Clinically Actionable Insights into Initial and Matched Recurrent Glioblastomas to Inform Novel Treatment Approaches

Glioblastoma is the most common primary adult brain tumour, and despite optimal treatment, the median survival is 12–15 months. Patients with matched recurrent glioblastomas were investigated to try to find actionable mutations. Tumours were profiled using a validated DNA-based gene panel. Copy number variations (CNVs) and single nucleotide variants (SNVs) were examined, and potentially pathogenic variants and clinically actionable mutations were identified. The results revealed that glioblastomas were IDH-wildtype (IDHWT; n = 38) and IDH-mutant (IDHMUT; n = 3). SNVs in TSC2, MSH6, TP53, CREBBP, and IDH1 were variants of unknown significance (VUS) that were predicted to be pathogenic in both subtypes. IDHWT tumours had SNVs that impacted RTK/Ras/PI(3)K, p53, WNT, SHH, NOTCH, Rb, and G-protein pathways. Many tumours had BRCA1/2 (18%) variants, including confirmed somatic mutations in haemangioblastoma. IDHWT recurrent tumours had fewer pathways impacted (RTK/Ras/PI(3)K, p53, WNT, and G-protein) and CNV gains (BRCA2, GNAS, and EGFR) and losses (TERT and SMARCA4). IDHMUT tumours had SNVs that impacted RTK/Ras/PI(3)K, p53, and WNT pathways. VUS in KLK1 was possibly pathogenic in IDHMUT. Recurrent tumours also had fewer pathways (p53, WNT, and G-protein) impacted by genetic alterations. Public datasets (TCGA and GDC) confirmed the clinical significance of findings in both subtypes. Overall in this cohort, potentially actionable variation was most often identified in EGFR, PTEN, BRCA1/2, and ATM. This study underlines the need for detailed molecular profiling to identify individual GBM patients who may be eligible for novel treatment approaches. This information is also crucial for patient recruitment to clinical trials.


Introduction
Gliomas are the largest group of intrinsic brain tumours with age adjusted incidence rates ranging from 4.67 to 5.73 per 100,000, causing more years of life lost compared with other cancers [1,2]. Glioblastoma (GBM) is the most malignant glioma and is classified molecularly as IDH-wildtype and IDH-mutant GBM [3][4][5][6][7][8][9][10]. During gliomagenesis, an array of genetic alterations may cause the dysregulation of cell growth signalling and cell cycle pathways [6,[11][12][13][14][15]. In particular, mutations in RTKs (receptor tyrosine kinases) and/or loss of PTEN (phosphatase and tensin homolog) alter the PI3K (phospinositide 3-kinase)/AKT cell growth pathway [11]. Further mutations in CDKN2A or CDK4 (cyclindependent kinase) lead to uncontrolled progression of the cell cycle, as do mutations in TP53 [16]. Neural stem cells in the subventricular zone may harbour recurrent driver somatic mutations that are shared with the tumour bulk (e.g., P53, PTEN, EGFR, and TERT) [17]. Telomerase (reactivation or reexpression) can occur in IDH wildtype and mutant GBMs driven either by telomerase reverse transcriptase (TERT) promoter mutations or other mechanisms [8,18]. e current standard-of-care for glioblastomas remains as maximal safe surgical resection with concurrent radiotherapy and temozolomide (TMZ) chemotherapy (Stupp protocol) [19,20]. Personalised therapies remain promising although trials have been unsuccessful to date [21][22][23]. For example, dysregulated PI3K and RTKs (EGFR, MET, PDGFR, FGFR, and BRAF) genes have been targeted with various small molecules, antibodies, and inhibitors [24][25][26][27][28][29]. To date, entry to clinical trials for GBM has not been based on a detailed molecular analysis of an individual patient's tumour using high throughput sequencing (HTS). HTSbased molecular diagnostics can aid the detection of genetic alterations, information required for personalised medicine [30,31]. Herein, initial and matched recurrent glioblastomas were examined using HTS with a validated DNA-based diagnostic panel. Potentially pathogenic variants and clinically actionable mutations were identified in different GBM subtypes. Findings were validated using TCGA-GBM and GDC datasets.

