The Use of Global Profiling in Biomarker Development for Gliomas

Abstract

The diffuse gliomas are a heterogeneous group of malignancies with highly variable outcomes and diagnosis is largely based on the histological appearance of the tumors. Tumor classification according to cello type and grade provides some prognostic information. However, the diversity of gliomas, within tumor type and grade categories, has made prognostic determinations based purely on clinicopathologic variables difficult. There is an increasing body of data suggesting a significant amount of molecular diversity accounts for the heterogeneity of clinical observations, such as response to treatment and time to progression. The last decade has witnessed an explosive advance in our knowledge of the molecular genetics of brain tumors, due in large part to the availability of high‐throughput profiling techniques, including new sequencing methodologies as well as multidimensional profiling by the Cancer Genome Atlas project. The large amount of data generated by these efforts has enabled the identification of prognostic and predictive factors and helping to identify pathways that are driving tumor growth. Identification of biomarkers, especially when coupled to clinical trials of newer targeted therapies, will enable better patient stratification and individualization of treatment.

INTRODUCTION

Neuroepithelial tumors are a diverse group of malignancies that are subdivided into distinct groups based largely on clinical characteristics. While this has been very useful in defining the prognosis and response to particular therapies, there remains a significant variability in the outcomes of patients and a need to further subdivide these diseases. Molecular profiling represents a promising opportunity to further define prognostic and predictive factors. It is therefore important to define the terminology before further discussing advances in this field. The terms prognostic and predictive are often incorrectly used interchangeably. We define a prognostic marker as a correlative factor associated with more or less favorable tumor biology, and hence a greater or lesser likelihood of responding to treatment compared with a typical tumor. Classic prognostic factors include such established clinicopathologic factors as stage and histological grade. Such factors are felt to be markers of the overall natural history of the tumor. In contrast, predictive markers are those that are predictive of a response to a particular therapy. An example is Her2 amplification in breast cancer. Her2 positive tumors are predictive of tumor response to trastuzumab and Her2 positive patients have significantly increased overall survival and time to failure compared with Her2 negative breast cancers following trastuzumab treatment (58).

Approaches to genomic profiling

Of all the methods to profile tumors at the molecular level, perhaps the most widely used to date is gene expression profiling. Prior to the advent of high‐throughput microarray technologies, profiling the expression of genes was limited to few or a single gene at a time, typically by using Northern blotting. This progressed to the development of arrays of complementary DNAs (cDNAs) on membrane substrates, which allowed examination of the expression of multiple target genes simultaneously, though with very little quantitative information. Two major technological breakthroughs were developed at Stanford University in the 1990s: the ability to miniaturize arrays and the use of fluorescent probes. This enabled collection of data on the expression of thousands of genes simultaneously and also provided quantitative data (61). The technology rapidly progressed with the use of oligonucleotide probes, high‐density arrays and precision manufacturing techniques such as the photolithographic techniques used to produce the modern Affymetrix GeneChip (Affymetrix, Inc., Santa Clara, CA, USA) and similar types of whole‐genome arrays that allow for comparison of expression among large numbers of tumors with little concern for technical, chip‐to‐chip variation 5, 48.

Some general approaches have been employed to analyze the large datasets generated by this method. Unsupervised clustering is a statistical method that determines correlation in expression among probes to bin targets into groups with similar patterns of expression (17). The groups that share molecular signatures often share an underlying biology and hence this method can potentially identify new members of pathways or identify genes that share a common function. Supervised clustering can be used to identify gene signatures for known classes of samples (3). Tumors can be grouped, for example, by patient outcome (survival vs. death) or by known molecular changes (1p/19q allelic codeletion in oligodendroglioma vs. not deleted) and signatures for each type can be identified. To identify genes which may be involved in tumorigenesis from a large dataset, it is necessary to perform statistical analyses to identify the false discovery rate and to set the threshold for discovery above this level. To guard against the inevitable false positives observed in large datasets, validation of the results is necessary (3). The use of independent sample sets is of critical importance, and an often overlooked aspect of microarray analysis. This minimizes false positives resulting from testing a large number of variables in a limited number of samples. With the use of multiple, large sample sets, the results can be generalized and valid gene “signatures” identified. These gene signatures have been increasingly used as biomarkers.

Gene signatures have been reported that are capable of distinguishing molecular subtypes of tumors that appear indistinguishable histologically but often have very different clinical outcomes. The groundbreaking study profiling diffuse large B‐cell lymphoma (DLBCL) is an example of such work (2). Extensive work of this type has also been reported for breast cancer, where gene expression profiling has identified five subtypes of tumors, each with different aggressiveness, and these signatures have now been incorporated into clinical diagnostic tools 4, 23.

