Ideas for next-gen GCNN charges | @Michael Shirts | NOT short term, just wanted to brainstorm what could be the next gen. Two options are retrain AM1-BCC to ESP’s and to include condensed phase simulations Alternatively For example, instead of reoptimizing BCC’s to match ESP’s could potentially train GCNN’s to fit to ESP’s directly. However, might still need retraining to condensed phase, which would require putting simulations inside the ML loop (already of interest, but a bit more work). The purpose of bringing this up is just to think about what the nearer-term roadblocks to making any options possible down the road. MG – Isn’t espaloma already predicting high-quality charges? MS – Currently it’s predicting AM1BCC. But the though is, if we were gonna retrain the BCCs to match something better, why not directly train a GNN to the better quality charges? The big requirement should just be that the information be written in a way that’s automatically differentiable. The other idea is to train to condensed phase sims at the same time. MG – This also requires a lot of data. So you could train BCCs to liquid state data as well as some nicer QM. MS – JC and I talked about this - It’s essentially the same cost. MG – One question is how many liquid property measurements we need, and whether/when we’ll have that. DM – Yesterday we were discussing some work by DCole down the line of “how polarized should small mols be”, and they arrived at a conclusion kinda like ipolq. I’m reasonably convinced by those arguments and they line up with the arguments from MG’s work with Schauperl. But even if that is good, it won’t be suitable as a deployed workflow. So one way would be to train a GNN to a large dataset like that. But you could also include some amount of information from condensed phase properties using transfer learning. MS – Even if we do GNN charges, there will still be people who want to do AM1BCC with better BCCs. So one thing we could do is “better BCCs” in the first pass, and then train a NN on that. Then we’d have both things available. LW – I like the idea of training to AM1, and then refitting BCCs. Seems like that’s the best of both worlds MS – YW showed the best performance training directly to AM1BCC, right? LW – MG – So it worked …; LW – I trained my NN to do AM1 charges, and then subsequently apply BCCs. DM – The reason AM1BCC kinda works for large areas of chemistry is because the AM1 will alwasys give a decent guess, but the BCCs wont systematically worsen the guess. LW – So I trained a model to predict AM1, and then retrained the BCCs to match the output of OE AM1BCC (?) JW – To clarify - Is this an NN that only predicts AM1, and then a charge increment model that applies BCCs, or is it a single NN that internally predicts AM1 and then internally applied BCCs as part of the NN? MS – So the current system started with simon’s BCCs, but had to retrain the BCCs because the originals don’t recapitulate good OE-AM1BCC, and the result looks like OE AM1BCC? LW – Yes. JW – So the BCCs have to be retrained because of systemic issues (like sulfur) can’t be resolved in the GNN and so the BCCs can cover for them? MG – I’m not sure how much I could expect the BCCs to fix weirdness if we don’t have said weirdness it in the training set for the GNN+BCC workflow put together. MS – Have you talked to SB about infrastructure to refit BCCs etc, and how to best use it? MG – Also there’s the question of whether to use something better than AM1BCC LW – SB was using HF-31G*. So we could use that or something even better. CBy – I’d be a fan of going past AM1. Maybe something like xTB. Or MG’s charge method. But the issue was that it’s hard to beat the coverage of AM1BCC. This keeps the door open for pharmaceutical chemistry and organic-friendly metals. But generally I’m a fan of moving past AM1 to a better semiempitical method. It’s not clear that 6-31G* is even the best, something like RESP2 or some of Nerenberg’s work could be further improvements. And whichever thing we choose here should be used for the GCNN fitting. MG – I think the reason to fit to point charges instead of ESPs is to save time. MS – And sticking with AM1BCC in the short term is safe. CBy – The more polar the molecule, the more accurate the model/fitting needs to be. MS – Yeah, once atoms already have a large absolute charge value, small errors have an outsized influence on the system energetics. LW – First iteration priorities are to get a fast AM1BCC-like model out asap. So the fitting to ESP stuff would probably be an improvement when we have more time+resources CBy – Fitting to ESP grids can be computationally simplified by reducing things to A and B matrix, which ends up just being an NxN matrix (far smaller memory footprint than brute forcing it). Hyesu’s package -
BS – The A matrix would have dimensions equal to the number of molecules, while the B matrix would scale with the number of grid points. CBy – Yeah, it just looks like a covariance matrix of 1/r terms. … (General) – This will be a good way to do fitting to ESPs, should make sure we reconstruct this method when we do that.
