Comparison and evaluation of statistical error models for scRNA-seq

Heterogeneity in single-cell RNA-seq (scRNA-seq) data is driven by multiple sources, including biological variation in cellular state as well as technical variation introduced during experimental processing. Deconvolving these effects is a key challenge for preprocessing workflows. Recent work has demonstrated the importance and utility of count models for scRNA-seq analysis, but there is a lack of consensus on which statistical distributions and parameter settings are appropriate. Here, we analyze 58 scRNA-seq datasets that span a wide range of technologies, systems, and sequencing depths in order to evaluate the performance of different error models. We find that while a Poisson error model appears appropriate for sparse datasets, we observe clear evidence of overdispersion for genes with sufficient sequencing depth in all biological systems, necessitating the use of a negative binomial model. Moreover, we find that the degree of overdispersion varies widely across datasets, systems, and gene abundances, and argues for a data-driven approach for parameter estimation. Based on these analyses, we provide a set of recommendations for modeling variation in scRNA-seq data, particularly when using generalized linear models or likelihood-based approaches for preprocessing and downstream analysis.