Single-Cell Technologies in Reproductive Immunology

Abstract

The maternal-fetal interface represents a unique immune privileged site that maintains the ability to defend against pathogens while orchestrating the necessary tissue remodeling required for placentation. The recent discovery of novel cellular families (innate lymphoid cells, tissue-resident NK cells) suggests that our understanding of the decidual immunome is incomplete. To understand this complex milieu, new technological developments allow reproductive immunologists to collect increasingly complex data at a cellular resolution. Polychromatic flow cytometry allows for greater resolution in the identification of novel cell types by surface and intracellular protein. Single-cell RNAseq coupled with microfluidics allows for efficient cellular trascriptomics. The extreme dimensionality and size of datasets generated, however, requires the application of novel computational approaches for unbiased analysis. There are now multiple dimensionality reduction (t-SNE, SPADE) and visualization tools (SPICE) that allow researchers to efficiently analyze flow cytometry data. Development of computational tools has also been extended to RNAseq data (including scRNAseq), which requires specific analytical tools. Here, we provide an overview and a brief primer for the reproductive immunology community on data acquisition and computational tools for the analysis of complex flow cytometry and RNAseq data.

Introduction

Pregnancy is a time of maternal adaptation to the semiallogeneic fetus, requiring a unique immunome to support the developing placenta^1,2. The maternal-fetal interface maintains a specific milieu of immune cells, consisting mostly of decidual natural killer cells (dNKs)^1–4, with smaller populations of antigen presenting cells, such as dendritic cells (DCs) and macrophages^5,6, and T cells^7,8, including unconventional subsets^9–11, and B cells¹². Further, there is evidence of cellular interactions between dNKs and DCs in early pregnancy^13–15, suggesting an important role in proper development of the placenta. However, the nature of these interactions remains unclear. Chemokine silencing is thought to limit T cell traffic into the decidua¹⁶, an important tolerance-inducing mechanism. Despite limited traffic, both CD4⁺ and CD8⁺ T cells have been identified at maternal-fetal interface^17,18 and implicated in preterm labor¹⁸. Recently, multiple groups have identified non-dNK members of the innate lymphoid cell (ILC) family at the maternal-fetal interface^19,20 with a specific tissue-resident NK cell (trNK) in the decidua^21,22. This adds a layer of complexity to our understanding of the decidual immunome. However, our ability to assign unique and unambiguous cellular identities depends on improved methods of assigning cellular phenotypes.

Technological advances allows researchers to collect increasing amount of data per cell²³. Advances in flow cytometry enable researchers to assess the expression of multiple proteins simultaneously from single cells. Similarly, advances in RNA-seq enable researchers to analyze the transcriptome of sorted cell groups or even individual cells. The tsunami of data, however, presents entirely novel challenges in visualization and require computational tools for thorough analysis.

This review aims to highlight some of the technological advances in (1) flow cytometry, including its sister platform mass cytometry (CyTOF™), (2) RNA-seq, and (3) data analysis workflows to handle complex data. We open with a discussion of data acquisition, outline best practices in flow cytometry and RNA-seq, and discuss some of the benefits and limitations of analysis tools for cytometry and sequencing data. We highlight some recent advances like expansion of fluorochrome detection in flow cytometry and single-cell RNA-seq (scRNA-seq). We then transition to data analysis and discuss dimensionality reduction tools that are available for the analysis of flow cytometry data, outline a workflow for RNA-seq data analysis, and discuss user-friendly packages available for the analysis of RNA-seq data. This review aims to provide reproductive immunologists with a go-to guide for single-cell data analysis to aid in the advancement of the field.

Data Acquisition

Flow cytometry

Flow cytometry is a method that allows rapid assessment of multiple parameters simultaneously for a single cell^23–28. The earliest version of flow cytometry was developed as a fast method for individual cell counting and separation using hydrodynamic flow focusing^29,30. Subsequently, cell sorting using a conducting medium was introduced²⁵. The introduction of fluorescent proteins to flow cytometry allowed for the development of fluorescence activated cell sorting (FACS)³¹, combining both parameter acquisition (i.e. surface marker expression) and live cell collection, giving rise to modern flow cytometry. Early applications of flow cytometry used one or a small number of fluorochromes, each with a specific absorption and emission spectrum, for the identification of cell populations. Development of additional fluorochromes and hardware (e.g. lasers, detectors, etc.) now allows detection of up to 18 parameters²³ by commercially available flow cytometers, with experimental cytometers able to detect more than 30 parameters³². In addition to traditional hydrodynamic focusing technology that employs laminar flow to focus cell suspensions through the flow cell^33,34, acoustic focusing cytometers have recently entered the market, with the advantage of faster analysis of large sample inputs^32,35. Flow cytometry is a powerful tool in immunology, including reproductive immunology, and has been broadly adopted, increasing our understanding of the maternal-fetal interface immunome^{19,20,36–41}.

