Single-cell RNA sequencing: an overview for the ophthalmologist

Abstract

Understanding the molecular composition of pathogenic tissues is a critical step in understanding pathophysiology of disease and designing therapeutics. First described in 2009, single cell RNA sequencing (scRNAseq) is a methodology whereby thousands of cells are simultaneously isolated into individual micro-environments that can be altered experimentally and the genome-wide RNA expression of each cell is captured. It has undergone significant technological improvement over the last decade and gained tremendous popularity. scRNAseq is an improvement over prior pooled RNA analyses which cannot identify the cellular composition and heterogeneity of a tissue of interest. This new approach offers new opportunity for new discovery, as tissue samples can now be sub-categorized into groups of cell types based on genome-wide gene expression in an unbiased fashion. As ophthalmologists, we are uniquely positioned to obtain pathologic samples from the eye for further study. scRNAseq has been applied in ophthalmology to characterize retinal tissue already, and it may offer the key to understanding various pathological processes in the future.

Introduction

Over time, medicine has shifted significantly toward targeted therapeutics. Biological macromolecules, stem cells and more recently gene therapy are all approaches that attempt to treat each disease in a manner specific to its pathogenesis¹. Such approaches have proven to be effective, and some have the benefit of minimizing off-target side effects. The purposeful design of each of these, however, required an understanding of the individual molecular processes driving disease. In order to design targeted therapeutics for the eye, one needs to understand the pathologic tissue at the molecular and cellular level. Identifying a molecular signature of each cell – one that defines its type and function – is a powerful way of characterizing a pathologic process in the eye.

To assay molecular signatures within tissue types of interest, pooling individual molecules is commonplace (gene sequencing, RNA, proteins, metabolites) because each individual molecule is difficult to accurately isolate and measure. Yet, to truly understand the cellular composition of the pathologic tissue at hand and advance existing therapeutics, one would ideally isolate different cell types to measure their constituents. For example, proliferative vitreoretinopathy (PVR), a troubling inflammatory fibrocellular process that occurs in a subset of rhegmatogenous retinal detachments, is currently lacking in terms of available medical therapies. Generic approaches, such as the use of oral and intravitreal steroids, have fallen short². In order to design targeted therapeutics, however, studying the exact cellular composition and molecular signature of PVR may prove worthwhile.

Traditionally, marker genes and proteins are used to assign cell types in a tissue sample. One of the most longstanding and reliable approaches is histology, whereby tissue samples are stained for proteins or molecules that are known to characterize particular cells. Histology is highly unique in its ability to resolve cell type and spatial structure simultaneously, and it will likely never find itself antiquated. RNA sequencing is an example of a technique that does not maintain information of tissue architecture. Yet, histology is limited in terms of its ability to narrowly characterize individual cells because usually only a handful of stains are employed. Function is one way that cell type can be distinguished while preserving architecture. This is particularly useful when studying neuronal tissue, such as the exquisite characterization of retinal layers described by Masland et al. in 2012³. However, the study of function is limited by the particular assay at hand, and only certain functional readouts can be measured.

One of the most common approaches for classifying cell type is fluorescence-activated cell sorting (FACS) which can be used to identify and sort sub-populations of cells based on size, morphology and surface proteins with the use of fluorescently conjugated antibodies. These sub-populations can be further characterized by pooled RNA sequencing or functional assays. However, marker-based approaches are inherently constrained by the availability and choice of markers and by our knowledge of how markers define cell types. Even within a seemingly narrowly defined group of cells based on cell surface markers, there is likely heterogeneity in gene expression signature⁴. More recently, mass cytometry has been employed, which involves cell characterization with antibodies labelled by heavy metal ions, and this has dramatically increased the number of proteins that can be assessed at one time by 5 to10 fold⁵. Still, it is challenging to assess the entire proteome all at once with flow cytometry.

