M2GMDA

M2GMDA predicts novel microRNA (miRNA)-disease associations using multiple meta-paths fusion graph embedding to prioritize candidate miRNA–disease pairs.

Key Features:

• Heterogeneous Network Construction: Constructs a heterogeneous network from verified miRNA–disease associations and miRNA and disease similarity data to support meta-path extraction.
• Meta-Path Extraction: Systematically extracts meta-path instances connecting miRNAs and diseases to capture semantic and structural relationships.
• Graph Embedding Model: Learns representations via linear transformations for miRNAs and diseases, a mean encoder for single meta-path instances, and attention-aware encoders to integrate multiple meta-path types.
• Attention Mechanisms: Applies attention to combine information from meta-path instances, meta-path-based neighbors, and intermediate nodes within meta-paths to weight relevant features.
• Performance Metrics: Reported AUCs of 0.9323 (global leave-one-out cross-validation) and 0.9182 (fivefold cross-validation) on the HDMM V2.0 dataset.

Scientific Applications:

• Neoplasm case studies: Validated in lung, breast, prostate, pancreatic, lymphoma, and colorectal cancers, with 47–50 of the top 50 predicted candidate miRNAs per case confirmed by biological experiments.

Methodology:

Construct a heterogeneous miRNA–disease network from verified associations and similarity data; extract meta-path instances and encode them; apply linear transformations, a mean encoder, and attention-aware encoders within a graph embedding framework, and use attention mechanisms to integrate information from meta-paths, meta-path-based neighbors, and intermediate nodes.