Sim2Real-PK framework for precision dosing: Integrating pharmacokinetic mechanisms and deep learning to decode clinical heterogeneity from sparse data

Abstract Precision dosing is a cornerstone of individualized therapeutics, yet its implementation is frequently hindered by clinical pharmacokinetic data that are characteristically sparse, irregularly sampled, and highly heterogeneous. Traditional mechanistic models provide structural interpretability but often lack the flexibility to capture complex patient-specific deviations, whereas purely data-driven deep learning models are prone to pharmacological inconsistency and poor generalization in low-data regimes. Here we present Sim2Real-PK, a hybrid neural-mechanistic framework that reconceptualizes pharmacokinetic forecasting as a time-aware latent state-space trajectory inference task. Sim2Real-PK first pretrains a neural dynamics core on massive synthetic datasets derived from mechanistic ordinary differential equation systems to internalize universal kinetic laws, then adapts to real-world clinical heterogeneity through a constrained residual correction pathway. Validated on individual participant data from the YODA Project for two drugs representing distinct pharmacological regimes—ibrutinib and daratumumab—Sim2Real-PK significantly outperformed conventional population-level baselines. The framework improved the F50 for ibrutinib from 32.1% to 39.9% and the F20 for daratumumab from 38.8% to 51.1%, while faithfully recovering expected linear and nonlinear pharmacological behaviors. Crucially, the model successfully resolved latent pharmacokinetic heterogeneity within patient cohorts that remained obscured under traditional analytical frameworks. These results demonstrate that mechanistically-anchored neural adaptation can bridge the gap between biological rigor and predictive agility, providing a robust foundation for next-generation model-informed precision dosing.