CollDTI: Dual-encoder collaborative learning for drug-target interaction prediction.

Abstract

Accurate identification of Drug-Target Interactions (DTIs) is essential for drug discovery, side effect evaluation, and drug repositioning. Although deep learning-based approaches have shown notable progress, the effective integration of intrinsic structural features with external relational features remains an open challenge. In this work, we propose CollDTI, a dual-encoder Collaborative learning framework, which improves high-level semantic integration by jointly modeling both intrinsic and relational features of drug-target pairs. Specifically, a multi-view heterogeneous graph neural network is first constructed to capture external relational features through a dual-view encoder with hybrid learning. Then, intrinsic structural features are extracted by combining character dictionary encoding with a graph convolutional network and a skip-connection strategy. Finally, a weighted residual cross-attention mechanism further enables collaborative fusion of multi-view embeddings to uncover fine-grained dependencies between drugs and targets. Extensive experiments demonstrate that CollDTI outperforms baseline models, while visualizations and case studies further confirm that its integration of structural and relational features improves the reliable identification of potential drug-target interactions.