Validation of AI-enabled surrogate models in quantitative systems pharmacology: a practical, context-of-use-driven review

Quantitative systems pharmacology (QSP) models support model-informed drug development but are computationally expensive, limiting workflows requiring many simulations. Artificial intelligence (AI)-enabled surrogates can accelerate these workflows, but require validation aligned with context of use and regulatory expectations. Here, we present an eight-step validation framework covering simulation-campaign design, baseline comparison, endpoint accuracy, trajectory fidelity, uncertainty calibration, distributional agreement, biological plausibility, and computational efficiency. The framework is illustrated through chimeric antigen receptor T cell (CAR-T) immunotherapy and hybrid neural ordinary differential equation (ODE) erythropoiesis examples and aligned with ICH M15 and ASME V&V40. Supplementary algorithms provide practical validation recipes.