Genomic copy number variation correlates with survival outcomes in WHO grade IV glioma

Allele-specific copy number analysis of tumors (ASCAT) assesses copy number variations (CNV) while accounting for aberrant cell fraction and tumor ploidy. We evaluated if ASCAT-assessed CNV are associated with survival outcomes in 56 patients with WHO grade IV gliomas. Tumor data analyzed by Affymetrix OncoScan FFPE Assay yielded the log ratio (R) and B-allele frequency (BAF). Input into ASCAT quantified CNV using the segmentation function to measure copy number inflection points throughout the genome. Quantified CNV was reported as log R and BAF segment counts. Results were confirmed on The Cancer Genome Atlas (TCGA) glioblastoma dataset. 25 (44.6%) patients had MGMT hyper-methylated tumors, 6 (10.7%) were IDH1 mutated. Median follow-up was 36.4 months. Higher log R segment counts were associate with longer progression-free survival (PFS) [hazard ratio (HR) 0.32, p < 0.001], and overall survival (OS) [HR 0.45, p = 0.01], and was an independent predictor of PFS and OS on multivariable analysis. Higher BAF segment counts were linked to longer PFS (HR 0.49, p = 0.022) and OS (HR 0.49, p = 0.052). In the TCGA confirmation cohort, longer 12-month OS was seen in patients with higher BAF segment counts (62.3% vs. 51.9%, p = 0.0129) and higher log R (63.6% vs. 55.2%, p = 0.0696). Genomic CNV may be a novel prognostic biomarker for WHO grade IV glioma patient outcomes.

gains and losses when tumors frequently deviate from a diploid state and populations of non-tumor cells are present within the tumor sample, however, has remained a persistent issue 13,14 . To overcome these challenges, a novel allele-specific copy number analysis of tumors (ASCAT) was developed by Loo et al. which can obtain accurate genome-wide ASCAT profiles while accounting for aberrant tumor cell fraction (non-tumor components), ploidy, gains, losses, loss of heterozygosity and copy number-neutral events 15 .
In glioblastoma, both TCGA and institutional data have been interrogated for CNV to evaluate different genes involved in oncogenesis and tumor behavior [16][17][18] . To our knowledge, genome-wide copy number analysis has not been correlated with clinical outcomes. Here, using data from the OncoScan SNP array, we applied the ASCAT algorithm to assess the impact of differences in ploidy, aberrant cell fraction (ACF), and CNV on progression-free survival (PFS) and overall survival (OS). We also generated a metric to assess the extent of copy number variation seen in the genome, a measure of overall genomic disorder, and correlated this with survival outcomes. 44 patients received radiotherapy of at least 60 Gy, typically using volumetric arc therapy (VMAT) per institutional standard; other patients received treatment to less than 60 Gy due to either advanced age, poor performance status, physiologic decline, or received radiotherapy at an outside facility. 44 (78.5%) patients received  Table 1.

Discussion
Appropriate patient selection for aggressive treatment strategies, and corresponding patient counseling are of paramount importance. Highly toxic treatments that confer little, if any, clinical benefit, but come at significant quality of life costs, should be avoided. A number of important molecular biomarkers that are prognostic and/ or predictive have already been identified in glioma patients including 1p19q co-deletion, IDH1 mutation, and MGMT methylation. Genome-wide level analyses can be complementary to these biomarkers, and may provide additional information on outcomes.
In this study, we demonstrated that higher whole-genome CNV correlates with longer survival outcomes on both our institutional and validation cohorts. Of note, due to differences in total probe number and location within the genome, the two input SNP arrays, Oncoscan (institutional data set) and Affymetrix SNP 6 (TCGA data set), could not be directly compared using the same ASCAT cut-points for the different parameters. Therefore, for future assay utilization, it is likely that the cut-point will be identified for the individual SNP array utilized. However, once these assay specific cut-points are established, we believe the clinical utility is clear. Additionally, despite these large differences between the two assays, they both demonstrated longer survival for   www.nature.com/scientificreports www.nature.com/scientificreports/ increased CNV. This suggests a deeper biological significance, clear value in utilizing this validation cohort, and it warrants further investigation.
