CB – When you say “reset to QM value of the equilibrium”, does that mean that a bond angle for a cyclopropane would be set to 60 degrees?
TG – It would get the average of the occurences of that parameter, so it would be 60 if the cyclopropane were split out by itself, but higher if other appearances were in there.
CB – It’s curious that the training performance is so similar to the testing performance - Previously we have been pretty sure that there was overfitting in the training performance.
TG – Important to note that test set is just from alkethoh, doesn’t have rings.
CB – I’d like to see a stronger test set in the future.
slide 5
CB – Wonder why exp2 had so many splits for bonds, whereas exp3 had so few.
TG – I didn’t show it on this slide but the force constants drove a lot of the splits.
Slide 6
CB – Important to note that QM value SHOULD be different from FF value, since other forces will compensate to make the FF wind up at a final geometry.
TG – Yes, I think we also see this in Sage.
CB – I’m not sure why we think that experiment 1 is overfit. Intuitively it should be overfit but the data shows it working well regardless. So I’d like to see how it performs “in the wild”
JW – Slide 4 shows that there are definitely local minima, but that resetting gets us to a better local minimum than not resetting.
Searching i3
TG – Almost all centers in test set are N, one center is S.
CB – The N-N bonds are unusual in actual drugs. The amide nitrogens are the biggest pain in the neck.
PB (in chat) – ring buckles too much when both N and S are in a ring, and doesn't when only one of them is in a ring?
JW + CB – The parameter split is concerning - Wouldn’t the new parameter capture different forms of sulfurs, unsaturated rings, different ring sizes, etc? Would that be desired?
CB – I also notice that there are molecules in the set with sulfurs OUTSIDE the ring, and the new parameter WOULDN’T apply to those cases
