Statistical and Bioinformatics Analysis of Data from Bulk and Single-Cell RNA Sequencing Experiments

Abstract

High-throughput sequencing (HTS) has revolutionized researchers’ ability to study the human transcriptome, particularly as it relates to cancer. Recently, HTS technology has advanced to the point where now one is able to sequence individual cells (i.e., “single-cell sequencing”). Prior to single-cell sequencing technology, HTS would be completed on RNA extracted from a tissue sample consisting of multiple cell types (i.e., “bulk sequencing”). In this chapter, we review the various bioinformatics and statistical methods used in the processing, quality control, and analysis of bulk and single-cell RNA sequencing methods. Additionally, we discuss how these methods are also being used to study tumor heterogeneity.

1. Introduction

Cancer research is becoming more computational focused in which researchers require working knowledge of data science methods to mine and extract new insights from large molecular datasets to derive novel hypotheses. Much of this need stems from the initial sequencing of the human genome to the advent of single-cell sequencing technologies, which has resulted in an explosion in the ability of cancer researchers to generate and acquire high-dimensional data. The most common type of ‘omic data currently generated in the study of cancer biological systems is mRNA gene expression or transcriptomic data.

Uncontrolled cell growth and proliferation is a hallmark feature of cancer. In doing so, genes that regulate cell growth and differentiation are altered, thus leading to the uncontrollable growth and oncogenesis. Often, these changes in gene regulation are caused by mutations that disrupt the “normal” function of the gene and the downstream protein, such as in the case of p53 [1]. In other cases, the gene regulation is changed by aberrant DNA methylation, where by silence the expression of a given gene [2]. Thus, studying the gene expression gives us clues into the biological mechanism of oncogenesis, along with tumor progression and growth. Additionally, gene expression has been used to subclassify a tumor class into smaller, more homogeneous molecular subtypes. For example, gene expression studies have been able to successfully subclassify breast cancer tumors into four to six molecular subtypes (luminal A, luminal B, HER2-enriched, triple-negative/basal-like, normal-like, claudin low) [3–7].

With the advances in technology from microarrays to high-throughput sequencing (HTS), one is able to detect other transcriptomic events, such as gene fusions or chimeras, isoform expression, and allelic expression [8]. Recently, HTS technology has advanced to the point that now one is able to sequence individual cells (i.e., “single-cell sequencing”) [9, 10]. Prior to single-cell sequencing technology, HTS would be completed on RNA extracted from a tissue sample made up of multiple cell types (i.e., “bulk sequencing”). The difference between single cell sequencing and bulk sequencing of RNA is that in the former the sequencing library represents a single cell while the latter represents a population of cells (Fig. 1). Single-cell technology allows researchers study the transcriptome of different cells within the same tissue type. This technology is particular useful in studying cancer immunology and the dissection of tumor heterogeneity, as tumors and the stromal component of tumors are a composition of (1) different cancer cells developed from different genomic events (i.e., clones, tumor heterogeneity) [11, 12] and (2) mixture of cancer cells and immune cells (i.e., tumor infiltrating lymphocytes (TILs)) [13, 14]. In the following sections we will outline the various bioinformatics and statistical methods used in the analysis of bulk and single-cell RNA sequencing.

Fig. 1.

Illustration of differences between bulk RNA sequencing (RNA-seq) vs. single-cell RNA sequencing (scRNA-seq)

2. Datasets Used to Illustrate the Methods

Melanoma is the fifth most common malignance in the United States and it is estimated that 96,480 individuals will be diagnosed with melanoma in 2019 and that an estimated 7230 will die from melanoma [15]. Many genomic studies have been conducted to understand the molecular features of melanoma. Data from The Cancer Genome Atlas (TCGA) determined four major subtypes of cutaneous melanoma: BRAF mutant (52% of tumors), RAS mutant, NF1 mutant, and Triple Wild-Type [16]. It was also found that the immune system plays a central role in the progression and treatment response in melanoma patients. To better understand the influence of immune system and tumor heterogeneity in melanoma, many studies have recently been completed to understand melanoma at the single-cell level [17–20]. To illustrate the bioinformatic and statistical methods used in the analysis of bulk and single-cell RNA-sequencing, we will use data from the TCGA melanoma study [16] and the Tirosh et al. study [20].

2.1. Bulk RNA-Sequencing Study: TCGA Skin Cancer Study

The RNA-seq summarized gene expression levels of skin cutaneous melanoma study (SKCM) using data obtained from TCGA project were downloaded via Genomic Data Commons (GDC) [16]. To illustrate differential expression analysis using RNA-seq data, set out to determine differentially expressed genes between primary tumors (N = 67) and metastatic tumors (N = 213). After filtering nonexpressed or low-expressed genes based on counts per million (CPM), 22,236 genes with CPM values above 1 in at least four libraries remain. To illustrate assessment of batch effects, technical artifacts were downloaded from MBatch (https://bioinformatics.mdanderson.org/tcgabatcheffects) for the SKCM TCGA data. This data set included the following variables: tissue source site (25 levels), plate ID (16 levels), batch ID (14 levels), and ship date (14 levels).

2.2. Single Cell Sequencing Study: Tirosh et al. Study

Tirosh et al. [20] measured single-cell RNA-seq gene expression of 4645 melanoma, immune, and stromal cells from 19 melanoma tumors. These tumors included one primary acral melanoma, ten metastases to lymphoid tissues, and eight metastases to distant sites. The immune (CD45⁺) and nonimmune (CD45⁻, including melanoma and stromal) cells were sorted into 96-well plates by flow cytometry (fluorescence-activated cell sorting). Single cell RNA was then isolated and sequenced with SMART-Seq2 protocol [21]. The gene expressions were quantified as y = log2 (TPM + 1), where TPM refers to transcripts per million. Cells with either fewer than 1700 detected genes or average housekeeping gene expression below 3 were excluded.

