Radiation-induced gliomas represent H3-/IDH-wild type pediatric gliomas with recurrent PDGFRA amplification and loss of CDKN2A/B

Long-term complications such as radiation-induced second malignancies occur in a subset of patients following radiation-therapy, particularly relevant in pediatric patients due to the long follow-up period in case of survival. Radiation-induced gliomas (RIGs) have been reported in patients after treatment with cranial irradiation for various primary malignancies such as acute lymphoblastic leukemia (ALL) and medulloblastoma (MB). We perform comprehensive (epi-) genetic and expression profiling of RIGs arising after cranial irradiation for MB (n = 23) and ALL (n = 9). Our study reveals a unifying molecular signature for the majority of RIGs, with recurrent PDGFRA amplification and loss of CDKN2A/B and an absence of somatic hotspot mutations in genes encoding histone 3 variants or IDH1/2, uncovering diagnostic markers and potentially actionable targets.

R adiation therapy (RT) constitutes an essential element in the standard treatment of many cancers, improving clinical outcomes of patients with childhood malignancies including leukemia and central nervous system (CNS) tumors 1 . However, radiation-induced malignancies are observed post-RT in a subset of patients 2,3 , especially in pediatric patients due to their prolonged follow-up in case of long-term survival 4,5 .
RIGs commonly display an aggressive clinical course associated with poor prognosis, making early and precise diagnosis (including distinguishing RIGs from recurrences of the primary tumor) crucial for optimal treatment planning 7,8 . This is particularly challenging in cases where putative recurrent CNS tumors are not biopsied or resected, and where material for histological or molecular confirmation is therefore lacking.
Histopathologically, most RIGs present as high-grade gliomas (HGG) reminiscent of their sporadic counterparts 7,8,13,14,19 , and histopathological features to distinguishing them from de-novo HGG are missing 6,7 . A report of morphological features of both MB and HGG within secondary tumors led to the hypothesis of transformation from MB (stem-)cells as a possible route for RIG development 11 , while the higher genetic homogeneity of RIGs compared to de-novo pediatric HGG is suggestive of a common origin for RIGs 7 . Genetic alterations including gain of chromosome arm 1q, loss of CDKN2A/B, and PDGFRA amplification or overexpression have been frequently observed in RIGs 13,[20][21][22] , and-in contrast to sporadic pediatric HGG-recurrent somatic hotspot mutations in genes encoding histone 3 variants or IDH1/ 2 were found to be uncommon 13,[20][21][22] . To date, however, most series have been relatively small, and there remains some lack of clarity about the origin of RIGs and the underlying molecular and biological mechanisms leading to their formation.
Here, we performed comprehensive molecular characterization of RIGs to detect distinctive, (epi-)genetic features which might predict or explain their formation after RT, and/or which may act as diagnostic or therapeutic markers.

Results
Clinical characteristics. Patients in our series developing a radiation-induced glioma (RIG, n = 32) were previously treated with cranial irradiation for MB (MB-RIG, n = 23) or ALL (ALL-RIG, n = 9). A variety of different RT regimens was administered, including craniospinal radiotherapy in MB patients or prophylactic cranial irradiation in ALL patients. In line with a higher probability of RIG occurrence in areas receiving a higher effective radiation dose, most MB-RIG tumors were encountered in the cerebellum (20/23, 87% vs. 3/23 in the cerebral hemisphere), the location with the highest dose in RT. ALL-RIG tumors predominantly appeared in the cerebral hemisphere (7/9, 78%, vs 3/ 23 in MB-RIGs, p < 0.01; Fig.1a), with only two cases arising in the posterior fossa. The gender ratio was slightly shifted towards males (male:female ratio: 1.28:1) (Fig. 1b). The latency for RIG occurrence ranged from 2 to 30 years in MB patients (n = 23, median: 5 years) and from 3-17 years in ALL patients (n = 9, median: 8 years; Fig. 1c).
For 16/23 MB-RIG and 5/9 ALL-RIG patients with available follow-up data, we observed an aggressive clinical course that was as bad as or even slightly worse than histone 3 K27M-mutant tumorswith 13 out of 16 MB-RIG and four out of five ALL-RIG patients showing fatal outcome during the follow-up period (OS: 6 months, range: 3-10 months vs 8.5 months, range: 2-18 months in MB-RIG patients; Suppl. Table 1). One ALL-RIG (RIG_09, follow-up period: 3 months) and 3 MB-RIG patients diagnosed at the age of 14 year (RIG_22, follow-up period: 9 months), 15 years (RIG_30, follow-up period: 3 months), and 39 years (RIG_14, follow-up period: 18 months) were known to be alive at last follow-up, with the latter showing signs of progression after 12 months (Fig. 2).
