Adaptive Global Innovative Learning Environment for Glioblastoma: GBM AGILE

Traditional phase II and III trials include two arms in preset patient populations with preset sample sizes and address a single question. A small number of phase II trials have departed from this traditional design and seek to address multiple hypotheses within a single trial. Some include many experimental arms, adding and dropping arms over time (1). Others strive to match treatment arms with patient subtypes (including those defined by biomarkers), adaptively randomizing patients based on accumulating results of the trial (2), and adapting the sample size to the results (3–5).

In this article, we describe an inferentially seamless (6) phase II/III platform trial for glioblastoma (GBM). GBM Adaptive Global Innovative Learning Environment (GBM AGILE) is a two-stage, multi-arm, platform trial. Arms enter the trial, are compared with a common control arm for impact on survival, and leave the trial when their evaluation is complete. The initial stage uses adaptive randomization among the experimental arms within clinical and biomarker patient subtypes. This screening stage evaluates many therapies (including combinations) and identifies indications for each promising arm. Highly effective therapies move to a second stage designed to confirm that signal and indication in a small cohort of patients using fixed randomization versus control to enable registration. The GBM AGILE trial design offers the opportunity to accelerate delivery of improved therapies to trial participants, whereas the broadly defined eligibility criteria will leverage information learned from more patients. The seamless inferential design means that highly effective treatment arms proceed rapidly through the trial, enabling faster registration, regulatory review, and adoption for routine clinical care. Promising arms that do not meet criteria for the confirmatory stage exit the trial with a wealth of data to refine biomarker hypotheses and enable go/no go decisions outside of the trial.

GBM AGILE is also a novel clinical research network designed to speed the process of developing therapies for patients with rare diseases. It focuses the therapeutic development process around a specific disease, leverages the expertise of the research community, and optimizes the clinical testing for that population. The planning processes that shaped the trial and its international scope comprises the efforts of over 130 oncologists, statisticians, pathologists, neurosurgeons, imagers, and translational and basic scientists.

GBM is a deadly disease with few effective therapies. There were an estimated 22,810 cases of primary malignant brain tumors in United States in 2014, of which GBM is the most common type (7). According to the International Agency for Research on Cancer, there are more than 250,000 tumors of the central nervous system worldwide each year, and approximately 190,000 deaths (8). Patients with newly diagnosed GBM are treated with maximal safe surgical resection followed by radiation and temozolomide. Median survival time for patients with tumors harboring methylation of the DNA repair gene O6-methylguanine DNA methyltransferase (MGMT) is 23 months with a 5-year survival of 14% (9). Patients with tumors that have unmethylated MGMT promoters fare worse with a median survival time of 13 months and a 5-year survival of 8% (10). Once GBM recurs, there are currently no options with meaningful efficacy.

Despite numerous phase II and III clinical trials performed over several decades, only minimal advances have been made, and little has been learned. This contrasts with the substantial molecular information available for GBM due to large-scale genome sequencing projects such as The Cancer Genome Atlas (TCGA; ref. 11) and others (12). In parallel, many new therapies have been developed for clinical testing. These scientific advances lead to optimism that molecularly based precision medicine may improve outcomes for GBM patients but they also highlight the limitations of current clinical trial designs that do not test multiple therapies and biomarker combinations simultaneously.

One potential solution for testing multiple hypotheses within the same clinical trial is a multi-arm, Bayesian adaptively randomized platform trial (13–15). These trials may incorporate common control arms for meaningful endpoints, a fluid infrastructure for adding or dropping experimental arms, and an ability to use data as it is available during the trial to alter decision-making in a prespecified manner. The most notable example is I-SPY 2 in breast cancer (3, 16, 17). Using such a design to evaluate new therapies for GBM requires some changes but the overall concepts and goals may still be applied (18). Outcome adaptive randomization for GBM would be more efficient than balanced randomization, even when longer time to event endpoint such as overall survival (OS) is used (19), and such designs have been advocated by expert panels (20).

GBM AGILE is the first global, disease-specific, platform trial for GBM designed to specifically capitalize on the growing knowledge base from the molecular sciences, incorporate novel clinical trial innovations, and leverage the emerging global capabilities to undertake more innovative and complex trial protocols. The trial will be initiated in the United States and Australia, followed by China and potentially others to accelerate recruitment of large numbers of patients. As a consequence of both the scope and innovative crowd sourced design, GBM AGILE will create a learning environment to identify effective therapies and biomarkers for GBM. By including patients with both newly diagnosed and recurrent tumors and accounting for their presentation in a statistical model, the trial design will facilitate integration of knowledge that might have otherwise been disparate. Importantly, although GBM AGILE is designed to identify effective therapies and develop biomarkers for GBM, the overall process and philosophy could also be adapted for other rare cancers and diseases.

