2025-03-12 BespokeFit Meeting notes

Procedural stuff

• FC will prepare meeting notes doc in the future (can use the “create” button here)

•  

How a force field fit to SPICE using smee performs on the problematic (non) planar molecules recently highlighted by Cresset in #ff-fitting

• …

• (slide 21) – Maybe Sage training set is missing the information it needs about how high energy these nonplanar geometries are. SPICE set has some of those. Also probably need parameter splits for ring angles/torsions. I think I saw this initially in BespokeFit.

  • DC – Agree.

• (slide 26)

  • DC – Reminder that these are outliers, and that aggregate performance was shown earlier.

  • JH – This is promising evidence that it’s not necessary to do our torsiondrive-style fitting, since these were fit all to confs.

  • DC – On slide 25, how are we getting the QM data on these plots without doing torsiondrives?

    • FC – I’m nabbing the data at angles that are available on QCA

    • JH – Would be better to re-minimize in MM. Can YAMMBS handle that?

    • JW – I think YAMMBS can handle torsiondrives, but you should check with MT.

• (slide 27)

  • DC – What are we doing with charged molecules? Slide 4 indicates that we filtered out charged mols.

  • FC – I saw SPICE2 had some problematic unconverged mols removed.

  • DC – Right, and MACE doesn’t support charged mols… And there are some high energy mols in QM that we don’t want to use (ex for high LJ energy). So we want one filter to remove large QM forces (some tips onthis in MACE-OFF paper), and another like a geometric filter for when SPICE mols have weird geometries (were generated by high-temp MD so some confs are weird, maybe AIMNET2 could help here).(some discussion of bad geometries from SPICE).

  • JH – David and Harry did some filtering for bad geoms, might be able to reuse that.

  •  

  •  

    •  

•  

• 
  
  

Plans for bespokefit_smee (fitting molecule-specific valence terms), as mentioned by Danny in #smee-valence

• JH – Been talking to cresset about how to reduce noise in FE calcs - How can we reduce noise in parameters? Ex thrombin dataset - if you use bespokefit you get lots of noise. Maybe using congeneric series would reduce this noise.

  • DC – Can they share data?

  • JH – Not yet. There’s some speculation that fragmentation is keeping long-ish chains on bespokefit fragments.

• FC will try using YAMMBS for torsionet500 (not on OpenFF QCA, but available in simialr format), reach out to MT/JW/JH when help is needed.

•  

•  

•  

Redo meeting description