MGCPI: A Multi-granularity Neural Network for Predicting Compound-Protein Interactions

The identification of compound-protein interactions (CPIs) is an essential step in the drug discovery process; however, existing sequence-based or graph-based single-granularity compound representations have difficulty in accurately predicting CPIs. In this paper, we propose MGCPI (Multi-granularity CPI), an end-to-end deep learning framework to predict the compound-protein interactions, which integrates the molecular features of both graph and sequence representation from the input and mines protein structure information by transformer and pre-training methods. Our experiments demonstrated that the multi-granularity molecular representation method is able to fuse protein information from multiple perspectives to enhance the predictive capability of the model and achieve competitive or higher performance compared to various existing CPI prediction methods. Additionally, the ablative analysis verified that the multi-granularity model is more robust than single representation-based models.