GBM AGILE has several important statistical innovations and is designed as a registration trial to accelerate availability of effective therapies and biomarkers for routine standard of care (SOC). As shown in Figs. 1 and 2, it is a Bayesian, adaptively randomized, multi-arm, platform trial. The primary endpoint is overall survival. GBM AGILE identifies and validates candidate biomarkers under a single platform master protocol. Experimental therapies can enter the trial at any time, accrual rate permitting. A therapy that is sufficiently promising in an adaptively randomized screening stage will move to a confirmatory stage with fixed randomization. Simulations ensure control of type I error to support registration.

Patients with a histopathologic diagnosis of GBM based on World Health Organization (WHO) criteria (21) will be eligible for GBM AGILE provided they are IDH R132H mutation negative by local IHC. IDH-mutant GBM has a sufficiently different genomic landscape and phenotypic behavior that the WHO has created separate classifications for IDH-mutant and IDH-wild-type GBM in the 2016 update. In particular, patients with either newly diagnosed or recurrent tumors will be included, regardless of MGMT promoter methylation status. Other standard clinical trial eligibility characteristics also apply.

GBM AGILE is also an efficient platform to explore the utility of various biomarkers. The trial will evaluate several kinds of biomarkers (22) as described below and summarized in Fig. 2. In contrast with the experience with other cancers, GBM has very limited well-defined a priori molecular biomarker subgroups with clinical utility (18). As stated above, IDH1 IHC will be used for diagnostic and eligibility purposes, whereas MGMT promoter methylation status will be used as a “stratification” variable to help assign patient subtype (Figs. 2 and 3). Clinical presentation as either newly diagnosed or recurrent disease (RD) will serve as the other stratification variable in addition to MGMT promoter methylation status. Stratification variables define three “subtypes” of GBM: newly diagnosed methylated (NDM), newly diagnosed unmethylated (NDU), and RD.

Some experimental arms will enter the trial with an associated “enrichment biomarker” identified through central testing that is specific to the arm. Enrichment biomarkers are hypothesized predictive biomarkers for the experimental agent. For that reason, enrichment biomarkers are “context dependent” and only considered with respect to the associated experimental arm; other arms are evaluated irrespective of that arm's biomarker. For example, an EGFR inhibitor might enter the trial with a proposed enrichment biomarker of EGFR mutation or amplification identified through next-generation sequencing. Testing for these alterations as enrichment biomarkers would continue as long as the EGFR inhibitor was being evaluated on the trial and be considered only for that arm. Enrichment biomarkers that define indications for effective therapies become “stratification” biomarkers when that new therapy becomes part of a new SOC. Each experimental arm can have at most one a priori defined enrichment biomarker. There may be other biomarkers that are better at finding responders that are unknown at the time of entry onto the trial, however. The wealth of biomarker data generated as part of GBM AGILE will therefore be a valuable resource for retrospective exploratory analyses to identify such biomarkers.

The addition of enrichment biomarkers to the stratification markers doubles the possible subtypes relative to the relevant experimental agent—NDM, NDU, and RD each have biomarker positive and biomarker negative subtypes (Fig. 3). “Subtypes” are characteristics specific to the tumor or patient; each patient belongs to one and only one subtype relative to a given experimental arm. This should be contrasted with biomarker “signatures,” described below, which are therapy-specific characteristics.

For patients with RD, contemporary tissue for biomarker assessment will not initially be required. There will be a subset of patients with RD that do have re-resection prior to enrollment, however. For such patients, the most contemporary tissue will be used for biomarker analysis. Furthermore, biomarker subgroups in such patients will be compared to tissue from the original diagnosis to analyze the stability of biomarker subgroups over time. If there is evidence of relevant biomarker subgroup change due to selection, requirements for contemporary tissue analysis will be revisited.

The three subtypes defined by stratification biomarkers have different SOC control arms and different ways that experimental arms are comprised. Control arms for the different subtypes are radiotherapy and temozolomide for NDM and NDU, and lomustine (CCNU) for RD. These controls may be updated through amendment if SOC changes over time. Experimental arms for NDM use radiotherapy and temozolomide as a backbone and add the experimental agent or combination whereas those for NDU may omit temozolomide due to limited efficacy in this population. For RD, the experimental agent or combination can be combined with CCNU or considered alone.

Control therapy is assigned to 20% of the patients within all patient subtypes throughout the trial. Experimental arms are compared against control therapy, and randomization probabilities are assigned accordingly within subtypes. Experimental therapies are assigned in proportion to their (Bayesian) probabilities of prolonging survival longer than control. Initially, randomization probabilities are equal. These probabilities are updated monthly based on the outcomes available in the trial at the time. Although overall survival (OS) was chosen as the primary endpoint for GBM AGILE, adaptive randomization is not reliant on OS in general. Other endpoints, such as progression-free survival or response rate, could also have been used to inform adaptive randomization. With faster time to event, such endpoints may lead to more efficiency gains. But there also exists the potential that a treatment might positively impact these endpoints without impacting survival. Because of this potential, GBM AGILE will initially use OS to inform adaptive randomization but leaves open the possibility of using earlier data through the longitudinal model.

