Shows b3 could be split into carboxylates vs non-carboxylates
DLM (in chat): Already I think I can guess where this is going: Carboxylates need their own bond parameter because resonance means it’s not fair to treat this as [#8X1+0]
specifically =[#8X1+0]
PB: Units?
BW: kcal/mol
b5: one group shows the C:C bond between two rings
CB: there are shoulders on the right-hand side, I think this is a goldmine
DLM: back to the big picture for a second, this is where Espaloma assigns different parameters to what we would assign. Do we have QM parameters for these? That may be the wrong question to ask, but what would MSM assign?
CB: TG, you said there weren’t sub-populations in the delta between the equilibrium in the deltas with the QM geometries – but it looks like BW is seeing that behaviour here in the Espaloma assignment of force constants
TG: this is pretty representative, but the problem is how we split the peaks out, and that’s hard to solve. My question here is – is there any reason we can’t see the equivalent Sage assignments?
BW: Sage assigns all of these the same value
TG: I agree that MSM could be a reasonable choice to get the spread of force constants with Sage
BW: Do we have Hessians in the test data?
PB: You can check on the training data
TG: with the MSM, could look at the geometry deviation and use that as a surrogate for this, since you can often kind of get the same answer by varying the force constant and the equilibrium length
PB: this is an aromatic bond – do you have points less than 400 kcal/mol?
CB: that central bond in napthalene is likely the least aromatic bond in napthalene and would have the longest bond length, would be interested in looking at bond lengths
LW (from Google Images):
CC (in chat): We found a similar problem where our FFs assigned different torsions to the equivalent carboxylate oxygens. Delocalized charges can't really be assigned to a bond order of 1 or 2, but they also aren't aromatic. It would be nice if we could specify a bond type in SMILES that means "not aromatic and not in [-, =, :]".
DLM: @Chapin Cavender — I think I’m missing something because … as far as the SMILES are concerned carboxylates are single and double, right? Like… are you just saying “I want an extension of SMILES that can represent resonance” [or delocalization]?
CC: Yes, an extension to SMILES. I'm making the argument that carboxylate bonds should not match the aliphatic - and = bonds in SMILES strings, and similarly for other non-aromatic resonance cases.
DLM: Yeah. But for now maybe we need to just say “the bond for carboxylates” and encode that in SMILES?
CB: How can we automate the determination of these different peaks/populations? Caitlin Bannan’s SMARTY tool was developed for this?
DLM: CBannan’s tool didn’t work for populations that had any sort of overlap, I think that would be a problem here
CB: could try clustering to address
TG: I’ve been working with BW to apply SMARTS on his Espaloma data. We’re now working on torsions
CB: I would love to see this analysis on equilibrium angles. e.g. bonds exocyclic to small rings. Also, looking at bond lengths, the force constants may form a narrower distribution, but particularly valence angles
CB (on espaloma FF slide): I wonder how much of these improvements is just improving e.g. the top 10% of results, I’m most interested in chemical understanding and the distributions you showed on the previous slide
Torsions
AMI
AMI: concluding that sterics are contributing a lot to the torsion profile shape
CB: for the Mol0 structure, at 180, that’s almost certainly an electrostatic clash between the oxygens. Raises the question of – when you want to filter out steric clashes, do you also want to filter out electrostatic clashes? But agree there’s definitely something steric happening at 0 degrees
CB: On t82 – we may have tried to put in too many kinds of chemistries here, e.g. not mix azides and the non-azides. Perhaps we should be asking if we should be putting them together
AMI: Yes, might start overlapping with BW’s project, that it may make more sense to figure out what to split up before moving forward with the project
CB: could you group by the minima, or the torsion values at the minima?
AMI: I’ve been trying to look for different ways to automate finding problems, that’s a good suggestion
AMI: planning to do deep dive into cyclopropane parameters
DLM: we have previously assumed that small rings would just end up with the correct geometries from bonds/angles, which is why we’ve previously excluded in-ring torsion drives of small rings. The takeaways here might be that 1) not worry too much about them, and 2) to treat in-ring torsions as a separate class
DLM: we have had issues with planarity of four-membered rings
CB: cyclopropane is kind of the worst-case scenario here, but moving up to cyclo4 and cyclo5, internal torsions may be a contributor, but with something as small as cyclo3, bonds and angles will be the biggest contributor here
AMI: looking at splitting out parameters, results should be in soon
TG: one thing I’ve learnt is splitting out 3-, 4-, 5-membered rings tends to happen first. So one thing you could look at is doing that as a first step, maybe by brute force
DLM: maybe the easiest way to check if we can generalize ring parameters is to de-generalize them
AMI: are you saying that some parameters that aren’t exclusive to rings?
TG: I’m saying there are, e.g. 4-membered rings, that are getting tossed into linear parameters. We have specific 3-membered ring parameters, but not so much for 4- and 5-membered rings
TG: Agree that this is very important, especially for valence angles. This should be differentiated completely
LW: TG, were you talking about bonds and angles, or also torsions?
TG: it would attack bonds and angles first, before getting to torsions, it ping-pongs between them.
CB: TG, does it really affect bonds much?
TG: it always wanted to do a 4-membered ring bond because it was “polluting” other bonds. They’re slightly tighter
TG: I have code to enumerate SMARTS, so if you need an enumerated set, just let me know
AMI: I’ll message you on Slack
Torsion MSM vs ab initio methods
CB
CB: wondering if there’s a difference from these two starting points
AMI: depending on where we start we can end up in a local minimum
(in chat) clarification that ModSeminario only does bonds and angles, not torsions so far
CB: my question is, if we start from a different initial value, will that have a strong impact on the solution
AMI: that’s what I’m trying to find out from this project, but I have trained a FF that only looks at small-molecule torsions, and then re-fit Sage from that starting point to see if that changes the performance of Sage
Filtering out steric clashes
CB
CB: we may need anharmonic angles to deal with steric clashes. I’m worried that our condensed phase is much nearer the minimum, I’m worried that our LJ doesn’t ask the right questions for the really repulsive parts. Where this affects torsions is the 1-4, they are scaled to ameliorate the steric element. When we start filtering out steric clashes, we don’t want to do it too early because it helps us deal with LJ, but we don’t want to do it too late, too high up the repulsive wall. How are we going to do this right?
AMI: Just to clarify, this isn’t part of my torsions initial values
CB: if we exclude over-repulsive LJ conformations, is it telling us that the equilibrium structure is not a good contributor, or that the LJ curve is too repulsive?
