id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
6605e28ce9ebbb4db9d320a5 | 5 | Water molecules were assigned to the nucleic acid hydration shell and to specific sites on the biomolecule following simple geometric criteria (see details in SI). Site-resolved hydration reorientation dynamics was then quantified with the reorientation time τ reor , obtained by numerical integration of the reorientation time-correlation function averaged over the ensemble of water molecules initially assigned to each site. The retardation factor ρ reor quantifies the slowdown in hydration dynamics with respect to the bulk. This strategy allowed us to obtain spatially resolved maps of hydration dynamics of the DNA and RNA double helices under study (Figure ). While earlier simulation studies, with limited statistics, relied on averaging water dynamics over equivalent sites in different base pairs along the sequence, our simulations are long enough to allow examination of converged dynamical properties at a single-site level, as demonstrated by the estimated error bars associated with the computed single-site jump and reorientation times (see SI), that are much smaller than the variations that we will discuss. |
6605e28ce9ebbb4db9d320a5 | 6 | The hydration dynamics of both DNA and RNA helices present common features. They are spatially very heterogeneous: while most of the hydration shell water molecules, in solvent-exposed sites, are only moderately slowed down compared to bulk water (τ reor < 10 ps, i.e. ρ reor < 6), water molecules located in the narrow grooves (DNA minor groove and RNA deep major groove) exhibit a dynamics up to 15-20 times slower than the one observed in the bulk. These findings are consistent, for DNA, with NMR MRD measurements 18 that evidenced a moderate 6-fold slowdown in most of the hydration shell, and the presence of a few water molecules in the grooves with longer ∼ 200 ps residence times. It contrasts with the bulk-like diffusion dynamics suggested by Overhauser effect dynamic nuclear polarization, which provides an estimate of the translational dynamics of water molecules within about 10 Å of the probe. The discrepancy between these recent experiments and other experimental and computational techniques 53 still remains to be explained, and could be related either to the impact of the probe on water dynamics, or to some assumptions used in the model to interpret the data. Interestingly, as noted in earlier computational studies, the location of the slowest water molecules differs between DNA and RNA sequences: in DNA, they are found at the floor of the groove, where they are H-bonded to DNA bases, whereas in RNA they lie on the side of the major groove, next to the phosphate groups. |
6605e28ce9ebbb4db9d320a5 | 7 | A better resolution is offered by the examination of H-bond jump times, which also allows examining the molecular determinants of the hydration dynamics spatial heterogeneity. As previously shown, water reorientation occurs mainly through large angular jumps that occur during exchanges of H-bond partner. Since a water molecule needs to exchange H-bond acceptor to move away from an initial H-bond acceptor site on the DNA or RNA surface, the jump time (see Supporting Information) reports on the H-bond dynamics at a very well defined location, or "site". Maps of the jump times on the nucleic acids surface (see SI Fig. ) show very similar characteristics as that of reorientation times, confirming indirectly that the jump is indeed the main ingredient of water reorientation in the hydration shell. |
6605e28ce9ebbb4db9d320a5 | 8 | To gain further insight into the spatial heterogeneity and examine the impact of different features of the DNA/RNA environment on the hydration dynamics, we split the overall jump time distributions (Figure ) according to the nature of the nearby nucleic acid site: H-bond acceptor sites (e.g., phosphate groups, O or N atoms on the base), H-bond donor sites (e.g., ribose OH groups), and hydrophobic groups (e.g., -CH 2 -groups). In line with the findings from earlier studies on proteins and DNA, the main peak of the jump time distributions is located at around τ jump ≃ 6 -7 ps, only moderately slowed down with respect to bulk water τ bulk jump ≃ 2.5 ps (i.e. a retardation factor of about ρ jump = 2.5 -3), and mostly comes from water molecules located next to the biomolecule hydrophobic and H-bond donor groups (Figure ). We specifically examined the hydration dynamics of the ribose 2 ′ OH group in RNA, where early structural studies have suggested long-lived hydration patterns, 31-33 that were not observed in subsequent molecular dynamics studies. Our simulations show no sign of long-lived hydration patterns, and the hydration dynamics at these sites is only moderately slowed down, with a jump time of about 5-6 ps (i.e. ρ jump = 2 -2.5). |
6605e28ce9ebbb4db9d320a5 | 9 | We now focus on the sites with slower hydration dynamics, where differences between This split between RNA phosphate O R and O S hydration dynamics has been noticed in previous works, with, however, very long residence times, of respectively 500 and 700 ps for phosphate O S and O R , two orders of magnitude slower than our jump times. |
6605e28ce9ebbb4db9d320a5 | 10 | While we report reorientation times (in agreement with NMR) and H-bond exchanges, these earlier works consider residence times-which rely on criteria (e.g., site definition, treatment of transient escapes) that can significantly affect the results All the key identified features-overall shape of the jump time distributions, nature of the sites with slowest dynamics, split in hydration dynamics of RNA phosphate oxygensare very similar in the two studied sequences, GGGG and GCGC (see Fig. and Fig. ). |
6605e28ce9ebbb4db9d320a5 | 11 | The average jump time at key hydration sites (e.g., 2 ′ OH, phosphate O S and O R ) is identical in both sequences within the computed error bars (see SI)). Interestingly, the only significant difference with alternating base pairs is observed at the slow O2(C) on the floor of the DNA minor groove, with an average jump time of 24.4 ± 1.1 ps in GGGG and 31.2 ± 2 ps in GCGC. This sequence dependence of the dynamics at slow sites in the groove is here specific to DNA, which is consistent with previous work. The slow hydration sites, phosphate groups and the floor of narrow grooves are also key interaction sites with ions and proteins, and thus often involved in the formation of protein-nucleic acid complexes and drug binding events. We thus want to go beyond the high-resolution site-level characterization of differences between DNA and RNA hydration dynamics. We now seek a physical understanding of the molecular origin of these differences, focusing on the slow sites and starting with phosphate hydration, where we found distinctly slower hydration dynamics next to phosphate O R sites in RNA. In A-RNA, water molecules have been found to bridge neighbor phosphate groups, and suggested by simulations to form long-lived water bridges. In our simulations, successive phosphate O R oxygen atoms along each strand are indeed bridged by a single water molecule 43% of the time. However, these water bridges are not responsible for the slow hydration dynamics encountered around the phosphate groups, since the jump dynamics of water molecules initially H-bonded to a phosphate O R atom are very similar irrespective of whether the second hydrogen atom is H-bonded to the neighbor phosphate group (bridge conformation) or to water (not bridge) (see SI Fig. ). In other words, the H-bonds formed by a water molecule in bridge between two phosphate groups are not longer lived than those of other water molecules bound to the phosphate O R . This finding nuances the usual picture of long-lived water bridges between RNA phosphates. However, the location of the "bridging" water molecules is much better defined by the 2 H-bonds with the phosphate backbone compared to singly-bound water molecules, which leads to a higher water density at those specific locations and explains their better resolution in X-ray experiments. Since the peculiar slowdown at RNA O R sites is not due to long-lived water bridges, we now use our jump model for water reorientation dynamics to identify the physical determinants of the slowdown next to these sites in DNA and RNA, and understand which physical factors cause the split in water reorientation dynamics between the two phosphate oxygen atoms in RNA. Previous works have shown that the effect of the environment on the jump time can be quantified through two main factors, ρ = ρ T SEV × ρ T SHB (details in SI). The entropic Transition-State Excluded Volume (TSEV) factor, ρ T SEV , quantifies the slowdown due to the presence of a nearby solute that hinders the approach of the new acceptor. The Transition State Hydrogen Bond (TSHB) factor ρ T SHB describes how the nature and restraints imposed by the solute modulate the free energy cost to elongate the initial H-bond to its transition-state geometry. |
6605e28ce9ebbb4db9d320a5 | 12 | The retardation factor predicted by our extended jump model compares well with that effectively observed in our simulations ρ jump (Fig. ), which demonstrates the validity of the model. While our model slightly underestimates the absolute value of the retardation factor, it captures almost quantitatively (see Table ) the split in hydration dynamics between the phosphate O R and O S atoms, both in RNA and DNA, which is key for our purpose here. Note that explicitly taking into account the excluded volume coming from background potassium counterions is required in RNA to capture the full extent of the split (see Table ) and recover a better estimate of the absolute value of the retardation factor, while it has only a very minor impact in DNA. The decomposition into the different molecular factors determining the jump time (Fig. ) shows that the split in RNA between the hydration dynamics of the two phosphate oxygen atoms originates both from the larger excluded volume next to O R -coming both from the RNA and the counterions-and from the stronger initial H-bond at this oxygen. |
6605e28ce9ebbb4db9d320a5 | 13 | These two effects can be linked to the distinct helical conformations of B-DNA and A-RNA helices, as the orientation of the phosphate groups strongly differs between the two double helical conformations (Figure ). The phosphate group points away from the helix principal axis in B-DNA, whereas it is more parallel to the axis in A-RNA (see α OP 2 angle in Figure and). This results in different orientations of the phosphate oxygen atoms: |
6605e28ce9ebbb4db9d320a5 | 14 | in DNA, the two oxygen atoms point outside the groove, thus resulting in similar hydration dynamics for the two phosphate oxygens due to similar environments, whereas in RNA, O R strongly points towards the inside of the groove (see Figure with very small values of α P R ). This logically results in larger excluded volumes (larger ρ T SEV ) at O R . It also imposes stronger restraints on the elongation of the initial water-phosphate H-bond-the H-bonded water in this case being confined inside the groove-thus leading to a larger ρ T SHB , and overall to a significantly more pronounced slowdown. |
6605e28ce9ebbb4db9d320a5 | 15 | In A-RNA, the close proximity of phosphate groups on both sides of the major groove, as visible on Figure and quantified by the width of the RNA major groove-which is even narrower than the DNA minor groove (see SI Figure )-also leads to a different structure of the ion cloud around RNA and DNA (see Fig. ), as already described in previous works. In particular, we note in RNA a larger density of K + counterions next to the phosphate groups, in the major groove and close the O R atom. In DNA, the cations are mainly at the bottom of the major groove next to the bases and along the helix principal , and that predicted with different components of the Extended Jump Model: ρ T SHB , ρ T SEV or ρ T SEV × ρ T SHB . ρ T SEV is calculated taking only into account the nucleic acid excluded volume (filled circles), or included that coming from the ions (empty circles). Plots were prepared using Matplotlib. axis, with a much smaller density next to the phosphate groups. This explains why inclusion of the excluded volume due to the counterions was necessary to capture the full extent of the slowdown and the split in hydration dynamics between the two phosphate oxygen atoms in RNA, but not in DNA. Note however that the increase of the slowdown due to the ions predicted by our jump model, while noticeable for RNA O R , remains very modest (factor 1.1 and GGGG-RNA (e) (light blue for the angle α P R and blue for the angle α P S ). The plots were prepared using Matlab. on average). In addition, the decay of the jump correlation function is independent of the initial presence of a counterion in the water hydration shell (see SI Figure ), which shows that the memory of the ion presence is lost faster than the typical jump time. Therefore, the presence of slow water molecules cannot be ascribed to that of ions near these locations. Now that we have rationalized the differences in hydration dynamics at the phosphate groups, we use our jump model to investigate why the acceptor sites with slow hydration dynamics in the DNA minor groove (cytosine O2, τ jump ∼24-31 ps), have a much faster hydration dynamics in RNA, where they are on the floor of the shallow groove (τ jump ∼9-10 ps) (see SI Figure ). There, the jump model does not work quantitatively, because it focuses on jumps to (bulk) water, whereas at these confined locations in narrow grooves about half of the jumps occur towards another acceptor site of the biomolecule (see SI for a detailed discussion). However, our model does qualitatively capture the difference in dynamics at these sites between DNA and RNA: |
6605e28ce9ebbb4db9d320a5 | 16 | The larger retardation factor predicted from our jump model at these sites in DNA comes both from a larger excluded volume at these sites (ρ DN A T SEV /ρ RN A T SEV ∼ 1.5) and from a stronger initial Hbond (ρ DN A T SHB /ρ RN A T SHB ∼ 2), so that jumps to water acceptors are slowed down. In addition, jumps to nucleic acid acceptor sites (mainly N3(G)) also appear less favored in DNA then RNA. This can also be linked to the different double helical geometries, that position N3(G) a bit further from O2 in B-DNA than A-RNA (on average 4 vs 3.6 Å). |
6605e28ce9ebbb4db9d320a5 | 17 | In contrast with this successful determination of the molecular origin of changes in hydration dynamics, the jump model cannot rationalize the slower dynamics observed in DNA at the O2(C) sites on the floor on the minor groove for the GCGC sequence (ρ jump (GCGC)/ρ jump (GGGG) ≃ 1.3), as the computed TSEV factor (with or without ions) predicts the reverse ordering (ρ T SEV (GCGC)/ρ T SEV (GGGG) ≃ 0.8). This shows the limits of our Jump Model which was not designed to work in such confined and restrained environments. Rationalizing at the molecular level the sequence-dependence at slow sites on the DNA minor groove would deserve additional investigation in future work. |