Clinical Specimens. Ethical approval was given by Brain
Tumour Bank South West and Brain UK (Ref: 14/010). All patients had been treated using the Stupp protocol [19]. A total of 72 formalin-fixed paraffin-embedded (FFPE) samples from 54 patients were identified (2009)(2010)(2011)(2012)(2013)(2014). Only FFPE slides with >30% tumour cells available for macrodissection were selected. Samples lacking cellularity or excessively necrotic were excluded. Following quality control, 67 samples for 46 patients and 19 with matched recurrent samples available were identified. Of these, a total of 49 samples were successfully sequenced for 41 patients (21 males; 20 females; mean age 55 years, range 16-78 years; see Tables 1 and S1). Matched initial and recurrent tissue samples were analysed for 8 patients (2 males; 6 females). Recurrent tumours all occurred locally to the initial tumour. Anonymised patient cases in the GBM cohort were numbered 1-11, 16-41, and 43-46, and "a" and "b" indicated initial and recurrent tumour samples, respectively (Table S1).

HTS Neuro-Oncology Gene Panel. A published HTS
DNA-based panel that uses targeted enrichment to examine exonic, selected intronic and promoter regions of 130 clinically relevant neuro-oncology genes was utilised (see Table S2) [30]. e diagnostic panel has been optimised for use either with fresh-frozen or FFPE tissue. Validation studies of the HTS panel analysing ∼200 single nucleotide variants (SNVs), gene fusions, and copy number variants (CNVs) showed 98% concordance with single marker tests [30]. Using the HTS panel, genetic alterations in tumours were characterized, and TERT promoter and IDH1/2 status confirmed.
2.3. DNA Extraction, HTS Library Preparation, Sequencing, and Analysis. Slides were deparaffinised and rehydrated using xylene and ethanol and left to dry. Tissue sections were then microdissected and placed into 180 uL ATL buffer. DNA was extracted from tissue sections (10 × 10 μm) according to manufacturer's instructions using the QIAamp DNA FFPE Tissue Kit (Qiagen, Manchester, UK). Following assessment of DNA quality and quantity, libraries were prepared using 200 ng of genomic DNA with an optical density 260/280 ratio between 1.8 and 2.0. Libraries were constructed using the SureSelect XT Target Enrichment System for Illumina Paired-End Multiplexed Sequencing Library protocol (Agilent). PCR master mixes were prepared using the SureSelect XT Library Prep Kit ILM following manufacturer's guidelines. In accordance with Illumina guidelines, libraries with a concentration of 4 nM were diluted to 20 pM, denatured, and sequenced on a NextSeq 500 (Illumina). HTS data were analysed following the pipeline described by Sahm et al. [30]. In brief, raw reads were demultiplexed, converted to fastq, quality checked, and manually trimmed when necessary. Paired-end reads were aligned to the human genome (version GRch37; hg19), and duplicate sequences were removed.

CNV Analysis in the GBM Cohort.
CNVs were investigated using a coverage analysis. e ratio of on-and offtarget reads, coverage per target region, and mean coverage per sample were estimated using the R package TEQC [32]. Measures provided an estimate of read depth, as the number of reconstructed strands across a region of interest, and this was utilised for CNV estimation of genes. Data normalisation and CNV comparison to a reference control were made using the R package seqCNA [33]. is method has previously been validated with 100% concordance for 47 GBM cases using 450 k data [30]. Potential CNV gain or loss is indicated by deviations from a proportional read depth of 50%, considered a normal gene copy number.  [30]. In brief, variants were called using SAMtools mpileup [34]. Variant calls were then filtered by (a) read depth ≥ 40, (b) genotype quality ≥ 99, (c) minimum allele frequency set at 10, and (d) at least 10% read coverage from each strand using the R package VariantAnnotation [35]. TERT promoter position calls were not filtered due to their low detection rate because of difficulties with their amplification as a GC-rich region [30]. Nonsynonymous filtered variants were annotated with the most up to date information including dbSNP and COSMIC identifiers using the online tool wANNOVAR [36]. Matched normal tissue was unavailable for comparison for the identification of germline mutations. us, to try to discern pathogenic from benign variants, the frequency of a variant in the general population was used as a key criterion in their clinical interpretation to try to exclude germline mutations. SNVs were filtered to those with a frequency of ≥0.01 in the 1,000 Genomes database and ≥0.05 in the Genome Aggregation Database (gnomAD), previously known as the Exome Aggregation Consortium database. gnomAD warehouses whole genome sequences from 15,496 unrelated individuals [37]. As the ethnicity of patients in the GBM cohort was unknown, SNV frequencies were compared to overall frequencies (rather than regional) of both databases. Filtered SNVs impacting genes were categorised into biological pathways using GeneCards [38]. SNVs occurring in the potentially clinically actionable genes: EGFR, PTEN, CDKN2A, RB1, TP53, ATM, ATR, MSH6, PDGFRA, PIK3CA, PIK3R1, SMO, PTCH1, BRCA1, BRCA2, and BRAF, were quantified in the initial and matched recurrent tumours. Further filtering was applied to SNV results to try to identify variants of unknown significance (VUS) that are possibly pathogenic and underpin gliomagenesis. VUS considered to be possibly pathogenic, were those that had no frequency recorded in the 1,000 Genomes database, and were predicted to be damaging by both LJB SIFT and FATHMM-MKL software [39]. All genomic positions listed for SNVs identified by this study are from the human genome version GRch37.

VUS and CNV Analysis in the TCGA-GBM and GDC
Datasets. VUS identified as possibly pathogenic mutations in the GBM cohort were further investigated for supporting evidence of their clinical significance using TCGA-GBM and GDC datasets. Frequencies of cases with mutations in genes were investigated in the GDC data portal. Abundance of mutations and copy number alterations within the TCGA-GBM dataset was visualised as an oncoprint plot generated using GlioVis, a data visualisation tool for brain tumour datasets [40].

Survival Analyses of IDH-Wildtype Glioblastomas.
A Cox proportional hazard regression analysis was implemented to determine the relationship between the total number of SNVs (median split) and overall survival. MGMT methylated and unmethylated GBMs were investigated separately. Survival analyses and plotting of results as Kaplan-Meier graphs were carried out using R software [41]. Of the 41 patients, univariate survival analysis was carried out on the 33 IDH-wildtype patients only. Omitted patients included the three IDH MUT patients and a further five patients lacking survival information.

SNVs Detected in Initial and Recurrent IDH MUT
Glioblastomas. SNVs detected in IDH MUT initial (n � 12) and recurrent tumours (n � 1; Tables S4, and S5) impacted IDH1 and 10 genes across 5 biological pathways (Figures 1 and 2; Table 2). Majority of initial tumours had SNVs in genes in the RTK/Ras/PI(3)K (66%; 2/3), followed by p53 (100%; 3/3) and WNT signalling pathway (33%; 1/3). All initial IDH MUT tumours (100%; 3/3) harboured at least one potentially actionable variation in TP53 (100%; 3/3), BRCA2 (33%; 1/3), and MSH6 (33%; 1/3; Table 4). Just 7 SNVs in 6 genes were VUS that were possibly pathogenic in IDH MUT initial tumours. ese included IDH1 and the p53 pathway genes MSH6 and TP53 and the RTK/Ras/PI(3)K genes KLK1 and TSC2 and the CREBBP gene in the WNT pathway (Table 3). e KLK1 variant was potentially pathogenic in IDH MUT but not in IDH WT . e recurrent IDH MUT tumour had SNVs in p53, WNT signalling, and G-protein pathway genes. Matched analysis revealed that seven genes had SNVs in the initial that were not observed in the recurrent tumour ( Figure 2). e recurrent tumour had SNVs in one gene not recorded in the initial (GNAS). No genes had SNVs that were potentially actionable in the recurrent IDH MUT tumour (Table 4). WT and IDH MUT Glioblastomas. CNVs were detected in IDH WT tumours only (Table S6). e results for CNVs in the corresponding genes in TCGA-GBM are presented in Figure S1. For sample 36, there appears to be a hemizygous deletion in BRCA2 in the initial, but a CNV gain in the recurrent tumour. Both trends were identified in TCGA-GBM, but predominantly BRCA2 had shallow deletions. ere were CNV gains in GNAS for recurrent sample 3b. TCGA-GBM results also predominantly indicate CNV gains for GNAS. In recurrent samples 1b and 7b, TERT appeared to have hemizygous deletions. TCGA-GBM had both TERT CNV losses and gains with no predominant trend evident. For SMARCA4, there appears to be a CNV gain in initial sample 1 but a hemizygous deletion in the recurrent sample. TCGA-GBM had mostly CNV gains with some losses for SMARCA4. Significant CNV gains in EGFR were observed for initial and recurrent sample 2 and similarly in TCGA-GBM cases.

