DPPE all exhibit normal
diffusion. In Droplet Interface Bilayers (DIBs),
4
+ SUVs
+ SUVs
C
+ SUVs
0.0 mol%
2.6 mol%
5.0 mol%
A
B
+ SUVs
B
C
FIG. S5. Length scales extracted from autocorrelation analysis of the AFM images. There
is a modest decrease in length scale as excluded areas fraction increases. Black data point is DOPC
only SLB. PEG(1K)-, PEG(2K)- and PEG(5K)-DPPE are represented in blue triangles, red circles
and green squares respectively.
FIG. S6. Schematic of a droplet interface bilayer (DIB). DIBs were prepared following our
previous protocol (Leptihn et al. Nature Protocols 8, 1048, 2013).
5
DPhPC
PEG-DPPE
Hexadecane
PMMA device
SUVs in buffer
Agarose
Coverslip
II. SUPPLEMENTARY TABLES
TABLE S1. Experiment sigmoid fit midpoints. Midpoints extracted from the sigmoid fit of the
α vs. mol% PEG-DPPE plots, Fig. 2D, for both fluorescence and iSCAT data.
In each case the
midpoint agrees within 6% and can be regarded as an indicator of the percolation threshold.
PEG / kDa
mol% (fluorescence)
mol% (iSCAT)
1
2
5
4.556 ± 0.004
2.700 ± 0.002
1.745 ± 0.001
4.785 ± 0.053
2.545 ± 0.019
1.745 ± 0.005
TABLE S2. Simulation sigmoid fit midpoints. Midpoints extracted from the sigmoid fit of the
α vs. EAF (Fig. 4C) and Γ vs. EAF (Fig. 4D) plots. Figure 4C was fitted with a single sigmoid,
whereas Figure 4D was better fit by a double sigmoid.
α (Fig. 4C)
Γ/ µm2 s−α (Fig. 4D)
EAF 1
-
0.277 ± 0.001
EAF 2
0.773 ± 0.002
0.737 ± 0.001
TABLE S3. Anomalous length scales (λ) for experiment and simulation.
Simulation
R /nm
λ /nm
Experiment
PEG /kDa
λ /nm
AFM FWHM /nm
150
65 ± 3
1
196 ± 4
54 ± 8
500
177 ± 7
2
145 ± 2
56 ± 2
1000
479 ± 24
5
160 ± 3
62 ± 8
6
III. SUPPLEMENTARY METHODS
A. Materials
1,2-dioleoyl-sn-glycero-3-phosphocholine (DOPC), 1,2-diphytanoyl-sn-glycero-3-phosphocholine
(DPhPC) and 1,2-dioleoyl-sn-glycero-3-phosphoethanolamine-N-cap-biotinyl (biotin-DOPE)
were purchased from Avanti Polar Lipids (Alabaster, AL). Texas Red 1,2-dihexadecanoyl-
sn-glycero-3-phosphoethanolamine (TR-DHPE) triethylammonium salt and 1,2-dipalmitoyl-
sn-glycero-3-phosphoethanolamine-N-[methoxy(polyethylene glycol) - 1000/2000/5000] am-
monium salt (PEG1/2/5K-DPPE) were purchased from Invitrogen (Eugene, OR). Goat
antibiotin-conjugated 20 nm gold nanoparticles were purchased from BBI (Cardiff, UK). Stan-
dard phosphate-buffered saline tablets (P4417), hexadecane (296317) and silicone oil AR20
(10836) were purchased from Sigma-Aldrich (St-Louis, MO). All chemicals and solvents were
of analytical grade and used without further purification. 18.2 MΩ cm water (Merck Millipore,
Billerica, MA) was used to prepare all solutions.
B. Supported Lipid Bilayers
Glass coverslips (VWR, Menzel Glaser #1) were rigorously cleaned using stepwise sonication
with detergent (Decon-90), water, and propan-2-ol. Immediately before use, the glass was dried
under nitrogen, cleaned with oxygen plasma for 3 minutes (Diener Electronic, Femto), and
stored under water before use. A well was created on each coverslip using vacuum grease (Dow
Corning high vacuum grease). 50 µL of SUV stock was diluted 1:1 in buffer (PBS pH 7.4)
and added to the chamber immediately. The vesicles were incubated for 60 minutes before the
membranes were washed with de-gassed Milli-Q.
SLBs were prepared on glass coverslips by vesicle fusion [? ]. Lipid mixtures (1 mg mL−1
DOPC; ∼ 10−6 mol% each TR-DHPE and biotin-DOPE) were doped with varying amounts of
PEG-lipids to span a range of 0 to 10 mol%. Lipid mixtures were dried under nitrogen and
placed under vacuum overnight. The dried lipids were hydrated in water and votexed before tip
sonication for 15 minutes. The resulting clear, vesicle suspension was centrifuged (3 minutes;
14000g) before the supernatant was separated and retained.
C. Droplet Interface Bilayers
DIBs were prepared following our previous protocol [? ]. Briefly, SUVs of DPhPC only
and DPhPC + PEG(2K)-DPPE were prepared as described above and diluted with buffer to
7
a final concentration of 0.5 mg ml−1. 100 nl droplets of SUV solution were pipetted into a
microfabricated tank containing hexadecane and incubated for 40 min. DPhPC (8.9 mg ml−1
in 9:1 hexadecane:silicone oil) was prepared from chloroform stock. A microfabricated PMMA
device containing the agarose layer on a glass coverslip was incubated with the DPhPC in oil
solution for 20 min. Incubated droplets were then added to the device and left for 10 min to
allow bilayer formation before imaging. The DIB experiment is represented schematically in
Figure S6.
D. Atomic Force Microscopy
AFM was carried out using a Bioscope (Bruker, MA). An SNL-10B cantilever was used with
a spring constant of 0.12 N m−1. The instrument was run in peak-force tapping mode with
a maximum applied force of 40 pN. Auto-thresholding was carried out on exported TIFs in
FIJI [? ]. Auto-correlation of the images was performed using the FD Math function. The
resultant plot was then radially averaged using a radial profile. The resultant plot was fitted
with a Lorentzian function to extract the FWHM.
E. Fluorescence Recovery After Photobleaching
Fluorescence recovery after photobleaching (FRAP) experiments were carried out as de-
scribed by Axelrod et al. [? ]. Briefly, an DOPC SLB was made with 0.5 mol% TR-DHPE,
and subjected to intense laser light through a small iris for 5 s. The iris was then opened,
the illumination intensity reduced, and the bilayer was imaged every 5 s for the following 20
min. The result was a video showing the fluorescence slowly recovering in the region where the
intense laser light had caused strong photobleaching. Image analysis of this recovery process
yielded the diffusivity of the TR-DHPE in the SLB. This was quantified by fitting with a mod-
ified Bessel function as outlined by Soumpasis [? ]. This fitting method is only applicable to
bilayers exhibiting normal diffusion and as such, the bilayer containing PEG(2K)-DPPE at 2.6
mol% was not quantified.
F. Total Internal Reflection Fluorescence Microscopy
A 532 nm continuous-wave laser was launched into an inverted microscope (Eclipse TiE;
Nikon) and focussed at the back aperture of the objective lens (60× TIRF oil-immersion NA
1.49, Nikon, ∼1.4 kW cm−2 at the sample) such that total internal reflection occurred. The re-
sultant fluorescence signal was transmitted through dichroic (ZT532rdc) and bandpass (605/55)
8
filters (both Chroma, Bellows Falls, VT) before being imaged onto an electron-multiplying CCD
(iXon+ 860; Andor) at 200 Hz.
G.
Interferometric Scattering Microscopy
A custom-built microscope was built to conduct these measurements as described [? ? ].
Briefly, a 639 nm laser beam (Toptica, Graefelfing, Munich) was directed through a polarizing
beam splitter and quarter wave plate (both Thorlabs, Newton, NJ) before reaching a focus at the
back aperture of an objective lens (100× oil immersion NA 1.49 Nikon). Light scattered from the
coverslip interface and the object of interest, produce an interference pattern, which is returned
through the beam splitter. The resultant signal was then magnified (overall magnification
174×) and imaged onto a CMOS camera (Phantom Miro 340). To track AuNP-labelled lipids,
a SLB was formed as described above and 5 µL of anti-biotin OD10 40 nm gold nanoparticles
were added. After 2 hours incubation, the SLB was washed thoroughly with buffer to remove
unbound nanoparticles before imaging. Image stacks of 10000 frames were recorded at 100 kHz.
Image stacks were then temporally-averaged to 5 kHz before tracking tracking.
H. Single-Particle Tracking
The movement of fluorescent lipids between frames of the recorded videos was tracked using
the TrackMate plugin for ImageJ [? ]. Briefly, spots were detected in each frame using a
Laplacian of Gaussian (LoG) filter. Tracks were then generated by linking these detected spots
together using a simple linear assignment problem (LAP) tracker. The output from this was a
collection of tracks containing the space-time co-ordinates of each point in the track. This data
was subsequently used to obtain mean-squared displacements for different observation times,
which in turn was used to obtain diffusivity values via the random walk model of diffusion.
I. Monte Carlo Simulations
For each of 200 simulated particles, two separate random walks were created (corresponding
to the displacement in x and y coordinates) using a pseudorandom number generator (Mersenne
Twister).
In order to account for the presence of obstacles to diffusion, periodic boundary
conditions were used. The unit cell of the simulation is shown in Figure 4A of the main
text. With each step of the random walk, the new coordinates were tested to see if they were
inside an obstacle or not. If they were, then that step was rejected, and a new one generated.
The resulting sets of co-ordinates for each simulated particle were subsequently processed in
9
the same way as experimental data obtained from tracking labelled particles. These tracks
represent diffusing particles that have been hindered due to the presence of obstacles, and
exhibited anomalous diffusion as expected.
10
|
1503.03344 | 1 | 1503 | 2015-03-11T14:12:36 | Flexibility of short DNA helices with finite-length effect: from base pairs to tens of base pairs | [
"physics.bio-ph",
"cond-mat.soft",
"cond-mat.stat-mech",
"physics.comp-ph",
"physics.med-ph"
] | Flexibility of short DNA helices is important for the biological functions such as nucleosome formation and DNA-protein recognition. Recent experiments suggest that short DNAs of tens of base pairs (bps) may have apparently higher flexibility than those of kilo bps, while there is still the debate on such high flexibility. In the present work, we have studied the flexibility of short DNAs with finite-length of 5 to 50 bps by the all-atomistic molecular dynamics simulations and Monte Carlo simulations with the worm-like chain model. Our microscopic analyses reveal that short DNAs have apparently high flexibility which is attributed to the significantly strong bending and stretching flexibilities of ~6 bps at each helix end. Correspondingly, the apparent persistence length lp of short DNAs increases gradually from ~29nm to ~45nm as DNA length increases from 10 to 50 bps, in accordance with the available experimental data. Our further analyses show that the short DNAs with excluding ~6 bps at each helix end have the similar flexibility with those of kilo bps and can be described by the worm-like chain model with lp~50nm. | physics.bio-ph | physics | Flexibility of short DNA helices with finite-length effect:
from base pairs to tens of base pairs
Yuan-Yan Wu, Lei Bao, Xi Zhang and Zhi-Jie Tan*
Department of Physics and Key Laboratory of Artificial Micro & Nano-structures of Ministry of
Education, School of Physics and Technology, Wuhan University, Wuhan 430072, China
ABSTRACT
Flexibility of short DNA helices is important for the biological functions such as nucleosome
formation and DNA-protein recognition. Recent experiments suggest that short DNAs of tens of base
pairs (bps) may have apparently higher flexibility than those of kilo bps, while there is still the
debate on such high flexibility. In the present work, we have studied the flexibility of short DNAs
with finite-length of 5 to 50 bps by the all-atomistic molecular dynamics simulations and Monte
Carlo simulations with the worm-like chain model. Our microscopic analyses reveal that short DNAs
have apparently high flexibility which is attributed to the significantly strong bending and stretching
flexibilities of ~6 bps at each helix end. Correspondingly, the apparent persistence length lp of short
DNAs increases gradually from ~29nm to ~45nm as DNA length increases from 10 to 50 bps, in
accordance with the available experimental data. Our further analyses show that the short DNAs with
excluding ~6 bps at each helix end have the similar flexibility with those of kilo bps and can be
described by the worm-like chain model with lp~50nm.
Keywords: short DNA, flexibility, finite-length effect, stretching, bending, persistence length
*
To whom correspondence should be addressed: [email protected]
I. INTRODUCTION
The structural flexibility and dynamics of DNAs play important roles in their biological
functions and are of fundamental importance for understanding the functions of DNAs.1,2 In the past
two decades, various advanced experimental methods have been developed to probe the structural
and dynamical properties of biomolecules, and the elastic properties of long DNAs of kilo base pairs
(bps) have been investigated extensively.1-9 The existing experimental measurements and theoretical
modelling indicate that a long DNA can be well described by the worm-like chain (WLC) model
with a persistence length of lp~50nm at physiological conditions.1-14
DNAs are often functional at the length of tens of base pairs,15-23 such as the formation of
nucleosome and DNA-protein recognition.15-23 Recently, a series of experiments suggest that the
flexibility of short DNAs may appear very differently from those in kilo bps.24-28 The pioneering
experiment on the cyclization of short DNAs of ~100 bps by Cloutier and Widom indicated that the
cyclization occurs ~1000 times more efficiently than the prediction from the WLC model.24 The
atomic force microscopy (AFM) experiment by Wiggins et al showed that the bending of DNAs at
short length scale is larger than that described by the WLC model for DNAs at long length scale.25
Another experiment with fluorescence resonance energy transfer (FRET) and small angle x-ray
scattering (SAXS) by Yuan et al also suggested the higher flexibility of short DNAs of 15-89 bps
which is beyond the description of the conventional WLC model.26 The recent SAXS experiment on
short DNAs of ≤35 bps with two linked gold nanocrystals by Mathew-Fenn et al has suggested that
short DNAs are at least one order of magnitude more extensible than long DNAs of kilo bps revealed
by the previous single molecule stretching experiments.27 The very recent experiment on the
cyclization of short DNAs of 67-106 bps by Vafabakhsh and Ha showed that the looping rate of short
DNAs could not be well described by the WLC model.28 To understand such high flexibility of short
DNAs, some possible mechanisms have been proposed based on the experimental findings.25-27 Yuan
et al modified the WLC model by incorporating base-pair-level fluctuation and found that the
possible base-pair flip out could lead to the high flexibility of short DNAs.26 Wiggins et al suggested
that spontaneous large-angle bends may be responsible for the high flexibility at short length scale.25
Mathew-Fenn et al proposed a mechanism of the long-ranged stretching cooperation over two
turns.27
However, another series of experiments and analyses show the opposite conclusion. Mastroianni
et al have studied the flexibility of short DNAs with finite-length of 42-92 bps with two linked gold
nanocrystals by SAXS and the WLC model, and concluded that the short DNAs of 42-92 bps can be
described by the WLC model with persistence length of ~50nm.29 Recently, Vologodskii and
Frank-Kamenetskii30,31 analyzed the experiment by Cloutier and Widom24 and that by Vafabakhsh
and Ha28. Their analyses showed that, for the former experiment,24 the very high ligase concentration
used in the experiment was considered to be responsible for the discrepancy with the WLC
model,30,31 and for the latter experiment,28 the possible sequence error in synthetic oligos and long
sticky ends very likely resulted in the significant difference from the WLC model30,31. Very recently,
Mazur and Maaloum32,33 employed the AFM-in-solution with high accuracy to re-examine the
2
experiment of Wiggins et al25, and showed that the flexibility of DNAs at length scale ≳ two turns
can be well described by the WLC model.32,33
Up to now, the previous experiments and analyses might have already led to the conclusion that
the flexibility of a short DNA at length scale approximately longer than two helical turn can be
described by the WLC model with persistence length of ~50nm.2-5,13,32,33 However, for the global
flexibility of short DNAs with tens of bps,24-31 there is still lack of consensus. First, the recent
analyses did not agree with the extremely high flexibility of short DNAs with length from ~60 to
~100 bps observed in the cyclization experiments.24,28,30,31 Second, the SAXS/FRET experiments
showed apparently higher flexibility for short DNAs with length ≤35 bps, 27 while other SAXS
experiments indicated that the flexibility of short DNAs of 42-92 bps can be described with the WLC
with persistence length of ~50nm.29 Therefore, there is still the controversy on the global flexibility
of short DNAs with tens of bps.24-31 To understand the elusive controversy and the relevant
mechanism, we will focus on the flexibility of short DNAs with length ≤50 bps.
Due to the small size of short DNA, it is rather difficult to accurately characterize their structure
flexibility by experiments, since certain inaccuracy would be generally involved via various
experimental techniques24-33 and may become nonnegligible for small-sized DNAs. For example, the
SAXS methods generally involve two linked gold nanocrystals which can be comparable to short
DNAs in size,26,27,29 and the AFM methods generally involve DNAs of kilo bps attached on a
substrate which would generally ignore the finite-length effect in DNA flexibility.25,32,33 To
circumvent the difficulty in studying short DNAs, the all-atomistic molecular dynamics (MD)
simulation could be a useful tool to probe and analyze the flexibility of short DNAs and RNAs.35-40
Recently, Noy and Golestanian have employed 130ns MD simulations with SPC/E water model to
investigate length-scale dependence of DNA mechanical properties, where they adopted two DNAs
of 56 and 36 bps.35 However, they did not obtain length-dependent flexibility of short DNAs since
the segments at helix ends were ignored in the data analysis and only two short DNAs were used in
the study.35-37 Therefore, it is still necessarily required to study the flexibility of short DNAs with
tens of base pairs.24-31
To study the length-dependent flexibility of short DNAs, we employ the all-atomistic molecular
dynamics (MD) simulations and Monte Carlo (MC) simulations with the WLC model, to
systematically investigate the flexibility of 8 short DNAs of 5 to 50 bps in 1M NaCl solutions. The
finite-length effect will be taken into account in the present work, since such effect was naturally
included in the relevant experiments24,26-29 and was generally ignored in previous MD and AFM
studies.32,33-37 Firstly, we will analyze the global flexibility of short DNAs with various lengths
through calculating the distributions of contour length and end-to-end distance. Afterwards, we will
analyze the detailed structure flexibility at base-pair level by calculating helical rise and bending
angle, helical twist for short DNAs of different lengths. Additionally, we will calculate the apparent
persistence length for various short DNAs and make comparisons with the existing studies. Finally,
we will quantitatively discuss the finite-length effect in the flexibility of short DNAs of different
lengths. Beyond the recent studies,35-38 the present work will be focused on the flexibility of short
DNAs of finite length rather than the local flexibility of a DNA at short length-scale, and will include
3
8 short DNAs with the wide length range from 5 to 50 bps to obtain the length-dependent flexibility
of short DNAs.
II. MODEL AND METHOD
A. All-atomistic molecular dynamics and data analysis
The short DNAs of finite-length used in the study are of 5, 10, 15, 20, 25, 30, 40 and 50 bps
which are displayed in Fig. S1 of supplementary material.41 The sequences of the short DNAs are
selected according to the sequences in recent experiments to yield normal B-form DNA helices,27
and contain all the dinucleotide base pairs42,43; see Table 1. All the DNA strands are perfectly
complementary.
The initial structures of short DNA helices were taken as the standard B-DNA fibers which were
immersed in rectangle boxes containing explicit water and ions. The rectangle boxes for the short
DNAs of different lengths are different and listed in Table 1, as well as the melting temperature
calculated from the nearest neighbor model.44-46 It is noted that the melting temperatures of the short
DNAs are significantly higher than room temperature (298K) and consequently the short DNAs
would generally keep their rather stable helix structures. The counterions of Na+ and 1M NaCl salt
ions were added with Amber LEaP Program47,48 to ensure that the negatively charged backbone of
DNA is nearly full-neutralized.45,46 The efficient Particle-Mesh-Ewald method was employed for
treating the electrostatic interactions, and the cut-off length of 1.4nm was used for treating
long-range interactions.49 The periodic boundary condition was also employed in all the MD
simulations.49 In each MD simulation, a DNA was initially placed in the center of simulation cell and
its axis was initially kept in parallel with z-axis. The distance between DNA and box edges was
initially kept larger than at least 1nm in z-axis. In the simulations, we used the Amber parmbsc0
force field and the TIP3P water model combined with parmbsc0 ion model,50-52 since the force field
has been shown to give good description for various nucleic acids35-40,53,54 and the usage of TIP3P
water model has been previously shown to be more efficient in convergence than the SPC/E water
model35-37.
All systems were optimized, thermalized and equilibrated with Gromacs 4.549,55 as follows.
Firstly, an energy minimization of 5,000 steps was performed with the steepest descent algorithm at
low temperature, and then the systems were slowly heated to 298K and equilibrated with the
Nose-Hoover temperature coupling until 0.5 ns. Afterwards, the subsequent NPT simulations of 20
ns (time-step 1fs, P=1atm) were performed with the Parrinello-Rahman pressure coupling and with
the short DNAs released. Finally, the simulations of the systems were continued for another 200 ns in
the isothermic-isobaric ensemble (P=1atm, T=298K).56 A time step of 1fs was used in the
conjunction with a Leap-frog algorithm,57 for capturing the detailed dynamic motion of atoms. For
each short DNA, three independent all-atomic MD simulations were performed. To test the
convergence of the MD simulations, we have examined the instantaneous value of end-to-end
distance versus MD time for two independent simulations until 300ns for the 50-bp DNA helix, as
shown in Fig. 1 & Fig. S2 in supplementary material.41 We found that the MD simulations are nearly
4
converged after ~100ns for the 50-bp DNA, since the relative difference between the mean values of
end-to-end distance in the time ranges of [100ns-200ns] and [100ns-300ns] is only ~0.1%. We have
examined the MD trajectories and found that the short DNAs (e.g. 50-bp DNA) mainly exhibit the
conformational fluctuations with apparent bending and moderate rotation rather than sharply rotate to
the direction of short box boundary; see the movies in supplementary material.41 Furthermore, for all
the MD trajectories used here, we have calculated the minimum distances between the two end
segments of short DNAs and found that they are always larger than the cut-off length of 1.4nm, as
shown in Fig. S2 in supplementary material.41 In addition, we have examined the effect of simulation
box size by performing an additional simulation for a larger box, and the box-size effect appears
negligible, as shown in supplementary material.41
In analyzing the conformational change of short DNAs at base-pair level, we first located a
local coordinate system with an orthonormal basis.58 Afterwards, the central axes of DNA helices
were derived with the use of the program Curves+,59 and then were used for analyzing the
length-dependent structural properties of short DNAs, such as contour length, end-to-end distance,
helical rise, helical radius, helical twist, bending and persistence length. In the simulations, due to the
very high melting temperature of the used short DNAs (see Table 1), we did not observe the opening
of the inner base pairs with all (two for A·T and three for C·G1) hydrogen bond (H-bond) opening
while the terminal base pair at each helix end could occasionally become open. Such few
conformations with the opening of two terminal base pairs have been ignored in the data analysis
with the program Curves+.59
B. Monte Carlo with worm-like chain model
In addition to the above described all-atomistic MD simulations, we have performed the MC
simulations with the WLC model. In the model, a N-bp short DNA helix is modeled as a simplified
linear chain of N sequential beads, and each bead represents one base pair. The energy U of the chain
is composed of two contributions: the bond-angle energy and bond-length energy which describe the
bending and stretching rigidities, respectively. Following the previous studies,29,60 the bond-angle
energy at a bead can be given by29,60,61
,
(1)
where lp is the persistence length of the chain. l0 is the mean bond length between two adjacent beads,
and θ is the local bending angle between the adjacent unit vectors. Our calculations showed that the
form of Eq. 1 gives the similar results with a harmonic form62 (data not shown). The bond-length
energy for a certain bond length l can be expressed as29,60-62
,
(2)
where k is the stretching modulus of the chain.
