Fig. 1. Schematic illustration of the attribution workflow.

a Overview of our workflow to interpret the Molecular Transformer. b Schematic of how the predicted probability difference between two products are attributed back to the reactant-reagent string in order to interpret the model’s understanding of selectivity. The IG attributions below the reactant SMILES are compared to the uniformly distributed probability difference (ua) below. c Schematic of how the latent space encoding of reactant-reagent strings are used to infer the learnt similarity between query reactants and those from the training set.