Flow cytometry has applications beyond cell classification and assessment of activation status and interfaces with many other single cell-based modern applications (Figure 1). FACS, for example, is used to obtain pure cell populations for downstream applications such as gene expression analysis (RNA-seq, qPCR)^23,32 and recovery of live cells to conduct in vitro experiments. Moreover, the versatility of flow cytometry has been leveraged to create flow bead-based variants of protein detection assays (cytometric bead analysis) using fluorescent sandwich-antibody capture beads³², analogous in principle to ELISA. Signaling pathway analysis using phospho-protein detection allows researchers to investigate signaling events in individual cells^42,43. Phospho-protein labeling can also be applied as a barcoding technique⁴⁴, allowing researchers to barcode different samples and pool them for staining and acquisition, thus minimizing reagent use and increasing reliability of comparative measures.

Figure 1. Flow cytometry applications.

Flow cytometry has many applications including: phenotyping, allowing for the identification of various cells in a specimen; sorting of specific cellular populations for downstream applications; cytometric bead arrays (CBA), which allows for the quantification of soluble proteins; analysis of signaling pathways, detecting phosphorylation events.

There are some limitations when conducting flow cytometry experiments. Spectral overlap between available fluorochromes limits the number and fidelity of channels available²³. Spectral overlap, or ‘spillover’, occurs when the same fluorochrome can be detected in more than one detector, due to wide emission spectra of the natural and especially tandem fluorochromes^45,46. Spectral overlap can be corrected for mathematically by applying a spillover coefficient within the analysis software with appropriate controls^45,46. Despite correction, designing and optimizing multispectral flow cytometry panels to achieve high levels of sensitivity remains a challenge as each additional color correction increases the coefficient of variance⁴⁷. To overcome this, a number of resources are available, including well-characterized Optimized Multicolor Immunofluorescence Panels, or OMIPs, that demonstrate optimized and validated panels⁴⁷. In addition, online tools, such as FluoroFinder (www.FluoroFinder.com), provide a panel building platform that allows researchers to visualize spectral overlap, match marker density to fluorochrome intensity, and display antibody availability with vendors, thus facilitating panel design.

Reproducibility constraints are a key and often underappreciated consideration in flow cytometry. To improve the integrity and quality of flow cytometry data the MIFlowCyt (The Minimum Information About a Flow Cytometry Experiment)⁴⁸ standards have been developed to address necessary reporting. Adhering to the MIFlowCyt guidelines allows for greater collaboration and improves reproducibility. In addition, the use of a standardizing method for flow cytometer set up/calibration should be considered, specifically when analyzing fluorescence intensities, as these values depend on laser intensity and detector settings that can vary between experimental runs.

Similar to flow cytometry, mass cytometry allows researchers to collect single cell data by combining both flow cytometry and mass spectrometry^23,32. Mass cytometry, or Cytometry by Time-Of-Flight (CyTOF™), uses heavy metal isotopes as tags attached to antibodies^49,50, thus allowing for assessment of up to 40 markers, achieving higher resolution compared to flow cytometry^32,50. However, the sensitivity and throughput of mass cytometry remains low^23,50. Furthermore, use of mass cytometry leads to loss of cellular material during data acquisition, with up to 70% of cells being lost²³, a particularly important limitation for reproductive immunologists who sample from very limited tissue sources. An additional obstacle are longer acquisition times required for mass cytometry compared to flow cytometry^23,32. Mass cytometry, however, has been successfully used to immunophenotype maternal and fetal peripheral blood at term pregnancy⁵¹, thus illustrating that mass cytometry can be a useful technique in long-term studies of immune changes across pregnancy.

RNA-seq

Cellular transcriptome has become an important tool to study the biology of cell populations. Recently, RNA-seq has emerged as a valuable tool that allows us to understand the transcriptome of tissue⁵² or cells⁵³ of interest. In the 1990s, microarrays were the standard analysis platform for gene expression analysis in large studies^52,54,55. However, microarrays have high degrees of background hybridization and relative gene expression analysis of different transcripts within the same array is difficult to interpret. Furthermore, scanning of transcripts is only possible for those whose probes are present in the predesigned microarray⁵², thus restricting discovery of novel observations. RNA-seq addresses some of the limitations of microarrays and allows researchers to sequence the entire transcriptome without limiting the scope of transcripts analyzed.

Experimental design is imperative in obtaining good quality data from RNA-seq. Traditionally, RNA-seq has been performed on bulk tissue⁵²; however, analysis of bulk tissue meant that cell heterogeneity was underappreciated. Depending on the research question or limitations in obtaining samples, RNA-seq on bulk tissue might be a viable option, although most studies now rely on purified cells. Whether performing RNA-seq on bulk tissue or purified cells, it is important that researchers follow the guidelines established by the ENCODE (Encyclopedia of DNA Elements) consortium^56–58. High-quality RNA is essential to the production of good poly(A) libraries⁵⁸; therefore, researches need to follow an RNA isolation protocol that minimizes introduction of additional biases⁵⁹, including the removal of ribosomal RNA^58,59 while maintaining RNA integrity.

