Deep learning of reaction properties via graph-convolutional neural nets

Abstract

Machine learning models are very successful in predicting various chemical properties. Graph-convolutional neural networks (GCNNs) are routinely used for the prediction of molecular properties, but their application to chemical reactions is largely unexplored. GCNNs allow for a learned extraction of important characteristics of a molecule and enable end-to-end learning, instead of relying on expert, system-dependent knowledge. However, the properties of chemical reactions, i.e. the combination of reactant and product molecules, are not readily accessible with current GCNNs which are designed to take molecular graphs as input. Recently, GCNNs based on the condensed graph of reaction (CGR) were shown to unlock the full potential of GCNNs also for reactions, where reactants and products are merged into a single pseudo-molecular graph, i.e. an artificial graph transition state. In this workshop, the anatomy of molecular GCNNs will be discussed in detail, as well as the changes necessary to encode reactions instead of molecules, including hands-on exercises to build your own reaction GCNN. Compared to previous approaches, GCNNs on CGRs offer a comparable or better performance with a lower number of parameters. We showcase the performance on different tasks, such as the prediction of barrier heights or rate constants, as well as the chemo- and regioselectivity of reactions.