id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
62db30f0a7d17e34206889cc | 3 | NMR dynamics experiments and molecular dynamics simulations with YopH and PTP1B have shown the acid loop samples both open and closed conformations throughout the catalytic cycle, although the equilibrium poise changes from favoring the loop-open form in the free enzymes, to the closed state when a ligand is bound. NMR data also reveal a correlation between the respective rates of WPD-loop closure and the cleavage rate (step 1 in Figure ) in YopH and PTP1B. The conserved Asp residue in the WPD loop serves critical roles in both chemical steps; as an acid in the first step, and a general base in the second. Why should an enzyme family evolve with a catalytic residue on a mobile loop? Evolutionarily, factors that affect loop motion could serve a regulatory purpose at the cost of maximally efficient catalysis. YopH, the fastest PTP yet characterized, exhibits the fastest WPD-loop dynamics measured to date. However, the "fastest" positioning of the Asp residue would be to place it on a non-mobile element, permanently in the optimal catalytic position. Such a stationary positioning is thought to be the case in many PTP family members on the basis of X-ray structures that show only a single conformation, although future work may reveal unsuspected mobility. Yet, those enzymes exhibit turnover numbers in the single digits, an order of magnitude slower than PTP1B, and two orders of magnitude slower than YopH. Variation in the rate of phosphoenzyme hydrolysis could reflect differences in optimal positioning of the Asp carboxylate, as well as the obvious role played by dynamics. Because only the WPD-loop closed conformation can perform catalysis, factors affecting the dynamic equilibria in Scheme 1, such as amino acid sequence within the WPD-loop, or interactions of the loop with surrounding regions of the protein, will affect both catalytic steps. |
62db30f0a7d17e34206889cc | 4 | The creation of loop-grafted chimeras is a proven tool to study key mobile portions of proteins, providing valuable insight into protein function and evolvability. In addition, it is an increasingly popular tool to generate new enzymes with novel reactivities and specificities. For example, the grafting of multiple loops has been used to design and evolve β-lactamase activity in glyoxalase II and the role of protein motions in β-lactamase activity has been studied using chimeras of two class-A β-lactamases, TEM-1 and PSE-4. A chimera of ribonuclease A and a homologue was used to identify the role of protein flexibility in the rate-determining step and to identify conserved dynamical traits in the RNAse superfamily. The exchange of surface loops was shown to convert trypsin to a chymotrypsin-like protease. Chimeras of the Pseudomonas and Burkholderia homologs of the Rh1A enzyme have been a key tool in protein engineering and directed evolution studies. Related to this, modifications to a key active site loop in the (βα)8barrel isomerase of histidine biosynthesis HisA during a real time evolution experiment allow this enzyme to diverge into HisA specialists, bifunctional HisA/TrpF specialists (the latter being an important step of tryptophan biosynthesis), and TrpF specialists. Finally, there is growing interest in the role of loop dynamics in protein evolvability and loop engineering and grafting as a tool for protein design. Most of the work in the field of loop dynamics has centered on surface and lid loops that cover the active site and their functional role in substrate and cofactor binding, protein-protein interaction, and stability. Crystal structures with bound analogs of peptide substrates show the WPD-loop does not act as a lid covering the PTP active site, and loop closure does not prohibit diffusion of the substrate and product (or water molecules) in and out of the enzyme. Rather, the primary purpose of WPD-loop motion is positioning of the key catalytic Asp residue into the active site from the side, allowing substrate binding or leaving group departure after the first step to occur from either loop position. But, as highlighted in the kinetic scheme (Scheme 1), only the loopclosed states perform catalysis. |
62db30f0a7d17e34206889cc | 5 | This is unique to the best of our knowledge, and provides a direct linkage between loop dynamics and the chemical steps of catalysis in PTPs. The work reported here is part of our broader efforts to understand factors affecting WPD-loop dynamics in the PTP family. Only the three WPD residues are highly conserved within the ~12-residues comprising this loop. To explore the role of non-conserved residues in the WPD-loop on catalysis and pH dependency, we have created and characterized a series of loop-swapped chimeras. In a previous report, we described chimeras in which the WPD-loop of PTP1B was exchanged into YopH, which resulted in unexpected and counterintuitive kinetic and structural results. Here, we report the characterization of reverse chimeras, based on the transposition of the faster, YopH enzyme loop into the slower enzyme, PTP1B. We began with a chimera consisting of the full YopH WPD-loop transposed into PTP1B, followed by a stepwise mutation back toward the native PTP1B loop sequence. Not all of the chimeras proved to be soluble and amenable for study, and we include those findings in the interest of including negative results. For the soluble, active Chimeras, enzymatic functions were examined by steady-state kinetics to obtain kcat values, and kinetic isotope effects with the substrate pNPP, which report on transition state details of the first step of catalysis. Degradation in the correct positioning of the Asp residue due to an altered WPD-loop conformation will affect the coordination of leaving group protonation with P-O bond fission in the first step of catalysis, and its role as a general base in the second step. |
62db30f0a7d17e34206889cc | 6 | Enzyme structures and dynamic behavior were characterized by X-ray crystallography, empirical valence bond simulations, and both conventional and enhanced molecular dynamics simulations, the former of which were in turn used to construct Markov State Models with which to compare loop transitions in the different enzyme variants. PTP1B is more tolerant of chimera substitutions than YopH, and these chimeras do not exhibit the structural disruptions found in the YopH-based chimeras. The combined results show how the non-catalytic amino acids within this mobile element of classical PTPs affect catalysis, particularly by altering WPDloop mobility and the equilibrium between open and closed states. More generally, the results highlight the potential for targeting non-catalytic residues as hotspots for manipulating the conformational equilibria of catalytic loops and thus regulating enzyme activity. 2 Materials and Methods |
62db30f0a7d17e34206889cc | 7 | Dithiothreitol (DTT) and ampicillin (AMP) were purchased from GoldBio. Restriction enzymes and primers were purchased from Integrated DNA Technologies. Protease-inhibitor tablets were purchased from Sigma-Aldrich. All other buffers and reagents were purchased from Sigma-Aldrich or Fisher. The substrate p-nitrophenyl phosphate (pNPP) was synthesized using published methods. Crystallography screens, trays, and coverslips were purchased from Hampton Research. |
62db30f0a7d17e34206889cc | 8 | The plasmid pEt-19b encoding the 37 kDa form of wild-type human protein PTP1B (residues 1 to 321) was provided by Dr. N. K. Tonks. The first chimera, designated Chimera 0, was made by substituting loop residues of wild-type PTP1B via the Q5-SDM kit (New England Biolabs), replacing with residues HVGNWPDQTAVS from the YopH WPD-loop in the corresponding region (Table ). Chimera 0 was used as the DNA template for the subsequent chimeras. Chimeras 1 to 7 were created using the polymerase chain reaction (PCR) with primers encoding residues before the mutation and the target mutation itself. The chimera DNA was then cleaved using restriction enzyme DpnI and ligated into the pEt-19b vector using T4 ligase. Each subsequent chimera was made using the previous chimera DNA as a template. The primers used are listed in Table . Chimeras studied in detail in this work are underlined in column 1. The designation "/" in the Activity column indicates an insoluble chimera for which activity could not be assessed. |
62db30f0a7d17e34206889cc | 9 | The DNA was transformed into BL21-DE3 cells and grown overnight at 37°C on an LB culture plate containing 100 µg/mL ampicillin. One colony was selected and placed into 10 mL SOC media containing 100 µg/mL ampicillin and grown overnight. The following morning, 1 L LB media containing 100 µg/mL ampicillin was inoculated with the 10 mL overnight growth and shaken at 170 rpm at 37°C until the OD600 reached 0.6-0.8 Abs. After reaching the optimal OD, the 1 L growth was induced by 0.1 mM isopropyl β-D-thiogalactoside (IPTG) and shaken at 170 rpm at room temperature overnight. The cells were harvested by centrifugation at 12000g for 30 minutes at 4°C and stored at -80 °C. |
62db30f0a7d17e34206889cc | 10 | All chimeras were expressed and purified as follows, based on the WT PTP1B protocol. The cells were thawed on ice and resuspended in 10X their equivalent volume of a lysis buffer, consisting of 50 mM imidazole pH 7.5, 1 mM EDTA, 3 mM DTT, and 10% glycerol with one crushed protease-inhibitor tablet for every 50 mL of solution. The cells were lysed by sonication at 60% power for 10 pulses then mixed on ice for 1 minute and repeated 4-6 times until completely lysed. The cell lysate was centrifuged at 4°C at 29000 g for 30 minutes. The supernatant was filtered with a 0.45-micron syringe filter. |
62db30f0a7d17e34206889cc | 11 | The cell lysate was loaded onto the columns at 1.5 mL/min, and the columns were washed with lysis buffer until the absorbance at 280 nm baselined. Elution was processed using a 100% gradient with elution buffer containing 500 mM NaCl, 50 mM imidazole pH 7.5, 1 mM EDTA, 3 mM DTT, and 10% glycerol. Eluted fractions exhibiting absorbance at 280 nm were collected and tested for phosphatase activity by addition of a few microliters of each fraction to a solution of pnitrophenylphosphate (pNPP), where an absorption increase at 400 nM indicated formation of pnitrophenol. Fractions that showed activity were assayed for purity on a 15% SDS-PAGE gel. |
62db30f0a7d17e34206889cc | 12 | The pooled fractions were loaded onto a desalting column and buffer exchanged into S-loading buffer (50 mM Bis-Tris pH 6.5, 1 mM EDTA, 3 mM DTT, and 10% glycerol). This solution was then loaded at 1.5 mL/min onto the equilibrated HiTrap™ SP HP column. The column was washed with the loading buffer until Abs280nm baselined, then eluted with S-elution buffer containing 500 mM NaCl, 50 mM Bis-Tris pH 6.5, 1 mM EDTA, 3 mM DTT, and 10% glycerol. Fractions with absorbance at 280 nm were collected and assayed for activity with pNPP, the fractions that showed activity were checked for purity on a 15% SDS-PAGE gel. |
62db30f0a7d17e34206889cc | 13 | The active fractions were pooled (ranging from 30-40 mL) and concentrated to <12 mL, loaded onto a pre-equilibrated HiLoad 26/60 Superdex 200 prepgrade column (GE) and purified using 10 mM Tris buffer pH 7.5, with 25 mM NaCl, 0.2 mM EDTA, and 3 mM DTT. Fractions were assayed with pNPP for activity and for purity on a 15% SDS-PAGE gel. Pure protein was concentrated to 10-35 mg/mL, and either immediately used for crystallization experiments or diluted with 10% glycerol and frozen with liquid nitrogen and stored at -80°C in aliquots. |
