Leveraging online resources to prioritize candidate genes for functional analyses: using the fetal testis as a test case

Abstract

With each new microarray or RNA-seq experiment, massive transcriptomic information is generated with the purpose to produce a list of candidate genes for functional analyses. Yet an effective strategy remains elusive to prioritize the genes on these candidate lists. In this review, we outline a prioritizing strategy by taking a step back from the bench and leveraging the rich range of public databases. This in silico approach provides an economical, less biased, and more effective solution. We discuss the publicly available online resources that can be used to answer a range of questions about a gene. Is the gene of interest is expressed in the system of interest (using expression databases)? Where else is this gene expressed (using added-value transcriptomic resources)? What pathways and processes is the gene involved in (using enriched gene pathway analysis and mouse knockout databases)? Is this gene correlated with human diseases (using human disease variant databases)? Using mouse fetal testis as an example, our strategies identified 298 genes annotated as expressed in the fetal testis. We cross-referenced these genes to existing microarray data and narrowed the list down to cell type specific candidates (35 for Sertoli cells, 11 for Leydig cells, and 25 for germ cells). Our strategies can be customized so that allows researchers to effectively and confidently prioritize genes for functional analysis.

Introduction

The process of sex determination is complex, involving multiple pathways that shape the identity of the gonads. Identifying the genes that make up these pathways is like solving a complex puzzle. We propose that by using public databases investigators can efficiently identify new candidate genes for their involvement in gonadogenesis. While scientists have gathered a number of missing puzzle pieces, unexplained cases of disorders of sexual development (DSD) and infertility clearly indicate that many key pathways have not yet been placed in the picture [Ono, and Harley, 2013]. In order to fill in the missing pieces, the field has used mouse models to generate numerous transcriptomic studies that reveal variations in gene expression between sexes, cell types, developmental stages and different genotypes [Jameson et al., 2012; McClelland et al., 2015; Beverdam, and Koopman, 2006; Nef et al., 2005; Bouma et al., 2010; Bouma et al., 2007; Albrecht, and Eicher, 2001; Rolland et al., 2011; Coveney et al., 2008; Munger et al., 2013]. Each of these studies had the goal to identify new genes that control key processes in sex determination of the gonads.

Over time the strategies used to generate new candidate genes has changed with the availability of different technologies. Initial cDNA array screens in the 1990s and early 2000s compared male and female gonads to search for genes enriched in either testis or ovary, with the assumption that enriched genes in one sex were important for development of that fate [Wertz, and Herrmann, 2000; Bowles et al., 2000; McClive et al., 2003; Menke, and Page, 2002]. With the advent of microarrays, comparing gene expression between multiple groups, such as testis and ovary at different developmental time-points, became feasible. This approach allowed the field to identify genes that were expressed at key developmental time-points, and begun the process of fitting together in pathways that were initiated by marshaling transcriptions factors such as SOX9 [Bouma et al., 2010; Coveney et al., 2008; Grimmond et al., 2000; Nef et al., 2005; Munger et al., 2009]. Subsequently, to move beyond male and female comparisons and to gain more cell type specific information, scientists generated transcriptomic data from enriched cell populations isolated from cell type-specific reporter mice. This strategy was used to isolate somatic cells (Sertoli, Leydig, and others) and germ cells to identify key differences between lineages in addition to sexes [Nef et al., 2005; Beverdam, and Koopman, 2006; McClelland et al., 2015; Rolland et al., 2011; Inoue et al., 2015; Bouma et al., 2007]. This approach was used in the GUDMAP (GenitoUrinary Development Molecular Anatomy Project) Consortium-backed microarray of four key cell types (supporting, germ, interstitial, and endothelial cells), aiming at differences in cell population gene expression over a developmental time-course [Jameson et al., 2012]. Over the past two decades, transcriptomic studies identified numerous factors and sex-linked genes that are critical for gonadogenesis. Efforts by other groups continued to fill in the puzzle with the advent of RNA-seq and whole exome sequencing. RNA-seq allowed gene expression to be assayed at a greater sensitivity in a more unbiased way than microarrays on whole gonads and sorted cells [McClelland et al., 2015; Inoue et al., 2015; Lindeman et al., 2015]. In tandem, the decreasing cost for next-gen sequencing has made whole exome sequencing of DSD patients economically feasible [for review see Ostrer, 2014; Ono, and Harley, 2013]. The function of genes from mutations identified in DSD patients can then be confirmed in the mouse model in complete knockout or exact recapitulations of the mutated human gene. One common feature of these approaches is that they produce an extensive list of candidate genes. The key question then becomes: how do we prioritize the list of candidate genes and determine which ones to pursue for further functional analyses?

Although the way we identify candidate genes continues to evolve, the way in which we prioritize candidate genes for further functional analysis remains rudimentary: the most obvious candidates were selected, based on whether they fit into existing pathways with known roles in gonadogenesis. From experience we know that transcription factors or sex-chromosome linked genes are good candidates for functional analyses. Traditionally, scientists tried to prioritize genes by targeting genes with a sexually dimorphic expression pattern, profiling gene expression over a developmental time-course, and then performing in situ hybridization or immunohistochemistry to the cell type in which the gene was expressed [McClive et al., 2003; Rolland et al., 2011; Menke, and Page, 2002; Bouma et al., 2004]. The gene with the most robust sex-specific expression, the best working in situ hybridization probe (or antibody), and some background information in PubMed goes forward as the candidate gene. This prioritization problem is not new: genes identified in early cDNA screens were present in subsequent microarrays, and found to be critical for gonadogenesis 20–25 years after identification. The problem was not the identification of these genes, the issue was to prioritize them or identify them as a good candidate for further functional analysis. Confronted by a list that contains hundreds of genes, the “gene prioritization problem” looms bigger than ever. We argue that current methods of selecting a candidate gene are unnecessarily expensive, ineffective and biased in today’s information rich world. We propose that the community could better prioritize genes for functional validation by utilizing publically available gene expression and transcriptomic databases. These resources represent a valuable pool of information that can be used freely by any scientist to move projects and hypotheses forward.

In this review, we discuss the development and use of gene ontology (GO) terms by databases and the key features of a “good” gene list before describing four key resources: expression atlases and databases, added-value transcriptomic databases, pathway or GO term enrichment analysis, and mouse knockout or human disease variant databases. We describe how to utilize these databases to prioritize genes, and discuss the strengths and limitations of each resource. Finally, we demonstrate how to use these resources as a discovery tool. As an example, we analyze “testis” annotated genes in the Eurexpress expression database. Although the reanalysis of existing data sets and in silico discovery is still a fairly new concept in the field of sex determination, we argue that by capitalizing on the wide range of publically available databases, scientists can prioritize genes in a candidate list without exhausting research dollars and time chasing ghosts.

I. Online tools for generating and analyzing the genes-of-interest list

While enormous amounts of functional data are available in published literature, each paper typically reports on a limited number of genes and developmental stages. However, there are numerous publically available databases that host a wide range of expression and transcriptomic data covering murine embryonic development. To increase utility of individual databases, most core databases are linked with specialist databases to create ways to springboard between resources. For example, Mouse Genome Informatics (MGI; http://www.informatics.jax.org/; Table 1) provides a comprehensive data page linking to most major resources for each gene. However, it is not yet possible to perform a bulk query and retrieve a set of metrics or a list of genes from a core database. How the user implements the strategies outlined in this paper will depend on the nature of the question and what type data are used. If the investigator has already had data from a patient or mouse model, prioritizing genes by focusing on key biological processes or signaling pathways might be the first step. Conversely, if the investigators are searching for new candidates, then reanalyzing published transcriptomic data may provide new genes of interest, followed by validating the expression patterns in the existing expression databases. In this section, we aim to guide the investigator through how each type of database works and what can be gained from each one so the investigator can customize their search to solve their “gene prioritization problem”.

Table 1: List of resources that can be used to prioritize candidate genes.

MGI is often considered the one stop shop when investigating a new gene. By using resources like the GO Consortium and AmiGO2 to understand the specifics on gene ontology the investigator can use tools like DAVID and GSEA to their advantage to uncover interesting processes and pathways present in the candidate gene list. These enrichment tools are underpinned by data from knowledgebases such as Entrez Gene, ENSEMBL, UniProt and TRANSFAC. The EMBL-EBI Expression Atlas provides a rich resource of reanalyzed transcriptomic data in mouse and human. The relevance of candidate genes in human disease can then be investigated using too such as OMIM and HGMD. URLs are current as on September 2016.

Resource	Functional Description	URL
MGI (Mouse Genome Informatics)	MGI is the international database resource for the laboratory mouse, providing integrated genetic, genomic, and biological data to facilitate the study of human health and disease.	http://www.informatics.jax.org
Gene Ontology (GO) Consortium	The GO Consortium develops up-to-date, comprehensive, computational models of biological systems, from the molecular level to larger pathways, cellular and organism-level systems.	Homepage: http://geneontology.org/ Descriptors: http://geneontology.org/page/ontology-documentation http://geneontology.org/page/development
AmiGO2	Searchable interface of GO Consortium	http://amigo.geneontology.org/amigo/landing
DAVID (Database for Annotation, Visualization and Integrated Discovery)	DAVID is a comprehensive set of functional annotation tools for investigators to understand biological meaning behind arge list of genes.	https://david.ncifcrf.gov/
GSEA (Gene Set Enrichment Analysis)	GSEA is a computational tool that determines whether an set of genes shows statistically significant, concordant differences between two biological states.	http://software.broadinstitute.org/gsea/index.jsp
Entrez Gene	Gene provides a record including nomenclature, Reference Sequences (RefSeqs), maps, pathways, variations,phenotypes, and links to genome-, phenotype-, and locus-specific resources.	https://www.ncbi.nlm.nih.gov/gene/
ENSEMBL	Ensembl is a genome browser for vertebrate genomes that supports research in comparative genomics, evolution, sequence variation and transcriptional regulation.	http://useast.ensembl.org/index.html
UniProt	UniProt is a comprehensive, high-quality and freely accessible resource of protein sequence and functional information	http://www.uniprot.org/
TRANSFAC	TRANSFARC houses data on eukaryotic transcription factors, their experimentally-proven binding sites, consensus binding sequences (positional weight matrices) and regulated genes.	http://www.gene-regulation.com/pub/databases.html (last updated 2005)
EMBL-EBI Gene Expression Atlas (GEA)	Gene expression data displayed in this resource is re-analyzed in-house to detect genes showing interesting baseline and differential expression patterns in different cell types and organs, in addition to different developmental stages, disease states and biological/experimental conditions	https://www.ebi.ac.uk/gxa/home
Online Mendelian Inheritance in Man (OMIM)	OMIM is a comprehensive online collection of records on human disease genes and genetic phenotypes.	https://omim.org/
UK10K Project	UK10K is an example of one of the many smaller efforts to understand the link between low-frequency and rare genetic changes, and human disease	http://www.uk10k.org
Human Gene Mutation Database (HGMD, from Cardiff University)	HGMD is a collection of germline mutations in nuclear genes that underlie, or are associated with, human inherited disease. -HGMD Cardiff: unlicensed and limited access (no access to mutations identified in the last 2.5 years) -HGMD Professional: licensed through Qiagen and complete access to entire collection	http://www.hgmd.cf.ac.uk/ac/index.php
Human Gene Mutation Database (HGMD) Professional		https://www.qiagenbioinformatics.com/products/human-gene-mutation-database/
IKMC (International Knockout Mouse Consortium) IMPC (International Mouse Phenotyping Consortium) IMSR (International Mouse Strain Resource)	-IKMC generates targeted ES-cells of all known protein coding mouse genes and companion Cre driver lines - IMPC generates the mouse strains and perform standardized phenotyping. - All gene trap alleles are housed at Jax (Jackson Laboratories) - ISMR is a searchable catalogue of over 2000 strains	IKMC: access through IKMC/IMPC hub IMPC: www.knockoutmouse.org IKMC/IMPC web portals were merged to create a central hub: http://www.mousephenotype.org Jax: https://www.jax.org IMSR: www.findmice.org;
DMDD (Deciphering the Mechanisms of Developmental Disorders)	Results from the Consortium to study knock-out lines where embryonic development is compromised.	https://dmdd.org.uk
HGNC (HUGO Gene Nomenclature Committee) Multi-Symbol Checker	This tool to check submitted gene names in a list against HUGO verified names and their known synonyms,	http://www.genenames.org/cgi-bin/symbol_checker

The Construction of a Gene Ontology Framework

The Gene Ontology (GO) framework [Ashburner et al., 2000] is an example of a flexible and rapidly developing textual framework for the molecular functions of gene products, including their sub-cellular localization, the biological processes in which they function, and the processes in which they participate during embryogenesis. The GO Consortium (http://geneontology.org/; Table 1), initially a collaboration between the Mouse Genome Informatics, FlyBase and Saccharomyces Genome databases, was formed to create descriptions of biological processes and establish standards for the community-wide organization of a set of frameworks in order to enable the integration of data from different model organisms [Ashburner et al., 2000; Blake et al., 2000; Ringwald et al., 2000; Ball et al., 2000; Consortium, 1999]. The structured language laid out by the Gene Ontology Consortium to describe the properties of gene products are called GO terms. GO terms are catalogued in a central thesaurus hosted by the GO Consortium (http://geneontology.org/page/ontology-documentation; Table 1). Databases using GO terms have these terms as set outputs (or answers). Therefore, it is important to understand how GO frameworks are constructed, the limitations of GO terms and how GO terms are utilized by different databases.