Investigation of the Corresponding Genes (with Mutations and CNVs in the GBM Cohort) in the TCGA-GBM and GDC Datasets.
e results of investigations in the TCGA-GBM and GDC datasets for the 21 genes identified with VUS that were possibly pathogenic in the GBM cohort are presented in Figure S2. A summary of SNVs identified from those corresponding genes in the TCGA-GBM dataset is provided in Table S7. TCGA-GBM cases in the mutation data included 6 verified and 2 ambiguous IDH-mutant individuals; however, majority of cases are unannotated. PTEN was the Journal of Oncology gene most impacted by mutations (34.86%) and shallow or deep deletions (Table S8; Figure S2). EGFR had mutations (26.97%) and CNV gains.  Table S9 and Figure S3. e WNT pathway genes DICER1 (2.29%), KLF4 (0.25%), and CREBBP (3.56%) had mutations and CNV shallow deletions, as well as low level gains and high level amplifications. TERT (2.80%) and KMT2D (3.05%) had mutations and CNV shallow gains and losses as well as deep deletions. APC (4.58%) and TCF4 (0.76%) had mutations, low level gains, and shallow deletions. e SHH genes, PTCH1 (3.56%), PTCH2 (1.78%), and SMO (1.02%) were impacted by mutations. Whilst the SMO gene had CNV gains, by comparison, the PTCH1 and PTCH2 genes had both CNV gains and losses. NOTCH genes, NOTCH2 (4.07%) and NOTCH1 (0.25%), had mutations and were impacted also by gains and losses in CNV.
Median survival was 13 months for unmethylated GBMs with ≤ 4 SNVs, compared to a median survival of 11 months for ≥ 5 SNVs (Figure 4; Table S10). Sample sizes were relatively small in these survival analyses; therefore, the observed trends would need to be confirmed using a larger cohort.

Discussion
e mutational landscape of the GBM subtypes in this cohort raises the possibility of new combinations of therapeutic approaches for individual GBM patients. Potentially actionable variation was most often identified in EGFR, PTEN, BRCA1/2, and ATM.

IDH WT Glioblastomas.
In IDH WT glioblastomas, SNVs impacted genes in the RTK/Ras/PI(3)K (79%), p53 (61%), Also included is a list of risk mutations related to heritable diseases. Genes identified with VUS that were possibly pathogenic in the GBM cohort are highlighted in bold. 8 Journal of Oncology Table  4: Summary of the proportion of initial and recurrent of IDH-wildtype and IDH-mutant glioblastoma patient tumours that had SNVs that could be assigned as potentially clinically actionable. WNT (58%), SHH (16%), NOTCH (8%), Rb (5%) and G-protein (5%) pathways. Potentially actionable mutations detected from initial IDH WT tumours included EGFR, PTEN, BRCA1, BRCA2, ATM, and ATR [54][55][56]. erapies for this subtype might include the EGFR-targeting antibodies, EGFR-targeting vaccines, TK inhibitors, erlotinib, and DNA damage repair inhibitors including olaparib and ATR inhibitors. Anti-EGFR-targeting antibodies to date have not shown clinical efficacy in GBM although trials are ongoing [57]. Similarly, trials of DNA damage repair inhibitors are underway, and the results are anticipated; however, patients have not been selected for these trials using molecular profiling with HTS.

