How polarised should charges be for condensed phase simulation | @Daniel Cole | new paper https://pubs.acs.org/doi/10.1021/acs.jctc.2c01123 DC – Aims to answer “what’s the best way to prepolarize your charges?”. Basically, in vaccuum, you’re guaranteed to underpolarize. But in water, you’ll overpolarize. So here we argue that about halfway between is about right. This is similar to RESP2, which I think uses a scaling of about 0.6. DM – This seems promising, but my concern is that, if we tell pharma people that they need to do a charge calculation with explicit water to get better values, they’ll think that’s too complicated and will continue using AM1BCC. So what action would you suggest we take? Could pre-train a GNN? DC – Yeah, I brought this up a lot in the writing process. I argued that we need to make this computationally more tractable for users. The GNN approach could work. Or if you do it for a few thousand compounds you could find some patterns like boosting the implicit solvent dielectic for certain molecules. DM – Cool, please post this in the slack. I think BSwope will be interested.
DM – We tried using iPolQ, but had usability issues. DC – Yeah, probably the same if you tried to use the scripts from the current work. DM – Could also think about refitting BCCs using a training set based on this. JW – We could probably start optimizing solvated small mol+water structures on the QC stack if it doesn’t require new code. DM – Yeah, could generate our own dimer energies as well DC – It’s QMMM, so maybe not supported. PB – We do have solvated amino acids and dimers in DES370K. DC – Do these have charges? PB – I don’t think so. JH – Did they get MBIIS charges? PB – They do have MBIIS, but not mulliken. DC – Right, it wouldn’t be clear which set of charges to store. We’d probably want to match whatever wa used for the GNN fitting set. JH – I tried doing that for the SPICE datasets, but there were some failures. I’m not sure what that would tell us. DC – I’ll talk to Miguel about this. JH – Thinking about the riniker multipole model - So you could have the solvated molecule in water, use that to predict point charges, and use that as a start for QMMM to accelerate the optimization. DC – How would that help with charges? JH – The charge could move around a bit using the multipole model, then that could be a better starting point for QMMM. JH – Like, you could generate starting confs using a cheaper method, and that could speed up the process. Like, if I understand correctly, the current model needs charges from the water model DC – Kinda, it’s iterative.
|
Interaction energies for nobonded training and evaluation
| @Joshua Horton | JH – Thought was to maybe fit to (other parameter sets that were fit to interaction energies) AND (interaction enegies directly). JW – Meaning of K index? JH – Triple bonded carbons are systematically bad. DM – Worth noting that Swope and others are really interested in dimer energies/polarization. DC – Yeah, I could see those being big problems and avenues for improvement. DM – I also wonder how much this could improve places where we lack coverage/physical properties. JH – I also wondered about that, for example the S44 dataset which has a lot of halogens. For examplehydration and sublimation free energies. DM – We may actually be able to get money from people like OpenEye to study that. They’re really interested in soft cores. DC – That’d be really interesting, we’d be happy to bring this up at a FFR meeting. PB (chat) – DES370K has more general chemistries as well JH – Yeah. I think at the moment I need to get FB working with interaction energies. Will submit a PR to LPW/ForceBalance DC – Should that go to OFF-FB, or LPW-FB? JW – Ah, but there’s the ecosystem split with OFFTK 0.11 and interchange… MT is updating interchange to use plugins, and I think he’s porting the existing ones probably including DEXP. So you should chat with him about the state of the plugin.
.
|
