April 2026 Science Update

Hi everyone. This is the science update for Q1 2026. Uh just a quick note, I am coming down with something, so I'll try and edit out all the sounds of me being sick from this video, but uh honestly, I probably won't do a very good job of it, so apologies in advance. Right, so like I said, this is the science update for Q1 2026. This will cover the work done by the science team at open force field largely from January to March this year. Although I'll add in some updates from early April as well for ongoing projects. Uh here's a a quick outline of some of the projects that we have to cover. Uh as you can see there's actually quite a lot to cover today. So I'll get right into it. First up, a manuscript on the development of Ashc 2.3 has been accepted for publication. Uh so that's exciting. It probably won't be online yet uh by the advisory board meeting, but we have been assigned a DOI. I think we've talked a lot about Ashg and Sage 2.3 in the adboard meetings already. There's not much more to add, so just quickly moving on. Uh next up is the water models co-op optimization project. So uh we've been spending most of Q1 on a road map goal of co-op optimizing a three sight water model with Sage that outperforms tip 3P and OBC3. In particular uh what we're aiming for here is a co-op optimized force field that improves performance on our typical physical property targets of densities and the mixing. uh in particular uh some of the outliers that we've identified. Our plan is to start from the tip 3PFB and OPC3 water models as starting points and optimize both the vanderals and geometry uh to both pure water and mixture properties. This is part of a wider collaboration between open FF uh the charts lab including Barbara and Lee Ping uh looking at both three sight and forsight water models. As an initial step, Barbara and Chris both looked into whether swapping water models would fix some of the allies that we see in our water mix properties. So Barbara benchmarked the performance of Sage 2.3 with a number of different threesight and foresight water models. Um and found that performance improved with Sage 2.3 if if we use better ones. So for example, uh tip 3P FB or OBC3 in place of uh tip 3P itself. And this is despite the stage 2.3 parameters being fit to tip 3P. Chris took it a bit further and looked to see if he could improve performance even further by refitting the bandw parameters uh to OPC3 water directly. Surprisingly in terms of B metrics, this actually didn't improve um overall compared to just using the existing stage 2.3 parameters and just swapping in OBC3 as the water model. That being said, we did improve of some of on some of the aiming water outliers uh down here, which is great to see. At this point, we also started stepping back and taking a closer look at some of these outlier points uh in more detail. And we identified a few further issues. So, for example, looking at one cluster of outliers up here, uh we realized that we had been simulating this lact form with water, which is not the stable preferred torma. uh instead here we should really have been using the lactam form. We eventually actually trace this back to a bug in evaluator itself which was inferring to automatic forms using the ini um instead of preferentially or uh incorporating the use of the pack name. Uh but more on infrastructure updates later. A second consideration was whether we may be ignoring meaningful protonation of amines which are a group that perform quite poorly with water uh with sage. We implemented a new check that such as Kemble or tries to predict pKa using the PK learn predictor and excludes amines with pKa over 11 where we might expect that a reasonably high proportion of the amines are proteinated in each with water and therefore not well represented by the approach that evaluator takes in uh estimating the enthaly of mixing using the same chem graphs over uh each component. And lastly as part of a general survey of the ML we identified and excluded some measurements. Um so for example uh these points that that are more than 10 sigma away from other uh apparently identical measurements um in the thermmo database. These actually haven't affected fit so far but better safe than sorry. So we really did kind of a general re-review of the data available in the public sement ML and we found that the current coverage of co solvents and chemical environments that we're interested in is fairly patchy overall. Uh so this figure is heat map of the co occurrence of uh chemical environments in binary mixtures. Um and on the top and the right hand side here we've labeled u not the specific functional groups but the group that they lie in. So for example different kinds of ethers, different kinds of alcohols and so on. Um and you can see that there's quite a lot of blank space here. Uh the heat map represents the number of data points available for that mixture in a general broad sampled set. um and gray represents a lack of coverage. So, we're also in talks with NIST to see if we can collaborate on another public release of data to augment our training and fitting data set uh and and expand our coverage of chemical space to fill in some of these