2024-05-30 Protein FF meeting note

Fits with alternate nonbonded parameters

Chapin Cavender

• CC will link slides here

• Slide 3: can compare objective functions with Null-0.0.3-SP

• JW: is ff99sb the same as 14sb NB parameters?

  • CC: Yes, they haven’t been changed since ff99

• JW – Kinda a bummer that the SP line of fits didn’t work

• LW – Might have missed this, but did Null-0.0.3-SP also maintain GB3 structure in benchmarks? Could the failure of the QAmb/NBAmb/NAGL NBs be more related to the SP protocol than the NB parameters?

  • CC – Yes. Null 003 SP shows unfolding of alpha helix in all water models. So it’s possible that the SP protocol is insufficient in all cases.

  • CC – So one idea is to redo all these fits using the full minimization routine. But that would take weeks for each FF. Instead, I could use a modified objective function that compares pairwise differences.

• JW – Why not just use the FF14SB objective function?

  • CC – That objective function was just used for sidechain optimization. Also it makes sense here to use RMSE instead of MAE, and use weights based on barrier height.

  • MG – Could do multiple things at the same time? But start with the one at the bottom of slide 15

  • CC – Yeah, would just take more of my time.

• JW – Possible that NAGL is doing something funny with larger molecules?

  • CC – Not sure if there’s a good way to check that other than what I’m doing.

  • DM – Could look at a by-residue charge difference as a function of sequence length

  • LW – I looked at 5-15 AA, but very few 15-AA-long chains due to the need to compare to OpenEye/AmberTools! We could do some more comparisons of larger proteins just comparing to LibraryCharges too.

  • CC – I can look into this. Based on these charts it’s not clear that NAGL is doing dramatically worse than anything else.

• CC – General consensus that next steps are to try modified objective function?

  • MG + JW – Yes

• LW – Just going back to the human vs computer time of redoing a NB fit with the minimisation protocol, is this something more compute would help with? e.g. would it be helpful to try doing a minimisation run on UCI resources in the background while Chapin focuses on other directions first?

  • CC – Yes, this would help, I could run the expensive minimization fits on UCI if that makes sense for everyone

Supplementing training dataset with PDB structures

Anika Friedman

• AF will link slides here

• MG – How many 4-mers are there in our original dataset?

  • AF – 30kish

  • CC – It’s 70 torsion scans, most of which are 2d with 24 points in each dimension, so 578. 70 x 578 = 35kish

  • MG – Another thing we could do is to take away some of the other ones.

  • CC – I think the overweighed bins in delta and beta are where I fixed the backbone to do sidechain scans.

  • MG –

  • CC – For the sidechain torsion scans, only the sidechain torsions should be allowed to vary, since the selection of parameters is based on which dihedral was driven.

  • MG – Do we know, if you’re fitting a sidechain torison

  • CC – Roughly 1/3rd of the 35kish confs are from backbone scan

  • AF – Maybe we should remove sidechain torsiondrives from this analysis if they’re not being used to inform backbone parameters.

  • MG – Another thing is, does it matter if we’re optimizing SC params in the presence of a alpha/delta vs. beta backbone?

  • CC – We did the sidechain scans in the context of two backbone confs to avoid biasing toward one, but also without needing to do a 4D scan.

• CC – I think we don’t want to have the sidechain scans in there, are we just want to look at the prevalence of the backbone scans.

• MG – It’s also worth considering how to mix in the new peptide structures.

• JW – How confident are we that the sidechain scans only affect the sidechain fits, and vice versa with backbones?

  • CC – About 50%, still need to look into this.

  • JW – Might be worth throwing out the sidechain scans and fits altogether, and just doing backbone fits to backbone scans.

  • LW – I hope I was getting the gist of the previous discussion correctly… but I don’t think there’s anything preventing side chain torsion drives from affecting backbone parameters and vice versa in ForceBalance — would have to check by looking at the code.

• AF – If the backbone torsions are just being optimized to backbone scans, then … the weighting of the fit could be thrown off by having lots of points in a certian region of backbone angle space (the sidechain scans)

  • …

  • MG – Maybe the sidechain scans should be redone sampling alpha helical backbone angles?

• MG – Could … compare more to backbone population distibution.

• AF – CC and MG, let’s schedule a call to discuss how to handle this in the future.