AFExNet

AFExNet leverages adversarial autoencoders (AAE) to differentiate breast cancer sub-types and extract biologically relevant genes from high-dimensional RNA-Seq data.

Key Features:

• Adversarial Autoencoder Architecture: Employs a dual-stage AAE architecture combining unsupervised pre-training with supervised fine-tuning to learn latent representations from high-dimensional genomic data.
• Classifier Independence: Demonstrated robust performance when evaluated using twelve different supervised classifiers on public RNA-Seq datasets of breast cancer.
• Feature Extraction Capability: Extracts meaningful latent features from genetic data that support characterization of breast cancer sub-types.
• TopGene Method: Identifies highly weighted genes from the model latent space as candidate biomarkers for downstream analysis.

Scientific Applications:

• Breast Cancer Sub-type Classification: Differentiates breast cancer sub-types using RNA-Seq–derived latent representations.
• Biomarker Discovery: Identifies highly weighted genes from the latent space that serve as potential diagnostic or therapeutic biomarkers.
• Mechanistic Characterization: Provides molecular feature sets that can inform studies of disease mechanisms and heterogeneity in oncology.
• Translational Research Framework: Offers a framework applicable to other complex diseases with genetic heterogeneity for feature extraction from high-throughput genomic data.

Methodology:

Two-stage approach comprising unsupervised pre-training with an adversarial autoencoder to learn a latent representation, followed by supervised fine-tuning using labeled data.