Designing RNA-seq experiments requires sequencing depth and number of biological replicates to be determined beforehand, as these two variables affect the power of statistical analysis. Increasing sequencing depth, or the numbers of reads for a given sample, allows for detection and quantification of a greater number of transcripts, especially less abundant transcripts^59,60. There is a trade-off between sequencing depth and number of replicates⁵⁸, with a minimum of three replicates needed for gene expression comparisons⁵⁹. Tools that help calculate the best experimental design based on researcher’s budget^59,61 are also available, allowing for the best balance between cost and statistical confidence.

Single-cell RNA-seq (scRNA-seq) has been developed to address cellular heterogeneity within tissues. The simplest workflow for scRNA-seq includes sorting of single cells, RNA isolation, library preparation, and sequencing. A major limitation is the difficulty in isolation of high-quality RNA, given that cells are subjected to staining and sorting pressures, consequently leading to some amount of cell death resulting in low RNA quality.

Microfluidic technologies, specifically droplet microfluidics, have been instrumental in the development of high-throughput protocols for scRNA-seq that requires fewer reagents and less time while increasing the amount of data acquired^62,63. Numerous droplet encapsulation protocols have been developed and applied to various biological systems^64–67. They all involve using beads to encapsulate cells in a high-throughput fashion and differ only in the encapsulation method^68,69. Droplet encapsulation methods for scRNA-seq are commercially available (reviewed in⁷⁰) thus making this technology accessible to many labs and centers. Encapsulation of doublets is a problem of droplet microfluidics. However, commercially available systems have been optimized to minimize doublet encapsulation, thereby allowing for true single-cell analysis, with the 10X Genomics platform providing the lowest doublet frequency^69,70.

Newer technologies can couple surface expression data with transcriptomics, thus allowing researchers to match protein expression with transcript levels^71,72. Cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq; available commercially as TotalSeq™ through BioLegend®), allows simultaneous acquisition of protein and mRNA expression⁷¹. CITE-seq technology uses antibodies tagged with an oligo containing a PolyA tail, thus mimicking an mRNA molecule and providing a unique ‘barcode’ for the antibody. The labeled single cells are then encapsulated with beads, using either the 10X Genomics and Drop-seq platforms, followed by reverse transcription of all mRNAs, including the antibody oligo tag, and sequencing of resulting full length and tagged cDNAs. Relative abundance of cDNA tags is used to approximate protein expression, thus providing both protein expression and transcriptome information for single cells. A similar approach is REAP-seq, which differs only in how the antibody is conjugated to the DNA barcode^72,73. Overall, these technologies will further our understanding of the maternal-fetal interface immunome.

Data Analysis

The pivotal challenge arising from acquisition of high complexity single cell data is data analysis. Advances in the data acquisition methods also require high computational resources. To address this, many computational algorithms have been developed to properly visualize, filter, and interpret high-dimensional data. Outlined here are some of the algorithms used to analyze such high-dimensional data sets.

Flow Cytometry

Flow cytometry has emerged as an important tool in single cell analysis. Advances in fluorochrome availability and detection capabilities by flow cytometers has led to an expansion of computational tools that allow researchers to analyze polychromatic flow cytometry data. Flow cytometry data is recorded in Flow Cytometry Standard (FCS) files which can be analyzed by multiple methods. Below, we discuss a series of computational tools that are available for visualization (t-SNE, SPADE, Cytosplore^+HSHE), clustering (DensVM, k-means), and category assignment (SPICE) (Figure 2).

Figure 2. Options for flow cytometry data analysis.

Flow cytometry data can be analyzed by various methods. For high-dimensional flow cytometry, the use of dimensionality reduction algorithms for visualization is recommended. Coupled with clustering algorithms, dimensionality reduction can aid in the identification of novel or rare cellular populations.

Manual analysis

Manual analysis involves the construction of one- or two-dimensional plots for a subset of markers and selecting cellular populations of interest (i.e. gating) in a sequential manner^27,28. Manual analysis remains the main method of analysis in many labs, however it has limitations which include user bias and operator variability²⁸. Multiple studies have shown a high degree of variability between different users, even when analyzing identical data sets^28,74–76. Having a pre-established gating scheme, based on well established phenotypes^77,78, as part of the analytical pipeline in the laboratory can help reduce variability between users. This can, however, inadvertently reduce exploration of the entire dataset, thus decreasing the possibility of identifying novel cellular subsets⁷⁹. Alternatively, laboratories and centers can assign one individual to analyze all data generated from one study^74–76, although this might no be feasible for small laboratories or longitudinal studies were personnel turnover maybe encountered. Despite its limitations, manual analysis remains an important component of data analysis and is required for data filtering and preprocessing in preparation for computational analysis.