While gene expression profiling has enjoyed the most exposure as the method of choice for identifying prognostic and predictive factors, this owes in large part to the longer time period during which this technology has been in wide use. The technology currently driving molecular diagnostics is DNA sequencing, in particular next‐generation sequencing using methods such as SOLiD® (Applied Biosystems, Carlsbad, CA, USA), Roche/454® (Roche Diagnostics Corp, Branford, CT, USA), or Illumina Genome Analyzer® (Illumina, Inc., San Diego, CA, USA) as well as others (37). It is beyond the scope of this review to describe the technical details of these sequencing methods, but they combine novel hardware, software engineering, chemistry, enzymology and high‐resolution optics to produce massive numbers of short sequence reads in parallel. The result has been a technique to generate a sequence output on the whole genome level (36). Initial large‐scale efforts focused on targeted sequencing, while other groups have taken the approach of sequencing entire genomes, or “deep” sequencing. The emergence of this technology has enabled us to not only better understand the landscape of the cancer genome, but also explore the epigenome by determining such things as DNA methylation, microRNA and histone modification, on a global scale. Finally, methods are being developed for the integration of multiple approaches (11). While methods of one‐dimensional analysis have been rapidly improving and leading to the identification of important cancer genes, genes affected by low‐frequency events still remain undiscovered. The use of multidimensional analysis may prove instrumental in identifying these low frequency, yet critical, genomic alterations. Gene expression profiling, deep sequencing and epigenomic profiling are increasingly being combined with techniques that address the amplification or deletion of genetic material, such as array comparative genomic hybridization (CGH) and single nucleotide polymorphism (SNP) analysis.

PROFILING OF HUMAN GLIOMAS

Within a few years of the development of the technology, the first studies of expression profiling of human brain tumors were reported. The initial studies utilized large, filter‐based arrays 9, 10, 22, 28, 54, 60, 64. In one such study commercial filters containing 597 genes were used to compare the expression of 20 tumors [5 glioblastoma (GBM), 5 anaplastic astrocytoma (AA), 5 anaplastic oligodendroglioma (AO), and 5 oligodendroglioma] to 2 normal brain specimens. The insulin‐like growth factor binding protein 2 (IGFBP2) gene was found to be upregulated in GBM and has since been proposed to have a biological function in GBM. While early profiling studies showed promise, they highlighted the need for high‐resolution arrays to allow for much larger sample sizes. An early study utilizing higher‐throughput arrays included 45 astrocytic tumors (19 grade I, 5 grade II and 21 grade IV) as well as 6 normal brain specimens were profiled on approximately 6800 probes (56). A signature of 360 genes was identified that distinguishes pilocytic (grade I) from GBM (grade IV). Five of the genes upregulated in GBM, but not previously known to play a role in the pathogenesis of this disease, were validated by quantitative real‐time polymerase chain reaction (qRT‐PCR) from tumor derived cDNA and immunohistochemistry with clinical specimens. Since the publication of this work, other groups have reported a role in the tumorigenesis and progression of gliomas for some of these genes, for example, FOXM1 15, 35, 66. Since the advent of cDNA microarrays in the 1990s, the technology has progressed significantly such that current arrays contain not only every gene in the genome, but also all potential splice variants, or so‐called “exon” arrays. Such arrays have been used for glioma profiling 12, 21. Interestingly, glioma‐specific alternative splicing appeared to be a somewhat rare event in one study, with only 14 genes identified with glioma‐specific splice variants (12).

A number of groups have attempted to identify individual genes as well as signaling pathways from microarray data that are prognostic in malignant gliomas 19, 30, 33, 34, 39, 42, 55, 59, 62, 65. Using a group of only 19 gliomas, the chitinase‐3‐like gene (CHI3L1), which encodes for the glycoprotein YKL‐40, was found to be upregulated compared with normal brain and its expression correlated with tumor grade (65). In another set of GBMs, this time with 34 tumors, CHI3L1 was again identified as a significant gene and its expression was prognostic with decreased overall survival and in vitro radioresistance (42). Subsequently, additional studies have also validated the role of CHI3L1 as a biomarker and suggested a functional role for YKL‐40 in glioma progression 45, 49, 50. In a study examining high‐grade pediatric astrocytomas, the epidermal growth factor receptor (EGFR) and hypoxia‐inducible factor‐2α (HIF‐2α) pathways were implicated in this disease (30). Another study revealed a 70‐gene signature in GBMs, which differentiated long‐term from typical survivors (33). One of the genes in this signature, fatty acid binding protein 7 (FABP7), was prognostic in an independent cohort of 105 patients. The ephrin receptor EPHA2 gene was found to be a prognostic marker based on an integrative analysis of mRNA profiling with CGH (34). Several other studies have attempted to combine classic profiling data with genomic structural information, in a similar way, to most accurately identify prognostic markers 34, 42, 51. Such combined approaches show great promise for the future of biomarker discovery.

