Multi-Modal Deep Learning for Accurate Drug-Target Interaction Prediction in Precision Medicine

Abstract

Precision medicine requires accurate prediction of drug–target interactions (DTIs) to enable individualized therapeutic selection and minimize adverse effects. This review synthesizes multi-modal deep learning frameworks for DTI prediction, positioning them as exemplar intelligent engineering systems (IES) empowered by artificial intelligence. Grounded in cyber-physical systems (CPS) theory and information theory, these frameworks integrate heterogeneous data modalities—molecular graphs, protein sequences, gene expression profiles, and clinical records—through graph neural networks, transformers, and cross-modal fusion architectures. The analysis traces four AI-driven transformation pathways: from population-level standardized screening to personalized, flexible prediction models; from reactive post-hoc validation to proactive interaction forecasting; from manual expert curation to autonomous intelligent inference; and from isolated computational tools to integrated cyber-physical ecosystems linking research, clinical decision support, and patient monitoring. Concrete examples include multi-modal transformer networks processing SMILES strings and amino acid sequences for binding affinity regression, and graph-based models fused with electronic health record data for adverse-event prediction. While these approaches substantially improve predictive accuracy and generalization to unseen compounds, critical limitations persist regarding data heterogeneity, model interpretability, and distribution shift across patient populations. Future prospects center on physics-informed multi-modal architectures, federated learning for privacy-preserving integration, and digital-twin-enabled virtual screening. The work provides a rigorous template for advancing trustworthy AI in precision medicine within broader IES paradigms.