Introduction

Neuroblastoma is a cancer typically affecting young children arising from the developing sympathetic nervous system, but can occasionally occur in adolescents and adults [1]. Over half of patients have widely metastatic disease at diagnosis where survival rates are less than 50% despite intensive therapeutic regimens including chemotherapy, radiation therapy and immunotherapy [2]. There is no standard therapeutic approach for relapsed disease [3]. Recent next generation sequencing (NGS) efforts of matched neuroblastoma samples collected at diagnosis and constitutional DNAs in 373 unique subjects across four studies has clearly documented a relatively low somatic mutation rate in the protein coding portion of the genome [4–7], challenging the concept of targeting oncogenic drivers with newly developed therapeutics. The data in neuroblastoma appears to be reflective of pediatric cancers in general [8]. However, recent studies of diagnostic-relapse-normal DNA “trios” from a limited number of neuroblastoma cases has shown that the mutation rate is much higher after exposure to genotoxic chemoradiotherapy, and that there may be an enrichment of previously subclonal mutations in pathways known to be therapeutically targetable in other diseases [9–11].

To better understand and characterize the landscape of potentially actionable mutations at both diagnosis and relapse, we analyzed targeted next-generation sequencing data for 151 neuroblastomas from 138 patients that were profiled either at diagnosis, in the midst of therapy, and/or at disease relapse. Our primary aim was to retrospectively determine the frequency by which a therapeutically relevant lesion was discovered at these time points and to infer if the biopsy procedure followed by DNA sequencing could provide the potential for patient benefit.

Results

We collected sequencing data and clinical information of neuroblastoma patients whose tumor biopsies had been sent to Foundation Medicine for molecular profiling (see Methods for sample processing details). The only inclusion criterion was the availability of high-quality sequencing data, and we did not filter the cohort further based on disease stage, risk group, age, or presence of molecular lesions. We analyzed data from 151 samples from 138 individuals at various time points during treatment (44 at diagnosis, 42 after the start of treatment, 59 at relapse, and 6 at unknown time points) and with varying risk status (Fig 1A, S1 Table). Samples labeled as “diagnosis” were biopsied before the start of treatment, “relapse” samples were taken at the time of disease relapse, and “post-treatment” samples comprised refractory disease, samples collected at definitive surgery, and any other non-relapsed tumors that had been exposed to treatment (generally 3–4 rounds of chemotherapy). Nine patients had serial biopsies sent for profiling (Fig 1B, S2 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

Study cohort overview A) Tabulation of Children’s Oncology Group (COG) risk classification and treatment time points of biopsy for 151 samples. (Intermed. = intermediate risk group) B) Number of samples taken at each treatment time point for nine patients with serial biopsies. (HR = high risk, IR = intermediate risk, LR = low risk at time of biopsy; further information in S2 Table) C) Tabulation of all variants identified (VUS: variants of unknown significance) D) Total number of variants identified per sample, stratified by COG risk group. Inset shows a similar calculation for suspected driver variants only. Heavy line represents the median of the data. “n” indicates the number of patients in each risk group. E) Total number of variants in each sample. Each bar represents an individual sample; color corresponds to risk group (red = high, blue = intermediate, green = low).

https://doi.org/10.1371/journal.pgen.1006501.g001

Across the cohort, we identified 1204 unique variants involving 352 unique genes (Fig 1C). We define “suspected driver” variants as any short variants (single amino acid substitution or short insertion/deletion) that appear in the Catalog of Somatic Mutations in Cancer (COSMIC), any copy number alterations (CNAs) that have been shown in the literature to be known oncogenic drivers, or any type of variant that disrupts a tumor suppressor gene, falls within a known hotspot, or is a fusion involving oncogenic driver kinases. Any type of variant uncharacterized in the literature is reported as a variant of unknown significance (VUS). It is important to note that every sample in this study was taken from a tumor biopsy without a paired germline sample from the same individual. Therefore, excluding the patients for whom we have data from multiple biopsies–and in whom mutations were only identified in a subset of biopsies–we cannot rule out the possibility that any reported mutation may be from the germline genome.

