Bayesian population PBPK approach for support of drug development

Low likelihood-of-approval rates of new drugs constitute a major problem in clinical development. Only one out of ten development programs entering the first clinical phase succeeds in being approved by the U.S. Food and Drug Administration (FDA) [1]. A main challenge is thereby an insufficient understanding and prediction of drug safety and efficacy, leading to the withdrawal of new drug candidates [2,3]. Here, model-based assessment of drug exposure and response can support the development process at all stages, starting from early preclinical to late clinical phases [4,5]. Thus, the quantification of interindividual variability in clinical outcomes and the identification of related sources of such variability are of utmost importance e.g. for individualized dosing strategies [6–8]. Furthermore, of particular interest are translational approaches that transfer and integrate knowledge of recent study programs or earlier steps of drug development in order to make improved conclusions about drug behavior in clinically-relevant populations [3,9]. In this thesis, we present a Bayesian population physiologically-based pharmacokinetic (PBPK) approach for assessment of interindividual variability and clinical translation. Therein, we combine large-scale mechanistic PBPK models describing the behavior of drugs within the body with a Bayesian statistical framework for efficient estimation of the parameter space. The parameter space consists of clearly deconvoluted physiological- and drug-specific parameters, which facilitates the use of large amounts of prior information about the parameters. Such prior knowledge is updated with information extracted from experimental data by considering the Bayesian theorem and in particular performing a Markov chain Monte Carlo (MCMC) approach. This allows to solve the inverse and strongly ill-posed parameter estimation problem and at the same time preserve the extrapolation capabilities of PBPK models. Our Bayesian population PBPK approach represents a specifically-designed workflow for whole-body PBPK models to take into account the distinct properties of such models and guarantee a generic form for broad applicability and to support translation of knowledge. The overall framework contains i.a. an hierarchical model to separate individual uncertainty about the parameters from population variability, and a covariate model to cope for systematic relationships of model parameters to age, gender and body height. We further provide a method for estimation of the a posteriori parameter dependency structure at the population level. A blockwise MCMC sampling structure reduces complexity and accounts for the different types of parameters that are estimated. Additionally, we present an adaptive sampling method that combines gradient-based sampling with continuous adaptation of the proposal scaling for a strongly improved performance of the MCMC approach compared to standard methods. Moreover, we establish a translational learning workflow, where our Bayesian population PBPK approach is iteratively conducted in several learning steps to finally predict an unsupervised scenario, e.g. the pharmacokinetic behavior of a diseased population after administration of a new drug candidate.Subsequently, three application examples represent how to support different phases of drug development by the developed workflow. In the first example we successfully identify clinically-relevant subgroups in a cohort of individuals. These findings can improve safety and efficacy assessment in a clinical phase I study. In the second example we determine the interindividual variability in the pharmacokinetic behavior and the underlying physiological parameters, and improve population simulations by adding estimated information about parameter dependencies. We further reveal the performance of our new adaptive MCMC approach. In the third example we predict the pharmacokinetic behavior in a cohort of patients after accumulation of available study data in three iterations of our Bayesian population PBPK approach. We here successfully demonstrate the concept of translational learning from a phase I to a phase II study, taking into account the derived pathophysiology of the population. Overall, these examples indicate the capabilities of our approach in accumulation of knowledge and extrapolation of drug behavior. Applied to drug development programs, our method could improve clinical trial design to increase the benefit/risk ratio of new compounds. Our model-based concept could hence give significant support to raise approval rates of new drugs in the future