Evaluation of various static and dynamic modeling methods to predict clinical CYP3A induction using in vitro CYP3A4 mRNA induction data.

Several drug-drug interaction (DDI) prediction models were evaluated for their ability to identify drugs with cytochrome P450 (CYP)3A induction liability based on in vitro mRNA data. The drug interaction magnitudes of CYP3A substrates from 28 clinical trials were predicted using (i) correlation approaches (ratio of the in vivo peak plasma concentration (Cmax) to in vitro half-maximal effective concentration (EC50); and relative induction score), (ii) a basic static model (calculated R3 value), (iii) a mechanistic static model (net effect), and (iv) mechanistic dynamic (physiologically based pharmacokinetic) modeling. All models performed with high fidelity and predicted few false negatives or false positives. The correlation approaches and basic static model resulted in no false negatives when total Cmax was incorporated; these models may be sufficient to conservatively identify clinical CYP3A induction liability. Mechanistic models that include CYP inactivation in addition to induction resulted in DDI predictions with less accuracy, likely due to an overprediction of the inactivation effect.

Evaluation of various static in vitro-in vivo extrapolation models for risk assessment of the CYP3A inhibition potential of an investigational drug.

Nine static models (seven basic and two mechanistic) and their respective cutoff values used for predicting cytochrome P450 3A (CYP3A) inhibition, as recommended by the US Food and Drug Administration and the European Medicines Agency, were evaluated using data from 119 clinical studies with orally administered midazolam as a substrate. Positive predictive error (PPE) and negative predictive error (NPE) rates were used to assess model performance, based on a cutoff of 1.25-fold change in midazolam area under the curve (AUC) by inhibitor. For reversible inhibition, basic models using total or unbound systemic inhibitor concentration [I] had high NPE rates (46-47%), whereas those using intestinal luminal ([I]gut) values had no NPE but a higher PPE. All basic models for time-dependent inhibition had no NPE and reasonable PPE rates (15-18%). Mechanistic static models that incorporate all interaction mechanisms and organ specific [I] values (enterocyte and hepatic inlet) provided a higher predictive precision, a slightly increased NPE, and a reasonable PPE. Various cutoffs for predicting the likelihood of CYP3A inhibition were evaluated for mechanistic models, and a cutoff of 1.25-fold change in midazolam AUC appears appropriate.