Logo: to the web site of Uppsala University

Introduction: Physiologically-based pharmacokinetic (PBPK) modeling and population pharmacokinetic (PopPK) modeling are two important pharmacometrics tools in drug development. PBPK modeling usually starts from pre-clinical study, while PopPK modeling is more frequently applied in clinical stage. Currently, these two kinds of models are developed and evaluated independently without any integration. This study aims to link the PBPK results to automatic PopPK model development to bridge the gap between these two approaches.

Methods: Three drugs from four published studies were selected as test drugs: vancomycin (IV drug, renal-eliminated), montelukast (IV drug, CYP2C8 substrate) and midazolam (oral drug, CYP3A4 substrate). The characteristics of study subjects, administration protocol and sampling schedule reported in the publication were applied in the software PK-sim to create corresponding datasets of each study. Next, the simulated PK data were modified by adding stochastic residuals as reported in the publication. The modified dataset was then split into 10 sub datasets and put into an automatic model development (AMD) tool to build a PopPK model. A bi-directional verification was then carried out by evaluating the models in three ways: (1) compare the final PK estimates in the AMD model: clearance (CL), volume of distribution at steady state (Vss) and mean absorption time (MAT) with the true PK parameters in PBPK model calculated from non-compartmental analysis (NCA). (2) compare the structural model and residual unexplained variability (RUV) model results between the AMD model and published model. (3) compare the NCA results of CL and Vss with the corresponding PK parameters in the publication.

Results: The AMD tool performed well in describing the PBPK simulated data. From the representative VPC plots, over 90% of the median, 2.5th, and 97.5th percentiles of observations were within the predicted interval in all four datasets. Meanwhile, the AMD tool successfully captured the PK parameter CL and Vss from the given dataset. In the 40 tested sub datasets (10 sub datasets for each study), 94.8% CL estimates and 87.18% Vss estimates were within a 2-fold error range, comparing with the NCA results. The datasets that fell outside of the 2-fold error range resulted from a combination of an uninformative study design and AMD model deficiency. However, the absorption parameter MAT of midazolam was hard to estimate and varied a lot among the ten datasets. In structural model comparison, PK-sim could simulate data with similar PK characteristics to the clinical PK data while it failed to reproduce all the PK parameters in the published models. The AMD tool was also able to identify the manually-added residual error correctly.

Conclusions: The PK-sim successfully created PK data close to the clinical PK data. AMD tool turned out to have a good performance in capturing the PK characteristics in the datasets simulated by PK-sim. In the long run, we anticipate applying this technique to bridge the gap between pre-clinical and clinical studies in drug development.