2022-11-07 Meeting notes

15 mins

Update on DEXP FF accuracy

Joshua Horton

• JH will upload slides here

• Slide 2

  • DM - What’s being compared here?

  • JH - Sage with fixed TIP3P water vs. DEXP functional form with reoptimized four-point water

  • DM - This is exciting, but it would be useful to let the water model reoptimize with Sage to make a fair comparison of only the functional form

• Slide 4

  • DM – Did you pull the hydration free energies from freesolv? Using other FF?

  • JH – These were done using absolv. Not sure if freesolv would be different.

  • …

  • DM – I don’t think this will change a lot using different methods… This was seen in previous work.

  • DM – Magnitude of offset?

  • JH – … Some low outlier is due to saturated ring

• Slide 5

  • JH – I think a lot of the error is contributed from sultur and phosphorous params that were inherited from GAFF and never optimized.

  • DM – In any comparison across force field, you want to make sure you’re doing it on the same compounds. So be sure to do the same filtering for cross comparison. We had this issue in SAMPL challenges where people where allowed to not submit predictions.

• Slide 7

  • JH – For some reason this won’t converge. MAybe big problems with sterics? Or priors are too large?

  • DM – Tobias had done work on inferring torsions from scratch. You may benefit from interfacing with that.

  • JH – That could be useful

  • DM – Tobias should be in a similar time zone to you. We’re not trying that yet for FF fitting.

  • …

  • DM – It wouldn’t surprise me if the periodicity and phase needed to be different now

  • JH – We could try that with bespokefit?

  • DC – What happened during training? Were torsion scans used or single points?

  • JH – Same training method as befspokefit uses.

  • DC – So there’s no DEXP torsion scan in the training.

  • JH – …

  • PB (chat) – Does Sage also show this kind of off-minima profile for this molecule?

  • JH – Yeah, Sage looks pretty bad here too.

  • CC – 1-4 scaling?

  • JH – Yup, that’s used. We optimized the scaling factor and didn’t see much improvement, so I don’t think that’s the problem

  • DC – Could be good to start with simpler molecules

  • JH – Agree

  • DC – I don’t understand why it doesn’t converge if it’s not getting better.

  • JH – Saw the same with the espaloma FF (?)

  • DC – But those were distorted structures, I don’t know why this would have the same.

  • JH – Are ring torsions optimized every time?

  • DM – We optimize the ones that are covered by the dataset, which may include ring- and non-ring torsions.

  • PB – I think sage doesn’t have too many in-ring torsions, and even if it has it they have separate parameters. An additional analysis for the DEXP paper could be dimer interaction energies. So Sage vs. DEXP could be handy

  • JH – Not sure how useful those would be if we’re aiming to model condensed phase, since dimer energies are vacuum.

  • DC – Bill Jorgensen starts his fits using dimer energies, but I don’t like that method. I think it’s useful to get the right order of magnitude, but in terms fo small charges to potential energy surface it’s not going to help much.

  • PB – Should I do another valence parameter fit with relaxed priors?

  • DC – You could do that. I’d suggest waiting until JH has done a wider analysis of current issues. I’d love to avoid redoing the transfer free energies. .

  • DM – I’d want to make sure that we have good evidence that doing something differently will make it better.

10 mins

Thoughts on LJ parameters on polar H atoms (don’t really have much data)

Daniel Cole

• DC – Thinking about starting a conversation on this in ff-release. (shows “QUBEKIT-FB” paper). Used this to make the case for vsites in force fields.

  • DC will link paper here.

  • DC – having zero or near-zero H parameters seems to make everything either the same or worse. So in the next generation it may be good to consider making those larger. The sigma ends up being