During each experimental arm's screening stage, its performance in comparison with control will be prospectively evaluated in predefined “signatures.” Signatures are groupings of the stratification and enrichment biomarker defined subtypes that are potential indications for the experimental therapy. Example signatures include “all newly diagnosed patients” or “patients with EGFR-positive recurrent disease.” There are 10 possible signatures for an arm with an enrichment biomarker and five possible signatures for arms without an enrichment biomarker. In contrast to subtypes, each patient or tumor belongs to multiple possible signatures, but effective therapies will graduate with only one signature. The signature is the biomarker-defined group for which there is the best chance of success.

An experimental arm that is performing sufficiently well during its initial stage will “graduate.” Graduation signals a seamless move into the arm's confirmatory stage within its graduating signature. During an arm's confirmation stage, it will be randomly assigned to a fixed proportion (40%) of the patients within its graduating signature, up to a maximum of 50 patients. The final and primary analysis of all experimental arms, whether or not they have a confirmatory stage, will be a comparison of the primary end point (OS) against control. The primary analyses will be the arm's (Bayesian) probabilities of superiority over control for each of the arm's signatures. All controls accrued to the trial up until the time the last patient was accrued to the experimental arm in question will be used in this comparison via a time-adjusted, covariate-adjusted, and arm-adjusted analysis that utilizes the results of all patients assigned to all arms in the trial.

The number of experimental arms will vary as arms are added or removed due to graduation or futility. An arm can be added to the trial at any time after it is approved by the relevant committees on GBM AGILE, provided the patient accrual rate is sufficient. Arms that do not graduate may still be worth further study. In those cases, data from GBM AGILE will be invaluable in making further go/no-go decisions, effectively powering follow up trials, and determining the value of biomarkers for eligibility decisions.

Biomarkers to assess response and monitor patients will also be collected. These may be in the form of pharmacodynamic (PD) or response biomarkers (22) that may factor into trial conduct if there is a potential association with OS through the longitudinal model (described below). PD/response biomarkers that show associations with treatment effects on OS may also generate data to support development along the surrogate endpoint hierarchy (i.e., reasonably likely to be validated; ref. 22). This is an example of how GBM AGILE's platform structure with continuous learning may be used to develop other biomarkers and improve the overall development process in addition to those associated with specific therapies.

Although there are no restrictions on the types of assessments that may be used as response biomarkers, imaging will play a major role. The Jumpstarting Brain Tumor Drug Development Coalition Imaging Standardization Steering Committee has undertaken an extensive effort to standardize MRI protocols for multicenter studies to maximize the potential of MR imaging techniques as both pre and post-treatment biomarkers (23).

The primary endpoint for GBM AGILE is OS. However, patients in the trial for the same length of time may have different future life expectancies. We are building a longitudinal model that will take each patient's current circumstances into account in predicting time of death. This model will be developed in coordination with regulators and will be incorporated via protocol amendment after the trial starts enrollment. Factors in the model include measurements over time using MRI, the patient's performance status, and importantly, the treatment arm. For example, immune-based therapy may have little effect on measurable tumor burden but still prolong survival. For such an arm, the model will learn that MRI measurements offer little help in predicting the patient's time of death. The various parameters in the model will have probability distributions that will be updated via Bayes' rule as OS information becomes available in the trial and potentially be utilized by the randomization algorithm for additional efficiency (24).

GBM AGILE will plan to add therapies and associated biomarkers during the course of the trial. Identification of robust data supporting these treatments and biomarkers is paramount. Potential experimental arms and associated enrichment biomarkers can be identified by investigators within GBM AGILE or proposed by outside investigators. These therapies and biomarkers are then prioritized and reviewed by the various GBM AGILE committees prior to inclusion. The treatment and biomarker selection processes will accord high priority to timely communication and transparency. Decisions for inclusion of potential therapies in the trial will be made on the quality of the science and the readiness for phase II testing.

GBM AGILE is a major departure from standard clinical trials. Several innovations are common to other platform trials: adding and dropping arms, adaptive randomization within biomarker-defined subgroups, and the ability to address multiple hypotheses in a single trial protocol. GBM AGILE takes these innovations a step further by including a seamless transition to a second confirmatory stage to enable registration. This could potentially cut years from the drug development process and substantially reduce cost. Even arms that do not progress to a confirmatory stage may generate valuable data to refine biomarker hypotheses and inform better decision making for trials outside of GBM AGILE. Including both newly diagnosed and recurrent patients and having an ongoing platform structure also enables more patients to participate. This results in more opportunities to learn from those who develop this deadly tumor and to offer better treatment options. These factors and an environment that fosters collaboration and innovation make GBM AGILE a model for the future development of new therapies for rare diseases.