Because measuring the entire proteome is challenging, in this review we will discuss its proxy – the transcriptome. A genome-wide transcriptional state is one way to ascribe signatures to cells. Initially, transcriptional analysis was done by assaying millions of transcripts in a tissue sample simultaneously using hybridization-based microarrays. Microarrays give a readout for the average tissue-wide level of pooled RNA for each target on the array, but they do not actually identify each individual transcript or cell. Massively parallel sequencing of cDNAs (mRNA-Seq) generates millions of short sequence fragments that are pooled and can be analyzed to accurately quantify expression at the level of the individual read across an entire tissue⁶. This approach gives a global gene signature that helps distinguish samples – for example, it has been used to predict treatment response in acute myeloid leukemia⁷. However, pooled RNA sequencing does not offer insight into which cells within the sample differ nor how the relative abundance of each cell type and its expression differs within an individual sample. The field has constantly been pushed to increase sensitivity in the setting of low amounts of RNA, ultimately to the single cell level and hence the birth of single cell RNA sequencing (scRNAseq) by Tang et al. in 2009⁸. With this approach, the RNA within each cell of a tissue sample is isolated and sequenced, allowing for genome-wide unbiased characterization of each cell’s transcriptional signature.

Since its discovery, the field of scRNAseq has undergone an explosion of innovation and improvement^4,9–19. The purpose of this review is to provide a general overview of scRNAseq for the ophthalmologist. As the technology has advanced to the point where it is relatively easy for any physician to send samples for RNA sequencing, ophthalmologists are uniquely positioned as surgeons to explore the transcriptome of ocular pathology and make great strides in our understanding of pathophysiology.

Why use single cell RNA sequencing

One of the main reasons to assay single cells is to understand transcriptional heterogeneity within a population. scRNAseq has demonstrated previously underestimated heterogeneity within embryonic, brain, liver and immune cells, among others^20–27. The findings in these studies are ones that pooled RNA sequencing would have been unable to detect. One of the first validation tissues for scRNAseq was the retina, where transcriptional profiling recapitulates our knowledge of the diversity of neuronal cell types in the retina but also offers new insight into potentially important pathways in the retinal pigment epithelium (RPE) and choroid related to disease, the details of which are discussed later^28–32.

scRNAseq is also capable of identifying rare but important subpopulations or even single cells within a larger pool. These cells would risk going undetected if looking at pooled expression. For example, single cells that underwent malignant transformation in melanoma are too rare to raise a signal in pooled sequencing but were identified when looking at the single cell level – such as subsets of cells that are unlikely to be responsive to treatment³³. Other uses of scRNAseq are in transcriptional regulation and how it differs between individual cells or how it leads to cell fate^34,35. Ultimately, its power is inherent in the fact that a far deeper understanding of individual cellular states and how they differ from or correlate with one another in the setting of disease and treatment is now possible.

How it works

Originally developed in the Surani laboratory in 2009⁸, scRNAseq has since gained tremendous popularity and has undergone dramatic improvements in sensitivity, scalability and technical noise reduction. The goal of scRNAseq is to count the number of mRNA transcripts per gene that are present within each cell, across thousands of cells. Each cell’s individual transcription profile will therefore define its type. The basic steps of scRNA sequencing are as follows:

1. Isolation of individual cells from tissue. First, intact viable cells are isolated from a tissue sample (Figure 1A). Note here that recent developments have allowed for effective (and possibly improved) RNA sequencing of nuclei that are not part of intact cells^36,37. Despite such developments, however, tissue dissection remains a crucial step that has the potential to alter the interpretation of transcriptional levels if not done correctly. Some tissues, such as immune or lymphoid cells, are easy to isolate from one another. In other tissues, such as tumor or epiretinal membrane, the cells are bound to each other by an extracellular matrix, and the processing of the tissue into individual cells can itself alter transcription or lead to degradation of mRNA. Minimal processing is therefore critical, and techniques to minimize artifacts during this step have been reported³⁸.

