CC: Could you interpret these plots on slides 2 and 3 more? What does the tan colour that corresponds to 7 mean?
TG: everything that’s tan is >7. The baseline value of no virtual site is ~7.1, so truncated the colour bar there.
DM (slide 2): So the virtual site performs better when it’s far away from the molecule?
TG: Yes, so this is clearly not physical
DM (slide 3): Chemically, taking charge from carbon 2 or from carbon 3 should be equivalent. So why aren’t the plots symmetric?
TG: answered on slide 4. I think the ESP grid points are not distributed symmetrically around the molecule.
LW – Method I was used could have masked this, there was some averaging happening… Treating symmetric atoms symmetrically would intuitively be an improvement.
TG – Yeah, some issues could be seen to arise from non-alignment with a grid. If I were to redo this, I’d take the SMIRNOFF types and see how much they differ, and then force equivalent atoms to be treated the same way.
DM - Enforcing symmetry seems desirable but potentially difficult. Should we change the grid to make it less sensitive to needing proper symmetry?
LW - We can change the grid without recomputing QM. My fitting method in NAGL already accounts for this, so I don’t think I’m worried about this being difficult.
TG - As a stopgap, we could add reflections of the grid across each Cartesian axis to force grid points to be symmetric.
LW - This might cause many grid points to cluster together in space.
BS - You should also be able to resolve this by using a finer grid. It might be a more precise test of the impact of the virtual site to only lookat grid points close to the site.
Sage 2.2 update
LM
JW (slide on 4-membered rings) - Is this an issue in Sage 2.1 or the release candidates for Sage 2.2?
LM - This is an issue in the Sage 2.1 release. Sage 2.1 improves metrics for entire molecules containing problem chemistries (e.g. sulfonamides and 4-membered rings with heteroatoms), but you can see the problem when you look at the sulfonamide or ring internal coordinates specifically.
LM - Conclusion is that reverting to Sage 2.1 angles gives better ddEs but reintroduces bad geometries for 4-membered rings.
JW - It seems like for molecules where there’s a small ring linking two other fragments, you can get the geometry of the ring wrong without screwing up the ddE for the entire molecule.
LW - RMSDs can mask small problems, so I think looking at internal coordinate RMSD is better
CC - BS and LA’s work showed that ddE stops being meaningful when the geometry changes too much. So I’m in favor of the variant with correct geometry over the one with incorrect geometry and better ddE
PB - Were the 4-membered rings modeled well in any of your FF variants?
LM - Yes, they are fixed in my Sage 2.2 variants, e.g. a7het.
PB - Can you transfer the parameters for 4-membered rings from that variant to Sage 2.1 but then don’t allow that parameter to change during subsequent optimization?
LM - I can do that, but I’m worried that this violates our fitting philosophy. Maybe it’s fine for diagnosis.
JW - I can’t imagine what term in our FF would be happy in the pyramidal geometry. Our angles have an equilibrium near 109.5 deg, so they should be unhappy there.
LW - Pyramidal is unfavorable, but there’s a local minimum compared to even more unfavorable intermediates
TG - Are methane hydrogens constrained?
LW - No, and the hydrogen bond lengths can become quite distorted.
DM/BS - This could be related to the explicit solvent/PackMol setup