OM – In “triangle plots”, make sure to fully explain right when you get to the slide
OM – On histogram slide, highlight the “percent-past-cutoff” number, maybe put a bright red box
Slide 13 and 15 – Fix modified method name in legend, cover figure title with something better not refer to “previous two plots”
Slide 21 – Change cutoff to 0.05 if it’s easy
Talk feedback
(Talk given in succeeding hour’s all hands meeting, notes taken at 2021-08-11 All-hands meeting notes ) (below points copied from those notes)
DM – I think the real “bake-in” will be in the vdW parameters. Maybe the thing to ask is “if I do an optimization with an OpenFF typical level of QM, then I fit RESP charges, how well do those correlate with different AM1 strategies (with some weight also toward considering keeping HFE accurate)?” Generally I’m nervous about using OE as a standard
OM – Previous reference for most of the field was HF-6-31G*. This has systematic problems like overpolarization in gas phase, but this is actually good because we DON’T run simulations in gas phase, we run them in solution phase. And when you use a BETTER QM method in solution phase, the results look like 6-31G* in gas phase. So it’s right for the wrong reason.
JW – Let’s think about how this fits into our current context. We have two problems that we can try to solve and a variety of methods to solve them:
Problem 1: What do we do when there would be a connectivity rearrangement?
Problem 2: What do we do when there WOULDN’T be a connectivity rearrangement?
JW – Is this in scope? If we decide that unrestrained AM1 in antechamber is the gold standard, then… we alreayd do that. IF we decide that something else is the gold standard, then how could we expect AM1 to reproduce something fundamentally different?
CD – Maybe our gold standard should be measured as “Getting correct MD results”
JW – Can we actually try to choose a good charge model to reproduce MD observables WITHOUT retraining the entire FF? It seems like any new charge model we try to propose will be plagued by the fact that the FF was trained with the OLD charge model.
CD – Right now, the FF was trained using OE charges. Long term, the only way would be to retrain FF repeatedly to try and reproduce observables.
CC – Also, I’m curious, when you do the MM constrained minimizations, I wonder how those results would change with a solvent model
DM – Could we set up a electrostatics meeting with Simon, Bayly, and Schauperl (if we can get him)?
OM – We should do this / get Bayly’s feedback on this.
Update project map?
Next steps
OM – Could pursue one question about reproducing “gold standard”, and another questions about “ideal restraint schemes”
CD – Agree.
JW – For first point – reproducing “gold standard” in cases where connectivity doesn’t rearrange – I’d need a lot of convincing that vanilla antechamber AM1 isn’t the gold standard. I don’t think there’s a “more golden” standard for AM1 charge generation than vanilla antechamber.
CD – Conformer dependence? If the user gives us one conformer, what if a different conformer would produce a better set of charges? Would a good partial charge method in that case choose a DIFFERENT conformer? How do we control for that?
JW – One assumption that we’re making in all of this work is that all conformers of a molecule should give “good” charges
CD – Maybe differences in the geometry minimization are the cause of all the partial charge differences that we see between antechamber and openeye.
JW – I think this is the case
OM – We should talk with Bayly, Boothroyd, Schauperl about this. Especially we should show bayly the cases where there aren’t connectivity rearrangements, and there’s a big difference between OE and original.
CD – I can show him the input conformers and the antechamber final geometries, but I can’t show him the OE final geometries.
OM – That’s fine, Bayly can look at those internally.
JW – We’ve reached a good point for a deliverable for the toolkit - the maxcyc=0 fallback is now statistically justified. CD made
CD – GeomeTRIC long term outlook?
GeomeTRIC isn’t looking great. Current implementation is very slow, tends to explode, and can’t handle many restraints. So it may be good to look into doing something different, and possibly make the tools to check these directions out. Or I could look into geomeTRIC and see if there’s some way to make it more stable, or make it run sqm/antechamber cycles faster.
JW – Kinda three major direction to proceed:
Fun, like cross-compiling geometric with antechamber/sqm to make something fast, but that may never reach production
