JW – Slide 37 - Why is only AM1BCC/NAGL seemingly symmetric, but others aren’t? Is there some source of chirality involved?
JH – I can chekc into it - Maybe P oxygens are getting different charges? Or maybe things get stuck in intiial configuration?
LW – Thanks, and the charge-swap section was really insightful. Could we look at the slipids and macrog torsion distributions as well?
JH – I can send these later.
LW – Yeah, I’m curious about the general comparison of torsion definitions.
(slide 31)
…
LW – Are macrog and slipid charges similar?
JH – They’re similar in that they’re more polarized than nagl and am1bcc. I have individual plots of those but they’re not in this presentation.
LW – For the final Os in phosphate group, what are those charges?
JH – I can get back to you
CC – Nonbonded cutoff of 0.9 nm for openff? What was used for the others?
PF – 0.9 for openff, 1.0 for others.
CC – There’s a paper from Angel García (link here) that showed that 1.0+ nm (??) cutoff was needed to get right lipid properties
MS – Our APL is already too small, making cutoff bigger would make it even smaller. We also tried LJPME.
CC – Bringing this up because I know the SMIRNOFF spec says to use 0.9, but I get the sense that we’ll need to revisit this to get lipids right.
MS – Right, we want to look at different cutoffs and LJPME. We discussed, and concluded that we should revisit it once we have clear data. There’s a problem with the fact that lots of codes don’t implement LJPME. But we need to show that lipids need LJPME to get engine makers to implement that
CC – Sounds good.
PF – Some of my original calcs used LJPME, I’d have to double check the full results but the torsions were roughly the same.
PB – These were run with GMX or OMM?
PF – GMX
JW – Do we think that it’s reasonably certain that headgroup issues are causing most of the problems, or might intermolecular tail interactions between lipids be too “sticky”?
MS – Looked at distributions, tails were a bit too gauche. But headgroups seemed like they’re less covered in our fitting process. We have lots of training data on alkanes so I get the sense that headgroups are more likely to be the issue.
BS – It seems like kinetics is bad with tails.
MS – Right, but is that downstream of the headgroups all being stuck to each other and keeping the tails from moving?
BS – Are we getting diffusion properties of alkanes right in our fitting/benchmarking?
MS – We’re getting torsional barriers wrong… (slide 15)
BS – If it’s an INTERmolecular issue, the torsions on slide 15 wouldn’t help us identify that.
JW – Based on videos, it looks like headgroups-glomping-togehter seems likely.
LW – Did the headgroups seem to be glomping up in the charge-swapped OpenFF sims?
JH – Yes, I can share those videos later.
MS – So, moving forward, I think we want more QM data for headgroups.
JW – We have a lot of compute capacity available if we want to submit a dataset.
LW – BW could help with dataset preparation, and we could work together on this.
CC – I also have torsion scans of nucleic acid backbones which may be useful. The AMBER FFs model phosphates the same with NAs and lipids
PB – Some of the phosphate torsions were fine-tuned in 2.1, might it be worthwhile to check against 2.0 as well? T159’s k2 becomes stiffer and there’s an additional periodicity term.
MS + PF – We looked at this, we were under the impression that they were mostly the same.
PF – We had a weird bug where one torsion was assigned two different parameters… I’ll open an issue for this.
MS – We could look at diffusion of decane. I doubt LJ is a problem, but maybe charges/headgroups. What benchmarks did PB’s posted paper do?
PB – Just conformer energies, not intermolecular.
LW – Also, comparing to a different OpenFF FF (say, 2.0) with different torsions could be a good idea.
(General) – 2.0 has a bunch of other issues so we shouldn’t use it out of the box.
CC – I can make a modified 2.0 that works better, will send this out.
PB – Also maybe switch cutoff in GMX?
MS – Using LJPME might be better here.
PB – I had some issues running SFEs with 0.9 nm cutoff, which went away with 1.2nm cutoff.
MS – Will do something like cutoff= 1.0, 1.2, 1.4, and LJPME.
MS – Also will try calcs on diffusion of alkanes. There’s lots of exptl data here.
MS – Can compare to GAFF+OPLS? And copy macrogs dihedrals.