Feature selection followed by a novel residuals-based normalization that includes variance stabilization simplifies...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Normalization is critical for single-cell RNA sequencing (scRNA-seq) data analysis, aiming to reduce technical biases and transform counts for statistical analysis.
• Current workflows typically perform feature selection after normalization to identify highly variable genes (HVGs).

Purpose of the Study:

• To propose and validate a revised scRNA-seq analysis workflow where feature selection precedes normalization.
• To introduce a novel method for identifying stable genes and a variance stabilization transformation inclusive residuals-based normalization method.

Main Methods:

• A new feature selection method is proposed, applied to observed counts before normalization.
• A novel normalization method is introduced, incorporating variance stabilization and residuals-based bias reduction, utilizing identified stable genes.
• The proposed workflow was implemented in an R package named Piccolo.

Main Results:

• The revised workflow demonstrated significant improvements in downstream clustering analyses.
• The method successfully identified both highly variable genes and stable genes.
• The novel normalization approach effectively reduced systematic biases by leveraging stable genes.

Conclusions:

• Performing feature selection before normalization offers a more effective approach for scRNA-seq data analysis.
• The proposed Piccolo R package provides a robust implementation of this novel workflow.
• This revised strategy enhances the accuracy and interpretability of scRNA-seq data, particularly for clustering tasks.