In a manner similar to some of the earliest studies mentioned above (56), a number of groups have attempted to predict tumor classification (eg, tumor grade) based on expression profiling 6, 14, 24, 31, 38, 44, 53, 63, 69. In one example, Affymetrix U95A GeneChips were used with 35 gliomas to identify a 170‐gene signature that accurately classified tumors based on grade. Another group attempted to develop a molecular classifier based on gene expression profiling that could be used to help classify gliomas (AO vs. GBM) with “nonclassic” histologies (44). Twenty‐one tumors were used to build a gene classifier that was then applied to an additional 29 tumors with ambiguous histology. The classifier was superior at being able to predict survival compared with histological diagnosis (44).

Phillips et al analyzed 76 high‐grade gliomas and classified survival‐associated genes into three groups: proneural, mesenchymal and proliferative based on analysis of gene ontology of the genes most high expressed into each group of tumors (52) (Figure 1). This classification proved to be highly prognostic in independent datasets. The proneural and mesenchymal groups have been subsequently identified in studies by other groups as major components of their classification schemes, including reports by The Cancer Genome Atlas (TCGA) (25). Genes within these classifications individually showed strong prognostic significance and a clinical test established to predict outcome for patients with GBM has included major proneural or mesenchymal signature genes despite the fact that the predictor was developed using a different statistical strategy (13). While a variety of molecular classification systems have been proposed, the collective evidence suggests that there are two major groups of gliomas. One group shows relative overexpression of genes with functional ontology relating to cell motility, extracellular matrix and cell adhesion (mesenchymal). The second demonstrates relative overexpression of genes known to be associated with neural development (proneural). Further subclassifications or different refinements have been proposed 1, 32, but methodologies for developing these classification systems differ and the optimal method for classification has not reached a consensus. With respect to the two major subtypes, there is a clear association of tumor grade with subtype: while GBMs are found to represent a mix of mesenchymal and proneural subtypes, grade II and grade III diffuse gliomas are almost invariably proneural. Consistent with these findings, those GBMs that are proneural tend to have independent clinical and molecular evidence of “secondary” GBM, including younger patient age, higher rates of p53 and IDH1 mutation and lower rates of EGFR amplification and chromosome 10 loss 32, 42, 52.

Figure 1.

Classification of high‐grade gliomas into proneural, mesenchymal or proliferative subtypes. A. Oligonucleotide‐based, gene expression profiles were obtained in a set of WHO grade III and IV gliomas and those genes most highly correlated with survival used to classify the tumors by hierarchical clustering. A unique expression pattern emerges separating the tumors into three distinct molecular subtypes. Genes with high expression are shown in red, while those with low expression are shown in green. Gene ontology analysis (ie, analysis of gene function) revealed that the highly expressed genes in each tumor group have either a neuro‐developmental (pronoeural), a mesenchymal or a proliferative function. These subtypes demonstrate statistically significant associations with patient outcome in both grade III and IV glioma (B) and in an independent dataset consisting only of grade IV glioblastoma (GBM) with necrosis (C). (Adapted from Phillips et al (52) with permission.)

While much of the work described above has focused on astrocytic tumors, there are also specific studies that have focused on oligodendroglioma 16, 20, 26, 29, 40, 41, 67, 71. In one study, an 1100 gene signature was capable of distinguishing WHO grade II from grade III tumors on the basis of the gene expression pattern (71). Response to treatment in oligodendrogliomas is related to deletion of human chromosome 1p/19q, however, as response is not restricted to a single modality, it remains more of a prognostic factor than able to identify the optimal treatment regimen (8). In an attempt to refine this marker, several groups have attempted to identify unique expression profiles based on 1p/19q status 16, 20, 41, 67. The expression of genes or gene signatures has been found to correlate with 1p/19q deletion, including the proneural signature 16, 41 or candidate tumor suppressor genes (67). The idea of functional gene signatures, which may be independent of histological grade but nevertheless reflect tumor biology has been developed by a number of groups and will likely gain increased use in the routine pathological characterization of human brain tumors.

