Intestinal absorption from structure
Draw a molecule and get the fraction of an oral dose a human is predicted to absorb. The answer arrives as a band, because a band is what the data supports. A second and entirely separate predictor reports cyclic peptide membrane permeability on two assays that are never pooled.
Twenty one absorption units is a wide band, and the page says so
A prediction of 70 means the compound is probably somewhere between 50 and 90. This site is useful for separating well absorbed compounds from poorly absorbed ones and for ranking a series against itself. It is not useful for telling 85 apart from 90, and any reading that suggests otherwise is wrong. Every prediction is drawn as a bar with a shaded interval for that reason, and no prediction is ever shown without it.
The ceiling is in the data, not in the model
Published absorption values for the same drug disagree with each other. Of the 783 compounds in the source compilation, 262 carry more than one published value. Sixty one of those disagree by more than 20 points and 31 by more than 40. Methotrexate appears in the literature at 20, 59, 65, 70 and 100. A model cannot be more precise than the measurements it is fitted to, so a 21 unit error is close to the resolution the field currently has rather than a shortfall against it.
What the algorithm chose, twenty eight years apart
A genetic algorithm searched a pool of 1,301 modern descriptors over 781 compounds, twelve times, each time on a different split. Mean partial charge was chosen in 12 of 12 repeats and topological polar surface area in 11 of 12. The 1998 model, searching a pool that no longer exists with software that no longer runs, chose hydrogen bonding and geometric descriptors. Two unrelated descriptor pools, twenty eight years apart, landed on the same physics: how much charge a molecule carries and how much polar surface it presents.
Applicability is a feature, not a disclaimer
The model fails on molecules unlike its training set, and it fails quietly. The single worst error in validation was salicylic acid, predicted at 4.6 against a measured 100. It broke no descriptor range. It was simply small and sparse in the region these descriptors resolve. Every prediction therefore carries a distance to the training set, and a prediction drawn from a sparse region is flagged in the interface rather than buried in a footnote. A flag is an answer. It is never hidden and never filled in.
Two peptide assays, never pooled
The cyclic peptide predictor has two arms. The Caco-2 arm is fitted to 1,281 peptides measured on a human intestinal cell line. The passive arm is fitted to 7,298 peptides measured on an artificial membrane. On the 585 peptides that both arms contain, the two assays correlate 0.557 and sit 1.15 log units apart. That is a systematic offset, not noise, and the two populations differ in size as well. So the peptide page shows two predictions side by side, each labeled with its assay, and offers no combined figure and no conversion between them.
What it is for
Triaging a series before any of it is made, and knowing when the model has nothing useful to say. The full method covers the 1998 model, what reproduced, what did not, and what the larger data changed.