2021-02-12 WBO/Impropers meeting notes
Participants
@Simon Boothroyd
@Christopher Bayly
@Trevor Gokey
@David Mobley
@Pavan Behara
Disclaimer: Not verbatim, might be missing some important details, feel free to edit anytime
Discussion
Slides link
docs.google.com/presentation/d/1aufSxexsZRedfwMWcG8tsgsQuUCylpX8kDuSgN5bg0o/edit?usp=sharing
PB: Removed the strong sterics by considering the LJ energy difference between the highest energy and lowest energy QM conformers with openff-1.3.0 at these geometries. This new refit without the strong sterics shows no improvement over previous fit, fit7, but rather degraded performance. It is tested on a subset of benchmark set by removing the strong steric cases by checking the above metric among conformers present.
DLM: I wonder what the statistical errors would be in each of those bins on benchmarking plots, that would be helpful in having a better perception of whether we are really doing bad or not when the results are so close.
CB: I would still like to see some of the fundamental questions answered about the general applicability of Chaya’s wbo interpolated parameters.
PB: Sure, iirc on the training sets the performance is good wrt interpolated parameters, I can share a concise set that shows improvement over using a general torsion parameter, may be from the simple fit1 case with t43,44,45 replaced with one single interpolated parameter.
PB: Moving on to gradient analysis Simon made a great visualization tool (https://github.com/SimonBoothroyd/graffan ) that can show the structures for all these plots of gradients. One issue for me is that each torsion target gives rise to non-zero gradients for rest of the parameters too even though the torsion driven corresponds to only one parameter.
SB: Yeah, forcebalance optimizes all the parameters present in the molecule even though we want to train only one parameter. I am looking at the internals and there is a chance that we can instruct it to consider only the parameter matched using the pgrad tag. I will touch base with Lee-ping and see what can be done.
CB/DLM: That’s a really cool tool you made Simon, a treat to the eyes. Regarding the fitting process that’s news to us, sounds like for Sage we may want to look into it. On this example PB is showing, the gradient values look valid since one of them is an in-ring torsion, a higher value for TIG2, and the other TIG1d is not. One initial task is to remove the in-ring torsion from training. Coming to the point of looking at two different values for the same molecule may be tagging the torsions on selected molecules might help.
PB: Curious on the status of ELF10 for wbos.
SB: Going through reviews mostly by next week it should be ready.
JM: I have the results from comparison of wbos using ambertools and openeye, most of them match and there are slight differences.
CB/DLM: It is good to see these results, some of those slight differences may change the interpolated torsion by 5 kcal/mol, so we need to find the origin of these differences. Whether ambertools produces the same wavefunction as openeye or it does something different under the hood that results in such discrepancy.
SB: I can help JM with that, I have a few questions on openeye implementation, whether any optimizations are performed within the module that is not exposed to the user
CB: That’s a good question, yeah we do restrained optimizations, the intent of the module is to provide charges so we kill that data, shoot an email to support asking for a user flag to access those geometries.
SB: Thanks, will check that.
JM: I got the bespoke fit script from Josh and I am working on it.
DLM: Nice.
SB: I want to update that I made changes in forcebalance to perform 1-4 scaling factor optimization, we would soon have that capability too.
DLM/CB: Excellent!