2025-08-27 Cole Lab Check-In Meeting Notes

Benchmarking of SPICE valence fit with new torsion types

• DC (6) – With unintentionally included constraints, did that affect the loss?

  • FC – Unknown, need to look into this

• DC(7) – Why so many NaNs?

  • FC – Unsure. Similar numbers to previous runs.

• DC(19) – Would be interesting to try something else that was trained on SPICE

  • DM – Would be interesting to try and fix this. There was a previous issue in AMBER protein FFs where peptide bond scans weren’t in training set, and when people started doing enhanced sampling they got peptide rotations.

• AA (19) – Is there a reason you want to include torsion scans in SPICE itself?

  • FC – … Wouldn’t be hard to add them in.

  • DC – We want to get away from running explicit torsion scans in fitting, since that’s super time-consuming. I don’t have a philosphical problem with taking QM singlepoints along a torsionscan and adding those. (introduces alice)

  • AA – I’m a group leader in Max Planck institute, inventor of modified seminario, formerly in FF development.

  • CC – Re: doing FB-style minimization along TD vs. singlepoints, the latter gave me bad results. But if I fit to PAIRWISE energies from single points it worked.

  • DM – Maybe this is an argument for revisiting Trevor's nudged elastic band work - getting a pathway across torsional space that you can do singlepoints on, maybe? Gonna have to do something about the barrier heights for "mostly non-rotable" bonds. I am not sure what.

• DC (conclusions) – One more round of fitting to clean up performance on phys prop data could be good. Could be good to decide then that this line of work is complete

• AA – FOr ML potentials, we have known uncertainty. Is there a similar thing for MM?

  • FC – I’ve been thinking about this. For the torsions there are different levels of specificity. So could have ensemble of torsion splits. Could also quantify how far mols are from a training set.

  • AA – Have you seen a variety of different parameters suggested by training data?

  • FC – … (recoding ~25 ? mins)

  • CC – This is problem that regularlization solves

  • FC – JH had tried regularizing to 0. But need to find the right strength.

  • CC – The classic way to do that is cross validation.

  • FC – Tried some regularization when I saw improper ks getting really high.

  • CC – May also be able to regularize bonds and angles to MSM values

  • …

• AA (33) – Is torsion angle coupling the most important?

  • FC – It seems to be likely, some studies suggest this is one of the more important routes. We also tried adding urey-bradley terms but didn’t really see benefit.

• DM – If you imagine pushing torsions to the extreme, then you should be able to continue making training set more detailed/larger, but you’d see test/train performance fall off. I wonder what point that happens at.

  • FC – RIght, eg OPLS3e has an incredible number of trorsion tiypes and does well.

  • DM – Maybe, I’d be carefil abvout that since it could just bethat it imporved when torsion…

  • JH – FF builder showed improved performance

  • DM – Possible that they’re so overfit that they need ff builder to recover against overfitting

• FC – Thoughts on aromaticity?

  • DM – We don’t like it, but kinda need it. Unless we use bond orders from a semiempirical calc, we’re stuck with graph info.

• BS – SPICE dataset is generated using molecular mechanscs using some FF. So presumably its boltzmann weighted toward that FF. Then you do QM. Are confs reweighted using QM?

  • DC – Good point, but I’d argue that QM geometry opts are important. But I’d argue that you don’t need to capture all minima since important information could be captured by another molecule finding min even if this molecule misses one.

•  

•  

•  

•  

•  

•