Dual-stream decoupled hypergraph learning for compound-target interaction prediction in traditional Chinese medicine

Traditional Chinese medicine (TCM) is a complex therapeutic system characterized by interactions among multiple herbal components. However, its multi-component and multi-target nature leads to highly complex biological relationships, making it difficult to systematically understand the underlying mechanisms and limiting its broader application in precision medicine. To address this issue, we propose DSDHG, a dual-stream decoupled hypergraph learning framework for therapeutic target identification of natural compounds in TCM. A four-layer heterogeneous biomedical network is constructed by integrating herbs, natural compounds, target proteins, and diseases. Global structural representations are initialized through random walk-based methods and Skip-Gram embedding. To capture higher-order biological associations, compound-herb and target-disease interactions are modeled as hypergraphs. Based on these structures, DSDHG adopts a dual-stream learning strategy to separate semantic and topological information. The semantic stream employs relation-aware global attention to capture long-range biological dependencies, while the topological stream uses linear propagation to preserve structural connectivity and alleviate over-smoothing. In addition, a cross-correlation decorrelation constraint is introduced to support disentangled representation learning, enabling latent factors to encourage representation diversity and reduce feature redundancy. Experiments on three benchmark datasets demonstrate that DSDHG outperforms existing state-of-the-art methods across multiple evaluation metrics. Case studies further show that the proposed framework can effectively identify potential therapeutic targets of natural compounds and facilitate target discovery in TCM. The code is publicly available at https://github.com/27167199/DSDHG.