Probabilistic Risk Assessment Using Integrated Preclinical-Clinical PK/PD Modelling in NONMEM
Philip J. Lowe (1) and William Sallas (2)
(1) Modelling and Simulation Section, Preclinical Safety, Novartis Pharmaceuticals AG, 4002 Basel, Switzerland; (2) Clinical Modelling and Related Technologies, Clinical Development and Medical Affairs, Novartis Pharmaceuticals Corporation, East Hanover, New Jersey, USA.
Objective: To assess the clinical risk of an adverse effect seen only in preclinical safety studies.
Methods: PK + PD data were obtained from toxicology studies (N=90) and PK from clinical Phase II and III studies (N=1543). For the PD, after checking that time delay hysteresis was not occurring, a sigmoid Emax function was used for the concentration-effect model. The clinical population PK utilised a one compartment system with first order absorption. For the risk assessment simulations, the clinical PK model for steady-state Cmax drove the preclinical PD model, with a team-agreed assumption of equal species sensitivity to the drug. Monte-Carlo simulations were carried out from a clinical population demographic database of 1117 patients who matched the dosing criteria. Subproblems=200 gave 223400 observations from which to count events of clinical concern. Uncertainty in the PK, PD and variance parameters was assessed by repeating the simulation 30 times with random parameter values drawn from distributions of those parameter values.
Results: In the clinical PK model, many covariates were found for patient demographic features. Only two of these were large enough to warrant changes in dose and regimen. Bodyweight and baseline disease status were used in the creation of a dosing table. The toxicology PD model was characterised by individual parameters for baseline, IC50 and Hill coefficient for each animal. Of 223400 simulated patients, none gave a clinically significant response, therefore the mean risk was judged to be less than 1 in 223400. After parameter uncertainty was taken into account the 95% prediction interval for the risk was estimated to be from 1/17000 to <<1/223400.
Conclusions: A scheme for assessing the probability of a toxicologically observed adverse effect being a risk to human health is presented, using an extension to the population PK modelling commonly carried out in drug development. The process relies on a number of assumptions which are either supported by data, or purposely set to be conservative. For drugs more complex than the case presented here, extra features could and should be built into the model to account for interspecies differences.