2020-09-14 Core-devs Meeting notes

Date

Sep 14, 2020

Participants

  • @Jeffrey Wagner

Goals

  •  

Discussion topics

Item

Notes

Item

Notes

F@H free energy spec

  • Matt Wittman walked through PR

  • JW – Would be good to kekulize SMILES, avoid aromaticity

  • JW – May want to include provenance for experimental data, since people will want to dig experimental data for outliers

  •  

Roundtable updates

Jeff Wagner

  • Chemical perception/AMBER FF porting stuff

  • openff-1.2.1 release

  • Pavan coming on board – Will be given a feature development task in OFFTK or other core infrastructure. After first 90 days will become more independent and start doing QM benchmarking and dataset creation.

  • Shirts has won a grant to do interoperability. Likely a combination of our current System goals and a parmed replacement. Room for about 1 more FTE. Meeting about this later today/this week

Matt Thompson--

  • Started lots of stuff. Approved vsites PR – Saw decreasing marginal returns. TG is still tidying up.

  • Lots of open OFFTK PRs. Need to either carry them forward or delegate.

  • Progress on bond WBOs. Kept track of the nice optional things that aren’t necessary for the functionality.Will handle those in separate PRs. Roughly ready for feedback.

  • Are bond WBOs ready for integration into a FF?

    • JW – Not sure. HJ’s training next week will get a lot more people ready to work on this.

    • SB – Do we have work happening for torsion WBOs?

    • JW – Unsure. I think JM’s started doing some work with torsion WBOs by hand, but hasn’t used ForceBalance yet.

  • Some work on conformer generation CLI tool.

  • Worked on OFFSystem.to_parmed. Thinking about how to divvy up OFFSystem ←→ other format work.

David Dotson

  • Had a wonderful vacation

  • Worked with Jeff Setiadi on pAPRika/Evaluator integration

  • Worked with Dominic Rufa on ligand benchmark set.

  • Addressed torchANI failures in QCEngine. Model files were being repeatedly loaded and this was causing file system slowdown issues

  • Working on JSON serialization.

Simon Boothroyd

  • Finished integration of Recharge into ForceBalance. Now can do ESP and electric field refits. Can do implicit solvent. Handles wavefunciton → ESP generation.

    • SB – Pulling wavefunctions form QCF will require further development. Current datasets don’t have wavefunctions, though they can be enabled.

    • SB – No implicit solvent yet

    • DD – do we want implicit solvent in the short term?

    • SB – It’d be good to start zeroing in on which implicit solvent model/settings we want.

    • DD – JH is looking at the technical side of this, but nobody is looking at the science side.

    • SB – Which calls?

    • DD – Tuesday QCF user group call, Friday QCF submission meeting

    • JW – Is timing of fitting session and chemical perception calls OK for UK?

      • SB – Mixed. would be good to move them to overlapping timezone-friendly times.

  • Automated a lot of the nonbonded fitting. Includes both phys pro and ESP data. Codecov >90%, numerous cleanups/refactors.

  • WRt to Dotson’s PR, fixed bug in evaluator re: mole fractions.

  • Fixed Evaluator bug with empty OMM checkpoint files. Fixes are in Evaluator 0.2.1 release.

Matt Wittmann

  • Working on maximum likelihood framework for FE transformations. Based on Hannah Bruce MacDonald’s code.

  • Getting ready to hand off projects, since I’m transitioning to a new job.

  • Integrated Cookiecutter structure into F@H schema repo.

 

MT – With 1.2.1 manual parameter fix, how are we going to keep this from happening in the future? Is there MM validation? Is there a general way we can prevent this?

  • JW – I don’t like the idea of MM test set, since it won’t inform us of what the specific problem is or how to fix it

  • JW – Optimistic that tighter restraints on k and more diverse training set

  • SB – Would be in favor of simple test/validation set that’s run before release

  • DD – Agree that these should be integration tests.

  • JW – It took a lot of weird permutations of events – HMR, 4fs timestep, vacuum simulation – to make this happen. Would we test just this permutation, or others?

  • SB – Think that HMR is a common thing in pharma workflows.

  • SB – Think that we should automate FF release benchmarks, like the sort that Hyesu does, but in a more automated, distributable way

    • DD – Automate using Evaluator?

    • JW – Agree – Could automate this or distribute a tarball of tests/benchmarks to run before each release.

    • DD – Could this run on GHA. We get 6 hours of walltime per workflow.

    • SB – GHA may have an option to run CI on a local machine

    • MT – Would like to establish an owner for this, and spec out what science it should cover.

    • JW – Could take a coverage set of molecules, pre-compute AM1BCC charges, and run HMR sims for 10k steps each. Shoudl take ~1 hour. Concerned about how the organization will handle it

    • MT – Not sure about the target number – 0/1000? 7/1000? What’s acceptable?I like the idea of the systematic sanity checks, instead of finding problems as they flare up.

    • SB – It’s concerning that we don’t have automated/comprehensive benchmarking. This is a major weakness until anyone can run benchmarks on a whim.

    • DD – Automating testing for the propyne case by itself its valuable, even without adding a more systematic solution

    • SB – Agree.

    • MT – Agree that the above would be a great starting point.

    • JW – If we started speccing out these tests, I’d think that the “coverage” set is something we can use immediately and get 80% of the quality of a perfect solution

    • Spec:

      • GH action

      • Runs on openforcefields PR that adds a new file

      • Does HMR sims w/ 4fs timestep if a new offxml is being added

      • Loads pre-charged molecule set

        • Molecule set lives in a separate repo

        • Should be seeded from s99F “coverage set”

        • Also add propyne molecule

        • Emphasize that tests are “canaries” – NOT meant to make rigorous benchmark data, just designed to prevent release of a bad FF.

        • These tests will NOT make data artifacts

      • Tests run

        • 4fs, HMR, vacuum

        • Verbose output – Each test should print its SMILES and success/failure, so that it’s easy to reproduce locally if a molecule fails.

Action items

Decisions