Purpose: The molecular heterogeneity of glioblastoma has been well recognized and has resulted in the generation of molecularly defined subtypes. These subtypes (classical, neural, mesenchymal, and proneural) are associated with particular signaling pathways and differential patient survival. Less understood is the correlation between these glioblastoma subtypes with immune system effector responses, immunosuppression, and tumor-associated and tumor-specific antigens. The role of the immune system is becoming increasingly relevant to treatment as new agents are being developed to target mediators of tumor-induced immunosuppression, which is well documented in glioblastoma.
Experimental Design: To ascertain the association of antigen expression, immunosuppression, and effector response genes within glioblastoma subtypes, we analyzed the Cancer Genome Atlas (TCGA) glioblastoma database.
Results: We found an enrichment of genes within the mesenchymal subtype that are reflective of antitumor proinflammatory responses, including both adaptive and innate immunity and immunosuppression.
Conclusions: These results indicate that distinct glioma antigens and immune genes show differential expression between glioblastoma subtypes and this may influence responses to immunotherapeutic strategies in patients depending on the subtype of glioblastoma they harbor. Cancer Immunol Res; 1(2); 112–22. ©2013 AACR.
Glioblastoma remains the most common malignant primary brain tumor in adults. Despite aggressive treatment with surgical resection, radio-, and chemotherapy, the tumor ultimately recurs. A major obstacle to treatment is the subversion of the immune system by the tumor to facilitate proliferation and malignant degeneration of tumor cells. Immunosuppression is thought to play a major role in the aggressive nature of gliomas and their resistance to current therapies. The presence of immunosuppressive infiltrates such as FoxP3+ regulatory T cells (Treg) and M2 macrophages has been documented in gliomas and in some cases, correlates with prognosis (1–3). In addition, immune responses to glioma can be impaired by the tumor itself through expression of immunosuppressive cytokines such as TGF-β and the upregulation of the STAT3 (4). Genomic profiling of glioblastomas has shown that up to 4 genotypic subtypes exist, with 2 showing marked differences in gene expression and patient survival (proneural and mesenchymal; ref. 5). The STAT3 pathway has been shown to be the key molecular driver of the mesenchymal transformation within glioblastoma (6, 7). In addition, STAT3 has been implicated in many mechanisms of tumor-mediated immunosuppression (8–13) and is a negative prognosticator for survival in mice (14) and in human patients with anaplastic astrocytoma (15). These data would suggest that the mesenchymal glioblastoma subtype may be more immunosuppressive compared with other glioblastoma subtypes and possibly more refractory to immunotherapy.
Immunotherapy is an appealing treatment of gliomas because it allows for tumor specificity while minimizing collateral damage to normal brain tissue. Clinical trials using dendritic cell or peptide vaccines to target glioma cells have shown promising results (16–19). However, as with other treatments, only subsets of patients respond. This may be due to molecular and genomic factors that affect interactions between the immune system and the tumor. A recent study using glioma lysate-pulsed dendritic cell vaccination showed that glioblastomas of the mesenchymal phenotype had higher levels of CD3+ and CD8+ tumor-infiltrating lymphocytes than glioblastomas of other subtypes (20). Furthermore, patients whose tumors had the mesenchymal gene signature, survived longer after dendritic cell vaccination than controls of the same genetic subtype (20). Although this finding tends to contradict what is thought about the immunosuppressive nature of mesenchymal gliomas with respect to aggressiveness and poor patient survival, it suggests that mesenchymal gliomas may be more immunogenic and more responsive to immunotherapy. Interestingly, robust immunosuppressive Treg infiltration into the glioblastoma microenvironment is almost always observed with a corresponding influx of T effector cells (2). Thus, one may speculate that the tumor becomes selectively more immunosuppressive as a reaction to antitumor effector responses. To investigate this paradox, we analyzed mRNA expression levels of immune system genes among the various glioblastoma subtypes using The Cancer Genome Atlas (TCGA) database and found not only a preferential enrichment of immunosuppressive genes but also an enrichment of immune effector genes within the mesenchymal subset.
