GM-GCN

GM-GCN predicts cancer driver genes by integrating gene–miRNA regulatory interactions with a graph convolutional network to capture miRNA-mediated regulatory influence on genes.

Key Features:

• Gene-MiRNA Network Construction: Constructs a network with nodes representing miRNAs and their target genes and edges denoting miRNA→gene regulatory interactions.
• Feature Initialization: Initializes node attributes for miRNAs and genes from their inherent biological properties.
• Graph Convolutional Network (GCN) Model: Uses a GCN to aggregate features from neighboring miRNA nodes and learn gene feature representations.
• 1D Convolution Module: Applies a 1D convolution module to adjust the dimensionality of learned feature representations after aggregation.
• Model Optimization and Prediction: Refines model parameters using optimized gene features and initial input features and employs a logistic regression model to predict driver-gene probabilities.

Scientific Applications:

• Pan-cancer and cancer-type driver prediction: Applied to pan-cancer datasets and individual cancer types to identify candidate driver genes, outperforming methods based solely on gene–gene networks.
• Benchmark performance: Demonstrates high area under the receiver operating characteristic curve (AUC-ROC) and area under the precision-recall curve (AUC-PR) in evaluations.
• Integrative regulatory discovery: Integrates miRNA regulatory information to discover novel driver genes that may be missed by protein–protein interaction or pathway-based methods and to inform personalized medicine and targeted therapy research.

Methodology:

Constructs a gene–miRNA regulatory network, initializes miRNA and gene node attributes, applies a GCN to aggregate neighboring miRNA features into gene representations, uses a 1D convolution module to adjust feature dimensionality, refines model parameters using optimized and initial features, and predicts driver probabilities with logistic regression.