Page Comparison

Sage-2.1.0-NAGL benchmark

Chapin Cavender

GB3 umbrella sampling

Chapin Cavender

• JW: Slide 10--why did you use OPC3?

  • CC--TIP3P is what ff14SB is trained with, but AMBER recommends “plug and play” e.g. use whatever water model works better. Only looked at OPC3 so far because it seems to perform better than TIP3P and we want to move away from TIP3P

• JW--Slide 13: Seems from this that this would be an issue with non-bonded interactions

  • DM/CC: Not sure you can tell that from this

  • CC: This could arise from torsion terms. Think of helix being made up of many little turns, if one turn is off by a 0.5 kcalmol, 15 turns would lead to ~7 kcal/mol error

  • MG: Also, we subbed in Amber nonbonded parameters [and re-trained] and still had the same issue

  • CC: Could still be some issues with nonbonded or other things to change, but based on literature it seems like issue with torsions and/or 1-4 parameters

• JW: slide 16--this will bias you toward a bin in native contact space?

  • CC--yes

• JW: slide 17--goal isn’t to take un-biased sim w/ one FF and re-weight for a new FF, instead it’s for umbrella sampling comparison?

  • CC: Say you have umbrella sampling with FF A, you change e.g. torsions and then want to benchmark new FF, can re-use old umbrella sampling but re-weight using info from new FF but don’t have to re-do umbrella sampling

  • MG: Also want to use it to find adjustment to torsions that will push us in the right diretion and improve frac in native state

• JW: how will you be able to validate the scoring of new FFs?

  • CC: validation would be estimate of observables

  • MG: wouldn’t the ultimate benchmark just be comparing to unweighted run?

  • DM: could compare to runs you’ve already done, e.g. re-weight an existing FF and compare to results of calc without bias

  • MG: going to try to keep compute time by focusing on tweaking torsions

• DM: how expensive is umbrella sampling

  • CC: can do it in 2 days, but need a lot of GPUs

  • DM: might be able to expedite by changing only valence params instead of other params.

  • CC: Not sure what you mean by that?
    DM: that’s ok, we can revisit it if needed

• JW: Love the idea, but just seems like a lot of room for indexing issues, that’s why I was asking

  • [more discussion in recording around 40 mins]

• DM: seems like you’re on to something, if you wind up blocked let us know and we can brainstorm

• JW: [chi2 plot, not in slides] Seems like variation at frac < 0.6 would be meaningful, e.g. something is off

  • MG/CC--agree

  • [disucssion around 45 mins]

• JW: We don’t have that much to run on alchemiscale and need new FFs to run on it. Maybe we could throw this on there. Could you pick one SMIRNOFF FF that should do well and one that should be bad and send to me?

  • CC: Not sure we have any that should do well…

  • JW: Could you send one that you know will be terrible, and one that will be middling

  • MG: Why?
    JW: to get protein-ligand binding benchmarks. partially for me to learn to do it and partially to get the benchmarks

  • CC: would it be with TIP3P?

  • JW: yeah I don’t think alchemiscale supports anythign else

  • CC: will send Null 0.0.2 family one which is doing the best, and FF doing the worst