| Interpolated parameter analysis
| @Pavan Behara | CB – Interesting that, in 2 of the tig2 molecules, the torsion applies to non-nitro groups CB - separation of TIG2 – Special-case nitros [#8X1:1]~[#7X3$(~[#8X1]):2]~[#6X3:3]~[*:4]
PB (confirming in chat): [#8X1:1]~[#7X3$(~[#8X1]):2]~[#6X3:3]~[:4] We should determine whether the $(~[#8X1]):2] pattern allows “backtracking” within the SMARTS
JW – I’ll open GH issues for the toolkit to specify what should happen when we roundtrip a nitro group. (possibly also linking to a parallel issue on openff-forcefields) CB – Re: TIG0, are the cases where it does well actually leveraging WBO dependence? Or is the range really tight and it’s just the parameter splitting that’s doing it? CB – I see this saying that the fit is better, and would be appropriate to use in the upcoming release. But it also indicates better ways to split chemistry/patterns. DM – It looks like, for the tig0 parameters, we want to split it for nitrogens in one terminal position. Should we also have one for nitrogens in both terminal positions? CB – I’d have one TIG that’s all carbons. Then possibly a different one with 3 carbons and one heteroatom (say, N, S, O). I’m hesitant about TWO terminal heteroatoms.
t96+t97 replacement CB – What are next steps for the tricky cases? PB – Split into a new term SB – When we look at the original figure that shows that t96 could be interpolated, it looks like a clear trend. So this makes it puzzling for why we see such disagreement for other molecules. So when we see these molecules that don’t the trend well, do they follow the trend? DM – (something about fitting to residual, and not just total energy?) CB + DM – This might be worth looking into further CB – We should split out a torsion that looks at a carbonyl group that’s attached to sp3 carbons. on mol ending in 404, it’d be a torsion for atoms id 9 8 5 6, where 9 has to have a double bond oxygen on it, and position 6 has to be an sp3 carbon (so not any other heteroatom). We’d need another split parameter where 9 8 5 6 where 6 is NOT a carbon) JW – It may not be necessary to create TWO new parameters, since the “not carbon” case could be handled by the base parameter.
DM – TIG2 and TIG0 are both in the category of “clearly does better on training data, and mostly does a bit better on test data” SB – I’d like to prioritize building out the test set. These seem to have poor generality.
|