Computational & AI Journal of Chemical Information and Modeling (ACS)

Clearance prediction is the main source of exposure error in ML–PBPK pipelines

  • Graph-based models predicted unbound fraction, clearance and volume of distribution, which fed fixed compound-specific PBPK models to simulate human exposure.
  • Swapping observed inputs for predicted ones one at a time, in a small compound set, showed clearance degraded exposure predictions most.
  • Graph-plus-descriptor models matched descriptor-only baselines, Cmax stayed underpredicted, and the authors expect optimism for new chemistry; retrospective only.
Points Based on the abstract

Clearance dominating the error is unsurprising; it always has in human PK. What I'd want is the error budget in fold-error terms versus a simple in vitro–in vivo extrapolation baseline using measured microsomal or hepatocyte data. If the ML clearance doesn't beat that, the pipeline adds complexity without payoff. And with a small set and no scaffold-held-out test, I wouldn't trust it on a new series.