/
2025-01-22 FF fitting meeting

2025-01-22 FF fitting meeting

Participants

  • @Lily Wang

  • Julia Rice

  • Julianne Hoeflich

  • @Jeffrey Wagner

  • @David Mobley

  • @Chapin Cavender

  • @Pavan Behara

  • @Matt Thompson

  • @Jennifer A Clark

  • @Michael Shirts

Recording: https://us06web.zoom.us/rec/share/c2Gh0crZNlI7EWvoA5bf4GNjIUXbRs5ScNGS1WV5LD3Ytk-ZLaiMFdQmYFmyMVct.xNLvPdsCoNukXqkw

Passcode: 3=cBGL1w

Discussion topics

Item

Presenter

Notes

Item

Presenter

Notes

Lipid data set fitting

JH

  • MT (in chat): Haven’t gotten to it but the conversion from existing QCArchive datasets - either opt or torsion drive datasets - to YAMMBS/YDS is on my plate right now

  • MT (in chat): These refits are “only” valence parameters, correct?

    • JH: yes

    • DM: guessing that this might not really affect diffusion coefficients, is there any particular reason it would?

    • JH: there’s an AMBER paper from 2017 where the C-C-C angle was re-fit for more flexibility and that increased diffusion in lipid tails

    • DM: how big were you looking?

    • JH: pentane to pentadecane

  • DM: with your MSD plots, I wonder if OpenFF 2.0 non-bonded re-fits made things too sticky

    • LW – Possible, hard to say without doing a refit. JH, when you first precented this work, you showed using amber changes swapped in. Maybe try vdW params as well?

    • JH – Could do that, we could follow up offline.

    • DM – Could also try with openff-1.x vdw (which will have nonbonded params very similar to amber).

    • JH – PF and I haven’t run any calcs using Parsley. Is the big difference between Parsley and Sage the nonbonded params?

    • LW – Yes, it was a full refit, but the major change was in changing the nonbonded params.

    • DM – So if using Parsley makes the lipid properties look better, that would tell us that the Sage nonbonded refit messed things up, maybe by making them too sticky.

    •  

  • JW (in chat): Brent’s Lipidmaps dataset is >95% done, working through some (hopefully) technical issues now but we’re not sure how long it will take to finish. Probably mostly usable in its current state though (modulo needing to re-export it once we complete the remaining jobs)

    • JH – maybe the partial completion was the cause of my issues.

    • JW: I would be surprised if that was the problem, many of our historical datasets are <100% complete

    • JH – Would this be worthwhile to benchmark against?

    • LW – Some knowledge about the exact composition lost with BW. But this dataset is intended to contain at least fragments of important lipids. README scripts should help explain composition and purpose.

    • JH – Ok, I’ll check these out.

  • LW – TDs might be supported in YAMMBS directly now. Can discuss more offline.

    • JH – That’d be helpful. Are there other lipid-like TD datasets I could use for benchmarking? My only one right now is the training set.

    • LW – Not aware off the top of my head, it couldn’t hurt to check TD profiles against the training set anyway.

  • JR – So you have a charged headgroup and an alkane tail - Do you have counterions in simulations?

    • JH – The lipid I’m showing is POPC (zwitterionic - phosphate+amine), and we’re running some sims with cationic lipids (POPS) which will have counterions.

    •  

    •  

    •  

  •  

  •  

  •  

  •  

Evaluator pre-equilibration

LW

  • MS – Worth noting that triethanolamine has a huge viscosity, so this is a very challenging case.

    • LW – It is kinda a worst-case scenario, but other mols in our datasets have similar issues.

  • MS – Might some of these systems be phase-separating? Maybe packing at a higher density might avoid this issue by removing bubbles or something?

    • LW – This used 0.9 g/ml (the default). Note that the OpenFF packmol wrapper scales the density a little. We underpack a little to keep packmol from failing. Not sure whether MT looked into how well different settings would work.

      • MT – Not really.

  • PB – Can you check whether ex. GROMACS gives the same equilibration behavior? Could be further misbehavior by openMM’s barostat.

    • LW – Could check this. Does GROMACS have montecarlobarostat implemented?

    • MS – Pascal Merz and my PR to add it is still waiting.

    • DM – Different barostats may still give useful info - like if it’s an order of magnitude faster there may be something wrong wtih OpenMM’s barostat.

  •  

  •  

  •  

  •  

  •  

  •  

  •  

 

 

 

Action items

Decisions