A Multimodal deep learning method for predicting drug-induced liver injury using structural and sequential molecular representations.

Drug-induced liver injury (DILI) remains one of the leading causes of drug development failure and post-marketing withdrawal. Despite recent advances in computational modeling, many existing approaches rely heavily on structural features such as molecular fingerprints, often neglecting the rich sequential information encoded in SMILES representations. In this study, we propose a multimodal deep learning framework that integrates both substructural features and sequential information to improve the prediction of DILI. Specifically, we utilize a diverse array of molecular fingerprints-KlekotaRoth, CDKExtended, MACCS, PubChem, Estate, Avalon, ECFP2, Morgan, and RDKit-to capture the structural aspects of molecules. In parallel, we use ChemBERTa, RoBERTa, ChemBERTaV2, and SMILES2Vec to extract high-level sequential information from SMILES strings. By fusing complementary modalities, our approach achieves superior performance compared to single modality methods. Furthermore, an ablation study reveals the synergistic contribution of both modalities in capturing the complex biological mechanisms underlying liver toxicity. Our findings underscore the importance of multimodal integration in developing more accurate and generalizable predictive models for DILI.