3. Statistical and Bioinformatics Methods for Analysis of Bulk RNA-Seq Data

Current RNA-seq protocols still possess several essential biases and limitations, such as nucleotide composition bias, GC content, and polymerase chain reaction artifact or contaminations [22, 23]. Raw RNA-seq data must be checked and processed by quality control (QC) procedures to ensure accurate transcript measurements. Initial steps in the QC process typically involve assessing such biases of the raw reads using metrics generated by the sequencing platform or calculated directly from the raw reads (Table 1). One of the most popular tools for the generation of these quality metrics is FastQC (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). The RNA-seq raw data and alignment files include various formats such as FASTA to store reference genome [24], gene transfer format (GTF) to store transcript/gene annotations, FASTQ to store raw read data [25], and the sequence alignment map (SAM/BAM) to store read alignments [26]. RNA-seq analysis typically is consisted of major steps including raw data quality control (QC), read alignment, transcriptome reconstruction, expression quantification, and end with downstream analysis (Fig. 2).

Table 1.

Selected list of RNA-seq analysis tools for preprocessing, read alignment, transcriptomic reconstruction, and expression quantification/abundance estimation

Workflow	Category	Software tools	Reference
Preprocessing	RAW data QC	FastQC	Babraham Bioinformatics website
		HTQC	Yang et al. [182]
	Read trimming	NGS QC	Patel and Jain [183]
		FASTX-Toolkit	Cold Spring Harbor Laboratory website
		Trimmomatic	Bolger et al. [184]
		SolexaQA	Cox et al. [185]
Read alignment	Spliced aligner	TopHat	Trapnell et al. [186]
		STAR	Dobin et al. [30]
		MapSplice	Wang et al. [29]
		GSNAP	Wu et al. [31]
	Unspliced aligner	Stampy	Lunter and Goodson [32]
		MAQ	Li et al. [33]
		BWA	Li and Durbin [34]
		Bowtie2	Langmead and Salzberg [35]
Transcriptome reconstruction	Reference-guided	Cufflinks	Trapnell et al. [36]
		StringTie	Pertea et al. [37]
	Reference-independent	Trinity	Grabherr et al. [39]
		Oases	Schulz et al. [40]
		transABySS	Robertson et al. [187]
Expression quantification	Gene-level quantification	featureCounts	Liao et al. [43]
		HTSeq	Anders et al. [44]
	Isoform-level quantification	Cufflinks	Trapnell et al. [36]
		StringTie	Pertea et al. [37]
		RSEM	Li and Dewey [41]
		Sailfish	Patro et al. [42]

Fig. 2.

Typical bulk RNA sequencing workflow

3.1. Quality Control of RNA-Seq Data

QC of raw data should be performed as the initial step which involves assessing such biases using metrics generated by the sequencing platform or calculated directly from the raw reads. In addition, depending on the RNA-seq library construction strategy and sequencing [27], trimming strategies include “adapter trimming” and “quality trimming” can be used to remove low-quality reads, trim adaptor sequences, and eliminate poor-quality bases. Adapter trimming is not necessary as most recent sequencers provide raw read in which the adapters are already trimmed, while quality trimming may be an important step depending on the analysis procedure used. Table 1 represents the widely used sequencing QC software tools.

3.2. Methods for Read Alignment

Reference-based alignment is the process used to determine the potential mapping locations by exact match or scoring sequencing similarity. Reads are typically aligned to either a genome or a transcriptome as a reference using two common approaches; (1) splice-aware read aligner or (2) nonsplice-aware read aligners (Table 1). The spliced read aligners use of a gapped or spliced mapper as reads may span splice junctions. Various spliced aligners have been developed including TopHat2 [28], MapSplice [29], STAR [30], and GSNAP [31]. Unspliced read aligners do not allow large gaps, such as those arising from reads spanning exon boundaries, or splice junctions including Stampy [32], mapping and assembly with quality (MAQ) [33], Burrow–Wheeler Aligner (BWA) [34], and Bowtie2 [35].

3.3. Methods for Transcript Reconstruction

Transcript reconstruction includes two approaches to identify all transcripts expressed in a specimen depends on the presence or absence of a reference sequence (Table 1). When the reference annotation information is well-known, then the reference-based approaches such as Cufflinks [36] and StringTie [37] are used to reconstruct transcripts by assembly of overlapping aligned reads. When a reference genome or transcriptome is not available or is incomplete, assembled de novo algorithm directly builds transcripts from short reads using platforms such as Trinity [38], transABySS [39], and Oases [40].

3.4. Methods for Gene Summarization or Abundance Estimation

One of the most widely used applications of RNA-seq is to quantify expression levels of genes and transcripts. Generally, the methods for gene quantification can be divided into two categories: “union exon”-based and “transcript”-based approaches (Table 1). Transcript-based approach fundamentally distributes reads among transcript isoforms including RSEM [41], Cufflinks [36], and StringTie [37]. However, some transcript-based quantification tools such as Sailfish [42] are alignment-free tools to estimate isoform abundances directly from a set of reference sequences. The “union exon”-based methods, such as featureCounts [43] and HTSeq [44], are widely used in RNA-seq gene quantification because of its simplicity to aggregate raw counts of mapped reads.