Based on the above bond-angle and bond-length energies, the MC method was employed to
sample the WLC model of short DNAs with pivot move algorithm which has been shown to be
rather efficient in sampling the conformations of a polymer.63-65 The conformation sampling of the
5
)cos1(0llTkUpBb200)(21lllkUs
bead chain was performed according to the standard Metropolis algorithm.64,65 In each MC
simulation, the WLC model could get equilibrated within 104 steps, and the following 2×107 steps
were employed to calculate the averaged properties of the WLC model. The efficient pivot move and
enough equilibrium MC steps would ensure the calculations of equilibrium properties for the WLC
model.
III. RESULTS AND DISCUSSION
In this section, we will analyze the flexibility of short DNAs with finite-length effect based on
the all-atomistic MD and MC simulations with the WLC model, and 8 short DNAs of 5-50 bps will
be covered. Firstly, we will focus on the global conformation fluctuation of short DNAs and
afterwards, we will make detailed analysis on length-dependent properties of helical structure at
base-pair level. Additionally, we will calculate the apparent persistence length of various short DNAs,
and make extensive comparisons with the previous studies. Finally, we will quantitatively discuss the
finite-length effect in DNA flexibility.
A. Global structural flexibility
The global size of a short DNA can be characterized by its contour length L and end-to-end
distance Ree. As shown in Fig. 1, to include the finite-length effect, the contour length L is taken as
the summation over all base pair steps and the end-to-end distance Ree is taken as the distance
between the centers of two terminal base pairs. The variance
of contour length of a short DNA
is expressed by
, and the variance
of Ree is given by
.
As shown in Fig. 2a, the MD trajectories for short DNAs of various lengths give the
distributions of contour length L and end-to-end distance Ree with approximately symmetric
fluctuations around their mean distance. It is noted here that the distributions of Ree slightly deviate
from the normal distribution for the 40-bp and 50-bp DNAs, which is in accordance with the
previous WLC analysis for semiflexible polymers with L<lp.66,67 We have examined the possible
statistical error and found that the relative difference between the maximum value and mean value of
Ree is small (<1% for the 50-bp DNA). Due to the small deviation and the wide usage in previous
related analysis,26,27,29 we still used the mean value and its variance to make analysis for the
flexibility of the 40-bp and 50-bp DNAs. As shown in Fig. 2b, the mean contour length
and
mean end-to-end distance
increase approximately linearly with DNA length. The slope for
with DNA length gives an average rise of ~3.32Å per base pair step, in agreement with the
crystallographic value of 3.32±0.19Å,68 and the slope of
value of ~3.22Å per base pair which is also close to the value of 3.27±0.1Å in the recent SAXS
experiment27. The structural fluctuations of a short DNA can be quantified by the variances
with DNA length gives an average
for
L and
for Ree, respectively. Fig. 3 shows the variances
and
as functions of DNA
length, which were derived from the all-atomistic MD simulations. As shown in Figs. 3a and b,
increases roughly linearly with DNA length N over the range of 5-50 bps. Fig. 3c shows that
, the
6
2L22)(LLL2R22)(eeeeRRRLeeRLeeR2L2R2L2R2L2R
variance of Ree increases quadratically with DNA length, which is qualitatively in accordance with
the recent SAXS experiments with two attached gold nanocrystal.27 Since the force field associated
with the thioglucose-passivated gold nanocrystals involved in the experiments is unavailable,69-71 the
strict and direct comparisons with the experimental data27 are absent in the present study and a
non-strict comparison with the experiments is given in Figs. S3 and S4 of supplementary material.41
Nevertheless, the accurate treatment on short DNAs with linked nanocrystals and the direct
comparison are still required in future works, not only due to examining the modeling results, but
also due to understanding the effect of linked nanocrystals on the flexibility of short DNAs.
The Monte Carlo simulations with the WLC model have also been performed for short DNAs,
in a comparison with the results from the all-atomistic MD simulations. As shown in Eqs. 1 and 2,
there are two (bending and stretching) parameters for the WLC model: persistence length lp and
stretching modulus k which describe the global flexibility of a DNA. The previous combination of
experiments and the WLC model gives lp~50nm and k~1400-1600pN for long DNAs of kilo bps at
high salt solutions.1-7,12,72,73 As shown in Figs. 3a and b, the variance
of contour length L is not
sensitive to persistence length lp while sensitive to stretching modulus k. This is reasonable since lp
mainly reflects bending rigidity while k describes stretching rigidity, and contour length L mainly
depends on the latter. As shown in Figs. 3b and c, the WLC model with the parameters (lp~50nm and
k~1500pN) for long DNAs apparently underestimates the conformational fluctuation of short DNAs.
Figure 3b also shows that, over the range of 10–50 bps, the length-dependent
curve of WLC
model is closest to that of all-atomistic MD at k~1320pN, which is not far away from the
experimental value of 1400-1600pN for long DNAs in high salt solutions.3,7,72 By taking k~1320pN
and different lp’s, we found that the short DNAs of ≤50 bps have higher bending flexibility than long
of end-to-end distance Ree in Fig. 3c; see
ones of kilo bps, which is reflected by the variance
also Fig. S5 for the distribution of Ree.41 In addition, Fig. 3c shows that for lp = 45nm (and
k=1320pN),
of the WLC model for the 50-bp DNA appears close to that of the all-atomistic
MD. For the DNAs with length <50 bps, the WLC model with lp=45nm and k=1320pN still slightly
underestimates the conformation fluctuation, indicating higher flexibility for shorter DNAs. For
example, as shown in the inset of Fig. 3c, the WLC model with lp = 30nm (and k= 1320pN) still
for DNAs length ≲10 bps as compared with the all-atomistic MD
visibly underestimates
simulations. It is noted that the WLC model does indeed reproduce the desired lp with a high degree
of accuracy, as shown in supplementary material.41
The above analysis on the global conformation fluctuation of short DNAs of different lengths
shows that shorter DNAs have higher global flexibility (lower apparent persistence length) and with
a certain pair of persistence length lp and stretching modulus k, the WLC model cannot reproduce the
results of all-atomistic MD over the DNA length range from 5 to 50 bps. Since stretching and
bending are two major contributions to the global flexibility of short DNA, in the following, we will
analyze the helical stretching and bending of short DNAs at base-pair level.
B. Helix stretching
7
2L2L2R2R2R
The helix stretching of short DNAs may be characterized by three parameters: rise per bp, helix
radius and twist angle per bp.58,59 In the following, we will analyze the length-dependence of rise,
helix radius and twist angle based on the MD trajectories for short DNAs of different lengths. Our
analysis shows that helix rise and radius are strongly correlated, i.e., the increase and decrease of rise
are generally accompanied with the decrease and increase of helical radius, respectively; see Fig. S6a
of supplementary material.41 That is to say, the helix stretching/shrinking of a DNA is generally
accompanied by helix thinning/thickening. It is reasonable since helix stretching/shrinking would
lead to the less/more crowding of atoms in a DNA, and consequently causes DNA helix
thinning/thickening.
Firstly, we calculated the distribution of base-pair rise and rise variance along the 20-bp, 30-bp,
and 50-bp DNAs, respectively. As shown in Figs. 4a and b, in spite of the fluctuation of base-pair
rise between ~3Å and ~3.6Å, the rises of base pairs near two helix ends appear slightly larger than
those in the middle of short DNAs. Moreover, the variance of rise also has slightly larger values for
the base pairs near two DNA ends than that for the middle base pairs. The larger variance of rise near
the two ends suggests that the segments at two ends have stronger stretching flexibility than those
middle base pairs, which would cause the length-dependent stretching flexibility for short DNAs. As
shown in Figs. 4c and d, with the decrease of DNA length, mean rise and rise variance increase
slightly, suggesting slightly higher stretching flexibility for shorter DNAs. The mean rise per bp over
different short DNAs of 10-50 bps is close to the value of 3.32±0.19Å from the crystallographic
experiments.68,74
Corresponding to the negative correlation between helix rise and radius, the segments near two
helix ends have smaller helical radius and larger variance, as shown in Fig. S8a of supplementary
material.41 Consequently, the mean helix radius increases slightly from ~9.3Å to ~9.6Å and its
variance decreases, as DNA length increases from 10 to 50 bps. Therefore, the analyses on helix rise
and radius show that shorter DNAs have stronger stretching-shrinking conformation fluctuation and
the mean values of helix rise and radius are ~3.32Å and ~9.5Å over 10-50 bps which are close to
those of the experiments.68-75 It is understandable that shorter DNA has higher stretching flexibility
because of the less spatial constraints for helix stretching/shrinking and consequently higher
stretching flexibility for the segments at two helix ends.
In addition to helix rise and radius, we also examined helix twisting. Our analysis shows that
there is strong fluctuation of twist angle, and the mean twist angle per base pair increases very
slightly from ~32º to ~33º when DNA length increases from 10 to 50 bps; see Fig. S9 of
supplementary material.41 Additionally, the variance of twist angle also appears larger for the base
pairs near two helix ends in spite of fluctuation. It is reasonable since the segments at helix ends have
less spatial constraints to retain the helical structure. Furthermore, we examined the coupling
between stretching and twisting. Our analysis shows that helix rise and twist are strongly coupled,
i.e., the increase and decrease of rise are generally accompanied with the increase and decrease of
twist angle, respectively; see Fig. S6b in supplementary material.41 Such negative twist-stretch
coupling is in accordance with the previous studies.72,76,77
In the subsection of “Global structural flexibility”, we mainly used the helical axis of a short
8
DNA to characterize contour length L and end-to-end distance Ree which quantify the global
flexibility of a short DNA as a linear polymer. The length-dependence of helix twisting shows that
the slightly smaller twist angle and larger variance of twist angle may also contribute to the global
flexibility of shorter DNAs.76,77
C. Helix bending and sharp bending
1. Helix bending
In addition to helix stretching, helix bending is another major contribution to the structure
flexibility of DNAs.30,31,78-83 Since bending is difficult to be characterized at base-pair level, we use
the bending over 6 base pairs to characterize the helical bending of short DNAs,79,80 and the bending
angle over 6 bps is calculated as the angle between the first unit vector and the last unit vector along
a DNA axis over adjacent 6 base pairs.79 At first, we calculated the distributions of bending angle
(over 6 bps) and bending angle variance along short DNAs of 20, 30, and 50 bps. As shown in Figs.
5a and b, the bending angle and its variance near two helix ends are distinctly larger than those in the
middle of short DNAs, suggesting that the base pairs near two ends have significantly stronger
bending flexibility. Beyond the bending distribution along short DNAs, we calculated the mean
bending angle and its variance as functions of DNA length. As shown in Figs. 5c and d,
corresponding to the stronger bending flexibility near two ends described above, the mean bending
angle and its variance both increase apparently as DNA length decreases, suggesting the stronger
bending flexibility for shorter DNAs. Physically, the segments near two helix ends of short DNAs
have stronger bending flexibility due to less (more) spatial constraints (freedom) for bending, which
would cause the stronger bending flexibility for shorter DNAs.
2. Helix sharp bending
To further understand the bending properties of a short DNA, we analyzed the degree of helix
bending along DNAs. Here, a kind of sharp bending was defined as those bending angles with >30º
over 6 base pairs. Figure 6 shows that, along a short DNA, the segments near the helix ends are more
likely to sharply bend and the mean probability of sharp bending (per 6 bps) for a short DNA
decreases with the increase of DNA length. Physically, due to the less spatial constraints from outer
segments at two ends, the segments near DNA ends would fluctuate more strongly and frequently,
and sharp bending would occur more probably near two ends. As a result, the bending angle is larger
for shorter DNAs. In addition, we calculated the probability of such sharp bending for the WLC
model with the same apparent persistence length. As shown in Fig. 6a, the probability of sharp
bending from the all-atomistic MD is apparently higher than that of the WLC near two helix ends,
suggesting that the bending properties along a short DNA are beyond the description of the WLC
model.
What causes the sharp bending of a free DNA? Is sharp bending caused by the opening of
H-bonds? To classify the microscopic mechanism, we analyzed the status of H-bonds between
base-pairing nucleotides when a finite-length DNA sharply bends. Firstly, we examined the averaged
H-bond opening probability along the 40-bp DNA, where we used the criteria of bond length >3.5Å
9
to determine the H-bond opening.84 As shown in Fig. S10 of supplementary material,41 only H-bonds
at major groove have visible opening probability, while those at minor groove and between two
grooves (for C·G base pairs) have very low (almost invisible) opening probability. The two terminal
base pairs have higher H-bond (at major groove) opening probability than those in the middle helix,
and A·T base pairs have higher H-bond opening probability than C·G base pairs. Physically, the
terminal base pairs have no constraints from outer base pairs and thus could fluctuate more strongly.
Consequently, the two terminal base pairs have higher opening probability of H-bond. In addition,
C·G base pairs have one more H-bond than A·T, thus the H-bond of C·G base pairs has more spatial
constraints and consequently has low opening probability than that of A·T base pairs. However, there
is no visible higher H-bond opening probability near the sharp bending sites, and for a sharp bent
conformation, H-bonds are not disrupted at the sharp bending sites, which suggests that H-bond
opening may not make the major contribution to the sharp bending for short DNAs. It is reasonable
since the melting temperatures of the short DNAs are significantly higher than room temperature
(298K); see Table 1.
In another way, the microscopic analysis on rise distribution along helix shows that at the
junction between stretching and shrinking segments on a DNA, the helix is shown to sharply bend.
As shown in Figs. 6c and d, there are two stretching-shrinking junctions along the 40-bp DNA, and
there are two sharp bending sites at the two corresponding junctions. Furthermore, our microscopic
analysis shows that, sharp bending is generally towards major groove and during bending, the base
pairs at bending sites would slide away from helix central axis to minor groove and major groove
would become deeper, which favors a sharp bending toward major groove in addition to the broad
width of major groove.79,80 Furthermore, such sharp bending would be aided by the asymmetrical ion
binding to bent DNAs.
3. Asymmetrical ion binding aids bending
Since DNAs are negatively charged macromolecules, metal ions can play an important role in
DNA structure deformation such as helix bending.79,80,85-96 As shown in Fig. 7, Na+ ions become
condensed around negatively charged DNA helix, and at the base pairs where the helix is sharply
bent, there are more Na+ ions accumulated to screen the Coulombic repulsion due to the bending. To
further analyze the role of Na+ during helix bending, we separated the condensed Na+ ions into those
binding in bending direction and in the opposite direction. As shown in Figs. 7c and d, it is clear that
the local density of condensed Na+ ions in bending direction is higher than that in the opposite
direction. Such asymmetrical ion binding would assist DNA bending, which is also in accordance
with the previous analysis.79,80,95 The salt effect, including ion concentration and ion valence, on
DNA bending is very important and deserved to be investigated separately in the future.
D. Apparent persistence length of short DNAs
Like other polymers, the global flexibility of a DNA can be quantified by its persistence length
lp.2-7,12,13,97-103 In experiments, the persistence length of a DNA can be determined by fitting an
elasticity model to stretching experimental data,2-5,12 or by measuring the bending angle along the
10
tangent vectors of a DNA25,32,33,99 or by measuring the size of a DNA.26,27,98,100 In the following, we
will calculate apparent persistence length of short DNAs with different lengths since DNA ends have
higher flexibility than the central base pairs. Based on the MD trajectories and the program
Curves+,59 the apparent persistence length lp for a short DNA of N bps can be calculated by 65,101,102
, (3)
where
is the unit vector along the central axis of a DNA which connects the centers of adjacent
base pairs. Meanwhile, the apparent persistence length for a short DNA can also be calculated
through the statistics of the mean square end-to-end distance
by2,65
. (4)
The two equations (Eqs. 3 and 4) were both employed in calculating lp for different short DNAs.
Figure 8a shows lp as a function of DNA length. With the increase of DNA length from 10 to 50 bps,
lp increases gradually from ~29nm to ~45nm and the two methods (Eqs. 3 and 4) give the similar
predictions on lp. For longer DNAs, lp is expected to become close to the value (~50nm) of long
DNAs of kilo bps.2-5,13,34 Our calculations also indicate that, lp’s calculated from MC with the WLC
model are just close to that set in the model (Eq. 1) and nearly independent of DNA length. This
suggests that the WLC model with a certain pair of parameters (persistence length lp and stretching
modulus k in Eqs. 1 and 2) cannot describe the length-dependent apparent persistence length for
short DNAs of different lengths.
1. Comparisons with previous studies
Firstly, we make comparisons with experiment data.26,29,35,80 Yuan et al have employed FRET
and SAXS to quantify the radius of gyration Rg of short DNAs with 16, 21, and 66 bps, and Rg can be
used to estimate the apparent persistence length26; see the caption of Fig. 8. As shown in Fig. 8, the
predicted lp’s agree well with the data from the experiments, except for N=16 bps. Our value of
~32nm for N=15 is higher than that (~16nm) of the experiments for N=16.26 Such deviation for the
16-bp DNA may possibly arise from the fact that the inaccuracy involved in the experiments may
become stronger for shorter DNA and the calculation of lp is sensitive to the measurement of Rg. Also,
the difference in sequence may contribute to the deviation. Mastroianni et al have also employed
SAXS to probe the conformation flexibility of DNAs of 42-94 bps with two linked end
nanocrystals.29 By combining with the WLC model, they concluded that the studied short DNAs
have approximately lp of ~50nm. Our prediction on lp does not differ much from the experimental
value since our predicted lp≳43nm for N ≥40 bps. Most of AFM experiments were performed for long
DNAs in kilo bps attached on a substrate and thus, the segments at two helix ends have generally
been ignored in the AFM measurements.25,32,33,103 Consequently, a direct comparison on lp with the
AFM experiments is absent in the present work. More extensive comparisons with experiments
require more accurate experiment data on lp for short DNAs. Secondly, we make the comparisons
with the previous simulational studies.35,80 Noy and Golestanian’s all-atomistic MD gives that, the
averaged apparent persistence length is ~43nm for a 56-bp DNA,35 a value close to our prediction
11
1))1()1(ln(NuuLlpu2eeRLlLlLlRpppee/))/exp(1(122
(~45nm for 50-bp). Spiriti et al employed the adaptive umbrella method to simulate DNA bending
and found that the 12-bp DNAs with 33% A·T bps and excluding two terminal bps have a lp of
~41nm at 150mM KCl.80 This value is a little higher than our value of ~32nm for the 15-bp with
40% A·T bps at 1M NaCl. Such slightly higher value of lp of Spiriti et al may come from the
exclusion of two terminal base pairs in the data analysis and much lower bulk salt employed in the
simulations.80
2. Empirical formula
Based on the apparent persistence lengths from the all-atomistic MD for the short DNAs of
10-50 bps in 1M NaCl solutions, we obtained an empirical formula for lp as a function of DNA length
N (bps)
,
(5)
where
(~50nm) is the persistence length of DNAs of kilo bps, and N is DNA length (bps). The
parameters A~450nm and B~10. As shown in Fig. 8, Eq. 5 fits the values from the all-atomistic MD
and experimental data well and would approach gradually to the value of long DNA as DNA length
increases. Here, it would be interesting to revisit the variance
of Ree discussed in the subsection
of “Global structure flexibility”. Can we reproduce the length-dependent
of all-atomistic MD by
the WLC model with the length-dependent apparent persistence length lp? As shown in Fig. 8b, with
the apparent persistence length lp from Eq. 5 and stretching modulus k=1320pN,
’s from the
WLC model are close to those of the all-atomistic MD over the DNA length range from 10 to 50 bps.
E. Finite-length effect
In the above, we have shown the apparently stronger bending and stretching flexibilities near
two ends of short DNAs, which causes the distinct finite-length effect in DNA flexibility. To
quantitatively examine the finite-length effect in the flexibility of short DNA, we make another
analysis on the all-atomistic MD data by comparing the variance
of isolated short DNAs and
that of the “inner” short DNAs with the same length which are taken from the middle of the 50-bp
DNA. As shown in Fig. 9a, we found that the isolated short DNAs have apparently higher structural
flexibility (
) than those “inner” short DNAs with the same length, verifying the higher flexibility
of shorter DNA which is attributed to the significantly higher flexibility of the segments at helix
ends.
Then how many base pairs at each helix end contribute to the higher flexibility of short DNAs?
In the following, we will quantify the number of base pairs which contribute to the strong
finite-length effect. Firstly, we calculated the variance
of end-to-end distance for the short
DNAs with excluding several (2, 4, 6, 8 and 10) base pairs at each end, and compared
with the
WLC model with lp=50nm (and k=1320pN). As shown in Fig. 9b, with the increase of the number of
excluded end base pairs,
of short DNAs decreases and approaches to that of the WLC model,
and when ~6 base pairs at each end are excluded from DNAs,
appears almost identical to that of
12
)/()(0NBAlNlpp0pl2R2R2R2R2R2R2R2R2R
the WLC model with lp=50nm. For these short DNAs with excluding end base pairs, we also
calculated apparent persistence length lp through Eq. 3. As shown in Fig. 9c, as more end base pairs
are excluded, lp would increase and when more than ~6 end base pairs are excluded, lp for short
DNAs of different lengths would converge to ~50nm, the value of long DNAs in kilo bps.
Therefore, Fig. 9 suggests that, the ~6 base pairs at each helix end are responsible for the
finite-length effect which results in the significantly higher flexibility (lower apparent lp) of short
DNAs. This is reasonable since base pairs at helix ends generally fluctuate with less spatial
constraints and consequently have the large bending/stretching flexibility. We expect that the
finite-length effect would become weak when DNA length is much longer than 6 bps, and the
number of ~6 (bps) would change when the solution ionic condition deviates away from 1M NaCl,
due to the significant electrostatic contribution to DNA flexibility.13,79,85-98,104
IV. CONCLUSIONS
In this paper, we have studied the flexibility of short DNAs with finite-length effect by the
all-atomistic MD simulations and MC simulations with the WLC model. The length of short DNAs
employed in the study covered the wide range from 5 to 50 bps. We investigated the flexibility of
short DNAs of various lengths by analyzing the end-to-end distance and contour length, and by
analyzing the stretching and bending of short DNAs of different lengths at base-pair level. The main
conclusions are as follows: (i) The short DNAs have lower apparent persistence length than long
ones, and such low apparent persistence length is attributed to the high flexibility of ~6 base pairs at
each helix end; (ii) Due to the strong finite-length effect caused by the high flexibility near two helix
ends, the WLC model with the persistence length (~50nm) of long DNAs in kilo bps would
underestimate the structure flexibility of short DNAs of 5-50 bps, and only with excluding ~6 base
pairs at each end, a short DNA can be described by the WLC model with persistence length of
~50nm; (iii) The short DNAs may sharply bend at stretching-shrinking junctions, and such sharp
bending occurs more frequently near two helix ends; (iv) The apparent persistence length of short
DNAs with finite-length effect increases gradually from ~29nm to ~45nm as DNA length increases
from 10 to 50 bps, and we obtained an empirical formula for the apparent persistence length as a
function of DNA length which may be practically useful.
