Deep feature extraction of single-cell transcriptomes by generative adversarial network.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables detailed cellular analysis but is limited by batch effects.
• Technical variations obscure true biological signals in cross-condition scRNA-seq studies.

Purpose of the Study:

• To introduce a novel computational method, single-cell Generative Adversarial Network (scGAN), for robust scRNA-seq data analysis.
• To mitigate batch effects and improve the accuracy of cell-type identification and gene expression analysis.

Main Methods:

• Developed scGAN, a deep learning model utilizing Generative Adversarial Networks for scRNA-seq data.
• scGAN projects cells into a latent space, modeling data likelihood while minimizing batch label correlations.
• Applied scGAN to three public scRNA-seq datasets for validation.

Main Results:

• scGAN effectively removes batch effects from scRNA-seq data.
• The method demonstrates superior performance in clustering known cell types compared to existing approaches.
• scGAN successfully identified known psychiatric genes linked to major depressive disorder.

Conclusions:

• scGAN offers a powerful tool for analyzing scRNA-seq data by addressing technical artifacts.
• This approach enhances the discovery of cell-type-specific gene expression patterns and disease-associated genes.
• The developed scGAN method and code are publicly available for broader research application.