2024-04-23 AMI/LW Check-in

General updates and discussions

• Benchmarked Parsley as well

  • AMI: All Sage seems to do better than Parsley

  • AMI: Filtered subsets of benchmark data to look closely at improved parameters. Bridgehead Ns covered many mols. I have phosphates and sulfamides

    • LW: with the wider range on bonds (sulfamides), you’re suggesting that it’s not the outlier points getting closer to the median (and therefore increasing the visual range) but some points near the median shifting out?

    • AMI: haven’t looked at this closely but at least some of them, yes – saw this with sulfamide angles at least

  • AMI: there’s some parameters introduced that benefit the protein FF but we don’t benchmark on those

    • Possible benchmark dataset: (Chapin’s training data)

    • AMI: 1-2 days to look into this

    •  

• Status of conformer ordering and ab initio target experiments

  • AMI: started analysing re-benchmarks, story looks similar to previous. Between forcefields, differences are ~100. Not a monotonic change. Don’t think data analysis itself is the most helpful. Perhaps cluster bad molecules to see if there are poorly-performing molecules. Need to re-do now with fewer bad molecules

  • PB: tried doing just energy and force re-fits in the past, found it gave him bad geometries. Tried doing torsiondrives, but didn’t spend too much time on it

  • LW: does FB use ultratight convergence tolerance?

  • AMI: uses a much tighter default but not ultratight. 0.001 kJ/mol/nm. When I tried this gave similar results as super tight one.

  • LW: easy experiment would be to change this

  • AMI: can look into that

• hypervalent S dataset

  • AMI: looked at size of ChEMBL molecules, they were pretty big

  • Started looking into this, still in progress

  • LW: BW also looking into fragmentation schemes

Event

• LW: will be in touch for slides