Computational Approaches

As data acquisition technology has advanced in recent years, computational analysis tools have become readily available to researchers. In flow cytometry, computational algorithms have allowed researchers to address the shortcomings of manual analysis^28,74,75. Additionally, use of unbiased computational algorithms give researchers the ability to discover novel cell types that they might have otherwise overlooked using manual analysis⁷⁹. More importantly, computational approaches are capable of recapitulating manual analysis results⁷⁵, thus reinforcing their applicability in our understanding of the immune system. Despite the advantages in implementing computational analysis, one needs to have the computational skills and resources required for program implementation. Fortunately, there are easy-to-implement computer programs that require minimal computer coding knowledge, thus making them accessible to novice users. Below, we outline programs that are useful to reproductive immunologists to further our understanding of the immunome at the maternal-fetal interface.

t-SNE

A dimensionality reduction method that allows for visualization of data points on a two-dimensional map is t-distributed stochastic neighbor embedding (t-SNE)⁸⁰. Unlike principal component analysis (PCA), which is well suited for linear data, t-SNE is better suited for the analysis of flow cytometry data, which is logarithmically distributed. Coupled with clustering methods, t-SNE allows for the partitioning and grouping of cells by similarity in an unbiased, data driven manner. Implemented in the immunophenotyping of human term decidua⁴¹, amniotic fluid³⁶, and the tracking of immunome changes across murine gestation³⁸, t-SNE visualization has proven to be a powerful tool in the understanding of the maternal-fetal interface immunome.

To implement t-SNE, researchers can use Cytofkit⁸¹, available as an R package through BioConductor^82,83. The user-friendly graphical user interface (GUI) allows easy implementation of Cytofkit in R, thus minimizing the need for extensive R language. Originally developed for the analysis of mass cytometry data, Cytofkit has been successfully used to adequately phenotype the mouse myeloid compartment⁸⁴ and human helper T cells⁸⁵. Cytofkit combines t-SNE with the machine-learning algorithm DensVM (density-based clustering aided by support Vector Machine) for the unbiased partitioning of cells, and it includes additional visualization algorithms such as phenograph and ClusterX⁸¹.

Further, t-SNE can be implemented using the Automatic Classification of Cellular Expression by Nonlinear Stochastic Embedding (ACCENSE) software⁸⁶. ACCENSE, like Cytofkit, is an easy-to-use tool that does not require understanding of programming language. Because ACCENSE uses k-means clustering to assign cells to a particular group, some cells remain unclassified⁸¹ thus limiting the power of ACCENSE. A fast, distributed version of t-SNE is viSNE⁸⁷. Once available as a stand-alone MatLab program called cyt, viSNE is now part of the software suite available through Cytobank©, which requires a paid subscription. Additionally, t-SNE is available as a plug-in feature in FlowJo® and requires only installation of required R packages.

Although t-SNE is widely available to researchers and is useful for the identification of rare cell types, it has considerable limitations. Experimental variability represents a major challenge in t-SNE analysis, and results in significant batching of experimental runs^41,86,88. This variability arises at multiple experimental stages, including minor deviations in staining procedures and set-up of voltages on the flow cytometer’s detectors. These can affect the fluorescence intensity measures of the markers assessed since these values depend on both antibody concentration and voltage levels. Laboratories can establish standard operating procedures for staining to minimize variability between experimental runs⁷⁴. For flow cytometer set-up, users can employ standardizing beads⁴¹ which allow detector voltages to be set at the same level across experiments. Because t-SNE is computationally expensive, minimizing data input by downsampling is necessary, thus limiting the power of t-SNE, specifically in identifying rare populations⁸⁹. Despite these limitations, t-SNE still represents a viable option for researchers in reproductive immunology.

SPADE

Spanning-tree progression analysis of density-normalized events (SPADE) allows multiple cell types to be visualized in a branched tree^90,91. Tree nodes represent cellular clusters connected in a minimum spanning tree based on phenotypic similarity⁹². The resulting SPADE tree allows for visualization of cluster density, with node size indicating the number of cells within clusters⁹² and color scale indicating marker expression levels⁹³. SPADE has been successfully implemented to understand changes in the immune system during cancer treatment⁹⁴, characterize the diversity of extracellular vesicles⁹⁵, and asses immune drug responses⁹⁶. SPADE is part of Cytobank’s suite of computational tools. It is also available as a standalone package that can be installed in both Macs and PCs and as a Matlab program (http://pengqiu.gatech.edu/software/SPADE/).

