MS – TB will be doing solvation and mixture calcs of small molecules. Would that be helpful?
LW – Yes, that’d be great. I can help set those up.
MS – Timing/details of leave?
LW – Out sept 9 - early October
JW – I think that the FE calcs may be so noisy that nagl charges wouldn’t make a significant difference
CB – ESP comparison is the major thing to look at. That’s what people are gonna look at to decide when choosing whether to use it.
LW – I previously looked at solvation free energies on bad outliers - That showed that the GNN charges do lead to a significantly worse prediction. So I think that PL FE calcs would be a useful source of information.
DM – I think, if the FE calc results look comparable, this should be good to go.
CC – Would we need to do a valence refit? It’s possible that we’d see worse performance. Would be useful to run the geometry/ddE benchmarking.
CB – Yeah, things like TFDs and dimer energies would be handy to check.
LW – You’ve convinced me that dimer energies should be part of our standard benchmark set.
MT – LW, what are your leanings?
LW – I see lots of the flaws with NN charges. I’m happy with peptide benchmarks but some small mols are worse. But the ones where we do poorly on are fairly exotic and I don’t think they’re realistically in scope. So PL benchmarks would make me feel a lot better. And valence benchmarks are much lower hanging fruit. But I’m so close to it that outside perspectives are helpful.
CB – People are going to say that this is being provided as a substitute for AT/OE AM1BCC. You already have HFEs calculated, you could plot GNN charge HFEs vs. OE charge HFEs. You’d hope to see a straight diagonal line, and that would convince users.
LW – We have done this for molecules with different charges - We see off-diagonals. The average RMSE for this plot is <1 kJ/mol. In this analysis it does do slightly worse than AT.
CB – I’d love to see that in the docs - And highlighting the molecules that did poorly.
