Answers are grounded in this paper and cite their sources.
End-to-end structure-to-exposure workflows combine molecular representation, pharmacokinetic (PK) input prediction, and physiologically based pharmacokinetic (PBPK) simulation, but aggregate accuracy does not reveal where error enters the workflow. We developed graph-based models for predicting plasma unbound fraction, systemic clearance, and steady-state volume of distribution and used paired input substitution in fixed compound-specific PBPK models to trace the downstream effects of these predictions. On held-out benchmark sets, models combining graph-derived and physicochemical features performed comparably to descriptor-based controls, with no detectable paired differences. In 41 compounds, replacing observed clearance with predicted intrinsic or systemic clearance caused the largest deterioration in exposure prediction. AUC0–t 2-fold coverage decreased from 82.9% to between 41.5 and 53.7%, whereas substitutions of physicochemical properties, plasma unbound fraction, and volume of distribution had smaller effects. The intrinsic-clearance and systemic-clearance workflows were not detectably different under matched input slots. Persistent Cmax underprediction with observed inputs indicated an additional limitation in distribution-model structure. When all three PK inputs were predicted, AUC0–t and Cmax were within 2-fold for 41 and 59% of compounds. Because encoder exposure may favor performance in this set, these coverage estimates are upper bounds for unseen chemistry. The paired-substitution design therefore identifies limited clearance prediction accuracy, rather than the PBPK clearance parametrization strategy, as the principal source of exposure error in the evaluated workflow and separates it from residual PBPK structural error.
No takes yet. Share an insight, caveat, or question.
Cai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: