scIGANs: single-cell RNA-seq imputation using generative adversarial networks.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) provides high-throughput transcriptomic data at single-cell resolution.
• scRNA-seq data is affected by technical noise, particularly dropouts (excess false zeros), which obscure true biological signals.
• Existing computational imputation methods can lead to oversmoothing and loss of natural gene expression stochasticity.

Purpose of the Study:

• To introduce scIGANs, a generative adversarial network (GAN)-based imputation method for scRNA-seq data.
• To address limitations of current imputation techniques, such as oversmoothing and removal of biological variation.
• To develop an imputation method that balances performance across major and rare cell populations.

Main Methods:

• Development of scIGANs, a novel imputation method utilizing generative adversarial networks (GANs).
• scIGANs employs generated cells instead of observed cells to mitigate oversmoothing and preserve cell-to-cell variability.
• Evaluation using diverse simulated and real-world scRNA-seq datasets.

Main Results:

• scIGANs demonstrates effectiveness in imputing dropout events in scRNA-seq data.
• The method enhances the performance of various downstream analyses.
• scIGANs shows robustness on small datasets and scalability to datasets exceeding 100,000 cells.
• The method performs consistently across different scRNA-seq protocols.

Conclusions:

• scIGANs represents a competitive and effective imputation method for scRNA-seq data.
• The application of GANs in omics data imputation is validated through scIGANs.
• scIGANs improves the accuracy and utility of scRNA-seq data for biological discovery.