Comprehensive comparison of six optimization algorithms for automated population pharmacokinetic model selection via two simulation studies and eleven real case studies
Abstract
Background and objectives: Prior efforts to automate population pharmacokinetic structural and statistical model selection have explored various algorithms with mixed results. This study is conducting a comprehensive comparison of six distinct optimization algorithms for selecting pharmacokinetic and target-mediated drug disposition models. Our objective is to determine which algorithms to use preferentially for different model search space sizes. Methods: We implemented six algorithms in Monolix R API: (i) Decision Tree, (ii) Genetic Algorithm, (iii) Simulated Annealing, (iv) Particle Swarm Optimization, (v) Ant Colony Optimization, and (vi) a Tournament Algorithm. Each algorithm's performance was evaluated on a first simulation involving a simple structural pharmacokinetic model, a second simulation focused on a more complex target-mediated drug disposition model and applied to 11 case studies representing diverse real-life situations. Results: In the first simulation with a small search space, Ant Colony Optimization, Particle Swarm Optimization, Genetic Algorithm, and Tournament Algorithm achieved 90% concordance with the gold standard. A Decision Tree algorithm reached the same performance, but three times faster. In the larger target-mediated drug disposition search space, Ant Colony Optimization consistently identified the best models, though with high computational costs. Tournament Algorithm balanced speed and performance. When applied to real case studies, the top-performing algorithms selected models with better cost function values than those previously published in 10 out of 11 cases. Conclusions: Optimization algorithms offer a robust and effective approach to model selection automation. To explore simple population pharmacokinetic models, we advise using a Decision Tree algorithm, while using Ant Colony Optimization for exploring more complex problems.
Related
- Comparison of Trough-Only and Peak–Trough Concentration Data for the Calculation of Vancomycin Area under the Concentration–Time Curve on Day 2 Using Previous and Updated Population Pharmacokinetic Models
- Population pharmacokinetic modeling of voriconazole based on real-world data for dosing regimen optimization
- Population pharmacokinetic modeling of adalimumab biosimilar SCT630 and Humira® in healthy participants and patients with moderate to severe plaque psoriasis to establish PK similarity
- Icariin, a Multi-Target Natural Neuroprotectant, and a Comprehensive Review on Its Synthetic Methods, Pharmacological Evaluation and Pharmacokinetic Characteristics
- Electronic-Nomogram-Guided Micafungin Dosing Regimen in Critically Ill Patients: A Population Pharmacokinetic Study
BibTeX
@article{pierre2026comprehensive,
title = {{Comprehensive comparison of six optimization algorithms for automated population pharmacokinetic model selection via two simulation studies and eleven real case studies}},
author = {Matthias Pierre and Frano Mihaljevic and Geraldine Celliere and Julie Bertrand},
journal = {Computer Methods and Programs in Biomedicine},
year = {2026},
month = oct,
doi = {10.1016/j.cmpb.2026.109649},
url = {https://inserm.hal.science/inserm-05780576v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Comprehensive comparison of six optimization algorithms for automated population pharmacokinetic model selection via two simulation studies and eleven real case studies
AU - Matthias Pierre
AU - Frano Mihaljevic
AU - Geraldine Celliere
AU - Julie Bertrand
PY - 2026
DA - 2026/10/06
JO - Computer Methods and Programs in Biomedicine
DO - 10.1016/j.cmpb.2026.109649
UR - https://inserm.hal.science/inserm-05780576v1
AB - Background and objectives: Prior efforts to automate population pharmacokinetic structural and statistical model selection have explored various algorithms with mixed results. This study is conducting a comprehensive comparison of six distinct optimization algorithms for selecting pharmacokinetic and target-mediated drug disposition models. Our objective is to determine which algorithms to use preferentially for different model search space sizes. Methods: We implemented six algorithms in Monolix R API: (i) Decision Tree, (ii) Genetic Algorithm, (iii) Simulated Annealing, (iv) Particle Swarm Optimization, (v) Ant Colony Optimization, and (vi) a Tournament Algorithm. Each algorithm's performance was evaluated on a first simulation involving a simple structural pharmacokinetic model, a second simulation focused on a more complex target-mediated drug disposition model and applied to 11 case studies representing diverse real-life situations. Results: In the first simulation with a small search space, Ant Colony Optimization, Particle Swarm Optimization, Genetic Algorithm, and Tournament Algorithm achieved 90% concordance with the gold standard. A Decision Tree algorithm reached the same performance, but three times faster. In the larger target-mediated drug disposition search space, Ant Colony Optimization consistently identified the best models, though with high computational costs. Tournament Algorithm balanced speed and performance. When applied to real case studies, the top-performing algorithms selected models with better cost function values than those previously published in 10 out of 11 cases. Conclusions: Optimization algorithms offer a robust and effective approach to model selection automation. To explore simple population pharmacokinetic models, we advise using a Decision Tree algorithm, while using Ant Colony Optimization for exploring more complex problems.
SN - 0169-2607
ER -