Prediction variability in physiologically based pharmacokinetic modeling of tissue disposition under deep uncertainty

Physiologically based pharmacokinetic (PBPK) models are increasingly invoked in virtual screening workflows for therapeutics. These mechanistic models project pharmacokinetic (PK) outcomes from molecular properties, with data-driven models acting as intermediaries to map molecular structure to PBPK input parameters. Errors in predicted parameters and unvalidated assumptions within PBPK models expose PK predictions to deep uncertainty. Herein, we examine how these uncertainties affect the prediction variability of dynamic, tissue-specific exposure. We validated four PBPK models against 1854 experimental datapoints-to establish their predictive fidelity before introducing parameter uncertainty typical of property-prediction models. Depending on molecule properties and model choice, the coefficient of variation under parameter uncertainty ranged from 10^-6 to 31 for predicted PK statistics. Further, we identified notable model disagreement for a subset of drug-like chemical space characterized by lipophilic, protonated molecules. Uncertainty quantification revealed biophysicochemical properties and parameter interactions that drove disagreement and highlighted model assumptions that exacerbated prediction variance. Our findings delineate the challenges presented by deep epistemic uncertainty in PBPK modeling.