Interpretable Protein-Nucleic Acid Binding Site Prediction via Domain-Guided Multimodal Learning

Accurate identification of protein-nucleic acid binding sites is essential for elucidating molecular mechanisms and for enabling structure-based drug discovery. Existing computational approaches struggle to jointly model sequence, structure, and domain knowledge, which limits both predictive accuracy and interpretability. We present BiMTBind, a domain-guided multimodal deep learning framework that integrates bidirectional state-space modeling with Transformer-based attention for residue-level binding site prediction. The model explicitly incorporates protein-domain annotations as biological priors to guide attention toward functionally relevant regions. Multi-source featurescontextual embeddings from protein language models and handcrafted structural descriptorsare fused via a learnable gating mechanism that adaptively weights each modality. This design enables the model to capture both global dependencies and local context. Extensive evaluations across diverse DNA and RNA benchmarks show that BiMTBind consistently outperforms state-of-the-art sequence- and structure-based methods on multiple metrics, including F1, MCC, and AUC. Case studies further show that domain-guided attention improves interpretability by aligning predicted binding residues with annotated functional domains. These results highlight the effectiveness of combining biologically informed priors with deep multimodal representation learning, offering a scalable and generalizable approach for protein-nucleic acid binding site prediction.