2023-04-13 Force Field Release Meeting notes

Sage 2.1

PB

PB – Nothing new since yesterday’s presentation. We’ve made a new release and conda package for openff-forcefields containing openff-2.1.0-rc.1.offxml.

BCC updates

LW

Open

• PB – BCCs here are assigned based on atom types, right?

  • LW – For OE and AT, yes. But for us no.

• DM – Why do we want to get AmberTools AM1BCC?

  • LW – Becasuse we train LJ to OE AM1BCC, and thenw e offer AT as an alterantive, I wondered if folks would get better results if they got charges closer to the OE charges. It turns out that AT and OE maybe apply the BCCs differently

• PB – Maybe this is due to the restrained minimization is OE that isn’t in AT?

  • LW – That’s possible.

• PB – …

• JW – Big picture - If we know we’ll eventually want to have some slightly different BCCs that compensate for odd pathologies in the GNN, we may not want to put in a ton of effort to debug OE vs OFF vs. AT BCCs

  • LW – Yeah, this is about as far as I want to go into debugging BCC differences.

  • DM – Agree, this seems like it’s close to good enough for release.

  • MT (chat) – As a downstream user with no say I’d be in favor of whatever produces a release on the quickest timeline

•  

•  

• .

GNN pathologies

Lorenzo D’Amore

JW – LD and YW met last week, LD sent a tarball that can reproduce the bad results we saw, YW should be working on it soon but may still be on vacation.

LD will link slides here

• DM – Slide 10 – at least one atom with 0.2 to 0.5 e seems to happen in ~25% of mols here. What might be different?

  • LW – LD’s work is showing single-conf AM1BCC. Also OE’s ELF10 yields different charges than our ELF10.

  • JW – Also OE’s ELF10 has some magic numbers inside, so we can’t replicate what they did.

  • LW – Right, that’s why I generated my own ELF10 confs outside OE.

  • (General) – Two possible sources of difference - 1) single conf AM1BCC vs. ELF10, and 2) different sorts of analysis (RMSE vs. max atom)

    • LW – Re: 2 - I also check max atom deviation and I didn’t see that many outliers.

  • PB – LD, are you comparing to AT or OE?

    • LD – AT

    • MS – We’d looked into differences between AT and OE before, could compare to that.

    • DM – The current data shows that Nagl is closer to OE AM1BCCELF10 than AT is. LD, do you have OE charges that you could run this analysis using?

    • LD – I do, I’ll run the analysis on the OE AM1BCCELF10 numbers.

  •  

  •  

  •  

  •  

Sage 2.1 and current force field benchmarks

PB

• DM – I think the next big question is the timescale for FE calcs. We’re doing beta secretase, and PB has passed the FF on to exscientia.

• LW – Agree with this timeline. Would exscientia need help from us to do this?

• DM – They’ll run FE calcs internally, we just have to provide them the FF.

• PB – Is MOsato running beta secretase?

  • DM – Yup, she’s working on it now. Have to swap in FF params.

• JW – Timeline-wise, as long as we have stuff out in RCs for the may meeting, we can show off cool new functionality and users will just have to use RC packages

  • DM – Agree, it’s not do or die, but it would be nice to get the final releases cut since we advertised “late march”.

• MS – Automated benchmarking will be a great boon here. Do we still do automated solvation free energies? What’s the full test that everything runs?

  • DM – When we developed Sage, we did free energy of solvation benchmarking. Since this is only a valence refit it’s probably not worth doing that. And we don’t have protein-ligand binding FE calcs automated yet. There’s not been a huge push from the ad board for binding free energy calcs, most of the feedback is about geometries. Could be a good topic of discussion at the ad board meeting.

  • MS – We have more personnel-time at CU that could be allocated to make sure OpenFF moves forward. We could talk about this in the coming month.