*transformGamPoi* provides variance stabilizing transformations to handle the heteroskedasticity of count data. For example single-cell RNA sequencing counts vary more for highly expressed genes than for lowly expressed genes. However, many classical statistical methods perform best on data with uniform variance. This package provides a set of different variance stabilizing transformations to make the subsequent application of generic statistical methods more palatable.

You can install *transformGamPoi* from Bioconductor after it has been accepted using the following command

```
if (!requireNamespace("BiocManager", quietly = TRUE))
install.packages("BiocManager")
BiocManager::install("transformGamPoi")
```

In the mean time or to get the latest development version, you can install *transformGamPoi* directly from GitHub using the devtools package

```
# install.packages("devtools")
devtools::install_github("const-ae/transformGamPoi")
```

The functions in *transformGamPoi* take any kind of matrix-like object (e.g., `matrix`

, `dgCMatrix`

, `DelayedArray`

, `SummarizedExperiment`

, `SingleCellExperiment`

) and return the corresponding transformed matrix objects. For sparse input the functions try to preserve sparsity. For container objects like `SummarizedExperiment`

, *transformGamPoi* extracts the `"counts"`

assay and returns an object of the same type as that `"counts"`

assay.

We start by loading the package to make the transformation functions available in our R session:

`library(transformGamPoi)`

In the next step, we load some example data. Here, we use a single-cell RNA sequencing experiment of 4096 blood cells. For convenience, we subset the data to 1,000 genes and 5,00 cells

```
# Load the 'TENxPBMCData' as a SingleCellExperiment object
sce <- TENxPBMCData::TENxPBMCData("pbmc4k")
#> see ?TENxPBMCData and browseVignettes('TENxPBMCData') for documentation
#> loading from cache
# Subset the data to 1,000 x 500 random genes and cells
sce <- sce[sample(nrow(sce), 1000), sample(ncol(sce), 500)]
```

If we take a look at the stored counts (mainly zeros) stored as a sparse matrix of type `DelayedMatrix`

. Fortunately, the precise meaning of that storage type is not important, because *transformGamPoi* handles this automatically.

```
assay(sce, "counts")[1:10, 1:5]
#> <10 x 5> sparse DelayedMatrix object of type "integer":
#> [,1] [,2] [,3] [,4] [,5]
#> ENSG00000167014 0 0 0 0 0
#> ENSG00000244575 0 0 0 0 0
#> ENSG00000118495 0 0 0 0 0
#> ENSG00000140479 0 0 0 0 0
#> ENSG00000236641 0 0 0 0 0
#> ENSG00000131174 0 1 2 0 0
#> ENSG00000250905 0 0 0 0 0
#> ENSG00000162040 0 0 0 0 0
#> ENSG00000140557 0 0 0 0 0
#> ENSG00000273523 0 0 0 0 0
```

To see what we mean by heteroskedasticity, let us compare the mean and variance for each gene across cells. We will use the *MatrixGenerics* package to calculate the row means and row variances. You might be familiar with the *matrixStats* package; *MatrixGenerics* provides the same set of functions but depending on the type of the matrix automatically dispatches the call to either *matrixStats*, *DelayedMatrixStats*, or *sparseMatrixStats*.

```
library(MatrixGenerics)
# Exclude genes where all counts are zero
sce <- sce[rowMeans2(counts(sce)) > 0, ]
gene_means <- rowMeans2(counts(sce))
gene_var <- rowVars(counts(sce))
plot(gene_means, gene_var, log = "xy", main = "Log-log scatter plot of mean vs variance")
abline(a = 0, b = 1)
sorted_means <- sort(gene_means)
lines(sorted_means, sorted_means + 0.2 * sorted_means^2, col = "purple")
```

The purple line shows a quadratic mean-variance relation (\(\text{Var} = \mu + 0.2 \mu^2\)) typical for data that is Gamma-Poisson distributed. For example a gene with a mean expression of 5 the corresponding variance is 10, whereas for a gene with a mean expression of 500 the variance ~50,000. Here we used an overdispersion of \(\alpha = 0.2\), *transformGamPoi* provides options to either fit \(\alpha\) on the data or fix it to a user-defined value.

