Recap | Monday: TG/OM/CD/JW meeting, wanted to find a reference dataset for charges, nothing was available Tuesday: OM/CD/JW/SB meeting, discussed: CD – Worked on getting GeomeTRIC to successfully have a proton transfer. I wasn’t able to do have geomeTRIC reproduce the transfer with Swope’s original molecule. But I was able to get it proton transfers to happen with other molecules. However these new cases have shown proton transfers, but then Antechamber/sqm has force errors. Could be a whitespace issue, where the force numbers get too large and exceed their intended column width. JW – If the proton migrates really far away and there are torsion restraints, it could lead to a really long “lever arm” for the torsion, leading to huge torques, or it could make the torsion kinda linear, leading to a “gimbal lock”ish problem
CD – I’m not sure that this would explain what I’m seeing. GeomeTRIC can only use hard constraints, not soft restraints. And it applies these constraints by ignoring some forces form the underlying engines that would lead the constraints to be violated. CD – So, to recap, one of our goals was to apply a soft restraint to bonds and angles. But I don’t see the functionality to do this in geomeTRIC. So I’d need another engine to add these soft restraints. (this seems to be the case) JW – This may not be high-priority. If this problem only appears when there’s a proton migration, and we know we’re going to throw out th calculaiton if there’s a proton migration, is this an important problem to solve? CD – If we fix the positions of the atoms involved in the migration, would this solve the problem? (this is equivalent to putting a hard distance constraint between the atom that loses the proton and the atom that ends up getting the proton) JW – This would be a good constraint scheme to try. Or constraining the distance between the mobile proton and the losing/recieveing atom, since that would be the most direct solution to the problem OM – Could it be a good idea to constrain the proton and the atom it initially belongs to?
CD – I double checked that the molecule we saw breaking into fragments the other day really was correctly input and charged. (CTerm His-Asp was input with a charge of -1)
|
Next steps
| If we use a hard constraint, which atoms should it involve? CD – I’m in favor of the heavy atoms that gain/lose the proton JW + OM – I’d like to include a constraint on the mobile proton and a heavy atom CD – What about heavy atom migrations? In that case, I’d probably need to constrain the bond that breaks, or the distance between the atoms for the bond that gets made.
Why isn’t geomeTRIC converging ever when a hard restraint is applied? How can we make geomeTRIC use soft restraints? Which connectivity change detection method is best? Is antechamber AM1 w/ maxcyc=0 giving the same answer as OE with optimize=False? What reference charges should we match to? What should CD present at Friday group meeting / what should be in Aug 20 project summary? OM – Good to present overall picture without getting bogged down in details. I think a good presentation would be starting with a video, saying “proton migration and heavy atom migration are bad and lead to bad AM1 charges”, then talk about how the dataset was made and the methods for automated detection, and then talk a bit about the methods that you’re using to try and correct it. The audience will generally be less FF experts than JW/SB/myself, so expect that a lot of the time will be spent on background, and that the feedback will be higher-level. CD – I’d like to explain why it’s hard to define “reference charges” JW – One component is that we basically know that AM1 is “wrong” compared to an ideal QM-based reference. But AM1 is cheap time-wise, and the corrections for its errors are “baked into” the other components. So it’s possible that, if we tried to use our FF with “better” charges, it’s likely that the results would be “worse”. OM – A similar thing happens with water models, where if you train with TIP3P, but try to do a simulation with a better water model alter, you’ll probably get worse results. CD – Also, it’s hard to define what a representative set of conformers is for a molecule, and the question is kinda circular because any practical method to explore conformers is dependent on the assigned charges. OM – Additioally, kinda building on JW’s initial point, if someone brings in a new molecule (that doesn’t look like our training set), we want a method that will do something consistent with it so that our parameters are likely to apply.
|