Celine Brochot 1, Pauline Bogdanovich 1, Anthonia Onasanwo 1, Hiroshi Momiji 1, Frederic Bois 1, Masoud Jamei 1
1 Certara (Sheffield, United Kingdom)
Physiologically based pharmacokinetic (PBPK) modeling is a powerful tool for integrating in vitro and in vivo data to predict drug behavior and support drug development. A major challenge in PBPK model development is parameter uncertainty, which may arise from limited experimental data or highly variable measurements. Parameter estimation using observed clinical data, such as plasma drug concentrations, is therefore often required to improve model prediction performance. Frequentist inference is often poorly suited to population-based PBPK models because it ignores prior distributions and can distort covariate relationships. Bayesian inference can overcome these limitations. In particular, Sampling Importance Resampling (SIR) [1,2] is well suited for population PBPK models, as it preserves inter-individual variability encoded in parameter distributions. In this work, we first demonstrate the SIR procedure to support the development of a PBPK model for theophylline using the Simcyp Simulator V24. Then the predictability of the resulting PBPK model is assessed by simulating theophylline plasma pharmacokinetics in the presence of an inhibitor of CYP1A2, the main metabolizing enzyme of theophylline.
A minimal PBPK model for theophylline in healthy volunteers was implemented in Simcyp V24. Clinical data from twelve healthy male subjects, including 24-hour plasma concentration profiles and body weights, were used. Default Simcyp distributions were applied for physiological parameters. Physicochemical properties of theophylline (logP and pKa) were assumed known. The initial values for the ADME parameters were either taken from literature or derived from a variability analysis that was run for the three most influential compound-dependent parameters on plasma pharmacokinetics, i.e., volume of distribution (Vss), first-order absorption rate (ka), and maximum CYP1A2 metabolic rate (Vmax) [3]. Prior lognormal distributions were set with means equal to these initial values and a 30% coefficient of variation.
The SIR procedure approximates a target posterior distribution by resampling weighted simulations drawn from an initial (prior) distribution. In this study, the prior distribution corresponds to multivariate parameter distributions defined in Simcyp, and the weights are proportional to the likelihood of observed clinical data given model predictions. Resampling from these weighted simulations produces a posterior sample that better reflects the individual clinical observations. SIR was applied to the theophylline PBPK model as follows: i) 10,000 prior PBPK simulations were generated per subject; ii) importance weights were calculated for each simulation using the data likelihood that was assumed to be a normal measurement error model with both constant and concentration-dependent components; and iii) 1,000 samples were drawn from the initial 10,000 according to the normalized weights to generate the posterior distributions.
Comparison of prior and SIR-filtered PBPK simulations demonstrated a marked reduction in variability, with simulated concentration-time profiles closely matching observed data. Comparison of the parameters prior and posterior distributions showed that the posterior distributions of a few parameters have been updated, including Vss, ka, Vmax and other metabolic parameters. Among the 12 subjects, the overall trend is that the posterior means of ka are lower than the prior value, greater for Vss, and at a similar level for Vmax.
SIR-derived weights were then used to predict theophylline pharmacokinetics under CYP1A2 inhibition by ciprofloxacin. A clinically relevant dosing regimen was applied [5]: a single oral dose of theophylline (3.4 mg·kg⁻¹) before and after 60 hours of ciprofloxacin therapy (500 mg twice daily). Simulated AUC ratios with and without inhibitor closely matched published clinical observations, demonstrating accurate prediction of drug–drug interaction outcomes.
In conclusion, SIR offers a robust and scalable approach to integrate clinical data into PBPK modeling. By updating joint parameter distributions while preserving inter-individual variability and covariate relationships, SIR enhances predictive model accuracy, supporting individualized dose optimization and rational design of clinical pharmacology studies.
References:
[1] Rubin DB Using the SIR algorithm to simulate posterior distributions. In: Bernardo JM, De Groot MH, Lindley DV, Smith AFM (eds) Bayesian Statistics 3. Oxford University Press, Oxford, 1988, pp 395–402
[2] Dosne AG, Bergstrand M, Karlsson MO. An automated sampling importance resampling procedure for estimating parameter uncertainty. J Pharmacokinet Pharmacodyn. 2017;44(6):509–20.
[3] Wedagedera JR, Afuape A, Chirumamilla SK, Momiji H, Leary R, Dunlavey M, et al. Population PBPK modeling using parametric and nonparametric methods of the Simcyp Simulator, and Bayesian samplers. CPT Pharmacometrics Syst Pharmacol. 2022;11(6):755–65.
[4] Trembath PW, Boobois SW. Pharmacokinetics of a sustained-release theophylline formulation. Br J Clin Pharmacol. 1980;9(4):365–9.
[5] Batty KT, Davis TM, Ilett KF, Dusci LJ, Langton SR. The effect of ciprofloxacin on theophylline pharmacokinetics in healthy subjects. Br J Clin Pharmacol. 1995;39(3):305–11.
Reference: PAGE 34 (2026) Abstr 12076 [www.page-meeting.org/?abstract=12076]
Poster: Methodology - Estimation Methods