Although SPADE has proven useful in visualizing flow cytometry data, there are some limitations to its use. First, unlike t-SNE/DensVM analysis which determines the number of clusters in a user-independent fashion, SPADE users have to specify the initial number of clusters believed to be in the data^93,97, potentially leading to underestimation of cellular clusters. If the number of clusters is overestimated, clusters can be expertly merged⁹⁷, by merging two clusters that seem phenotypically similar; however, leading to the introduction of user bias. Second, SPADE down-sampling occurs in a stochastic manner, thus leading to potential exclusion of rare cellular populations. Users interested in identifying rare populations should indicate a low down-sampling percentage, to increase the likelihood that low density regions are assigned their own cluster^23,90,97. Third, SPADE does not allow for group-level statistical comparisons when analyzing multiple FCS files simultaneously⁹⁸. The most recent version of SPADE includes several improvements, including a deterministic density-dependent down-sampling method and the implementation of a deterministic k-means clustering algorithm, reducing the randomness of down-sampling and clustering in the original implementation of SPADE⁹¹. This newest version also includes a tree-partitioning algorithm that automatically suggests a tree partition based on the largest phenotypic difference in high-dimensional space. The user can either accept or reject the suggested partition, thus allowing for a semiautomated interpretation of the SPADE tree⁹¹. Similar to t-SNE, SPADE uses fluorescence intensity as data input, therefore experimental standardization should be adopted to maintain consistency across experiments and minimize any batch effects.

Despite its limitations, SPADE is a useful analytical tool in reproductive immunology. Because SPADE allows the user to predetermine the number of clusters (i.e. categories), allowing additional datasets to be added and fitted into the existing clusters. This can prove useful in longitudinal studies involving the immune profiling of women across pregnancy and into the postpartum period, thereby allowing researchers to map peripheral immune changes across pregnancy progression.

Cytosplore^+HSNE

Cytosplore^+HSNE implements Hierarchical Stochastic Neighbor Embedding (HSNE)⁹⁹ in an integrated fashion, thus allowing for the interactive exploration of high-dimensional single-cell datasets and identification of distinct populations in a data-driven fashion^89,100. Cytosplore^+HSNE has been applied to the analysis of innate lymphoid cells in fetal human intestine¹⁰¹. One of the limitations of t-SNE is the computational time required to analyze a dataset, which often requires down-sampling. To address this, Cytosplore^+HSNE uses A-tSNE, a modified version of t-SNE that aims at minimizing the precomputation times required to analyze high-dimensional datasets¹⁰². Cytosplore^+HSNE also uses SPADE clustering which allows for high-level partitioning of the data. Coupling SPADE with A-tSNE reduces the input size of each embedding and makes analysis feasible. To address some of the shortcomings that come with the use of ACCENSE, Cytosplore^+HSNE uses Gaussian Mean Shift (GMS), which can create arbitrarily shaped clusters and leads to all the available data being clustered.

One of the main benefits of Cytosplore^+HSNE is that it can be applied to many different types of data, such as that from flow and mass cytometry and scRNA-seq. Because of the computational benefits designed into the program, downsampling is not required, allowing the analysis of the entire data set. This added benefit allows one to identify rare populations that would have been missed using traditional t-SNE⁸⁹.

SPICE

SPICE, or Simplified Presentation of Incredibly Complex Evaluations, allows users to evaluate multiple parameters, such as age or treatment, using an easy-to-use graphical interface¹⁰³. Unlike the computational tools described above, SPICE is not a dimensionality reduction tool. Rather, SPICE allows grouping of samples based on any number of categories to be displayed graphically, either in pie charts or bar graphs. Pestle© is an accompanying program that allows users to reformat FlowJo® data tables for SPICE analysis¹⁰⁴. Although Pestle© is not necessary for the current version of SPICE (v.6), it is useful to subtract background noise, thus allowing for cleaner visualization¹⁰³. SPICE includes multiple statistical tests such as permutation tests to compare pie charts and Student’s t tests to compare different groups. SPICE has been applied to understand the functional properties of mucosal associated invariant T (MAIT) cells in the female genital mucosa¹⁰⁵ as well as analyzing the immune modulating effects of progesterone during pregnancy¹⁰⁶.

SPICE is freely distributed by the NIH and is presented in a user-friendly interface that is easy to implement. However, a drawback is that it is only supported in Mac with no plans to support a PC version. Formatting data for SPICE can also be a challenge but can be overcome with a working knowledge of table formatting in FlowJo. Additionally, SPICE depends on Boolean gates of manually gated populations, which can lead to biases¹⁰⁷, as manual gating can be subjective as discussed above. To address this, it is possible to integrate dimensionality reduction techniques and draw Boolean gates on identified clusters, thus making SPICE analysis data-driven and user independent.

RNA-seq

Data analysis for RNA-seq experiments can be broken down into the following steps: quality control, preprocessing, alignment, alignment quality control, and differential gene expression (DGE) (Figure 3; Table 1). Below we discuss the analysis of scRNA-seq which follows a similar approach. We then provide an overview of additional analysis options, including pathway analysis and deconvolution approaches. We end discussing user interface packages that might prove useful and accessible.

Figure 3. Steps for RNA-seq data analysis.

Proper RNAseq data analysis requires multiple steps. Outlined in this figure are the six basic steps: (1) Quality control check of raw reds; (2) Preprocessing of reads, including the removal of adaptor sequences, poor quality reads, and trimming of reads; (3) Alignment of reads to the reference genome; (4) Quality control check of the resulting read alignment; (5) Quantification of the read alignment and count table generation; and (6) Differential gene expression. Table 1 describes programs of interest for each of the steps outlined above.

