Abstract
The count table, a numeric matrix of genes × cells, is the basic input data structure in the analysis of single-cell RNA-seq data. A common preprocessing step is to adjust the counts for variable sampling efficiency and to transform them so that the variance is similar across the dynamic range. These steps are intended to make subsequent application of generic statistical methods more palatable. Here, we describe four transformation approaches based on the delta method, model residuals, inferred latent expression state, and factor analysis. We compare their strengths and weaknesses and find that the latter three have appealing theoretical properties. However, in benchmarks using simulated and real-world data, it turns out that a rather simple approach, namely, the logarithm with a pseudo-count followed by principal component analysis, performs as well or better than the more sophisticated alternatives.
Software The R package transformGamPoi implementing the delta method- and residuals-based variance-stabilizing transformations is available via Bioconductor. We provide an interactive website to explore the benchmark results at shiny-portal.embl.de/shinyapps/app/08_single-cell_transformation_benchmark.
Contact constantin.ahlmann{at}embl.de
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
The manuscript was fully revised to follow the more familiar structure of Introduction, Results, Discussion, and Methods. The benchmarking section was significantly expanded and now encompasses three different benchmarks and compares 22 different transformations (representing 4 major transformation classes). We also now consider alternative performance metrics in addition to the overlap of k nearest neighbor graphs. The discussion of how to select an appropriate overdispersion was shortened and complemented by an empirical evaluation.
https://shiny-portal.embl.de/shinyapps/app/08_single-cell_transformation_benchmark