Malignant gliomas are a group of intracranial cancers associated with disproportionately high mortality and morbidity. Here, we report ultradeep targeted sequencing of a prospective cohort of 237 tumors from 234 patients consisting of both glioblastoma (GBM) and lower-grade glioma (LGG) using our customized gene panels. We identified 2,485 somatic mutations, including single-nucleotide substitutions and small indels, using a validated in-house protocol. Sixty-one percent of the mutations were contributed by 12 hypermutators. The hypermutators were enriched for recurrent tumors and had comparable outcome, and most were associated with temozolomide exposure. TP53 was the most frequently mutated gene in our cohort, followed by IDH1 and EGFR. We detected at least one EGFR mutation in 23% of LGGs, which was significantly higher than 6% seen in The Cancer Genome Atlas, a pattern that can be partially explained by the different patient composition and sequencing depth. IDH hotspot mutations were found with higher frequencies in LGG (83%) and secondary GBM (77%) than primary GBM (9%). Multivariate analyses controlling for age, histology, and tumor grade confirm the prognostic value of IDH mutation. We predicted 1p/19q status using the panel sequencing data and received only modest performance by benchmarking the prediction to FISH results of 50 tumors. Targeted therapy based on the sequencing data resulted in three responders out of 14 participants. In conclusion, our study suggests ultradeep targeted sequencing can recapitulate previous findings and can be a useful approach in the clinical setting.
This article is featured in Highlights of This Issue, p. 871
Malignant gliomas are a group of intracranial tumors associated with disproportionately high mortality and morbidity. The most common and aggressive subtype is grade IV glioblastoma (GBM), which has a dismal median survival of 15 months despite an intrusive standard of care comprising surgery, radiotherapy, and chemotherapy (1). Lower-grade gliomas (LGG, grade II to III) usually have a much better outcome, but they eventually progress to higher grade. Glioma represents one of the most challenging malignancies due to their anatomic location in the brain. Treatment options for glioma have not significantly changed in the past two decades.
Recent genomic profiling efforts such as The Cancer Genome Atlas (TCGA) have revealed the genetic and epigenetic landscape of glioma. Integrated DNA sequencing and copy-number analysis pinpointed potential driver genes including not only known players such as EGFR, TP53, IDH1, CDKN2A, etc., but also novel genes such as LZTR1 and FGFR-TACC fusions (2–5). Unsupervised clustering of gene expression and DNA methylation profiles led to the discoveries of three robust transcriptome-based GBM subtypes and a hypermethylation phenotype (G-CIMP) featuring recurrent mutations in IDH genes (6, 7). The TCGA pan-glioma analysis identified six methylation subtypes that demonstrate significant overlap with genetic alterations (4). Importantly, multiple groups independently reported a highly consistent genetic classification of LGGs using IDH mutations and 1p/19q codeletions (2, 8). In this classification, IDH mutations are the top-tier marker that separates LGGs into mutant and wild-type, followed by 1p/19q codeletion as the second marker that further divides mutant tumors. The new World Health Organization classification has incorporated this schema to mitigate variations in pathologic review (9). These advances collectively demonstrate that molecular profiling not only can provide insights into glioma biology but also represents an important tool toward personalized diagnosis and therapeutic interventions for glioma.
Panel sequencing has been tested in the clinical setting for selecting the most probable responders to certain cancer drugs, as the returned results inform on alterations of the targetable genes, mostly established cancer genes, in the tumor. These alterations can be subsequently matched to one or multiple therapeutic agents, as illustrated in several large-scale clinical sequencing efforts (10–12). Recently, the American Association for Cancer Research (AACR) coordinated with six major cancer centers to release its consortium panel sequencing project Genomics Evidence Neoplasia Information Exchange (GENIE; ref. 13). Similar efforts on childhood cancers such as pediatric glioma and neuroblastoma were also reported (14, 15). These studies laid the foundation for the approval of clinical use of high-throughput sequencing–based gene panels such as the Foundation Medicine panel and Memorial Sloan Kettering IMPACT panel by the FDA.