There are three central themes defined by GO terms based on the function of the protein. 1) Terms that describe the Cellular Component, the parts of a cell or its extracellular environment. 2) Terms that describe the Molecular Function, the basic functions of a gene product at the molecular level. 3) Terms that describe Biological Process, which is defines as the molecular events necessary for the function of cells, tissues, organs, and organisms [Ashburner et al., 2000]. Within the ontology framework each term has defined relationships to one or more other terms in the same classification, and sometimes to terms in other classifications; some terms are parent terms to a host of more specific terms like a branching tree to describe things in greater and greater detail. For an example of GO annotation, the gene product “Sox9” can be described by the Molecular Function term “transcription”, the Biological Process terms “cell differentiation” and “system development”, and the Cellular Component term “nucleus”. To look up terms for specific genes, a brief overview can be found in the Gene Ontology Classifications section of MGI and more detail can be obtained by querying AmiGO2 (http://amigo.geneontology.org/amigo/landing; Table 1).

The consistent use of GO terms for cellular and biological processes and for anatomical structures throughout development is essential for the success, development and maintenance of biological databases. Creating and maintaining ontologies seem straightforward; however, the creation and maintenance of ontologies in the context of developmental biology poses some interesting problems. The subtleties between different ontology terms used to by the Biological Process Ontology Guidelines for Development to describe development and differentiation of lineages and tissues can be found in the GO Consortium resources (http://geneontology.org/page/development). The goal of ontology is to represent knowledge in a computer-interpretable manner and that can (hopefully) be used to cross-reference data among different databases. The success of a database and the ontology depends greatly on the needs of the end user of the database, the developmental biologist wants to know about the genes and proteins involved and to map the expression of the genes associated with a tissues and cell populations at a different developmental stages. For example for the process of differentiation of one cell type into a different cell type, there are multiple ways this process could be represented by the GO terms. Consider these three similar yet distinct terms that could be used to describe the differentiation process (GO term in italics):

[cell type] cell fate commitment —

The process whereby the developmental fate of a cell becomes restricted such that it will develop into a [cell type] cell.

[cell type] cell fate specification —

The process whereby a cell becomes capable of differentiating autonomously into a [cell type] cell in an environment that is neutral with respect to the developmental pathway. Upon specification, the cell fate can be reversed.

[cell type] cell fate determination —

The process whereby a cell becomes capable of differentiating autonomously into a [cell type] cell regardless of its environment; upon determination, the cell fate cannot be reversed.

Each of these terms, although similar on first glance, encapsulates important information about the cell and the process that takes place, and they are not interchangeable. Knowing what pertinent GO terms mean in detail and how those GO terms are nested within other GO terms is critical to extract important information about function of a gene product.

GO terms can be a powerful tool to describe biology and make findings accessible across data formats and platforms. The core item the investigator needs to begin is a “good” candidate gene list, but what defines a “good” list?

Key features of a “good” candidate gene list

Lists of ‘interesting’ genes, which may range in size from hundreds to thousands of genes, can reveal patterns, signaling cascades and processes important for organogenesis. The analysis of these data sets has gradually become the responsibility of the biologist, rather than the bioinformatician, as more biologist-friendly tools become available. In addition to analyzing new primary data, reanalysis and inclusion of published data should be considered to bolster a new analysis. Before considering how to analyze the genes-of-interest list, the quality of the list itself should be assessed. Realistically, any correlations and directions indicated by a list should be confirmed by bench work and functional analysis. Regardless of this, as a general rule a ‘good’ gene list has a few key features [modified from Huang et al., 2009a]:

1. A reasonable number of biological (not technical) replicates and the reproducibility of the list before analysis, either from independent experiments or by statistical testing (appropriate p-values and test parameter should be chosen for each experiment), are important considerations, especially when re-analyzing data.

2. The genes-of-interest list contains a series (not just one or two) bona fide marker genes (i.e. Sox9, Amh, and Dhh for the fetal testis) that would be expected as a result for the given experiment or analysis. When looking at reprocessed data, it is important to remember that the statistical approaches for microarray and RNA-seq analysis have changed over time; this is especially important when processing data from pre-2012. Simple reprocessing of experiments using newer analysis methodology may result in subtle or more dramatic changes in the final genes of interest list compared to the original analysis. If using reprocessed data, checking for known marker genes remains a simple, yet effective metric to determine if the analysis and data is clean.

3. The number of genes on the list is big enough for pathway analysis (around 100 genes) but not so big that it is not comparable (around 2000–3000 genes). The size of a list can affect the ontologies that are selected for, especially for smaller lists. As a test, the enriched terms should appear in the queried list and not in a random list of approximately the same size (number of genes).

Expression atlases and databases

The first question the investigator often asks is in which cell type/s a gene is expressed during gonadogenesis. In the past, in situ hybridization on embryonic gonads was used to validate the expression of candidate genes from mRNA expression studies, and determine in which cell type the gene was expressed. This first pass analysis should now be considered redundant for many developmental biologists with the completion of large-scale in situ hybridization atlases that aim to capture gene expression throughout the murine embryo on a gene-by-gene basis. Several databases used high-throughput robotic technology to conduct RNA in situ hybridization on sectioned or whole murine tissues [Geffers et al., 2012]. The consortiums then manually mapped and annotated gene expression patterns on standardized images using anatomical GO terms (Fig. 1A; [Reymond et al., 2002; Visel et al., 2004; Gitton et al., 2002]). These databases can play a critical role in validation of the quality of a gene list, taking an in-depth look at many genes in a list with a focus on a specific process or pathway, or can be used as a discovery tool to uncover new candidate genes and expression patterns (discussion on using these databases as a discovery tool can be found in Section II)

Fig. 1: Overview and examples of gene entries in the Eurexpress, Gene Expression Atlas, and OMIM databases.

(A) In situ hybridization images of 14.5 dpc embryos are compiled in the Eurexpress Transcriptome Atlas Database for Mouse Embryo (http://www.eurexpress.org). Expression observed in an image is annotated and encoded as text that can be queried. The more general anatomical terms branch into more specific terms. For example, the nested terms that describe the testis are displayed. (B) The Eurexpress viewer scrolls through sections in a “virtual embryo environment” and zoom into regions as similar to a digital microscope. For the gene Star, the testis can be visualized in the 6^th section. The testis region is demarcated by a dashed line; a, adrenal; k, kidney; m, mesonephros. (C, D) The Gene Expression Atlas houses easy to access transcriptomic data for a wide variety of species from all major consortia and sequencing efforts. For the gene Sox9, detailed information are available on expression in the (C) mouse and (D) human. Data is presented in a tabular format that can be sorted using a number of parameters (graphics shown here are part of the full table). Selecting data from a project in the table, such as 16.5 dpc 49 FANTOM (C, in mouse) or 68 FANTOM (D, in human), highlights the organs (in our case, the testis) that are included in that dataset in purple on the interactive body-map. (E) The OMIM database (Online Mendelian Inheritance in Man) catalogues human disease variants. Entries can be either by phenotype or gene. In the example of a search for the gene NR5A1 (also known as SF1), there are many alternative gene symbols that have been used historically; however, the disease information is catalogued under the HUGO approved symbol (NR5A1). Information on the location of the gene, all phenotypes caused by mutations in NR5A1 are listed in a table with ID numbers for each phenotype and additional information on inheritance. The entry number and mapping key encode additional information about the disorder.

In the race to map the expression pattern to the transcriptome throughout development, specialized knowledgebases, such as the Allen Brain Atlas (ABA) ([Lein et al., 2007; Jones et al., 2009]; http://brain-map.org/), which catalogues gene expression in brain sections throughout development and adulthood, have developed alongside with more general whole embryo atlases, such as Eurexpress (mainly whole mid-gestation murine embryos; http://www.eurexpress.org/ee/; [Diez-Roux et al., 2011]) and the fledgling 3D Atlas of Human Embryo Development developed by the Academic Medical Centre in the Netherlands (Carnegie Stage 7–23 (15–60 days); http://www.3dembryoatlas.com/; [de Bakker et al., 2016]). Specialized databases cataloguing expression of genes in specific murine tissues such as the ABA, the Gene Expression Nervous System Atlas (GENSAT; the nervous system; http://www.gensat.org/; [Gong et al., 2003]), the GenitoUrinary Development Molecular Anatomy Project (GUDMAP; the gonads, reproductive tract, kidney and urinary tract; http://www.gudmap.org/; [Brunskill et al., 2008]) and FaceBase (curated craniofacial data from mouse, human and zebrafish; https://www.facebase.org; [Van Otterloo et al., 2016]) have now been integrated into broader databases. These collaborations bring together specialist expression data with that from projects on the whole embryo such as the Edinburgh Mouse Atlas of Gene Expression (EMAGE; [Hill et al., 2004]; http://www.emouseatlas.org/emage/) and the Mouse Genome Informatics (MGI) Gene Expression Database (GXD; [Christiansen et al., 2006]; http://www.informatics.jax.org/expression.shtml). Generally, each database uses a set of standard anatomical ontology to annotate gene expression in each image (for details see: [Hayamizu et al., 2015]).

As an example database, the Eurexpress transcriptome atlas catalogues annotated RNA in situ hybridization expression patterns for approximately 18,000 Mus musculus protein-coding genes at 14.5 days post coitus (14.5 dpc; http://www.eurexpress.org). Using Eurexpress sections taken throughout the entire embryo, annotated gene expression patterns can be visualized in an online viewer. This atlas aims to achieve complete representation of all embryonic tissues throughout the 24 representative sagittal sections. For example, the gene Star is expressed in the adrenal and fetal Leydig cells at 14.5 dpc; in the Eurexpress assay it is annotated as expressed in the testis on section 5–8 and 14–16. Flipping through the online slide deck and zooming in, expression of Star is clearly detected in the adrenal and testis in section 6 (Fig. 1B). The quality of the in situ hybridization in Eurexpress is high but not always consistent. The Eurexpress Project reports that 18% of genes tested were not detected at 14.5 dpc; whether this is due to ineffective probe design, the experimental conditions, or the gene is not expressed in the 14.5 dpc embryo is unknown. Therefore, a negative result does not necessarily mean that a gene is not expressed in vivo; false negatives cannot be excluded. In our experience, among all the genes annotated in the testis, 83% of the in situ hybridization results with annotated testis expression were considered “publication quality” (see Section II). However, in some cases we found that expression was often detected even though it was not annotated. If the structure of interest is not annotated as “expressed” or the assay is classified as “not detected”, we recommend the readers to look through the entire slide deck.

Eurexpress and other expression databases allow basic and advanced queries based on annotated anatomy, gene name, gene symbol, template, and gene sequence. Entries in the Eurexpress database are linked to other databases, such as the ABA [Jones et al., 2009; Lein et al., 2007], EMAGE [Christiansen et al., 2006; Ringwald et al., 1994], and the GENSAT [Gong et al., 2003], and to informational resources such as Entrez Gene (https://www.ncbi.nlm.nih.gov/gene/), ENSEMBL (http://useast.ensembl.org/index.html), and MGI. The key drawback of these atlases is the lack of the ability to process a batch of genes generated by approaches such as microarrays and RNA-seq. The lack of an option to do bulk queries on a gene list, instead of on individual genes, in many of these databases means that the biologist has to manually go through digital slide decks for each gene to find the structures of interest. Although structures typically occur in the same 3–5 slice window of the 24 slice slide deck, each slice image must then be manually downloaded at high resolution and processed. Currently, the most efficient way of examining batches of genes within most expression databases is to automatically extract information from the webpage rather than displaying it to the investigator using a third party program. This way the queried image files for each gene can be downloaded as a batch; the images can then be scrolled through on a host computer. The introduction of a tool that allows users submit lists of genes and return batch searches within the database will make accessing the information stored within expression databases far easier and increase the utility of expression databases for biologists. Until this kind of tool is implemented, the investigator has two choices: 1) harvest/extract the image files from the database website by writing an automation script or, 2) go through the browser manually. Even if the investigator choses the more laborious manual approach, there is great value to be gained from trawling through expression atlases.

Transcriptomic databases (for non-bioinformaticians)

The abundance of transcriptomic data stored in the Gene Expression Omnibus (GEO, a public functional genomics data repository) provides many opportunities for the investigator; however, many biologists do not have the programming skills to exploit this resource. Luckily, there are some useful databases that can assist the biologist querying transcriptomic data. Transcriptomic microarray data covering various stages of gonad development and a variety of enriched cell types has been generated by a number of groups [McClelland et al., 2015; Inoue et al., 2015; Jameson et al., 2012]. These data complement the spatiotemporal expression data in databases like Eurexpress. It is now standard for all the raw data for these types of studies to be deposited in GEO, so they are available for reanalysis. However, mining published transcriptomic data (such as microarray and RNA-seq) can be challenging without appropriate analysis frameworks. Many “at the bench researchers” lack the expertise to extensively mine the RNA-seq and microarray data stored in GEO. This means researchers are not able to fully utilize published data in their field.

To use data from primary archives, a certain level of expertise is needed — raw or processed data must be downloaded, and then the data can be analyzed independently, or in combination with other data. For microarray data many biologists have some expertise in using software programs such as Partek to analyze and reanalyze data from CEL files; however RNA-seq data is more complicated to analyze. Added-value databases make the biological content of the expression data more accessible to non-bioinformaticians. These tools aggregate data stored in repositories such as GEO by extracting relevant information from the raw primary data, and therefore allow the user to ask biological questions through a user-friendly interface. For example, the user can determine in which samples their gene-of-interest is expressed, or which genes are differentially expressed between the two samples without having to handle the raw data directly.

One of the key added-value aggregators of transcriptomic information is the Gene Expression Atlas (http://www.ebi.ac.uk/gxa; [Kapushesky et al., 2009]), which provides information about gene expression in different cell types and organs, in addition to different developmental stages, disease states and biological/experimental conditions. This atlas has expression data from a large number of species including all common model organisms. The user can query individual genes looking for differential gene expression by gene names or by searching for genes correlated with an attribute such as cell types. For example, Sox9/SOX9 expression can be queried in a general (to pull up all available species) or in a specific species (Fig. 1C & D). In mouse, expression data is available for a variety of embryonic and adult stages from the FANTOM 49 Consortium project and other individual projects. Examining expression in the testis reveals expression during embryonic development, specifically at E16 or 16.5 dpc expression in the testis, pancreas, kidney and other organs (purple in Fig. 1C). In human, expression data is available for a number of Consortiums, including FANTOM 65, which examines expression in adult tissues. SOX9 is expressed in the testis, skin and a number of distinct brain regions (purple in Fig. 1D).

Pathway analysis: GO term enrichment analysis

Pathway analysis is one of the most biologist-accessible ways to look for patterns in candidate gene lists. Many similar publicly available analysis software and tools that were developed in the early 2000s can be used to functionally analyze large genes-of-interest lists [Huang et al., 2009b; Khatri et al., 2012]. However, many of these databases are no longer fully updated and maintained; for this reason, out of the freely available tools we recommend DAVID (Database for Annotation, Visualization and Integrated Discovery; Table 1; [Huang et al., 2007; Huang et al., 2009b]) and GSEA (Gene Set Enrichment Analysis; Table 1; [Subramanian et al., 2005]). DAVID is currently upgrading: v6.8 with updated knowledgebase will be available on October 17, 2016 (data here is analyzed in v6.8Beta), and v6.7 will be available for continued use until January 15, 2017. GSEA and DAVID use the same core approach of searching a genes-of-interest list, and then systematically map the list against a bank of GO terms in order to identify the most overrepresented or enriched terms out of all the linked terms that associate with the genes on the list. This kind of enrichment analysis strategy allows investigators to identify biological patterns and processes that may be relevant to their area of study that would never be discovered by looking at the list with the naked eye. Knowledgebases, such as DAVID and GSEA, draw on different repositories (including NCBI, https://www.ncbi.nlm.nih.gov/) and UniProt (http://www.uniprot.org/)), therefore hosting multiple GO terms for a single gene in an attempt to increase the comprehensiveness of the query output. A single gene can be mapped to many different and redundant terms, just as a single term maps to many genes. DAVID deals with this GO term redundancy by clustering and classifying the redundant terms into themed “annotation clusters” that can be searched by the user.

When submitting a query to a pathway analysis tool, the lists of genes can be in a number of formats. DAVID’s flexible input allows for a broad range of identifiers to be uses as initial search criteria. However, it is recommended that unique universal gene identifiers are used as inputs so that redundant gene names are not confused. GSEA also has a series of curated gene lists that cover a curated range of gene sets drawn from chemical and genetic perturbation experiments from PubMed, in addition to genes that share conserved cis-regulatory motifs [Xie et al., 2005] and transcription factor targets (using TRANSFAC, BioBase licensed through Qiagen; http://www.gene-regulation.com/pub/databases.html; [Matys et al., 2003; Wingender et al., 2000]). In both GSEA and DAVID the output or results of the search can be viewed and exported in a variety of ways. Be aware that these tools will restart to the main page after a period of inactivity, regardless of which step the analysis is paused at (you can set the timeout in GSEA, but for DAVID it is automatically set at 20 min inactivity). Both tools have features that link pathways and genes to disease associations, in addition to more protein-based tools that highlight protein functional domains and motifs.

For enrichment analysis it is important to remember that size matters: a larger gene list generally results in higher statistical confidence and more significant P-values in lowly enriched terms and more specific ontological terms at the ends of the ontological branches. Conversely, the broader and more general terms ontological terms are less enriched. The effect of list size on the absolute enrichment P-values means that it is not recommended that the users directly compare the absolute enrichment P-values across gene lists [Huang et al., 2009c]. In cases where well under 100 genes are in the genes-of-interest list, tools like DAVID can still be used. But using the statistical P-values as metrics of significance must be used cautiously, as the statistical power behind the enrichment analysis is limited by the small number of genes [Huang et al., 2009a]. Searching such a small list will produce a very focused list of ontologies and annotations that can be thoroughly explored by the user; however, the statistics produced by the software are largely meaningless.

The basis of enrichment analysis is that there are differences between the biological processes in an abnormal or perturbed state (or in the case of developmental biology often a different cell type or time point or genotype). The assumption is that the co-functioning genes (or related genes, which can be determined by looking at GO terms) should be enriched together and that these terms will therefore be selected as a relevant or significant group [Huang et al., 2009b; Huang, and Yao, 2010; Huang et al., 2007; Huang et al., 2009a]. The degree of enrichment depends on the background of gene expression, or the noise. The background is a key factor that can influence the conclusions drawn from the data and the certainty with which we can say genes are enriched [Huang et al., 2009a]. The background must be set up to perform the comparison in tools such as DAVID and should be carefully considered. The background gene set for analysis should only include the genes that have a chance of being selected. For this reason, choosing the whole genome when the whole genome is not represented on the Affymetrix Chip can skew the data. Similarly, when analyzing RNA-seq data, including all genes in the genome instead of discarding those genes for which no counts were recorded will skew the data [Huang et al., 2009a]. Carefully selecting a background should be a priority for each study. Customized Chip background lists are available for all commonly used microarray platforms and individualized background lists can be easily imported to meet the user’s individual needs.

Mouse knockout databases

In order to build a comprehensive functional catalogue of the mammalian genome, a collaboration was launched to create a comprehensive library of knock-out/conditional allele mouse models for researchers to utilize. In mice, this project has been spearheaded by the International Knockout Mouse Consortium (IKMC; https://www.mousephenotype.org/; [Bradley et al., 2012]). The aim of the IKMC is to generate targeted ES-cells of all known protein coding genes in mice and companion Cre driver lines [Rosen et al., 2015]. The current design used for targeting vector allows the production of reporter, conditional and knockout alleles, and provides researchers with flexibility in the design of their experiments and the ability to complement CRISPR/Cas9 strategies [Rosen et al., 2015]. The International Mouse Phenotyping Consortium (IMPC; www.knockoutmouse.org; [Ring et al., 2015]) builds on the work of the IKMC to generate the mouse strains and perform standardized phenotyping. The production of mouse strains from these ES cells are tracked and this information is freely available to the research community. The IKMC/IMPC web portals were merged to create a central hub (http://www.mousephenotype.org; [Rosen et al., 2015]) that has detailed information about the available resources including a catalogue of the targeting vectors, targeted alleles, ES cell clones, and mutant mouse strains generated and links out to other repositories. All gene trap alleles are housed at Jax (http://www.informatics.jax.org/allele). In addition, the International Mouse Strain Resource (IMSR; www.findmice.org; [Eppig et al., 2015; Eppig, and Strivens, 1999]) has a searchable catalogue of over 2000 Cre strains produced by the scientific community that links to the MGD and the repository holding the material.

At least 30% of the targeted knockouts generated in mouse by programs such as the IKMC and IMPC result in embryonic or perinatal death [Adams et al., 2013]. This lead to the inception of the Wellcome Trust-funded research program Deciphering the Mechanisms of Developmental Disorders (DMDD; https://dmdd.org.uk; [Mohun et al., 2013; Adams et al., 2013; Wilson et al., 2016]) that aims to characterize these lethal mutants further. The DMDD focuses on phenotyping embryonic lethal mutants to shed light on the genetic regulation of tissue differentiation, organ formation and embryo morphogenesis [Adams et al., 2013]. This resource is designed for developmental biologists and clinicians; it also complements existing United Kingdom clinical programs focused on better understanding low-frequency and rare genetic changes leading to human disease, such as, the Deciphering Developmental Disorders (https://www.ddduk.org) and UK10K (http://www.uk10k.org) projects.

Human disease variant databases

Online Mendelian Inheritance in Man (OMIM; http://www.omim.org/; [Amberger et al., 2011; Amberger et al., 2009; Hamosh et al., 2005]) is a resource, published since the 1960s, aimed at cataloguing known human disease variants and improving disease classification with a focus on diseases that have a significant genetic basis. One of the goals of the OMIM is to develop a standard nomenclature for features of a disorder (traits) through the Human Phenotype Ontology [Amberger et al., 2011]. OMIM is a curated resource based exclusively on the biomedical literature that links to genomic databases and model organism information, as well as other clinical resources. OMIM serves both molecular biologists and healthcare providers by classifying disorders and biological variation reported in the literature[Amberger et al., 2011]. Entries in OMIN can be classified under the phenotype or the gene. For example, the gene NR5A1 (also known as SF1), resides on Chromosome 9q33.3 (Fig. 1E). Mutations in this gene result in four known phenotypes: 46XY sex reversal, premature ovarian failure, spermatogenic failure, and adrenocortical insufficiency (Fig. 1E). The allelic variants associated with each phenotype listed for a gene are briefly described under the gene entry. More detailed information about the disorder is provided in a separate descriptive entry corresponding to the phenotype (not a unique locus), this entry has a separate identifier. The entry number and mapping key encode additional information about the disorder; more information about the assignment of different MIM numbers can be found at http://www.omim.org/help/faq.