D.A. Berry holds ownership interest (including patents) in Berry Consultants, LLC. S.M. Chang reports receiving commercial research grants from Novartis. T.F. Cloughesy is a consultant/advisory board member for AbbVie, Agios, Alexion, Boston Biomedical, Bristol Myers-Squibb, Celldex, Cortice, Cytrx, Human Longevity, Insys, Medqia, Merck, Newgen, Notable Labs, Novocure, Novogen, Oxigene, Pfizer, Pronai, Roche, Sunovion, Tocagen, Upshire Smith, and VBL. M. Khasraw reports receiving other commercial research support from AbbVie and Specialised Therapeutics Australia and is a consultant/advisory board member for AbbVie, Bristol Myers-Squibb, and Eli Lilly. G.H. Poste is a consultant/advisory board member for Caris Life Sciences. P.Y. Wen is a consultant/advisory board member for AstraZeneca, Aurora Biopharma, Cavion, Genentech/Roche, Insys, Monteris, Novartis, Novogen, Vascular Biogenics, and VBI Vaccines. W.K.A. Yung reports receiving speakers bureau honoraria from DNAtrix and Merck and is a consultant/advisory board member for DNAtrix. A.D. Barker is a consultant/advisory board member for Caris Life Sciences. No potential conflicts of interest were disclosed by the other authors.

The GBM AGILE Network includes the following members: P. David Adelson, Brian M. Alexander, Joe Alper, Michelle M. Arnold, David F. Arons, David N. Ashley, Sujuan Ba, Anna D. Barker, Mitchel S. Berger, Donald A. Berry, Jerrold L. Boxerman, Daniel J. Brat, Cameron W. Brennan, Michael Buckland, Kenneth Buetow, Meredith Buxton, Lewis C. Cantley, Webster K. Cavenee, Susan M. Chang, Ling Chen, Lynda Chin, E. Antonio Chiocca, Timothy F. Cloughesy, Darrel P. Cohen, Howard Colman, Carolyn Compton, Jason Connor, Laurence James-Neil Cooper, Vladimir Coric, Joseph F. Costello, John F. de Groot, Jayesh Desai, Giulio Draetta, Benjamin M. Ellingson, Laura Jean Esserman, Howard A. Fine, Evanthia Galanis, Hui Gan, Val Gebski, Elizabeth R. Gerstner, Robert Gillies, Jonathan G. Goldin, Charles Goldthwaite, Federico Goodsaid, Todd Graves, Steven Gutman, Wendy Hague, James R. Heath, Amy B. Heimberger, Jonathan Hirsch, Kyle Holen, Jason T. Huse, Nola Hylton, Tao Jiang, Mustafa Khasraw, Alec Kimmelman, Robert Latek, Sean Leong, Wenbin Li, Ping Li, Linda Liau, Michael Lim, David N. Louis, Pedro Lowenstein, Wenbin Ma, Mark Maclean, Carlo Maley, Zhiqi Mao, Qing Mao, Ingo K. Mellinghoff, Tom Mikkelsen, Paul S. Mischel, Robert Mittman, Sarah J. Nelson, Hideho Okada, Melissa C. Paoloni, Luis Parada, Tom Parke, David R. Parkinson, Kristyn Pineda, Whitney B. Pope, George H. Poste, Michael D. Prados, Xiaoguang Qiu, John Quackenbush, David A. Reardon, Gregory J. Riggins, Mark Rosenthal, John H. Sampson, Jann N. Sarkaria, Trevor Saxman, R. John Simes, David Spetzler, Catherine Stace, Michelle Stewart, Robert Strausberg, Daniel C. Sullivan, Erik P. Sulman, Kristin Swanson, Donald E. Thornton, Carrie Treadwell, Laura van 't Veer, Scott R. Vandenberg, Roeland Verhaak, Inder Verma, Max Wallace, Renzhi Wang, Xiang Timothy Wang, Xinghe Wang, Yi Michael Wang, Yu Wang, Anthony Weeks, Michael Weller, Patrick Y. Wen, Forest White, Wolfgang Wick, Otmar Wiestler, Nicole Willmarth, Elizabeth M. Wilson, Benjamin Winograd, Ji Xiong, Hai Yan, Yu Yao, Mao Ying, Xinguang Yu, W. K. Alfred Yung, Wei Zhang, Xin Vincent Zhang, Zhenyu Zhao, and Fan Zhao.

This work was supported by the National Biomarker Development Alliance, the Cure Brain Cancer Foundation, the National Foundation for Cancer Research, the National Brain Tumor Society, a Burroughs Wellcome Innovations in Regulatory Science Award, and anonymous patient advocate donors.

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.