DNA methylation profiling of radiation-induced gliomas and primary medulloblastomas. DNA methylation patterns of both MB-and ALL-RIGs were analyzed by unsupervised clustering and t-SNE alongside 120 reference samples from pediatric HGG subgroups with known molecular features. The majority of RIGs after both primary indications closely resembled those of pedGBM_RTK1 tumors (29/32, 91%; Fig. 3a, b) 32 . Two MB-RIG and one ALL-RIG showed a methylation pattern more similar to pleomorphic xanthoastrocytomas (PXA) (Fig. 3b). Thus, RIG DNA methylation patterns suggest a high degree of similarity (and likely common origins) between tumors, and were clearly distinct from any other HGG DNA methylation class.
The 11 pre-RIG primary MBs for which DNA was available were also profiled, and found to represent all four major MB subgroups (MB-WNT: 1/11; MB-SHH: 3/11; MB-Group 3: 2/11, MB-Group 4: 5/11), with no obvious enrichment for a particular primary subtype relative to their overall incidence rate. DNA methylation and CNV profiling confirmed the distinct biological nature between the primary MB and secondary RIG using t-SNE and copy-number analyses (Fig. 4a, b). Histological staining for hematoxylin and eosin (Fig. 4c) of two paired primary medulloblastoma and radiation-induced glioma couples (MB/ RIG_28, MB/RIG_29) display distinct morphological characteristics.
Genomic profile of radiation-induced gliomas. RNA sequencing (n = 9), gene panel (n = 5), and whole-exome (n = 13) sequencing with matched blood samples were performed for a subset of RIG samples. None of the RIGs harbored somatic hotspot mutations in genes encoding histone 3 variants, IDH1/2, BRAF, or in the TERT promoter region. Four out of 18 tested samples (22%) exhibited somatic mutations in the TP53 gene. Additional somatic mutations included CBL, PDGFRA, NTRK2, EGFR, RAF1, ATRX, and BCOR (Suppl. Table 1). One patient (RIG_25) harbored a TP53 germline splice site mutation. One patient (RIG_14) was clinically diagnosed with Li-Fraumeni syndrome, but tumor material was insufficient for molecular testing. Further germline alterations associated with cancer predisposition syndromes were absent in the remainder of cases with germline material available (n = 13).
Analysis of RNA sequencing data (n = 9) revealed relevant gene fusions including PTPRZ1:MET (RIG_11, RIG_25), CAP-ZA2:MET (RIG_09), FYCO1:RAF1 (RIG_08), and GFAP1:NTRK2 (RIG_13) (Fig. 6). Interestingly, the FYCO1:RAF1 and GKAP1:NTRK2 fusions were detected in two of the three tumors from the PXA-like subgroup based on DNA methylation profiling (Fig. 3). Both rearrangements are predicted to lead to a constitutive activation of the kinase domain, via loss of the N-terminal regulatory domain (RAF1) or constitutive dimerization (NTRK2). Deriving from a complex structural rearrangement, PRPRZ1:MET fusion proteins harbor an almost full length MET protein driven by the highly active PTPRZ1 promoter. All MET-fused cases (RIG_09, RIG_11, RIG_25) demonstrated amplification of MET on copy-number profiles derived from DNA methylation arrays. While all four genetic rearrangements, which result in aberrant MAPK/ERK pathway activation, were reported in high-grade gliomas, none of them were previously discovered in RIGs [34][35][36] .

Discussion
RIGs represent a fatal long-term side effect of cranial irradiation 2,3,19,37-43 . A better understanding of the occurrence of    32 . This suggests that both radiation-induced and sporadic pedGBM_RTK1 tumors might share a common cell of origin, which could be particularly vulnerable to ionizing radiation. However, the outcome of RIG patients is particularly poor, and worse than what has previously been described for sporadically occurring PDGFRA-amplified, 9p21-deleted pediatric GBM RTK1 tumors 32 . The discrepancy could suggest that additional oncogenic mechanisms might play a role during RIG formation, e.g. the trend towards greater genomic instability in the RIG cohort, as indicated by the elevated frequency of chromosomal breakpoints compared to their sporadic counterparts, potentially caused by radiation-induced DNA double-strand breaks. On the other hand, the RIG patient cohort also typically represents children with a long history of often aggressive prior treatment for their primary malignancy, which means that certain treatment options typically available at first diagnosis of GBM (e.g. radiotherapy) may no longer be an option for RIG patients 44,45 . It is also anticipated that the general health status of RIG patients is likely to be worse than that of primary-diagnosis counterparts.
