The human infertility single-cell testis atlas (HISTA): an interactive molecular scRNA-Seq reference of the human testis

Abstract

Background:

Single-cell RNA-seq (scRNA-Seq) has been widely adopted to study gene expression of the human testis. Several datasets of scRNA-Seq from human testis have been generated from different groups processed with different informatics pipelines. An integrated atlas of scRNA-Seq expression constructed from multiple donors, developmental ages, and fertility states would be widely useful for the testis research community.

Objective:

To describe the generation and use of the human infertility single-cell testis atlas (HISTA), an interactive web tool for understanding human spermatogenesis through scRNA-Seq analysis.

Methods:

We obtained scRNA-Seq datasets derived from 12 donors, including healthy adult controls, juveniles, and several infertility cases, and reprocessed these data using methods to remove batch effects. Using Shiny, an open-source environment for data visualization, we created numerous interactive tools for exploring the data, some of which support simple statistical hypothesis testing. We used the resulting HISTA browser and its underlying data to demonstrate HISTA’s value for testis researchers.

Results:

A primary application of HISTA is to search by a single gene or a set of genes; thus, we present various analyses that quantify and visualize gene expression across the testis cells and pathology. HISTA also contains machine-learning-derived gene modules (“components”) that capture the entire transcriptional landscape of the testis tissue. We show how the use of these components can simplify the highly complex data in HISTA and assist with the interpretation of genes with unknown functions. Finally, we demonstrate the diverse ways HISTA can be used for new data analysis, including hypothesis testing.

Discussion and conclusions:

HISTA is a research environment that can help scientists organize and understand the high-dimensional transcriptional landscape of the human testis. HISTA has already contributed to published testis research and can be updated as needed with input from the research community or downloaded and modified for individual needs.

1 |. INTRODUCTION

In the last decade, single-cell RNA-Seq (scRNA-Seq) has become widely used to quantify cellular heterogeneity and transcriptional landscapes; several studies have focused explicitly on the testis tissue,^1–7 providing insights and data. The complexity and size of scRNA-Seq datasets are beneficial innate features, but they limit accessibility and translatability to such data by clinicians and researchers. We introduce the human infertility single-cell testis atlas (HISTA) for the human testis tissue to overcome such limitations. This interactive web portal is a repository of curated scRNA-Seq datasets (Figure 1A). It offers a unique approach and sets of analytical tools for investigating cell types, gene expression patterns, and potential regulatory elements in the human testis^1,8 (Figures 1B and 2–4). Broadly, there are two general ways to use HISTA: (A) as a web tool to explore testis scRNA-Seq data (https://conradlab.shinyapps.io/HISTA) or (B) to download for specific customized analyses.

FIGURE 1.

(A) The human infertility single-cell testis atlas (HISTA) web portal (2022 data release) (https://conradlab.shinyapps.io/HISTA/) (shown version 2.9.7 Jan 2024) enables exploration and hypothesis testing across 26,093 high-quality cells derived from testis biopsies (two juveniles, six normal adults, one adult with azoospermia, one adult with ejaculatory dysfunction, and two adults with Klinefelter syndrome (KS)). (B) The main tab of HISTA holds the primary set of analyses to examine gene expression, identify and visualize component scores, and access driving genes and their GO enrichment analysis, guiding the translation of the identified gene modules. The magnitude and the direction of the gene weights (loadings) drive the scoring pattern observed; the top absolute loaded genes are shown relative to chromosomal organization. The Input box is the key interactive region of each tab. Searching a gene of interest will not only visualize the expression in 2D (selectable through data origin buttons) but also provide a list of components in which the searched gene is found. The cell score pattern is also visualized on a 2D map by typing these component numbers. Several metadata are also selectable. The cell scores across the donors are a quantitative alternative to evaluate donor-level scoring patterns.

FIGURE 2.

