PGCNMDA: Learning node representations along paths with graph convolutional network for predicting miRNA-disease...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Identifying microRNA-disease associations (MDAs) is vital for disease diagnosis and treatment.
• Experimental methods for MDA identification are costly and time-consuming.
• Computational approaches, particularly Graph Convolutional Networks (GCNs), show promise for MDA prediction.

Purpose of the Study:

• To propose a novel computational method, PGCNMDA, for enhanced inference of miRNA-disease associations.
• To leverage GCNs with a learned spatial operator from paths for improved MDA prediction.
• To validate the effectiveness and feasibility of PGCNMDA in practical applications.

Main Methods:

• Developed PGCNMDA, a method employing graph convolutional networks (GCNs).
• Incorporated a learned graph spatial operator derived from paths within the GCN framework.
• Evaluated PGCNMDA performance using 5-fold cross-validation (5-CV), 10-fold cross-validation (10-CV), and global leave-one-out cross-validation (GLOOCV) on HMDD v2.0 and HMDD v3.2 datasets.

Main Results:

• PGCNMDA achieved high performance, with AUCs around 0.923 and AUPRCs around 0.921 on HMDD v2.0.
• On HMDD v3.2, PGCNMDA demonstrated superior performance with AUCs around 0.941 and AUPRCs around 0.942.
• Case studies confirmed a high percentage (up to 50/50) of top predicted miRNA-disease links for pancreatic neoplasms, thyroid neoplasms, and leukemia.

Conclusions:

• PGCNMDA significantly outperforms existing methods in predicting miRNA-disease associations.
• The novel approach of learning spatial operators from paths enhances GCN performance for MDA inference.
• PGCNMDA shows strong potential for practical applications in disease diagnosis and treatment development.