Abstract
Background: Quantitative Systems Pharmacology (QSP) models are increasingly used to investigate drug effects in complex biological systems. A key component of QSP analyses is the generation of virtual populations (VPops), which reproduce observed biological variability and enable the quantification of uncertainty in model predictions. However, VPop generation is often computationally intensive due to the large number of model simulations required to identify plausible patients.
Methods: We propose a surrogate-assisted framework to accelerate the generation of plausible patients and virtual populations in QSP models. The approach employs Gaussian Process (GP) surrogate models trained on quantities of interest (QoIs) generated by the original mechanistic model. The surrogate is then used within established VPop generation workflows to guide parameter selection and reduce the number of expensive model evaluations required. The methodology was evaluated on three QSP models of increasing complexity and compared against conventional approaches, including Simulated Annealing, Nested Simulated Annealing, and Markov Chain Monte Carlo methods.
Results: Across the tested models, surrogate-assisted strategies substantially reduced computational costs associated with plausible patient and virtual population generation while maintaining comparable performance in terms of goodness-of-fit and preservation of target variability distributions. The balance between surrogate training cost and predictive accuracy was found to depend on model complexity, highlighting the importance of selecting an appropriate surrogate strategy for each application.
Conclusions: Gaussian Process surrogate models provide an effective means of accelerating virtual population generation in QSP workflows. By reducing computational burden while preserving population quality, the proposed framework can facilitate the application of increasingly complex mechanistic models in systems pharmacology and model-informed drug development.