Logotyp: till Uppsala universitets webbplats

Population pharmacokinetic (PK) models inform dose optimization and exposure prediction in drug development, with parameter estimates directly determining dosing recommendations. Reliable parameter estimation is essential for patient safety yet challenging when automated model development workflows generate numerous candidate structures. The choice of initialization strategy can significantly influence estimation stability, convergence success, computational cost, and model ranking.A simulated moxonidine population PK dataset was analyzed using 36 candidate model structures generated by an automated model development workflow. Each model was fitted using a baseline estimation using default, rule-based starting values generated by the automated workflow, and multiple alternative initialization strategies, including four heuristic approaches: Latin Hypercube Sampling (LHS), Simple Random Sampling (SRS), Saddle Reset, and Monte Carlo ETA (MC-ETA) with four sample sizes (10, 100, 1,000, and 10,000) as well as a metaheuristic global optimization approach based on Particle Swarm Optimization (PSO) implemented via pyPESTO with NONMEM as the local solver. Methods were compared in terms of changes in objective function value (ΔOFV) relative to baseline, runtime, convergence success rates, and identification of highest-ranked models.Across most model structures, heuristic methods (LHS, SRS, and Saddle Reset) produced ΔOFV values near baseline, indicating stable parameter estimation. MC-ETA showed greater variability, with larger sample sizes occasionally yielding lower OFV values at increased computational cost. PSO achieved successful optimizations for all models but exhibited substantially longer runtimes and variable ΔOFV outcomes. Highest-ranked models identified by heuristic methods remained largely consistent with baseline results, whereas PSO selected different optimal structures in several cases.Initialization strategy influenced estimation robustness and computational efficiency more than final model selection for most candidate structures. Heuristic methods provided computationally efficient and stable performance suitable for routine drug development, while PSO offered exhaustive global exploration at substantially higher cost. These findings support evidence-based selection of initialization strategies in pharmaceutical practice: heuristic methods are appropriate for routine model development, whereas metaheuristic approaches may be justified for complex models or regulatory submissions requiring maximum parameter confidence. The results highlight trade-offs between efficiency and global exploration in population PK model estimation and support complementary use of heuristic and metaheuristic strategies in automated model development for model-informed drug development.