DM/JW: how many torsions are being subtracted from the MM energy?
AMI: just one instance of each torsion – if it gets applied multiple times across the same central bond, it still only gets subtracted once
(end of part 1)
CC – One slide showed that, if you take general FF and fit a torsion specifically to a molecule, you see that the torsion gets a certain value. How will this transfer back to the rest of the FF?
LM – Thought is to fit to the smallest molecule that exhibits a torsion, and to use that as the starting point for the whole FF fit. This should help us keep from getting trapped in a bad local minimum that we can’t escape.
TG – This may have an issue if you fit to multiple torsions about the same rotatable bond (/in the same “smallest molecule”). So it’s possible that they’ll be free to vary in opposition to each other - One might get large and the otehr small, and this could be problematic when the two parameters have to interact in other mols in the full training.
LM – Good point, I’ll look into that.
(first slide in part 2)
DM – Hard to disentangle effects from sterics vs. bad torsion params
DM – Wouldn’t some of the problem cases get fixed by the switch to explicit-multipliciy-on-central-atom change?
BW – …
LM – That may work for some of them, but not all.
DM – Is the root cause of this that these two things shouldn’t be sharing a parameter? Like, maybe the things on the left should be flat but the things on the right shouldn’t
…
LW – Could mine our QM data to figure out SMARTS that classify flat vs. tetrahedral
TG – I could give this a shot
LM – I’ll send you SMARTS patterns that may work for that.
CC (chat) – From the SMIRKS and shape of the FF parameters, it seems like t16 is meant to handle the case where all three carbons are in the same cyclopropane ring, not the torsion between two cyclopropane rings. Probably could be fixed by specifying ring bonds between the r3 carbons in the SMIRKS pattern.
LM – Good call
WS – One of the things that make s torsion parameters less transferrable than we’dlike is that contributions from the 1-4 interactions. In cases where you’re seeing k=0 works fine, it could mean that the potential is captured in the 1-4 interactions. Different forcefields over the decades have used different scaling values for the 1-4s, which can radically change the k values (for example, flipping the sign). So that may be one thing to think about. Biosim had done a radical experiment around this.
DM – We may have some of the folks who did that work in our partner companies - eg olga inqvist.
Matt Thompson to Everyone (Oct 4, 2023, 1:45 PM) Somebody in OpenFF has done this at some point - I can dig up the paper trail if this is a path you all are interested in going down
David Mobley to Everyone (Oct 4, 2023, 1:45 PM) Ah yeah I was thinking someone had done that.
Pavan Behara to Everyone (Oct 4, 2023, 1:47 PM) for double-exponential if we optimize scale14 with QM data it drops to 0.33 from 0.5
LPW – (see recording ~45 minutes) … Interested to see what would happen if someone did a totally fresh start from like substituted ethanes. If you go from ethane to chloroethane, you know that there’s only one additional term… you could tease out some linear dependence in the torsion terms
Infrastructure needs
Jeff
Virtual sites – ForceBalance > 1.9.3 can’t work with virtual sites, restricting any fitting stack to the toolkit 0.10.x and old Evaluator
LW – Not sure where root issue here is - just something I’ve observed. I’m using the last version of evaluator compatible with OFFTK 0.10.x. This is throwing an error
LW will send MT and JW a reproducing example.
Essential for Thyme
Evaluator
HFEs, solvation FEs, transfer free energies
LW – These are all essential to a vdW refit/benchmark
Essential for Rosemary/Sage point release
Gradients?
LW – Not blocking research or releases
vsites?
(see above)
QCSubmit
BP said that he may make changes in minor releases that will prevent old QCPortal versions from talking to server. We may need to start storing datasets locally to do reproducible work.
LW – Have been exploring dimer energies (e.g. for ForceBalance) on an ad-hoc basis, but probably need a go-to dataset structure (akin to OptimizationDataset or TorsionDriveDataset) for continuous applications/integration into our software stack, e.g. benchmarking?
Not immediately essential
LPW – I think FB may already be able to do dimer energies
LW – It’s in for everything except SMIRNOFF. JHorton has a PR open extending thsi to SMIRNOFF.
LW – Just flagging – BW has been having difficulty reproducing the Sage 2.1.0 fit, possibly because of software versioning. This may be important to address.
BW – Still looking into this, maybe somehting with omega conformer generation.
LW – filtering out ~10% molecules of our training dataset, ~20 batches/180 (each batch with 30 molecules)
BW will continue investigating and present when he has conclusions to share