Isabelle ALAKIKI 1,2, David Marchionni 1, Gilles Tiraboschi 1, David Fabre 1
1 Sanofi R&D Montpellier, PMx-D, QP, TMU, France (, ), 2 Master of Pharmacokinetics, Aix-Marseille University, Marseille, France (, )
Objectives
Population pharmacokinetic (PopPK) modeling plays a central role in model-informed drug development (MIDD); supporting decision-making throughout clinical development and regulatory evaluation [1]. For several decades, NONMEM has constituted the reference platform for nonlinear mixed-effects modeling (NLME) and remains extensively used across academia and the pharmaceutical industry.
Recent advances in scientific computing have led to the emergence of modern pharmacometric platforms designed to improve computational efficiency, transparency, and reproducibility of modeling workflows. Pumas, developed within the Julia programming ecosystem, provides a high-performance environment integrating model development, parameter estimation, and simulation capabilities [2,3]. Demonstrating reproducibility of established PopPK analyses across computational platforms is
therefore an important methodological challenge for the pharmacometric community.
In parallel, regulatory agencies increasingly emphasize robustness, transparency, and reproducibility of pharmacometric analyses supporting drug development and evaluation [4,5]. This project aimed to reproduce a complete PopPK analysis of a monoclonal antibody, originally developed in NONMEM version 7.5.1, within the Pumas framework. The objectives were to reproduce the full model development and qualification workflow, evaluate cross-platform
consistency of parameter estimation and predictive performance, and establish a reproducible approach for translating NONMEM models into Pumas.
Methods
Pooled pharmacokinetic data from five clinical studies including 280 individuals were analyzed, including three Phase I studies in healthy participants and two Phase II studies in patients. The doses were administered intravenously or subcutaneously.
The reference PopPK analysis was conducted in NONMEM version 7.5.1 using the first-order conditional estimation method with interaction (FOCE-I). The final model consisted of a twocompartment model with first-order absorption and linear elimination parameterized using clearance (CL), central volume of distribution (Vc), intercompartmental clearance (Q), peripheral volume of distribution (Vp), absorption rate constant (Ka), and bioavailability (F). Allometric scaling based on body weight was applied to CL, Vc, Q, and Vp using estimated allometric exponents. Additional covariate effects included a linear effect of body mass index (BMI) on Vc, a power effect of BMI on Vp, multiplicative effects of race on Vc and immunogenicity status on CL, as well as effects of sex and race on bioavailability. Interindividual variability was modeled exponentially, with correlated variability between CL and Vc implemented through a multivariate normal distribution, while independent variability was estimated for Vp, Ka, and bioavailability. Residual unexplained variability was described using a combined proportional and additive error model.
The complete modeling workflow was reproduced in Pumas through an iterative translation process ensuring statistical equivalence between platforms. Dedicated Julia functions were developed to automate model execution, diagnostic plots and quality criteria generation. Model evaluation reproduced standard qualification procedures including goodness-of-fit diagnostics and predictioncorrected visual predictive checks (pcVPC). Cross-platform comparisons focused on parameter estimates, variability components, likelihood metrics, model diagnostics, and execution time.
Results
Structural parameter estimates obtained in Pumas were highly consistent with those obtained in NONMEM. Relative differences between platforms were below 5% for all parameter estimates, including structural parameters, interindividual variability, and residual error parameters.
Likelihood metrics were comparable across platforms. Objective function values differed between NONMEM and Pumas, as expected due to differences in likelihood constant handling across software implementations. Despite these differences, parameter estimates and diagnostic evaluations remained consistent between platforms. Diagnostic plots generated in Pumas were visually consistent with NONMEM outputs, with CWRES and IWRES distributions centered around zero and showing no evidence of systematic bias.
Prediction-corrected visual predictive checks demonstrated equivalent predictive performance, with observed concentrations contained within simulated prediction intervals in both implementations, confirming preservation of model qualification results after translation.
From a computational perspective, Pumas showed performance comparable to NONMEM for parameter estimation while providing reductions in computation time for several workflow components, particularly simulation-based diagnostics. The developed functions enabled standardized execution of analyses and facilitated efficient cross-platform comparison, demonstrating that the entire model development pathway could be reliably reproduced.
Conclusions
This work demonstrates cross-platform reproducibility of a PopPK analysis across all stages of model development and qualification.
Beyond software comparison, this work demonstrates an approach supported by dedicated computational tools for translating pharmacometric models toward modern computational environments while maintaining consistency with established regulatory expectations, thereby contributing to improved reproducibility, transparency, and scalability of pharmacometric analyses within MIDD.
References:
References
[1] Mould DR, Upton RN. CPT Pharmacometrics Syst Pharmacol. 2012.
[2] Rackauckas C et al. bioRxiv. 2020.
[3] Bezanson J et al. SIAM Review. 2017.
[4] FDA. Population Pharmacokinetics: Guidance for Industry. 2022.
[5] European Medicines Agency. Guideline on reporting population pharmacokinetic analyses. 2007.
Reference: PAGE 34 (2026) Abstr 12064 [www.page-meeting.org/?abstract=12064]
Poster: Methodology - Other topics