6605e28ce9ebbb4db9d320a5 | 18 | In summary, we obtained an unprecedented high resolution mapping of the hydration dynamics around analogous ds-RNA and ds-DNA 18-mers, which allowed us to highlight, beyond apparently similar features with an average moderate slowdown, key differences. We rationalize these differences and obtain a physical understanding of their molecular determinants thanks to an analytic jump model. In RNA, the slowest water molecules belong to the phosphate hydration, while they are located close to the bases in the minor groove in DNA. |
6605e28ce9ebbb4db9d320a5 | 19 | Focusing on phosphate hydration, the two phosphate oxygen atoms O R and O S exhibit very distinct dynamics in RNA, which is not the case in DNA. The specific slowdown of O R hydration is not due to the presence of water molecules bridging successive phosphate groups, that have similar dynamics to that of non bridging hydration water at these sites. Instead, we showed, using our Extended Jump Model, that it is caused in RNA by both a larger excluded volume and a stronger initial H-bond next to O R . These two factors originate from the markedly different phosphate orientations in RNA, which result from the different double helical conformation of B-DNA and A-RNA. The same framework was used to rationalized the absence of slow sites in the RNA grooves. |
6605e28ce9ebbb4db9d320a5 | 20 | While this work has focused on two specific GC-rich sequences, the main features of DNA and RNA hydration dynamics (e.g., heterogeneity and location of fast or slower hydration sites, split of RNA phosphate hydration dynamics, nature of the slowest sites) have been linked to the overall shape of the double helix (A-form vs B-form) rather than to specific chemical features of the bases, so we expect all these observations to be similar in AT/Urich sequences. The main difference can be expected at slow sites at bases on the floor of the DNA minor groove, where we already see here a difference between GGGG and CGCG sequences. In addition, we know from previous work that AAAA-rich sequences have narrow minor grooves, where we expect much slower hydration dynamics. The dynamics in the grooves could be further sensitive to small changes in helical conformation, that can be experimentally modulated for instance by the background salt concentration and nature, and in simulations can depend on the force field used for water, nucleic acids and ions. Hydration dynamics at these sites and its sequence-dependence will be investigated in future work. |
6605e28ce9ebbb4db9d320a5 | 21 | The molecular-level understanding of RNA hydration dynamics provided by such approaches will be important in a number of contexts. These include drug binding and the formation of protein-nucleic acid complexes -key for the regulation of gene expression-as well as the the formation of RNA-protein biomolecular condensates through liquid/liquid phase separation, as these processes all involve dehydration of the slowest RNA hydration sites, phosphate groups and grooves. Depending on the relative timescale of binding and water dynamics, hydration dynamics can play an important role in those essential biochemical processes. |
67d95a696dde43c9083f7c98 | 0 | RNA has for long been considered as one of the most essential biomolecules necessary for the emergence of life on Earth, owing to its capability to carry genetic information and perform catalytic activity such as generation of activated nucleotides and polymerisation of nucleic acids . Considering that synthesis of complex proteins first required the emergence of RNA-based translation , it is likely that early RNA sequences operated in the absence of protein-based enzymes. Therefore, self-replication is considered as one of the most fundamental functions of prebiotic RNA sequences and, in fact, it was demonstrated that ribozymes can not only catalyse template-directed RNA primer extension but also support the evolution of catalytic RNA assemblies . However, before the formation of complex rybozymes, first short RNA oligomers could only undergo a more primordial form of nonenzymatic template copying, which enables the incorporation of activated nucleotides into the copied strand with minimal catalytic support of metal cations. |
67d95a696dde43c9083f7c98 | 1 | One of the first successful examples of nonenzymatic RNA self-replication was demonstrated by Lohrmann and Orgel in 1980 , in which they used guanosine-5 ′ -phosphoroimidazolides to synthesize oligoguanylic acids on the poly(C) template. Unlike nucleotide triphosphates, phosphoroimidazolides proved to offer much higher reactivity and enabled the so-called primer extension reaction with the catalytic support of Pb 2+ and Zn 2+ cations . This process was improved a year later by applying a more reactive 2-methylimidazole activating group instead of imidazole . However, this improvement did not resolve the original challenge of copying adenosine and uridine, which almost completely blocked further RNA strand synthesis. Enzyme-free replication of nucleic acids has been extensively studied by numerous groups since then. These efforts included further developments of nonenzymatic primer extension of canonical and alternative nucleic acids with phosphoroimidazolides on different heterogeneous templates or immobilization of RNA and other or other forms of activation, like adenine derivatives . It included also various approaches to templated ligation of RNA oligonucleotides, involving condensing agents and organocatalysis . The computational work presented here focused on several variants of nonenzymatic template copying involving phosphoroimidazolides, hence, we offer more background on the associated scenarios below. |
67d95a696dde43c9083f7c98 | 2 | The rate and fidelity of RNA self-replication based on nucleotide phosphoroimidazolides was further improved by the consideration of short helper oligonucleotides, which could bind downstream of the activated nucleotide monomers which take part in the primer extension reaction . The initial mechanistic rationale was that such a downstream oligonucleotide (or even mononucleotide) could bring closer phosphate group of incoming nucleotide to the 3 ′ end of a primer by conformational constraints inducing the formation of 3 ′ -5 ′ phosphodiester linkages . Furthermore, it was demonstrated that monomers and and helper oligonucleotides activated with 2-aminoimidazole yielded the highest rates of primer extension reaction (Fig. ) and b)). However, 2-aminoimidazole itself did not improve the rates and fidelities of nonenzymatic RNA self-replication of the A and U components. |
67d95a696dde43c9083f7c98 | 3 | 2,6-diaminopurine and 2-thiouridine . In particular, 2,6-diaminopurine can interact with canonical uridine (or its C5-(1-propynyl) derivative) by forming a third hydrogen bond in the Watson-Crick base pairing pattern which could result in stronger binding and increased self-replication rate . Similarly, greater thermodynamic stability of the 2-thiouracil:adenine base pair (Fig. ) and g)) was suggested as the reason for improved kinetics and fidelity of nonenzymatic RNA template copying with 2-thiouridine monomers being present both in the templating strand as well as among the activated monomers . More recently, Kim and co-workers demonstrated that highly efficient nonenzymatic RNA template copying could be achieved with activated inosine monomers (Fig. f), while activated 8oxopurine nucleotides proved to perform poorly under the same conditions. It is worth emphasizing that as opposed to the G:C Watson-Crick base pair, inosine interacts with cytosine through only two hydrogen bonds and not three. Therefore, the observed comparable performance of inosine and guanosine monomers challenges the assumption that nonenzymatic RNA template copying is largely regulated by the thermodynamic stability of specific base pairs and the molecular mechanism of the process is more complex. |
67d95a696dde43c9083f7c98 | 4 | Mechanistic aspects of nonenzymatic RNA self-replication involving phosphoroimidazolide monomers were previously investigated with X-ray crystallography at different stages of the primer extension process . The acquired crystal structures confirmed the previous interpretation that the process involves the formation of imidazolium bridged dinucleotide intermediates and also offered information about the binding mode and conformation of activated guanosine nucleotides . However, X-ray crystallography does not offer any information about the dynamical behavior of the system and about minor conformations which might play an important role in the mechanism. Furthermore, it is limited to systems which can be readily crystallized, which is often challenging for nucleic acids with non-biological components . Therefore, to offer a comprehensive description of the mechanism of non-enzymatic RNA template copying, a more precise examination of the structures and dynamics of the binding of activated nucleotides to the template is necessary. |
67d95a696dde43c9083f7c98 | 5 | In this article, we report on a computational investigation of the binding of activated ribonucleotides in nonenzymatic RNA template copying. In particular, we employ molecular dynamics (MD) simulations based on classical force fields to investigate the binding of different imidazole-bridged dinucleotide intermediates to the templating strand, just below the 3 ′ -end of the copied strand. These results demonstrate that the rate and fidelity of nonenzymatic template copying is governed by more subtle structural and dynamical effects than simply the number of hydrogen bonds in the Watson-Crick base pairing pattern, or the thermodynamic stability of a given base pair. |
67d95a696dde43c9083f7c98 | 6 | We performed classical molecular dynamics simulations for 9 variants of a self-replicating RNA system using the crystal structure structure deposited by Zhang et al. [PDB:6C8K] as the initial model system. Our self-replicating RNA model system consists of a templating strand having 14 nucleotides, the copied strand of 12 nucleotides and a imidazole-bridged dinucleotide intermediate located near the 3 ′ -end of the copied strand, which interacts with the terminal nucleotides of the templating strand through Watson-Crick base pairing (see Fig. ). The 9 variants of this system include different compositions of the imidazolium-bridged dinucleotide intermediate including constituent nucleotides such as: guanosine, cytidine, uridine, adenosine, inosine and 2-thiouridine (see Fig. ). Each of these variants also contained the appropriate complementary nucleotides on the templating strand. After the equilibration runs which maintained all the Watson-Crick base pairing interactions, we performed production MD simulations with the cumulative time of 1.5 µs for each of these systems (more details can be found in the Methods section). |
67d95a696dde43c9083f7c98 | 7 | For each of the systems, we next performed clustering of the structures present in the trajectories based on RMSD to evaluate and quantify the contribution of different binding modes. For this purpose we used two different algorithms, namely DBSCAN and K-means. The DBSCAN algorithm reveals clusters characterized by a broader average distance between points within each cluster (AvgDist) compared to the K-means analysis. This allows for a more comprehensive overview of the data. However, K-means provides more detailed insights, offering a higher level of specificity in the clustering results. We also performed additional structural analyses of the trajectories which offer further insights into the possible reactive conformations which can enable the primer extension reaction. We start the discussion with the clustering analysis below. Within this discussion we put emphasis on the binding site and the activated dinucleotide intermediates with the naming scheme of the key nucleotides presented in Fig. . In particular, we focused the analyses on the structural arrangements of three key nucleotides, namely the nucleotide located at the 3 ′ -end of the templating strand, the nucleotide of the imidazolium bridged dinucleotide intermediate that is adjacent to it (named simply adjacent), and the remaining nucleotide of the dinucleotide intermediate referred to as terminal (Fig. ). |
67d95a696dde43c9083f7c98 | 8 | We first focused on simulating the parent system which contains the imidazolium-bridged guanosine dinucleotide (GpAIpG) as in the original crystal structure acquired by Zhang et al. . Activated guanosine nucleotides were previously shown to offer the highest rate and fidelity in nonenzymatic primer extension reactions and here we this system as the reference point in the structural analyses. The geometry clustering analysis based on the DBSCAN algorithm found only one cluster which was populated for 86.0% of the simulation time. The structure with the lowest cumulative distance to all remaining frames within this cluster (our representative frame) is characterized by the anti conformation of both guanine nucleobases of GpAIpG and the C3 ′ -endo conformation of the corresponding sugar moieties. GpAIpG is bound to the template with Watson-Crick (WC) base pairing interactions and it is additionally stabilized by stacking with the guanine nucleobase in the 3 ′ -end of the copied strand (see Fig. )). The K-means clustering analysis revealed several cluster structure which differ by minor details, but possess the same conformational features and binding mode as the structure representing the cluster identified with the DBSCAN algorithm (see Supporting Information Fig. ). The overall population of such bound structures from the K-means procedure closely resembles the population of the main cluster from DBSCAN (94%). Overall, these results demonstrate that high rates and fidelities of nonenzymatic template copying are well reflected by stable binding of the dinucleotide intermediate through Watson-Crick base-pairing interactions and high population of the corresponding bound structures. |