In this study, we describe our effort on sequencing a cohort of 234 prospectively collected gliomas using the T200 gene panel developed at MD Anderson Cancer Center. Our goal was to test the utility of panel sequencing in predicting glioma prognosis and matching patients to therapeutic agents based on their genetic profiles.
Materials and Methods
Sample and clinical data collection
Collection and analysis of patient samples were approved by the Institutional Review Board at MD Anderson Cancer Center in accordance with U.S. Common Rule. Written-informed consent was obtained from all participating donors. Clinical annotations, including follow-up, were collected by attending oncologists and were deposited into a central electronic patient portal. Tumor slides were reviewed by our pathologist team. An extra nonoverlapping tissue was collected for three tumors in order to assess intratumoral heterogeneity.
For each patient, we collected formalin-fixed paraffin-embedded tumor tissues and peripheral blood for sequencing. The inclusion of matched germline DNA allows us to identify somatic mutations with high confidence. Our sequencing protocol was described in detail elsewhere (16, 17). Briefly, DNA was extracted and sheared followed by library preparation using KAPA library preparation kit. After PCR primer removal, we used biotin-labeled probes for capturing target regions. The captured libraries were sequenced on an Illumina HiSeq 2000 machine following the manufacturer's recommendations. The resulting FASTQ file was demultiplexed using CASAVA 1.8.2 with no mismatches.
Mutation and copy-number calling
Fastq files were aligned to hg19 using BWA followed by duplicate reads removal using Picard. Somatic single-nucleotide variation and small indels were called using VarScan2 (18). We called copy-number alteration by comparing sequencing coverages in matched tumor and normal samples normalizing for the total sequencing depth. We defined a gene as H.AMP if its estimated copy number (ECN) was 5, and a gene as H.DEL if its ECN was 0.6. Mutation, copy number, and clinical data are publicly accessible on synapse (syn18410809).
TCGA glioma mutation data were downloaded from a published study (4). The χ2 test was used to compare mutation prevalence between TCGA and T200. Cox proportional hazard model was used to assess associations between clinical factors and overall survival (OS). Backwards elimination procedure was used to identify final model with only significant factors, starting from factors including histology, age, grade, and extent of resection. Survival analysis was carried out using SAS version 9.4 (SAS Institute). For mutation signature analysis, we counted the types of substitutions as a surrogate for mutation signature instead of applying a deconvolution algorithm because the number of mutations in nonhypermutators were too low, and for hypermutators, there were not sufficient mutations in every transition and transversion group except the dominant type such as C>T. Linear regression was performed using the R package lm. Naïve Bayesian prediction and cross-validation were done using the R packages naivebayes and caret.
Patient cohort and assay overview
We prospectively collected 237 tumor samples and their matched peripheral blood from 234 patients (Table 1). One ganglioglioma case was dropped for further analysis because no somatic mutations were detected. Six grade II or III cases lacked clear evidence of pathologic classification. The remainder of the cohort consisted of 22 oligodendrogliomas, 7 mixed gliomas, 35 astrocytomas, and 164 GBMs. In this study, we grouped the tumors into GBM (grade IV, 72%) and LGG (grades II–III, 28%) for comparison with TCGA. However, an important distinction between our cohort and the TCGA cohort is that 105 of the 234 patients in our cohort were pretreated by a systemic therapy, whereas all TCGA patients were treatment naïve, i.e., no treatment before surgical resection. Despite a modest sample size, this distinction allows us to examine the effect of systemic treatments on tumor genomic profiles. Overall, high-grade tumors (grades III and IV) tend to have more recurrent samples in our cohort collection (29%) compared with grade II tumors (12%), likely reflecting more aggressive diseases and referral patterns for high stage tumors. Of the 44 recurrent GBMs, 22 had a history of a previous lower-grade lesion and thus were secondary. These secondary GBMs had similar age with LGGs at the time of diagnosis (P = 0.93, Wilcox rank-sum test). Mutations in IDH1 and IDH2 (collectively referred to as IDH hereafter) were found with high prevalence in LGGs (83%) and secondary GBMs (77%) but much less commonly in primary (9%) and recurrent GBMs (9%). The similar frequencies of IDH mutation in primary and recurrent GBMs corroborate the contention that these mutations are acquired early in gliomagenesis (19, 20).