62db30f0a7d17e34206889cc | 14 | Crystals for Chimera 3 and Chimera 4 were grown by hanging drop vapor diffusion at 4 ºC using 10-15 mg/mL protein and a precipitant solution of 0.1 M tris hydrochloride pH 6.5-8.5, 0.2 M magnesium acetate tetrahydrate, and 20-25 % PEG 8000 at a 2:2:0.5 protein:well:20% benzamidine hydrochloride drop ratio. Both the tungstate and vanadate bound structures were obtained by adding either 3.9 mM sodium tungstate (Na2WO4) or 1 mM sodium metavanadate (Na3VO4) to the protein for co-crystallization. Crystals grew in 24 hours and were transferred to a cryo-protectant solution containing mother liquor, 20% benzamidine hydrochloride, and 50% sucrose before flash freezing in liquid nitrogen. |
62db30f0a7d17e34206889cc | 15 | Crystals for Chimera 7 were grown by sitting drop vapor diffusion at 4 ºC using 12 mg/mL protein. The crystallization drop was prepared by mixing 2 μL of protein solution, 0.5 μL sucrose 30 % (w/v) and 3 μL of precipitant solution (0.1 M HEPES pH 7.5, 0.2 M magnesium acetate and 15-20 % polyethylene glycol 8000). Single crystals were visible after three days. Cryoprotection was performed by transferring crystals stepwise into stabilization solution with increasing glycerol amounts to a final concentration of 15% and the respective initial concentrations of ligands present in the protein and precipitant solutions, and then flash-cooled in liquid nitrogen. |
62db30f0a7d17e34206889cc | 16 | Diffraction data for Chimeras 3 and 4 were collected on the Stanford Synchrotron Radiation Lightsource (SSRL) beamline 9-2, and diffraction data for Chimera 7 were collected on a home source (Rigaku Micromax 007/Raxis IV++) (Table ). Data were indexed and processed using DENZO and SCALEPACK in the HKL2000 program suite. Molecular replacement was performed with Phaser-MR as implemented in Phenix 57 and CCP4 55 using WT PTP1B (PDB ID 3I80 ) with active-site water molecules manually removed as a search model. Phenix.refine was used for refinement. Model building was performed using Coot. All figures of the enzyme structures and structural alignments therein were made using PyMOL (The PyMOL Molecular Graphics System, Version 1.2r3pre, Schrödinger, LLC.). |
62db30f0a7d17e34206889cc | 17 | Steady-state kinetic parameters were measured at 25 °C. Concentrated protein aliquots were thawed on ice and diluted with a buffer base mix (BBM) containing 50 mM sodium acetate, 100mM tris, and 100 mM bis-tris from pH 4.35-7.5. This buffer system maintains constant ionic strength throughout the pH range examined. A 50 mM solution of the dicyclohexylammonium salt of pNPP was prepared in the buffer base mix. Reactions were run on 96-well plates, diluted enzymes were added to reactions with substrate concentrations from 0.76-22.73 mM and allowed to proceed for 10 minutes. The reactions were quenched using 50 µL of 10 M NaOH, and the amount of the product was assayed from the absorption at 400 nm using the molar extinction coefficient of 18,300 M -1 cm -1 for p-nitrophenolate. Reaction blanks were made using identical conditions replacing the enzyme with buffer, to correct for non-enzymatic hydrolysis of the substrate. The amount of product released and elapsed time were used to calculate the initial rates. |
62db30f0a7d17e34206889cc | 18 | The 18 O KIE in the scissile P-O bond and the 15 N KIE in the leaving group were measured using the competitive method and isotope ratio mass spectrometry. Figure shows the positions in the substrate where KIEs were measured and the isotopic isomers used. KIEs measured by the competitive method are isotope effects on V/K, the part of the mechanism up to the first irreversible step, which is the first step shown in Figure , cleavage of the pNPP substrate. Concentrated protein aliquots were thawed on ice and diluted to 9.26 and 17.45 µM for Chimera 3 and Chimera 4 respectively in reaction with the buffer base mix. Natural abundance pNPP was used for measurements of 15 (V/K). N, O-labeled pNPP used for measurement of O isotope effects by the remote label method were synthesized using previously published methods. Isotope effect determinations were carried out in triplicate, at 25 °C in 100 mM Bis-tris at pH 5.5, by adding 200 µL enzyme and 108 micromoles pNPP in 5 mL of buffer and allowed to react until approximately 50% completion, approximately 2-3 hours. The enzymatic reactions were then stopped by titration to pH 4 with HCl, and the p-nitrophenol product extracted using diethyl ether and purified by sublimation. The residual substrate in the aqueous layer was completely hydrolyzed using bovine alkaline phosphatase at pH 9, and the p-nitrophenol released was isolated in the same way. The KIEs were calculated from the nitrogen isotopic ratios in the p-nitrophenol product at partial reaction (Rp), in the residual substrate (Rs), and the starting reactant (Ro). Eq. 3 was used to calculate the observed KIE from Rs and Ro at fraction of reaction f, and Eq. 4 from Rp and Ro. These were the same within experimental error and averaged to give the final results. |
62db30f0a7d17e34206889cc | 19 | Molecular dynamics (MD) simulations of wild-type (WT) PTP1B as well as the soluble proteins Chimeras 1, 3, 4 and 7, were performed using the Amber 2018 62 simulation package. All simulations were performed in the phosphoenzyme intermediate, which represents the reactive species for the rate-determining second chemical step (Figure ). NMR dynamics measurements on this species are not feasible due to its transient existence in solution. Simulations were performed for each enzyme starting from both the WPD-loop closed and open conformational states, using the PDB 63 structures outlined in Table for each system and state. Simulation starting structures of WT PTP1B were prepared as described previously, and Chimeras 1, 3, 4 and 7 were also prepared using the available crystal structures from both this and a previous study where possible. For simulations of systems with no available crystal structure, the most closely related (in terms of WPD-loop sequence) crystal structure was modified (through PyMOL mutagenesis) to generate the starting structure (no more than three mutations were required for any system, see Table for further details). |
62db30f0a7d17e34206889cc | 20 | All simulations were performed under periodic boundary conditions with octahedral water boxes using the ff14SB force field and TIP3P 66 water model alongside our previously developed parameters to describe the phosphorylated cysteine. Following the equilibration of each system in the NPT ensemble (298 K, 1 atm, see the Supporting Information), an initial round of production MD simulations of each enzyme in both WPD-loop conformational states were performed of each enzymes for twenty replicas of 1 µs length. Following this, k-means clustering was performed on the Cα RMSD of the WPD-loop of each system to identify 25 unique clusters. |
62db30f0a7d17e34206889cc | 21 | MSMs were generated using PyEMMA, with the inter-and intra-Cα-atom distances between the WPD-and E-loop residues of each system (for all pairs of residues that were at least 3 residues apart from one another) used as input features. Simulations from all 5 systems (WT PTP1B and all 4 chimeras) were combined for dimensionality reduction analysis via time-lagged independent component analysis (TICA), meaning all systems share the same time-lagged independent components (TICs). At this point each system was analyzed separately, first by clustering each system using the k-means clustering algorithm to produce between 125-200 clusters/microstates (depending on the system, see Table ). The MSMs generated were then validated by analysis of their implied timescales plots (Figure ), and this was also used to select an appropriate lagtime and number of metastable states for each system. |
62db30f0a7d17e34206889cc | 22 | Lag-times were between 25-100 ns whilst the number of metastable states chosen were between 4 and 7, see Table . These parameters were validated by a Chapman-Kolmogorov test for each system (Figures ). Metastable states were identified using the perron-cluster cluster analysis (PCCA+) were then used to determine the mean first passage times (MFPTs) for each system as implemented in PyEMMA. The free energy differences were obtained from the MFPTs by calculating the equilibrium constants from the forward and backward rates (and, from that, ΔG). |
62db30f0a7d17e34206889cc | 23 | catalyzed by Chimeras 1, 3, 4 and 7. All simulation setup, equilibration and EVB simulations were performed using the same setup as in our prior work, using the revised parameters provided as Supporting Information to ref. . The starting point for all simulations was crystal structures of each relevant chimera in its respective closed conformation, using the structures summarized in Table . |
62db30f0a7d17e34206889cc | 24 | water molecule were constructed/placed manually into the active site in order to optimize the position of the water molecule for nucleophilic attack on the phosphate. In Chimera 3, the catalytic aspartic was found in the crystal structure in an unproductive conformation and thus was rotated into a productive conformation taking the highest probability rotamer from the Dynameomics rotamer library 72 (probability 0.66, top-ranked out of 9 possible rotamers) as implemented in UCSF Chimera, v. 1.14. In addition, the side chain of Gln261 was rotated to match the rotamer found in PDB ID: 3I80, which is the conformation it is expected to take during the hydrolytic step of catalysis. |
62db30f0a7d17e34206889cc | 25 | A complete list of all ionized residues and histidine protonation patterns for each system can be found in the Supporting Information (Table ). Empirical Valance bond (EVB) calculations performed using the Q6 simulation package and the OPLS-AA force field, for consistency with our previous work. Each reaction step/system was simulated for 30 replicas, using an initial 30 ns of equilibration starting from the approximate transition state (λ = 0.5, Figure ). Production simulations were then propagated downhill from the transition state using 51 mapping windows in total, each with a simulation time of 200 ps. Simulation analysis was performed using CPPTRAJ. 3 |
62db30f0a7d17e34206889cc | 26 | The presence of a key catalytic residue on the WPD-loops of classic PTPs provides a direct link between loop motion and chemistry, and previous results have shown a correlation between WPD-loop dynamics and catalysis in YopH and PTP1B. Enzymes in the PTP family have variable residues within this mobile loop except for the WPD-residues themselves (Figure ). A suggestive hypothesis is that the identity of the intervening residues affects loop dynamics, and thus the catalytic rate, by altering hydrogen bonding interactions. Changes in populations of catalytically functional and nonfunctional states can also alter pH-rate dependencies, protonic equilibria, and the relative contribution of individual steps to the rate across the pH range. One goal of this study was to ascertain whether a WPD loop can be transposed intact into a related PTP and bring with it the rate and pH dependency of its source. If so, exchanging the YopH WPD-loop into PTP1B should confer a faster rate and altered pH dependency. Or, if interactions with neighboring regions of the protein are important, such a chimera might differ from either parent. In this study, the initial Chimera 0 with the YopH WPD-loop residues transposed into PTP1B was systematically restored to the native PTP1B. |
62db30f0a7d17e34206889cc | 27 | Table shows which residues were initially swapped from YopH and their mutation back to native PTP1B residues in subsequent chimeras; the active site hydrogen bonding patterns in wildtype PTP1B are shown in Figure . In the initial chimera construct, designated Chimera 0, eleven residues in the PTP1B WPD-loop region were swapped for the corresponding residues from YopH. |
