Sparsity-Penalized Stacked Denoising Autoencoders for Imputing Single-Cell RNA-Seq Data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for transcriptome analysis but is hampered by excessive zero counts, including technical zeros.
• These technical zeros represent missing data and impede accurate downstream analyses, necessitating robust imputation methods.

Purpose of the Study:

• To develop and evaluate a novel deep learning-based method, sparsity-penalized stacked denoising autoencoders (scSDAE), for imputing technical zeros in scRNA-seq data.
• To assess the efficacy of scSDAE in recovering true gene expression values and enhancing the performance of downstream analytical tasks.

Main Methods:

• Implementation of stacked denoising autoencoders with a sparsity penalty and a layer-wise pretraining strategy.
• Application of scSDAE to simulated bulk sequencing data with noise, RNA mixture datasets with varying dilutions, and CITE-seq data.

Main Results:

• scSDAE demonstrated effectiveness in recovering true gene expression values and sample-sample correlations in simulated noisy data.
• The method accurately imputed technical zeros in RNA mixture datasets and improved the consistency between RNA and protein levels in CITE-seq data.
• scSDAE imputation improved the results of downstream clustering analysis.

Conclusions:

• scSDAE is a powerful deep learning tool for imputing technical zeros in scRNA-seq data.
• The method successfully recovers true expression values and significantly benefits downstream analyses, including clustering.