Background

DNA copy number and gene expression studies have highlighted key distinctions between high grade gliomas (HGG) arising in childhood and far more commonly, much later in adult life [1–4]. Indeed, recent exome-level sequencing initiatives have conclusively shown the existence of subgroups of HGG marked by distinct driver mutations [5], which are significantly enriched in young children (H3F3A K27M), teenagers and young adults (H3F3A G34R/V), and middle-aged adults (IDH1/2) [6]. Specific driving events for infants and elderly patients remain to be elucidated, however they too represent biological sub-entities, with infants having few genomic alterations [4], and elderly patients harbouring frequent amplification of EGFR and other genomic events [2, 3].

The identification of driving genetic alterations at the DNA copy level are necessarily most commonly focussed on assessing the amplification/deletion of genes in their entirety, and approaches to ascribe significance to genomic events make use of overlapping regions across multiple samples to find genes consistently within regions of gain/loss [7]. This approach has the result of ignoring genes for whom the breakpoint, i.e. the specific location of copy number change, is found within the coding regions. Such events may be more than mere bystanders of the “driving” aberration, and may themselves play significant roles in tumour initiation and maintenance.

One key implication of copy number breakpoints occurring within genes is the possibility of generating novel fusions. Gene fusions can occur through both intra- and inter-chromosomal translocations, bringing together coding regions from two or more genes within a single reading frame allowing expression of a novel protein. Such gene fusions are common in cancer, but have historically been thought to be largely restricted to haematological malignancies and selected solid tumours such as sarcomas. Recent evidence has overturned this, with numerous novel gene fusions being discovered in a wide range of cancer types, exemplified by the identification of common TMPRSS2:ERG fusions in prostate cancer [8] and the EML4:ALK fusion in non-small cell lung cancer [9].

The first fusion gene found in glioblastoma was the rearrangement located at an amplified region at chromosome 4q12, resulting in the fusing of the kinase domain of PDGFRA with the regulatory domains of KDR (VEGFR2) [10]. This KDR:PDGFRA was found to be activating and tumorigenic, however to date only a single additional case has been found, in a paediatric high grade glioma (pHGG) [11], and thus these fusions do not represent a common event. Another low frequency fusion has more recently been identified in approximately 3% of adult HGG, involving FGFR1 or FGFR3 partnering with TACC1 or TACC3[12]. These FGFR:TACC fusions have been shown to localize to mitotic spindle poles, have constitutive kinase activity and induce mitotic and chromosomal segregation defects and aneuploidy [12]. The types of integrated analysis that identified these mutations have also begun to identify more common rearrangements, such as numerous fusions involving EGFR, the most frequently seen partner producing the EGFR-SEPT14 fusion demonstrated to activate STAT3 signaling and confer mitogen independence and sensitivity to EGFR inhibition [13].

Such analyses are clearly proving extremely valuable in furthering our understanding of HGG biology and generating novel targets for therapeutic intervention. As similar approaches are yet to be undertaken in the paediatric setting, we have applied an algorithm designed to identify intragenic copy number breakpoints in our previously published study of DNA copy number [4]. We identify numerous potentially functional gene disruptions and a novel validated complex fusion, DHX57:TMEM178:MAP4K3.

Methods

Published DNA copy number data

We previously carried out a DNA copy number profiling study of 100 pHGG cases on Affymetrix 500 K SNP arrays [4]. The data have been deposited at the Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) with accession number GSE19578. Copy number assignment was carried out as per the original publication, using Affymetrix Genotyping Analysis Software (GTYPE version 4.s) improved using Bayesian Linear Model with Mahalanobis distance classifier algorithm (BRLMM) and standard dChipSNP normalization and model-based expression algorithms [4]. Log2-transformed data was used for all subsequent analysis in the present study.

