Count Matrix Normalization

The simplest approach scales each cell's total counts to a common value, typically 10,000 or 1 million counts per cell.

• Mechanism: Divides each gene count by the cell's total UMI count, then multiplies by a scale factor
• Output: Counts-per-10K (CP10K) or counts-per-million (CPM)
• Limitation: Assumes all cells have identical total RNA content, which is biologically false
• Use case: Quick exploratory analysis before applying more sophisticated methods

This method fails when a few highly expressed genes dominate a cell's transcriptome, artificially suppressing all other gene measurements.

A robust normalization method originally developed for bulk RNA-seq that estimates size factors by comparing each cell to a synthetic reference.

• Mechanism: Computes the median ratio of each gene's count to its geometric mean across all cells
• Key insight: Uses the median to resist outlier gene influence
• Assumption: Most genes are not differentially expressed between conditions
• Output: Size factors applied to raw counts for downstream negative binomial models

While powerful for bulk data, this method struggles with the extreme zero-inflation characteristic of single-cell data where >90% of genes may have zero counts in a given cell.

A single-cell-specific method that models gene expression using a regularized negative binomial regression to account for sequencing depth while stabilizing variance.

• Mechanism: Fits a generalized linear model per gene with UMI count as response and log(total UMI) as predictor
• Regularization: Shares information across genes with similar expression levels to prevent overfitting
• Output: Pearson residuals that represent normalized, variance-stabilized values
• Key advantage: Removes the depth-dependent variance while preserving biological heterogeneity

This method produces values centered around zero, where positive residuals indicate expression above the expected level for that gene given the cell's sequencing depth.

A normalization strategy that computes cell-specific scaling factors by pooling cells with similar library sizes to overcome the sparsity problem.

• Mechanism: Groups cells into pools, sums counts within each pool, then estimates size factors by deconvolving pool-level factors back to individual cells
• Key innovation: Pooling creates pseudo-bulk profiles where zeros are less problematic
• Assumption: Most genes are not differentially expressed across the pooled cells
• Output: Size factors that can be applied to log-normalize the data

This method is particularly effective for datasets with highly variable sequencing depths across cells, as pooling ensures robust factor estimation even for shallowly sequenced cells.

A probabilistic approach using a variational autoencoder that learns a latent representation of gene expression while explicitly modeling technical covariates.

• Mechanism: Encodes each cell's expression profile into a low-dimensional latent space while conditioning on batch, library size, and other technical factors
• Zero-inflation handling: Models counts using a zero-inflated negative binomial distribution
• Key advantage: Simultaneously performs normalization, batch correction, and dimensionality reduction
• Output: Denoised, normalized gene expression values reconstructed from the latent space

Unlike linear methods, scVI captures non-linear dependencies between genes and technical artifacts, producing normalized values that better preserve complex biological structures.

A normalization method that computes size factors based on the median fold change of each gene relative to its geometric mean across all cells.

• Mechanism: For each gene, calculates the ratio of its count in a cell to its geometric mean across all cells; the cell's size factor is the median of these ratios
• Robustness: Median operation provides resistance to differentially expressed genes
• Relationship: Equivalent to the DESeq2 median-of-ratios approach
• Implementation: Available in the scater and scran R packages

RLE normalization assumes that the majority of genes have similar expression across cells, making it suitable for homogeneous populations but potentially problematic in highly heterogeneous single-cell data.