2. Generate single cell suspensions. Cells are then isolated into single cell partitions where they are lysed to release mRNA contents (Figure 1B). Over the last 5 years, different methods have been explored to isolate cells in a high-throughput manner, including FACS, microfluidics and droplet-based instruments^11,17. If a tissue is too delicate or scarce and will not survive this processing, manual micromanipulation is still an option, though laborious and low-throughput³⁹.

3. Amplification. Poly(T) primers are used to isolate mRNA from ribosomal RNA and the poly-A mRNAs are converted to cDNA (Figure 1B). These primers can have adapters for next generation sequencing (NGS) and unique molecular identifiers (UMIs) to signal one unique RNA molecule. The cDNA is amplified within each individual partition, either by polymerase chain reaction (PCR) or by linear amplification by in vitro transcription (IVT), to create an RNA library for that cell. Some protocols amplify full-length transcripts, which allows for the sequencing of splice isoforms, sometimes at the cost of genomic coverage. More commonly, other protocols sequence partial transcripts (3’ end) and are able to better quantify abundance and cover more genes, at the cost of losing isoform data and incurring a 3’ bias. Like the cellular isolation step, the amplification step is a point at which there is risk of distorting the readout. Because RNA cannot be directly sequenced (it needs to be converted into cDNA and amplified), and because the absolute level of RNA is very low in a single cell, an amplification bias can lead to a misunderstanding of relative expression levels. The issue of amplification bias has been helped by the use of UMIs which enable computational removal of PCR duplicates⁴⁰. Furthermore, the use of poly(T) primers will exclude small non-coding RNAs that may be clinically relevant, such as microRNA and snoRNA. These hurdles have been overcome by ligating 3’ and 5’ molecular adaptors to small RNAs and amplifying the ligated product^41,42. This is particularly important in ophthalmology, where siRNA therapeutics may be a real possibility⁴³.

4. Sequencing. The amplified and tagged cDNA is then pooled across all cells and sequenced using next generation sequencing (NGS) protocols (Figure 1B). Note that multiple samples can be combined into one experiment by using genetic variation (single nucleotide polymorphisms or haplotypes) to later segregate the samples.

Figure 1. Schematic of single-cell RNA sequencing pipeline.

A. Tissue is isolated surgically (for example vitreous fluid), transported in saline and digested to release individual cells. B. Cells are isolated into individual micro-environments through various techniques (flow cytometry shown here) and each cell is lysed to release RNA, which is then converted to cDNA and amplified. Amplified libraries (blue circle) are then submitted for next-generation sequencing and the sequence reads are aligned to the human genome. C. A matrix containing the number of reads per gene for each cell is obtained, and after normalization and dimensionality reduction, the cells are clustered into groups that are transcriptionally similar.

The kits needed to prepare the cDNA libraries for NGS are now commercially available and accessible, opening up the door to clinicians and researchers who may not have experience in each step. The throughput of available kits has expanded dramatically over the last 10 years; Tang et al.’s initial experiment in 2009 was on 8 cells⁸, and now 10X Genomics has released a data set of mouse 1.3 million brain cells from mouse (https://support.10xgenomics.com/single-cell-gene-expression/datasets ³²).

As mentioned, numerous RNA-seq protocols have been published and made commercially available. Technical challenges in coverage (i.e., how many genes are measured), efficiency and amplification bias (certain genes amplified more than others) have been overcome in the last decade as well^8,10,16,39. There are ~20 available protocols (such as CEL-seq2, Drop-seq, inDrop, MARS-seq, SCRB-seq, Smart-seq, Smart-seq2, Gemcode, Chromium, SMARTer and others^{10,11,15,32,36}), and they vary in their individual steps, such as platform (plate-based, microfluidics, droplet-based), transcript data (3’ end of the transcript only versus full length sequencing) and read depth (number of reads generated per cell, on the order of 10⁴ to 10⁶). There have been large studies that compare amongst the various techniques with a focus on sensitivity (ability to detect gene expression for a given depth of sequencing) and accuracy (how accurate the expression value of a given gene is relative to others within a cell based on spiked-in experiments). The methods vary in sensitivity, while most have an acceptable accuracy. The variability in sensitivity mainly has to do with genes with low levels of expression; for mid- or high-level expressed genes, all methods are highly sensitive and accurate⁹. There are a number of published resources that summarize each protocol^11,14. Note that despite good accuracy and sensitivity, technical artifacts still complicate downstream analysis, as discussed below.

