Abstract
<title>Abstract</title> <p>Automated population pharmacokinetic (PopPK) modelling has emerged as a promising approach to reduce the time and expertise required for model development while enabling exploration of large model search spaces. However, existing studies have primarily evaluated automated methods using information criteria or by comparing them with expert-developed models, providing limited evidence of whether the selected model is truly optimal. A more rigorous assessment requires comparison against exhaustive search, which identifies the best model within a predefined search space. This study tested nlmixr2auto, an end-to-end automated PopPK modeling framework that integrates automated generation of initial parameter estimates, model evaluation, and model selection. The framework was evaluated using 22 real clinical datasets. To establish a ground truth model, exhaustive searches were performed across predefined search spaces containing up to 711 candidate models for intravenous datasets and 1,422 candidate models for oral datasets. The performance of metaheuristic algorithms implemented in nlmixr2auto was then compared with exhaustive and conventional stepwise model building. The results showed that metaheuristic algorithms outperformed the stepwise model building. Genetic algorithms, ant colony optimization, and tabu search successfully recovered the exhaustive-search optimum in 21, 20, and 19 of the 22 datasets, respectively, compared with only 7 datasets for the stepwise model building. In 63.6% (14 out of 22) cases, the metaheuristic algorithms completed the modeling task within 24 hours using a 4-CPU, 2-GB computing environment. These findings suggest that nlmixr2auto offers a transparent and efficient open-source framework with considerable potential to support PopPK modeling.</p>