Materials and Methods
TCGA data acquisition
It should be noted that the composition of the analyzed glioblastoma tissue from the TCGA database contains genetic information from both glioma cells and tumor-supportive stroma cells including infiltrating immune cells because these components were not selectively eliminated from the specimen, which was used in the construction of the database. The glioblastoma cancer study set from the TCGA database consists of defined glioblastoma subtypes (21). The TCGA database was analyzed in 2 distinct ways: (i) using online knowledge bases [Ingenuity Pathway Analysis (IPA) and Uniprot Protein Knowledgebase (www.uniprot.org)] to define genes associated with immune responses; and (ii) using a collated list of immune response genes from the literature with an emphasis on those previously documented to have a role in glioblastoma. For the first analysis, mRNA expression levels (available from the TCGA Research Network website: cancergenome.nih.gov), of the proneural (n = 107), mesenchymal (n = 119), classical (n = 115), and neural (n = 58) glioblastoma subsets were used as the source data available as of February 2013. This first analysis compares mRNA expression levels using the mesenchymal subset as the reference relative to the other subsets. The second analysis was through the open access cBio Cancer Genomics Portal at www.cbioportal.org (22), in which proneural (n = 141), mesenchymal (n = 160), classical (n = 147), and neural (n = 96) were used as the source data. This second analysis compares specific mRNA expression levels between glioblastoma subsets. To analyze mRNA expression based on Agilent microarray of selected genes in the 4 subtypes of glioma, the z score threshold for all genes (the number of SDs above the mean expression level of the selected gene) was set to more than 1 for each of the subtypes [an example of the syntax used for the search is: interleukin (IL)-10: EXP>1].
Compilation of immune genes and tumor antigens
The immune effector gene sets were defined by:
The IPA tool to identify immune activators with the following characteristics: cytokine, enzyme, G-protein–coupled receptor, growth factor, kinase, peptidase, phosphatase, transcription factor, transcription regulator, transmembrane receptor, transporter, and unknown and secondarily the Uniprot knowledge base (www.uniprot.org) for the terms “activation of immune response” and “immune effector process.” Of 814 genes identified by IPA and Uniprot, 734 could be found in the TCGA database and were used for the analysis.
By collating a list from the literature of documented proinflammatory effector cytokines (IFN-γ, IL-1, IL-2, IL-4, IL-7, IL-12, IL-15, and TNF-α); surface markers reflective of the presence of immune effector cell responses (CD3 and CD8) and their associated immune-activating markers (CD80, CD86, CD40, and HLA); signaling pathways reflective of adaptive immune activation (NF-κB, STAT1, IRAK, STAT4, T-bet, DNAM1, and IRF7); and innate immunity and markers reflective of activation (NKp46, NKG2D, NKp30, NKp44, KLRD1, TREM1, TREM2, MIF, TLR2, TLR3, TLR4, and TLR9) that participate in antitumor effector responses.
The immunosuppressive genes were defined by:
The IPA tool to identify mediators of immunosuppression with the following characteristics: cytokine, enzyme, G-protein–coupled receptor, growth factor, kinase, peptidase, phosphatase, transcription factor, transcription regulator, transmembrane receptor, transporter, and unknown and secondarily the Uniprot knowledge base for the search terms related to tumor-mediated immunosuppression including: “inhibition of immunity,” “inhibition of immune system process,” “inhibition of effector immunity,” and “inhibition of immune response to tumor cell.” Of 235 genes identified by IPA and Uniprot, 218 could be found in the TCGA database and were used for the analysis.
By collating a list from the literature of documented immunosuppressive cytokines and mechanisms (galectin-3, VEGF, IL-10, IL-23, TGF-β, PD-1, PD-L1, and CTLA-4); chemokines (CSF-1, CCL2, and CCL22); tumor-supportive and immunosuppressive myeloid- and monocyte-related genes (CD163, CD204, MIC-1, arginase, and CD47); immunosuppressive signaling pathways (IL-6, gp130, Jak2, STAT3, Pim-1, SOCS3, STAT5A, and STAT5B); and markers related to Tregs (CD4, ICOS, IDO, and FoxP3) that participate in tumor-mediated immunosuppression.
The glioma antigen selection was based on the previously analyzed and compiled list by Zhang and colleagues (23) but was further expanded to include other known potentially overexpressed antigens such as EGF receptor (EGFR; 24), and nucleolin (25).
A χ2 test was used to compare the number of patients between glioblastoma subtypes in which mRNA expression levels of a selected gene were at least one SD above mean expression in a particular subtype. To identify differentially expressed genes between the subtypes, a modified two-sample t test using the limma package was applied. Genes that were differentially expressed (DE) on the basis of the false discovery rate (FDR) of at least <0.01 in the mesenchymal subset relative to the proneural (∼7,000), neural (∼9,000), and classical (∼6,000) subsets were then identified. The β-uniform mixture (BUM) model, described by Pounds and Morris (26), was used to control FDR. This list of differentially expressed genes was then dichotomized into 2 subgroups: genes (mRNAs) overexpressed in the mesenchymal subtype [designated DE(m>x); where m = mesenchymal and x = subtype] and vice versa [DE(x>m)]. Pearson χ2 with Yates' continuity correction and hypergeometric tests were used to determine the enrichment of the immune gene sets within these 2 groups.
Immune genes have differential expression in glioblastoma subtypes
Analysis of the IPA-selected immune activators and suppressors revealed that the greatest immunologic diversity exists between the proneural and mesenchymal subsets (Fig. 1). Only 17% (n = 123) of the immune activators and 19% (n = 41) of the immunosuppressive genes shared differential expression among all glioblastoma subtypes (proneural, classical, and neuronal) relative to the mesenchymal subset. This diversity of the immune activating and immunosuppressive genes can distinguish the various glioblastoma subtypes on heatmaps (Fig. 2). To ascertain which immune genes were differentially enriched in the mesenchymal subset in comparison with the proneural subset, the immune genes were ranked and then compared with the overall 17,000 genes from the TCGA dataset (Table 1). Using both χ2 and hypergeometric tests to ascertain if the immune genes were selected by chance, a preferential enrichment of both immune activators (Pearson χ2 test with Yates' continuity correction, P < 2.2 × 10−16; hypergeometric test, P < 2.2 × 10−16) and immunosuppressors (Pearson χ2 test with Yates' continuity correction, P = 3.5 × 10−5; hypergeometric test, P = 2.04 × 10−5) was found in the mesenchymal subset relative to the proneural subset. Similarly, in the comparison between the classical and mesenchymal subset, immune activators (Pearson χ2 test with Yates' continuity correction, P < 2.2 × 10−16; hypergeometric test, P <2.2 × 10−16) and immunosuppressors (Person χ2 test with Yates' continuity correction, P = 2.2 × 10−08; hypergeometric test, P = 4.88 × 10−8) were found preferentially enriched in the mesenchymal subset. However, this preferential enrichment was not as evident in the comparison between the mesenchymal and neural subsets [immune activators (Pearson χ2 test with Yates' continuity correction, P = 0.2501; hypergeometric test, P = 0.8915) and immunosuppressors (Pearson χ2 test with Yates' continuity correction, P = 0.5736; hypergeometric test, P = 0.2847)].
Although 4 subtypes of glioblastoma were identified by Verhaak and colleagues (21) and 3 subsets by Phillips and colleagues (5), the proneural and mesenchymal subsets identified using distinct methodologies and sample sets seem to be the most robust and concordant (27, 28). Because the greatest immunologic diversity exists between the proneural and mesenchymal subset, we compared the immune genes of interest between these subsets. Many of the top tiers of preferentially expressed immune genes in the mesenchymal glioblastoma subtype relative to the proneural subtype have been shown previously to have a biologic role in glioblastoma (Table 1).
Immunosuppression predominates in mesenchymal glioblastomas
To determine if the list of genes generated from IPA and Uniprot was artificial or if it contained candidates that might correlate to the growth and maintenance of mesenchymal glioblastoma, we generated a list of known immunosuppressive genes associated with glioma biology and then surveyed the TCGA database for overexpression within subsets. To ascertain if there was a selective enrichment of the immunosuppressive genes among the overexpressed genes in the mesenchymal subset or if this was merely by chance, we compared the immunosuppressive gene set with the overall set of differentially expressed genes from the TCGA data set using the χ2 and hypergeometric tests and found selective enrichment of immunosuppressive genes within those overexpressed in the mesenchymal subset relative to the other subsets (P < 0.05), similar to the findings described earlier.
We parsed the TCGA dataset using the cBio cancer genomics portal analysis tool to show that the proportion of patients with glioblastoma in the mesenchymal subset had significantly higher (relative to the proneural subtype) mRNA expression of previously well-characterized glioma-mediated immunosuppressive cytokines such as galectin-3 (χ2 test; P < 0.0001), IL-10 (P < 0.0001), IL-23 (P = 0.0012), TGF-β (P < 0.0001), and the immune activation inhibitor PD-L1 (P < 0.0001) relative to expression in the proneural subtype (Table 2). Many of these immunosuppressive genes are associated with the monocyte and macrophage family including: colony-stimulating factor 1 (CSF-1; P < 0.0001), a cytokine that controls the production, differentiation, and function of macrophages; CCL2 (P < 0.0001) and CCL-22 (P = 0.0008), chemokines that attract monocytes; CD163 (P < 0.0001), a marker of immunosuppressive M2 macrophages; CD204 (P < 0.0001), a macrophage scavenger receptor; and macrophage inhibitory cytokine-1 (MIC-1; P < 0.0001). CD47, a block used by solid cancers to prevent phagocytosis (29), was not preferentially expressed in the glioblastoma subtype. Arginase, produced by myeloid-derived suppressor cells and an inhibitor of T-cell responses in the glioma microenvironment (30), is also enriched in the mesenchymal subset (P = 0.0248). Genes in the IL-6/STAT3 immunosuppressive signaling axis was preferentially expressed in the mesenchymal subset; they include: gp130 (P < 0.0001), part of the IL-6 receptor family; IL-6 (P < 0.0001); STAT3 (P = 0.0004); Pim1 (P < 0.0001), a STAT3 downstream-regulated proto-oncogene; and SOCS (P < 0.0001), a negative regulator of cytokine signaling.
Proinflammatory responses are more frequent in the mesenchymal glioblastoma subtype
We generated a list of key immune genes that reflect immune stimulation/effector responses of both adaptive and innate immunity. Again, we found a selective enrichment of the overexpressed immune effector genes in the mesenchymal subset relative to all other subsets (P < 0.05). Focusing on the differences between immune effector genes expressed between the mesenchymal and proneural subset, genes encoding key receptor interactions for T-cell activation, including CD3 (P < 0.0001), CD40 (P < 0.0001), CD80 (P ≤ 0.0001), CD86 (P < 0.0001), MHC HLA-B (P = 0.0001), MHC HLA-DRA (P < 0.0001), MHC HLA-DQA1 (P < 0.0001), and MHC HLA-DPB1 (P < 0.0001) and downstream signaling pathways reflective of T-cell activation, effector function, and immune-activating transcription factors, such as STAT4 (P = 0.0034), DNAM-1 (P < 0.0001), and IRF7 (P = 0.0088), were enhanced in the proportion of patients with glioblastoma with the mesenchymal phenotype (Table 3). Although there was a trend toward preferential expression of the adaptive proinflammatory cytokines in the proportion of patients with glioblastoma with the mesenchymal subtype, only IL-1 (P = 0.0007), IL-15 (P < 0.0001), and IL-7 (P < 0.0001) were significantly expressed at the mRNA level.
In contrast, genes associated with innate immunity, specifically natural killer (NK) cell markers such as NKp46, NKG2D, NKp30, NKp44, and KLRD1 were more uniformly distributed across the subtypes, with the exception of the NK-activating receptor CD244 (P < 0.0001) that was preferentially expressed in the proneural subtype. Other genes involved in innate immune response, such as Toll-like receptor (TLR) 2 (P < 0.0001) and 3 (P < 0.0001) were preferentially overexpressed in patients with glioblastoma with the mesenchymal subset. The triggering receptors expressed on myeloid cells (TREM) 1 (P < 0.0001) and 2 (P < 0.0001) that enhance monocyte/macrophage–inflammatory responses were also enhanced in the mesenchymal subset. Cumulatively, these data suggest that there is a preferential distribution of both proinflammatory and immunosuppressive genes within the mesenchymal subset.
Glioma antigens segregate differentially within glioblastoma subtypes
We postulated that the mesenchymal subset may have greater incidence of tumor-associated and -specific antigens that could contribute to the increased propensity of the immune effector genes present within the mesenchymal subset. Furthermore, there would be a predicted enhancement in immunosuppression to counteract the antitumor effector responses. Thus, to evaluate if tumor antigens are selectively enriched in the mesenchymal subset, a list of known glioma antigens was compiled, and the frequency of each mRNA overexpression was determined within the various glioblastoma subtypes (Table 4). Although there were predilections of specific tumor antigens for glioblastoma subtypes, there was not a preferential enrichment of antigens found within the mesenchymal subset. Specifically, the EGFR family of antigens including EGFR and ERBB2 were frequently found to be overexpressed at 80% and 24%, respectively, within the classical subtype, consistent with previous reports (21). Within the neural subtype, EGFR (54%) was frequently expressed, whereas within the proneural subtype survivin (32%) and Sart-1 (25%) were more commonly expressed. SART-2 (36%) was the most commonly expressed antigen within the mesenchymal subset. Thus, preferential antigen enrichment in the mesenchymal subset does not seem to be the underlying etiology for enrichment of the immune activator and suppressor genes within this subset.
The TCGA database consists of an unselected cellular population that includes glioma-infiltrating immune cells within the genetic composition. The analysis of the mRNA overexpression of immune genes may reflect either gene amplification (increased expression on the immune population) or a relative increase in a designated infiltrating immune population. The interaction between the immune system and molecular subtypes of glioma remains largely unstudied. Although a previous report identified an immune signature that is prognostic for survival in patients with glioblastoma, especially within the proneural subtype (31), this is the first report to show the concordant association of proinflammatory and immunosuppression in the mesenchymal subset. Specifically, patients with glioblastoma that have a preexisting induced anti-glioma effector response and an actionable immunosuppressive target (i.e., the mesenchymal subset) may be “immune reactive” and therefore particularly amenable to immunotherapeutic approaches including those targeting immune checkpoints. This is further supported by a retrospective analysis of patients with glioblastoma receiving dendritic cell immunotherapy who were more likely to have a mesenchymal subtype glioblastoma (20). These data indicate that the glioblastoma subtype may be a confounding variable in therapeutic response analysis and should be considered in stratification. In addition, our data indicate that glioblastoma subtype may influence the interpretation of posttreatment analysis of immune infiltration. Specifically, because glioblastomas are more likely to transition to the mesenchymal subtype upon recurrence (5), the intratumoral immune analysis after immunotherapy may be more reflective of the biology of the underlying subtype rather than a direct immunotherapy-induced response.
The proinflammatory and immunosuppressive gene sets derived from online-curated databases identified a set of common genes (e.g., TIMP1). In some cases, the role of these genes is contextual based on the disease state (i.e., autoimmunity vs. malignancy) or model system studied. For example, ANXA1 has been shown to both inhibit (32, 33) and enhance inflammation (34, 35). In other instances, that attribution to a particular category simply seemed erroneous (i.e., LTBP1 that activates the immunosuppressive cytokine TGF-β or TGFBR2 that encodes the TGF-β2 receptor). Thus, we selected a proinflammatory and immunosuppressive list of genes based on the documented roles of these genes in the setting of malignancy with a special emphasis on those operational in glioblastoma. This defined gene list also showed a preferential enrichment of both proinflammatory and immunosuppressive genes within the mesenchymal subtype. Although we had concerns about the selection of genes by the online-curated databases, they did provide novel insights into immune genes that had, thus far, not been appreciated as playing a role in glioblastoma immune biology. Furthermore, these databases revealed the marked diversity of the preferential enrichment of proinflammatory or immunosuppressive genes in the mesenchymal subtype relative to the others.
In many instances the findings of the immune genes associating with a particular glioblastoma subtype were consistent with and validated previous observations. For example, as STAT3 has been previously shown to be a key molecular hub driving the mesenchymal transformation (6), STAT3 and related downstream targets, such as Pim-1 and VEGF, were found to be enriched in the mesenchymal subtype relative to the other subsets. Furthermore, multiple monocyte genes were enriched within the mesenchymal subtype as reflected in the expression of CCL2 (a known monocyte chemotactic protein), CD163 (a marker of cells of the monocyte/macrophage lineage), and CD204 (a marker of macrophage scavenger receptors). These data are consistent with our findings in genetically engineered murine models in which the mesenchymal transition was shown to correlate with increased macrophage infiltration (7, 14). In addition, the proinflammatory T-cell cytokine IL-2 was minimally enriched in any subtype, consistent with our previous reports showing that while T cells may be active in the periphery, upon encountering the local tumor microenvironment, this immune response is markedly downregulated (3). Interestingly, markers of Tregs such as CD4, ICOS, IDO1, and CTLA-4 seemed to be preferentially enriched in the mesenchymal subset. Because robust immune effector response is also present in the mesenchymal subset, these patients may be specifically predisposed to respond to immunotherapeutics targeting the Treg population. However, while these markers have been shown to be important for Treg function and are associated with immune inhibition, they do not independently define the Treg population and mechanistic conclusions cannot be drawn from this type of analysis.