iCNA algorithm

We implemented the iCNA algorithm developed as part of the GTS package under R2.11.0 (cbio.mskcc.org/~brennan) [14]. Breakpoints are calculated according to user-defined ‘delta’ values representing shifts in log2 ratios between two contiguous genomic regions after segmenting the copy number data using circular binary segmentation (cbs) [15]. Using a delta of 0.4, breakpoint boundaries are identified and errors estimated by permutation-based calculations of neighbouring probe data. Confidence intervals are assigned and those falling within the 95% window considered ‘high confidence’. An estimate is calculated for the expected rate of breaks for each gene based upon gene size and rate of breaks per sample, with a p value obtained based upon (observed-expected)/standard error. A corrected p value of < 0.05 is considered significant. Manual inspection of copy number plots was undertaken to ensure sufficient probe coverage was present at identified loci in order to prioritise the most convincing breakpoints. Those with substantial gaps at either side of the break were excluded.

Custom oligonucleotide array CGH

We designed two fine-tiling oligonucleotide microarrays to cover the specific amplicons observed at chromosome 2p22.1 and 10q11.21 This was undertaken using the Agilent custom array design tool e-Array (Agilent, Santa Clara, CA, USA; https://earray.chem.agilent.com/), and comprised 700 probes covering 43.56–43.70 Mb on chromosome 10 and 5000 probes covering 39–40 Mb on chromosome 10 with a median probe interval of 200 bp on 2x105K microarray. Due to limited amount of material, DNA was whole genome amplified (WGA) using the GenomePlex® Complete Whole Genome Amplification Kit (Sigma, Gillingham, UK) starting with 10 ng of sample and control DNA, and following the manufacturer’s protocol. WGA DNA was labelled using the Agilent Genomic DNA ULS labelling kit, hybridised as per manufacturer’s instructions, and scanned on the Agilent 2505B Microarray Scanner System. Data has been submitted to ArrayExpress with accession number E-MTAB-2340.

siRNA knockdown

siRNA was carried out using a Dharmacon SMARTpool™ (Dharmacon, Lafayette, CO, USA) against MAP4K3 (#003588) with paediatric glioma cells SF188, KNS42, UW479, Res259 and Res186 [16] and a panel of breast carcinoma lines. Cells were plated and transfected 24 hours later with siRNA using Lipofectamine RNAiMax™ (Invitrogen, Paisley, UK) as per manufacturer’s instructions, alongside transfections of siControl. Twenty four hours following transfection, cells were trypsinised and media replenished after 48 hours and 96 hours, with cell viability assessed after seven days using CellTiter-Glo™ Luminescent Cell Viability Assay (Promega, Madison, WI, USA) as per manufacturer’s instructions.

Results

Intragenic copy number breakpoints in paediatric HGG

We have previously carried out DNA copy number profiling of a large series of pHGG samples using Affymetrix 500 K SNP arrays, and reported numerous genes encompassed within areas of focal amplification and deletion [4]. We now applied an algorithm (iCNA [14]) designed to identify copy number breakpoints contained within the sequence of known genes. A full schema of the analytical process is given in Figure 1.

Figure 1
figure 1

Schema of iCNA algorithm applied to paediatric high grade glioma. 1099 intragenic breakpoints were initially identified in a published series of 100 pHGG. From these, known copy number polymorphisms were excluded, as were these seen in a series of 26 matched normal DNA samples. After filtering for statistical significance and manual inspection of copy number plots, a series of 19 intragenic deletions, 18 amplicons, and 2 candidate fusions were identified.

Full size image

This algorithm was applied to 100 pHGG and 26 matched normal DNA samples, resulting in the identification of 1099 unique DNA copy number breaks contained within gene sequences across all tumour samples (Additional file 1: Table S1). Of these, 479 were found to map to known regions of copy number variations found commonly in the germlines of the general population by cross-referencing the breakpoints with The Centre for Applied Genomics Database of Genomic Variants [17], leaving a total of 620 events.

These were filtered to 500 after excluding those with p values > 0.05, and further reduced to 388 with at least one sample harbouring a given aberration at ‘high confidence’. A further 100 of these were excluded as they were also found in at least one of the normal samples profiled, representing either technical artifacts associated with the array platform used, or low frequency normal copy number polymorphisms. A final list of 288 iCNA is provided in Additional file 2: Table S2.

Most pHGG samples harboured at least one iCNA (median = 3), although seven cases were found to contain none. Several cases were found to contain many more aberrations (maximum = 19), though these were in the minority (Figure 2a). The number of iCNA events per sample was found to be prognostic in this multi-institutional series of cases, with pHGG containing more than 10 iCNA (n = 9) found to have a significantly poorer survival (median = 7.8 months), and those with no iCNA a better survival (median = 24 months) than the rest of the tumours (median = 13.2 months) (p = 0.026, log-rank test) (Figure 2b).

Figure 2
figure 2

Clinicopathological correlates of intragenic copy number breaks in paediatric high grade glioma . (a) Number of intragenic breaks per sample in a series of 100 pHGG. (b) Kaplan-Meier plot of overall survival stratified by number of intragenic breaks per sample. Pink: no breaks; Red: more than 10 breaks per sample; Grey: rest. (c-f) Boxplot of number of intragenic breaks per sample separated by - (c) WHO grade. Orange: grade III; Brown: grade IV. (d) Previous radiation treatment for an earlier malignancy. Grey: primary pHGG; Purple: post-irradiation; (e) Age at diagnosis. Yellow: infant (<3 years old); Dark yellow: older children (>3 years old). (f) H3F3A status. Blue: G34R/V; Green: K27M; Grey: wild-type.

Full size image

There were no differences in the number of iCNA between grade III (n = 20, median = 3, range 0–7) and grade IV tumours (n = 58, median = 3, range 0–19) (p = 0.456, t-test), though the cases with the highest number of iCNA were all grade IV glioblastoma (Figure 2c). Similarly, there were no differences between primary tumours (n = 68, median = 3, range 0–19) and those which arose as secondary malignancies after cranio-spinal irradiation (n = 10, median = 2.5, range 0–11) (p = 0.698, t-test) (Figure 2d). Infants (less than 3 years at diagnosis) had significantly fewer iCNA (n = 10, median = 2.5, range 0–6) than older children (n = 68, median = 4, range 0–19) (p = 0.050, t-test) (Figure 2e). Tumours with the K27M mutation in the gene encoding the histone variant H3.3, H3F3A, harboured significantly more iCNA (n = 5, median = 6, range 1–16) than either G34R/V mutant tumours (n = 4, median = 4.5, range 3–7) or wild-type (n = 14, median = 3, range 0–11) (p = 0.043, ANOVA) (Figure 2f). This was independent of location of tumour, with no differences in number of iCNA between supratentorial GBM (n = 51) and DIPG (n = 7, p = 0.684, t-test) (data not shown).

Intragenic amplifications and deletions

The 288 iCNA were further subjected to individual manual inspection of the data plots in order to identify the most robust copy number shifts associated with intragenic breaks. This resulted in a list of 39 unique events in 51 samples (Table 1). The recurrent changes included copy number loss, resulting primarily in either the absence of the 3′ end of a gene or small deletions wholly within the coding sequence. These intragenic deletions included those targeting known tumour suppressors in glioblastoma such as NF1 (17q11.2, n = 2) (Figure 3a) and RB1 (13q14.2, n = 1) (Additional file 3: Figure S3), as well as putative novel GBM-associated genes including FAF1 (1p33, n = 2) and MTAP (9p21.3, n = 2) (Additional file 3: Figure S3). In addition, there were novel deletions in the protein phosphatase PTPRE (10q26.2, n = 2) (Figure 3b) and recurrent internal microdeletions in the gene CSMD3 (CUB and Sushi multiple domains 3) (8q23.3, n = 3), all of which overlapped to result in the loss of exon 4 (Additional file 3: Figure S3).