Table 1.

Programs used to complete steps in RNA-seq data analysis outline in Figure 3.

Step	Tool	Description	Ease of Use	Ref.
1	FastQC	Provides a report of raw read quality. Implemented in JAVA and accepts BAM, SAM, and FastQ file formats. Available at (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/).	GUI*	152
	FastX	Part of the FastX-Tool kit. Implemented either through Galaxy or through command-line. Pre-compiled binaries are available in Linux and MacOS X platforms.	Command line	153
	PRINSeq	Used to check quality of RNA-seq data. Can also filter, reformat, and trim reads. Provides summary statistics of the reads in both graphical and tabular format.	GUI	154
2	Trimmomatic	Is a flexible trimmer that can handle paired-end data. It is implemented through Java and is available at www.usadellab.org/cms/index.php?page=trimmomatic. Only works with Illumina generate data. Does not automatically detect the PHRED score automatically.	Command line	155
	AdapterRemoval	Trimming tool that can remove adaptor sequences. Implemented in C++. Useful in processing large data sets, with longer reads, on a desktop machine.	Command line	156
	TagCleaner	Automatically detects an adaptor sequence. It is available at (http://edwards.sdsu.edu/tagcleaner) and is implemented using Perl 5.8, through a user web-interface.	GUI	157
3	BWA	Burrows-Wheeler Aligner’s (BWA) is able to align both short and long reads. It allows for mismatches and gaps. Performance is faster compared to other aligners such as MAQ. Available at http://bio-bwa.sourceforge.net	Command line	158,159
	Bowtie	Aligns shorts reads and requires less memory allowing implementation in a desktop computer. It is faster than comparable programs. It is available at http://bowtie.cbcb.umd.edu	Command line	160
	STAR	Aligns non-contiguous sequences directly to the reference genome. Is able to detect splice junctions, multiple mismatches, and indels. Benefits include its ability to accurately align long reads, having the lowest false-positive rate while maintaining high sensitivity, and being fast. Implemented in C++.	Command line	161
4	RNA-SeQC	Provides important measures of alignment quality including: yield, alignment and duplicate rates, GC bias, rRNA content, regions of alignment, continuity of coverage, 3’/5’ bias, and count of detectable transcripts. Implemented through Java or through the GenePattern web interface (www.GenePattern.org).	GUI	162
	RSeQC	Can evaluate sequence quality, GC bias, PCR bias, nucleotide composition bias, sequencing depth, strand specificity, coverage uniformity, and read distribution over the genome structure. It is the most comprehensive and efficient program.	Command line	163
	Qualimap 2	Can compare multiple sequencing data sets and includes a novel mode that aids in the discovery of biases and problems specific to RNA-seq technology. It is available in a user-friendly interface at http://qualimap.bioinfo.cipf.es	GUI	164
5	Flux Capacitor	Quantifies the abundance of annotated alternatively spliced transcripts by distributing the reads mapping to a given splice junction among the transcripts including the exon. Written in Java; requires a Java Virtual Machine; platform independent.	Command line	165
	Cufflink	Allows for the probabilistic deconvolution of RNA-seq fragment densities and accounts for cases in which genome alignments of fragments do not uniquely correspond to source transcripts. It is an open-source C++ program and can be implemented in Linux and Mac OS X.	Command line	166
	HTSeq	Using the Htseg-count function, it counts the overlap between reads and genes, and counts only reads that map unambiguously to a single gene. Implemented in Python.	Command line	167
6	EBSeq	Uses an empirical Bayes hierarchical model approach to identify differentially expressed isoforms. It can compare two or more biological conditions. It is a robust method for identifying differentially expressed genes. Implemented in R and can be implemented through a user-friendly interface available at https://www.biostat.wisc.edu/ñingleng/EBSeq_Package/EBSeq_Interface/	GUI	168
	DESeq2	Uses shrinkage estimators for dispersion and fold change which improves its stability and reproducibility. Ideal for analysis of small studies with few replicates. Allows for a more quantitative analysis focused on the strength rather than the mere presence of differential expression. Implemented in R.	Command line	169
	Limma+Voom	Transforms the normalized counts to logarithmic base 2 and adds a precision weight for each observation. Can model the data in normal Gaussian distribution, thus allowing the data to be tested statistically. It is computationally fast and can be used with small sample sizes, with a minimum of two replicates per group. Implemented in R.	Command line	170

^*GUI = Graphical user interface.

Analysis of scRNA-seq

Single-cell RNA-seq has emerged as a powerful tool that provides the ability to analyze the transcriptome at the single-cell level¹⁰⁸, thus allowing researchers to identify rare cell populations, infer cell lineage trajectories, and assess cellular differentiation¹⁰⁹. The workflow for scRNA-seq data analysis follows the same steps outlined above. However, there are special considerations for scRNA-seq data analysis. For instance, scRNA-seq data is sparse, compared to bulk RNA-seq, either because of the nature of transcription and temporal gene expression or event dropout resulting from inefficient reverse transcription reactions of low abundance genes⁷⁰. As a result, assumptions about data distribution in bulk RNA-seq do not apply to scRNA-seq⁷⁰.