BEYOND GENE EXPRESSION PROFILING: THE USE OF MULTIPLE PLATFORMS

A more robust analysis is often possible by combining profiling platforms. For example, several of the groups mentioned above have attempted to combine genomic copy number data with mRNA expression data from microarrays 16, 29, 34, 40, 42, 51, 67. The availability of multiple tools to profile tumors, including mRNA expression, microRNA profiling, SNP genotyping, array CGH, direct deep sequencing of tumor samples and epigenetic evaluation by assessment of promoter CpG island methylation provides the opportunity to generate the most robust classifier. The multiplatform approach has been reported by several groups of investigators 1, 47.

Perhaps the most well‐known example of multiple‐platform profiling is the TCGA project, sponsored by the National Cancer Institute in the United States. One of the goals of this project is to profile a large cohort of GBMs (approximately 500) at the DNA, mRNA, microRNA and epigenetic (DNA methylation) levels (1). In an initial report, the preliminary analysis of the first 206 GBMs has already provided validation of known genes and pathways previously implicated in GBM as well as identified new targets (1). More recently, this work has been extended by identifying the previously described proneural, and mesenchymal subtypes, as well as describing additional subtypes (neural and classical). In this dataset, the subtype assignment was not associated with patient outcome, and further work is required to address the clinical significance of these subtypes (70) (Figure 2).

Figure 2.

The Cancer Genome Atlas (TCGA) classification of glioblastoma (GBM) based on gene expression based and correlation to known genetic alterations and patient outcome. Gene expression data from 202 GBM samples was used to identify four molecular subtypes, the previously reported proneural and mesenchymal types (52), and two additional subtypes, neural and classical. For the subset of 116 cases with both mutation and copy number data, correlations to TP53, IDH1, PDGFRA, EGFR, NF1, and CDKN2A were reported. Below are Kaplan‐Meier graphs reported to show the correlation between molecular subtype and survival based on intensity of therapy. In the molecular classification reported by the TCGA, only the classical and mesenchymal groups showed a significant association with survival.

Another group took a more comprehensive sequencing approach (eg, “deep” sequencing) and sequenced >20 000 genes in 22 tumors. The sequencing data from this study was correlated with gene expression profiling and array CGH data and identified the isocitrate dehydrogenase 1 gene (IDH1) as a novel site of mutation in GBM (47). Of note, this gene was not included in the initial analysis by the TCGA. In a follow‐up study, the group then showed that the majority of several types of malignant gliomas, including secondary GBMs, harbored mutations in the IDH1 gene or the closely related gene IDH2 (72). This topic will be covered in more depth by another review article in this issue, but highlights advantages of unbiased approaches toward screening for biomarkers and new biologic concepts in gliomas.

Another recent study highlighted the use of targeted proteomic analysis combined with analysis of genomic and expression data (7). This group performed targeted proteomics on 27 surgical glioma samples using Western blots probed with antibodies for pathways of interest such as the PDGF pathway, EGFR pathway and Notched pathways. They used specific antibodies to activated forms of some members of these pathways (eg, phospho‐EGFR), highlighting the ability to integrate post‐translational modification into tumor profiling. Using this data, they identified patterns of coordinate activation among these pathways and compared these results with integrated analysis of expression and genomic data from 243 GBM samples from the TCGA dataset. Three important pathways emerged from this integrated analysis of proteomic and genomic data, the EGFR signaling pathway, the PDGF pathway and the NF‐1 pathway (7).

A significant departure from the previously mentioned glioma classification schemes, which have relied on mRNA expression or genetic alterations, has been the recent analysis of epigenetic changes, namely DNA promoter CpG island methylation. In the first study of this modality, analysis of methylation profiles with 272 GBMs from TCGA found that a distinct subset of samples displayed concordant hypermethylation at a large number of loci, thus demonstrating a glioma‐CpG methylator phenotype (G‐CIMP, Figure 3) (43). They demonstrated that G‐CIMP tumors are associated with IDH1 somatic mutations, belong to the proneural subgroup, display distinct copy‐number alterations and are more prevalent in low‐grade gliomas. Furthermore, patients with G‐CIMP tumors were found to be younger at the time of diagnosis and experienced significantly improved outcome across all grades.

Figure 3.