62db30f0a7d17e34206889cc | 28 | This protein was insoluble, potentially due to unfavorable interactions involving the hydrophobic sidechain of Val176. The corresponding residue in native PTP1B is Tyr176, whose hydroxyl group forms a hydrogen bond with the sidechain of Ser190. Val176 was restored to Tyr as is native in PTP1B, yielding Chimera 1, which was soluble but inactive. Some of the subsequent chimeras were insoluble, for reasons that are not evident, and were not characterized. Chimeras 3, 4 and 7 |
62db30f0a7d17e34206889cc | 29 | were soluble, catalytically active, and were kinetically and structurally characterized. These chimeras have 5, 4 and 1 amino acid substitutions, respectively, as shown in Table . The final Chimera 7 carries a single amino acid substitution F182Q. This residue is adjacent to the general acid; previous work has hypothesized that the reorientation of the peptide bond joining these residues affects the energy barrier for loop movement. |
62db30f0a7d17e34206889cc | 30 | The The crystal structure of Chimera 7 showed the presence of a HEPES molecule from the crystallization buffer at the active site (Figure ). The sulfonyl group makes hydrogen bonds to P-loop residues and to Arg221, analogous to those of a phosphate ester substrate, and the WPDloop is in the closed conformation. The alkyl ring of the HEPES molecule is in a position that would clash with the side chain of F182 in native PTP1B. Other details of HEPES interactions are shown and discussed in the Supporting Information. |
62db30f0a7d17e34206889cc | 31 | Vanadate is notable for its ability to adopt a trigonal bipyramidal geometry, providing a structural transition state analog (TSA) for phosphoryl transfer. The vanadate-bound crystal structures show an apical interaction with the sulfur of cysteine and a trigonal bipyramidal geometry, an analog of the transition state for the second step shown in Figure . The active sites in the vanadate-bound complexes of Chimeras 3 and 4 are highly superimposable, with an RMSD of 0.07 Å when aligning 8 P-loop residues, H214-R221.The vanadium-sulfur distance is 2.6 Å and the apical V-O distance is 1.8 Å. These compare well to the corresponding distances in the vanadate structure of WT PTP1B, where the same TS analog has a V-S distance of 2.5 Å and V-O (apical) distance of 2.1 Å. These results are consistent with conclusions from KIE data that the transition states for the chimera-catalyzed reactions are similar to the native enzyme. |
62db30f0a7d17e34206889cc | 32 | and YopH is oriented towards the P-loop, where it protonates the leaving group oxygen of the bound substrate. However, in the tungstate-bound chimera structures, this side chain is pointing away from the active site, despite the lack of any significant change in the conformation of the WPD-loop backbone. This atypical aspartate sidechain conformation is facilitated by hydrogen bonding with Lys116 and Gln182, and by coordination with a magnesium ion from the crystallization solution. Magnesium is reported to have a modest activating effect on native PTP1B. This unproductive orientation provides another possible rationale for their reduced rates. |
62db30f0a7d17e34206889cc | 33 | This led us to consider the possibility that these chimeras might exhibit inhibition by Mg 2+ , in contrast with the accelerating effect of Mg 2+ ion for WT PTP1B. The interaction of Asp181 with Mg 2+ ion is not seen in WT PTP1B tungstate-bound structure, nor with the corresponding Asp356 in tungstate-bound YopH. To assess whether this unproductive conformation of Asp181 occurs in solution, and whether the presence of magnesium ion contributes to its population, kinetic data were collected on Chimera 4 in the presence of Mg 2+ and showed an inhibitory effect (Figure ). The inhibition data were fitted to the competitive, uncompetitive, and mixed inhibition models. The mixed inhibition model gave the best fit results and yielded a Ki value of 4.8 ± 0.2 mM. The structural data explain why this nonproductive conformation is unique to the chimeras. |
62db30f0a7d17e34206889cc | 34 | In WT YopH, the Lys loop providing one of the Mg 2+ -coordinating residues is absent, and no other residue is in close proximity for hydrogen bonding with this nonproductive conformation of Asp181. In WT PTP1B, residue 182 is phenylalanine, which lacks a hydrogen bond acceptor for this conformation of the Asp sidechain. Hence, the network for coordinating a magnesium ion can only occur in the chimeras. Only Chimera 4 was tested, but Mg 2+ inhibition is expected in both Chimeras 3 and 4 based on their structural data. This is another example of how changes to residues in the WPD-loop that are not directly catalytic can indirectly affect catalytic characteristics with potential regulatory consequences. |
62db30f0a7d17e34206889cc | 35 | Chimeras 3, 4 and 7 are slower than either of the parent enzymes, but all exhibit the bell-shaped pH-rate profile characteristic of the PTP family (Figure ). In Table , the pH optima and kinetic constants for the chimeras are compared with their parent enzymes, and with the most active YopH-based chimera from a previous investigation. The Chimera 3 and 4 turnover rates (kcat) |
62db30f0a7d17e34206889cc | 36 | The residue at this position was recently shown to affect the conformational dynamics of native PTP1B, causing a population shift towards the catalytically-active closed conformation of the WPD loop even in the ligand-free form of the enzyme. Chimera 7 (the F182Q point mutant) is about an order of magnitude slower than WT PTP1B in k cat , with no shift in the pH optimum but a broader maximum. The kinetic pKa values obtained from fits of the pH-rate data using Eq. 1 reflect the nucleophilic Cys215 (pKa1), and the general acid, Asp181 (pKa2). The kinetic pKa values for the Cys residues in both native enzymes (5.1 in PTP1B, 4.6 in YopH, Table ) are reduced from that of cysteine in solution by a hydrogen bonding network that stabilizes the thiolate anion. The kinetic pKa values of the aspartic acid in PTPs are less perturbed from the solution value. The kinetic pKa values of the chimeras differ from those of the WT parent PTP1B (Table ), particularly for the aspartic acid. Several factors can cause kinetic pKa values extracted from pH-rate data to be distorted from the true thermodynamic pKa values of ionizable catalytic residues. Because WPDloop motions correlate with catalysis in PTP1B and YopH, mutations that affect loop motion may not only affect the rates of the chemical steps, but also change the degree to which particular steps are rate-limiting within the pH range examined, leading to changes in pH-rate profiles such as those observed. The modest changes in the kinetic pKa values of the Cys residue in the Chimeras are most likely due to such effects. The greater differences in the D181 data (Table ) arise from differences in the hydrogen bonding network to D181, due to changes within the grafted loop that affect its thermodynamic pKa. The results here are consistent with the notion that differences in the loop residues and the altered hydrogen bonding patterns affect loop dynamics, and thereby alter the pH profile in ways that include broadening or moving the pH optimum. This provides an explanation for the fact that the broadness/narrowness and the optima of pH-rate profiles within the PTP family vary despite conservation of their ionizable catalytic residues. |
62db30f0a7d17e34206889cc | 37 | Loop closure is necessary for general acid catalysis, and the pH-rate profiles indicate the chimeras retain this function. The 18 O kinetic isotope effect (KIE) in the scissile P-O bond and the Computational analysis of the reaction energetics by EVB calculations, and molecular dynamics to compare conformational differences between the chimeras, provided insight into the origins of the experimental observations. |
62db30f0a7d17e34206889cc | 38 | Figure and Tables and show a summary of experimental and calculated free energies for the rate-limiting hydrolysis of the phosphocysteine intermediate (Figure ) by PTP1B, YopH and each of Chimeras 1, 3, 4 and 7, obtained from EVB calculations, as well as the corresponding electrostatic contributions to the calculated activation free energies. Figure and Table show the structures of representative stationary points from calculations for each Chimera, as well |
62db30f0a7d17e34206889cc | 39 | as key distances at each stationary point. From this data, it can be seen that all chimeras are less active than the parent enzymes, although, similarly to our prior work, the changes in the calculated values are not large enough to account for the larger observed changes in kcat (Table ), further indicating the putative importance of dynamical effects, as in our study of the two parent enzymes. In particular, of the residues identified from our EVB simulations as having larger differences in their electrostatic contributions to catalysis (Figure ), R112, E115, K116 and K120 are all residues found on the E-loop. As in our prior computational work, the calculated transition states for each system are very similar; however, as can be seen from Figure , there are in particular differences in the electrostatic contributions from the side chains of K120 and F182/Q182, as well as the side chain of E115. kcal mol -1 ) side chain contributions to TS stabilization can be found in Table . |
62db30f0a7d17e34206889cc | 40 | We note here that the EVB approach is an extensively validated approach for studying chemical reactivity in enzymatic systems, and a well-calibrated EVB potential typically provides activation free energies within 1-2 kcal mol -1 from experimental values across enzyme (variants). Our group has substantial experience of applying this approach to a range of phosphotransferases, including PTP1B and YopH, and the fact that we obtain lower activation free energies than would be expected from the changes in turnover number is unlikely to be a computational artefact of model parameterization. Taken together, this further emphasizes that the contribution of the E-loop to catalysis is both dynamic and electrostatic, and that factors such as WPD-loop sequence that alter E-loop dynamics also alter the electrostatic contributions of key E-loop residues, thus impacting activity. |
62db30f0a7d17e34206889cc | 41 | Molecular dynamics simulations provide insight into how the conformational dynamics of the WPD-loops in the chimeras differ from the WT enzymes, and to the origins of their slower catalytic Focusing first on Chimera 7, our ΔRMSF data from our MD simulations demonstrate the flexibility/rigidity of the WPD-loops of Chimera 7 and WT PTP1B to be highly similar in their open conformations (Figure ). Likewise, our H-bonding analysis only identifies one major difference, which is the newly formed hydrogen bond (H-bond) between F182Q and D181 in Chimera 7. This new H-bond does not appear to notably alter the conformational dynamics or protein interaction network in PTP1B, however given its direct interaction with the catalytic acid/base, it may have an important electrostatic effect on both chemical steps of catalysis, as also suggested from our EVB simulations (Figure ). Chemical intuition would suggest that mutation to glutamine would increase the acidity of the catalytic aspartic acid by placing more negative change on the acid's oxygen atoms through a side chain H-bond. This would be expected to have a beneficial effect on the rate of the first step in which the WPD-loop acts as an acid. However, in the second (rate-limiting) step where the WPD-loop acts as a base, the decreased basicity on the aspartic acids' oxygen due to mutation of the adjacent residue to glutamine could ultimately result in an increased activation free energy, which would explain the decreased kcat values observed for Chimera 7 compared to the WT PTP1B. Indeed, our EVB simulations (Figure and Table ) |