We detected an average of 9.8 total variants per sample, with a range of 2–20. The number of suspected driver variants was lower, with an average of 1.28 variants per patient, and a range of 0–6. High risk tumors had, on average, a higher total number of variants (10.1 per sample for high risk vs. 6.13 per sample for low risk, P = 5.49x10^-6) and a higher number of suspected driver variants (1.33 per sample for high risk, 0.625 per sample for low risk, P = 0.0058) (Fig 1D and 1E). There was no statistically significant difference in the number of mutations between high and intermediate risk tumors.

Among the entire cohort of 151 unique neuroblastoma cases, we detected suspected driver single nucleotide variants or short insertions/deletions in 60 genes (Fig 2A); amplifications in eight genes (Fig 2B), homozygous deletions in nine genes (Fig 2C), and genomic rearrangements or fusions in eight genes (Fig 2D). ALK was the most commonly mutated gene, with variants occurring in 20% of patients overall (suspected driver variants in ALK occurred in 16% of patients). Suspected driver ALK variants were present in 3/43 (7.0%) of samples at diagnosis, 7/41 (17%) post-treatment samples, and in 11/54 (20%) of samples at relapse. We observed a similar trend for suspected driver short variants in other genes (41/60 genes were mutated in a higher fraction of patients after treatment than at diagnosis; P = 0.0062; Fig 2A) and for suspected driver high-level copy number amplifications (8/8 genes were amplified in a higher fraction of patients after treatment than at diagnosis; P = 0.0078; Fig 2B). However, homozygous deletions and genomic rearrangements were present at a similar frequency before and after treatment (Fig 2C and 2D).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

Suspected driver variants in diagnostic, treated, and relapsed neuroblastomas A) Frequency of all suspected driver short variants (single nucleotide variants, insertions, deletions) per gene in the entire cohort. Bars represent the number of patients with at least one lesion in a given gene, normalized to the total number of patients at each time point. Patients with multiple lesions in the same gene were only counted once per gene. B) All suspected driver amplification events with copy number >10. Bars represent the number of patients at each time point with a given gene amplification, normalized to the total number of patients at that treatment time point. C) All suspected driver homozygous deletion events. Bars represent the number of patients at each time point with a given gene loss, normalized to the total number of patients at that treatment time point. D) All suspected driver gene rearrangements and fusions. Bars represent the number of patients at each time point with a given genomic rearrangement, normalized to the total number of patients at that treatment time point. E) Number of variants in each sample, stratified by type of variant (“All variants” includes suspected driver variants as well as VUSs; “Suspected” includes suspected driver variants only; “Actionable” are suspected driver variants that have FDA approved or investigational therapy matches) and disease time point (blue, diagnosis; yellow, post-treatment; red, relapse). P values calculated using Welch’s T-test. F) Fraction of samples containing at least one variant of each type (variant types as described in (E)), stratified by disease time point. P values calculated using Fisher’s exact test.

https://doi.org/10.1371/journal.pgen.1006501.g002

The observation that a majority of genes were mutated or amplified in tumors that had undergone treatment led us to ask the inverse question; specifically, whether tumors that had undergone treatment were more likely to have a higher number of suspected driver variants. Although the total number of variants per sample was similar at all treatment time points (P = 0.25 between diagnosis and relapse), we found that the average number of suspected driver variants increased from diagnosis to relapse (1.1 variants per sample at diagnosis vs. 1.7 variants at relapse; P = 0.038; Fig 2E). The likelihood of a tumor to harbor at least one suspected driver variant also increased from diagnosis to relapse, although to a less significant degree (66% of tumors showed at least one variant at diagnosis, while 80% did at relapse; P = 0.17; Fig 2F).

With a relatively higher frequency of genomic alterations at relapse, it seems likely that there may be more targeted therapeutic options available for patients who have previously undergone conventional treatments. We therefore identified 40 of the 80 genes containing suspected driver variants as “potentially actionable” by annotating those that have an existing FDA approved or investigational therapy matches (see Methods and S4 Table). On average, samples taken at relapse had a higher number of potentially actionable variants (0.57 at diagnosis vs. 0.95 at relapse; P = 0.048; Fig 2E), and each relapse sample was more likely to have at least one actionable variant (41% of tumors had at least one actionable variant at diagnosis vs. 58% at relapse; P = 0.11; Fig 2F).

