2025-10-07 FF Fitting Meeting

LW – Decide which FF to release as 2.3.0rc2 candidate

Open

JW (slide 3) – Interesting that 2.2.1+NAGL had problems with RBFEs, but something that changed in 2.3.0rc1 recovered from it.

LW – We see densities improve in both fits. I think the problems with mnsol are due to the amine composition ofthe tarining.

CB – Since MNSol is in differnt solvents, maybe there’s something to do with properties in non-water solvent.

RBFE slides: https://docs.google.com/presentation/d/1oMDOQ0TlXjGyv5LUcmegfxqsQLvHA2FVxdk_HbTp_QA/edit?usp=sharing

JAC – Do we use tip3p water consistently in all benchmarks?

• LW – Yes in all the ones shown in slide 3 of my presentation

LW – Q: were there version differences in the 2.2.1 repeats?

JR – What is v3-k100?

• LW – That’s the FF version we’re considering for 2.3.0rc2 🙂

CB: Are brown bar and red bar repeats? No, they’re rc1 and rc2. I do see that red bar (rc2) and purple bar (rc2 + NAGL) are quite close, suggesting that ELF10 and NAGL are performing pretty similarly.

• IA: seems to me that …

• CB: between 2.3rc1 and 2.3rc2, is it just the vdW that differs or everything?

  • LW: everything

• JW: so SAge (ref) and Sage are the same parameters on 2 different days?

  • IA – Yes. Though I’ve run some of these with more repeats.

• CB – IA, in your opinion, having seen lots of these sorts of results, does there seem to be significance in these results?

  • IA – For overall all-ligands-for-a-target plots, no. But looking at the deviations, some of them could be significant.

• JR – What about BACE?

  • IA – There is a big difference, but I’m not sure that it’s significant.

• LW – On the two age versions - Were there any version differences between sage 2.2.1s?

  • IA: there was a stack update here, the toolkit changed. The charges would have been the same.

• LW – Deviations/stats are computed on 3 replicates?

  • IA – These are calculaetdon the dGs. Based on all data points in a set. Error bars mostly based on bootstrapping data points, a little based on reported error.

• JH – Maybe I missed this, what is the reason for the force field names ending in rc2 rc1. Are those the NAGL charge models that should be used with the force field?

  • LW – The rc stands for “release candidate”, these are different candidates we’re considering for the full 2.3.0 force field release. We usually release release-candidates first for people to try before full release 🙂The NAGL model is the same across both

  • JH – And these are all using the openff-gnn-am1bcc-1.0.0.pt NAGL model?

  • LW – Yup

• CB – Re: 2.3.0-rc2 with NAGL doing worse than ELF10 for some systems. Could this be tracked down to a diffrence in charges?

  • IA – I didn’t look at this in particular. JHorton had looked at this and found its ligands to be more susceptible to conformer-dependent charges. We haven’t looked at PTP1B before, it may be one of the ones flagged by meghan.

  • LW – IIRC, PTP1B has charged ligands. I recall NAGL having more trouble with charged ligands.

  • CB – I wonder if one result could be ”we recommend using NAGL charges, except on charged mols”.

  • LW – There are some recommendations we can make, we considered making similar recs for small molecules and we put in a lookup table instead. Since we don’t intend to change the charge model for 2.3 at this time, in terms of communicaitons we can make around rel;eases, that’s probably the only safeguard we can put in.

  • IA – The thing that’s mostly encouraging from these results is that using ELF10 won’t cause serious regressions (some of our partners will be hesitant to switch to NAGL, so it’s good to see.

• JR – What fraction of NAGL training set is charged?

  • LW – A small fraction, details here: https://docs.openforcefield.org/projects/nagl-models/en/latest/models/openff-gnn-am1bcc-1.0.0/index.html#training

• LW – Given how common TYK2 is a very commonly used test system, are regressions on there especially important?

  • IA – It was pretty sensitive to small FF parameter issues before (when ff14sb parameters were being assigned to an ACE-like moeity in some ligands).

• LW – for p38, you said JH looked into it and found some conformer-dependence. I guess I’d expect NAGL to perform better since it isn’t conformer dependent.

  • IA – When I say conformer dependence here (and I may be misremembering),JH had run this system with NAGL and with ELF10 and JH had found a difference. One thing here is that the ligands have some complex rings (like a cyclehexane-ish moeity).

  • LW – I also recall seeing larger discrepancies between AT and OE on these sorts of moeities. So if this gives a “free tier” access to ELF10 style charges, even if they’re slightly worse, it’s a big improvement for most users.

• LW – Do you have data on these systems with AmberTools

  • IA – Yes,but not ready to present today.

• IA – Re: running on more systems - I would advocate for this. There's a lot more diversity in the rest of the Ross set.

• CB – So we think v3-k100 RBFEs are AT LEAST comparable

  • IA – Generally agree

  • LW – Same

  • CB – User-facing consequences of v3-k100’s decrease on mnsol?

  • LW – Decreased performance on nonaqueous solvents, possibly hydrophobic binding pocket

  • IA – Most JACS set ligands are quite charged, not enough hydrophobic info to say with much certainty.

  • CB – So it could be good to do more systems

  • JW: we would like to get a release out

  • LW – It would be great to get as much benchmark data as we can, but within constraints. We can expect to get more feedback from community and that will help us find targeted improvements.

  • IA – From OpenFE, the answer to "is there time and resource" would be "if there's a graduate student willing to help".

  • CB – So, it sounds like the consensus is that, even if there’s an issue with hydrophobic binding sites, the RBFE data says it’s comparable. But since a FF release with NAGL charges will have so much value to users, we should getthat in their hands.

  • LW + JW + BS – Agree

  • BS – Problem with MNSol could be vdW issue.

  •  

•  

•  

LW – Assorted items on why 2.3.0 rcs are having trouble with mnsol

• JR – Is it obvious why the transfer free energy (v3-k100) is slightly improved but MNSol is substantially worse?

  • LW – It’s kinda half and half

• IA – When comparing MNSOL, did anyone look at the gas phase vs solvent phase legs?

  I.e. is there a shift specific to one?

  • LW – I looked at this but forgot the conclusion.

• CB – If LJ terms make things smaller, then that may overstabilize hbonds.

  • BS – Was thinking the same

  • CB – Slide 13’s decreases in SFE would support this

• JR – Do polar hydrogens have particular vdw params?

  • LW – Yes

• IA – A bit of a shot in the dark, do you see any shift in the SFE estimate errors?

  • LW – I haven’t looked at this

• JR – Have you tried letting ES and vdW to change at the same time?

  • LW – Not yet, that’d be quite difficult. JH is trying to co-optimize valence and vdW, and it’s a bit difficult.

• CB (13) – I like this slide because it shows that NAGL charges aren’t the problem with the mnsol performance. We could say that, if mnsol is having trouble with nonpolar solvents, then it’s not a bad enough problem to appear in nonpolar active sites. in RBFE calcs.

  • LW (shows some figures of performance grouped by solute vs. solvent)

•  

•