67d95a696dde43c9083f7c98 | 9 | For comparison, we also considered the opposite configuration with the imidazolium bridged cytosine dinucleotide intermediate (CpAIpC) and two complementary guanosine nucleotides in the templating strand. The DBSCAN analysis returned a single cluster with the population of 99.9% (Fig. )). As demonstrated through the representative frame, the adjacent cytosine nucleobase is positioned at the binding site, forming WC hydrogen bonds with the complementary guanine nucleobase. However, the terminal cytidine adopted a non-canonical orientation (an O-stacked structure, see below), which maximized the stacking between the two cytosine nucleobases of CpAIpC. Consequently, the terminal nucleotide of CpAIpC interacts with the complementary guanine through two hydrogen bonds: one connecting the N-H group of guanine with the carbonyl oxygen of cytosine, and the other one connecting the 2 ′ -OH group of cytidine with the carbonyl group of the guanine nucleobase on the template. Both nucleobases of CpAIpC adopted the anti conformation, with the adjacent cytidine having a C2 ′ -endo conformation of the sugar, while the terminal nucleotide was in C3 ′ -endo conformation. It is important to emphasize, though, that this frame presented in Fig. ) does not perfectly represent all of the structures included in this cluster. Consequently, the K-means allowed us to better discriminate between the different arrangements of CpAIpC bound to the template, which particular emphasis on the terminal nucleotide. In particular, we found similar structures in the O-stacked geometry, and related conformers having the templating terminal guanosine nucleotide rotated further away from the CpAIpC dinucleotide (see Supporting Information Fig. ). Additionally, K-means identified two clusters, representing 31.4% of all the analysed frames, which closely resemble the crystal structure of GpAIpG (Fig. )) and maintain WC interactions as well as. In these clusters, the ribose moieties are predominantly in the C2 ′ -endo conformation, with the exception of the terminal nucleobases in one cluster, along with the anti orientation of cytosine We next considered two systems with mixed activated dinucleotide intermediates which consisted of one cytidine and guanosine monomer, namely CpAIpG and GpAIpC. The former one contained cytidine as the adjacent nucleotide, while GpAIpC was oriented in the opposite direction on the template. In both cases we included the complementary WC nucleotides on the templating strand. For CpAIpG, K-means algorithm identified clusters with structurally similar arrangements, from which three prominent structural patterns can be distinguished (Fig. )-f)). The major pattern populated in 41.8% resembles canonically stacked GpAIpG along with anti configuration of nucleobases and WC hydrogen bonding interactions with the complementary nucleobases on the templating strand. In this case, the dominant ribose conformation within the CpAIpG dinucleotide is C3 ′ -endo. Another structural pattern populated in 32.7%, which can be seen for two other clusters from the K-means analysis involves the terminal activated guanosine nucleotide displaced sideways from the templating strand, with no stacking interactions between the two terminal nucleotides within the template. The final structural motif populated in 21.5% is characterised by the nucleotide at 5 ′ terminus of the templating strand acting as a loose overhang and the activated guanosine nucleotide of CpAIpG taking its place in the stacking interaction with the second to the last guanosine nucleotide of the template. Importantly, in all of these clusters the activated adjacent cytidine nucleotide of CpAIpG maintains a WC base pairing interaction with its partner on the template. |
67d95a696dde43c9083f7c98 | 10 | In the reverse system (GpAIpC), the K-means clustering analysis showed that 30.0% of the structures sampled canonical structure of the GpAIpC dinucleotide involving WC base pairing interaction with the complementary nucleotides on the templating strand which resembles the major con- ). An alternatively folded conformation of the GpAIpC dinucleotide is populated in 19.1% and is referred to as an O-stacked structure, due to its ring-like, circular arrangement (see Fig. ) and Fig. )). Consequently, the binding of the activated GpAIpC dinucleotide to the template is clearly less specific than for the CpAIpG dinucleotide discussed above. |
67d95a696dde43c9083f7c98 | 11 | To compare the above systems to activated dinucleotide intermediates which were suggested as weaker binders, we next investigated the binding of activated uridine dinucleotide intermediate (UpAIpU) to adenosine nucleotides on the template containing complementary adenosine residues. The structure of the UpAIpU dinucleotide representing the highest populated cluster generated by the DBSCAN algorithm (68.3% of the population) involves only one WC base-pairing interaction which is formed between the adjacent uracil of UpAIpU and the complementary adenine (see Fig. )). The terminal uracil of UpAIpU is oriented away from the interaction site and forms the so-called T-shape interaction with the complementary adenine nucleobase of the template, where the terminal uracil is arranged in a perpendicular manner with respect to it . Both uracil nucleobases exhibited the anti orientation with respect to the ribose moieties, with terminal nucleotide in the C2 ′ -endo conformation and the adjacent adopting the C3 ′ -endo conformation. In the second highest cluster from the DBSCAN analysis (28.8% of population), and also four clusters from the K-means analysis (total population of 29.9%), the UpAIpU dinucleotide diffused away from the binding site and interacted with the sugar phosphate backbone on the opposite side of the RNA double helix (Supporting Information Fig. and). Importantly, the K-means clustering procedure revealed that even though the UpAIpU dinucleotide can be found near the interaction site for ∼69% of the time, its structure tends to largely sample random coil arrangements or O-stacked structures. Reasonable recognition for at least one A and U pair was found for merely 14.5% of the simulation frames (Supporting Information Fig. and). |
67d95a696dde43c9083f7c98 | 12 | In the case of the opposite system, containing the imidazolium bridged di-adenosine intermediate (ApAIpA) with two complementary uridine nucleotides within the templating strand, we did not observe any clusters that could enable accurate recognition for both WC contacts at the same time. For all the identified clusters with the DBSCAN and Kmeans procedures the 5 ′ -end of the templating strand was largely disordered, often sampling random coil arrangements. The K-means clustering algorithm returned one cluster char-acterized by a Hoogsteen base pairing interaction of the adjacent adenine nucleobase and a WC base pair formed by the terminal nucleobase of ApAIpA with the template. However, this cluster represents only 8.0% of the analyzed trajectory (Supporting Information Fig. ). Formation of only one WC interaction between ApAIpA and the template can be assigned to merely 31.2% of the analyzed trajectory frames, which demonstrates lower binding specificity of this activated dinucleotide when compared to the GpAIpG parent system. Instead, 67,5% of the frames from K-means analysis displayed G-stacked structure, similar to what was observed for GpAIpC (Fig4 e)) and 19.5% adopted the O-stacked geometry (Fig4 f)). In the average structures representing these motifs, the C3 ′ -endo conformation was predominantly observed, accompanied by the anti orientation of the nucleobases. |
67d95a696dde43c9083f7c98 | 13 | To evaluate the binding of activated dinucleotide intermediates containing alternative nucleobase we next turned to systems containing inosine (I) and 2-thiouridine (2-tU) nucleotides. We started the investigation with the system containing an imidazolium-bridged inosine dinucleotide intermediate (IpAIpI) interacting with two cytidine nucleotides at the 5 ′ -end of the templating strand. The highest populated cluster generated by DBSCAN analysis (55.5%) is associated with a structure that closely resembles the canonically stacked GpAIpG dinucleotide (Fig. )). Consequently, both inosine residues exhibit a canonical Watson-Crick basepairing interaction, anti conformation of the nucleobases and C3 ′ -endo conformation of the ribose fragments. This canonical structure is also represented by 2 clusters from the K-means procedure with the overall population of 22.8% (see Supporting Information Fig. ). In addition, we see a single related cluster charaterised by a less specific interaction formed between the terminal hypoxanthine (inosine) and cytosine nucleobases, with the WC base pair retained for the adjacent inosine residue (14.4%). However, a considerable proportion of clusters from the K-means procedure (27.7% of the trajectory) are characterized by nonspecific interactions of the backbone of IpAIpI with the complementary strand which do not allow for selective recognition of the activated inosine nucleotides and result in higher distances of the adjacent phosphate group of IpAIpI from the C3 ′ terminus of the primer. The latter feature was suggested to effectively hinder the primer extension reaction . Given that, experimentally activated inosine nucleotides were observed to undergo practically as efficient primer extension reactions as guanosine nucleotides, the observed lower population of the reactive conformation for IpAIpI, requires further analysis and considerations. Importantly, the clustering analysis based on the K-means algorithm revealed another structural arrangements of the IpAIpI dinucleotide (Fig. and)), which includes the hypoxanthine nucleobases folded into one another. We denoted one of these as G-stacked, as its structure resembles the shape the capital letter G (Fig. a)) while the other is termed O-stacked (Fig. )). The Gand O-stacked arrangements of IpAIpI are usually associated with the formation of a single WC base pair for the adjacent inosine nucleotide. |
67d95a696dde43c9083f7c98 | 14 | In addition, we performed simulations involving the substitution of uridine nucleotides with 2-tU both for the dinucleotide intermediate and the template. The system containing the ApAIpA dinucleotide with two tU nucleotides on the template exhibited much more specific binding than its counterpart containing biological uridine nucleotides within the template. In particular, we identified two clusters characterised by WC base pairing interactions for both components of ApAIpA with the total population of 19.4%, based on the K-means procedure (Fig. )). Similar to the parent system containing the ApAIpA dinucleotide, the tU modification of the template still resulted in substantial population of G-stacked structures (77.6%), part of which also involved a single WC base-pairing interaction of the adjacent A nucleobase (18.5%). Qualitatively, the results demonstrate that substituting the oxygen atom in the 2 position with a sulfur atom stabilizes the ApAIpA dinucleotide on the template and improves the specificity. This observation is consistent with the experimental data . |
67d95a696dde43c9083f7c98 | 15 | We also considered the reverse structural arrangement containing the imidazolium bridged 2-thiouridine dinucleotide (tUpAIptU) intermediate and two adenosine nucleotides in the complementary position of the templating strand. However, both clustering algorithms returned clusters with very unspecific binding of the tUpAIptU dinucleotide, which either interacts with the sugar phosphate backbone of the major grove of RNA double helix (23.9%) or it partly stacks with the adenine nucleobases at 5 ′ -end of the templating strand (32.3%). The clusters found with the K-means procedures display the O-stacked arrangement of tUpAIptU being dominant, and the canonical arrangement of the dinucleotide was found only for one cluster with 6.9% of population. This latter cluster also involves a single WC base pair between the adjacent tU nucleobase and the complementary adenine nucleobase. Overall, the very low binding specificity of tUpAIptU found in our simulations seems inconsistent with the improved rate and fidelity of nonenzymatic RNA template copying observed experimentally . This result seems somewhat surprising when compared to increased rate and fidelity of nonenzymatic template copying involving activated tU nucleotides. This inconsistency could originate from weaknesses of the nucleic acid force field, particularly since our extended parameter library includes non-canonical phosphoroimidazolide groups and alternative nucleobases, which still require much more thorough benchmarking and testing. |
67d95a696dde43c9083f7c98 | 16 | To provide a more quantitative overview of the structural arrangements and interactions formed between the activated dinucleotides and the templating strand, we carried out additional global analyses of the trajectories. Consequently, we evaluated the percentage of Watson-Crick base pairing interactions, the conformations of the sugar rings within the activated dinucleotide and the terminal nucleotide of the primer as well as the distance between its associated 3 ′oxygen and the phosphorus atom of the adjacent nucleotide. As explained below, these structural parameters are of particular importance for the rate and fidelity of nonenzymatic RNA template copying. We first focused on the population (in %) of WC base pairing interactions near the dinucleotide binding site, with the focus on the two activated nucleotides and their partners within the templating strand. The NH• • • N hydrogen bond distance in such interactions was estimated to 2.02 Å, while NH• • • O bonds were shown to reach up to 2.17 Å . Given that crystal structures are affected by crystal packing effects and do not account for the oscillation of hydrogen bonding contacts we applied slightly more relaxed thresholds of 2.2 Å for NH• • • N distances and 2.5 Å for NH• • • O distances, to identify frames containing WC interactions for the two nucleotide pairs. In the case when all of the necessary hydrogen bonding distances were below these thresholds, given structure was counted towards the population of WC base pairing interactions. Otherwise the WC interaction was treated as absent. This analysis illustrated in Fig. demonstrates that the GpAIpG dinucleotide achieved the highest overall percentages, with 89.0% for adjacent nucleobase and 81.9% for terminal nucleobase. Similarly, the adjacent nucleotides of CpAIpC, CpAIpG and GpAIpC dinucleotides exhibited a comparably high population of the WC base pairing interactions, namely, 76.7%, 82.8% and 82.5%, respectively. However, despite the fact that the nucleobases of these three dinucleotides are involved in three WC hydrogen bonding interactions, the terminal nucleotides exhibited much lower degree of WC base pairing (24 -31%) than in the case of the parent GpAIpG system. This indicates that proper binding of the imidazole-bridged dinucleotide intermediate is in fact supported by stacking interactions. The highest degrees of stacking can be generally achieved for dinucleotides cotaining only purine nucleobases, while we showed that in the presence of pyrimidine nucleobases the dinucleotides tend to fold into O-stacked and often G-stacked structures. |
67d95a696dde43c9083f7c98 | 17 | As expected, the percentage of WC interactions with the template calculated for the adjacent residues of the ApAIpA and UpAIpU dinucleotides were much lower, that is, 13.8% and 12.4%. The associated terminal nucleotides of ApAIpA and UpAIpU exhibited negligible WC base pairing. However, as demonstrated for the ApAIpA dinucleotide, the inclusion of complementary (alternative) tU nucleotides within the template increased the population of WC base pairs up to 32.8% for the adjacent and 12.3 % for the terminal nucleotides. The opposite system containing the tUpAIptU dinucleotide is characterised by the lowest population of WC base pairing interactions for both nucleotides, with 8.8 % for the adjacent nucleotide and virtually non formed for the terminal nucleotide. Given that activated tU nucleotides have been generally shown to improve the rate and fidelity of nonenzymatic RNA template copying, the outcome of our final simulation is inconsistent with the experimental observations. Similar to the clustering results, this indicates potential imbalances in the force field parametrization, which is particularly exposed for the terminally bound tUpAIptU dinucleotide. |
