Prediction of drug-induced liver injury (DILI) using graph convolutional network on heterogeneous network

Abstract

In this paper, we propose a drug-induced liver injury (DILI) prediction model that uses a graph convolutional network with novel architecture on a heterogeneous network consisting of drug-target and gene-gene interactions. Our model uses not only drug structure information but also information about interactions between drugs and genes. The model was trained on 1006 drugs, and the performance was evaluated by nested cross-validation. The model achieved Matthew’s correlation coefficient of 0.327 outperforming baseline models that use structure information only, and distinguished structurally similar but having different bioactivity drug pairs. DILI-related genes were selected based on the attention-based novel architecture of the model, and suggested previously undefined DILI mechanisms. Our model successfully integrated various information for predicting DILI and can be applied easily to other research fields due to its simplicity.