Single Cell RNA Sequencing in Atherosclerosis Research

Abstract

Technological advances in characterizing molecular heterogeneity at the single cell level have ushered in a deeper understanding of the biological diversity of cells present in tissues including atherosclerotic plaques. New subsets of cells have been discovered among cell types previously considered homogenous. The commercial availability of systems to obtain transcriptomes and matching surface phenotypes from thousands of single cells is rapidly changing our understanding of cell types and lineage identity. Emerging methods to infer cellular functions are beginning to shed new light on the interplay of components involved in multifaceted disease responses, like atherosclerosis. Here, we provide a technical guide for design, implementation, assembly, and interpretations of current single cell transcriptomics approaches from the perspective of employing these tools for advancing cardiovascular disease research.

[1]. INTRODUCTION

Atherosclerosis is a major cause for many cardiovascular disease pathologies, such as coronary artery disease (CAD) leading to myocardial infarctions (MI), cerebrovascular disease (CVD) leading to stroke, and peripheral artery disease (PAD) leading to amputated limbs. Together, MI and stroke are the most common cause of death, ahead of cancer, accounting for 31% or deaths in the US and a similar percentage worldwide¹. Atherosclerosis manifests as cell- and lipid-rich plaques in mid- and large-sized arteries. After initiation, infiltrating leukocytes promote vascular inflammation, plaque growth, calcification, instability, plaque rupture or erosion, ultimately resulting in catastrophic arterial thrombosis that fully or partially occludes the affected artery and leads to clinical manifestations^{2, 3}.

It is now clear that inflammation is a key factor in human atherosclerosis. In the CANTOS clinical trial, more than 10,000 subjects with prior MI and high serum C-reactive protein were treated with the anti-IL-1β monoclonal antibody (mAb) canakinumab or placebo⁴. Canakinumab treatment significantly reduced the incidence of major adverse cardiovascular events (MACE). In the COLCOT trial, a cohort of patients with previous MI receiving low-dose colchicine as a broad anti-inflammatory treatment were protected against recurring ischemic cardiovascular events⁵. Inflammation is associated with the infiltration of various immune cells, most prominently macrophages, T cells, natural killer (NK) cells and B cells. Understanding the diversity of the cellular components of atherosclerotic plaques and unraveling the immune response in atherosclerosis holds the promise of enabling novel therapeutic strategies.

Traditionally, the composition of the immune cell infiltrate in atherosclerotic plaques was assessed by immunostaining⁶ or flow cytometry (FACS)^{7, 8}. The number of parameters that can be studied simultaneously is limited to about 3 to 8 for immunostaining⁹ and about 8 to 16 for flow cytometry. New high-parametric flow cytometry allows to measure up to 28 fluorescent parameters¹⁰ and will soon increase to 40 (Cytek Aurora). Implementation of new approaches for expanded multi-dimensional analysis has expanded our understanding of cellular diversity present within atherosclerotic plaques. An intermediate step in broadening the number of antigens that could be assessed on a single cell was the development of mass cytometry¹¹, which allows assessment of ~35–42 surface markers¹² including intracellular cytokines or transcription factors¹³, but provides no information on mRNAs.

The entire repertoire of all mRNAs present in a cell is called the transcriptome. This review will provide a description of the current single-cell technologies for assessing transcriptomes and cell surface markers and outline emerging technologies that will continue to propel the field forward.

Primer to single cell approaches

Development of next-generation sequencing (NGS) approaches, along with the advent of microfluidics devices suitable for single cell encapsulation¹⁴ and advanced multi-well plates¹⁵, have enabled the interrogation of single cells for transcriptome analysis. NGS refers to DNA sequencing approaches that utilize a fluorescence- or ion-based multi-strand analysis approach, allowing for simultaneous measurements (sometimes referred to as “shotgun” sequencing) across the entire transcriptome at once¹⁶. Compared to traditional Sanger sequencing, NGS has dramatically reduced the need for large quantities of DNA template, increased read reliability and reduced cost and time associated with RNA sequencing. Complementing NGS technology was the development of microfluidics devices with capillary-like constraints able to enrich molecules or cells of a desired size, weight, or charge within a closed system¹⁷. Microfluidics platforms are widely used across many fields allowing for flow of reagents across a cell or substrate. They dramatically reduce sample and reagents needed for experiments. Single cell RNA sequencing (scRNA-seq) technologies typically use microfluidics platforms for separating cells of interest and performing the initial steps of sample preparation. Alternatively, multi-well plate systems are used for the same purpose with 250,000 (BD Rhapsody) to 1 million (Celsee) wells per plate. Together, NGS and modern microfluidics or multi-well approaches have enabled the single cell revolution and unveiled a new layer of complexity within biological systems.

Commercial scRNA-seq approaches have only been available since 2017 (Table 1), but experimental single cell transcriptomics have been attempted since the early 1990s^{18, 19}. Early efforts were restricted by low cell numbers, as well as the number of transcripts that could be examined in a given experiment. Using modern approaches, transcriptome coverage with high sensitivity across tens of thousands of cells is readily achievable. The ability to identify rare cell populations and the implications of our new understanding of population heterogeneity have been immediately scientifically impactful. scRNA-seq approaches were used to identify the precursors of plasmacytoid dendritic cells (DC) in the bone marrow and for the first time unraveled detailed human thymus organogenesis and early T cell development ^{20, 21}. Transcriptional signatures of murine aortic T cells or tumor-associated macrophages obtained from scRNA-seq were used to predict outcome in patients with recurrent ischemic events or prognosis in breast cancer patients, respectively^{12, 22}. Combining transcriptomics with assessment of cell surface phenotype using oligonucleotide-labeled mAbs²³ has allowed for the development of better identification markers for discrimination between cell populations and identified many new cell types. As an example, scRNA-seq was recently used to develop lineage-tracing networks within the zebrafish brain. These approaches allow for the identification of novel cell types and their cellular origins during embryonic development^{24, 25}.

Table 1.

Comparison of scRNA-seq methodologies.

