Prescreening of hematoxylin-eosin cryosections from 11 fetal adrenal glands demonstrated that large neuroblast clusters of more than 100 cells were predominantly found in adrenal glands at 19 and 20 weeks' gestational age (Figure 1a). To verify that these cell clusters indeed represent neuroblasts and to estimate the degree of intermingled chromaffin cells, cryosections were stained for the neuronal and chromaffin marker TH (tyrosine hydroxylase), the chromaffin marker CHGA (chromogranin A; which also has low expression in neuroblasts), and the neuronal markers BCL2 (B-cell CLL/lymphoma 2) and HNK1 (carbohydrate epitope) 8. As shown in Figure 1, the clusters of neuroblastic cells stained positive for all markers and, in particular, these cells were positive for BCL2 and HNK1. The majority of chromaffin cells, identified by their strong CHGA and TH expression, were found to be scattered throughout the adrenal cortex (these cells coalesce and form large islands of chromaffin cells later during development), whereas a few cells were located in or adjacent to the neuroblast clusters.

Neuroblast clusters and adjacent cortical cells (used as controls) were isolated using laser capture microdissection from stained cryosections from three different fetal adrenal glands (glands 1, 2 and 3, which were of gestational ages 20, 19 and 19 weeks, respectively) (Figure 2) and immediately lysed in RNA extraction buffer. In order to obtain a sufficient amount of good quality neuroblast RNA for oligonucleotide chip analyses, we applied a previously validated protocol for tissue sectioning, staining, and microdissection 9 (Additional data file 1(a)). By pooling different isolates of the same adrenal gland, between 2.5 and 15 ng total RNA could be obtained for each of the three neuroblast samples (Additional data file 1(b)). After two-round amplification and labeling of three neuroblast, three cortex, and 18 neuroblastoma RNA samples, hybridization was performed on HG-U133A Affymetrix oligonucleotide chips. Real-time polymerase chain reaction analysis of selected genes showed that there was no RNA amplification bias in the chip data (Additional data file 1(c)).

The expression profiles of the neuroblast and cortex samples were compared using the rank product nonparametric method, which is particularly suited for extracting significantly differentially expressed genes in a limited number of samples 10. Two lists of 156 and 86 unique genes were established with significantly higher expression in neuroblast and adrenal cortex cells, respectively (multiple testing corrected P < 0.01; Additional data file 2). Gene Ontology (GO) analysis identified those classes of genes that are significantly over-represented in the cell specific gene lists (P < 0.01; Table 1). As expected, the neuroblast gene list is enriched for genes that are involved in catecholamine metabolism, neurogenesis and other neural processes, whereas cortex cells specifically express genes involved in steroid and cholesterol metabolism.

To further test the validity of the neuroblast gene expression profile, we evaluated the expression of known neuronal and chromaffin markers that were previously studied in human fetal sections 8. High expression (among the 10% most abundant genes) of neuronal markers (BCL2, GAP43, and NPY) together with chromaffin (and to a lesser extent neuronal) markers (CHGA, CHGB, DBH, DDC and TH) and an adrenal chromaffin marker (PNMT) in the microdissected cell clusters is in keeping with our observation that the neuroblast isolates are pure, with only rare intermingled chromaffin cells (Figure 1).

Gene set enrichment analysis 11 based on expression of the 156 neuroblast-specific genes in 79 human tissues 12 was performed in order to explore whether the microdissected neuroblasts indeed have neural characteristics. The neuroblasts exhibit a significant overlap in expression with various nervous system tissues (P < 0.05; fetal brain, prefrontal cortex, brain amygdale, whole brain, occipital lobe, and hypothalamus), further demonstrating that the proper cells were microdissected.

Although multiple lines of evidence indicate that neuroblastoma originates from immature sympathetic neuroblasts, the mRNA expression repertoire of these neuroblasts and neuroblastomas have not yet been compared. Before our analysis, we assumed that, in addition to differences resulting from oncogenic transformation, both cell populations would exhibit many cell type specific similarities.

Three data mining strategies were employed to investigate this hypothesis. First, an unbiased multidimensional scaling of all genes on the chip showed that the neuroblasts cluster close to the neuroblastoma tumors and that both groups cluster far away from the fetal adrenal cortex cells (Figure 3a). Second, we extended our dataset with publicly available expression profiles (measured on the same platform) from 79 normal tissues 12and three neural stem cell cultures 13. Based on the genes that are differentially expressed between the neuroblasts and cortex samples (156 and 86 genes, respectively), multidimensional scaling showed again that the neuroblastoma tumors cluster close to the neuroblasts and further away from the other normal tissues. Interestingly, the neural stem cells also cluster close to the neuroblastomas and neuroblasts (Figure 3b). These findings further support the notion that adrenal neuroblasts are indeed of neuronal origin with possible neuronal stem cell features, and the observed considerable similarities to neuroblastomas in terms of expression give further strength to the 'cell of origin' hypothesis for neuroblastoma development.

Third, we looked for similarities in mRNA expression between neuroblast and neuroblastoma by cataloging their expression repertoire. We defined a reasonable cut-off to determine whether a gene is expressed or not in a given sample (the mean percentage of present calls for the various chips; Additional data file 1(d)). As such, the 36% most highly expressed probe IDs in the cortex, neuroblast and neuroblastoma cells, were selected and compared in a Venn diagram. This analysis clearly shows that neuroblasts have more expressed genes in common with neuroblastoma than with the cortex cells (432 versus 292; Figure 4a). GO analysis on the common 432 genes revealed an expected over-representation of neurogenesis genes (P < 0.01; data not shown). Next, we zoomed in on neurogenesis and transcription factor ontology classes by performing a similar Venn diagram for these gene sets, assuming their putative importance in neuroblastoma development. Interestingly, the similarities between neuroblast and neuroblastoma are even more pronounced for these two GO classes (Figure 4b, c).

In the final and most challenging part of our data mining approach, we aimed to identify genes that are under-expressed or over-expressed in neuroblastomas compared with neuroblasts, because these genes and the pathways that they govern might be involved in neuroblastoma development or represent markers for the stage of developmental arrest of neuroblastomas. Rank product analysis (multiple testing corrected P < 0.01) yielded a list of 71 genes that were more highly expressed in neuroblasts and 565 genes that were more highly expressed in neuroblastomas (Additional data file 2).

A first crucial step in our data mining strategy to identify genes that are putatively involved in neuroblastoma was a meta-analysis of our generated gene lists in published neuroblastoma microarray data. We used the Neuroblastoma Gene Server (NBGS) which was developed in-house (see Additional data file 3 for detailed information) to compare the neuroblast-specific and neuroblastoma-specific gene lists with genes that have been reported as differentially expressed in 25 previous gene expression profiling studies conducted in neuroblastoma (Additional data file 3). We found that as many as 17 of the 71 genes (24%) that are over-expressed in neuroblasts relative to neuroblastomas were reported in the NBGS, mainly annotated as genes that are more highly expressed in maturing, differentiating, or localized neuroblastomas. Likewise, 102 out of the 565 genes (18%) that were over-expressed in neuroblastoma were previously identified in other gene expression studies on neuroblastoma. The high overlap of our gene lists with published gene lists demonstrates the validity of our lists, which were subsequently further explored in chromosomal mapping, GO, and pathway analysis.

Chromosome 17q gain is the most frequent genetic aberration in neuroblastoma and is assumed to play a crucial role in its pathogenesis through a dosage effect of one or more genes. This critical region still comprises 25 megabases (Mb) 14, precluding straightforward candidate gene identification. Here we apply an alternative, intuitive strategy to pinpoint putative critical dosage sensitive loci. Using positional gene enrichment analysis (De Preter and coworkers, unpublished data) 15, we sought chromosomal loci that are significantly over-represented in the list of genes that are over-expressed in neuroblastoma relative to their normal cells of origin (Figure 5). We found two peaks on chromosome 17q, with high significance for a locus on 17q21.32-q22 that coincides with the consistently gained segment just distal from the most distal breakpoint in a series of high-resolution copy number profiles (Vandesompele and coworkers, unpublished data).

Apart from over-expressed genes, we also sought positional tumor suppressor genes by mapping under-expressed genes (relative to normal neuroblasts). The following positional candidates could be identified, located within or very close to the known shortest regions of overlap in neuroblastoma: CASP9 on 1p36; CACNA2D3, TDGF1 and NKTR on 3p21-p22 (SRO (shortest region of overlap) from 16); IGSF4, APOA1, MLL and RDX on 11q23 171819; and MEG3 and DLK1 on 14q32 20.

