| Torsions: QUBEKit compared to BespokeFit | @Daniel Cole @Joshua Horton | JH – I’ve got both pathways working. Some systematic differences due to inability to fragment automatically, different regularization, DC – Using modified seminario in bespokefit would be equivalent to fitting bonds, angles, torsions all at the same time? DM – Simon Boothroyd is looking at changing 1-4 scaling factors during fits. JH – I had seen that work and could give it a try. DM – Thought is that you could refit the LJ terms, but maybe refitting 1-4 scaling factors would have the desired effect. TG – I had found that, if I try to refit LJ terms, the values would change by unacceptably large amounts (relative differences of like rmin_half>1.5x) DC – I could see this getting bad if we don’t have liquid properties in the mix. DM – Agree that this would invite overfitting.
DC – It’s good to see that QUBEKit and Bespokefit are giving approximately the same results. It’d be good to have a closer look at this in a subsequent meeting. It’d be good to look at a large variety of molecules early on. JH – Not sure whether bespoke nonbonded terms would make a difference. DC – I’d expect bespoke nonbonded terms to do OK in liquid simulations, though I’m not sure that they’d be better than the original. JH – Would we consider including liquid properties in bespoke fits? DC – That’s an interesting possibility.
DM – So, at a high level, we have bespoke fitting working, but we don’t have many examples of clear improvement? JH – Yes. JW – Would it be possible to reuse the pharma benchmarking public set to measure performance of bespokefit? Or are underlying mechanics so similar that it wouldn’t be a valid comparison? JH – Seems like it’d be a good comparison, and it’s what I’m planning on. DM – Could segment pharma benchmarking set into “molecules for which FFs do well”, “molecules where FFs do OK”, and “molecules where FFs do poorly” → Could focus on the latter DC – I think a key dataset would be the FEP results from Chodera’s group (the same as was used for ANI study). IIRC OpenFF was ~1kcal/mol error, and ANI was 0.5 kcal/mol error, so I’d be interested to see where bespokefit falls. JH + JW – That would be a much better dataset to start with. JW – What would be the training data vs the test data in this case? JH – The plan is to use the train a bespoke-fitted FF using the existing QCA data, and then test it using the FE calculations. DM – If we don’t want to redo the whole simulation, we could just do endpoint perturbation calculations. This would be interesting because instead of changing the ligand, you’d be changing the force field.
DM – JM, have you been able to get bespoke fitting running locally? JM – Some environment issues, would like to do screensharing with JH JW – I’d also like to join that session (working session scheduled for 8 AM Pacific on Friday)
|
5 mins | MBIS electron density partitioning available in psi4 | @Daniel Cole |
DC – This support for electron partitioning lets us directly get bespoke nonbonded parameters. Previously our electron partitioning code was hard to deploy/resistant to conda-installation. DM – Esteban Vohringer-Martinez is on the OpenFF slack and is also interested in this area. They did some reparameterization based on MBIS partitioning, which lets them do as well as AM1BCC, without fitting directly to experimental data. So this could be a good way to directly get charges without the use of a fitted method in the middle. DC – Yes, I know Gilson also used MBIS in one of his papers. JH – This is promising, since we’re starting to store the wavefunction for QM calculations, so we’ve already done most of the work for these. DC – That’s exciting, though it’s unfortunate that they’re not in implicit solvent. DM – If we were going to get implicit solvent outputs for the same dataset, do we have any advantage from having done the optimizations in vaccuum before? JH + DC – Possible, some complications. But having the optimized geometries ahead of time should in theory provide a lot of savings.
|
10 mins | Check on content for advisory board meeting | @Daniel Cole | |