Recurrent IDH WT Glioblastomas.
Interestingly in this cohort, no tumours exhibited a TMZ-induced hypermutated phenotype. Tumours did not have mutations in TERT promoter regions. Kim et al. found that a TMZ-induced hypermutated phenotype was rare in IDH-wildtype primary glioblastomas [76]. Acquired resistance in glioma has been attributed to dysregulated pathways (signalling and DNA repair), persistence of cancer stem cell subpopulations, and autophagy mechanisms [77]. In this cohort, only the RTK/ Ras/PI(3)K, p53 DNA damage repair, WNT signalling, and G-protein pathways were impacted by genetic alterations and not the SHH, NOTCH, and Rb pathways, despite their association with glioma resistance. Whilst fewer pathways were impacted, intertumour heterogeneity between initial and recurrent IDH wildtype tumours was nevertheless observed, similar to previous studies [76,78]. Indeed, recurrent tumours can diverge to such an extent that they are no longer recognised as lineal descendants of the dominant clone identified initial at diagnosis [78,79]. Potential signatures of IDH WT recurrent tumour resistance included VUS that were possibly pathogenic in PTEN. PTEN mutations cause activation of the PI3K/AKT survival pathway and chemoresistance in GBM [80]. Other possible signatures of recurrent tumour resistance in this GBM cohort included CNV gains in the genes (chromosome), BRCA2 (Chr13), GNAS (Guanine nucleotide-binding protein G(s) subunit alpha; Chr20), and EGFR (Chr7). Copy number gains are thought to impact driver genes to initiate tumourigenesis. e oncogene EGFR is located on chromosome 7, which frequently has CNV gains in IDH-wildtype glioblastomas  (∼70%) [5,6]. Gains in the chromosome 20 arm containing GNAS are frequently observed in pituitary brain tumours (adenomas) and may exert a mitogenic influence on the WNT signalling pathway via cAMP activation, which may provide a proliferative advantage for resistance [81]. However, GNAS has not been identified as a prognostic in dicator implicated in GBM [82]. CNV losses observed in the GBM cohort included SMARCA4 (Chr19) [47] and TERT (Chr5). CNV losses may be concordant with gene expression downregulation [83].

IDH MUT Glioblastomas.
Results for IDH MUT glioblastomas comprised three initial and one recurrent case only. Pathways impacted by genetic alterations included the RTK/ Ras/PI(3)K (66%), p53 (100%), and WNT pathways (33%). Possibly pathogenic VUS identified herein included those co-mutated in both subtypes as well as KLK1 (kallikrein1). e kallikreins KLK6, KLK7, and KLK9 have been shown to have higher protein levels in Grade IV glioma compared to Grade III tumours and consequently may have utility as prognostic markers for patient survival [84]. All initial IDH MUT tumour samples harboured potentially actionable variation in at least one of the genes TP53, BRCA2, and MSH6.
e recurrent tumour had fewer pathways (p53, WNT, and G-protein) impacted by genetic alterations. Matched analysis revealed intertumour heterogeneity. e recurrent IDH MUT tumour lacked potentially actionable variation that could be targeted. Given the small sample size for this subtype all trends reported here would need to be confirmed in a larger cohort.

Conclusion
Our study reveals that matched initial and recurrent GBM samples harbour potentially actionable variations, and these were most often identified in EGFR, PTEN, BRCA1/2, and ATM. ese genetic alterations could potentially be targeted by novel approaches with EGFR-targeting antibodies, tyrosine kinase inhibitors, and DNA damage repair inhibitors either singly or in combination. is study underlines the need for detailed genetic analysis of GBM patients to identify individuals that might benefit from novel therapeutic approaches that are becoming available in the near future. is information is also important for patient recruitment to clinical trials.

Data Availability
Data are available upon request from the Dept. of Neuropathology, Ruprecht-Karls University of Heidelberg.

Ethical Approval
Ethical approval was given by BRAIN UK and Brain Tumour Bank South West.  Figure S1: oncoprint plot of mutations and copy number alterations identified in the TCGA-GBM dataset for 8 corresponding genes impacted by CNVs in the GBM cohort. Genes are represented as rows, and individual patients are represented as columns.

