Varistran: Anscombe's variance stabilizing transformation for RNA-seq gene expression data


RNA-seq measures RNA expression levels in a biological sample using high-throughput cDNA sequencing, producing counts of the number of reads aligning to each gene. Noise in RNA-seq read count data is commonly modelled as following a negative binomial distribution, where the variance is a quadratic function of the expression level. However many statistical, machine learning, and visualization methods work best when the noise in a data set has equal variance. Varistran is an R package that uses Anscombe's (1948) variance stabilizing transformation for the negative binomial distribution to transform RNA-seq count data, so that the noise has equal variance across all measured gene expression levels. The transformed data may be treated as log2 transformed gene expression levels, but with variability reduced at low read counts. Varistran also includes a function to open a Shiny report with simple diagnostic visualizations, including a plot to assess how effective the variance stabilization was, a biplot of samples and genes, and a heatmap. This allows defective samples, sample mislabling, and batch effects to be easily identified.


Anscombe, Francis J. 1948. “The Transformation of Poisson, Binomial and Negative-Binomial Data.” Biometrika 35 (3/4): 246–54.