62db30f0a7d17e34206889cc | 42 | Furthermore, were the WPD-loop able to close normally, our EVB simulations predict the hydrolysis step of catalysis catalyzed by Chimera 1 to have a similar activation free energy to that of other Chimeras modeled here (Figure and Table ). It is possible, however, that dthe WPD-loop of Chimera 1 can adopt another conformation that is more favorable than the closed WPD-loop conformation (that is separated by a reasonably large energetic barrier), which would preclude it from being able to attain a catalytically competent loop-closed conformation. |
62db30f0a7d17e34206889cc | 43 | Molecular dynamics (MD) simulations (50 µs of accumulated sampling per system) were used to generate Markov State Models (MSMs) describing how the conformational sampling of the WPD-loop is altered for each chimera. The MD simulations were performed at the phosphocysteine intermediate (the starting point for the rate-limiting hydrolysis step of the reaction, Figure ), We note that WPD-loop closure is essential for correct positioning of the catalytic aspartic acid (Figure ) in the active site, therefore conformational fluctuations of the WPD-loop at this intermediate have the potential to adversely affect turnover. MSMs for each system were constructed using the Cα-atom distances of the WPD-loop and neighboring E-loop residues. The E-loop residues were included due to the importance of the "correct" sampling of several E-loop residues to provide transition state stabilization to WT-PTP1B. MSMs for each system are shown in Figure , with the free energy landscapes (FELs) of each system described by the first two time-lagged independent components (TICs), which represent the slowest motions of the systems (see the Materials and Methods). We note that the TICs used to describe each system are identical (i.e., TIC1 in WT-PTP1B is identical to that of TIC1 in all the chimeras). Analysis of the FELs of WT-PTP1B and Chimera 7 (which differ only by the F182Q substitution, see Focusing first on changes in the closed WPD-loop conformation for each chimera as compared to the WT (Figure ); the ΔRMSF values between WT PTP1B and Chimera 7 which differ from one another by a single point mutation (F182Q) are highly similar with the exception of the Eloop. The H-bonding analysis identifies a newly formed hydrogen bond (H-bond) between the side chains of F182Q and D181 in Chimera 7 (Figure ) and a reduction in the occupancy of the Hbonds between D181 to R221 (on the P-loop) and K120 (on the E-loop), which may have therefore altered the stability of the E-loop. For Chimeras 1, 3 and 4, however, there are differences in the rigidity of several regions around the WPD-loop as well as the WPD-loop itself, including the Eloop, Loop 11 and the N-terminal section of the α7-helix (Figure ). Both Loop 11 and α7-helix are known parts of the PTP1B allosteric network that regulates the WPD-loop conformation, suggesting the mutations present in these chimeras have also impacted this allosteric network. The new intra WPD-loop H-bond between G183T and P180 is likely partially responsible for this rigidification alongside the new side chain H-bond between T178N and T177 (Figure ). RMSFs in the open WPD-loop conformation and their H-bonding interaction networks are virtually identical (Figure ). Given the single point mutation separating the two enzymes (F182Q) points out to solvent in this conformation, these results are perhaps unsurprising. |
62db30f0a7d17e34206889cc | 44 | Significant differences in both the flexibility and hydrogen bonding network were observed for Chimeras 1, 3 and 4, and these changes were not just limited to the WPD-loop, but also some surrounding regions, most notably the E-loop and Loop 11 compared to WT PTP1B (Figures and). The similar stability observed for the WPD-loop across the WT and Chimeras is likely an artifact of the measurement process, in which only frames with a WPD-loop RMSD ≤ 1.5 Å to the open crystal structure conformation was taken forward for the RMSF calculations (Figure ). Analysis of the differences in the H-bonding networks of these Chimeras to WT PTP1B (Figure ) suggest the mutation G183T (present in Chimeras 1, 3 and 4) is primarily responsible for altering the stability of these regions through a new hydrogen bond between the side chain alcohol of G183T to D265. This appears to "pull" the WPD-loop, E-loop and Loop 11 apart, substantially weakening many key interactions (relative to WT PTP1B) between these loops (Figure ), thereby increasing their observed flexibility, and likely giving rise to the overall reduced stability of this conformation as seen in our Markov state models (Figure and Table ). |
62db30f0a7d17e34206889cc | 45 | The sensitivity of the E-loop to changes in the WPD-loop is consistent with a recent NMR study and our recent work in which we identified that the dynamics of these loops were correlated with one another. Likewise, Loop 11 is part of a previously characterized allosteric network which modulates the WPD-loop conformation. Moving from Chimera 4 to Chimera 3, we observe further destabilization of both the WPD-and E-loop, and in particular at the N-terminal portion of the WPD-loop. Given the nature and location of the mutation that separates Chimeras 4 and 3 (T177G, located at the N-terminus of the WPD-loop) it is perhaps not surprising to see this effect. Finally, differences in stability are observed between Chimera 3 and Chimera 1 and are located towards Loop 11 and the α7-helix. It is unclear from the H-bond networks how these two mutations (P185V and E186S) that separate Chimera 3 and Chimera 1 have resulted in altered stabilities at Loop 11 and the α7-helix, these changes may be due to alterations in a different type of non-covalent interaction(s). |
62db30f0a7d17e34206889cc | 46 | Although the WPD-loop backbones in the open and closed conformations of YopH and PTP1B are highly superimposable in crystal structures, our results show that the WPD-loops are not simple transposable elements. The sequence variation within the WPD-loops of various PTPs is a source of variations observed in the pH dependency and maximal catalysis rates in this enzyme family. |
62db30f0a7d17e34206889cc | 47 | The "information" that programs their loop dynamics is not contained within their sequences alone. The exchange of the WPD-loop from the faster YopH enzyme into the corresponding loop of PTP1B does not result in a faster enzyme; all of the PTP1B chimeras exhibited slower rates than either WT parent (Table ). |
62db30f0a7d17e34206889cc | 48 | The kinetics and KIE data indicate that their reduced catalytic rates compared to the parent enzymes, and differences between the chimeras themselves, do not arise from changes in the mechanism or to the rate-determining step of catalysis. Even moderate differences in WPD-loop sequence identity between the chimeras and native PTP1B result in significant effects on their state, affects catalytic turnover, as was similarly observed, for instance, in prior work on TIMbarrel proteins. By influencing the population of catalytically competent enzyme-substrate complexes, the pH profile of catalysis can be affected, as well as the limiting rate. Thus, variations in sequence of noncatalytic residues within the WPD-loop provides a means for nature to fine tune these enzymes, and likely contributes to the variation in rate and pH-dependency (broad versus narrow) in the PTP family despite strict conservation of the ionizable residues directly involved in the chemical steps. In another example, the change of the residue following the general acid from phenylalanine to glycine, the residue found in this position in WT YopH, switches the effect of Mg 2+ from an activator to an inhibitor. |
62db30f0a7d17e34206889cc | 49 | Finally, we demonstrate that although the WPD-loop may appear to be a simple decorating loop that positions a key catalytic residue, in fact, changes in one part of the protein (substitutions on this loop) affect dynamics of other parts of the protein (e.g. the coupled dynamics of the E-loop and loop 11). Such coupled dynamics can be exploited in protein engineering, through the introduction of mutations distal to the active site, that can shift the overall conformational ensemble of the protein, including potentially controlling the dynamics of key catalytic loops. This is also significant in light of increased awareness of the role of the dynamics of decorating loops in the natural evolution of enzyme function, and the potential of altering loop dynamics and loop grafting as a powerful tool in protein design. Here, we demonstrate that non-catalytic loop residues are potential mutational targets for manipulating the conformational equilibria of key catalytic loops in PTPs, in addition to analogous effects from mutation of allosteric residues. |
63760db7be365e235822c184 | 0 | As the sulfur analogous of lactones, sultones are an important class of sulfur heterocycles that used widely in the synthesis of natural products, pharmaceuticals, biologically active molecules and advanced functional materials . Due to the remarkable synthesis importance, considerable efforts have been exerted to develop efficient methods for the synthesis of these valuable building blocks. Over the past decade, transition-metal catalyzed C-H insertion reactions 2 , cycloaddition reactions , radical reactions and metathesis reactions have been developed for the construction of sultones. Recently, Lupton and Qin 7 developed a tandem annulation reaction of easily tautomerizable ketones (or their TMS enol ethers) and ethenesulfonyl fluorides, which provide a novel protocol for the synthesis of δ-sultones. Despite these important progress, development of efficient and mild method for the construction of different functionalized δ-sultones is still highly significant. |
63760db7be365e235822c184 | 1 | On the other hand, allyl ketones are another type of synthetically valuable reagents that used widely in organic synthesis. Owing to the steric effect, these deconjugated carbonyls usually serve as vinylogous nucleophiles to undergo aldol reactions or Michael additions at γ-position selectively . In addition, the terminal carbon-carbon double bond of allyl ketones can also participate in Diels-Alder cyclization to form heterocycles . During our research on SuFEx click reactions 10 , we found that allyl ketones do not undergo the anticipated γ-selective Michael addition with the active ethenesulfonyl fluoride. Interestingly, an unexpected α-selective Michael addition-SuFEx cyclization reaction performed to produce γ-alkenylated δ-sultones efficiently. Herein, we would like to report this result. |
63760db7be365e235822c184 | 2 | Initially, the reaction of phenyl allyl ketone 1a and β-phenyl-substituted ethenesulfonyl fluoride 2a was examined. In the presence of 20 mol% DBU, γ-ethenylated δ-sultone 3a was obtained in 48% yield in DMSO at room temperature (Table , entry 1). We concluded HF generated in the reaction neutralized DBU and thus quenched the reaction. Increasing the amount of DBU to 1.0 equivalent, the yield of 3a was improved to 60% ). However, triethylamine and DIPEA promoted the reaction in low efficiently (Table and). Pleasingly, Inorganic bases such as Na 2 CO 3 , K 2 CO 3 , Cs 2 CO 3 and NaHCO 3 mediated the reaction in high yields (Table ). Further study showed that the combination of catalytic amount of DBU and stoichiometric amount of inorganic base can promote the reaction to afford the desired product in excellent yield (Table and). Interestingly, when 4 Å Molecular sieves was used instead of inorganic bases, DBU can also catalyze the tandem reaction in 96% yield (Table , entries 11-13). A brief screening of the reaction solvent showed that high polar solvents such as DMSO, DMF and acetonitrile give the desired product in high yield (Table , entries 12, 14 and 15), while DCM, THF, ethyl acetate showed low efficiency (Table , entries 16-18). Finally, control experiment showed that in the absence of DBU, no desired product was formed (Table , entry 19). Having evaluated the optimal reaction conditions, we then investigated the substrate scope of this tandem annulation reaction. As shown in Table , both electron-withdrawing and electron-donating substituents substituted aryl allyl ketones participated in the reaction efficiently, producing the corresponding δ-sultones in high yields (3a-3f). However, very strong electron-withdrawing cyanide group substituted allyl ketone only gave the desired product in moderate yield (3g). Different positions of the substituents have no obvious effect on the reaction yield (3h-3k). |
