2020-11-12 Force Field Release meeting notes

wbo interpolation

dataset question

pavan/jessica

• JM – We’ve been working on fitting torsion terms, based on SMIRKS proposed by CBy.

PB – Looking at WBO vs. barrier heights

• CBy – TIG3 barrier separation may be due to SHORT conjugated alkenes, where some bond remain more double-ish and single-ish.

• CBy – Reminds me of tetraphenylbenzene, where naive ffs kept barrier pretty low

  • DM – Yes, but important to recall that we’re splitting out ring bonds.

• CBy – Largely concerned by the scale – The lower WBO cluster for TIG3 ranges from barrier heights of 5 to 80 kJ/mol, and we’ll assign all of them a barrier height of 25

• CBy – TIG7: The WBO range is super narrow in these datasets. Maybe there really isn’t WBO-dependence in these sets/for this torsion.

  • CBy – I'm thinkng that we do want a WBO dependence for TIG7.

  • CBy – From the Rowley data I’ve seen, TIG7 should be one of the really valuable central bonds.

• CBy – Two big principles

  • THe ones jessica is working on/the relationship with other valence terms

  • When we see a cluster of points like in the TIG7 plot, it means they’re not very dependent on WBO, and the parameter may need to be separated out

TIG8

• CBy – This is a 5-membered aromatic ring connected to a pyrrole.

• CBy – Looks like we don’t have enough data.

Additional datasets?

• JW – Protein fragment datasets may be good

  • DM – There will be a ton of redundancy there, right?

  • JW – Yes, so maybe it’s not a very good set

• CBy – A dataset that we’re going to use should pass some conditions:

  • It should have a decent range of WBO values for the bond of interest (>0.2)

  • Should have a torsion barrier range that’s worth looking at.

  • All involved training sets should show the same variation (the range shouldn’t be due to systematic difference between datasets)

  • Is one dataset’s X and Y range strictly inside of another’s?

• DM – Should split this into two problems:

  • Which SMARTS should we use?

  • Which QM dataset should we use for fitting it?

    • Basically make sure we ingest a lot, but avoid Lim benchmark molecules

  • CBy – Agree. On TIG6 plot, I’m optimistic about the green/”Substituted phenyl” set. Those don’t have many substituents at the atoms adjacent to bridgeheads.

• (General) – These “barrier heights” are the relative QM energies – No correction has been made for MM energy components!

  • LPW – We’ve often discussed subtracting certain MM energy terms from QM energy profile to try and distill out contribution from other terms.

JM – My plans

• JM – Calculate out other energy components, and do single-molecule fits

• DM – If we use FB to fit, then the torsion barrier height is implicitly done including the other MM energy components. Then we could plot the barrier height from the MM energy profile.

• LPW – That sounds like a reasonable approach.

• LPW – One idea that I had is: We could fragment an individual molecule for which we’re trying to calculate a torsion barrier (every point in the torsion drive), then cleave the molecule and calculate their noncovalent energy.

  • DM – How would we deal with dangling bond? There wouldn’t be room for capping groups.

  • LPW – Agree that we’d need to cap, but not sure how.

  • JW – Use SAFT/SAPT?

  • LPW – Agree that that would work. That would do 3 calcs – fragment 1, fragment 2, and both together. Then it tries to decompose energy into components, but which may not be meaningful.

  • DM – Agree that LPW’s approach would be valuable, though I’m not sure how to implement

  • CBy – Agree. Would like to think more about decomposition idea. Very concerned about error in energies due to sterics. Maybe a first pass would just remove high-steric-energy mols. Could use MM energies to determine that.

• HJ – Wanted to ask about fitting plan. If you fit parameters to torsion profile, there are a ton of knobs to set (eg, can other bonded terms change?).

  • DM – We only wanted to fit one torsion in each case.

• DM – Plans for moving forward:

  • Energy decomposition idea is good but won’t be ready on sage timeline.

  • Do MM calcs and filter out molecules with high-magnitude vdW or Electrostatic energies.

  • Could do the following:

    • Start doing test fits.

    • JM’s work on getting MM-fitted barrier heights.

cis/trans preference of amide

@Hyesu Jang