Drug–target affinity (DTA) prediction is crucial for understanding molecular interactions and aiding drug discovery and development. While various computational methods have been proposed for DTA prediction, their predictive accuracy remains limited, failing to delve into the structural nuances of interactions. With increasingly accurate and accessible structure prediction of targets, we developed a novel deep learning model, named S2DTA, to accurately predict DTA by fusing sequence features of drug SMILES, targets, and pockets and their corresponding graph structural features using heterogeneous models based on graph and semantic networks. Experimental findings underscored that complex feature representations imparted negligible enhancements to the model’s performance. However, the integration of heterogeneous models demonstrably bolstered predictive accuracy. In comparison to three state-of-the-art methodologies, such as DeepDTA, GraphDTA, and DeepDTAF, S2DTA’s performance became more evident. It exhibited a 25.2% reduction in mean absolute error (MAE) and a 20.1% decrease in root mean square error (RMSE). Additionally, S2DTA showed some improvements in other crucial metrics, including Pearson Correlation Coefficient (PCC), Spearman, Concordance Index (CI), and R², with these metrics experiencing increases of 19.6%, 17.5%, 8.1%, and 49.4%, respectively. Finally, we conducted an interpretability analysis on the effectiveness of S2DTA by bidirectional self-attention mechanism. The analysis results supported that S2DTA was an effective and accurate tool for predicting DTA.

Advances in protein structure prediction have transformed the landscape of structure-based drug discovery, enabling deep learning models to incorporate spatial constraints into the design of target-specific ligands. This review provides a comprehensive synthesis of structure-aware molecular modeling from a task-centric perspective, focusing on binding pocket identification, interaction prediction, pose estimation, and complex modeling. We highlight the technological evolution from traditional docking and scoring frameworks toward geometry-informed deep learning architectures that encode protein structures via surface geometry, equivariant representations, and multi-modal embeddings. Special attention is given to recent progress in structure-conditioned molecular generation. We classify generative approaches into four core strategies: sequence-based generation with 3D conditioning, fragment-based linking and growing, graph-based generation under structural constraints, and 3D coordinate–based generation including diffusion models. Each paradigm balances chemical validity, spatial fidelity, and computational tractability in distinct ways, with diffusion-based and point cloud models emerging as powerful tools for synthesizing pocket-complementary molecules in full 3D space. We also discuss the emergence of co-folding models, which unify protein folding and ligand binding into a single predictive framework, bridging the gap between sequence-level learning and structural resolution. Finally, we examine the key challenges of data scarcity, generalization, and multi-objective control, and outline future directions toward scalable, interpretable, and physically plausible generation pipelines. By tracing how structural knowledge is reshaping AI-driven drug design, this review aims to provide both a conceptual roadmap and practical insight into next-generation molecular modeling.

Drug-target affinity (DTA) prediction is crucial for understanding molecular interactions and aiding drug discovery and development. While various computational methods have been proposed for DTA prediction, their predictive accuracy remains limited, failing to delve into the structural nuances of interactions. With increasingly accurate and accessible structure prediction of targets, we developed a novel deep learning model, named S2DTA, to accurately predict DTA by fusing sequence and structural knowledge of drugs, targets, and pockets using heterogeneous models based on graph and semantic networks. Experimental findings underscored that complex feature representations imparted negligible enhancements to the model’s performance. However, the integration of heterogeneous models demonstrably bolstered predictive accuracy. In comparison to three state-of-the-art methodologies, the supremacy of S2DTA became strikingly apparent. It showcased a noteworthy 25.2% reduction in Mean Absolute Error (MAE) and an impressive 20.1% decrease in Root Mean Square Error (RMSE). Furthermore, S2DTA exhibited substantial advancements in other pivotal metrics, including Pearson Correlation Coefficient (PCC), Spearman, Concordance Index (CI), and R2. These metrics experienced remarkable increments of at least 19.6%, 17.5%, 8.1%, and a remarkable 49.4%, respectively. Finally, we conducted interpretability analysis on the effectiveness of S2DTA by bidirectional self-attention mechanism, fully proving that S2DTA is a valuable and accurate tool for predicting DTA. For further exploration, the source data and code repository can be accessed at https://github.com/dldxzx/S2DTA.