3.5. Normalization

Variability in measurement can be attributed to both the biological and technical factors. Sources of technical variation, involving, differences in library preparation across samples, sequencing error, mapping and annotation bias, sequencing composition and similarity, gene length, and sequencing depth [45–47] that can significantly reduce the accuracy of statistical inferences and also prevent researchers from properly modeling biological variation and group-specific changes in gene expression [48–51]. Some sources of between-sample technical variation are due to differences in library size or sequencing depth [47]. To correct for library size, most of methods use a common scaling factor per sample to normalize genes such as upper quartile (UQ) [45], median (Med) [45], relative log expression (RLE) [52], trimmed mean of M-values (TMM) [53], and quantile (Q) [54, 55]. Many of these methods are implemented in the edgeR and DESeq2 packages.

Additionally, gene length impacts the comparison of abundance estimates between genes [56], as longer genes contribute more sequenced fragments compared to shorter ones [47]. Most commonly used methods for gene length correction include TPM (transcripts per million) [57] and RPKM/FPKM (reads/fragments per kilobase per million mapped reads) [36, 47], where the former one is considered more robust to differences in RNA library size [58, 59]. However, it has been shown that scaling by gene length cannot entirely remove the positive association between gene size and read counts, and can introduce new biases to the estimates of differential expression [45, 60]. Usual normalization approaches mostly adjust the sequencing depth or/and gene length and fail to correct known or unknown technical artifacts due to the complex. To adjust known batch effect, appropriate statistical models were proposed such as linear regression model or flexible empirical Bayes method (ComBat) which is more robust and appropriate for small sample sizes [61]. ComBat can be implemented in R using the sva package and the function “ComBat.”

In addition to correcting for known artifacts, assessment and adjustment for potential latent factors is also warranted where mostly rely on singular value decomposition (SVD) or some other factor analysis approaches (e.g., Remove Unwanted Variation (RUV) in R package “ruv” [50] or surrogate variable analysis (SVA) [49] in R package “sva”). Finally, principal component analysis (PCA) is often used in which a subset of the principal components are used to normalize the data [62–65]. Note that the choice of normalization method to remove technical artifacts can affect noticeably the results of differential gene expression analyses [45, 53, 66]. For the TCGA skin cancer study, PCA was completed using filtered raw counts to assess the effects of the batch ID (Fig. 3a) and biological factor (i.e., primary and metastatic tumors groups) (Fig. 3b). Figure 3a represents the effect of batch ID with high proportion of variation for top principal component (15%) which leads to consideration of batch ID as a technical effect for normalization.

Fig. 3.

Plots from Principal Component Analysis to assess technical batch and biological factor effects globally in bulk RNA-seq experiments for the TCGA skin cancer study. (a) Fourteen levels of known batch ID, where batch ID was downloaded from https://bioinformatics.mdanderson.org/BatchEffectsViewer/; and (b) primary factor of interest (primary tumor and metastatic tumor) using filtered raw counts data

3.6. Methods for Differential Gene Expression Analysis

The main methodologies for differential gene expression analysis for RNA-seq data are categorized by the distributional assumptions (Table 2). Models for read counts originated from the idea that the number of reads for each gene can be approximated by a Poisson distribution where log-linear or generalized linear model were proposed to model the mean difference between samples along with using test statistics such as likelihood ratio test, exact test, and score test for hypothesis testing are implemented in DEGseq [67], Myrna [68], and PoissonSeq [69] packages. The extended Poisson models, two-stage Poisson model [70] and generalized Poisson model [71] are also considered to adjust for overdispersion issue.

Table 2.

Statistical methods to identify differential gene expressions based on RNA-seq data

Model	Software	Reference
Poisson	DEGseq	Wang et al. [67]
	Myrna	Langmead et al. [68]
	PoissonSeq	Li et al. [69]
Negative binomial	edgeR	Robinson et al. [73]
	DESeq	Anders and Huber [52]
	DESeq2	Love et al. [74]
	NBPSeq	Di et al. [75]
Beta-binomial	BBSeq	Zhou et al. [76]
Bayesian and empirical Bayesian	ShrinkSeq	Van de Wiel et al. [77]
	baySeq	Hardcastle and Kelly [78]
Normal	limma+voom	Smyth [54, 79]
		Law et al. [81]
Nonparametric	SAMseq (samr)	Li and Tibshirani [82]
	NOIseq	Tarazona et al. [83]

Poisson and negative binomial distributions are the two widely used models [52, 72], whereas the higher variability between biological replicates leads to incorporate negative binomial distribution to accommodate overdispersion [45]. The dispersion parameter estimation can be based on the conditional maximum likelihood, pseudo-likelihood, quasi-likelihood, local regression, and conditional inference using hypothesis testing approaches such as Wald test, likelihood ratio test, and exact test. Such methods are included in edgeR [73], DESeq [52], DESeq2 [74], and NBPSeq [75] packages. A beta-binomial model is implemented in BBSeq [76] package which accommodates the overdispersion using logistic regression where maximum likelihood approach is applied to estimate overdispersion parameter. Moreover, full or empirical Bayesian frameworks are included in ShrinkSeq [77] and baySeq [78] packages. Lastly, methods have been proposed that allow RNA-seq to be modeled using a linear model framework (i.e., Gaussian distribution) through limma package [79, 80] using voom transformation [81] or approaches that do not assume any distribution assumption (nonparametric approaches) such as SAMseq [82] and NOIseq [83]. It should be noted that if the modeling assumptions are valid the parametric methods will be more powerful than the nonparametric methods. However, if the modeling assumptions are not valid, nonparametric methods would be advisable. In practice, it is difficult to assess the modeling assumptions and thus parametric methods are mostly often used, particularly when the sample size is small. In skin-TCGA study, to identify the differentially expressed genes between primary and metastatic tumors groups (Fig. 3b), the voom-transformed UQ normalized data is considered where the design matrix contains the estimated latent artifact and the batch ID along with the tumor groups using limma package.

3.7. Methods for Correcting for Multiple Testing

With thousands of genes to test, controlling both the overall Type I error rate and the desired statistical power becomes important. Multiple comparison adjustment approaches can control the familywise error rate (FWER), false discovery rate (FDR) (i.e., Benjamini and Hochberg approach [84]), and Bayesian FDR (q-values) [85, 86]. Various approaches control FWER and compute the adjusted P values such as Bonferroni [87], Holm [88], and Hochberg [89]. In contrast to the strong control of FWER, the FDR-based control is less conservative with the increased gain in power and has been widely used in cases where a large number of hypotheses are simultaneously tested. Figure 4 represents the heatmap of top differentially expressed genes (n = 8928) after correcting multiple testing, FDR_BH < 0.05.

Fig. 4.

Heatmap of the top differentially expressed genes (FDR_BH < 0.05) from the analysis of the TCGA skin cancer study

3.8. Studying TME Using RNA-Seq in Bulk Samples

A software tool CIBERSORT is widely used to estimate fractions of multiple cell types using gene expression data in bulk samples [90]. It is commonly used to characterize global immune landscape by estimating different proportions of different immune cells. For instance, in a recent large-scale study to characterize immune landscape by analyzing 10,000 tumors comprising 33 diverse cancer types, CIBERSORT was used to estimate immune infiltration fractions for understanding tumor–immune interaction [91]. Six immune subtypes identified are wound healing, IFN-g dominant, inflammatory, lymphocyte depleted, immunologically quiet, and TGF-β dominant across cancer types and provided this as a source, iAtlas (https://www.cri-iatlas.org/), for researchers to understand tumor–immune interaction and potential therapeutic opportunities. In addition to cell type identification, the cell–cell interaction from the ligand and receptor database is also incorporated. Although CIBERSORT is widely used, its performance potential is affected by statistical multicollinearity due to the inclusion of highly correlated immune cell types, and also was developed using expression on microarrays. TIMER is developed to select genes, which are negatively correlated with tumor purity for each cancer type, and then apply constrained least squares fitting to expression to predict the abundance of a subset of TILs: B cells, CD4 T cells, CD8 T cells, macrophages, neutrophils, and dendritic cells [92]. To capture the complexity in TME better, xCell attempted to infer 64 immune and stromal cell types by using gene set enrichment analyses and deconvolution method to analyze the 1822 harmonized human pure cell type transcriptomics samples [93].

4. Statistical and Bioinformatics Methods for Analysis of Single-Cell RNA-Seq Data

Single-cell RNA-sequencing (scRNA-seq) has been playing important roles in the study of tumor heterogeneity and tumor evolution. In contrast to the bulk RNA-seq where the average gene expressions are measured across a large population of cells, scRNA-seq quantifies transcriptome of individual cells. With the newly developed high-throughput cell separation technologies, thousands of cells per tumor can be profiled in parallel to capture intra-tumor heterogeneity at an unprecedented resolution. Multiple different platforms have been developed for scRNA-seq including SMART-seq [21], CEL-seq [94], Fluidigm C1 [95], Smart-seq2 [96], and more advanced droplet-based platforms including Drop-seq [97] and Chromium 10X [98]. In droplet-based platforms, cells are encapsulated in water-based droplets together with unique molecular identifiers (UMIs), a cell-specific and transcripts-specific barcoding system. These barcodes help to diminish the sequencing reads representation biases due to library amplification.

The considerable differences in cell isolation and molecule capture lead to large variations in sensitivity, specificity, and capacity of these platforms [99, 100]. However, they all rely on similar computational pipelines to reveal transcription dynamics. In the following sections, we will review the algorithms in major steps of scRNA-seq data analysis using the most commonly used droplet-based platforms, but the discussion applies to all platforms.

4.1. Quality Control

The droplet-based scRNA-seq platforms encapsulate thousands of cells individually into barcoded-droplets and sequence their RNA material simultaneously. All computational analyses and interpretations of results reply on the assumption of single-cell behavior, such that only one living cell exists in a single droplet. However, even for the most sensitive protocols, it is inevitable to have dead cells and doublets (multiple cells encapsulated in one droplet) [101]. Therefore, it is essential to apply quality control (QC) to identify the low-quality droplets/barcodes/cells which ought to be excluded from downstream analyses [97, 101–105].

A common QC metric for scRNA-seq is the number of transcripts/UMIs detected per droplet. A small number of transcripts detected per barcode are often an indicator for poor droplet capture, which can be caused by cell death and/or capture of random floating RNA molecules released by dead cells. Inversely, a considerable large number of transcripts with the same barcode can suggest doublets or floating RNA encapsulated together with a living cell. Percentage of mitochondrial transcripts is another common QC metric. A high number of mitochondrial transcripts suggest the cells might be undergoing stress, for example, from cell isolation and sorting process. It is advised to remove these cells since stress level is usually not the interests of scRNA-seq analysis. In addition to the cell-level QC, gene-level QC is often performed to exclude genes expressed in only a very small proportion of cells. Removing such genes can decrease technical noise, and speed up the downstream normalization and clustering.

When deciding on the cutoffs for scRNA-seq QC, it is important to take into the consideration the cell compositions of the samples being analyzed. Different cell types actively express different number of genes, different number of mitochondrial transcripts, and different sets of transcripts, especially when comparing tumor to normal cells. A too stringent cutoff can remove a cell population of interests if this population is rare in sample of cells. Therefore, we suggest that if the researchers have no prior knowledge about the cell compositions, a possible solution is to carry out initial cell type identification and use that information to guide the QC process.

4.2. Drop-Outs, Normalization, and Spike-Ins

When analyzing scRNA-Seq data, normalization is a critical step to adjust for unwanted biological effects and technical noise collectively known as “batch effects” that mask the real signal. Similar to bulk RNA-seq, scRNA-seq batch effects can come from the variations in handling protocols, library preparation, sequencing platforms, and sequencing depth. In addition to these variations commonly seen in bulk RNA-seq, a prominent characteristic of scRNA-seq data is zero inflation, where the expression count matrix of single cells is mostly filed with zeros [106–108] (Fig. 5). There are two sources of zero inflation: (1) biological reason—the real zeros, where the cells are in transient state of transcript bursting [109] or the genes simply do not express in a subpopulation of cells; (2) technical reason—dropout, which is caused by the inefficiency of mRNA capture such that a large percentage of mRNA molecules are not captured and consequently not sequenced. Besides the dropout events, individual cells can show stronger overdispersion than typically observed in bulk RNA-seq data, even for genes with median-to-high expression levels [108]. Furthermore, compared to bulk RNA-seq, scRNA-seq data is much more heterogeneous since the sequenced cells are usually of different populations, cell types, or statuses. Even the cells from the same population but undergoing different cell cycles can show very different expression profiles [110, 111]. To uncover the cellular heterogeneity, scRNA-seq studies often start with choosing the HVG (highly variable genes) that are most informative in distinguishing cell populations. It has been shown that the choice of HVG is highly affected by normalization [112, 113]. Therefore, it is important for scRNA-seq normalization to retain the cell-to-cell biological heterogeneity while removing the cell-specific noise at the same time.

Fig. 5.

Cell-to-cell variability observed in scRNA-seq data of two single B cells

4.2.1. Normalization Methods

The global-scaling normalization methods inherited from bulk RNA-seq analysis have been widely used in scRNA-seq. However, these methods cannot accommodate the cell-specific variability in scRNA-seq data and can lead to biased estimation of scaling factors [113–115]. To optimize the modeling of the cell-to-cell variability, multiple normalization methods specifically tailored for scRNA-seq data have been developed. Table 3 provides a detailed summary of their main features, statistical models, and special considerations when modeling cell-specific biological and technical variations. In the following sections, we will briefly discuss the key characteristics of these methods.

Table 3.

Normalization methods used in scRNA-seq data analysis

Method	Features	Statistical models	Notes on biological variation	Notes on technical variation	Other factors
SCDE [108]	A two-component mixture model is used to capture dropout events and amplification events; Differential expression is evaluated by Bayesian approach	1. Dropouts: Poisson 2. Amplification: Negative binomial
TASC [116]	An empirical Bayes approach models the cell-specific dropout rates and amplification bias	1. Biological variation: Poisson 2. Technical variation: Logistic regression		Cell-specific technical variation is modeled by spikes-ins	Cell cycle and cell size are considered as covariates
MAST [106]	Two-part generalized linear regression models dropout and amplification events	1. Probability of detection: logistic regression 2. Expression level: Gaussian	Fraction of genes detected in each cell (CDR) is used as proxy for both biological and technical variation		Cell size is captured by CDR
BASiCS [117]	Use integrated Bayesian hierarchical model to simultaneously quantify unexplained technical noise and cell-to-cell biological heterogeneity	1. Expression: Poisson 2. Random effect and size-specific factor: Gamma	Biological cell-specific variability is quantified by gene-specific parameters borrowing information across all cells	Technical variability is quantified based on spike-ins	Cell cycle and cell size are captured
SCALE [109]	Allele-specific transcription expression is modeled by a systematic statistical framework	1. Alleles expression status: empirical Bayes method 2. Allele-specific transcription kinetics: Poisson-Beta hierarchical model 3. Allelic difference: resampling-based test	Requires allele-specific read counts to start with	Technical variability is adjusted through spike-ins	Spike-ins/total reads servers as a proxy for cell size
scran [107]	Deconvolution frameworks estimates cell-specific factor by pooling multiple cells to deal with zero inflation		A ring arrangement by library size and sliding window is used to select random pool of cells; pool-based scaling factors are then deconvolved to yield cell-specific factors
SCnorm [115]	Use quantile regression to estimate gene-specific scaling factors based on sequencing depth	Quantile regression	Groups genes with similar on sequencing depth and then estimate scale factors within each group; Spiked-ins are not required, but can be helpful
BISCUIT [118]	Employs Bayesian inference to cluster single cells considering both biological and technical variation in parallel	Hierarchical Dirichlet Process mixture model (DPMM)	The model simultaneously estimates the heterogeneous clusters through the DPMM and infers the technical variation parameters for imputing dropouts

The first category we consider detects differential expressions among cells while adjusting technical variations that are specific to scRNA-seq data. This category includes SCDE [108], TASC [116], and MAST [106], which all employs empirical Bayesian frameworks to estimate dropout events and real amplification of transcripts. More specifically, TASC takes the cell size and cell cycle as covariates, where the former is estimated from the ratio of endogenous RNA reads to spike-ins and the latter is represented by the expression of curated genes [20]. Alternatively, MAST uses a fraction of genes detected in each cell as a proxy for technical and biological variation. Although most of these methods focus on the differential expression at the gene level, some go beyond and study the difference in allelic expression. One such example is SCALE, which models the allele-specific transcription bursting with both technical variation and cell size differences accounted for. However, SCALE requires input as allele-specific read counts at heterozygous loci, which can be challenging for many scRNA-seq platforms, especially the tag-based quantification methods including Drop-seq and 10X.

