DM – Is this a new problem in 2.1.0-rc1? Or was it also there in 2.0.0?
PB – It was also there in 2.0.0.
DM – So, could we fix this in a 2.2.0 instead of 2.1.0? We just want to make sure that we don’t make anything worse, we can always make more releases.
PB – This should be simple to fix, can retrain this param without changing the rest of the FF.
CBy – This reminds me of phosphonate problems…. I wonder if this is an electrostatics problems. This seems like a small angle.
DM – If this is a chemical perception issue, and we can fix this by splitting a parameter to avoid sharing a parameter a similar but physically different moiety, what is that similar moiety?
PB – a31 hold on to everything with double bonds. a32 then applies to groups with at least one single bond. So I think this will “put things back” in a31 which didn’t fit with the angle described in a32.
DM – Are you sure those belong back in a31?
PB – Yes, they have more similar angles to a31.
PB – The a32 redefinition+training seems like a hack. It wouldn’t be a general refit. Is this suitable for a FF refit?
CBy – In the case of NS(=O)(=O)([O-]), the angle assigned between all of those should be equal. In the proposed a31/a32 split the three inter-O angles would get different values. This seems wrong.
PB – A31 should be looking for a S with single bonds, whereas …
CBy – I’d use [#8]~[#16]~[#8] to capture partial double bonds
PB – But there are some weird molecules where we have symmetricish [=N]s instead of [=O]s.
CBy – …
DM – Agree that the SO3 case should be split, and that [#8]~[#16]~[#8] is the better way forward
CBy – It’s tricky that S(=O)(=O)([O-]) can also be represented with dative bonds, like [S+3]([O-])([O-])([O-]).
DM + JW – Maybe [#8X1]~[#16X4]~[#8X1] ?
DM – LW, should this be rolled into 2.1.0rc2? or made into a separate release? Our previous statement had been that we won’t make a 2.1.0rc2 unless we find that it has made something worse. We’ve already run 2.1.0rc1 through the FE calc pipeline. Since this problem was around in 2.0.0, should this wait until 2.2.0?
LW – Since we hear about people continuing to use 1.3.1 because the sulfonamide problem is fixed, that’s an argument that we should really get sulfonamide geometries fixed in the next release.
…(see recording, around 35 mins)
CBy – …
DM – How about we prepare the next release asap, and …
JW – Can we set a plan now for scheduling the release?
PB – I could have the refit done by Monday
LW – We’ll stick with a monday May 1 release for 2.1.0-rc.1 unless major new problems appear. We can make a follow-up 2.2.0 release soon after if the refit is promising.
Good benchmarks to run for Espaloma - 10 min
Ken Takaba
KT – We refit espaloma using energies and FFs, we think it’s a lot better now. Which benchmarks should we be aiming to use to prove the value of Espaloma?
DM – The suite of scripts that made the plots you saw is our standard benchmark. CC also is looking at protein benchmarks
KT – I’ve simulated GB3 and looked at RMSD, there’s kinda a lack of automated NMR benchmarks.
CC – I’d be happy to share my NMR observable benchmark code - I’m still in the process of writing it but I’d love feedback. Currently have a set of short peptides with J coupling, and some small proteins like GB3 where we have NMR data. We’re trying to expand that to more protien systems. But right now I just have scalar couplings. But in the future I want to add more observabled like chemical shift, order parameters. But if you can get your starting structures to OpenMM systems, you should be good.
DM – Tobias Hufner would also be a good person to talk to about this. He’s doing crystal benchmarks.
CBy – Electrostatic potential and electric field are good to compare to. Just for electrostatics.
MS – Trained both valence and charge parameters?
KT – Both, in different packages - Espaloma (regular) for valence, espaloma_charge for electrostatics
MS – Would be good to evaluate against density as a funciton of temperature, but it’s expensive.
CBy – …
DM – We’d looked at dipole moments, passed over the materials for that to YW last week.
BS + KT – We had had talked about this - There were two ways to get espaloma charges - we were using the espaloma_charge package, which was making weird charges, however using the refitted espaloma does better on the test molecule.
MS – We’re trying espaloma charges for polymers - Is there a way to get just charges from the good model (the espaloma package)?
KT – That should be possible - We could use the new version and then just strip out the charges.
KT – Agree with improper problems. We’re using n=6 multiplicity for impropers now but it seems like we should restrict that to 1 or 2.
DM + JW – Depending on how roadmap plays out, we may be rolling out a big generalized benchamrking infrastructure in the coming year that will let FF researchers calculate these metrics themselves.