gaps. This is still ongoing, so it's likely that our three site model will be restricted to public data only, but will be really beneficial to future fits. Uh also as part of this effort, the science team has been working with both the infrastructure team and leaping to get the uh necessary software stack back in working shape. Prim primarily this focused around updates to open FF evaluator and force balance where we needed to get virtual site refitting up and running again uh in tandem with each other. We've also fixed some other bugs to do with fitting in evaluator as of version uh 0.5.3 including some issues that touched other projects around open FF such as the lipid force field. We put in a fix for inferent automatic forms as mentioned previously and we're also making u evaluator a bit easier to use for the water optimization project in force balance. We've also put in a few updates that we'll hopefully get released shortly. Uh so this includes updates to the virtual site refitting modernizing to work with the um modern open FF toolkit including Naga charges as well as some unrelated updates to the GMAX and PIM and buy interfaces. And lastly, interchange is largely the domain of the infrastructure team. But I thought I would just mention here that we did find and fix um a virtual site geometry bug. So if you're using virtual sites with open FF, please do use open FF interchange version 0.5.1 and above. Uh so we generally at a point where we're about to start with a site MVP where we'll go parameters to existing public ser ML data. As part of this, Michael led a grant for uh getting us some compute time on the Oakidge frontier cluster. So we're in the process of setting up both FFitting experiments and refining technical workflows to prepare for a larger grant application down the line. And lastly, this has just started, but Chris is also looking more closely at alcohol and water mixtures. Uh, as you can see here, this subset also performs particularly poorly in both densities and enthalpies of mixing. The direction that we're currently investigating is whether rescaling alcohol charters will improve enthalpies of mixing. Uh, this is still in progress, so I won't talk too much about it, but you'll probably hear more about it next quarter. On a very related note, we have continued with our remark goal of moving away from force valance by constructing a new library for co-optimizing veance and vanderal parameters with smear and descent. While the call group have been successfully using sme and descent for optimizing veance parameters, fitting parameters to simulation data such as physical properties requires more management of uh distributed compute and we have determined that putting in this new library would be easier to maintain than upgrading the evaluator library. We're also taking this opportunity to add uh some quality of life improvements and support for anticipated fitting needs. While this currently falls under the umbrella of scientific code, progress here is actually being driven by Matt Thompson from the infrastructure team. In terms of other uh updates to SME and descent, which currently are also regarded as scientific code, we have put out a new release version 0.17 and are working on updates to descent. SMEI version 17 or 0.17 uh corrects a few nasty bugs affecting fits in particular with periodic systems on CUDA and also removes the current dependence on amber tools. In descent we're aiming to add a few features that will also be useful to periodic fits and remove the uh amber tools dependence as well. This progress is mostly work that is being driven by Javal and Finlay in the collab. uh this ties in well towards progress on our our next major effort which is refitting a stage 2.4 force field using smear and descent. So this covers a number of experiments where the overall goal is to work out a good fitting workflow for stage 2.4. Unlike our force balance workflow where we fit to optimize geometries and torsion drives, smear descent currently only support fits to energies and forces of individual snapshots. Uh therefore there's still quite a lot of open questions uh on which data is best used with being descent whether we should use the existing spice 2 data set of of equilibrium snapshots or whether we should use our existing data set uh corpus of equilibrium data as Jennifer Finlay's comparison of per per atom force magnitudes uh these data sets do differ quite a bit we ran a few experiments here that I won't go into but the overall takeaway is that trying to use only our existing data set of of optimized data without off equilibrium data results in force fields where the bond and angle force constants trend continuously towards zero. This also leads to drift in the equilibrium bond and angle values and the overall result is much worse performance on benchmarks like relative uh confirmer energies. Uh in parallel Jennifer Clark has also been experimenting with fits just as spice to in particular investigating whether finite partitioning of chemical space improves performance. Jen has also run a lot of experiments here that I also won't get into this time. They're not quite ready yet to talk about. But the overall pattern again is that the force