Mamba-Sinter: Enhancing Drug-Target Binding Affinity Prediction via Multimodal Feature Synergy and Interactive Attention

Accurate Drug-Target Affinity (DTA) prediction is crucial for accelerating drug discovery. To overcome the limitations of current models in multimodal integration, interaction modeling, and long-sequence handling, we propose Mamba-Sinter (Mamba-Synergistic Interaction Network), a novel deep learning framework. It features a hybrid encoder that synergistically processes drug topology (GNN) and sequence (Mamba). For targets, it incorporates biological priors using pre-trained ESM-2 embeddings, which are then contextually refined by Mamba. The framework’s core is a Symmetric Cross-Attention Module for fine-grained, bidirectional interaction modeling at the atomic level. An Attention-Weighted Pooling (AG-W Pooling) module then uses interaction-derived attention weights to focus the final representations on key binding substructures. Experiments on the Davis, KIBA, and Metz benchmarks show Mamba-Sinter significantly surpasses state-of-the-art models. Ablation studies and visualizations further validate the synergy of its components and its interpretability.