Subgraph-Aware Dynamic Attention Network for Drug Repositioning

Drug repositioning, recognized for its potential to identify new applications for existing medications, faces challenges in deeply unraveling the complex, dynamic interactions between drugs and diseases through traditional computational methods. Addressing this challenge, we introduce an innovative model, termed SubDR, meticulously crafted to accurately predict the potential associations between drugs and diseases. SubDR leverages known drug-disease pairs to construct a heterogeneous network enriched with relational information and dynamically identifies subgraph structures directly relevant to the drugs using neighborhood information. By utilizing the advanced dynamic attention mechanism, it not only efficiently highlights interaction features crucial for prediction but also flexibly adjusts the weight distribution among nodes within the network, thus accurately capturing the intricate interplay between drugs and diseases. Furthermore, leveraging the hierarchical structure of the DiffPool technique, SubDR adeptly encodes and aggregates the network structure across multiple levels of abstraction, finely discerning subtle differences between nodes. The final association predictions are generated using a multilayer perceptron. The evaluation conducted via 10-fold cross-validation on three extensively utilized benchmark datasets demonstrates that the SubDR model significantly outperforms existing leading technologies across multiple key metrics, showcasing its robust application potential and significant advantages in the field of drug repositioning. Moreover, through detailed case studies and visualization of attention mechanisms, this study further confirms the practicality of the SubDR model and its high interpretability in revealing the biological mechanisms of drug action.