Integrating multi-type features and knowledge graph for graded prediction of drug-induced liver injury in humans

Drug-induced liver toxicity poses a threat to human health and remains a significant reason for drug withdrawal from the market. Therefore, early identification of drug-induced liver injury (DILI) during drug development is crucial. However, most studies on hepatotoxicity prediction are limited to single type of features or binary toxicity assessment. In this study, we propose a novel liver toxicity prediction model called MolFPKG-DILI (Molecular Graph, FingerPrint and Knowledge Graph-based DILI), which integrates multi-type compound features and knowledge graph for assessing DILI severity. Molecular fingerprints and molecular graphs capture different information of compounds, and models using individual features alone have shown limited performance. Our model incorporates an attention mechanism to effectively fuse the information from molecular fingerprints and molecular graphs. Furthermore, we leverage the relationship between drugs and other entities from the knowledge graph to achieve liver toxicity grading. Experimental results demonstrate that our proposed method exhibits highly competitive performance in both DILI/No-DILI and Most-DILI/Less-DILI classification. External validation conducted on an independent set of benchmark drugs yields satisfactory results, demonstrating the robustness of our approach. Additionally, we employ a series of interpretability methods to investigate the relationship between the different types of data utilized by the model and toxicity outcomes. These analyses highlight the interpretability of our method, providing valuable insights and support for drug toxicity evaluation.Author summary: Drug-induced liver injury (DILI) is a critical concern in drug development, often resulting in the withdrawal of numerous medications due to their hepatotoxic effects. Traditional methods of assessing DILI are not only costly but also labor-intensive. While recent computational strategies leveraging molecular fingerprints and descriptors have enhanced efficiency, they typically rely on single data types. This reliance can neglect the complex interactions among various molecular features, thereby constraining the predictive accuracy and effectiveness of these models. Our research introduces MolFPKG-DILI, an advanced model that integrates multiple feature sources—molecular fingerprints, molecular graphs, and expansive knowledge graphs with diverse drug-related data—to enhance prediction accuracy and detail. This model not only predicts the presence of DILI but also discerns varying degrees of liver injury severity. Utilizing attention-based feature fusion and relational graph convolutional networks, MolFPKG-DILI outperforms existing models, offering nuanced insights into DILI levels. External validation on 17 drugs and a series of interpretable analyses have demonstrated the competitiveness of the model in practical applications. This method is anticipated to serve as a dependable tool for toxicity risk assessment in early drug development, reducing the likelihood of development failures attributed to liver toxicity.