2021-09-22 AM1 Restraints Meeting notes

CD Updates

• Implementing ELF1 for the past two weeks, trying to see if we got more consistent partial charges.

• Broke this up into two datasets:

  • RDKit-generated conformers

  • OpenEye-generated conformers

  • CD – I noticed that OE generated way more conformers than RDKit

    • JW – Could either be that OE conformer generator is better, or that they’re using a different definition of RMS in their RMS cutoffs (all-atom vs. heavy-atom)

    • LW: RDKit is heavy atom only

    • LW: OE docs are not super helpful in explaining what goes into the RMS threshold. https://docs.eyesopen.com/toolkits/python/omegatk/OEConfGenClasses/OESliceEnsembleOptions.html#OEConfGen::OESliceEnsembleOptions::SetRMSThreshold

      • I don’t quite understand this option SetMaxTerminalHeavy but it affects the RMS calculation: https://docs.eyesopen.com/toolkits/python/omegatk/OEConfGenClasses/OESliceEnsembleOptions.html#OEConfGen::OESliceEnsembleOptions::SetMaxTerminalHeavy

    • JW – This lines up with OE doing all-atom RMSD (including hydrogens).

      • LW – There’s an EmbedParams keyword for onlyHeavyAtomsForRMS Can we pass EmbedParams to EmbedMultipleConfs in the python layer?

      • https://rdkit.org/docs/source/rdkit.Chem.rdDistGeom.html#rdkit.Chem.rdDistGeom.EmbedParameters

    • JW – May also generate more conformers by setting useRandomCoords=True in EmbedMultipleConfs

    • LW – Or setting useExpTorsionAnglePrefs=False, useBasicKnowledge=False

• Tested whether Simon’s RDKit-based ELF implementation selects the same conformer as OE’s ELF conformer selection. In the 36 test molecules I tried, both methods selected the same conformer every time.

• Open

  

  • JW – I don’t think the difference between these is super significant. The MSE would be the best metric here, but it’s not that dramatically different.

    • CD also calculated RMSE – got 3.5e-2 vs 3.7e-2

  • CD – I may try doing OE AM1 calcs for both conformers.

  • JW – It seems like the problem must be that OE and RDKit are coming up with substantially different conformers to pick from. Maybe we should have RDKit make more conformers? It could eb that there’s an in-built energy window in RDKit that we could expand.

  • OM –

  • CD – I’d like to compare conformer RMSD to partial charge deviation to become more sure that the difference is due to conformers

    • (LW+JW) Resources for calculating RMSDs:

      • https://rdkit.org/docs/source/rdkit.Chem.rdMolAlign.html#rdkit.Chem.rdMolAlign.CalcRMS

  • LW: Energy windows:

    • RDKit, can’t see any options for changing; might have to poke around in the source

      • Breadcrumb in a recent PR: https://github.com/rdkit/rdkit/pull/973/files#diff-f8a94dcd981c541fcfc9a6461afec491f4172b4d75a488b759a9b50a0d1a2c83R341-R344

    • OpenFF toolkit sets energy window to 15 kcal/mol