Background
Peripheral neuroblastic tumours (NT’s) are derived from developing neuronal cells of the sympathetic nervous system and are the most frequent extracranial solid tumours of childhood. NT’s are composed of variable proportion of neuroblasts (neuronal lineage) and Schwannian cells (glial lineage), and are classified into histopathological categories according to the presence or absence of Schwannian stromal development, differentiation grade of the neuroblasts, and their cellular turnover index. According to the International Neuroblastoma Pathology Classification (INPC - Shimada system) [
1], the three subtype categories and their subtypes are: 1) Neuroblastoma (NB), Schwannian stroma-poor; 2) ganglioneuroblastoma (GNB), intermixed (Schwannian stroma-rich) or nodular (composite Schwannian stroma-rich/stroma-dominant and stroma-poor); 3) ganglioneuroma (GN), Schwannian stroma-dominant. Neuroblastoma exhibit an extreme clinical and biological heterogeneity, and patients are assigned to risk groups based on several criteria including stage [
2,
3], age [
4], histological category and grade of tumour differentiation (histopathology) [
5], the status of the
MYCN oncogene [
6], chromosome 11q status [
7], and DNA ploidy [
8] as the most highly statistically significant and clinically relevant factors [
9]. One-half of NB patients have metastatic disease at diagnosis (INSS stage 4 or INRGSS stage M). All metastatic tumours with
MYCN amplification (MNA) are aggressive and considered being high-risk tumours [
9], whereas children with metastatic disease without MNA (approximately 65%) have variable clinical behaviours depending on age at diagnosis, histopathology, and other genetic factors. Based upon cytogenetic profiles, previous studies have categorized NB tumours into three major subtypes [
10,
11]: Subtype 1 representing favourable tumours with near triploidy and high expression of the Neurotrophic receptor TrkA (or NTRK1), mostly encompassing non-metastatic NB stages 1 and 2; subtype 2A representing unfavourable NB stages 3 and 4, with 11q deletion (Del11q) and 17q gain (Gain17q) but without MNA; subtype 2B representing unfavourable widespread NB stages 3 and 4 with MNA often together with 1p deletion (Del1p) and Gain17q. Several gene sets are shown to discriminate the molecular subgroups and risk groups by mRNA and microRNA expression profiling in neuroblastic tumours [
12‐
21]. A recent expression analysis by our research group identified the three cytogenetically defined subtypes (1, 2A, and 2B) by unsupervised clustering, but further indicated the existence of a fourth divergent subgroup [
12]. Moreover, we identified a 6-gene signature including
ALK,
BIRC5,
CCND1,
MYCN,
NTRK1, and
PHOX2B to successfully discriminate these four subgroups [
12]. The fourth (r4) subgroup encompassed tumours characterized by Del11q and high expression of genes involved in the development of the nervous system, but showed low expression of
ALK. Approximately 7-9% of sporadic NB cases show inherent
ALK mutations [
22,
23], and
ALK overexpression, both in its mutated and wild type form, is demonstrated to define a poor prognosis in NB patients [
24]. In relation to this our previous findings suggests the Type 2A (r2) and Type 2B (r3) subgroups, which both display high
ALK expression, to be driven by the ALK pathway. In contrast, the r4 subgroup displaying low expression of all six genes of the signature, is suggested to be driven by an alternative oncogenesis pathway.
In the present study we aimed to further investigate the expression profiles of the four subgroups, and r4 in particular. By differential expression analysis and reverse engineering we found ERBB3 and its network members to be significantly overrepresented within the r4 tumour subgroup. Moreover, two other ErbB family members, ERBB2 and EGFR, were found to show concurrently higher expression. In contrast, unfavourable neuroblastoma subgroups (r2 and r3) were specifically characterized by G2/M cell cycle transition and mitotic regulating genes. By expression analysis of histopathology categories (i.e. NBs, GNBs, and GNs) we found the r4 subgroup to show an identical expression profile to GNB/GN types, and overexpression of ErbB3 was also confirmed at the protein level in GN tumours. We conclude that the ERBB-profile (high expression of EGFR, ERBB2 and ERBB3) defines a ganglion-rich neuroblastic tumour sub-set.
