2023-03-23 Protein FF meeting note

Modeling of amide torsions

Chapin Cavender

• Slides will be uploaded

• CC (slide 20) – Hypothesis: energy mismatch between QM and SMIRNOFF profiles arises from different charges. In training data, charges come from OpenEye ELF10. In test data, charges come from LibraryCharges

  • PB – did you

  • LW – did you compare charges between the LibraryCharges and OpenEye?

    • CC – they are more different to AMBER 14SB than to OpenEye. I can have a look at the magnitudes of the specific charges

• CC – are we worried about these energy profiles as a problem? If so, we could re-fit the FF and include the current test data as training data

  • MG – could it be the Lennard-Jones? How much data does this add? Would it be enough to move the dial?

    • CC – I have omega scans for 24 peptides, flanked by alanine. I had 100 small molecule amides previously.

  • TG – 2 questions:

    • Is attenuation used in any of these?

      • CC – yes, v0.0.2 has higher attenuation weights. But doesn’t seem to change results much

    • What are the deviations in values, since you’re using a regulariser? Maybe the regulariser is preventing the k value from getting the right magnitudes?

      • CC – Can get barrier heights well on small molecules.

      • TG – if you have to change the k value a lot, the penalty will scale quadratically, so starting point matters. Curious about deviations in k values.

      • MG – Clarification?

        • TG – Improvement depends on penalty of changing k value, so initial conditions matter a lot. Would want to see if the penalty at the end is quite large.

      • MG – is this getting into “wizard” territory where we set priors in an ad-hoc way?

        • DM – instead of wizarding on parameters, we’re wizarding on where we’ve gone wrong in fitting

        • TG – in true wizard land, sometimes we turn the regulariser off to see where we end up

        • MG – in theory I’d be happy with having the regulariser off. Let the data speak for itself

        • TG – PB is working on values where the starting point is the MSM value. That’s not being incorporated here, right?

        • CC – yes

        • PB – the starting point Sage-CC was initialised with MSM values (although final values are the result of the fit)

      • PB – what were the initial force constants for torsion parameters?

        • CC – final values of Sage-CC

        • PB – I mean for the new parameters you introduced

        • CC – I haven’t introduced any new values for the omega torsion. It’s all one parameter

• CC – let me summarise discussion outcomes and to dos

  • Look at magnitude of charge differences between ELF10 and LibraryCharges

  • Look at m values from ForceBalance for proper torsion amplitude

  • If those diagnostics point to the problem being with the torsion parameter, I’ll do a new fit targeting the validation data I’m using here

    • MG – if it’s overly restrained to the wrong value, wouldn’t it be a good idea to relax this and re-fit? I’m not disagreeing with you, just thinking aloud

    • CC – Yes, that’s a more precise statement of what i said

    • MG – Ok, so if it turns out to be overly restrained, you’ll re-fit without additional data?

    • CC – yes

    • MG – and if charges are significantly different?

    • CC – that tells me that having one parameter will not be enough, I’ll split it off into a different parameter just for proteins

    • MG – charges may be different but not responsible. Could you just look at the electrostatic component of the energy to see if it’s in the right direction to be the problem? It’s conceivable they could even be helping you, or only wrong by a fraction of kcal/mol.

    • CC – yes

    • PB – could you pick one or two torsions from this set and just see how much the k values change from changing the priors?

      • CC – yes

      • TG – I did something similar and the solution wound up being adding a bunch of periodicities. Maybe we can compare our k values for omega torsions and see if they’re similar.

      • CC – sounds good

NMR benchmarks for short peptides

Chapin Cavender

• CC – Tier 1 can be done in a day

• CC – Tier 2 will take 20 days to run

GB3 trajectories

Chapin Cavender

• CC – in v0.0.1, an alpha helix unfolds in a trajectory

• CC – same in v0.0.2

• CC – the unfolding happens around the C-terminus

• CC – I’m trying to diagnose the exact failure mode. I’m looking at populations of Ramachandran clusters in phi/psi angles, and clusters that show up in crystallographic data, and seeing how often simulations sample those clusters

  • MG – Could it be the omegas?

  • CC – I don’t see any omegas deviating from 180

  • TG – Could you look at the total torsion force on some of these torsions, and compare them to 14sb? We mix torsions in a way that could be surprising, so looking at the total torsion force we might see that it’s pushing it somewhere we don’t expect

  • TG – We want to look at the sum of torsion profiles for different torsions, e.g. phi/psi/omega. I’m not familiar with the space of parameters Sage assigns to these torsions. Sometimes we’re surprised a particular torsion gets applied to a peptide bond. For me, Sage looks like it’s about k = 10 for the torsion (sum of 2 torsions with k= 6 and k = 4).

  • TG – could tie into previous issues. e.g. omega issue is the barriers are low, maybe that’s a systemic issue.

  • MG – clarification about the energies to look at

  • TG – just the torsion energy

  • MG – ultimately its placement on the Ramachandran map is from the total energy

  • TG – yes, but torsion energy could identify outliers, or surprising torsions. What I’m saying is that we may be looking at one torsion term for diagnostics, but the problem may lie in another

  • TG – going back to FB, when looking at the mvals, you’ll want to look at all the mvals for the all the torsions

• CC – once we’re happy I’ll pass it off to Lily for larger benchmarks on lilac, we will iterate on a release candidate, and then I’ll pass the release candidate back off to Lily

• PB – why do you think Sage-CC is doing better than Specific on slide 6, with the geometry targets?

  • CC – We don’t see this happening on the ddE targets. I’m not very worried about this

  • PB – just wondering why adding more specific torsions degrades performance

  • CC – I added a bespoke parameter for phi/psi/chi1 but all the other parameters are shared. I tried to only add one bespoke parameter per rotatable bond. I also upweighted protein data to get these – maybe we should have the small molecule data be more dominant.