Computational considerations and biases

The breakthrough in scRNAseq was preceded by a critical advancement in bulk RNA sequencing. RNA expression was previously assayed via microarrays, which did not allow for a precise quantitative count or characterization of mRNA transcripts. In the late 2000s, bulk RNA sequencing was developed and measured the average expression level across thousands of cells in a sample. This transition avoided some of the constraints of microarrays (probe redundancy, annotation, cross-hybridization, etc) and allowed for the flexibility to discover novel genes, splice isoforms and allele-specific expression. However, the transition to bulk RNA sequencing came with new computational challenges particularly regarding normalization based on sequencing depth (for between sample variation) or gene length and GC content (for within sample gene-wise variation). Similiar challenges are met with scRNA sequencing and they are discussed below.

As with DNA sequencing or bulk RNA sequencing, scRNAseq generates an enormous amount of data. Technical biases that are unique to single cell RNA sequencing must be considered in order to properly interpret the results. The analysis of such data requires some background in bioinformatics or statistics, and more recently off-the-shelf tools have become available but even these require a background in data management and analysis (such Partek software https://www.partek.com/about-us/, BISCUIT⁴⁴, and MAST⁴⁵).

After the cDNA reads are aligned to the genome and assigned to genes (to generate a cell by gene matrix with counts as each entry, Figure 1C), quality control and normalization are paramount. First, poor quality data are removed from single cells based on library size cutoff, gene coverage cutoff and mitochondrial DNA (a proxy for cell viability – more mitochondrial DNA indicates less viability). After quality control, a number of non-biological sources of variation must be accounted for in the normalization step. Batch effects are the first and most obvious source, and these can be modeled as a covariate later in the analysis, or one might choose to analyze each sample separately. The data then needs to be normalized within each cell, but there is no perfect consensus on how to accomplish this. Cell-specific library complexity (i.e., the amplified cDNA for a single cell) and the amount of cellular RNA in a cell varies significantly and not necessarily in relation to biological signal. One example is transcriptional bursting, in which genes are transcribed in pulses of transcriptional activity followed by inactive refractory periods⁹. Many approach this problem with the unit of analysis being “transcripts per million” to normalize to the individual transcriptional activity of that cell. Other sources of variation include differences in cell integrity, lysis, RNA capture and amplification which lead to apparent differences between cells that do not correlate with biology³⁵.

After quality control and normalization, one typically performs dimensionality reduction [principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), uniform manifold approximation and projection (UMAP) and Gaussian process latent variable modeling] to transform the data from 20,000 variables (one for each gene) to a set of vectors that explain most of the variability. The first couple principal components are usually technical factors and some argue to remove them³⁵. Finally, the normalized data are clustered along the principal components of variation, and these clusters are presumed to represent different groups of transcriptionally similar cell types (Figure 1C). Various approaches have been employed to identify which genes or pathways are driving the difference between the groups of cells such as gene set enrichment analysis. The distribution of transcript quantities usually follows a negative binomial distribution or a multimodal distribution (if the population of cells is heterogeneous) and this requires careful non-parametric consideration¹².

A main technical challenge is that the data is sparse due to a zero or dropout problem: the efficiency with which mRNA is captured, converted into cDNA and amplified is unclear, and can be as low as 10–40%⁹. Therefore, transcripts that are expressed at very low levels have a real probability of going uncaptured. The percentage of zero counts in a scRNAseq dataset are much more than in bulk RNA sequencing. One way to maximize capture is by increasing sequence depth, though studies have shown that most protocols are saturated at 1,000,000 reads¹¹. The other way to maximize capture is to sequence more cells to increase population level coverage (assuming that some of the zeros are from stochasticity in gene expression or gene capture). Importantly, Townes et al. proposed a technique that uses generalized PCA on non-normalized UMI counts and found that it minimizes zero inflation⁴⁰.