Table 1 Nominated intragenic copy number aberration candidates
Full size table
Figure 3
figure 3

Intragenic deletions in paediatric high grade glioma. (a) Recurrent copy number breakpoint within NF1 on chromosome 17q11.2 in two cases of pHGG. (b) Recurrent copy number breakpoint within PTPRE on chromosome 10q26.2 in two cases of pHGG. Dark pink: confirmed region of loss; Light pink: region within which breakpoint lies, as defined by the resolution of probes on the array.

Full size image

Copy number gains within gene coding regions tended to be associated with regions flanking known oncogenic amplicons. These included amplification of the MYCN locus at chromosome 2p24.3, which in case HGG077 breaks within the coding region of the kinase D-interacting substrate KIDINS220 (Figure 4a); amplification of PDGFRA at 4q12, harbouring an iCNA in CHIC2 in the same case (though only covered by two probes); and recurrent breakpoints in the gene encoding the potassium voltage-gated channel KCND2 at 7q31.31 in association with amplification of MET, though curiously this targeted either 5′ or 3′ ends in two different cases (Figure 4b). Similarly, common amplification events encompassing EGFR (7p12) and CDK4 (12q14) had intragenic breakpoints at both ends in cases HGG060 (LANCL2 and ECOP) and HGG029 (MYO1A and XRCC6BP1), respectively (Additional file 4: Figure S4).

Figure 4
figure 4

Intragenic amplifications in paediatric high grade glioma. (a) Copy number breakpoint within KIDINS200 on chromosome 2p25.1, flanking the MYCN amplicon in a case of pHGG. (b) Recurrent copy number breakpoint within KCND2 on chromosome 7q31.31 in two cases of pHGG, in both cases part of the MET amplicon, though targeting either the 5′ or 3′ end of the gene. Dark green: confirmed region of gain; Light green: region within which breakpoint lies, as defined by the resolution of probes on the array.

Full size image

Identification of novel fusion genes

For the most part, iCNA events resulted in an imbalance of certain regions of coding genes in isolation, with the predicted consequence a disruption of full-length gene expression. For certain events however, a 5′ end of one gene was found amplified at a similar copy number to a 3′ end of a second gene within the same case. We reasoned that such instances may represent candidate fusion genes, and we identified two such examples in our cohort.

The first was at chromosome 10q11.21 and reflected a single amplicon, breaking within the genes RET and CSGALNACT2 such that we propose a hypothetical fusion gene encompassing the 5′ regulatory regions of CSGALNACT2 and the 3′ kinase domain of RET. In order to determine the precise breakpoints to allow validation of this novel fusion, we designed custom oligonucleotide arrays spanning the amplicon in order to carry out high-resolution array CGH on the reference case HGG139, a relapse sample of glioblastoma in which this genomic event was not present in the primary tumour. Although the breakpoint for CSGALNACT2 was identified within intron 2, leaving the catalytic domains intact, the breakpoint within RET could not be accurately determined to closer than 10 kb between introns 1 and 2 (Additional file 5: Figure S5). As material was limited for this case, we were unable to confirm the precise nature of the putative CSGALNACT2:RET fusion by PCR-based techniques.

The second fusion candidate was located at an amplified region of chromosome 2p22.1 in case HGG063, an anaplastic astrocytoma. At Affymetrix 500 K SNP resolution, this appeared to be a single amplicon with breaks within the coding regions of the RNA helicase DHX57 and the transmembrane protein TMEM178 (Figure 5a). Applying the same approach as above, using custom-designed oligonucleotide arrays for high-resolution array CGH revealed two amplicons within this structure, with further intragenic breakpoints within the mitogen-activated protein kinase MAP4K3 (Figure 5b). Designing PCR primers to amplify across the highly specific breakpoints confirmed the presence of the fusion, which was further validated by direct sequencing (Figure 6).

