JH – Pulled into some benchmarking stuff. I expect that participation in this will be useful to the bespoke workflow
DM – Victoria Lim has been doing some quality checking using OE geometry based tools
HX – We don’t do these quality checks, though I’d like to go back and look at this in our datasets
DM – You’ll be able to see the WBOs on all of our entries in QCArchive
HX – Where will this functionality live? OFFTK?
JW – I’m not sure
JH – WBO checks will be in QCSubmit. But they’re a bit slow due to having to do multiple queries to dig down into WBOs
DM – Trevor gokey knows about faster downloading
DM – HX, where do the molecules in the torsion set you gave us come from?
HX – It’s a synthetic druglike set, where we’ve reassembled fragments of druglike molecules. We’re trying to diversify the types of torsion that we’re considering in our FF.
JA – Is the “benchmarking” infrastructure just for benchmarking, or is it also for fitting?
DM – We anticipate it to fill the need for benchmarking, but it will be useful other places
HX – We hope to run enough quantum calcs to train a neural network that can generate parameters
JA – Have people tried doing tabulated torsion potentials?
HX – In our experience, it doesn’t give a big advantage over fourier series
JW – We’re using the word “benchmarking” in a weird way ehre, and it’s coming up in this meeting because we mean “the infrastructure to pull down QM molecules into MM-land, and compute their energies”. In the future, we’ll have a better word for this, and it will probably correspond to a concrete piece of infrastructure.
(Discussion about use of WBOs to sidestep problems in graph representation and interpolate parameters)
JA – With the bonds, there’s a synthetic perspective and a phys chem perspective. A synthetic chemist will say that a double bond can’t rotate. A physical chemist will say that it has a well-measured barrier to rotation. We want to move toward the physical chemist perspective, and consider the barrier profile as more important than the “bond order”
DM – TFD is an accuracy metric for molecular geometry like RMSD or shape tanimoto. RMSD doesn’t have the same meaning for large and small molecules. Shape tanimoto is highly sensitive to movements in central “hinge” regions. TFDs measure torsion differences, with extra weighting toward central torsions.
Add Comment