63760db7be365e235822c184 | 3 | Bulkyl naphthyl substituted allyl ketones underwent the reaction to produce the corresponding δ-sultones in high yield (3l and 3m). Heteroaryl substituted allyl ketones coupled with 2a to furnish the desired ethenylated δ-sultones in good yields (3a-3o). β-Alkyl substituted vinyl ketones were also proved to be competent reactants for the reaction, providing the corresponding δ-sultones 3p and 3q in 63% and 87% yield, respectively. |
63760db7be365e235822c184 | 4 | We next explored the scope of ethenesulfonyl fluorides and the results are summarized in Table . A variety of electron-withdrawing, -neutral and -donating substituents substituted β-aryl ethenesulfonyl fluorides smoothly underwent the tandem reaction to produce the corresponding products in high yields (3r-3ab). In addition, varied positions of the substituents can be well tolerated for the reaction (3ac-3an). However, when phenolic hydroxy group substituted ethenesulfonyl fluoride was used to react with allyl ketone 1a, the reaction was complex and no desired product was obtained (3ao). We assumed that the phenolic hydroxy group can undergo SuFEx click reaction or other side reactions, which restricted the desired tandem annulation reaction. In contrast, when the phenolic hydroxy group was protected with -SO 2 F group, the desired product was obtained in 84% yield (3aa). Naphthyl-substituted ethenesulfonyl fluorides were suitable reactants for the reaction, affording the corresponding sultones in excellent yields (3ap and 3aq). Heteroaryl substituted ethenesulfonyl fluorides can also undergo this reaction, albeit with relatively low yields (3ar-3at). The reaction involving ESF only gives 38% yield (3au). The structure of 3ab and 3au were determined via X-Ray crystallographic analysis. |
63760db7be365e235822c184 | 5 | In summary, we have demonstrated a tandem annulation reaction of allyl ketones and ethenesulfonyl fluorides. The mild and transition-metal free conditions. Simple procedure, generally high reaction yield and broad substrate scope provide a new method for the synthesis of γ-ethenylated δsultones. Further study of the applications of this method are ongoing in our laboratory. |
60c744babb8c1a21303da553 | 0 | Developments in computer assisted synthesis planning (CASP), specifically retrosynthetic analysis have gained considerable interest in recent years. The resurgence of artificial intelligence (AI) in computer aided drug design (CADD) has driven the shift from more traditional expert systems, built around a manually encoded set of reactions as templates, to data-driven approaches. Recent successes have been reported coupling neural networks to Monte-Carlo tree search (MCTS), and within reinforcement learning frameworks, deviating from more traditional expert systems. Their ability to rationalize a set of promising synthetic routes from reaction data, has been realized in the framework of Design, Make, Test, Analyze (DMTA) cycles, in which they have played an integral role for coupling to automation platforms. However, despite recent achievements in the field to advance predictive capability, little attention has been paid to the underlying datasets, the size of the dataset required, an assessment criteria specific to the template prioritization method and overall model performance. Retrosynthetic planning or analysis refers to the technique used by chemists to recursively deconstruct a compound into its simpler precursors, until a set of known or commercially available building blocks is reached. After an initial pattern recognition step, a chemist works in the reverse direction, using a knowledge-base of synthetic transformations ('synthetic tool-box') obtained through years of experience and exposure to a variety of both successful and failed chemistry, to intuitively identify and prioritize a promising set of forward transformations required to synthesize a given compound. To complement this process, computer assisted synthesis planning (CASP) tools are desired that can rapidly consider a vast body of chemical knowledge, effectively prioritize a set of reactions, and develop synthesis plans that can be tailored for the domain in which they will be applied. These have been reviewed extensively elsewhere. With the rise of automation, de novo design, and more extensive virtual libraries, such a tool has the added requirement that it must be able to pre-filter compounds prior to synthesis, thus reducing experimental failure and accelerating Design, Make, Test, Analyze (DMTA) cycles prevalent in molecular design. Herein, we investigate the role of the template prioritization method and the tree search algorithm derived from the work of Segler and Waller. Template prioritization is framed as a multi-class classification problem, for which we employ a neural network which outputs the probability of applying any given template, henceforth referred to as the filter network. This constitutes the machine learning (ML) part of the process, which we couple to a search strategy and decision-making process in the form of a tree search. Together these constitute an AI driven model for retrosynthetic planning. We examine this model in the context of the underlying datasets, pooling from internal AstraZeneca ELN, publicly available USPTO, proprietary Reaxys and Pistachio data. The overlap and relations between the datasets are examined. The final model's performance is tested on a set of 1,731 compounds from a set of 41 virtual libraries designed at AstraZeneca between October 2017 and January 2019, in relation to filter network accuracy, percentage of routes found, and the number of compounds synthesized experimentally. Thereby, demonstrating the potential use for such tools in DMTA cycles, and how datasets with known experimental results can be used to assess model performance and improvement of CASP tools. As such, we relate our findings of model performance to the underlying datasets. |
60c744babb8c1a21303da553 | 1 | Given the variety of data sources, patents (USPTO and Pistachio), literature and patents (Reaxys), and industrial data (AstraZeneca ELN), it is interesting to note that a comparable number of templates were extracted from the Reaxys and Patent datasets (Table ). However, whilst both template sets are similar in size they differ in their coverage of the reaction space as highlighted in Figure . The inclusion of the Reaxys data offers a greater breadth of unique reaction templates, accounting for 41.1 % of the overall combined dataset. The comparably high number of unique templates extracted from the combined patents data (32.5 %), suggests that a considerable portion of patents data covered are not present in Reaxys (7.4 % overlap), or that the structural components that make up the templates are unique to Reaxys. The exact differences between the patent coverage of the patent datasets (USPTO and Pistachio) and Reaxys is not clear with regards to the templates that can be obtained. Furthermore, the increased number of structural components and templates unique to the Reaxys dataset may be a residual artefact of multi-step reaction pathways. In this regard, we have filtered for all multi-step reactions, such that they have been removed from the dataset to the best of our knowledge. |
60c744babb8c1a21303da553 | 2 | The discrepancy between the two patent sets can be rationalized by the time-period over which the data was collected. The USPTO dataset accounts for reactions published up to September 2016 whereas Pistachio includes reactions until 17th Nov 2017. Further differences in the Pistachio and the public USPTO set arise from the inclusion of ChemDraw sketch data, and text-mined European patent office (EPO) patents which are included in Pistachio. The sketch data may be missing agent and condition details, as they are 'as drawn', and do currently not incorporate information from the accompanying text. Therefore, species that contribute a changing atom or bond may be absent and would not be incorporated in the template extraction. As this information cannot be included in the templates, the reaction is discarded, and no template is extracted. |
60c744babb8c1a21303da553 | 3 | The subset from the AstraZeneca ELN data accounts for 1.5 % of unique templates. Additionally, we observe that there is a greater overlap with Reaxys than the Patent data. These do not necessarily correspond to novel reactions, but rather are an artefact of the structural diversity present in the AstraZeneca collection. For instance, the synthesis of a novel lead compound could have different atomic environments around the reaction center compared to the literature or patent precedent on which it was based, thus leading to a new reaction template. Similarly, 2 % of all templates are common between the datasets, thus there is a small degree of structural overlap as might be expected. |
60c744babb8c1a21303da553 | 4 | In previous studies, accuracy has been used as a metric to gauge the network's performance for the task of retrosynthetic planning. The accuracy of the filter network reflects its ability to correctly predict a reaction template. However, for the task of retrosynthetic planning the aim is to predict several applicable templates, not just the one recorded in the dataset. Given the underlying data describes a one to one mapping of product to template and the task is to predict a one product to many templates' relationship. High accuracy values are associated with the model's ability to predict the template or reaction center from which it was originally extracted, thus overfitting the data by creating a like for like mapping to the underlying dataset. Additionally, the accuracy does not account for the applicability of the predicted template, for which we and others have found high failure rates owing to an inability to match the template substructure to the target for which it was predicted. This is illustrated in Figure , whereby the increased specification of the molecular environment surrounding the reaction center (radius) leads to a higher rate of failure for its application, and translates to decreased model performance. In contrast, the test accuracy does not highlight the extent of the performance decrease, but rather increases as more of the environment surrounding the reaction center is considered, thus is misleading. We propose that in conjunction with the accuracy, the more task-specific specific measure of the number of applicable templates be used for policy assessment, and a more holistic view be taken of overall model performance. In all datasets examined, on average less than 1 % of all templates were applicable for any given compound. Whereby, only ca. 0.00035 % of all templates were applicable and in the top 50 templates prioritized by the network for any given compound. Increasing template specificity further reduces the number of templates that can be applied in a given context. Therefore, to balance specificity with generalizability we propose that templates considering the reaction center and the first degree nearest neighbors be used, in conjunction with the specification of a variety of functional and protecting groups, to maintain chemical integrity. |
