Abstract
<title>Abstract</title> <p>Automating population pharmacokinetic (PopPK) modeling can reduce manual intervention, saving time and effort. Existing automated approaches within the nonlinear mixed-effects (NLME) framework have typically followed either modular stepwise strategies, in which model components are evaluated sequentially, or global search strategies, with genetic algorithms (GA) as a representative example. Stepwise approaches may become trapped in local optima, whereas GA-based approaches can be affected by epistasis arising from interactions among model components. This study aimed to develop a fully automated, end-to-end framework in R within the nlmixr2 ecosystem that integrates automated model building, initial estimate generation, parameter estimation, and model selection. The framework implemented a conventional stepwise model-building approach together with three metaheuristic algorithms: GA and two newly introduced algorithms, ant colony optimization (ACO) and tabu search (TS), and evaluated their performance against exhaustive search results as the ground truth across 144 simulated datasets. Results showed that metaheuristic algorithms consistently outperformed the stepwise approach for identifying ground-truth models. TS and ACO achieved the highest overall model retrieval rates (91%, 90%), followed by GA (87%), whereas the stepwise algorithm achieved approximately 60%. The proposed framework provides an open-source solution for fully automated PopPK model development and is available as a CRAN package “nlmixr2auto”. Our findings highlight the potential of ACO and TS as promising optimization strategies for automated NLME model selection.</p>