Roundtable updates JW – Worked through a little bit of PR backlog. Initial Amber FF release. Preparing for interoperable molecule. SB – Would be good to make UML diagram of new molecule class (maybe UML diagram, could have design meeting based on that). I have some comments about aspects of the pydantic design – Specifically around units ( handling multiple unit registries in different areas of our codebases). Also there’s different options about serialziing units into JSON and validating them. MT – This seems to have three faces that may be separable:
PB – Started working on issue #502 (Writing Topology out to PDB) – offtop.to_file('test.pdb', positions, file_format='PDB') DD – Lots of planning for benchmarking, especially review of current tooling with TG. Focusing on data access layer around QCA – Looking to be QCSubmit for the time being. Concerned about slow data access. Also chatting with stakeholders to help plan. Value stream diagram. Supporting DR and JM’s QCA use. Would like to re-align priorities – Focus on QCA and adjacent areas, and distributed computation. Would like to resolve benchmarking “owner”. MT – Worked on psi4/OpenFF package incompatibility – Figured out where the conflicting pins/dependencies were on mac – Psi4 pins an outdated hdf5. This will be resolved when psi4 does a new mac build. Psi4 will have trouble due to some of their dependencies like their preference for OpenBLAS (numpy needs to connect to a BLAS library – Could use MKL or Intel – numpy needs thread safety, which MKL guarantees). Generally, the issue is that psi4 is very conservative with pinning. So the upcoming release will smooth things out, but if they continue being slow on stable releases, we’ll start having pinning problems again. DD – Can confirm. But using dev makes it hard to maintain provenance. But our QCF workers pin to a psi4 dev version. We should update our psi4-dependent build to install from dev
Worked on atom/bond.is_in_ring – Ended up being a rabbit hole, but I’ve opened a PR with the most likely functionality. Chargeincrement N-1 issue discussion – Opened PR to implement. Lots of interoperability work/discussion. Had initial meeting with Vanderbilt on Friday – Wanted to steer away from “compromise” object or “separate but interconvertible” and toward “just use OpenFF System”. Collaboration seemed to move in this direction, and leadership was interested in this route. Engineering meeting coming soon. Looking to break this down into discrete tasks that we can track individually. Also started researching InterMol / how much we want to resuscitate it / what we can reuse from it. Has a bad/nonexistent core object model, but a lot of readers/writers and single-point energy functionality.
SB – Fixed bug in recharge where implicit solvent calcs weren’t converging because of a grid issue. Looking at rewriting data models to use pydantic. Rewriting how gradients are evaluated in Evaluator to slot in more nicely with timemachine. JW – How does timemachine fit into our future? SB – I’ve heard a lot about how it’ll tie in, but it never seems ready. I’d like to use it for phys prop simulation but it doesn’t support PME. JW –
DH – Working on many smaller things – Conf benchmark repo with Lim. Protein-ligand benchmark calcs Call with Vytas – He just ran a big benchmark with OpenFF and other FFs. Looks like we’re comparable with GAFF and CGenFF. Some cases of comparable performance with FEP+ Last week, joined benchmarking call with DD and pharma partners. Have a call with Gary and others to show how to deploy.
Holiday schedule JW – Week of thanksgiving, at least two weeks in december, contiguous with Christmas DD – Unknown, will update later. MT – Unknown, will update later. DH – At least two weeks contiguous with christmas. Some scattered other days. SB – Unknown, will update later. PB – Unknown, will update later.
We will update the OpenFF calendar as we figure out holiday schedules Benchmarking planning JW – Looking forward to using abstract molecule dataset class SB – Nonbonded has some infrastructure that looks similar to what JW is talking about. Also think carefully about dataset classes, since one-size-fits-all is likely to be painful if you’re not careful. If you’re planning on this going to databases, think about what shape it should take, and how that’s influenced by choice of SQL/noSQL. JW – Timeline-wise, we might see how well our plans here align with the deliverables for the industry benchmarking push. If they’re separate that’s OK. DD – It will likely be the case that immediate needs will be largely based on QCSubmit. So current analysis code would go into QCSubmit. DD – Will host another round of design review in the future. Do we want to make a wholly separate package for benchmarking or bundle everything into QCSubmit?
JW – Janssen asked for a single point of contact, and I didn’t want to identify a single team member since we’re all oversubscribed DD – I can take this on. Many of these areas are aligned. JW – Let me know if you need help prioritizing – This is more important than QCA dataset submission and evaluator distribution. Also let me know if you’re being tasked with more than your current commitment. Decision – DD will become the point of contact for industry benchmarking
0.7.2 release? SB – Would really like chargeincrement changes in this if it happens MT – We have two options here: Could cut 0.7.2 immediately and then cut 0.7.3 when CIMH N-1 PR is merged Could cut this release once CIMH N-1 is merged, but likely that this will end up being a week or two, and then we’ll struggle with timing vs. vsites.
JW – If we get release automation oiled and working well in a 0.7.2 release, then we could do nightly builds Decision – We’ll cut 0.7.2 immediately, then JW will take over on CIMH N-1 PR and push it to completion ASAP for 0.7.3 release
|