Cby – Re: Should we switch to xTB?: What were our findings there? DM – This looked promising compared to AM1BCC. MS – If we’re going to end up with a GNN, then computation time doesn’t matter as much. JW – Agree
… CBy – If you’re gonna train a GNN to ESPs, then you don’t need to do any cheap stuff like AM1BCC or xTB, you can just train directly to good methods. MS – CBy – If the GCNN stuff works by training to ESPs directly, then the entire world will use that instead of AM1BCC. MS – Right, but what about training condensed phase properties? CBy – Even if, ultiamtely, GNN charges win, until that’s been shown to be applicable across all domains of chemistry, it would be wise to also have a semiempirical BCC charging method to maintain the robustness/coverage of our current approach. MG – But our domain needs to also cover large polymers and things like that. MS – But we’ll need to have a physics-based method that we can fall back to 0 the GNN should try to replicate that model at a larger scale, but we should be able to go back to it. CBy – You can always cap and charge covalent adducts to polymers. MG – I thought the goal of the GNN route was to not to do this. CBy – The conformational dependence is tricky, and I tried to mitigate that with ELF10 MG – Do we want to provide cap-and-charge? MS – I’m imagining that we have something we can verify against CBy – I second that we also want to have a physics method that we can use to verify results on. And once we have that then we can validate the GNN on things like that. LW – But any GNN would need to train to a physics based method. MS – We also want to train to condensed phase. CBy – … MS – Two choices: You have ground truth data. You fit your GCNN directly to that. You have a first-step charge method trained on QM, then you do a fit of BCCs to condensed phase data
CBy – I’d love to have a Rosemary-like FF like the AMBER FF that was trained to RESP, but still did fine with AM1BCC. This is a world I’d like, where if it looked like the GCNN was misbehaving I could go run the physics based charge assignment.
DM – These are interesting thoughts for next steps - Will be good to revisit this once the .
|
| | WW – We’re developing a polarizable version of BCCs, using openff-recharge. I’ve got some additions that let us fit to MP2, so far we just have a small training set. Right now our AM1BCC-dpol model is producing a more accurate charge surface compared to RESP, but it’s not getting better for liquid properites. If we wanted to retrain the rest of the FF to liquid properties, what data/method/properties should I use? CBy – I’m thinking of dipole? Do we have a mix of polar and nonpolar solvents? MS – If we have miscible mixture data, we’ll have dipoles that aren’t too different. But I think we have like cyclohexane-hexanol data. CBy – So, is the question, “what is the target level of theory to which we should fit”? WW – We also want to trian the BCCs along with liquid properties instead of just to QM DM – I’d want to run two experiments in a perfect world - Assuming you have a general charge model and you want to fit a FF to it. I’d suggest using: that charge model in the condensed phase to do fitting for LJ and valence parameters. Other thing to compare is doing the same thing, but where your condensed phase fits also get to tweak the charge model somewhat.
WW – That’s what I was guessing as well. We’re currently running some condensed phase calcs… DM – Also good to consider molecule size/conformational landscapes. CBy – I largely agree. We’d largely want BCCs to be stable and independent of polarization. But you’ll need to retrain the torsions anyway MG – To clarify, we do have new BCCs once we put the polarization in. CBy – Ideally the BCCs would be independent of whether they’re in a polar/nonpolar environment. MG – On a larger level, maybe we should focus on “how do we compare this to another FF?” - We’d probably need to refit some terms to condensed phase. Fit to enthalpies, hvaps, etc? CBy – SB had a paradigm about retraining LJ LW – We haven’t changed from SB’s paradigm for fitting nonbonded. MS – JHorton and DCole sent their new FF through the same workflow, right? LW – Yes MS – So they could give some pointers about how to run the workflow locally. CBy – I’d think torsions are the thing to refit, not angles or bonds. WW – Great, I think that answers my questions. Thanks!
|