GAMSAP-DTA: Graph-Augmented Multi-View Substructure Attention for Drug-Target Affinity Prediction

Accurate Drug-Target Affinity (DTA) prediction is a cornerstone of computational drug discovery, enabling the identification of high-affinity drug candidates with greater precision. This study introduces GAMSAP-DTA, a novel framework that unifies multi-scale molecular representations with an advanced Multi-View Substructure Attention mechanism to enhance predictive fidelity. Drug molecules are characterized using a dual-representation strategy, where Morgan and Avalon fingerprints capture topological and substructural patterns, while molecular graphs undergo systematic substructure perturbation to enhance robustness and generalizability. Protein representations are formulated through a hybrid approach: amino acid sequences are processed via a Stacked Convolutional Neural Network (Stacked CNN) to extract hierarchical sequence dependencies, while protein graphs are modeled using a Stacked Graph Neural Network (Stacked GNN) to capture molecular motifs spanning atomic interactions to macrocyclic architectures. The heterogeneous representations derived from these modalities are then integrated through the Multi-View Substructure Attention module, enabling an enriched, high-resolution feature space tailored for affinity prediction. Comprehensive evaluations on the Davis, KIBA, and Metz benchmark datasets confirm the proposed framework’s superior performance compared to state-of-the-art models.