|Histology .||Grade .||Cohort size .||Median age (range) .||Median OS (days) .||Median follow-up (days) .||IDH mutation .|
|Histology .||Grade .||Cohort size .||Median age (range) .||Median OS (days) .||Median follow-up (days) .||IDH mutation .|
We sequenced a list of 202 genes (T200 panel), which were later expanded to 262 (T200.1 panel; Supplementary Table S1). T200 and T200.1 targeted 4,874 and 5,003 exons, covering cancer-associated genes with an emphasis on targetable genes with a total length of 0.9 and 2.0 million base pairs, respectively. The most frequently mutated genes in glioma including TP53, IDH1/2, ATRX, CDKN2A, EGFR, NF1, NOTCH1, PIK3CA, and PTEN were encompassed by both panels. Comparing with a list of 75 computationally determined driver genes from the TCGA pan-glioma study (4), no driver genes mutated in more than 5% of GBM or LGG were omitted in our panel except FUBP1, mutations of which were observed in 9% of LGGs. CIC, whose mutations were observed in 16% of LGGs, was included in T200.1 but not T200.
We sequenced 204 tumor/blood pairs with T200, 21 pairs with T200.1, and 12 pairs with both panels. For three tumors, we sequenced two nonoverlapping specimens with T200 to evaluate intratumoral heterogeneity. The median depths of coverage on targeted exons were 1,032 for tumor and 781 for normal with T200, 368 for tumor, and 469 for normal with T200.1. We evaluated the consistency of the two panels using the 12 replicates sequenced by both panels. Eighty-five percent of mutations found by T200 were also detected on T200.1. When germline variants were included, this percentage rose to 95%. T200.1 panel generally identified more mutations due to its larger targeted territories, but this difference was not statistically different (P = 0.4, Wilcox rank-sum test).
For the three patients for which we sequenced two distinct specimens, in two of them, both specimens yielded the same mutation calls. These two cases included an IDH mutant, 1p/19q noncodeletion grade II glioma, and an IDH wild-type primary GBM. The last case was also an IDH wild-type primary GBM. Interestingly for this case, we detected five mutations in the first specimen and four mutations in the second specimen, and BRAF V600E mutation was the only mutation shared between the two specimens (Supplementary Fig. S1). This pattern of limited genetic similarity was reminiscent of branched evolution reported by us and others elsewhere (20, 21). Both specimens harbored mutations in PTEN, but the mutations occurred at different loci in distinct forms (missense point mutation vs. a two-nucleotide deletion), suggesting a remarkable selective pressure-driven evolutionary convergence on this important tumor-suppressor gene.
Patterns of cancer gene mutations across glioma groups
We designed and deployed a rigorously tested protocol that automates reads alignment, mutation and indel calling, filtering, and annotation (17). Using this centralized pipeline, we detected a total of 2,485 somatic mutations, including single-nucleotide point mutations and small indels. The vast majority of the mutations cause changes to the encoded protein products, predominantly missense (n = 2,151, 87%). Only 35 mutations (1.4%) were synonymous, compared with 24% observed in TCGA (P < 0.05, t test), suggesting a strong selection for nonsynonymous mutations in the cancer genes sequenced on T200. Tumor suppressors, including PTEN, RB1, and ATRX, were enriched for deactivating mutations such as frameshift indels and gain of stop codons, illustrated by a panel of manually curated driver genes (Fig. 1). The median number of mutations per case was 4 (0–427, SD = 35), but a few cases were evident outliers indicative of a hypermutation phenotype. We examined the distribution of mutational loads and manually selected a cutoff 30 to identify 12 hypermutators (Supplementary Fig. S2). Combined, these 12 samples, which accounted for 5% of the entire cohort, contributed 61% of the total mutations (n = 1,506).
