Deep Learning Strategies for Predicting Drug–Target Interactions: Advances, Challenges, and Future Perspectives

Drug–target interaction (DTI) prediction lies at the heart of modern drug discovery, determining whether a candidate small molecule will bind to and modulate a biological macromolecule of therapeutic relevance. Traditional experimental high-throughput screening is expensive, time-consuming, and constrained by library size, while classical computational approaches—docking, pharmacophore modelling, and quantitative structure–activity relationship (QSAR) modelling suffer from limitations in scalability and generalizability. The emergence of deep learning (DL) has fundamentally transformed the field, enabling end-to-end learning of molecular representations and interaction patterns from heterogeneous, large-scale biomedical data. This review provides a comprehensive synthesis of DL-based DTI prediction methodologies, covering convolutional neural networks (CNNs), recurrent neural networks (RNNs), graph neural networks (GNNs), transformer-based architectures, autoencoders, and multi-modal fusion frameworks. We discuss the critical role of molecular representation—from one-dimensional SMILES strings and circular fingerprints to three-dimensional molecular graphs and protein contact maps. Benchmark datasets including Davis, KIBA, BindingDB, ChEMBL, and PDBbind are reviewed with respect to their composition, metric conventions, and appropriate use. We critically examine key challenges: data scarcity and imbalance, negative-sample bias, interpretability deficits, cold-start generalization, and the limited availability of experimental three-dimensional protein structures. Emerging solutions—pre-trained chemical language models, AlphaFold3 integration, federated learning, knowledge graph-augmented GNNs, and causal interpretability methods—are discussed as future directions. This review aims to serve as an authoritative reference for computational chemists, bioinformaticians, and medicinal chemists seeking to leverage DL for accelerated, cost-effective drug discovery.