2025-07-02 Cole Lab/OpenFF Meeting notes

Fitting using smee

• (see recording before 10 mins for notes before this)

• DC 10 – if you only have a small number of torsions to train, you want electroststics and sterics to do some of the work for you

  • JH – So 1-4 electrostatics scaling is going down, so this is contributing less?

  • FC – Yes

•  

•  

•  

•  

Bespokefit-smee

• JW: Could you explain the meta dynamics strategy?

  • FC: We are varying the torsion angles and adding biases to explore more of the conformer space

• JR – Does OFF 2.2.1 include fragmentation as usual?

  • FC – No, other than the testing dataset being fragments from a datasets of larger molecules (fragmentation performed BEFORE tests are run)

• DC: Was excluding the amides a conscious decision?

  • FC: Yes, I excluded because I expected high barriers that we wouldn’t prioritize sampling, but as you see it is important

  • JW: Yes they are usually 8 or 10 kcal/mol, is it that bad to give it a barrier of 20?

  • FC: No that would be fine, but I’m concerned that we are overfitting some low energy regions but excluding others with this high barrier

  • JW: Slide 28, the minimum shows a metastable minimum region that is not as broad as the in the QC region, is there a metric we can use as a penalty to achieve that?

  • FC: I’m not sure what the motivation of that strategy would be

  • DC: We definitely want a better fit in that region

  • FC: Im mostly concerned that it gets worse with additional SMEE fitting

  • DM – Amides are particularly tricky because of the are-they-puckered-or-not question, which kinda plays with impropers too. In practice they end up being a mix of wanting-to-be-puckered and wanting-to-be-planar, giving them a “squishy” wide basin, but the functional form of the impropers is limiting. We have played with alternative functional forms to resolve that.

• (slide 31) DM: I had a conversation with the espaloma folks on impropers, are they using significantly more? Particularly for amides?

  • FC – Haven’t looked into this/can’t remember.

  • DC – Worth a look. You’re fitting impropers if they appear in the FF, right?

  • FC – Right

  • DM – But worth checking whether espaloma assigns a lot more impropers in general, I wonder if this is true in a comparison of OpenFF vs Espaloma treatment of amides.

• (Slide 25) JR: With the high barriers the shape of the torsion seems to change at the bottom of the well. I expect that there is an indirect effect on sampling.

  • FC: Sure, something to keep an eye on, but I’m using 500K that should be thermally accessible.

  • JR: Sure, but I think with high thermal noise the actual high temperature simulations would show significantly different profiles in comparison to QM without that noise.

• DC: Did you say you’re only going through the meta dynamics cycle once?

  • FC – Al of the metadynamics runs are from one cycle, tried doing multiple cycles on biaryl but didn’t get any improvement.

  • DC – I’d keep doing multiple cycles I’m not concerned about time now that we are operating within a small number of minutes, let’s focus on robustness first, and then accuracy.

• JW – I wonder if the high energy regions are looking bad because some other term (bond/angle) are getting really strained, and the high energies that should be accounted for by those are getting dumped into torsions?

  • FC: That’s a good point, I’ll have deeper look into the outliers. May also be intramolecular hbonds that are complicating energy profiles.

  • DC: I agree that if the outliers aren’t accounted for in fitting that could be an issue so we can iterate on the lowest performing class of functional groups iteratively.

• JH – Good to see espaloma being competitive here - Would be good to keep comparing to it and see if we can find regions where it falls down.

  • FC – I was looking at the ata and there are some outliers where espaloma gives crazy outputs. But those are pretty rare.

  • JH – It would be great if we could use this to characterize where espeloma falls short to improve it considering that it’s faster than OpenFF and Bespokefit.

  • FC – Would be interesting to use something like an ensemble of nagl models that try to predict things, and if there’s disagreement between then then it triggers bespokefit

  •  

  •  

  •  

  • .