Bidirectional cross-modal fusion with tensor interaction for drug-target binding prediction

Accurate prediction of drug-target binding affinity is central to computational drug discovery, yet it remains difficult because binding is governed by complex, non-linear interactions between chemical substructures and protein residue environments. Many existing deep learning approaches learn drug and protein representations largely in isolation and combine them only at the final stage, which can weaken their ability to capture informative cross-modal dependencies. To address this limitation, we propose Bidirectional Cross-Modal Fusion with Tensor Interaction (BiT-Fusion), a framework that strengthens interaction modeling through bidirectional fusion and multiplicative coupling between drug and protein features. BiT-Fusion enables more effective information exchange across modalities while preserving complementary signals from molecular graphs and protein sequences. Experiments on the widely used Davis and KIBA benchmarks show that BiT-Fusion delivers competitive and consistent improvements across multiple evaluation metrics. Ablation analyses further verify that bidirectional fusion and tensor-based interaction are the main contributors to the performance gains. Overall, these results suggest that enhancing cross-modal interaction learning is a practical and interpretable direction for improving drug-target binding prediction, with potential relevance to AI-enabled drug discovery, precision medicine, and the United Nations Sustainable Development Goal 3 on good health and well-being.