Tetrapeptide torsion profiles | @Chapin Cavender | DM – For the profile comparison, could we also compare the ff14SB port? Not super important but it would be a good point of reference. JW – What if some of these spurious MM minima are caused by long-range contacts? Should we do a modified minimization to keep MM from introducing other contacts? What is is we’re trying to measure, and how can we improve our pipeline to ensure we get that? CC – Could apply different restraint schemes. Maybe restrain all heavy dihedrals in the whole molecule? JW – It seems like there are two different FF quality questions to answer ehre: 1) Are QM minima also MM minima? Can test this by starting MM minimizations from QM minima. 2) Do we get the relative energies of QM conformers right? Can test this by doing heavily restrained minimizations of QM conformers and compare energies. CC – Good idea. We can fit to a mix of torsiondrive targets and optgeo targers
DM – How do AMBER FF fits handle this confusion? CC – AMBER uses dipeptides. They also don’t minimize the conformers. Some exceptions to this in the cornell fits in the 90s. DM – In some sense we want to allow a “tiny” amount of relaxation for torsiondrives. But we also want to include tests to ensure that MM minimizations don’t walk away from QM minima. SB – I recall a conversation a while ago about adding “canary” alarms/tests for forcebalance if a geometry does minimize too far away from the QM geometry. CC – I’ll ping JHorton to see if he recalls what came of that.
CC – Should we use cartesian or dihedral restraints on heavy atoms? DM – I’d be in favor of dihedral restraints, though cartesian restraints CC – I think a lot of the clashes are when eg sidechain methyls get rotated into the backbone. JW – If it’s a problem due to clashes, we may want to look back on PB’s work to prune high steric energy conformers CC – Forcebalance does reduce the weight on high-energy conformations. JW – It’s rough, because small errors in our vdW terms will cause much higher energy discrepancies than small errors in our torsion terms. There are a lot of unknowns here “large molecule” problems vs. “minimization strategy” problems. So maybe this will make more sense once we’ve seen how the similar plots look for dipeptides.
DM – It seems like there are a lot of different directions this can go, and it’ll be important to put a lot of them in “the bucket” for later consideration, so they don’t get in the way of current work. CC – That’s a good idea.
|
Protein library charges
|
| JW – We’re also looking at doing ELF1 for AM1 charge assignment, which may reduce the inconsistency in charges in the coming weeks MG – So, we’re seeing effects on the central residue due to interactions with the +-{1,2} residues in these sequences. These are especially pronounced when the neighbors are charges. I think these are due to conformation-dependent polarization issues. What that info is valuable, I don’t think we need to include quite so much variety in the +-{1,2} positions. DM – Agree. We could temporarily exclude that data from our fitting/benchmarking and consider coming back to it later. CC – So, there probably are interesting questions to ask about the charging models here, but we don’t need to solve those to make a first pass at the protein FF.
MG – I don’t feel super strongly about variability at +-1, since I doubt they’ll make a big difference in central residue charges CC – I think we’ll distribute the first generation of this with librarycharges. So we want to do a lot of work to get these right. While we don’t need to make these neighbor-dependent, I do think we should make our librarycharges be an average of the possible neighbors. MG – Fitting librarycharges to just a few QM conformations that include polarization effects may introduce a conformational preference that we don’t want to introduce. Maybe we should take a look at how charges for a particular molecule change with conformation. If this difference is large, then I’d be concerned about exactly how we select our training conformations. LW – I don’t have data on that precisely, but I did compare charges on the 0-residue depending on whether it had flanking charged vs. uncharged residues. The thing I found is that ELF10 is really important to get consistent charges. MG – Given the success of ELF10 in making charges more consistent, should we do charge generation using many conformations of neutral capping residues? LW – That would make sense. I’m not sure how much the improvement in consistency is due to averaging over distinct conformers vs. picking the ELF conformers. JW – CDavel is looking at ELF1 for now, and will later look at ELF10. CC – Initially using ELF1 is mostly for computational cost reasons. It’s pretty clear that we’ll want to use something like ELF10. MG – Given that LW’s ELF10 results are unaffected by the charge of flanking residues, maybe we don’t need to do anything clever with weighting. So we could either use just one or a few flanking residues, and it won’t make a huge deal. (The conclusion from LW’s slide is that, using ELF10 and scanning across a variety of neighboring residues (charged and uncharged) and comparing the central residue’s per-atom charges between flanking neighbors JW – (confused rambling that threw the conversation off) MG – What exactly is the approach we’ll use to go from QM results to partial charges? RESP? CC – We need to answer whether we want to reproduce AM1 charges or a higher level of theory. MG – I’d caution against a bit against using a higher level of theory+RESP, because it introduces a lot of tricky conformational dependence and other complications. So we should use AM1BCC biopolymer charges. SB – I think it’ll be best to AM1BCC biopolymer charges. We can also include some protein-specific BCC terms if we have large errors in the protein FF. This will ensure that our charges are consistent. This will also keep us consistent when we go to post-translational-modification world. CC – Agree JW – My mistake, I was confused and thinking that we’d do RESP fitting for the biopolymer FF.
JW – What’s between LW’s results from last week and the charges in our first biopolymer FF release? SB – Also, it’s worth keeping in mind that we may use graph models in the future. MG – What’s the timeline on this? JW – I don’t know – It’s kinda chicken and egg – SB – The important thing is that espaloma doesn’t need to be super accurate, they just need to capture the salient features. We wouldn’t need to use espaloma exactly, and I’ve made a similar protoyype. The major infrastructure blocker is that the DGL project isn’t on conda-forge – We may benefit from getting some PIs to chime in on that discussion to see if they can move it along.
PB – Could we make an espaloma prototype that’s easier to use? JW – Yes, we could make a new toolkitwrapper and parameterhandler in the smirnoff-plugins repo. I’m just very hesitant about putting it into the production toolkit because that will be 100x the work to implement, validate, and maintain.
|
