Answers are grounded in this paper and cite their sources.
Review reveals an architectural evolution toward structure-aware prediction in computational drug discovery, highlighting persistent data bias and evaluation challenges.
Background Drug-target interaction prediction is an important task in computer-aided drug discovery. It supports virtual screening, drug repurposing, and lead optimization. Recent advances in deep learning have improved the representation of drug molecules and protein targets, but differences in datasets, model design, and evaluation protocols still make direct comparison difficult. Methods This review summarizes recent studies from four aspects: data resources and task settings, drug and target representations, prediction methods, and evaluation strategies. It covers molecular strings, fingerprints, molecular graphs, protein sequences, pretrained representations, and structural information. Traditional machine learning, sequence-based deep learning, graph neural networks, Transformer-based methods, and structure-aware approaches are reviewed and compared. Results The reviewed studies show a gradual shift from handcrafted features to learned representations and multimodal modeling. Graph-based methods describe molecular topology more directly, while pretrained and structure-aware methods make better use of protein sequences, binding pockets, and three-dimensional structures. However, current studies still face data bias, similarity leakage, limited generalization to unseen drugs or targets, high computational cost, and weak interpretability. Reported performance is also strongly affected by dataset construction and data-splitting strategies. Conclusion Future studies should use stricter evaluation protocols and give more attention to cold-start prediction, structural information fusion, model interpretability, uncertainty estimation, and experimental validation. More consistent benchmarks are also needed to support reliable comparison and practical use in drug discovery.
No takes yet. Share an insight, caveat, or question.
Wan Qin (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: