A meta-analysis of human embryonic stem cells transcriptome integrated into a web-based expression atlas

Lists of genes differentially expressed

Analyzing 38 original studies using transcriptome analysis to study hESC, we were able to collect 20 lists of transcripts that were upregulated in hESCs compared to differentiated cell types and 11 lists of transcripts that were downregulated (see Supplemental Tables S1, S2 and S3). We only selected transcripts lists that provided a fold ratio of the mean expression in hESCs to that in differentiated cells. Each list was mapped to Unigene build 176. When the mean value of expression in differentiated cells was zero, which occurred with ESTs and SAGE, the hESC/differentiation ratio was arbitrarily set at 50. Only genes with a fold ratio greater or equal to 2 (“hESC” genes), or lower or equal to 0.5 (“differentiation” genes) were selected. The Gene Ontology annotation analysis was carried out using the Fatigo+ tool on the Babelomics website (http://babelomics.bioinfo.cipf.es) using gene symbols [12]. Only annotation with a false discovery rates (FDR) adjusted P-value < 0.05 were considered significant.

Integrating Affymetrix GeneChip datasets obtained from distinct studies

In order to compare the transcriptome of hESCs to that of differentiated cell populations, we built an expression compendium by combining the U133A (Affymetrix, Santa Clara, USA) microarray data from 8 publications [13–20]. Indeed, we and others have shown that the GeneChip system (Affymetrix) allows direct comparison between datasets obtained in different centers, provided that the same chip and the same normalization are used [10, 21, 22]. The number of samples amounted to 217, including 24 hESC samples (11 different hESC lines). All samples were normalized before analysis with the GCOS 1.2 software (Affymetrix), using the “global scaling” method with a TGT value set to 100. The “ detection call” can either be “present”, when the perfect match probes are significantly more hybridized than the mismatch probes, “absent” when both perfect match and mismatch probes display a similar fluorescent signal, or “marginal” when the probeset does neither comply to the present nor to the absent call criteria. When several probesets measured the same gene, only the probeset with the maximal number of present detection call across all samples was selected. This step reduced the list of probesets to 14 074. This dataset is available as Supplemental Table S4 and can be accessed on our website http://amazonia.montp.inserm.fr.

Hierarchical clustering

Hierarchical clustering was carried out on the detection call data with the CLUSTER and TREEVIEW software packages [23]. The value 1 was assigned to present calls, −1 to absent calls and 0 to marginal calls. This matrix was clustered without further mathematical transformation. Only genes were clustered, i.e. the order of the samples is the order used on the Amazonia! website, and is based on grouping according to the embryonic germ layer origin of the sample.

hESC lines and karyotype

The list of hESC lines used and their respective karyotype are listed in Table 1.

hESC culture

After approval from the French Ministry of Research, the French Ministry of Health and the Agence de la Biomédecine, HUES1 and HUES3 hESCs were imported from Douglas Melton’s laboratory (Harvard University, MA, USA) and cultured as described [24]. Cells were passaged every 3–4 days enzymatically with 0.25% trypsin/EDTA (Invitrogen, Cergy Pontoise, France) and cultured in knockout Dulbecco’s modified essential medium (Invitrogen) without plasmanate, with 10% KO-SR (Invitrogen), 2 mM L-glutamine, 1x nonessential amino acids, 0.05 mM β-mercaptoethanol, 10 ng/ml FGF2 (Abcys, Paris, France). Medium was replaced daily. HUES1 and HUES3 were cultured on murine embryonic fibroblasts (MEF) obtained from E13 ICR mice embryos (Harlan, Gannat, France) or on human foreskin fibroblasts (hFF) in porcine skin gelatin coated (Sigma-Aldrich, St. Quentin. Fallavier, France) 6 wells dishes. The hFF cell lines SM1 and SM3 were derived from respectively 75 and 3 years old patients undergoing foreskin reduction. Informed consent was obtained from the patient or the patient’s parents. MEF were cultured in DMEM with 10% fetal calf serum (FCS) and hFF in DMEM with 20% FCS. MEF and hFF were mitotically inactivated by mitomycin-C (2h at 10 μg/ml). These hESC expressed POU5F1, TRA-1-60, TRA-1-81, displayed phosphatase alkaline activity, and were able to differentiate into embryoid bodies that expressed differentiation markers of astrocytic lineage (GFAP) or endodermal lineage (alpha-feto protein) (see Supplemental Figure S1 and data not shown).

