Machine learning for predicting drug–drug interactions: Graph neural networks and beyond

Identification of interacting drugs before application would be imperative to mitigate the serious risk represented by drug–drug interactions for patient health. Machine learning–based methods are increasingly recognized by regulatory agencies as tools with a central role in drug development, including the identification of novel interactions. In recent years, graph and hypergraph neural networks delivered promising performance improvements compared to non–graph-based methods on the field. In this primer, we discuss recent developments of graph and hypergraph neural networks and highlight the potential of incorporating protein and metabolite data into the identification task to provide a new, more comprehensive, systems biology–based perspective on drug–drug interactions.