Pauline Bogdanovich 1, Céline Brochot 1, Masoud Jamei 1, Sibylle Neuhoff 1, Oliver Hatley 1, Iain Gardner 1, Karen R. Yeo 1
1 Certara Predictive Technologies, Cerata UK (Sheffield, United Kingdom)
As part of the qualification of the Simcyp Simulator V19 by the European Medicine Agency (EMA) in 2025, a statistical model was developed to quantify bias and uncertainty of the Simulator in physiologically based pharmacokinetic (PBPK) predictions of drug-drug interactions (DDIs) mediated by cytochromes P450. This model accounted for systematic differences between inhibition types and was used to analyze the potential discrepancies between reported clinical data on DDIs and corresponding Simcyp predictions.
A Bayesian hierarchical meta-regression model was developed to estimate: (1) mean bias in geometric mean ratio (GMR) predictions, accounting for the influence of the type of inhibition mechanism i.e., competitive inhibition (CI) or mechanism-based inhibition (MBI); (2) between-study variance as an upper limit of prediction imprecision; and (3) bias in between-subject variability (BSV) predictions. The data considered in the analysis were derived from scientific public literature that reported GMRs and associated population SDs for clinical DDI studies involving CYP1A2, 2D6, 2C8, 2C9, 2C19, and CYP3A4 enzymes. For each clinical study, a corresponding set of data was generated by simulating the DDI studies through the Simcyp Simulator to predict the GMRs and population SDs. In the end, the model was applied to 220 studies reporting AUC GMRs and 160 studies reporting Cmax GMRs. Bayesian inference for the model was performed using Markov chain Monte Carlo (MCMC) sampling.
The mean GMR biases were found to be small and negative (ranged from -6% to -4% for both AUC and Cmax), indicating a slight overprediction of the effect of DDIs. The between-study variance values corresponded to CVs that ranged from 18% (CI) to 25% (MBI) for AUC and 19% (CI) to 31% (MBI) for Cmax. MBI predictions exhibited marginally greater uncertainty than CI predictions. These CVs represent upper bounds of Simcyp imprecision, as between-study variability cannot be separated. The mean BSV bias factor was approximately 2 for AUC and 3 for Cmax, indicating that Simcyp predictions tend to underestimate the variability.
The calibrated model can then be applied to evaluate how uncertainty influences future DDI predictions for CYP-mediated inhibition for regulatory decision-making. Predictive visualizations generated from the model-based discrepancy analysis can be used to inform drug development strategies. These predictive capabilities were demonstrated for a hypothetical scenario in which: (1) the compound and the inhibitor under study are both CYP substrates with adequate PBPK models and the anticipated DDI is mediated by one of the CYP enzymes considered in this study; (2) the therapeutic index of the compound is assumed to be known and set to a 0.5- to 2-fold range relative to the expected geometric mean exposure at the therapeutic dose; (3) a hypothetical DDI is predicted using the Simcyp Simulator following concomitant administration with the CYP inhibitor. Accounting for the inhibition mechanism, posterior predictive analyses generated 90% credibility intervals around hypothetical predicted GMRs and calculated the probability of exceeding therapeutic index thresholds as a function of predicted GMR.
The posterior predictive analyses revealed wider credibility intervals for MBI compared to CI predictions, reflecting greater uncertainty. For a compound with a predicted GMR of 1.5, the 90% credibility intervals were: [1.06, 1.91] (CI) and [0.95, 2.18] (MBI) for AUC, and [1.05, 1.94] (CI) and [0.84, 2.35] (MBI) for Cmax. When overlaid on a hypothetical therapeutic window (0.5- to 2-fold), MBI predictions showed a portion of the 90% credibility interval extend beyond the therapeutic range for both AUC and Cmax, whereas CI predictions remained within the bounds. Risk-based thresholds indicated that for <5% probability of exceeding a 2-fold upper limit of the therapeutic window, the maximum acceptable predicted GMRs were 1.5 (CI) and 1.3 (MBI) for AUC, and 1.5 (CI) and 1.2 (MBI) for Cmax. The statistical model developed provides a robust framework for bias and uncertainty quantification in PBPK-based DDI predictions. The low bias and bounded imprecision support regulatory confidence in Simcyp V19 predictions. The posterior predictive analysis allows for determination of acceptable predicted GMR limits based on drug-specific therapeutic indices and risk tolerance, supporting informed regulatory decision-making.
Reference: PAGE 34 (2026) Abstr 12294 [www.page-meeting.org/?abstract=12294]
Poster: Methodology - Model Evaluation