In contrast to our study, CNV has also been evaluated in a number of other malignancies and a higher CNV burden is most commonly associated with worse outcomes 19,20 . CNV has also been shown to be important in increasing cancer risk and accelerating disease progression [21][22][23] . In this context, our results are intriguing. Interestingly, in the context of glioblastoma, a reduction in copy number segments of specific genes confer a shorter survival 24 . Lehrer et al. showed a longer median survival for patients with increased copy number segments of SGK1, a gene required for growth and survival of glioblastoma stem-cells 24 . Others have shown the converse, that total CNV was shown to be a prognostic factor for worse outcomes in adult astrocytoma, especially in the IDH-mut group 25 . In the context of our study, this subgroup represented a very small fraction of the patient samples we evaluated and may therefore not be directly applicable. Others have also showna strong correlation between CNV and shorter disease-free and overall survival 26 . It remains interesting that our study stands in contrast to the work of a few other groups. However, the clinical management of patients in the other studies was often not well described. It is possible that higher CNV reflects a poor prognostic factor in the absence of optimal therapy, but similar to Her2neu + breast cancer, it may serve as a predictive biomarker for benefit with specific therapy, in this case chemoradiation. This questions, unfortunately, could not be addressed in the current study.
The data concerning radiation sensitivity and CNV is mixed. Yard et al. showed that in irradiated colorectal, uterine and ovarian carcinoma human cell lines, there was a positive correlation with increased somatic copy number alterations and survival 27 . They argued that the same mechanism that generated the copy number alterations initially also increased a tumor cell's capability to repair double-stranded breaks following radiotherapy. Other studies have shown that the sensitivity to radiation depends on gene dosage and, specifically, which gene copy numbers are altered. A reduction in Rad51 expression, for example, in glioma cells significantly increases radiosensitivity 28 . In this case, our results are hypothesis generating. It may be that the broad and extensive coverage of the genome with copy number variation tips the balance towards radiosensitivity, but further study is warranted.
The increased aberrant cell fraction (ACF) also correlated with both longer PFS (p = 0.001) and OS (p < 0.001) in our institutional dataset. The non-aberrant cells reflect the tumor microenvironment, and these results suggest that non-aberrant cells are either tumor promoting or are involved in treatment resistance. The non-tumor glioma microenvironment is a complex interplay of stem cells, astrocytes, and immune cells 29,30 . It has been shown that the accumulation of microglia/brain macrophages around the tumor correlates with poor clinical prognosis 31 . In addition, tumor infiltrating lymphocytes (TILs) have also demonstrated a correlation with survival. Han et al. showed using immunohistochemistry for CD4+ FoxP3+ and CD8+ TILs, there was an inverse correlation between tumor grade and the number of CD8+ TILs. Additionally, they showed that FoxP3+ T-cells (regulatory T-cells) were only observed in glioblastomas, and not in lower grade gliomas 32 . In glioblastoma patients, a high CD4 to CD8 ratio was a predictor of worse survival outcomes 32 . In contrast, another study demonstrated that higher levels of CD8+ TILs had shorter survival, however, this was confounded by the association between low CD8+ TILs and methylation of MGMT 33 . These data combined with our observations suggest that the worse outcomes with lower ACF may be a function of increased myeloid or lymphoid immunosuppressive infiltrates. These results also have implications for the on-going checkpoint blockade clinical trials in high-grade gliomas with interim results from CheckMate-143 (nivolumab versus bevacizumab for recurrent GBM) failing to show an OS benefitfor nivolumab 34 . It is possible that more myeloid-derived suppressor cells or FoxP3+ CD4+ T-cells are responsible for the worse outcomes observed in patients with lower ACF, and may impact responses to immunotherapy [35][36][37] . Notably, this idea is supported by a retrospective analysis of the nivolumab study which showed decreases in MDSCs in peripheral blood correlated with treatment response 37 .
Our study's relatively small size and retrospective nature are inherent limitations. Additionally, due to the small sample size, we were not able to incorporate all potentially relevant variables into our multivariable analysis including radiation dose, timing of temozolomide administration, and use of other therapies such as NovoTTF. Use of a WHO grade IV glioma cohort represents some inherent heterogeneity. The molecular definition of WHO grade IV gliomas makes distinctions based on IDH1 mutation and 1p19q co-deletion status. Here, the proportion of patients with these 2 mutations account for a relatively small minority. The inclusion of IDH1 mutation and 1q19q co-deletion status in the multi-variable models take into account their contribution towards heterogeneity. Notably, 1p19q co-deletion status and IDH1 status were not significantly linked to PFS in the multi-variable model when ASCAT variables were present; and IDH1 mutation was similarly non-significant in the OS multi-variable model (which has substantial overlap with 1q19q co-deletion as a sub-population). Despite these limitations, the results remain thought provoking and warrant further validation and investigation.  www.nature.com/scientificreports www.nature.com/scientificreports/ Methods patients. Records of patients treated at a single academic institution between 1/2009 and 8/2015 were evaluated and reviewed; this was performed under a protocol approved by the Winship Cancer Institute institutional review board. Data were de-identified according to the Health Insurance Portability and Accountability Act, and all methods were performed in accordance with the relevant guidelines and regulations. Informed consent was obtained for tissue sample banking; informed consent for this study was waived by the Institutional Review Board that approved this study's protocols. Inclusion criteria included a pathologic diagnosis of a primary WHO grade IV glioma, no prior brain radiotherapy and an OncoScan analysis for ASCAT inputs. treatment. Patients with newly-diagnosed WHO grade IV gliomas were discussed at a multi-disciplinary tumor board, including surgical oncology, neurosurgery, pathology, radiology, and medical and radiation oncology services. Decisions for therapeutic management were done jointly. Recorded baseline characteristics included age, gender, tumor location, performance status, MGMT methylation, 1p19q co-deletion and IDH1/2 mutation status. Radiation characteristics were also recorded, including radiation total dose, dose per fraction, and receipt of concurrent and adjuvant temozolomide.