Supplementary Materials
e right barplot displays the number and type of alterations to each gene, categorised as AMP: high level amplification, GAIN: low level gain, HETLOSS: shallow deletion, HOMDEL: deep deletion, and MUT: SNV mutation event (green). Figure S2: oncoprint plot of mutations and copy number alterations identified in the TCGA-GBM dataset for the 21 corresponding genes impacted by VUS that were possibly pathogenic in the GBM cohort. Genes are represented as rows, and individual patients are represented as columns. e right barplot displays the number and type of alterations to each gene, categorised as AMP: high level amplification, GAIN: low level gain, HETLOSS: shallow deletion, HOMDEL: deep deletion, and MUT: SNV mutation event (green). Figure S3: oncoprint plot of mutations and copy number alterations identified in the TCGA-GBM dataset for 12 WNT/Notch/SHH pathway genes impacted by SNVs in the GBM cohort. Genes are represented as rows, and individual patients are represented as columns. e right barplot displays the number and type of alterations to each gene, categorised as AMP: high level amplification, GAIN: low level gain, HETLOSS: shallow deletion, HOMDEL: deep deletion, and MUT: SNV mutation event (green). Table S1: demographic data for the IDHwildtype (n � 38) and IDH-mutant glioblastomas. Clinical records are for case ID, age, sex, tumour location on the MRI scan, IDH1 R132H hotspot mutation status, patient survival in months, and samples with matched initial and recurrent tumours. Table S2: list of the clinically relevant neuro-oncology genes that were analysed by the HTS-based diagnostic panel used in this study that was developed in Ruprecht Karl-University Heidelberg, Germany (see Sahm et al. [30]). Table S3: summary of the possibly pathogenic VUS identified in initial and recurrent IDH-wildtype and IDH-mutant glioblastoma tumours. e exonic nonsynonymous SNVs were predicted to be damaging by both LJB SIFT and FATHMM-MKL tools and had not been recorded by the 1000G database. Descriptive information for tumour, IDH status, genomic position, affected gene and pathway, available dbSNP and COSMIC identifiers, functional impacts predicted by LJB SIFT and FATHMM-MKL, and a shortened description from InterPro domain are provided. NA; not applicable (see Supplementary Tables  Excel File). Table S4: summary of SNVs identified in initial tumours. Descriptive information for tumour, IDH status, genomic position, reference, and alternative variant alleles, affected gene, and pathway, ClinVar significance, functional impacts as predicted by LJB SIFT and FATHMM-MKL and available dbSNP and COSMIC identifiers and InterPro domain description are provided (see Supplementary Tables  Excel File). Table S5: summary of SNVs identified in recurrent tumours. Descriptive information for tumour, IDH status, genomic position, reference and alternative variant alleles, affected gene and pathway, ClinVar significance, functional impacts as predicted by LJB SIFT and FATHMM-MKL and available dbSNP and COSMIC identifiers and InterPro domain description are provided (see Supplementary Tables Excel File). Table S6: summary of CNVs identified in initial and recurrent IDH-wildtype glioblastomas. CNV estimation is based on the read depth (%) of the variant (V) compared to a reference control (R; see Methods). Table S7: summary of the SNVs in TCGA-GBM dataset identified for the corresponding genes with VUS that were possibly pathogenic in the GB cohort. Descriptive information for tumour sample, gene, mutation type, amino acid change, genomic position, reference, and alternative variant alleles is provided (see Supplementary Tables Excel  File). Table S8: number of cases in TCGA-GBM and GDC mutation datasets affected by mutations in the genes identified to have VUS that are possibly pathogenic in the GB cohort. According to TCGA, a total of 393 cases were tested for somatic mutations. TCGA-GBM comprises a small number of verified (n � 6) and ambiguous IDH-mutant cases (n � 2; see ). Table S9: number of cases in TCGA-GBM and GDC datasets affected by mutations in the WNT, notch, and SHH genes identified to have somatic mutations in the GB cohort.