The HS293 and HS235 hESC lines were cultured in the Department of Obstetrics and Gynecology, (CLINTEC) at the Karolinska University Hospital as described [25]. Briefly, hESC were cultured on hFF (CRL-2429; American Type Culture Collection, Manassas, VA) mitotically inactivated by irradiation (40 Gy), in KO-DMEM 20% knockout SR, 2 mM Glutamax, 0.5% penicillin–streptomycin, 1% nonessential amino acids, 0.5 mM β-mercaptoethanol (all from Gibco Invitrogen Corporation, Paisley, Scotland), 1% insulintransferrin-selenium (Sigma-Aldrich) and 8 ng/mL of bFGF (R&D Systems, Oxford, UK). For karyotype analysis, hESC cells were treated with 75ng/ml final Karyomax Colcemid (Invitrogen) for 1h, trypsinized, incubated in 0.0375M KCl for 20 min, and fixed in fresh 3:1 methanol/acetic acid solution. 1 to 2 spreads were counted for chromosome number and 12 to 16 banding patterns were analyzed at 300–500 bands resolution.

Flow cytometry analysis

Human ES cells and fibroblasts were dissociated with trypsin (0.25%)-EDTA (1mM) (GIBCO) for 3 min. Cells were then washed with PBS and incubated for 30 min at 4°C in PBS with the corresponding monoclonal antibody (MAb): anti-CD24 MAb conjugated to phycoerythrin (PE) (dilution 1:50) (clone ALB9, Immunotech, Marseille, France) and/or anti-CD44 MAb conjugated to fluorescein isothiocyanate (FITC) (dilution 1:50) (clone J-173, Immunotech). After PBS washes, cells were suspended in Facsflow (Becton Dickinson, San Jose, CA) and fluorescence was analyzed with a FACSCalibur flow cytometer (Becton Dickinson) or sorted with a FACSAria cell sorter (Becton Dickinson). Appropriate isotype controls were included in all analyses.

Immunofluorescence

hESCs cultured on coverslips were fixed for 20 min in 4% paraformaldehyde and washed three times in PBS. Cells were permeabilized with 0.1% Triton X-100. After blocking at room temperature for 60 min in PBS with 5% donkey serum (S30, Chemicon international, Temecula, CA), cells were incubated for 1 hour at room temperature with primary antibody diluted in PBS with 5% donkey serum: POU5F1/OCT3/4 (sc 9081, Santa Cruz Biotechnology, Santa Cruz, CA; 1:300) and SEMA6A (AF1146, R&D, Abington, United Kingdom; 1:50). Cells were washed three times in PBS and incubated for 1 hour at room temperature with Alexa Fluor® 488 Donkey anti-Rabbit (A-11034; Molecular Probes; 1:1000) and Cy3 Donkey anti-Goat (Jackson ImmunoResearch; USA; 1:400) secondary antibodies, for POU5F1 and SEMA6A respectively. Unbound antibodies were removed by three washes in PBS. Hoechst staining was added to the first wash (Sigma, 5 μg/ml).

Compiling hESC expression profiles

As of October 1^st, 2006, we identified 38 original studies, 1protocol description and 4 reviews analyzing the transcriptome of hESCs (Table 2). The original studies used various hESCs, control cells and gene expression analysis techniques (summarized in Supplemental Table S1). 28 different hESC lines were used and transcriptome techniques included microarrays (7 different types of chips), ESTs scanning, SAGE, MPSS and Illumina beads. One study compared chromosome immuno-precipitation (ChIP) chip data with transcriptome data in hESCs [10]. Nevertheless, some common features emerged such as the frequent investigation of the H1, H9, and BG01 hESCs (in 12, 12 and 11 studies respectively), and the use of the GeneChip (Affymetrix) microarray system (in 16 studies) (see supplemental table S1).

Meta-analysis of genes differentially expressed between hESCs and non-pluripotent cells

One main objective of large scale gene expression analyses of hESCs is to identify the set of genes that are overexpressed in this unique cell type (“hESC” genes) or underexpressed (“differentiation” genes). We reasoned that bona fide “hESC” or “differentiation” genes would be repeatedly uncovered by independent groups, regardless of the hESC lines, the control cells, the assay format or the statistical method that had been used. We collected, from these 38 original studies, 20 lists of transcripts overexpressed in hESCs (“hESC genes”) and 11 lists of genes underexpressed in hESCs (“differentiation genes”).

The 20 lists of hESC genes comprised 5567 different genes. As illustrated in Figure 1A, we observed a marked heterogeneity between these lists, with only 1076 genes found overexpressed in hESCs by three or more independent studies, 48 genes by 10 or more studies and only 1 gene by all 20 lists (see Supplemental Table S2). The 48 genes found overexpressed in at least 10 studies are listed in Table 3. Of note, the pivotal ES transcription factor POU5F1/OCT3/4 is the one gene found by all 20 lists, whereas genes found in at least 10 studies include the transcription factors NANOG and SOX2, and the growth factors TDGF1/CRIPTO and Galanin (GAL) that are known to be highly expressed by hESCs. Thus, according to hESC transcriptome analyses published to date, this list of 1076 genes found overexpressed in hESC cells by three or more studies can be viewed as a “consensus hESC gene list”.