Discussion
Neuroblastic tumours (NT’s) represent a spectrum of disease, from undifferentiated and aggressive NB to the differentiated and largely quiescent GN tumours. NB tumours are commonly categorized into three main types based on numerical and structural genomic alterations, as well as expression of the neurotrophin receptor TrkA [
10]. In a recent study using Principal Components Analysis (PCA) however, our data indicated the existence of four molecular tumour groups, r1-r4 [
12]. In the current study we aimed to further characterize these four molecular subgroups, and investigated the divergent r4 group in particular. While the r2 (Type 2A) and r3 (Type 2B) tumour subgroups were dominated by cell cycle-related genes and networks, those were completely absent in the r4 subgroups (data sets 1–3) and GNB or GN subtypes (data set 4). The vast majority of the cell cycle-related genes were linked to the G2/M transition and spindle assembly checkpoint (
e.g. BIRC5,
BRCA1,
BUB1B,
CCNA2,
CCNB1,
FANCI,
HMMR,
KIF15, and
MCM2), many of which were found to belong to the ARACNE-modelled BIRC5-network. Overexpression of genes involved in mitotic regulation is typical for rapidly proliferating tumours and would also be expected to be enriched in the aggressive NB subtypes when compared to more differentiated quiescent GNB and GN tumours. The BIRC5 protein is found to stabilize the microtubules in the chromosomal passenger complex, and knockdown of
BIRC5 causes apoptosis in NB via mitotic catastrophe [
28]. Also, a previous publication show that NB tumours with genomic aberrations in G1-regulating genes leads to S and G2/M phase progression [
20]. Interestingly, the fork head box (FOX) gene
FOXM1 encoding a protein phosphorylated in M phase was significantly up-regulated in r2 and r3 subgroups. FOXM1 activates the expression of several cell cycle genes,
e.g. AURKB,
CCNB1,
CCND1,
MYC, and is involved in cell proliferation and malignancy [
29]. Several cell cycle and DNA repair genes, including
BIRC5, are suggested to act downstream of N-myc [
21,
30,
31]. In addition, most of the studied tumour suppressor (TS) candidates were specifically down-regulated in the r3 subgroup, which is probably explained by them acting downstream of N-myc. Several of the TS candidate genes are also located in the 1p36 chromosomal region (
e.g. CHD5 and
KIF1B[
32‐
34]), and Del1p is a well-known prognostic marker highly correlated to
MYCN-amplification in NB [
35]. One such N-myc-regulated and 1p36-localized TS candidate is
CDC42, encoding a small GTPase protein. This protein have a function in cell polarization and growth cone development in NB cell differentiation, similar to Rac1 and Cux-2, and is suggested to inhibit neuritogenesis in NB [
36]. In concordance to this, we found
CDC42 to be the 14
th most significantly down-regulated gene in the MNA subgroup (r3) compared to subgroup r2.
