| | @Jessica Maat (Deactivated) | JM: I looked at QCA and ChemBL, and for both the wbos are dense around 1.01 to 1.02. DM: Yeah, it doesn’t look like it spans a wide range. JM: There are 100,000 matches in e-molecules, will be doing that now. DM: So, there are only a few matches in ChemBL. JM: Actually, there are some failures in the ChemBL matches. DM: Do you know the reason? If the failure rate is higher there is chance of unspecified stereochemistry, or failure to generate conformers, etc. Pick up some of those smiles and see why it might be occurring since you’re seeing a large fraction of failures. JM: Yeah, will check on that. SB: Sometimes the toolkit makes one up if it encounters unspecified stereochem, may be try-except when it fails and try to generate a conformer before trying to compute the charges. DM: Yeah, may be use omega and generate conformer. SB: Yeah, you can use toolkit generate conformers before calling compute_charge. DM: Yeah, even with e-molecules there should be less than 1% failure rate IIRC. Also, if it is still very high let’s look at the chemistry.
|
|
|
| PB: I did some fits with the binned targets and biphenyls, still not a good interpolated fit. DM: Did you check the priors? PB: Ah, I used the defaults, will redo these. That’s definitely a blocker for picking up higher values of k in high wbo region. PB: Apart from wbo work, I have some other fitting experiments with ICH, amides, etc., going. And I wrapped up the QM theory benchmark work, here are final results. SB: One idea about benchmarking, different from this. I have read complaints about training and testing on the same theory level QM data would be a bit skewed, it would be interesting to see how that would affect our benchmarks. May be we can switch to more accurate level of theory for our training set. DM: That's a great idea. We could think about it in near future. PB: The new openmm QM datasets would be at a different level of theory and we may use some of them for testing. SB: Josh Horton is also doing some benchmarks with XTB and ANI for bespoke stuff, it is ongoing and keep an eye out on that. DM: Also, when it comes to polarizability the answer may not be the same Paul Nerenberg might have something to say on this, post the results on slack and check with the electrostatic folks.
|