In addition to the above described general conclusions, the present work would be very helpful
for understanding the elusive controversy on the global flexibility of short DNAs arising from the
existing experiments.24-31 Our results show that, compared with long DNAs, short DNAs have
apparently higher flexibility for DNA length N≤ ~30 bps while slightly higher flexibility for N≥ ~40
bps, which might help to bridge the gap between the observations from the experiments of
Mathew-Fenn et al for N≤35 bps27 and those of Mastroianni et al for N>40 bps.29 The nanocrystals
linked with short DNAs
the experimental
observations,26,27,29,105,106 while was not involved in the present work. However, the present work
could not explain the cyclization experiments of short DNAs with ~60-100 bps24,28 since the end
effect would become small for DNAs of ~60-100 bps. Furthermore, DNA cyclization generally
involves the formation of DNA loop in which the ends are not fully free.24,28,30,31
the experiments may also contribute
in
to
13
The present work is apparently different from the previous MD works, e.g., the work of Noy
and Golestanian,35 at least in the following four aspects: (i) The present work covers the wide length
range of short DNAs to systematically study the length-dependent global flexibility of short DNA
with finite-length effect which was often ignored in the works of Noy and Golestanian and others
25,32,33,35; (ii) The length-dependent apparent persistence length of short DNAs has been obtained in
the work; (iii) The WLC model has been extensively employed in parallel with the all-atomistic MDs,
to make more comprehensive analyses on the flexibility of short DNAs; (iv) The present work is
extended to longer simulation time and three independent MD runs to warrant the converge and the
stable equilibrium properties. Moreover, the conclusion that shorter DNAs have higher flexibility
(lower apparent persistence length) is not contradictory with the recent AFM experiments (e.g. that
of Mazur and Maaloum32,33) since it is the segments near helix ends which results in the higher
flexibility of short DNAs while such effect was generally ignored in the AFM studies.25,32,33,103 Our
analyses show that the flexibility of short DNAs with excluding ~6 base pairs at each end would
behave like long DNAs of kilo bps, and the finite-length effect would become weak for long DNAs.
The present work also has some approximations and limitations. Firstly, we employed the
quadratic function in analyzing the distributions of bending and its variance (and rise and its variance)
along short DNAs in spite of certain fluctuation. However, the overall larger values of bending angle
and variances of bending angle and rise at helix ends for all short DNAs would indeed suggest the
higher flexibility of DNA ends. Secondly, the analyses and predictions in the work would strongly
depend on the force field employed in the MD simulations, while our analyses show that the
employed force field should be reliable since the predicted apparent persistence length approaches to
~50nm (the well-accepted experimental value) for the short DNAs without >~6 base pairs at each
helix end. Thirdly, we ignored the few conformations with the terminal base pair opening since the
program Curves+ would produce abnormal analysis for such conformations, while such ignorance
should not notably affect our results due to the very small amount of the conformations. The
inclusion of such states with frayed terminal base pairs at the ends of short DNAs would further
increase the flexibility of base pairs at two ends. Fourthly, although there is the direct comparison
with experimental persistence length, the present work did not involve the direct comparisons with
the experiment on length-dependent end-to-end distance and its variance due to the lack of the
associated force field for the linked nanocrystals27. Such direct and strict comparison is still
necessary and significant in the future study, which would be helpful for examining the simulation
analysis, and understanding the effect of the labeling nanocrystals on probing the flexibility of
biomacromolecules.105,106 Finally, the flexibility of DNAs is also dependent on sequence,41,42,45,107
temperature108,109 and ionic condition5,13,92,99, which is beyond the scope of the present work and
would be deserved to be discussed separately. Nevertheless, the present work would be very helpful
for understanding the flexibility of short DNAs and the controversy on the high flexibility of short
DNAs. The obtained empirical formula of apparent persistence length might be practically useful.
ACKNOWLEDGEMENT
We are grateful to Profs. Shi-Jie Chen, Haiping Fang and Wenbing Zhang for valuable
14
discussions, and Chang Shu for facility assistance. We are also grateful for Prof. Xian-Wu Zou and
Ya-Zhou Shi for critical reading the manuscript. This work was supported by the National Key
Scientific Program (973)-Nanoscience and Nanotechnology (No. 2011CB933600), the National
Science Foundation of China grants (11074191, 11175132 and 11374234), the Program for New
Century Excellent Talents (Grant No. NCET 08-0408). One of us (Y. Y. Wu) also thanks financial
supports from the interdisciplinary and postgraduate programs under the Fundamental Research
Funds for the Central Universities.
15
References:
1. V. A. Bloomfield, D. M. Crothers, and I. Tinoco, Jr., Nucleic Acids: Structures, Properties, and Functions, 1st
ed. (University Science Books, 2000).
2. P. J. Hagerman, Annu. Rev. Biophys. Biophys. Chem. 17, 265-286 (1988).
3. S. B. Smith, Y. J. Cui, and C. Bustamante, Science 271, 795-799 (1996).
4. P. Cluzel, A. Lebrun, C. Heller, R. Lavery, J. L. Viovy, D. Chatenay, and F. Caron, Science 271,792-794
(1996).
5. C. G. Baumann, S. B. Smith, V. A. Bloomfield, and C. Bustamante, Proc. Natl. Acad. Sci. USA 94, 6185-6190
(1997).
6. Z. Bryant, M. D. Stone, J. Gore, S. B. Smith, N. R. Cozzarelli, and C. Bustamante, Nature 424, 338-341
(2003).
7.
J. R. Wenner, M. C. Williams, I. Rouzina, and V. A. Bloomfield, Biophys. J. 82, 3160-3169 (2002).
8. W. B. Fu, X. L. Wang, X. H. Zhang, S. Y. Ran, J. Yan, and M. Li, J. Am. Chem. Soc. 128, 15040-15041
(2006).
9. X. Zhang, H. Chen, H. Fu, P. S. Doyle, and J. Yan, Proc. Natl. Acad. Sci. USA 109, 8103-8108 (2012).
10. Y. Zhang, H. Zhou, and Z. C. Ou-Yang, Biophys. J. 81, 1133-1143 (2001).
11. J. F. Marko, and E. D. Siggia, Macromolecules 28, 8759-8770 (1995).
12. F. Massi, J. W. Peng, J. P. Lee, and J. E. Straub, Biophys. J. 80, 31-44 (2001).
13. G. S. Manning, Biophys. J. 91, 3607-3616 (2006).
14. J. P. Peters, S. P. Yelgaonkar, S. G. Srivatsan, Y. Tor, and L. J. Maher, Nucleic Acids Res. 41, 10593-10604
(2013).
15. P. Ranjith, J. Yan, and J. F. Marko, Proc. Natl. Acad. Sci. USA 104, 13649-13654 (2007).
16. W. K. Olson, N. Clauvelin, A. V. Colasanti, G. Singh, and G. H. Zheng, Biophys. Rev. 4, 171-178 (2012).
17. T. J. Richmond, and C. A. Davey, Nature 423, 145-150 (2003).
18. J. Y. Zhang, M. J. McCauley, L. J. Maher, M. C. Williams, and N. E. Israeloff, Nucleic Acids Res. 37,
1107-1114 (2009).
19. S. C. Schultz, G. C. Shields, and T. A. Steitz, Science 253, 1001-1007 (1991).
20. W. Li, P. Wang, J. Yan, and M. Li, Phys. Rev. Lett. 109, 218102 (2012).
21. J. E. Coats, Y. Lin, E. Rueter, L. J. Maher, and I. Rasnik, Nucleic Acids Res. 41, 1372-1381 (2013).
22. J. H. Bredenberg, C. Russo, and M. O. Fenley, Biophys. J. 94, 4634-4645 (2008).
23. H. Lei, J. Sun, E. P. Baldwin, D. J. Segal, and Y. Duan, Adv. Protein Chem. Struct. Biol. 94, 347-364 (2013).
24. T. E. Cloutier, and J. Widom, Mol. Cell 14, 355-362 (2004).
25. P. A. Wiggins, T. Van Der Heijden, F. Moreno-Herrero, A. Spakowitz, R. Phillips, J. Widom, C. Dekker, and P.
C. Nelson, Nature Nanotech. 1, 137-141 (2006).
26. C. L. Yuan, H. M. Chen, X. W. Lou, and L. A. Archer, Phys. Rev. Lett. 100, 018102 (2008).
27. R. S. Mathew-Fenn, R. Das, and P. A. Harbury, Science 322, 446-449 (2008).
28. R. Vafabakhsh, and T. Ha, Science 337, 1097-1101 (2012).
29. A. J. Mastroianni, D. A. Sivak, P. L. Geissler, and A. P. Alivisatos, Biophys. J. 97, 1408-1417 (2009).
30. A. Vologodskii, Q. Du, and M. D. Frank-Kamenetskii, Artif. DNA PNA XNA 4, 1-3 (2013).
31. A. Vologodskii, and M. D. Frank-Kamenetskii, Nucleic Acids Res. 41, 6785-6792 (2013).
32. A. K. Mazur, and M. Maaloum, Phys. Rev. Lett. 112, 068104 (2014).
16
33. A. K. Mazur, and M. Maaloum, Nucleic Acids Res. gku1192 (2014).
34. E. Herrero-Galan, M. E. Fuentes-Perez, C. Carrasco, J. M. Valpuesta, J. L. Carrascosa, F. Moreno-Herrero, and
J. R. Arias-Gonzalez, J. Am. Chem. Soc. 135, 122-131 (2012).
35. A. Noy, and R. Golestanian, Phys. Rev. Lett. 109, 228101 (2012).
36. A. K. Mazur, Phys. Rev. Lett. 111, 179801 (2013).
37. A. Noy, and R. Golestanian, Phys. Rev. Lett. 111, 179802 (2013).
38. S. Xiao, and H. Liang, J. Chem. Phys. 136, 205102 (2012).
39. Y. Zhang, J. Zhang, and W. Wang, J. Am. Chem. Soc. 133, 6882-6885 (2011).
40. Z. Gong, Y. Zhao, C. Chen, Y. Duan, and Y. Xiao, PloS one 9, e92247 (2014).
41. See supplementary materials at XXX for an additional text with 10 figures, and two movies. The two movies
illustrate the dynamic structure fluctuations of the 50-bp DNA from two 200ns MD simulation trajectories.
42. J. Widom, Q. Rev. Biophys. 34, 269-324 (2001).
43. L. Czapla, D. Swigon, and W. K. Olson, J. Chem. Theory Comput. 2, 685-695 (2006).
44. J. SantaLucia, Jr., Proc. Natl. Acad. Sci. USA 95, 1460-1465 (1998).
45. Z. J. Tan, and S. J. Chen, Biophys. J. 92, 3615-3632 (2007).
46. Z. J. Tan, and S. J. Chen, Biophys. J. 90, 1175-1190 (2006).
47. C. Schafmeister, W. S. Ross, and V. Romanovski, LEAP, (University of California, 1995).
48. S. Lee, Y. Song, and N. A. Baker, Biophys. J. 94, 3565-3576 (2008).
49. B. Hess, C. Kutzner, D. Van Der Spoel, and E. Lindahl, J. Chem. Theory Comput. 4, 435-447 (2008).
50. A. Pérez, I. Marchán, D. Svozil, J. Sponer, T. E. Cheatham III, C. A. Laughton, and M. Orozco, Biophys. J. 92,
3817-3829 (2007).
51. B. Tarus, J. E. Straub, and D. Thirumalai, J. Am. Chem. Soc. 128, 16159-16168 (2006).
52. F. Massi, J. W. Peng, J. P. Lee, and J. E. Straub, Biophys. J. 80, 31-44 (2001).
53. M. Feig, and B. M. Pettitt, Biophys. J. 75, 134-149 (1998).
54. Z. Gong, Y. Zhao, and Y. Xiao, J. Biomol. Struct. Dyn. 28, 431-441 (2010).
55. D. Van Der Spoel, E. Lindahl, B. Hess, G. Groenhof, A. E. Mark, and H. J. Berendsen, J. Comput. Chem. 26,
1701-1718 (2005).
56. U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee, and L. G. Pedersen, J. Chem. Phys. 103,
8577-8593 (1995).
57. S. Miyamoto, and P. A. Kollman, J. Comput. Chem. 13, 952-962 (1992).
58. X. J. Lu, and W. K. Olson, Nucleic Acids Res. 31, 5108-5121 (2003).
59. R. Lavery, M. Moakher, J. H. Maddocks, D. Petkeviciute, and K. Zakrzewska, Nucleic Acids Res. 37,
5917-5929 (2009).
60. A. K. Mazur, J. Phys. Chem. B 112, 4975-4982 (2008).
61. D. A. Sivak, and P. L. Geissler, J. Chem. Phys. 136, 045102 (2012).
62. A. K. Mazur, Biophys. J. 91, 4507-4518 (2006).
63. N. Madras, and A. D. Sokal, J. Stat. Phys. 50, 109-186 (1988).
64. Y. Z. Shi, F. H. Wang, Y. Y. Wu, and Z. J. Tan, J. Chem. Phys. 141, 105102 (2014).
65. F. H. Wang, Y. Y. Wu, and Z. J. Tan, Biopolymers 99, 370-381 (2013).
66. J. Wilhelm, and E. Frey, Phys. Rev. Lett. 77, 2581 (1996).
17
67. D. Thirumalai, and B. Y. Ha. In A. Grosberg (Ed.). Theoretical and mathematical models in polymer research,
San Diego, CA. Academic Press. Vol. 1, pp135 (1998).
68. W. K. Olson, A. A. Gorin, X. Lu, L. M. Hock, and V. B. Zhurkin, Proc. Natl. Acad. Sci. USA 95, 11163-11168
(1998).
69. A. C. Yang, and C. I. Weng, J. Phys. Chem. C 114, 8697-8709 (2010).
70. S. A. Alsharif, L. Y. Chen, A. Tlahuice-Flores, R. L. Whetten, and M. J. Yacaman, Phys. Chem. Chem. Phys.
16, 3909-3913 (2014).
71. Y. Y. Wu, F. H. Wang, and Z. J. Tan, Phys. Lett. A 377, 1911-1919 (2013).
72. P. Gross, N. Laurens, L. B. Oddershede, U. Bockelmann, E. J. Peterman, and G. J. Wuite, Nature Physics 7,
731-736 (2011).
73. P. C. Nelson, Science 337, 1045-1046 (2012).
74. A. Wynveen, D. J. Lee, A. A. Kornyshev, and S. Leivin, Nucleic Acids Res. 36, 5540-5551(2008).
75. R. R. Sinden, DNA Structure and Function, (Academic Press, INC., 1994).
76. R. D. Kamien, T. C. Lubensky, P. Nelson, and C. S. O'Hern, Europhys. Lett. 38, 237 (1997)
77. J. Gore, Z. Bryant, M. Nöllmann, M. U. Le, N. R. Cozzarelli, and C. Bustamante, Nature 442, 836-839 (2006).
78. H. Zhou, Y. Zhang, and Z. C. Ou-Yang, Phys. Rev. Lett. 82, 4560 (1999).
79. Z. J. Tan, and S. J. Chen, Biophys. J. 94, 3137-3149 (2008).
80. J. Spiriti, H. Kamberaj, A. M. de Graff, M. F. Thorpe, and A. van der Vaart, J. Chem. Theory Comput. 8,
2145-2156 (2012).
81. M. Muthukumar, Annu. Rev. Biophys. Biomol. Struct. 36, 435-450 (2007).
82. B. Ha, and D. Thirumalai, Macromolecules 36, 9658-9666 (2003).
83. B. Ha, and D. Thirumalai, Macromolecules 28, 577-581 (1995).
84. G. A. Jeffrey, An Introduction To Hydrogen Bonding, (Oxford University Press, 1997).
85. G. S. Manning, Q. Rev. Biophys. 11, 179-246 (1978).
86. L. Pollack, Annu. Rev. Biophys. 40, 225-242 (2011).
87. Z. J. Tan, and S. J. Chen, Nucleic Acids Res. 34, 6629-6639 (2006).
88. Z. J. Tan, and S. J. Chen, Biophys. J. 103, 827-836 (2012).
89. E. Stellwagen, J. M. Muse, and N. C. Stellwagen, Biochemistry 50, 3084-3094 (2011).
90. L. D. Williams, and L. J. Maher, Annu. Rev. Biophys. Biomol. 29, 497-521 (2000).
91. S. Xiao, H. Zhu, L. Wang, and H. Liang, Soft Matter 10, 1045-1055 (2014).
92. M. O. Fenley, G. S. Manning, and W. K. Olson, J. Phys. Chem. 96, 3963-3969 (1992).
93. A. Savelyev, and G. A. Papoian, Proc. Natl. Acad. Sci. USA 107, 20340-20345 (2010).
94. P. Ren, J. Chun, D. G. Thomas, M. J. Schnieders, M. Marucho, J. Zhang, and N. A. Baker, Q. Rev. Biophys. 45,
427-491 (2012).
95. K. M. Kosikov, A. A. Gorin, X. Lu, W. K. Olson, and G. S. Manning, J. Am. Chem. Soc. 124, 4838-4847
(2002).
96. N. A. Simonov, M. Mascagni, and M. O. Fenley, J. Chem. Phys. 127, 185105 (2007).
97. J. K. Weber, and V. S. Pande, J. Chem. Phys. 139, 121917 (2013).
98. G. Caliskan, C. Hyeon, U. Perez-Salas, R. M. Briber, S. A. Woodson, and D. Thirumalai, Phys. Rev. Lett. 95,
268303 (2005).
99. A. Savelyev, Phys. Chem. Chem. Phys. 14, 2250-2254 (2012).
18
100. X. Shi, D. Herschlag, and P. A. Harbury, Proc. Natl. Acad. Sci. USA 110, E1444-E1451 (2013).
101. A. Fathizadeh, B. Eslami-Mossallam, and M. R. Ejtehadi, Phys. Rev. E 86, 051907 (2012).
102. Y. Lu, B. Weers, and N. C. Stellwagen, Biopolymers 61, 261-275 (2002).
103. F. Faas, B. Rieger, L. J. Van Vliet, and D. I. Cherny, Biophys. J. 97, 1148-1157 (2009).
104. Z. J. Tan, and S. J. Chen, J. Chem. Phys. 122, 044903 (2005).
105. A. K. Mazur, Phys. Rev. E 80, 10901 (2009).
106. G. Zheng, L. Czapla, A. R. Srinivasan, and W. K. Olson, Phys. Chem. Chem. Phys. 12, 1399-1406 (2010).
107. V. Ortiz, and J. J. de Pablo, Phys. Rev. Lett. 106, 238107 (2011).
108. S. Geggier, A. Kotlyar, and A. Vologodskii, Nucleic Acids Res. 39, 1419-1426 (2010).
109. N. Theodorakopoulos, and M. Peyrard, Phys. Rev. Lett. 108, 078104 (2012).
19
Figure Captions
Figure 1 (a) An illustration for the end-to-end distance Ree and the contour length L of the 50-bp
atomistic DNA helix; see also Table 1 and Fig. S1. (b) The instantaneous end-to-end distance Ree (red
and green lines) versus MD simulation time from two independent MD simulations for the 50-bp
DNA helix. (c) The averaged value of end-to-end distance over every Δt by
(red
and green lines) versus MD simulation time from two independent MD simulations for the 50-bp
DNA helix. Here, Δt =500ps. The central line denotes the mean value
of end-to-end distance in
equilibrium (MD time is larger than 100ns for the 50-bp DNA). The analyses on the DNA
conformations were performed with the program Curves+.59
Figure 2 (a) The normalized probability distributions of end-to-end distance Ree (red) and contour
length L (blue) distribution curves for the 5-bp, 10-bp, 15-bp, 20-bp, 25-bp, 30-bp, 40-bp, and 50-bp
DNA helices, respectively (from left to right). Note that the distributions are not perfectly normal
and the deviation from normal distribution is slight, as discussed in main text. (b) The mean
end-to-end distance Ree (red) and contour length L (blue) as functions of the DNA length (bps). The
mean values of Ree and L are simply calculated by averaging over all the possible conformations in
equilibrium. The slopes of end-to-end distance Ree and contour length L are ~3.22Å and ~3.30 Å per
base pair, respectively.
Figure 3 The variances
of contour length L (a, b) and the variances
of end-to-end
distance Ree (c) as functions of DNA length (bps). Solid lines: all-atomistic MD; dash lines: MC
simulation with the worm-like chain model where the stretching modulus k and persistence length lp
are shown in the figures.
and
were calculated by averaging the
values over all the conformations in equilibrium. It is noted that the WLC model does indeed
reproduce the desired lp with a high degree of accuracy, as shown in supplementary material.41
Figure 4 (a) The distribution of rise per base pair along the 20-bp, 30-bp and 50-bp DNAs (from
top to bottom). (b) The variance distribution of rise per base pair along the 20-bp, 30-bp and 50-bp
DNAs (from top to bottom). For panels (a) and (b), the mean rise
and rise variance
along the short DNAs were calculated by averaging the values over all the
conformations in equilibrium. (c) The mean rise per base pair for short DNAs as a function of DNA
length (bps). (d) The mean variance of rise per base pair for short DNAs as a function of DNA length
(bps). It is noted that in panels (a) and (b), the lines were fitted with quadratic function rather than
linear horizontal function since the former can give better fits with smaller deviation for the points.
For the 50-bp DNA, the fitting deviations with quadratic function are 0.11432 and 0.01818 for the
rise per base pair and the rise variance, while the fitting deviations with the linear horizontal function
20
/2 /2(')'/tteettRtdtteeR2L2R22)(LLL22)(eeeeRRRir22)()(iirrri
are 0.12907 and 0.020695, respectively. The distribution of rise per base pair and variances
(averaged over 6 bps) along the 20-bp, 30-bp, and 50-bp DNAs were also shown in Fig. S7 of
supplementary materials.41
Figure 5 (a) The distribution of bending angle (over 6 bps) along the 20-bp, 30-bp, and 50-bp
DNAs (from top to bottom). (b) The distribution of bending angle variance (over 6 bps) along the
20-bp, 30-bp, and 50-bp DNAs (from top to bottom). For panels (a) and (b), the bending angle
and rise variance
along the short DNAs were calculated by averaging the values
over all the conformations in equilibrium. (c) The mean bending angle (over 6 bps) for short DNAs
as a function of DNA length (bps). (d) The mean variance of bending angle (over 6 bps) for short
DNAs as a function of DNA length (bps).
Figure 6 (a) The sharp bending (≥30º over 6 bps) probability along the 30-bp and 50-bp DNAs
from all-atom MD simulations (blue symbols) and MC simulations with the WLC model with the
same apparent persistence length from the MD simulations (red lines; lp=45nm and 38nm for the
50-bp and 30-bp DNAs, respectively). It is noted that in panel (a), the blue lines were fitted with
quadratic function rather than linear horizontal function since the former can give better fits with
smaller deviation for the points. For the 50-bp DNA, the fitting deviation with quadratic function is
0.00471391 while the fitting deviations with the linear horizontal function is 0.00558103; (b) The
mean sharp bending (≥30º over 6 bps) probability for short DNA helices as a function of DNA length
(bps). The sharp bending probability over 6 bps for a short DNA was calculated with averaging the
probability in each 6-bp section over the whole length of the DNAs. (c) A snapshot of the 40-bp
DNA to show the sharp bending sites denoted by the circles. (d) The base-pair rise distribution along
the 40-bp DNA helix for the conformation shown in panel (c). The central axis of the DNA is
determined with the program Curves+.59
Figure 7 (a) An illustration for the region of high ion charge density (larger than 0.02 e/Å3) around
a 40-bp DNA helix with a sharp bending. (b) The distribution of binding ion charge per bp in the
cylindrical cell with radius of 12Å around the bent DNA helix. (c) The distribution of binding ion
charge in concave side in the cylindrical cell with radius of 12Å around the bent DNA helix. (d) The
distribution of binding ion charge in convex side in the cylindrical cell with radius of 12Å around the
bent DNA helix.
