Predict absorption

Draw a structure, or paste a SMILES and it will be drawn. The answer comes back as a bar with a shaded interval, because that is the width the data supports.

Reference compounds, and where they sit

Where these compounds sit, by what was measured in people. Your own result is drawn on the same ruler when you run one, so a prediction can be read against compounds you know rather than on its own. Press Load on any of them to put it in the editor.

Percent of an oral dose absorbed. Every measured value is read from the released dataset. Press Load to put a compound in the editor, then predict it.

SMILES

How to read the bar

The shaded band is the interval, and it is set by the error this model makes on compounds it never saw, measured over 250 repeated extractions of the descriptor set that ships. Across all of them the root mean square error is 20.47 plus or minus 1.71 percentage points and the correlation is 0.76 plus or minus 0.05.

The interval you see is not that overall figure. Error depends on where a compound sits, so the band drawn around a prediction is the error measured on the predictions that landed in the same range: 18.23 points for a compound predicted below 30 percent, 29.71 between 30 and 80, and 15.67 at 80 and above. Those bands are grouped by the predicted value, because that is the only thing known about a compound nobody has measured.

The bar is a decision rather than a measurement. At a 30 percent threshold the call is right 92.64 percent of the time. Two compounds that are both well absorbed cannot be ranked against each other: the gap between them is smaller than this error, and smaller than the disagreement between published measurements of the same drug.

The deployed absorption model reads 25 descriptors, the fixed consensus set the search settled on. Every one of them is listed on Data.

The bar is useful for telling a well absorbed compound from a poorly absorbed one and for ranking a series against itself. It is not useful for telling 85 apart from 90. If two compounds in a series have overlapping bands, the model has not separated them.

A prediction from a sparse or distant region of the training set is flagged in a box above the result, not in a footnote. The flag is part of the answer: it says the model has nothing reliable to offer for this structure. It is never hidden and never quietly filled in. The full method is on How it works and the training data on Data.

Caco-2 permeability, small molecules

A small molecule is scored by both models that apply to it in one request: percent absorbed above, and apparent permeability across a Caco-2 monolayer here. They are different quantities in different units and are never combined. Each says for itself whether the structure falls inside the region it was fitted in.

Ten compounds from the small molecule Caco-2 set, one per decile of what was measured, so they span the range rather than cluster in it. They carry the identifiers the released challenge data gives them and no common names, so point at or tap a name on the ruler to see its structure. Your own result is drawn on the same ruler when you run one. Press Load on any of them to put it in the editor.

Apparent permeability across a Caco-2 monolayer, apical to basolateral, as log10 Papp with Papp in cm/s. A less negative number is more permeable. Every measured value is read from the released data.