We next examined trends in the data that might be of clinical relevance. First, we looked at differences in the mutational profiles between MYCN amplified and non-amplified tumors. ALK mutations were found in both MYCN amplified and non-amplified cases (Fig 3A), but several other variants were observed only in one context or the other (i.e. ATRX mutations only in MYCN non-amplified cases, as reported previously [4,7], and STAG2 mutations only in MYCN amplified cases). The majority of these variants only appear in a single patient in the cohort, so these trends towards mutual exclusivity will need to be confirmed in large case series. We also examined the frequency of alterations in the MAPK pathway (gene set in S6 Table) at diagnosis and relapse. Fifteen unique patients harbored suspected driver alterations in the MAPK pathway at relapse, while only five tumors had a similar aberration at diagnosis (P = 0.076, Fig 3D). These data support our prior observation that mutations predicted to activate the MAPK pathway are enriched at disease relapse (Fig 3B) [9].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3.

Mutation frequencies in MYCN amplified and non-amplified tumors, and in the MAPK pathway A) Frequency of suspected driver genomic alterations (short variants, copy number changes, and genomic rearrangements and fusions; MYCN amplification events excluded) in patients with and without MYCN amplification. Bars represent number of patients with at least one lesion in a given gene, normalized by the number of patients in each category. B) Percentage of patients with any suspected driver variant in the MAPK pathway at diagnosis, after treatment, and at relapse. Patients with multiple mutations in the same gene are only counted once.

https://doi.org/10.1371/journal.pgen.1006501.g003

Nine patients had samples submitted for sequencing at various times in therapy (S2 Table). Of these, two patients did not have tumor samples with driver mutations. However, paired samples from the remaining seven patients provided insight into neuroblastoma tumor composition and evolution. Notably, we observed that definitive clonal driver mutations such as MYCN amplification and ALK mutation identified in diagnostic samples frequently persisted as drivers following exposure to chemotherapy. Interestingly, Patient 82 presented with bilateral adrenal disease, and biopsies were obtained at diagnosis from both the right and left adrenal glands. Of the 13 total variants detected in both of these samples, 10 variants were shared in common between the two tumors, suggesting that both tumors shared a common cell of origin. Patients 1, 3, 4, and 36 had multiple biopsies from the same tumor location at different time points during therapy (Patient 1: cervical lymph node metastasis at diagnosis and relapse; Patient 3, 4: adrenal gland primary tumor at diagnosis and definitive surgery; Patient 36: tibial metastasis at two relapses). Unsurprisingly, three of the four patients showed similar mutational profiles with only slight differences at each time point; however, Patient 3 had no variants in common between the two biopsies, demonstrating the potentially significant effects of spatial tumor heterogeneity and evolution. This phenomenon was further illustrated by Patient 2, who had four serial biopsies at different anatomic sites (Table 1). This patient’s diagnostic tumor contained multiple driver lesions, but some appeared clonal (heterozygous ARID1A mutation with allelic fraction approximating 50%), while others appeared to be present in only a subclonal fraction of cells studied (incomplete ATRX deletion in this male patient and inferred FGFR1 mutation in 30% of cells). Interestingly, two unique ALK mutations emerged at the first and second relapse in two different metastatic locations, and in each case these were subclonal or absent at the time of final relapse following crizotinib treatment, whereas the ATRX and CDK4 copy number lesions appear to have been evolutionarily selected for and were enriched in the final relapse specimen. This single case demonstrates the absolute requirement of treating physicians to understand that mutations are not binary events, and consideration of whether or not a lesion is clonal or subclonal may have major implications in the outcomes of targeted therapeutic interventions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Genetic variants from a single patient at different treatment time points.

Each biopsy was at a different anatomic site. Red denotes suspected driver variants; gray denotes variants of unknown significance. Letter preceding tumor location indicates primary (P) or metastatic (M) site. Number in parentheses indicates inferred allelic fraction for mutation calls, or inferred copy number for amplification or deletion calls. See S2 Table for additional details. Note that this patient was treated with crizotinib following the 5^th relapse.

https://doi.org/10.1371/journal.pgen.1006501.t001

Of the 138 unique patients in this cohort, 15 were reported to have received targeted therapy based on FoundationOne sequencing results and 13 had reported outcomes. Targeted treatments included ALK inhibitors, CDK4/6 inhibitors, MEK inhibitors, Raf inhibitors, mTOR inhibitors, and antibody therapies. One patient currently has an ongoing complete response, two have ongoing stable disease (>1 year and >6 months), two patients exhibited stable disease (4 months and 1 year) followed by progressive disease, and the remaining nine patients showed progressive disease despite treatment.