Figure 8 (a) The apparent persistence lengths
of short DNAs as a function of DNA length
(bps). The apparent persistence length
was calculated from all-atomistic MD with Eq. 3 (●) and
Eq. 4 (◆), respectively. The experimental apparent persistence length
(■) was calculated from
the experimental data of
the
radius of gyration of
the
short DNAs by
(26)
, and
is corrected by involving DNA
21
i22)()(iiiplplpl22342/32/2(1exp(/))/gpppppRLlllLlLlL2gR
radius RDNA:26
, where the radius of DNA
=1.1nm.26 The dash line is the
persistence length calculated from the WLC model with lp=45nm. The solid line is the empirical
formula of Eq. 5. (b) The variance
of end-to-end distance Ree of short DNAs as a function of
DNA length (bps). Solid line: from all-atomistic MD; Dash line: from the WLC model with
k=1320pN and a length-dependent
of Eq. 5.
Figure 9 (a) The variance
of end-to-end distance Ree as a function of DNA length for isolated
short DNAs (full symbols) and the “inner” short DNAs with the same length which are taken from
the middle of the 50-bp DNA (open symbols), respectively. (b) The variance
of Ree as a
function of DNA length for the short DNAs with excluding 2, 4 and 6 bps at each helix end. Here,
DNA length indicates that of the short DNAs after 2, 4, and 6 bps at each end were removed. (c) The
apparent persistence length of the short DNAs with excluding 2, 4, 6, 8 and 10 base pairs at each
helix end. Here, DNA length indicates the original length of the short DNAs before 2, 4, 6, 8 and 10
bps at each end were removed.
22
222,/2gDNAgDNARRRDNAR2Rpl2R2R
Figure 1
23
Figure 2
24
Figure 3
25
Figure 4
26
Figure 5
27
Figure 6
28
Figure 7
29
Figure 8
30
Figure 9
31
Table 1 DNA sequences used in the study.a
helix
length
(bps)
5
10
15
20
25
30
40
50
Sequences
5’-GCAGC-3’
CGTCG
5’-GCATCTGGGC-3’
CGTAGACCCG
5’-CGACTCTACGGAAGG-3’
GCTGAGATGCCTTCC
5’-CGACTCTACGGCATCTGCGC-3’
GCTGAGATGCCGTAGACGCG
5’-CGACTCTACGGAAGGGCATCTGCGC-3’
GCTGAGATGCCTTCCCGTAGACGCG
5’-CGACTCTACGCAAGGTCTCGGACTACGCGC-3’
GCTGAGATGCCTTCCAGAGCCTGATGCGCG
Simulation box size
(Lx×Ly×Lz) Å3
CS (M)
b
Predicted
Tm (ºC)
c
72×72×74
0.00866
50.5
72×72×74
0.00866
72×72×74
0.00866
77.8
81.2
80×80×90
0.00577
90.1
85×86×102
0.00446
92.8
91×91×132
0.00304
93.6
5’-CGACTCTACGGAAGGGCATCCTTCGGGCATCACTACGCGC-3’
GCTGAGATGCCTTCCCGTAGGAAGCCCGTAGTGATGCGCG
91×91×167
0.00240
96.8
5’-CGACTCGACTCTACGGAAGGGCATCCTTCGGGCATCACTACGCGCCGCGC-3’
GCTGAGCTGAGATGCCTTCCCGTAGGAAGCCCGTAGTGATGCGCGGCGCG
91×91×192
0.00209
100.2
aThe sequences of short DNAs are selected according to the recent SAXS experimental sequences27
to yield normal B-form DNA helices, and contain all the dinucleotide base pairs;44-46
bCS is the strand concentration (in M) which was calculated according to the corresponding
simulation box.
cThe melting temperatures were estimated from the nearest neighbor model based on the measured
thermodynamic parameters.44-46
32
|
1310.0830 | 1 | 1310 | 2013-10-02T20:02:30 | Diffusion, subdiffusion, and trapping of active particles in heterogeneous media | [
"physics.bio-ph",
"cond-mat.soft"
] | We study the transport properties of a system of active particles moving at constant speed in an heterogeneous two-dimensional space. The spatial heterogeneity is modeled by a random distribution of obstacles, which the active particles avoid. Obstacle avoidance is characterized by the particle turning speed $\gamma$. We show, through simulations and analytical calculations, that the mean square displacement of particles exhibits two regimes as function of the density of obstacles $\rho_o$ and $\gamma$. We find that at low values of $\gamma$, particle motion is diffusive and characterized by a diffusion coefficient that displays a minimum at an intermediate obstacle density $\rho_o$. We observe that in high obstacle density regions and for large $\gamma$ values, spontaneous trapping of active particles occurs. We show that such trapping leads to genuine subdiffusive motion of the active particles. We indicate how these findings can be used to fabricate a filter of active particles. | physics.bio-ph | physics |
Diffusion, subdiffusion, and trapping of active particles in heterogeneous media
1Odessa National University, Department for Theoretical Physics, Dvoryanskaya 2, 65026 Odessa, Ukraine
Oleksandr Chepizhko1, 2 and Fernando Peruani2, ∗
2Universit´e Nice Sophia Antipolis, Laboratoire J.A. Dieudonn´e,
UMR 7351 CNRS, Parc Valrose, F-06108 Nice Cedex 02, France
(Dated: April 29, 2014)
We study the transport properties of a system of active particles moving at constant speed in an
heterogeneous two-dimensional space. The spatial heterogeneity is modeled by a random distribution
of obstacles, which the active particles avoid. Obstacle avoidance is characterized by the particle
turning speed γ. We show, through simulations and analytical calculations, that the mean square
displacement of particles exhibits two regimes as function of the density of obstacles ρo and γ. We
find that at low values of γ, particle motion is diffusive and characterized by a diffusion coefficient
that displays a minimum at an intermediate obstacle density ρo. We observe that in high obstacle
density regions and for large γ values, spontaneous trapping of active particles occurs. We show
that such trapping leads to genuine subdiffusive motion of the active particles. We indicate how
these findings can be used to fabricate a filter of active particles.
PACS numbers: 05.40.Jc, 05.40.Fb, 87.17.Jj
Locomotion patterns are of prime importance for the
survival of most organisms at all scales, ranging from
bacteria to birds, and often involving complex processes
that require energy consumption: i.e. the active motion
of the organism [1, 2]. The characterization and study
of these patterns have a long tradition [1, 2] and the
experimental observation of subdiffusion, diffusion, and
superdiffusion has motivated the development of power-
ful theoretical tools [3].
It is only in recent years that
the (thermodynamical) non-equilibrium nature of these
active patterns has been exploited, leading to the study
of the so-called active particle systems [4]. Exciting non-
equilibrium features have been reported in both, inter-
acting as well as non-interacting active particle systems.
For instance, large-scale collective motion and giant num-
ber fluctuations have been found in interacting active
particle systems [5–8].
In non-interacting active parti-
cle systems, the presence of active fluctuations leads to
complex, non-equilibrium transients in the particle mean
square displacement [9, 10] and anomalous velocity dis-
tributions [11], and the lack of momentum conservation
induces non-classical particle-wall interactions, which al-
lows, for instance, the rectification of particle motion [12–
16].
The study of active particle systems has recently wit-
nessed the emergence of a promising new direction: the
design and construction of biomimetic, artificial active
particles. The directed driving is usually obtained by
fabricating asymmetric particles that possess two dis-
tinct friction coefficients [18–20], light absorption coeffi-
cients [21–24], or catalytic properties [10, 25–29] depend-
ing on whether energy injection is done through vibra-
tion, light emission, or chemical reaction, respectively.
One of the most prominent features of these artificial ac-
tive particles is that their motion is characterized by a
diffusion coefficient remarkably larger than the one ob-
(a)
Dx
10
(b)
Dx
10
γ=1
simulations
LD approx.
HD approx.
0.01
0.1
γ=5
1
Trapping
Subdiffusion
simulations
LD approx.
5
η=0.1
γ
1
1.0
0.5
η=0.01
ρ
o
Trapping
Subdiffusion
η=0.01
(c)
5
4
γ
3
2
1
0.01
ρ
o
0.1
1
0
0.5
ρ
o
1
No Trapping
Diffusion
FIG. 1: Diffusive and subdiffusive regimes. (a) For low values
of the turning speed γ, the motion is diffusive and charac-
terized by a diffusion coefficient Dx that exhibits a minimum
with the obstacle density ρo, as expected by combining the
low-density (LD) and high-density (HD) approximation. See
Eq. (9), (11) and text.
(b) For large values of γ, diffusive
motion occurs at low ρo values only, while for large ρo values,
particle motion becomes subdiffusive. (c) The boundary be-
tween the diffusive and subdiffusive regime is given, for a fixed
noise η, by ρo and γ, see inset and [34]. Parameters: R = 1,
L = 100, Np = 104, and η = 0.01 (inset in (c), η = 0.1).
tained using symmetric particles [10].
The rapidly expanding study of active particles has fo-
cused so far almost exclusively, theoretically as well as
experimentally, on the statistical description of particle
motion in idealized, homogeneous spaces. However, the
great majority of natural active particle systems takes
place, in the wild, in heterogeneous media: from active
transport inside the cell, which occurs in a space that is
filled by organelles and vesicles [30], to bacterial motion,
which takes place in highly heterogeneous environments
such as the soil or complex tissues such as in the gastroin-
testinal tract [31]. While diffusion in random media is a
well studied subject [32, 33], the impact that an hetero-
geneous medium may have on the locomotion patterns of
active particles remains poorly explored.
We address this fundamental problem by using a sim-
ple model in which the active organisms move at constant
speed in an heterogeneous two-dimensional space, where
the heterogeneity is given by a random distribution of
obstacles. An “obstacle” may represent the source of a
repellent chemical, a light gradient, a burning spot in a
forest, or whatever threat that makes our (self-propelled)
organisms to move away from it once the danger has
been sensed; with obstacle avoidance characterized by
a (maximum) turning speed γ. Our analysis shows that
the same evolution equations (behavioral rules) lead to
very different locomotion patterns at low and high den-
sity of obstacles. In the dilute obstacle scenario, there
is no conflicting information and organisms can easily
move away from the undesirable area they find in their
way. On the other hand, when we stress the environ-
mental conditions, such that organisms sense several re-
pellent sources simultaneously, the processing of the in-
formation is no longer simple. Organisms compute the
local obstacle density gradient and use this information
to move away from higher obstacle densities. Since the
distribution of obstacles is random, as the overall obsta-
cle density increases, this task becomes increasingly more
difficult. As result of this, no strategy guaranties how to
escape away from obstacles and the organisms behave
more and more as if there were no obstacles in the sys-
tem. For low γ values, we find that the above described
change of behavior is reflected by the minimum exhibited
by the diffusion coefficient at intermediate obstacle den-
sities ρo, Fig. 1(a). For large γ values, particle motion
is diffusive at small densities ρo, while for large enough
densities a new phenomenon emerges: spontaneous trap-
ping of particles, Fig. 1(b). These traps are closed or-
bits found by the particles in a landscape of obstacles,
Fig. 2. The time particles spend in these orbits is heavy-
tailed distributed, and particle motion is genuinely (i.e.,
asymptotically) subdiffusive. The boundary between the
diffusive and subdiffusive regime depends on γ and ρo as
illustrated in Fig. 1(c).
Our results open a new route to control active particle
systems. For instance, the appearance of spontaneous
trapping as a dynamical phenomenon that depends on
the intrinsic properties of the particles allows us to design
a generic filter of active particles.
Model definition.– We consider a continuum time
model for Np self-propelled particles moving in a two-
dimensional space of linear size L where No obstacles are
placed at random [35]. Boundary conditions are periodic.
In the over-damped limit, the equations of motion of the
i-th particle are given by:
xi = v0V(θi)
θi = h(xi) + ηξi(t) ,
(1)
(2)
where the dot denotes temporal derivative, xi corre-
sponds to the position of the i-th particle, and θi to its
moving direction. The function h(xi) represents the in-
teraction with obstacles and its definition is given by:
2
h(xi) =( γ
0
n(xi)PΩi sin(αk,i − θi)
if n(xi) > 0
if n(xi) = 0 ,
(3)
where the sum runs over all neighboring obstacles Ωi such
that 0 < xi − yk < R, with yk the position of the k-th
obstacle, and the term αk,i the angle, in polar coordi-
nates, of the vector xi − yk. The term n(xi) denotes the
number of obstacles located at a distance less or equal
than R from xi.
In Eq. (1), v0 is the active particle
speed and V(θ) ≡ (cos(θ), sin(θ))T . The additive white
noise in Eq. (2) is characterized by an amplitude η and
obeys hξi(t)i = 0 and hξi(t)ξj (t′)i = δi,jδ(t − t′), which
leads to an angular diffusion Dθ = η2/2. Notice that for
γ = 0, equations (1) and (2) define a system of persistent
random walkers characterized by a diffusion coefficient
Dxo = v2
0/(2Dθ), see [1, 2, 9].
Continuum description.– We look for a coarse-grain
description of the system in terms of the concentration
p(x, θ, t) of particles at position x and orientation θ at
time t. The evolution of p(x, θ, t) obeys [37]:
∂tp + v0∇. [V(θ)p] = Dθ∂θθp + F [p(x, θ, t), ρo(x)] , (4)
where Dθ is the angular diffusion as defined above, and
F [p(x, θ, t), ρo(x)] represents the interaction of the self-
propelled particles with the obstacles. The term ρo(x)
refers to the obstacle density at position x [38]. Here,
we discuss two clear limits where F [p(x, θ, t), ρo(x)] can
be specified. We refer to these limits as the low-density
(LD) and high-density (HD) (obstacle) approximation.
Low-density approximation.– We consider that the ac-
tive particles move most of the time freely, bumping
into obstacles only occasionally. More specifically, we
assume that η, ρo << 1 and approximate the interaction
with obstacles, for time-scales much larger than 2R/v0,
as sudden changes in the moving direction of the par-
ticle. Let T (θ, θ′; x) be the rate at which a particle at
position x and moving in direction θ turns into direc-
tion θ′. To compute T (θ, θ′; x) we need to estimate the
frequency at which particles encounter obstacles as well
as the scattered angle after each obstacle interaction. If
D−1
, we can approximate particle motion,
in between successive encounters with obstacles, as bal-
listic. In this limit, the obstacle encounter rate can be
estimated as λ(ρo) ≈ voρoσo, where σo = 2R is the
associated scattering cross section. The absence of the
classical constants of motion such as angular momentum
and mechanical energy prevents us from deriving an ef-
fective potential formalism from which to estimate the
scattered angle. To simplify the calculations we approx-
imate the scattered angle distribution by a simple top-
hat functional form. Putting all this together, we express
θ v0 >> ρ−1/2
o
T (θ, θ′; x) ≃ λ(ρo)T (θ, θ′) ≈ [λ(ρo)/(2ǫθ)] Θ(ǫθ −θ − θ′)
and express F as:
F [p] = −λ(ρo) p(x, θ, t) +Z 2π
dθ′T (θ, θ′)p(x, θ′, t)
0
∂θθp ,
(5)
λ(ρo)ǫ2
θ
6
≈
where ǫθ is numerically obtained from the study of the
scattering process. Expression (5) allows us to rewrite
the r.h.s. of Eq. (4) as Dθ∂θθp, where Dθ is defined
as Dθ = Dθ + λ(ρo)ǫ2
θ/6. By performing a moment
expansion of Eq. (4), where we define ρ(x, t) = R dθp,
Px(x, t) = R dθ cos(θ)p, Py(x, t) = R dθ sin(θ)p, and
Qs(x, t) = R dθ sin(2θ)p, and Qc(x, t) = R dθ cos(2θ)p,
we arrive to the following set of equations:
(8)
(7)
(6)
∂tρ = −v0∇.P
∂tPx = −
∂tPy = −
v0
2 ∇. [Qc + ρ, Qs] − DθPx
v0
2 ∇. [Qs, ρ − Qc] − DθPy ,
It can be
where we assumed that ∂tQc = ∂tQs = 0.
shown that the temporal evolution of Qc and Qs is faster
than the one of Px and Py, which in turn is faster than
the one for ρ. Since we are interested in the long time
behavior of ρ(x, t), and there is no induced order, we
take Qc = Qs = 0 and use the fast relaxation of Eqs. (7)
and (8) to express Px and Py as slave functions of ρ and
its derivatives [39]. This procedure leads to the following
asymptotic equation for ρ(x, t):
∂tρ = ∇.(cid:20) v2
2 Dθ ∇ρ(cid:21) .
0
(9)
From Eq. (9), it is evident that the spatial diffusion coeffi-
cient Dx takes the form: Dx = v2
0/ [2 (Dθ + Λ0ρo)], with
Λ0 = v0σ0ǫ2
θ/6. Notice that Dx is a decreasing function
of ρo [40].
High-density approximation.– At large obstacle densi-
ties, particles always sense the presence of several obsta-
cles around them. This means that we cannot think of
collisions as rare sudden jumps in the moving direction.
Thus, we replace Eq. (5) with a direct, local, coarse-
grained expression for the interactions. This means that
we leave the Boltzmann-like for the Fokker-Planck ap-
proach where the interaction with obstacles is expressed
by F = ∂θ [Ip(x, θ, t)]. The term I represents the (aver-
age) interaction felt by a particle at x moving in direction
θ, which takes the form:
I =
γ
n(x)Xj
sin(θ − αj) =
γΓ(x)
n(x)
sin(θ − ψ(x)) , (10)
where j is an index that runs over all neighboring obsta-
cles, αj is the polar angle associated to the vector x− yj,
3
and Γ(x) and ψ(x) are the modulus and phase, respec-
Eq. (10) by its average and use the fact that it repre-
sents a sum of n random vectors of magnitude 1 in the
tively, of Pj exp(ı(αj)), see [41]. We now approximate
complex plane, to express I ∼ sin(θ − ψ(x))/√n, where
n ≈ πR2ρo. Inserting this approximated expression into
Eq. (4) and performing the moment expansion and ap-
proximations that led us from Eqs. (6)-(8) to Eq. (9), we
arrive to:
∂tρ =
=
v2
0
2Dθ ∇2ρ −
v2
0
2Dθ ∇2ρ −
γv0
2DθR√πρ0 ∇. [(cos(ψ), sin(ψ))ρ]
2DθR√πρ0 ∇.(cid:20) ρ∇ρo(x)
∇ρo(x)(cid:21) ,
γv0
(11)
have
approximated
vector
field
where we
the
If we replace
(cos(ψ), sin(ψ)) ∼ ∇ρo(x)/∇ρo(x).
our current definition of ρ by a local average over a
volume of linear dimensions much larger than R and
look for the long-time dynamics of this redefined density,
by applying homogenization techniques, we expect
to recover a diffusive behavior with a new effective
diffusion, whose explicit form depends on the statistical
properties of the random field ρo(x) and is proportional
to the square of the constant in front of the convective
term. While according to Eq. (9) (LD approx.), Dx → 0
as ρo → ∞, Eq. (11) (i.e., the HD approx.) indicates
that in the limit of ρo → ∞, Dx → Dx0, where Dx0 is
the diffusion coefficient in absence of obstacles defined
above. These two results necessarily imply the existence
of a minimum in the spatial diffusion coefficient Dx as ρo
is increased from 0 to ∞. Moreover, this minimum has
to be located at the crossover between the LD and HD
approximation, which can be roughly estimated to occur
at ρc ∼ 1/(πR2), around which, Dx ∼ 1/ [ρc − Λ1ρo],
with Λ1 a constant. All these findings are confirmed by
particle simulations as shown in Fig. (1).
Trapping.– Eq. (11) indicates that the vector field
∇ρo(x) governs the long-term dynamics at high obsta-
In particular, the random distribu-
cle concentrations.
tion of obstacles, together with the compressible nature
of ρ(x, t), may result in the spontaneous formation of ac-
tive particle sinks. These topological defects, which we
refer to as “traps”, are indeed observed in simulations for
large values of γ, Fig. 2(a) and (b). Inside traps particles
form vortex-like patterns. The average time hτTi spent
by a particle inside a trap depends on the precise con-
figuration of the obstacles that form the trap. We find
that the presence of traps can lead to a genuine subdif-
fusive behavior, with particles exhibiting a mean-square
displacement σ2(t) = hx2(t)i that grows slower than t.
To test this observation, let us assume that particle mo-
tion can be conceived as a random walk across a two
dimensional array of traps such that σ2(t) ∝ nJ (t) where
nJ (t) represents the number of jumps from trap to trap
the random walker performs during t. To estimate nJ (t),
(a)
(b)
(c)
(d)
4
ρ
A(x)
ρ
B(x)
initial condition
steady state
x
x
FIG. 2: Trapping and filtering. For large values of γ and ρo, spontaneous trapping of particle occurs, see (a) where γ = 5,
ρo = 0.5, η = 0.01, and L = 30. Obstacles are indicated by red dots while black arrows correspond to SPPs. Inside traps
particles self-organize into vortex-like structures, (b) [42]. Spontaneous trapping leads to subdiffusion as indicated in Fig. 3
and can be used to design filters. Starting with an initial condition as in (c), with two types of particles, characterized by
γA = 5 and γB = 1, we quickly arrive to a steady state where A particles are confined to the left half of the box, while B
particles diffusive freely over the system. The lower panels in (c) and (d) indicate the density projected on the x-axis of A and
B particles.
(a)
σ2
(t)
t/ln(t)
η=0.1; L=100
η=0.1; L=200
η=0.05; L=100
η=0.05; L=200
10-2
To=103
To=105
τ
T
103
(b)
P(τΤ)
10-3
104 10-6
103
<τΤ>
(c)
102
103
t
106
102
103
104
To
105101
FIG. 3: Genuine subdiffusive behavior.
(a) scaling of the
mean square displacement σ2(t) with t for two system sizes,
with Np/L2 = 1, ρo = 0.5 and γ = 5. Notice that the growth
of σ2(t) is even slower than t/ ln(t). (b) the distribution of
trapping times P (τT ) is power-law distributed for long enough
To. (c) the average waiting time hτT i is an increasing function
of To. The red dashed curve corresponds to a fit ∝ ln(To).
Measurements in (b) and (c) were performed on a particular
trap for η = 0.1.
we study the distribution to trapping times P (τT ) of a
given trap, see Fig. 3(b). Subdiffusion can only occur if
P (τT ) is asymptotically power-law distributed and such
that hτTi grows with the observation time To as ln(To)
or faster [43], see Fig. 3(b) and (c). Within this sim-
plified picture, the behavior of hτTi shown in Figs. 3(c)
suggests that σ2(t) ∝ nJ (t) ∼ t/ ln(t). By taking t → ∞,
we expect σ2(t)/[t/ ln(t)] to approach a constant value.
Fig. 3(a) clearly shows that σ2(t) is slower than t/ ln(t).
Arguably, this is due to the fact that once a particle es-
capes from a trap, typically it performs a small excursion
before being reabsorbed by the same trap.
Discussion.– The spontaneous trapping of particles de-
pends not only on ρo but also on γ, which is an intrinsic
property of the particle. This means that given the same
spatial environment, two particle types, say characterized
by γA and γB, will respond differently. We can make use
of this fact to fabricate a simple and cheap filter as indi-
cated in Fig. 2(c) and (d). Notice that trapping of one
of the particle types is required to obtain this effect. A
different diffusion coefficient for A and B particles on
the left half of the box does not suffice to induce higher
concentration of one particle type to the left.
Trapping, rectification, and sorting have been reported
for a particular kind of active particles: chiral, i.e. circu-
larly moving particles [44, 45]. By placing L-shaped ob-
stacles on a regular lattice, the motion of such particles
can be rectified [45], while elongated obstacles arranged
in flower-like patterns can be used to selectively trap ei-
ther levogyre or dextrogyre particles [44]. For non-chiral
active particle, trapping and rectification can be achieved
by using V-shape objects. Kaiser et al. in [46, 47] showed
that self-propelled rods can be trapped by placing V-
shape objects. These traps provide a geometrical con-
strain to the active particles that end up being blocked
in the V-shape devices. On the other hand, by arranging
in line V-shape objects, but with their tips open, rectifi-
cation of particle motion can be achieved [12, 15]. These
V-shape objects, either with their tip closed or opened,
cannot be used to produce a filter of active particles.
