NAGL update | LW | JR – Why size 3-6 for rings? What happens with rings of 7 or 8? LW – That’s not included - There’s a separate feature for each ring size. DM – Idea is that, once rings get big enough, it becomes irrelevant whether it’s in a ring or not. LW – Earlier iterations included ring sizes up to 8, but we foundd that 7 and 8 weren’t important. MG – And this uses electronegativity and hardness? LW – Yes, and additionally initial partial charge guesses
(slide 6) … PB – How big of a difference is bad? LW – Answered on next slide. Under 0.01 is probably negligible DM – MOsato in my group sometimes finds that even small differences in charges can have big effects on properties.
(slide 8) DM – … PB – So is this still ELF10? LW – Yes PB – Is it possible that the ELF10 conf gen might include both diastereomers? LW – OE should respect the stereochemistry, but I’m using an older version of the toolkit that may not respect this. JW – MS – Shape can also contribute to HFE a lot LW – Wanted to quantify the scale of differences within OE ELF10 charges for different diastereomers. CC – What’s the precision on these calcs? Eg, if you did diastereomer 1 twice, what’s the RMSE? PB – Do differences scale with molecule size? DM – Is this result surprising? LW – AGobbi has asked what the scale of these differences is at the leiden meeting. DM – This might suggest two experiments. This has changes from charges and geometry. So what if you swapped charge sets, so diastereomer 2 conf would have diastereomer 1 charges. Right now you can’t tell which contributions to these values come from electrostatics OR shape, so this would let you measure effects from those. LW – That’d be a good experiment to run.
(slide 9) (General) – Enantiomers should have identical HFEs. DM – I’d guess this is a sampling problem. And I don’t recommend going too much deeper on this. MG – Could use additional torsion constraints to fix the carboxyl torsion so sampling converges faster. LW – Could remove carboxylic acids to simplify the experiment. DM – Worth considering. MS – Would also be good to look at more enantiomer pairs to assess dataset quality or get some rough validation. DM (chat) – You could restrain the carboxylic acid -OH to point one specific direction so it stays stuck in a consistent direction. At least that would be the easiest way to “fix” the sampling problem. DM (chat) – Figure 4, this paper, has the PMF: https://dx.doi.org/10.1007/s10822-010-9343-7
JW – These are very interesting results - Raise questions about whether we were right to require stereochemistry in our infrastructure (since it’s really only used in charge assignment and conf gen)
(slide 14) (slides 18-19) - See recording, ~43 mins MS – We should make firmer plans to write everything up MG – Looks like it’s performing well. Is there a plan for systematically finding more issues? LW – Can keep exploring, depends on what people want to see. I’m mostly GPU-limited right now. MG – Is there any way the model could emit some measure or trustworthiness? LW – Could look into that
JR – Could compare dipole moments? Might be a validation test. DM – I’m thinking this looks very good, and we probably don’t need more HFEs. We’ve shown pretty well that this is better than our open-source alternative JR – How well do you think you’re doing with hypervalent S and P? How many are in the training set? PB – Right now we have the openff-nagl and openff-nagl-models repos. Do you think making a monorepo is a good idea? Or is there a reason to keep them separate? LW – Good question. I’ve kept them separate for similar reasons to the distinction between openff-toolkit and openff-forcefields. For example for making a NEW network, you don’t need the models repo, and the models are about 10MB a piece, so that could add up after a while. JW – Also can get confusing when code and models live togeteher, and a release of a new model implies a change to the code, though none is present.
|