To examine gene expression differences between neuroblasts and neuroblastomas from a different perspective, we mapped the neuroblast-specific and neuroblastoma-specific gene lists to the biologic process GO classification (Table 2). This revealed that the neuroblasts express significantly (P < 0.01) more genes that are involved in steroid and catecholamine metabolism compared with neuroblastomas. Neuroblastomas are characterized by an over-representation of genes that are involved in immune response, cell growth, and cell cycle. The immune response gene signature may be due to infiltrating immune cells, whereas the over-representation of cell growth and cell cycle genes in neuroblastomas is in perfect concordance with the hyperproliferative character of tumors.

We then specifically looked at genes belonging to GO terms neurogenesis, transcription factor activity, and apoptosis; these three processes can be assumed to play an important role in neuroblastoma pathogenesis (Table 3). This analysis identified the following interesting genes from the neuroblast-specific and neuroblastoma-specific gene lists: transcription factors involved in neurogenesis TFAP2B (6p12.3; more highly expressed in neuroblasts); ASCL1 (12q23.2), SIX3 (2p21) and STAT3 (17q21.2; more highly expressed in neuroblastoma); and APOE (19q13.31) and INHBA (7p14.1; more highly expressed in neuroblastoma), which are involved in both apoptosis and neurogenesis.

Thus far, most published microarray studies on neuroblastomas mainly compared favorable with unfavorable neuroblastomas in order to identify prognostic markers or pathways that are involved in these clearly different neuroblastoma tumor types. In order to add value to such an analysis, we contrasted similar differentially expressed gene lists with the normal neuroblast expression profile (Additional data file 2). In a first step, we compared the differentially expressed genes between these two tumor types with published prognostic gene lists. We found that 25 of the 194 genes on our list were previously reported, including the well established markers MYCN, NTRK1, and CD44 (see NBGS analysis in Additional data file 3). This overlap demonstrates the validity of the selected neuroblastoma panel and their expression profile. Subsequently, we sought the corresponding gene expression levels of the differentially expressed genes in the normal counterpart cells, aiming to select neuroblastoma candidate genes. Of the 95 genes that are more highly expressed in favorable tumors (versus unfavorable ones), 37 also have significant differential expression (either higher or lower) compared with neuroblasts, whereas 41 out of the 101 genes that are more highly expressed in unfavorable tumors exhibit differential expression compared with the neuroblasts (Table 4).

From this analysis, a few putative positional tumor suppressor candidates emerge: CDC42 on 1p36, CACNA2D3 on 3p21, and DLK1 on 14q. The latter two genes are of particular interest because they are highly expressed in neuroblasts and favorable neuroblastomas, and their expression is significantly lower in unfavorable neuroblastomas. Among the genes that are more highly expressed in unfavorable neuroblastomas than in favorable ones and neuroblasts, the proven oncogenic transcription factor MYCN emerges (and putative downstream genes KIFAP3, OPHN1, RGS7, ODC1, TOP2A, TWIST1 and TYMS, according to NBGS), as do several other genes that have been identified or studied within the context of neuroblastomas such as ALK and PRAME, and positional candidates on 17q including BIRC5, RNU2 and TOP2A.

Although this was not the primary aim of the present work, the neuroblast expression profile provides a unique resource for the investigation of gene expression in human sympatho-adrenal progenitors. In a first attempt, we made an inventory of the genes that belong to the neurogenesis GO class, or that have been described to play a role in neural crest formation and migration, or that have proneural activity (Additional data file 4). This analysis showed that human fetal neuroblasts of 19 weeks' gestational age expressed 174 of the 359 genes in the neurogenesis GO class, and 26 of 89 proneural genes and genes involved in neural crest formation/migration.

To obtain possible clues on the developmental origin of neuroblastoma we compared the expression profiles of the neuroblastoma tumors with those of normal neuroblasts and neural stem cell cultures. Intersectional Venn diagram analysis of expressed genes shows that neuroblastomas have many genes in common with neuroblasts, as already shown above (Figure 6). Interestingly, when compared with neuroblasts, the neuroblastomas have more genes in common with the self-renewing neural stem cells (535 versus 145), among others the neurogenesis genes ASCL1, GSS, STAT3, UTP11L, ENAH, APBB2, CDK5RAP2, and LARGE.