Closed V-shape objects collect, by imposing geometric
obstruction, any kind of self-propelled particle, while in
opened V-shape objects clogging of large size particles
necessary occurs, preventing particle flow. Notice that
the novel trapping and sorting mechanism reported here,
which is based on the obstacle avoidance response time, is
generic and should apply to all kind of active particles, in-
cluding interacting, non-interacting, chiral or non-chiral
active particles.
Finally, it is important to mention that genuine sub-
diffusion occurs for fixed obstacles only. For slowly dif-
fusing obstacles, i.e. for a (slow) dynamic environment,
the asymptotic behavior of the active particles is diffu-
sive. The low and high density approximations, given
by Eq. (9) and (11), provide a reasonable description
of particle motion even for dynamical environments as
long as obstacle diffusion remains substantially smaller
than the active motion. Furthermore, trapping and sub-
diffusive behavior is also observed for interacting active
particles as those studied in [36] for values of the interac-
tion strength significantly smaller than those associated
to obstacle avoidance [42].
Numerical simulations have been performed at the
‘Mesocentre SIGAMM’ machine, hosted by Observatoire
de la Cote d’Azur. We thank F. Delarue and R. Soto
for valuable comments on the manuscript and the Fed.
Doeblin for partial financial support.
∗ Electronic address: [email protected]
[1] H. Berg, Random walks in biology (Princeton University
Press, Princeton, 1983).
[2] A. Okubo and S. Levin, Diffusion and ecological problems
(Springer, New York, 1980).
[3] R. Metzler and J. Klafter, Physics Reports 339, 1 (2000).
[4] P. Romanczuk et al., Eur. Phys. J. Special Topics 202,
1 (2012).
5
[24] J. Palacci et al., Science 339, 936 (2013).
[25] W. Paxton et al., J. Am. Chem. Soc. 126, 13424 (2004).
[26] N. Mano and A. Heller, J. Am. Chem. Soc. 127, 11574
(2005).
[27] G. Ruckner and R. Kapral, Phys. Rev. Lett. 98, 150603
(2007).
[28] J. Howse et al., Phys. Rev. Lett. 99, 048102 (2007).
[29] R. Golestanian, T.B. Liverpool and A. Ajdari, Phys. Rev.
Lett. 94, 220801 (2005).
[30] B. Alberts et al, Molecular biology of the cell (Garland
publishing, 1994).
[31] M. Dworkin, Myxobacteria II (Amer Society for Micro-
biology, 1993).
[32] J.-P. Bouchaud and A. Geoges, Physics Reports 195, 127
(1990).
[33] H. Berry and H. Chat´e, arXiv:1103.2206 (2011).
[34] The boundary has been numerically estimated by:
i)
Studying Dx(t → ∞) = hx2(t)i/4t – for subdiffusive par-
ticles Dx(t) never saturates. ii) By measuring the asymp-
totic fraction of particle on the left half of the box in sys-
tems prepared as indicated in Fig. 2(c) and (d) – values
of this quantity above 0.5 can only occur due to the pres-
ence of traps. Both methods, i) and ii), led to the same
boundary estimate.
[35] Collective effects of an interacting version of this model
were studied in [36]. Here, we focus on individual loco-
motion patterns.
[36] O. Chepizhko, E.G. Altmann, and F. Peruani, Phys. Rev.
[5] T. Vicsek and A. Zafeiris, Physics Reports 517, 71
Lett. 110, 238101 (2013).
(2012).
[37] C. W. Gardiner, Handbook of Stochastic Methods
[6] S. Ramaswamy, Annual Review of Condensed Matter
(Springer, Heildelberg, 2004).
Physics 1, 323 (2010).
[38] We reserve the symbol ρo, without x dependency, to the
[7] F. Peruani et al., Phys. Rev. Lett. 108, 098102 (2012).
[8] F. Ginelli et al., Phys. Rev. Lett. 104, 184502 (2010).
[9] F. Peruani and L. Morelli, Phys. Rev. Lett. 99, 010602
(2007).
[10] R. Golestanian, Phys. Rev. Lett. 102, 188305 (2009).
[11] P. Romanczuk and L. Schimansky-Geier, Phys. Rev.
Lett. 106, 230601 (2011).
[12] P. Galajda et al., J. Bacterial. 189, 8704 (2007).
[13] H. H. Wensink and H. Lowen, Phys. Rev. E 78, 031409
(2008).
[14] M. Wand et al., Phys. Rev. Lett. 101, 018102 (2008).
[15] J. Tailleur and M. Cates, Europhys. Lett. 86, 60002
(2009).
[16] P. Radtke and L. Schimansky-Geier, Phys. Rev. E 85,
051110 (2012).
[17] Kudrolli et al., Phys. Rev. Lett. 100, 058001 (2008).
[18] A. Kudrolli, G. Lumay, D. Volfson, and L. Tsimring,
Phys. Rev. E 74, 030904(R) (2006).
[19] J. Deseigne, O. Dauchot, and H. Chat´e, Phys. Rev. Lett.
global obstacle density.
[39] Alternatively, multi-scaling analysis can be used to go
from Eqs. (4) and (5) to Eq. (9).
[40] There are qualitative differences with the Lorentz gas
model, see [43]. Firstly, in our model angular diffusion
acts on the moving particles in between collisions. Sec-
ondly, the scattering process is not an instantaneous pro-
cess and leads, given the absence of conserved quantities,
to a non bijective relationship between the impact pa-
rameter and the scattered angle.
[41] Y. Kuramoto, in International Problems in Theoretical
Physics, edited by H. Araki, Lecture Notes in Physics
(Springer, New York, 1975).
[42] See supplemental material at XXXX for movies.
[43] P. L. Krapivsky, S. Redner, and E. Ben-Naim, A kinetic
view of statistical physics (Cambridge University Press,
New York, 2010).
[44] M. Mijalkov and G. Volpe, Soft Matter 9, 6376 (2013).
[45] C. Reichhardt, C.J. Olson Reichhardt, arXiv:1307.0755
105, 098001 (2010).
(2013).
[20] C. A. Weber et al., Phys. Rev. Lett. 110, 208001 (2013).
[21] H.-R. Jiang, N. Yoshinaga, and M. Sano, Phys. Rev. Lett.
[46] A. Kaiser et al., Phys. Rev. Lett. 108, 268307 (2012).
[47] A. Kaiser, K. Popowa, H.H. Wensink and H. Lowen,
105, 268302 (2010).
Phys. Rev. E 88, 022311 (2013).
[22] R. Golestanian, Phys. Rev. Lett. 108, 038303 (2012).
[23] I. Theurkauff et al., Phys. Rev. Lett. 108, 268303 (2012).
|
1006.3959 | 2 | 1006 | 2011-07-23T16:00:41 | Molecular Communication Using Brownian Motion with Drift | [
"physics.bio-ph",
"cond-mat.mes-hall",
"cond-mat.soft",
"cs.IT",
"cs.IT"
] | Inspired by biological communication systems, molecular communication has been proposed as a viable scheme to communicate between nano-sized devices separated by a very short distance. Here, molecules are released by the transmitter into the medium, which are then sensed by the receiver. This paper develops a preliminary version of such a communication system focusing on the release of either one or two molecules into a fluid medium with drift. We analyze the mutual information between transmitter and the receiver when information is encoded in the time of release of the molecule. Simplifying assumptions are required in order to calculate the mutual information, and theoretical results are provided to show that these calculations are upper bounds on the true mutual information. Furthermore, optimized degree distributions are provided, which suggest transmission strategies for a variety of drift velocities. | physics.bio-ph | physics | Molecular Communication Using Brownian Motion
with Drift
Sachin Kadloor, Raviraj S. Adve, and Andrew W. Eckford
1
1
1
0
2
l
u
J
3
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
2
v
9
5
9
3
.
6
0
0
1
:
v
i
X
r
a
Abstract
Inspired by biological communication systems, molecular communication has been proposed as a viable scheme
to communicate between nano-sized devices separated by a very short distance. Here, molecules are released by
the transmitter into the medium, which are then sensed by the receiver. This paper develops a preliminary version
of such a communication system focusing on the release of either one or two molecules into a fluid medium with
drift. We analyze the mutual information between transmitter and the receiver when information is encoded in the
time of release of the molecule. Simplifying assumptions are required in order to calculate the mutual information,
and theoretical results are provided to show that these calculations are upper bounds on the true mutual information.
Furthermore, optimized degree distributions are provided, which suggest transmission strategies for a variety of
drift velocities.
I. INTRODUCTION
Communications research has almost exclusively focused on systems based on electromagnetic prop-
agation. However, at scales considered in nano-technology, it is not clear that these methods are viable.
Inspired by the chemical-exchange communication performed by biological cells, this paper consid-
ers molecular communication [1], in which information is transmitted by an exchange of molecules.
Specifically we consider the propagation of individual molecules between closely spaced transmitters and
receivers, both immersed in a fluid medium. The transmitter encodes a message in the pattern of release
of the molecules into the medium; these molecules then propagate to the receiver where they are sensed.
The receiver then attempts to recover the message by observing the pattern of the received molecules.
It is well known that microorganisms exchange information by molecular communication, with quorum
sensing [2] as but one example, where bacteria exchange chemical messages to estimate the local popula-
tion of their species. The biological literature on molecular communication is vast, but there has been much
recent work concerning these systems as engineered forms of communication. Several recent papers have
described the design and implementation of engineered molecular communication systems, using methods
such as: exchanging arbitrary molecules using Brownian motion in free space [3]; exploiting gap junctions
between cells to exchange calcium ions [4], [5]; and using microtubules and molecular motors to actively
drive molecules to their destination [6], [7]. A comprehensive overview of the molecular communication
system is also given by [8], [9] and the references therein.
Given this engineering interest, it is useful to explore the theoretical capabilities of molecular commu-
nication systems. To the authors’ knowledge, the earliest effort towards information-theoretic analysis of
these channels was given in [10], which examined information flow in continuous diffusion. In [11], [12],
physical models and achievable bounds on information rate were provided for diffusion-based systems.
Manuscript received Jan. 12, 2010; revised Jun. 23, 2011; accepted Jul. 19, 2011. The associate editor coordinating the review of this
paper and approving it for publication was Dr. M. Hughes. The material in this paper was presented in part at the International Conference
of Computer Communications and Networks (ICCCN), San Fransisco, CA, 2009.
S. Kadloor was with The Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto. He is now with
the Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, 1308 West Main Street, Urbana, Illinois, USA 61801-2307.
Email: [email protected]
R. S. Adve is with The Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, 10 King’s
College Road, Toronto, Ontario, Canada M5S 3G4. Email: [email protected]
A. W. Eckford is with the Department of Computer Science and Engineering, York University, 4700 Keele Street, Toronto, Ontario, Canada
M3J 1P3. Email: [email protected]
2
Information rates were provided in [13], [14] for the case where the receiver chemically “reacts” with the
molecules and form “complexes”. In [15], it was shown that the additive white Gaussian noise (AWGN) is
appropriate for diffusion-based counting channels. Information-theoretic results have also been obtained
for specific systems, such as propagation along microtubules [16], [17] and continuous diffusion [18].
All these studies indicate that useful information rates can be obtained, although much lower per unit
time than in electrical communication; this is not surprising, since chemical processes are far slower than
electrical processes. It is worth pointing out that these results build on theoretical work in Poisson and
queue-timing channels [19], [20], which is an active area of research in information theory.
In any communication system, the potential rate of communication is determined by the characteristics
of the channel. We consider molecular propagation in a fluid medium, governed by Brownian motion and,
potentially, a mean drift velocity. Our model is therefore applicable to communications in, e.g., a blood
vessel. This drift velocity is a key difference between our work and [11], [12], which considered a purely
diffusion channel. Furthermore, we consider two cases - a single and two molecules being released. In
this regard, it is worth emphasizing the preliminary and conceptual nature of this work. The long-term
goal of this work is to understand the role of both timing and the number of molecules (‘amplitude’).
Thus, the contributions of this paper include:
• Calculation and optimization, under some simplifying assumptions, of mutual information in Brow-
nian motion with drift, where the transmitter uses pulse-position modulation;
• Optimization of the degree distributions related to two transmit molecules; and
• Demonstration (via theoretical results) that our simplified mutual information calculation is an upper
bound on the true mutual information of any practical implementation of this system.
Our optimized degree distributions reveal interesting features of these channels, suggesting transmission
strategies for system designers.
The paper is organized as follows. In Section II we describe the system under consideration, in which
the propagation of the molecule is analyzed and the probability distribution function of the absorption
time is derived. In Section III, we characterize the maximum information transfer per molecule, for the
case where information is encoded in the time of release of the molecule, and the case of two molecules.
In Section IV, numerical and theoretical results arising from these models (including optimized degree
distributions) are presented. The paper wraps up with some conclusions and suggestions for future work.
A. Communication model
II. SYSTEM MODEL
The communication model we consider is shown in Figure 1. The subsystems which make up the
molecular communication system are:
1) Transmitter. The transmitter is a source of identical molecules. It encodes a message in the time
of dispersal of these molecules. We will assume that the transmitter can control precisely the time
of dispersal of these molecules but does not influence the propagation of these molecules once
dispersed.
2) Propagation medium. The molecules propagate between transmitter and receiver in a fluid medium.
The propagation is modeled as Brownian motion, and is characterized by two parameters: the drift
velocity and the diffusion constant. These in turn depend on the physical properties of the fluid
medium. The trajectories of different molecules are assumed to be independent of one another.
3) Receiver. In this paper, the propagation of the molecule is assumed to be one dimensional. When it
arrives at the receiver, the dispersed molecule is absorbed by the receiver and is removed from the
medium. The receiver makes an accurate measurement of the time when it absorbs the molecule.
It uses this information to determine the message sent by the transmitter.
4) Transmission of information. The transmitter can encode information in either the time of dispersal
of the molecules, or the number of molecules it disperses, or both.
3
y
x
Receiver
Transmiter
Molecules
Fig. 1. An abstract model of the molecular communication system. One or more molecules are released by the transmitter. These molecules
then travel through the fluid medium to the receiver, which absorbs them upon reception. If all the molecules are identical, then information
is conveyed from transmitter to the receiver only through the times at which the molecules are released.
The motion of the dispersed molecule is affected by Brownian motion; the diffusion process is therefore
probabilistic and, in turn, the propagation time to the receiver is random. Even in the absence of any im-
perfection in the implementation of a molecular communication system, this uncertainty in the propagation
time limits the maximum information rate per molecule. In this paper, we study the maximum information
per molecule that the transmitter can convey to the receiver, for a certain velocity and diffusion in the
fluid medium. Before proceeding to do so, we need to characterize the propagation of the molecule in
the medium.
B. Diffusion via Brownian motion
Consider the discrete-time, discrete-space propagation model in Figure 2(a). Let X(n) denote the
position of the particle at time n. Let PX (x, n; xo, no) denote the probability mass function (pmf) of
the position of the particle at time n, given that it was dispersed in the fluid medium at position xo at
time no. Assume that the fluid medium is static, and so the particle disperses in either of the directions
with equal probability. If p is the probability that the particle moves from position x to position x + l in
one time unit, and q is the probability that it moves from position x to x − l, then this situation is the
case when p = q = 0.5. It is easy to see that PX (x, n; xo, no) obeys the equation
PX(x, n + 1; xo, no) =
1
2
PX(x − l, n; xo, no) +
1
2
PX(x + l, n; xo, no),
(1)
which states that if a particle at time n + 1 is at position x, then at time n, it should have been at position
x − l or x + l, where l is the distance between two slices of space. This formulation of Brownian motion
is analogous to a Wiener process, where distinct increments of the motion are independent from each
other.
Equation (1) can be re-written as
PX(x, n + 1; xo, no) − PX(x, n; xo, no)
=
1
(PX(x − l, n; xo, no) − PX(x, n; xo, no)) +
2
l2
2 (cid:18) 1
l (cid:18)PX (x − l, n; xo, no) − PX(x, n; xo, no)
1
2
(PX(x + l, n; xo, no) − PX(x, n; xo, no))
(cid:19)(cid:19) .
(2)
=
l
4
(a) Modeling the motion of the particle as a one dimensional random
walk
l
e
c
i
t
r
a
p
e
h
t
f
o
n
o
i
t
i
s
o
p
e
h
t
f
o
i
t
e
a
n
d
r
o
o
c
−
x
50
40
30
20
10
0
−10
0
200
400
Time
600
800
1000
(b) Sample paths of six particles in the same fluid medium, three released
at t = 0, three at t = 400.
Fig. 2.
If the size of the receiver is several orders greater than the size of the molecule, and if the velocity of the fluid in the ‘y axis’ is
negligible compared to the velocity of the fluid along ‘x axis’ (Refer Fig. 1), then, one can ignore the y-coordinate of the position of the
molecule and consider only the x-coordinate. The position of the molecule along the x axis is modeled as a Markov chain, specifically, as
a one dimensional random walk. The bias of the walk (the values of p and q) depend on the velocity of the fluid medium along the x-axis.
When n ≫ 1 and x ≫ l, the difference equation becomes a continuous time differential equation, yielding
a probability distribution function (pdf) for the position of the particle, given by,
∂
∂n
PX(x, n; xo, no) =
l2
2
∂2
∂x2 PX(x, n; xo, no).
(3)
Now, considering a continuous time Brownian motion X(t), the probability density function of the position
of the particle can be modeled by the diffusion equation
∂2
∂x2 PX(x, t; xo, to),
PX(x, t; xo, to) = D
∂
∂t
(4)
where D = l2/2 is the diffusion constant, whose value is dependent on the viscosity of the fluid medium.
Note that the above equation characterizes only the ‘x-coordinate’ of the position of the molecule. Solutions
to this equation are well known.
5
Equation (4) characterizes the motion X(t) of the particle in a macroscopically static medium. The
more general and useful case is that of a fluid medium is in motion with a mean drift velocity v. Consider
a frame of reference which is moving with the same velocity. In this frame, the fluid medium is static
and hence the diffusion of the particle should obey Equation (4). Let
x′ = x + vt,
t′ = t
be the new coordinate system, and without loss of generality, assume to = 0. Let
PX(x, t; xo, 0) = P ′
X′ (x′, t′; xo, 0),
then
X′ (x′, t′; xo, 0).
In the static frame of reference, the differential equation can be written as
X′ (x′, t′; xo, 0) = D
P ′
∂2
∂x′2 P ′
∂
∂t′
∂
∂t
X′ (x′, t′; xo, 0)
P ′
D
∂
∂x′ (cid:18)(cid:18) ∂x
∂x′
∂t
∂t′
∂
∂x
+
+
∂
∂x
∂t
∂x′
∂x
∂t′
=
X′ (x′, t′; xo, 0)
P ′
∂
X′ (x′, t; xo, 0)(cid:19) ,
∂t(cid:19) P ′
∂x(cid:19) PX(x, t; xo, 0).
∂
which simplifies to
∂
∂t
PX(x, t; xo, 0) =(cid:18)D
∂2
∂x2 + v
Assume that there is no absorbing boundary (receiver) and that the fluid medium extends from −∞
to +∞. The probability density function of the location of the particle can be obtained by solving the
differential Equation (6) with boundary conditions PX(x, 0; xo, 0) = δ(x− xo) and PX(±∞, t; xo, 0) = 0.
The solution to (6) is given by [21]
(5)
(6)
(7)
PX(x, t; 0, 0) =
1
√4πDt
exp(cid:18)−
(x − vt)2
4Dt (cid:19) .
Equation (7) states that, for every t, the probability density function (pdf) is a Gaussian centered at vt
with variance 2Dt. As expected, the expected location of the particle drifts along the direction of flow
of the fluid medium with velocity vt. Figure 3 plots P (x, t) for the case when v = 3 and D = 0.3.
Furthermore, for any transmitter point ζ and transmit time t0, we have that
PX(x, t; ζ, t0) =
1
p4πD(t − t0)
exp(cid:18)−
((x − ζ) − v(t − t0))2
4D(t − t0)
(cid:19) .
(8)
As expected, Brownian motion X(t) satisfying (7)-(8) is a Wiener process with drift.
Now, consider the case when there is an absorbing surface (receiver) at x = 0. The particle is absorbed
and is removed from the system when it hits the absorbing surface. For such a system, to solve for
PX(x, t;−ζ, 0)), we need to solve the differential equation in (6) with the following boundary conditions.
• For x < 0, PX(x, 0;−ζ, 0) = δ(x+ζ). The probability density function has a physical interpretation
• PX(−∞, t;−ζ, 0) = 0,
• PX(0, t;−ζ, 0) = 0,
The solution to the differential equation can be computed using the method of images, it is given by:
only for x < 0. In this region, we require it to be a delta function at t = 0 centered at x = −ζ.
∀t.
∀t. Condition imposed by the absorbing surface.
exp(cid:18)−
(x + ζ − vt)2
exp(cid:18)−
√4πDt
(cid:19) −
4Dt
1
1
√4πDt
(x − ζ − vt)2
4Dt
(cid:19) exp(cid:18)vζ
D(cid:19) (9)
PX(x, t;−ζ, 0) =
Simulation parameters:
Velocity, v=3
Diffusion constant, D=0.3
6
5
4
3
2
1
f
o
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
P
l
l
e
u
c
e
o
m
e
h
t
f
o
n
o
i
t
i
s
o
p
e
h
t
0
−4
−2
0
2
Position of the molecule, x
6
T=0.01
T=0.25
T=0.5
T=0.75
T=1
4
6
Fig. 3.
The pdf of the position of the molecule P (x, t) for different values of t, when it is released at time t = 0 at position x = 0.
Because of positive drift velocity, the mean of the pdf travels in the positive direction, and because of the diffusion, the variance of the pdf
grows with time.
C. Distribution of Absorption Time
Recall from Section II that the receiver senses the particles only when they arrive, at which time they
are absorbed and removed from the system. Thus, for the purposes of this paper, the most important
feature of the Brownian motion X(t) expressed in (6)-(8) is the first passage time at the destination. For
a Brownian motion X(t), and an absorbing boundary located at position ζ, the first passage time τ (ζ) at
the barrier is defined as
τ (ζ) = min
t {X(t) : X(t) = ζ}.
(10)
In Figure 2(b), the simulated trajectories of six particles, modeled as a random walk, through a medium
are plotted. The particles were all released at x = 0, three at time 0 and three at time 400, into a fluid
medium that had a positive drift velocity. The receiver is located at x = 50. Notice the large variation
in the absorption times. Among the particles released at t = 0, one gets absorbed at t ≈ 360, other at
t ≈ 500, and another does not get absorbed even by t = 1000. Furthermore, this plot shows how particles
can get absorbed in an order different from the order in which they were released. It is therefore important
to understand the variation in the propagation times of the particle.
