Latent variable modelling and variational inference for scRNA-seq differential expression analysis

Abstract

Disease profiling, treatment development, and the identification of new cell populations are some of the most relevant applications relying on differentially expressed genes (DEG) analysis. In this context, three leading technologies emerged; namely, DNA microarrays, bulk RNA sequencing (RNA-seq), and single-cell RNA sequencing (scRNA-seq), the main focus of this work. Although scRNA-seq tends to offer more accurate data, it is still limited by many confounding factors. We introduce two novel approaches to assess DEG: extended Bayesian zero-inflated negative binomial factorization (ext-ZINBayes) and single-cell differential analysis (SIENA). In addition, we benchmark the proposed methods with known DEG analysis tools for single-cell and bulk RNA data, using two real public datasets. One contains house mouse cells of two different types, while the other gathers human peripheral blood mononuclear cells divided into four types. The results show that the two procedures can be very competitive with existing methods (scVI, SCDE, MAST, and DEseq) in identifying relevant putative biomarkers. In terms of scalability and correctness, SIENA stands out from ext-ZINBayes and some of the existing methods. As single-cell datasets become increasingly larger, SIENA may emerge as a powerful tool to discover functional differences between two conditions. Both methods are publicly available (https://github.com/JoanaGodinho/SIENA, https://github.com/JoanaGodinho/ext-ZINBayes).