Intelligence on the Graph: Graph Neural Networks for Mechanistic Drug Target Discovery

Drug discovery is increasingly challenged by rising costs, long development cycles and high attrition rates, with accurate target identification remaining a critical bottleneck. Although artificial intelligence (AI) has demonstrated transformative potential, the systematic application of graph neural networks (GNNs) to drug target discovery remains underexplored. To address this gap, this paper provides a comprehensive and structured analysis of recent advances in GNN-based methods for drug-target interaction (DTI) and drug-target affinity (DTA) prediction. We dissect the methodological foundations of representative architectures including graph convolutional networks (GCNs), graph attention networks (GATs) and graph autoencoders (GAEs), and compare their mechanisms, advantages and applicable scenarios in modeling complex molecular and biological systems. Also, we synthesize frontier paradigms such as multimodal data fusion, high-order graph reasoning and dynamic GNNs, which enable the capture of atom-residue interactions, multi-target coordination mechanisms and cross-scale biological features. By systematically mapping methodological innovations to biological applications, this paper offers both theoretical guidance and translational insights. The key contributions of this paper include: (1) establishing a comparative framework that clarifies when and how different GNNs architectures can be applied in drug target discovery; (2) integrating cutting-edge paradigms rarely addressed in prior reviews, such as multimodal fusion and high-order graph modeling; and (3) highlighting representative case studies that bridge algorithmic innovation with practical drug discovery outcomes. Collectively, this work provides an authoritative and forward-looking reference, promoting the development of AI-driven, efficient and interpretable drug discovery pipelines. • We comprehensively summarize recent advances in Graph Neural Networks (GNNs) for drug–target interaction (DTI) and affinity (DTA) prediction. • Comparative analysis of GCNs, GATs and GAEs reveals their strengths in modeling molecular and biological network complexities. • Emphasis on multimodal data fusion, high-order graph reasoning and dynamic learning enhances DTI/DTA predictive performance. • Representative breakthroughs include atom-residue interaction modeling, synergistic multi-target discovery and cross-scale feature extraction. • Key challenges such as data sparsity, computational load and interpretability are discussed with proposed solutions. • A closed-loop framework integrating generative modeling, lightweight GNNs and experimental validation is proposed for optimized drug target prediction. • The paper advocates a paradigm shift toward mechanism-informed AI to advance precision medicine and rational drug design.