The derivation of the first passage time for our case is given in [22]. Here, we repeat briefly the steps
involved. At a given time t, the probability that the particle has not yet been absorbed is given by
7
1
−∞
PX(x, t;−ζ, 0)dx
exp(cid:18)−
¯F (t) = Z 0
= Z 0
√4πDt
D(cid:19)Z 0
−exp(cid:18)vζ
= 1 −Z ∞
1
√2π
−(vt−ζ)
√2Dt
−∞
−∞
1
4Dt
(x + ζ − vt)2
exp(cid:18)−
√4πDt
2 dx! − exp(cid:18)vζ
e −x2
(cid:19) dx
(x − ζ − vt)2
(cid:19) dx
D(cid:19) 1 −Z ∞
4Dt
1
√2π
2 dx!
e −x2
−(vt+ζ)
√2Dt
¯F (t) is the probability that the particle has not been absorbed until time t. The probability that the particle
has been absorbed before t is given by F (t) = 1 − ¯F (t). Hence, the probability density function of the
absorption time is f (t) = F ′(t) = − ¯F ′(t).
d ¯F
dt
f (t) = −
= d
dtZ ∞
1
√2π
= −
exp(cid:18) vζ
√4πDt3
=
ζ
1
√2π
2 dx!
e −x2
e −x2
2 dx! − exp(cid:18)vζ
(cid:19)(cid:18) −v
√2Dt
exp(cid:18)−(vt + ζ)2
D(cid:19) d
dtZ ∞
2√2Dt3(cid:19) +
(vt − ζ)
(cid:19)(cid:18) −v
√2Dt
4Dt
+
+
−(vt+ζ)
√2Dt
(vt + ζ)
2√2Dt3(cid:19)
1
√2π
−(vt−ζ)
√2Dt
4Dt
exp(cid:18)−(vt − ζ)2
D(cid:19) 1
√2π
exp(cid:18)−(vt − ζ)2
4Dt
(cid:19)
(11)
To summarize, (11) gives the probability density function of the absorption time of a particle released in
a fluid medium with diffusion constant D, at a distance ζ from the receiver, when the fluid has a constant
velocity v. Note that this equation is valid only for positive drift velocities, i.e., when the receiver is
downstream from the transmitter. Since our communication is based largely on the time of transmission
(and reception), this pdf characterizes the uncertainty in the channel, and plays a role similar to that of the
noise distribution in an additive noise channel. Some example plots of this function are given in Figure 4.
III. MUTUAL INFORMATION
The transmitter encodes the message in the time of release of molecules and possibly the number
of molecules. Based on the number and the time of absorption of the molecules, the receiver decodes
the transmitted information. This section develops the mutual information between the transmitter and
receiver for two cases: with a single transmitted molecule and two molecules whose release times can be
chosen independently. For a given information transmission strategy at the transmitter (called the input
distribution in the information theoretic literature), the mutual information is also the maximum rate at
which information may be conveyed using that strategy. (Mutual information is related to but distinct
from the capacity, which is the maximum mutual information over all possible input distributions.)
A. Overview
In a traditional wireline communication system, receiver noise causes uncertainty in the reception,
limiting the rate at which information can be conveyed. However, as discussed before, the uncertainty
in the propagation time is a major bottleneck to the information transfer in molecular communication.
8
Probability density function of the absorption time
v=0.1
D=0.01
v=0.01,D=0.1
v=0.01,D=0.01
v=0.1,D=0.1
v=0.1,D=0.01
10
20
30
Time of absorption
40
50
0.16
0.14
0.12
0.1
0.08
0.06
D=0.1
0.04
0.02
0
0
Fig. 4. The time at which the molecule gets absorbed by the receiver, given that it was released at time 0, is a random variable. This is
a result of the diffusion of the fluid medium. Here, we plot the probability distribution function of the absorption time for different sets of
velocity and diffusion. For this plot, the distance between the transmitter and the receiver is set at 1 unit.
This uncertainty in the propagation time also means that the order in which molecules are received at the
receiver need not be the order in which they were transmitted. This will result in “inter-block interference”.
This is a serious impairment in the low velocity regime, where the pdf of the absorption time decays very
slowly, making inter-block interference more likely.
In this paper, we ignore inter-block interference and assume that the clocks are synchronized. Developing
techniques for both issues are significant works in themselves and outside the scope of this paper. So our
results are most relevant to system in a fluid with some significant drift; further, as we show in Section
V, our results can be used to obtain upper bounds on both mutual information and capacity for any drift
velocity.
The channel here falls under a class of timing channels, channels where the mode of communication
is through the timings of various events. The capacity of such channels are usually more difficult to
characterize. A celebrated result in this field is the computation of the capacity of a single server queuing
system [20]. The molecular communication channel can be modeled as a ·/G/∞ queuing system, i.e., an
infinite server queuing system where the service time of a server is a random variable with distribution
same as the pdf of the absorption time. To our knowledge, the exact capacity of such a channel has not
been computed to date.
B. Single molecule: Pulse position modulation
We first analyze the case of the transmitter releasing just a single molecule. In such a scenario, it can
encode information only in the time of release of the molecule. The transmitter releases the molecule (or
not at all) in the beginning of one of N time slots, each of unit duration (i.e., Ts = 1 in arbitrary units);
this action on the part of the transmitter is called a channel use. This molecule then propagates through
the medium and is absorbed by the receiver in a later time slot. The receiver then guesses the time slot
in which the molecule was released. This is a form of pulse-position modulation (PPM).
Given that it has (N + 1) choices, the transmitter can encode a maximum of log2(N + 1) bits of
information per channel use, though in practice much less due to the uncertain arrival times of the
9
molecules. For instance, suppose the velocity of the fluid medium is high enough so that the particle gets
absorbed by the receiver in M ≃ N time slots with very high probability. In this case, one transmission
strategy would be to emit a molecule in one of N/M time slots (each separated by M slots), since inter-
block interference would thus occur with very low probability, and the transmitted information would arrive
without distortion. For such an ideal system, we can transmit information at a rate of log2N/M + 1 bits
per channel use. However, more practical and interesting is the less than ideal case with lower velocities.
In this paper we neglect inter-block interference, i.e., we assume that the receiver waits for enough time
slots M, to ensure that the molecule propagates to the receiver with high probability. Here M is chosen
such that this probability is 0.999. Further, we assume that the receiver sampling rate is Tr = Ts/5. This
provides both a digital input/output system while maintaining fairly high accuracy of the received time.
Both these parameters could be changed as required.
C. Mutual information as an optimization problem
Having dealt with preliminaries, we now derive the maximum possible mutual information, here as
an optimization problem. Define a random variable X to denote the time slot in which the transmitter
releases the molecule. Assume that the transmitter releases the particle at the beginning of the ith slot
(1 ≤ i ≤ N) with probability pi. With probability p0 = 1 −PN
i=1 pi, the transmitter does not release
the particle. Let Y denote the time slot in which the receiver absorbs the molecule. For the time being,
we allow Y to range between 1 and ∞, we will see shortly that this is not required. Also, let Y = 0
denote the event that the molecule is never received. Since in our idealized case, the receiver waits for a
sufficiently long time, the event of Y = 0 is the same event that the molecule is not transmitted. Assume
that the duration of the time slot is Tr.
Let F (t) denote the probability that the particle gets absorbed before time t given that it was released
at time 0, i.e., F (t) is the integral of the pdf in (11). Denote by αj the probability that the particle arrives
in the jth time slot, given that it was released at time 0, which is equal to F (jTr) − F ((j − 1)Tr) ;
αj = 0, j ≤ 0. Let H(X) denote the entropy of random variable X and let entr(x) represent the binary
entropy function, where
entr(x) =(−xlog2x x > 0
x = 0.
0
(12)
(13)
(14)
We now proceed to calculate the mutual information between the random variables X and Y .
N
Xi=1
H(Y X) = H(Y X = 0)p0 +
Xj=i+1
entr (P (Y = jX = i)) =
H(Y X = i)pi
pi
∞
N
pi
N
Xi=1
∞
Xj=i+1
entr (αj−i)
H(Y ) = entr(P (Y = 0)) +
entr(P (Y = j))
Xi=1
= 0 × p0 +
Xk=1
= (1 − p0)
∞
= entr(p0) +
= entr(p0) +
∞
∞
Xj=1
Xj=1
entr (αk) ,
∞
Xj=1
entr N
Xi=1
entr N
Xi=1
P (Y = jX = i)pi!
(αj−i) pi!
I(X; Y ) = H(Y ) − H(Y X)
= entr(p0) +
∞
Xj=1
entr N
Xi=1
piαj−i! − (1 − p0)
∞
Xk=1
10
entr (αk)
(15)
As seen in Figure 4, the sequence {αj} is a an eventually decreasing sequence. The rate of decay depends
on the values of the drift velocity v and the diffusion coefficient D. The summations in (15) can, therefore,
be terminated for some large enough M.
The expression for mutual information is a non-negative weighted sum of concave functions plus a
constant. Hence, the mutual information is a concave function of the input distribution {pi, i = 1, . . . , N}.
Finding the degree distributions, the values for pis which maximize the entropy, is therefore a concave
optimization problem. Standard convex optimization techniques can therefore be used to solve for the
input probability distribution which maximizes the mutual information efficiently, in particular, the Blahut-
Arimoto algorithm [23], [24].
As a special case, suppose that we were to convey information only in the time of release of the
molecule, i.e., we require the molecule to be transmitted. The derivation of mutual information is very
similar to the derivation above. Mutual information can then be expressed as
I(X; Y ) =
M
Xj=1
entr N
Xi=1
piαj−i! −
M
Xj=1
entr (αj) ,
(16)
and, again, the optimal degree distribution can be obtained through concave optimization.
D. Two molecules
In the work so far we have considered only the propagation of a single molecule and the focus was on
PPM-based communication. We now take a step toward involving amplitude wherein the transmitter can
release two identical molecules. The analysis is simplified by assuming that the propagation paths of these
two molecules are independent. The transmitter releases each of these molecules in one of the N time
slots or chooses not to release it. Based on the arrival times of these molecules at the receiver, the receiver
estimates their release times. However, because of the nature of the diffusion medium, different molecules
can take different times to propagate to the receiver. Hence, the molecules can be absorbed in a different
order than in which they were released: a key difference between this channel and traditional additive
noise channels. As a result, the amount of information that can be conveyed through the medium with
two indistinguishable molecules, as we will shortly see, is less than twice the amount of the information
that can be conveyed using a single molecule.
To obtain the maximum mutual information, let X1 ∈ {1, 2, . . . , N} be the time slot in which the first
particle is released, X2 ∈ {X1, X1 + 1, . . . , N} be the time slot in which the second particle is released.
Let Y1, and Y2 be the time slots in which the first and second particles are received. For notational
convenience, if a particle is not released, we denote it by a release in slot 0. Likewise, if a particle is not
received at the receiver, we denote it by a reception in time slot 0.
The probability mass function of the reception times (P (Y1, Y2)), and the conditional probability mass
function of the reception times given the transmission times (P (Y1, Y2X1, X2)) can be expressed in terms
of the conditional probability mass function of the reception time of one molecule, given its transmission
time (P (Y1 = y1X1 = x1) = αy1−x1). Let px1x2 represent P (X1 = x1, X2 = x2).
P (Y1 = y1, Y2 = 0X1 = x1, X2 = 0) = αy1−x1,
P (Y1 = y, Y2 = yX1 = x1, X2 = x2) = αy−x1αy−x2,
P (Y1 = y1, Y2 = y2X1 = x1, X2 = x2) =
x1, y1 > 0
x1 ≥ x2, y > 0
αy1−x1αy2−x2 + αy1−x2αy2−x1,
P (Y1 = 0, Y2 = 0) = p00
N
11
x1 ≥ x2, y1 6= y2 > 0
P (Y1 = y1, Y2 = 0) =
P (Y1 = y, Y2 = y) =
P (Y1 = y1, Y2 = y2 6= y1) =
N
Xx1=1
Xx1=1
Xx1=1
N
N
px10(cid:0)αy1−x1(cid:1)
Xx2=x1
Xx2=x1
N
px1x2(cid:0)αy−x1αy−x2(cid:1)
px1x2(cid:0)αy1−x1αy2−x2 + αy1−x2αy2−x1(cid:1)
The term αy1−x2αy2−x1 in the above equations accounts for the event that the molecule released later
gets absorbed before the molecule which is released earlier. The mutual information between the variables
(X1, X2) and (Y1, Y2) can now be written in terms of these probability mass functions. Note that in the
above derivation, we have assumed that αk for k ≤ 0 is defined as zero.
Using these equations, we can frame the mutual information maximization as another optimization
problem. The optimization is to be done over the upper triangular N × N matrix PX1,X2(x1, x2), where
each entry in the matrix is a positive number and all the entries sum to one. The mutual information is
a concave function of the optimization variables {px1x2 : x1 ∈ {1, 2, . . . , N}, x2 ∈ {x1, x1 + 1, . . . , N}}.
The exact expression is tedious to write, and is omitted here.
The well known Blahut-Arimoto algorithm [23], [24] is used to compute, numerically, the input
distribution that maximizes the mutual information in each of the different scenarios. The distance from
the sender to the receiver, ζ is set to one unit in all the results presented here.
IV. RESULTS
A. Release of a single molecule
When one molecule is to be released, information can be conveyed in whether it is released or not,
and if released, the slot number in which it is released.
Case when the molecule can be released in one of the N slots, or not at all: In Figure 5, we
plot the mutual information as a function of velocity, for two different sets of diffusion coefficients,
0.05, representing the low diffusion scenario, and a high diffusion constant 0.2. We have two sets of
plots in the figure, one for the case where we have two slots in which we can release the molecule, or
choose not to release it, and another, where we have four time slots. Also, we give the input distribution
(p0, p1, p2, p3, p4) at which the mutual information is maximized at the two extreme values of velocity.
From the figure, it is evident that the mutual information increases with an increase in velocity and
saturates to a maximum of log2(N + 1) bits. This trend is as expected. At high velocities, the optimal,
information maximizing, distribution is uniform. This is because the receiver can detect, without error,
the slot in which the transmitter disperses the molecule. Also, because the receiver waits for a sufficiently
long time, we can detect, without error, if a molecule was transmitted or not. Therefore, a lower limit
on the mutual information is one bit. At lower velocities, timing information is completely lost and the
mutual information is marginally greater than one bit.
12
2.5
2
N=2, D=0.05
N=2, D=0.2
N=4, D=0.05
N=2, D=0.2
1.5
0.4874,0.1995,0.0000,0.0000,0.3131
0.4548,0.1734,0.0449,0.0467,0.2803
s
t
i
b
n
i
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
M
1
10−2
10−1
100
Velocity
101
Fig. 5. Variation of mutual information (which measures in bits, the information that can be conveyed from the transmitter to the receiver)
with velocity. There are 2 sets of curves corresponding to the number of slots in which the molecule is released, N = 2 and N = 4. For
N = 4, we also list the p.m.f. of the release times which maximizes the mutual information.
The diffusion constant is a measure of the uncertainty in the propagation time. Hence, we would expect
the mutual information to be lower when the diffusion constant is high. This is indeed the case at high
velocities. However, it is surprising that a higher diffusion constant results in higher mutual information
at low velocities (Also refer Figure 6). This is because, at low velocities, it is the diffusion in the medium
which aids the propagation of the molecule from the transmitter towards the receiver. This is illustrated in
the pdf of the absorption time, shown in Figure 4. Compared to the case when the diffusion in the medium
is low, the probability distribution function is more “concentrated” (lower uncertainty) when the diffusion
in the medium is higher. Unfortunately, there does not seem to be a single parameter that characterizes
the resulting interplay between velocity and diffusion.
Case when we do not permit the transmitter to not transmit the molecule: The information, in
this scenario, is conveyed only in the time of release of the molecule. We find the input distribution
which maximizes (16). The mutual information in this case is plotted in Figure 7. The maximum mutual
information is now log2(N) bits, which is achieved at high velocities. However, it is in the low velocity
regime where the mutual information is significantly lower than the case where the transmitter is allowed
to not transmit the molecule. Figure 8 compares the two scenarios.
From the results, we see that the velocity-diffusion region can be roughly classified into three regimes:
• A diffusion dominated region, where mutual information is relatively insensitive to the velocity; this
corresponds to v < 10−1 in Figure 5.
• A high-velocity region where the mutual information is insensitive to the diffusion constant; this
corresponds to v > 3 in Figure 5.
• An intermediate regime, where the mutual information is highly sensitive to the velocity and diffusion
constant of the medium, 10−1 < v < 3 in Figure 5.
In the low velocity regime, we see no significant improvement in the mutual information when we
increase the number of time slots in which we can release the molecules. As expected, very little
information can be conveyed in the time of release of the molecule when there is high uncertainty in the
propagation time. Hence, we need to explore alternative ways of encoding message in this regime.
13
s
t
i
b
n
i
n
o
i
t
a
m
o
f
n
i
l
a
u
t
u
M
2.5
2
1.5
1
101
100
Diffusion
10−1
10−2
10−2
101
10−1
100
Velocity
Fig. 6. A grid plot denoting the mutual information for a range of different velocities and diffusion constants, for the case when N = 4.
Observe that at lower velocities, more information can be transferred in a medium with higher diffusion constant.
(0.2500,0.2500,0.2500,0.2500)
N = 2, D = 0.05
N = 2, D = 0.2
N = 4, D = 0.05
N = 4, D = 0.2
(0.3181,0.0825,0.0855,0.5140)
(0.3905,0.0000,0.0000,0.6095)
101
100
10−1
s
t
i
b
n
i
n
o
i
t
a
m
r
o
n
f
i
l
t
a
u
u
M
10−2
10−2
10−1
100
Velocity
101
Fig. 7. Variation of mutual information with velocity when the transmitter must disperse the molecule (p0 = 0). The scenario is similar to
the one used in plotting Figure 5, with the difference being that the transmitter is not permitted not to transmit a molecule.
14
Diffusion constant is 0.2 for all the curves
100
s
t
i
b
n
i
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
M
10−1
10−2
N = 2, p0 = 0
N = 4, p0 = 0
N = 2, p0 ≠ 0
N = 4, p0 ≠ 0
10−1
100
Velocity
101
Fig. 8. A comparison of the information bits conveyed in the scenarios when the transmitter must (p0 = 0) or may not release the molecule.
Plots from Figures 5 and 7 are compared here.
B. Release of multiple molecules
Here, we present the results of the scenario in which the transmitter is allowed to transmit at most 2
molecules. The results are presented in Figure 9. We have two sets of plots, one where the transmitter
can release the molecule in one of two time slots, other, where the transmitter can release the molecule
in one of four time slots. At low velocities, the mutual information is close to log23 bits. This is because,
at low velocities, any information encoded in the time of release of the molecule is lost. The receiver can
however accurately estimate the number of molecules transmitted. With two molecules, the receiver can
decode if the number of molecules transmitted was one or two or zero. However, this is because, we wait
for infinite time at the receiver. The probability distribution function which attains the maximum mutual
information at low velocities assigns, roughly, a probability of 1
3 to the events of releasing one or two or
no molecules.
At very high velocities, information encoded in both the time and number of molecules released is
bits can be conveyed at high
retained through the propagation. Hence, a maximum of log2
velocities.
(N +1)(N +2)
2
In Tables I, II and III we list the mutual information maximizing input distributions for the case of
release of two molecules in two time slots. Tables IV, V and VI list the input distributions for the case
of release of two molecules in four time slots. As expected, at low velocities, the total probability of
releasing one, two or zero molecules is roughly one third each. The molecules, to minimize uncertainty,
are transmitted ‘far apart’.
It is however surprising to note that for reasonable velocities when two molecules are released, they
are both to be released in the same time slot. This may be explained by the fact that, due to diffusion,
molecules can arrive out of order and the timing information is lost. Transmitting both molecules at once
avoids this confusion. This is also an important result; if this trend is to hold true for the release of
multiple molecules, then we could consider only those schemes wherein all the molecules are released in
one of the time slots, and where information is encoded only in the time slot in which all the molecules
are released.
MUTUAL INFORMATION MAXIMIZING INPUT DISTRIBUTION WHEN TWO MOLECULES ARE RELEASED IN ONE OF THE TWO POSSIBLE
TABLE I
15
SLOTS OR NOT RELEASED AT ALL, v = 10−2, d = 0.05
P (X2 = 1)
P (X2 = 2)
P (X2 = 0)
P (X1 = 1)
P (X1 = 2)
P (X1 = 0)
0.1424
0
0
0
0.1939
0
0.1412
0.1921
0.3303
TABLE II
v = 10−2, d = 0.2
P (X2 = 1)
P (X2 = 2)
P (X2 = 0)
P (X1 = 1)
P (X1 = 2)
P (X1 = 0)
0.1382
0
0
0
0.2113
0
0.1299
0.2035
0.3171
TABLE III
v = 10, d = 0.2 OR 0.05
P (X2 = 1)
P (X2 = 2)
P (X2 = 0)
P (X1 = 1)
P (X1 = 2)
P (X1 = 0)
0.1667
0
0
0.1667
0.1667
0
0.1667
0.1667
0.1667
MUTUAL INFORMATION MAXIMIZING INPUT DISTRIBUTION WHEN TWO MOLECULES ARE RELEASED IN ONE OF THE FOUR POSSIBLE
SLOTS OR NOT RELEASED AT ALL, v = 10−2, d = 0.05
TABLE IV
P (X2 = 1)
P (X2 = 2)
P (X2 = 3)
P (X2 = 4)
P (X2 = 0)
P (X1 = 1)
P (X1 = 2)
P (X1 = 3)
P (X1 = 4)
P (X1 = 0)
0.1395
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.2094
0
0.1313
0
0
0.2022
0.3176
TABLE V
v = 10−2, d = 0.2
P (X2 = 1)
P (X2 = 2)
P (X2 = 3)
P (X2 = 4)
P (X2 = 0)
P (X1 = 1)
P (X1 = 2)
P (X1 = 3)
P (X1 = 4)
P (X1 = 0)
0.1345
0
0
0
0
0
0.0305
0
0
0
0
0
0.0129
0
0
0
0
0
0.2052
0
0.1256
0.0122
0
0.1964
0.2827
TABLE VI
v = 10, d = 0.2 OR 0.05
P (X2 = 1)
P (X2 = 2)
P (X2 = 3)
P (X2 = 4)
P (X2 = 0)
P (X1 = 1)
P (X1 = 2)
P (X1 = 3)
P (X1 = 4)
P (X1 = 0)
0.0667
0
0
0
0
0.0667
0.0667
0
0
0
0.0667
0.0667
0.0667
0
0
0.0667
0.0667
0.0667
0.0667
0
0.0667
0.0667
0.0667
0.0667
0.0667
4
3.5
3
2.5
2
s
t
i
b
n
i
n
o
i
t
a
m
r
o
f
n
I
l
a
u
t
u
M
1.5
10−2
16
N=2, D=0.05
N=2, D=0.2
N=4, D=0.05
N=4, D=0.2
10−1
100
Velocity
101
Fig. 9. Variation of mutual information with velocity for the case when the transmitter is allowed to release at most two molecules. The
mutual information maximizing input distributions at the extreme points of the graph are given in Tables I, II, III, IV, V and VI.
V. RELATIONSHIP TO ACHIEVABLE INFORMATION RATES AND CAPACITY
When pulse-position modulation is used, symbols are normally transmitted consecutively. That is, if
the duration of a symbol is T , then the first symbol is transmitted on the interval [0, T ), the second on
the interval [T, 2T ), and so on. However, for the Brownian motions considered in Section II, molecules
transmitted during a given interval may arrive during a later interval, causing inter-block interference. In
this paper, we have avoided this problem by only considering symbols transmitted in isolation, disregarding
inter-block interference.
In fact, for a fixed input distribution, our information results lead to an upper bound on the mutual
information under consecutive symbol transmission. To show this, let X represent the alphabet of allowed
symbols. For simplicity, suppose that a symbol is composed of the release of a single molecule, although
this assumption can be relaxed without changing the argument. Then we will assume that X is a finite,
discrete list of allowed molecule release times on the interval [0, T ); the cardinality X gives the number
of allowed release times. Further, there exists a discrete input distribution, with pmf pX (x), over X . Let
Y represent the corresponding set of channel outputs, given a single symbol input to the channel, and
disregarding inter-block interference. Since Y is the arrival time of a single molecule transmitted on the
interval [0, T ), then clearly Y = [0,∞), and nothing changes if Y is quantized.
