(JW taking notes, but joined late. But missed first ~20 minutes of meeting)
CBy – WBO concerns – Happy to see that the WBO DOES make same relative ranking for most compounds, even though there are different absolute values when comparing between AM1 and more detailed QM.
CBy – Would like to see plot of of QM energy versus MM energy with everything except the torsion being driven (full QM energy - full MM optimized energy with specific torsion zeroed out)
LPW – I’m not sure that this is how we should do things. I havent seen evidence that the errors that this could bake into the resulting profile are small enough to not mangle everythign. This is why I rpoposed doing the interaction energies from fragments.
DM – On a practical level, the approach that msot people have taken is fitting to the residual. If we’re going to do something different we will need to show data on why we do something different.
CBy – As a practitioner of the “classic method”, I actually agree with LPW’s points. But I think it is more likely to get us to a self-consistent FF, and then we can start modifying the details of the method.
LPW – We use this approach to get cis/trans energy differences and then subsequently use that to see how well WBO correlates with that difference. This is distinct form cases where we’re fitting to the residual.
CBy – For something that we want to put into production, we should stick to the classical approach.
BS – This questions has been going on for 30 years. The question is whether the torsion terms are a “garbage heap” that clean up errors in other terms, or if it has another meaning? It seems like torsions are largely there to compensate for errors in 1-4 interactions. So this is also coupled to the 1-4 scaling factors. But that’s something that needs to be considered in the torsion energy plots/fitting.
CBy – We should be practical, and use something “tried and true” for now. I’d consider the “classic approach” to be the null hypthesis, and it’s what we should use until we have evidence that other approaches are better.
LPW – The context of this whole discussion is “amides”. In this particular discussion, the question that we’re trying to answer is whether there’s a correlation between bond order and torsion barrier. We’re already seeing subtle energy differences between cis and trans. So I’m looking beyond “getting a specific version out” and looking at “answering the basic question”. ???… Would like to see these approaches tried on larger molecule sets
DM – I think it’s worth doing it the old way primarily, and it’s fine if we study new ways to do it at the same time. Also, we should look more deeply into correlations between AM1 and higher-level wibergbond orders
SB – Agree
CBy – NCNC torsion about amide from scan 2 – What is the WBO when that is changed from 0 to 180 degrees?
15 min?
new ideas from bespoke fitting
@Joshua Horton
Insights from recent bespoke fitting work which may be worth testing in a fit of a general FF
DM – This is a proposal for fitting FFs using a method where we don’t restrain the geometry to the QM minimum
CBy – So the restraints are only there to remove cases where there are pathological cases, where MM learns a different minimum than QM is trying to communicate. In the case where the minima are the same, the restraints will decrease the quality of the fit. Is there some way where we could only apply restraints if they’re needed?
JH –
CBy – So, I’d propose not using restraints, but just throwing out points where the RMSD changes a lot.
JH – MAkes sense. That’s how we do torsion profiles.
LPW – I like CBy’s idea of detecting when we’re optimizing to the wrong local minimum. But that would be very complex, and I’m not sure how we’d do it. But one thing that we could do is perform a non-parameter-changing calculation (optimiztion) where we don’t optimize any parameters but we do look at the final RMSD, and then choose to remove the ones that have a high RMSD from the actual fitting step.
JH – Sounds good.
TG – (Something about detecting torsion shifts to identify discrepancies in torsion fits?)
LPW – So what you’re saying is that, if there are torsions other than the one being fit that are bad, this may affect the torsion being fit by throwing out their data points? So you’re asking whether there’s a way we could avoid those cases?
TG – Yes
LPW – So maybe you’re saying, instead of using RMSD as an objective function in a torsionsdrive, we use something more like an internal coordinate metric. Not sure how to predict the cost of that.
DM – Is there an action item from this?
JH – I’ve almost got this working. I can work with TG on it.
LPW – If a torsiondrive now includes orthogonal degrees of freedom, it’s going to touch on a lot more parameters than just the torsiondrive itself. This is probably a good thing, but that’s just my prediction.
DM – Yeah, let’s treat this as an experiment and not a plan.
30 min?
WBO progress
@Jessica Maat (Deactivated)
DM – We’d hoped that WBOs would more clearly show improvement, but it’s turning out to be complex.
(Insert JM slides here)
JM – Feedback on plans?
CBy – Chaya’s dataset had really nice correlation. But I deliberately chose those because they were particular clean (no big 1-4 interactions, etc). Do we have residual energy plots (ie, residual MM energy w/ torsions in question 0’d out and then optimized vs torsion)?
JM – I don’t think we’ve done that analysis yet.
PB – Perhaps we did effectively did this, when doing the fitting of k terms for particular torsions
DM – One confounding factor with these plots is that we can only perform this sort of fit on molecule with only one instance of this bond (eg *-CX3~CX3-*)
CBy – I’m curious about how the negative k/torsion barrier values come about.
JM – Note that there are two Y axes – The torsion barriers are still positive
DM – These plots also exclude cases with strong sterics
CBy – …?
JW – Based on the SMIRKS, these k values would be applied 4 times to a rotatable bonds. So to compare the relative differences between the red and the blue, the relative differences between red and blue will be off by a factor of 4.
CBy – The range around the single bonds seems to have a nice trend, but the double bonds are much lower than I expected.
JM – Agree
DM – Yes, when we start doing this using the bespoke fitter this agreement should get a lot better
CBy – I’m optimistic that we could craft SMARTS that capture the “nice correlation” range around a WBO of 1, and exclude the tricky cases around 2.
DM – Another route forward is to make another, “cleaner” congeneric series to play with.
PB – The error bars on alide 16 – These relect the difference between k value and QM barrier. Can we determine the cause of this?
JW – I think it’s hard to say whether it’s caused by actual chemistry, or if it’s an artifact introduced by how this dataset had to be selected
CBy - Some interesting additional chemistries would be deprotonated hydroxyls wherever they appear, and possibly alkoxides (eg CCO-) substituted onto rings
DM – One question is what to do if we get good results on this “clean” molecule set?
CBy – Then we need to tailor our SMARTS to distinguish between “good” and “bad” torsions.
DM – One reason it may not perform better than 1.3 in benchmarking is that the methods used to train them are slightly different.
HJ – What’s the starting point FF for these fits?
PB – 1.3.0
HJ – I think there may be some relaxation introduced by angles and bonds. You might get better results if you let them optimize as well, though the priors for those should be high so that they don’t vary too much.
CBy – Is the objective function being affected by the reduction in number of parameters?
DM – Yes
CBy – So, getting similar objective function values with less parameters IS a victory. But now we want to analyze whether interpolating the barrier actually helped. So, there are two question: 1- Is WBO giving us the physical behavior we want to capture and 2- Are we separating chemistries correctly?
DM – Would like to try a refit all the parameters with this new term
SB – Agree. We need to test out the infrastructure anyway so this wouldn’t be a waste.
PB – Interested in determining what a “significant difference” looks like. The difference when using my WBO parameters is similar in size to the difference between 1.2.0 and 1.3.0.