Searching HISTA by a gene of interest can be done in several tabs, beyond what is provided on the main tab. (A) The gene expression per condition (boxplot) tab allows the user to obtain a statistical comparison of gene expression by experimental conditions (CNT, INF1, INF2, KS, JUV); some conditions may lack certain cell types thus are automatically excluded based on the cell type selection. (B) The gene expression per cell type (boxplot) tab allows the user to quantify gene expression by cell types, in specific conditions such just the adult controls. Due to the large set of contrasts, a statistic is not provided, thus interpretation of the provided boxplots is needed for expression differences. (C) The gene expression per cell type (2D) tab similar to the main tab visualizes the gene expression on a select 2D representation, however in this tab the user can additionally select specific cell types. (D) The gene expression on pseudotime (germ cell only) tab provides a unique perspective on gene expression across spermatogenesis. By computing a pseudotime through the transcriptionally organized trajectory of germ cells from early germ cells (see Vignette 2 for more details on these) to late-stage round spermatids. HISTA, human infertility single-cell testis atlas.

FIGURE 4.

Searching components, which are machine-learning derived signatures, capturing gene modules of interest, across all available testes cell types and detectable transcriptome, is possible through several tabs of HISTA. A component of interest can be found by searching gene(s) in HISTA or by reading the provided index on the components. The main tab allows the user also to navigate and explore each component. (A) The Fingerprinting heatmaps identify similarity across components by assessing enrichment (see methods) of the absolute score (thus one for positive and one for negative cell scores). It can be useful to identify similar scoring signatures to contrast the signatures captured between two or more components. (B) Another method to compare component similarity is to use the gene loadings from the top-weighted genes; the number of which can be determined using the sliding scale. (C) The SDA score per cell type (2D) tab allows the user to visualize an SDA score pattern on a 2D projection of specific cell types. (D) The SDA score per cell type (boxplot) enables the user to use boxplots to quantify score distributions by cell types, in specific conditions such as only adult controls. (E) SDA scores are a complex multi-gene signature, which can be visualized across spermatogenesis using pseudotime in the SDA Score on Pseudotime (Germ Cells Only) tab. Often a striking wave pattern is depicted which highlights the exact nature of the signal across spermatogensis.

In this report, our main objective is to broaden and encourage the usage of HISTA by discussing its technical construction and application. HISTA leverages machine learning and linear algebra to facilitate a deeper understanding of testis biology and infertility. That is because scRNA-Seq analysis involves navigating a vast and noisy transcriptional landscape across numerous cells, transcriptional programs, and biological/clinical covariates. Various existing approaches group cells by transcriptional similarity,⁹ performing rigid clustering in the reduced top-variable gene space. However, HISTA takes advantage of an alternate approach described as ‘soft clustering’ that examines the entire detectable transcriptome, dissecting out specific signatures called components. These components are identified by a machine-learning algorithm called sparse decomposition of arrays (SDA).¹⁰ Utilization of SDA is one of the fundamental distinctions of HISTA from other methods and platforms hosting testis scRNA-Seq data.

There are a total of 150 components in HISTA. Each component comprises gene loadings that give weights to genes and cell scores that rank the activity of each component across cells. In generating HISTA, SDA searched the entire detectable transcriptome and all the cells to find relationships in high-dimensional gene space, creating sparse gene modules that are more interpretable than similar methods.² The gene weights provide magnitude and direction that effectively rank genes toward positively or negatively scored cells. The highest absolute weighted genes and cells are the most ‘important’ to understanding a component; for example, highly ranked genes can be passed to enrichment analysis software such as gene ontology (GO) analysis with specific cellular functions. HISTA provides precomputed GO results for each component’s top 150 positive and negative genes, giving general insights into the functional coherence of each component (if there are any).

To support this report, we develop two short research vignettes that illustrate how we, as expert users, might use HISTA to study open questions of interest to us. While preliminary in nature, their results highlight the application and translatability of HISTA for pre-clinical research. In the first investigation, we use HISTA to test whether long non-coding RNA (lncRNA) molecules are randomly expressed across different stages of germ cell development (Figures S1–S4, Tables S1 and S2). Next, we use our amalgam dataset to describe functional characteristics and transcriptional signatures of undifferentiated spermatogonia (Figures S5–S8). We hope that these computational investigations highlight how HISTA is a valuable resource for the research community and how to use HISTA for hypothesis generation and to plan for follow-up and validation studies to understand testicular biology and infertility.

