Normalization

The simplest approach scales each cell's total counts to a common value, such as counts per 10,000 (CP10K). It assumes that differences in total RNA content between cells are purely technical. While computationally efficient, it fails when cells genuinely differ in size or transcriptional activity. This method is often used as a quick initial step before applying more sophisticated normalization.

• Formula: ( X_{norm} = \frac{X_{raw}}{\sum X_{raw}} \times 10,000 )
• Assumption: All cells contain equal total mRNA
• Vulnerability: Distorted by a few highly expressed genes dominating the library

After library size scaling, a log1p transformation (log(1 + x)) is applied to reduce the skewness of the expression distribution and mitigate the influence of highly expressed genes. This transformation makes the data more normally distributed, satisfying assumptions of many downstream statistical tests. The pseudo-count of 1 prevents taking the log of zero.

• Transformation: ( Y = \ln(1 + X_{norm}) )
• Effect: Compresses extreme values and reduces heteroscedasticity
• Limitation: Introduces a non-linear relationship between variance and mean for low counts

This method forces the distribution of expression values to be identical across all cells by matching quantiles. It ranks genes within each cell, then replaces values with the mean of genes at the same rank across all cells. While powerful for removing technical variation, it assumes global expression differences are purely technical, which can erase genuine biological signals.

• Mechanism: Sorts and averages expression at each rank position
• Assumption: No global biological differences exist between cells
• Risk: Can over-normalize and remove true biological variation

Non-biological systematic variation introduced by technical factors—different experimental runs, reagents, or sequencing lanes—that confounds biological signal. Normalization within a single batch cannot correct for batch effects; this requires data integration algorithms like Harmony or scVI. Failure to address batch effects leads to clustering artifacts where cells group by technical origin rather than biological identity.

A sparse numerical matrix where rows represent genes and columns represent cell barcodes, storing the number of transcripts detected per gene in each cell. Normalization transforms this raw count matrix into a scaled representation suitable for comparison. The matrix is typically stored in formats like AnnData (Python) or Seurat objects (R), with normalization results saved as additional layers or slots alongside the raw counts.

A random barcode sequence incorporated during library preparation to tag individual transcripts, enabling absolute molecular counting and PCR duplicate removal. UMI counts form the basis of modern normalization methods. Unlike read counts, UMI counts represent absolute molecule counts and follow a negative binomial distribution, which informs the choice of normalization and variance-stabilizing transformations like sctransform or scran.

Genes exhibiting greater expression variance across cells than expected by technical noise, selected as informative features for downstream dimensionality reduction. Normalization is a critical prerequisite for HVG selection because unscaled data conflates biological variation with sequencing depth differences. Methods like vst in Seurat model the mean-variance relationship on normalized data to identify truly variable genes.

Computational alignment of multiple single-cell datasets to remove batch effects while preserving true biological variation. Normalization is performed within each batch before integration, but cross-batch normalization alone is insufficient. Integration methods like Harmony, scVI, and MNN correct for systematic shifts in the latent space, enabling joint analysis of cells from different experiments, donors, or technologies.

Statistical comparison of gene expression levels between cell groups to identify significantly up- or down-regulated transcripts. Normalization directly impacts differential expression results—library-size normalization assumes most genes are not differentially expressed, while more sophisticated methods like DESeq2 or edgeR use internal robust scaling. For single-cell data, pseudobulk aggregation after normalization provides the most robust statistical framework.