Quality control, alignment, and quantification of scRNA-seq reads can be done with the same programs used for bulk RNA-seq¹¹⁰. Because of dropout in scRNA-seq data, it is important that additional data filtering is done. Gene filtering allows for the removal of low quality genes and samples¹¹⁰. The best way to filter out poor quality genes and samples is to use External RNA Controls Consortium (ERCC) spike-ins that are added at the start of the experiment, thus providing calibration of the relative amount of starting material^111,112. OEFinder is a program that can be used for the removal of poor quality genes¹¹³. Certain confounders such as batch effects and cell-cycle induced variation also need to be removed from scRNA-seq data for adequate analysis¹¹⁰. For most studies, down-sampling can eliminate batch effects, however this reduces the complexity of the data¹¹⁰. There are packages that can be used for the removal of batch effects, such as COMBAT¹¹⁴, and for cell-cycle induced variation, scLVM¹¹⁵ and ccRemover¹¹⁶.

Normalization of scRNA-seq data, as with bulk RNA-seq, is an important step in data analysis. For experiments that did not use ERCC spike-ins, DESeq2 and edgeR can be used for normalization, just as in bulk RNA-seq¹¹⁰. For scRNA-seq experiments that do contain ERCC spike-ins, programs such as GRM¹¹⁷, BASiCS¹¹⁸, and SAMstrt¹¹⁹ can use spike-ins as internal controls for normalization¹¹⁰.

Although DGE is not usually the main objective of scRNA-seq experiments, there are programs that are capable of doing this analysis¹¹⁰. There are specific programs that address dropout in scRNA-seq data, such as Single-Cell Differential Expression (SCDE)¹²⁰, PAGODA¹²¹, and MAST¹²². Other programs available, such as Monocle¹²³, address the bimodality of gene expression. Programs used for bulk RNA-seq, such as edgeR and DESeq2, can be applied to scRNA-seq data and have been shown to perform better than programs specific for scRNA-seq data^110,124.

The power of scRNA-seq comes from the ability to identify rare subpopulations and to infer developmental trajectories of single cells¹¹⁰. For subpopulation identification, dimensionality reduction algorithms such as PCA and t-SNE can be applied^{110,125–127}. There are specific dimensionality reduction algorithms for scRNA-seq such as Zero-inflated factor analysis (ZIFA)¹²⁸ and PAGODA¹²¹ that can be implemented for the identification of subpopulations. Furthermore, there are multiple programs, such as Monocle¹²³, Waterfall¹²⁹, SLICER¹³⁰, and SCUBA¹³¹ that can infer differentiation pathways from scRNA-seq data.

Gene Pathway and Deconvolution

Gene pathway analysis takes into account the relationship between expressed genes, thus allowing researchers to determine whether biological processes differ between groups¹³². Gene pathway analysis workflow involves a list of genes of interest from the RNA-seq data set and the application of statistical methods to test for gene enrichment^132,133. There are multiple packages that can perform pathway analysis, such as GSEA¹³⁴ and GSVA¹³⁵. Recently, programs have been developed to account for the pathway topology¹³². Multiple programs, such as SPIA¹³⁶ and TAPPA¹³⁷, provide pathway topology information by taking into account gene interactions and weighing interacting genes more than non-interacting genes, while DEAP¹³⁸ identifies the pathway that is most differentially expressed.

Deconvolution programs are capable of determining cellular composition of heterogeneous samples based on gene expression patterns across different cell subtypes^139,140. In most instances, these programs require gene expression datasets as a reference to determine the composition of bulk RNA-seq samples¹⁴⁰. There are multiple programs that are capable of determining the composition of bulk samples all using a specific mathematical approach¹³⁹. CIBERSORT¹⁴¹, for example, is a partial deconvolution algorithm that uses nu support vector regression to estimate the cellular fractions in bulk samples¹³⁹. However, the effectiveness of CIBERSORT depends heavily on the reference profiles used¹⁴¹. Multi-Subject Single Cell (MuSiC) is another deconvolution algorithm that uses scRNA-seq data obtained from a cross section of individuals as a reference to then analyze bulk RNA-seq data from different individuals¹⁴². MuSiC might prove of particular interest to reproductive immunologist as only a small set of samples would be required for scRNA-seq, thus limiting costs, with the remaining of the study samples requiring bulk RNA-seq.

Integrated Graphical User Interface Packages

Most of the programs/packages outlined here require some coding knowledge. Fortunately, there are some software packages that combine some of the tools listed above that make RNA-seq data feasible to users who might not have extensive coding skills. Here we cover Chipster and Galaxy.

