Separating measurement and expression models clarifies confusion in single cell RNA-seq analysis

Abstract How to model and analyze scRNA-seq data has been the subject of considerable confusion and debate. The high proportion of zero counts in a typical scRNA-seq data matrix has garnered particular attention, and lead to widespread but inconsistent use of terminology such as “dropout” and “missing data.” Here, we argue that much of this terminology is unhelpful and confusing, and outline simple ways of thinking about models for scRNA-seq data that can help avoid this confusion. The key ideas are: (1) observed scRNA-seq counts reflect both the actual expression level of each gene in each cell and the measurement process, and it is important for models to explicitly distinguish contributions from these two distinct factors; and (2) the measurement process can be adequately described by a simple Poisson model, a claim for which we provide both theoretical and empirical support. We show how these ideas lead to a simple, flexible statistical framework that encompasses a number of commonly used models and analysis methods, and how this framework makes explicit their different assumptions and helps interpret their results. We also illustrate how explicitly separating models for expression and measurement can help address questions of biological interest, such as whether mRNA expression levels are multi-modal among cells.