DHF-DTI: a dual-modality hybrid fusion framework for drug-target interaction prediction

Abstract Background Efficient computational methods for predicting drug–target interactions (DTI) are crucial for drug discovery, shortening the development cycle, and reducing costs. Although deep learning has achieved remarkable progress in this field, existing methods often struggle to fully exploit complementary information between different modalities when handling multimodal data. Moreover, relying solely on classification loss makes it difficult to learn highly discriminative feature representations in the latent space, which severely limits the robustness and domain generalization ability of the model. Results To address these issues, this study proposes DHF-DTI, a dual-modality neural network model that integrates sequence and graph structure information of drugs and proteins for DTI prediction. To achieve deep information interaction between modalities, this paper designs the Cross-Attention and Bilinear Fusion Module (CABFusion), which enhances fine-grained fusion of local features while capturing global contextual dependencies. To further improve generalization and inter-class separability, supervised contrastive learning is introduced to optimize the feature space distribution. Experiments in intra-domain, cross-domain, and cold-start scenarios demonstrate that DHF-DTI significantly outperforms state-of-the-art methods in prediction performance and generalization. Conclusions Case studies, drug feature visualization, and error analysis are also conducted in this paper to fully validate the application potential of the model in practical drug discovery. Overall, DHF-DTI effectively overcomes the challenges of insufficient dual-modality feature fusion and limited discriminability, paving the way for constructing a more efficient and accurate drug discovery pipeline.