2 |. MATERIAL AND METHODS

2.1 |. Data acquisition and preprocessing

HISTA contains data from several published studies six normozoospermic adult controls (“CNT”, GEO accessions GSE109037 & GSE120508), two juveniles (“JUV”, GSE120506), a patient with idiopathic azoospermia (“INF1”, GSE169062), two Klinefelter patients (“KS1” and “KS2”, GSE169062), and a patient with secondary infertility treated as a control (“INF2”/CNT-U4, GSE169062). The bioinformatics and clinical aspects of these data have been previously described.¹

2.2 |. Sparse decomposition of array and component annotation

SDA is a machine-learning algorithm that finds meaningful relationships in sparse, high-dimensional data. SDA is a form of soft clustering and data reduction that has been demonstrated to be instrumental in untangling the complex scRNA-Seq landscape.¹⁰

The SDA run housed by HISTA was set to identify 150 components. These were individually examined to determine cell type of action and the function of associated genes,¹ and these human interpretations are displayed in the information box across HISTA, including the ‘Main’ tab of HISTA. Beyond our previous description of how SDA was implemented for HISTA,¹ we elaborate on this machine-learning approach in detail in the Supporting information.

Several key features of the SDA algorithm make it unique from other factorization methods,¹⁰ mainly when applied to scRNA-Seq data.^1,2 In short, SDA helps to break down the >20,000 detectable gene space into a manageable set of components, thus reducing the dimensionality of the original cells by the gene matrix (DGE). This approach contrasts with other pipelines that initially reduce the gene space to 2000−3000 top detectable genes before a linear Principal Component Analysis (PCA) dimensionality reduction. In our application, SDA inputs the DGE and outputs two sparse matrices: the gene loadings matrix (genes by component) and the cell scores matrix (cells by components). By definition, their dot product reproduces the original gene expression matrix. The key takeaway is that SDA finds relationships that go beyond the current pipelines of scRNA-Seq (found in almost all other web portals of scRNA-Seq data) that rely on performing differential expression (DE) analysis on clusters identified in a space defined by a heuristically reduced set of principal components. Another key takeaway from component annotations is that each SDA component scores the cells in HISTA. This score represents a spectrum in which the cells relate to the function of that component.

2.3 |. Cell type annotation

Annotation of cell types in HISTA has previously been described in detail.¹ Each cell was projected onto two dimensions using t-SNE, and this summarized dataset was clustered using k-means clustering. The resulting clusters were manually annotated into “cell types” by inspection of marker genes.

2.4 |. Shiny App

HISTA was developed using the “shinydashboard” package¹¹ on the Shiny framework¹² written in R¹³ with the help of several essential packages, including ggplot,¹⁴ dplyr,^9,15 data.table,¹⁶ and Seurat.⁹ The code is available on Github (https://github.com/eisascience/HISTA).¹⁷

2.5 |. lncRNAs

To identify if a given component is enriched or depleted of lncRNAs (or any gene set from the user), a background distribution is obtained by repeated (N = 3) random sampling of an equal number of (or at least 1,000) genes. Because the randomly obtained sets reflect a normal distribution, we can conservatively use a threshold of three standard deviations from the mean to identify components highly enriched for lncRNAs; such clustering on a component would be observed in less than 0.3% of random draws.

To compare the similarity of germ cell-associated components concerning lncRNAs found in each, we computed Pearson’s correlation coefficients from an identity matrix of these genes across the SDA loadings with hierarchical pair-wise clustering (Figure 2B).

In the supplement, we also report additional findings about sense and antisense paired expression patterns, using a specialized tool in HISTA that maps the expression profiles across pseudotime (spermatogenesis trajectory) for each pair. A simple subtraction of the antisense from the sense expression is also shown as a simplistic measure of their delta expression.

