Predicting human clearance of new compounds from structure within 2-fold for most compounds

• IVIVE accuracy. Reviews report that most hepatic IVIVE predictions miss by more than 2-fold: a weighted analysis found about 66.5% of predictions fell outside 2-fold of the observed in vivo clearance for both microsomes and hepatocytes, and Wood et al. (Drug Metab. Dispos. 2017) reported about 75% outside 2-fold; systematic under-prediction of 3- to 10-fold is common, corrected empirically by regression offsets. For aldehyde-oxidase substrates the reported geometric mean fold errors are large (roughly five- to ten-fold depending on the in vitro system), and empirical scaling factors are reported to raise the share of predictions within 2-fold substantially; the exact values are review-reported and were not re-checked against the underlying datasets for this entry.
• Machine learning. Reported within-2-fold rates for structure-based models are typically 40-60%, usually from retrospective cross-validation on public data: one published multivariate model on several hundred compounds reported roughly 60% within 2-fold (figures as reported by the authors, retrospective). A 2025 ChemRxiv preprint (FLEX-SSS, rat-based allometry plus random forest) reported about 42% within 2-fold and GMFE 2.7 on an external set of 62 compounds. These studies differ in inputs and are not comparable, and none was a large prospective blinded test.
• Not shown. No published method meets the target above prospectively; no standard blinded community challenge for human clearance prediction is established as of October 2026.

Accurate prediction would let teams prioritise compounds with human-appropriate clearance before synthesis, reduce animal studies, improve first-in-human dose selection and reduce late exposure-related attrition. Even calibrated uncertainty would be valuable in deciding which experiments to run.

A refereed, prospective, blinded study (for example a consortium challenge with undisclosed clinical clearance values) in which one method meets criteria (i)-(v) across at least 200 compounds and the five mechanism classes, with the method reproducible by outside groups.