The main focus of the study was to define the underlying regulatory networks of the r4 subgroup. In contrast to the other three well-known subgroups of NB, the r4 tumours showed high expression of embryonic development and nervous system signalling genes. One of the most prominent genes from the differential expression analysis was
ERBB3, encoding a member of the epidermal growth factor receptor (EGFR) family of receptor tyrosine kinases (RTK’s). The ARACNE-modelled ERBB3-network was significantly enriched in the differentially expressed gene lists of the r4 subgroups (data sets 1-3), and this enrichment was also found in the GNB and GN histopathology categories of data set 4. Two members of the ERBB3-network,
S100B and
SOX10, were among the ten most significantly up-regulated genes in the r4 subgroups. The S100 calcium binding protein B (S100B) has long been reported as a prognostic biomarker of malignant melanoma [
37], and a paired down-regulation of
ERBB3 and
S100B is observed in malignant peripheral nerve sheath tumours confirming their functional relationship [
38]. Interestingly, the S100 beta protein, mapping to chromosome 21, has been proposed to be responsible for the lack of NB in Down syndrome patients by producing growth inhibition and differentiation of neural cells [
39]. The SRY box 10 transcription factor (Sox10) is a key regulator of the developing nervous system, and has been shown to control expression of ErbB3 in neural crest cells [
40,
41]. A paired overexpression of ErbB3 and Sox10 has been observed in pilocytic astrocytoma (PA) a common glioma of childhood, which verifies their network connection found in the current study [
42]. Also, Sox10 and S100 are routinely employed in the pathological diagnosis of neural crest-derived tumours [
43], and Sox10 serves as an embryonic glial-lineage marker in NT’s [
44]. By immunohistochemistry assessment, we found Sox10 to be expressed in both the schwannian cells and ganglion cells, whereas ErbB3 was found mainly in the mature ganglion cells. We could also verify the GN-specific expression of ErbB3 by immunoblot analysis.
ErbB3 is activated through ligand binding of neuregulin (NRG), leading to heterodimerization of ErbB3 to other ErbB members and subsequent phosphorylation. Activated ErbB3 regulates proliferation through downstream signalling of the phosphoinositol 3-kinase/AKT survival/mitogenic pathways [
25]. In the current study we found a significant correlation of
ERBB3 to its family members
EGFR and
ERBB2 in all four independent data sets.
EGFR and
ERBB2 were also both found to be significantly up-regulated in all r4 subgroups as well as in the GNB and GN tumours. Amplification of
ERBB3 and/or overexpression of its protein has been reported in numerous cancers, including prostate, bladder, and breast. Moreover, loss of ErbB3 function has been shown to eliminate the transforming capability of ErbB2 (also known as HER-2) in breast tumours [
45]. Although the extent of the role of ErbB3 is emerging, its clinical relevance in different tumours is controversial. There are a few studies of ErbB/HER receptor expression in neuroblastoma, showing that ErbB/HER family members in neuroblastic tumour biology is interrelated and complex, but their expression level may present a prognostic factor for patients outcome [
46‐
48].
The heat map of 25 genes including the 6-GeneSig genes,
ERBB-genes and TS-genes showed a very specific expression pattern among the different r-subgroups and histopathology categories. The similarity of expression profiles between the four data sets was striking. The correspondence of the r4 subgroups to the GNB and GN histopathology subtypes was obvious, and
ERBB3 appeared as a clear-cut marker for a GNB/GN-like expression profile. To demonstrate this further, a new 7-GeneSig (including
ERBB3) was constructed and used in a histopathology reassignment classification test. The 7-GeneSig successfully assigned 100% NB tumours, 62,5% GNB tumours, and 90% GN tumours into the correct histopathology category (Kappa measure of agreement p = 7.489E-17, data set 4). Also, all r4-tumour types from data sets 1–3 were categorized as GNB or GN tumours according to the 7-GeneSig. By these facts we conclude that the NB tumours previously assigned to the r4 subgroup by the 6-GeneSig, most likely represent more differentiated NT’s and are seemingly GNB/GN tumours types. Our study brings out the complexity in classifying neuroblastic tumours. The previously described unfavourable characteristics and poor outcome of the r4 tumour group is puzzling [
12], but can be explained by the fact that prognostic subsets of GNB’s exist [
49]. Historically, GNB’s have been the most difficult of the NT’s to define in a consistent and uniform fashion, because the number and degree of differentiation of the neuroblastic cells tend to vary between cases as well as between different microscopic fields in the same tumour [
1]. Moreover, the data sets used in the current study are probably not truly population-based, and the r4 subgroups found probably consist of different proportions of F/UF subsets. In addition, some tumours may previously have been misclassified as NB, or the tumour tissue part analysed by microarray may not be the same as the tissue part that underwent histopathology assessment. Furthermore, it is not clear whether differentiation markers are superior to other prognostic factors in defining outcome. Unfavourable markers such as MNA and clinical stage may also be present in or among differentiated cells, and mark a poor prognosis by themselves.