OMIM now facilitates a series of more advanced search options, such as retrieval of similar concepts, clinical or anatomical features. Currently the user is still restricted to querying a single gene at a time through the OMIM interface. However, OMIM does actively encourage the large-scale mining of its repository; API (Application Program Interface) access to the entire OMIM repository (updated nightly) is freely accessible for individual research use with a reasonable fair-use license signed upon download (https://omim.org/api/). Once downloaded, tools such as “R” can be used to query batches of genes using gene names. For example, users can query all the genes on their genes-of-interest list against the OMIM database and pull out features such as the gene name, MIM number and OMIM description into a searchable Excel file. The OMIM descriptions can then be searched as a batch for pathologies and disorders of interest.

The Human Gene Mutation Database (HGMD) represents the other comprehensive collection of mutations that underlie or are associated with human inherited disease. From its inception the HGMD (which is run out of Cardiff University, Wales, UK) was part of a commercial agreement meaning it runs two versions. The online HGMD version that is “free” for academics (http://www.hgmd.cf.ac.uk/ac/index.php; [Stenson et al., 2008; Cooper et al., 2006]) is available via the Cardiff University and a licensed version of the database, “HGMD Professional” (see Table 1) is available through QIAGEN. Newly added mutational information is available to paid users for 2.5 years from the date of initial inclusion in the database before it can be accessed in the free academic resource [Stenson et al., 2009]. Therefore, we recommend using OMIM if you do not have access to the HGMD Professional version of the database.

II. Using databases as a discovery tool to construct a genes-of-interest list for testis development

In addition to interrogate genes-of-interest lists, databases can also be used as a discovery tool to identify new candidate genes. By querying annotated “testis expressed” genes generated by the Eurexpress expression screen, we identified a pseudo-candidate list with genes expressed in the testis at 14.5 dpc. This list contains 298 entries with 289 protein-coding genes, 6 unannotated transcripts, and 3 microRNAs (see Table S1). Among the 289 annotated protein encoding genes, 34 of them are still listed under Rik ID numbers (numbered genes ending with ‘Rik’ are annotated genes without a canonical name yet) although the gene has subsequently been renamed. To ensure the remaining 255 genes are listed under the current approved gene names, we ran them through the HGNC (HUGO Gene Nomenclature Committee) Multi-Symbol Checker (http://www.genenames.org/cgi-bin/symbol_checker). This tool checks all the names in the submitted list against HUGO verified names and their known synonyms. This tool does not cross reference Rik IDs, we therefore cross-referenced the 34 genes under Rik IDs in the MGI database to locate the current gene identifier (Table S1). As many of the databases began cataloguing entries over 15 years ago, not all gene are listed under the current HUGO approved name. This means that the investigator may overlook entries that remain curated under defunct synonyms of IDs. Once we confirmed/identified the gene names for all 295 genes (289 protein-encoding genes + 6 unannotated transcripts), we obtained the in situ hybridization images, and score them on an arbitrary metric of 1–4 (1-very high publication quality; 2-high publication quality; 3-moderate quality (requires confirmation); 4-low quality (not publication quality)). For every gene that scored a 1 or 2 (142 genes; Table S2), we downloaded a representative full embryo image containing the testis and cropped a 600 × 600 pixel section containing the testis in Photoshop. In most cases the first testis of the pair appeared in slice 3–7 of the slide deck (before the kidney) whereas the second testis appeared between slide 11–21 (after the kidney).

Table S1: A full list of 298 entries under the annotation of expressed in the “testis” from the Eurexpress database.

This table contains the complete list of entries listed in the Eurexpress databases under the anatomical annotation of “testis” for embryos at 14.5 dpc on May 1, 2016. Non-HUGO approved symbols listed in Eurexpress have been replaced with HUGO approved symbols for data analysis.

