Background <p>Predicting drug-target interactions (DTIs) is essential for accelerating drug discovery, yet traditional experimental methods are time-consuming and costly. Computational approaches, especially those using machine learning and deep learning, offer a more efficient alternative.</p> Results <p>This paper presents the BiMA-DTI framework, which integrates Mamba’s State Space Model (SSM) with multi-head attention mechanisms. This combination maximizes Mamba’s ability to process long sequences while taking advantage of the attention mechanism’s strength in handling short sequences. We designed a hybrid Mamba-Attention Network (MAN) and a Graph Mamba Network (GMN) for processing multimodal inputs, including protein amino acid sequences, Simplified Molecular Input Line Entry System (SMILES) strings, and molecular graphs of drugs, enabling comprehensive feature extraction and fusion. To enhance the complementarity between features extracted from sequences and graphs, BiMA-DTI performs a two-step weighted fusion of sequence features of drugs and proteins with molecular graph features of drugs. Finally, these fused features are concatenated and passed through a fully connected network to predict DTIs.</p> Conclusions <p>BiMA-DTI demonstrates its potential to discover new drugs, offering a powerful tool for drug discovery. Experimental results show that BiMA-DTI outperforms state-of-the-art competing methods on benchmark datasets. Additionally, ablation experiments validate the rationality of BiMA-DTI’s architecture and its generalization ability. Visualization studies provide interpretability of biological mechanisms. Finally, case studies further confirm that BiMA-DTI is a reliable drug-target interaction prediction tool.</p>