ErbB3 also has an important role in differentiation of Neural crest cell (NCC) lineages during the embryonic development [
50]. Although ErbB receptors are also found to mediate proliferation and survival [
47,
48], the ERBB-profile found in this study is likely to reflect the phenotype or differentiation stage of developing neuronal progenitors. Upon induction of differentiation, neuronal progenitors may follow a variety of stages of NCC lineages. For example, neuroblasts in culture are shown to represent an immature bilineage stage able to progress towards neuronal and glial fates [
44]. Schwannian cells are the principal glia of the peripheral nervous system, whereas neuroblasts differentiate from neural stem cells and exhibit variable degrees of differentiation up to ganglion cells. In this context, the ERBB-profile seems to be a marker of ganglionic-neuronal differentiation. A recent immunohistochemistry study of ErbB2 in neuroblastic tumours supports this conclusion [
51]. However, it still remains uncertain whether the r4 subgroup of datasets 1 and 3 are indeed GN or GNB, or if the ERBB expression profile just marks the gradually differentiated NB tumours (encompassing increased levels of mature ganglion cells). Nevertheless, the results from all data sets are consistent in regards to the expression profile of the 25 genes selected for the heat map, strengthening the robustness of the suggested 7-gene signature. Accordingly, we propose ErbB3 as an excellent marker of neuronal differentiation, and suggest mRNA expression profiling by the 7-gene signature as a complement to histopathological assessment. However, the exact cut-off expression levels for classification needs to be worked out in more detail, and classification must be based on international standard cases and assays.
Methods
Pre-processing microarray data
Data from five published neuroblastoma expression microarray studies run on three different Affymetrix platforms (HU133A, HGU95Av, and HU133plus2) were used in this study (Table
1). Raw data files were obtained from Array Express (
http://www.ebi.ac.uk/microarray-as/ae/) and Gene Expression Omnibus (
http://www.ncbi.nlm.nih.gov/geo/), or directly from collaborators. Expression data files were normalized by gcRMA using Bioconducter (library BioC 2.4) in R 2.9.2 [
52] in four separate groups; 1) the De Preter [
53] data set run on the HGU133A Affymetrix platform (17 samples, preamplified), 2) the McArdle [
54] and the Wilzén [
55] data sets run on the HGU133A Affymetrix platform (30 samples, not pre-amplified), 3) the Wang [
56] data set run on the HGU95Av2 platform (102 samples, not pre-amplified), and 4) the Versteeg [
57] data set run on the HU133plus2 platform (110 samples). For each probe-set, the maximum expression values over all samples were determined, and probe-sets which showed very low or no detectable expression levels were filtered out (log2 expression <5). Next, the mean log2 expression level for each Gene symbol was calculated to generate “mean-per-gene” data files: 7439 genes in data set 1, 8106 genes in data set 2, 7542 genes in data set 3, and 15614 genes in data set 4.
Differential expression analysis
NB samples from the DePreter and McArdle/Wilzén data sets were divided into four r-subgroups by a 6-gene signature (further referred to as the“6-GeneSig”) according to Abel et al., 2011 [
12] (Additional file
1). From these two data sets, 14 (preamplified, De Preter) and 23 (non-preamplified, McArdle/Wilzén) cases respectively were successfully assigned into one of the four r-groups (Table
1). Differential gene expression analysis was performed by a two group unpaired Significance Analysis of Microarray (SAM) test (
i.e. six comparisons) [
58]. Gene lists comprising the 1000 most significantly differentially expressed genes (sorted after the d-statistic) with a fold change above 2 were exported from each comparison, from each direction (up or down), and from each data set, separately (resulting in 12 SAM gene lists per data set). Next, SAM gene lists from the two different data sets were compared, and 12 intersection gene lists (SAM
intersect) were created. Too minimize the variance, a combined fold change (FC
comb) for each gene in the SAM
intersect gene list was calculated as follows:
where FC
i
is the fold change in data set
i and
where SE
i
is the standard error of the mean log2 expression values in data
set i.
