Objective: To perform a comparative quantitative evaluation of the prediction accuracy for human hepatic metabolic clearance of 5 different mathematical models: allometric scaling (multiple species and rat only), physiologically based direct scaling, empirical in vitro-in vivo correlation, and supervised artificial neural networks.
Methods: The mathematical prediction models were implemented with a publicly available dataset of 22 extensively metabolised compounds and compared for their prediction accuracy using 3 quality indicators: prediction error sum of squares (PRESS), r2 and the fold-error.
Results: Approaches such as physiologically based direct scaling, empirical in vitro-in vivo correlation and artificial neural networks, which are based on in vitro data only, yielded an average fold-error ranging from 1.64 to 2.03 and r2 values greater than 0.77, as opposed to r2 values smaller than 0.44 when using allometric scaling combining in vivo and in vitro preclinical data. The percentage of successful predictions (less than 2-fold error) ranged from 55% (rat allometric scaling) to between 64 and 68% with the other approaches.
Conclusions: On the basis of a diverse set of 22 metabolised drug molecules, these studies showed that the most cost-effective and accurate approaches, such as physiologically based direct scaling and empirical in vitro-in vivo correlation, are based on in vitro data alone. Inclusion of in vivo preclinical data did not significantly improve prediction accuracy; the prediction accuracy of the allometric approaches was at the lower end of all methods compared.