Featurising Reactions for Deep Learning


I am looking to create a dataset of Hydrogen Abstraction Reactions to feed into the DMPNN. I see that there is a featuriser class called DMPNNFeaturizer, however, from my understanding it only deals with SMILES of molecules. While I want to be able to represent a reaction SMILES.

So, for example, if I have:
OO + [CH2]CCCO >> [O]O + CCCCO

Does anyone have advice on how to tackle this?

DeepChem doesn’t have great support for reaction featurization at present (something we need to improve). One way you could handle reactions with DMPNN would be to write a custom featurizer that featurizes and combines each reactant and product together into a joint graph