HUGO approved gene name	Non-HUGO symbols listed in Eurexpress	HUGO approved name
1700011H14Rik	1700011H14Rik	*
Aatf		Apoptosis Antagonizing Transcription Factor
Abhd17c	2210412D01Rik	Abhydrolase Domain Containing 17C
Abl2		ABL Proto-Oncogene 2, Non-Receptor Tyrosine Kinase
Abtb2		Ankyrin Repeat And BTB (POZ) Domain Containing 2
Acsbg1		Acyl-Coa Synthetase Bubblegum Family Member 1
Adamts16		ADAM Metallopeptidase With Thrombospondin Type 1 Motif 16
Adamts2		ADAM Metallopeptidase With Thrombospondin Type 1 Motif 2
Adh1		Alcohol Dehydrogenase 1A (Class I), Alpha Polypeptide
Ajap1		Adherens Junctions Associated Protein 1
Akr1cl	4921521F21Rik	Aldo-Keto Reductase Family 1, Member C-Like
Amh		Anti-Mullerian Hormone
Anxa11		Annexin A11
Anxa4		Annexin A4
Ap3b2		Adaptor Related Protein Complex 3 Beta 2 Subunit
Apc		APC, WNT Signaling Pathway Regulator
Apex1		Apurinic/Apyrimidinic Endodeoxyribonuclease 1
Apln		Apelin
Arl6ip2		Atlastin Gtpase 2
Art5		ADP-Ribosyltransferase 5
Ash2l		ASH2 Like Histone Lysine Methyltransferase Complex Subunit
Aspa		Aspartoacylase
Atp1b1		Atpase Na+/K+ Transporting Subunit Beta 1
Atp6v1b1		Atpase H+ Transporting V1 Subunit B1
Atp6v1e1		Atpase H+ Transporting V1 Subunit E1
Atp8b2		Atpase Phospholipid Transporting 8B2
Bbx		BBX, HMG-Box Containing
BC034902	BC034902	*
Bex2		Brain Expressed X-Linked 2
Brca1		BRCA1, DNA Repair Associated
Btf3		Basic Transcription Factor 3
Bub1		BUB1 Mitotic Checkpoint Serine/Threonine Kinase
Bub1b		BUB1 Mitotic Checkpoint Serine/Threonine Kinase B
Camkk2		Calcium/Calmodulin Dependent Protein Kinase Kinase 2
Casp6		Caspase 6
Cd34		CD34 Molecule
Cdca7l		Cell Division Cycle Associated 7 Like
Cdk5rap1		CDK5 Regulatory Subunit Associated Protein 1
Cenpj		Centromere Protein J
Cenpo	2810429O05Rik	Centromere Protein O
Cep68	BC027174	Centrosomal Protein 68
Clca1		Chloride Channel Accessory 1
Clca4		Chloride Channel Accessory 4
Clcn2		Chloride Voltage-Gated Channel 2
Cldn13		Claudin 13
Clmn		Calmin
Cobll1		Cordon-Bleu WH2 Repeat Protein Like 1
Col4a1		Collagen Type IV Alpha 1 Chain
Col9a3		Collagen Type IX Alpha 3 Chain
Copb2		Coatomer Protein Complex Subunit Beta 2
Csrp1		Cysteine And Glycine Rich Protein 1
Ctnna2		Catenin Alpha 2
Cx3cl1		C-X3-C Motif Chemokine Ligand 1
Cxadr		Coxsackie Virus And Adenovirus Receptor
Cxcl12		C-X-C Motif Chemokine Ligand 12
Cxxc6		Tet Methylcytosine Dioxygenase 1
Cyp11a1		Cytochrome P450 Family 11 Subfamily A Member 1
Cyp51		Cytochrome P450 Family 51 Subfamily A Member 1
D930030O05Rik	D930030O05Rik	*
Daam2		Dishevelled Associated Activator Of Morphogenesis 2
Dcun1d5	4833420K19Rik	DCN1, Defective In Cullin Neddylation 1, Domain Containing 5
Ddx41		DEAD-Box Helicase 41
Ddx48		Eukaryotic Translation Initiation Factor 4A3
Dnmt3b		DNA Methyltransferase 3 Beta
Dppa3		Developmental Pluripotency Associated 3
Dppa5		Developmental Pluripotency Associated 5
Drg1		Developmentally Regulated GTP Binding Protein 1
Dsp		Desmoplakin
E430025E21Rik		*
Eif2b1		Eukaryotic Translation Initiation Factor 2B Subunit Alpha
Eif2s3y		Eukaryotic Translation Initiation Factor 2, Subunit 3, Structural Gene Y-Linked
Elavl2		ELAV Like RNA Binding Protein 2
Eps8		Epidermal Growth Factor Receptor Pathway Substrate 8
Eps8l2		EPS8 Like 2
Ets1		ETS Proto-Oncogene 1, Transcription Factor
Etv5		ETS Variant 5
Fads1		Fatty Acid Desaturase 1
Fam214a	BC031353	Family With Sequence Similarity 214, Member A
Fasn		Fatty Acid Synthase
Fastkd2	2810421I24Rik	FAST Kinase Domains 2
Fbrsl1	2410025L10Rik	Fibrosin-Like 1
Fiz1		FLT3 Interacting Zinc Finger 1
Fosl2		FOS Like 2, AP-1 Transcription Factor Subunit
Foxn2		Forkhead Box N2
Frzb		Frizzled-Related Protein
Fthfd		Aldehyde Dehydrogenase 1 Family Member L1
Fzd3		Frizzled Class Receptor 3
Gas6		Growth Arrest Specific 6
Gem		GTP Binding Protein Overexpressed In Skeletal Muscle
Gjb2		Gap Junction Protein Beta 2
Gmnn		Geminin, DNA Replication Inhibitor
Gnl3		G Protein Nucleolar 3
Gpc4		Glypican 4
Gpr37		G Protein-Coupled Receptor 37
Gpr56		Adhesion G Protein-Coupled Receptor G1
Gprk5		G Protein-Coupled Receptor Kinase 5
Gramd1b		GRAM Domain Containing 1B
Grrp1		Family With Sequence Similarity 110 Member D
Gsta4		Glutathione S-Transferase Alpha 4
Gstm1		Glutathione S-Transferase Mu 1
Gstm4		Glutathione S-Transferase Mu 4
Gstm5		Glutathione S-Transferase Mu 5
Gstm7		Glutathione S-Transferase, Mu 7
Gtf2b		General Transcription Factor IIB
Gtf2e2		General Transcription Factor IIE Subunit 2
Gtf3c3		General Transcription Factor IIIC Subunit 3
Gucy1a3		Guanylate Cyclase 1 Soluble Subunit Alpha
Gucy1b3		Guanylate Cyclase 1 Soluble Subunit Beta
Hadhsc		Hydroxyacyl-Coa Dehydrogenase
Hat1		Histone Acetyltransferase 1
Heatr1	B130016L12Rik	HEAT Repeat Containing 1
Hmg20a		High Mobility Group 20A
Hmga1		High Mobility Group AT-Hook 1
Hoxd9		Homeobox D9
Hs6st1		Heparan Sulfate 6-O-Sulfotransferase 1
Hsd17b7		Hydroxysteroid 17-Beta Dehydrogenase 7
Hsd17b7		Hydroxysteroid 17-Beta Dehydrogenase 7 Pseudogene 2
Hsd3b3		Hydroxy-Delta-5-Steroid Dehydrogenase, 3 Beta- And Steroid Delta- Isomerase 3
Hsp90aa1		Heat Shock Protein 90 Alpha Family Class A Member 1
Hspa1a		Heat Shock Protein Family A (Hsp70) Member 1A
Hspa2		Heat Shock Protein Family A (Hsp70) Member 2
Iah1	4833421E05Rik	Isoamyl Acetate-Hydrolyzing Esterase 1 Homolog
Idh1		Isocitrate Dehydrogenase (NADP(+)) 1, Cytosolic
Idi1		Isopentenyl-Diphosphate Delta Isomerase 1
Ier5l		Immediate Early Response 5 Like
Ifi27	D12Ertd647e	Interferon, Alpha-Inducible Protein 27
Ifitm7
Immt		Inner Membrane Mitochondrial Protein
Irf1		Interferon Regulatory Factor 1
Itga9		Integrin Subunit Alpha 9
Itgb8		Integrin Subunit Beta 8
Kcnt1		Potassium Sodium-Activated Channel Subfamily T Member 1
Kctd14		Potassium Channel Tetramerization Domain Containing 14
Kif2c		Kinesin Family Member 2C
Kirrel2		Kin Of IRRE Like 2 (Drosophila)
L3mbtl		L(3)Mbt-Like 1 (Drosophila)
Lba1		Tetratricopeptide Repeat And Ankyrin Repeat Containing 1
Lin28		Lin-28 Homolog A
Liph		Lipase H
Lisch7		Lipolysis Stimulated Lipoprotein Receptor
LOC620538		*
LOC673219		*
Lyar		Ly1 Antibody Reactive
Mafg		MAF Bzip Transcription Factor G
Mcm8		Minichromosome Maintenance 8 Homologous Recombination Repair Factor
Melk		Maternal Embryonic Leucine Zipper Kinase
Mettl25	BC067068	Methyltransferase Like 25
Mmp8		Matrix Metallopeptidase 8
mmu-miR-291a-5p		^
mmu-miR-291b-5p		^
mmu-miR-696		^
Mod1		Chromobox 1
Mphosph6		M-Phase Phosphoprotein 6
Mpp3		Membrane Palmitoylated Protein 3
Mrpl23		Mitochondrial Ribosomal Protein L23
Mrpl24		Mitochondrial Ribosomal Protein L24
Mrps27		Mitochondrial Ribosomal Protein S27
Mtif2		Mitochondrial Translational Initiation Factor 2
Mybbp1a		MYB Binding Protein 1a
Mybl1		MYB Proto-Oncogene Like 1
Nat10		N-Acetyltransferase 10
Ndufa9		NADH:Ubiquinone Oxidoreductase Subunit A9
Ndufab1		NADH:Ubiquinone Oxidoreductase Subunit AB1
Ndufs1		NADH:Ubiquinone Oxidoreductase Core Subunit S1
Ndufs3		NADH:Ubiquinone Oxidoreductase Core Subunit S3
Ndufv1		NADH:Ubiquinone Oxidoreductase Core Subunit V1
Nepro	BC027231	Nucleolus And Neural Progenitor Protein
Nfe2		Nuclear Factor, Erythroid 2
Npr1		Natriuretic Peptide Receptor 1
Nr4a1		Nuclear Receptor Subfamily 4 Group A Member 1
Nrg1		Neuregulin 1
Nsmaf		Neutral Sphingomyelinase Activation Associated Factor
Nts		Neurotensin
Nupl1		Nucleoporin 58
Nutf2		Nuclear Transport Factor 2
Orc5l		Origin Recognition Complex Subunit 5
Osbpl10		Oxysterol Binding Protein Like 10
Parm1	9130213B05Rik	Prostate Androgen-Regulated Mucin-Like Protein 1
Pcanap6		Solute Carrier Family 45 Member 3
Pcmtd1	A030012M09Rik	Protein-L-Isoaspartate (D-Aspartate) O-Methyltransferase Domain Containing 1
Pcsk6		Proprotein Convertase Subtilisin/Kexin Type 6
Pdgfa		Platelet Derived Growth Factor Subunit A
Pdzd4		PDZ Domain Containing 4
Per2		Period Circadian Clock 2
Pgpep1		Pyroglutamyl-Peptidase I
Pink1		PTEN Induced Putative Kinase 1
Pip5k2a		Phosphatidylinositol-5-Phosphate 4-Kinase Type 2 Alpha
Pla2g5		Phospholipase A2 Group V
Plekha1		Pleckstrin Homology Domain Containing A1
Plekha2		Pleckstrin Homology Domain Containing A2
Polr2h		RNA Polymerase II Subunit H
Polr2l		RNA Polymerase II Subunit L
Polr3f		RNA Polymerase III Subunit F
Por		Cytochrome P450 Oxidoreductase
Por		Porcupine Homolog (Drosophila)
Pou6f1		POU Class 6 Homeobox 1
Ppm1a		Protein Phosphatase, Mg2+/Mn2+ Dependent 1A
Ppt1		Palmitoyl-Protein Thioesterase 1
Prkar2b		Protein Kinase Camp-Dependent Type II Regulatory Subunit Beta
Prlr		Prolactin Receptor
Prlr		Prolactin Receptor
Prss15		Lon Peptidase 1, Mitochondrial
Psma1		Proteasome Subunit Alpha 1
Ptpro		Protein Tyrosine Phosphatase, Receptor Type O
Ptpro		Protein Tyrosine Phosphatase, Receptor Type U
Rad52b		RAD52 Motif Containing 1
Rbbp7		RB Binding Protein 7, Chromatin Remodeling Factor
Rcc2		Regulator Of Chromosome Condensation 2
Rcl1		RNA Terminal Phosphate Cyclase Like 1
Reep2	BC020184	Receptor Accessory Protein 2
Ren1
Rgs11		Regulator Of G-Protein Signaling 11
Rhpn1		Rhophilin Rho Gtpase Binding Protein 1
Rmdn2		Regulator Of Microtubule Dynamics 2
Rnf138		Ring Finger Protein 138
Rnf181	2500002L14Rik	Ring Finger Protein 181
Rnf213	D11Ertd759e	Ring Finger Protein 213
Rnf34		Ring Finger Protein 34
Rnmt		RNA Guanine-7 Methyltransferase
Rpp30		Ribonuclease P/MRP Subunit P30
Rps27l		Ribosomal Protein S27 Like
Sc4mol		Methylsterol Monooxygenase 1
Scarb1		Scavenger Receptor Class B Member 1
Schip1		Schwannomin Interacting Protein 1
Scrn2		Secernin 2
Sct		Secretin
Sec11l1		SEC11 Homolog A, Signal Peptidase Complex Subunit
Senp8		SUMO/Sentrin Peptidase Family Member, NEDD8 Specific
Serpinb6a		Serine (Or Cysteine) Peptidase Inhibitor, Clade B, Member 6a
Sgk3		Serum/Glucocorticoid Regulated Kinase Family Member 3
Sil1		SIL1 Nucleotide Exchange Factor
Slc16a1		Solute Carrier Family 16 Member 1
Slc20a1		Solute Carrier Family 20 Member 1
Slc25a31		Solute Carrier Family 25 Member 31
Slc25a39	D11Ertd333e	Solute Carrier Family 25, Member 39
Slc29a1		Solute Carrier Family 29 Member 1 (Augustine Blood Group)
Slc38a1		Solute Carrier Family 38 Member 1
Slc44a1		Solute Carrier Family 44 Member 1
Slc7a5		Solute Carrier Family 7 Member 5
Smarca4		SWI/SNF Related, Matrix Associated, Actin Dependent Regulator Of Chromatin, Subfamily A, Member 4
Smoc1		SPARC Related Modular Calcium Binding 1
Snrpn		Small Nuclear Ribonucleoprotein Polypeptide N
Snw1		SNW Domain Containing 1
Sostdc1		Sclerostin Domain Containing 1
Sox2		SRY-Box 2
Sox4		SRY-Box 4
Spice1	D16Ertd480e	Spindle And Centriole Associated Protein 1
Srfbp1		Serum Response Factor Binding Protein 1
Srpx2		Sushi Repeat Containing Protein, X-Linked 2
Star		Steroidogenic Acute Regulatory Protein
Star		Guanylate Cyclase 2C
Star		Steroidogenic Acute Regulatory Protein
Stc1		Stanniocalcin 1
Stc2		Stanniocalcin 2
Stk25		Serine/Threonine Kinase 25
Stxbp2		Syntaxin Binding Protein 2
Tacr3		Tachykinin Receptor 3
Tacstd1		Epithelial Cell Adhesion Molecule
Taf7		TATA-Box Binding Protein Associated Factor 7
Tbc1d8		TBC1 Domain Family Member 8
Tcf21		Transcription Factor 21
Tde1		Serine Incorporator 3
Tdrd12	2410004F06Rik	Tudor Domain Containing 12
Tesc		Tescalcin
Tex2		Testis Expressed 2
Tex261		Testis Expressed 261
Tgif		TGFB Induced Factor Homeobox 1
Timm8b		Translocase Of Inner Mitochondrial Membrane 8 Homolog B
Tle6		Transducin Like Enhancer Of Split 6
Tmcc3		Transmembrane And Coiled-Coil Domain Family 3
Tmem22		Solute Carrier Family 35 Member G2
Tmem35		Transmembrane Protein 35A
Tmem64	9630015D15Rik	Transmembrane Protein 64
Tmem86a	1810054O13Rik	Transmembrane Protein 86A
Tpd52		Tumor Protein D52
Tpr		Translocated Promoter Region, Nuclear Basket Protein
Trim71		Tripartite Motif Containing 71
Trip13		Thyroid Hormone Receptor Interactor 13
Trpv2		Transient Receptor Potential Cation Channel Subfamily V Member 2
Trrap		Transformation/Transcription Domain Associated Protein
Ttc39c	2810439F02Rik	Tetratricopeptide Repeat Domain 39C
Txnrd2		Thioredoxin Reductase 2
Tyro3		TYRO3 Protein Tyrosine Kinase
Ube2t		Ubiquitin Conjugating Enzyme E2 T
Uchl1		Ubiquitin C-Terminal Hydrolase L1
Uchl5		Ubiquitin C-Terminal Hydrolase L5
Usf1		Upstream Transcription Factor 1
Utrn		Utrophin
Wdr10		Intraflagellar Transport 122
Wdr21		DDB1 And CUL4 Associated Factor 4
Wdr31		WD Repeat Domain 31
Wdr36		WD Repeat Domain 36
Xrcc6		X-Ray Repair Cross Complementing 6
Yars2		Tyrosyl-Trna Synthetase 2
Zdhhc24		Zinc Finger DHHC-Type Containing 24
Zfp451		Zinc Finger Protein 451
Zfp516		Zinc Finger Protein 516

^*indicates that a transcript but no gene is annotated (6 genes).

^{^}indicates a pre-miRNA transcript.

Most analysis was performed on a list of 295 genes excluding the 3 pre-miRNAs or 289 genes when including only known protein-coding transcripts.

Table S2: 142 Genes Annotated As Expressed In The “Testis” Ranked 1 Or 2 (Publication Quality) In The Eurexpress Databases.

Of the 295 transcript entries (excluding the 3 pre-miRNAs ) in the Eurexpress database, 142 were considered publication quality. We scored the in situ hybridization images for each of the genes on a metric of 1–4. 1-very high publication quality; 2-high publication quality; 3-moderate quality (reconsider as publication quality); 4-low quality (not publication quality). Genes scoring a 1 or 2 are listed in this table. These data were obtained from Eurexpress Transcriptome Atlas Database for Mouse Embryo (http://www.eurexpress.org) on May 1, 2016.