Another major category is to generate normalized gene expression matrix that can be used as input for downstream analysis. Based on how the scaling factors are modeled, we further stratify these methods into two groups: cell-specific methods including scran [107] and BASiCS [ 117], and gene-specific methods including SCnorm [115]. BASiCS estimates the cell-specific biological variations by borrowing information across all cells and all genes while quantifies the technical variations replying on spike-ins. Alternatively, scran first clusters the cells into more homogenous groups and then deconvolutes the pooled cells to yield cell-specific factors. On the other hand, SCnorm groups genes with a similar dependence on sequencing depth and then estimates the scale factor within each gene group.

In addition to above approaches that focus on differential expression and normalized gene expression matrix, which are concepts adopted from bulk RNA-seq, several methods have been developed to specifically target the downstream heterogeneity studies in scRNA-seq data. BISCUIT [118] is a cell-type dependent normalization which uses a Bayesian probabilistic model to iteratively normalize and cluster cells. It simultaneously assigns cells to cluster and learns cell-dependent parameters within each cluster. The inferred parameters are then used to generate cell-type dependent normalization that can be fed back to improve clustering. It has been shown that BISCUIT can identify more refined subtypes of cells than global normalization methods [119].

4.2.2. Drop-Out Imputation

Instead of directly normalizing endogenous genes, many chose to impute the drop-out events prior to normalization. There are mainly two strategies used in scRNA-seq data imputation:

1. To distinguish the biological zeros from technical zeros using models of expected gene expression, which is usually obtained either from borrowing information across cells or from spike-in sequences. These methods include DrImpute [120], SAVER [121], McImpute [122], scImpute [123], ALRA [124], and scRMA [125].

2. To reduce the noise by using information from neighboring data. This category includes MAGIC [126], netSmooth [127], and knn-smooth [128].

Although imputation can rescue missing information that is important to study cellular heterogeneity, concerns have been raised about their sensitivity and specificity. A positive-control based benchmarking study assessed these methods and concluded that most methods only provide small improvement [129]. Andrews and Hemberg [130] evaluatedthe false discovery rate of these methods using negative controls. They found that SAVER performed well on simulated data compared to others; however, all methods introduced false signals at various levels in the permuted real data. These limitations of imputing scRNA-seq data are probably due to the lack of a comprehensive and independent reference, for example, as in GWAS imputation. Until such reference is generated, caution should be used when imputing scRNA-seq data.

4.2.3. Spike-Ins

As shown in Table 3, many scRNA-seq normalization methods rely on the spike-ins to estimate the cell-specific technical variations. Spikes-ins are a set of synthetic RNA sequences added to the samples in a theoretically constant and known amount, in order to calibrate the gene measurement and distinguish the biological vs. technical variations in RNA-seq experiments. The most commonly used spike-ins are the 92 External RNA Control Consortium (ERCC) molecules [46] and the Spike-in RNA variant control mixes (SIRVs, Lexogen). These extrinsic control sequences have also been used in scRNA-seq experiments [131–133], where the spike-ins with different concentrations are added with a constant amount across all cells. Vallejos et al. [113] and Lun et al. [134] have discussed the benefit of extrinsic control sequencing in scRNA-seq. However, the use of spike-ins remains challenging.

Although often neglected, calibrating the amount of spike-ins sequences is critical and should depend on the endogenous mRNA content [50]. However, due to the large and unknown heterogenous among tumor microenvironment, it is difficult to obtain prior knowledge about the cell-type specific endogenous mRNA content before sequencing. In addition, the spike-in sequences do not reflect the gene-length and GC content in the mammalian transcriptome, such that the technical effects may be different for the extrinsic and intrinsic genes [113]. Moreover, it has been shown that spike-ins signals can vary across technical replicates [50], and only partial spike-in sequences can be actually sequenced and aligned [113]. Furthermore, the use of spike-ins in the recent developed large-scale droplet-based platforms is not as cost-effective as in small scale platforms. For example, to reduce the doublet rate in Chromium 10X, the percentage of cell-containing droplets is deliberately designed as low as 1–10%. The spike-ins are added evenly across all droplets, not just the cell-containing ones and consequently takes up the vast majority of sequencing reads. Due to these limitations and challenges, caution should be used when employing spike-ins for technical variation estimations. Additionally, efforts should be made to design spike-ins sequences accommodating the unique characteristics of scRNA-seq experiments.

As discussed above, normalization methods that are specially tailored for scRNA-seq are theoretically and operationally superior over the global-scaling normalization inherited from bulk RNA-seq. However, these methods vary substantially in terms of their assumptions and their models, where none of them outperform others under all scenarios in the performance assessment [114, 135]. Great efforts are actively underway to develop more efficient and robust normalization for scRNA-seq.

4.3. Data Integration and Batch Correction

Due to the complexity of scRNA-seq experiments, it is often difficult for a study to process all samples at the same time and/or using the same protocols. In such situations, it is necessary to integrate samples of different batches or even of different scRNA-seq platforms. Due to the unique data structure of scRNA-seq, batch correction methods designed for bulk RNA-seq are not suitable. Several approaches have been recently proposed to deal with sample-level batches in scRNA-seq data. kBET (k-nearest neighbor correct effect) [136] quantifies the sample-level batch effects using a χ²-based test. MNNs (mutual nearest neighbor) [137] corrects the effect using only a subset of populations shared between batches and has been implemented in scran [102] as the mnnCorrent function. The Seurat group [138] uses cell pairwise correspondences between single cells across datasets, termed as “anchors,” to integrate gene expressions across technologies and batches. All these approaches have been shown to have better performance than bulk RNA-seq batch-correction methods.