60c744babb8c1a21303da553 | 5 | Templates were extracted using an adaptation of Coley et al.'s implementation for rule extraction, which only contain the immediate neighborhood of the reaction centers, thus do not capture the extended environment required to account for leaving and protecting groups. In addition, the algorithm failed to account for reactive species, without specification of which, the reactants would not be regenerated. This has since been corrected by Coley et al. in RDChiral and has been extended in this study to encompass ca. 75 functional and protecting groups commonly used in organic synthesis. These were determined by analysis of frequently used reactions in the underlying datasets. |
60c744babb8c1a21303da553 | 6 | We found that half of the top 10 templates across all datasets, and 25 % of the Pistachio dataset accounted for protections and deprotections. This value is similar across all datasets examined in this study and demonstrates the utility of protecting group strategies in organic synthesis. Furthermore, we determined that these improvements translate into the model being able to account for the extended molecular environment for the groups specified. However, whilst the model can employ protections and deprotections, their use is not necessarily strategic. Further work is required to allow the model to learn their most appropriate use and incorporate them for maximal effect into synthetic route planning. The model is also limited in that in cannot learn the form of new protecting and functional groups from additional data and is restricted to those specified. |
60c744babb8c1a21303da553 | 7 | Figure shows the top-1 accuracy computed for the hold out test set for a range of library sizes using templates obtained from the USPTO dataset, as compared to the ability to predict full synthetic routes to 1,731 compounds in a series of 41 virtual libraries designed at AstraZeneca. We observed that the accuracy decreases with increasing template library size, where the size of the template library reflects the top N templates in the USPTO dataset. In comparison the average predictive ability of the model increases, reflecting a more task specific measure of model performance. Of note is the increasing difference between the accuracy and overall predictive performance as the library size increases. |
60c744babb8c1a21303da553 | 8 | Whilst the test accuracies have been measured for a baseline template-based CASP tool, templatefree models are also prone to misleading accuracy values. In both cases the task is to predict a series of viable outcomes, however the accuracy reflects the ability to predict the 'ground truth' from the underlying dataset, which inherently accounts for only one 'true' value, thus is partially known. In a similar work, Segler and Waller used the top 1, 10 and 50 accuracies to gauge the performance of their network, and showed that a model trained on 17,134 rules extracted from Reaxys, covering 52 % of the dataset, was able to predict the reaction center with accuracies of 50.1 %, 89.1 %, and 96 % respectively. In an extension of the work considering only single step reactions Baylon et. al. reported an accuracy of 81 % on 129 rules compared to 83 % on 137 rules by Segler and Waller. However, we have found that accuracy can be misleading when used for the assessment of overall model performance as shown in Figure and Figure , and specifically for the assessment of whether the network is able to correctly predict applicable reaction templates for single step reactions. |
60c744babb8c1a21303da553 | 9 | The virtual library set can be further broken down into libraries designed using a 'combinatorial' approach, and a broader set of reactions using more 'bespoke' chemistry, which covers the reaction space more extensively. This enabled consideration of domain dependency with respect to template library size. We found that virtual libraries designed using a combinatorial approach benefited marginally from increasing the template library size. With the 1,064 most frequently occurring templates in the USPTO dataset, routes could be found for 65 % of the compounds in the virtual libraries designed using a combinatorial approach. This increased to a maximum of 72 % when the 25,126 most frequently occurring templates were used. This is in line with what would be expected, as combinatorial libraries employ frequently used and robust reactions in their design. |
60c744babb8c1a21303da553 | 10 | Using the 1,064 most frequently occurring templates in the USPTO dataset, the model predicted synthetic routes to 50 % of the compounds in the 'bespoke' library, increasing by 19 % to a maximal value of 69 % when using 77,281 reaction templates. This alludes to the point that increasing the number of templates increases the chemical diversity of the templates, thus more synthetic routes can be found than with smaller template library sets. The increase in diversity of the templates originates from the fact that no two templates are the same, as they account for different sub-structural patterns. |
60c744babb8c1a21303da553 | 11 | Increasing the template library size, also increases the probability of finding a sub-structural match to the product to which the template is applied. On the other hand, the 'combinatorial' libraries are less diverse, arising from the fact that a limited number of reactions were used to make them. Therefore, templates matching sub-structural patterns occurring within 'combinatorial libraries' are also limited. |
60c744babb8c1a21303da553 | 12 | There is a balance between the number of reaction templates and the reaction space they represent, which is specific to the domain in which the tool is applied. However, increasing the number of reaction templates also introduces noise and does not necessarily mean that there is an increase in the diversity of templates, nor coverage of the reaction space. This can be seen in Figure , where the overall predictive performance falls by 4 % and 6 % for the 'combinatorial' and 'bespoke' libraries respectively, when increasing the template library size from 77,281 to 285,018 reaction templates. the test set, as compared to the overall performance. The overall performance is with respect to the ability to predict full synthetic routes to a set of 1,731 compounds from 41 virtual libraries designed at AstraZeneca. The experimental average refers to the percentage of compounds synthesized out of those sent for synthesis after refinement of the virtual library. The accuracy decreases with increasing template library size, whereas the overall predictive performance increases up to a library size of the 77,281 most frequently occurring reactions. b) The virtual library set can be further broken down into libraries designed using a 'combinatorial' approach, and a broader set of reactions using more 'bespoke' chemistry. The overall model performance increases marginally for the 'combinatorial' libraries with increasing template library size. Whereas, the libraries requiring more 'bespoke' chemistry for their synthesis benefit from the inclusion of additional reactions. |
60c744babb8c1a21303da553 | 13 | Whereas, the number of compounds that could be synthesized for the 'bespoke' library was consistently under-predicted. This highlights that only considering the number of compounds for which routes can be predicted does not afford enough granularity for the assessment of synthetic routes, and CASP tools. Furthermore, the reasons for a 'failed' synthesis are not always known and can be dependent on the nature of the project, the skill of the chemist, and the conditions used, to name a few factors influencing the outcome of a synthesis. These factors cannot always be quantified or considered qualitatively, thus both the predictions and 'true' experimental results have an associated degree of uncertainty which proves difficult to measure. |
60c744babb8c1a21303da553 | 14 | We compared the predictive performance of models trained on each reaction dataset, and combinations thereof, on 1,731 compounds from 41 virtual libraries at AstraZeneca and the top 125 small molecule therapies of 2018 (Figure ). The models, regardless of reaction dataset, consistently over-estimate the number of compounds that can be synthesized in the case of the virtual libraries, and under-estimate with regards to the top 125 small molecule therapies. For both cases, the average number of steps taken to synthesize a molecule is 4, however the average time taken to solve each molecule varies considerably with the dataset size (Figure ). The smaller datasets are faster at finding routes to a given compound (< 4 seconds) owing to a smaller search space in comparison to the larger search spaces associated with the larger datasets (Pistachio and Reaxys). The simple architecture used is not able to handle the large search space and is biased towards frequently occurring reactions, which are augmented by the additional data in the larger sets. In the case of the top pharmaceutical compounds, the lower predictive performance may arise from more sophisticated ring systems, and natural product like structures upon which the final compound is based. Reactions of this nature are not prioritized by the network as they are infrequent, thus become difficult to separate from the noise. |
60c744babb8c1a21303da553 | 15 | The average number of successfully applied templates of the top 50 predicted templates for one-step synthesis per compound varies considerably across the reaction datasets examined (Figure ). The model built on a subset of the AstraZeneca ELN appears to be worse than the models built on other reaction datasets by this measure. However, we have found that the number of options the network suggests for one-step synthesis does not impact overall model performance in this case. Thus, as Segler and Waller suggested in a previous study examining training set size, models competitive with those built on larger reaction sets can be obtained with datasets as small as an internal ELN. The subset of the AstraZeneca ELN accounts for 4.5 % of the template library obtained from a combination of all datasets examined, yet is capable of providing sufficient training data to train filter networks and resulting models which are competitive with those of larger proprietary datasets. However, we expect that this is domain specific and reflects that the subet of the AstraZeneca ELN is tailored to the medicinal chemistry domain in comparison to the patent and Reaxys datasets, which are more extensive in their converage (Figure ). This further demonstrates that there is a balance between the type of chemistry covered by the template library set, and the size of the template library. An optimal set would be domain specific, and cover enough examples of sufficient diversity, that the output space would be managable by the filter network. In the current approach we have found that as the dataset size increases, so does the output space of the filter network (Table ). This increases the time taken to train the network, and makes it increasingly difficult for the network to prioritize appropriate reactions as seen when increasing template library size in Figure . |
60c744babb8c1a21303da553 | 16 | Previous studies have demonstrated that models built on the USPTO dataset, can predict one-step synthesis. We show that despite the seemingly lower amount of data in the USPTO dataset compared to Reaxys (Table ). The USPTO dataset accounts for 44.8 % of the template library obtained from a combination of all datasets examined, in comparison to 53.7 % which comes from Reaxys. Whilst there is a 8.9 % difference and the coverage of the reaction space that the templates encode varies (Figure ), this does not appear to be a limiting factor for route prediction in the medicinal chemistry domain. |