Let x = [x1, x2, . . . , xn] ∈ X n and y = [y1, y2, . . . , yn] ∈ Y n represent vectors of channel inputs and
outputs, respectively, for n uses of the channel in isolation; throughout this section, we will assume that
xi is independent and identically distributed (IID) for each i. Suppose the symbols in x are transmitted
consecutively. Then the resulting sequence of molecule release times can be written r = [r1, r2, . . .], where
(17)
Since xi ∈ [0, T ), clearly ri ∈ [(i−1)T, iT ). Note that the mutual information per unit time of the channel
is given by I(X; Y )/T , which is calculated for given X and pX(x).
ri = xi + (i − 1)T.
Let ui represent the arrival corresponding to ri. Since ri is a time-delayed version of xi, and yi is the
arrival corresponding to xi, from (17) we have
17
ui = yi + (i − 1)T.
(18)
The corresponding vector is u = [u1, u2, . . . , un]. However, the receiver does not observe u directly –
instead, it observes w, where
(19)
and where the function sort(·) sorts the argument vector in increasing order. That is, while information
symbol xi corresponds to arrival time ui, it is potentially unclear which element of x corresponds to
arrival time wi.
w = sort(u),
Since the length-n vectors of consecutive input symbols r and sorted outputs w are random variables,
we can write the mutual information between them as I(R; W). However, we are more interested in the
mutual information per unit time I(R; W ), which is given by
I(R; W ) = lim
n→∞
= lim
n→∞
1
nT + δ
1
nT
I(R; W)
I(R; W) − ǫ,
ǫ =(cid:18) 1
nT −
1
nT + δ(cid:19) I(R; W).
where nT represents the total time to transmit n symbols, δ is the extra time after nT required to wait
for all remaining molecules to arrive, and
We let δ = log n, so that limn→∞ δ = ∞ (which is long enough time for all molecules to arrive with
probability 1). With this in mind, we have the following result:
(20)
(21)
(22)
(23)
(24)
(25)
Theorem 1:
1
T
Proof: From (20), since ǫ is positive,
I(X; Y ) ≥ I(R; W ).
I(R; W ) ≤ lim
n→∞
1
nT + δ
I(R; W).
Then we have that
I(X; Y) = I(R; U)
≥ I(R; W),
where the first equality follows from (17)-(18), since r and u are bijective functions of x and y,
respectively; and the second inequality follows from the data processing inequality (e.g., see [25]) and
(19). Finally, since xi, yi and xj, yj are independent for any i 6= j, I(X; Y) = nI(X; Y ), and the theorem
follows.
Note that the result in Theorem 1 bounds the mutual information, and thus applies to each set X and
input pmf pX(x); however, we can also show that the result applied to capacity. Let Cm represent the
maximum of I(X; Y ) where X = m, i.e.,
The capacity of the channel uses in isolation is then given by
Cm = max
pX (x):X =m
I(X; Y ).
It remains to show that this limit exists, which we do in the following result.
C = lim
m→∞
Cm.
(26)
Theorem 2: C exists, and is finite, if 0 < v, D, ζ, T < ∞. Furthermore, if maxpX (x) I(R; W ) represents
the capacity of I(R; W ) under IID inputs, then
max
pX (x)
I(R; W ) ≤
C
T
.
(27)
Proof: To prove the first part of the theorem, we proceed in two steps.
18
1) Cm is a nondecreasing sequence. For each m, either: the maximizing distribution pX(x) (or every
maximizing distribution, if not unique) satisfies pX(x) > 0 for all x ∈ X ; or pX(x) = 0 for at
least one x ∈ X (in at least one maximizing distribution, if not unique). If the former is true, then
Cm > Cj for all j < m; if the latter is true, then Cm = Cm−1. Thus, Cm is nondecreasing in m.
2) Cm is upper bounded. We can write I(X; Y ) = H(Y ) − H(Y X) = H(Y ) − H(N), where H(N)
is the entropy of the first arrival time. We can upper bound H(Y ) independently of m as follows.
If the pdf of y is fY (y), then H(Y ) = E[log2 1/fY (y)], where E[·] is expectation. If g(y) is any
valid pdf of y, then by a well-known property of entropy, H(Y ) ≤ E[log2 1/g(y)] (with equality
when g(y) = fY (y)). Pick g(y) = e−y (supported on y = [0,∞)), the exponential distribution with
unit mean. Then H(Y ) ≤ E[y] log2 e, which is finite if E[y] is finite. Finally, E[y] = E[x] + E[n] ≤
T + E[n], and E[n] is known to be finite if 0 < v, D, ζ < ∞ [22].
Since Cm is a nondecreasing, upper bounded sequence, it must have a finite limit.
To prove the second part of the theorem, note that Theorem 1 applies to all input distributions pX(x);
thus, it also applies to the one maximizing I(R; W ). As a result, since C exists and is finite (from the
first part of the theorem), it is a nontrivial upper bound on maxpX (x) I(R; W ).
In [12], it was shown that the mutual information cannot be tractably computed in general for “sorting”
channels, i.e., those with outputs given by (19). Since I(X; Y ) can be calculated relatively easily, the
results from Theorems 1-2 give us useful information about the capacity of a practical system.
VI. CONCLUSIONS AND FUTURE WORK
In this paper, a framework was constructed to study data rates that can be achieved in a molecular
communication system. A simple model was considered for the communication system, consisting of a
transmitter and receiver separated in space, immersed in a fluid medium. The rates achieved by a simple
pulse-position modulation protocol were analyzed, where information is encoded in the time of release of
the molecule. These results were extended to two molecules wherein the optimal distribution reverted to
the PPM protocol. While preliminary, the results of this work suggest practical data transmission strategies
depending on the value of the drift velocity.
Given the preliminary nature of this work, there are many interesting related problems. For example,
it would be useful to consider the limitations inherent in molecular production and detection: precise
control over release times and amounts, and precise measurement of arrival times, may not be possible;
more realistic communication models could be produced. Furthermore, the communication architecture
of molecular communication systems may be considered; for instance, in order to achieve the mutual
information results given in this paper, error-correcting codes must be used; an appropriate modulation and
coding strategy for molecular communication needs to be identified. Finally, channel estimation techniques
need to be derived in order to cope with unknown parameters, such as an unknown drift velocity. Much
work remains to be done to understand molecular communication from a theoretical perspective, which
presents many interesting and exciting challenges to communication researchers.
[1] S. Hiyama, Y. Moritani, T. Suda, R. Egashira, A. Enomoto, M. Moore, and T. Nakano, “Molecular communication,” in Proc. 2005
NSTI Nanotechnology Conference, Anaheim, CA, 2005, pp. 391–394.
[2] S. P. Brown and R. A. Johnstone, “Cooperation in the dark: Signalling and collective action in quorum-sensing bacteria,” Proceedings
of the Royal Society of London B, vol. 268, pp. 961–965, 2001.
REFERENCES
19
[3] A. Cavalcanti, T. Hogg, B. Shirinzadeh, and H. C. Liaw, “Nanorobot communication techniques: A comprehensive tutorial,” in IEEE
Intl. Conf. on Control, Automation, Robotics and Vision, Singapore, 2006.
[4] T. Nakano, T. Suda, M. Moore, R. Egashira, A. Enomoto, and K. Arima, “Molecular communication for nanomachines using intercellular
calcium signalling,” in Proc. 5th IEEE Conference on Nanotechnology, Nagoya, Japan, 2005, pp. 478–481.
[5] T. Nakano, T. Suda, T. Kojuin, T. Haraguchi, and Y. Hiraoka, “Molecular communication through gap junction channels: System design,
experiments and modeling,” in Proc. 2nd International Conference on Bio-Inspired Models of Network, Information, and Computing
Systems, Budapest, Hungary, 2007.
[6] A. Enomoto, M. Moore, T. Nakano, R. Egashira, T. Suda, A. Kayasuga, H. Kojima, H. Sakibara, and K. Oiwa, “A molecular
communication system using a network of cytoskeletal filaments,” in Proc. 2006 NSTI Nanotechnology Conference, Boston, MA,
2006, pp. 725–728.
[7] S. Hiyama, Y. Moritani, and T. Suda, “A biochemically engineered molecular communication system,” in Proc. 3rd International
[8]
Conference on Nano-Networks, Boston, MA, USA, 2008.
I. Akyildiz, F. Brunettib, and C. Bl´azquezc, “Nanonetworks: A new communication paradigm,” Computer Networks, vol. 52, no. 12,
pp. 2260–2279, 22 August 2008.
[9] S. Hiyama and Y. Moritani, “Molecular communication: Harnessing biochemical materials to engineer biomimetic communication
systems,” Nano Communication Networks, vol. 1, no. 1, pp. 20–30, Mar. 2010.
[10] P. J. Thomas, D. J. Spencer, S. K. Hampton, P. Park, and J. P. Zurkus, “The diffusion mediated biochemical signal relay channel,” in
Proc. 17th Annual Conference on Neural Information Processing Systems, Vancouver, BC, 2003.
[11] A. W. Eckford, “Nanoscale communication with Brownian motion,” in Proc. Conf. on Information Sciences and Systems, Baltimore,
MD, 2007, pp. 160–165.
[12] ——, “Molecular communication: Physically realistic models and achievable information rates,” arXiv:0812.1554v1 [cs.IT] 8 December
2008. Submitted to IEEE Transactions on Information Theory.
[13] B. Atakan and O. Akan, “An information theoretical approach for molecular communication,” in Proc. 2nd Intl. Conf. on Bio-Inspired
Models of Network, Information, and Computing Systems, Budapest, Hungary, 2007.
[14] ——, “On channel capacity and error compensation in molecular communication,” in Transactions on Computational Systems Biology
X. Springer, 2008, pp. 59–80.
[15] M. J. Garvey, Diffusion Mediated Signaling: Information Capacity and Coarse Grained Representations. M.Sc. Thesis, Case Western
Reserve University, 2009.
[16] A. W. Eckford, “Timing information rates for active transport molecular communication,” in Proc. 4th International Conference on
Nano-Networks, Lucerne, Switzerland, 2009.
[17] M. J. Moore, K. Oiwa, and T. Suda, “Molecular communication: Modeling effects on information rate,” IEEE Trans. Nanobioscience,
vol. 8, no. 2, pp. 169–180, Jun. 2009.
[18] M. Pierobon and I. F. Akyildiz, “A physical end-to-end model for molecular communication in nanonetworks,” IEEE J. Sel. Areas in
Commun., vol. 28, no. 4, pp. 602–611, May 2010.
[19] Yu. M. Kabanov, “The capacity of a channel of the Poisson type,” Theory of Probability and Applications, vol. 23, no. 1, pp. 143–147,
1978.
I. Karatzas and S. E. Shreve, Brownian Motion and Stochastic Calculus (2nd edition). New York: Springer, 1991.
[20] V. Anantharam and S. Verdu, “Bits through queues,” IEEE Trans. Inform. Theory, vol. 42, no. 1, pp. 4–18, Jan. 1996.
[21]
[22] R. S. Chhikara and J. L. Folks, The Inverse Gaussian Distribution: Theory, Methodology, and Applications. Marcel Dekker, 1989.
[23] R. E. Blahut, “Computation of channel capacity and rate-distortion functions,” IEEE Trans. Inform. Theory, vol. 18, pp. 460–473, 1972.
[24] S. Arimoto, “An algorithm for computing the capacity of arbitrary discrete memoryless channels,” IEEE Trans. Inform. Theory, vol. 18,
no. 1, pp. 14 – 20, Jan. 1972.
[25] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. Wiley, 2006.
|
1901.09841 | 1 | 1901 | 2019-01-28T17:35:33 | A Mechano-Chemical model for tumors growth | [
"physics.bio-ph",
"q-bio.TO"
] | In this paper we present a study of local dynamics of the growth of cancer tumor and healthy cells considering the presence of nutrients in the system. We also analyze the evolution of system if we take indirectly into account the level of alkalinity (pH) in the system which shows that influences in tumor growth. The model consists in a set of differential equations of first order that involves a mechanical model added by a pair of differential equations for local oxygen and glucose. | physics.bio-ph | physics |
A Mechano-Chemical model for tumors growth
Cristian C. P´erez ´Aguila1, Maura C´ardenas G.2, and J. Fernando Rojas1
1Facultad de Ciencias F´ısico Matem´aticas., Benem´erita Universidad Aut´onoma de Puebla,,
Av. San Claudio y 18 sur, Ciudad Universitaria, Col. San Manuel., C. P. 72570. Puebla,
2Facultad de Medicina. Benem´erita Universidad Aut´onoma de Puebla, Calle 13 sur 2702,
Col. Los Volcanes, C. P. 72420 Puebla, M´exico
M´exico.
Abstract
In this paper we present a study of local dynamics of the growth of cancer tumor and healthy
cells considering the presence of nutrients in the system. We also analyze the evolution of system
if we take indirectly into account the level of alkalinity (pH) in the system which shows that
influences in tumor growth. The model consists in a set of differential equations of first order
that involves a mechanical model added by a pair of differential equations for local oxygen and
glucose.
Keywords: Tumor growth, Tumor cells, Healthy cells, Extracellular matrix, Extracellular liquid,
Enzymes, Nutrients
1 Introduction
Cancer is a disease of uncontrolled cell proliferation [1]. These cells exhibit metabolic alterations
[2] that distinguish them from healthy tissues and make their metabolic processes susceptible to
pharmacological targeting [2]. Successful cell proliferation is dependent on a profound remodeling of
cellular metabolism, required to support the biosynthetic needs of a growing cell [3]. In the literature
we find that the conditions of the tumor microenvironment are important in the metabolism of
the tumoral cells [2], this allows cells to use available nutrients such as glucose and glutamine
[2, 1, 3, 4] for its survival and propagation. Some tumours show increased glucose uptake, that
can be promoted by gene expression or oncogenic signalling changes [2, 3]. Signaling through
oncogenes allows cancer cells to perpetually receive pro-growth stimuli, which for normal tissues
are only transient in nature. As a consequence, cancer metabolism is constitutively geared toward
supporting proliferation [3].
Some types of cancer show the nutrients are insufficient. In recent work has shown that cancer
cells are able to overcome this nutrient insufficiency by scavenging alternative substrates as proteins
and lipids [1].
Another problem that arises is the reduction of oxygen supply or hypoxia. Lack of sufficient oxy-
gen has a profound effect on cellular metabolism as it inhibits those biochemical reactions in which
oxygen is consumed [3]. One of the most important phenotypic changes is the cancer cells ability
to change the extracellular acidity due to an altered glycolysis pathway. The ability to expel pro-
tons from cells and, therefore, change the extracellular pH provides an advantage to cancer cells [5].
1
In recent works it is clear that pH plays an important role in the survival mechanisms of
cancer cells [5, 3, 2, 4]. The acidic tumour microenvironment has also been associated with the
degree of cancer aggressiveness. It has been reported that the lower the pH surrounding the cell,
the more likely is the chance of malignancies with higher degree of invasiveness [5]. A recent
therapeutic approach takes advantage of the altered acidity of the tumour microenvironment by
using proton pump inhibitors (PPIs) to block the hydrogen transport out of the cell. The alteration
of the extracellular pH kills tumour cells, reverses drug resistance, and reduces cancer metastasis.
Targeting tumour pH can be therefore considered a therapeutic strategy [5].
2 Components of system.
A tumor is an example of a complex system composed of cells. Each of the the reproduced cells
follows a set of rules and responds to local interactions with other cells either from the tumor or
with healthy cells.
Extracellular matrix.
The Extracellular matrix (ECM) represents a network that includes to all organs, tissues and cells
of the body, in this medium the cells survive, multiply and perform their vital functions [6].
The balance of extracellular matrix is maintained by a regulation between synthesis, training and
reshuffle. Also the extracellular matrix provides oxygen and nutrients to the cell and the elimina-
tion of CO2 and toxins under normal conditions [6].
The extracellular matrix of animals need to reorganize and also need a constant regeneration
through of the degradation of components and the production of new components by cells. The
degradation of ECM causes enzymes production such as metalloproteinases [7].
Extracellular liquid.
The extracellular liquid is body fluid not contained in cells. Often it is secreted by cells for an
suitable environment. It is transported across the body in two stages, the first is the movement of
blood by circulatory system; the second is movement between blood capillaries and cells.
When blood flow through the capillaries, a large portion of the liquid containing is diffused into and
out of interstitial liquid that found between cells, allowing the continuous exchange of substances
between the cells and interstitial liquid and between the interstitial liquid and the blood.
Source of nutrients in the extracellular liquid.
• The respiratory system, it provides oxygen to the body and removes carbon dioxide.
• The digestive system, digests food and absorbs nutrients including carbohydrates, fatty acids,
and amino acids, for to bring into the extracellular liquid.
The blood.
Transports nutrients, electrolytes, hormones, vitamins, antibodies, heat and oxygen, its composed
of plasma, red and white blood cells, platelets. The principal function of red blood cells, transports
oxygen and carbon dioxide. Glucose is a sugar which comes of food we eat, this glucose is stored
in the body, is the main source of energy for body cells and reaches every cells through the blood-
stream.
2
3 Mathematical model
We consider a mixture in which the essential components are cells, extracellular matrix, and extra-
cellular liquid with dissolved chemicals. Hypotheses that are assumed are as follows:
• The components of the extracellular matrix form a complex network and move together.
• Cells responds mechanically to compression from cells around, regardless of the type of cell.
• The presence of nutrients in the mixture is essential for growth and vital functions of cells [8].
For this model is defined φ ∈ [0, 1] as the volume occupied by cells and either m ∈ [0, 1] as
the volume occupied by extracellular matrix. The mixture is saturated if the remaining space is
filled with extracellular liquid then we define l ∈ [0, 1] as the volume occupied by extracellular
liquid. Considering normalized amounts for the sample volume we have:
φ + m + l = 1
(1)
In the cellular component of tissue, the differential equation that describes the change in mass of
the cellular component is
+ ∇ · (φ−→v φ) = Γφ
∂φ
∂t
(2)
which considers that cell density is constant ([9], [10], [11]). Here Γφ corresponds to rapid growth
or death for cell mass, −→v φ is the cell rate and the term ∇ · (φ−→v φ) it is associated with the flow of
cell component
In the same way, the corresponding equations for the components of the mixture
+ ∇ · (φ−→v φ) = Γφ
+ ∇ · (m−→v m) = Γm
+ ∇ · (l−→v l) = Γl.
∂φ
∂t
∂m
∂t
∂l
∂t
(3)
(4)
(5)
For the system of interest it is needed to distinguish different types of cells such as tumor cells,
endothelial, epithelial, fibroblasts, etc. However, for simplicity, in the cell mass only we only
distinguish two subpopulations that correspond to healthy cells and tumor cells. In the extracellular
matrix components all are considered as a single assembly.
and considering the result obtained in [9] and [10], cell velocity is:
i = n, t,
(6)
(7)
i = n, t.
+ ∇ · (φi
∂φi
∂t
(cid:18)
−→vi = −Kim
1 −
−→v φi) = Γφi,
(cid:19)
σim
∇ · (φiTφ)
∇(φiTφ)
In this equation it is considered σim = 0 ([11]), that is associated with frictional forces and it´s
expected to depends on adhesion mechanisms of volume occupied by cells. Also it is supposed that
the coefficient of mobility Kim = Ki is constant, while the term Tφ expresses stress to which the
cells are subjected.
3
Now if we divide the sample volume in small cubes in which only there is a cell population healthy
or tumor, then φi is constant in each cube, and in this way the cell velocity can be expressed as
−→v i = −Ki∇(φiTφ)
i = n, t
where, the term associated with the flow can be expressed as follows
∇ · (φi
−→v φi) = −φiKi∇2(φiTφ).
It is necessary an equation that describes the response of cells and extracellular matrix to stress,
this equation is constructed assuming that they behaves like an elastic fluid ([9], [12], [13], [10]), so
we consider
(8)
(9)
(10)
(11)
and for the function Σ, we use the expression proposed by Chaplain et. al. [11]
Tφ = −ΣI,
(cid:18) 2 − ψ0
1 − ψ0
(cid:19)(cid:18) ψ − ψ0
(cid:19)
1 − ψ
,
Σ(ψ) =
where ψ = φn + φt + m, measure indirectly, the free space available locally and it can be use for
calculate the stress exerted by the environment on cellular matter. Also ψ0 ∈ (0, 1), is identified as
the volume free of stress.
3.1 Term of growth or death cell.
Cells live in a crowd environment and they perceive the presence of other cells. This fact is fun-
damental in controlling cell concentration and implies to stop mitosis (cellular division) when the
volume exceeds a threshold, so the behavior of cells in terms of growth and movement depends
crucially on how they perceive the presence of other cells and its signal is translated.
For the terms of growth, we consider ([9], [11], [10]), the next
Γi = [γiHσ(ψ − ψi) − δi(ψ)]φi,
(12)
where Hσ(ψ − ψi) is a amendment to step function, and it is equal to 1 for ψ smaller that the
threshold value ψi and is equal to zero for ψ > ψi + σ. Also it must be satisfied that ψn < ψt, and
we consider constants δi, and γi > 0 which represent coefficients of death and cell reproduction
respectively. The parameter σ > 0 controls the transition between Hσ(s) = 1, and Hσ(s) = 0
([11]), therefore it controls the transition on/off in the cells reproduction, so the expression used in
this work is
i = n, t
1, if s ≤ 0
0, if s > σ
1 − s
Hσ(s) =
σ , another case.
(13)
3.2 Remodeling of the extracellular matrix.
As mentioned on the biological part of extracellular matrix, now we consider a concentration of
enzymes that degrades the ECM. The enzymes are produced by tumor or healthy cells[7] and so
the process of remodelling can be described by the equation for m as follows:
= µnφnΣ(ψ) + µtφtΣ(ψ) − νem.
∂m
∂t
(14)
4
Here µn and µt corresponds to speed of production in the extracellular matrix by healthy or tumor
cells respectively, ν is the degradation coefficient due to action of enzymes that degrade the matrix
and e is the concentration of said enzymes [9].
Now, we introduce an reaction-diffusion equation which aims to describe the evolution in con-
centration of enzymes that degrade the extracellular matrix
∂e
∂t
= κ∇2e + πnφnΣ(ψ) + πtφtΣ(ψ) − e
τ
.
(15)
In this equation πn and πt correspond to the rapid production of enzymes that degrade the extra-
cellular matrix produced by healthy and tumor cells respectively; τ is the average lifetime and κ
the diffusion coefficient of enzymes [9].
3.3 Nutrients.
In tumor growth, as well as vital functions of an healthy cell it is necessary the presence of dissolved
chemicals in the liquid component of the mixture considered, such as nutrients, growth factors, etc.
These chemicals are "absorbed" by cell populations to perform some vital functions such as prolif-
eration, growth or intercellular communication [9]. For the consideration of nutrients will assume
that there is an source that supplies the nutrients, that is to say, there is a capillary in which blood
flow is constant that transports nutrients.
Considering only oxygen and glucose [8], we propose the following equations [10]:
= Dc∇2c − βnφnc + fc
∂c
∂t
(16)
= Dg∇2g − βtφtg + fg
∂g
∂t
(17)
in which c ∈ [0, 1] is the oxygen concentration and g ∈ [0, 1] is the glucose concentration with Di
(i = c, g) are the diffusion coefficients of nutrients and βj (j = n, t) corresponds to the absorption
rate. Also fi represents the continuous supply of nutrients (i.e. a source). In the supply of nutrients
must be met fc + fg < 1.