Ethical approval was obtained for the collection of fetal adrenal glands from fetuses aborted for clinical reasons (Ethics committee Erasme Hospital, Brussels, Belgium; approval no.: OM021). The induced abortion was performed by prostaglandin instillation to the patient. The adrenals were removed during necropsy and snap-frozen in liquid nitrogen within 3 hours after delivery. Neuroblastoma tumors were collected in the Center for Medical Genetics (Ghent, Belgium; n = 12), in the National Center for Medical Genetics (Dublin, Ireland; n = 1), and in the University Children's Hospital of Essen (Essen, Germany; n = 5). For this study, we preferentially selected tumors that were localized in the adrenal gland (11/18). Based on INSS stage (international neuroblastoma staging system), MYCN status, ploidy and age at diagnosis, and for some cases pathologic rapports, samples were divided into favorable or unfavorable neuroblastoma (Table 5).

Fetal adrenal glands were embedded in Tissue-Tek OCT compound (Sakura, Torrance, CA, USA). Immunohistochemical staining was performed as described previously 8. For microdissection, cryosections were first stained with hematoxylin and eosin, and mounted in order to scan for neuroblast clusters. When neuroblast clusters were found, stained but unmounted cryosections were prepared for laser capture microdissection. Embedding, sectioning, staining, and laser capture microdissection of neuroblast clusters and surrounding cortex cells was performed as described previously 9.

Microdissected cells were collected in RNA extraction buffer, followed by RNA extraction and DNase treatment on column (Qiagen, Venlo, Netherlands). RNA of the tumor samples was extracted using the RNeasy Mini kit (Qiagen), in accordance with the manufacturer's instructions. Four of the neuroblastoma tumor pieces were first mixed with Lysing Matrix D microbeads (Qbiogene, Illkirch, France) and 700 μl RTL buffer (Qiagen), and homogenized using FastPrep FP220 (Qbiogene). A fraction of the RNA was used for cDNA synthesis after DNase treatment (described by Vandesompele and coworkers 39). RNA quality was measured with the RNA Nano or Pico LabChip kit (Agilent, Diegem, Belgium) using 1 μl of the RNA isolates.

For each of the three fetal adrenal glands, the different neuroblast RNA isolates were pooled, amplified using a two-round labeling protocol, and hybridized to HG-U133A oligonucleotide chips (Affymetrix, Santa Clara, CA, USA), containing 18,400 transcripts including 14,500 well characterized human genes (protocol described previously 40). The same amplification protocol was applied to RNA of three cortex samples and approximately 100 ng RNA of 18 neuroblastoma tumors. The homogeneity of the subgroups (neuroblast, cortex, favorable and unfavorable neuroblastoma) allowed us to use a limited number of samples for expression profiling. Several technical parameters demonstrate that the hybridization was of good quality (Additional data file 1(d)).

We obtained the raw data from the Genomics Institute of the Novartis Foundation compendium of normal tissues consisting of 79 normal tissues assayed in duplicate using the Affymetrix HG-U133A array 12. Raw HG-U133A Affymetrix array data from three neural stem cells were kindly provided by Wright and coworkers 13.

CEL files were loaded in the R-Bioconductor (BioC) software and normalized with the Robust Multi Chip Average (RMA) method 41. Identification of differentially expressed genes for pairwise comparisons were performed using the RankProd R-package, which is based on the Rank Product principle 10. We used the GoHyperG function from the BioC project to find over-represented biologic process GO categories from the gene lists using hypergeometric test for significance. KEGG pathway analysis was performed with the Webgestalt web interface using hypergeometric test for significance 42.

Meta-analyses of published neuroblastoma microarray data were performed with the NBGS (see Additional data fiile 3 for detailed information).

Positional gene enrichment analysis was performed with in-house developed R-Bioconductor script PGE (De Preter and coworkers, unpublished data) 15. PGE scans the entire genome using a moving window with a user-defined width (5 Mb) and step size (1 Mb). In each window, the -10log(p) of the Fisher Exact Test is calculated. This test was used to investigate whether there is an association between the gene list and a particular chromosomal region (the window under investigation). As such, it will identify regions that contain more (or less) genes in the gene list than expected by chance. The (known) unequal distribution of the genes along the chromosomes is taken into account, because the number of genes from the list that are located in the region is compared with the total number of genes in that particular region. Correction for multiple testing is performed using the false discovery rate method of Benjamini and Hochberg 43, using the R-multtest package.

Expression microarray data were submitted to ArrayExpress 44, accession number E-MEXP-669.