We next counted the prevalence of genes mutated in our cohort excluding hypermutators (Fig. 2A). TP53 was the most frequently mutated gene, followed by IDH, EGFR, and PTEN. PI3K pathway genes, including PIK3CA, PIK3CG, and PIK3R1, were collectively mutated in 37 patients (16%), suggesting that targeting this pathway could have important potential in treating glioma. Surprisingly, we found 10 patients harboring BRAF mutations, 9 of which were V600E. These patients could potentially benefit from receiving BRAF inhibitors (see below).
Because our samples were collected from various histopathologic subtypes, we compared prevalence of gene mutation across these groups. For this analysis, we initially focused on the 144 genes that were sequenced by both T200 and T200.1. This list allowed us to use all our cases except hypermutators, which were excluded in the comparison. We also excluded genes that were mutated in less than 5% of the cohort because it was challenging to confidently identify group preference with too few mutations, and furthermore, excluding them reduced the multiple testing problem by narrowing search space. Not surprisingly, ATRX, IDH1, RB1, PTEN, and TP53 were all significant, suggesting they were differentially mutated across the groups (FDR < 0.01, Fig. 2B). We verified their significance by repeating the analysis using cases sequenced by T200 only and found one more significant gene, filaggrin (FLG), underrepresented in lower-grade tumors (0%). In contrast, it was mutated in around 10% of GBMs, including secondary GBM (11.8%). In TCGA, this gene was mutated in 2% of LGGs but 8% of GBMs (Supplementary Table S1). FLG encodes an intermediate filament-associated protein that aggregates keratin intermediate filaments in mammalian epidermis, and it was identified as a significantly mutated gene in esophageal adenocarcinoma (22).
Higher frequency of EGFR mutations was previously reported in GBM than LGG (2). We found EGFR mutations evenly distributed across the groups. Unlike TCGA that identified EGFR mutations in 6% of LGGs, we observed 23% mutation in our cohort (FDR = 1.6e–4, χ2 test, Fig. 2B). The mutation frequencies were similar in primary GBM between our cohort and TCGA (FDR = 0.32, 26% vs. 19%). We speculate that the elevated prevalence of EGFR mutation compared with TCGA in lower-grade tumors might be due to the different composition of patients in the two studies, and the fact that our sequencing data have a higher depth of coverage. Consistent with TCGA, EGFR is more frequently amplified in GBM (34%) than LGG (4%) in our dataset.
To gain further insights into the elevated mutation frequency, we compared EGFR variant allele fraction (VAF), i.e., the fraction of reads carrying variant allele over the total number of reads, across glioma groups (Fig. 2C). VAFs in astrocytoma and oligodendroglioma, particularly oligodendroglioma, were significantly lower than that in GBM (P < 0.001, t test). We speculate that the lower VAFs suggest EGFR mutations in these groups were likely present in a small subset of cells. It was unclear if these mutations endowed the carrying cells with growth advantage over other clones.
IDH mutation and 1p/19q codeletion as biomarkers for glioma classification
IDH mutation and 1p/19q codeletion as biomarkers have been adopted in the WHO glioma classification system (9). We tested the utility of T200 panel sequencing as a tool to call these markers. Mutations in IDH1 and IDH2 predominantly affect amino acid 132 of IDH1 and the analogous amino acid 172 of IDH2 (23). We detected mutations in IDH1 or IDH2 in 88 tumors, including four hypermutators (Table 1). As expected, low-grade gliomas were predominantly IDH mutants (83%–100%), compared with grade III astrocytoma (70%) or primary and recurrent GBM (9%). All seven mixed gliomas harbored IDH mutations. The 84 nonhypermutators all had only one IDH mutation, affecting either R332 in IDH1 or R172 in IDH2. The only exception was a recurrent GBM case where a I98T mutation was found in IDH2. The four hypermutators harbored more than one mutations in IDH, but IDH1 R132H was found in all four with higher VAFs than other IDH mutations with borderline significance (P = 0.15, t test), suggesting the hotspot mutation was acquired before the hypermutator phenotype.
