2021-04-09 QCA Submission Meeting notes

Industry Benchmark Set

David

• DD: do we include Bayer? Some may have been used in fitting, and not clear which ones from existing set

  • also could get 5k compound Bayer patent space from Katharina; pursue as separate set

  • JH: sounds good; leave it out of this submission and can pursue later

• DD: how do we want to do MM? Match protocol for Season 1 (final_mols as starting points for MM)?

  • JH: Could do both. Add MM specs to this set. Once QM completes, create set that does MM only from final_mols

OpenFF Amide Torsion Set v1.0

Simon

• Moving along, SB waiting for it to be complete to begin use

• Had production issue on GreenPlanet with jobs being consumed and dying quickly from preemption

• JH: have some compute on Newcastle HPC can donate. Best starting point for manager?

  • TG: multiple ways to do it; depends mainly on if you’re allowed to run long-running processes on head node

  • DD: check out benchmarking protocol for two approaches for manager, also QCFractal docs; hit up Trevor, DD for advice

Torsiondrive benchmark update

Pavan

• DD: almost complete, we’ll let the remaining optimizations continue to error cycle until it does

• PB: sufficiently complete to begin use

1st round Gen 3 torsiondrive dataset

Hyesu

• Working on number of torsiondrives; should be around 3000

  • current set doesn’t cover all the torsions we want yet

  • want this to finish within 1 month

  • will be used for Sage fit

• Validation error currently

  • JH: stereochemistry error can happen when you use openeye to generate dataset but rdkit used on GHA

  • TG: re-perception of SMARTS-input by rdkit causing issues

    • https://github.com/openforcefield/qca-dataset-submission/pull/201/checks?check_run_id=2306704206#step:8:73

    • TG: we need to know which molecules are failing here

  • DD: can we try to build this dataset with rdkit first? If that isn’t possible, can resort to openeye

    • HJ: sounds good, will try to build with rdkit and let you know if that isn’t possible

• JH: dihedrals are tagged; when you enumerate in qcsubmit, tags are lost

  • EnumerateStereoisomers will lose the tags, so want to tag the dihedrals you want to drive after

  • SB: you may not need to enumerate stereoisomers for this one; may make sense to not use it at all here

• SB: once we have the steps hammered out, we’ll have probably 2000 mols, so up to 2000 torsiondrives

  • might be able to pare it down to 500

  • if after full dedup, applying an hbond filter, aim to have 2000 torsiondrives to do

    • might be about 2 months worth of compute

• TG: one thing that can improve the speed of torsiondrives: if your range of conformers span the dihedral range, that gives the service better starting points

  • that would be an optimization to the process that can really help

• SB: TG do you have another HPC resource?

  • switched from 100 to 300 on HPC3

OpenFF Aniline 2D Impropers

Simon

• SB: submit following the torsiondrives set above