Host-guest validation for bespoke workflow | JH – OpenEye toolkit can’t fragment cyclodextrin. DM – Showed “drugs of abuse” host-guest data slides, GAFF vs. Parsley comparison. JW – Could we run bespoke workflow on one monomer? JH – Also looked at TYK2 system (part of FEP benchmark set). Thought is that I’d make a full FF based on ANI2x scans and see if performance improves. Unfortunately the core structure that’s shared between the series doesn’t converge in ANI2x. Which is unfortunate, because the scans that DO converge look really good. JW – Could we massively relax convergence criteria to get ANI scans to work? DM – What about XTB? DM – We should make a plan that doesn’t rely on ML potentials.
Congeneric series. JH – In my current test sets, many ligands have a common core, but they’re redundantly parameterized. However, ForceBalance sometimes produces slightly different results for identical inputs. It’s not clear how we can combine these results into a single JW will send JH a code snippet for concatenating FFs which allows duplicate parameters (ignoring subsequent matches). DC – JH – (Describes how chemper works to make the SMARTS for the desired torsion, by asking for the overlap between the whole molecule and the fragment)
JH – There’s sometimes a large number of parameters to optimize (60+ in some cases). Could be reduced by having logic to identify common cores and only do them once. I also think it’d be helpful to use Parsley parameters as an initial guess (as opposed to 0) DM – Using Parsley as initial guess may not be perfect, but it’s probably a better guess than 0. JH – Also using Parsley to have initial guess at periodicities and phases. DM – Probably also a good guess JH – Could allow for all periodicities, but this makes the search/optimization space really large. DM – Would be nice to have some measure of difference between Parsley and QM, and use that to guide parameter fitting/derivation strategy. JH – Agree, I was thinking along these lines. I’ll do an experiment where I refit them using periodicities from Parsley, and another where I use all periodicities. DC – We’ve also found that regularization to OPLS is generally useful, but it would be interesting to do this experiment.
JH – Regarding idea for benchmarking where we fit FFs to ANI and then use it to rerank conformers, I think the ANI instabilities make this unfeasible in the short term. ANI issues? DM – Yeah, short-term work shouldn’t rely on ANI. Could use default method, or anything else kinda fast and “accurate enough” JH – Will continue testing with XTB. JW – Could submit calculations on default level of theory on common ligand set
DM – Re: deciding on default level of theory – LPW just needs to make this decision, but it seems to have fallen off his radar. Discussion is in #benchmarking-partners. Dotson is pushing for a decision by Dec 1. JH – This sounds good, we can change the method and basis later if needed. DM – For background, we had decided on a “default method” (B3LYP-DZVP-BJ(?)) for all our datasets. But we acknowledged that this was a short-term decision and that we’d need to revisit it for the long term. Especially with regard to polarization + charged mols.
DC – As of a few days ago, CR is working on functionality for vsites in GeomeTRIC. This is so we can do torsion scans in OpenMM w/ vsites. DM – If QCEngine is using geomeTRIC for torsion drives, then we’ll need it to support vsites. Could loop in Trevor Gokey to help with this. DC – As of yesterday, geomeTRIC scans with vsites weren’t converging. DM – CR should tag Trevor Gokey for code review or feedback on this. DC – We’re using the LocalCoordinatesSite type of vsite. DM – Would also be good to loop in Simon Boothroyd on this, since he’ll also be interested in how vsites slot into fitting machinery. JW will loop together CR, TG, SB, JH to touch base about vsites in geometric/other fitting machinery.
|