A Comprehensive Survey on Graph-Structured Small-Molecule Drug Discovery and Development with Deep Learning

Small-molecule drug discovery remains a cornerstone of modern therapeutics, yet traditional development pipelines are often limited by high costs, long timelines, and low success rates. Recent advances in graph-structured deep learning have significantly transformed the modeling of molecular structures, biological interactions, and pharmacological responses. By naturally representing molecules as graphs, these approaches provide an effective framework for capturing structural, geometric, and physicochemical information across multiple biological scales. In this survey, we present a comprehensive review of recent advances in graph-structured small-molecule drug discovery and development. To provide a unified perspective, we organize the literature into a four-layer framework covering data resources, representation learning, task formulation, and downstream applications. We first summarize the evolution of molecular representations from molecular sequences and topological graphs to geometric structures, molecular dynamics, and electronic field representations. We then introduce a unified task formulation covering six major tasks: molecular property prediction, drug-target interaction and affinity prediction, drug response and perturbation prediction, drug-drug interaction prediction, molecular generation, and molecular optimization. Building upon this framework, we review recent developments in graph neural networks, geometric learning, diffusion models, and foundation models. Finally, we summarize representative benchmark datasets and evaluation resources, and discuss several key challenges, such as out-of-distribution generalization and data scarcity, together with promising directions for future research.