2.6 |. Undifferentiated Spermatogonia

In R, we subset to the Undifferentiated Spermatogonia (USgs) cells of control adults (representing the torus and its neck structures in Uniform Manifold Approximation and Projection (UMAP) space) and reprocessed these with Seurat⁹ to obtain the top 500 variable genes. This set was used to compute K-means clustering with K heuristically set to provide equal-size clusters. The clusters were then used to train a random-forest classifier (80% training, 20% testing scheme) scheme, yielding an importance ranking of genes using the Gini index. Broadly, these steps intend to provide a parallel approach to SDA in identifying clusters of cells in an unbiased fashion and the most important genes driving those clusters. Since USgs, like all other cells, are indeed a spectrum, we used Scorpius¹⁸ to determine the gene expression pattern of the pseudotime through the K-means clusters identified.

3 |. RESULTS

HISTA can be navigated through its website, or, for those with an R (or another coding) environment, the data object can be downloaded for deeper analysis and customization of visualizations. The HISTA website is an interactive application with multiple analytical tools that interrogate the testis single-cell transcriptional landscape. Below, we highlight specific features, translational capacity, and possible navigation of HISTA while providing applicable examples of interest to the research community.

HISTA is organized by “Tabs,” which are separate pages linked in an index on the left-hand side of the site. The HISTA homepage shows a screenshot of the “Main” tab, annotated with descriptions of buttons, labels, and plots. The central data type of HISTA is the single cell. A key concept in HISTA is “metadata”—additional descriptive information about each cell in the atlas. Perhaps the most important cell metadata is the cell type. There are 17 cell “types” annotated in HISTA, including 8 germ cell types, Sertoli cells, Leydig cells, B cells, T cells, M1 macrophages, M2 macrophages, peritubular myoid cells, endothelial cells, and an uncharacterized cell population expressing neuronal markers that we assume are nerve cells. Other metadata available on cells are: donor ID, conditions (CNT, INF1, INF2, KS, JUV), experiments (origin of data), or cell cycle stage (G1, G2M, S). These can be viewed on the “Metadata (2D)” panel of the Main tab, which shows a 2D plot of cells generated by either uMAP or t-SNE data reduction.

3.1 |. Search by gene

A typical application of HISTA is to search for a gene. In our group, this is often a gene of unknown relevance that has been identified from genomic analysis of a patient with azoospermia. In HISTA, multiple tabs provide various visualizations that capture the gene expression of interest.

On the ‘Main’ tab (Figure 1B), after typing a gene name in the text input box, its expression is visualized on a 2D (t-SNE or UMAP) representation of all the testis cells.

Through the ‘Gene Expression per Condition (box)’ tab (Figure 2A), box plots with a distribution statistic (Wilcox) are available that contrast gene expression across the available conditions; a selection button set enables choosing specific cell types (such as Leydig or undifferentiated-Sg cells) or all cells to compare. For this example, we have selected Protamine 1 (PRM1) across all germ cells. This analysis demonstrates that the control adults (CNT) and an individual identified clinically with retrograde secondary azoospermia (INF2) have the most spermatids, which can be quantified by the number of cells with PRM1 expression. However, INF2 is observed on average to have a higher expression of PRM1 than CNTs, which is likely linked with INF2’s perturbed spermatogenesis having proportionally higher spermatids than earlier germ cell types (Figure 2A).¹

In the ‘Gene Expression per Cell type’ tab (Figure 2B), box plots quantify gene expression distributions by cell type; a selection button can filter the testis cells of HISTA by the various conditions. The expression of PRM1 across cell types is visualized for only the CNT group, demonstrating the highest expression in spermatids.