Snp array.
For the primary (institutional) dataset, an Affymetrix OncoScan FFPE single nucleotide polymorphism (SNP) array was used for raw genomic data acquisition. Two data tracks were produced from the SNP array: total signal intensity and allelic contrast. Log ratio (R) represents the total signal intensity, which reflects the total copy number on a logarithmic scale. The B-allele frequency (BAF) represents the allelic contrast and demonstrates the relative presence at each SNP locus evaluated of the two alternative nucleotides. For The Cancer Genome Atlas (TCGA) dataset, Affymetrix Genome-Wide Human SNP Array 6.0 was utilized for the raw genomic data.
AScAt analysis. Following acquisition of the raw genomic data from biopsy or resected tumor samples, for each microarray case in Affymetrix Chromosome Analysis Suite v3.0, the Export Probe Level Data function was used to create a text file with the microarray probe set name, log 2 ratio signal, and BAF signal suitable for processing with the ASCAT v2.4.4 library in R v3.4.1. TCGA data set, for which the raw data was previously acquired with the Affymetrix Genome-Wide Human SNP Array 6.0, was processed using an R-package "Rawcopy" to create the probe level BAF and log R data.
These data were used as inputs of the ASCAT algorithm. As a function of the SNP data, ASCAT models the allele-specific copy number and arrives at the solution that is closest to integer copies at all assessed loci. To do this accurately, the ASCAT algorithm accounts for important factors including polyploidy and aberrant cell fraction when calculating the CNV profile 8 . Aneuploidy, a deviation from the normal chromosomal number, can confound copy number analysis. Non-aberrant cell admixture or aberrant cell fraction reflects the non-tumor component of the sample and can differ significantly between samples again confounding copy number studies. As an intermediate output of the ASCAT algorithm, we quantified the number of segments with changes in the log R and BAF. A segment is defined as a region of the genome with a fixed copy number. For a representative log R data sample, there were an average of 4,866,610 base pairs and 332 probes per segment with a range of 109,439-8,276,5227 base pairs and 7-5,997 probes. The total number of segments reflects the number of CNV inflection points or changes in copy number in the genome. This was defined as the segment counts. Statistical analysis. PFS was defined as time in months from diagnosis to progression, death or last follow-up, where those alive without progression were censored at last follow-up. Progression was defined radiographically based on the T1 post-contrast MRI, as assessed by post-treatment neuroradiology reports and supplemented with information from tumor board evaluation, specialized MR sequences, and tissue confirmation where available. OS was defined as time in months from diagnosis to death or last follow-up, where those who survived were censored at last follow-up. Genomic characteristics considered for analysis were ASCAT ploidy, ASCAT aberrant cell fraction, ASCAT log ratio segment counts, and ASCAT BAF segment counts. These variables were analyzed as continuous variables, and as categorical variables dichotomized at their median values. Cut point analyses were also performed to identify statistically significant cut points for each genomic variable and each endpoint using an outcome-oriented approach 38 . All survival data with stratification were reported using the optimal cut point. The log-rank statistic was maximized, and the significance of each cut point was assessed. Statistically significant cut points were considered for further analysis.
Descriptive statistics were reported for each variable. PFS and OS curves were estimated using the Kaplan-Meier method, and compared using log-rank tests. Median follow-up was estimated using the reverse Kaplan-Meier approach, where the censoring indicator is switched such that deaths are censored. Univariate Cox proportional hazards models were also fit, modeling PFS and OS as a function of age, gender, Karnofsky performance status (KPS), surgery type, the four genomic characteristics, IDH1 mutation (positive vs. negative), and MGMT (methylated vs. unmethylated). Multivariable Cox proportional hazards models were fit for OS and PFS as function of the statistically significant ASCAT variable, IDH1, MGMT, KPS and 1p19q co-deletion. Multivariable Cox models also were fit for OS as a function of the statistically significant ASCAT variables, IDH1, MGMT, KPS, and surgery type. Model assumptions were checked and verified. Statistical analyses were performed using SAS 9.4 (SAS Institute Inc., Cary, NC), and statistical significance was assessed at the 0.05 level.