Zhonghui Huang 1, Joseph Standing 1, Matthew Fidler 2, Frank Kloprogge 3
1 Infection, Immunity and Inflammation Research & Teaching Department, UCL Great Ormond Street Institute of Child Health (London, UK), 2 Novartis Pharmaceuticals (Fort Worth, USA), 3 UCL Institute for Global Health, University College London (London, UK)
Introduction/Objectives. Population pharmacokinetic modelling (PopPK) is an essential tool in drug development and clinical practice, supporting dose optimisation and individualised dosing decisions. nlmixr2 [1] is an open-source R [2] package for nonlinear mixed-effects (NLME) modelling that is increasingly used within the pharmacometrics community. In routine use, it involves model specification, parameter initialisation, and model evaluation, with user interaction at each step. We therefore developed nlmixr2autoinit [3] and nlmixr2auto [4] to support automated nonlinear mixed-effects model development by providing end-to-end workflow automation within the nlmixr2 ecosystem. This work aimed to present an automated framework and associated workflow built on these tools, summarise their functional modules and implementation, and provide practical operational guidance.
Methods. The nlmixr2auto family comprises two core packages: nlmixr2autoinit, which automatically generates initial parameter estimates based on data-driven methods, covering both fixed-effect and random-effect parameters, and nlmixr2auto which performs automated model fitting, evaluation and selection. Model parameter estimation is performed using the SAEM algorithm. The search component integrates four built-in algorithms: stepwise model building, genetic algorithms, ant colony optimisation, and tabu search. A unified fitness framework defaults to Bayesian Information Criterion (BIC) with step penalties, while allowing adjustment of Akaike Information Criterion (AIC) or Objective Function Value (OFV), and detailed diagnostics, including relative standard error (RSE) thresholds, inter-individual variability (IIV) correlation control, and covariance step success. Penalties can be configured as step (default) or binary to support flexible model selection. The default search space currently encompasses 711 models for intravenous administration and 1,422 models for oral administration, and extension interfaces are provided to support customised model development.
Results. Using three example datasets [5,6], namely theo_sd, pheno_sd and tobramycin, we systematically presented the core functions and typical workflow of nlmixr2autoinit and nlmixr2auto. The results showed that the tool supports fully automated model search and fitting under default settings, while also allowing stepwise execution of individual search modules to achieve more refined structural control and strategy adjustment. On the example datasets, the three metaheuristic algorithms consistently identified the same models as those obtained by exhaustive search within the predefined model space. In addition, stepwise model building was able to recover the same reference models when a customised strategy was applied. For model evaluation, key components of the fitness function, including the penalty term, penalty type and penalty strength, are configurable to accommodate different research objectives. In addition, a user-specified starting model can be defined as the basis for the search to enhance the use of prior knowledge. When confronted with complex pharmacokinetic structures that exceed the coverage of the built-in search space, custom models can be parameterised and tested, thereby extending structural exploration to more advanced scenarios. Overall, these results demonstrated that the tool achieves a balance between automation and flexible configuration, highlighting its feasibility and applicability in complex pharmacokinetic modelling contexts.
Conclusion. nlmixr2auto integrates data-driven initialisation, multi-algorithm search and a multidimensional evaluation framework to enable automated exploration and optimisation of NLME models. This workflow provided an overview from both functional and applied perspectives, offering practical strategies and guidance for users in the pharmacometric community in applying nlmixr2auto.
References:
[1] Fidler, M., Wilkins, J. J., Hooijmaijers, R., Post, T. M., Schoemaker, R., Trame, M. N., Xiong, Y., & Wang, W. (2019). Nonlinear Mixed‐Effects Model Development and Simulation Using nlmixr and Related R Open‐Source Packages. CPT: Pharmacometrics and Systems Pharmacology, 8(9), 621–633. https://doi.org/10.1002/psp4.12445
[2] R: The r project for statistical computing [Internet]. Available from: https://www.r-project.org/
[3] Huang, Z., Fidler, M., Lan, M., Cheng, I. L., Kloprogge, F., & Standing, J. F. (2025). An automated pipeline to generate initial estimates for population Pharmacokinetic base models. Journal of Pharmacokinetics and Pharmacodynamics, 52(6), Article 60. https://doi.org/10.1007/s10928-025-10000-z
[4] PAGE 32 (2024) Abstr 11272 [www.page-meeting.org/?abstract=11272]
[5] Schoemaker, R., Fidler, M., Laveille, C., Wilkins, J. J., Hooijmaijers, R., Post, T. M., Trame, M. N., Xiong, Y., & Wang, W. (2019). Performance of the SAEM and FOCEI Algorithms in the Open‐Source, Nonlinear Mixed Effect Modeling Tool nlmixr. CPT: Pharmacometrics and Systems Pharmacology, 8(12), 923–930. https://doi.org/10.1002/psp4.12471
[6] Lixoft. (2024). Tobramycin data set. https://monolixsuite.slp-software.com/monolix/2024R1/tobramycin-data-set
Reference: PAGE 34 (2026) Abstr 12080 [www.page-meeting.org/?abstract=12080]
Poster: Methodology - New Modelling Approaches