Virtual sites | LW will link slides here Update Move this topic (+ other non-release topics) to FF fitting meeting? JW – Is brown line clearly the best on aryl chrloide plot? (next slide, dipole comparison) Next slide (Br ESPs) CBy – So my interpretation is that, for chrloine, vsites didn’t really do anyhting since the lines basically overlapped. But for Br there’s a significant improvmenet. LW – I’d argue that the chlorine ESPs also showed significant improvement. DM – And I suspect that the benefit of chlorine vsites would increase with more heavily polychlorinated compounds (based on my previous work)
(next slide) JW – This seems like the first clear example of refitting making something worse MS – Why do we think that is? LW – We’re only training to ESPs, which might be it. MS – Interesting that training to ESPs makes dipoles worse. CBy – Agree, that seems odd LW – Could mean that we should expand our datasets to avoid discrepancy between training and test. BS – Could depend on where the points are measured. If the points are far out the dipole is probably being well measured. If the points are close in, then they may be struggling with the limitations of multipole expansion. LW – It’s a standard MSK CBy – That’s something like {1.2, 1.4, 1.6, 2.0} x vdW radius outside of mol LW – Are you suggesting measuring potential at points further away? BS – Maybe, or just weighting how much points at the different distances are scored.
(next slide: Pyr: vsites) CBy – Are pyr vsites allowed to change distance? LW – Not really, constrained around 0.3A. CBy – … LW – The ones where I didn’t restrict it (the green one) found a very quirky minimum CBy – This seems concerning, this is a strong qualitative contrast to my prior experience (which was admittedly on a much smaller scale)… … CBy – One thing that I did was not actually looking at electrostatic potential around pyridine, I’d just look more qualitatively at charge distribution. Never done as much of a large analysis as this. I’d expect to be able to look at some flagship molecules and ensure that positive and negative regions are correctly placed. LW – Interesting. I’ve looked at this a little, and SBoothroyd wrote a library called Molesp and I glanced as a few using that. The mols with vsites from my candidate FFs qualitatively matches the correct charge distribution. CBy – Interesting that the ESPs would visually look good but that the plots here don’t show a clear improvement. LW – I like your take on “making sure that parameters that get assigned are transferrable”. The way we ended up with v1 was when the vsite “pushed” a bunch of charge on to other atoms. Whereas … CBy – But the CH BCCs were retrained in all of them, so why would it be different in the application? LW – …
(halfway overview slide) CBy – As a big advocate for vsites, I’m asking myself how I rationalize these results. I remember I TRAINED ESPs to vsites, and TESTED on dimer energy. But since here you’re looking at dipole direction and magnitude… For me, the vsites are most important are in structured environments like binding sites. So I’m interested to see dimer energies. MS – Could it be that the closer ESP points should have higher weight in training/scoring? CBy – You could weight ESP by distance to the nearest atom? BS mentioned this earlier. LW – Clarify? CBy – Since the ESP magnitude is larger at close range, it would be relatively larger in the score. MS – Are tehre more points at larger distances? JW – Is point density constant? or angle density from the atom’s perspective? DM – Can’t recall if it’s constant density. LW – IIRC, the implementation in openff-recharge is constant-density. But generally the idea of different weighting would be good.
(Enthalpies of mixing improved) (Br: enthalpies of mixing are worse..) JW – Oxygen lone pairs may make ester enthalpies happier? CBy – Rabbit ear vsites on oxygen are kinda significant on carbonyls, but not on esters. RAbel/Schrodinger didn’t find them this important, even though they added vsites elsewhere. LW – I think the oxygens are pretty undifferentiated (there’s only one O vdW parameter in Sage), so I was thinking about splitting those out. CBy – AMBER always had two O vdW types. Could be good to try, though note that we’d expect adding parameters and retraining would be expected to improve things, so be sure to run control w/ and w/o vsite. MS – Generally concerning that we’re seeing such different things from people who looked at this before. CBy – As one of the people who looked at this before, what’s different this time is that we’re looking at global properties… It’s also important to keep in mind that liquid properties may not improve, but accurate modeling of binding is what we want to hit, and we’ll need to measure dimer energies to see that. LW – Thinking about Schrodinger work, they did add a vsite for Br, and IIRC the phys prop benchmark showed it getting worse. DM – … MS – Were phys props for nitrogen compounds outliers? Or were they basically OK? LW – We see they mostly improved. The outliers we see mostly improved. MS – Did the props involving N vsites improve as much as props with other sorts of vsites? Is there a signal that our FFs were always bad at nitrogen, and so there’s more of a challenge around them and also more of an opportunity for improvement. MS – We know, for water, vsites make a big difference. So moving a dipole a little can make a big difference. So I’d expect these vsites to have a bigger effect. CC – But water vsites are trained to phys props, whereas this work is trained to ESPs. MS – True, but the point is that vsites can sometimes have a really big effect. CBy – A comparable case to water could be hydrogen chloride, hydrogen bromide, or ammonia - similar in scale. But for larger mols, where can we find good exptl data? I think xtal structures, maybe CSD data. DM – I’m hesitant to send OpenFF off in the direction of xtal assessment/free energy assessment. So I wonder if we could quietly send a candidate FF out and when groups like mine have trouble with specific targets in sims we could try the vsite FF out. LW – That makes sense. MS – Or do some citizen science and open it up to everyone. DM - THuefner might find this useful. And this could be cool for annual report. CBy – That’d be cool for industry.
BS – I think in this data we’re looking at metrics that are relatively insensitive to vsite change. So I think we need to look at local things where we’d expect vsites to be more significant. For me that’d be dimer energies and dimer structures. DM – Agree, would be interesting to see the effect on dimer energies and geometries. Not sure about workload. CBy – … I think what we want to do here is to scan some displacements/angles for dimers … (see recording ~60 mins)
JW – Potential roadmap item? CC – CHARMM folks have looked at osmotic pressures for nonbonded fitting. They used this for the DRUDE paper. We might want to pull that data and try fitting to it.
(Summary) CBy – Key difference is that dimer interactions are often very local, interesting to see that the global metrics don’t vary much. MS – Mixing enthalpies with water all got better. DM – This is all kinda consistent with “vsites are doing some good, it’s just hidden wehn we look at bulk properties not involving water” CBy – Agree, would think that you should also look at methanol, since that’s a strong hbond donor. And pyrrole (which I think you did show). JW – So is the plan to go forward something like “try splitting oxygen and refit, if that works, and see if something great happens in protein-ligand FE calcs?” CBy – Could we whittle down which permutations of FFs/electrostatics you’re testing? … And scrunch down things like dimer datasets to focus on things where we expect a response? LW – Yeah, those sound good. For the “assay”, DES370k covers that subset well. So I could look at that more closely. And I want to try splitting out vdW parameters with vsites. And maybe dropping the Br vsite? MS – Do we expect Br to have a vsite? We might drop that and move forward. CBy – Br vsite should have a bigger effect than the Cl vsite. And I’m glad that the analysis here included looking directly at the N vsites to ensure charge distribution made sense. Also, is the level of QM theory sufficient to get the sigma hole distribution right? BS – These are all being compared with DFT? LW – The dimer energies yes. And the ESPs are coming from HF 6-31G CBy – I’m not sure HF 6-31G would put the sigma hole on Br LW – Visually, it does. CBy – Randomly pulling some compounds from the validation set to ask whether this makes sense… If not that, xtal structure stuff or dimer energies would be good.
|