2025-04-17 Protein FF meeting note

GB3 NMR fits

Open

@Chapin Cavender

• (slide 14) DM – Do you think the umbrella sampling protocol is just undersampling and not visiting some important confs?

  • CC – Yes, though note that we needed the umbrella sampled confs to get to the other FFs

  • DM – Is it a reasonable hypothesis to say that you’re using umbrella sampling to find unfolded states to help it learn to … so we shouldn’t trust the free energy curve coming out of umbrella sampling too much since it doesn’t sample enough in unfolded-land?

  • CC – Yes, I think one of the big issues is that we havent sampled unfolded space enough.

  • MS – It’s also likely that some confs around 0.7 and 0.8 are kinetically trapped so we’re not getting good sampling there as well. But unbiased sampling is giving us a lot of info about >0.8 region.

  • CC – Yeah, we might only really be thoroughly sampling space down to 0.925 or something.

  • MS – Yes, though this might be good enough to get our first protein FF, and then we can investigate doing better sampling later.

• (slide 15) JW – I’m wondering if a “strategy” for a FF to do well is to not stay in any basin and all and be totally structureless. So FFs with high barriers/torsion values all get bad chi2 here, and ones with weak torsions get good chi2. I’d be kinda curious what an all-0 torsion force field would look like here

  • CC – An unstructured FF wouldn’t score well, there’s a specific basin that these short peptides should fall into

• PB – Explain comments on FE curves again? Why is blue ff14SB lower at extremely high (>0.98) native contacts?

  • CC – Sampling is probably only thorough/adequate above 0.92, so our fits start showing really good free energy there. Our FFs are probably somewhat weak at high native contact because we can’t yet sample thoroughly above that high. (I think I transcribed this right, see 30 mins)

•  

•  

PDB 4-mer QM fits

@Chapin Cavender

• (slide 21) MG – Would it make sense to benchmark old FFs on 4-mers dataset? Could be a good reality check.

  • CC – Yes, that could be cool to do next. Somewhat cautious about interpretation, since the only thing we can compare to is the FF that we trained on.

• JW – Were 4-mers geometry optimized?

  • CC – Yes, with backbone dihedral restraints

• MT – Anticipating criticism for ad-hoc nature of the path taken to get to final parameters?

  • CC – Possibly, but it’s worth mentioning that since adding 4-mers is bringing down objectives for other confs, it looks like…

  • CC – Also we’re going to benchmark on QM again after we do the NMR fits, and we expect to see performance on QM remain good. So I think that will eliminate some of the criticism we might get.

  • …

  • MS – Yeah, possible we’re just picking wrong QM data.

  • MT (chat) – I ought to add I care less about referees and a little more about reproducibility of a hypothetical 3.0, in part given some issues we’ve have reproducing sage fits. Not a criticism, just trying to double check that this is being thought about (and it obviously is)

  • CC – Hoping that adding in 4mers like we’re doing here will make the QM optimizaiton land closer to the end point before we do the NMR fitting. Right now our path is doing a lot in NMR fitting, but with better QM from the start we might be better off.

  • MG – If there are enough parameter that anything we get to will have a risk of having a nonreproducible optimization path, unless it’s highly restrained against the original. so it might be unavoidable that things will depend on the starting point. Eg WW’s work on water models had very few parameters and still had trouble finding an optimum without human intervention. So I expect in this project too it’ll be hard to avoid history dependence.

  • CC – Right, kinda tiered goals here. Top goal is to make a protein FF at all, but after that it’s very important to do this in a reproducible and iterable way

• JW – Would even more 4-mer data help? Could test that by throwing out half the 4-mer data and see if fit doesn’t end up being as good.

  • CC – That could work, and would indicate whether we should expect more 4-mer data to help

• PB – Stupid question, is anyone gatekeeping protein FF training? What do we expect reviewers/critics to say?

  • CC – Criticism would be that we don’t know to what extent the choices that we made were dependent on our starting point, and if someone else tried the same steps as I took from a different starting point, it may not work for them.

•  

•