Gene	Chromosome	Gene Name
Acsbg1	9	Acyl-Coa Synthetase Bubblegum Family Member 1
Adamts16	13	A Disintegrin-Like And Metallopeptidase (Reprolysin Type) With Thrombospondin Type 1 Motif, 16
Adamts2	11	A Disintegrin-Like And Metallopeptidase (Reprolysin Type) With Thrombospondin Type 1 Motif, 2
Adh1	3	Alcohol Dehydrogenase 1 (Class I)
Ap3b2	7	Adaptor-Related Protein Complex 3, Beta 2 Subunit
Apc	18	Adenomatosis Polyposis Coli
Apex1	14	Apurinic/Apyrimidinic Endonuclease 1
Apln	X	Apelin
Arl6ip2	17	Atlastin Gtpase 2
Atp6v1e1	6	Atpase, H+ Transporting, Lysosomal V1 Subunit E1
Atp8b2	3	Atpase, Class I, Type 8b, Member 2
Bex2	X	Brain Expressed X-Linked 2
Bub1	2	Bub1, Mitotic Checkpoint Serine/Threonine Kinase
Cd34	1	Cd34 Antigen
Cdca7l	12	Cell Division Cycle Associated 7 Like
Cdk5rap1	2	Cdk5 Regulatory Subunit Associated Protein 1
Clca1	3	Chloride Channel Accessory 1
Clca4	3	Chloride Channel Accessory 4a/4b
Clcn2	16	Chloride Channel, Voltage-Sensitive 2
Col9a3	2	Collagen, Type Ix, Alpha 3
Copb2	9	Coatomer Protein Complex, Subunit Beta 2 (Beta Prime)
Ctnna2	6	Catenin (Cadherin Associated Protein), Alpha 2
Cxadr	16	Coxsackie Virus And Adenovirus Receptor
Cxcl12	6	Chemokine (C-X-C Motif) Ligand 12
Cyp11a1	9	Cytochrome P450, Family 11, Subfamily A, Polypeptide 1
Cyp51	5	Cytochrome P450, Family 51
Rnf213	11	Ring Finger Protein 213 (Rnf213)
Ifi27	12	Interferon, Alpha-Inducible Protein 27 (Ifi27 )
Spice1	16	Spindle And Centriole Associated Protein 1 (Spice1 )
Daam2	17	Dishevelled Associated Activator Of Morphogenesis 2
Dppa3	6	Developmental Pluripotency-Associated 3
Dsp	13	Desmoplakin
Elavl2	4	Elav (Embryonic Lethal, Abnormal Vision, Drosophila)-Like 2 (Hu Antigen B)
Ets1	9	E26 Avian Leukemia Oncogene 1, 5' Domain
Etv5	16	Ets Variant 5
Fads1	19	Fatty Acid Desaturase 1
Fthfd	6	Aldehyde Dehydrogenase 1 Family, Member L1
Gas6	8	Growth Arrest Specific 6
Gem	4	Gtp Binding Protein (Gene Overexpressed In Skeletal Muscle)
Gjb2	14	Gap Junction Protein, Beta 2
Gnl3	14	Guanine Nucleotide Binding Protein-Like 3 (Nucleolar)
Gpc4	X	Glypican 4
Adgrg1	8	Adhesion G Protein-Coupled Receptor G1 (Adgrg1 )
Gramd1b	9	Gram Domain Containing 1b
Grrp	4	Glycine/Arginine Rich Protein 1
Gsta4	9	Glutathione S-Transferase, Alpha 4
Gstm1	3	Glutathione S-Transferase, Mu 1
Gstm4	3	Glutathione S-Transferase, Mu 4
Gstm7	3	Glutathione S-Transferase, Mu 7
Gtf2b	3	General Transcription Factor Iib
Gtf2e2	8	General Transcription Factor Ii E, Polypeptide 2 (Beta Subunit)
Gtlf3b	11	N-Acetyltransferase Domain Containing 1
Gucy1a3	3	Guanylate Cyclase 1, Soluble, Alpha 3
Hat1	2	Histone Aminotransferase 1
Hmg20a	9	High Mobility Group 20a
Hmga1	17	High Mobility Group At-Hook 1
Hoxd9	2	Homeobox D9
Hs6st1	1	Heparan Sulfate 6-O-Sulfotransferase 1
Hsd17b7	1	Hydroxysteroid (17-Beta) Dehydrogenase 7
Hsd3b3	3	Hydroxy-Delta-5-Steroid Dehydrogenase, 3 Beta- And Steroid Delta-Isomerase 3
Hsp90aa1	12	Heat Shock Protein 90, Alpha (Cytosolic), Class A Member 1
Hspa1a	17	Heat Shock Protein 1a
Idi1	13	Isopentenyl-Diphosphate Delta Isomerase
Ifitm7	16	Interferon Induced Transmembrane Protein 7
Irf1	11	Interferon Regulatory Factor 1
Itga9	9	Integrin Alpha 9
Kcnt1	2	Potassium Channel, Subfamily T, Member 1
L3mbtl	2	L(3)Mbt-Like (Drosophila)
Trank1	9	Tetratricopeptide Repeat And Ankyrin Repeat Containing 1 (Trank1 )
Liph	16	Lipase, Member H
Lsr	7	Lipolysis Stimulated Lipoprotein Receptor (Lsr)
Melk	4	Maternal Embryonic Leucine Zipper Kinase
Mmp8	9	Matrix Metallopeptidase 8
Me1	9	Malic Enzyme 1, Nadp(+)-Dependent, Cytosolic (Me1)
Mphosph6	8	M Phase Phosphoprotein 6
Mrpl23	7	Mitochondrial Ribosomal Protein L23
Mybbp1a	11	Myb Binding Protein (P160) 1a
Npr1	3	Natriuretic Peptide Receptor 1
Nrg1	8	Neuregulin 1
Nts	10	Neurotensin
Nupl1	14	Nucleoporin Like 1
Osbpl10	9	Oxysterol Binding Protein-Like 10
Slc45a3	1	Solute Carrier Family 45, Member 3 (Slc45a3)
Pcsk6	7	Proprotein Convertase Subtilisin/Kexin Type 6
Pdzd4	X	Pdz Domain Containing 4
Pgpep1	8	Pyroglutamyl-Peptidase I
Pla2g5	4	Phospholipase A2, Group V
Plekha1	7	Pleckstrin Homology Domain Containing, Family A (Phosphoinositide Binding Specific) Member 1
Polr3f	2	Polymerase (Rna) Iii (Dna Directed) Polypeptide F
Por	5	P450 (Cytochrome) Oxidoreductase
Ppt1	4	Palmitoyl-Protein Thioesterase 1
Prkar2b	12	Protein Kinase, Camp Dependent Regulatory, Type Ii Beta
Prlr	15	Prolactin Receptor
Psma1	7	Proteasome (Prosome, Macropain) Subunit, Alpha Type 1
Rad52b	11	Rad52 Motif 1
Rcc2	4	Regulator Of Chromosome Condensation 2
Rcl1	19	Rna Terminal Phosphate Cyclase-Like 1
Ren1	1	Renin 1 Structural
Rgs11	17	Regulator Of G-Protein Signaling 11
Rhpn1	15	Rhophilin, Rho Gtpase Binding Protein 1
Rnf138	18	Ring Finger Protein 138
Msmo1	8	Methylsterol Monoxygenase 1 (Msmo1 )
Scarb1	5	Scavenger Receptor Class B, Member 1
Schip1	3	Schwannomin Interacting Protein 1
Scrn2	11	Secernin 2
Sct	7	Secretin
Serpinb6a	13	Serine (Or Cysteine) Peptidase Inhibitor, Clade B, Member 6a
Slc20a1	2	Solute Carrier Family 20, Member 1
Slc25a31	3	Solute Carrier Family 25 (Mitochondrial Carrier; Adenine Nucleotide Translocator), Member 31
Slc29a1	17	Solute Carrier Family 29 (Nucleoside Transporters), Member 1
Slc38a1	15	Solute Carrier Family 38, Member 1
Slc7a5	8	Solute Carrier Family 7 (Cationic Amino Acid Transporter, Y+ System), Member 5
Smoc1	12	Sparc Related Modular Calcium Binding 1
Snrpn	7	Small Nuclear Ribonucleoprotein N
Sostdc1	12	Sclerostin Domain Containing 1
Sox2	3	Sry (Sex Determining Region Y)-Box 2
Stc1	14	Stanniocalcin 1
Stc2	11	Stanniocalcin 2
Stxbp2	8	Syntaxin Binding Protein 2
Tacr3	3	Tachykinin Receptor 3
Epcam	17	Epithelial Cell Adhesion Molecule (Epcam)
Tbc1d8	1	Tbc1 Domain Family, Member 8
Tcf21	10	Transcription Factor 21
Tcl1	12	T Cell Lymphoma Breakpoint 1
Serinc3	2	Serine Incorporator 3 (Serinc3 )
Tesc	5	Tescalcin
Tex2	11	Testis Expressed Gene 2
Tle6	10	Transducin-Like Enhancer Of Split 6
Tmcc3	10	Transmembrane And Coiled Coil Domains 3
Tmem35	X	Transmembrane Protein 35a
Tpd52	3	Tumor Protein D52
Trip13	13	Thyroid Hormone Receptor Interactor 13
Trrap	5	Transformation/Transcription Domain-Associated Protein
Tyro3	2	Tyro3 Protein Tyrosine Kinase 3
Ube2t	1	Ubiquitin-Conjugating Enzyme E2t
Uchl1	5	Ubiquitin Carboxy-Terminal Hydrolase L1
Yars2	16	Tyrosyl-Trna Synthetase 2 (Mitochondrial)
Zdhhc24	19	Zinc Finger, Dhhc Domain Containing 24
Zfp451	1	Zinc Finger Protein 451
Zfp516	18	Zinc Finger Protein 516

We first identified previously described marker genes and validated their expression. We compared the list of 289 protein-encoding genes and 6 unannotated transcripts (a total of 295) from the Eurexpress database to published microarray data for the genes enriched in five key testicular cell types (supporting or Sertoli cells, germ cells, interstitial cells, Leydig cells, and endothelial cells; Table S3) at 13.5 dpc [Jameson et al., 2012]. For the supporting or Sertoli cell-enriched genes, among 295 testis-expressing genes from the Eurexpress database and 491 genes from the Sertoli cell enriched genes from the Jameson et al microarray, 35 genes were found in both datasets (Fig. 2A), including 5 known Sertoli cell genes Amh, Ptgds, Etv5, Tyro3 and Col9a3 (Fig. 2B). Sixteen of the 35 genes were of publication quality (Fig. 2C). For the male germ cell genes, 25 genes were shared by both the Eurexpress database and the Jameson et al microarray (Fig. 3A), including Sox2, Dazl, and Dppa3 (Fig. 3B). Among the 25 genes, 10 were of publication quality in situ hybridization images from the Eurexpress database (Fig. 3C).

Table S3: List of genes represented in both the list of Eurexpress testis annotated genes and the gene lists from Jameson et al., 2012 microarray.

To validate the expression of previously reported genes expressed in Sertoli, germ, interstitial, Leydig and endothelial cells, we compared the list of 295 entries from the Eurexpress database to published Jameson et al microarray data. Genes in both datasets are listed by cell type in this table. (295 entries is 289 protein-encoding genes and 6 unannotated transcripts). These data were obtained from Eurexpress Transcriptome Atlas Database for Mouse Embryo (http://www.eurexpress.org) on May 1, 2016.

Supporting (Sertoli) cells	Leydig cells	Mixed interstitial cells	Germ cells	Endothelial cells
Gene name	Gene name	Gene name	Gene name	Gene name
Adamts16	Cyp11a1	Clca1	Abl2	Acsbg1
Amh	Fads1	Cxcl12	Ap3b2	Nr4a1
Atp8b2	Gramd1b	Gucy1b3	Bub1b
Bex2	Hsd17b7	Itga9	Cdk5rap1
Camkk2	Prkar2b	Itgb8	Ddx41
Col9a3	Prlr	Nrg1	Dppa3
Ctnna2	Ren1	Sostdc1	Elavl2
Cxadr	Scarb1	Stc1	Epcam
Dsp	Sct	Tacr3	Gnl3
Eps8	Star		Gtf3c3
Fzd3	Stc1		Irf1
Gjb2			Liph
Gpr37			Mybbp1a
Gstm7			Nsmaf
Lsr			Pip4k2a
Mybl1			Pou6f1
Nfe2			Rcc2
Npr1			Rcl1
Pdgfa			Rdm1
Pgpep1			Srfbp1
Pla2g5			Taf7
Ppt1			Tet1
Schip1			Trip13
Sgk3			Uchl1
Sil1			Wdr31
Slc20a1
Slc38a1
Slc7a5
Stc2
Tbc1d8
Tesc
Tmcc3
Tpd52
Trank1
Tyro3

Fig. 2: Sertoli cell-expressed genes identified from “testis” annotated entries in the Eurexpress database.