We examined the prognostic value of IDH mutations (Supplementary Table S2 and Supplementary Fig. S3). A univariate analysis using Cox proportional hazard model suggests IDH wild-type patients had significantly worse OS than mutants [P < 0.0001; HR, 8.8; 95% confidence interval (CI), 4.6–16.8]. A multivariate analysis controlling for tumor grade, pathology, and age confirmed the significant association between IDH mutation and better outcome (P = 0.003, HR = 0.25).
We sought to determine 1p/19q codeletion status out of the panel sequencing data, though profound noise in copy-number prediction has been noticed previously (24). We compared coverages of captured regions between tumor and normal samples for 1p and 19q, correcting for GC content and total sequencing coverage of the sample. We also calculated B allele fraction (BAF) for each capture region. In addition, we performed FISH on 50 cases (35 LGGs and 15 GBMs) to provide a benchmark dataset for this prediction.
As expected, copy-number signal from panel sequencing was noisy, with large variations observed in essentially all chromosome arms tested (Fig. 3A). BAF and copy number of 1p and 19 were positively correlated (r = 0.64, P < 0.001, Pearson correlation; Fig. 3B), raising the possibility that using 1p or 19q alone can be sufficient to determine 1p/19q status. We tested this approach by using an empirical threshold for 1p (CN < –0.7 or BAF > 0.7). Benchmarking against FISH data, we obtained a sensitivity of 0.64 and a specificity of 1.0, respectively. The low sensitivity was due to nine cases that demonstrated relatively high copy numbers despite their positive results by FISH (Fig. 3B and C).
We next applied Naïve Bayes classifier considering copy number and BAF of both 1p and 19q as features to predict their status. Leave-one-out cross-validation suggests an accuracy of 0.82 of the model. Nine cases were misclassified, eight of which overlapped with the misclassified cases from the empirical approach. All eight were predicted to be 1p/19q codeletion negative, although FISH data suggested otherwise. Among them, three were IDH wild-type, thus molecularly resembling high-grade astrocytoma and GBM. We speculate that FISH was able to capture subtle 1p/19q codeletion signals beyond the resolution of the panel platform. Overall, the modest performance of our data in predicting 1p/19q status was consistent with previous reports (24), raising caution and reiterating the importance of orthogonal validation.
Hypermutation phenotype is more frequent in recurrent lower-grade tumors
Despite the limited coverage offered by targeted sequencing, we observed 12 tumors with excessive mutational load (Supplementary Fig. S2), including five lower-grade tumors (7%), four secondary (18%), two primary (2%), and one recurrent GBMs (5%). All 12 cases received some form of systematic treatment, including radiation, chemo, angiogenesis inhibition, and in some cases, miscellaneous antineoplastic drugs. Seven of the 12 cases demonstrated the temozolomide signature, with C>T substitutions accounting for 95% of all point mutations (Fig. 4A; Supplementary Fig. S4). No evidence of altered polymerase POLE-associated hypermutation (featuring C>A transversion at CpT dinucleotide and C>T transition at CpG dinucleotide) was observed in any of the hypermutators. We note that deconvolution of the mutations into 96 trinucleotide context signatures, as described previously (25), was not successful due to the sparsity of mutations other than C>T. Notably, two secondary GBMs demonstrated almost entirely C>A mutations. C>A mutations are best known to be caused by tobacco consumption and are abundant in lung cancer (25, 26). Retrospectively examining their records, we found that 1 patient was a light smoker, whereas the other was a nonsmoker. More data are warranted to explain what caused the predominant C>A mutation pattern.
