| | |
|---|
Project plan | @Chapin Cavender | MS – Re: “Prioritize consistency with small molecule FF” – We want nonbonded to be the same, as well as bonds and angles. But we’d kinda expect dihedrals to differ. CC - Agree SB – Agree JW – Agree MS - They way this is described to other groups is going to be critical MKG - shouldn’t be too hard to justify as people always ask, well, is you SMFF compatible with X protein FF?
(protein ff models slide) SB – We should put the protein-specific LJ refit goal above CMAP – LJ compatibility should be a high goal. CC – So protein FF should be an extension of small molecule FF MS – Yes. One question is “how much surrounding chemical context needs to be present for a ‘protein’ parameter to be applied?” MG – I think it’s the specificity… CC – So maybe the right way to brand it is “polypeptide torsion”? JW – CMAPs at least 6+ months out – Current infrastructure goals are pretty ambitious and don’t include CMAPs yet. SB – People previously have also expressed hesitation about CMAPS.
(Proposed timeline slide) MS – What’s the difference between new torsions and torsion CMAPs? MS – And should the QC dataset for training torsion parameters be different for training CMAPs, or should they be the same? JW – It seems like they should be the same. CC – I think they could be different – There may be a need to train on a different number of grid points for torsion parameters vs CMAPs. MS – So, maybe the initial scans should be done really finely, so we can subsample more easily. SB – The QC* infrastructure is very slow for 2D scans, so we may want to do widely-spaced grid points initially.
MS – We’ll want the fitting infrastructure+benchmarking infrastructure to be very automated. So we should coordinate with infrastructure team. JW – This will depend on where we’re running. Will lilac be our compute center for this? SB – Lilac is probably the largest cluster we’ll have access to. Folding@Home will be for RBFE calcs. MS – Could use compute at oak ridge if needed, but it’d be on powerpc. MG – Would XSEDE time help? Could get startup allocations very quickly. CC – I put through an XSEDE proposal in my PhD, so I could take the lead on this.
JW – If we want CMAP parameters ready for august, when would we want the infrastructure to be ready?
|
Infrastructure | @Chapin Cavender
| JW - currently have library charges available for all natural amino acids which may be alternative to generating charge. Reach out if you want more info. CC – Do we need stereo-specific SMIRKS? JW - where this will likely have most impact are partial charges, that will then have knock on effect for torsions. If all the stereocenters flip then the partial charges should be the same, so the parameters could be the same. But if only some stereocenters flip then the parameters wouldn’t transfer. (General) – We’ll keep looking into this
MT – Are CMAPs implemented the same way in all engines? Or is there a significant risk that the energies in other engines would be different? CC - not sure, but people probably looked into this. Details about CMAP implementations in AMBER are likely in the chamber program. MT – Thanks. My biggest concern is about interoperability - This is separate from the openff toolkit implementation
|
Training datasets | @Chapin Cavender | |
Validation datasets | @Chapin Cavender | MG – CC and I talked for a while about didpeptides, tripeptides, tetrapeptides, and the time-cost of the runs. And can we reduce the combinatorial explosion by flanking the AA of interest with “representative” JW – We can also submit partially-overlapping datasets with varying “priority” levels for computation. Then we can reduce the effect of uncertainty wrt runtime – We just do test fits with whatever’s available when we reach the deadline.
|
LiveCoMS article | @Chapin Cavender | |
