KansFormer network based on a multimodal cross-attention mechanism for drug-target interaction prediction.

The prediction of drug-target interactions (DTIs) is a critical step in drug development and repositioning. Traditional wet lab experiments are constrained by high temporal and economic costs, which impede research and development efficiency. Deep learning models based on chemical genomics can significantly reduce R&D expenditures. However, existing multimodal models face challenges such as inadequate modality synergy and difficulties in precisely capturing highly non-linear fused features, thereby compromising DTI prediction accuracy. To address these limitations, we propose KansFormer, a multimodal cross-attention mechanism-based network for drug-target interaction prediction (KMCA-DTI). In KMCA-DTI, a fusion contact map is constructed using residue token features, contact map features, and category scores, which provides more informative protein structural representation. In addition, a multilayer cross-attention mechanism is employed to strengthen the interaction learning between multimodal drug and protein features, thereby improving modality synergy. Furthermore, a KAN-based prediction module replaces the conventional MLP layer, which enhances the ability of the model to capture high-dimensional nonlinear fused features. Evaluations across multiple public datasets demonstrate that KMCA-DTI outperforms current state-of-the-art models. Case studies further validate the reliability of KMCA-DTI in practical applications.