Impact of input parameters on the prediction of hepatic plasma clearance using the well-stirred model.

The in vitro metabolic stability assays are indispensable for screening the metabolic liability of new chemical entities (NCEs) in drug discovery. Intrinsic clearance (CL(int)) values from liver microsomes and/or hepatocytes are frequently used to assess metabolic stability as well as to quantitatively predict in vivo hepatic plasma clearance (CL(H)). An often used approximation is the so called well-stirred model which has gained widespread use. The applications of the well-stirred model are typically dependent on several measured parameters and hence with potential for error-propagation. Despite widespread use, it was recently suggested that the well-stirred model in some circumstances has been misused for in vitro in vivo extrapolation (IVIVE). In this work, we follow up that discussion and present a retrospective analysis of IVIVE for hepatic clearance prediction from in vitro metabolic stability data. We focus on the impact of input parameters on the well stirred model; in particular comparing "reference model" (with all experimentally determined values as input parameters) versus simplified models (with incomplete input parameters in the models). Based on a systematic comparative analysis and model comparison using datasets of diverse drug-like compounds and NCEs from rat and human, we conclude that simplified models, disregarding binding data, may be sufficiently good for IVIVE evaluation and compound ranking at early stage for cost-effective screening. Factors that can influence prediction accuracy are discussed, including in vitro intrinsic clearance (CL(int)) and in vivo CL(int) scaling factor used, non-specific binding to microsomes (fu(m)), blood to plasma ratio (C(B)/C(P)) and in particular fraction unbound in plasma (fu). In particular, the fu discrepancies between literature data and in-house values and between two different compound concentrations 1 and 10 µM are exemplified and its potential impact on prediction performance is demonstrated using a simulation example.