Vendor	Method	# of Cells Loaded1	Wet Lab Duration hrs	First Break hrs	Bioinformatics Pipeline	Runtime	Cost per Cell2	CITE-seq	WTA7 or Targeted	# of Genes	Sequencing Depth Required3 per cell
10X Genomics	3’ v3	16,000	22	4	Cell Ranger	1 Week	$0.684	Yes	WTA	1,600	50,000
	5’	17,400	24	4	Cell Ranger	1 Week	$0.305	Yes	WTA	1,200	50,000
BD Biosciences	Rhapsody	20,000	16	6	Seven Bridges Genomics	1 Day	$0.364	Yes	Targeted	500	20,000
	Rhapsody (WTA)7	10,000	16	6	Seven Bridges Genomics	1 Day	$0.754	Yes	WTA	800	50,000
Takara/Illumina	SMART-seq	96	22	7	Open source	1 Week	$80	No	WTA	8,0006	2 million

¹Based on manufacturer’s maximum loading capacity (input).

²Cost excludes tax, shipping, labor, and sequencing.

³Based on manufacturer’s maximum recommendations for transcriptome only.

⁴Based on 50-plex oligonucleotide antibodies used.

⁵Based on 20-plex oligonucleotide antibodies used.

⁶Referencing⁶²

⁷Whole transcriptome analysis

Initial microarray approaches were utilized to determine bulk transcriptomes on tissues such as atherosclerotic plaques or mixtures of cells like peripheral blood mononuclear cells (PBMCs)²⁶. Deeper transcriptomes of FACS-sorted cell populations such as aortic macrophages²⁷ or newly identified atherosclerosis-relevant Treg subsets²⁸ were obtained by bulk RNA-seq of sorted cells. Understanding gene expression derived from whole tissues or PBMCs is a complicated task, because measured gene expression is impacted by cellular composition. It was not until scRNA-seq data became available that the averaging problem of bulk RNA-seq became obvious. Even within one cell type, the transcriptome varies with activation, cell cycle, apoptosis, stress, or time of day (circadian)²⁹. Deconvolution methods, such as Cibersort³⁰, allow to dissect the cellular components found in a bulk transcriptome, but only for known cell types with known transcriptomes. Complete deconvolution methods of bulk RNA-seq data sets are being developed without the need for prior tissue composition knowledge. Most of such algorithms assume that a dataset obeys a linear mixing model, which refers to the linear relationship between the proportion of a cell-type specific gene in a bulk transcriptome to the abundance of this particular cell-type in the probed complex tissue. Multiple cell-type specific genes can show mutual expression linearity or collinearity, which is used to determine the varying cell types in bulk transcriptomes without the knowledge of its composition. For real-life datasets, an algorithm would need large enough input data to robustly determine the linear component of variability across multiple samples to faithfully determine the cellular composition e.g. in healthy and diseased tissue³¹. This requires a sizeable sample cohort and is limited by low abundance of a given cell type in a complex tissue. scRNA-seq allows us to identify the gene signatures required for accurate gene expression deconvolution within bulk RNA-seq datasets, so recently generated scRNA-seq datasets can be used to reanalyze bulk RNA-seq from previous studies^{32, 33}. Using scRNA-seq will continue to help build our understanding of the relationship between cells within a given group and allow to predict differentiation trajectories within the bounds of a single snapshot in time.

Recently, several studies have introduced mass cytometry and scRNA-seq technologies into the cardiovascular field to study a diverse array of cell types, including stromal, endothelial, and immune cells^34–42. In addition, fate-mapping approaches that insert artificial genes can also be used for tracking newly recruited cells or to map the differentiation of a heterogeneous population throughout disease kinetics⁴². Unbiased dimensionality reduction and clustering algorithms have resulted in the discovery of new cell types and better resolution of cell subsets. The results of such studies pave the way for identifying novel pathways for the regulation of atherosclerosis. Here, we review existing studies and aim to provide a guide for scRNA-seq experimental design, library preparation, processing, and bioinformatics approaches to developing reliable and interpretable results.

[2]. SAMPLE PREPARATION

Obtaining a viable population of target cells in suspension and minimizing population bias or alterations in gene expression (Box: The promise of CITE-seq (or AB-seq) combined with scRNA-seq) is the first step in performing effective scRNA-seq. Cell liberation from dense tissue using enzymatic and mechanical isolation methods can often be a primary culprit in the introduction of variability within transcriptome analysis approaches. The investigator must determine whether tissue processing, digestion, or even cellular sorting are necessary, as each additional step introduces possible errors into the final dataset⁴³.

Box. The promise of CITE-seq (or AB-seq) combined with scRNA-seq.

Discovery	FACS panels are focused on known cell types. Unbiased approaches such as CITE-seq with 50 antibodies yield 1225 dot plots, many of which have never been seen before. This will lead to the discovery of new cell types in blood and tissues of humans and model animals in health and disease. The surface phenotype is superior to transcriptomes for cell identification, because it takes advantage of knowledge gained in 30 years of flow cytometry.
Complexity	Classical flow cytometry allows for detection of 16 markers. New analyzers promise up to 50 markers (Cytek Aurora, BD Symphony). Mass cytometry resolves up to 40 markers. Due to uniquely barcoded antibodies, CITE-seq is virtually unlimited. 100-plex panels are on the horizon.
Uncovering true heterogeneity	Machine learning algorithms perform dimensionality reduction (UMAP, tSNE) and group cells into clusters of similar surface antibody expression (Louvain). In a second analysis step, the single cell transcriptomes refine the clusters.
TCR, BCR	T and B cell receptor sequences can be assembled, for example using the 5’ solution by 10x Genomics, and combined with transcriptomes and surface phenotype.
Leveraging transcriptomes	Pathway analysis allows to determine the functionality of a given cell cluster, including activation, proliferation and apoptosis status.
Developmental cues	Single cell transcriptomes can be subjected to algorithms such as Monocle and RNAvelocity, which render data as pseudotime plots based on the expression status of a cell. This allows to infer developmental trajectories and lineage branching of cells in a complex environment.
Limitations	All scRNA-seq approaches require enzymatic and mechanical tissue dissociation, which induces artifacts. CITE-seq or AB-seq are not compatible with intracellular staining. The assays for CITE-seq or AB-seq are more elaborate and require a higher technical skill compared to sample preparation for flow cytometry. The workflooxw requires accessibility to multiple instruments including a FACS sorter, single cell platforms such as the 10x Chromium Controller and a sequencer. The biggest bottleneck today is the bioinformatic analysis of the generated data.

For cell isolations from suspensions like blood, cells merely require a brief wash before being enrolled into a desired scRNA-seq platform. In some soft tissues like spleen or lymph nodes, lymphocytes can be isolated from single cell suspensions generated by only mechanical means (typically by a 70 μm nylon mesh) without enzymatic digestion. However, for the remainder of cells within tissues, typically some form of processing is required to liberate cells for single cell suspensions. Optimized protocols for liberating cells from mouse aortas^{8, 44} have been published and extensively tested. Enzymatic digestion with blends of collagenases and DNAses are the most common approach, particularly for cardiovascular disease-related tissues. For each tissue, the conditions (enzymes, concentrations, time, temperature, ion composition) need to be optimized and validated. Cold digestion protocols taking advantage of proteases with activity at 4–6°C are effective for tumor tissue dissection and revealed activation of stress response genes as a result of collagenase digestion at 37°C⁴⁵. This stress response was particularly elevated when digestion persisted beyond 1 hour. Few genes have been associated with a cold-protease digest approach⁴⁶. An alternative for cold-protease includes the use of a transcriptional inhibitor, such as actinomycin D, to prevent stress response elements from being activated during the isolation procedure⁴⁷.

A reasonable validation step is using the enzyme cocktail on blood cells and comparing the phenotype of the cells exposed to the enzymes with that of untreated cells⁸. Protocols for large arteries and heart biopsies (unpublished) are being developed. Whether samples need to be digested in an enzyme cocktail and whether digestion must be performed at elevated temperatures (like 37°C), where RNA expression levels may be modified more rapidly and more profoundly, should be investigated on a case-by-case basis due to the potential to develop biased or non-reproducible data⁴⁸. A recent study showed that a two-hour collagenase treatment of muscles to isolate muscle-resident satellite cells induced expression of immediate early genes including Fos, Jun, Socs3, and diverse heat shock proteins, marking a cellular stress response⁴⁸. These genes were not detectable in cells derived from a one-hour collagenase digest. Thus, published data sets need to be carefully evaluated for the tissue dissociation protocols applied and might need to be reinterpreted in view of cell stress. A further source which may lead to decreased quality and yield of cells isolated from a complex tissue are lengthy manual processing times at the bench, such as removing perivascular adipose tissue or cutting samples for rapid digestion. Commercially available platforms including the Miltenyi gentleMACS allow for a more efficient and standardized tissue dissociation and homogenization.

In addition to potential processing artifacts from mechanical dissociation and enzymatic digestion, any given protocol may result in a misrepresentation of the actual cellular diversity present in the tissue. This is due to intrinsic predisposition of the preparation method to target some cell lineages over others. Often a historically established protocol was designed for the study of a specific cell type of interest, without consideration for other non-target cells that may be left to succumb to cell damage or death. An example would be that in our previous experiments to retrieve cholesterol-loaded foamy macrophages from within atherosclerotic plaque in the artery, an aggressive digest of the tissue with a relatively complex cocktail of enzymes was necessary³⁶. This approach led to efficient recovery of the cells of interest but left other cells, particularly arterial endothelial and smooth muscle cells, almost uniformly unrecoverable. Using this same digestion approach for examining stromal cells within the arterial wall would result in poor quality transcriptomes that are difficult to interpret. With this possibility understood, the fraction of cells within a given sample may therefore be over- or under-represented. Validation may involve cell imaging in the tissue environment, but microscopy is by nature low-dimensional (few fluorochrome channels). Cell proportions obtained by scRNA-seq can be corrected by deconvolving bulk transcriptomes^{34, 49}. When digesting tissues that will be subsequently stained for surface antigens, for cellular indexing of transcriptome and epitopes (CITE)-seq or FACS sorting, it is important to test for the loss of surface markers due to cleavage or shedding. For example, L-selectin (CD62L) and CD4 are lost after enzymatic treatment, but CD4 will recover after short-term (30 min) culture.

scRNA-seq approaches require a high viability (ideally >95%) of cellular input. Dead cells can be removed by bead-based approaches or FACS sorting, which is an approach favored by immunologists to isolate leukocytes or other cells of interest from tissues and organs. FACS sorting can also be used to narrow down the cell populations to be studied by scRNA-seq or bulk RNA-seq. Single cell suspensions are antibody-labeled on ice, undergo multiple buffer washes, and are pulled through a sorter. Narrowing down populations by pre-sorting can improve cell type resolution, but comes at the expense of potentially missing new cell types.

Different choices can be made for cells to be analyzed by scRNA-seq approaches. Broad populations from complex tissues have been chosen, for example arterial cells with no enrichment⁴¹, total hematopoietic cells expressing CD45-antigen from atherosclerotic aortas^{34, 35} or specific immune cell subpopulations such as regulatory T cells obtained from transgenic fluorescent reporter mice⁵⁰. Advantages of using FACS sorting include the removal of dead cells and the ability to condense a viable, but relatively rare population or group of populations into a single sample. For example, purifying heart leukocytes is a laborious process and provides very few hematopoietic cells. Without sorting, there may not be sufficient cells to identify the diversity of macrophage populations within the tissue³⁷. However, high pressure, shear stress and osmotic changes during the FACS sorting process can lead to cellular stress and altered gene expression⁵¹. To circumvent this problem, it is it is advisable to block the cells’ ability to produce new mRNAs by adding RNA polymerase inhibitors^{52, 53}. Also, newer generations of low pressure-based cell sorters such as the MACSQuant Tyto may reduce these problems. In addition, faster and less invasive methods for cell enrichment include magnetic bead isolation approaches where cells of interest are negatively-selected by using antibody-magnetic bead labeling approaches to enrich for cells of interest by removing other cells from a mixed population. This procedure can be performed completely on ice in less than one hour and with low-budget equipment, making it a suitable replacement to sorting, which is time intensive and requires well-maintained and expensive equipment. Positive bead isolation strategies for cells of interest should be avoided as they can activate cells and lead to stress response signatures⁵⁴. Approaches are being employed to limit cellular changes by shortening digestion periods, keeping cells on ice during digests, or even supplementing buffers with RNA polymerase inhibitors. Optimizing many of these methods appears to be ongoing in many laboratories and standards have not been established yet.

In some scenarios it is not feasible to isolate viable cells from tissues without substantial damage. For some clinical samples that are isolated following very long procedures like transplantation, were previously frozen or collected post mortem (such as brain aneurysm samples), it may be advantageous to take on an alternative isolation approach known as single-nuclei RNAseq (snRNA-seq)⁵⁵. This approach takes advantage of the resistance of nuclei to degradation during freeze-thaw cycles. A typical protocol and work flow for snRNA-seq is highlighted here⁵⁶. However, the quantity of mRNA available in the nucleus is limited. More unspliced mRNA may be recovered. Additionally, snRNA-seq is not compatible with CITE-seq, so no simultaneous protein expression information can be obtained. However, comparative analysis has shown that snRNA-seq faithfully replicates overall transcriptomes of whole-cell lysates^{57, 58}.

[3]. scRNA-seq METHODOLOGY

scRNA-seq refers to a broad class of techniques and protocols that each contain their own inherent strengths and weaknesses. Depending on the readout being sought or the availability of resources, certain methods may be more desirable than others. The primary difference relates to the ability to process large (tens of thousands) or small (hundreds) numbers of cells. Higher cell numbers come at the expense of lower read depth within individual cells and impaired ability to decipher mRNA details such as splice variants or expression isoforms. However, all approaches share some universal steps. First is the ability to separate individual cells (Figure 1), either through microfluidic chambers, nano-droplet formation or plate-based systems. Generating the library entails lysing the cells, synthesizing the RNA into cDNA and amplifying the cDNA. Due to mRNA degradation after cell lysis and an inefficient reverse transcription reaction, only 10–20% of total transcripts are synthesized into cDNA^{14, 59, 60}. Because amplification is not uniform, the number of reads mapping to a given transcript does not indicate the amount of transcript in the sample. This problem was solved by introducing unique molecular identifiers (UMIs)⁶¹. Counting UMIs is closely related to gene expression^{62, 63}. Initially, the commercially available platform for scRNA-seq by 10x Genomics allowed to capture below 10% of a single cell’s transcripts¹⁴. This was significantly improved to 30–32% by new reagents and chemistry according to the manufacturer. All samples are processed for NGS. Illumina’s HiSeq or NovaSeq are the most commonly used platforms. In this section, we will reference three of the most popular approaches to get from single cells to a sequencing library. More approaches are reviewed here⁶⁴. Single cell technologies continue to evolve and change.

Figure 1.

Experimental workflow for CITE-seq and AB-seq experiments.

Single cell suspensions derived from artery wall (top) or PBMCs (bottom) are sorted for live cells and other desired markers (here: CD45 for leukocytes), hash-tagged, washed and incubated with the oligonucleotide-marked antibody panel (currently 50 mAbs are possible). After 3 more washes, the cells are loaded into the BD Rhapsody scanner or the 10x Chromium controller. Beads are retrieved and processed for library preparation. Pre-sequencing quality control (QC) is done by TapeStation. DNA fragments are removed, and the library is sequenced using NGS sequencers like the Illumina NovaSeq. The figure was created with BioRender.

Drop-seq

Droplet-based sequencing approaches, which include the popular 10X Genomics Chromium¹⁴ and Drop-seq^{65, 66} approaches and others take advantage of nanoliter droplet formation through microfluidics cartridges to place individual cells into an oil-based medium to separate cells for micro-reactions. These approaches are currently the most high-throughput with tens of thousands of cells being assessed in a single run (Table 1). They utilize UMIs that are associated with individual transcripts. UMIs are molecular barcodes that are attached on the surface of a bead to identify each transcript of a cell. Once cells are lysed, sample libraries can be combined after multiplexing. Multiplexing and sequencing multiple samples in one run reduces batch effects. Sequencing is limited to the cell barcode, multiplexing tags, UMI tags, and read-length size portions of transcripts that can be read. This limits the ability to detect many splice variants within Drop-seq approaches when compared to full-transcript sequencing approaches like SMART-seq. This makes it impossible to detect splice variants. Drop-seq provides high throughput, but low read depth. It is usually sufficient to obtain 40,000–50,000 reads per cell. This approach is useful for identifying rare cell populations and to map cellular diversity within whole organs.

SMART-seq

SMART-seq can be paired with index sorting or microfluidics platforms like the Fluidigm C1. Index sorting is a FACS-based approach of separating single cells into multi-well plates for subsequent single cell analysis. Fluidigm C1 is a fully automated system for separation of single cells that includes an automated platform for cell lysis, cDNA synthesis, and tagging procedures. The Fluidigm platform is suitable for analysis of surface protein expression by imaging, genomic DNA, epigenetic, or micro RNA abundance all within the microfluidics chamber. Whereas the first C1 generation allowed for the assessment of up to 96 cells, the C1 high-throughput integrated fluidics circuits platform allows to analyze mRNA of up to 800 individual cells The distribution of cells in the microfluidic chambers is gentler than flow cytometry sorting and may influence transcriptomes less. Optical assessment of the microchamber after cellular loading provides unparalleled confidence in the number of cells that are subjected to single cell RNA sequencing protocols. However, this approach comes at the cost of a high cellular input for loading.

Following separation of individual cells using cell sorting or microfluidics platforms, cells are lysed and RNA is hybridized to an oligo-dT containing primer after first strand cDNA synthesis to act as a barcode⁶⁷. Current versions of this approach, termed SMART-seq2, have optimized the nucleic acid linking approach to improve kit performance^{68, 69}. Full cDNA transcripts are then amplified and sequenced. This approach requires read depths beyond 1,000,000 reads per cell and can detect isoform variants and even single nucleotide polymorphisms (SNPs) in transcripts. Full sequence analysis improves mRNA sensitivity and provides great depth within individual cells. This comes at the cost of being low-throughput. The cost is high at >$50 per cell. Sophisticated new approaches based of SMART-seq2 protocols have been used to determine cellular heterogeneity in the developing mouse brain and an entire nematode using Split Pool Ligation-based Transcriptome sequencing (SPLiT-seq)⁷⁰ and single-cell combinatorial indexing RNA sequencing (sci-seq)⁷¹, respectively. Both depend on multiple rounds of combinatorial indexing with oligonucleotide barcodes leading to uniquely barcoded individual cells and transcriptomes. For this, 100–1000 cells are sorted into 96- or 384 well-plates and kept intact. An in situ reverse transcription with barcoded primers will be performed, all cells will be collected, and randomly distributed into wells of another plate. This is followed by 2^nd strand synthesis, tagmentation and a PCR to introduce a 2^nd barcode. This technique allowed for the parallel analysis of 150,000 cells and drastically reduced cost per cell. Generation of SMART-seq3 is on the horizon; it detects transcripts in an allele-specific way and at increased sensitivity, which may lead to the detection of thousands of more genes per cell than are currently available with commercial products⁷².

BD Rhapsody

The BD Rhapsody approach allows for the targeted detection of a few hundred known targets. Several gene panels are available. Rhapsody performs paired barcoding in microwells with magnetic beads. Beads and attached mRNA are retrieved and synthesized into cDNA. cDNA is then amplified using targeted primers to specific genes of interest and labeled with library index barcodes. Samples are then sequenced and target gene expression is assessed. This approach leads to the ability to run more cells at a lower cost than other approaches. Rhapsody is compatible with the detection of surface antigens (by AB-seq) and hence provides protein information. A newly developed Rhapsody approach is designed to provide full (not targeted) transcriptomes.

CITE-seq, integrating transcriptome and proteomics

CITE-seq detects surface antigen abundance in tandem with scRNA-seq through the use of oligo-barcodes attached to antibody conjugates (instead of traditional flow cytometry or CyTOF, which use fluorochromes or rare metals, respectively). An analogous approach, REAP-seq (RNA expression and protein sequencing), has also been developed using an alternative method for conjugating oligonucleotide probes to antibodies⁷³. Oligonucleotide barcodes can then be interrogated for enrichment within populations of cells to assist in identifying populations of interest. A typical workflow is shown in Figures 1 and 2. Matched transcriptome and cell surface phenotype information is retrieved, which allows for a more fine-grained analysis of cells. A first report in 2017 provided information of single cell transcriptomes and simultaneous assessment of 17 surfaces markers²³.

Figure 2.

Bioinformatic analysis workflow of CITE-seq and AB-seq experiments

The output of the sequencer is a FASTQ file with information about read quality. Further post-sequencing QC steps are indicated. If batch effects are detected, they can be removed if the experiment was designed to do so. Missing transcripts may be imputed. Dimensionality reduction (here by UMAP) and clustering (here by Louvain) result in 2 D maps of cells segregated by surface markers. Each antibody and transcript can be displayed as a heat map projected onto the UMAP. Differential gene expression is determined between clusters, or between samples from healthy and diseased individuals, or one cluster against all other cells. From the gene signature, pathway analysis is performed. The figure was created with BioRender.

The BioLegend brand of CITE-seq, called Total-seq, uses poly(dT) oligos and Drop-seq approaches utilizing 3’ or 5’ sequencing. Fernandez et al. recently used a 21-surface marker CITE-seq panel to describe leukocyte heterogeneity in human atherosclerotic plaques^{39, 74}. Antibody numbers are limited only by the number of oligo barcodes developed by manufacturers, with custom options available. Therefore, expect antibody CITE-seq panels to dramatically expand in the near future.

[4]. PRACTICAL CONSIDERATIONS

scRNA-seq costs have dropped dramatically since 2017, but remain a limitation for researchers to access the technology. scRNA-seq is 10–200 times more expensive than bulk RNAseq approaches. Multiplexing scRNA-seq samples slightly reduced experimental costs, but most importantly helped to normalize between samples and control batch effects. These approaches utilize molecular barcoding at the sample level in addition to the cellular level. This allows for multiple samples to be combined and run on the same lanes, which becomes particularly important when collecting clinical samples with limited sample size, or when large numbers of samples need to be assessed. An intriguing new approach suitable for scRNA-seq or snRNA-seq approaches localized DNA-barcodes to lipid anchors on membranes termed MULTI-seq (Multiplexing using lipid-tagged indices) may be particularly helpful for sample barcoding approaches and discrimination of cell viability and endogenous gene expression patterns⁷⁵. Antibody mediated multiplexing approaches have also been developed and modified for use with snRNA-seq analysis⁷⁶.

Limitations of scRNA-seq include the limited depth of reads that can be detected per cell, often 10 times lower than what is obtained in bulk RNA-seq of sorted cells. Most bulk RNA-seq experiments are sequenced at millions of reads per sample. Shallow sequencing approaches are thought to underrepresent the transcriptional information. A greater sequencing depth increases the number of detectable transcripts. However, this is not linear and only marginally scalable, which means that sequencing depth needs to be increased tremendously to obtain information of additional transcripts. At a certain sequencing depth sequencing saturation occurs, which means that no additional unique transcripts will be detected.

Saturation is a function of both library complexity and sequencing depth. RNA content varies between cell types and their activation status, which will be represented by different numbers of transcripts in a library, called the complexity. The library complexity limits detection of transcripts even with increasing sequencing depths. As increasing sequencing depth comes with higher experimental costs, alternative strategies might be favored. A high-throughput scRNA-seq experiment can be used to uncover cellular heterogeneity in a complex tissue under homeostatic or pathologic conditions. Once an unusual cell subpopulations or transcript combinations within a cell population has been discovered, cell surface marker of this population can be identified. Subsequently, this population can be FACS-isolated and subjected to deep bulk RNA-seq to obtain a better picture. This was recently implemented for vascular macrophages in atherosclerotic arteries²³.

In addition to sequencing depth, intrinsic bias between approaches has been reported. Specifically, full-length sequencing libraries are associated with increased sensitivity for longer mRNA fragments. By contrast, droplet approaches using UMI tagging acquire 3’ or 5’ ends for analysis, thus avoiding length biasing⁷⁷.

[5]. POST-SEQUENCING PROCESSING

Post-sequencing processing for large data sets requires significant computational resources and experienced specialists, which remain limiting factors for many laboratories performing scRNA-seq (Figure 2). In droplet-based technologies, after paired-end sequencing is performed, two FASTQ files are obtained. The first contains cellular and UMI barcodes, which identifies the cellular source, the second contains sequence and quality scores of the sequenced cDNA. Subsequently, reads are mapped to the reference genome/transcriptome of the target species^{78, 79}. Pseudoaligners instead of classical aligners can speed up the alignment process⁸⁰. Droplet-based technologies, after alignment and quantification, yield a large matrix of detected gene expression in each cellular barcode.

A substantial proportion of droplets contain a bead but no cell. Other droplets contain one bead and one cell (desired). Yet other droplets contain one bead and two or more cells (multiplets). All multiplets must be identified and removed to prevent flawed results. Multiplets, unless removed, will look like unique cell types. There are three fundamental ways to remove multiplets: based on SNPs⁸¹, cell transcriptomes⁸², or hashtags⁸³. Several methods to identify multiplets have been described⁸⁴. Dead cells should also be removed. Overrepresentation of mitochondrial transcripts is commonly used to identify dead cells, but there is no consensus on the cutoff.

Regardless of the scRNA-seq strategy, the filtered count matrix must be normalized, scaled, and subjected to dimensionality reduction. Since scRNA-seq data contain extensive and highly variable data, visualization requires dimensionality reduction approaches for removal of noise and simplification of variation between samples. The variability between tens or hundreds of genes needs dimensionality reduction for an understandable display. Using algorithms such as tSNE, UMAP, or PCA, complex data can be generated into 2-dimensional models for simplified visualization. Additional downstream analyses include differential gene expression, followed by pathway/gene set enrichment, trajectory analysis, and others⁸⁵.

Batch effects refer to preparation-dependent changes in how cells were isolated, RNA was processed, or any other unwanted experiment-to-experiment variations. Batch effects make it difficult to compare transcriptomes collected from different laboratories or at different times. Common batch effects in scRNA-seq datasets include protocol variations, variations among donors, and operator variations. Depending on the severity, batch effects can be so large as to completely obscure the biological information contained in the data. Principal Component Analysis (PCA) is often used to visually inspect whether cells cluster by some experimental co-variate. Recently, Büttner et al⁸⁶ designed a test to quantify the extent of batch effects that is inherently more accurate than simple visual inspection. Although the optimal approach to prevent batch effects is following a balanced experimental design^{87, 88}, balanced designs are rarely feasible or cost-effective. Hence, several batch correction algorithms have been proposed to correct for batch effects^89–94, some of which have been initially designed for microarray and bulk RNA-seq⁸⁹, whereas others have been specifically designed for scRNA-seq^90–94. Generally, batch effect correction is applied to the normalized data before dimensionality reduction and clustering. These algorithms differ in the level at which they correct batch effects. Combat⁸⁹ and MNNCorrect⁹² operate by directly correcting gene expression in each cell, whereas Seurat⁹³ generates corrected low-dimensional reductions, similar to principal components, for clustering and visualization. BBKNN⁹⁰ generates an adjusted neighbor graph taking into account the different batches for subsequent clustering and visualization. Other batch correction algorithms are embedded within frameworks specifically designed for scRNA-seq analysis⁹⁵.

Dropouts and imputation of data

Dropouts are transcripts that are expected in a cell, but for which no reads were detected. The rate of dropouts is dramatically higher in scRNA-seq data than in bulk approaches, because the starting nucleotide concentration in a single cell is limiting. For a low-expressed gene, a transcript may be present in the cell or not at any given time. Even if present, transcripts might not be detected. This risk increases if the sequencing depth was insufficient, thus creating false negatives. The loss of “true” gene expression can lead to biased results and might necessitate the removal of cells or samples from further analysis. Conversely, similar dropouts in a subset of cells might aggregate them in the same cluster, producing false apparent cellular diversity. Multiple approaches to address this problem and fill the blanks in single cell data have been developed^96–100. Each of these methods outperformed several others depending on the benchmarks and datasets used. Fundamentally, these approaches can be used to impute all gene expression values (e.g. MAGIC⁹⁸) or only dropout genes (e.g. DrImpute⁹⁷). Another approach adopted by SCRABBLE¹⁰¹ attempts to borrow information from a matched bulk RNA-seq sample in order to constrain imputation. This strategy is considered superior and outperforms five competitive platforms. The assumption underlying imputing all genes is that dropout is a broad phenomenon that affects all genes; albeit to varying degrees. These algorithms are not part of the necessary and routine scRNA-seq analysis workflow and should only be implemented when expected genes in cell populations are missing. On a cautionary note, Andrews et al¹⁰² recently suggested that imputation may be beneficial to visualization and clustering, but may introduce false positives when used for cell type markers and differential expression. New bioinformatic workflows allow to normalize and remove technical variation in molecular counts while preserving the true biological variation¹⁰³. This procedure is independent of pseudocount addition or imputation and log transformation, thus improving the analysis of scRNA-seq data.

Data deposition

Typically, most investigators deposit their raw data on the National Center for Biotechnology Information Gene Expression Omnibus database (NCBI GEO). Raw data from NCBI GEO can be downloaded using the freely available SRA toolkit. Some groups prefer to publish their data on their own websites, which usually have added features including the ability to browse individual genes and visualize clustering, e.g. Tabula Muris¹⁰⁴ (https://tabula-muris.ds.czbiohub.org/) and the Human Cell Atlas (https://data.humancellatlas.org/). Other repositories are available, including the Single Cell Expression Atlas (https://www.ebi.ac.uk/gxa/sc) or the Pangloa database (https://panglaodb.se). When depositing scRNA-seq data, it is important to deposit at least the raw data and an accurate, detailed and complete metadata file. Although NCBI GEO provides a template for researchers to fill out the metadata information related to their data, it is up to the researcher to include as much detail as they want. This has not been standardized yet. Some GEO entries will have fastq files as raw files while others will have 10x bams files. Supplementary data is not standardized in GEO and can contain either count files, normalized expression values, h5 molecule matrices from the 10x cell ranger pipeline, or all of the above. Standardized and detailed data deposition is crucial for reproducibility, particularly due to the inherent variation observed in single cell studies. When using deposited data, it is very critical to ensure that data has been properly processed, especially when integrating multiple datasets. Processing the raw fastq data avoids many potential sources of error. The main challenges for using repositories is the ability to integrate the data in a way that retains as much biological information as possible while reducing the unwanted variation. A lot of progress has been made in developing tools for the integration of different datasets while accounting for potential batch effect and avoiding confounded experimental designs. An important consideration when integrating datasets is having the complete metadata, with as much detail as possible. Additionally, as more and more datasets are being generated, it is important to consider scaling both the hardware requirements, i.e. hard-disks and computing cores, and software requirements, i.e. tools and packages for the preprocessing and analysis of such data.

[6]. UNSUPERVISED CLUSTERING AND MACHINE LEARNING

Using unbiased computational approaches where gene expression similarities between individual cells leads to a clustering within the veil of a shared expression makeup has become the dominant approach for data analysis. The most popular algorithm is Louvain, which is a graph-based method for community detection. In-depth discussion of unsupervised clustering can be found elsewhere¹⁰⁵. By allowing algorithms to sort the data, analytical bias can be avoided. These unbiased analysis approaches have proven that unique clusters of cells found in specimens can be reproduced in different labs using different implementations of this approach. Determining the boundaries between clusters in an unbiased way is still challenging. It is up to the researchers to identify or name clusters, based on prior knowledge or systematic naming conventions. Researchers often have to choose clustering resolution: leveraging this parameter will either merge similar clusters into one bigger cluster or split a bigger cluster into smaller, transcriptionally different clusters. While choosing this parameter is not straightforward, good practice needs confirmation of transcriptionally diverse cellular clusters with protein-based methods such as FACS. Also, in our experience, clusters are not always restricted to a single cell type. This is best exemplified by proliferating cells, which share a specific and dominant transcriptional profile. Thus, they cluster together although they are of different cell types. New bioinformatic tools including scmap¹⁰⁶ and SingleR¹⁰⁷ are aiding to call cell clusters. Both methods rely on reference data. They automatically choose sets of genes that identify a given cell type. However, both algorithms are prone to misclassifications if the transcriptome originating from a particular phenotype or cell of interest is not included in the reference. Particularly, aberrant cells or cells obtained from pathological conditions can harbor a heterogenous transcriptome which is not included in the reference. In case cells are not definable by the reference, scmap leaves these cells unclassified. A different algorithm, CHETAH¹⁰⁸, also utilizes reference data, but infers classification on intermediate or unassigned cells by hierarchical clustering.

Analysis tools

Close to 500 scRNA-seq analysis packages are available by now, impossible to review in a short article. The expression of ligand-receptor pairs in single cell data obtained from complex tissue can be harnessed to predict how tissue cells interact and communicate with each other¹⁰⁹. The cellular RNA content is a great indicator of the state of the cell at the time of analysis. However, dynamic processes such as embryogenesis cannot be addressed by a static method. Approaches to overcome the static nature of scRNA-seq include utilization of nucleotide analogs that can be pulsed prior to processing, allowing for temporal resolution of newly expressed RNA to be tracked within single cells, termed scSLAM-seq¹¹⁰. RNA velocity¹¹¹ – which represents the time-resolved transcriptome state – can be predicted by analyzing unspliced and spliced mRNAs. In other words, RNA velocity allows to predict the future cell state given that unspliced mRNA has yet to be processed and expressed within the cell. Monocle^{112, 113} allows for unsupervised trajectory analysis of single cells by pseudo-time ordering. The second version of monocle⁶⁸, a machine-learning technique based on reverse graph embedding (RGE) and parsimonious principal graphs, can be used on all scRNA-seq data sets and does not require input information about cell fates or branch points. Another algorithm available in the monocle2 package is Census⁵⁹ . This algorithm converts fragments per kilobase per million reads/transcripts per million reads (FPKM/TPM) gene expression values in single cells to relative transcript counts without the need for spike-in standards or UMIs. For easier modeling of the data compared to normalized read counts, Census-obtained counts can be used to determine splice isoforms and provides allelic information which can be used to determine developmental regulation. An in-depth review of algorithms inferring trajectories of single cells can be found here¹¹⁴. Further useful analytical tools allow researchers to investigate not only the different transcriptional profiles of the cells but also their regulatory networks (SCENIC)¹¹⁵ and their potential interaction via ligand receptor relationships (CellPhoneDB¹¹⁶, NicheNet¹¹⁷). Multiple algorithms are particularly useful for studying adaptive immune responses in complex tissues. We find the most powerful tool in this regard to be TraCeR¹¹⁸, which can be used to reconstruct paired and full-length TCR chains from transcriptomes of single T cells. This allows to study the TCR V, D and J segment usage and clonality in homeostasis or under pathological conditions. It should be noted that TraCeR requires full-length transcriptomes. We have been able to assemble TCRα and β sequences from 10X Genomics 5’ sequencing.

Multi-omics approaches

The next step in single cell platform analysis includes broader multi-omics approaches like DNA methylation by bisulfite conversion sequencing or TAB-seq, non-coding RNAs, chromatin accessibility by ATAC-Seq, histone modification by CHIP-seq, and protein expression levels by CITE-seq (also called AB-seq). Integrating these data with ever-changing bioinformatics approaches enables more powerful analysis. For example, expressed quantitative trait loci (eQTLs) can help identify underlying regulatory networks associated with targets of interest.

