Dipeptide 2-D TorsionDrives with sidechain dihedral constraints | @Chapin Cavender | JC – Are sidechains CONstrained or REstrained? If CONstrained, we could get unnaturally high energies CC – CONstrained, I think. It’s the same style of constraint as the driven torsions DM – What do other people do here? CC – In most other existing protein FFs, they don’t do anything to the sidechains. They have all their degrees of freedom free to minimize
JC – (In tryptophan rotamer 2 slide) - Doesn’t this show that energies are highly sensitive to sidechain position? I think this means we shouldn’t totally constrain sidechain rotamers JW: The discussion above seems to assume that high energies are bad. But isn’t that fine if the actual molecular energetics SHOULD give a high energy? Is the implication that we know that sterics will be incorrectly modeled, so we want to avoid large steric terms? JC – In the MM constraints, are the torsion ANGLES constrained, or are the ATOM POSITIONS constrained? JC – On Alanine MM taget slide, it looks like things are coupled diagonally, which would imply that we need a CMAP that depends on both phi and psi MG + MS – Agree MS – Could include coupling terms to become more accurate DM – Would be better to do conventional torsions in the MVP. JC – May be able to do custom 1-4 scaling to compensate for this as well.
JC – The energy difference plot - Does this equally weight high-energy/low-probabilty regions? Or is the MM difference already boltzmann-weighted? JC – Do we know whether the moving minima are due to the constrained sidechain position, or (something else)? We could look at the forces on the constrained atoms to measure the effect of the constraints CC – I could look into this. Previous MacKerrell paper said that common sidechain rotamers nearly always correspond to QM minima. JC – It could be good to double check this. Maybe we could run some quick ANI calcs? DM – So, you’re taking a single set of sidechain confs and constraining them to stay there throughout the whole torsiondrive. This could mean that there’s a total clash somewhere in the landscape on a single sidechain conf. Thinking out loud, we’re trying to fix this phenomenon where the sidechain confs jump throughout the torsiondrive. MG – This is why we checked and found that there aren’t big steric bumps in these results. DM – Another way we could have done this is with REstraints instead of CONstraints. DM – Thinking about the minimum viable product - We were trying to fix a problem that we observed without any -straints. JC – The data that we have available can help us here: The sidechain forces can tell us how unhappy they are We have an optimization dataset for each of these, and they should be at a minimum for each of these. So we can see whether those went far from the torsiondrive results.
MG – Again, we didn’t see huge steric energies, so I don’t get the sense that any of these are invalid. They’re all achievable by thermal sampling. DM – … MG – This is coming close to asking for a 4D torsiondrive to ensure we find the best sidechain position for everything JC + DM – We’re more of advocating that we use REstraints instead of CONstraints. DC – In previous work, we had not restrained sidechains at all, we just seeded a few different sidechain rotamers in the scan points and then selected the lowest energy for each grid point. MG – So maybe a plan of action would be to run a coarser 2D scan, with sidechain REstraints seeded at a few different positions JC – People generally use three sidechain rotamers, and you could come up with a force constant based on some boltzmann stats. SB – I could take a look at which angular restraint strengths I used. (general) – CAN we use REstraints in QCA? SB – I don’t think so. GeomeTRIC may allow this using “optking”. Maybe a PR to torsiondrive or geometric would enable this. JC – We may have the data we need here - Can compare the results of the optimizationdataset to the results of the torsiondrive. Then, if the energy differences are large, we know that Restraints should be used. CC – I don’t think we have that info. JC – I thought we’d generated optimizationdatasets for all the torsiondrive endpoints. That’s how we’d get bond lengths and angles (General) – The plan for rosemary hadn’t been to make protein-specific bonds or angles. Just torsions. But new torsions would be added, and all the numerical values (JW edit on 2022_01_18: Apologies, I think that what was discussed was actually “all k values for these new parameters”) will be refit.
JC – So, we could do a new optimizationdatasets seeded with the protein SMILES, and then for each optimized geometry, compare it to the closest grid point from this study, and see if the energy difference is large. This will tell us whether the sidechain-constrained TD gets close enough to the minima
(Next steps slide) MG – A third option would be what Danny said - Not constraining sidechain rotamers and seeding multiple starting points could work. JC – What about seeding from all the scan points? JW – I’d be skeptical of the Cerutti set - There are several chemistry deficiencies and a high error rate. MG – We could save some time by increasing the scan step to something like 30 degrees. Could experiment by doing a fit to the current data, then do a fit to the data but only as if we had submitted a 30 degree increment (so slice the existing data).
SB – The sooner we can identify the benchmarking observables, the sooner we can get those implemented.
|