A combined p-value (P
comb) for each gene in the SAM
intersect gene list was calculated as follows:
where N
i
is the total number of samples of the two groups compared by the
d-statistic in SAM, and P
i
the corresponding p-value for dataset i. Φ is the cumulative distribution function of the standard normal distribution and Φ-1 is its inverse function.
Based on an approximation of 8000 multiple tests (i.e. 8000 genes), a nominal p-value <6.25E-06 was found to correspond to an adjusted p-value <0.05 (according to Bonferroni correction) and was subsequently used as a cut-off for significance in SAM.
Gene network modelling
A large gene regulatory network was constructed from an independent data set (Wang) of 102 expression profiles [
56]. Mutual information values were estimated with the ARACNE (Algorithm for the Reconstruction of Accurate Cellular Networks) algorithm using a p-value cut-off of 1E-7 [
26]. The data processing inequality (DPI) was applied with a tolerance of 0.15. Gene networks of seven selected genes were extracted from the global network together with their immediate gene neighbours. The gene networks of nearest neighbours were visualized in Cytoscape 2.8.2.
Gene ontology (GO) and Gene Set enrichment analysis (GSEA)
Ranked SAM gene lists (by d-statistic) from the separate data sets were investigated for Gene Ontology terms using BiNGO 2.4 (Biological Network Gene Ontology,
http://www.psb.ugent.be/cbd/papers/BiNGO/). The Gene Set Enrichment Analysis (GSEA,
http://www.broad.mit.edu/gsea/) software was used to investigate whether a gene network was significantly overrepresented in the different r-subgroups. The enrichment tests were performed using seven ARACNE-constructed gene networks ALK (n = 12 genes), BIRC5 (n = 45 genes), CCND1 (n = 22 genes), ERBB3 (n = 38 genes), MYCN (n = 40 genes), NTRK1 (n = 62 genes), and PHOX2B (n = 67 genes), as well as 4850 MSigDB-curated gene sets (c2,
http://www.broadinstitute.org/gsea/msigdb/index.jsp, Additional file
6). The GSEA according to Subramanian
et al.[
59], was run on the “mean-per-gene” data files using the following settings: number of permutations = 1000, permutation type = gene-set, chip platform = GENE_SYMBOL.chip, enrichment statistic = weighted, metric for ranking genes = Signal2Noise, gene list sorting mode = real, gene list ordering mode = descending, max gene set size = 500 (the default), min gene set size = 10 (the default is 15). In addition, the r3 versus r1 comparisons in data sets 1–3 were investigated according to the gene list sorting mode = abs.
Human tissue samples used for protein expression validation
Tumours histopathologically classified as GN and NB (data set 5, Table
1) were used for immunohistochemistry (4 NB and 4 GN), and immunoblot analysis (4 NB and 5 GN). Tissue from patients was obtained during surgery and stored in -80°C. Ethical approval was obtained from the Karolinska University Hospital Research Ethics Committee (Approval no. 2009/1369-31/1 and 03–736). Informed consent for using tumor samples in scientific research was provided by parents/guardians. In accordance with the approval from the Ethics Committee the informed consent was either written or verbal. When verbal or written assent was not obtained the decision was documented in the medical record.
