Normalization by distributional resampling of high throughput single-cell RNA-sequencing data

MotivationNormalization to remove technical or experimental artifacts is critical in the analysis of single-cell RNA-sequencing experiments, even those for which unique molecular identifiers (UMIs) are available. The majority of methods for normalizing single-cell RNA-sequencing data adjust average expression in sequencing depth, but allow the variance and other properties of the gene-specific expression distribution to be non-constant in depth, which often results in reduced power and increased false discoveries in downstream analyses. This problem is exacerbated by the high proportion of zeros present in most datasets. ResultsTo address this, we present Dino, a normalization method based on a flexible negative-binomial mixture model of gene expression. As demonstrated in both simulated and case study datasets, by normalizing the entire gene expression distribution, Dino is robust to shallow sequencing depth, sample heterogeneity, and varying zero proportions, leading to improved performance in downstream analyses in a number of settings. Availability and implementationThe R package, Dino, is available on GitHub at https://github.com/JBrownBiostat/Dino. Contactbrownj@biostat.wisc.edu, kendzior@biostat.wisc.edu