scGGAN: single-cell RNA-seq imputation by graph-based generative adversarial network.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) data frequently contain numerous missing values, hindering critical gene signaling information and limiting downstream analyses.
• Existing deep learning imputation methods often overlook inherent gene-to-gene relationships, which are crucial as gene expression is regulated by other genes.

Purpose of the Study:

• To develop a novel imputation model, scGGAN, that addresses missing values in scRNA-seq data by incorporating regional gene-to-gene relationships.
• To enhance the accuracy and biological relevance of scRNA-seq data imputation for improved downstream applications.

Main Methods:

• scGGAN utilizes Graph Convolutional Networks (GCN) to learn gene-to-gene relations and Generative Adversarial Networks (GAN) to model global scRNA-seq data distribution.
• The model integrates scRNA-seq data, gene sequencing data, and a gene relation network, built using both single-cell and bulk genomics data, for imputation through adversarial learning.

Main Results:

• scGGAN effectively identifies dropout events and recovers biologically meaningful gene expressions in simulated and real scRNA-seq datasets.
• The imputation method improves the determination of subcellular states and types, enhances differential expression analysis, and aids temporal dynamics analysis.

Conclusions:

• The proposed scGGAN model demonstrates superior performance in imputing scRNA-seq data by effectively leveraging gene-to-gene networks and adversarial learning.
• Both the gene relation network and gene sequence data significantly contribute to the imputation accuracy and biological interpretability of scRNA-seq data.