Immunohistochemistry
Formalin-fixed and paraffin-embedded (FFPE) tissue slides were deparaffinized in xylol and rehydrated in graded alcohols. For antigen retrieval, slides were boiled in a sodium citrate buffer (pH 6.0) for 10 min, in a microwave oven. After blocking in 1% bovine serum albumin (BSA) for 20 min, the tissue sections were incubated with primary antibody overnight, Sox10 ([N-20], Santa Cruz Biotechnology) and ErbB-3 ([RTJ2], Abcam) respectively, diluted 1:50 in 1% PBSA. Thereafter slides were rinsed in PBS and endogenous peroxidases were blocked in 0.3% H2O2 for 10 min. As a secondary antibody, anti-mouse-horseradish peroxidase (HRP) and anti-goat-horseradish peroxidase were used (Invitrogen, Paisley, UK). All slides were counterstained with haematoxylin. To control for non-specific binding, antibody specific blocking peptides and isotype-matched controls were used. For colocalization studies of Erb3 and Sox10, tumor tissue sections were simultaneously stained with primary antibodies and for fluorescence visualization, anti-goat Alexa Fluor 594 and anti-mouse Alexa Fluor 488 were used, respectively.
Immunoblot analysis
Tumours were homogenized in RIPA buffer (20 mM Tris–HCl, pH 7.5, 150 mM NaCl, 1 mM Na2EDTA, 1 mM EGTA, 1% NP-40, 1% sodium deoxycholate, 2.5 mM sodium pyrophosphate, 1 mM beta-glycerophosphate, 1 mM Na3VO4, 1 ug/ml leupeptin) with protease inhibitor cocktail (Roche), 42 mM DTT and 1 mM PMSF. The total protein concentration was determined using A280 absorbance readings and 100 ug of total protein was diluted in NuPAGE® LDS sample buffer (Invitrogen) with 50 mM DTT and denatured for 10 min at 70°C. The samples were then loaded with a prestained Page Ruler protein ladder (Thermo-Scientific) on a 4-12% NuPAGE® Bis-Tris polyacrylamide gel (Invitrogen) and separated using MOPS buffer at 200V for 50 min. The proteins were transferred to PVDF membranes using NuPAGE® transfer buffer (Invitrogen) and 10% methanol. Following Ponceau staining to ensure equal loading, membranes were washed with TBS-T (Tris-buffered saline containing 0.1% Tween 20) and blocked with blocking buffer (5% milk/TBS-T) for 1 h. The primary antibodies were added to the membranes and incubated overnight at 4°C. The following day, membranes were washed with TBS-T and incubated with secondary antibodies. Following final TBS-T washes, protein detection was achieved with Pierce Super Signal® West Pico or Femto Chemiluminescent Substrate (Thermo-Scientific). The primary antibodies used were anti-ErbB3 [RTJ2] (Abcam, 1:200) and anti-Gapdh (Abcam, #ab8245, 1:10000). The secondary antibodies used were anti-mouse IgG HRP linked antibodies (Cell Signaling, #7076, 1:5000), anti-rabbit IgG HRP linked antibodies (Cell Signaling, #7074, 1:5000). All antibodies were diluted in blocking buffer.
Statistical analyses
The expression relationship of ERBB3 to the discriminative 6-GeneSig (ALK, BIRC5, CCND1, MYCN, NTRK1, and PHOX2B) and the ErbB family members EGFR, ERBB2, and ERBB4 were investigated by a Pearson correlation test. The statistical significance of expression levels of ERBB genes (i.e. EGFR, ERBB2, ERBB3, and ERBB4) were investigated by Welch t-test. Inter-rater reliability of group assignments was tested by the Kappa statistic on crosstabs in SPSS (version 20.0).
Competing interests
The authors declare that no competing interests exist.
Authors' contributions
FA formulated the study design, and performed the microarray data pre-processing. AW and FA accomplished the analysis of SAM gene lists, GSEA, GO, and heat maps. AW and FA drafted the manuscript. CK performed the immunoblot analyses, and revised the manuscript. BS performed the immunohistochemistry analyses, and revised the manuscript. EK performed SAM analysis, and revised the manuscript. DD performed network modelling, and revised the manuscript. IØ performed the histopathology assessment of the Versteeg110 data set, and revised the manuscripts, KD, RS, JM, PK, and RV provided histopathology data as well as clinical data in terms of status of prognostic marker and survival of patients, and revised the manuscript. SN supervised the statistical analysis and interpretations of results, and revised the manuscript. All authors read and approved the final manuscript.