GBM AGILE Phase II/III Bayesian Randomized Platform Trial

Key Takeaways

• In GBM AGILE, regorafenib did not improve overall survival versus control in either newly diagnosed unmethylated or recurrent glioblastoma and enrollment stopped early for limited efficacy.
• The trial used a phase II/III Bayesian adaptive platform design with a common control arm, prospectively defined subtype signatures, and monthly interim decision-making.
• Regorafenib was associated with increased toxicity relative to control, although the negative arm also illustrated the operating features of the GBM AGILE platform.

In the regorafenib report, investigators assessed the drug as the first investigational arm in the platform after prior phase II data had suggested an overall survival benefit in recurrent disease. The study included both newly diagnosed unmethylated glioblastoma and recurrent glioblastoma, allowing the drug to be tested across predefined patient subtypes within a shared trial structure.

The regorafenib arm enrolled patients with 1 of 2 trial-defined subtypes: newly diagnosed unmethylated glioblastoma and recurrent glioblastoma. Prospectively defined arm signatures were newly diagnosed unmethylated, recurrent disease, and the combined population of both groups. Because regorafenib was the first investigational arm in GBM AGILE, it was equally randomized against control. In newly diagnosed unmethylated disease, the control regimen was temozolomide plus radiotherapy; in recurrent disease, the control regimen was lomustine. Efficacy was evaluated using the hazard ratio for overall survival, comparing investigational arm with control.

The platform’s decision framework was explicitly Bayesian and adaptive. Benefit was to be declared if the Bayesian probability that the hazard ratio was less than 1.00 reached at least 98%. Monthly analyses also assessed predictive power. If Bayesian predictive power fell below 25% across all prespecified signatures, the arm stopped enrolling for limited efficacy. Follow-up then continued for 12 months after accrual stopped. This design was intended to allow earlier decisions than a fixed traditional trial while preserving registration intent.

The regorafenib arm did not meet efficacy criteria. At the August 2021 interim evaluation, predictive power fell below 0.25 in all 3 signatures, and enrollment stopped for limited efficacy. Final intention-to-treat sample sizes were 176 patients assigned to regorafenib and 179 assigned to control. In the final analysis, regorafenib did not improve overall survival in either subtype. Median hazard ratios were 1.05 for newly diagnosed unmethylated disease, 1.07 for recurrent disease, and 1.07 in the combined population. Final probabilities of benefit, defined as a hazard ratio below 1.00, were 0.421 in newly diagnosed unmethylated disease, 0.312 in recurrent disease, and 0.296 overall, all well below the prespecified threshold for success.

The subgroup-level survival findings were similarly negative. In newly diagnosed unmethylated glioblastoma, Bayesian-modeled median overall survival was 14.28 months with regorafenib versus 14.56 months with control, with a hazard ratio of 1.05. In recurrent glioblastoma, modeled median overall survival was 9.20 months with regorafenib versus 9.69 months with control, with a hazard ratio of 1.07. A modified intention-to-treat analysis was consistent with the intention-to-treat results and also did not show an overall survival benefit.

Safety favored control. The investigators reported no new safety signals with regorafenib, but treatment-related toxicities were higher with regorafenib than with control. In recurrent disease, grade 3 or 4 treatment-emergent adverse events occurred in 45% of regorafenib-treated patients versus 32% of controls. In newly diagnosed unmethylated disease, the corresponding rates were 74% with regorafenib and 30% with control. Treatment-related serious adverse events, dose interruptions, dose reductions, and discontinuations were also generally more frequent with regorafenib. The most common treatment-emergent adverse events with regorafenib included fatigue, headache, hypertension, nausea, constipation, increased ALT, and palmar-plantar erythrodysesthesia, varying somewhat by disease setting.

The authors concluded that GBM AGILE did not show superiority of regorafenib over control in either recurrent glioblastoma or newly diagnosed unmethylated glioblastoma and that regorafenib caused increased toxicities. They also noted that regorafenib has been removed from NCCN guidelines as a treatment option for recurrent disease after failure to confirm the earlier REGOMA signal. At the same time, the regorafenib arm was presented as a demonstration of the GBM AGILE platform’s operating model: shared controls, adaptive randomization, predefined signatures, and Bayesian stopping rules allowed the trial to stop this ineffective arm earlier and continue evaluating other therapies within the ongoing platform.