constants here seem to trend towards zero over the course of fit. And again, the equilibrium values can drift quite a bit. Our overall hypothesis from these results and from other results that Finlay and Danny have shown is that the optimal training data set would include a mix of equilibrium and of equil equilibrium geometries. To test this hypothesis, we want to run an experiment fitting to a data set combining spice and optimized geometry data. So to that end, Jen is applying a library that she developed for the Genentech collaboration for robust chemical perception to the OML 25 data set. We're choosing to use OML 25 instead of combining the existing spice and goosey archive data sets to avoid any possible issues that might arise from using different levels of theory. OML 25 recmp computed a number of existing public data sets including spice and helpfully they also took 30% of the GM data set and geometry optimized the subset which provides us equilibrium data. It took a non-trivial amount of work for us to convert this data. So Jen is uploading our results to hugging face as a complete public use and you can find uh you can find these data sets here at these links. Uh all righty on to the protein force field to recap the goals outlined for this project. We're aiming for co-op optimized parameters for both proteins and small molecules that achieve similar accuracy to existing protein force fields. And we're using amber force fields, most specifically amber 14 SP as our reference. As a reminder, we are evaluating the the performance of each protein forcefree candidate on several tiers of benchmarks. The first is uh a rapid evaluation of QM data and smaller structure peptides. The second tier focuses on maintaining protein structure of folder proteins in simulation as well as maintaining the hicity of a fifa. And of note, this tier includes the GB3 protein as well as head and egg white lysosyme just because I will be mentioning those later. And the third tier of benchmarks includes a wider variety of protein structures as well as binding free energy targets. And I won't get into the fourth tier because we're not really there yet. And uh as a brief recap of the history of the project and where we were at the start of the year, Chapen has done many different experiments fitting SMOF candidates to various data. Typically the trend that we've seen with our candidate force fields which descend from the stage 2.1 small molecule force field is that they do better than amber on QM small molecule and unstructed peptide benchmarks. However, that doesn't translate to good performance on larger folder proteins and for all fits shape and tried where the force field was just trained to small molecule and peptide QM data. Uh we would see secondary structure unfolding with the folded protein benchmarks. The rosemary alpha force we had released last year extended the small molecule and peptide QM data set to add GB3 and the helico 15 as training targets and this did result in good performance on these targets across the entire tier 2 benchmark suite which includes four folder proteins. The rosemary alpha performed within range of amber 14 SB on three of them. The one exception was for hen and lysosyme where the performance was worse for rosemary alpha than for 14b. So we also wanted to benchmark on more diverse structures to characterize whether this was indicative of rosemary alpha performing poorly or whether hen egg is just a difficult target to hit. So at the start of Q1 we had just finished running additional tier three folded protein benchmarks although we had yet to look at the results. And lastly uh last year we also benchmarked rosemary alpha on our standard small molecule benchmarks and we discovered that that it introduced some issues with sulfur and phosphorus parameters. So at the start of Q1 2026, we were also beginning to think about what the release process would look like of taking rosemary alpha to a full force release and whether we could fix these small molecule issues in a refit without ruining performance on proteins. Okay, on to some of the work that we did in Q1. Firstly, the tier 3 benchmark results. This is not the complete tier three set, but the results that Chapen looked at seem to indicate that the rosemary alpha in general has problems with keeping folded proteins stable. out of the eight proteins that he looked at, three of them deviated more from the NMR model uh than Amber, four were similar and only one was better. So that's still a substantial proportion of the data set where the rosemary alpha is not doing well, suggesting that it might be more of a general false problem. Uh on the right hand side here is an example of one of the aggregated RMSD plots that J showed for the proteins. So instead of show showing a typical time series trace per replicate, this plot combines data from three different replicates and plots a cumulative distribution function instead where a force field having a line more to the left um indicates closer agreement to the NMR structure and more to the right in uh indicates worse agreement. Since we saw such improvement from incorporating GB3 and the helical 15 into our fitting targets last time, Chapen then investigated