Excellent. I'd been wanting to implement something like this but these guys beat me to the punch.<p>Most existing research on NNs in chemistry uses the SMILES string representation and feeds it into a recurrent neural network. The problem with SMILES in my opinion, is that SMILES is not unique, and there are a gazillion ways to represent the same molecule. Thus, the RNN is probably overfitting to the quirks of the SMILES serialization algorithm used by the US Patent Office and/or ChemDraw and/or whatever program was used to generate the library of molecules.<p>Convolutional graph networks are the obviously correct way to handle molecules.