Molecular predictors of the mechanism of chemotherapy failure in malignant gliomas

Primary glioblastoma, the most-commonly diagnosed brain tumor, remains nearuniformly fatal regardless of treatment. With the recent in-depth molecular characterization of glioblastoma by the NIH Cancer Genome Atlas (TCGA) and others, these tumors can now be robustly categorized amongst distinct molecular subclasses (1-4). However, the clinical impact of this upfront classification is yet unknown. At the same time, a massively elevated mutation frequency has been uncovered within recurrent glioblastomas by large-scale genome sequencing (5-7). These tumor recurrences evade chemotherapy through somatic inactivation of the DNA mismatch repair (MMR) pathway (8-10). Intriguingly, detailed examination of TCGA datasets reveals that hypermutator recurrences arise uniformly from a specific subclass of primary glioblastomas. This preliminary observation raises the possibility that a tumor's response to treatment could be predicted in part based upon pre-treatment molecular classifiers, and holds the promise for upfront individualized therapy to close off potential treatment escape pathways. Herein, we propose to analyze the relationship between treatment response, and upfront versus recurrent tumor-specific characteristics, to better understand the development of therapeutic resistance.

As the most common primary human brain tumor, glioblastomas are associated with a median survival of less than two years (11). The standard-of-care treatment is surgical resection followed by radiation therapy and chemotherapy with the DNA-alkylating agent temozolomide. Prospective randomized clinical trials have shown that this combined adjuvant regimen results in a statistically-significant prolongation of life (12, 13). Nonetheless, virtually all patients suffer tumor regrowth that is resistant to prior therapies and ultimately leads to the patient's death.

Glioblastoma has been an important focus of the recent National Cancer Institute's TCGA. This critical infrastructure investment has led to a significant advance in our knowledge of the fundamental genomic alterations in this cancer. As a result, we now have detailed information about a broad study sample of glioblastomas prior to initial therapy, and subsequently when they recur after treatment failure. These resources provide the basis for the identification of the genetic mechanisms which link these two scenarios (upfront presentation and recurrence) in individual patients.

Amongst the key advantages of TCGA datasets are the multi-platform, multidimensional molecular analyses that are performed across the common core tumor sample set. The rationale for this design was to provide a framework to allow for the generation of hypotheses that would not otherwise be apparent from isolated data collections. Ongoing large-scale characterization of glioblastoma gene expression profiles has categorized the TCGA tumors into different subclasses within a proneural mesenchymal axis (2). However, while there is good evidence demonstrating a distinct genetic identity between subclasses (2, 16), there is controversy whether clinicallyrelevant differences exist between the groupings that cannot be explained by known differences between grade and histological categorization (17, 18).

At the same time, large-scale genome sequencing of recurrent glioblastomas by multiple groups (including our own) has discovered a hypermutation phenotype from somatic MMR pathway inactivation, linked to alkylator chemotherapeutic resistance and treatment failure (5-10). Approximately one-quarter of glioblastomas recurrent in patients receiving temozolomide have somatic hypermutation across the genome. The local sequence context of this hypermutation is indicative of a mutator phenotype arising during treatment. Furthermore, somatic inactivating mutations of the MMR genes MSH6, MSH2, and MLH1 are found in these recurrences, but not in any untreated tumors (5, 6, 8, 10).

Interestingly, the earliest rounds of TCGA data acquisition inadvertently included posttreatment recurrent tumors in the analyses. This fortuitous oversight provided the unique opportunity to carefully analyze this abundance of large-scale sequencing information with regards to the timing of MMR inactivation, firmly establishing the finding of MMR inactivation in a significant proportion of alkylator treatment failures. Interestingly, these observations in human tumors mirror studies from earlier decades which identified MMR pathway inactivation as a mediator of alkylator resistance in vitro (15). Accordingly, MMR-deficiency serves as an in vivo route of emergent alkylator resistance in these patients (10). This common molecular defect driving chemoresistance in human tumors in vivo represents an exciting opportunity for the design of improved treatment strategies to exploit this defective pathway.

With robust molecular sub-classifiers now available for both upfront primary glioblastomas and the subsequent emergence of treatment resistance, it is crucial to now understand the relationship between these groups in individual patients, as well as the relationship between sub-class and clinically-meaningful outcome measures. A strong association with treatment outcome would have significant translational impact, and provide a basis for the promise of personalized therapy. In our analyses of TCGA hypermutator datasets, we have found that the recurrences examined from primary glioblastomas seem to arise uniformly from the mesenchymal subclass. While this observation is preliminary and needs to be confirmed in more samples, it raises the possibility that a tumor's response to treatment can be predicted upfront. Indeed, if the mechanism of treatment failure for a given glioblastoma subclass can be predicted to be MMR-deficiency, that treatment escape pathway could potentially be closed off with targeted salvage therapies or an upfront combination regimen. We therefore hypothesize that it may be possible to identify treatment-related factors that differ between subclasses amongst otherwise histologically-indistinguishable primary glioblastomas. To test this hypothesis, we propose to assess treatment response to both standard-of-care and experimental salvage therapies, and overall outcomes, within the different sub-classes of both de novo primary glioblastomas and post-treatment recurrences.