The number of available platforms and computational tools and considerations is large and can be challenging to navigate. The best practices for the analysis of scRNA-seq data has been nicely outlined by Luecken et al. in June 2019 with a helpful step-by-step guide⁴⁶.

Single cell sequencing in ophthalmology

Now that the field has advanced and optimized scRNAseq to the point where it is accessible to the clinician, ophthalmologists are now adapting the technology to interrogate retinal tissue. The retina has long been dissected in terms of its numerous neuronal cell types, and therefore some of the earlier scRNA sequencing experiments sought to recapitulate and enrich our knowledge of these cellular groups in mouse retinas. Macosko et al. was the first to employ scRNAseq on the eye in 2015¹⁹. In this paper that describes Drop-Seq (a droplet-based approach), ~44,000 mouse retina cells were sequenced. 39 clusters were identified, and this corresponded to all known cell types – including rods, cones, bipolar cells, muller cells, ganglion cells, horizontal cells, amacrine cells, astrocytes, resident microglia, endothelial cells, pericytes, and fibroblasts. The amacrine cells were further sub-categorized based on neurotransmitter, and all groups recapitulated the known abundance and type. Subsequently, Rheaume et al. performed deeper sequencing on mouse retinas and identified 40 different subtypes of retinal ganglion cells that recapitulated prior electrophysiologic experiments but that offered new molecular characterization for many, thereby increasing the known diversity of retinal ganglion cells considerably⁴⁷. These transcriptional signatures are published online and will provide a helpful reference for future experiments. Similarly, Shekhar identified 15 subtypes of bipolar cells that recapitulated known rod, cone, ON and OFF subgroups, but also included 2 novel subtypes that they categorize as “amacrine-like” and show that they have lost their apical processes and migrated to the amacrine cell layer⁴⁸. Therefore, scRNAseq has been used both to confirm the known diversity of retinal cell types as well as distinguish new subgroups that were previously unknown or less well understood. Already an exquisitely complex and well-studied tissue in the human body, the retina’s known neuronal and cellular diversity is being rapidly advanced by scRNAseq. It should be noted that unlike histology, spatial resolution cannot be maintained with the described scRNA sequencing methods, and therefore while cellular signatures are defined, the physical location of the original cells with respect to each other in the retina can only be speculated.

Following the studies in mouse, others have gone on to study post-mortem human retinas^26,28,29. Lukowski et al. analyzed retinas from 3 donors²⁸; Voigt et al. analyzed retinas from 3 donors²⁶; and Menon et al. extracted retinas from 6 donors²⁹ (all three studies discussed below). Lukowski et al. was one of the first to look at human retinas and, like the mouse experiments, recapitulated the known retinal cellular diversity in type and abundance. They also found that MALAT1 is a gene that was upregulated in degenerating rod photoreceptors (due to increased post-mortem time), providing a potential candidate molecular driver for photoreceptor degeneration. MALAT1 is a long non-coding RNA (lncRNA) and is well known for its upregulation in metastatic solid tumors, but in the eye it is thought to lead to CREB signaling activation and thereby lead to increased stress and increase neurodegeneration⁴⁹.

Voigt et al. separated tissue samples into foveal and peripheral tissue and also recapitulated known diversity in the retina. Interestingly, Voigt et al. found that a distinguishing gene between peripheral and foveal retina was BCO2 (the gene encoding Beta-Carotene Oxygenase 2)³¹. Deficiency of this enzyme in the fovea may account for the accumulation of carotenoids that cause yellow pigment deposition within the macula seen in a number of macular degenerations. They also found that across all cells, the gene most enriched in peripheral retina was the gene for serum iron binding protein transferrin (TF), and this distinction has also been found in primates⁵⁰. They speculate that relative absence of foveal transferrin may make the central retina more susceptible to iron-mediated oxidative damage³¹. These are early results, but they demonstrate how scRNAseq in peripheral versus central retina are extremely useful, as further analysis of fovea-only expression versus peripheral expression may help explain the striking macula-only involvement of macular dystrophies and degeneration.

