2021-05-05 Benchmarking for Industry Partners - Development Meeting notes

Updates

• LD: doing torsion profiling, completed benchmarking at Janssen, including OPLS3 (defaults) optimizations

  • next doing the same with OPLS3 custom params, will be running ffbuilder next; unclear how long that will take

  • DD: is there a path-dependence to ffbuilder in terms of the params you get out?

  • LD: unclear

  • DH: the params for e.g. torsions tend to be very specific, so less opportunity for variation

• DH: not too much this week, worked with Lorenzo to get ffbuilder running

  • DD: if you folks are finished with results, can you export and add them to the drop zone?

  • DH: will get clearance path from Gary tomorrow

  • DD: does it make sense to re-engage Kaushik, Xavier for testing? Do we have bandwidth?

    • DH+LD: yes, think so

    • DD: cool, I’ll reach out to them and we’ll get it going

• JW: could not reproduce the issue raised from Thomas Fox; we need to regroup on that front

  • on openff-toolkit#924, not clear how Bill is calculating the dipole moment; unclear if AM1BCC is also problematic in what it’s giving

    • This code snippet illustrates what he should use:

      import json from openforcefield.topology import Molecule from openforcefield.typing.engines.smirnoff import ForceField with open('TST-00000-00.json', 'r') as f: res = json.load(f) # get final molecule mol = Molecule.from_qcschema(res['final_molecule']) # get conformer geometry in angstroms conf = mol.conformers[0] geometry = conf/conf.unit # get partial charges ff = ForceField('openff-1.3.0.offxml') sys, returned_top = ff.create_openmm_system(mol.to_topology(), return_topology=True) partial_charges = list(returned_top.reference_molecules)[0].partial_charges

    • We have a SMILES string from him to work directly on.

    • JW: I will take point on interfacing with Bill on this.

  • Fox issue: unable to reproduce; perhaps something up in the workflow that doesn’t preserve stereochemistry

    • DD: I will reach out to Thomas again, ask if he can create a sharable fragment if there’s some exotic chemistry there

    • JW: current hypothesis is that RDKit is encountering a case where the stereochemistry doesn’t change when it tries to figure out which to set for a given bond

    • JH: Simon shared what might be a similar failure in #qcfractal; doesn’t give error with most recent toolkit

    • JW: will give this a play with the version of the toolkit we’re using with the benchmark

• JH: no real change, talked about benchmarking infrastructure in a different call

  • Concern raised that we want to be able reproduce the Lim work; don’t know if we can currently

  • DD: need someone to take on running the Season 1 protocol on the Lim paper molecule set

    • JH: only need to run the MM step; the QM data is already in QCArchive

    • LD: I’m willing to take this on

      • DD: I will assist; have a fairly clear idea as to how to do it, possible problems we’ll run into

      • DD: 5/12 9amPT / 6pm Barcelona

    • DH: in Victoria’s repo, there is also a script to extract the QM data from QCArchive:

• DD: created openff-gopt; two initial issues:

  • optimization executor

  • torsiondrive executor