2022-09-14 Thompson/Wagner Check-in meeting notes

General updates

• JW – Followup workshop announcements made - How can I help preparations?

  • MT – Not sure how to fill 3 hours. Plan to steer away from machine learning stuff - Basically just want to communicate that you can get matrix representations out of this without talking specifically about JAX. Want to avoid talking about ForceBalance replacement or ML force fields.

  • MT – My understanding is that the greatest interest will be around getting protein-ligand simulations to run. So I can run over something like the toolkit showcase. Not sure if the best idea is to cover a lot of topics in a little depth, or to talk about protein-ligand stuff in a lot of depth.

  • JW – Toy FF training example is pretty cool, though I agree it risks the workshop falling into ML gulch

    • MT – I didn’t totally understand the inner workings of the previous example, so I’d be concerned if people followed up with questions.

    • JW – Could be cool to use something simple like scipy.optimize, jax is probably too deep.

    • MT – I’d strongly prefer that we don’t have industry people using our jax infrastructure - I wouldn’t think that this is in the top tier of valuable things for us to present.

  • Good topics

    • Parameter inspection + manipulation

      • Are there cases where we can show parameter manipulation?

        • Hard to define because the only thing that should be manipulating parameters is a next-gen fitting engine, but it doesn’t exist.

      • Are there cases where we can show parameter inspection?

        • Could get a molecule that improves a lot due to a parameter split between openff-1.2, 1.3, and 2.0, and show it minimzing correctly in the later ones, and inspecting the parameter ID and values where it got better.

    • Component combination

    • Different format exports

    • Nascent import

    • Vsite shenanigans - Vsite on sulfur

  • Could also be good to have cookbook examples that have interchange solving common problems

    • Component combination - Components need to be assigned using FFs that can’t coexist, so we make them separately and then put them together -

      • Maybe trying to use a flexible water model at the same time as Sage?

      • Ligand-in-water example - Since vsites now don’t live on molecules, how about we show the ligand-in-water example, but switching out water models for TIP3/4/5P?

    • Residue confusion resolution - Like, if parameterizing a protein-ligand system, filling in residue info for the ligand so it doesn’t get auto-set.

    • Code block that can solvate 90% of PL complex structures from protein-ligand-benchmarks repo (code from toolkit showcase may already do this and can be copy-pasted)

• JW – PDB loading chat with BRies and RGowers - New slack channel on OpenFE - Interested in being involved?

  • MT – No, just interested in staying in the loop.

  • JW – Basically:

    • Due to a misunderstanding several years ago, Chodera and Mobley told Peter that OpenMM could be natively interoperable with OpenFF if it stored bond orders. However, they didn’t yet realize that it would need formal charges. So Peter added a field for bond orders but not formal charges to the Topology class, and the work never progressed past that.

    • OpenFE is likely going to vendor OpenMM’s PDBFile and update the PDB reader to assign formal charge.

    • OpenFE is a bit resistant to upstreaming this and importing it from OPenMM, both because of the friction if changing the OpenMM API and because of the additioanl OpenMM dependencies.

    •  

• MT – Met with MHenry, things are progressing along previously discussed lines toward typing and unit serialization. They’re more cavalier than us with importing code from small projects. Makes me a bit uncomfortable but we can let them lead the way and see how it goes.

Trello update

(Done)

Issue/PR resolution

• No-Radicals PR

  • JW – I started working through test failures on this, got into a rabbit hole. Not sure what’s going on, things are making an assumption somewhere and I’m not sure if it’s a technical or fundamental problem.

• MT – I’d like tests to run faster - The big molecule test sets are really frustrating.

  • JW – I’d be happy to have the big-molecule-test-sets in openff to only be run when a PR is marked ready for review.

  • MT – This particularly a problem while I’m trying to fix openmmforcefields - So far I’ve probably spent more than 20 hours, less than 50. The CI takes an hour to get back to me, and I while I can work on other things in that time I can’t focus deeply, which is preventing me from doing certain tasks that are normally my job. The openff toolkit update portion of this was likely a single digit number of hours.

• JW – I’d like to force through the new forcebalance release but we probably don’t have time today.

  • MT – This is complex because of interdependencies - Bespokefit needs forcebalance needs evaluator.

  • JW – Ok, I’ll be pushing on maintainers to review PRs (or will do it myself) to get the main branches ready.

• MT – Looking forward to evaluator release, ETA next two days. Some bumpiness around release - confusion about when/where PRs were merged and an accidental c-f build on the main label.

  • MT – Next big problem for Evaluator will be YANK. But I won’t take action on that for the moment. Also troublesome that it’s not compatible with python 3.10. Should be a really hyper-simple fix - just a collections.abc update. But when MH tried to fix that, JC asked for several more fixes and the PR stalled.

  • JW – I’ll keep prodding JC about getting more attention on it. I’ll also ask Lily about YANK’s position in our future - Will we continue to need it? Do we know of a replacement?

•