2022-01-06 Force Field Release meeting notes

20 mins

@Willa Wang

• Update on the polarizability work

  • transferability of fixed charges across conformations with and without direct polarization,

  • and applications using Sage with OpenMM

• SLIDES - https://docs.google.com/presentation/d/1bUhM3a8HOWY-c9gWdPSKevwtqdTfd3IX0KIAniAHoDE/edit?usp=sharing

• slide 2

  • CB – So, each point on the x axis is a separate conformation?

  • WW - Yes

  • CB – Does each conformer get different changes in any lines/series?

  •  

  • CB – So, you took the charges from the most stable conformer, and applied it to all the confromers?

  • WW – Yes

  • CB – …?

  • WW – Multiconformer RESP was calculated using psi4 plugin.

  • CB – I’m not familiar with that plugin specifically, but my thinking is that multiconformer resp should use the best… conformers for both methods

  • WW – I think I didn an average of the charges

  • CB – It’s important not to do a linear average of the charges, but rather fit to the electrostatic potentials of the conformers simultaneously.

  • WW – I’ll look into the psi4 plugin to see how it does this

  • CB – Main point seems to be that, with direct polarization, there’s a big advantage for the charge models.

• Slide 3

  • CB – For green line on the left, are these charges recalculated for each electric field strength, or is it a single set of charges?

    • WW – Single set of charges

• Slide 4

• Slide 5 (Solvation FEs for ethanol)

  • PN – What’s being varied here?

  • WW – 1-4 electrostatic scaling

• Summary

  • DM – Do you plan to do more HFE comparisons?

  • WW – Yes, I’m planning to do more experiments in the future, have a plan that I didn’t show.

  • DM – If you switch the charge model it may require compensatory changes in LJ parameters.

  • PN – Agree.

  • PN – Question about first results for alanine dipeptide. Ignoring the solid lines, it looks like we get a big advantage by ignoring point dipoles at low electric field. But then at high charge, the advantage goes away. But maybe this would change in molecules with more conformational flexibility. I’m curious as to what’s going on here.

  • DC - I wonder if that’s similar to the LJ parameter correlation comment above. Like, if this is in a big bucket of terms that are combined to reproduce strain

  • PN – I don’t think so.

  • CB – This reminds me of the problems I had to solve by introducing ELF charges. Because, if I’m going to come up with a multiconformer RESP, it matters a lot which conformers you use. So I wonder why the quality-of-fit falls off in the right graph, and it’s really important to understand how the conformers that are selected for multiconformer RESP affect the results. Also, these results show that the contribution from the fixed direct polarization is way more important than the charges from the partial charges that vary with field strength. In short, we could have just set the partial charges to 0 and seen the same results, since it’s the point polarizaiton contribution that must be doing all the work here.

  • CB – Maybe use HJang’s respyte package to have better control over how this works

20 mins

@Pavan Behara

• Discussion on XTB as a potential replacement for AM1 - SLIDES

• Slide 2

  • CB – I’d like to reiterate that it’s possible that GFN2-xtb won’t need bond charge corrections, and that we’ll know this is true if we train it with BCCs and they end up being 0.

  • DM – The 1.2*CM5 thing is incredibly sketchy, especially because they’re not compensating for that change in other FF terms. So it’s not clear at all that these results will generalize.

  • JW – How will this interact with current plans to do NN-based charges?

  • DM – Maybe two ideas to pursue - something-QM-BCC, and neural-net-BCC.

  • SB – I don’t think there’s a big distinction - We can train an NN to XTB charges OR AM1 charges. So it’s not an either-or.

  • SB – These ideas aren’t mutually exclusive. The NNs are just a replacement for the ENGINE that yields charges.

  • CB – Maybe there are three ideas - We could just come up with a way to invent charges that yield correct observables. Or we can fit to partial charges derived from highly accurate QM and RESP. Or we can come up with a way to reproduce electrostatic potential

  • CB – So a decision needs to be made about the above.

  • SB – It’s hard to just say “we want to go for X” or “we want to go for Y”. Even the “get the ESP as right as possible” leaves a lot unstated. Then there’s also the new AMBER charge model that is kinda like AM1BCC, but with BCCs fit to solvation FEs instead. So I’m working on deciding which data we should be training to, and in what balance? So I’m looking into the idea that we should be making a charge model that’s regularized to ESPs and then trained to solvation FE.

  • DM – One tricky thing there is that hydration FEs are super noisy, so I’m kinda skeptical about their value as fitting data.

  • SB - I wonder why we’re having to so solvation FE corrections in the first place. Is it something with 6-31G*. Are amides systematically underrepresented?

  • CB – Need for vsites, aliphatic amines. So trivalent nitrogens have this lone pair. And when you fit point charges and BCCs, the BCCs have to eat this fundamental deficiency. In my PM3-BCC work, there was some improvement observed but we probably need vsites.

  • DM – I’m wondering what the proof of concept for this idea would be. Maybe an experiment where we switch out AM1 for XTB charges?

  • CB – I envision a situation where in thee short term we use XTB-BCC. Then a few vsites. Then further down the road, direct polarization.

  • SB – Two questions:

    • How do we build a better electrostatics model? This is where vsites come in.

      • For this one, I don’t think we have anyone available to work on this.

      • DM – Agree

      • SB – If we even just do XTB-BCC, we’ll need a way to train the BCCs. I’m not totally equipped to handle this, but I’m generating data for this now using 6-31G*.

    • Which charge model should we use for rosemary for biopolymer charges?

      • For this one, we want to have something in the short term, and that’s where I’m thinking we want a scalable AM1 for a fundamentally AM1-BCC approach.

      • DM – Agree

  • DC – My group is generally interested in NN-based approaches for LJ fitting. So if there’s infrastructure developed for that, please keep me looped in!

    • SB – Sounds good - I’ll loop you in!

20 mins

@Simon Boothroyd

• Discussion on resonance issue

  • resonance enumeration in case of charged species, esp. for heteroatoms

• SB – I’ll be organizing a dedicated call on charge models in a few weeks, will push this topic to that meeting.