*transformGamPoi* implements two approaches for variance stabilization: (1) based on the delta method, (2) based on model residuals.

The delta method relates the standard deviation of a transformed random variable \(g(X_i)\) to the standard deviation of the original random variable \(X_i\). This can be used to find a function such that \(g(X_i) = \text{const}\). For a quadratic mean variance relation this function is \[ g(x) = \frac{1}{\sqrt{\alpha}} \text{acosh}(2 \alpha x + 1). \]

We can apply this transformation to the counts:

```
assay(sce, "acosh") <- acosh_transform(assay(sce, "counts"))
# Equivalent to 'assay(sce, "acosh") <- acosh_transform(sce)'
```

We plot the variance of the `acosh`

transformed counts and find that for \(\mu < 0.5\) the variance still increases for higher average gene expression. However, for larger expression values the variance for a gene is approximately independent of the corresponding average gene expression (note that the y-axis is not log transformed anymore!).

```
acosh_var <- rowVars(assay(sce, "acosh"))
plot(gene_means, acosh_var, log = "x", main = "Log expression vs variance of acosh stabilized values")
abline(h = 1)
```

The most popular transformation for single cell data is \(g(x) = \log(x + c)\) with pseudo-count \(c=1\). It turns out that this transformation is closely related to the `acosh`

transformation. When we choose \(c = 1/(4\alpha)\) the two converge rapidly, only for small values the `acosh`

is closer to \(g(x) = 2\sqrt{x}\):

```
x <- seq(0, 30, length.out = 1000)
y_acosh <- acosh_transform(x, overdispersion = 0.1)
y_shiftLog <- shifted_log_transform(x, pseudo_count = 1/(4 * 0.1))
y_sqrt <- 2 * sqrt(x) # Identical to acosh_transform(x, overdispersion = 0)
```

The plot looks as follows:

```
plot(x, y_acosh, type = "l", col = "black", lwd = 3, ylab = "g(x)", ylim = c(0, 10))
lines(x, y_shiftLog, col = "red", lwd = 3)
lines(x, y_sqrt, col = "blue", lwd = 3)
legend("bottomright", legend = c(expression(2*sqrt(x)),
expression(1/sqrt(alpha)~acosh(2*alpha*x+1)),
expression(1/sqrt(alpha)~log(x+1/(4*alpha))+b)),
col = c("blue", "black", "red"), lty = 1, inset = 0.1, lwd = 3)
```

The offset \(b\) for the shifted logarithm has no influence on the variance stabilization. We choose \(b\) such that sparsity of the input is retained (i.e., \(g(0) = 0\)).

An alternative approach for variance stabilization was suggested by Hafemeister and Satija (2019). They used the Pearson residuals from a Gamma-Poisson generalized linear model fit as the variance stabilized values. The advantage of this approach is that the variance is also stabilized for lowly expressed genes unlike the delta method-based transformations:

`assay(sce, "pearson") <- residual_transform(sce, "pearson", clipping = TRUE, on_disk = FALSE)`

```
pearson_var <- rowVars(assay(sce, "pearson"))
plot(gene_means, pearson_var, log = "x", main = "Log expression vs variance of Pearson residuals")
abline(h = 1)
```

Pearson residuals are by definition a linear transformation. This means that for genes with strong expression differences across subgroups they cannot achieve variance stabilization. As an alternative, *transformGamPoi* provides randomized quantile residuals which are non-linear and exploit randomization to work around the discrete nature of counts:

`assay(sce, "rand_quantile") <- residual_transform(sce, "randomized_quantile", on_disk = FALSE)`

```
rand_quant_var <- rowVars(assay(sce, "rand_quantile"))
plot(gene_means, rand_quant_var, log = "x", main = "Log expression vs variance of Randomized Quantile residuals")
abline(h = 1)
```