The experimental setup of the study is summarized in Figure 1. In order to identify DNA methylation biomarkers in NB, we first applied two genome-wide methylation screening methodologies on eight NB cell lines: microarray after re-expression analysis and massively parallel sequencing after capturing with a methyl-CpG-binding domain (MBD- seq). The genome-wide assessment of gene expression reactivation upon DAC treatment is an indirect method to detect DNA methylation as the influence of the demethylating effect is measured at the transcriptional level using oligonucleotide chips. Out of 54,675 probes, a total of 3,624 were upregulated after DAC treatment compared to untreated controls (RankProd false discovery rate (FDR) <5%), of which 1,665 were upregulated at least two-fold in at least one cell line. Using a cutoff of at least a two-fold difference between the DAC-treated and the untreated sample, 989 probes were re-expressed in at least 2 cell lines. In order to select specific and sensitive methylation biomarkers from this high number of reactivated probes, an integrated bioinformatics approach was applied. The 1,665 upregulated probes identified by RankProd analysis were further filtered using a genome-wide promoter alignment strategy, referred to as the 'broad approach' in Hoque et al. 20. This strategy consists of a genome-wide multiple alignment of promoter regions, where similar sequence regions thus cluster together and where the 'distance' (the number of nodes in the hierarchical alignment model) is shown to be able to predict novel biomarkers. Such approaches using DAC re-expression data have previously been successfully applied to enrich towards truly methylated genes 2021. We selected 150 genes that were either in the 'neighborhood' (less than 8 nodes away) of a known methylation marker or that clustered together in the promoter sequence alignment with a high number of reactivation events (at least two genes in the cluster showed at least three reactivation events). Integration with (NB) literature, using an in-house developed text-mining-based approach (using NCBI E-Utils to query PubMed, using all known gene aliases in combination with either DNA methylation-related or NB-related search terms), and selection for genes located in genomic regions reported as recurrently affected by DNA copy number changes in NB, eventually led to the selection of 120 candidate biomarkers, comprising 30 novel candidate markers and 90 known methylation markers in other tumor types. To obtain direct evidence for DNA methylation and to further select prognostic biomarkers, the selected 120 candidate biomarkers were tested on the DAC-treated and untreated NB cell lines CLB-GA, LAN-2, N206, SH-SY5Y and SJNB-1, and primary NB samples (9 low-risk survivors (LR-SURV) and 11 high-risk deceased (HR-DOD) patients; for details see Material and methods), using high-throughput methylation-specific PCR (MSP). In the NB cell lines, the DAC-treated samples show less methylation calls in comparison to untreated samples (130 MSP assays (64%) are more frequently methylated in the untreated samples), and taking all MSP assays into account the average number of methylated samples per assay is 0.39 for the DAC-treated cell lines versus 1.47 for the untreated cell lines (P = 0.0002), revealing dense methylation in genes upregulated upon DAC treatment and efficient demethylation by DAC (data not shown). The complete results of the initial high-throughput MSP screening on the primary NB samples can be found in Additional file 1.

The second genome-wide DNA methylation screening methodology we applied, to the same eight NB cell lines, was MBD-seq: massively parallel sequencing of methylation-enriched DNA fragments, whereby the enrichment is based on the capture of methylated sheared DNA using the high affinity of the methyl-CpG-binding domain () of the protein MBD2 towards methylated cytosines. Sequencing yielded 4.4 to 8.6 million paired-end reads, depending on the cell line, and after peak calling 70,816 to 112,412 peaks were detected, representing genomic regions methylated in the corresponding cell line. Between 7,612 and 11,178 of these peaks (around 10% of all identified peaks) are located in promoter regions of annotated genes (-1,500 bp to +1,000 bp around the transcriptional start site (TSS)). These 'methylation peaks' were visualized in the Integrative Genomic Viewer 22, showing that promoter regions that are well known to be heavily methylated in NB were confirmed - for example, the protocadherin β (PCDHB) family cluster (Additional file 2) 2324. In some regions (for example, in the promoter regions of HIST1H3C and ACSS3) it was also possible to distinguish different DNA methylation profiles between MNA (IMR-32, LAN-2 and N206) and MYCN single copy (SH-SY5Y, SK-N-AS, CLB-GA and SJNB-1) NB cell lines (Additional file 2). Using the R/BioC package DESeq 25, 510 regions were identified as differentially methylated between MYCN amplified and single copy cell lines, of which 95 are in close proximity to an annotated TSS (-1,500 bp to +1,000 bp). Also, some miRNAs appeared to be methylated in their promoter region.

After peak calling, we also performed gene set enrichment analysis 26, using a custom, ranked list of genes with at least one MBD peak present in a region -1,500 bp to +500 bp around its TSS, in order to explore whether promoter regions that are enriched after MBD capture are often re-expressed as well upon DAC treatment. This analysis clearly showed a high enrichment score for each cell line (enrichment scores from 0.32 to 0.36; FDR q-value <0.01), demonstrating that a large portion of methylated regions (captured by MBD) are indeed reactivated upon DAC treatment. The overlap between the two genome-wide datasets can be further explored by intersecting them. In total, 183 genes are both re-activated upon DAC treatment (at least 1 log2 difference after and before treatment) and have an MBD peak in their promoter regions (-1,500 bp to +1,000 bp around the TSS) in at least 2 of the 8 investigated NB cell lines. Of these 183 genes, 46 are both re-expressed and methylated in 3 cell lines, 9 in 4 cell lines and 5 in at least 5 cell lines.

As we feared that only using cell lines in the selection phase of potential prognostic DNA methylation biomarkers would lead to the identification of methylated markers not necessarily related with prognosis, six publicly available mRNA expression studies 2728293031323334 were included in the analysis. In these studies, which comprise mRNA expression data of 380 primary NB tumors, identifying differentially expressed probes (genes) between prognostic groups would allow us to pinpoint potential prognostic methylated promoter regions in our methylome maps. Finally, a rank-based scoring system was used to prioritize genes that show methylation, re-expression after DAC treatment and differential expression (related with risk) across the prognostic groups. This score scheme uses the individual ranks of each analysis. In brief, DAC reactivation is ranked according to FDR rate (as determined by RankProd analysis), MDB-seq data are ranked according to peak P-values and expression data are ranked according to FDR (determined by RankProd analysis). Each data source is given the same weight and a combined rank is calculated (for details, see Materials and methods). This scoring system combined all generated data and allowed us to select 43 top-ranking and thus strong prognostic methylation candidate genes without the need to use rather artificial threshold values for the different datasets.

For these 43 genes, 48 MSP assays were designed and tested on 3 NB cell lines (IMR-32, SK-N-AS and SH-SY5Y) and the HCT-116 DKO cell lines, along with an independent cohort of 89 primary NB samples. Within the 89 primary NB sample set all three prognostic groups (LR-SURV, HR-DOD and high-risk survivors (HR-SURV); for details see Material and methods) were approximately equally represented. The complete matrix with all MSP results of all samples and a global overview of the MSP results per assay can be found in Additional file 3. Over 60% of the designed assays indeed detected methylation for the respective marker in at least 10% of the selected NB tumors. Ten MSP assays (COL6A3, miR-1225, miR-3177, PCDHA6, PLXNC1, ANKRD43, ADRB2, APOE, miR-671 and QPCT) revealed methylation in at least 75% of the patient samples, and the MSP assays for KCND2, PRPH, KRT19 (assay 83159) and TNFRSF10D were methylated in 50% to 75% of the patient samples. We could also detect DNA methylation in the promoter region of miR-1225, miR-3177, miR-671 and miR-663, methylated in 99%, 99%, 79% and 4% of the patient samples, respectively.

Unique in this study is the use of three discrete prognostic patient groups, which allowed us to assess differential methylation across all these prognostic groups. Therefore, we performed hierarchical cluster analysis on the methylation data of all 48 MSP assays on the entire NB tumor cohort, revealing two clusters with a separation between high-risk (HR) and low-risk (LR) patients (heatmap in Additional file 3). Furthermore, the overall methylation pattern in the primary NB tumor samples was compared by calculating the number of methylation events for each sample. This indicates that HR patients show, on average, more methylation events compared to LR patients (P < 0.001; HR-DOD, 17.21 methylation events (95% confidence interval (CI) 15.62 to 18.81); HR-SURV, 17.13 methylation events (95% CI 15.81 to 18.46); LR-SURV, 13.00 methylation events (95% CI 11.86 to 14.14)). Also on the individual marker level, some MSP assays are differentially methylated across the prognostic patient groups: KRT19 and ACSS3. These genes are more frequently methylated in HR patients compared to LR patients (Table 1). Within the HR group, HIST1H3C shows a tendency to be more frequently methylated in HR-DOD compared to HR-SURV samples (21% in HR-DOD versus 7% in HR-SURV), while KRT19 (32% versus 48%) and ACSS3 (25% versus 47%) show the inverse pattern.

Some individual MSP assays were also associated with one or more NB risk factors (stage, MYCN status and age at diagnosis), and are thus potential prognostic biomarkers in NB (Table 1). In this analysis, the age at diagnosis was tested using two different age cutoffs. The 12 months cutoff was chosen as it was used for therapy stratification and as a criterion in the sample selection. The more recently established cutoff of 18 months 13536 was also taken into account. Newly discovered methylated markers are FAS, PRPH, CNR1, QPCT, HIST1H3C, ACSS3 and GRB10, methylation of which is associated with at least one of the NB risk factors. Table 1 further indicates that the difference in the methylation status of HIST1H3C and ACSS3 between MYCN single copy and MNA NB cell lines as detected by MBD-seq is reflected in the MSP results of the primary tumors as well, as HIST1H3C and ACSS3 are almost exclusively methylated in MNA samples.

Survival analysis using the complete MSP data set indicates that patients with less methylation events showed better survival rates than patients with a high number of methylation events (P = 0.01; Additional file 3), as this analysis principally discriminates HR and LR patients. In order to assess to what extent our MSP data set is able to predict overall survival (OS) in HR-SURV versus HR-DOD patients, leave-one-out decision tree analysis was performed and repeated 58 times (the number of HR patients). For this analysis, we only included the data from MSP assays (un)methylated in at least three samples. Comparison of the 58 generated decision trees showed that 4 DNA methylation biomarkers (CNR1, ACSS3, HIST1H3C and PRPH) are included in at least 50% of the resulting classifiers. Then, leave-one-out decision tree analysis was redone, but this time using only the methylation data of CNR1, ACSS3, HIST1H3C and PRPH. Afterwards, the predictions for all 58 HR samples were visualized in a Kaplan-Meier plot (Figure 2). This analysis indicates that the combined methylation status of CNR1, ACSS3, HIST1H3C and PRPH has the potential to discriminate between HR-SURV and HR-DOD patients (P = 0.058).

Survival analysis was also performed on the individual marker level. We first tested differences between the HR-DOD and LR-SURV groups using the univariate log-rank test (with multiple testing correction). This first analysis indicates that six genes (KRT19, FAS, CNR1, HIST1H3C, ACSS3 and GNAS) are significantly related to survival when comparing these patient groups. As we also want to discriminate the HR patient groups (HR-DOD and HR-SURV), we then used the entire dataset (all samples) to assess which of these six genes were associated with survival (in a specific stratum only, such as only in MYCN single copy samples). These results are shown in Table 2. According to log-rank tests, HIST1H3C methylation is associated with both OS and event-free survival (EFS), while GNAS methylation is associated with EFS. As NB is a heterogeneous disease, these biomarkers may be suited to a specific subgroup of patients for predicting survival. For example, HIST1H3C methylation only occurs in high-stage tumors with MNA (6/17 (35%) in HR-DOD patients versus 2/22 (9%) in HR-SURV patients). Figure 3 shows the Kaplan-Meier plots for HIST1H3C and GNAS methylation (OS or EFS and OS in specific strata related to one of the risk factors).

