Accounting for technical noise in differential expression analysis of single-cell RNA sequencing data

Abstract

Recent technological breakthroughs have made it possible to measure RNA expression at the single-cell level, thus paving the way for exploring expression heterogeneity among individual cells. Current single-cell RNA sequencing (scRNA-seq) protocols are complex and introduce technical biases that vary across cells, which can bias downstream analysis without proper adjustment. To account for cell-to-cell technical differences, we propose a statistical framework, TASC (Toolkit for Analysis of Single Cell RNA-seq), an empirical Bayes approach to reliably model the cell-specific dropout rates and amplification bias by use of external RNA spike-ins. TASC incorporates the technical parameters, which reflect cell-to-cell batch effects, into a hierarchical mixture model to estimate the biological variance of a gene and detect differentially expressed genes. More importantly, TASC is able to adjust for covariates to further eliminate confounding that may originate from cell size and cell cycle differences. In simulation and real scRNA-seq data, TASC achieves accurate Type I error control and displays competitive sensitivity and improved robustness to batch effects in differential expression analysis, compared to existing methods. TASC is programmed to be computationally efficient, taking advantage of multi-threaded parallelization. We believe that TASC will provide a robust platform for researchers to leverage the power of scRNA-seq.

Using multiple measurements of tissue to estimate subject- and cell-type-specific gene expression

“…For each subject and gene, MIND's CTS estimate represents the average 13 expression of the gene for fundamental cell types, such as neurons, astrocytes and 14 oligodendrocytes in brain. 15 1 Two ideas are key for obtaining CTS gene expression from tissue. First, because a tissue 16 sample's bulk transcriptome is a convolution of gene expression from cells belonging to various 17 cell types, deconvolution methods 7,8,9,10 can estimate the fraction of each cell type within this 18 tissue.…”

“…MIND also assumes CTS expression is similar across brain regions of the 57 134 Remark: While our estimates of the abundance of pyramidal neurons, for example, match 135 previous findings, such estimates can be inconsistent with those from neuroanatomical and 136 other direct studies of cell representation 13,14 . To better understand the estimated cell type 137 fractions, we studied the relationship between cell size and gene expression in GTEx data using 138 techniques in Jia et al 15 and results from Zeisel et al 2 . We find that the estimated cell size is 139 highly positively correlated with level of gene expression ( Supplementary Fig.…”

“…For the Remark about cell size and level of gene expression, our analyses made use 376 of the scRNA-seq data in Zeisel et al 2 , which also contained spike-in information. We 377 leveraged the spike-in information to estimate cell size 15 for neurons and non-neurons 378 (Supplementary Fig. 5) and thus interpreted the impact of size versus cell type composition 379 in the deconvolution of bulk transcriptomes.…”

Latent periodic process inference from single-cell RNA-seq data

“…This is a tremendous advantage over other model-free approaches, such as reCAT, for the purpose of characterizing cell cycles. Evidently, Cyclum's demonstrated robustness to a reduced number of cells and genes makes it desirable to analyze current scRNA-seq datasets, which often suffer from cell-specific dropout and amplification bias [29]. Cyclum also appeared to work better on data that was heavily confounded by cell cycles.…”

SCRIBE: a new approach to dropout imputation and batch effects correction for single-cell RNA-seq data

“…However, the existence of batch effects makes data integration difficult since they are often associated with an outcome of interest [3]. Since the earliest observation of batch effects in microarray experiments [4], an explosion of batch effects correction methods for both bulk RNA-seq data [5][6][7][8][9][10][11][12][13] and scRNA-seq data [14][15][16][17][18][19] have been developed over the past two decades. Moreover, the inflation of zero caused by dropout events [20] renders more noise to scRNA-seq data than bulk RNA-seq data.…”