In order to get further insights into this hESC list, we built an expression compendium by combining the data from 5 publications using the U133A GeneChip microarray to analyze hESC transcriptome and 3 publications providing the transcriptome of various normal fetal and adult tissues (see material and method and Supplemental Table S4). This compendium included the gene profiling of 24 hESC samples and more than 190 various fetal and adult tissues samples. A heat map was generated for the “consensus hESC gene list” in this expression compendium based on the detection call provided by the GCOS 1.2 software. The detection call is a way to evaluate whether a gene is expressed or not in a given sample. Hierarchical clustering (Figure 1C) delineated four major clusters of genes: cluster (a) was a group of 40 genes specifically detected in hESCs (“hESC specific” genes, Table 4), including expected genes such as POU5F1/OCT3/4, NANOG, TDGF1, LIN28, CLDN6, GDF3, DNMT3A, but also genes such as CYP26A1, HELLS or GPR19, cluster (b) featured genes that were detected in both hESCs and CNS samples such as the GABA receptors GABRB3 and GABRA5, and the growth factor FGF13, cluster (c) genes detected in samples characterized by a high mitotic index such as SKP2, MYC, the cyclin CCNA2 and the MCM genes MCM2, MCM5, MCM6 and MCM7, cluster (d) genes overexpressed in hESCs but also expressed in a majority of the tissues included in this dataset such as PGK1, HSPA9B, and the ribosomal genes RPLP0, RPL6, RPL7 and RPL24. The complete lists of genes composing these clusters are available as Supplemental Table S5. The expression histograms of the 40 hESC specific genes are shown in Supplemental Figure S2. These results show that, though all 1076 genes of the “consensus hESC gene list” have been found overexpressed in hESCs compared to non-hESC samples by at least three different studies, 40 genes are indeed hESC specific (cluster a) but most are nonetheless also expressed in adult tissues to various extent.

The 11 lists of “differentiation” genes summed up to 4798 genes and we noted a similar heterogeneity as in the “hESC” lists (Supplemental Table S3). Out of these 4798 genes, only 783 were found underexpressed in hESCs by at least 3 different studies, composing a “consensus differentiation gene list”, three genes, lumican, collagen 1A1 and 3A1, by nine studies and none by all 11 studies (Figure 1B). Table 5 shows 30 selected genes found by at least 6 studies, which include the bone morphogenetic proteins (BMP) 1 and 4, the keratins 7 and 18, insulin-like growth factor 2 (IGF2), Heart and neural crest derivatives expressed 1 (HAND1) and the transducing chain IL6ST/GP130. The complete “consensus differentiation gene list” can be found in Supplemental Table S3.

We compared functional Gene Ontology (GO) annotations of the hESC genes to the differentiation genes. Several functional annotations were more represented in each category (Figure 1D). There were significantly more genes involved in “metabolism”, “mitosis”, “RNA splicing”, “nuclear pore”, “DNA repair” in the hESC gene list, reflecting the intense proliferation, DNA replication and DNA remodeling taking place in these cells. Conversely, GO annotations such as “organ development”, “skeletal development”, “extracellular matrix”, “cell adhesion”, “cell communication”, “integral to plasma membrane”, “signal transduction”, were significantly more frequent among genes upregulated in differentiated tissues, in agreement with the idea that hESC differentiation mimics early organogenesis, which is associated with the development of complex cell-cell communication and cell-extra cellular matrix (ECM) interactions.

CD24 and SEMA6A are hESC markers

We next inquired whether we could find, among the hESC genes, new markers that may be useful to identify, isolate and qualify hESCs in vitro. We focused on two cell surface hESC genes: CD24 and SEMA6A. These two genes have been found to be overexpressed in hESCs by respectively nine and fifteen studies, and are therefore good candidates as new hESC markers.