4.4. Dimension Reduction, Clustering, and Cell Type Identification

One of the most popular uses of scRNA-seq is to identify and characterize cell types within the heterogeneous tissues or samples. The de novo identification of putative cell types has been considered as an unsupervised clustering problem. In this section, we will discuss the main classes of clustering methods having been applied to scRNA-seq, as well as the remaining issues and challenges. We will also briefly touch on the supervised and semisupervised clustering.

4.4.1. Dimension Reduction and Feature Selection

The high dimensional transcriptome data generated by scRNA-seq provides tremendous information for uncovering the biology of cells. However, it also introduces challenges to statistical analysis which is often referred to as the “curse of dimensionality.” With a large number of genes measured in scRNA-seq, the distance between individual cells can become small and thus make it difficult to distinguish between-population differences and within-population differences [139]. The two main approaches to deal with the issue of high dimensionality are feature selection and dimension reduction.

Feature selection removes the uninformative genes in terms of their ability to distinguish cells, in order to reduce the dimensions used in analysis and speed up calculations. The most commonly used feature selection in scRNA-seq is to select the highly variable genes (HVG), assuming that genes with high variance are more likely due to biological signals rather than technical noise [140]. The normalization and usage of spike-ins can facilitate the selection of HVG as we discussed in the previous section. Other feature selection approaches include identifying biological relevant genes based on expression correlation between cells [97, 141, 142] and using the magnitude and/or significance of the correlation to select genes.

Dimension reduction, on the other hand, completes a projection of the gene expression data onto a lower dimensional space. There are many generic dimension reduction methods that can be applied to any high dimensional data, including principal component analysis (PCA), independent components analysis (ICA) [143], Laplacian eigenmaps [144], and t-distributed stochastic neighbor embedding (t-SNE) [97, 145]. In this section we will focus on PCA and t-SNE due to their popularity in scRNA-seq data analysis. PCA uses the orthogonal transformations to project the gene count matrix onto a reduced number of linear independent dimensions called principal components. The advantage of PCA is that it is relatively fast and can preserve the distance information among cells. However, PCA is restricted to linear combinations of variables, which can be inappropriate in the context of scRNA-seq. In particular, it has been reported that the first components are often related to the number of genes detected per cell rather than the biological signals [106, 146]. To improve PCA, Risso et al. [147] proposed ZINB-WaVE which uses a zero-inflated negative binomial to deal with the dropouts in scRNA-seq data.

T-distributed stochastic neighbor embedding (t-SNE) is a stochastic method that reduces high dimensions to two or three embeddings while preserving the local structure among cells, such that neighbor cells stay close and distant cells remain distant. Due to the probability distribution used in embedding estimation, t-SNE follows two rules: (1) all points (cells) repel each other, and (2) each point (cell) attracts its nearest neighbors [148]. Therefore, t-SNE can specifically project cells into more distantly isolated clusters, making it almost the standard choice in visualization exploration of scRNA-seq data. Figure 6 projects 4645 single cells extracted from 19 melanoma tumors [20] onto t-SNE 2D planes. The clusters formed by t-SNE are in good agreement with the cell types identified by the original study, and show a high degree of intra-tumor heterogeneity for malignant cells but not for immune/ stromal cells. However, t-SNE has its limitations. First, t-SNE is computationally expensive. This is particularly problematic in large-scale scRNA-seq studies which require analyzing hundreds of thousands of cells simultaneously. In addition, although t-SNE captures the local structure it often fails to preserve the global geometry of the data. When t-SNE places cells into distinctive clusters, the relative position of these clusters is almost arbitrary and with little biological meanings [148]. Moreover, the embedding is governed by a parameter “perplexity,” which controls the number of nearest neighbors each point is attracted to. Different perplexity choices can lead to different degrees of separation and the judgment of the appropriate perplexity is subject to the analysts. Several reviews have provided in-depth discussion and practical suggestions for using t-SNE in scRNA-seq analysis [148–150]. We would recommend readers consult with these reviews before making conclusions with t-SNE results.

Fig. 6.

t-SNE projection of single cells from melanoma tumors colored by patient origin (left) and cell type (right)

More recently, a nonlinear dimension reduction method, uniform manifold approximation and projection (UMAP) [151], implemented in R package “umap,” was developed as an alternative to t-SNE. It is claimed to preserve as much of the local structure and more of the global geometry with a shorter run time than t-SNE. Becht et al. [152] performed a systematic evaluation using well-annotated scRNA-seq data, and concluded that UMAP provided faster run times, higher reproducibility, and more meaningful cell clusters. Recognizing its advantage, several scRNA-seq analysis tools incorporate UMAP into their standard dimension reduction and visualization pipeline. Feature selection and dimension reduction are not necessarily mutually exclusive. As a matter of fact, dimension reductions are susceptible to the batch effects caused by cell-specific technical variations [106, 153]. Performing feature selection that removes genes with little biological relevant prior to dimension reduction can greatly reduce the technical noise [154]. The combination of these methods has been widely adopted by scRNA-seq analysis pipelines.

