Analytic Pearson residuals for normalization of single-cell RNA-seq UMI data

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Standard single-cell RNA-seq UMI data preprocessing involves normalization and nonlinear transformation.
• Recent studies propose statistical count models like negative binomial regression (Pearson residuals) and generalized PCA.
• This work investigates the theoretical and empirical connections between these statistical modeling approaches.

Purpose of the Study:

• To compare the effects of Pearson residuals and generalized PCA on downstream single-cell RNA-seq data processing.
• To analyze the performance of different statistical count models for UMI data.
• To identify the most effective method for identifying biologically variable genes and capturing meaningful variation.

Main Methods:

• Theoretical analysis of statistical count models for UMI data.
• Empirical investigation of Pearson residuals from negative binomial regression and generalized PCA.
• Benchmarking using single-cell RNA-seq datasets with known ground truth.
• Estimation of technical overdispersion using negative control data.

Main Results:

• The Hafemeister and Satija model, when specified parsimoniously, is equivalent to the Poisson GLM-PCA proposed by Townes et al.
• Per-gene overdispersion estimates in the Hafemeister and Satija method are biased; data suggest overdispersion is independent of gene expression.
• Negative control data indicate UMI counts are close to Poisson with moderate overdispersion across protocols.
• Analytic Pearson residuals demonstrate superior performance in identifying biologically variable genes and capturing meaningful variation.

Conclusions:

• Analytic Pearson residuals are highly effective for identifying biologically variable genes in single-cell RNA-seq data.
• This method captures more biologically meaningful variation compared to other approaches, particularly for dimensionality reduction.
• The findings support the use of analytic Pearson residuals for robust scRNA-seq data analysis.