Blog by: Kert Viele
News from the Precision Promise platform trial.
Precision Promise is an adaptive platform trial investigating potential therapies for pancreatic cancer, an indication with very few options available. Given this history, we may have a low expectations apriori for any potential new therapy in this area. When looking for a new therapy under these conditions, we want to explore as many potential therapies as possible. A platform trial does this naturally, exploring multiple arms simultaneously, stopping therapies which are not sufficiently promising quickly, and replacing those dropped therapies with new potential options. By exploring many arms quickly, we minimize our time to finding an effective therapy. For more see this paper by Saville and Berry on efficiencies of platform trials.
Therapies in Precision Promise proceed in stages. Data is obtained and therapies must be sufficiently promising to advance. The press release notes pamrevlumab passed stage 1 by obtaining at least a 35% predictive probability of eventually demonstrating superiority on overall survival. We don’t know whether that predictive probability was 35.01% or 99.99%, but it had to be over 35%.
Predictive probabilities incorporate both the sampling variability in future data (the additional 75 patients described in the press release have some survival distribution) and the uncertainty in the parameter value (here the hazard ratio). They differ from conditional power in that conditional power typically assumes a parameter value, often the currently observed value, and only incorporates the sampling variability. Predictive probabilities, in contrast, add the uncertainty about the parameter and can be thought of, mathematically, as conditional power averaged over the current Bayesian posterior distribution. For more information see these two papers (one more technical, the other a JAMA Guide to Statistics and Methods).
The predictive probability threshold of 35% doesn’t guarantee success of course, but is viewed as sufficiently promising to trigger an additional investment of resources. Precision Promise is a seamless 2/3 platform, and this second stage collects enough patients to allow confirmatory level evidence for regulatory bodies. This trial has been through regulatory review for this purpose, and includes various features such as including both the stage 1 and stage 2 data in the final analysis as well as a model incorporating the totality of the data in Precision Promise, including all arms of the study and all controls. This model is referred to as a “Bayesian Time Machine”
We will find out the final results in due time. We hope the therapy is effective for the benefit of patients with pancreatic cancer, and hope this experience continues to help us learn how to investigate therapies both rigorously and efficiently. While we learn about oncology, we also “experiment with experimentation” to accelerate learning for pancreatic cancers and many other indications.