Not much on the coding side. Evaluator can now handle fitting based on handler attributes (like 1-4 scale factor). This should be useful for JHorton’s custom potentials. ForceBalance now has PRs ready to ahndle some other fitting stuff.
Did some pytorch testing, largely to give MT feedback. Includes a continuously differentiable way to convert between cartesian and internal coordinates. So this should be able to do fitting to hessians (defined in cartesian) using using something like pytorch.
MT – Open question re: whether I should be making an interface to stuff that smirnoffee does well, or if I shouldn’t try to replicate smirnoffee. I’ve bee thinking about this a bit the last few weeks and was uncertain if that’s worth it.
SB – Good question. smirnoffee is about playing with vectorized representations of a system.
MT
OpenMM
Experimenting with OMM as an interoperability layer. It has a ton of readers that are similar to parmed. I found it difficult to get simple data out - Getting the charge of a particle could require looking through a variety of forces
OpenMM + ParmEd
ParmEd and OpenMM were briefly incompatible. This has been resolved now.
GROMOS
Chatted with Sereina Riniker lab. Some confusion about scope - They’re using OpenFF + GROMOS, and want more (this ecosystem is different from GROMACS, and so support isn’t on our roadmap)
Interchange + ParmEd/OpenMM improvements
For some systems, Interchange can go both to and from OpenMM.
Opened code cleanup PRs to the toolkit - Lots of debt that I’ve been looking at resolving.
CC
Worked a lot on LiveCOMS review
Started looking at how OpenFF forcefields perform on DCerutti’s torsiondrive datasets. I was able to set up an optimization in forcebalance without running any cycles, and using that to probe FF quality. Still working on results, it seems to do better on phi than psi.
JW – Benchmarking infrastructure could be helpful here, though it would have trouble extracting torsiondrives
SB – There’s probably good code in openff-sage repo, in the benchmarks folder, to do specifically this. I can point you to this if there’s trouble finding this.
LW – Scripts in openff-sage that would be good for comparing charges?
SB – I think the code for benchmarking transfer free energies/hydration free energies could be good here. This will be missing the MNSol data but I can send that to you directly.
JW
Was out last week. Did some work on biopolymer stuff, but didn’t push to GH.
LW
Rolling back to fundamentals and trying to think of a way to compare different charge models. Tried looking at AM1 antechamber vs. openeye ELF10 charges look, and how charges change in small polypeptides. Will present a kinda comprehensive overview on Thursday.
PB
Touched base with TFox on his reports regarding problem chemistries. I attempted to reproduce his described problems by crafting molecules with the features he described and managed to find some failing cases. Some of these are resolved in Sage. https://openforcefield.atlassian.net/wiki/spaces/~426615872/pages/2089484385
Found a probelm with IC hessian fitting, worked with JHorton to resolve.
SB – What was the ultimate issue?
PB – After doing an MD run, some coordinates were not updated properly. Then there was a factor-of-10 error in the coords
Started writing up theory benchmark
Started looking at using MD trajectories to find diverse conformers
Worked with Mobley lab grad student on running Evaluator.