In the ‘Gene Expression per Cell type (2D)’ tab, a similar gene expression 2D plot as the ‘Main’ tab is initially shown. However, specific cell types can be selected in this tab, providing a zoomed-in view (Figure 2C). Typing in PRM1 as the input gene and choosing to show a UMAP of only the germ cells, we can visualize where PRM1 is expressed across spermatogenesis. Similar to the 2D (t-SNE or UMAP) gene expression visualized in the ‘Main’ tab, this plot lists the components in which the selected gene is most weighted in order of absolute magnitude; therefore, searching these components will provide gene modules correlated with the gene of interest (see Search by Gene Modules (SDA)).

The ‘Pseudotime Gene’ tab can be used to visualize gene expression across spermatogenesis using pseudotime as the x-axis (Figure 2D). “Pseudotime” is a term used by the scRNA-Seq field to describe the ordering of cells in a developmental process. Cells with smaller pseudotime values appear to be earlier in the developmental process. This highly striking and unique visualization of gene expression goes beyond box plots that discretize a continuous process. An emergent property of this visualization is that transcription is not Boolean across the tightly regulated spermatogenesis trajectory, that is, on/off states; rather, inclines and declines suggest transition dynamics. For example, PRM1 demonstrates a ‘wave’ expression pattern across pseudotime-ordered germ cells with a gradual increase at the start of spermiogenesis and a sharp decline toward the end of the signal where we find late-state round spermatids. Overlaying multiple genes of interest, for example, from a single component containing a module of correlated genes, the patterns demonstrate a complex dynamic relationship that fine-tunes cellular behavior and function instead of differentially expressed discrete groupings.

In the typical use case of following up a lead, for instance, a gene identified by sequencing a male infertility patient, a user may want to look at HISTA to better understand what the gene “does” and whether it is relevant to their interests. For this purpose, we find that looking at the SDA components associated with the gene to be very helpful. In the “Gene Expression (2D)” panel of the Main tab, the top 10 components associated with the current gene will be listed in order of decreasing gene weight. More extreme weights (positive or negative) typically indicate components where the gene is most active. In the case of PRM1, SDA component 49 is the top component (−0.914), followed by 102, 92, etc. The top component(s) can often provide important insights into the function of a gene, since, in our experience, genes in the same component often have similar functions. In Section 3.3, we describe in more detail how to use HISTA to search for and interpret SDA components.

3.2 |. Search by gene sets

HISTA also allows the use of gene sets to find SDA components most enriched with the set or to visualize gene–gene correlations in specific cell types.

To find SDA components enriched with a small gene set (~2–50 genes), the ‘Enrichment Analysis tab identifies statistically significant SDA components enhanced by the provided set (Figure 3A). As an example, we have pre-selected several genes of interest expressed in spermatids (i.e., PRM1, SPATA42, SPRR4, NUPR2, HBZ, DYNLL2); we find three SDA components significantly enriched (hypergeometric test, False Discovery Rate-adjust p < 0.01) with these genes: SDA1, SDA49, and SDA103 which can be investigated further in other HISTA tabs (see Section 3.3).

FIGURE 3.

Searching HISTA by gene sets is also possible. (A) Since the SDA components organize transcriptional programs and their driving genes and gene modules, we can identify which components are enriched with a particular set of genes. Due to the statistical nature of this comparison (see methods) the input gene could identify multiple components that capture subsets of the input genes. However, the more overlap the input set has with the top loaded components (positive or negative), the smaller the adjusted p-value obtained (significance, i.e., adjusted p-value less than 0.01 is identified by a star on top of the components). (B) To examine if a family of genes (e.g., anti-sense genes, lncRNAs, etc.) is found enriched in particular components, we deployed an alternate statistical testing approach to compare the input set with a random set of equal length. The overlap distribution compares the overlap of the input genes with the random set. The enrichment across components bar plot ranks the SDA component by the enrichment overlap. The loding corellations of input genes heatmap clusters the components by similarity based on the input given set of genes. (C) The gene expression correlation tab enables the user to identify pair-wise gene expression correlation within specific cell types. HISTA, human infertility single-cell testis atlas; SDA, sparse decomposition of arrays.