As it is known that promoter methylation may cause transcriptional silencing of the gene, we further measured the mRNA expression levels of five promising DNA methylation biomarkers that were methylated in a substantial fraction of HR patients (CNR1, GRB10, KRT19, PRPH and QPCT). Quantitative RT-PCR assays were developed and tested on 366 primary NB tumor samples. Table 3 displays the results of the comparisons of the expression levels of each DNA methylation biomarker between the different NB tumor stages, MYCN single copy and MNA tumors, the two age groups (using both the 12 and 18 months cutoff), and surviving and deceased patients. As an example, the mRNA expression levels of these genes across the NB tumor stages are depicted in Additional file 4. Out of the 366 primary NB tumors, 245 could be assigned to one of the prognostic groups defined in this study (Additional file 5), which allowed us to asses differential mRNA expression between these groups as well. For all genes mRNA expression levels were significantly higher in the LR group compared to the HR groups. As methylation of these genes was mainly detected in the HR groups, this suggests that methylation may contribute to the transcriptional silencing of these genes.

Survival analysis using Cox proportional hazards further shows that low mRNA expression levels of CNR1 (hazard ratio (HR) 0.768; 95% CI 0.619 to 0.953; P = 0.028), GRB10 (HR 0.613; 95% CI 0.433 to 0.866; P = 0.015) and PRPH (HR 0.714; 95% CI 0.566 to 0.922; P = 0.015) were significantly associated with poor survival. After dichotomization of the mRNA expression data, using the median relative mRNA expression value as a cutoff, Kaplan-Meier survival curves were plotted (log-rank test; Additional file 5).

An interesting observation in our MBD-seq and MSP data is the fact that HIST1H3C and ACSS3 are differentially methylated between MYCN single copy and MNA NB cell lines and primary tumors (Table 1; Additional file 2). To further explore this finding, the HIST1H3C and ACSS3 MSP assays were tested on 31 NB cell lines, of which 10 were MYCN single copy and 21 MNA (Additional file 4). In addition, we also profiled HIST1H3C and ACSS3 mRNA expression levels in these cell lines, in order to assess the direct relationship between promoter methylation and mRNA expression and to compare this relationship between MYCN single copy and MNA cell lines. The significant differential methylation status of HIST1H3C and ACSS3 between MYCN single copy and MNA samples was confirmed in the NB cell lines (HIST1H3C, methylated in 15/21 (71%) MNA cell lines and in 2/10 (20%) MYCN single copy cell lines, P = 0.018; ACSS3, methylated in 20/21 (95%) MNA cell lines and in 3/10 (30%) MYCN single copy cell lines, P < 0.001). Moreover, expression of HIST1H3C mRNA was significantly lower in methylated samples compared to unmethylated samples, both in MNA (P = 0.005) and MYCN single copy (P = 0.044) cell lines (Figure 4). These data support the idea that HIST1H3C promoter methylation contributes to transcriptional silencing of the gene. Figure 4 further indicates that the MYCN status itself is not significantly associated with HIST1H3C mRNA expression levels (P = 0.204). As ACSS3 is expressed at very low mRNA levels, we could not correlate its mRNA expression data with the methylation data (data not shown).

In total, 33 well-characterized NB cell lines, authenticated using array comparative genomic hybridization and short tandem repeat genotyping, were included in this study (Additional files 2 and 4). DNA was isolated using the QIAamp DNA Mini Kit (Qiagen, Venlo, The Netherlands). In addition, 109 primary tumor samples of NB patients were collected prior to therapy at the Ghent University Hospital (Ghent, Belgium), the University Children's Hospital Essen (Essen, Germany), Our Lady's Children's Hospital Dublin (Dublin, Ireland) or the Hospital Clínico Universitario (Valencia, Spain). Informed consent was obtained from each patient's guardian and the study was approved by the ethical committee of the Ghent University Hospital (approval number B67020109912). Clinical characteristics of the patients are shown in Additional files 1 and 3. All NB patient samples were assigned to one of three defined risk groups based on risk parameters (tumor stage, MYCN status and age at diagnosis) and disease outcome. First, HR patients that died of disease (HR-DOD) as defined by stage 2/3, MNA, DOD; stage 4, age at diagnosis <12 months, MNA, DOD; or stage 4, age at diagnosis >12 months, DOD (n = 39). Second, HR patients alive (HR-SURV) after follow-up time >1,000 days (n = 30). Third, LR patients alive (LR-SURV) defined by stage 1/2, MYCN single copy, follow-up time >1,000 days; stage 3, MYCN single copy, age <12 months, follow-up time >1,000 days (status at last known follow-up is alive; n = 40). The clinical data of the 366 primary NB tumors (SIOPEN/GPOH cDNA library 6), used to test the mRNA expression levels of the most promising DNA methylation biomarkers, can be found in Additional file 5.

