Evaluation of commonin vitroassays for the prediction of oral bioavailability and hepatic metabolic clearance in humans

ABSTRACT Introduction Intrinsic hepatic metabolic clearance (CL int ) measured with human hepatocytes, apparent intestinal permeability (P app ) obtained using the Caco-2 model, unbound fraction in plasma (f u ) and blood-to-plasma concentration ratio (C bl /C pl ) are commonly used for predicting the hepatic clearance (CL H ) and oral bioavailability (F) of drug candidates in humans. The primary objective was to select drugs whose in vitro hepatocyte CL int , Caco-2 P app , f u and C bl /C pl have been measured in various laboratories and studies, and estimate correlation coefficients (R 2 ) for predicted and observed F and log plasma CL H . Secondary aims were to estimate the laboratory/study variability and its impact on predictions and to compare results to in silico and animal model-based predictions. Materials and Methods A literature search was done in order to find unbound hepatocyte CL int , (and corresponding predicted in vivo CL int ), Caco-2 P app , f u and C bl /C pl data. Compounds with multiple measurements for the four assays, without significant in vivo solubility/dissolution limitations and with known in vivo CL H and F, were selected. Min, max and mean estimates were used in the analysis. Results and Discussion Thirty-two compounds with data (in total 561 estimates) produced by 21 major pharmaceutical companies and universities met the inclusion criteria. The predicted vs observed R 2 for log mean CL int , log mean CL H and mean F were 0.32, 0.08 and 0.20, respectively. Exclusion of atenolol increased the R 2 for CL H to 0.20. R 2 -values were considerably lower than those presented in many studies, which seems to be explained by selection bias (choosing favorable reference values). There was considerable interstudy variability for measured and predicted CL int (80- and 1,476-fold mean and max differences, respectively) and measured f u (6.6- and 50-fold mean and max differences, respectively). For F, higher predictive performance was found for in silico (Q 2 =0.58; head-to-head) and animal in vivo models (R 2 =0.30). Conclusion The combination of data from many laboratories and the use of mean values resulted in reduced selection bias and predictive accuracy. Overall, the predictive accuracy (here R 2 ) for log CL int , log CL H and F was low to moderately low (0.08-0.32). The halved R 2 compared to individual studies where high performance was demonstrated seems to be explained be selection bias (enabled by large data variability). Animal in vivo models, and in particular, in silico methodology, outperformed in vitro methodology for the prediction of F in man.