Further, some RIGs arising after MB were treated with an MB relapse protocol on the assumption that the new lesion was a recurrence of the primary tumor in the absence of biopsy confirmation, posing a possible risk that these protocols may be less efficacious against high-grade glioma than upfront GBM protocols. Thus, there are a number of key clinical reasons why an inferior survival might be expected, in addition to molecular mechanistic differences. A small subset of RIGs with a more PXA-like profile harbored potentially druggable alterations in RAF1 and NTRK2, and it will be of interest to investigate the precise characteristics of this group in larger cohorts. Our findings also do not at present suggest a substantial contribution of germline alterations in known tumor-associated genes in the occurrence of RIGs. However, further comprehensive germline analyses with matched blood samples are required to elucidate in full the potential role of hereditary predisposition in RIG formation.
RIG following MB of various subtypes as well as after ALL displayed converging fingerprints, despite the diversity of applied radiation protocols (MB-RIGs arose mostly at the site of local boost, while ALL-RIGs occurred throughout the cranial radiation field). All of the four main molecular MB subgroups were represented in the primary lesions, with no recurrent alterations observed in the primary MBs. Our findings unfortunately suggest that no conclusions can currently be drawn regarding predictive features in the patient or primary tumor for assessing the probability of subsequent RIG formation. An important question to address in future will be the precise role of radiotherapy protocols (craniospinal vs local boost, fractionation schema, photons/protons/heavy ions) in determining the risk of secondary malignancy [46][47][48][49][50] . The ALL-RIGs included in our series were all diagnosed with their second tumor between 9 and 18 years of age, suggesting a particularly vulnerable but relatively narrow time frame for RIG occurrence after ALL. If confirmed in larger series, this could be an important variable for the planning of surveillance in this patient population. In contrast, both the age at diagnosis and latency after the primary tumor were much more varied for medulloblastoma. The earliest RIG occurred after a latency period of only 2 years following treatment for MBwithin a timeframe which could easily allow misidentification as a recurrence of the primary tumor. Molecular characterization of 31 clinicallypresumed relapse MBs within the prospective INFORM study 51 revealed that 6/31 'relapsed' MBs (19%) were in fact secondary HGG (unpublished data). In light of their aggressive clinical course and fundamentally different biology, it is essential to distinguish newly-arising RIGs from true recurrences in order to adjust treatment planning as early as possible to increase treatment efficacy. Thus, biopsy of recurrent lesions after MB is clearly warranted and should be considered whenever surgically feasible. Furthermore, the observation that a subset of prognostically favorable MBs (e.g. the MB-WNT subgroup) are also at risk of developing RIGs, highlights the urgency of current efforts to further examine the application of alternative treatment modalities (e.g. proton/heavy ion beam therapy) or the general deintensification of radiation therapy in these patients 47,[52][53][54][55][56][57][58] .
In summary, our findings demonstrate that RIGs are an aggressive, relatively homogenous group of CNS tumors with recurrent amplification of PDGFRA and loss of CDKN2A/B in the absence of somatic H3/IDH hotspot mutations. Our study uncovers possible similarities in origins with the pedGBM_RTK1 group of sporadic HGG, which will be important for further understanding the mechanisms by which these secondary tumors arise.

Methods
Patient population and tumor samples. Criteria for patient selection adapted from Cahan et al. 59,60 included (1) a glioma emerging from the previously irradiated field, and (2) a non-glial primary malignancy, in this series MB (n = 23) and ALL (n = 9), to exclude potential tumor progression/recurrence not directly related to radiotherapy 61-63 . This study is covered by the ethical approval of the Heidelberg University's medical faculty. Additional clinical data and tumor material were collected by international collaborating centers according to local ethical and institutional review board approval and collected at the German Cancer Research Center (DKFZ, Heidelberg, Germany). Informed consent for molecular profiling was obtained from all patients and their legal representatives.  merged to a combined data set by selecting the intersection of probes present on both arrays (combineArrays function, minfi). Each sample was individually normalized by performing a background correction (shifting of the 5th percentile of negative control probe intensities to 0) and a dye-bias correction (scaling of the mean of normalization control probe intensities to 10,000) for both color channels. Subsequently, a correction for the type of material tissue (FFPE/frozen) and array (450k/EPIC) was performed by fitting univariate, linear models to the log2transformed intensity values (removeBatchEffect function, limma package version 3.34.5). The methylated and unmethylated signals were corrected individually.