67d95a696dde43c9083f7c98 | 18 | Another set of structural parameter, which were suggested to play an important role for nonenzymatic RNA primer extension are the conformations of the key ribose rings, which were shown to correlate with the distance between the O3 ′ atom at the terminus of the primer and the P atom of the adjacent activated nucleotide . In particular, previous works suggested that high population of the C3 ′ -endo ribose conformation within the imidazole-bridged dinucleotide intermediate could facilitate efficient formation of the new phosphodiester linkage . To establish this, we measured the value of the C1 ′ -C2 ′ -C3 ′ -C4 ′ dihedral angle (see Fig. ) throughout each trajectory for which values greater than 20 • were attributed to the C3 ′ -endo arrangement and values lower than -20 • were classified as the C2 ′ -endo conformation. The relative populations collected in Fig. demonstrate that the 3 ′ -terminal nucleotide of the primer predominantly adopts the C3 ′ -endo conformation, which aligns with the canonical A-form of RNA. |
67d95a696dde43c9083f7c98 | 19 | Interestingly, the analysis of the conformation of the adjacent residue within the activated dinucleotides demonstrates that the C3 ′ -endo is dominant (92.5 %) only for the parent GpAIpG system, which is consistent with the recent work of Mittal et al. . The population of the C3 ′ -endo conformation in the adjacent nucleotide is much lower for all remaining activated dinucleotides. Most of these dinucleotides are characterised by somewhat higher population of the C2 ′ -endo arrangement in the adjacent nucleotide. Only the ApAIpA dinucleotide interacting with the tUtU templates does not sample the C2 ′ -endo conformation for the adjacent nucleotide, with 31.5 % of the population reaching the C3 ′ -endo conformation and the remaining frames sampling values of the C1 ′ -C2 ′ -C3 ′ -C4 ′ dihedral angle, which are between the two ribose conformers. |
67d95a696dde43c9083f7c98 | 20 | The geometry of all terminal nucleotides is more variable and fluctuates throughout the simulation, as demonstrated in Fig. for GpAIpG. The averaged value of the considered dihedral angle for this system oscillates around 0 deg throughout the simulation, indicating no evident preference for one conformation over the other (Fig. ). Higher population the C3 ′ -endo conformation among different terminal nucleotides was observed only for the activated CpAIpC and CpAIpG dinucleotides with the latter system reaching the population of nearly 70%. We observed the opposite preference for the population of the C2 ′ -endo arrangement of the terminal nucleotide for GpAIpC and UpAIpU dinucleotides (59.5% and 64.9 %). The remaining four systems maintain an almost equal population of both sugar forms in the terminal nucleotide (48 -49%). Overall, the studied dinucleotide intermediates do not retain well C3 ′ -endo conformations associated with the canonical A-RNA structure, when bound near the 5 ′ -end of the template. This is also consistent with previous experimental works, which showed that downstream helper oligonucleotides facilitate efficient nonenzymatic RNA template copying, which was assigned to stabilization of the A-RNA structure . Finally, to better understand how the conformation of the three key ribose rings affects the O3 ′ • • • P distance between the 3 ′ -end of the primer and the adjacent nucleotide, we plotted three correlation plots for each dinucleotide system (see Fig. and the SI for more details on other systems). The parent GpAIpG system demonstrate a strong correlation of the C3 ′ -endo conformation of the adjacent ribose ring with a possibly short O3 ′ • • • P distance, that below 4.0 Å. Similar trend can be observed for the activated dinucleotides GpAIpC, CpAIpG IpAIpA and ApAIpA on the tUtU template, all of which exhibit at least a considerable population of the C3 ′ -endo conformation within the adjacent nucleotide. In contrast, stronger preference for the C2 ′ -endo arrangement of the adjacent nucleotide found for the CpAIpC, UpAIpU and ApAIpA (UU template) dinucleotides results O3 ′ • • • P distances often exceeding 4.0 Å(see Fig. ). This ca be very clearly illustrated based on the correlation plot for the bound GpAIpC dinucleotide, which maintains a good balance between both C3 ′ -endo and C2 ′endo ribose conformations. When the adjacent guanosine of GpAIpC adopts the C3 ′ -endo conformation, the O3 ′ • • • P distances center around ∼3.8 Å, whereas population of C2 ′endo sugar arrangement leads to a diffused population of O3 ′ • • • P distances ranging from 3.8 Åup to 8.0 Åand beyond. Overall, this analysis further demonstrates that retention of the A-RNA conformation of the primer end and the adjacent activated nucleotide could facilitate the formation of a new phosphodiester linkage in nonenzymatic RNA selfreplication . Importantly, the ribose conformation within the terminal nucleotide exhibits little or no direct correlation with the with the monitored O3 ′ • • • P distance. |
67d95a696dde43c9083f7c98 | 21 | In summary, we performed force-field based classical molecular dynamics simulations for nine different imidazoliumbridged dinucleotide intermediates bound to a templating strand with complementary WC base pairing partners, downstream of the 3 ′ -end of the copied strand (primer). These simulations allow us to shed more light on experimentally observed rate and fidelity of nonenzymatic RNA template copying reactions and explain the differences between different systems containing biological and alternative ribonucleotides. |
67d95a696dde43c9083f7c98 | 22 | In particular, we show that activated dinucleotides containing G and C monomers retain the highest percentage of Watson-Crick base pairing interactions with the template which can be both correlated to higher rates and fidelities of the primer extension reactions with these two RNA components. It is worth emphasising though that, when bound to the 5 ′ -end of the template, activated dinucleotides containing C more frequently adopt C2 ′ -endo ribose conformation which could hamper primer extension owing to longer O3 ′ • • • P distance between the adjacent nucleotide and the 3 ′ -end of the primer. Low rate and fidelity for the nonenzymatic template copying with A and U monomers is very well reflected by very low binding specificity of the dinucleotide intermediates containing these monomers in our simulations. In fact the ApAIpA and UpAIpU dinucleotides demonstrate a preference to fold into O-stacked and G-stacked structures of the imidazole bridged intermediate, which in most cases cannot bind specifically to the template. In addition, higher contribution of the C2 ′ -endo ribose conformation within these dinucleotides result in much less favorable (longer) O3 ′ • • • P distances. |
67d95a696dde43c9083f7c98 | 23 | A very interesting picture emerges from the simulations performed for the systems with alternative RNA monomers. In particular the IpAIpI dinucleotide exhibits a decent pro-portion of WC base pairing interactions with the template, in spite of the formation of only two hydrogen bonds per each I:C pair. We tentatively assign this property to the lower stability of G/O-stacked structures when compared to the systems containing activated A and U nucleotides. Such a preference for the formation of a canonical A-form stack within the IpAIpI dinucleotide could explain its higher binding specificity and fidelity of the process , whereas the reasonable proportion of C3 ′ -endo ribose conformations correlates with the reported high rate of the reaction . This finding for the first time demonstrates that high rate and fidelity of nonenzymatic RNA template copying with nucleotide phosphoroimidazolides is not only modulated by the stability of WC interactions but also by the conformational stability of the canonically stacked quasi-A-form of the imidazolium-bridged dinucleotide intermediate. It is worth noting though that our simulations indicate visibly lower population of IpAIpI fully bound to the template with WC interactions when compared to the GpAIpG dinucleotide. While these results are of high importance for future experimental explorations of this process, more accurate computational data is needed to fully confirm this hypothesis, for example with the support if density functional theory calculations. |
67d95a696dde43c9083f7c98 | 24 | Our simulations also allowed to demonstrate that replacing oxygen with sulfur in uridine enhances the binding of activated adenosine dinucleotides. Consequently the ApAIpA:tUtU system exhibits somewhat higher binding specificity through Watson-Crick base pairing interactions that its counterpart containing biological U nucleotides within the template. This result is qualitatively consistent with the experimental improvement of the yield and fidelity of the reaction with tU nucleotides, however, the simulations do not indicate a substantial increase, which is otherwise seen from the experiments . Furthermore, MD of the opposite system containing the tUpAIptU dinucleotide bound to the AA template resulted in less specific binding than in the case the UpAIpU. This inconsistency between the experimental result and the simulations may arise from various factors, including potential deficiencies and limitations of the classical RNA force field in accurately describing the binding of imidazolium-bridged dinucleotide intermediates on the RNA template, especially when they consist of non-standard and newly parameterized residues, which were not extensively tested so far. Therefore, we underscore that our classical MD simulations offer very valuable qualitative insights into the process on the molecular level. Nevertheless, a more picture would require the application of more accurate methods of quantum chemistry, which are often limited by very unfavorable scaling with the system size. |
67d95a696dde43c9083f7c98 | 25 | Finally, we emphasize that our simulations focused on the location of the imidazole bridged dinucleotide intermediate near the 5 ′ -end of the template. In fact, the more labile and flexible activated dinucleotides could exhibit somewhat higher binding specificity when placed in the middle of the template and accompanied by a downstream helper oligonucleotide . This selection of the model was largely dictated by the initial usage of the crystal structure for our parent GpAIpG system, which was originally acquired by Zhang et al. . However, despite this possible limitation of the model, it complements the previous results of Mittal et al. by showing the molecular intricacies of the process at the end of the template, which were not addressed up to now. Importantly, it demonstrates the fragility of the processes, emphasizing that some oligonucleotide sequences cannot be easily copied in full, particularly if the templating strand does not contain two C or G nucleotides at its termini. This could result in the possible prebiotic pre-selection of sequences which posses a higher potential to be copied in full, to prevent fractionation into a diverse pool of too short and non-functional oligomers. |
67d95a696dde43c9083f7c98 | 26 | The initial structure for this study was sourced from an article by Zhang et al. [PDB:6C8K]. Hypoxanthine and 2-thiouracil were parameterized according to the protocol given in the Supporting Information Section 2. Each system was solvated in truncated octahedral water box (SPC/E water model )with minimal distance between RNA molecule and border set up to 12 . To the solution were added 25 K + counterions for neutralization and 0.15M of KCl (Joung and Cheatham parameters ). The AMBER parmOL3 force field was applied, and an equilibration protocol was used. Initially, 500 steps of steepest descent minimization and 500 steps of conjugate gradient minimization were performed with 25 kcal•mol -1 position restraints on the RNA atoms. Subsequently, the system was heated to 300K, starting at 100K, with 25 kcal•mol -1 restraints during a 100 ps molecular dynamics simulation at constant volume. The third step involved minimization with 5 kcal•mol -1 , 500 steps of steepest descent and 500 steps of conjugate gradient minimization. Following this, a 50 ps equilibration was conducted with 5 kcal•mol -1 restraints, at a constant temperature of 300K and pressure of 1 bar. These last two steps (minimization and equilibration) were repeated for 4, 3, 2, and 1 kcal•mol -1 restraints. The final equilibration included 0.5 kcal•mol -1 position restraints on the solute molecule. The final molecular dynamics simulation was performed for 50 ps at constant temperature 300K and pressure 1 bar, without any constraints. The production runs were divided into 3 different simulations for every system, with randomized initial conditions and each production simulation was performed for 500 ns. Consequently, it gave 1.5 µs of simulation time for each system. All of the production simulations were performed at 300K and 1 bar with use of the Langevin thermostat and the Monte Carlo barostat . The hydrogen mass repartitioning , SHAKE and SETTLE algorithms were applied to allow setting time step to 4 fs as indicated by Hopkins Chad W., et al. . The cuttoff for non-bonding interactions was setl to 9. The CUDA accelerated pmemd module of AMBER22 was used to perform molecular dynamics production simulations. |
67d95a696dde43c9083f7c98 | 27 | RMMD-based clustering procedure was used to analyse structures obtained from simulations. Two algorithms were implemented: DBSCAN and K-means. DBSCAN is a density based clustering algorithm for which the parameter minpoints was set to 25 and the ϵ varied between 1.2 and 2.8Å(estimated with help of Kdist plot and for exact values see Table in Supporting Information) whereas K-means divides all obtain frames into N clusters . The number of clusters N was set to 10. Every 20-th frame was analyzed to make it more efficient and water molecules along with ions were excluded from the analysis. |
67d95a696dde43c9083f7c98 | 28 | The distances between atoms involved in Watson-Crick interactions, as well as the O3 ′ • • • P atoms and dihedral angles for each frame of the trajectories, were analyzed using the Python library MDAnalysis () . These data were plotted against simulation time or analyzed using the Python hexbin function to correlate angles with distances. |
6142fc6965db1e0764b853b8 | 0 | Many useful reactions are underutilised in synthetic organic chemistry because of an inability to predict the regioselectivity of the reaction and there is thus an increasing interest in developing regioselectivity prediction methods for such reactions. Recent examples include nucleophilic and electrophilic aromatic substitution reactions, Diels-Alder reactions, Heck reactions, 12 radical C-H functionalisation of heterocycles, and reactions such as as alkylations, Michael additions, and aldol condensations that proceed through proton abstraction. These methods have been based on either quantum chemical (QM) calculations, machine learning (ML) trained on experimental data, 8,10-12 or a combination of the two where QM has either provided descriptors for the ML model or used to augment the training data. However, these approaches have rarely been compared on the same dataset. In this paper we present an ML model (RegioML) that predicts the regioselectivity of electrophilic aromatic substitution (EAS) reactions using QM charges. We compare the performance of RegioML to RegioSQM20 6 -a QM-based predictor for EAS regioselectivity -for the same dataset and discuss how QM-based predictors can be used to augment sparse experimental datasets. We focus in particular on the precision and recall of these methods for in-and out-of-sample datasets. |
