Detection of Differentially Expressed Genes in Discrete Single-Cell RNA Sequencing Data Using a Hurdle Model With Correlated Random Effects

Abstract

Single-cell RNA sequencing (scRNA-seq) technologies are revolutionary tools allowing researchers to examine gene expression at the level of a single cell.Traditionally, transcriptomic data have been analyzed from bulk samples, masking the heterogeneity now seen across individual cells. Even within the same cellular population, genes can be highly expressed in some cells but not expressed (or lowly expressed) in others. Therefore, the computational approaches used to analyze bulk RNA sequencing data are not appropriate for the analysis of scRNA-seq data. Here, we present a novel statistical model for high dimensional and zero-inflated scRNA-seq count data to identify differentially expressed (DE) genes across cell types. Correlated random effects are employed based on an initial clustering of cells to capture the cell-to-cell variability within treatment groups. Moreover, this model is flexible and can be easily adapted to an independent random effect structure if needed. We apply our proposed methodology to both simulated and real data and compare results to other popular methods designed for detecting DE genes. Due to the hurdle model's ability to detect differences in the proportion of cells expressed and the average expression level (among the expressed cells), our methods naturally identify some genes as DE that other methods do not, and we demonstrate with real data that these uniquely detected genes are associated with similar biological processes and functions.

Differential expression of single-cell RNA-seq data using Tweedie models

“…We compared the performance of our proposed methods to a range published differential expression analysis tools. Specifically, we considered (i) four representative methods from the scRNA-seq literature: MAST (Finak et al, 2015), monocle (Trapnell et al, 2014), scREhurdle (Sekula et al, 2019) (NRE version), and DESingle (Miao et al, 2018); (ii) two methods designed for bulk RNA-seq differential expression: edgeR (Robinson et al, 2010) and DESeq2 (Love et al, 2014); (iii) one method specially geared towards metagenomics differential abundance analysis: metagenomeSeq (Paulson et al, 2013), and (iv) finally, a commonly used non-parametric method: Wilcoxon test. DE genes were determined at a FDR of 0.05 for each method.…”

“…Recent work has demonstrated that data generated from different scRNA-seq experimental protocols can result in differences in the distribution of gene expression, for example, between UMI counts and read counts, including differential zero-inflation or mean-variance patterns across experimental platforms (Vieth et al, 2017; Townes et al, 2019; Hafemeister and Satija, 2019; Svensson, 2020; Cao et al, 2021). This work emerged, in part, because historically many bioinformatics tools and statistical methods have been broadly proposed to model this technological variation in downstream analyses such as (i) zero-adjusted or zero-undjusted continuous models (Paulson et al, 2013; Korthauer et al, 2016; Soneson and Robinson, 2018), (ii) two-part hurdle models (Finak et al, 2015; Sekula et al, 2019), and (iii) countbased models such as Poisson, negative binomial, or multinomial models with (or without) zero-inflation components (Risso et al, 2018; Alessandrì et al, 2019; Hie et al, 2020).…”

“…Specifically in the context of differential expression (DE), many early methods designed for non-UMI counts focused on the use of zero-inflated or hurdle models (Finak et al, 2015; Korthauer et al, 2016; Miao et al, 2018; Sekula et al, 2019). Several of these methods require a data transformation including the use of a pseudocount and log-transformation, but this has been recently shown to introduce false variation in downstream analyses (Townes et al, 2019; Hafemeister and Satija, 2019).…”

Combinatorial and statistical prediction of gene expression from haplotype sequence

“…Both low and high expression groups in RNA-seq data have previously been associated with disease risk ( Gamazon et al , 2015 ; Zheng et al , 2018 ) and cancer prognoses ( Tichỳ et al , 2019 ). Further, allele specific expression and single-cell RNA-seq data commonly include genes with multimodal expression ( Kharchenko et al , 2014 ; Shalek et al , 2013 ), and recently, methods have been developed to detect differential expression in discretized expression data ( Sekula et al , 2019 ).…”

Bayesian Correlation is a robust similarity measure for single cell RNA-seq data

“…Noise in gene expression measurements has been modeled and studied to identify differentially expressed genes [36,31,70]. Recently, uncertainties have been incorporated in methods to study differential expression in RNA-seq experiments [52] and a Bayesian scheme has been proposed to identify differentially expressed genes in scRNA-seq [63]. Noise is especially important in scRNA-seq because the low number of read counts.…”