2025-03-26 Meeting notes

2025-03-26 Meeting notes

Participants

  • @Matt Thompson

  • @Jennifer Clark

  • @Lily Wang

Discussion topics

Item

Presenter

Notes

Item

Presenter

Notes

Update on outlier helper tool

MT

  • MT – A new YDS feature could be merging different datasets together (since we often test each FF once, so we have lots of sqlite files with one set of unique MM info and a set of oft-repeated QM data)

  • MT – How important to know which angles specifically in a molecule are problematic, vs. knowing which angle it is?

    • LW – It’d be really helpful to know exactly which mesurements/atom indices are problematic, so we can trace it back to parameters

    • MT – So Q was: Is it OK that ICRMDs are summed across whole mol, and not per bond/angle/torsion? And the conclusion is that this is ok for now but in the future we should plan to split it out.

    • LW – Agree

  • LW – How does summing across whole mol work? Like if everything was good except one angle, what would that look like? And how would it change for small vs. large mol?

    • MT – Unsure.

    • LW – Could also try finding the worst outliers by reporting maximum.

    • JC – What if there were two really bad angles? Would it be possible to have a threshold value and then report all angles/atoms above that threshold?

    • MT – Probably. Current kinda averaging RMSD-ish approach should be sufficient to find really problematic stuff. But the max approach may be good too. But doing all interactions for all mols in dataset might be a lot. So this could be handled in a subsequent step.

  • MT – Is there a standard way to present structural discrepancies?

    • LW – Not really, RMSD alignment is the most standard. If you wanna be more fancy you could try to align to a scaffold that doesn’t change much.

    • rdkit’s GetBestRMS is probably the best for this

  •  

  • LW – I’ve been having trouble with nglview, is there a way to work around that?

    • MT – Can dump out to PDB and use other tools

    • JC – When I was using RDKit mols, I was using py3dmol

  •  

Restraints in TD MM opts

JW

  • In the newcastle meeting, there was some discussion about whether restraints should be used during torsiondrive MM optimizations. The current YAMMBS implementation does apply restraints but DCole doesn’t think it should.

  •  

  • LW - why does DC prefer this?

    • JW - FB applies restraints to non-torsion atoms. DC says that (in benchmarking) these restraints shouldn’t be applied

    • LW - Brent and I saw that without these restraints, MM would optimize to a very different structure sometimes. Open to having a toggle, maybe need to talk with DC.

    • JW - LW & DC might want to follow-up on slack

  • JC – Feedback from cresset on torsion contributions might be a result of having resraints on non-torison atoms?

    • LW – The reason we have those is to fit torsion profiles of strained torsions… ForceBalance can’t weight things by RMSD

  • LW will follow up with DC on slack

  •  

  •  

Action items

Decisions