Page Comparison

Sage 2.2.1 valence re-fit and performance on quick QM benchmarks

LW

• View file
  name 2024-07-03_FFF_Sage-2.2.1.pdf

• Azide issues

  • JW – So the fit got worse?

  • LW – We fixed the MM angle to 180, but the QM thought it should be 174ish.

  • MG – Might be good to do QM calcs of azides with less context to see if they really should be linear.

• PB – Can get a good match with QM if the force constant isn’t too high.

  • LW – Good point, the force constant was allowed to change in the fit.

• JW – Assuming that real azides are 174ish, then I’m in favor of putting this in a release with the azides being non-180

• BS – Even if they’re 174 in nature, having an angle near (but not equal to) 180 will put a discontinuity in the FF.

  • MG + BS + JW – Should do gas-phase QM calcs.

  • LW – Could start a fit with a 180 degree equil angle and lower spring constant.

  • BS – Good idea.

  • LW – Will do.

Sage 2.1 performance on physical property benchmarks, and plans for a vdW re-fit

LW

• View file
  name 2024-07-03_FFF_Sage-2.1-condensed-phase.pdf

• MG – Before we delved into this, WW and I lookedinto this, and we wondered how there were densities near 0.

  • LW – Right, this is a very broad grab of ThermoML data, in a later slide I filter to only the data that we used for our fits

• Should we retrain vdW from 2.1.0, or can we just do 2.2.0?

  • WW – I’d love a refit vdW of either, happy to do my work with the next gen of OpenFF.

  • LW – We have a 2.2 refit in the works, but the valence terms are changed between 2.1 and 2.2. I assume you want to continue from something like 2.1 since that’s what you started with. So I’ll do a 2.2 refit first, and 2.1 refit after that.

  • WW – If the 2.2.0 refit will be what OpenFF invests in for the long term, then that’s what I want to base my work off of.

  • LW – And you’d be OK to use that for your research?

  • WW – Yes.

  • MG – So you’d like to build your dPol work off the “latest, greatest” OpenFF model?

  • WW – Yes.

• MG – TH discovered an error in one of the barostats in OpenMM. I doubt that’s the case here since he was using an anisotropic one.

  • LW – We’re using the same barostats - Monte carlo in OpenMM and (parinello-raman?) in Gromacs

• MS – What constraints are being used?

  • LW – Hbond constraints.

  • MS – I recall there was an issue in Gromacs when hbond constraints were used, that manifested in some bromine-containing compounds.

• DM – PB had an issue recently where gromacs did weird stuff for particular combinations of timestep+barostat.

• MS – If it’s a constraints issue, try running with 1fs timestep. Maybe 2fs is right at the edge of stability. So maybe we need to make constraint tolerances higher.

  • BS + DM + JW – Agree

• MS – For mols that vaporized, did valence terms (ie force constants) get bigger?

  • LW – Not obvjously

• PB – The bond force constants in 2.2 should be quite different, right? Since they started from MSM values.

  • LW – Right, I’m going to check 2.2 to see if the same problem exists.

• JW – LW, can you pull out the temperature from these sims? Things like non-180 angles and HMR-induced bond vibrations may be messing with temperature.

  • MG – Could also turn down thermostat strength and check that way.

• PB – if the issue is with simulation setup does a vdW refit help?

  • LW – We’ll have to see

Additional molecules to get phosphate coverage in upcoming QM fit.

MRS/PF

(list of molecule to be submitted 7/2/2024 to be discussed, will post link here).

(context from Slack) https://openforcefieldgroup.slack.com/archives/C8P54J2JY/p1719947737286229

Image Added

• PF – Building a QM dataset based on lipids and phosphates. Interested to take advice on how to improve dataset.

• MS – LW, for the non phosphate molecules, is the coverage good enough already? As in, do we just want to focus on lipid headgroups if the sidechains/tails are sufficient?

  • PB – I think so.

  • DM – We could look at our coverage of smaller sugars - I’m not sure how many hydroxyl-rich alkanes we have.

  • LW – Without looking at the data, I think it’s a relatively small number.

  • DM – For cholesterol, I’d think it’d be important to have the ring system. How confident are you that this QM data will resolve the problems we’ve seen in lipids in earlier presentation? As opposed to condensed phase.

  • MS – We know that our headgroups differ strongly from their confs in slipids and macrogs. We tried substituting in the charges from those FFs and it didn’t fix the issue. We know that alkane-alkane interactions are in our condensed phase sets, but we don’t know that we fit to headgroups. So we suspect it’s headgroup torsions.

  • DM – Speaking without thinking much, this seems important. My question would be whether there’s more stuff to enumerate.

  • MS – Yes, and I mentioned phosphodiester backbones and ATP/ADP like things.

• BS – What do you want to get for the charge models? One thing you could imagine doing is making the three Os equivalent.

  • MS – We’ll be using AM1BCC, which should be treating all three equivalent.

  • PF – could put methyl caps

  • LW – I think CC’s fixes have ensured that the valence params are applied correctly.

  • MS – Should also ensure…

  • MS – Should also have some structures where we add a methyl, so that we have references with 2 and 3 monovalent Os.

  • PF – Adding methyls to headgroups would be best.

  • MS – Would be good to do both.

  • LW – Yeah, we have a lot of room to expand this dataset.

  • BS – In reality, how are those negative charges neutralized?

  • PF – They’re deprotonated at neutral pH, exist as zwitterions.

  • MS – What’s the pKa? Phosphoric acid pkas are like “always deprotonated”, 8, …

  • MG – There’s a history that some phosphate params lead to making phosphate-sodium “crystals”. So phosphate ion by itself in solution can be tricky.

  • MS – “just phosphate” will have very low conc, most with be H2PO4

  • MG – IIRC, this problem happens with many protonation states of phosphate

  • PF – ROP(O2)OR is always deprotonated

• LW – I think you can include the “full POPC” headgroup.

  • …. one for each???… (one torsiondrive for each structure? )

• PF will send mols to LW to craft QC submission

• LW – Should we be looking at this more? Phosphate hadn’t been heavily represented in training data before now.

• MS – Timeline?

  • LW – A week to do the fit, mostly human time.

  • JW – Maybe as quick as a weekish to compute if just 1D torsiondrives

  • JW – Is this high priority enough to displace NAGL work and the other stuff on LW’s plate? Could be an opportunity to train more people how to do QM dataset gen and fits.

  • MS – We do have funding to get this done.

  • DM – We’ll discuss in lead team meeting.

Density and dHmix properties with dPol model and Sage 2.1 valence terms

WW

• (only partially presented in this meeting)

• BS – Are condensed phase sims electrically neutral?

  • WW – There’s not an accurate way to define systems that have variable protonation.

  • MG – These were simulated in the neutral form. Which we know is wrong, and it’s why we advocate not including it.

NAGL small molecule lookup table implementation

LW

• (Not presented in this meeting)