CD24 is a sialoglycoprotein known to be expressed on mature granulocytes and B cells subpopulations [26]. In addition to these hematopoietic cell types, microarray data from the expression compendium evidenced a high CD24 mRNA expression in keratinocytes, pancreas and thyroid, whereas most neural tissues, muscle, liver and testis did not express CD24 (Figure 2A). Upon differentiation of hESCs into embryoid bodies (EBs), CD24 expression is markedly downregulated (Figure 2B). Importantly, human foreskin fibroblasts (hFF) samples did not express CD24 (Figure 2C, 2F). Hence, we investigated whether CD24 could discriminate hESCs from fibroblasts in culture. We analyzed by flow cytometry the hESC lines HUES1, HUES3, HS293 and HS235 [24, 25] cultured on hFF and evidenced two distinct cell populations, one CD24+ and one CD24− (Figure 2E). To demonstrate that the CD24+ population corresponded to hESCs and the CD24− population to hFF, we took advantage of CD44, a strong fibroblast marker not expressed by hESCs (Figures 2D, 2G). Double staining of hESCs cultured on hFF showed that these two markers were expressed in a mutually exclusive manner and delimited a hESC CD24+CD44− population that did not overlap with the fibroblast CD24−CD44+ population (Figure 2H). Using CD24/CD44 labeling, we were able to separate hESC from fibroblasts and obtain pure hESC populations that recovered the cell sorting procedure and grew in vitro while retaining POU5F1 and cardinal cell surface markers of pluripotency expression (Supplemental Figure S3 and data not shown).

SEMA6A is a class 6 semaphorin [27], i.e. transmembrane with cytoplasmic domain, known to be expressed in developing neural tissue. We recently reported the expression of this semaphorin in cumulus oophorus cells [28]. Comparison of RNA expression in hESCs and normal adult tissues showed that in addition to hESCs, SEMA6A was also expressed at high level in adult samples from the central and peripheral nervous system and placenta (Figure 2I). As for CD24, SEMA6A mRNA is downregulated upon EB differentiation and is not detected in human fibroblasts (Figure 2J and K). Immunofluorescence analysis showed a membrane localization of SEMA6A on hESCs, in contrast to POU5F1/OCT3/4 which had a strict nuclear localization (Figure 2L-M). In summary, we showed that CD24 and SEMA6A have indeed a preferential expression in hESCs, that this expression is confirmed at the protein level, and is declining upon EB differentiation. Thus these markers can be used to discriminate hESCs from feeder fibroblasts

hESC transcriptome data visualization through an open access web interface

We developed a website, Amazonia! (http://amazonia.montp.inserm.fr), to allow the scientific community access to these public data. This website is dedicated to the visualization of large, publicly available, human transcriptome data (Le Carrour et al., manuscript submitted). A main topic of this website is human embryonic stem cells. Data are visualized as expression histograms with a color code facilitating the recognition of cell type. Genes are accessed either by key words or through lists of genes. Most interestingly, when data were obtained using the same platform format, sample labeling and data normalization, it was possible to combine different experiments in one single virtual experiment. The U133A expression compendium comprises for example more than 200 different samples from 8 publications including hESCs and normal adult tissues. Thus, Amazonia! provides the expression profile of about 15 000 different genes in about 100 different tissues types or purified cell populations including 11 hESCs (H1, H9, HS181, HS235, HS237, FES21, FES22, FES29, FES30, ,I6, HES-2). Figure 3A illustrates this feature of our website with the expression histograms of five hESC specific genes (POU5F1, NANOG, GPR19, Helicase lymphoid-specific (HELLS) and the cytochrome P450 CYP26A1), two genes expressed in nonlineage-differentiated cells (HAND1 and IGF2), two factors highly expressed by human fibroblasts and smooth muscle that may contribute to the supporting properties of fibroblast to hESC culture (Gremlin (GREM1) and Matrix metalloproteinase 1 (MMP1)), one hematopoietic marker (CD45), one central nervous system marker (Glial fibrillary acidic protein (GFAP)) and one ubiquitously expressed gene (Ribosomal protein L3 (RPL3)).

Another important feature of Amazonia! is the possibility to compare on the same web page the expression of a gene of interest in various datasets. Figure 3B shows that the expression of Frizzled 7 (FZD7) was markedly upregulated in hESCs, was downregulated during non-lineage differentiation into EBs, and was also highly expressed in embryonal carcinoma and Yolk sac carcinoma samples. The combination of these three histograms evidences a preferential FZD7 mRNA expression in normal and malignant embryonic cells, suggesting that FZD7 may play a major role in these pluripotent cell types.

In order to facilitate access for the scientific community to the lists of genes from published transcriptome analyses, we implemented a list manager in our Amazonia! website. Thus, one can access these lists straightforwardly and obtain an expression histogram for each gene in various public transcriptome collections. This feature is of particular interest to challenge a list of genes, for example a “hESC” gene list, with other hESC datasets or even with non-hESC datasets such as cancer datasets. Indeed, once a gene is selected, the user can navigate between various thematic pages, switching for instance from the “stem cells” page to the “leukemia and lymphoma”, “lung cancer” or other pages. Using this feature of the Amazonia! website, we could observe that the sialoglycoprotein CD24 was also highly expressed in acute lymphoblastic leukemia, lung cancer and glioma samples when compared to the corresponding normal samples (Figure 3C and data not shown).