Figure 5
figure 5

Identification of a novel complex fusion DHX57:TMEM178:MAP4K3. (a) Affymetrix 500 K SNP array of chromosome 2, highlighting two amplicons, the most telomeric encompassing MYCN, the more centromeric as 2p22.1 involving intragenic breakpoints in DHX57 and TMEM178 (green). (b) Custom oligonucleotide array of the 2p22.1 amplicon, revealing two amplified structures and three intragenic breakpoints, in DHX57, MAP4K3 and TMEM178 (green).

Full size image
Figure 6
figure 6

Cartoon depicting the structure of the DHX57:TMEM178:MAP4K3 fusion. A complex microhomology-mediated rearrangement of exons 1–12 of DHX57 (blue), exons 2–4 of TMEM178 (green) and exons 13–34 of MAP4K3 (orange) was confirmed by direct sequencing. Regions of microhomology are highlighted in grey.

Full size image

The resultant fusion gene, DHX57:TMEM178:MAP4K3, is a complex three gene fusion formed from a series of intragenic breaks, amplifications and inversions to produce a sequence comprising exons 1–12 of DHX57, exons 2–4 of TMEM178 and exons 13–34 of MAP4K3, associated with regions of microhomology (Figure 6). This would produce a protein with the zinger finger and DEAD-like helicase domains of DHX57, the claudin family transmembrane domains of TMEM178 and the citron domain of MAP4K3. S elective knockdown of MAP4K3 by siRNA leads to a significant reduction in cell viability in five paediatric glioma cell lines as assayed by CellTiter Glo, an effect not seen in 18/20 breast cancer cells (p = 0.0017, pHGG vs breast cancer, t-test) (Additional file 6: Figure S6).

Discussion

Comprehensive copy number profiling of adult and paediatric high grade gliomas was among the first data to demonstrate the biological differences between these similar-looking histological malignancies [18]. In this context, the focus has been on large-scale genomic copy number changes. A more refined analysis of copy number and exon-level expression data has identified new insights into genomic architecture and novel fusion proteins in adult glioblastoma [12, 13]. Here we leverage a large dataset we have previously generated [4] in the paediatric disease to carry out a scan of intragenic breakpoints, leading to the identification of novel gene disruptions and candidate gene fusions.

The presence of intragenic copy number aberrations was confirmed in the vast majority of pHGG cases, and was itself prognostic, with an absence of iCNAs conferring a longer overall survival in paediatric patients. This was associated with the infant age group, known to have a better clinical outcome than older children [19], and further highlights the biological distinctiveness of this age group. By contrast, the presence of large numbers of intragenic breaks conferred a shorter survival time, but was not a result of the grade of the tumour, nor associated with a second malignancy due to radiation treatment for an earlier cancer. We had previously reported an association of post-irradiated HGG with PDGFRA amplification and chromosome 1q gain [4], so it appears these are relatively selective radiation-induced changes, rather than reflecting a generalised genomic instability in secondary tumours from these patients. Importantly, we identified an increased number of iCNA in tumours harbouring an H3F3A K27M mutation, regardless of anatomical location. This is a group of thalamic and pontine HGG associated with a particularly dismal prognosis [18], for whom understanding the mechanisms of genomic instability and the identification of novel gene disruptions is of considerable interest.

The majority of intragenic breakpoints we identified were associated with gene disruption. This includes deletions of known tumour suppressors such as RB1 and NF1, but also more novel glioblastoma associated genes. FAF1 and MTAP were both recurrently targeted by intragenic deletion events in pHGG. These genes are localised close to known cyclin-dependent kinase inhibitors and tumour suppressors CDKN1C and CDKN2A/B, respectively, but both FAF1 and MTAP have recently been proposed to harbour tumour suppressor activity in their own right. FAF1 is associated with a FAS-mediated apoptosis response and restoration of the FAF1 protein in adult glioma cell lines significantly increases cell death [20], whilst in MTAP-deficient cells, methylthioadenosine, generated during polyamine biosynthesis, is not cleaved and the salvage pathway for adenine and methionine is absent [21]. It seems that such mechanisms are also likely in a subset of paediatric tumours.