For larger (~ greater than 10) gene sets, the “Top Loaded Components” tab ranks the SDA components by enrichment, identifying components of interest for deeper investigation (Figure 3B). As a demonstration, we have pre-filtered all anti-sense genes (164 detectable) in HISTA to find components enriched with this category of transcripts. This analysis compares the enrichment of the input set of genes and a random set of equal length with each component’s top 200 weighted genes; this provides a statistical approach to quantifying enrichment. The histogram plot is the distribution of the number of anti-sense genes per component compared to a random set, demonstrating that anti-sense genes are found less frequently, that is, are depleted relative to the random set.

Using this analytical tool to prepare this paper, we also examined the enrichment of lncRNA and found several components enriched with these molecules. In exploring these components, we found novel signatures that group lncRNAs with other genes with known functions, which can unravel functional annotation of the lncRNAs as well as molecular machinery and function they participate in. For example, in a recent application of HISTA, we found a component (SDA59) that identifies disruption of piRNA biogenesis as a cause of spermatogenic failure in men.⁸ Specifically, SDA59 exhibits significant co-expression of established piRNA processing genes with multiple lncRNAs as top-loaded genes (lncRNAs categorized as pre-pre-piRNAs). Nagirnaja et al. identified 11 NOA patients with damaging mutations in SDA59 genes. Small RNA-seq from testis biopsies collected from six of these patients confirmed disruption of piRNA processing, validating our interpretation of the component. This example shows how HISTA can be used to generate hypotheses about the function of uncharacterized genes, and how to select validation assays to follow-up these hypotheses; valuable advantages for both research and patient management. To stay within the scope of this paper, we separately have summarized our findings about the expression of other lncRNAs in the testis in an analytical vignette and components that capture them as a demonstration of an application of HISTA in uncovering hypotheses to design experiments around (Figures S1–S4).

The ‘Gene Correlation’ tab provides a heatmap and cell-type selection tool that computes and visualizes pair-wise gene expression correlations within specific cell types. The hierarchical clustering also identifies gene modules within the provided input set of genes (Figure 3C). As a demonstration, we have pre-selected USg and a signature of interest we have identified in HISTA. Briefly, in these cells, this signature identifies at least two major gene expression modules that groups NANOS1 and NANOS3 with DNMT1^19,20 and DMRT1,^21–23 critical genes in epigenetic reprogramming of primordial germ cell (PGC) and male fertility. We also find NANOS2 to correlate most with the Rhesus factor RHCE gene outside the two major modules. Interestingly, RHCE has also been implicated with male infertility.^24,25 However, HISTA enabled finding it in correlation with NANOS2 in USg. In developing this paper, we further investigated gene expression patterns for early germ cells with HISTA. We reported our findings in a separate analytical vignette found in the Supporting information.

3.3 |. Search by gene modules (SDA)

There are several approaches to searching and finding SDA components of interest. Initially, the ‘Index of Component’ tab provides the full granular annotations of the SDA components. Another common method to find components of interest is to search by a gene(s) of interest, as previously described.

To find related SDA components, the ‘Fingerprinting’ tab (Figure 4A) is a dynamic quantification of the cell score enrichments relative to selectable metadata (e.g., cell types, donors, conditions, etc). Clustering is available to group similar SDA components; this clustering may help find additional related components of interest. The heatmap shown uses the score binned by the selected metadata to compute an enrichment statistic (Chi-squared residuals). This approach allows us to visualize which categories are deleted or enriched of high-scoring cells.

Another way to find similar components is using the ‘Component Correlations’ tab that computes a correlation heatmap on the top N (slide bar enables N = 10 to N = 150) gene loadings (Figure 4B). This approach highlights SDA’s ‘soft-clustering’ nature, where genes may be highly weighted in two or more components, but the captured module (set of genes and weights) is unique to the component.

