Rowan Docs

ADME-Tox Prediction

Python API · stjames models · API example

How it works

Predict absorption, distribution, metabolism, excretion, and toxicity properties from molecular structure using ADMET-AI. Use the results to compare compounds and flag properties that need experimental follow-up.

Settings

The model and prediction protocol are fixed; there are no workflow-specific settings.

Notes

Rowan uses the v2 checkpoints of ADMET-AI, developed by Kyle Swanson and co-workers. Predictions are averaged over five models per endpoint. The workflow also calculates physicochemical descriptors such as molecular weight, logP, hydrogen-bond counts, and polar surface area from the structure.

Results are grouped into absorption, distribution, metabolism, excretion, toxicity, and physicochemical properties:

  • Classification outputs, such as CYP inhibition, blood–brain barrier penetration, and hERG inhibition, are probabilities from 0 to 1. A higher value indicates a greater predicted likelihood of that endpoint; whether that is desirable depends on the property and your goal.
  • Continuous outputs have property-specific units or scales. For example, half-life is in hours, while aqueous solubility is reported as log(mol/L). Check the units before comparing values.
  • The Lipinski score counts the number of rules satisfied, from 0 to 4; it is not a count of violations.

For unphysical negative predictions, the results table applies a post hoc cutoff: half-life and clearance values are displayed as zero. Raw API results can retain negative values. This display adjustment does not improve prediction accuracy.

These predictions can help prioritize compounds; they do not replace experimental measurements. Colored result categories are screening guides, and a favorable prediction does not establish safety or efficacy.

Submission video

Further reading