Application of Population-Based Meta-Heuristic Algorithms for Robust Initialization in Mechanistic Pharmacometric...

Mechanistic PK/PD models represent complex biological systems in which model parameters are in nonlinear, multidimensional parameter spaces and often require robust optimization informed by experimental data, making parameter estimation a critical yet challenging aspect of pharmacometric analysis. Moreover, optimization methods for nonlinear-mixed effects PK/PD modeling are highly sensitive to initial parameter values, and poor initial estimates frequently result in convergence failure or suboptimal fits. To address the challenges, this study introduces a systematic framework that integrates population-based meta-heuristic algorithms to identify optimal or near-optimal initial estimates that facilitate subsequent nonlinear mixed-effects modeling. Nineteen distinct population-based meta-heuristic algorithms, including evolutionary, swarm-based, and bio-inspired methods, were evaluated on two mechanistic ODE-based models: (1) a two-compartment pharmacokinetic model with linear and Michaelis-Menten elimination and (2) the Friberg myelosuppression model. The temporal evolution of model goodness-of-fit over iterations was evaluated for each algorithm, and the effects of population size and iteration number on the algorithm performance were also examined. The results demonstrate population-based optimization algorithms efficiently and autonomously refine parameter estimates over iterations. Jellyfish search optimizer, symbiotic organisms search, and memetic algorithm showed strong performance compared to others with respect to optimization accuracy and computational efficiency across both models. As expected, increasing population size and iteration number enhanced the optimization performance. Overall, population-based meta-heuristic algorithms demonstrate superior capability in navigating nonlinear, high-dimensional parameter spaces, generating robust initial estimates for subsequent modeling workflows, and enhancing convergence stability, computational efficiency, and parameter identifiability.