Once a component of interest is found, the cell score plot on the ‘Main’ tab visualizes the cell scores on a 2D plot. The distribution of the cell scores per donor is linearly visualized by indexing each cell on the x-axis (Figure 1B). One can think of the “cell score” as approximately the “expression” or activity level of a component in each cell. This information is helpful in decoding potential functions of a component—for example, when a component is scored more highly in adult cells compared to JUV, it might be related to hormonal regulation or feedback from spermatogenesis. The LH receptor gene (LHCGR) has its highest gene loading on SDA118; this component has high cell loadings on adult Leydig cells but not JUV Leydig cells. Each component’s top-loaded genes (positive and negative weighted) are shown in the panels “Pos. Top Genes” and “Neg. Top Genes”. In the case of SDA118, many other genes known to be involved in the adult Leydig cell function appear in the top positive genes: INSL3, HSD17B6, and CYP17A1 are all in the top 5. The top genes are then used to obtain GO annotations, providing guidance on biological function, which are visualized in the “Pos. Loadings GO” and “Neg. Loadings GO” panels. In the case of SDA118, many recognizable GO categories are present—“sterol metabolic process,” “androgen biosynthetic process,” etc. Lastly, these genes and their associated loading weights are listed as a table and visualized per chromosome using a genomic coordinate system.

The ‘Cell Score per Cell type (2D)’ tab visualizes a similar 2D score plot as in the ‘Main’ tab; however, a selection is provided to zoom in on specific cell types (Figure 4C). For example, we select to visualize SDA59, which we previously introduced as a piRNA biogenesis component⁸ on a UMAP representation of only the germ cells.

The ‘Cell Score per Cell type (box)’ tab quantifies each component score using boxplots across selectable conditions (Figure 4D). In adult controls, SDA59 is the most variable in three key spermatogenesis stages; it is primarily positive late differentiating spermatogonia and mostly negative in spermatocytes in mitotically active cells that are in and passed pachytyne. Spermatocytes entering the miosis pre-pachytene stage observe both positive and negative score distributions.

The ‘Pseudotime Meta’ tab visualizes cell scores relative to pseudotime (Figure 4E), which provides a striking pattern that boxplots and colorized 2D representations can misinterpret. Because gene expression signatures across cells from a continuous process are complex and dynamic signals, the cell scores identified by SDA that capture the nature of cells relative to each signature are also complex and ongoing signatures. For example, SDA59 across pseudotime shows a ‘wave’ score pattern early in spermatogenesis. To facet this score by available metadata, a selection is provided for further investigation.

4 |. DISCUSSION

Single-cell transcriptomic analysis has become widespread in biomedical research, offering a powerful approach to unraveling cellular heterogeneity and identifying rare or elusive cellular populations and transcriptional signatures. Despite its immense potential, translating molecular findings in the high-dimensional space of single-cell data has posed significant challenges. To address such limitations, several tools and frameworks have been proposed.^9,10,26 These developments have paved the way for deeper insights and broader applications in single-cell transcriptomics. However, such software can mask the complexity of the data and computational methods and provide confounded results, as initially, they were designed and parameterized with defaults for specific tissues or cell types such as PBMCs.

HISTA remains the only interactive tool with a comprehensive representation of spermatogenesis, focused on the pathology of infertility as well as development, curated with multiple types of analyses and figures to provide translational tools examining the transcriptomic landscape of the testis. Furthermore, using HISTA’s data, specifically the SDA¹⁰ model, new data (single-cell or spatial omics) can be annotated by transferring the components; the supplemental section on SDA describes specifically the application of SDA to scRNA-Seq data^1,2 and the transfer of HISTA’s model to new data. Since the development of HISTA several new scRNA-Seq dataset of the testes have been published which we intend to package and release in future versions of HISTA. As with any new data, the current SDA model of HISTA (i.e., the components) can be easily projected and examined. Additional SDA models could be trained on the new data to discover new undiscovered gene modules.

