2025-09-09 FF Fitting Meeting

2.3.0rc2 fitting update

• DM (2) - does “solvation free energies” mean non-water solvents?

  • LW - I’m using “solvation free energies” to mean water solvation, non-water solvation, and transfer. I’ll call out subsets of those when needed.

• CB (9) – Were there any trends WRT functional groups? Eg. that would tell us it’s not something with torsions but rather with electrostatics(?)

  • LW – Only 87 HFEs on this slide (9) but slight decrease in accuracy with ketones on SFEs (11)

• DM – So this is transfer free energies of MNSol? (yes) OMadin had found that if we fit to densities and mixture properties, we saw significant benefits on transfer free energies and madebetter FFs, but it didn’t help solvation free energies.

  • LW - The new Sage 2.3 candidates improve densities and mixing enthalpies but are worse on MNSol, so I think this agrees with OMadin’s finding

• CB – When MNSol MSEs is negative, it means it’s overstabilizing the hydration/solvation?

  • LW – Yes.

  • CB – We’d previously thought, since we were neglecting polarization, we needed to pre-polarize mols. And the idea was that, if you used overpolarized charges in a solvent less polar than water, you’d expect SFEs to be too negative. I wonder if this is telling us that the charge model is too negative/polarized. But blue/orange points are the different charge models, and they have the same MSE, so the difference must come from vdW fits. Maybe vdW radii are too small.

  • LW – We have a lookup table in our charge model for very small mols, and I wondered if that would assign different charges for some mols in NAGL vs. ELF10 comparison, but there weren’t many mols in this dataset that got charges from lookup table

  • LW – vdW didn’t change much in magnitude in RC1, but did change more in rc2.

• CB (11) - Are ketone molecules simple, e.g. acetone? If so, you could get a useful readout from the molecular dipole of whether the error comes from electrostatics vs vdW.

  • LW – Good point. Mols range in complexity, and the ketone designation can be met by EITHER the solvent OR solute. Can look more and report next week.

• JW (9) - does CB’s point about overpolarizing fixed charges in water explain the results for water (slide 9) vs non-water SFEs (slide 10)?

  • LW - Yes, water SFEs have near-zero MSE. Non-water SFEs have negative MSE. This would be expected if NAGL-charged mols were overpolarized

• DM (18) - what does explicit torsions mean?

  • LW - specifying multiplicity (?) based on BWestbrook’s work

• DM (19) - Are transfer free energies better for v3?

  • LW - Not really, error bars overlap

• DM – Enthusiastic about this, since getting the better charge model out there is a good idea. And results look good.

  • LW – Agree, this looks better than rc1

  • CB – Agree

• JW (24) - Why is 2.0 better than 2.2.X?

  • LW - Starting from MSM made a difference in geometries, and we’ve also changed the way we calculate RMSDs

  • CB - Why would so many structures move to a different conformer? It’d be concerning if there were minima that the QM saw as minima, and the MM didn’t recognize them at all. Do you know whether this is the case?

  • LW – Haven’t looked into this yet. Also interested to see how different the minima are.

  • DM – Anecdotally, we’ve looked at this before. Hard to draw significant conclusions from anecdotes though, and optimizing to different minima could be caused by small changes in the peak location.

  • CB – Having a qualitative difference in the shape of the potential energy surface is significant. Maybe a two-stage minimization where the first one is restrained would be informative, since that would guide the mol toward staying in the local min.

  • LW – The ddE analysis would hopefully help identify this

  • CB – Size inconsistency with RMSD would make it hard

• JW (chat) – Also worth mentioning that 0.4 A rmsd is very small, and would be hard to spot visually, so I wonder if we're talking about meaningful minima there.

• PB (chat) – just speculating about issues with ketones and amides (seen in earlier iterations), whether splitting the generic [#8X2H0+0] vdW term into carbonyl and (heteroatom double bonded to oxygen) terms make any sense, don't know if anyone looked into it before and arrived at the conclusion that it doesn't matter

  • LW (42) – the carbonyl or ketone oxygen parameter changed more in rc2 than rc1

  • CB (40) – I want to make sure we’re not comparing point charges, but rather something like dipoles. So where there are troublesome things with ketones… hard to say whether it’s uniquely due to charges or vdW.

  • LW – Looking at dipoles is a good idea.

  • CB – Would help for sulfones, ketones, etc.

• LW – Might be better to benchmark on a different subset of freesolv, though this would cost GPU time that could otherwise be used protein FF efforts and binding free energies.

  • CB – Could cheat with freesolv by using quick PB approximation, and look for systematic discrepancies.

  • DM – That’d only give us electrostatics, not vdW

  • CB – Yeah… changing radius is possible but would be new science.

  • LW - In early NAGL benchmarks, we found it was hard to shift the error across FreeSolv between force fields

  •