PB – Last week, was talking about training to ddE, instead of differences with respect to a fixed QM minimum. Still working on this. But back in April, I transformed a torsion target into a ddE target, and am still looking into how it performs.
DM – Restraint in MM optimizations?
PB – Yes
DM – And the logic is that, if the restraint is removed, the MM could stray away?
PB – …
CC – Are restraints applied only to driven torsion atoms, or all heavy atoms?
PB – All heavy atoms.
CBy – Are you enforcing on the QM that the potential is a minumum at the points where you’re comparing? So, you’re fitting a MM potential function , where the QM knows that the points are minima (have first derivative of 0), but does the MM know?
PB – In the case I’m showing we are doing force matching. (?)
CBy – It’d make sense that you get spurious minima if you’re not fitting to forces
JW – If PB …
TG – It is free to mis-order the configurations, even when fitting to ddE RMS.
BW update
DM (chat) – (History) Meghan did a lot, and handed to Jessica who did some more with Pavan but left before finishing. I think a lot of it is still in a similar place to where Meghan had it. So Brent is not far off; probably leading contributors were Meghan and Pavan.
JW – Note - Matt’s benchmarking infrastructure hasn’t been quantitatively validated against previous benchmarks, though we do see qualitative agreement (like, it ranks the FFs in the same order)
CBy – In intro piperazine slide. where is there a mutiplicity of 9?
BW – The pictures are arbitrary, I don’t think any of these actually have 6 or 9 multiplicity.
PL – I’m not sure I get the issue - This isn’t counting the lone pairs. Isn’t the issue just solved by counting the X3s and X4s in the rule?
DM – Right, this project came about when we realized that the central atoms in torsions should have explicit valences. So that’s what we’re seeing.
LW – TFDs actually seem improved here (peak shifted left). So this could be entirely due to changing the fitting data. But we do have to add these new targets so that there’s coverage of everything in our new FF.
BW – Yeah, I should just benchmark on the old test set instead of expanding it.
DM – (chat) – We already know changing the dataset changes FF quality, so when we compare HOW we fit we should hold the dataset fixed. (or in this case comparing WHAT we fit)
CBy – When curves look this similar, it’d be good to start focusing on individual cases that show improvement/degradation.
PB – Looking at individual torsion profiles would be informative - Could see whether they actually improve in terms of fit. This is because the benchmark set may not have many molecules exercising new parameters.
DM – Agree, this kinda mirrors CBy’s point.
TG update
(TG shows some slides from previous talks for intro)
TG – Update this week is that I’m basically unable to find eigenvectors of (big matrix), so I’m going to have to do this by regression.
PL – Goal is to fit to a torsionscan without refining other parameters? Like, fitting torsionscans without worrying about other force components?
TG – Basically, yes.
Open discussion
PL – I’m working on a project on to mix and match FFs. In theory things are transferrable in general FFs. But I’m finding complications…. I’m finding improvements, and it seems like things like torsion fitting can be improved…
DM – The way we got to where we are is that existing FFs like GAFF are combinatorially hard to extend, and you end up with tons of redundant parameters. So we figured out independent parameter assignment and that’s reduced complexity a lot. But it’s clear that we should be adding parameters.
PL – We have some internal results that I’d love to share… But there are substituent effects like aniline that will always require independent parameters. I don’t think that explicit SMIRKS will be suitable for this, there will need to be some other sort of rules.
DM – We’re assembling automated benchmarking
PL – Could this be implemented without using SMIRKS?
JW – This should be possible in a parameterhandler plugin. You can run pretty much arbitrary code in there. But the benchmarking suite isn’t ready for external use.
PL – …
CBy – … We tried looking at WBO interpolation, it’s a promising concept but results were mixed. I think it was limitations of datasets and infrastructure at the time.
PL – Not criticizing the existing work. But based on my recent results, I think there’s additional structure that can inform parameter assignment. Right now there’s a lot of informatics that would need to be compressed into SMIRKS.
DM – One reason I’m glad you’re here is the WBO thing - I liked the idea but …
PL – Yeah, would need an expanded SMIRKS language with an open specification for conjugation/some other descriptor… I’d love to join meetings on how to make validation sets. I can present on my work next week.
DM – It’d be promising to plug this in to the automated benchmarking once it’s ready.
PL – Not sure about developer-time on that, but I’m happy to continue this conversation.
CBy – If we had a s
Could we connect a plugin that uses a function to get Paul’s work in?
PL: it’s not just based on one SMIRKS pattern
PL: a lot of our work is dependent on knowing the detailed hybridization of every bond (nearby? in the molecule?
DLM: as part of the AM1 calculation we can get partial bond orders for the molecule. Things similar to this could be within scope. We could shove a molecule into a tool and then it could come out with information on which parameters should get assigned
JW – could put together a proof of concept pretty easily, as long as we think the idea is promising. Would help if you could send over code that runs (in any language, we can use Docker)
PL – Also would like to discuss validation sets
DM – Yes, we can do that too.
…
CBy – Good to think of SMIRKS parameters are “bracketing” the chemistry that needs the most attention. So we should expect there to be relatively few parameters.