Started setting up protein force field benchmarks. Looking to run small peptides, disordered proteins, and folded proteins. Trying to figure out how to handle protonation and ion concs. Once that’s set up I should be able to extend/reuse these scripts
JW – It looks like there are a few options for expanding our GPU compute if the benchmarks show we need ti - Argonne?, PRP, Google/AWS/Oracle donations?,
PB – Lilac?
DD – Lilac has a lot of GPUs
JW – Lilac would be a great choice.
CC – This should be largely automated. So if we need to use pre-emptible compute we can checkpoint frequently.
MT – Out of curiousity, how are these sims getting set up?
CC – Will simulate using OpenMM (on development branch, they support setting up boxes using rhombic dodecahedron). Will use PropKa to assign protonation (heuristic, but the same thing as DH used for FE benchmarks - Generally considered to be good but not as good as eg Schrodinger tools). Simulations workflow is minimization, equilibration, and then multi-replicate simulation.
Met with Brendon Duggan, MG, DM on Friday. He’s measured a wide variety of small molecule NMR observations. We’re trying to figure out whether we can use that to fit/test the FF. Still determining feasibility, will plan to meet a few more times to work out details.
Â
MT
Backporting/forwardporting PRs between biopolymer-topology-refactor and master branches in Toolkit. So all PRs into master got forwardported, expected it to take days, instead just took hours. Pleasantly surprised.
Interchange regression tests
Did a lot of digging to learn that 200-300 failing single-molecule charge assignment comparisons were failing because of trans-cooh behavior in OpenEye wrapper. Ported this fix into the other branch and then only 30-40 molecules are failing (max partial charge on any atom deviation between 1e-6 and 1e-2). Really hard to isolate these last few. JW looked at this and the SMILES used an aromaticity model other than MDL, which may have led to it being interpreted oddly. Possible it’s due to duplicates in the dataset, hashing+caching, runtime non-determinism, or aromaticity issue above.
CC – It sounds like this is all OE - Do you see similar discrepancies in AmberTools?
MT – We’re just testing OE (since it’s fast enough to go through this industry benchmark set). OE is also more straightforward - It’s a single library to assign charges, rather than relying on the combined behaviors of RDKit and AmberTools. So this route prioritizes finding bugs in Interchange and the Toolkit rather than the external libraries.
Condensed phase tests - Did these tests a while ago. Didn’t see any differences. This was a little bit surprising since the single-molecule tests showed differences. But I looked again and there were no COOHs in the condensed phase work.
Value propagation tests - Looking good
WBO tests - Looking good
Vsites - Depends on SB refactor getting merged
JMitchell has been working on interchange docs for much of the last two weeks - Made a big set of changes to user guide. Last week he finished overhauling examples, and they’re in a much better state now. I’m really happy witht he state of Interchange docs now.
Â
Â
DD
Protein-ligand benchmarks
presented what we are doing as a F@H Science talk; positive feedback and interest
JW – Who is F@H community?
slower progress this last week on data model;
performed a research cycle on current state of Interchange (new docs ), current OpenFE example notebook (https://try.openfree.energy )
spending today using my notes to put out an iteration of Transformation, Protocol, AlchemicalNetwork
Have to work on Sage paper and blog post this week.
JW –
Worked half days Mon+Tues
Good amount of project-ish planning - Working on how we plan + communicate with OMSF and ad/gov boards.
Prioritizing vsite rework review and topology refactor
No longer checking numerical value of outOfPlaneAngle for DivalentLonePair to determine allowed values of match
SB – As the FF fitting user I’d prefer to compare the value numerically and get an error if the vsites are going to be basically on top of each other.
Enforcing correct length of charge_increment list for each vsite type (turns out one of my test scripts was feeding in lists of the incorrect length and the handler was fine with this)
SB – Timeline for 0.10.5?
JW – I’ll push to merge vsites today
MT – I can make relesae as soon as it’s merged.
JW – So plan on Tuesday release.
SB –
Been doing some vsite fitting. It’s super easy to see improvements on single molecules like pyridine. But whenw e try to do vsites on chemcical series (like larger molecules that include pyridine) then the improvement becomes a lot less clear. So it could be that comparing to electric potential at the level of detail we’re using isn’t sufficient to show improvement. Still need to look into why we’re not seeing improvement.
JW – Are BondCharges doing any better?
SB – Not sure - I’m seeing improvement for single molecules, but not when I try to generalize parameters (RMSE on larger sets doesn’t improve significantly).