Single-cell RNA sequencing (scRNA-seq) aims to characterize how variation in gene expression is distributed across cells in tissues and organisms. Yet, effective comprehension of these extremely high-dimensional datasets remains a critical barrier to progress in biological research. In standard analyses of scRNA-seq data, feature selection steps aim to reduce the dimensionality of the data by focusing on a subset of genes that are the most biologically variable across a set of cells. Ideally,

more »

... ese features provide the genes that are the most informative for partitioning groups of transcriptionally distinct cells, each representing a different cell type or identity. In this work, we propose a simple feature selection model where a binomial sampling process for each mRNA species produces a null model of technical variation. To compare our model to existing methods, we use scRNA-seq data where cell identities have been established a priori for each cell, and characterize whether different feature sets retain biologically varying genes, distort neighborhood structures, and allow popular clustering algorithms to partition groups of cells into their established classes. We find that our model of biological variation, which we term "Differentially Distributed Genes" or DDGs, outperforms existing methods, and enables dimensionality reduction without loss of critical structure within the data set.