GAEMDA

GAEMDA predicts potential associations between microRNAs (miRNAs) and diseases using a graph auto-encoder model to identify disease-related miRNAs.

Key Features:

• Graph neural network encoder: Leverages graph neural networks with an aggregator function and a multi-layer perceptron to gather neighborhood information and generate low-dimensional embeddings for miRNA and disease nodes.
• Heterogeneous information fusion: Integrates diverse data types to fuse heterogeneous information and enhance predictive capability.
• Bilinear decoder: Applies a bilinear decoder to embeddings to infer potential links between miRNA and disease nodes.
• End-to-end prediction: Operates end-to-end from input data to association prediction without intermediate manual steps.
• Performance: Achieves an average area under the curve (AUC) of 93.56 ± 0.44% in a 5-fold cross-validation setup.

Scientific Applications:

• miRNA–disease association prediction: Predicts candidate miRNA associations with diseases to support studies of miRNA regulatory roles in human complex diseases.
• Experimental prioritization: Ranks predicted miRNA–disease associations to guide and prioritize biological validation experiments.
• Case-study validation: Demonstrated validation in case studies on colon neoplasms, esophageal neoplasms, and kidney neoplasms, with 48 of the top 50 predicted miRNAs confirmed by the database of differentially expressed miRNAs in human cancers and the microRNA deregulation in human disease database.

Methodology:

Uses a graph neural network encoder with an aggregator function and a multi-layer perceptron to produce low-dimensional miRNA and disease embeddings, fuses heterogeneous information, decodes links with a bilinear decoder, and evaluates performance via 5-fold cross-validation (AUC 93.56 ± 0.44%).