2023-03-30 Force Field Release Meeting notes

General updates

• MG – Presented OFF at ACS meeting, at a Rom Levi symposium. Had good interactions with Carlos Simmerling and Alex MacKerrell. Richard Friesner also said nice things. I think my slides are available on slack - Where should I upload these to, DM?

  • DM – I’ll message you with the proper spot

GNN update

LW

Link slides here.

Open

RECORDING: https://drive.google.com/file/d/1yn61ebmdEXqZfbKRazSlsZUfYvumr326/view?usp=share_link

Slide 2 (RMSE scatter plot)

• MG – clarification (see recording)

• PB – Are you refitting BCCs here or using the standard BCCs?

  • LW – Refitting

  •  

  •  

Slide 3

• MG – Why do you think these results are better?

  • LW – I think they’re more flexible this way…

• MG – Maybe the big molecules are like a small molecule with a big alkane attached. So if there’s a big “easy” part being added, then getting a specific funcitonal group wrong would seem less impactful.

• Jw – Would be helpful to have something like a dotted line for the y=0 line, since it’s hard to really compare the average errors here.

• MG – Looking at largest errors?

  • LW – There are some molecules with big differences.

  • MG – Could be good to report those - One of our big objectives will be to build trust in the GNN.

  • LW – Yeah. Also worth mentioning that I test on peptides and those look quite good.

• PB – The GNN largest error looks like 0.25. AmberTools is around 0.3.

  • LW – Right, “better than AT” is our target. Currently about 70% of molecules are better than AT in every metric. There’s also complexity in the fact

• JW – Does refitting BCCs lose us a big part of the advantage of doing a two-step AM1+BCC?

  • …

  • General – No, we should do what the data shows us is most accurate.

•  

Charge model assessment

Lorenzo D’Amore, Bill Swope

Link slides here

Open

Slide 3

• PB – How do I interpret the Y axis (“Density”)?

  • WS+LD – about 10k mols, 70k mols

• MG – Which espaloma is this? Is it trained to AM1BCC?

  • LD – It’s trained to AM1BCC directly

• JW – Would be handy to have a line with random charges or all-0 charges to help determine whether Espaloma is doing better than random.

End of slides

• JW (chat) – Espaloma having large errors for 5-membered aromatic rings makes me think this could be an aromaticity assignment issue. The toolkit will try to re-perceive everything using the MDL aromaticity model… Maybe espaloma was trained using a different model? Aromaticity Perception — Toolkits -- Python

• WS – Which nagl model was LW showing?

  • LW – …

• …

• LD – Did you use SPICE dataset for training?

• LW – No, just testing.

• PB – Explicit handling of resonance in nagl could be making the difference here.

• DM – I’m noticing that the big espaloma discrepancies are in the immediate proximity of a nitrogen, where it might be unclear what the formal charge of the nitrogen is.

• MG – I was worried about electrons “sloshing” around with electron equalization methods, since there’s nothing really that constrains charges to be reasonable. But the analysis of charges on regions of molecules look good.

• WS – I think the resonance averaging could be responsible for smoothing out some of the sloshing.

• DM – I think it’s cool that the dipole analysis looks good, given that we didn’t even test this.

• LD – Is there another research direction that I can check that would be helpful for nagl development?

• LW – This is already really helpful. Which dataset was this on?

  • LD – Industry benchmarking public set, filtered for molecules with 0 net charge.

  • MG – Could look at computing dipole moments relative to center of charge

  • WS – Sometimes the center of charge is outside the molecule. Only about 10% was removed by this filter.

  • WS – Could skip the dipole analysis and still do the others for the charged mols.

  • MG – I’d also expect zwitterions to be tricky

• PB – Is there interest in trying out xTB as well to see if it does well relative to quantum?

  • DM – That’s a good question, but right now we’re aiming to reproduce AM1BCC. Though that could be useful further in the future.

  • PB – Right, I meant it’d be handy for followup work.

• WS – We’ve done the same analysis with OPLS4 charges. They have a default CM1 model with and without vsites.

  • LD – The OPLS+vsite FFs do extremely well for dipoles (relative to quantum)

  • MG – OPLS non-vsite is CM1?

    • LD – CM1-BCC, I think

    • MG – We may have some data on that - THufner ran some stuff with OPLS-AA and saw bad things. Then we dropped to AM1BCC charges and things improved.

  • JW – And OPLS+vsites uses?

    • WS – It runs a quantum opt on each conformer and fits to ESP.

    • MG – Ah, that’s why it looks so good. Seems kinda impractical.

  •  

• LD – Should I keep looking for pathological problems with functional groups in nagl?

  • LW – This would be great. What you found today was really interesting and I’d love to know about more.

  • LD – Sure, let me know how to install the new model as well.

• JW – Should reach out to Yuanqing to make sure that the espaloma interface is correct.

  • LD – Sure, I’ll reach out on the #espaloma channel

  •  

• .