A tool for First-in-human PK Prediction Incorporating Experimental Uncertainty
Lindauer, A (1), Yee, K (1), Guo, D-L (2), Lee, B (2), Ivashina, J (3), Deshmukh, SV (4), Kothare, P (1), Martin, I (4), Gibson, C (4)
Merck & Co. Inc. Departments of: (1) Quantitative Pharmacology & Pharmacometrics - Oss, NL and West Point, US; (2) Translational Medicine Research Centre - Singapore; (3) Information Technology - West Point, US; (4) Preclinical ADME - Boston and West Point, US
Objectives: To develop a user-friendly application that facilitates the routine prediction of human PK with uncertainty on the basis of preclinical data measured with error.
Methods: Recommended methods for human PK predictions at Merck are being refined to more systematically incorporate experimental uncertainty. Equations for first-in-human PK predictions (e.g. well-stirred liver model, allometric scaling of volume of distribution) were coded in R-scripts (R v3.0.2). To conveniently enter experimental data for users not familiar with R, a graphical user interface (GUI) was developed in ASP.NET MVC 4 framework as a front-end. Information exchange between the GUI and R is handled via XML files.
Results: Depending on the type of parameter, the preclinical data can be entered in different formats such as geometric mean and standard error of the measurement or as typical estimate and correlation matrix in case of in vivo PK parameters obtained from fitting a compartmental model to animal data. Upon execution of the R-scripts in the background, random samples are drawn from a distribution determined by the input format (e.g. log-normal, multivariate-normal). In case of parameters that are bound between values of 0 and 1 (e.g. fraction unbound) samples are drawn from a logistic distribution. The equations for preclinical-to-clinical extrapolation are then applied on these sets of random values to predict the distribution of human PK parameters (i.e. bioavailability, volume of distribution, clearance). Simulations of the expected concentration-time profiles are automatically conducted to determine the likelihood of achieving a target concentration (e.g. trough level, AUC; provided by the user) for a range of different doses. The results are visualized as PK curves with confidence intervals, bar-plots displaying the probability of exceeding the target at different doses, and a tornado plot showing the contribution of each input parameter to the overall uncertainty of the prediction of the PK parameter of interest.
Conclusions: Accounting for experimental uncertainty facilitates a transparent team discussion around confidence associated with human predictions. Instead of obtaining a prediction of ‘the’ efficacious dose, a probabilistic statement can be made about the likelihood that a certain dose will hit the target. Importantly, from the tornado plots it is immediately visible which parameters are most influential and may require additional experiments to improve the precision of the predictions.