Noise regularization removes correlation artifacts in single-cell RNA-seq data preprocessing.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-dimensional data with technical noise.
• Existing scRNA-seq preprocessing methods aim to correct expression levels but may impact gene-gene association inference.
• Systematic evaluation of these methods' effects on gene network reconstruction is lacking.

Purpose of the Study:

• To benchmark the impact of five representative scRNA-seq normalization/imputation methods on gene-gene association inference.
• To develop a method for mitigating correlation artifacts introduced during scRNA-seq data preprocessing.
• To reconstruct gene co-expression networks using noise-regularized associations.

Main Methods:

• Benchmarking of five scRNA-seq normalization/imputation methods using Human Cell Atlas bone marrow data.
• Evaluation of method-induced changes in gene-gene correlations.
• Development and application of a model-agnostic noise-regularization technique.
• Gene co-expression network reconstruction using regularized correlations.

Main Results:

• Preprocessing steps, particularly oversmoothing, introduced significant spurious correlations in gene-gene associations.
• The proposed noise-regularization method effectively eliminated these correlation artifacts.
• Reconstructed gene co-expression networks using noise-regularized data successfully identified known immune cell modules.

Conclusions:

• Standard scRNA-seq preprocessing methods can distort gene-gene association inference, impacting network analysis.
• Noise regularization is a crucial step for accurate gene network reconstruction from scRNA-seq data.
• The developed method enhances the reliability of inferring gene interactions and biological modules from single-cell data.