2020-09-18 QCA Submission Meeting notes

geomeTRIC combinations

David

• Can we do combinatorics for e.g. phi and psi?

  • JH: yes, limited to two combinatorics

  • probably not advised, FB doesn’t accept 2D target

• TG: can select all rotateable bonds, just toss them in

• JH: could do different priorities for different torsions as separate compute calls

• TG: all these are small amino acids, just grab all rotateable bonds, maybe filter out the carboxyl at the end of e.g. aspartate

• TG + JH : make sure you (DD) submit the dihedrals separately with their own index in dataset.add_molecules

Upcoming FF fit datasets

Josh

• Have we heard anything about incoming datasets for the next fit?

  • DD: I keep expecting to see PRs come through from e.g. Hyesu, nothing yet

  • TG: Next fit (1.3) looks like it will mostly be an adjustment of priors

Implicit solvent usage

Josh

• All keywords now supported on psi4 nightly

  • TG: we’ll need to update our psi4 version pinning around when we need to use this functionality

  • DD: Josh, can you test the most recent linux build (https://anaconda.org/psi4/psi4/1.4a2.dev1013+fb7f385/download/linux-64/psi4-1.4a2.dev1013+fb7f385-py38hd3583e2_0.tar.bz2) and see if it meets our needs here? Simon may want to use it within a month or so, and we’ll want to verify we have a version we can pin to for production

ESPs

Josh

• Need this PR to proceed.

  • cannot rebuild psi4 wavefunction without this

• There is a hack we can do, but not sure if it’s sufficient

• BP : will talk to Lori about this PR, see if this is something she is willing to hand off or if it’s basically finished

• DD : we’ll loop back on this next week for path forward

Uploading datasets from private instance

Trevor

• Would like to be able to submit dataset computed on private QCA instance to prod instance

• BP: the export part of that is actually implemented; import part needs more work

• TG: For bespoke workflow would be nice to have

• JH: will definitely be helpful and is of interest in bespoke workflow

Benchmarking

Trevor + David

• TG: is there benchmarking tooling that can help with batch processing on the QCPortal client?

  • BP: ideally the client already does this, but if it doesn’t, we’ll want it to

• DD: should be able to support benchmarking needs for batch processing; Ben and I embarking on this in the next few months

• TG: arbitrary lambda operations on data possible?

  • DD: would have to happen in the python layer on the server

    • perhaps submitted function could be dispatched to a worker process on the server, operations data local, then shipped by the server to client

• TG: richer filtering would be a higher priority than the above

QCSubmit for automated FF coverage improvement

Josh

• JH: Would be welcome to have this functionality included in QCSubmit

• DD: would like everyone to to chew on what would be needed for the feedback loop of automated FF coverage gap identification, torsion prioritization, submission generation over the next week

• JH: Is the feedback loop approach similar to something the ANI folks used (including automated dataset selection)? Isn’t this called “active learning”?

• DD: I can reach out to Adrian Roitberg and friends for advice on their approach.

  • even if we partially achieve an autonomous (with human approval) feedback loop, we will have also improved much of our data access and benchmarking infrastructure to support the decision-making access patterns required

  • we already have large candidate datasets to draw samples from

  • we already have good submission automation tooling via QCSubmit