(A) The list of 295 entries annotated as testis-expressed in the Eurexpress database were compared to 491 Sertoli cell expressed genes from a microarray of sorted cells at 13.5 dpc (Jameson et. al., 2012). 35 genes were represented in both sources (see Supp Table 2). (B) Out of these 35 genes, 5 known Sertoli cell marker genes were identified: Amh, Col9a3, Etv5, Ptgds, and Tyro3. 16 novel Sertoli cell candidate genes with a quality ranking of 1 or 2 were annotated in the Eurexppress database: Adamts16, Atp8b2, Ctnna2, Cxadr, Dsp, Gjb2. Gstm7, Lsr, Mpr1, Pgepep1, Slc20a1, Slc38a1, Tesc, Tmcc3, Tpd52a and Trank1. The testis region is demarcated by a dashed line. a: adrenal; k: kidney; m: mesonephros. Scale bar: 0.5 mm.

Fig. 3: Germ cell-expressed genes identified from “testis” annotated entries in the Eurexpress database.

(A) The list of 295 entries annotated as testis-expressed in the Eurexpress database were compared to 433 germ cell-expressed genes from a microarray of sorted cells at 13.5 dpc (Jameson et. al., 2012). 25 genes were represented in both sources (see Supp Table 2). (B) These genes included the known germ cell markers Sox2 and Dppa3. The germ cell marker Dazl was also included as a positive control. In the Eurexpress database, there were 10 novel germ cell genes that were also represented in the microarray: Ap3b2, Bub1, Cdkrap1, Irf1, Liph, Mybbp1a, Rcc2, Rcl1, Trip13 and Uchl1. The testis region is demarcated by a dashed line. a: adrenal; k: kidney; m: mesonephros. Scale bar: 0.5 mm.

The testis interstitium houses heterogeneous populations of cells including vasculature, steroidogenic fetal Leydig cells, and non-steroidogenic interstitial cells. In the Jameson et al microarray [2012], a Mafb-eGFP line was used to isolate both steroidogenic and non-steroidogenic interstitial cells and a Flk-mCherry line was used to isolate endothelial cells. However, expression of steroidogenic Leydig cell genes was found in cells isolated from both these lines. As a result, a mixed interstitial cell list without steroidogenic genes and a steroidogenic Leydig cell list were generated by the authors. Of the 130 mixed interstitial cell genes identified in the Jameson et al microarray [2012], 9 genes were also represented in the Eurexpress database (Fig. 4A). In addition to 4 publication quality in situ hybridization images for known interstitial marker genes (Fig. 4B), we identified images for 5 of the 9 overlapping interstitial cell genes (Fig. 4C). RNA-seq data on sorted non-steroidogenic interstitial cells and fetal Leydig cells at 12.5 dpc suggests that the expression of 3 of these genes (Clca1, Itga9 and Nrg1) may be in fetal Leydig cells [McClelland et al., 2015]. The Jameson et al microarray [2012] produced a list of Leydig cell-specific genes, 11 of which were also represented in the Eurexpress database (Fig. 4D); 8 of them are of publication quality (Fig. 4E).

Fig. 4: Interstitial and fetal Leydig cell-expressed genes identified from “testis” annotated entries in the Eurexpress database.

(A) The list of 295 entries annotated as testis-expressed in the Eurexpress database were compared to 130 interstitial cell-expressed genes from a microarray of sorted cells at 13.5 dpc (Jameson et. al., 2012). 9 genes were represented in both sources (see Supp Table 2). (C) Among these genes, 4 known interstitial/Leydig cell marker genes were identified: Cyp11a1, Prlr, Star and Tacr3. (C) high quality in situ hybridization images were available for 5 of the 9 overlapping interstitial genes: Clca1, Cxcl12, Itga9, Nrg1 and Sostdc1. (D) 11 genes were represented in both the Leydig cell microarray and the Eurexpress database (see Supp Table 2); (E) high quality in situ hybridization images were available for 8 genes: Fads1, Gramd1b, Hsd17b7, Prkar2b, Ren1, Scarb1, Sct and Stc1. . The testis region is demarcated by a dashed line. a: adrenal; k: kidney; m: mesonephros. Scale bar: 0.5 mm.

In addition to verifying microarray and other transcriptomic data, the data from expression atlases can be used to identify novel putative markers of the different cell lineages. The resolution of the Eurexpress images is sufficient for characterizing the subdivision of organs or mapping regional differences within structurally complex organs [Diez-Roux et al., 2011; McClelland et al., 2015; Yang, and Chen, 2014]. We subsequently examined the images with a quality score of a 1 or 2 that were not represented in the Jameson et al microarray and categorized them based on testis structures. Inside the testis cords, we found 28 genes with putative Sertoli cell expression and 24 genes with putative expression in germ cells. Twenty-six genes were found in the interstitium, and 9 genes were expressed in the entire testis, and 3 genes were expressed putatively in the vasculature. For the genes expressed in the testis cords, we could not determine with certainty whether the expression was localized to the Sertoli cells or the germ cells. Likewise, interstitium-expressed genes could be expressed in the fetal Leydig cells or in non-steroidogenic interstitial cells.

This list of candidate genes (295 genes) can be considered as a small randomized data set similar to an experiment examining the expression patterns of genes in a 14.5 dpc testis. As a first pass for the following pathway analysis, we used the “Tissue Expression” feature in DAVID to get a quick look at a list of reported expression patterns in different tissues for each gene [Huang et al., 2009a]. Using this tool we were able to interrogate expression patterns of 162 genes and determined that 55 genes had expression reported in the fetal or adult testis (Table S4). We then queried the complete list of genes in DAVID (V6.8Beta) to search for enriched associations among these genes. As expected, because this gene list was randomly assembled, not a complete representation of gene expression at 14.5 dpc and not directly testing a hypothesis, there were few strongly enriched processes or pathways. Subsequently, as an example, we performed pathway enrichment analysis on the list of 289 protein-encoding genes to look for enriched clusters; we identified an overrepresentation of genes involved in Glutathione metabolism (6 genes) and transcriptional regulation (39 genes) in the gene list (Table S5). When looking at a newly generated gene list, an initial submission to DAVID after validating expression of a few genes in the list can provide directions for specific niche searches. Once a pathway or process is identified as being enriched using DAVID, expression of other pathway components can be validated using expression databases and transcriptomic resources. These pathways can then be queried in OMIM to look for association with human disease. For example, from RNA-seq data profiling the interstitial cells of the fetal testis, “neuroactive ligand interaction” is overrepresented by analysis using DAVID. By identifying the receptor/ligand pairs in different testicular cell populations, a putative model could be constructed. Several components of this model had a known association with disorders of sex development [Diez-Roux et al., 2011; McClelland et al., 2015]). By reconstructing the pathways in silico, expression of the gene Frem2/FREM2 (Fras1 related extracellular matrix protein 2) and its family member Fras1 (Fraser syndrome 1 homolog), known DSD genes (OMIM:219000), were identified in the developing gonad [McClelland et al., 2015; Jadeja et al., 2005].

Table S4: “Tissue Expression” feature in DAVID identifies 55 genes with expression annotated in the testis.

162 of the Eurexpress genes had annotated expression in the DAVID “Tissue Expression” tool. Of these 162 genes, 55 had expression documented in the testis. These 55 genes are listed in this table. DAVID was accessed on September 12, 2016

Anti-Mullerian Hormone(Amh)

ADP-Ribosyltransferase 5(Art5)

G Protein-Coupled Receptor Kinase 5(Grk5)

Glutathione S-Transferase, Mu 5(Gstm5)

Homeobox D9(Hoxd9)

Hydroxysteroid (17-Beta) Dehydrogenase 7(Hsd17b7)

Heat Shock Protein 2(Hspa2)

Heat Shock Protein 90, Alpha (Cytosolic), Class A Member 1(Hsp90aa1)

Ly1 Antibody Reactive Clone(Lyar)

Myeloblastosis Oncogene-Like 1(Mybl1)

Phospholipase A2, Group V(Pla2g5)

POU Domain, Class 6, Transcription Factor 1(Pou6f1)

Prolactin Receptor(Prlr)

Secretin(Sct)

Scavenger Receptor Class B, Member 1(Scarb1)

Steroidogenic Acute Regulatory Protein(Star)

Stanniocalcin 1(Stc1)

Syntaxin Binding Protein 2(Stxbp2)

Testis Expressed Gene 2(Tex2)

Testis Expressed Gene 261(Tex261)

Utrophin(Utrn)

Proteasome (Prosome, Macropain) Subunit, Alpha Type 1(Psma1)

Serine Incorporator 3(Serinc3)

Coatomer Protein Complex, Subunit Beta 2 (Beta Prime)(Copb2)

Apoptosis Antagonizing Transcription Factor(Aatf)

Ring Finger Protein 138(Rnf138)

Sclerostin Domain Containing 1(Sostdc1)

Ring Finger Protein 181(Rnf181)

RNA (Guanine-7-) Methyltransferase(Rnmt)

Solute Carrier Family 25, Member 39(Slc25a39)

PTEN Induced Putative Kinase 1(Pink1)

NADH Dehydrogenase (Ubiquinone) 1, Alpha/Beta Subcomplex, 1(Ndufab1)

Bobby Sox Homolog (Drosophila)(Bbx)

Aldo-Keto Reductase Family 1, Member C-Like(Akr1cl)

WD Repeat Domain 31(Wdr31)

Solute Carrier Family 25 (Mitochondrial Carrier; Adenine Nucleotide Translocator), Member 31(Slc25a31)

Kinesin Family Member 2C(Kif2c)

Interferon Induced Transmembrane Protein 7(Ifitm7)

Fatty Acid Desaturase 1(Fads1)

Inner Membrane Protein, Mitochondrial(Immt)

Mitochondrial Translational Initiation Factor 2(Mtif2)

Ring Finger Protein 34(Rnf34)

Intraflagellar Transport 122(Ift122)

Lin-28 Homolog A (C. Elegans)(Lin28a)

Calmin(Clmn)

Zinc Finger Protein 451(Zfp451)

Ets Variant 5(Etv5)

Heat Shock Protein 1A(Hspa1a)

Calcium/Calmodulin-Dependent Protein Kinase Kinase 2, Beta(Camkk2)

Centrosomal Protein 68(Cep68)

Basic Transcription Factor 3(Btf3)

General Transcription Factor IIB(Gtf2b)

GRAM Domain Containing 1B(Gramd1b)

Lipase, Member H(Liph)

Ring Finger Protein 213(Rnf213)

Table S5: Example of annotation clustering from DAVID describing two enriched processes, Glutathione metabolism and transcriptional regulation.

Querying the Eurexpress list of genes using the annotation clustering tool in DAVID can define related processes. For example, the processes of Glutathione metabolism and transcriptional regulation were over represented in the Eurexpress gene list. DAVID was accessed on September 12, 2016.