Considering the nutrients that influence in the evolution of cell population, we assume that the
rate of reproduction in cells is proportional to nutrients concentration, that would result replace
γn by cγn (γn −→ cγn) for healthy cells and γt −→ gγt for tumor cells, in the case of rate cell
death is obtained δn −→ (1 − c)δn and δt −→ (1 − g)δt for healthy and tumor cells respectively,
the reason for these substitutions is that for example if considered to healthy cells, with an optimal
concentration of oxygen these are reproduced and the term of rate of death must decrease, this
situation is described by the term (1 − c). We must mention that competition is not considered
between cell populations for nutrients. Thus, the complete mathematical model for the simulation
5
of the system it is as follows:
∂e
∂c
∂g
∂m
∂φt
∂φn
∂t = φnKn∇2(φnΣ(ψ)) + cγnφnHσ(ψ − ψn) − (1 − c)δnφn
∂t = φtKt∇2(φtΣ(ψ)) + gγtφtHσ(ψ − ψt) − (1 − g)δtφt
∂t = µnφnΣ(ψ) + µtφtΣ(ψ) − νem
∂t = κ∇2e + πnφnΣ(ψ) + πtφtΣ(ψ) − e
∂t = D∇2c − βnφnc + fc
∂t = D∇2g − βtφtg + fg.
τ
(18)
which will be solved in an homogeneus space.
4 Results
In this section we show, the results of numerical solution for the system of equations obtained,
some parameters were obtained from [11], and we propose some parameters for the simulations.
We consider the cell population is aproximately 50%, the extracellular matrix occupies a 20% and
extracellular liquid a 30% of the total volume.
We will take as reference the following parameters:
φn = 0.45 φt = 0.005 m = 0.2
e = 0.3
c = 0.25
g = 0.16
Table 1: Initial condition.
Kn = 0.1
πn = 6000000
δt = 0.03
γn = 0.746
βn = 1.2
µt = 0.05
ψn = 0.6
Kt = 0.3
πt = 3000000
ν = 0.000016
κ = 0.00000734
τ = 0.005
δn = 0.1
γt = 0.97
βt = 1.3
D = 1.0
µn = 0.1
ψt = 0.8
σ = 0.2
ψ0 = 0.75
fc = 0.25
fg = 0.16
Table 2: Set of parameters.
Following graphs show the results of the local dynamics of the system considered.
Figures 1 and 2 show dependence of cell populations in nutrient uptake, for which is varied the
absortion rate of tumor and healthy cells
In the graphs 3, 4, correspond to growth of cell population, varing the concentration of oxygen,
also we considered that fg = 0.1.
6
Figure 1: Evolution of healthy cells varying the rate of oxygen uptake.
Figure 2: Evolution of tumor cells varying the rate of glucose uptake.
The parameters used in the simulation remain constant. But if we consider the parameters of
the source for oxygen and glucose (which are associated with pH levels), these parameters may vary
depending on the diet in each person.
For this situation we use uniform random numbers for these parameters, only in the interval [0,
0.5].
5 Conclusions
From the numerical results that we get can say the following: the model presented in this paper is
congruent with Gompertz's model for tumor cells, in which it is observed a stage of uncontrolled
growth and then they stabilize at a value, also can be seen in graphics that they are suscepti-
ble to the presence of nutrients, these are two cases, if the absorption rate is low the occupied
volume increases and if absorption rate is high the occupied volume is less. With respect to the
variation in the oxygen concentration, can be seen two situations, if the local concentration of
7
Figure 3: Growth of healthy cells varing the oxygen concentration.
Figure 4: Growth of tumor cells varing the oxygen concentration.
oxygen is low, the volume in stabilizing the tumor cells is high and when the local concentration
of oxygen is high the volume of tumor cells is lower compared to the previous situation, this sug-
gests that can intervene in tumor growth increasing the local concentration of oxygen in the system.
Acknowledgments
We are grateful for the facilities provided by the Laboratorio Nacional de Superc´omputo (LNS) del
Sureste de M´exico to obtain these results.
8
References
[1] Michalopoulou E., Bulusu V., and J. Kamphorst J. Metabolic scavenging by cancer cells: when
the going gets tough, the tough keep eating. British Journal of Cancer, 115:635 -- 640, 2016.
[2] Anastasiou D. Tumor microenvironment factors shaping the c´ancer metabolism landscape.
British Journal of Cancer, 116:277 -- 286, 2017.
[3] Pavlova N. N. and Thompson C. B. The emerging hallmarks of cancer metabolism. Cell Metab,
23:27 -- 47, 2016.
[4] Swietach P., Vaughan-Jones R. D., Harris A. L., and Hulikova A. The chemistry, physiology
and pathology of ph in cancer. Phil. Trans. R. Soc. B, 2014.
[5] Walsh M., Fais S., Pierluigi S. E., Harguindey S., Abu I. T., Scacco L., Williams P., Allegrucci
C., Rauch C., and Omran Z. Proton pump inhibitors for the treatment of cancer in companion
animals. Journal of Experimental & Clinical Cancer Research, pages 34 -- 93, 2015.
[6] ´Alvaro N. Tom´as, Noguera S. Rosa, and Farinas G. Fernando. La matriz extracelular: mor-
folog´ıa, funci´on y biotensegridad (parte 1). Rev. Esp. Patol., 42:249 -- 261, 2009.
[7] Meg´ıas Manuel, Molist Pilar, and A. Pombal Manuel. Atlas de Histolog´ıa vegetal y animal:
Matriz Extracelular. 2014.
[8] Munoz P. Cristina. El metabolismo del c´ancer. SEBBM Divulgaci´on, 2013.
[9] Astanin S. and Preziosi L. Multiphase Models of Tumor Growth. Ed. Springer New York, 2002.
[10] Preziosi L. and Tosin A. Multiphase modeling of tumor growth and extracellular matrix
interaction: Mathematical tools and applications. J. Math. Biol., in press???, 2007.
[11] Chaplain M. A. J., Graziano L., and Preziosi L. Mathematical modelling of the loss of tissue
compression responsiveness and its role in solid tumour development. Math. Med. Biol., 23:197 --
229, 2006.
[12] Astanin S. and Tosin A. Mathematical model of tumour cord growth along the source of
nutrient. Math. Mod. Nat. Phen., pages 153 -- 177, 2007.
[13] Pijush K. Kundu and Ira M. Cohen. Fluid Mechanics. Academic Press, second edition edition,
2002.
9
|
1005.5314 | 1 | 1005 | 2010-05-28T15:24:39 | Dendritic Actin Filament Nucleation Causes Traveling Waves and Patches | [
"physics.bio-ph",
"cond-mat.soft",
"q-bio.SC"
] | The polymerization of actin via branching at a cell membrane containing nucleation-promoting factors is simulated using a stochastic-growth methodology. The polymerized-actin distribution displays three types of behavior: a) traveling waves, b) moving patches, and c) random fluctuations. Increasing actin concentration causes a transition from patches to waves. The waves and patches move by a treadmilling mechanism which does not require myosin II. The effects of downregulation of key proteins on actin wave behavior are evaluated. | physics.bio-ph | physics | Dendritic Actin Filament Nucleation Causes Traveling Waves and
Patches
Anders E Carlsson
Department of Physics, Washington University,
One Brookings Drive, Campus Box 1105, St. Louis, MO 63130
(Dated: June 21, 2018)
Abstract
The polymerization of actin via branching at a cell membrane containing nucleation-promoting
factors (NPFs) is simulated using a stochastic-growth methodology. The polymerized-actin dis-
tribution displays three types of behavior: a) traveling waves, b) moving patches, and c) random
fluctuations. Increasing actin concentration causes a transition from patches to waves. The waves
and patches move by a treadmilling mechanism which does not require myosin II. The effects of
downregulation of key proteins on actin wave behavior are evaluated.
0
1
0
2
y
a
M
8
2
]
h
p
-
o
i
b
.
s
c
i
s
y
h
p
[
1
v
4
1
3
5
.
5
0
0
1
:
v
i
X
r
a
1
The dynamic behavior of the intracellular protein actin in cells is crucial in directing cell
shape changes and migration. Actin occurs in both monomeric ("G-actin"), and polymerized
("F-actin") forms. F-actin consists of polar semiflexible filaments which grow preferentially
at their "barbed" ends. The filaments assemble into supramolecular structures which provide
forces to move cells or form dynamic protrusions along the cell periphery. A widely observed
type of supramolecular structure is a branched dendritic network, where new filaments grow
as branches on existing ones [1, 2]. Several investigations have shown that F-actin in cells
displays spontaneous dynamic behavior including traveling waves and patches [3–6]. The
waves and patches in Dictyostelium typically move at about 0.2 µm/s. A closely related
phenomenon is Hem-1 waves in neutrophils, where Hem-1 acts upstream of actin polymer-
ization [7]. Actin waves, when they impinge the membrane, cause cellular protrusions to
form [8]. Therefore they may be involved in oscillatory and/or random motions of the cell
membrane, which allow a cell to explore its environment. Thus understanding the origins
of actin waves can help clarify the mechanisms by which cells determine key properties of
their motion.
The molecular mechanisms underlying actin patches and waves are not well known. How-
ever, a partial molecular inventory of the waves has been obtained [8], including Arp2/3
complex, which nucleates new filaments by branching, myosin-1B, which may link the actin
network to the membrane, and coronin, which disassembles actin networks. The presence
of Arp2/3 complex suggests that patches could consist of dendritic actin networks; patches
in yeast do in fact consist of such networks [9].
Actin waves and patches might result [3, 10–12] from a reaction-diffusion mechanism
[13], based on spatially varying concentrations of an "activator" and an "inhibitor". The
activator grows at the front of the wave by positive feedback, and spreading (for example by
diffusion) drives the wave forward. The inhibitor then builds up, suppressing the activator
at the rear of the wave. Actin waves and patches could contain positive feedback either in
F-actin ("actin-first"), in membrane-bound filament nucleation-promoting factors which act
upstream of actin polymerization ("NPF-first"), or both. Ref. [11] treated a coarse-grained
model of F-actin dynamics with delayed membrane-induced inhibition and positive-feedback
of F-actin on itself, which could implicitly contain contributions from NPFs. Assuming
spontaneous polarization of actin filaments and diffusion-like spreading of F-actin led to the
formation of patches and waves in appropriate parameter ranges. Refs. [7] and [12] found
2
spontaneous waves of F-actin and the NPF Hem-1 in models based on NPF cooperativity
and inhibition of Hem-1 by F-actin; Ref. [12] also found actin patches under some conditions.
This Letter uses a detailed 3D dendritic network model to treat the spatial dynamics of
F-actin, coupled to a NPF which shuttles between the cytoplasm and the membrane. This
allows a treatment of actin waves and patches based on well-established molecular mecha-
nisms, permits exploration of the possible dynamic phases of F-actin, and enables prediction
of the dependence of F-actin dynamics on key biophysical rates and protein concentrations.
The main finding is that known molecular features of dendritic actin nucleation embody
mechanisms of positive feedback, spreading, and delayed negative feedback which generate
traveling waves and patches. This model differs from that of Ref. [11], in that polarization
of filament orientations is not required. Three broad types of F-actin dynamics are found:
1) traveling waves, 2) moving patches, and 3) random fluctuations with occasional moving
patches.
The simulation method extends one used previously to treat dendritic actin networks
[14, 15]. I treat network growth on a 3µm × 3µm piece of membrane. Explicit coordinates
are stored for each filament. Polymerization, filament nucleation by branching, and capping,
which prevents barbed-end polymerization, are treated stochastically. New filaments are
generated near the membrane as branches on the sides of existing filaments, and by a slower
non-branching nucleation process. Branching is allowed only within 10 nm of the membrane,
so that most branches form near filament tips. The method is extended by including the
following effects (see Supplementary Material for more details):
(i) Filament severing, which is the primary actin disassembly mechanism in the model.
It is also an important actin filament nucleation mechanism [16].
(ii) Attractive filament-membrane forces, with binding energy Eb, required because the
actin network is attached to the membrane [8].
(iii) Translational and rotational motion of filaments, which are important for filament
detachment from the membrane and for "docking" of filament clusters to the membrane.
Because this is very computationally demanding, I simplify the calculation by treating con-
nected dendritic filament clusters as rigidly moving units. The translational cluster motion
has a deterministic part from membrane forces, and a random part from Brownian motion.
(iv) Membrane attachment, detachment, and diffusion of NPFs. Detachment reduces or
eliminates NPF activity, and F-actin enhances the detachment rate [7]. Taking the mem-
3
brane as the x-y plane, the attached (active) NPF density na(x, y) and spatially averaged
detached activator density ¯nd satisfy
∂na
∂t
d¯nd
dt
= −kdetF (x, y)na + katt¯nd + Dmemb∇2na
= −(cid:104)∂na
∂t
(cid:105),
(1)
where kdet, katt are constants, Dmemb is the attached NPF diffusion coefficient, (cid:104)(cid:105) denotes
2D spatial averaging, and F (x, y) is the 2D density of F-actin (see Supplemental Material).
Detached-NPF diffusion is assumed fast enough that nd(x, y) is constant. Because dimer-
ization greatly accelerates the activity of attached NPFs [17], I take the branching rate for
a filament impinging the membrane at (x, y) to be proportional to na(x, y)2.
The parameters (see Supplemental Material) are taken from experiments where available
and otherwise are chosen to obtain a realistic network structure. The on- and off-rates, the
severing rate, the branch dissociation rate, and Dmemb are obtained from experimental data.
The value of [A] is varied within biologically reasonable limits, and the branching rate and
[CP] are chosen to give reasonable network structures; Eb, katt, kdet, and the non-branching
nucleation rate knuc are treated as independent variables.
The simulations begin with only G-actin present and are run out to 400s. Polymerization
peaks and later reaches a steady-state dynamic regime, which is my focus. Fig. 1 shows
snapshots of the 3D actin network and the simulated fluorescence intensity (with a spread
of 200 nm), under conditions favoring wave formation.
I define a wave to be a moving
feature having a large aspect ratio, while features with smaller aspect ratios are patches;
a quantitative description is given in the Supplementary Material. A wave of polymerized
actin moves across the membrane at a speed of about 0.2µm/s. The wave's shape changes as
it moves, due to the randomness of polymerization and detachment. Although the F-actin
features are mainly wavelike, switches to a patchy state and back are sometimes seen in the
simulations (see end of Video 1).
At lower [A], a patch initiates (left frame, Fig. 2) as a localized focus of actin polymer-
ization. It remains stationary for about 4s and then moves to the right at about 0.15µm/s,
maintaining a roughly constant size. The patch forms because G-actin becomes depleted
before a complete wave can be formed. This result is consistent with previous studies of
reaction-diffusion systems [18], which showed that global inhibition such as that mediated
by G-actin depletion causes transitions from waves to patches. The reduced [CP] pictures in
4
FIG. 1: 3D network structure (left) and simulated fluorescence (right) for actin wave (red) propa-
gating across three-micron square area of membrane (green). Time difference between frames is 6
sec. Parameters are: katt = 0.025s−1, kdet = 0.009µM−1s−1, Eb = 2.8kBT , knuc = 0.003µM−1s−1,
[A] = 22.5µM , and [CP ] = 0.25µM . See Supplemental Video 1 for complete time evolution.
Fig. 2 show a more random distribution of F-actin, with some patchy features which move
and give the appearance of broken wave fronts.
The motion of the waves and patches results from three factors operating jointly:
Positive feedback of F-actin. This is casued by autocatalytic branching from existing
filaments, which by itself gives linear positive feedback [19] that is confirmed by in vitro
polymerization dynamics[20]. Branching is modulated by filament detachment from the
membrane. If Eb is small, most single filaments leave the membrane before they branch, so
that formation of a dendritic cluster consisting of several filaments, by a statistical fluctua-
tion, is required before continued growth occurs. The requirement for such a critical cluster
at small Eb values results in a sudden onset of polymerization, which favors wave/patch
formation. This effect is similar to that of nonlinearity in the positive feedback.
Spreading of F-actin. Fig. 3 shows a filament growing from a patch by polymerization
and nucleating several branches, thus providing new polymerization nucleui and spreading
the patch to the right. This mechanism does not require the polarization of filament ori-
5
FIG. 2: Simulated fluorescence of patches obtained with parameters of Fig. 1 but with [A] reduced
to 15µM (top row) or [CP] decreased to 0.10µM (bottom row). Images are 12 sec (top row) and
18 sec (bottom row) apart. See Supplemental Videos 2 and 3 for complete time evolution.
entations suggested in Ref. [11], since a broad range of filament orientations, including the
vertical orientation, can generate new branches in the direction of motion (see Supplemen-
tary Material). However, the patches and waves are polarized in that there are more free
barbed ends at the leading edge.
Delayed negative feedback. This effect is, as mentioned above, grounded in exper-
imental observations. The delay has a large effect on the F-actin dynamics when 1/katt
exceeds the lifetime of polymerized actin, about 5 sec in the present simulations.
These three factors can give rise to traveling waves and, when combined with global
inhibition such as that arising from G-actin depletion, patches [18]. Small values of the
non-branching nucleation rate knuc are also required in order to avoid "parasitic" nucleation
interfering with wave or patch formation.
This model of actin waves and patches makes several experimentally testable predictions:
1) Increasing G-actin concentration leads to a transition from stationary patches, to moving
patches, to waves (cf. Figs. 1 and 2). This prediction is consistent with experiments [6]
that treated Dictyostelium cells with latrunculin, which depolymerized actin by sequestering
6
FIG. 3: Closeup of edge of actin patch at top of Fig. 2 moving to right. Frames are 0.2 sec apart.
G-actin monomers. The latrunculin was then removed from the growth medium. During
the resulting recovery of polymerizable G-actin, the F-actin distributions showed transitions
from stationary patches, to moving patches, to waves. The patch-to-wave transition may
well result from changes in [A], as in the present model. Because intracellular latrunculin
is strongly bound to G-actin, the latrunculin exit required for G-actin recovery could take
minutes or longer. The coexistence of waves and patches seen in Ref. [6] is also consistent
with the wave-patch switching seen in the simulations.
2) Colliding waves annihilate each other, since each depletes the NPF required for the other
to persist. This is consistent with the experiments of Ref. [8]
3)The waves and patches move by a treadmilling mechanism independent of myosin II (cf.
Fig. 3). This is consistent with photobleaching experiments [8], in which a bleached spot
in the F-actin remained stationary during wave motion and myosin II was not required for
wave motion.
4) Arp2/3 complex is uniformly distributed throughout the wave (see Supplemental Mate-
7
FIG. 4: Total number of polymerized subunits Np vs time: a) Baseline parameters as in Fig. 1; b)
[CP] decreased to 0.10µM .; c) knuc increased to 0.03µM−1s−1; d) Eb increased to 4.5kBT ; e) katt
increased to 0.05s−1
rial), consistent with the fluorescence measurements of Ref. [8].
5) Uncapped barbed ends are polarized toward the front edge of patches and waves. This
prediction could be tested by standard techniques for measuring the spatial distribution of
free barbed ends.
6) The actin filament orientations are isotropic rather than being biased along the direction
of motion. This prediction could be tested by cryo-electron microscopy measurements of
wave or patch structure.
7) Downregulation of capping protein, increased nucleation activity in the membrane (as
8
from formins), increased filament-membrane attachment strength (controlled by the protein
VASP [23]), or more rapid reattachment of NPFs to the membrane, will prevent wave
or patch formation. Fig. 4 shows the effects of such interventions on the dynamics of
the number of polymerized subunits Np, used here as an approximate descriptor. Large
fluctuations in Np (as in frame a) correspond to traveling-wave or moving-patch states
seen in the simulation videos, while smaller fluctuations correspond to randomly fluctuating
states. All of the interventions reduce wave behavior. A recent study [21] suggested that
CP enhances branching in vitro, so that reduced CP levels might reduce polymerization.
However, in Dictyostelium reduced CP enhances polymerization [22], consistent with the
assumptions made here.
If CP enhances branching, the effects seen in Fig. 4b will be
reduced.
The main conclusion of this work is that essential mechanisms needed for formation of
traveling waves and patches of F-actin follow from known properties of actin filament nucle-
ation at membranes. Thus the predictions are not limited to the particular mathematical
model employed, but rather follow from the general mechanisms of F-actin positive feedback,
F-actin spreading, and delayed negative feedback. The first two of these follow immediately
from the nature of branching nucleation at membranes, while delayed negative feedback
follows from plausible interactions mechanisms between F-actin and NPFs. In combination,
these mechanisms lead to dynamic features closely paralleling those seen in experiments.
I appreciate informative discussions with Alexander Mikhailov, Martin Falcke, and
Karsten Kruse. This work was supported by the National Institutes of Health under Grant
R01 GM086882.
[1] R. D. Mullins, J. A. Heuser, and T. D. Pollard, Proc. Natl. Acad. Sci. USA 95, 6181 (1998).
[2] T. D. Pollard and G. G. Borisy, Cell 112, 453 (2003).
[3] M. G. Vicker, FEBS Lett. 510, 5 (2002).
[4] M. G. Vicker, J. Expt. Cell Res. 275, 54 (2002).
[5] T. Bretschneider, S. Diez, K. Anderson, J. Heuser, M. Clarke, A. Mueller-Taubenberger,
J. Koehler, and G. Gerisch, Curr. Biol. 14, 1 (2004).
[6] G. Gerisch, T. Bretschneider, A. Mueller-Taubenberger, E. Simmeth, M. Ecke, S. Diez, and
9
K. Anderson, Biophys. J 87, 3493 (2004).
[7] O. D. Weiner, W. A. Marganski, L. F. Wu, S. J. Altschuler, and M. W. Kirschner, PLoS
Biology 5, 2053 (2007).
[8] T. Bretschneider, K. Anderson, M. Ecke, A. M. Muller-Taubenberger, B. Schroth-Diez, H. C.
Ishikawa-Ankerhold, and G. Gerisch, Biophys. J. 96, 2888 (2009).
[9] M. E. Young, J. A. Cooper, and P. C. Bridgman, J. Cell. Biol. 166, 629 (2004).
[10] H. Le Guyader and C. Hyver, C. R. Acad. Sci. Paris. Life Sci. 320, 59 (1997).
[11] S. Whitelam, T. Bretschneider, and N. J. Burroughs, Phys. Rev. Lett. 102, 198103 (2009).
[12] K. Doubrovinski and K. Kruse, Europhys. Lett. 83, 18003 (2008).
[13] A. S. Mikhailov, Foundation of Synergetics I: Distributed Active Systems (Springer-Verlag,
New York, 1990).
[14] A. E. Carlsson, Biophys. J. 81, 1907 (2001).
[15] A. E. Carlsson, Phys. Rev. Lett 92, 238102 (2004).
[16] H. Yamaguchi and J. Condeelis, Biochem. Biophys. Acta 1773, 642 (2007).
[17] S. B. Padrick, H.-C. Cheng, A. M. Ismail, S. C. Panchal, L. K. Doolittle, S. Kim, B. M.
Skehan, J. Umetani, C. A. Brautigam, J. M. Leong, et al., Molecular Cell 32, 426 (2008).
[18] K. Krischer and A. Mikhailov, Phys. Rev. Lett. 73, 3165 (1994).
[19] A. E. Carlsson, Biophys. J. 84, 2907 (2003).
[20] D. Pantaloni, R. Boujemaa, D. Didry, P. Gounon, and M.-F. Carlier, Nat. Cell Biol. 2, 385
(2000).
[21] O. Akin and R. D. Mullins, Cell 133, 841 (2008).
[22] C. Hug, P. Y. Jay, I. Reddy, J. G. McNally, P. C. Bridgman, E. L. Elson, and J. A. Cooper,
Cell 81, 591 (1995).
[23] S. Samarin, S. Romero, C. Kocks, D. Didry, D. Pantaloni, and M.-F. Carlier, J. Cell Biol.
163, 131 (2003).
10
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.