We examined mutational load as a function of molecular markers and histologic groups using multivariable linear regression. Interestingly, disease status, i.e., disease at initial diagnosis (primary disease) or recurrence, explained as much variance in mutational load as histology (6% for both), only slightly lower than ATRX mutations (7%). When separating tumors into primary and recurrent groups, we found a significant enrichment of hypermutators in the recurrent group (P = 1.2e–4, χ2 test). For recurrent lower-grade tumors, 21% were hypermutators, compared with none in the primary group. This percentage is similar to secondary GBM (18%) where all tumors by definition were recurrent. In contrast, only 4% of recurrent GBMs were hypermutators. We noticed that hypermutators in recurrent LGGs and secondary GBMs were typically years apart from their initial diagnosis when their tumors were sequenced, much longer than the single recurrent GBM case (Fig. 4B). We were able to collect information on duration of temozolomide treatment for 204 patients. The rest were either treated before being referred to MD Anderson or did not respond to follow-up. Hypermutators received significantly more cycles of temozolomide than nonhypermutators (Wilcoxon rank-sum test, P = 0.01), suggesting prolonged exposure to exogenous agents might catalyze the generation of hypermutation phenotype.
We next examined whether hypermutation was associated with outcome in glioma, especially given that a subset of these hypermutators seemed to be caused by chemotherapy (temozolomide) observed here and elsewhere (27). Ten of 12 hypermutators were alive at the latest follow-up, an outcome comparable with the average cases in the respective pathological group (P = 0.08, Cox proportional hazard model correcting for pathology and age; HR, 0.16; 95% CI, 0.02–1.2; Fig. 4C). Despite a small sample size, these data suggest increased mutational load did not significantly promote tumor aggressiveness.
Clinical utility for directing targeted therapy
Of the 202 T200 genes, 114 were targetable according to our proprietary Personalized Cancer Therapy knowledge base (http://pct.mdanderson.org). Informed of their tumor profiling results, 14 patients consented for receiving targeted therapeutic agents and eventually ten received them. These agents included those targeting PDGFRA, BRAF, PI3K, and IDH mutations. Of the 10 patients, 3 remained currently on treatment after 38 months (BRAF inhibitor), 21 months (IDH inhibitor), and 35 months (IDH inhibitor). These 3 patients included a recurrent GBM, a grade III oligoastrocytoma, and a grade III oligodendroglioma. Both lower-grade tumors were IDH mutants and had a history of disease progression.
The most exceptional responder out of our 10 patients is a 34-year-old female patient. The patient was diagnosed with a primary GBM, for which she underwent a subtotal resection followed by adjuvant concurrent radiation and low-dose temozolomide. After completion of 12 cycles of temozolomide, progressive disease was noted on imaging and was confirmed by biopsy. T200 sequencing revealed the presence of a BRAF V600E mutation in the recurrent tumor. This patient was subsequently treated with the BRAF inhibitor, Dabrafenib, in combination with the MEK inhibitor trametinib. The patient has remained without evidence of progressive disease at the latest follow-up, after the completion of 38 cycles of treatment.
In this study, we report target sequencing of 234 gliomas including low- and high-grade, primary, secondary, and recurrent tumors. This broad histopathologic spectrum reflects our intent to recapitulate patient cohort seen in the clinic. Developing upon our institutional efforts, we aimed at addressing two questions: (1) to assess panel sequencing as a prognostic and diagnostic tool in glioma, and (2) to establish a pipeline that could allow us to select glioma patients for clinical trials. Though the genes on the panel were not elected specifically for glioma, our platform included almost all known targetable glioma driver genes. Importantly, some genes that may not be elected for glioma such as BRAF were on the pan-cancer panel and proved informative for patient management and outcomes.
