Bayesian adaptive trial designs for neoantigen based immunotherapy and borrowing strength across subpopulations within the trial and from external controls
CERSI Collaborators: Joseph Ross, MD, MHS, , Molly Jeffery, PhD, Jun Yin, PhD (formerly at Mayo Clinic, now at Moffitt Cancer Center) (PI)
FDA Collaborators: Adnan Jaigirdar, MD, Pourab Roy, PhD, Rajeshwari Sridhara, PhD, Zhenzhen Xu, PhD, Rebekah Zinn, PhD
Project Start Date: October 2020
Project End: June 2026
Regulatory Science Challenge
Cancer immunotherapy has changed the landscape of modern oncology. Immune checkpoint inhibitors have emerged as an effective form of immunotherapy, and the development of novel therapies targeting tumor-specific antigens (or neo-antigens) has been an active area of investigation. Many cancer immunotherapy development programs use a basket trial/master protocol strategy where a single investigational agent (or drug combination) is tested on multiple cancer populations as defined by specific characteristics such as disease state, histology, or biomarkers (e.g., neoantigens). These basket trials typically use single arm designs and may be prone to selection bias and confounding due to the lack of randomization.
The focus of this research was to create a Bayesian statistical method to enable faster identification of promising immunotherapies to potentially accelerate the development of safe and effective treatments for patients with cancer. Bayesian approaches allow incorporation of accumulating evidence from the ongoing trial and sources external to the trial. PDUFA VI identified developing novel clinical trial designs (including Bayesian methodologies) as a priority area for regulatory science research. This study developed a Bayesian model that borrows information from the different single arms /disease populations within a master protocol to efficiently identify promising drugs for further development by classifying active and inactive response clusters of patients. .
Project Description and Goals
This study developed a mathematical model that will make sequential adaptive subgroup-specific decisions while clustering subtypes that have similar responses to treatment. The model was evaluated through simulation and by analyzing data from real basket trials available at the Mayo Clinic. This research also tried to minimize the potential bias/confounding in a basket trial due to the lack of randomization by using external controls (generated using matching techniques such as propensity score analysis). The synthesized datasets were compared to determine the treatment effects both through simulations and by utilizing available clinical trial data.
Research Results
We found that statistical methods that combine information across related patient groups can improve the efficiency of basket trials used in cancer research. By jointly analyzing two measures of treatment activity—tumor response and an early biomarker—we showed that trials can detect promising therapies earlier without increasing the risk of false-positive results. Our results also showed that using carefully selected external control data from previous clinical trials or real-world data can help strengthen comparisons and reduce the number of patients needed in new trials. Simulation studies and analyses of completed clinical trials demonstrated that these approaches can maintain reliable statistical performance while improving the ability to identify effective treatments. Overall, the findings suggest that these methods may help accelerate the evaluation of new cancer therapies while maintaining rigorous scientific and regulatory standards.
Research Outcomes
Research Outcome 1 I. Foundational Requirements for Regulatory Science Research Projects
This project developed a statistical method that jointly analyzes two measures of treatment effectiveness in cancer basket trials: tumor response and an early biomarker of treatment activity. The framework allows information to be shared across related patient groups within the trial while evaluating both outcomes together. By modeling the outcomes jointly, the approach improves the efficiency of the analysis and allows earlier evaluation of treatment activity without increasing the risk of false-positive findings. These methods can help researchers design more efficient basket trials and may support better evaluation of targeted cancer therapies in biomarker-defined patient populations.
Research Outcome 2 IV. Inform Regulatory Decision Making
This project evaluated statistical methods for using external control data to support clinical trial analyses. We studied how data from previously completed clinical trials and real-world sources, such as electronic health records, can be used as comparison groups when a randomized control arm is limited or unavailable. Using case studies from completed oncology trials and extensive simulation studies, we examined when borrowing information from external controls can improve statistical efficiency while maintaining reliable results. The findings provide practical evidence and methodological guidance that may help regulators and researchers evaluate when external control data can be appropriately used in clinical trials.