(General) GAFF performs really well here. What specifically is it doing well?
(General) – Bond and angle terms?
DM – We haven’t done hessian-based fitting. That could help a lot. Also using optimization trajectories. Also using high-temperature MD is likely to stress bond and angle geometries a lot.
DC – Interesting that GAFF is doing better against OUR QM data than we are.
JH – Agree that using optimization trajectories should improve the quality of these fits. I think the torsions here are nearly perfect – I doubt that we can improve any more without refitting bonds and angles.
DM – we should try to fit to the optimization trajectories
JH – will try.
DM: Try to modify the k values for the bonds/angles; use existing parameters
DR – Do you minimize snapshots from MD before getting MM and QM energies?
JH – No minimization
DR – IIRC, normal mode sampling was used to generate confs for ANI (maybe with minimization in the middle)
DC – We’re using a lot of ANI for our methods, generating confs using up to 1000K, seeing good agreement between ANI and QM energies.
DR – mol size?
DC – Small fragments
JW – what FF was used for the high temperature MD
JH – 1.0.0
JW – this might bias the conformations generated. Maybe we should use something like UFF instead to compare with oFF and GAFF
DM – GAFF might also be doing the same
JW – doing minimized conformers would remove bias, but maybe expensive
DM – how quick can you generate these plots?
JH – about 3000 single-points/day
Should I try doing the MD with ANI instead?
DM – pursue it if experiments won’t take much time
Comparison of ANI-2x energies and QM
ANI2X comparison to QM energy looks better, but there is an issue with convergence
DC – What’s the status of issues with ANI torsiondrives?
JH – Complicated. They’re retraining it.
JW – What was the problem with this?
DR – The CELU function in one layer of their network led to nonsmooth second derivatives, which choked geomeTRIC up.