Random matrix theory-guided denoising of single-cell RNA-seq data

Reproducible analysis of single-cell RNA-seq data is challenging due to various sources of variability, including biological noise, PCR bias, limited sequencing depth, and low RNA capture rates. Existing denoising methods often rely on specific noise models, limiting their adaptability to changing technologies or data integration. Consequently, noise is frequently managed in a case-specific manner, and principal component analysis (PCA) remains the primary dimensionality reduction technique. In our study, we leverage the high dimensionality of single-cell RNA-seq data to develop a parameter-free sparse PCA method, guided by random matrix theory. We introduce a novel biwhitening procedure that simultaneously stabilizes noise variance across cells and genes, ensuring our method's versatility with changing technologies and datasets integration methods. We show that our approach enhances principal subspace reconstruction accuracy while minimizing data distortion.