Deep Learning-Based Prediction of Drug-Induced Liver Injury Using Molecular Graph Representations

Abstract

Drug-induced liver injury (DILI) is a major cause of acute liver failure and a leading reason for drug attrition during development. Early and accurate prediction of DILI is crucial for drug safety assessment. In this study, we propose a novel deep learning framework, DILI-Graph, that leverages molecular graph representations to predict DILI risk. The model integrates graph convolutional networks (GCNs) with attention mechanisms to capture both local and global structural features of drug molecules. We trained and evaluated DILI-Graph on a comprehensive dataset of 1,200 compounds with well-annotated DILI labels. Our model achieved an area under the receiver operating characteristic curve (AUC) of 0.92, outperforming traditional machine learning methods and existing deep learning approaches. Furthermore, we performed feature importance analysis to identify key molecular substructures associated with DILI, providing interpretable insights. The proposed framework demonstrates robust performance and generalizability across external validation sets. Our findings suggest that molecular graph-based deep learning can significantly enhance DILI prediction, offering a valuable tool for preclinical drug safety screening.

1. Introduction

Drug-induced liver injury (DILI) represents a significant challenge in drug development and clinical practice, accounting for approximately 50% of acute liver failure cases in the United States. The complex mechanisms underlying DILI, including oxidative stress, mitochondrial dysfunction, and immune-mediated responses, make it difficult to predict using conventional approaches. Traditional in vitro and in vivo models are time-consuming and often fail to recapitulate human-specific toxicity. Consequently, there is an urgent need for computational models that can accurately and efficiently assess DILI risk early in the drug discovery pipeline.

Recent advances in deep learning have opened new avenues for toxicity prediction. In particular, graph neural networks (GNNs) have shown promise in learning from molecular graphs, which naturally represent the structure of chemical compounds. Unlike traditional fingerprint-based methods, GNNs can capture the spatial arrangement of atoms and bonds, enabling more nuanced structure-activity relationship modeling. In this study, we introduce DILI-Graph, a GNN-based framework that incorporates attention mechanisms to focus on toxicologically relevant substructures. Our approach aims to improve prediction accuracy and provide interpretable insights into the molecular determinants of DILI.