Design and computational analysis of single-cell RNA-sequencing experiments.

Bacher R, Kendziorski C. · 2016

Read Design and computational analysis of single-cell RNA-sequencing experiments. by Bacher R, Kendziorski C. (2016) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc4823857.txt.

Introduction

Rhonda Bacher and Christina Kendziorski review the computing side of single-cell RNA sequencing in Genome Biology in 2016. The method reads gene activity in one cell at a time, answering questions that pooled samples blur. It also brings fresh statistical trouble: many genes read as zero in a cell, scatter runs high, and patterns arrive in multiple humps.

The review maps the tools on offer for planning such studies and for each analysis step, from quality checks to finding cell groups and comparing conditions. It notes where each tool shines or strains, and where no good method exists yet. The authors close with the developments they expect as the method matures, aiming the piece at both bench scientists and statisticians.

Catalog record

CREATORBacher R, Kendziorski C.
YEAR2016
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC4823857

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Design and computational analysis of single-cell RNA-sequencing experiments By Rhonda Bacher, Christina Kendziorski Published in: Genome Biology DOI: 10.1186/s13059-016-0927-y ABSTRACT Single-cell RNA-sequencing (scRNA-seq) has emerged as a revolutionary tool that allows us to address scientific questions that eluded examination just a few years ago. With the advantages of scRNA-seq come computational challenges that are just beginning to be addressed. In this article, we highlight the computational methods available for the design and analysis of scRNA-seq experiments, their advantages and disadvantages in various settings, the open questions for which novel methods are needed, and expected future developments in this exciting area. BACKGROUND The ability to derive genome-wide mRNA expression data from a population of cells has proven useful in thousands of studies over the past two decades. In spite of their utility, traditional expression experiments are limited to providing measurements that are averaged over thousands of cells, which can mask or even misrepresent signals of interest. Fortunately, recent technological advances now allow us to obtain transcriptome-wide data from individual cells. This development is not simply one more step toward better expression profiling, but rather a major advance that will enable fundamental insights into biology. While the data obtained from single-cell RNA-sequencing (scRNA-seq) are often structurally identical to those from a bulk expression experiment (some K million mRNA transcripts are sequenced from n samples or cells), the relative paucity of starting material and increased resolution give rise to distinct features in scRNA-seq data, including an abundance of zeros (both biological and technical), increased variability, and complex expression distributions (Fig. 1). These features, in turn, pose both opportunities and challenges for which novel statistical and computational methods are required.Fig. 1Prominent features in single-cell RNA-seq data relative to bulk RNA-seq include an abundance of zeros, increased variability, and multi-modal expression distributions. a Boxplots of the gene-specific proportion of zeros in a bulk (bulk1) and single-cell (sc1) dataset stratified by percentile of median gene expression. Sequencing depth ranges from 420,000 to 16.6 million in bulk1 and 385,000 to 16.4 million in sc1 (samples were chosen to have comparable depths; see the “Data” section). b Densities of gene-specific log variance for all genes in three bulk and three single-cell RNA-seq datasets. Densities are also shown for the single-cell datasets for log variances calculated following the removal of zeros, emphasizing that the increased variability observed relative to bulk is not entirely due to the presence of zeros. c For each dataset shown in b, 1000 genes were selected at random from the list of genes for which at least 75 % of cells showed non-zero expression. For each gene, zeros were removed and Mclust [92] was applied to log expression to estimate the number of modes. Because zeros were removed prior to Mclust, a mode at zero will not contribute to the total number of modes shown In this review, we discuss such statistical and computational methods in detail. We begin with an overview of practices for robust experimental design, quality control, and expression estimation, where the principles and methods used in bulk experiments, perhaps slightly modified, apply directly. We then discuss methods for normalization, noting that features that are unique to scRNA-seq pose challenges for existing approaches. Methods developed