MODIG

MODIG integrates high-throughput multi-omics data and multiplex gene association networks using a graph attention network to identify cancer driver genes.

Key Features:

• Multi-omics inputs: Ingests mutations, copy number variants, gene expression, and methylation levels as multi-omics input types.
• Multi-dimensional gene networks: Constructs gene relationship maps from protein-protein interactions, gene sequence similarity, KEGG pathway co-occurrence, gene co-expression, and Gene Ontology annotations.
• Multiplex network representation: Represents approximately 20,000 genes as nodes interconnected by five distinct types of gene associations acting as multiplex edges.
• Graph Attention Network (GAT): Applies a graph attention network to each network dimension to produce dimension-specific gene representations.
• Joint learning module: Fuses dimension-specific representations into unified general gene representations via a joint learning module.
• Semi-supervised driver identification: Performs semi-supervised learning on the fused representations to identify cancer driver genes.
• Performance evaluation: Assesses effectiveness using area under precision-recall curves (AU PRC) and area under receiver operating characteristic curves (AU ROC) against baseline models.

Scientific Applications:

• Driver gene prioritization: Prioritizes candidate cancer driver genes from integrated multi-omics and network data.
• Comprehensive gene representation: Generates gene representations that combine mutation, CNV, expression, and methylation information across network contexts.
• Network-aware functional inference: Models multiplex gene-gene relationships (PPI, sequence similarity, KEGG co-occurrence, co-expression, GO) for functional and interaction inference.
• Quantitative benchmarking: Enables comparison of driver gene identification performance using AU PRC and AU ROC metrics.

Methodology:

Constructs multi-dimensional gene networks from PPI, gene sequence similarity, KEGG pathway co-occurrence, gene co-expression, and Gene Ontology annotations and integrates them as a five-type multiplex network of ~20,000 gene nodes; applies a graph attention network (GAT) to each dimension to derive dimension-specific gene representations; fuses these representations via a joint learning module to obtain general gene representations; performs semi-supervised learning on the fused representations to identify cancer driver genes and evaluates results using AU PRC and AU ROC.