Sage torsion parameter issues | MM | MMackey intro: Cresset “Xed” FF is polarisable and has vsites everywhere For Cresset v4, thought we’d use OpenFF machinery and data, but keep the philosophy of Cresset v3 with a more complicated charge model Using deep learning to reproduce ESPs on molecular surfaces for small organic molecules. After that, used standard OpenFF training pathway to build the FF. TP looked at reparametrizing L-J potentials, found that they did need some tweaking for the new charges. Found areas where our FF wasn’t working well, and noticed that Sage was struggling with those as well, so I thought we’d let you know in this presentation. Found that Xed and Sage-2.2.1 torsion profiles were so similar that all results in this presentation will be of Sage 2.2.1. There are some cases where behaviour diverges, particularly with intermolecular cases, but not shown here.
Failure type 1 JW: are these torsion profiles just the torsional contribution? MH: “Sage-2.2.1” line is the full potential of the molecule at a constrained minimization, modulo energy of minimum. “Periodic” is is the periodic part of the potential, i.e. all torsional/improper contributions. “HCCO” is just the function for that one torsion type.
JW: are you saying the scanned and complementary torsions cancel out if they have the same barrier height? Or that the purple line is trained properly but the orange line is cancelled out? MM: the issue is lack of data coverage for the orange (scanned) torsion. Because the single molecule torsion has both torsions, and the purple torsion gets fit to other molecules, the orange torsion is not trained properly.
MM – Also, the training molecule may not be a good fit for this torsion - thiophene and ethene might not be interchangeable. CC – I may have some insight here - In earlier openFF versions we had torsions getting assigned differently to the single and doule bonded oxygen in the carboxylate. So we recentl;y split these off from t48 (which had a 3-fold torsion), but we didn’t add more training data for that torsion.
Failure type 2 MM – Maybe threefold term should also be removed from aromatic carbon. (but may be needed for other uses of this term) DC – We’ve done some experiments in NOT using forcebalance for training, and not using any pre-conceived periodicities (basically allowing fitting of all periodicities up to 3 or 4)
Failure type 3 Failure type 4 Anomeric effect MM: haven’t done free energy simulations in solution, these results are vacuum single points. MH: QM calculations do line up as expected, which Sage fails to reproduce. MM: likely would need more specific torsions to accomplish this. Xed FF accomplished this with 1-3 electrostatics, which we’re not including here. If not electrostatics, has to be included in torsion potentials.
LW – This is super helpful, we’d love to act on all of this. Could I get a copy of these slides? DM – This is really great, this is really perfect for us. DC – With all the vsites, are they included in the torsion scans? MH – They are - vsites are included as part of the atom they’re attached to (for bookkeeping on atom removal). There’s not many for small molecules, but… They mostly perform the same as Sage, but there are a small number of cases where Zed does a lot better. DC – And have you been able to convert vsites between openmm/other engines? MH – The vsites that we’re using don’t work in the openff ecosystem. TPotter has but a lot of work into getting a layer in the middle to get the types that we need.
MM – Primary use for Zed FF is small molecule confs. Haven’t designed it for use in simulations/free energy calcs. MM – We do have ideas for more torsiondrives to run - Is it possible for us to add molecules to the queue if you have spare compute? DM – IME the easy part is running, the hard part is figuring out what to run. LW – Hundreds of molecules are doable for torsiondrives. At the moment our external compute is broken but we’re working on fixing that.
|
