Prudencio Tossou

Interpretable Neural machine translation for reaction prediction and drug discovery

In this talk, we explore how to use recent advances in neural machine translation to tackle a rate-limiting step in drug discovery, namely forward reaction prediction. We propose a novel end-to-end translation method which receives as input the molecular graphs of the reactants and reagents and outputs the SMILES representation of the products. Named Hybrid Graph-SMILES Transformer (HGST), the method addresses the weaknesses of existing reaction predictors while taking advantage of their strengths. In fact, compared to existing models, HGST is faster at training and inference and more accurate, surpassing in all regards the models behind the IBM RXN for Chemistry. But more importantly, the model comprises an explainer that is able to automatically highlight atoms that are important and involved in the reactions. These reaction centres are often provided as inputs for existing models but in our case, we get these probable reaction centres and their associated electron movements simply as a byproduct of our modelling choices.