6142fc6965db1e0764b853b8 | 1 | The reaction data are extracted from Reaxys using a set of queries (see supporting information), which results in a total of 30,368 bromination reactions. A thorough dataset curation is then performed to obtain a set of unique SMILES (simplified molecular input line entry system) of the reactants and their corresponding site of bromination, which reduces the total number of reactions to 21,896. For example, a reaction is discarded if there is not an exact one-to-one mapping between the heavy atoms of the reactant and the product excluding the reacting bromine(s), or if a reacting bromine forms a bond to something other than a cyclic sp 2 hybridized carbon atom (accounting for 5,314 reactions). Furthermore, reactions with unique reactions IDs in Reaxys but identical reactants are merged (accounting for 3,158 reactions). |
6142fc6965db1e0764b853b8 | 2 | Recently, we published the RegioSQM20 method, 6 which predicts the regioselectivities of EAS reactions from semiempirical calculations of proton affinities. The single-tautomer version of this method is applied to the 21,896 reactions to get proton affinities for all of the unique reaction sites. An extension of this method is also applied in which the Re-gioSQM20 calculations are followed by single point density functional theory (DFT) calculations in methanol (MeOH, dielectric = 33.6) using the PBEh-3c composite electronic structure method 15 and the conductor-like polarizable continuum model (C-PCM) as implemented in the quantum chemistry program ORCA version 4.2. 18 A few of the calculations resulted in extreme proton affinities corresponding to outliers in the dataset that complicated the development of regression models. However, the calculated proton affinities for both the original and extended RegioSQM20 calculations follow a Gaussian distribution (see supporting information), which enables the use of Chauvenet's criterion to remove these outliers in the dataset. In the Chauvenet's criterion the probability of the farthest point is calculated under the assumption of a Gaussian distribution. If this point is below some predefined value then the point is removed, and the procedure is repeated until no more points are removed. In our dataset molecules are removed if at least one atom in the molecule has a proton affinity corresponding to a probability below 1 %. |
6142fc6965db1e0764b853b8 | 3 | We investigate seven different atomic descriptors as input to the ML models (details on the descriptors are given in Table in the supporting information). The atomic descriptors are developed by Finkelmann et al. and have been successfully applied to the prediction of site of metabolism, hydrogen bond donor and acceptor strengths, and Ames mutagenicity of primary aromatic amines. Almost all of the descriptors depend on charge model 5 (CM5) atomic charges, which are obtained from a single point calculation using GFN1-xTB as implemented in the open source semiempirical software package xtb version 6.4.0. 26 This particular charge scheme has been shown to be largely conformation-independent and to correctly reflect changes in the chemical environment i.e. substituents effects. Hence, only a single conformer is generated for each molecule using ETKDG versions 3 27 with us-eSmallRingTorsions=True as implemented in RDKit version 2020.09.4. This is the key to using quantum chemical derived descriptors as the computational cost is kept at a minimum (details about computational timings are provided in Results and discussion). The atomic descriptors are generated fully automatically from a SMILES representation of a given molecule. |
6142fc6965db1e0764b853b8 | 4 | From the screening of the seven atomic descriptors, we find that a charge shell descriptor with 5 shells and values sorted according to the Cahn-Ingold-Prelog (CIP) rules is particularly good for predicting the regioselectivty of bromination reactions (see Table and Figure in the supporting information). An illustration of this 485-dimensional descriptor can be seen in Figure . |
6142fc6965db1e0764b853b8 | 5 | We utilize an unsupervised learning procedure similar to the one found in the MOLAN workflow by Sivaraman and Jackson et al., which resembles the ButinaSplitter from DeepChem. The procedure is as follows: SMILES representations of each molecule are converted into extended connectivity (Morgan) fingerprints 30 with a radius of 2 and 1,024 bits (ECFP4). The ECFP4 fingerprints are then used to construct a Tanimoto similarity matrix, which enables a clustering of the molecules using the Butina clustering algorithm 31 with a radial cutoff of 0.6 as implemented in RDKit. Clusters with at least 7 molecules are included in the training/test set and otherwise in the out-of-sample set to explore how well the trained machine learning models generalize. For some molecules either the descriptor or RegioSQM20 calculations fail, or the molecules are excluded due to the Chauvenet's criterion, which left us with 21,201 reactions corresponding to 100,588 unique reaction sites. Thus, applying the above procedure results in a training/test set and an out-of-sample set of 15,246 and 5,955 molecules, which corresponds to 73,123 and 27,465 unique reaction sites, respectively. Uniform stratified and random splits are then used to obtain a 80:20 ratio between the training and test sets resulting in 12,196 and 3,050 molecules corresponding to 58,384 and 14,739 unique reaction sites in each set, respectively. For the uniform stratified split, each of the individual clusters are randomly split and hereafter combined to ensure that both the training and test sets have similar representations of the underlying data distribution. On the other hand, the random split is indeed completely random with respect to all of the molecules obeying the cluster size cutoff. |
6142fc6965db1e0764b853b8 | 6 | In order to learn and predict the regioselectivity of EAS reactions, we explore various regression and classification models with respect to both the experimental and calculated data described above. Initially, a screening of 17 regression models and 13 classification models using PyCaret version 2.3.2 is conducted (details can be found in the supporting information). This allows us to quickly find promising machine learning methods, which are then thoroughly examined in terms of finding optimal hyperparameters. The hyperparameter optimizations are carried out using a tree-structured Parzen estimator (TPE) as implemented in Optuna version 2.5.0. All training and evaluation are done using either a normal or a stratified 5-fold cross-validation of the randomly shuffled training set in the case of the regression and classification models, respectively, and only the models with the best validation performance are saved for testing. As we show in Table the best performance for both regression and classification is the ensemble decision tree variant called light gradient boosting machine (LightGBM) version 3.1.1 34 using the sorted-shell atomic descriptors with a shell radius of 5. We refer to this method as simply "LightGBM" hereafter. Furthermore, we examine the imbalance in the dataset using a "Null model", where all sites are predicted to be non-reactive. And we employ a 1-nearest neighbors (1-NN) classifier as a baseline model using the brute-force search algorithm and the Jaccard metric as implemented in scikit-learn, which corresponds to a perfect memorization of the training set. Table : Comparing different methods for predicting the reactivity of the 14,739 unique reaction sites in the test set and the 27,465 unique reaction sites in the out-of-sample set. The reported metrics are accuracy (ACC), Matthew's correlation coefficient (MCC), precision (PPV or positive predictive value), recall (TPR or true positive rate), specificity (TNR or true negative rate), and negative predictive value (NPV). |
6142fc6965db1e0764b853b8 | 7 | The results we present here only involve the random splitting of the training and test set as similar performances are observed for both the stratified and random splits as seen in Table in the supporting information. Unless otherwise noted all machine learning models are classifiers that output a value between 0 and 1 for each atom, where a value greater than 0.5 indicates that an atom should be reactive. |
6142fc6965db1e0764b853b8 | 8 | In this section, we train and evaluate machine learning classifiers on experimental data collected from Reaxys consisting of 58,384, 14,739, and 27,465 unique reaction sites in the training, test, and out-of-sample sets, respectively. The experimental data often contain just single or a few reported reactive sites among all reaction sites in the reactant, i.e. there are significantly more negatives (N) than positives (P) in the dataset. Consequently the accuracy (the proportion of correct predictions, ACC = TP+TN P+N ) can be a misleading metric. For example, a "Null model", where all sites are predicted to be non-reactive achieves a respectable accuracy of 76 % (Table ) for both the test and out-of-sample sets, but this just reflects the fact that 76% of the sites in both the datasets are unreactive. The Matthews correlation coefficient 37 (MCC) is a more robust metric to assess the model performance, since it also considers false positives (FP) and false negatives (FN) in addition to true positives (TP), true negatives (TN). |
6142fc6965db1e0764b853b8 | 9 | where PPV = TP TP+FP , TPR = TP P = TP TP+FN , TNR = TN N = TN TN+FP , and NPV = TN TN+FN are also known as precision, recall, specificity, and negative predictive value, respectively. The MCC values for both the test and out-of-sample sets are zero, which clearly shows that the Null model lacks any real predictive power. |
6142fc6965db1e0764b853b8 | 10 | As a baseline model, we trained a 1-nearest neighbors (1-NN) classifier corresponding to a perfect memorization of the training set. The data shows that an impressive-looking 86% accuracy can be achieved for the test set by simple memorisation of the training set. In contrast, the MCC value is only 0.62 for the test set and considerably lower (0.49) for the out-of-sample set. These values primarily reflect a low precision where only 71% and 59% of the positive predictions are actually correct. |
6142fc6965db1e0764b853b8 | 11 | Our best machine learning model (LightGBM) achieves a considerably better precision of 88% and 80%, respectively. Note that while there is only a 3% drop in accuracy on going from the test set to the out-of-sample set, there is an 8% drop in the precision (and a concomitant drop in MCC). Hereafter, we refer to this method (i.e. LightGBM trained on experimental data) as RegioML. |
6142fc6965db1e0764b853b8 | 12 | The test set MCC of RegioML is virtually identical to the Weisfeiler-Lehman neural network (WLN) architecture specifically trained to predict the regioselectivity of EAS reactions by Struble et al. . While the precision is 4% higher for WLN, the recall (the fraction of positives that are predicted correctly) is 5% lower leading to an nearly identical overall performance. WLN performs better on the out-of-sample set, with an MCC value that is nearly identical to that off the test set. However, it should be noted that many of the molecules in these two sets are likely included in the set used to train the WLN method, which is likely to inflate the WLN MCC values. For example, we are able to achieve MCC values of 0.98 on both the test and out-of-sample sets by training the LightGBM model on the entire collection of data using 10-fold cross-validation (the MCC value is for the best performing model). |
6142fc6965db1e0764b853b8 | 13 | RegioSQM20 predicts the regioselectivity of EASs by finding the reaction center with the highest proton affinity. For computational efficiency the proton affinities are computed using the semiempirical tight binding method GFN1-xTB and a continuum solvent model of methanol. The centers with proton affinities within 1 kcal/mol of the maximum are considered reactive. The method thus has only a handful of hyperparameters (choice of computational method, solvent, energy cutoff, conformational search method) and these are chosen based on a dataset of 532 experimental measurements, some of which are included in the current training set. |
6142fc6965db1e0764b853b8 | 14 | For the test set, the recall of RegioSQM20 is similar to RegioML (81% vs 83%) but the precision is significantly worse (75% vs 88%). For the out-of-sample set, the recall is somewhat better for RegioSQM20 (80% vs 76%) but the precision is still worse (73% vs 80%), leading to a slightly smaller MCC value of 0.69 compared to 0.72 for RegioML. In contrast to RegioML, the overall performance of RegioSQM20 is very similar for the test and out-of-sample set, as one would expect from a more physics based method. However, RegioSQM20 does not offer an advantage over RegioML for the out-of-sample dataset, while being computationally much more demanding (see below). |
6142fc6965db1e0764b853b8 | 15 | The main advantage of the RegioSQM20 approach is that it may offer an accuracy similar to RegioML based on a much smaller training set. Indeed a LightGBM model trained on the same 532 reactions used to develop RegioSQM20 results in MCC values around 0.5 for both the test and out-of-sample set. While the precision is quite good for this model, the recall is less than 50% due to a large proportion of false negatives. Thus, in cases where little experimental data is available physics-based methods like RegioSQM20 is likely to outperform ML based methods, even if the latter rely on quantum descriptors such as atomic charges. |
6142fc6965db1e0764b853b8 | 16 | The computational expense of the physics-based methods can be mitigated by using them to generate synthetic data for the machine learning model. Indeed, a LightGBM classifier trained on RegioSQM20 predictions for the large training set of 58K reaction centers offers the same performance as RegioSQM20 for the test set. Of course, the performance is worse for the out-of-sample set just like for RegioML, but the training dataset can now easily be expanded to ensure a better coverage of chemical space. Furthermore, since RegioSQM20 is not used to offer real-time predictions to a user, more accurate and computationally expensive methods can be explored. For example, the precision of RegioSQM20 can be increased by 6% by using PBEh-3c single point calculations to compute the proton affinities -an increase that is reflected in the corresponding ML model. The overall performance of RegioSQM20 PBEh-3c is now identical to the RegioML for the out-of-sample dataset, with MCC values of 0.72. |
