v1.2.0 release candidates | @Hyesu Jang | MM value discrepancy from QM CB – It’s possible that we’re going into different minima for some bonds and angles. CB – It’s possible that, because our FF has an artifact (intramolecular hbond), we might be stuck with having some cases where we can fix the artifact, or capture the chemical reality better. CB – I agree with HJ’s assessment that “here’s where we see an artfact due to a intramolecular hbond, which affecrts an urelated parameter”. But one of the questions is “what do we do about that?”. CB – The third tagged atom in a15, we could make that #8X2H1 to differentiate out the hbonding case. CB – Interesting to use two optimizations to zero in on a parameter that needs a chemical perception change. I think this artifact may be addressable to have a “charge penetration” term. That would be “spreading out the partial charge on an atom” DM – It could be that LJ changes in this cse could also address things. DM – A course of action here could be to make a 1.2.1 release that splits a15.
RMSE plots JW – Is it basically a direct conclusion that high-RMSE torsions are the ones we should focus on splitting first? CB – Is it possible that the torsions that improved and the ones that got worse are interconnected? CB – We might pick which FF becomes our “official” release by identifying which one performs best on “pharmaceutically relevant” compounds. DM – There may be one or two quick tests that we can do, but we probably don’t want to do anything too time-consuming here. We could test against the FDA approved drug set and use that to break the tie. CB – I agree with that. However the problematic angle (a15) might be worth a point fix – could we fix that before the final release? LPW – This would be moderately complex, and we’d never know how much better we could have made the FF if we’d done a full optimization. So I’d prefer we move forward with RC1. Making a local fix would probably take a few days. HJ – We’ll want to come up with an initial guess for the new parameter values CB – When we had the artifact with the phosphate and hydroxyl, David case mentioned that the same thing was a long-running issue in AMBER, which has always been tolerated. So, in terms of facing the competition, we don’t need to win on this front. But in terms of phosphate in particular, the single-protonated state is probably a lot more abundant in nature. DM – It’s worth noting that we dind’t aim to fix phosphates in particular DM – How long would it take to benchmark FDA-approved drugs? HJ – It would take a few hours HJ will benchmark on FDA-approved drug set, and pick the RC that does better, and make the release this week.
LPW – At the moment, I agree that we’re comparing the benefits and drawbacks of the release candidates. The biggest thing to me is the large force constant of a15, which I agree is artifactual. For some reason, in rc2 the artifact seems less bad DM – In rc2, the a15 artifact is just transferred to the equilibrium angle instead of the force constant DM – I also think we shouldn’t consider ring parameters at highly. Topological constraints in rings make the exact parameter values less important
LPW – Looking at gradients to inform our choice of which parameters to work on is a good direction. But I don’t think we’re using gradients correctly now. What’s shown in this plot is the gradient of the final objective function wrt the parameter. But it may be more informative to look at gradient contributions per molecule, but right now we can only see it per target, which lumps together many molecules. If we looked more closely, we might see parameters with bimodal distributions, and those are the ones we should split. DM – When SB discussed nonbonded parameter gradients, he found that everything was opposed to ketones, in terms of the direction they want to move the parameters. DM – Would be good to see red/blue curve FF comparison plots for 1.2.0 RC2.
|
|
| CB – Last time we talked, we talked about amides. We had seen an artifact for amides that didn’t have two hydrogens, but were dialkyl. To see whether the ketone problem generalized to amides, we proposed doing liquid simulations of dialkyl amides. If SB tries to use our FFs to model these, we might get LJ parameters offsetting our artifacts. DM – Maybe in the future, we should start optiizations from 1.2.0 CB – I’d be concerned that we’d get stuck deeper in a local minimum DM – But if we converge faster with a more optimized FF, then we could run optimizations from many starting points in the same amount of time. HJ – I DID see that there were some unphysical ks in RC2, so I’d be concerned about using those as a starting point. CB – Over the course of this call, the pattern that we’ve folowed is to do a global optimization, and then look into specific pathologies. When we do that, we’d expect it to have a relatively small effect on the global optimization What we want to do is to identify the part of the objective function. We should develop infrastructure to account for the fact that we’re improving specific subspaces.
|