2020-12-09 BespokeFit meeting notes

Reviewing TYK2 ligand performance

• JH – Did test again using GAFF-generated confs. Showed that all performances got worse, GAFF is still best.

  • TYK2 Refit Validation

  • DM – Would be good to show the slope and RMS error on these plots as well

• JH – After doing bespoke fitting, we systematically UNDERestimate energies (they’re below the diagonal)

  • JW – Is the issue that points are “under the diagonal”, or that the slope of the cloud of MM points is too shallow

  • JH – That the slope is too shallow

• JH – I tried recomputing conformer energies using charges calculated at those confs and it looks a lot better, if anything MM energies are now above the diagonal

  • DM – Is there some directionality to this? Like, is this process biased by the fact that the MM is trained from QM energies, but the training doesn’t have any penalty for missing QM minima.

    • (General) – The QM geometries are not minimized in QM, so there’s no “asymmetry” being introduced by training on these.

  • DM – Should talk to Simon about this – See whether refit nonbonded parameters could improve this a lot

• DM – Could be good to remove conformers with big steric clashes here

• JW – The energy ranges on these plots are really large. Is the contribution coming from “interesting” things like torsions, or “uninteresting” things like bond/angle distortions?

  • JH – A bit of both. Thinking about generating geometries from lower-temp MD with rotatable torsions zeroed out. What’s the best way to zero rotatable torsions?

  • DM – Could make a new torsion term where at least one atom in the central bond isn’t part of a ring, and assign that 0-valued ks

• JH – Interestingly, ANI vs. QM comparisons are better using openff-generated confs than GAFF-generated confs

  • (General) – Is it possible that one FF crossed torsional barriers and another didn’t?

  • JW – Could be a consequence of equipartition principle? If one FF has stretchier bonds or spinnier (lower barrier) torsions, then it’ll have more “distorted” geometries.

  • DM – Different geometry generation methods can lead to systemaic “metric” differences in a study

• JH – Looked at charge variation by geometry.

  • (General) – Super interesting. May be more related to QUBE work than openff.

  • JA – The J shape of this curve – Is this because the partial charges are weighted toward lower energy geometries, so the charge error is higher at high-energy geometries.

  • JA – If we were to look at a certain group of these molecule, and looked at “for which molecules are partial charges dpeendent on conformation”, and we separate a dataset into molecules which vary by conf, and molecules which don't, would we expect to see different patterns of charge variation?

    • JH – I think so. That would be my expectation.

  •