2023-07-27 Protein FF meeting note

Protein FF refit

Chapin Cavender

• Chapin will link slides here

• Slide 26

  • JW – Has “Null” on this slide been retrained with the new points? Or is it just being evaluated?

    • CC – No, it’s the old Null FF being rescored.

• Slide 27

  • MS – Where did new fits start?

    • CC – Started at previous fitting result. It’s taking about as long as an entirely new refit.

  • LW – Are individual steps taking longer, or are more steps being taken?

    • CC – More steps are being taken - I’m still seeing the objective function dropping a lot at each step.

  • CC – A little concerned that we may be overweighting the new points. I’ll have to see how the new FF performs. But if the new data is overweighted we may need to do this again.

  • PB – Could you also change the magnitude of the priors/change the starting values to the small mol FF?

    • CC – Worth considering - If we have to do another round we could look at that.

  • PB – What are the relative weights of the targets?

    • I’ve weighted protomers to they add up to the same as a single mol. And I’ve made the protein torsiondrives have the same total weight as the previous torsiondrives.

    • PB – Between abinitiotarget and torsiondrivetarget, whata re the relative weights/

    • CC – TDs are weighted between 13 and 80 depending on number of targets, and abinitio are weighted around 4 (since there’s more of them). … (see recording ~22 mins)

  • PB – …

    • CC – Yeah, we’re expecting that the problem is in the torsions, not the bonds and angles.

  • CC – For next steps, I could remake this scatter plot using an intermediate FF and make sure they’re looking good. This would help us ensure that we’re not finding other spurious minima.

  • CC – Once this converges, we can run peptide benchmarks. Then I’ll pass the force fields on to Anika if they look good.

  • PB – CC, do you have compute to do simultaneous iterations?

    • CC – I can really only do one fit at a time on TSCC. I could spin up more fits on the UCI cluster if needed.

  • LW – Is the new fit using Null or Specific?

    • CC – Null. I can spin up a new fit on UCI using specific.

  • LM – Difference between Null and Specific?

    • CC – Null has the same torsions as Sage. Specific has special new torsions that should only hit proteins.

  • PB – For abinitiotargets, are you matching both energies and forces?

    • CC – Just energies

Benchmarking w/ GROMACS Quick Update

Anika Friedman

• AF – Background: Currently all benchmark sims are run in OpenMM. But we have access to lots of CPUs, which GROMACS can take better advantage of. So right now we’re running everything on GROMACS using ff14 and a few water models. I’ll have the control data by next week and plan to present at the next meeting. This will let us evaluate whether the observables come out within error using the two engines.

• LW – Could you remind me which water models?

  • AF – TIP3P, TIP3P-FB, OPC3,

• MS – Tim Bernat will be running density and mixture calcs for data that run into the sage fit. That’s been delayed by a week or two but we’re looking forward to starting those. So that can guide us to which 3 point water model to optimize for.

• CC – AF, do you think you’re ready to make a pull request for your branch?

  • AF – Working on some tests, will talk to you via DM to coordinate the PR and formatting review.

  • JW – Please loop me in too - Happy to check the test format and stuff

  • AF – Also we’re seeing better than anticipated performance so we’re not using all the CPU time. We’re using HMR though so I hope that we get similar results.

• JW – Lots of enthusiasm at CADD GRC for RNA FFs, and people trust OpenFF

  • MS – It’d be good to channel that into letters of support.

  • JW – I’ll reach out to folks and loop you into a conversation with the mRNA-modeling GRC attendees.

• MS – Also, we’ll want to have espaloma proteins tested the same way. I saw there’s a discussion on CC’s repo for doing that.

  • CC – I recall that discussion, I think we have a plan laid out, just needs to be executed.

  • JW – Could a chodera lab person do that?

  • MS – I’ll start a discussion to ask if they’ll do it.