Our analysis suggests T200 can recapitulate patterns reported by TCGA and others, most notably histological group associated mutations in glioma genes. These associations can be used in combination with pathologic review to help clarify disease diagnosis. Using prospectively collected tumors, we show that IDH mutation is a prognostic marker after adjusting for age, grade, and histology. Predicting 1p/19q status using T200 data yielded only modest performance, especially low sensitivity, reflecting the technical challenge of calling copy-number signals from targeted sequencing. However, the specificity was high, suggesting that when T200 indicates that 1p/19q is lost, it is likely valid. Discrepancies with FISH result, which is commonly regarded as golden standard, indicate that T200 is not yet ready to replace this cytogenetic technique but that it could be used as a preselection approach with only 1p/19q normal on T200 requiring FISH which is a more challenging assay.
The hypermutation phenotype recently received much attention (28, 29), as higher mutational load partially predicts positive response to immune checkpoint blockade (30, 31). We identified 12 hypermutators from our cohort, 7 demonstrated temozolomide-related mutational signature, suggesting cytotoxic agents can serve as potent mutagens. Accordingly, we observed an enrichment of these hypermutators in recurrent tumors. Interestingly, the hypermutators had comparable outcome than nonhypermutators, in fact, slightly better (P = 0.08, Cox proportional hazard model, correcting for pathology and age; HR, 0.16; 95% CI, 0.02–1.2), suggesting the additional mutations acquired did not confer significant growth incentive or evolutionary advantage to the tumor. Recent studies suggest excessive somatic mutations may facilitate the generation of more neoantigens and thus elicit stronger immune cell infiltration to the tumor niche, resulting in a more hostile growth environment (32, 33). Nevertheless, these hypermutators could be excellent candidates for immune therapy trials should other clinical options be exhausted.
In gliomas, molecular profiling is not yet mainstream practice as the clinical relevance is still to be determined. In current clinical practice, most treatment decisions for gliomas are largely made based on the patient's age and performance status. However, the recent inclusion of IDH mutation and 1p/19q loss into classifiers for gliomas suggests that these do report on patient outcomes. To a large extent despite our growing understanding of the extensive heterogeneity of gliomas, a one-size-fits-all approach is employed. Presently, there are no targeted agents that have led to improved clinical outcomes in gliomas. Overall, single-agent targeted therapy to date has been extremely disappointing due to the inherent heterogeneity of the disease and due to multiple driver mutations in different cell populations within a tumor. Many agents have been developed, but overall response rates continue to be poor with some exceptions such as our 3 patients who have demonstrated durable responses on a biomarker-driven clinical trial. These patients demonstrating longer than expected responses underscore our need to better understand the importance of patient selection and enrichment for clinical trials in order to identify the subset of patients most likely to benefit from treatment. For those patients who do not respond as well to biomarker-driven therapy tailored to the profiling of their tumor, efforts need to be employed to better understand mechanisms of resistance to avoid the pitfall of trials failing time and time again in gliomas.
Moving forward, rational combination therapy may be superior to monotherapy, and consideration will need to be given to combinatorial strategies utilizing immunotherapy, radiotherapy, targeted therapy, antiangiogenic therapy, and/or chemotherapy. Combinatorial strategies will need to be balanced with the potential risk and likelihood of increased toxicity often referred to as “toxic synergy.” Future therapeutic strategies in clinical trials should ideally be informed by molecular profiling results and employ combination therapies with targeted drugs to attack subsets of tumor cells harboring different driver mutations while using treatments with less specificity to target cells with passenger and/or secondary mutations.
