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).