6142fc6965db1e0764b853b8 | 17 | We also explore whether it is better to predict the proton affinities using regression and use them to identify reactive centers, rather the classification approach. Although the Light-GBM RegioSQM20 regression model is able to achieve a mean absolute error (MAE) of 2.00 kcal/mol on the test set, the accuracy is not good enough to distinguish between reactive and non-reactive sites compared to the LightGBM RegioSQM20 model, as evidence by the low recall values of 71%-73%. However, the LightGBM RegioSQM20 regression model can be used to predict low, medium, or high reactivity as we showed in the RegioSQM20 paper. In fact, by combining the classification model and the regression model one gets both regioselectivity predictions and a qualitative prediction of the reactivity of a molecule with almost no additional cost as the atomic descriptors only have to be calculated once. Examples of the output of RegioML can be seen in Fig. . |
6142fc6965db1e0764b853b8 | 18 | Table : Timings of the RegioSQM20 method, the RegioML model, and the WLN architecture for predicting the regioselectivity of the 3,050 molecules in the test set given a SMILES representation as input. For the RegioML model and the WLN architecture, the timings include descriptor creation and model prediction for all reaction sites in the given reactant. |
6142fc6965db1e0764b853b8 | 19 | In Table we compare the timings of the RegioSQM20 method, the RegioML model, and the WLN architecture by Struble et al. 8 for the 3,050 molecules in the test set. We report the median CPU time and the mean CPU time for predicting the regioselectivity of a molecule given a SMILES representation as input. For the RegioML model and the WLN architecture, the timings cover descriptor creation as well as model prediction for all reaction sites in the given reactant. |
6142fc6965db1e0764b853b8 | 20 | The results show that the median CPU time requirements of the RegioSQM20 method is 48 s per molecule on four Intel(R) Xeon(R) CPU E5-2643 v3 @ 3.40GHz cores. The RegioML model is almost 100 times faster on just a single Intel(R) Xeon(R) CPU X5550 @ 2.67GHz core with a median CPU time of less than half a second per molecule. The WLN architecture is able to achieve a mean CPU time of just 0.03 s per molecule on the single Intel(R) Xeon(R) CPU X5550 @ 2.67GHz core. The main reason for the slower performance of RegioML is the GFN1-xTB single point calculation needed to compute the atomic charges. |
6142fc6965db1e0764b853b8 | 21 | We present RegioML, an atom-based machine learning model for predicting the regioselectivities of electrophilic aromatic substitution (EAS) reactions. The model relies on ultra fast quantum chemical descriptor calculations combined with an ensemble decision tree variant called light gradient boosting machine (LightGBM). The atomic descriptors are based on CM5 atomic charges obtained from a single conformer embedding with RDKit 28 and a single point calculation using the open source semiempirical tight binding method GFN1-xTB. The model is trained and tested on 21,201 bromination EAS reactions and 101K reaction centers, which is split into a training, test, and out-of-sample datasets with 58K, 15K, and 27K reaction centers, respectively. The accuracy is 93% and 90% for the test and out-of-sample set, respectively, but this is not a good measure of performance due to the preponderance of non-reactive sites. For example, the precision (the percentage of positive predictions that are correct) is 88% for the test set, but only 80% for the out-of-sample set. |
6142fc6965db1e0764b853b8 | 22 | The final RegioML model released to users is trained on the entire data set and we expect similar performance for molecules in-sample and out-of-sample for this large dataset. The test-set performance is very similar to the graph-based WLN method developed by Struble et al. 8 though the comparison is complicated by the possibility that some of the test and out-of-sample molecules are used to train WLN. RegioML out-performs our physics-based RegioSQM20 method 6 where the precision is only 75%. Even for the out-of-sample dataset, RegioML slightly outperforms RegioSQM20. |
6142fc6965db1e0764b853b8 | 23 | The good performance of RegioML and WLN is in large part due to the large datasets available for this type of reaction. For example, if we retrain the RegioML model on the same 532 reaction we used to develop RegioSQM20, the performance is much worse due to a large increase in the false negative rate leading to a recall (the percentage of positives that are predicted correctly) below 50% compared to 80% for RegioSQM20. Thus, one use of physics-based approaches like RegioSQM20 is to generate synthetic data for ML model for reactions where there is little experimental data. We demonstrate this by showing that the performance of RegioSQM20 can be reproduced by a ML model trained on RegioSQM20generated data. |
6142fc6965db1e0764b853b8 | 24 | PyCaret version 2.2.0 32 is used as an initial screening of 17 regression models and 13 classification models to find promising models. To evaluate the performance of the models, we use a 5-fold cross-validation scheme of the atomic data for atoms in molecules belonging to the randomly split training set. The atomic data consist of the sorted-shell descriptor with 5 shells in combination with either a binary label corresponding to whether or not a bromination reaction has been experimentally observed on the specific site or the calculated proton affinity obtained by the original RegioSQM20 method. The sorted-shell descriptor with 3 shells and the combinatorial descriptor were also tested in this initial screening, but the ranking of the different models were similar to those presented in Figures and. . |
677ea1d281d2151a025203f0 | 0 | Point-of-care tests (POCT) are a class of diagnostics capable of conducting medical tests within the immediate patient vicinity. Among such tests, lateral flow assays (LFAs) have long represented an archetypical, low-cost and rapid POCT, meeting many of the REASSURED criteria for diagnostics. The user-friendly and equipment-free virtues of these tests enable LFAs to be conducted in a variety of decentralised settings, such as clinics, pharmacies and homes. The impact of LFAs is perhaps most evident in resourcelimited countries as a means of offering accessible and delocalised tests, where access to quality healthcare can be limited. The utility of LFAs was recently highlighted during the COVID-19 pandemic where tests were adopted on an unprecedented scale in efforts to monitor and control the spread of SARS-CoV-2 worldwide. The performance of a LFA is often criticised as lacking in sensitivity and thus strategies to improve this are desirable. Realising high sensitivity in LFAs is frequently underpinned by a need for excellent recognition components for biomarker detection. In a typical test, the timeframe in which analyte flows past the test line is relatively short, presenting a small opportunity for antibodies to bind target epitopes. In sandwich LFA formats, antibodies decorating the test line and nanoparticle labels must bind the analyte with fast association kinetics and high affinity to ensure maximal analyte detection and LFA sensitivity. Achieving this often necessitates screening monoclonal antibodies, which are generally expensive and time consuming to produce, despite no guarantee of identifying candidates with high affinity and specificity towards biomarker targets. To this end, the ability to enhance the sensitivity of these tests -especially where low-affinity or poorly performing antibody reagents fall short of clinical or early-detection criteria -in a costeffective way whilst maintaining maximal user-friendliness remains crucial. |
677ea1d281d2151a025203f0 | 1 | Capture-and-release strategies are ubiquitous in a variety of analytical or separation techniques, such as chromatography or immunoprecipitation, often with the aim of sequestering and enriching a target from its matrix, and subsequently triggering its release in a controlled manner. In the context of LFAs, the ability to enrich biomarkers in a sample volume represents a powerful opportunity towards bypassing poor assay kinetics. In the case of nucleic acid detection, biomarker enrichment has been demonstrated with the use of recombinase polymerase amplification (RPA) workflows as a sample pre-processing step. The application of such strategies for antigen-based LFAs, however, remains limited and comparatively difficult. Key examples of this include an immobilised metal-affinity chromatography (IMAC) inspired 'capture-and-release' approach where magnetic beads were utilised to manipulate sample concentration prior to LFA detection. Here, Bauer et al. relied on the affinity between Zn(II) and a characteristic histidine-rich motif in the structure of a malarial HRP2 antigen to afford specific, non-covalent capture prior to triggered elution using competitive ligands (such as imidazole). The application of this IMAC approach has been further explored by utilising non-covalent aptamer binding and hexahistidine-tagged antibodies to mediate 'capture-and-release' of non-histidine rich antigen targets in a similar fashion. Finally, studies by Moore et al. have shown that Zn(II) functionalised cellulose can be used to perform 'capture-and-release' of histidine-rich HRP2 antigens without relying on magnetic bead technologies, demonstrating improved potential for field-based testing. In each of these examples, sample pre-processing workflows require users to elute and collect the concentrated analyte sample as a flowthrough volume, which is then applied to a LFA test. Whilst efforts have sought to realise this approach with improved userfriendliness, for instance by moving from a tube containing magnetic beads to a cellulose membrane platform, these 'capture-and-release' examples are limited to being performed externally to the LFA device itself, ultimately necessitating users to handle and operate additional equipment to produce a LFA readout. Furthermore, the ability to selectively bind protein-based biomarkers in orthogonal ways without 'using-up' target epitopes has been highlighted as being non-trivial, presenting additional challenges when considering In this work, we describe an alternative LFA 'capture-and-release' methodology which seeks to address limitations regarding poor assay binding kinetics and generate signal amplification for enhanced sensitivity. To accomplish this, we introduce a folding 'twostrip' LFA architecture where antigen complexes are initially sequestered in a capture strip via the binding of immunoproteins modified with cleavable linkers. Following triggered linker cleavage, immunocomplexes are released and subsequently rebound with higher affinity in a detection strip where a higher density of particles and subsequent signal is obtained. The capabilities of this 'capture-and-release' strategy, which we term as the "AmpliFold" approach, is highlighted by augmenting the analyte binding region within capture strips. In the AmpliFold approach, a large surface area of capture receptors can be implemented to expand the opportunity for target capture, thus enabling a greater proportion of sample analyte to be sequestered and labelled. We demonstrate that the benefits of an enhanced degree of analyte capture could be realised within our AmpliFold platform where, through triggered linker cleavage, antigen-bound complexes could be redistributed and concentrated at the test line of the detection strip. For a model system, where low capture receptor density was used to emulate poor capture affinity, the AmpliFold platform was shown to yield a 16-fold sensitivity enhancement when compared to a traditional LFA format. The application of the AmpliFold approach was further demonstrated in a LFA system utilising large (150 nm) nanoparticle labels. Specifically, large capture areas were shown to address limited binding kinetics -associated with poor particle diffusivity -within an AmpliFold approach, resulting in a 12-fold improvement in sensitivity when compared to traditional LFA methods. |
677ea1d281d2151a025203f0 | 2 | In conventional LFA design, antigen binding can typically be described as a two-antibody system which relies on orthogonal epitopes to simultaneously capture and label the target analyte. In our AmpliFold approach, illustrated in Figure , 'capture-and-release' is performed entirely across a lateral flow architecture -comprising a capture and detection strip -and utilises additional affinity interactions to accomplish orthogonal binding mechanisms beyond that of a traditional immunoassay. In the first instance, the selection of polystreptavidin (PSA) as a capture receptor presented opportunities to develop a 'capture-and-release' mechanism via the medium of a cleavable linker. Here, the utilisation of a cleavable linker provides a discrete alternative from initiating analyte release directly from the antibody-antigen complex itself, which often requires harsh immunodisruption conditions. Aside from its exceptional affinity for capturing biotinlabelled targets, the choice of PSA also represents a commonly employed LFA receptor commercially regarded for its stability and robustness against thermal and chemical influences. In a pre-assay step, the model target analyte, HER2 antigen, was pre-mixed with, and bound by, a modified anti-HER2 Fab fragment (FabHER) and a fluoresceintagged anti-HER2 antibody-AuNP conjugate, forming a sandwich immunocomplex structure. When run up AmpliFold capture strips, antigen-bound sandwich immunocomplexes are initially sequestered by PSA test line receptors due to chemical modification of anti-HER2 Fab fragments (FabHER) with a cleavable linker bearing a biotin tag. The cleavable linker features a disulfide bond, which can undergo triggered cleavage using reducing reagents, thus enabling the release and elution of captured sandwich immunocomplexes from initial PSA capture. Disulfide-based cleavage was selected as a promising candidate as it is well-validated within cleavable antibody-drug applications and benefits from a prevalence of commercial disulfide-containing linkers capable of enabling facile protein biotinylation. Beyond this, disulfide-based cleavage chemistry was anticipated to provide suitably fast cleavage kinetics for developing a capture-and-release strategy, which could be performed under a paper microfluidic flow regime and in a comparable timeframe to typical LFAs (ca. 10-20 minutes). 26 |
