bioconductor-transformgampoi

name	bioconductor-transformgampoi
description	Variance-stabilizing transformations help with the analysis of heteroskedastic data (i.e., data where the variance is not constant, like count data). This package provide two types of variance stabilizing transformations: (1) methods based on the delta method (e.g., 'acosh', 'log(x+1)'), (2) model residual based (Pearson and randomized quantile residuals).

transformGamPoi

Workflows

Standard Workflow

Apply delta method-based transformations (acosh or shifted logarithm) to stabilize the variance of count data.

library(transformGamPoi)
library(SingleCellExperiment)

# Load the data
sce <- TENxPBMCData::TENxPBMCData("pbmc4k")
sce <- sce[sample(nrow(sce), 1000), sample(ncol(sce), 500)]

# Apply the acosh transformation to the count matrix
assay(sce, "acosh") <- acosh_transform(assay(sce, "counts"))

# Alternatively, apply the shifted log transformation with a specified pseudo-count
y_shiftLog <- shifted_log_transform(assay(sce, "counts"), pseudo_count = 1/(4 * 0.1))

Inputs are a SingleCellExperiment or count matrix; outputs are variance-stabilized values added as an assay or returned as a matrix.

Model Residuals Transformation

Apply model residuals-based transformations (Pearson or randomized quantile residuals) to stabilize variance across all genes, including lowly expressed ones.

library(transformGamPoi)
library(SingleCellExperiment)

# Load the data
sce <- TENxPBMCData::TENxPBMCData("pbmc4k")
sce <- sce[sample(nrow(sce), 1000), sample(ncol(sce), 500)]

# Apply the Pearson residuals transformation with clipping
assay(sce, "pearson") <- residual_transform(sce, "pearson", clipping = TRUE, on_disk = FALSE)

# Alternatively, apply the randomized quantile residuals transformation
assay(sce, "rand_quantile") <- residual_transform(sce, "randomized_quantile", on_disk = FALSE

Inputs are a SingleCellExperiment or count matrix; outputs are Pearson or randomized quantile residuals.

When to Use

• When analyzing heteroskedastic count data (e.g., single-cell RNA-seq counts) where variance increases with mean expression.
• To prepare count data for classical statistical methods (like PCA or clustering) that perform best on data with uniform variance.
• When you want to stabilize variance for lowly expressed genes using model residuals (residual_transform).

When NOT to Use

• Do not use on already normalized or log-transformed values; this package requires raw, heteroskedastic count data.
• For differential expression testing directly; use glmGamPoi or DESeq2 which model the count distribution directly.

Data Requirements

• Raw count matrix-like objects (e.g., matrix, dgCMatrix, DelayedArray, SummarizedExperiment, SingleCellExperiment).
• Sparsity is preserved for sparse inputs where possible.
• Typically filtered to exclude genes where all counts are zero (e.g., using rowMeans2(counts(sce)) > 0).

Key Parameters

• overdispersion (0.1): The Gamma-Poisson overdispersion parameter $\alpha$ used in acosh_transform.
• pseudo_count (1/(4 * overdispersion)): The pseudo-count $c$ used in shifted_log_transform.
• clipping (TRUE): Whether to clip extremely large Pearson residuals in residual_transform.
• on_disk (FALSE): Whether to perform calculations on disk for large datasets in residual_transform.

Best Practices

• Exclude genes with zero counts across all cells using rowMeans2 before applying transformations.
• Use acosh_transform or shifted_log_transform to retain sparsity of the input data (ensuring $g(0) = 0$).
• Use randomized quantile residuals (residual_transform with "randomized_quantile") to handle the discrete nature of counts and stabilize variance for lowly expressed genes.

Common Pitfalls

• Loss of sparsity: Choosing an offset in log-transformation that shifts zero counts to non-zero values. Fix: Use shifted_log_transform or acosh_transform which are designed to preserve sparsity ($g(0) = 0$).
• Poor stabilization of lowly expressed genes with delta method: Delta method-based transformations (like acosh) still show increasing variance for low mean expression ($\mu < 0.5$). Fix: Use residual_transform with Pearson or randomized quantile residuals instead.

Alternatives

• scry: For deviance residuals as a feature selection and dimension reduction tool.
• sctransform (Seurat): For regularized negative binomial regression-based normalization.
• scran: For pooling-based size factor normalization.

Citations

• Hafemeister, C. and Satija, R. 2019. "Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression." Genome Biology.

References

• Homepage: bioconductor.org/packages/transformGamPoi
• Vignette: https://bioconductor.org/packages/release/bioc/vignettes/transformGamPoi/inst/doc/transformGamPoi_Quickstart.html