60c744babb8c1a21303da553 | 17 | Figure shows that the model trained on Reaxys marginally outperforms that trained on the USPTO dataset, at the expense of longer prediction times. Furthermore, we show that as the size of the dataset increases to a combination of both Reaxys and the combined patents data (USPTO and Pistachio), the overall performance of the model decreases with regards to both time and number of routes identified. |
60c744babb8c1a21303da553 | 18 | We noted that the fingerprint size used to encode the product had a marginal effect on the ability of the model to predict full synthetic routes for the internal virtual library dataset (Supplementary). In addition, we found that increasing the size of the stock library to include the ACD catalogue, increased the ability of the model to predict full synthetic routes to compounds in the virtual library. For both the 'Combinatorial' and 'Bespoke' libraries, the model was able to reduce the average time taken to predict full synthetic routes with the ACD catalogue, as well as reduce the average number of steps by one. The reduction in the average number of steps is more pronounced for the 'Bespoke' libraries, whereby it is consistent over both the USPTO and Reaxys datasets. This is in comparison to the 'Combinatorial' libraries whereby the reduction in the number of steps is not observed for the combined Reaxys and patent data (Supplementary). |
60c744babb8c1a21303da553 | 19 | Comparison to existing literature in the domain showed that the model trained solely on the USPTO dataset was competitive with that reported in the literature (Figure ), and was able to find a route to the target compound in 4.26 seconds. This was also observed for models trained on the subset of the AZ ELN, Pistachio and Reaxys datasets. We found that the model was able to suggest an alternative route in addition to that reported, involving a ring formation (Figure ). Furthermore, we show that the model can predict routes to the top 125 pharmaceutical products, where the performance is dependent on the stock set of compounds. Examples of which have been given in the supplementary information. The route predicted using the model trained on the USPTO dataset to Amenamevir is compared to the literature route. Both routes vary in the order of the steps they take, with the predicted route preferring a standard amide coupling over the amide Schotten-Baumann. The model trained on the USPTO dataset, finds an alternative route to that in the previous study, and finds synthetic routes to the target compound in 4.26 seconds. The model can prioritize and apply ring formations as demonstrated in Step 4. b) Comparison of the route found by the model trained on the USPTO dataset with the literature route for Amenamevir. The model can suggest a route comparable to the literature, differing in the sequence of steps and using similar reactions to those in the literature. The predicted route is found in 3.26 seconds. |
60c744babb8c1a21303da553 | 20 | We have developed and implemented a baseline retrosynthetic tool with only a single neural network, to investigate the role of the ML template prioritization method in the tree search algorithm derived from the work of Segler and Waller. We have found that models trained on datasets as small as the internal ELN (4.8 % of all templates) and USPTO datasets (44.8 % of all templates), are sufficient for the prediction of synthetic routes to compounds found in medicinal chemistry pipelines. Furthermore, we demonstrated the potential use for such tools in compound selection and prioritization in DMTA cycles and suggest that datasets with known experimental results can be used to assess model performance. |
60c744babb8c1a21303da553 | 21 | In addition, we demonstrate that accuracy can be a misleading measure for the performance of the filter network and final tree-search model. Thus, we propose an alternative approach to assessing the ability of the filter network, based on the number of templates that can be successfully applied in the top N predictions, for a given context. We demonstrate that the specificity and generalizability of the extracted templates must be balanced such that, the first degree nearest neighbors to the reaction center, are used in conjunction with the specification of functional and protecting groups that are common in organic chemistry. |
60c744babb8c1a21303da553 | 22 | We have found there is a dependence between the size and content of the template library used, and the domain in which it is applied. We found that syntheses of compounds originating from combinatorial libraries could be predicted using the most frequently occurring reactions. In contrast, compounds originating from libraries requiring more complex syntheses, required an expanded template set for their successful prediction. Further work is required to make use of the broad selection of reactions available to improve the variety and complexity of routes suggested. Further investigations into the template extraction process are also required to determine their descriptive limits and how this translates into route prediction. |
60c744babb8c1a21303da553 | 23 | Of the datasets used, only the United States Patent Office extracts (USPTO) ranging from the years 1976 to 2016 is publicly available. This is split into granted and applied patents and is openly available for use by the community. A subset of the AstraZeneca Electronic Notebooks (ELN) were mined (May 2019) to yield the internal proprietary dataset, considering only positive reactions, classified as those with a yield greater than 1 % and having a conclusion statement. The Pistachio (2017-11-17) and Reaxys datasets are commercially available, provided by NextMove software and Elsevier respectively under licensing agreements. The Reaxys dataset was filtered for multi-step reactions to yield only the intermediate single step records for which templates were extracted. Full details of the number of reactions and unique extracted templates can be found in Table . |
60c744babb8c1a21303da553 | 24 | All reactions were atom-mapped and classified using the commercially available Filbert and HazELNut packages (v. 3.1.8) provided by NextMove software. These were subsequently processed using RDKit and RDChiral for template extraction, in conjunction with a custom reaction class developed by the authors to facilitate reaction processing. The reactions are parsed as reaction SMILES, along with the ID linking back to the data source, and classification code or textual classification obtained from the NameRxn software. The reaction SMILES are of the form: |
60c744babb8c1a21303da553 | 25 | Where the reactants, agents, and products are separated by '>' and the individual non-covalently bound species represented by a '.' according to the Daylight SMILES specification. The definition of reactant and agent is ambiguous, as agents may participate in the reaction and contribute mass to the products. Additionally, as the templates are extracted based on atom-mapping, only the species contributing to the product or changing during the reaction were considered in the process. Thus, we have moved all agents into the reactants to give a reaction SMILES of the form: |
60c744babb8c1a21303da553 | 26 | Reactions leading to more than one product, incomplete reactions (i.e. missing reactants or products), or reactions in which the reactants and product were equivalent were removed. Equivalence was determined by converting the reactants and products to InChI and comparing. Permutations in the ordering of reactants and products were accounted for, however this was not significant in this case as we only account for reactions with one product. |
60c744babb8c1a21303da553 | 27 | Reaction templates were extracted as SMIRKS patterns using RDChiral, which we modified to consider an additional ca. 70 commonly occurring functional and protecting groups as determined by an analysis of the underlying datasets and extended to commonly used protecting groups in the wider literature. These are automatically identified through a substructure search of the encoded protecting groups and included in the templates alongside the reaction center and first degree nearest neighbor atoms. The reaction center is defined as atoms and bonds that change during the reaction. |
60c744babb8c1a21303da553 | 28 | Owing to the number of variations sharing the same core structure for some protecting groups i.e. silyl ethers, esters, but varying in alkyl chain length, we have refrained from an exhaustive encoding of all possible protective groups. Rather, we have focused on those we found to be commonly occurring in the dataset and cover the main form of the protecting group, leaving the decision of the exact form to the chemist. |
60c744babb8c1a21303da553 | 29 | The extracted templates were parsed and checked for validity in RDKit, 37 following which the template was applied to the product of the reaction from which it was extracted to determine if an outcome could be generated. The outcomes were assessed using the definitions shown in Figure , and the quality of the template extraction process quantified. Selective) The template generates the reactants from the reaction from which the template was extracted in addition to other possible precursors that are not part of the original reaction. Unselective) The template generates reactants that do not correspond to any of the reactants in the reaction from which the template was extracted. These may or may not be viable reactants. |
60c744babb8c1a21303da553 | 30 | The reactions and resulting templates were hashed individually following a hashing scheme developed by the authors inspired by the reaction InChI (Figure ). This was also used to identify duplicate reactions and templates and can be used as an identifier for database lookups. The reactants and products are converted to a RDKit mol objects in without separation of the individual species. Conversion to InChI for the reactants and products respectively is carried out in RDKit. This is order invariant and overcomes the issue of having multiple SMILES representing the same molecular structure. The resulting InChIs are concatenated and hashed. |
60c744babb8c1a21303da553 | 31 | The datasets used in this study and their respective sizes, given as the raw dataset size without filtering are shown in Table . To our knowledge, the combined dataset is the largest reported to date. To enable clarity in the task specific curation process, the reduction in size through extraction and validation, followed by duplicate removal has been shown. Extraction refers to the extraction of reaction templates from the reaction SMILES, and validation refers to the application of the extracted template to the product of the reaction from which it was extracted, to determine if the corresponding reactants can be generated. Duplicates were identified as reaction SMILES consisting of identical reactants, agents, and products, using an order invariant hashing scheme accounting for variance in atom-mapping as developed by the authors. Unique reaction templates were also identified in the same manner. The overlap of reaction templates extracted from the respective datasets was ascertained by using the in-built set methods in Python. We have observed that some of the noise associated with automatic template extraction originates from incorrect mapping, text-mining errors, and human-error from manual curation. There are several variations of these cases including, incorrect recording of functional groups, incorrect mapping of reactive components (i.e. substructures present in the reactive center may also be present in the solvent or reagents, for instance the incorrect mapping of an amine in both the reactant and base), accidental extension of alkyl chains, representation of catalysts and incomplete reactions, examples of which can be found in the supplementary information. Whilst our approach to curation can identify such inconsistencies and disregard their associated reactions, further efforts are required to improve catalyst representation, text-mining, template SMIRKS generation and atommapping. |
