scSemiGAN: a single-cell semi-supervised annotation and dimensionality reduction framework based on generative...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Cell-type annotation is vital for single-cell RNA sequencing (scRNA-seq) data analysis.
• Automated label transference methods are preferred over manual annotation due to increasing reference datasets.
• Existing methods often lack integration of cell-type annotation with dimensionality reduction and deep latent representation learning.

Purpose of the Study:

• To introduce scSemiGAN, a novel framework for scRNA-seq data analysis.
• To address the limitations of current methods in unifying annotation, dimensionality reduction, and latent representation.
• To model scRNA-seq data from a data generation perspective.

Main Methods:

• Developed scSemiGAN, a semi-supervised framework utilizing a generative adversarial network.
• Integrated deep latent representation learning with cell-type label prediction.
• Modeled scRNA-seq data generation for enhanced analysis.

Main Results:

• scSemiGAN demonstrates competitive or superior performance compared to state-of-the-art methods.
• Achieved high accuracy in cell-type annotation across simulated and real scRNA-seq datasets.
• Successfully performed latent representation visualization, confounding factor removal, and enrichment analysis.

Conclusions:

• scSemiGAN effectively unifies cell-type annotation and dimensionality reduction for scRNA-seq data.
• The framework provides robust deep latent representation learning and accurate cell-type prediction.
• scSemiGAN offers a powerful tool for comprehensive scRNA-seq data analysis.