2024-10-09 FF fitting meeting

NAGL2 charge method

LM

• DM/BS: have you reported this to Psi4? What have they said?

  • AMI: not too much response. SCF is probably an external package, I may report to Psi4

  • Existing issue on Psi4:

• DC: Some QM people I talked to weren’t very surprised that diffuse + PCM didn’t work.

  • AMI: a lot of my PhD/postdoc colleagues use PCM and had better results, but it is a huge basis set so not fully surprised there were issues

• CB: If 60% molecules converge, is that all the chemical environments you want to characterise, or are there entire chemical spaces that fall into the 40% failures?

  • AMI: haven’t analysed the split. JH saw a lot of iodine failures, but I have seen non-iodine failures. Haven’t really analysed functional group presence in successes/failures

• CB: do I understand correctly that the MBIS charges are related to the issues?

  • AMI: SCF convergence is unrelated

• BS: CB, the danger might be that there’s a bug in the code that affects everything and just doesn’t manifest on 60% of them, or there’s an issue with the methodology. I’d be skeptical of this method

  • CB: if you ditch diffuse functions, the wavefunctions you end up with won’t move as much charge onto the relevantly charged atoms with high charge density

  • CB: back in the day I would make locally dense basis sets with diffuse functions on the atoms bearing charge density. This probably won’t save this problem.

  • CB: suggests a delta-ML model that trains to the 60% data with diffuse functions

• DM: Next steps?

  • AMI: Thinking of giving up on diffuse functions for now

• BS: suggests switching up radii and cavity sizes

  • AMI: that was one thing that worked with Josh’s test molecules. Not sure it was a generalisable solution

• TG: do you opt geometries prior? I used to do a multi-step approach with an easy basis set and increasing the method later (e.g. HF → add diffuse → add PCM). That usually helps with initial guess

  • AMI: I tried reading in vacuum initial guess and that didn’t help. It’s also not supported by QCA.

  • AMI: geometries are optimized in vacuum with B3LYP. Issue is not with geometries, but with SCF convergence of the optimized structure.

• DC: if you’re not using a charge scheme, how are you getting around buried atoms in the molecules when you fit to ESP?

  • AMI: thinking of a RESP-like restraint. One concern was that if we just train on ESPs, we may just get RESP charges out

  • BS: is there a problem converging the wavefunction with PCM?

  • AMI: yes. Not specific to MBIS

• CB: going back to larger radii on negative atoms, AMI, you said that worked but seemed dodgy. Could we use some of the converged calculations to check if there are artifacts? Would that be a general solution?

  • AMI: do you mean to compare results between calculations?

  • CB: yep. sounds like this could be a path forward

  • AMI: I can compare this. It did work for a couple test cases. You can use UFF and Bondi radii for PCM. Typically UFF is 1.1x, Bondi 1.2x scaled. I scaled UFF to 1.2x in my attempts, which I’ve seen in other codes like Gaussian, so not crazy.

  • BS: Speaking to artifacts DC mentioned, we saw the dipole moment was very sensitive to the radii.

  • CB: Bondii radii are pretty old. There’s other radii sets we could use. Not sure we have a gold standard. Can you read in your own vdW radii? Bondii not necessarily constructed for dipoles/ESPs.

• AMI: happy to try vdW scaling quickly. Looking for custom vdW sets could be tricky.

  • DM: does having diffuse functions with different radii always improve results over not having diffuse functions?

  • AMI: I’m not an expert in PCM so I can’t necessarily speak to that. RESP2 averages PCM and vacuum. Vacuum will be quite substantially improved for anions.

  • CB: solvation response will be less pronounced. (see recording ~36 min…)

  • DM: ok, so not an easy answer, a research problem.

  • DM: is there a simple test we could run to check if this improves results?

• DC: seeing differences around 0.1e with higher level charge model vs AM1-BCC. Might be able to drop diffuse functions without much effect.

  • DM: I would also try to reach out to PNerenberg for electrostatics for charge things. He’s looked a lot at dipole moments.

• AMI: I’ve only looked at neutral molecules so far. Can look at anions too. It might be that the errors introduced by ML and FFs might outweigh everything.

  • BS: diffuse functions help a lot with multipole moments.

• AMI: will choose example anion molecule and compare differences.

• DC: would decide beforehand what change in charge I’d worry about.

  • CB: I would not be worried about changes in charges on atoms. I’d be worried about final ESP changes.

NAGL Zinc charges

• CB: recently, with cofactors (e.g. zinc), we found that valence FF needed to incorporate zinc tetrahedrally and topologically to get charge transfer happening

  • CB: At Merck Frosst I made BCCs. Does NAGL account for cofactors bound to these key residues? That would be fantastic if so.

    • LW: not currently accounted for in NAGL1

  • CB: How does OFF handle metals now?

    • DM: amber parameters, still rather painful in other force fields too (zaff)

  • CB: BCCs were essential for accurate handling of zinc. NAGL could be a beautiful, general way to handle this if included early enough