| | |
|---|
Torsion refit targeting MM minima (new slides from 26) | @Chapin Cavender | CC, please upload slides here (slide 27) DLM: does this mean there’s not enough attention being paid to the diverse QM minima at the moment? CC: the “MM Minima” target here is trying to address that – if we have MM minima that aren’t true minima in QM DLM: my qn is about the reverse - Are there QM minima that we’re not finding? CC: not sure we have specific evidence that we’re missing QM minima, vs. introducing minima that shouldn’t be there DM – When you drop the opt geo target, aren’t you missing some QM minima? At some level there should be a symmetry - Whatever we’re doing to find QM energies of the MM minima, we should also do in reverse. CC – Agree. Was hoping taht, by dropping optgeo targets and only fitting torsions, we’d keep things on the rails. But it looks like we do need to include all valence terms in fitting and keep optgeo in there.
PB – Cycle 1 is starting point, and cycle 2 is after you included the QM energies of MM minima? CC – Yes PB – Instead of mixing torsion profiles and abinitio target, can you just take the torsiondrive profile and make abinitio targets from that? Then you could use them as single point energy targets. CC – So you’re suggested that I take out “td proteins” column and just group them with “MM minima”? Then you’d have no torsiondrive targets, and only abinitio targets. PB – Yes. That could take care of the issues with protein torsion weighting. CC – Doesn’t that asusme that things should be equally weighted? PB – Kinda, then it will depend on the number of points in each set. But then it should run faster as well.
MS – One idea that came up at ACS is “do we want to try something along the lines of some sort of last optimization to NMR observables, using some sort of dihedral weighting. Could use Vincent Voelz’s BICEPS package. AF could be the one who tries this. Would probably want to get this to work on smaller peptides, then validate on larger proteins. CC – I’ve discussed this with MGilson - We’d agreed that doing something like BICEPS makes sense. There’s also something from Giovanni Bussi that does something similar. So I’d looked into doing that, but it would be hard to get it to work in a ForceBalance fit. However I think I could run this in a scipy optimizer outside of ForceBalance as a proof of concept. But when MG and I talked we thought this would be a big deviation from the normal workflow, so I’ve been focusing on studies like this. MS – So, question is “when is the right time to switch from one approach to the other?” AF could potentially help out, but I’d defer to CC to decide when to switch tracks. CC – Gotcha. Since we have some low-effort ideas here to investigate, we should go forward with that. But if we end up at a dead end, we could look at fitting to NMR directly. MS – Other direction is to decide which software framework to do this in. CC – Met with MT a few weeks ago to look at this, it seems like BICEPS is the better path, since we can likely get support from VVoelz lab. Also it seems like it may be best to make some one-off code involving a scipy optimization.
PB – Remind me of the plan for water models? CC – Not going to use a 4-point water model for infra reasons. So we’ll probably do a 3-point water model that’s better than TIP3P. I think MShirts lab was looking into exploratory studies with different water models. MS – Right - We’ve got a sizable backlog here - Will look at getting results on this.
|
Protein benchmarks in GROMACS | @Anika Friedman | AF, please upload slides here JW – What were differences between this and the large-error plots we saw last wee? DLM – but there is systematic error visible. Due to HMR? MS – Maybe. Could try rerunning without HMR. AF – Depends on whether this will cancel out … DM – It’s just interesting that there’s systematic differences LW – Could it be worth doing OpenMM sims with HMR in case it’s differences in the software? MS – Or could do GROMACS without HMR. LW – In terms of computational expense, the openmm sims would be cheaper. DM – Yes, if there’s space compute elimitating one of the differences - OpenMM or HMR - would be helpful CC – Could use my GPU allocation on it … CC – I’ll run the equivalent HMR sims using GPU and report back.
|
Virtual sites current benchmarks | @Lily Wang | LW, please upload slides here JW – Were vsite charges allowed to eb 0? Would this mean that results never get worse? MS – were vdW params retrained as well? MT – Which parameters are being trained? Charge+distance? MS – LW – In my training, vdW was trained to phys prop. Vsites and BCCs trained to ESPs and (something) DM – … PB – Optimizing valence params to phys prop? DM – CBayly will have strong feelings/advice about pyridines. CC – Do you think we may need additional LJ types for atoms with vsites? Eg, right now we just have one N type DM – IIRC, CBayly said that AMBER ended up with specific vdW parameters for some atoms in pyridines, I think specifically carbons. Possibly to account for the lack of vsites. LW – I was surprised that Br didn’t do as well as Cl. It’s possible this is because of a differnet amount of training data. I can augment ESP training data pretty easily but experimental measurements are hard. MS – SO, if densities didn’t change… it might be interesting if there’s pure physical properties for these, and we could determine whether they’re more polar or less polar… Would be an interesting benchmark. The trajectories may be in there already, because, for the mixing calcs you have to do pure sims. LW – Unfortunately DM – Could be lots of density data for Br compounds so that’s pinning down vdW for those. … LW – I’m experimenting more wiht nonbonded fitting. I noticed that the sage fitting never converged. MS – Yeah, I noticed that pure chloroform density got a lot worse from openff-1.3 to 2.0. DM – I like the idea of the experiment where we fit to the heat of mixing. That may solve this sort of possible “double minimum” problem. MS – It is concerning that low molecular weight bromine compounds do poorly in 2.1….
LW – I’ll go ahead with experiment on fitting to enthalpies of mixing, and calculate more QM ESP data. Also looking at benchmarking with … can probbaly do this with my own compute. I think we’re going to have a problem with a lack of data. May look at benchmarking against pure compound properties. DM – someone in my group has been using BigSolDB, not sure if it has anything useful here. Solvation energies in diverse solvents. PB – Re pyridine – Are you using more training data than SB used for experimenting with vsites? LW – I think I’m using the same data - I don’t think NIST has updated. PB – It looks like these plots have more data points than SB’s - he picked out heterocycle-heterocycle mixing data. LW – I think he picked out specific interactions to show
|
| | |
| | |
| | |