Disclosure of Potential Conflicts of Interest
F. Meric-Bernstam serves in a consulting/advisory role for Genentech, Inflection Biosciences, Pieris Pharmaceutical, Xencor, Samsung Bioepis, Sumitomo Group, Spectrum Pharmaceuticals, OrigiMed, Aduro Biotech, and Mersana; and has research funding support from Novartis, AstraZeneca, Taiho Pharmaceutical, Genentech, Calithera Biosciences, Debiopharm Group, Bayer, Aileron Therapeutics, PUMA Biotechnology, Cytomx Therapeutics, Jounce Therapeutics, Zymeworks, Curis, Pfizer, eFFECTOR Therapeutics, Abbvie, Daiichi Sankyo, and Guardant Health. R.G. Verhaak owns stock in Pretzel Therapeutics. G.B. Mills is a consultant for AstraZeneca, Catena Pharmaceuticals, Critical Outcome Technologies, ImmunoMET, Ionis, MedImmune, Nuevolution, Pfizer, Precision Medicine, Signalchem Lifesciences, Symphogen, Takeda/Millenium Pharmaceuticals, and Tarveda; owns stock in Catena Pharmaceuticals, ImmunoMET, SignalChem, Spindletop Ventures, and Tarveda; has licensed technology of an HRD assay to Myriad Genetics; and has research sponsored by Abbvie, Adelson Medical Research Foundation, AstraZeneca, Breast Cancer Research Foundation, Critical Outcome Technologies, Horizon Diagnostics, Illumina, ImmunoMET, Ionis, Karus Therapuetics, Komen Research Foundation, Nanonstring, Ovarian Cancer Research Foundation, Pfizer, Prospect Creek Foundation, Takeda/Millenium Pharmaceuticals, and Tesaro. W.K.A. Yung is a paid consultant for DNATrix and NBTS and receives honoraria from DNATrix and NBTS. J.F. de Groot has research support from Sanofi-Aventis, AstraZeneca, EMD-Serono, Eli Lilly, Novartis, Deciphera Pharmacetuicals, and Mundipharma; is a paid consultant for Celldex, Deceiphera Pharmaceuticals, Abbvie, FivePrime Therapuetics Inc., GW Pharma, CArthera, Eli Lilly, Boston Biomedical Inc., Kairos Venture Investments, Syneos Health, and Monteris; serves on the advisory board for Genentech, Celldex, Foundation Medicine, Inc., Novogen, Deciphera, AstraZeneca, Insys Therapeutics, Kadmon, Merck, and Eli Lilly; and owns stock in Ziopharm Oncology and Gilead. J.F. de Groot's spouse is employed by Ziopharm Oncology. No potential conflicts of interest were disclosed by the other authors.
Conception and design: G.B. Mills, W.K.A. Yung, J.F. de Groot
Development of methodology: S. Zheng, K. Chen
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K. Alfaro-Munoz, A.K. Eterovic, K.R.M. Shaw, F. Meric-Bernstam, G.N. Fuller, W.K.A. Yung, S.-P. Weathers, J.F. de Groot
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S. Zheng, W. Wei, X. Wang, F. Wang, A.K. Eterovic, K. Chen, R.G. Verhaak, G.B. Mills, W.K.A. Yung, S.-P. Weathers, J.F. de Groot
Writing, review, and/or revision of the manuscript: S. Zheng, K. Alfaro-Munoz, W. Wei, F. Meric-Bernstam, G.N. Fuller, G.B. Mills, S.-P. Weathers, J.F. de Groot
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K. Alfaro-Munoz, K.R.M. Shaw, K. Chen
Study supervision: S. Zheng, K. Chen, J.F. de Groot
We thank all patients who donated their samples to this project. This work is funded by Sheikh Khalifa Bin Zayed al Nahyan Institute for Personalized Cancer Therapy (IPCT) and Khalifa Bin Zayed Al Nahyan Foundation Grant (F. Meric-Bernstam, K.R.M. Shaw), MD Anderson Glioblastoma Moonshots Program (J.F. de Groot), Marnie Rose Foundation (J.F. de Groot), and Naz Moez Sarrafzadeh Glioma Research Account (W.K.A. Yung). S. Zheng is supported by CPRIT (RR170055) and a Career Enhancement Award of the Brain SPORE at MD Anderson (P50 CA127001).