2020-09-16 BespokeFit meeting notes

Pharma partner/Roche benchmarking

• DM – Xavier Lucas (Roche) wants to run local ANI jobs

• JH – Should be possible. Will meet with DD and XL soon

• JW – JH, are you interested in being involved to the tune of up to 5 hours/week with industry benchmakring paper?

  • JH – Yes – Seems to overlap with my goals, and I could push the bespoke workflow as well.

• HX – What is scope of benchmarking?

  • DM – Similar to preprint on chemrxiv. Different pharma partners would do the same study on internal datasets.

  • https://chemrxiv.org/articles/preprint/Benchmark_Assessment_of_Molecular_Geometries_and_Energies_from_Small_Molecule_Force_Fields/12551867

  • HX – Will read. What’s the scope of test set?

  • DM – 2000ish molecules, 20,000ish geometries

  • HX – We’re doing something similar with CNO compounds. Chemistry is a bit more constrained than paper set.

  • DM – We could use more simple molecules for our fitting, if you’re willing to share.

  • HX – Our dataset is focused on small fragments. Allows simpler enumeration of hierarchical torsions. Would love to check out your molecule set. Ours et has fewer molecules.

    • OpenFF molecule set:

  • HX – When doing torsion drives, how are structures propagated along the scan?

    • JH – TorsionDrive method/wavefront propagation. https://chemistry.ucdavis.edu/news/driving-torsion-scans-wavefront-propagation

    • HX – Is there an initial MM minimization step before running QM?

    • JH – No, we take starting structures directly from RDKit/OpenEye.

    • HX – Is that done for each structure in the scan point?

    • JH – No, only the first scan point is done using a cheinformatics-derived starting structure. After that, the scan points are propagated from the QM-minimized structure of a neighbor.

    • JW – For generating starting structures, we’re making “OpenFF conformer generation” as the first tool in our CLI

    • JA – Could we do semiempirical initial minimization to get a fast, good starting point for more details QM?

    • DM – Somewhat concerned that we’d be led into local minima

    • JH – Unsure whether this would be a systematic problem.

General Check in

• JH – Polishing up bespoke workflow.

  • Currently can submit molecule from commandline, will trigger ANI scans, unsure about output spec.

  • Need to optimize FB settings – Some outputs seem wonky.

  • Fragmentation/graph matching still has some bugs.

  • Large dependence on initial conformer method (OE vs. RDK) – How to control for this?

    • DM – Ask Relay about how they handle this? They’re big RDK users.

    • JH – Would like to check whether full QM shows the same problem as ANI.

    • JW – Start recording these cases in an issue on bespokefit so that we can squarely attack this problem later.

  • DM – Ready for beta testers? Lots of pharma folks would be interested.

    • JW – Should we meet about the spec for the minimum viable product?

    • JH – Yes

    • (General) – JH should be involved in spec discussion for atom map refactor/implementaion

• JW – Atom map spec?

  • JW – System object development will reuqire more extensive atom mapping

  • JH – Don’t really store atom map in molecule, store it nearby in a separate dictionary. Currently lose track of atom mapping during fragmentation. Would be preferable to keep track of this.

    • (General) – Broad adoption of bespoke workflow will require removal of fragmenter OE dependecny. But in short term, fragmentation isn’t required.

      • For fragmenter backend replacement – JACS benchmark set has good diversity.

Galileo meeting prep

Give example of bespokefit?

• No easily-deployable conda package built, so let’s skip this.

QCSubmit demo

PB’s current understanding

• Submissions begin as QCA dataset submission PR

• DD has a workflow that prepares and submits those jobs.

JH – The problem that QCSubmit solves

• Helps build dataset for submission. QCArchive is new technology and there were no input format converters, and ambiguity about standards.

• QM representations of molecules lose track of graph. This led to ambiguous interpretation of output molecules when we need to recover their graph.

• QCSubmit objects are heavily-validated python objects. Includes logic for avoiding doing torsion scans on, eg, linear torsions.

• Torsions are the most important thing to get right when looking at a molecule’s energy surface.

• Users have inputs in a variety of formats.

  • If they have SMILES, then we need to generate “good” input coordinates.

  • If they have other 3D formats, the conversion pathway to QCSpec input is different.

• Install instructions and source code for QCSubmit

• Initial tasks for Pavan

  • Reproduce previous submission – OpenFF biaryl set