2023-11-02 Protein FF meeting note

NMR fit

• JC (slide 18): which experiment are these points arising from? Are the simulations also done at 274K?

  • CC: Yes

  • CC: in general, easier to fit the computed dihedrals requirement than the H-bonds.

  • MS: which should match the experimental values more?

  • JC: neither of the dihedral or H-bond measures would correlate smoothly with the chemical shifts

  • MS: where does your exp data come from?

  • CC: in the NMR experiment what they report is the metric of helicity here. Raw chemical shifts aren’t given in the paper

  • MS: your NMR simulations don’t have C=O shift? Perhaps that would be interesting as a measure of the fhelix?

  • CC: Yes, I can do that

• KT (slides 18-21):

  • How long are each simulation?

  • CC: 2 us right now, would like to get them out to 10. We’d expect if a peptide is going from an extended to helical region, it would pass through delta, as Espaloma is doing here

• MS (slide 23): AF generated some data on GB3

  • H-bond Scalar Couplings:
    O.135 = ff14sb
    O.215 = Espaloma

    Scalar Couplings:
    1.009 = Espaloma
    0.782 = ff14sb
    0.782 = ff14sb

  • 

• DLM: have you looked at differences between Espaloma assignment and your force fields?

  • CC: yes, I was going to propose this as a next direction

  • DLM: could be typing, could be you’re stuck near a bad minimum

  • JC: wouldn’t bespokefit be a better comparison, of a helical fragment? That could tell you if it’s the torsions that are the problem, or LJ? There’s no guarantee Espaloma is correct

  • CC: I’m not certain our fits to QM data are giving us what we need

  • MG: but training more carefully to local QM may not give us what we need to solve this problem

  • CC: if we follow the Stonybrook line of FFs, they found that training to gas-phase QM did not help condensed-phase peptides very well, and they had to fit to peptide data. In ff19sb they didn’t have to fit to NMR data, but they had CMAPs and continuum solvent QM

• KT: …

  • CC: I’m showing that resampling works well for short peptides, but we don’t get as good performance on larger ones

• MT (in chat): Naive question … are torsion parameters the only ones being re-fit? And, if so, how much improvement over these long-range properties can be obtained without also fitting vdW parameters? (One could ask the same question about this work, the Stony Brook work, and maybe other efforts)

  • CC: yes, only 6 torsion types.

  • CC: I did run an experiment with NB parameters from AMBER (but SMIRNOFF valence) and didn’t observe differences in behaviour, so not expecting this to lie with NBs

  • MT: What if you refit it?

  • CC: MG and I have discussed this. It is a non-trivial amount of work and it seems that other FFs have been able to solve this problem without fitting vdW parameters to peptide data.

  • MT: …

  • DM: the AMBER people built their FFs without re-fitting LJs in that context

  • CC: only exception is ipolq style FFs

• JC which dataset are you using for the (slide 33):

  • CC: Kyle Beauchamp, Pande, …?

  • JC: did you look at https://pubs.acs.org/doi/full/10.1021/ja0660406 ?

  • CC: I’m using Ala3 to Ala5 from this paper

  • CC: I could add Ala6 and Ala7

  • MG: any idea how helical you’d expect them to be?

  • JC: conclusion is that Rama populations are very similar from Ala3 to Ala7

  • MS: Pure ala is not that helical

• MS: how different are the torsiondrives of ff14sb and proto-Rosemary? Not the parameters but the potentials, but it would include the NB interactions too

  • …

  • (around 43 min into recording)

• CC: things I’m doing to move forward:

  • Benchmarking the experiment where I train just to Ala5

  • Taking previous version of NMR fit and iterating on it until parameters stop changing

NIH resubmission

• MS: suggests showing peptide NMR data for Espaloma and some selection of FFs (maybe null)

  • MG: as long as it’s not intellectually dishonest or misleading. Fitting to NMR led to improvements on peptides. We can say we’re working on benchmarking on full-size proteins

  • CC: (shows slide 16) is this what you’re after?

    • MS: yes, let’s simplify, either 1 or 2 FFs

    • MG: don’t need so many water models

    • CC: … suggests FF selection

    • MS: Null is almost as good as specific and is the more general FF

    • MG: we could comment on generalisability

Espaloma manuscript

• Tabled for next time