Recap of last week | Bayly feedback/conclusions CD – We were approaching from “universe 1” (the QM-for-one-conformer view) rather than “universe 2” (the best-general-FF view). So moving forward we need to do something with multiple coformers if we’re going to remain relevant. OM – That makes sense, and we did initially (mistakenly) frame the question as “solve this problem starting with one conformer”, when really we shouldn’t have introduced that constraint. So we DO still want to check for connectivity rearrangement, but BEYOND that we want to be looking at multiple-conformer approaches. CD – It’s still an open question how to restraint the optimization using the smallest number of constraints possible. CBayly’s idea to restrain the dihedrals is a good call on this front. OM – Also could look into whether ELF conformers even experience proton migrations – If there were closely-separated large charges then they’d fail the ELF test. CD – A lot of the feedback we had was very broad, but one thing we need to note is that, even if we’re not dealing with proton rearrangements/electrostatic collapse, the partial charges will change if the geometry changes. So we should kinda ask whether we should ALLOW the conformer to change at all. So basically, “what does the user expect?”. JW – I think the “do antechamber original, and then if there’s a connectivity rearrangement then do maxcyc=0” is a 90% solution to everything in universe 1. So let’s focus on gains to be made in universe 2 OM – Would we expect to be able to get equivalent ELF results from the OpenEye-omega-quacpac stack and the AmberTools/RDKit-embedconfs-antechamber-sqm stack? JW – Could test whether ELF1 results from both packages are similar/equivalent CD – If we start futzing around with solution to “universe 2” problems, what “universe 1” method should we use?
CD – What about looking further into possible solutions to the “universe 1” problems? Like a more accurate or performant constrained AM1 method. OM – Looking at endgames for more work into EITHER universe 1 or 2, we want good reference/test data. Would we need to make a pipeline for hydration free energy to move further here JW – It does seem like we’ll benefit from having larger-scale benchmarks. CD – But this will require refitting the entire FF/also vdW terms, right? JW – But beyond having good benchmarking data, getting CONSISTENT partial charges alone should improve the quality of the FF fits. RIght now conf gen is basically a random noise generator n the middle of our fitting and application process. So decreasing the level of noise would by itself improve the quality of the FF in the next fitting cycle. OM – Having inconsistent internal charging methods decreases the meaninfgulness of our own benchmarking
CD – If we look at solutions to universe 2, how can we compare whether RDKit and OpenEye are giving us the same ELF conformer? JW – Conformer RMSD calculations are really complex, so I would try to sidestep comparing the geometries until we have to. CD – If we’re looking at universe 2 solutions, which universe 1 method should we use? Fully constrained in both OE and AT?
|
Next steps | Universe 1 solutions CD will look into more perofrmanty ways to incorporate restraints/constraints into AM1 optimizations CD will begin putting antechamber-original-then-maxcyc=0 into production, using RDKit to check for connectivity changes CD – Should I also look at avoiding internal hbonds (that aren’t proton transfers) or other “electrostatic collapse” JW – That could be cool to look at, but any way that we use to detect this will be heuristic, and heuristics are harmful just by themselves. So I’d be interested to see how much these sorts of rules help, but that benefit will need to be weighed against the harm of using heuristics OM – Internal hbonds/elestrostatic collapses are kinda a universe 2 question, so we can keep looking into it even after we choose an answer for universe 1.
Universe 2 solutions Reference/benchmark data
|