Of note we identified novel deletions in the protein phosphatase epsilon, PTPRE. This has not been reported previously, although there are several reports of the tumour suppressive capacity of the related PTPRD[22, 23]. This gene also appears targeted by intragenic deletions, and human astrocytes lacking PTPRD exhibited increased growth, as it is thought the protein usually functions to dephosphorylate the oncoprotein STAT3 [23]. The wholly intragenic microdeletions observed in CSMD3 in four cases may represent another novel mechanism of gene disruption. CSMD3 encodes a gene with multiple CUB and Sushi domains whose function is poorly understood. Recently, CSMD3 was identified as the second most frequently mutated gene (next to TP53) in lung cancer, where it was demonstrated that loss of CSMD3 results in increased proliferation of airway epithelial cells [24].

Gene disruption may also play a significant functional role when known gain-of-function oncogenes are amplified. We report numerous intragenic breakpoints which may have been overlooked in the context of identifying the ‘driver’ event within a common amplicon, but which may themselves be tumorigenic. These include disruptions of KIDINS220, a functional mediator of multiple receptor signalling pathways and essential for cortical development [25, 26]; CHIC2, frequently deleted/rearranged in myeloid malignancies [27]; and KCND2, encoding a potassium voltage-gated channel, which is expressed in both neuronal and glial cells and has been shown to regulate ERK signalling in ganglioglioma [28]. All of these gene disruptions represent novel avenues for understanding the underlying biology of pHGG.

Of most interest was the use of the iCNA algorithm to identify potential novel fusion genes, as was demonstrated in adult glioblastoma with the identification of the KDR:PDGFRA fusion [10], which we also found in a case of pHGG [11]. Our analysis nominated two potential candidates – the first we were unable to conclusively validate, CSGALNACT2:RET. Such a putative fusion would retain the kinase domain of the RET oncoprotein, but would lose the autoregulatory portion of the protein, instead fusing it to the N terminal of chondroitin sulfate N-acetyl-galactosaminyltransferase 2. Although a precise cancer-related function has not been ascribed to the latter enzyme, it is though to play an important role in morphogenesis in zebrafish models [29, 30]. Whilst not validated, oncogenic RET rearrangements and fusions are common in thyroid and lung cancer [31, 32], and the presence of infrequent activating fusions in HGG do not seem unlikely.

We were able to validate a novel complex fusion involving three genes with intragenic breakpoints and amplification/rearrangement on chromosome 2p22.1. The resulting fusion gene, DHX57:TMEM178:MAP4K3 encompasses key regulatory domains from all three proteins, though a specific function is hard to predict. The helicase properties of the DHX57 component may be a candidate for oncogenicity, with numerous other DEAD-box helicases appearing to play a role in regulation of DNA repair, apoptosis and drug sensitivity [33]. MAP4K3 has been associated with several malignancies in both an oncogenic and tumour suppressor capacity [34, 35]. In particular, one function that has been ascribed includes activation of mTOR signalling via the TORC1 complex [36], a pathway commonly activated by diverse mechanisms in pHGG [18].

In the context of pHGG, although the kinase domain is not retained in the fusion, MAP4K3 plays some functional role as selective knockdown by siRNA leads to a significant and selective reduction in cell viability in paediatric glioma cell lines. Thus we hypothesise that the DHX57:TMEM178:MAP4K3 is activating as disruption of the protein would otherwise seem incompatible with tumour cell growth and proliferation.

Conclusion

In summary these data represent a key addition to our understanding of the genomic alterations driving pHGG and provide novel avenues for developing sorely-needed novel therapeutic strategies for children with these otherwise incurable tumours.