Aromaticity in representation

[Dom1L]

@kvkarandashev pointed out some problems to me for a simple benzene molecule

Training on random 3000 qm7 molecules gave a benzene molecule that was a bit off (see image). Especially the hydrogens were quite wrong. Increasing training set size improved the heavy atom coordinates but the hydrogens stayed out of plane.

An issue could be that aromaticity is not considered in the representations.
Theoretically, it isn't a problem to include this as long as there are parameters for it and it is represented in the graph via 1.5 bonds.
The problem is more on a technical side that getting a dataset and/or query that contains 1.5 bonds is more difficult.
RdKit has an aromatic bond type, but this requires some tweaking of xyz2mol. I'm also not sure whether xyz2mol actually detects aromaticity.
If someone queries a SMILES it would be easy to extract aromatic components, I see the problem more on the training data and parameter side.

Another way would be to augment the training data with resonance structures, but this would be more of a deep learning approach than KRR...

Any ideas/opinions @ferchault ?

[Dom1L]

Actually the resonance structures wouldn't fix the hydrogen problem.
One hydrogen can't see a full ring, it only sees 5 atoms at most and the representation covers only bond length paths from heavy atoms to hydrogen.
Including the heavy atom bond paths to each other would bring more uniqueness and could also cover rings. This could help with this issue.

[ferchault]

Consider benzene and cyclohexane. Wouldn't this difference already be covered by g2s both in bond order and bond length approaches? Perhaps it does make sense to include aromaticity, but I'm not sure about the 1.5 solution, since that implies interpolation along the bond order. Did we try one-hot encoding for bond orders? In any case, is aromaticity not implicit in the bond orders of the whole molecule?

The other issue is constructing the molecule bottom-up: in benzene, all heavy sites are symmetric, in cyclohexane, they're not. Could it be the order in which we build the molecule?

For the hydrogens: how about symmetrizing them subject to the heavy atom scaffold?

[ferchault]

re implicit: perhaps then it's still worth making aromaticity explicit.

[Dom1L]

could you elaborate on the hydrogen part? I would think that making the connection between the heavy atoms clear in the hydrogen representation (changing 2-3 lines of code) helps with uniqueness.

re implicit: yes it's implicitly covered, but since an aromatic ring is represented as 3 double bonds, these bonds can alternate in the ring. So for rings with different substituents which are not symmetric anymore this yields 2 possible/valid graphs per aromatic 6 ring. Right now I assume its more or less random which one we get from xyz2mol.
(Also something to consider for localG2S)

[ferchault]

re Hydrogens: Generally, I'd expect that the symmetry of the H locations is the same symmetry group as the heavy atoms. The representation fix is clearly preferable, since it should work more generally. If we redesign how the Hs are done, then symmetry (be it in the representation, be it afterwards) might help.

re implicit: How about averaging the representations from resonance structures instead? That makes things unique and works generally because it also fixes the benzene issue. I guess the question is whether two resonant structures should have the same geometry in our predictions. What is the chemists' answer to this? My understanding would be that it's still just one molecule, so one geometry.

[Dom1L]

The geometries have to be the same for two resonant structures, so averaging is a good idea.
We would have to make sure that we get all resonance structures for the training set as well as queries, which is a problem on its own..

I'll take a look if I can gather some insights of the impact of this. Could be that it lowers the learning curve offset.

[ferchault]

resonance structures could come from rdkit.Chem.rdchem.ResonanceMolSupplier - even if rdkit is wrong from time to time, that should still help