4.4.2. Unsupervised Clustering

Recent single-cell sequencing technologies enable the study of tumor or TME heterogeneity at a single cell level [155, 156]. One of the ongoing challenges on scRNA-seq data analyses is that cell type recognition or subpopulation classification since the tumor–stromal interaction has been shown to be important. The diverse cell types would often be visualized in t-SNE space (Fig. 6) or other similar linear or nonlinear projections. Unsupervised clustering is a central component of scRNA-seq analysis, as it can identify cell populations and thus strongly affects any downstream analysis. Although many methods have been developed and applied to scRNA-seq, clustering and the interpretation of clusters are still facing great biological and computational challenges. Kiselev et al. [139] and Andrews and Hemberg [ 154] reviewed the commonly used clustering methods in the context of scRNA-seq and summarized their advantages and limitations. Duo et al. [157] systematically evaluated 15 clustering methods using simulated data and found substantial differences in their performances. In this section, we briefly review the four main classes of clustering methods.

K-means clustering iteratively assigns cells to the nearest cluster center and recomputes the new center. Although this method is fast, it requires a predetermined number of clusters and assumes the clusters are of equal sizes, which can be easily violated in TME studies. Tools that implement K-means include SC3 [158], SIMLR [159], RaceID [160], and pcaReduce [161]. The next method is hierarchical clustering which sequentially merges cells into larger clusters (bottom-up) or divides clusters into smaller communities (top-down). This method is deterministic but more computational expensive than k-means. Many scRNA-seq tools have modified hierarchical clustering either to accommodate low-depth samples by adding imputation of zeros [162] or to improve identification of small clusters by iteratively performing dimension reduction [161, 163]. Hierarchical clustering-based tools include CIDR [162], BackSPIN [163], pcaReduce [161], SINCERA [164], mpath [165], and ascend [166].

The third type of clustering methods is density-based, which can identify dense clusters without any assumption on the shape or size on the clusters. However, it assumes equal homozygosity (density) of the clusters, requires a predetermined density cutoff, and works better with datasets with a large number of cells (e.g., droplet-based scRNA-seq assays). Such methods include DBSCAN [167], GiniClust [168] which employs DBSCAN, and monocle [169]. Lastly, there is graph-based clustering, which identifies cells densely connected by edges. Compared to k-means and hierarchical clustering, graph-based clustering does not require any predefined parameters, makes the minimal assumption on cell populations, and can scale to a large number of cells [170, 171]. The most population application of graph-based clustering combines k-nearest-neighbor graphs and Louvain community detection [171, 172]. However, the main drawback of graph-based clustering is that it heavily relies on how well the scRNA-seq is translated into graph space. Therefore, it is often necessary to perform dimension reduction or feature selection beforehand to boost the search for nearest neighbors. Graph-based clustering has been implemented in multiple tools including Seurat [97], Pheno-Graph [173], densityCut [174], SNN-Clip [175], SACNPY [176], and MetaCell [177].

4.4.3. Supervised Classifier and Cell Type Identification

To characterize the sample heterogeneity, the groups of cells defined by unsupervised clustering are often assigned to different cell types based on enriched canonical markers. In cancer researches, the single-cell type identification has been focusing on the subpopulations of immune cells, stromal cells, and tumor-related cells present in TME and/or the circulating system. However, this canonical workflow has its limitations. First, as summarized above, each clustering method has its own drawbacks and different similarity metrics usually result in different cluster separations. Second, this process relies on the researchers’ knowledge on the signature genes, and it can become arbitrary when making conclusions based on only a handful of genes. Only very recently, alternative methods have been proposed. SuperCT [178] is a supervised classifier (SC)-based machine learning method. It trains a nonlinear SC from bulk and single-cell RNA-seq data of pure cell types covering a wide range of immune and stromal cells, and then uses the trained classifiers to reveal cell types of any scRNA-seq data provided as new input. Another single-cell identifier is SingleR [179]. It constructs a reference database by collecting bulk RNA-seq data from over 1000 samples with pure cell types, and then determines the type of a single cell in scRNA-seq experiment by its Spearman correlation with each sample in the reference database. Although these methods are still immature for application in cancer research due to the limited cancer-specific reference database, they defiantly opened up new avenues for cell type classification in scRNA-seq.

4.5. Studying Heterogeneity Using scRNA-Seq

Tumor heterogeneity is commonly observed with wide range of infiltrations, as illustrated in Fig. 7 for 19 melanoma tumors from the Tirosh et al. study [20]. As the algorithms for cell type or subtype classification has been developed and improved in recent years. Some algorithms have been focused on how to quantify cellular heterogeneity. SinCHet estimates cellular heterogeneity using Shannon index over the all-possible clustering resolutions is developed to analyze cellular heterogeneity and characterize subpopulation composition [180]. A recent paper proposes a general diversity index (GDI), a generalized form of ecological diversity index, to quantify heterogeneity on multiple scales and relate it to disease evolution [181]. The index takes the generalized from the low diversity order, the clonal richness, to intermediate diversity, Shannon or Simpson’s indices, to higher order of diversity. The results showed that healthy individuals had lower diversity than cancer patients and little difference in diversity between pre– and post–bone marrow samples from AML patients.

Fig. 7.

Cellular composition of 19 melanoma tumors showing tumor heterogeneity

5. Conclusions

High-throughput sequencing (HTS) has revolutionized the study of the transcriptome and its relationship with disease. Two types of transcriptomic studies are now possible, bulk or single cell sequencing studies. With the advances in technology, many bioinformatics and statistical methods have been developed to process and analyses data from bulk sequencing (RNA-seq) and single-cell sequencing (scRNA-seq), with more methods currently being developed for scRNA-seq studies.