A deep generative model for single-cell RNA sequencing with application to detecting differentially expressed genes

Single-cell RNA sequencing (scRNA-Seq) is a revolutionary technology, which allows studying fundamental biological questions that were previously out of reach [1, 2]. It allows, for the first time, to reveal a cell’s identity and characterize its molecular circuitry in an unbiased, data-driven way. The product of a scRNA-Seq experiment is a data matrix X where entry Xng approximates the number of transcripts of gene g in cell n. Careful computational analysis allows deriving from such data exciting insights in diverse biomedical fields [3, 4]. While it is typical to observe thousands of gene products per cell, many transcripts are observed very infrequently, and for technical reasons related to the method of sequencing these are particularly prone to high variance. Additionally, due to the limited transcript capture efficiency inherent to RNA-Seq protocols, entries of X are typically zero-inflated [5].