for downstream analysis are also considered with specific focus on methods for clustering and sub-population identification, assessing differences in gene expression across conditions, pseudotime ordering, and network reconstruction (a summary is provided in Table 1). We conclude with a discussion of the open questions facing computational scientists, as well as those that will soon arise as datasets quickly become ever larger and more complex.Table 1Statistical methods for single-cell RNA-seq experimentsNameDescriptionRequirements/deliverablesNormalizationGRM [57]Fits polynomial gamma regression model to FPKM data from spike-ins; estimated parameters are used to convert FPKM of endogenous genes to an absolute scale within each cellPerforms within cell normalization and may be used with FPKM, RPKM, or TPMSAMstrt [56]The resampling-based bulk normalization method in SAMseq is applied to spike-insAssumes that an equal number of spike-in control RNA molecules have been added to all samplesIdentifying highly variable genesBrennecke et al. [48]A gamma generalized linear model fit to the mean-variance relationship quantified by the square of the coefficient of variation (CV2) of the spike-ins estimates technical noise parameters. These parameters are then used to estimate technical variability for endogenous genes and to test whether each gene exceeds a variability thresholdSpike-ins and endogenous genes are normalized separately using the median normalization method. Gene specific P values are provided to identify highly variable genesKim et al. [63]Uses spike-ins to estimate parameters related to technical variance, allowing for differences in variability across cells. Estimates gene-specific biological variability by subtracting technical variability from total varianceNormalization factors are estimated using the median normalization method. A simulation based framework to test for highly variable genes is providedBASiCS [54]Jointly models spike-ins and endogenous genes as two Poisson-Gamma hierarchicalmodels with shared parametersEstimates normalization parameters jointly across all genes. Gene-specific posterior probabilities are provided to identify both lowly and highly variable genesNoise reductionscLVM [47]Uses a Gaussian Process Latent variable model to estimate the covariance matrix associated with latent factors. Residuals from a linear mixed model with the covariance term represent de-noised expression estimatesRequires genes associated with the latent factor to be identified a priori. Normalization factors are estimated using the median normalization methodOEFinder [12]Uses orthogonal polynomial regression to identify genes whose expression is associated with position on the C1 Fluidigm integrated fluidic circuit (IFC)Gene-specific P values are provided to identify genes affected by the artifactSub-population identificationZIFA [70]Models dropout rate as a function of expression in a factor analysis (linear dimension reduction) frameworkRequires normalized, log-transformed estimates of gene expression (zeros are not transformed)Destiny [81, 82]Extends diffusion maps (a non-linear dimension reduction approach) to handle zeros and sampling density heterogeneities inherent in single cell dataRequires variance-stabilized gene expression estimates; works best with a large number of cellsSNN-Cliq [71]Clusters cells by identifying and merging sub-graphs (quasi-cliques) in a shared nearest neighbor (SNN) graph; the number of clusters is chosen automaticallyRequires a reduced set of genes. Xu and Su [71] recommend using genes with average RPKM >20 and using a log transformation to reduce the effect of outliers. Relies on a valid choice of graph parametersRaceID [59]Uses k-means applied to a similarity matrix of Pearson’s correlation coefficients for all pairs of cells; the number of clusters is chosen using the gap statistic. Outlier cells are those that cannot be explained by a background model that accounts for technical and biological noise. In a second step, rare subpopulations can be identified and outlier cells may be merged to an outlier cluster; new cluster centers are then computed and each cell is assigned to the most highly correlated cluster centerRequires a reduced set of genes. Grün et al. [59] consider genes with a minimum of five transcripts in at least one cellSCUBA [73]Uses k-means to cluster data along a binary tree detailing bifurcation events for time-course data. Models expression regulation along the tree using bifurcation theoryRequires a reduced set of genes. Marco et al. [73] recommend using the 1000 most variable genes that are expressed in at least 30 % of cellsBackSPIN [60]Iteratively splits a two-way sorted (by both genes and cells) expression matrix into two clusters containing independent cells and genes, for

Read the complete text