Multidependency Graph Convolutional Networks and Contrastive Learning for Drug Repositioning.

Area of Science:

• Pharmacology
• Computational Biology
• Bioinformatics

Background:

• Drug repositioning accelerates therapeutic development by identifying new uses for existing drugs.
• Computational methods, including graph-based approaches, are used for predicting drug-disease relationships.
• Existing graph-based methods often neglect node-specific information and struggle with data noise and sparsity.

Purpose of the Study:

• To develop a novel computational drug repositioning method to improve drug-disease association prediction.
• To address the limitations of existing graph-based methods, such as ignoring semantic influences and data sparsity.

Main Methods:

• Proposed MDGCN, a method integrating multidependency graph convolutional networks and contrastive learning.
• Constructed multidependency graphs using drug/disease similarity and drug-disease relationship matrices.
• Employed graph convolutional networks for information propagation and contrastive learning for aligning node embeddings across views and layers.

Main Results:

• MDGCN demonstrated superior performance in drug-disease association prediction compared to seven advanced methods.
• The method effectively handles weak supervision in drug-disease connections through contrastive learning strategies.
• Experimental results validate the efficacy of MDGCN in identifying potential therapeutic indications.

Conclusions:

• MDGCN offers a robust approach for computational drug repositioning.
• The method improves the accuracy of predicting drug-disease associations.
• MDGCN provides valuable support for discovering novel therapeutic applications for drugs.