677ea1d281d2151a025203f0 | 3 | As a binding system, Pertuzumab and the Fab fragment of Herceptin (FabHER) were selected as a matched pair for binding the HER2 antigen. Pertuzumab and Herceptin are cost-effective FDA-approved antibodies that are known to bind HER2 via different epitopes, with low nanomolar binding affinities, providing an adequate matched pair for R&D purposes. The capture protein FabHER was modified with biotin via multiple strategies allowing us to study the influence of linker length, flexibility, and modification approach when evaluating the efficiency of 'capture-and-release'. A range of thiolcleavable commercially available lysine-reactive linkers including sulfo-NHS-SS-biotin, NHS-SS-PEG4-biotin were employed in reaction with FabHER in a non-site-selective modification strategy, with NHS-PEG12-biotin as a non-cleavable control (Figure , parts i-iii). This generated FabHER-linker-biotin conjugates 1-3. A further strategy utilised pyridazinedione (PD) chemistry to site-specifically modify the FabHER's single solvent accessible disulfide using Br2PD-BCN to site-selectively install a single strained alkyne (bicyclo[6.1.0]non-4-yne, BCN), which was then reacted with an N3-PEG3-SS-PEG4-biotin linker (see Figure for synthesis) via a quantitative 'click' chemistry reaction to generate FabHER-PD-PEG3-SS-PEG4-biotin (4) (Figure ). Representing the other half of the immunoassay, Pertuzumab was employed as a detection antibody for the functionalisation of gold nanoparticle (AuNP) labels. Aside from binding the HER2 antigen, our design criteria comprised a particle label which could be orthogonally re-bound at the test line of the detection strip. The method of particle rebinding itself was anticipated to necessitate a high-affinity interaction to ensure maximal recovery of particle complexes -following release by linker cleavage -and detection with superior signal-to-noise, when compared to initial analyte capture. To develop a 'dual affinity' nanoparticle label, Pertuzumab was first pre-modified with fluorescein isothiocyanate (FITC) to install specific and orthogonal affinity tags. The resulting fluorescein-Pertuzumab (FluoroPer) conjugates were then used to functionalise 40 nm AuNPs via passive physisorption (Figure ). The physisorption of FluoroPer to citrate capped 40 nm AuNPs was confirmed by UV-Vis (Figure ) and dynamic light scattering (DLS) (Figure ) where upon binding of the protein, a peak plasmon shift and increase in hydrodynamic diameter of 12.1 nm respectively were observed. The increase in hydrodynamic diameter reflected the thickness of a protein layer consistent with expected antibody dimensions. 40 nm FluoroPerAuNPs blocked with BSA were shown to have a further increased hydrodynamic diameter suggestive of corona formation. Zeta potential measurements (Figure ) indicated the blocked FluoroPerAuNPs exhibited a negative surface charge -22.8 mV, comparable with the bare citrate capped AuNPs, however during the process of FluoroPer physisorption, an intermediate surface charge closer to neutral was measured. To evaluate the affinity of these particles as a re-binding target, LFA studies were performed to benchmark the performance of FluoroPerAuNPs against a biotin-streptavidin model AuNP system. Here, nitrocellulose membranes printed with either anti-FITC or PSA test lines were used to bind and detect 40 nm FluoroPerAuNPs and biotinylated-AuNPs, respectively. As a general procedure, test line signal measurements were achieved by imaging test strips under controlled lighting conditions and by using ImageJ software to extract pixel intensity values (full procedure described in the Materials & Methods, as well as in Supplementary Information Figure ). In these experiments, 40 nm FluoroPerAuNPs were shown to be detected down to concentrations of 274 fM (95% CI of 150 to 533 fM), and 40 nm biotinylated-AuNPs down to concentrations of 94.8 fM (95% CI of 52.6 to 169 fM). Whilst lesser than its model comparison, the detection of our FluoroPerAuNPs down to femtomolar sensitivities likely reflects a significant degree of decoration on the particle surface by FluoroPer antibodies. |
677ea1d281d2151a025203f0 | 4 | Given that the biotin-streptavidin interaction is commonly regarded as the highest affinity interaction, the comparable performance of FluoroPerAuNPs was further rationalised based on the availability of fluorescein tags on the particle surface. Using mass spectrometry analysis, the average fluorescein-to-antibody ratio was determined to be 4:1, enabling functionalised particles to present numerous tags across its surface for rebinding by test line receptors (Supplementary Information Figure ). |
677ea1d281d2151a025203f0 | 5 | The viability of the 'dual-affinity' 40 nm FluoroPerAuNPs was further studied by observing the stability and performance of these particles in the presence of disulfide linker cleavage reagents. The incorporation of thiol-based reducing reagents, such as dithiothreitol (DTT), presented potential off-target consequences within assay reagents, such as protein denaturation and undesirable thiol-gold bond formation on the particle surface. To mimic the proposed assay workflow, FluoroPerAuNPs were pre-mixed with DTT at pH 10 prior to detection using anti-FITC pAb printed test strips. This format enabled us to primarily assess particle stability but also the retention of functionality of test line itself while challenged to exposure to the cleavage mixture over a prolonged 10-minute period. No notable losses in signal intensity -compared to a negative DTT control -was observed when measuring test line signal (Figure ). Whilst the reduction of structural elementssuch as antibody disulfide bridges -cannot be ruled out, these results echo findings by Shlyapnikov et al. who utilised reducing conditions to cleave disulfide-linked hydrophobic blocking agents from the surface of microarray-based immunoassays. Authors noted that no significant effect on antibody binding after a 5-minute incubation period in cleavage reagent, offering similar insights into the robustness of recognition components under reducing conditions within brief assay timeframes. FITC pAbs and PSA, respectively. Data was collected in triplicate and was fitted using a 4-parameter logistic (4PL) regression fit. f. The mean signal intensity (N = 3) of anti-FITC test lines where 40 nm FluoroPerAuNPs were directly mixed with DTT in cleavage conditions prior to running. |
677ea1d281d2151a025203f0 | 6 | An ideal linker cleavage was envisaged to have high efficiency (i.e. nearly quantitative release of captured material) and to occur in a rapid timeframe (10 minutes). To appraise both cleavage efficiency and kinetics, a simple, multi-step procedure was performed where, in short, a pre-mixture of HER2 antigen, FabHER-linker-biotin and 40 nm FluoroPerAuNPs were carried up PSA-printed membranes to generate a HER2 antigen dependent test line signal. Following a wash step, biotin anchored immunocomplexes were then cleaved using various concentrations of DTT in a pH 10 cleavage buffer (see Supplementary Information Figure for pH optimisation). After 10 minutes, a thiolreactive, N-methylmaleimide solution was then wicked through the membrane to quench the reaction. A control in the absence of DTT was used to assess the release of bound complexes, via triggered linker cleavage (Figure ). Interestingly, linker cleavage was not as facile as anticipated. In examples featuring complexes captured via FabHER-SSbiotin (2) (a short linker conjugate), a maximal signal loss of only 40.6% (± 0.9%) was observed despite a DTT concentration of 100 mM being used. When compared to supplementary studies using mass spectrometry (Supplementary Information Figures ), quantitative linker cleavage of unbound FabHER-SS-biotin (2) proceeded at a 30fold lower concentration of DTT within the same timeframe was observed. Based on the differences between bound and unbound linker cleavage efficiencies, steric hinderance at the linker cleavage site may have been induced between the interface of this large proteins -in this case, the Fab fragment and PSA. Further cleavage experiments involving an alternative modified fragment with a longer linker to the biotin entity, i.e., FabHER-SS-PEG4-biotin (3), showed an improved maximal cleavage of 63.1% (± 2.5%). Here, the elongation of the PEG spacer in the linker was thought to provide some flexibility, thus alleviating some steric hinderance and improving cleavage efficiency. In an effort to achieve near-quantitative linker cleavage, FabHER-PD- PEG3-SS-PEG4-biotin conjugate (4) was also employed. Cleavage experiments featuring this conjugate demonstrated the highest cleavage efficiency within 10 minutes with a reduction of signal by 82.1% (± 7.4%) upon elution with 100 mM DTT. As with previous iterations, reduction of steric hinderance by virtue of linker elongation may be offered as an explanation for improved linker cleavage, yet the drastic enhancement in cleavage efficiency is also theorised to be the result of site-specific linker installation. Previous FabHER-linker-biotin iterations conjugated through lysine modifications were observed by mass spectrometry to have a large distribution of FabHER-to-linker loading ratios which may enable a single complex to bind numerous PSA receptors multivalently, such that multiple cleavages are required per complex for release. Using PD modification, linker installation occurs at the site of a disulfide bridge which, for a Fab fragment, results in a 1:1 loading ratio and a single cleavage target per complex. In this way, the release of captured antigen-bound complexes is expected to be relatively facile -compared to constructs with the potential for multivalent anchoring to the immobilised PSA test-lineand may further explain the marked difference in performance between the studied examples. |
677ea1d281d2151a025203f0 | 7 | As an 'non-cleavable' control, FabHER-PEG12-biotin (1) was similarly studied in cleavage experiments to ensure that the basis of signal loss was as a result of linker cleavage and not by the interference of the complexes themselves. Signal loss was observed when employing DTT concentrations above 20 mM, despite complexes being captured via a non-cleavable linker (Figure ). At these concentration regimes, a modest mean average signal loss of just 5.6% (± 2.5%) was observed, indicating that -for the example of FabHER-PD-PEG3-SS-PEG4-biotin -complex release is predominantly the result of linker cleavage and triggered by specific conditions. |
677ea1d281d2151a025203f0 | 8 | Longitudinal measurements of test line intensity over 10 minutes were performed and indicated that maximal cleavage of complexes formed using FabHER-PD-PEG3-SS-PEG4biotin occurred in under 5 minutes (Figure ). More specifically, 82.3% (± 11.8%) test line cleavage was observed within the 1.5 minutes of test strips being interfaced with cleavage reagent, showcasing the rapidity of the cleavage chemistry for triggered release within LFA relevant timeframes. |
677ea1d281d2151a025203f0 | 9 | Beyond assay reagents, efforts to engineer our AmpliFold approach -comprising sequential capture, cleavage, and re-binding operations -within a single, comprehensive paper microfluidic framework necessitated the ability to perform multi-step LFA processes with controlled liquid flow. For the purposes of demonstrating our 'capture-and-release' approach as a proof-of-concept, a rudimentary 'two-strip' LFA procedure was employed where the capture (PSA-printed) strip was manually interfaced with the detection (anti-FITC pAb printed) strip using a ribbon of adhesive tape to maintain membrane-tomembrane contact. In essence, this assembly would be performed following the capture of material and wash steps, prior to the introduction of a thiol-based cleavage buffer. A full step-by-step schematic outlining a typical AmpliFold workflow is shown in the Supplementary Information (Figure ). This procedure enabled us to explore the capabilities of our AmpliFold approach within a proof-of-concept design which, at this stage, necessitated additional user steps, such as the manual removal of the capture strip absorbent pad and the transfer of strips 'well-to-well'. Ultimately, it is envisioned that future iterations of an AmpliFold device could simplify the manual assembly of capture surface by anti-fluorescein antibodies. This binding process is different from that of the capture step as our FluoroPerAuNP conjugates demonstrate high degrees of fluorescein presentation across the entire particle surface, vastly improving the rate of successful collisions at the anti-FITC test line (Figure ). As a result, the re-binding of immunocomplexes during the detection step occurs with faster association kinetics and higher affinity, thus yielding sharper test line profiles and amplified signal-to-noise compared to initial capture stages. |
677ea1d281d2151a025203f0 | 10 | Having demonstrated the concept of 'capture-and-release' within a full AmpliFold format, our attention turned towards establishing a model example where the limits of signal amplification could be explored. Here, the receptor density of printed PSA was modulated to achieve a regime where captured sandwich immunocomplexes were uniformly distributed across a large capture area. Whilst poor capture affinity typically detriments the sensitivity of traditional LFAs, we hypothesise that such conditions represent an opportunity where enhanced analyte recovery, over a large capture area, can be leveraged to afford maximal signal amplification, when using a 'capture-and-release' approach. |
677ea1d281d2151a025203f0 | 11 | To probe the effect of capture affinity on complex distribution, capture strips -printed with 6 test lines of varying PSA concentrations (0.1, 0.5, 1 mg/mL) -were used to detect a high concentration of HER2 antigen (500 ng/mL). When examining the signal intensities for each test line, it was observed that strips printed with low receptor densities (0.1 mg/mL PSA test lines) resulted in similar levels of signal across all test lines. This highly uniform distribution of signal represents the ideal conditions for our 'capture-and-release' approach where a large receptor area circumvents poor test line affinity. In this scenario, a significant proportion of analyte target, which would otherwise go undetected in traditional LFA formats, can be realised through the AmpliFold 'capture-and-release' approach, whereby triggered release enables the distributed immunocomplexes to be rebound during detection strip in a concentrated manner. This is in contrast to assays under 'high sensitivity' regimes (1 mg/mL PSA test lines) where the majority of signal tended sharply towards the front of the capture area. Under these conditions, the benefits of a large capture area were diminished with signal distribution more closely resembling a typical LFA test line. In this scenario, the amount of additional target analyte recovered by employing a large capture area is less impactful, reducing the scope for signal amplification by 'capture-and-release', when compared to a traditional LFA format. |
677ea1d281d2151a025203f0 | 12 | Using a low receptor density regime to emulate an ideal distribution of analyte capture across a large capture area, dose-response studies were conducted to evaluate the optimal surface area of printed PSA when eliciting signal amplification in 'capture-andrelease' AmpliFold assays (Figure ). In dose-response studies, serially diluted samples of HER2 antigen were run up assays and test line signal were measured in order to determine the minimum concentration of antigen, or the limit-of-detection (LOD), that is detectable either a traditional or AmpliFold LFA system. |
