Time | Item | Presenter | Notes |
|---|
10 mins | | Pavan Behara | Results of fitting impropers along with other valence parameters. CC – How is this different from Sage? CB – Was there a prior on these before? PB – Yes, I used 15 CB – It’s strange that i3 became so stiff. SB – This means that there’s a harmonic spring with a strength proprotional to 15, restraining it to its original value.
MS – What happens if we start a new optimization starting at these final values? CB – What happened when this was run without restraints? CB – I look at this refit, and it really looks like many of the handwavey values from the early FFs are very accurate. CB – The initial work on this were achieved by taking eg. benzene and doing normal mode analysis with HF-6-31G*, and then doing binning of the results into cheminformatics. DM – I looked into the history of these parameters, and it looks like they were all descended from 4 impropers in s99F. i1-3 are identical, and i4 is now i7. CB – When I see the “refit” values, I want to trust them. But what would you identify as the big limitations? Would things be better with less regularization? MS – A few thoughts about impropers MS – I’m reminded of some issues from when we were getting odd results in nonbonded refits for acetic acid, which were ultimately due to a cis/trans transition for carboxylate groups. So it may be worth considering kinetics to ensure the transition frequency to understand that things remain accurate CB – For aryl amines we should expect inversion frequently. But this point raises a spectrum, where carboxylates and amines vary from “never flip” to “frequently flip”. Basically, the lower the barrier, the more important it is to get it right
JW – For the training set here, were the impropers driven to ensure a diversity of angles, or was this the vanilla sage training set? PB – Do we match the same atom twice in an improper? Like, if I have 1-2-3-4 (PB shows a case where an improper is like 4-5-6-4) JW – Oh, that looks like a bug. Could you send me the code to reproduce this?
PB – Should every possible improper center get parameters assigned? DM – No, impropers are the one type of parameter that won’t apply to everywhere they could in the FF. Historically they were only used to keep things flat, but we’re free to try using them for non-flat centers
|
JMaat’s slides | | | CB – It’d be nice to show the impact of these changes on the overall accuracy of the FF DM – This info wasn’t intended for presentation, so I wouldn’t read too far into it
|
Biopolymer plan clarification | | | DM – CC, are there any outstanding questions about the slack thread? CC – No, I think things were resolved on the slack thread. DM – Great. I basically wanted to say that you shouldn’t worry about refitting other parameters. SB – For sage, we did the vdw, then we did a valence fit. So once we start doing the next gen, we should do it as electrostatics → vdW → torsions → valence, and we can group adjacent steps. MS – With the difference that, before we release this, we need to do benchmarks to prove that we’re not inferior to other protein FFs. SB – Maybe we want to see the results – If we have 5 situations where the performance gets better, and 5 situations where it gets worse, how do we decide if that’s acceptable? MS – We’ll have to discuss that at the time. CC – We got Christ Oostenbrink? to write about how to compare against different topics.
CC – Strategy-wise, it seems like the charge model should have primary, since we have an independent charge model that will hit an absolute target. So I think that everything else should be refit in light on that, which would include rerunning LJ optimizations CB – Generally FFs have been fit in the order of electrostatics → vdw → valence. There are arguments now that they could be fit simultaneously, or that the ES could be fit to produce the correct overall answer. I think we should do the physics based approach SB – Agree. I think we should plan on having AM1 be “ground truth”, and then we can fit BCCs to condensed phase properties.
CB – We’re going to hit an accuracy limitation until we get vsites in. Can vsites be in JW – It’s not clear that we could get support for vsite export to other ecosystems into interchange this year. SB – Did MT complete looking into the ability/effort estimate to export vsites to other formats? JW – I don’t think that completed. SB – I have a new dataset that could be suitable for vsite fits.
|