Leveraging complicated PK/PD models for the development of a Bayesian adaptive Dose-ranging design.
B. Boulanger, A. Jullion and M. Lovern
Pharmacometrics, UCB Pharma, Belgium
Background: When designing a Dose-ranging study, a common approach is to envisage the use of a Bayesian dose adaptive design, where the dose will progressively be adapted as patients are enrolled, to optimize a specific efficacy end-point and the confidence on the dose that gives the minimal required efficacy. Frequently, adaptive trial designs have been based upon simulations from assumption-rich, simplistic models relating dose to one or more efficacy end-points to optimally select the rules of adaptation, such as allocation and stopping rules. Our objective is to show how to leverage a sophisticated PK/PD model, relating dose to exposure (PK) and exposure to response (PD), including covariate effects. In so doing, we hope to obviate the limitations inherent in empirical dose-response models and allow greater flexibility in exploring alternative trial design scenarios.
Methods: The method envisaged here can be decomposed into two steps. First, using Pharsight Trial Simulator (TS2), we simulate thousands of "virtual" patients, at a wide range of doses, relying on the PK/PD model. Those "virtual" patients are stored into a database. By changing the parameters of the PD component of the model, various scenarios, from no efficacy to high efficacy as a function of dose are envisaged. Second, a cohort-based adaptive design is established where 20 patients in each cohort were allocated to placebo and up to four doses. Patients are allocated in order to minimize the variance of the smallest dose that gives an expected proportion of 80% of patients with an efficacy score improved by at least 50% at 2 weeks. To relate the efficacy end-points to the Dose, a Bayesian Normal Dynamic Linear Model (NDLM) model, implemented in Winbugs, is considered. The adaptive design is then intensively simulated by drawing, with replacement, virtual patients within the data base created by TS2.
Conclusions: The proposed approach is found meaningful and relevant, in comparison with the traditional approach based on empirical models that relate dose to efficacy response, because it integrates all the richness and uncertainty up to the present stage of development. This alleviates the need to make strong assumptions for the model that relates dose and response. In addition, as far as PK/PD models enable the use of appropriate covariates, it's even possible to investigate and optimize the population of interest to be included into an adaptive study.