Eight NB cell lines (CHP-902R, CLB-GA, IMR-32, LAN-2, N206, SH-SY5Y, SK-N-AS and SJNB-1) were grown in the presence of 3 µM DAC (Sigma, Bornem, Belgium)) for 3 days, as previously described, and untreated controls were also prepared 61. After harvesting, RNA was extracted with the RNeasy Mini kit (Qiagen), accompanied by RNase free DNase treatment on column (Qiagen). After RNA quality check on the Experion (Bio-Rad, Nazareth, Belgium), sample preparation, hybridization to Affymetrix Human Genome U133 Plus 2.0 oligonucleotide chips and scanning were carried out according to the manufacturer's protocol at the VIB MicroArray Facility. Standard quality metrics (simpleaffy BioC package 62 - boxplots, visual inspection of the slides, 5'-3' degradation plot) demonstrated that the oligonucleotide chip data were of good quality. The BioC affy package was used to normalize (gc-RMA normalization) the expression levels and to obtain present/absent (expression/no expression) MAS 5.0 calls for each probe set. For all cell lines and for each probe set, the number of reactivation events was counted (absent in untreated cells and present in treated cells). Expression data (before and after DAC treatment) have been deposited into the Gene Expression Omnibus [GEO: GSE31229], according to the MIAME guidelines.

DNA samples (1 µg DNA) of the eight NB cell lines were sheared (Covaris S2) to an average length of 200 bp. Fragment distribution was determined by the Agilent 2100 Bioanalyzer and the concentration was determined using the Quant-iT PicoGreen dsDNA HS Assay Kit (Invitrogen, Ghent, Belgium). Starting from 200 ng sheared DNA, the MethylCollector Kit (ActiveMotif, La Hulpe, Belgium) was used to enrich for methylated fragments. Library preparation for multiplex Illumina sequencing was done by combining the DNA Sample Prep Master Mix Set 1 (New England Biolabs, Frankfurt am Main, Germany) and the Multiplexing Sample Preparation Oligo Kit (Illumina). Size selection of the library was done on a 2% agarose gel. Fragments of around 300 bp (±50 bp) were excised and purified. Illumina library amplification (21 cycles) was performed and concentration was determined. Paired-end sequencing was used for high confidence mapping of captured fragments (2 × 45 bp sequencing - Illumina GAIIx, NXTGNT). Paired-end reads were mapped on the human reference genome (GRCh37) using Bowtie 0.12.7 and peaks were called using MACS 1.4beta. For differential methylation analysis, PCR duplicates were removed and sequence tags counted by using the BioC packages ShortRead and rtracklayer 6364. Sequence tag counts per sample were used to compose a count matrix that could be processed by the BioC package DESeq 25. Sequencing data (raw sequence files, WIG files for visualization of the mapping results and the BED peak files as determined by MACS) have been deposited into GEO [GEO:GSE31353].