Studies in heritable ocular disease will likely make large strides in terms of understanding pathogenesis. In age-related macular degeneration (ARMD), for example, 52 independent genetic loci have been reported based on genome-wide association studies (GWAS)⁵¹. However, understanding how these variants contribute to disease requires knowledge of the exact ocular cell type whose gene expression is affected by associated variants, and single-cell RNA sequencing is one path forward. Menon et al. is the largest study to date of scRNAseq in human retinas, and in addition to recapitulating the known cellular diversity of the retina, they importantly linked these findings to the known genetics of ARMD²⁹. To this end, Menon et al. tested for enrichment in genes within known AMD GWAS associated loci and found that glia, vascular cells, and cone photoreceptors are all enriched for expression of genes lying within associated loci, suggesting that these cell types are important in disease. Pauly et al. studied complement gene expression at the cellular level in mouse retinas and found that Muller cells are major contributors to complement activators (c1, c3, c4, cfb) while RPE mainly expresses CFH, suggesting that the known role of the complement cascade in ARMD likely extends beyond the RPE³⁰.

Others have studied the choroid as well as the RPE. Voigt et al. subsequently went on to perform RNA sequencing of the choroid and RPE of human donor eyes, including one donor with ARMD, and they identified highly specific expression of RGCC in the choriocapillaris and upregulation in the eye with ARMD, a gene that interacts with the complement cascade²⁶. Lehmann et al. isolated RPE and choroid of mouse eyes and revealed a Hedgehog-regulated choroidal immunomodulatory signaling circuit that is responsive to changes in complement⁵². Together, the aforementioned studies help put the complement cascade in the context of relevant cell types.

While most studies so far have aimed at recapitulating the known cellular diversity of the retina, very early studies of scRNAseq in pathologic tissue to identify novel cellular characterization are emerging. Kim et al. recently obtained PVR membranes from 4 patients with complex retinal detachments and generated scRNAseq data⁵³. They found 4 groups of cells recapitulated in all samples, including a dominant microglia population, a fibroblast/mesenchymal group and a smaller population expressing genes characteristic of RPE. Importantly, the fibroblast/mesenchymal group expressed high levels of RUNX1 and SNAI1, suggesting that these genes play a role in epithelial to mesenchymal transformation⁵³.

Ultimately, like other fields, ophthalmology is in the very early stages of adopting this technology. Now that the biochemical and technical aspects of embarking on such studies has been packaged into existing toolkits, ophthalmologists can consider interrogating intraocular tissues, including tumors, membranes, infections and inflammatory conditions. And, ophthalmology is not alone; in 2017, Regev et al. announced the initiation of The Human Cell Atlas, an international consortium with the goal of defining all human cell types using scRNAseq⁵⁴. This will prove to be the ultimate resource – much like the Human Genome Project, The HapMap Project and the Ten Thousand Genomes project are for genetics.

Conclusion

Our ability to tackle disease relies on an accurate understanding of pathologic cell type and tissue composition. Distinctions based on morphology, histology, function and global gene expression are commonplace, but it is more difficult to achieve the finer divisions of cellular subtypes with precision. scRNAseq has revolutionized the field and opened up the possibility of microdissection at a molecular level. This will likely lead to critical discoveries in the field – such as which cell types are truly driving ARMD, identification of different levels of malignancy among tumor cells and what the cellular composition of inflammatory processes and fibrocellular membranes are. Studies of the intraocular inflammatory cellular composition in mice with uveitis and retinitis have been undertaken⁵⁵, for example, but future directions will involve sampling human intraocular inflammation directly. It will be important in the future to develop spatial resolution in addition to expression signatures in order to align expression findings with location in the retina. As a physician who can directly visualize pathology in the clinic and access live human tissue samples relatively easily, the ophthalmologist is uniquely positioned for such a task.