JH – Tried to visualize trajectories to catch sampling issues.
DM – Rotamer sampling issues for rgroups?
(Needs to do more edges? How to do this?)
MB – Uses RDKit to make the MCSes.
DM – Could switch which mapping you get, and/or switch the coordinates of atoms
DM – One thing I’m thinking about is that TYK2 is “the system where everything works”, so I’m wondering what we’d see on the other systems.
LPW – An RMSE of 0.64 kcal/mol is really good. The parsley paper got around 1.0.
DC – Yeah, JChodera’s paper with ML FFs also showed about half a kcal per mol.
JH – All the software is in place to run the whole JACS dataset, we’d just need more compute time.
JH – It’s frequently not worth it to drive amide bonds. Not sure how to automate skipping this.
DM – It could be complex to figure out which ones to skip scanning
JH – I’m thinking of caching torsion parameters for all fragments from QCArchive, and distributing with software. This could play well with the caching functionality that I recently added.
SB – What’s the cost comparison between nromal FFs, XTB, and full QM?
JH – XTB is about 1 minute/fragment, full QM is hours. If you get clever with caching you can run the QM faster, but we haven’t incorporated that into bespokefit.
SB – Do we ever want to run full QM at this point? XTB seems plenty performant.
SB – PB, you’d experimented with XTB1 vs XTB2. What’s the tradeoff?
PB – On the set of our QM benchmark molecules (59 of them), wrt df-CCSD(T)/CBS//MP2-heavy-aug-cc-pVTZ baseline
| Specification | RMSE in kcal/mol | MAE in kcal/mol | | gfn1xtb | 1.5271 | 1.0309 | | gfn2xtb | 1.2481 | 0.8554 |
JH – Orbnet paper compares GFN1 vs. 2. on torsionnet. GFN2 is a bit better, but it’s debatable whether it’s worth the cost.
LPW – Agree, it’s not clear how representative these plots are.
Bespokefit conda-force release
JW – Bespokefit is now on conda -forge. Looking for an initial tester before we ship it to swope!
DC – I tried it yesterday and
I needed mamba to get it to install, is that a problem?
JW – I used to think it was, but I’m finding more and more that I need to do everything with mamba now. So I don’t think it’s that bad any more.
JH – Did you try to do psi4?
DC – No
SB – Current bespoke package doesn’t require psi4, targets toolkit-base to avoid AT/psi4 conflict. There was another dependency that had lots of pins, I talked to them and got that loosened up. I think it’s ready to send to Bill.
DC – Time to announce this to the world?
SB – Let’s get Bill to test it for a few days before we announce.
We’ll have Bill test bespokefit for 3 days, then JH will announce release.
It wasn’t clear to me how to have bespokefit use XTB instead of full QM.
SB – Agree – We need to document how to go through python to get XTB. We also need to add a commandline argument to get XTB.
JH – Agree, we’ll want to incorporate some pre-built commandline arguments.
JW – Just made emergency bugfix release for OFF Toolkit, unfortunately an unrelated change broke forceblaance, which means it broke bespokefit. So we need to make a new FB release.
Overview of vsites and functional form work
DC – Vsite work was started before we started with OpenFF.
DC – sulfur lone pair vsites show good improvement
DC – F-C bond charge vsite keeps falling inside the fluorine, with a positive charge
LPW – Would that basically be a sigma hole?
DC – It’s the opposite of what we’d expect for a sigma hole - The region that should be getting positive is instead becoming negative. But fluorine is said to see the least sigma hole effect.
DC – Adding vsites reduces errors for densities and heats of vaporization
DM – Is it accurate to say that, if someone were interested in working with this in the main OpenFF force fields, they could start with your placement of vsites?
MS – There are two questions for each vsite -
A discrete question of “should this vsite exist?”
Numerical questions of “what should the values be?”
MS – I know SB has infrastructure to do the fitting.
DM – So it seems like this is ready to start in OpenFF.
MS – Maybe, my student Yu-Tang may be able to start here, but he’s just beginning grad school so it’ll take a good amount of training.
MS – I think we need to pin down our electrostatics fitting procedure first. Will it be AM1, then fitting new BCCs? Without this we don’t have a foundation to begin fitting vsites.
MS – Is there a heuristic to decide which vsites should exist?
DC – We only have a ~100 molecule set to help us decide, but these have informed us about where to put vsites. They largely indicate we need things like O and S lone pairs.
CC – Which MD engine did you implement vsites in?
DC – OpenMM, using localcoordinatesite
CC – We might think about, at OpenFF, that we’re concerned about the ability to export to other engines…
(general – Support for our vsites in GROMACS is fine, but AMBER is unknown)
DM – I don’t think this should block our research. If we fit vsites that are good thenw e can push to get them implemented in engines.
MS – If we agree that the goal of the study is to figure out which vsites SHOULD exist, then … Also figure out what the numbers should be.
DC – We’ve experimented with replacing LJ potential with double-exponential.
JH – Here are my fits with DExp - This is very preliminary! Mostly proof of concept to show that it could work.
Original sage fit
Densities (normal on left, DExp on right)
JH – Some stuff with absolv, not finished yet
MS – I know how to fix that, can chat offline - Will send DC and JH a series of papers on it
DC – I’ll take a look at this before the ad board meeting
MS – A few thoughts
it would be nice to jump from 2 to 3 parameters instead of 2 to 4 parameters
We want to get to the point where we can do bayesian inference on several different proposed functional forms (long term, 2 years). If you’re working along these lines it may be good to discuss this. But we basically need irrefutable evidence to convince people to switch funcitonal fomrs
Where this matters is pressure dependence. This will make a difference as we model the repulsive component. So with ions and high pressure systems.
MS – To dos for me are
send papers to DC and JH
begin long-term planning for working together on improved nonbonded modeling.