Category	Term	Count	PValue	List Total	Fold Enrichment
INTERPRO	IPR003081:Glutathione S-transferase, Mu class	4	9.17E-05	275	40.58805195
INTERPRO	IPR004046:Glutathione S-transferase, C-terminal	5	0.001125184	275	10.76198347
INTERPRO	IPR004045:Glutathione S-transferase, N-terminal	5	0.001125184	275	10.76198347
GOTERM_BP_DIRECT	GO:0018916~nitrobenzene metabolic process	3	0.001306243	249	50.02710843
GOTERM_MF_DIRECT	GO:0004364~glutathione transferase activity	5	0.001350374	250	10.20451613
GOTERM_MF_DIRECT	GO:0043295~glutathione binding	4	0.001854759	250	15.817
KEGG_PATHWAY	mmu00480:Glutathione metabolism	6	0.002019562	128	6.569318182
INTERPRO	IPR010987:Glutathione S-transferase, C-terminal-like	5	0.003592444	275	7.892121212
GOTERM_BP_DIRECT	GO:0042178~xenobiotic catabolic process	3	0.004438218	249	28.5869191
GOTERM_BP_DIRECT	GO:0006749~glutathione metabolic process	5	0.004473115	249	7.411423472
KEGG_PATHWAY	mmu00980:Metabolism of xenobiotics by cytochrome P450	5	0.006215115	128	6.690972222
KEGG_PATHWAY	mmu00982:Drug metabolism - cytochrome P450	5	0.007815539	128	6.272786458
UP_SEQ_FEATURE	domain:GST N-terminal	4	0.008489343	255	9.420130719
UP_SEQ_FEATURE	domain:GST C-terminal	4	0.019907987	255	6.892778575
KEGG_PATHWAY	mmu05204:Chemical carcinogenesis	5	0.06493113	128	3.272758152
INTERPRO	IPR012336:Thioredoxin-like fold	5	0.122225117	275	2.630707071
GOTERM_BP_DIRECT	GO:0008152~metabolic process	10	0.291498366	249	1.378157257
Annotation Cluster 2	Enrichment Score: 2.202438644893318
Category	Term	Count	PValue	List Total	Fold Enrichment
UP_KEYWORDS	Nucleus	86	2.44E-04	281	1.432714499
UP_KEYWORDS	Transcription	42	8.85E-04	281	1.694999913
UP_KEYWORDS	Transcription regulation	39	0.002932861	281	1.626656414
GOTERM_BP_DIRECT	GO:0006351~transcription, DNA-templated	40	0.012933994	249	1.466801787
UP_KEYWORDS	DNA-binding	33	0.014036324	281	1.541673065
GOTERM_MF_DIRECT	GO:0003677~DNA binding	37	0.041367714	250	1.372971261
GOTERM_BP_DIRECT	GO:0006355~regulation of transcription, DNA-templated	39	0.080401663	249	1.285281442

Tools like DAVID rely on querying what is known about the function of a gene. As a result, although powerful in uncovering functionality, this approach cannot make inferences or uncover unknown functionality. This limitation is illustrated by the example of Pdgfa (platelet-derived growth factor-alpha), which encodes the PDGF-A and is annotated in GO for the molecular function of platelet-derived growth factor receptor binding. However, although it is annotated for the biological processes of lung alveolar development (GO:0048286), salivary gland morphogenesis (GO:0060683) and bone development (GO:0060348; among others), it is not annotated for the biological process of male gonad development (GO:0008584). To a researcher in the field of sex determination and germ cell development, it may be obvious that Pdgfa should be involved in male gonad development, as we know this pathway is important for gonadogenesis [Brennan et al., 2003; Cool et al., 2011; Schmahl et al., 2008]. However without appropriate annotation for this biological process, the program cannot computationally retrieve Pdgfa as associated with testicular development. The tool is only as complete and up-to date as the GO terms it uses.

III. The power of annotation: the case of the missing ovary

If 298 entries are annotated as expressed in the testis in the Eurexpress database, we would assume that the same discovery capacity should be found in querying the ovary. When queried under the anatomy search for “ovary”, only 28 entries were retrieved. These numbers of genes are far less than the numbers of gene catalogued from the microarray data [Jameson et al., 2012; Nef et al., 2005; Liu et al., 2010]. We have discussed that many genes are annotated in several biological processes, yet we know that some biological processes are better studied than others. This should be a consideration when interpreting query results relying on GO terms, as some processes and gene families are more thoroughly annotated than the others largely due to historical reasons or the size of the research field. Similarly, the expression of genes annotated in some organs, for example the ovary, depends on semi-experts being able to accurately identify an organ in a tissue section [Diez-Roux et al., 2011]. In the 298 entries annotated for the testis, 11 (3.7%) genes were annotated incorrectly, where expression was actually in the neighboring pancreas or kidney. In an organ like the ovary, which doesn’t develop apparent distinguishing features like testis cords, it is a lack of annotation in databases that hampers researchers, rather than a lack of expression of ovarian genes.

In some cases, a lack of annotation can become a serious problem that skews results. All GO terms tend to be weighted equally in enrichment tools whereas in reality certain ontologies are better described than others [Huang et al., 2009b]. This is often more of a problem in emerging or smaller fields. A better-described ontology, based on gene families or specific processes where there is more available data, tend to have a higher chance of being associated with any list. The search is only as good as the ontology and the annotation behind it.

Conclusion

We are surrounded by transcriptomic data telling us what genes go up and what genes go down. Thanks to the affordability of transcriptomics, we have more data than we could have ever imaged. Each microarray or RNA-seq experiment produces an extensive list of candidate genes. What we haven’t had is a way to prioritize the genes on these lists for functional analysis. In this review we have provided strategies to narrow down the genes-of-interest list and identify which genes in a candidate list to pursue for further functional analyses. The key frameworks used by databases revolve around standardization of terms; both gene ontology (GO) terms and anatomical descriptions. Understanding how GO term descriptors work and are used by different databases is a critical first step for the researcher. Coupled with an understanding of GO terms, we provide a strategy using 4 types of databases to help researchers prioritize candidate genes. 1) Expression databases can be used to verify the expression of genes of interest in the different organs throughout development. 2) After confirming expression of the genes of interest, investigators can use databases like DAVID and GSEA to uncover interesting processes and pathways present in the candidate gene list. Focused searches that expand out from these pathways can identify candidate genes that may be involved in central processes in the system of interest. 3) Resources such as that EMBL-EBI Expression Atlas that annotate reanalyzed transcriptomic data can provide important information about gene expression patterns across developmental time points, during disease states, and in different species including human. 4) Finally, the importance of a gene in human disease can then be investigated by querying knowledgebases such as OMIM and HGMD. Using this approach a candidate gene list can be whittled down to a selection of promising candidates that fit the investigator’s criteria for strong candidate for functional analysis.

This prioritizing approach requires the investigator to move between multiple databases. Luckily, there are new efforts underway to centralize the data housed by different Consortiums, and to better store and annotate the kind of visual data that developmental biologists generate. The best example of this evolving database and framework is the Edinburgh Mouse Atlas Project (EMAP), which aims to build a 3-D anatomical atlas of mouse embryo development using a flexible anatomical ontology. The ontology uses a thesaurus of alternative anatomical terminologies that was built upon the Kaufman “Atlas of Mouse Development” and allows flexibility within queries. This framework, by enabling visual tracing of progenitors and derivatives for each structure or organ, aims to provide the kind of fate tracing map that are missing in other databases. The structures in the digital model are anchored by key anatomical terms, and the data are then associated with that tissue ‘lineage’ within the spatial framework. This kind of framework provides information about lineages belonging to structures, like an organ, over time. The EMAP database currently links out to, and draws data from Eurexpress, and both EMAP and Eurexpress link out to interface with resources such as MGI.

Ultimately, the community needs these resources to become more centralized. To this end, the MGI Gene Expression Database (GXD) project and the Edinburgh Mouse Atlas (EMA) project are working together to develop a comprehensive resource called the Mouse Gene Expression Information Resource (MGEIR) (for more information see: http://www.emouseatlas.org/emage/about/mgeir.html). MGEIR aims to provide a central resource that will house text-and spatial-based gene expression data. Gene expression data will be acquired from the literature and by direct data submission from external labs/consortia. This resource merges ‘spatial’ in situ hybridization data into the EMAGE’s 3-D format and flexible anatomical framework with the ‘non-spatial’ expression data (i.e. RT-PCR) collected in the GXD [Christiansen et al., 2006; de Boer et al., 2009; Hill et al., 2004; Geffers et al., 2012]. The development of MGIER and the creation of centralized databanks will provide the investigators with an amazing array of information at the click of a mouse.

The next big challenge for expression databases will be the onslaught of single cell RNA-seq data; this new technology offers an unprecedented amount of detail about the transcriptome of any tissue. It’s reasonable to expect that in the next few years, high quality single cell datasets will be available for many organs, including the developing mouse and human gonad. How these datasets are integrated into expression databases will be a challenge for the use of ontologies. The flexible anatomical design, used by databases like EMAP, may provide enough flexibility to trace progenitor cell populations and population derivatives for each tissue or organ (for a detailed overview of current eMouseAtlas capabilities see: [Armit et al., 2015]). How tissue ‘lineages’ are managed within the spatial framework of the expression atlas will provide a challenge for database custodians.

The next time when the investigator is facing a “gene prioritization problem”, we encourage the investigator to step away from the bench and use the tools and databases described here to move their projects and hypotheses forward. Publically available data can be used to efficiently narrow down a candidate gene list to produce strong candidates for functional analysis. Using databases, genes can be selected that have known expression in the organ of interest and are involved in important signaling pathways or biological processes. The genes of interest can represented in other transcriptomic studies or associated with human disease. So, before you decide to a make a genetically modified mouse, sit down at your desk and put the other type of mouse to work.

Table 2: Expression databases and resources that can be used to prioritize candidate genes.

Individual specialty resources such as ABA, GENSAT and GUDMAP provide specialized data and resources to their respective communities. Broader resources looking at embryonic development include Eurexpress and GenePaint. The data from both these resources feeds into the larger EMAP project and the MGI-GXD. The development of the MGEIR database, which is a collaboration between the EMAP and GXD projects, will provide a centralized hub for murine gene expression data.

Expression databases	Stages/Tissue	Aim and Scope of the Database	URL
ABA (Allen Brain Atlas)	Embryonic to adult brain and spinal cord	Database of gene expression patterns in the mouse brain and spinal cord.	http://brain-map.org/ For mouse: http://mouse.brain-map.org/
GENSAT (Gene Expression Nervous System Atlas)	- 10.5/15.5 dpc heads - P7, adult brain	Map the expression of genes in the central nervous system of the mouse, using both in situ hybridization and transgenic mouse techniques. The EGFP BAC-transgenic mice created by this project are available to the scientific community (278 new Cre driver lines).	http://www.gensat.org/
GUDMAP (GenitoUrinary Development Molecular Anatomy Project)	- 14.5 dpc whole embryos	Consortium of laboratories working to provide the scientific and medical community with tools to facilitate research into the GU system.	http://www.gudmap.org/
FaceBase	- various mouse, human and zebrafish	Consortium generating data in support of advancing research into craniofacial development and malformation. Comprehensive Includes data from high-throughput genetic, molecular, biological, imaging and computational techniques.	https://www.facebase.org
Eurexpress	- 14.5 dpc whole embryos	A transcriptome atlas database for mouse embryo.	http://www.eurexpress.org/ee/
GenePaint	- 10.5/14.5 dpc whole embryos - 15.5, P7, adult (P57) brain	Digital atlas of gene expression profiles in the mouse.	http://www.genepaint.org/
The e-Mouse Atlas Project (EMAP) which includes: EMA (The e-Mouse Atlas) EMAGE (Edinburgh Mouse Atlas of Gene Expression) eHistology (Online Atlas of Mouse Development)	- 1 cell – 18.5 dpc embryo - P0-P57	EMA, the e-Mouse Atlas. A 3-D anatomical atlas of mouse embryo development including detailed histology. EMAGE, the e-Mouse Atlas of Gene Expression. Anatomy is reconstructed from serial sections of single embryos at each representative developmental stage enabling 3D graphical display and analysis of in situ expression data eHistology resource, allows interactive exploration of cellular- resolution color images detailing mouse development with annotations from the Kaufman “Atlas of Mouse Development”.	EMA: http://www.emouseatlas.org/emap/ema/ EMAGE: http://www.emouseatlas.org/emage/ eHistology: http://www.emouseatlas.org/emap/eHistology/
MGI-GXD (Mouse Genome Informatics Gene Expression Database)	- 1 cell – 18.5 dpc embryo - adult	GXD collects and integrates the gene expression information in MGI. Its primary emphasis is on endogenous gene expression during mouse development. It integrates different types of data and provides links to other resources to place the data into the larger biological and analytical context.	http://www.informatics.jax.org/expression.shtml
MGEIR (Mouse Gene Expression Information Resource)	Pulls data from GXD and EMAP	MGEIR is a collaboration between the GXD project and the EMAP project.	Under development
3D Atlas of Human Embryo Development	Morphological data from Carnegie Stage 7–23	This resource is a morphological (not an expression) atlas. It has 3D fully reconstructed human embryos covering the phase of organogenesis, between Carnegie stage 7 (15–17 days old embryo) and 23 (56–60 days).	http://www.3dembryoatlas.com/