Chipster¹⁴³ is a collection of up-do-date analysis and visualization tools and is available as a graphical Java desktop application. Chipster easily integrates tools regardless of how the tool is implemented¹⁴³. Chipster allows users to track their analysis workflow and save any series of steps¹⁴⁴. In addition to RNA-seq data, Chipster is able to analyze other types of data, including microarrays, miRNA-seq, ChIP-seq, and whole genome sequencing^143,145. Chipster is freely available and can be used with a free short-term evaluation account through Chipster’s server¹⁴³. Additional benefits of Chipster include the ability for users to save their workflows and share them with other users, thus allowing for collaboration and reproducibility¹⁴³. Users are able to view the original tool code and integrate additional code¹⁴³. Chipster, however, is not designed to be integrated to a laboratory information management system (LIMS), limiting Chipster’s use to laboratories that do not have a large amount of NGS data¹⁴⁵.

Galaxy^146–148 is a popular, web-based framework¹⁴⁸ that provides computational tools with an intuitive user interface^146,149. Most importantly, Galaxy does not require substantial programming skills¹⁴⁷. Galaxy includes a unique history system that allows the user to organize and save workflows, and maintain quality control of the analysis performed¹⁴⁹, thus allowing for reproducibility¹⁴⁶. Although written in Python, Galaxy does not require the user to have Python programming knowledge. Galaxy hides the computational details from the user, thus eliminating the need for extensive bioinformatics expertise^146,149. Despite this, Galaxy is limited because pipeline output is not standardized and depends on the user input¹⁴⁵. Because of the high level of flexibility within Galaxy, its usage might be limited towards those that have a good knowledge of the tools being applied, compared to Chipster¹⁴⁵.

Verifying biological veracity of single cell data sets

Single cell data presents are wealth of information. Despite its advantages there are some considerations to keep in mind, both as users and readers. Below, we outline advice on how to minimize common problems and how to improve data quality.

To address batch effects and document the extent of experimental variation in flow cytometry, users should create additional parameters in the FCS files that document user/operator and experiment number. Downstream analysis would reveal whether there is grouping based on experimental runs and/or operator, allowing researchers to properly interpret their data.

We recommend that researchers take advantage of user-friendly programs such as Chipster to check the quality of their RNA-seq data. At a minimum, researchers would be able to check read quality and determine the presence of outliers. This would allow the researchers to make experimental adjustments if needed. Having knowledge of outliers or batching effects would also allow researchers to have an informed conversation with bioinformaticians that could then guide them on how to handle downstream analysis. In addition, researchers should perform some validation of their RNA-seq data, either by performing qPCR or protein level detection (flow cytometry, imaging). As readers, we should also expect some level of validation on reported RNA-seq results.

When working with bioinformaticians, we recommend that researchers request all downstream data files (count tables, DE analysis tables). This is particularly important when looking at gene pathway analysis, were the bioinformatician might be unfamiliar with the biology and might consider some results irrelevant.

Finally, we recommend that researchers become familiar with basic bioinformatics language (Table 2) and basic data analysis terms (Figure 3). This will facilitate conversations with bioinformaticians and will make undertaking RNA-seq experiments manageable.

Table 2.

Common useful terms

Term	Definition
Alignment	The process of matching reads to a particular region of the genome.
Counts	Number of reads that map to a particular gene.
Coverage	The extent to which a genomic region of interest was sequenced.
Dimensionality reduction	The processed by which the most important features of a high-dimensional data set are extracted thus resulting in a data set with reduced dimensions.
Dropout events	Missing data values due to low transcript expression and the stochastic nature of gene expression.
Embedding	Mapping of high-dimensional features onto a low dimensional space.
Feature extraction	Transformed data set built with the most predictive features of a high-dimensional data set.
Feature selection	Selection of most predictive features in a high-dimensional data set.
Features	Variables of a particular data set (i.e. fluorescence intensity or gene counts).
Machine learning	Process of building mathematical models that allow computers to be trained to be predictive. Requires training with a “training dataset”.
Partitioning	Division of data.
Quantitation	Generation of count table that integrates that number of reads per gene that were aligned successfully.
Reads	Nucleotide sequences as a result of sequencing.
Sequencing depth	The number of reads per genomic region.

Conclusion

Expanding our knowledge of the reproductive immunome presents a great opportunity to improve fetal and maternal health. Many immunological questions remain obscure at the maternal-fetal interface, including (1) how decidual immune cells interact with non-immune cells (trophoblasts, stromal cells, endothelial cells) and other decidual immune cells, (2) if and how decidual immune cell dysfunction lead to pregnancy pathologies, and (3) how leveraging single-cell data could lead to reliable diagnostic tools. Already, tools discussed here, such as the 10X Genomics platform and scRNA-seq, have been leveraged to map cellular interactions in first trimester decidua^150,151, providing insight into these important cellular relationships. This review, although not exhaustive, highlights practical tools that will help the reproductive immunology community decipher the mysteries of immune cellular networks governing gestational outcomes.