Although CCL22 (a chemokine that attracts Tregs; ref. 36), IDO and TGF-β (inducible Treg factors; ref. 37), and ICOS (critical for the functional stability of Tregs; ref. 38) were present in gliomas, no differential expression of FoxP3 was observed in glioblastoma subtypes. This seems inconsistent with the immunohistochemical data showing a great deal of heterogeneity in the presence of Tregs within glioblastomas (2). However, because the TCGA database is generated by cDNA microarray analysis, which represents gene regulation at the transcriptional level, immune genes that are regulated at the posttranscriptional, translational, or posttranslational level may not appear through this screening. A specific example of this is FoxP3, which is regulated either at the mRNA transcription level or at its protein stability level, which is degraded through the proteasome rapidly (39). Thus, the discrepancy between the immunohistochemical data and the TCGA data can not only be explained on the basis of FoxP3 mRNA and protein stability, but also shows a limitation of solely relying on analysis of mRNA data.
Another limitation of the current study is that the analysis does not directly reflect the systemic immune status or functional status of these immune responses. Some of the immunosuppressive targets may be markedly enhanced within the tumor microenvironment, with marginal elevations systemically, and agents without significant glioma penetration may fail to show a correlation of treatment response with tumor expression levels. Furthermore, the analysis of the frequency of overexpression is relative to the mean expression levels of a marker across all gliomas, and thus more clinical responders may be identified within any given phenotype because the threshold of minimal expression necessary to result in a therapeutic response has not yet been characterized. In addition, the mRNA expression levels analyzed in our study may not correspond to the protein expression levels especially in the case of B7-H1, which has been shown to be posttranscriptionally regulated by cytokines (40). Finally, although there is an association of immune effector and suppressor genes within the mesenchymal subtype, this has not yet been immunologically functionally validated. Not previously reported, but as part of a previous analysis, we found that 26% of newly diagnosed glioblastoma tumors have very little immune infiltration, whereas approximately 50% have both effector (CD8+ T cells) and suppressor (FoxP3+) immune responses in the glioblastoma (2). The correlation of these immunohistochemical findings with glioblastoma subtype is an area of future investigation along with the association of observed immune gene signatures and patient outcome.
Although the mesenchymal glioblastoma subtype may be “immunologically reactive,” as defined by the preexistence of immune effector responses and possessing the therapeutic targets of immunosuppressive modulators, it is unlikely that this is the sole subtype that can potentially benefit from immunotherapeutic strategies. We have previously shown that a peptide vaccine targeting the EGFR variant III (EGFRvIII) in newly diagnosed patients with glioblastoma significantly increased median survival time to more than 26 months (17, 18). On the basis of the TCGA data, it is likely that most of these patients have glioblastomas within the classical subtype because EGFRvIII expression almost always coincides with EGFR amplification (24, 41). This indicates that with specific immunotherapeutic strategies, patients who have distinct glioblastoma subtypes may preferentially benefit.
In conclusion, the analysis of these selected immune effector and suppressor genes may provide a way to programmatically prioritize different types of potentially competing immunotherapeutic strategies, and to identify subsets of patients that may respond to a particular strategy and thus selectively enrich for potential responders during early or small-scale clinical trials.
Disclosure of Potential Conflicts of Interest
M. Gilbert is a consultant/advisory board member of Merck, Genentech, EMD Serono, and Novartis. No potential conflicts of interest were disclosed by the other authors.
Conception and design: G. Rao, A.B. Heimberger
Development of methodology: T. Doucette, G. Rao, A. Rao, L. Shen, K. Aldape, J. Wei, A.B. Heimberger
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): G. Rao, L. Shen, K. Aldape, K. Dziurzynski, A.B. Heimberger
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): T. Doucette, G. Rao, A. Rao, L. Shen, K. Aldape, M. Gilbert, A.B. Heimberger
Writing, review, and/or revision of the manuscript: T. Doucette, G. Rao, A. Rao, L. Shen, K. Aldape, K. Dziurzynski, M. Gilbert, A.B. Heimberger
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A.B. Heimberger
Study supervision: A.B. Heimberger
Grant support was provided by the NIH RO1-CA1208113 (to A.B. Heimberger), 5P50 CA127001 (to T. Doucette, G. Rao, K. Aldape, and A.B. Heimberger), Dr. Marnie Rose Foundation (to A.B. Heimberger), the Cynthia and George Mitchell Foundation (to A.B. Heimberger), and the Vaughn Foundation (to A.B. Heimberger).
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
We thank Audria Patrick and David M. Wildrick for editorial assistance.
- Received March 25, 2013.
- Revision received May 2, 2013.
- Accepted May 15, 2013.
- ©2013 American Association for Cancer Research.