There are significant limitations to the current version of HISTA. Some of these are limitations linked to our use of scRNA-Seq. The testis datasets that we used in HISTA were based on random sampling of cells from testis biopsies. As a result, the sampling of somatic cells in HISTA is highly skewed to JUV, KS, and NOA samples, which have few or no germ cells. Likewise, the sampling of germ cells is highly biased toward CNT samples, as the majority of cells in healthy adult testis are meiotic and post-meiotic germ cells. Another limitation of scRNA-Seq is that lowly expressed genes are not robustly detected. Our expression quantifications are whole gene summaries—they do not consider splice isoforms. We have small numbers of donors for each condition (e.g., only 2 KS donors), so some signatures that seem specific to a condition may not be. The samples lack histological characterization. Future versions of HISTA, or HISTA-like resources, can address these limitations by use of technological innovations, such as long-read sequencing and spatial transcriptomics, and by increasing the study size with more well-characterized patients.

Other websites that allow interaction with scRNA-Seq data of human testis samples have been recently summarized.²⁷ None provided a multi-dataset integrated resource carefully curated to minimize technical noise and batch effects. No functional clusters of genes were identified across the various testis cell types and spectra of states. Furthermore, none provide an interpretable, scalable, and generalizable approach to interrogate and characterize new omics data. Several reported websites²⁷ are simply post-processed data on the UCSC cell browser. The Human cell landscape (https://bis.zju.edu.cn/HCL) houses the data from a recent pan tissue study,²⁸ which unfortunately excluded the testis, so this atlas utilizes another previously published dataset,⁶ as their testis dataset (https://bis.zju.edu.cn/HCL/dpline.html?tissue=Testis). Websites like PanglaoDB²⁹ and ReproGenomics³⁰ are great resources for exploring and downloading individual datasets; however, cross-dataset integration is a current feature. The Germline Atlas (https://germline.mcdb.ucla.edu/) is another resource that houses in vitro and in vivo data on specific developmental stages. More recently, a website by our collaborators (https://humantestisatlas.shinyapps.io/humantestisatlas1/) enables general exploration of individually processed data of their “puberty atlas,” “young adult atlas,” or “adult SSC states.” Lastly, Deeply Integrated human Single-Cell Omics (DISCO) (https://www.immunesinglecell.org/) is a highly efficient and generalizable framework with a comprehensive set of human tissues that can integrate limited datasets for downstream analysis. DISCO currently features about 40 QC passing samples (of 60 total) for the testis. Their current atlas (V1.0) has focused mainly on early human life from earlier weeks of gestation to infancy and adulthood (GSE161617, GSE143381, GSE124263, GSE120508, GSE149512, and E-MTAB-10551); GSE120508 has three adults and two juvenile that overlap with HISTA. DISCO provides interaction to search gene expression on a 2D projection across several metadata options. Although their data are integrated, it still appears batch is a driver of some of the clusters; HISTA was manually and carefully curated to minimize such unwanted noise. More importantly, DISCO lacks a full spectrum of germ cells representing the entire spermatogenesis trajectory.

SDA components are useful for interpreting genes of unknown function or relevance to spermatogenesis. As we have demonstrated (see results and supplemental vignettes), using machine-learning-derived signatures, we are able to gain insights into antisense, lncRNA, and piRNA processing. Such methods emphasize the potential for future diagnostics in detecting molecular defects related to fertility. The challenges that remain today include: (A) how to infer molecular mechanisms and pathways from single-cell data remains underexplored. (B) The specific roles of coding and non-coding genes, such as lncRNAs in infertility and reproductive disorders, need further investigation, including their potential as therapeutic targets or predictive markers. (C) A crucial missing piece of the single-cell picture of the testis is the spatial landscape, connecting cellular heterogeneity (including large and irregularly shaped cells) with proximity. We aim for such considerations and data to be part of future releases of HISTA.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are openly available in Zenodo at https://zenodo.org/records/8206603, reference number 10.5281/zenodo.8206603. HISTA code is available on GitHub (https://github.com/eisascience/HISTA), the processed R Shiny object can be found (https://conradlab.shinyapps.io/HISTA/), and the raw data used to generate these data can be found via our original manuscript.