JH – For simple molecules, there isn’t much of an improvement, but the energy profiles were already pretty good.
JH – For somewhat congested molecules, the restraints don’t make much of an improvement
DM – Interesting that such small changes to torsion parameters lead to large (0.5A) RMSD differences.
JH – One hard case seems to be things like phenyl rotations, which encounter symmetry problems.
DM – This suggests a general fitting experiment that we might be able to do – What if we use RMSD in a target? Could you add this to the FF reelase call agenda tomorow?
TG – Have you tried combining this with optgeo targets?
JH – Not yet, I’m adding that and hessian fits to bespokefit.
JW – In general, we have this question where “FF gets the right geometry” vs. “gets the right relative energies”. I wonder if the use/strength of restraints lets us scale between those.
DC – I wonder what the outputs would look like if the restained fits were evaluated using nonrestrained minimization
DM – We have to avoid two cliffs: One if where we don’t let the fits resolve steric clashes, but we wind up with terrible energetics. The other is where we get the energies to agree at the QM minima, but the MM FF has minima somewhere else.
DC – Would be good to analyze a alrger set to be able to speak with more confidence
TG – Peptides are tricky on this front.
TG – Can we say that the only difference is that the third case on the page is due to the symmetric ring?
JH – Yes, pretty much.
DC – could you clarify what optgeo fits mean?
TG – It does an unrestraind minimization and tried to match the QM structure. Also it uses internal coordinates to compare geometries.
Cyclohexane
DC – When we do torsion fits, we inherit torsion parameters from the parent force field. We’ve been looking into hvap and other properties of saturated rings. We checked out the torsion profiles of the ring conformations and both QUBE and OpenFF get good conformational energetics. So the issue is likely bad nonbonded parameters.
(clarification) Used QUBE nonbonded parameters and OpenFF valence parameters
DC – I suspect this is because RDKit was making strained initial conformers that can’t convert to the lower energy conformers. So we’ll try again with omega conf gen.
DM – Re: boat/chair transitions, I’ve seen that sometimes the barriers to interconversion are so high that you can’t get them to happen by just driving a single torsion
DM – Also, in your particular case, the nitrogen in the ring can be very hard to model. Nitrogens can switch their preference between planar/tetrahedral based on a number of complex factors. There may be some value to having bespoke parameters for nitrogens.
DC – One thought it that we could make all the propers that terminate at the N have phases of 180 and let the improper can overwhelm them if a different geometry is preferred.
DM + JM – We found it’s good to have the fits for tricky nitrogens be allowed to change both the propers and impropers.
Further bespoke fits for OpenFF studies
JM – Rough computational cost?
JH – For whole torsion set, about 10 minutes per molecule
JM – Planning on using about 1000 torsiondrives
JH – May be possible to break it into a supercomputer job
DC – Could the barrier height be extracted from the QM data?
DM – Hard to de-convolute contributions from eg sterics and electrostatics
JM – When doing bespoke fitting, is it possible to fit torsions for a single rotatable bond or does it always do every torsion for all rotatable bonds in the molecule?
JH – It is possible to do just one rotatable bond. I’ll show how to do this in the notebook I send.