Angle parameters analysis @Lorenzo D'Amore | (Post slides here after talk) Slide 3 CBy – Is that (far right sulfonamide molecule, RCH-01236-06) a divalent or trivalent nitrogen? a22 should be divalent, so there’s no H on there. So that should be a negatively charged N. However, a22 is a neutral smirks pattern. Perhaps this broken valence situation is throwing off your analysis LD – I’ll double check this in the notebook CBy – We’re saying that the far right case is “wrong” - Is that coming from psi4? If so, then there must be some amount of impurities in our QM JW – far left 7-membered ring (JNS-00387-03), is that a ring pucker at an odd viewing angle, or is that just a very strange ring geometry? LDA – Bill Swope said perhaps the ring minimization had an issue? WS – not sure – let’s look at it in 3D and just check it’s not a bad viewing angle LDA – in 3D, even the hydrogens on the carbon pointing inwards, also point inwards General – that’s unexpectedly strange geometry CBayly – perhaps we should exclude weird geometries from training PB – QM can optimize to a higher energy conformation DM – but can there be a minimum with Hs pointing into the ring? That’s the worrying fact. If there is a minimum, perhaps we should filter it out, but we should check that something isn’t wrong with the optimization
Slide 4 Slide 5 JW – Woudl it be possible to test whether it’s the basis vs. the optimizer by using a more detailed basis/method? CBy – I wonder if the flattening of the NPN angle is the only way that Sage can relieve an issue with sterics/torsions? The QM structure looks reasonable, the Sage one looks unreasonable, so at least in this case the large deviation in angle is the right symptom to notice. BS – Another thing that could help would be to start a quantum opt from the Sage structure and see if it’s really a minimum at all CBy + JW – Good call
CBy – So it looks like there are 3 types of problems: QM going wrong Broken valence like the rightmost on slide 3, RCH-01236-06 QM going right and the Sage parameters having trouble
LD – Agree JW – For #2, it’s hard to know what we’re actually dealing with becuase the Hs are hidden PB – the smiles entry of RCH-01236-06 shows a double bond to sulfur, [H]c1c(c(c(c(c1C([H])([H])[H])[H])[H])S(=NS(=O)(=O)c2c(c(c(c(c2[H])[H])C([H])([H])[H])[H])[H])(=O)[O-])[H] CBy – ok, this is an unlikely resonance structure, not a broken valence
DM – I wonder if we should run something on our fitting data that looks for this weird charge rearrangement. So we have special parameters assigned for carboxylates and stuff, and maybe we have some sort of campaign to find cases like this. CBy – So we could like make special parameters/checks for specific parameters? DM – Not quite, more like “all exotic resosnace forms are forbidden, except ones that we explicitly allow to pass” CBy – This is hard, I think maybe we should keep this in mind and be grateful that it’s not too common. JW – Maybe we put in a filter in QC submission? This would save us a headache in fitting and benchmarking, but it wouldn’t help the user. CBy – SO2 double bound to N would raise a red flag for me. So that’s the simplest filter. DM – Important to remember that we’re seeing three cases here: DM – Maybe science twitter knows a way to detect weird resonance structures, seems like a cool ML problem CBy – I also really likes BSwope’s idea of submitting the weird final geometries from one method to an optimization with the other to see if they both at least recognize a local minimum. BS – JNS-00387-03- That particular molecule had 9 conformers, and 8 of them were fine, but this one was bad, and the energy was astronomical. So some sort of window for delta E could also have removed this from the scoring. CBy – At some point we’ll have to deal with repulsive sterics beyond r^12 And we’ll want sterically congested molecules to study that. So things like this case may be helpful there. LW – I prefer the energy window idea - We exclude high energy confs from training, so we should exclude them from benchmarking as well. DM – LD, could you throw out the high-energy conformers from this analysis and see if the conclusions change? LD – Sure, I can do that - I’ll check the conformers with really high relative energies. DM – Even filtering at like 100kcal/mol would help. LD – We already excluded a lot of outlier molecules. 99.5% of conformers were in an energy window within 60 kcal/mol, and everything beyond that was excluded. JW – maybe window should be a little bit narrower LD – the 7-membered ring was definitely outside the window LW – Yeah, then it might be worth using the window again and regenerating these graphs
|