Beta-values were calculated from the retransformed intensities using an offset of 100 (as recommended by Illumina). Before further analysis, the following filtering criteria were applied: Removal of probes targeting the X and Y chromosomes (n = 11,551), removal of probes containing a single-nucleotide polymorphism (dbSNP132 Common) within five base pairs of and including the targeted CpG-site (n = 7998), probes not mapping uniquely to the human reference genome (hg19) allowing for one mismatch (n = 3965), and 450k array probes not included on the EPIC array. In total, 428,230 probes were kept for downstream analysis [64][65][66] .
To perform unsupervised dimension reduction, the remaining probes were used to calculate the 1-variance weighted Pearson correlation between the samples. The resulting distance matrix was used as input for t-SNE analysis (t-Distributed Stochastic Neighbor Embedding; Rtsne package version 0.13). The following nondefault parameters were applied: theta = 0, pca = F, max_iter = 2500 perplexity = 20.
To perform unsupervised hierarchical clustering, the 10.000 probes with highest standard deviation were selected to calculate the Euclidean distance between samples, followed by applying Wards linkage method for sample clustering. In the heatmap, representation probes were reordered by complete linkage hierarchical clustering of the Euclidean distance between probes.
To evaluate focal amplifications and deletions and chromosomal gains and losses, we visually inspected copy-number profiles of each case. Candidate genes and their 3′ and 5′ intergenic neighborhood were further investigated using the Integrative Genomic Viewer (IGV) for the presence of breakpoints, as an indication for potential gene fusions.
Gene expression profiling. Tumor samples with sufficient high-quality RNA were analyzed on the Affymetrix GeneChip Human Genome U133 Plus (v.2.0) Array (Affymetrix, Santa Clara, USA) at the Microarray Department of the University of Amsterdam, the Netherlands or the Genomics ad Proteomics Core Facility of the German Cancer Research Center (DKFZ). Subsequent library preparation, hybridization, and quality control was conducted following the manufacturer's guideline. Expression data were normalized using the MAS5.0 algorithm. Identification of significantly differentially expressed genes between subgroups (test: anova; correction for multiple testing: false discovery rate, transform: z-score) and unsupervised hierarchical clustering were performed within the R2: Genomics Analysis and Visualization Platform (http://r2.amc.nl).
Next-generation gene panel, whole-exome, low-coverage whole genome, and RNA sequencing. For the detection of single-nucleotide variations (SNVs), small insertions and deletions (indels), and gene fusions, a subset of tumor samples (n = 5) were analyzed via a customized enrichment/hybrid-capture-based nextgeneration sequencing gene panel 67 . DNA from FFPE tissue was extracted using the Promega Maxwell device (Promega), sheared on Covaris M220 (Covaris), according to the manufacturer's guidelines. Following a successful quality control using a Bioanalyzer 2100 (Agilent), sequencing was performed on a NextSeq 500 instrument (Illumina). Paired-end sequencing was applied to increase the detection sensitivity of duplicates and possible gene fusions.
For whole exome (WES; n = 13) and RNA sequencing (n = 9), library preparation, sequencing and data processing were conducted following the pipeline established in the INFORM study 51 . In brief, library preparation for WES was performed using the Agilent SureSelectXT Human V5 kit. Prepared libraries were sequenced together with a tumor cDNA library (poly(A) + RNA, Illumina TruSeq RNA Kit v2) on an Illumina HiSeq. The 1000 Genomes phase 2 human reference assembly (NCBI build 37.1) was selected for mapping the sequencing reads using BWA (version 0.6.2). For detection of SNVs and indels, custom pipelines were used 68 . In brief, the computational analysis was performed using SAMtools mpileup and bctools version 0.1.19 to detect somatic variants. To annotate variants, ANNOVAR was applied with Gencode version 17. All high confidence coding or splice site germline variants in a selected panel of cancer predisposition genes were extracted using a custom Python script. To call short insertions or deletions (InDels) in tumor and control blood BAM files, the extracted data was annotated using Platypus (version 0.5.2/0.7.4 with parameters genInDels = 1, genSNPs = 0, ploidy = 2, nIndividuals = 2). Poor genotype quality and low variant counts in the tumor were excluded for the subsequent analysis. RNA sequencing data were analyzed with deFuse to detect gene fusions 69 .
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
Methylation array data from this study have been deposited in GEO: GSE131482. Gene expression array were deposited in GEO: GSE168457. DNA sequencing data are available from the European Genome-Phenome Archive (EGA) under accession EGAS00001005243. Source data are provided with this paper.