Discussion

Recent papers have demonstrated that there are relatively few somatic genetic alterations in neuroblastomas at diagnosis, but also that the frequency of mutations and other genetic aberrations increases at disease relapse [9,10]. Here we performed a retrospective study of a larger cohort of 138 patients whose tumors had been biopsied and sequenced at diagnosis, second look surgery, relapse, or some combination of these, to gain a better understanding of the mutational landscape of neuroblastoma throughout treatment and disease progression. The current study provided validation of our original observation that relapsed high-risk neuroblastoma harbor an increased mutational burden [9]. Gene panel sequencing is not designed to infer mutational burden, but our data are consistent with an enrichment of mutations in known cancer driver genes at relapse, consistent with clonal evolution under the selective pressure of chemoradiotherapy. We also showed that the prevalence of “potentially actionable” mutations increased at relapse. Our data provide support for the hypothesis that the neuroblastoma genome evolves under the selective pressure of chemoradiotherapy [9,10], and It is necessary to understand the current genomic landscape of a tumor rather than that at diagnosis in order to make the most informed treatment decisions. Thus, the clinical practice of acquisition and analysis of relapsed tumor material, with a view to designing prospective clinical trials to determine the impact of matched target therapy on patient outcomes should be encouraged.

In parallel with several recent reports of prospective collection of tumor genomic data using gene panel [12] or whole exome [13–15] sequencing across pediatric cancer, with or without transcriptome sequencing and at diagnosis or relapse, a few general conclusions are apparent. First, while not evaluable in this retrospective study, the prospective studies noted above all clearly demonstrate that the acquisition of high dimensional genomics data from pediatric cancers is feasible, and to date has been demonstrated to be safe to incorporate into clinical practice. This is relevant for this retrospective study of neuroblastoma because the majority of relapses in this disease are in the bone or bone marrow compartment, and often require skilled interventional radiologists and pathologists to safely access, process and prepare often small samples for sequencing. Second, together the data support the conclusion that a significant proportion of patients at relapse will have a potentially actionable oncogenic driver lesion. In this study, 58% of subjects studied at relapse showed a somatic mutation in one of several genes that are major targets for drug development in other cancers, consistent with other reports [12,14,15]. Third, this study did not study the germline genome, but certainly many of the mutations identified in this study have the potential to be germline [16], and recent studies have shown that up to 10% of pediatric patients harbor likely relevant germline mutations in known predisposition genes [12–14,17]. Paired germline and somatic DNA sequencing should be considered in the genetic evaluation of pediatric cancer subjects. Finally, our and recent studies together highlight the current dismal reality that, at present, pediatric cancer sequencing efforts are rarely leading to patient benefit. The reasons for this are multifactorial and applicable to cancer in general, such as lack of adequate mutation-drug matches for many driver lesions. Further, there is a need to look beyond mutations in the coding genome, and forays into clinical epigenomic profiling and mRNA profiling may yield other targetable genomic alterations. However, there are also likely childhood cancer specific explanations as well, such as lack of pediatric dosing information and perhaps some degree of risk aversion on the part of physicians and industry partners to expand targeted therapy options in this population of children who most likely will die from their disease. Taken together, recent efforts in next generation sequencing of pediatric cancer highlight the potential clinical significance of these efforts, but more importantly the urgent need to credential precision medicine approaches to childhood cancer in well designed prospective clinical trials. Such clinical trials are currently underway, and are investigating mutations not only through next-generation sequencing of tumor samples, but also through collecting and sequencing cell free DNA and circulating tumor DNA [18,19].

Materials and Methods

Supporting Information

Acknowledgments

The authors would like to thank Drs. John Anderson, Ellen Basu, Satheesh Chonat, Steven Dubois, William Clay Gustafson, Giannoula Lakka Klement, Brian Kushner, Brian Ladle, Stephen Liu, Stephen P. Naber, Navin Pinto, Stephen Roberts, Anita Sengupta, Peter Shaw, Diana Steppan, Kimo Stine, Brian Turpin, Johannes Wolff, and Teresa York for providing patient data.