Identification of the CpG island methylator phenotype (CIMP) in glioblastoma (GBM). DNA methylation data from 91 tumors profiled by The Cancer Genome Atlas (TCGA) was subjected to unsupervised clustering. Three distinct methylation clusters were identified, with Cluster #1 (red in top bar) designated as G‐CIMP+ due to the high‐frequency of methylation. The genes whose promoters are methylated in that cluster comprise the G‐CIMP signature. The gene expression classification and gene mutation status (EGFR, IDH1, NF1, PTEN, and TP53) is shown for above each tumor signature. G‐CIMP‐positive tumors (labeled below the heatmap) were correlated with IDH1 mutations and exhibited a significantly improved survival. Controls include fully methylated DNA (by SssI methyltransferase treatment) and whole‐genome ampified (WGA) unmethylated DNA. (Adapted from Noushmehr et al (43) with permission).

The use of multiplatform and/or multimarker profiles will not only improve our understanding of tumor biology, but may potentially improve upon treatment decisions. Several studies in breast cancer suggest that a multimarker panel is more robust with respect to prediction of outcomes than single gene makers on clinicopathologic predictors 46, 68. In GBM, a multigene profile compatible for formalin‐fixed paraffin embedded (FFPE) samples is currently used as a stratification factor in a large phase III clinical trial (RTOG‐0825) (13). The use of FFPE‐based assays is critical to the wide scale acceptance of a biomarker due to the limited availability of fresh/frozen tissues for most brain tumor types. The 9‐gene set was validated with an independent sample set and was shown to be an independent predictor of clinical outcome after adjusting for clinical factors and as well as methylation of 0⁶‐methylguanine methyl transferase (MGMT) status, a known prognostic molecular marker in GBM. Interestingly, this 9‐gene profile was also positively associated with the glioma stem‐like cell markers CD133 and nestin. This study opens the door for the development of biomarker panels that could predict patients whose tumors are unlikely to respond to standard therapy, enriching for a population that may specifically benefit from clinical trials with new agents in the up‐front setting. Additionally, given the difficulty of distinguishing true progression from pseudoprogression following combined radiation/temozolomide, a biomarker panel which could predict response to this therapy could add additional data to help distinguish pseudoprogression from true progression in the clinical setting. Future trials could incorporate molecular inclusion criteria to identify refractory tumors, and targeted therapies for resistant tumors, allowing for the personalized treatment of brain tumors.

PREDICTIVE MARKERS IN BRAIN TUMORS

Prognostic markers themselves do not necessarily impact patient care unless they alter treatment choice and very few predictive markers have been identified in gliomas and other brain tumors. Perhaps the best example in brain tumors of a predictive marker is the 1p/19q codeletion in oligodendroglioma mentioned above, where this codeletion is a predictive marker for sensitivity to chemotherapy. However, as the 1p/19q co‐deletion also predicts for radiation sensitivity, it does not meet the full criteria for a predictive factor. In the case of GBM, methylation of the promoter of MGMT, a DNA repair gene, is potentially both a prognostic and predictive marker to response to alkylating agent. A report in 2000 found that methylation of the MGMT promoter was associated with prolonged overall and disease‐free survival independent of age and performance status (18). In the companion report to the phase III trial establishing the use of temozolomide in the treatment of primary GBM, patients who had MGMT promoter methylation and treated with temozolomide exhibited improved survival (27). However, even patients with unmethylated MGMT promoter had an improved survival outcome when temozolomide was added suggests that MGMT methylation status is more of a prognostic factor than a predictive one. A separate analysis of patients treated for primary GBM with radiation in the pre‐temozolomide era found MGMT methylation to be a prognostic factor, suggesting that its role in predicting response to temozolomide may be overstated and rather reflect the inherent prognostic role of this marker (57). Future studies will help clarify whether MGMT methylation is a truly predictive marker, whether it should influence therapeutic options for the individual patient, and how it can best be incorporated with robust multigene predictors.

CONCLUSIONS

Human brain tumors represent a heterogeneous group of tumors, both with respect to tumor characteristics and patient outcome. While classification of human brain tumors to date has largely relied on histological characteristics, emerging tools for molecular analysis are rapidly advancing the field of molecular pathology for brain tumors. Transcriptional profiling of these tumors has identified molecular subtypes with complex gene signatures resulting in the identification of prognostic markers independent of existing clinical and pathological criteria. While the discovery of prognostic markers is rapidly progressing, few predictive markers have emerged to date. Multiplatform profiling approaches, including deep sequencing and epigenetic profiling, will provide more robust biomarkers, and this approach is poised to rapidly advance the field of molecular neuropathology. The further identification of prognostic and predictive markers allows for customization of therapy to match an individual patient's genetic makeup.