MSP assays were designed to only amplify the bisulfite-converted target region of interest and do not anneal to genomic DNA. As each primer contains at least two CpG sites, this means that a PCR product will only be generated if the template is methylated. We choose not to design the according U primers (that would amplify the non-methylated bisulfite-converted DNA) as we do not assess methylation in a quantitative way. After in silico assay evaluation, 48 selected MSP primers (including the ACTB control; Additional file 3) were empirically validated on the Roche LightCycler 480 (LC480) using the in vitro methylated HCT-116 DKO (positive control), the HCT-116 DKO (negative control) and NB cell lines. Based on melting curve and amplicon size analysis, all assays were considered amplicon specific. The MSP assays were tested on 89 samples, selected from the previously described patient groups (31 LR-SURV patients, 28 HR-DOD patients and 30 HR-SURV patients; Additional file 3). A no template control (NTC) sample was loaded as well. For all samples, 500 to 1,000 ng DNA was bisulfite-treated (EZ DNA Methylation Kit, Zymo Research) and eluted in 40 µl elution buffer. Prior to MSP, bisulfite-treated DNA (BT-DNA) was amplified using the EpiTect Whole Bisulfitome Kit (Qiagen), starting from 100 ng BT-DNA. After amplification, the yield was determined by the Qubit 2.0 fluorometer in combination with the Quant-iT PicoGreen dsDNA BR Assay Kit (Invitrogen). The MSP was performed on the LC480 and plates were prepared using the Tecan freedom Evo robot, using a design that assures that all samples were tested for the same assay in the same run 65. MSP amplifications were performed in 10 µl containing 5 µl LC480 SYBR Green I Master Mix (2×; Roche, Vilvoorde, Belgium), 1 mg/ml bovine serum albumin (Roche), 1 mM MgCl₂ (Roche), 125 nM forward and reverse primer (IDT, Leuven, Belgium), sample (20 ng amplified BT-DNA) and nuclease-free water (Sigma). MSP conditions were as follows: activation for 10 minutes at 95°C, 45 amplification cycles (10 s at 95°C, 30 s at 60°C and 5 s at 72°C), followed by melting curve analysis (5 s at 95°C - melting curve from 60 to 95°C) and cool down to 45°C. Afterwards, the size of the amplicons was determined using the Caliper LabChip GX. A MSP assay was considered methylated if (1) its Cq value <35 (calculated by the LC480 software using the second derivative maximum method), (2) its melting temperature (Tm) differed no more than 2°C from that of the positive control sample, and (3) the amplicon length differed no more than 10 bp from the band size of the positive control sample. In addition, the band height, as determined by the LabChip GX software, was required to be higher than 20.

The mRNA expression levels of CNR1, GRB10, KRT19, PRPH and QPCT were profiled on the NB SIOPEN/GPOH cDNA library generated from 366 primary NB tumor samples (Additional file 5) 6. For each DNA methylation marker a qPCR mRNA assay was designed and validated in silico and in vitro (Additional file 5) 66. PCR plates were prepared as described in the previous section and RT-qPCR was performed on the LC480 as described in 6. Relative gene expression levels were then normalized using the geometric mean of five reference sequences (HPRT1, SDHA, UBC, HMBS and AluSq) 67. For HIST1H3C and ACSS3, a qPCR mRNA assay (Additional file 4) was designed and tested on 31 NB cell lines on which the corresponding MSP assay was tested as well. Here, qPCR amplifications were performed in 5 µl containing 2.5 µl SsoAdvanced SYBR Green Supermix (2×; Bio-Rad), 0.25 µl forward and reverse primer (5 µM each) and 2 µl cDNA sample (corresponding to 5 ng cDNA). Relative gene expression levels were normalized using the geometric mean of the reference sequences SDHA, UBC and AluSq. All RT-qPCR data analysis was done in qbase^PLUS version 2.0 (Biogazelle, Ghent, Belgium) 65. Logged and normalized qPCR data can be found in Additional file 4 and 5.

Statistical analyses were performed using IBM SPSS software version 19.0. All statistical tests were two-sided and P-values <0.05 were considered statistically significant. Differential methylation across the prognostic groups was determined by the Chi square test. The relationship between the methylation status and NB risk factors was determined using Fisher's exact test. Univariate survival analysis was performed with the Kaplan-Meier method and log-rank statistics (Mantel-Cox) to determine the impact of methylation status on EFS and OS. EFS was defined as the time between initial diagnosis and relapse or death of disease, or time between diagnosis and last follow-up if no event had occurred. OS is the time to disease-related death or last follow-up. Hierarchical clustering and leave-one-out decision tree analysis were performed using R 2.13.0 (rpart package). The relationship between logged mRNA expression levels and the prognostic groups, OS status and NB risk factors was determined using the nonparametric Kruskal-Wallis test or Mann-Whitney test. Hazard ratios between logged mRNA expression data and survival were estimated using the Cox proportional hazard model. Kaplan-Meier curves were created by dichotomizing the logged mRNA expression data, using the median mRNA expression value as a cutoff. For HIST1H3C, the relationship between logged mRNA expression levels and the methylation status of the gene, and the MYCN status, was determined using the Mann-Whitney test. For all the above mentioned statistical tests, multiple hypothesis testing correction was performed (Benjamini-Hochberg method by using the R function p.adjust).