scSemiGAN

scSemiGAN integrates semi-supervised learning and generative adversarial networks to jointly perform cell-type annotation and dimensionality reduction of single-cell RNA-seq (scRNA-seq) data.

Key Features:

• Semi-Supervised Learning: scSemiGAN employs a semi-supervised framework that leverages both labeled and unlabeled scRNA-seq data and available reference datasets for training.
• Generative Adversarial Network (GAN) Architecture: scSemiGAN uses a GAN-based model to simulate the generation process of scRNA-seq data and to learn deep latent representations.
• Unified Annotation and Dimensionality Reduction: scSemiGAN integrates cell-type annotation and dimensionality reduction within a single framework to produce latent representations suitable for labeling and visualization.
• Competitive Performance in Downstream Tasks: scSemiGAN demonstrated competitive or superior performance relative to four state-of-the-art annotation methods on simulated and real-world scRNA-seq datasets across tasks including cell-type annotation, latent visualization, confounding factor removal, and enrichment analysis.

Scientific Applications:

• Cell-Type Annotation: Assigning cell-type labels to single cells based on gene expression profiles using labeled reference data.
• Dimensionality Reduction: Producing lower-dimensional representations of high-dimensional scRNA-seq data for visualization and interpretation.
• Latent Representation Learning: Capturing deep latent structures that reflect underlying biological states or processes in scRNA-seq data.
• Confounding Factor Removal: Identifying and mitigating confounding variables that obscure biological signals in scRNA-seq datasets.
• Enrichment Analysis: Supporting downstream enrichment analyses using annotated cells and learned latent representations.

Methodology:

scSemiGAN models scRNA-seq data from a generative perspective in a semi-supervised setting, using adversarial training between a generator and a discriminator to learn deep latent representations while simultaneously predicting cell-type labels.