2024-03-06 FF Fitting Meeting

Using Sage to estimate partition coefficients

Steven Goold

• Slides/recording

• Three largest outliers contain tertiary amines--> suggests issue with how functional group is represented

  • i4 in Sage 2.0.0 covers both planar and pyramidal N geometries, has a wide distribution

  • Looked at geometries, for first molecule find ok results for some N’s but bad for another in the molecule (part of aromatic system) compared to QM

  • Trazodone has particularly evident issues--QM says should be pyramidal, MM says planar

  • 2 of problem molecules have i4, one (quinine) doesn’t

  • Want to expand dataset to see whether the i4 is likely to be the culprit

• LW: Sounds like a real problem, sounds like i4 would hold a molecule planar even if it shouldn’t be. This is Sage 2.0 but Sage 2.1 probably would make it worse. Should work on fixing this

• VV: These are anecdotal results, also not sure if fixing it would actually affect the log P values that they are trying to predict

  • CC: thinks issue with LogP could be driven by water interactions, might want to try another water model. Even though OpenFF trained with TIP3P, still usually get better results with better water models

• VV: Do you think BespokeFit might help?

  • CC: have had good luck in the past with BespokeFit patching small issues with individual molecules that we don’t treat correctly, shouldn’t be too hard since we only have a few molecules, and would be good to see because it might tell us whether fixing i4 would fix this.

• VV: If we compare to GAFF, what would you expect to see? Expect it might be a widespread problem

  • LW: non-bonded parameters are re-trained in Sage, but impropers are from Amber so should be similar to GAFF. But GAFF might have more impropers, might have different result

• BS: Should charge model be different for polar/nonpolar solvent? Could try doing QM and ESP calculations to see

  • SG: haven’t thought about solvent, just gas phase, but could look into it

  • VV: Is there any recommended approach for this within OpenFF?

    • LW: Don’t think we have an official endorsed way to do this systematically, have OpenFF recharge that can fit charges to ESPs, or could use RESP, but no way to go beyond AM1BCC easily within our toolkit

  • BS: there are some theoretical issues to consider regarding changing the charge model

    • [discussion around 30 mins in recording]

• CC: were you using single-conformer AM1BCC or ELF10?

  • SG: single-conformer

  • CC: we’ve found much better performance with ELF10, much more consistent. Updated toolkit uses ELF10 by default

• VV: trying to publish smaller amount of data in special issue (ASAP), next follow up would be expanded, hopefully could explore some of these issues in the follow up

  • CC: BespokeFit probably the best bet for fixing issue as quickly as possible

  • LW: yes, assuming the geometries are the problem with the LogP