DBG-DDA: A Dual-Branch Graph Learning Method for Drug−Disease Associations Prediction

Drug repositioning allows for the identification of potential alternative therapeutic uses for existing drugs, thereby accelerating the drug development process. In this field, there are numerous methods that utilize deep learning techniques for Drug−Disease Association (DDA) prediction. However, capturing complex features of biological entities and graph nodes still presents challenges. To tackle these challenges, we proposed a Dual-Branch Graph (DBG) representation learning approach for DDA (DBG-DDA). Specifically, we integrate features obtained from graph learning algorithms based on random walks and convolutions for subsequent classification operations. Firstly, we construct a drug−Protein−Disease Heterogeneous Information Network (HIN). Then, we employ the Meta-path Aggregation HIN method (MAHIN) based on meta-path to extract high-level features from the graph network, followed by utilizing the graph attention network for graph transformer structure algorithm to extract low-level features, and fuse the features by applying nonlinear transformations followed by summation. Finally, the fused features are inputted into a classifier to predict DDAs. A series of experimental results demonstrate that DBG-DDA outperforms several state-of-the-art DDA prediction methods on multiple public databases, providing effective predictions for discovering new drug indications and novel therapeutic approaches for diseases. Overall, the results indicate that DBG-DDA offers a promising direction in drug repositioning. Our code is available at https://github.com/guanzhenghua/DBG-DDA.