BAR12: Bayesian Autoregressive Phase 1-2 Design for Cell Therapy Trials With Manufacturing Changes

An early phase dose-finding trial of a new cell therapy may involve one or more manufacturing modifications made during the trial, known as "tweaks," to improve the cell product quality. For example, a tweak may change the cell culture duration, cytokine cocktail, or donor criteria. Ideally, a tweak can be done without changing the treatment so much that the trial must be restarted as if the treatment were entirely new. However, a statistical design still should account for changes in the dose-response distribution due to the tweak. For such settings, we propose a Bayesian AutoRegressive Phase 1-2 (BAR12) design that accounts for manufacturing tweaks during a phase 1-2 trial by using a first order autoregressive model with spike-and-slab priors having components corresponding to pre- and post-tweak distributions of toxicity and efficacy parameters. Simulations under a broad range of dose-response functions were conducted to compare BAR12 to conventional phase 1-2 designs that either assume the tweak had no effect and use all available data, or ignore the pre-tweak data. The simulations show that BAR12 has superior operating characteristics, including higher probabilities of correct dose selection, allocation of more patients to optimal doses, and more efficient monitoring for identifying unsafe or ineffective doses.