10 mins | General status of wbo work | Pavan | PB will link slides here Slide 1 CBy – Is the main message “Just by introducing new non-interpolated parameters (column2 vs 3) do we improve the FF?” or “by introducing interpolated torsions (column 2 vs the third from last) do we improve the FF”? CBy – My big question is whether we improve the FF or reduce the number of parameters SB – Are these the number where we look at the test set and see how many molecules exercise the parameters, and the answer ended up being “2” SB – When we looked at molecules and the WBOs they produces, it would often be in a narrow range, like between 0.9 and 1.0. And the gradient would be very steep. CBy – Does that indicate that WBO dependence only matters when there’s a large (non-normative?) conjugation effect at play? SB – I think the hard thing is that the molecules that we’re testing against don’t have a strong conjugation effect. CBy – I’d expect to see the largest impact of FF calculations would be for a conjugated single bond, where the sp2-sp2 Cs and Ns have conflicting effects – One based on sterics and electrostatics driving it out of plane, and one for conjugation trying to keep it in plane – things like biphenyl mols. DM – That might be the problem. We don’t have super diverse test sets that do span a range of conjugated environments.
CBy – The nitrogen pyramidialization slide shows a urea, which is does have both flat + pyramidal character. The final preferred geometry is decided by finding a balance between improper strength, proper strength, and valence angles around the urea N. DM – Based on this,w e want to make sure we don’t rush through and make a bad solution to this. CBy – I wonder if this is because the biphenyl systems never go nonplanar, but that’s where our proof of concept comes from. So maybe we don’t try to fit ureas for now? DM – Maybe that means we should avoid conjugated bonds with sterics involved in their profile for now. So “suppose I want to fit the cental torsion parameter for two six memebered aromatic rings that only have carbon, and all of the carbons except the para position have no other substituents attached.” CBy – We could easily extend that to things with no ortho substituents [#6X3aH1:1]~[#6X3aH1:2]-[#6X3aH1:3]:-[#6X3aH1:4]
SB – If this is going to work, it’ll work here. I don’t want to change the current plan, though, so this should be kept as PB’s side project. But I think there are other, lower hanging fruits in our FFs. Especially because there are so many other technical hurdles around getting WBOs matching between OE and AT. Also, if we end up not being able to use WBOs for more chemistry, then it will be weird that our whole FF only has one WBO parameters several releases later. CB – I think this isn’t a blocker. The magnitude of changes is small. Especially if the AT implementation can be modified to avoid catastrophic failures/rearrangements. JW – Disagree. The magnitude of the difference is too large. We’re also supposed to be advocates of reproducibility, and introducing a new implementation-dependent feature is not good.
DM – I agree with SB that this shouldn’t be rushed. (General) – Should we have “guard rails” on the possible k values that could come out of interpolation? CBy – Can we get anything toward WBOs going here? SB – Trouble with infrastructure Lack of fast charging/WBO calcs for large molecules Lack of proof of generality past biphenyl
CBy – Could include both WBO value guardrails in the interpolation and chemical space guardrails in the SMIRKS SB – That could be possible. It depends how quickly we can become confident in the new feature’s release readiness. SB – I’ve been thinking about how we could have “experimental” ffs that are lower stakes.
CBy – “What if t69 is WBO dependent, but the FF additionally has a “typical” value that will be used in the case that a WBO can’t be determined. JW – I think this is a great engineering idea, but this would likely be unacceptable to project leadership because of the lack of reproducibility. SB – Agree that this is very interesting. I’m a bit cautious about what should happen if observed distribution is bimodal and the default value either needs to be in a valley or select one of the peaks.
To dos: AM1-Wiberg equiv. between OE and AT Testing guardrails on WBO’s using either very limited chemical substructures (that avoid sterics eg above) or flat regions of interpolations. Make “clean biphenyl” dataset for QCA submission w/ medium priority.
|