60c744babb8c1a21303da553 | 32 | Template libraries were constructed by filtering the respective dataset for templates that occurred a minimum of N times. In all cases duplicate reactions were removed prior to filtering. Products were represented as extended connectivity fingerprints (ECFP) with a radius of 2, using the Morgan algorithm in RDKit. Whereas, templates were represented as binarized labels in a one-vs-all fashion using the scikit-learn library using the 'LabelBinarizer'. Both the input ECFP4 and output vectors were precomputed. Training, validation, and test sets were constructed as a random 90/5/5 split of the datasets, using a random state of 42, where the datasets were shuffled prior to splitting. This was conducted using the scikit-learn library. The policy networks framed as supervised multiclass classification problems were trained using Keras with Tensorflow 48 as the backend, the Adam optimizer with an initial learning rate of 0.001, and categorical cross entropy as the loss function. The learning rate was decayed on plateau by a factor of 0.5, where the plateau was considered as no improvement of the validation loss after 5 epochs. The |
60c744babb8c1a21303da553 | 33 | A random subset of 200 and 20,000 compounds from ChEMBL (v. 24.1) were used to assess the baseline number of applicable templates and the applicability of the top N templates respectively, unless otherwise stated. Salts were removed from the ChEMBL dataset using RDKit. Random subsets were drawn from the resulting dataset using a random state of 1. |
60c744babb8c1a21303da553 | 34 | The model to be assessed was loaded into Keras and the compounds to be queried converted into ECFP4 fingerprints prior to passing to the model for prediction. The top N predictions sorted in order of decreasing probability were used for each compound. The templates were applied to the compound in turn using RDChiral to determine if an outcome was generated. Templates leading to an outcome were classed as successful. |
60c744babb8c1a21303da553 | 35 | The tree search was implemented as a simplification of the algorithm described by Segler et. al. The MCTS algorithm was simplified with regards to the policy (or filter) network. The same network was used for both the expansion and the roll-out. The prior probabilities were not used by default during the selection of leaf nodes for expansion, but the Q value was initialized at 0.5 and N at 1, as expansion counts as a first visit. |
60c744babb8c1a21303da553 | 36 | The search tree is built up from nodes that contain states with current molecules of the route. The root node contains one molecule, which is the target molecule of the algorithm. Other nodes can contain states with one or more molecules. Each node is bound to others in a directed way as parent-child nodes, with actions as edges. The action is the retrosynthetic reaction performed on one of the molecules of the parent state, to yield the molecules of the child node state. The search algorithm starts with the expansion of the root node (see below). |
60c744babb8c1a21303da553 | 37 | In each iteration the search tree is traversed using the upper confidence bound (UCB) scores of the nodes (eq1). Starting from the root node, the UCB scores of the children are calculated. Here Q is the current sum of previous rewards. N is the number of times the child state has been visited, N-1 is the number of times the parent state has been visited. C is a tunable parameter balancing exploitation and exploration which was set to 1.4 by default. If the selected child is already expanded (i.e. has child nodes), the UCB scores of these are then calculated and the next child selected in an iterative way until an unexpanded leaf node is selected. Actions are stored at the parent level, and the child nodes are first instantiated as node objects by applying the associated action when visited (see below). |
60c744babb8c1a21303da553 | 38 | Expansion is performed by employing the expansion filter neural network for each of the molecules present in the state of the selected node. The top scored reaction templates are filtered to retain the top 50 or until a cumulative filter network score of 0.995 is reached. The possible actions (molecule + reaction) for all molecules are stored at the parent level, and vectors of associated Q and N values initialized (0.5 and 1 respectively). |
60c744babb8c1a21303da553 | 39 | The action with the highest UCB score is selected for the roll-out. In case of multiple actions sharing the largest score, random selection is performed. The child state is instantiated and added to the search tree by employing the associated reaction template to the molecule specified in the action using RDKit. In case the reaction did not give any output, the action Q is given a value of -10 6 , effectively preventing reselection. If no actions are available, the state is marked terminal and the state evaluated with the reward function (see below). |
60c744babb8c1a21303da553 | 40 | Expansion of new child nodes during roll out is similar to the above, except the selection is done by random among the available actions. After each roll-out step the State was evaluated and the roll out stopped if either the state was solved (all compounds found in stock) or the maximum tree depth reached, or no valid actions are available. |
60c744babb8c1a21303da553 | 41 | The reward function for the final state is then calculated (eq 2) and the score back propagated through the tree, updating the Q and N values of all parent states between the final state and the root state (target compound). N is the total number of compounds in the state, Nin_stock Is the number of compounds that are in stock. |
64526d5527fccdb3ea6e13c8 | 0 | Cellulose is one of the most abundant biopolymers and can be produced from many sources, including plants and bacteria (1). Cellulose nanocrystals (CNCs) are made from cellulose by selectively degrading the cellulose-disordered and amorphous regions using strong acids and oxidizers (1). CNC exhibits high crystallinity and hydrophilicity, robust mechanical strength, excellent biodegradability, low cytotoxicity to a range of animal and human cell types (2) and favorable interactions with cell lines (3). These characteristics made CNC an ideal candidate for applications in biological systems (2), such as hosts for drug delivery (4, 5) and cell-supporting scaffolds for culturing or tissue engineering . In addition, incorporating magnetite nanoparticles (MNP) into the nanocellulose networks helps expand the applications and better realize some existing functionalities . CNC-Magnetite nanoparticle nanocomposites have been demonstrated as contrast agents for magnetic resonance imaging . MNPs were also found to promote the differentiation, proliferation, and growth in certain types of cells when present in scaffolds or loaded into cells (13 -16). Studies have demonstrated that applying an external magnetic field offers positive effects on cells (13 -16), however, even without an applied magnetic field, the incorporation of superparamagnetic magnetic nanoparticle impacts cell behaviors and the bioactivity of the scaffolds in beneficial ways . |
64526d5527fccdb3ea6e13c8 | 1 | The porosity of cell scaffolds is an important consideration for applications in both tissue engineering and drug delivery. In tissue engineering, it is critical to control cell location in the body and their functioning after they are injected or implanted into human bodies. Cell-hosting scaffolds have the important job of providing a framework to help the cells attach and proliferate, while also affecting stem cell differentiation and development into desired tissues (17). The porosity of the scaffolds, including pore sizes, pore shape, and interconnectivity strongly influence cell adhesion, cell-cell interaction, cell transmigration, and cell development and growth (17). |
64526d5527fccdb3ea6e13c8 | 2 | Nanopores in scaffolds that are <1 μm can be applied to regulate cell-surface interaction due to the interactions between nano-topography and integrins that governed cell binding (2, 17). Cell-cell communication requires slightly larger pore size (around 1-3 μm), while pores above 3 μm in sizes enable cell interactions and migrations (17). Pore sizes of 100 μm and above are favorable to promote cell proliferation and achieve the regeneration of tissues, as this size range enables the passage of more functional units necessary for tissue regeneration, although the optimal pore sizes for cell growth also depend on cell types and sizes (17). For example, the regeneration of neurons requires long and narrow pore spaces that are hundreds of microns or even millimeters in length, but only 20-70 μm in width, while microvascular epithelial and smooth muscle cells can be accommodated by the small pores of 38 μm (17). Besides pore sizes, pore interconnectivity is also an important parameter because the exchange of nutrition, oxygen, and metabolic wastes must be sufficient to improve further cell functioning . In drug delivery, the porosity level and pore size of the drug-hosting scaffolds are especially important when the system is based on cellulosic materials. Drug molecules are typically loaded onto cellulosic scaffolds through physical entrapment and binding interactions through functional groups such as -OH, accomplished by soaking . Increased porosity, which led to a higher surface area and more abundant surface binding sites, thus have determining effects on the loading and release of the drugs . The fibrillar structure of CNC also created a physical barrier to drug diffusion and therefore slows down the drug release process . |
64526d5527fccdb3ea6e13c8 | 3 | CNC is known to create self-supporting percolated networks due to hydrogen-bonding interactions between individual particles in the colloid form . This rigid percolation is different from the entangled mesh network created by polymer chains bending and wrapping around each other, as in the case of many natural or synthetic polymers . This percolation property of the CNCs has been harnessed to produce hydrogels and aerogels via freeze-thawing or freeze-casting , although the porosity of the resulting CNC networks has not been studied in detail, and no work has been done to examine the freeze casting of MNP-decorated CNC scaffolds. The experiments described herein aim to provide a reference for the freeze-casting fabrication and characterization of cryogel scaffolds using the colloidal gels of CNCs and MNP-decorated CNCs (CNC-MNP). |
64526d5527fccdb3ea6e13c8 | 4 | In freeze casting, the solidification of the solvent induces a physical separation of nanocellulose from the suspending medium. Crystals of the frozen solvent serve as sacrificing templates, allowing physical or chemical crosslinking of nanocellulose around the crystals. The frozen crystals are then removed by freeze-drying, revealing scaffolds of nanoparticle networks with an interconnected porous structure . Three freezing modes were tested: one-step low temperature freezing at -196 ℃, one-step high temperature freezing at -12 ℃, and two-step freezing by a freeze-thaw cryogelation cycle at high temperature, followed by freezing at low temperature. The two-step freezing aims to investigate how a change in freezing pattern affects pore structures. In addition, inspired by the debates on the effects of static magnetic fields (SMF) on freezing , an SMF was applied to the cryogelation step of some samples to examine possible changes to the final pore structures. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.