Multidependency Graph Convolutional Networks and Contrastive Learning for Drug Repositioning.

The goal of drug repositioning is to expedite the drug development process by finding novel therapeutic applications for approved drugs. Using multifeature learning, different computational drug repositioning techniques have recently been introduced to predict possible drug-disease relationships. Nevertheless, current graph-based methods tend to model drug-disease interaction relationships without considering the semantic influence of node-specific side information on graphs. These approaches also suffer from the noise and sparsity inherent in the data. To address these limitations, we propose MDGCN, a novel drug repositioning method that incorporates multidependency graph convolutional networks and contrastive learning. Based on drug and disease similarity matrices and the drug-disease relationships matrix, this approach constructs multidependency graphs. It subsequently employs graph convolutional networks to spread side information between various graphs in each layer. Meanwhile, the weak supervision of drug-disease connections is effectively addressed by introducing cross-view and cross-layer contrastive learning to align node embedding across various views. Extensive experiments show that MDGCN performs better in drug-disease association prediction than seven advanced methods, offering strong support for investigating novel therapeutic indications for medications of interest.