Individual updates | DN We delivered the annual report to industry partners! 🎉 I’m really appreciative for the help that everyone gave when pulling this information together. This week, I’ll be updating the website. Previously I was making updates to align our public-facing material with the new annual report. I’d love ideas about how to improve the website JSetiadi is working on his blog post, I expect to review that in the next month.
MT Likely offline Thursday OpenMM import - Working on this for a few reasons - No RC feedback so far. Also this is a highly requested feature - This will be an important structural component for other importers (AMBER, GROMACS, CHARMM, probably others). So the exciting thing is that this could enable usage of externally-parameterized components in OpenFF workflows. Though I expect this to be a lot of work, so I’ll be on this for a few weeks. Pymbar release with breaking changes, minimal impact so far but tag me if any issues arise. Progress on EPs based on OpenFE feedback. Currently stalled, however, and I don’t see it as my responsibility to un-stall it. Virtual site position math stalling, could use 1-2 hours of pair programming with somebody. Maybe I’m setting particle positions incorrectly? JW – I’d love to work on this. Let’s do this at our working session wednesday. DN – I’m happy that this is progressing - The ad board really wants vsites JW – Two things to keep in mind We’re talking about the infrastructure to apply vsites, there’s also the issue of having infrastructure for a force field trained with vsites, which is the second half f what the ad board wants, and now something we’re guarnateeing The OpenFE infrastructure doesn’t support vsites, and won’t support them in the near future. So we won’t be able to run PL benchmarks on Rosemary if it has vsites.
DN – To clarify, our plans haven’t changed for Rosemary - It won’t have vsites, even though everyone keeps asking. DD – From my perspective, I think we’d just need to add vsite handling to the atom mapper. Then everything might just fall into place.
Otherwise quiet week and/or forgot what happened
DD PB Mostly fitting related work helping JM setup her FB runs for torsion multiplicity work, debugging failures dug a bit into creep in some angle parameters over generations of force fields and my hypothesis is ignoring Gen1 sets completely post v1.2.0 fits might’ve caused this. There were some torsion parameters, coupled with those angle parameters that had discrepancies, that didn't have any torsion profile training targets but were optimized just with optgeo targets. Some smaller fits with additional targets from Gen 1 sets seem to work better, going to do a larger fit to see if there would be any improvement. One example, a40 and t123, for phosphorous related molecules. Another, a31, a32 and t155 for hypervalent sulfur. t123, t155 (and a few other parameters) are completely missing from Gen2 sets post filtering out strongly H-bonded molecules. PB – After filtering the 800+ torsiondrives, there are only 700 left, and so some parameters are trained to only 1 or 2 torsiondrives, and other don’t have any targets at all.
Sage manuscript revisions based on DM's feedback, have few more sections to overhaul. Valence parameter optimization for Double-exponential fits had a problem with few targets failing consistently and Josh found the source of error to be the OH vdW parameter, which is supposed to be a non-zero small value (a fix used in LJ versions) also, we decided to reduce the scope to trained vdW parameters, and excluded S, P, F and I from the valence training set
JW – We got the annual report out the door! 🎉 DN – I think we invested this year in making a process for putting out the report, and that next year we’ll have this ready to go. DD – Could I take a look at it? DN – Sure. I’ll share the PDFs.
Working on fixes for biopolymer rc. Think I fixed OE atom indexing in PDB. This week will go full steam towards biopolymer, working with MT on interchange, reviewing+approving interchange regression testing
|
Add Comment