Looks like there are two families of charges that come out of RDKit/AmberTools charge calc, whereas there’s only one family of charges that come out of OpenEye
Can we assign a magnitude of badness to charge inconsistency? Like how much it gets us the wrong answers to observables relative to errors contributed by other FF tems
SB – This will be super hard to decompose, because of how coupled different FF terms are during training.
Slide 15
JW – It’s concerning that all the AT densities are higher than all the OE densities. This may indicate that there’s a systematic difference between OE and AT charges (rather than a random one), which would mean that we train with OE, and introduce a systematic error for users that are using AmberTools.
SB – Have the simulations converged?
LW – Not yet.
SB – Let’s wait until the final data come in
SB – It’d be good to do error bars on these, since some of the observables have super high error.
SB – Re: How to compare charges more directly? Could compare against RESP charges. I can help get a pipeline with Recharge set up pretty quickly for that.
SB – CBayly had looked at dimer energies as well. And there’s ESP comparisons that could we done. So we could build up a hierarchy of different tests for these sorts of ideas.
LW – I’ve been thinking about doing a weighted average of charge surfaces/ESPs, weighted by energy
SB – I made “openff-bcc-study-2” qca dataset, which is several hundred molecules with several conformers each - You could use this as a reference set or a way to calibrate timing.
SB – What are next steps?
LW – For my library, I want to be able to come up with good charges for a polymer made of building blocks. I’m still thinking about immediate next steps though.
CC – I’m interested in knowing whether increasing the number of conformers will smooth out the resulting charges. Like, for pentapeptides, just sampling a few backbone positions has an exponentially large number of
JW – In a sci-fi world, polymeterizer and Rosemary fitting could use the same API points, just with different keyword arguments. So it’s probably worth keeping in mind that there should be a lot of underlying machinery that could be shared here.
SB – One potentially high-impact avenue for partial charges would be graph-based conv NNs. My understanding is that they’re deficient in is a training set. This should entirely sidestep conformer dependence, but they need more personnel-time and data.
PB – What are plans for graph based methods?
SB – There’s the vcharge method from GIlson et al, which does some kinda ML stuff but forces it through a functional form with “softness” and “hardness” to arrive at reasonable charges.
