Page Comparison

Peptide and GB3 benchmarks in OPC water

Chapin Cavender

• Slides will be uploaded

• Meeting is recorded

• Slide 7: error bar is confidence interval that comes from standard error

  • DLM – evidence seems to support need for us to co-optimise our own water model

• Slide 8: aside from Val3, OPC generally improves results for SMIRNOFF

• Slide 9:

  • MS: SMIRNOFF RMSDs in GB3 look higher than 14SB

    • CC: they are slightly (2.0 vs 1.7), but not clear to me that’s a problem

• Slide 10:

  • CC: no unfolding yet at 3 us – in TIP3P these unfolded at 2 us, so cautiously optimistic

    • MS: points out that one replicate is going up a bit in RMSD

    • CC: we’ll run it for longer

• WS: this looks great. TIP3P has an unrealistically large diffusion constant – I’ll bet OPC is much closer to experiment, so maybe kinetics are slower, so anything crazy might take longer to happen

  • MS: confirms diffusion constant is lower

• LPW: what are you looking at?

  • CC: I’m looking at scalar couplings:

    • 3-bond Karplus

    • interresidue scalar coupling between amide NH to C=O of another. On Slide 12 I’m showing backbone dihedrals (1-3), side-chain dihedrals (3-6), last is inter-residue H-bond

    • LPW: I’ve done these with RMSEs in hertz. I haven’t done chi squared, and I don’t know how to convert to RMSE, so I’m not sure how to compare our results. (shares results)

    • CC: I should be able to convert my chi2 values into RMSE but I don’t know the number off the top of my head

    • LPW: I’m curious about the spread of values – high chi2 suggests to me that it is, but my results should have a chi2 of <= 1

    • MS: is the question here if the J-couplings differences are small as you’d expect?

    • BS: you showed experimental values

    • CC: the number I’m using for exp uncertainty is the RMSE between parameterised Karplus couplings and what they use to do the parameterisations… they’re generally around 0.25-0.6 hz. This also depends on which set of Karplus eqns or coefficients you use. The BBs have multiple options. Do you recall what you used LPW?

    • LPW – it’s in the paper, not quite sure now (paper from 2017). (checks paper) I’m using Schmidt et al. for backbones and Perez et al. for sidechains which should have RMSEs between 0.6 and 1, but Vuister et al. had similar RMSEs

    • CC – I’m using Perez et al. for sidechains and Vogeli et al. for backbones

• CC: recommend using OPC from now on

  • MS: what about tip4p-fb and other four-point models? Others may also balk at extra computational cost

  • CC: good point

  • MS: want to do subset of mixture properties with a bunch of water models. OM is gone, LW has other work to do, maybe I can find someone in my group to do it. e.g. Anika

  • BS: agree that’s a great idea

  • MS: enthalpies of mixing, densities that involve water, transfer solvation energies. Maybe can coordinate with LW to see how much there is to do. MS and Lily can coordinate

  • CC: can repeat with tip4p-fb since we already have the port in SMIRNOFF

  • LPW: the parameters between OPC and TIP4P-FB are very different. Similarly, OPC3 vs TIP3P-FB are very similar.

  • LPW: could see if you could reproduce tipxp-fb in evaluator with tipxp data set, then add mixture properties

  • MS: I think we can test water models before we re-fit models

  • JW: wrt virtual sites, OpenFE may not have infrastructure to roll out virtual sites for a while. We may not be able to easily do FE calcs with four-site models.

  • CC: we thought we would wait until release candidate to do protein-ligand benchmarks, which will be when I hand off to Lily for refit

  • MS: we don’t want to do PLBs until we are comfortable about protein stability

  • JW: i’ll keep them up to date and ask when they may have vsites implemented

  • BS: I thought OpenFE discussion was to do with vsites on ligands, not water model

  • JW: I think water model might be a bit simpler. In my understanding, the difficulty is with alchemical transformation bookkeeping.

  • MS: if vsite has no charge on it, the statistical mechanics are identical as if there’s no vsite

  • DLM: from OpenMM POV, it should just be a matter of running with the water model, whichever one it is. Unless they’ve done something really weird, it should be quite simple

  • MS: since water model isn’t changing, yes, it should be fairly simple

  • BS: agree

  • JW: I’m just being cautious in case OpenFE can’t handle it

  • DLM: worth putting on their radar

  • (general agreement 4-site water should be straightforward)

• CC: summarise to do:

  • Run benchmarks with tip4p-fb

  • MS will see if someone can run small molecule properties

    • MS: won’t be before OpenFF meeting, so let’s nail the plans down at the meeting in a week. We can aim for results at end of May. Maybe Tim could do it

    • CC: we can use these results to decide further protein benchmarks

    • DLM: can use agenda on Monday to figure this out

  • MS will sync up with Lily to figure out what properties to use

  • CC will sync up with LPW about couplings

• Update on AF:

  • She’s expanding dataset to look at more protein systems

  • CC – I’ll continue running other water models and new observables for folded proteins

• MS: on Monday is when we have updated NIH information

  • MS: we can put a list of accepted publications. Anything that’s been accepted in the past two months?

    • Sage

    • Owen’s surrogate model

  • MS: covid regulations still in place, so we can submit up to 1 page of preliminary data

    • Since our results are middling, not sure this is the best idea. Should we try to show tentative results?

      • DLM: I don’t think I’ve ever submitted anything

    • Consensus: don’t submit tentative results

• CC: thanks JW and MT for helping with OPC!

• MT: TIP4P (EW and FB variants) are not in the forcefields repo but the work is done, just needs a review/merge/release when that’s important