All of today’s scRNA-seq methodologies require the preparation of single cell suspensions, which leads to complete loss of spatial information of cell types within a given tissue. There are some limited workarounds, like injecting an antibody intraveneously to label intravascular cells¹¹⁹. New approaches aim to fully retrieve spatial information. Recent publications introduced two new methodologies, seqFISH+¹²⁰ and Slide-seq¹²¹, which allow for in situ spatial identification of transcripts at near-single cell resolution across a wide array of RNA probes. 10x Genomics offers a commercially available solution called Visium which allows for whole transcriptome detection, but not at single-cell resolution. Cartana provides a product based on in situ sequencing¹²². The technology allows to detect over 100 genes within tissue sections at single cell resolution¹²³. In seqFISH+, tissue sections are incubated with transcript-specific probes containing a complementary sequence to an oligonucleotide attached to a fluorochrome. Key is the multiplexing with three different fluorochromes and multiple hybridization rounds, allowing to image multiple transcripts in every round. The images of every hybridization round are bioinformatically parsed together; detection of 10,000 transcripts in a complex tissue has been achieved. An alternative approach was used where barcoded-beads were spotted on a functionalized glass slide and sequenced subsequently to define areas of beads and establish a map of barcodes spotted on the slide¹²¹. Tissue sections were then incubated with the functionalized slide, RNA was released, hybridized with the spotted barcoded oligonucleotides, and subjected to sequencing. However, this approach does not provide true single cell resolution, because RNA from more than one cell may hybridize to each spot. Transcript detection at a resolution of 100 μm² (10×10 μm) has been achieved.

The future of single cell methodologies

We are only in year 3 of commercially available scRNA-seq. Innovation continues at a rapid speed and new products are released weekly. CITE-seq and AB-seq are only the beginning of multi-omics approaches. Spatially resolved and high-throughput strategies are on the verge of becoming practical. ScRNA-seq approaches are being applied to high-throughput chemical screens to assess hundreds of thousands of cells in individual runs, allowing for molecular insights to be gained from large scale pharmacologic perturbations¹²⁴. More than 100x cost reduction per cell has already been achieved compared to the pioneer days of sorting single cells into single wells. Sensitivity is improving, but still not satisfactory. Standard protocols for removing low quality transcriptomes, doublets and dead cells are sorely needed. Imputation and batch effect removal software programs exist, but there is no consensus in the field which ones should be used. Another important direction is the integrated and comparative analysis of single-cell data between different laboratories, donors, batches, single-cell platforms or even species. The goal of these methods is to identify common cell types/states that are present in different scRNA-seq samples while rendering dimensionality reduction and clustering robust to unwanted variation. While first techniques were only recently developed^{93, 125}, enormous progress has already been made showing the possibility to extend these techniques to single-cell protein, chromatin and spatial data¹²⁶. With more labs using scRNA-seq, the demand for unified processing strategies has increased. The biggest bottleneck in scRNA-seq today is the post-sequencing bioinformatic analysis, which is currently carried out by well-trained experts who are in high demand in academia and industry. Future strategies will aim to create user-friendly data analysis interfaces, databases and standard file formats that will streamline analysis and data sharing.

We expect methods and algorithms that will try to identify possible regulators, key transcription factors and epigenetic data. A single cell assay for transposase accessible chromatin (ATAC-Seq), a method to map open chromatin, is already available from 10x Genomics. Cell type calling and phenotypical associations based on scRNA-Seq data may use association rule mining or other rule-based machine learning techniques. These algorithms assist in generating hypotheses that can be tested in the lab.

Conclusion

Single cell RNA-seq with CITE-seq or AB-seq is a powerful methodology for cell type discovery, lineage relationships, homeostasis, development and disease. Terabytes of data are generated every day. Single cell RNA sequencing is beginning to reveal the complex interactions between cell types in atherosclerosis initiation, progression, regression and plaque rupture as well as other fields of cardiovascular research. New insights will help to tailor new and innovative therapeutic strategies for atherosclerosis beyond controlling lipid levels.

Nonstandard Abbreviations and Acronyms:

ATAC-seq

Sequencing protocol which assesses open chromatin regions

BCR

B cell receptor

CHIP-seq

Chromatin precipitation sequencing, maps transcription factor binding

CITE-seq

(also AB-seq) cellular indexing of transcriptome and epitopes, allows simultaneous assessment of antigens and transcriptomes

CyTOF

Mass cytometry

Deconvolution

Bioinformatic analysis of bulk RNA transcriptomes for cellular composition

Drop-seq

Droplet-based sequencing

Dropouts

Expected transcript for a given cell, but no read was detected

eQTL

Expressed quantitative trait loci

FACS

Flow cytometry

FASTQ

File format obtained after sequencing

FPKM

Fragments per kilobase per million reads

Imputation

Algorithms that infer the expression of a dropout

MULTI-seq

Multiplexing using lipid tagged indices-sequencing

REAP-seq

RNA expression and protein sequencing

RGE

Reversed graph embedding, used to resolve single-cell trajectories

Saturation

Sequencing depth at which no additional transcripts will be detected

Sci-seq

Single-cell combinatorial indexing RNA sequencing

scRNA-seq

Single cell RNA sequencing

scSLAM-seq

Single cell, thiol-(SH)-linked alkylation of RNA for metabolic labeling sequencing

SMART-seq

Protocol to perform deep transcriptional profiling

snRNA-seq

Single nuclei RNA sequencing

SPLiT-seq

Split pool ligation-based transcriptome sequencing

TAB-seq

Sequencing protocol which utilizes bisulfite sequencing and TET proteins to study 5-hydroxymethylcytosine DNA modifications

Tagmentation

Sequencing library preparation with the Tn5 transposase

TCR

T cell receptor

TPM

Transcripts per million reads

Trajectory analysis

Pseudotime projection obtained from single cell transcriptomes allowing to predict cell states and developmental relationships

UMI

Unique molecular identifier; molecular tag used to detect and identify unique mRNA transcripts