various avenues of adding hen egg white lysosyme data into the training set as well. Uh but neither way that he investigated offer clear directions forward. So some experiments for example fitting to NOE distances uh resulted in force fields that maintain good performance on GB3 and the helical feet 15 month but using the same iterative approach that worked well for for the GB3 NMR fitting did not correspondingly improve performance on head equating to hen egg white lysosine PMFs also resulted in force fields with somewhat patch performance for example improving performance on J couplings but degrading performance for residual dipole couplings and NOA targets even for hen of white lysosyme itself. Chapen also recently took a closer look at PMFs over fine s for caps reus. Interestingly, this revealed that this man force fields consistently have higher energies near the alpha basin uh than the cure profile and the amber force fields here in in blue and purple reproduce this the shape of the profile much better. Chapen tried a couple of things to see if they would shift the PMF and the profile is improved if he um if he swaps in ambber and non-bonded parameters. uh that's a brown line here although not to the extent of reaching the amber force fields the other approach was adding cap streamers into the training data set itself which resulted in the pink force field null streamer QM disturbingly this didn't seem to shift the profile at all uh resulting in questions about how we could improve our workflow this is going to come up later and we're still starting investigations so I won't get into that much here but here is the state of things at the end of Q1 on a similar topic though we have been exploring the general question of QM fits with rosemary uh with our small molecular benchmarks we noticed that the rosemary alpha has some problems with uh sulfur and phosphorus parameters resulting in geometry optimizations with very odd angles the main question we want to answer is if we can refit rosemary alpha to QM data to fix this behavior without degrading protein image marks so ran uh a couple of experiments where he refit the rosemary alpha to QM data in two ways one where the torch were all free to move and one where the protein specific torions were frozen. Both fits unfortunately resulted in worse performance on tier 2 benchmarks and would probably require another round of NMR fitting to get back to rosemary alpha level performance. So you can see these two fits in olive green and red here and you can see that the the uh kai squared value is much higher than the rosemary alpha value in orange. The next experiment we're planning to try next quarter is to refit with tight price across all torsions, not just the protein specific ones since some of the general torsion parameters in the rosemary force field are also involved in secondary structure. Last but not least, I just wanted to introduce another effort that has started up in collaboration with the infrastructure team with Ashley, which is starting peptide benchmarks with smart candidates. One of the big advantages that Rosemary will bring is being easily applicable to non-standard peptides, including cyclic peptides. So Ashley is starting to set up a small benchmark set of peptides to compare performance with other peptide and protein force fields on these targets. Um all righty onto another ongoing project in the consortium which is improving lipid parameters with the goal of reproducing membrane properties in simulation. This is being led by Julian in the SH lab and she's starting off with phosphoippids which has several key parts. The head group with a phospho group, a glycerol linker and hydrocarbon tails. So as a sub goal, this project aims to improve alkan, alkan and phosphate parameters. A targeted approach has worked well for alkane and alkan parameters. Julian was able to optimize the force fields starting from stage 2.2.1 to alcan density and enthalpy of vaporization data which substantially improved performance in those areas and porting those vandal parameters to stage 2.3 retained performance. One interesting outcome from this experiment was that the force field that performs best on alkan torsion profiles relative to QM overestimates the transcal ratio in simulation more so than a force field that slightly underestimates the relative minima underscoring the need for fitting and benchmarking on condense phase data. Um however she ran into difficulties extending this approach to phosphates. Even an optimization and benchmarking on a single phosphate torsion drive became complicated as the constraint optimization that's used in torsion drives often results in a phosphate geometry that was quite different from the QM structure. We're still experimenting with alternate approaches such as converting the constrained optimization to one that uses restraints instead. Uh so we'll have updates on this in the next quarter. Uh okay so a quick update on QM data sets run by a data star Jennif Clark the stage 2.3 optimization and torsion drive fitting and benchmarking data sets are now available on QC archive and zunotto for easy download we've also added a few new data sets and additional torsion drive data