2023-10-05 Protein FF meeting note

NMR fit

Chapin Cavender

• Slides will go here

• Cross-validation study of regularization parameter alpha shows good performance with alpha ~ 1

• Also find improved performance using regularized NMR data vs just QM

• PB: Are you fitting all parameters?

  • CC: No, only fitting N_param parameters, includes number of cosine terms in torsion. Basically optimizing 6 parameter types, with between 2-4 cosine terms each. Correspond to phi/psi in 3 different peptide groups.

  • Don’t want parameters that are trained to peptide NMR to affect small molecule fits, so have created new params even in Null FF for backbone.

• JW: For black rows (NMR fit--slide 12), most chi^2 are in the range of 0.7-0.8, is that difference important?

  • CC: chi^2 is basically first term of loss function, square error from right answer divided by variance in right answer. Would expect chi^2 to be about 1 for a “good” model, anything less than that means you’re matching the reference well but the reference is imprecise, doesn’t matter how much less than 1 you are

  • CC: Based on that, don’t really have a way to decide between these 4 NMR fit force fields.

• RMSE from QM torsion drives agree about as well as non-NMR fit FF’s in most cases

• Free energy for beta → alpha generally still overestimated by NMR-trained FF

• Re-sampling leads to worse chi^2 vs expt, except specific OPC3

• DM: Strategy did work in at least one case, not a total miss on this strategy

• CC: Decision now is: is specific OPC3 good enough to move forward with larger protein benchmarks

  • DM: if it doesn’t take too much time, just go ahead with benchmarks, and at the same time try to get alpha-beta difference down while waiting. Maybe specific OPC3 will be good enough once we see results

• MG: Not clear how much this measure corresponds to what we’re looking for. By these metrics we’re already beating AMBER, why did we think that improving these metrics would improve performance?

  • See recording around ~29 mins

  • CC: Free energy beta → alpha, we’re overestimated compared to ff14sb by 1-2 kcal/mol. Having better chi^2 isn’t everything

  • MG: Does NMR data have to do with alpha helix issue?

    • CC: Peptides do sometimes have alpha helix structures, but don’t spend a lot of time there

    • MG: Is it possible to do NMR refitting but with larger proteins that do spend more time in alpha helix?

    • CC: Yes, that’s the next step if this doesn’t work

• JW: TIP3P-FB looks good for most of the metrics except the alpha helix, important to look at all metrics

• MG: Tough because we’re fitting very small differences in functions