Individual updates | DN Worked almost entirely on OpenFE tasks. After last week’s core-devs, I talked with MT and DM about ff-release meeting. I realized that we should address that again. DM took the lead again and at last week’s ad board meeting, he brought up the topic of publishing water models in SMIRNOFF format and it was well-received. So I told MT that we should go ahead and publish water models. MT – I’m happy with how this turned out. JW – I’m glad that this order ultimately came from DN (at DM’s suggestion), since that’s how it’s supposed to work in the org chart.
MT Past week or two, we’ve had small contributions to interchange from outside folks - some from industry in non-partnering companies and one from NASA - that have been really high-quality and encouraging. NASA person said that our FF is viable for modeling epoxies, which I don’t think we ever explicitly thought about. I’m probably due for another interchange release this week. Some moving parts with GROMACS, can’t remember all the details right now. Made Toolkit 0.11.2 release. Two or three small but important changes. None were critical but they were pretty important for quality-of-life. JM has been updating QCSubmit and Bespokefit to work with new toolkit API. Both of those are pretty much done and need a little more work. QCSubmit needs a new reelase of QCEngine, and I’ve been in contact with LBurns about this and it’s on the radar. Also some other small deps need advancing. Continuing to work with LW on benchmarking refactor. Currently working toward a more modular design that reaches parity with the previous openff-benchmark season 1 as validation DD – I’m happy to chat about openff-benchmark and the background of the project if you like. JW – Worked with MT last week to try and figure out where MM optimization actually happens. It’s a big stack of deps. DD – That helped us reach parity between QM and MM minimization settings.
DD CC Protein parameter fits for the null model (no protein-specific SMIRNOFF types) are plowing ahead (12 h to 18 h per ForceBalance iteration). Attributing this progress to: Expanding number of new parameters to address incorrect torsion assignment for delocalized charges Removing chemistries that cause ForceBalance failures ([#15:1]=[*:2]). Not sure about root cause, punting for now and we can resolve this later. Only removes ~45 opts out of ~4500 from training set
Identified protein-specific proper torsion SMIRKS that hit protein residues but no small molecules in QC training set. This was the final barrier to running fits for the protein-specific model. This week will focus on benchmarking software for peptides and proteins.
PB Last week worked a bit on a new energy levels target (based on ddE) in ForceBalance from Bill Swope's suggestions in FFR call. Made some progress on preliminary implementation, will work out some finer details like weights and other issues, meeting LPW for his feedback as well. JW – Since we’re using this for fitting, is the ddE thing continuous/differentiable? PB – There are 3 levels - Rank ordering correctly Getting the right ddE (thinking of using a switching function to keep this from getting out of control, maybe maxing penalty at a diff of 10kcal/mol or something) CC – Might be able to take inspiration from FB’s torsiondrive scoring? PB – Yes, but one issue is the matching of minima to define “0”. Some of the improvements I’m trying here may actually get propagated into changes back to torsion objective func weighting.
Getting the “same” coordinates from opt
JW – Do you think this is a promising direction relative to the other fitting changes you were looking at
Some work on Sage manuscript. Uploaded benchmarking data to Sage repo (release assets on rc1 release) We’ll eventually need to make a fork of nonbonded under the openff org JW – I’m happy to help with this. I recall talking to SB about the future of the fitting repos and he said that some could be moved to the openff org wholesale, and the others should just be forked when changes are needed. So this is probably a “fork” one. MT – Are changes needed? PB – MShirts commented on the manuscript that we should have an openff-namespace version of nonbonded to cite MT + JW – It’s probably counterproductive to fork it just for the sake of citation. We can ask SB to make a zenodo entry if we need a citable artifact, and that way he’ll get the credit. PB – Ok, when it goes up for coauthor review it would be great if you could reply to MS’s comment with this info.
JW – RDKit UGM was great. Really enjoyed the vacation as well. My lightning talk slides went fine but the live demo failed because too many people clicked on the binder link (in the future, I should send out the binder link AFTER the talk!) This week, I’m going to continue clearing my message/org backlog and push to update our fitting stack. I’ll try to also tackle some issues in the process. Overall hoping to do lots of quality improvement. Probably a lot of Fridays off for the rest of the year.
|