set for lipids multiple data sets expanding the chemical diversity of sulfur nitrogen and phosphorus groups and uh a data set reoptimizing torsion net 500 at the openf level of theory The first two are largely for improving for bits and the last is for use as a benchmark set for the in progress section. We're really excited that the architected data set for diverse organometallic uh lian complex structures is finally up and running. Uh this was the result of a huge amount of work from Jen and would really expand coverage for the genetic collaboration. She has also recently folded a subset of problematic geometries from Spice which are largely rare chemistries that Sage doesn't perform well on and is running geometry optimizations on these as additional fitting data. While we're on the topic of collaborations, Jen has also been working closely with Open Fe to collaborate on building out a more standardized benchmarks repo. This includes support for both protein lean binding for energies and absolute solvation for energies. uh includes uh both a set of standardized systems as well as code for discovering these stores a standardized set of partial chargers uh in the repo and contains examples of building or using previously generated networks. We're really excited that this is nearing completion um as this repo and this project and collaboration will make it much easier for open FF to run our own benchmarks for future for candidates. And lastly onto a few odds and ends. One item that Jen uncovered during her collaboration with the open FE benchmarks reburn is that our current scripts actually interpret the molecules from the Minnesota salvation set incorrectly. Um in particular during the conversion step where we infer the chemical structure from the molecule name. Some of this is because the molecule name itself is underspecified. It doesn't specify uh which is to use and we wind up using a different one than is in the paper. In other cases though, um it's likely due to an underlying issue with how we're passing back names. It turns out this is very sensitive to any modifications to the input. Uh so for example, adding a space to fend formate results in a different structure and actually the correct structure here. Uh we did a preliminary check and none of the incorrectly passed molecules were present in our stage benchmark sets and we have now identified and removed them for future benchmarks. While we're on the topic of bugs, um a somewhat Sonia bug is uh one that we've recently discovered impacting our calculation of A1 BC charges and this is specifically using open eye software. This bug affects atom types that should get assigned um atom type uh 23 uh which is a highly deoized lone pair nitrogen and the cononical example given is the nitrogen in indul. With this bug, however, this type actually gets assigned type 21 instead. This error can result in very different chargers on the nitrogen such as a point4 elementary charge difference. Um, and we notified open eye about this last year. Uh, since then they've gone back to us and there should be a fix coming in the next release of the software. But this bug has now been present in uh open eye charge assignment for multiple force field fits over a number of years. We wanted to look at whether Sage had been negatively impacted by this bug and whether users should upgrade their open eye versions when the fix came out. So we did a quick analysis of the Sage 2.2.1 benchmarks and we broke down molecules by their BCC differences by open eye and ambitals. And we did this because Amptors doesn't assign the indol nitrogen incorrectly which is actually how we discovered the issue in the first place. We were looking into a systematic discrepancy between the two toolkits. We found that the ring nitrogen differences typically fell into the bin between 0.4 and 0.6 of an elementary charge difference uh which is this purple line here on this plot comparing um relative confirm energies. You can see that this purple line actually arguably performs better across DDE. It has lower differences in relative complment energy than the bins with lower charge differences in blue and red. So our main takeaway here is that saved was not negatively impacted by the bug and it's likely that the parameters had been fit to compensate for it. So when the next release of open eye is out, we actually recommend not updating to maintain consistent charge behavior with Sage. Okay. Uh thanks for holding with me to the end here. u major focuses of the second quarter will be continuing on with the projects that I've gone over uh throughout as well as some that I haven't gone into like the next generation espala work um we will of course have the annual workshop in May some of the conferences and I'll also be offline for a while so in general we're anticipating that Q2 will be a bit slower pace than Q1 but as always we'll aim for concrete progress on road mapap items such as refitting a water model uh working towards a protein and lipid force build continuing collaborations around open FF and OMSF and core projects such as constructing the dim sim package and progressing towards stage 2.4. So you can hear all about this next time. Uh thanks for listening.