FKAN-MoCBDTI: Multimodal drug-target interaction prediction via collaborative bilinear attention and knowledge...

Drug-target interaction (DTI) prediction plays a pivotal role in drug discovery and drug repurposing, yet current deep learning approaches still suffer from limitations in multimodal fusion, fine-grained cross-modal interaction modeling, nonlinear representation learning, and inductive generalization. To overcome these issues, we develop FKAN-MoCBDTI, a multimodal framework that unifies heterogeneous feature extraction, collaborative bilinear attention, Fourier Kolmogorov-Arnold Networks (FKAN), teacher-student knowledge distillation and conditional domain adversarial learning (CDAN). On the drug side, SMILES sequences are encoded by a Transformer and molecular topology is modeled by an FKAN-based graph neural network. On the target side, M-gram embeddings are combined with multi-layer graph convolutional networks to capture complementary sequential and sequence-derived topological information. A collaborative bilinear attention module is designed to explicitly characterize local pairwise associations between drug molecular and protein residue features, improving both feature fusion and qualitative attention visualization, while FKAN is used as the predictor to enhance high-order nonlinear modeling. Furthermore, we integrate a teacher-student KD mechanism with a CDAN module to effectively bridge domain discrepancies under cross-domain scenarios. Comprehensive evaluations across six benchmark datasets demonstrate the competitive performance of FKAN-MoCBDTI. Specifically, standard pair-wise random splitting validates in-domain pair interpolation, while evaluations across cold-entity settings (cold-drug and cold-both) and clustering-based cross-domain transfers demonstrate its strong attribute-based inductive generalization on unseen chemical compounds and novel interaction pairs, alongside nuanced performance trade-offs under unseen target protein (cold-target) scenarios. Ablation studies, visualization analyses, and retrospective case studies further confirm the effectiveness of the proposed framework, providing structural and literature-based support for the plausibility of selected predictions.Overall, FKAN-MoCBDTI offers a robust and generalizable approach for DTI prediction.