Accounting for uncertainty in the application of generic PBPK workflow to candidate selection
Matthieu Jacobs (1), Sylvain Fouliard( 1), Nicolas Couvreur (1)
(1) Translational Medicine, Servier, France
Introduction:
During drug discovery, a high number of drug candidates are screened for their ADME (absorption, distribution, metabolism and excretion) properties using in vitro assays(1). Integration of these ADME parameters into predicted PK profile in animals and humans allows to predict the therapeutic dose, which can be further utilized to optimize compound design, prioritize series/compound and identify potential series-related ADME risks(1,2). Ultimately this reduces PK-related failure in drug development.
For an efficient workflow, 2 challenges need to be overcome. Firstly, the human PK prediction based on screening-level in-vitro data only: most PBPK software allows to predict human PK without any in-vivo input (1). Secondly, at very early stage, PK/PD relationship is often unknown and therefore the PK metrics (AUC, Cmax, Cthrough,…) driving the pharmacological effect remain unknown. This step bears high risk as a wrong hypothesis can lead to a biased compound selection and a waste of resources (money, time, and most importantly unnecessary in-vivo experiments). In order to quantify and mitigate this risk, PKPD simulations can test hypotheses in-silico and provide insightful information for decision-making on compound selection, accounting for the uncertainty in the PKPD relationship.
Objectives:
To evaluate a Pharmacometrics workflow informing compound selection in a real drug discovery program through:
- Assessment of IVIVE in animals using a generic PBPK model and a minimal data package
- Integration of PBPK & PKPD to predict active dose
- Assessment of the uncertainty on PK/PD relationship for compound selection
Methods:
Simulations in humans and mice were performed for 155 compounds from a Servier discovery program with Simcyp software, using a minimal PBPK model with linear clearance. The screening objective was to select the 20 best compounds to move forward. PBPK simulations were based on a minimal screening data package consisting in physicochemical (molecular weight, logP, acid-basic type, pKa) and in-vitro ADME data (e.g. microsomal clearance) i in both species. In-vitro/in-vivo extrapolation assessment was performed comparing PBPK simulation with in-vivo PK profiles from a 10 mg/kg IV administration in mice (n=25). Assessment was based on prediction accuracy for AUC, fu and goodness of fit plots. All dose predictions were performed using PBPK simulations and the pharmacological IC50 on a relevant model was considered as a target concentration. Doses were computed in different PKPD scenarios based on the PK metric of interest (AUC, Cthough, Cmax or time over threshold), the plasma protein binding and in-vitro protein binding of IC50. Overall, 24 scenarios (i.e. 24 doses) for each compound and species were calculated. For each scenario, each compound was ranked based on its dose by species. Then the top 20 compounds of each scenario were considered for selection. Finally, the overall probability of selection (number of times one compound reached the top 20 divided by total number of scenario) was calculated and used as a score for each compound.
Results:
IVIVE using generic PBPK model was good overall with 84% of AUC prediction within into 3 folds and good fitting on in-vivo mice data. Human fu prediction based only on physicochemical data was good with 87.5 % into 3 folds.
On the 155 compounds, none reached 100% of selection (i.e. selection in all scenario). Maximum selection score was 68.7%, interquartile range was [3%, 30%] and minimum was 0%. Interestingly, the best compound had a IC50 only in the 43rd percentile (i.e: 66 compounds add a lower IC50), meaning that its potency was not in top 20. This highlights the importance of PK for compound selection. Furthermore, the scenario where it was not in top20 were in mice, due to higher clearance (46th percentile), which is informative on the developability in preclinical species. Overall, dose prediction for each scenario showed variable rankings, informing on the weaknesses of each compound and highlighting the interest in the approach.
Conclusion:
PBPK modeling associated with simulations of scenario is a powerful tool in a discovery setting. It allows to support compound selection based on early dose prediction as an objective assessment of each candidate. Our approach integrates a risk analysis to account for uncertainty on the PKPD relationship in this decision-making. Also this implementation proved to be compatible with rapid turnaround time and lead to a high adhesion from project team.
References:
[1] Jones HM, Dickins M, Youdim K, et al. Application of PBPK modelling in drug discovery and development at Pfizer. Xenobiotica. 2012;42(1):94-106. doi:10.3109/00498254.2011.627477
[2] Parrott N, Jones H, Paquereau N, Lavé T. Application of full physiological models for pharmaceutical drug candidate selection and extrapolation of pharmacokinetics to man. Basic Clin Pharmacol Toxicol. 2005;96(3):193-199. doi:10.1111/j.1742-7843.2005.pto960308.x