Subtype-GAN

Subtype-GAN applies deep adversarial learning using multiple-input multiple-output neural networks to integrate multi-omics data (copy number variations, mRNA expression, miRNA, DNA methylation) for molecular cancer subtyping.

Key Features:

• Multi-Omics Integration: Processes copy number variations, mRNA expression profiles, miRNA levels, and DNA methylation to represent tumor molecular landscapes.
• Deep Adversarial Learning: Employs a Generative Adversarial Network (GAN)-based feature extraction module with multiple-input multiple-output neural networks to derive latent variables for each omics type.
• Consensus Clustering and Gaussian Mixture Model: Uses consensus clustering combined with a Gaussian Mixture Model to determine the optimal number of subtypes and assign cluster labels to samples.

Scientific Applications:

• TCGA benchmarking: Evaluated on approximately 4,000 tumors from 10 cancer types sourced from The Cancer Genome Atlas (TCGA), producing clustering results comparable to NEMO, MCCA, and VAE and distinct from AE.
• BRCA subtyping: Applied to 1,031 BRCA tumors where it identified clinically relevant molecular subtypes exhibiting distinct feature-space patterns.

Methodology:

GAN-based feature extraction using multiple-input multiple-output neural networks to obtain latent variables per omics type, followed by consensus clustering and a Gaussian Mixture Model to determine subtype number and assign labels.