677ea1d281d2151a025203f0 | 13 | In general, larger capture areas -represented by increasing numbers of test linesresulted in notable sensitivity enhancement within an AmpliFold approach. When compared to a 'traditional' LFA example (i.e. a single, uncleaved test line), no significant difference in LOD was determined when detecting HER2 antigen using AmpliFold assays featuring single (1×) PSA test line capture strips. In this scenario, the amount of initial analyte captured is equivalent in both LFA formats, though some additional sensitivity loss is incurred by the 'capture-and-release' approach due to a small percentage of LFA format against the 'capture-and-release' AmpliFold approach under model capture affinity conditions (0.1 mg/mL PSA, low capture receptor test lines). Capture strips were printed using a varying number of test lines to probe the effects of larger capture areas on signal amplification. Data was collected in triplicate (except blanks where N=6) and plotted using a four-parameter logistic fit. d. A table collating the limit of detection for AmpliFold assays featuring different sized capture areas and a 'traditional' LFA format. |
677ea1d281d2151a025203f0 | 14 | So far, the utility of the 'capture-and-release' strategy has been demonstrated to highlight the improvements in sensitivity that can be obtained when low affinity surface capture limits assay performance. This same issue also affects the surface anchoring of larger, higher signal nanoparticles in LFAs where reduced diffusivity and sterically hindered collisions between immunoassay components at the solid phase reduce sensitivity. Here, we assess the performance of large nanoparticles, and particularly their deployment as detection labels in the AmpliFold architecture. |
677ea1d281d2151a025203f0 | 15 | demonstrated that, when measuring the intensity of different AuNP sizes, the LOD for visual detection of 115 nm AuNPs was roughly 467-fold lower than that of smaller 16 nm particles, when spotted directly onto membrane. Nanoparticle size also significantly influences a particle's binding affinity towards its target by virtue of the available surface area for receptor functionalisation. For antibody-nanoparticle conjugates, this enables greater numbers of antibodies per particle and increased particle valency when binding target. |
677ea1d281d2151a025203f0 | 16 | Whilst larger nanoparticles represent greater assay sensitivity in theory, their benefit in practice is offset by poor kinetics. As modelled by Zhan et al., large particle complexes (>100 nm) were shown to have slower diffusion rates which limited LFA sensitivity due to poor test line capture. Authors also noted that larger particles were associated with increased non-specific binding, limiting their effective concentrations in assays. To demonstrate this in the context of our work, a high concentration of HER2 antigen (500 ng/mL) was detected across test strips -printed with a capture area of 9 PSA test lines -using 20, 40, and 150 nm FluoroPerAuNPs (Figure ). Particles were prepared using identical physisorption conditions and similar antibody per surface area loading ratios (12.6 nm 2 / Ab), whilst test lines were printed using a high receptor density of PSA (1 mg/mL). When normalising the signal intensity at each test line, we observed that assays deploying 150 nm FluoroPerAuNPs yielded a more even distribution of signal across the capture area. By comparison, the use of smaller FluoroPerAuNPs (20 and 40 nm) resulted in signal distributions which instead tended sharply towards the front of the capture area, further indicating the trade-off between particle size and capture affinity. |
677ea1d281d2151a025203f0 | 17 | These signal artifacts were noted as absent in HER2-negative controls, indicating perturbation of mass transport at the edge regions of test strips. We theorise that this originates from the compression of pores which results from membrane cutting during test strip preparation. Using 150 nm FluoroPerAuNPs to highlight the potential of our 'capture-and-release' AmpliFold approach, we hypothesised that, through using a large capture area to circumvent poor capture affinity, we could subsequently capitalise on our cleave and re-bind strategy to yield significantly improved signal-to-noise. When evaluated against the 'traditional' LFA approach, our optimised AmpliFold assay was shown to detect HER2 antigen with a roughly 12-fold sensitivity enhancement at a LOD of 0.716 ng/mL (95% CI of 0.45 to 1.14 ng/mL) (Figure ). (1 mg/mL PSA). c. Dose-response data comparing the use of 150 nm FluoroPerAuNPs in a traditional 'uncleaved' LFA format and an optimised 'capture-and-release' AmpliFold assay when detecting HER2 antigen spiked in controlled buffer conditions. Data was collected in triplicate (except blanks where N=6) and plotted using a four-parameter logistic fit. |
677ea1d281d2151a025203f0 | 18 | The application of our AmpliFold format was further evaluated in experiments where HER2 antigen was spiked into a human serum. Although this work aims to report AmpliFold primarily as a proof-of-concept using a model antigen system, these experiments sought to provide some insight into the robustness of AmpliFold chemistries and performance when detecting analyte targets within a complex sample matrix where the additional step of incorporating control lines was performed. |
677ea1d281d2151a025203f0 | 19 | For traditional LFAs and AmpliFold capture strips, anti-FITC control lines were printed after PSA test line regions and used to bind excess 150 nm FluoroPerAuNP conjugates (examples shown in Figure and). To introduce a control line in AmpliFold detection strips, a chicken IgY (CIgY) and anti-chicken IgY (anti-CIgY) antibody binding pair was utilised, yielding red signal at the control line (examples shown in Figure ). Studies of the 40 nm CIgY and anti-CIgY control line binding system were performed to characterise particle functionalisation (Supplementary Information Figure ) as well as validate the performance of the binding system when integrated into AmpliFold workflows (Supplementary Information Figures ). |
677ea1d281d2151a025203f0 | 20 | Following the incorporation of LFA control lines, we initially conducted a matrix effect study in diluted human sera to identify an appropriate sample dilution that could be used to effectively compare the traditional and AmpliFold formats of the assay. Immunoassays typically observe matrix effects in complex physiological fluids due to competing interactions between various components of the assay and the sample. In this study, the assays were run as previously, but with the volume fraction of human serum to running buffer adjusted to cover a range of human serum fractions from 0 (pure running buffer) to 50% v/v. These samples were prepared such that the concentration of HER2 added was |
677ea1d281d2151a025203f0 | 21 | In traditional LFA experiments, sample mixtures containing a total of 12.5% v/v human serum was shown to produce the greatest signal-to-noise, as well as the least variation between measured test line signals (Figure ). AmpliFold sample mixtures containing a total 12.5% v/v were similarly shown to result in the greatest signal-to-noise, when compared to AmpliFold experiments conducted solely in running buffer conditions (Supplementary Information Figure ), and negligible false positive test line signals (Supplementary Information Figure ). In both LFA formats, increasing the percentage of human serum in sample mixtures was noted to have a detrimental impact of test line signal intensities for HER2-positive (500 ng/mL) tests as well as increased variation in the signals generated. We theorised that the detrimental impact on test line signal for HER2positive tests could reflect the non-specific interference of FluoroPerAuNP and FabHERbiotin conjugate anti-HER2 binding sites by human serum components, such as albumins, hormones, enzymes, and IgG antibodies. Although endogenous free thiols could also play a role in the observed matrix effects -particularly regarding the undesirable cleavable of FabHER-biotin conjugates -these would be expected to be present at drastically lower concentrations than DTT in these experiments. Ultimately, for a fair comparison of the traditional and AmpliFold LFA formats, the percentage of human serum in assay sample mixtures for both formats was kept the same, following this study, in order to match and matrix effects in the two formats as best as possible. In subsequent dose-response experiments, a dilution series of HER2 antigen spiked into human serum was prepared. A total human serum concentration of 12.5% v/v in assay sample mixtures was achieved by prediluting the typical antigen volume by half, prior to mixing in with FabHER-biotin (4) and 150 nm FluoroPerAuNP conjugate volumes. When testing serial dilutions of HER2 antigen in human serum, our AmpliFold assay was shown to detect the target analyte with a LOD of 3.35 ng/mL (95% CI of 2.37 to 4.74 ng/mL) (Figure ). This was shown to be a 12-fold enhancement in sensitivity when compared to traditional LFA formats where the LOD for HER2 antigen detection was determined to be 38. Overall, in both running buffer and real matrix conditions, our AmpliFold approach was shown to greatly enhance LFA performance when using large 150 nm FluoroPerAuNP labels. Despite this, the magnitude of sensitivity enhancement observed in these studies (ca. 12-fold) was lesser than that of our 40 nm FluoroPerAuNP model example where a 16-fold improvement was exhibited at minimal capture efficiency (Figure ). To rationalise this, we noted that the signal distribution of capture strips featuring 150 nm FluoroPerAuNPs (Figure ) remained comparatively uneven, when compared to the ideal uniform distribution obtained under low receptor density capture regimes (Figure ). We anticipate that larger particle systems, beyond what we shown, would tend towards more evenly distributed capture regimes based on diffusivity arguments. This further highlights the potential of our 'capture-and-release' AmpliFold approach in combination with existing LFA enhancements, such as using flurorescent nanodiamonds or latex microspheres, where larger particles play key roles in obtaining higher sensitivity. We envisage that there is also scope for future work to further optimise the workflows presented in this work, particularly with a focus on reducing the number of wash steps and the overall need for manual operation steps. Addressing these limitations are key towards affording faster assay runtimes and greater ease-of-use for future generations of AmpliFold design, ultimately bringing the technology closer towards viability at the point-of-care. |
677ea1d281d2151a025203f0 | 22 | Linkers were evaluated in lateral flow assay studies to determine the rapidity and efficiency of thiol-induced release from streptavidin bearing test-lines on nitrocellulose membranes. We show that an elongated azido-PEG3-SS-PEG4-biotin linker exhibited particularly rapid release kinetics within minutes which, when coupled with the exceptional binding capabilities of our 'dual-affinity' labels, enabled high-affinity re-binding of analytebound complexes at anti-FITC test lines of detection strips. Demonstrating the efficacy of our approach, we illustrate substantial sensitivity enhancements of 16-fold in a model system probing streptavidin capture receptor density and binding kinetics. Finally, we applied our approach to address a pervasive challenge in lateral flow assay development in the form of the limited diffusivity and capture kinetics of larger nanoparticles where, in a high sensitivity LFA system using 150 nm 'dual-affinity' labels, a 12-fold enhancement in limit of detection was realised when detecting antigen spiked in both buffer and human serum samples. Our 'capture-and-release' methodology emerges as a robust signal amplification technique with broad applicability across diverse LFA platforms, facilitated by adaptable protein modification strategies. Moreover, its rapid turnaround time (within 30 minutes) positions it as a promising diagnostic tool within the point-of-care testing landscape. |
677ea1d281d2151a025203f0 | 23 | Protein LoBind® microcentrifuge tubes were used to handle and contain samples containing proteins and reagents. Lateral flow immunoassays were conducted in nonbinding flat-bottom microplates (Grenier). Unless otherwise stated, all chemical reagents and proteins were purchased from Sigma Aldrich and did not undergo further purification before use. Concentrated buffer solutions, such as phosphate buffered saline (PBS, Gibco) and borate buffer (Thermo Fisher), were diluted to desired concentrations prior to use. Fab(Her) fragments (FabHER) were produced via the antibody digestion of Herceptin antibody (University College London Hospital) following a protocol by Szijj et al.. Nylon syringe filters (0.45 μm; Nalgene) were used to filter solutions containing semi-skimmed milk (Marvel) as blocking agent prior to assay use. |
677ea1d281d2151a025203f0 | 24 | The UV-Vis absorption spectra of nanoparticles were measured using a NanoDrop One (Thermo Fisher). Dynamic light scattering (DLS) and electrophoretic light scattering (ELS) was performed using a Litesizer DLS 500 (Anton Parr). Particle size was measured by using DLS to determine hydrodynamic diameter, whilst ELS was used to determine zeta potential. |
677ea1d281d2151a025203f0 | 25 | The following represents a typical physisorption methodology used for the functionalisation of 20, 40, and 150 nm AuNPs; exact quantities relating to each particle size can be found in the Supplementary Information. Particle-to-antibody stoichiometries were maintained such that a particle surface area to provide an antibody molar ratio of 12.6 nm 2 /antibody during conjugation. |
677ea1d281d2151a025203f0 | 26 | To an aqueous solution of fluorescein-tagged Pertuzumab conjugates (FluoroPer), borate buffer and AuNP solution (NanoComposix) were added sequentially prior to incubation under agitation. Following this, the physisorption mixture was blocked with blocking buffer and further incubated under agitation. Blocked mixtures were then centrifuged, forming a free-flowing pellet of colloid from which supernatant was extracted. The colloidal pellet was then washed by re-suspension in blocking buffer and then centrifuged prior to supernatant extraction. This washing procedure was repeated twice more prior to the characterisation of FluoroPerAuNPs. |
677ea1d281d2151a025203f0 | 27 | Details relating to assay components, test strip specifications, and operating procedures for particular experiments can be found in the Supplementary Information. The following is a typical protocol for half-stick LFA experiments: To a microplate well, assay components (antigen, FabHER conjugates, nanoparticles) were added and pre-mixed (5 min). To each well, printed test strips were added and allowed to wick (5 min). Strips were added into wells containing wash solution and allowed to wick (5 min). |
677ea1d281d2151a025203f0 | 28 | Capture strips were cut 17 mm from the bottom of the test strip, removing the absorbent pad. As depicted in the Supplementary Information (Figure ), a strip of adhesive tape (10 × 30 mm) was then used to fix and assemble the AmpliFold test, such that 2 mm from the top of the capture strip interfaced the detection strip, 2 mm from the bottom, with aligned, membrane-to-membrane contact. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.