2024-01-03 Wagner/Alibay Check-in meeting notes

Antechamber keyword pass-through

• https://github.com/openforcefield/openff-toolkit/pull/1801

• Opening positions:

  • JW – Basically, I want to help this m,ove forward. But I don’t want to change public APIs. Happy to put this in private API or a side-loaded toolkitwrapper.

  • IA – Currently in the middle of trying to work out what’s going on. Quite messy. Likely we won’t return with a pure recommendation - In some cases ELF1 helps, in others it doesn’t. In some cases SQM keywords help, in others it doesn’t. So we don’t yet have recommendations for solid improvement. The thing we’ll need in the future is the ability to play around with sqm keywords. So rather than shimming more into the API, maybe a custom wrapper would be better.

• Resolutions

  • Options for custom wrapper

    • Could go in a separate package/repo on GH

    • Could go in an existing package that gets development-installed (might be better since things get installed on remote hosts)

  • Notes on implementaiton

    • Will need to do tookit_registry_manager to select the custom/experimental wrapper – Details here

  • IA will make a custom wrapper, reach out to JW if there’s any trouble/questions

Planning study conclusions/handover

• IA -- What kinds of information do you need to motivate the change? What’s a useful point to reach if I’m going to wrap this up in the next few weeks?

  • JW – Bar’s not super high - we want to do ELF1 anyway. The things that would help would be:

    • Establishing that current workflow leads to random charge differences of X magnitude

    • Establishing that charge differences of X magnitude mess up FE calculation results

    • Showing that using ELF1 reduces the value of X

    • Showing that antechamber keywords reduce the value of X

  • IA – Current status is that, in the cases that we’ve noticed, there’s significant randomness. Setting conf gen seed to 1 is not sufficient - Even GH runners with ubuntu vs. macos lead to differences. It’s not clear to us that ELF1 fixes this problem, since RMS pruning can choose different confs due to ordering, even if two machines make similar ensembles. .

  • (IA shows a case where ELF1 conformers of a series of ligands have larger RMSDs (0.0 A -2ish, median 0.7ish) when generated on different hosts than the “current” method (same seed) (0.0 A - 1.5ish, median 0.3ish).

  • IA – antechamber keywords seem to be less ambiguous and generally good for reproducing on the same machine (tight keyword and diagonalization routine).

  • …

  • JW – Goal doesn’t have to be 100% elimination of the problem. Right now prevalence seems to be 25%ish, and if we can drop that to 10% or 5% that’s a win. And in the end Nagl will take over.

  • …

  • OpenFE-OpenFF lead team agenda item - Do gov boards have different standards for NAGL benchmarking? Should we commit to shared benchmarking that would satisfy both?

  •  

  • JW will ask OFF lead team what we’d want to see if OpenFE is only going to spend a few more days looking into AM1BCC charge assignment variability. .

  • JW will look into validating ELF implementation briefly today.