ADMET-AI 2.0 and RDKit notebook tutorial profiles absorption, metabolism and toxicity for CDK2 inhibitors
- The tutorial walks through an annotated Jupyter notebook that combines RDKit preprocessing with a graph neural network ADMET predictor, covering drug-likeness, structural alerts, metabolism, excretion and toxicity.
- It is demonstrated on a set of CDK2 inhibitors curated from ChEMBL, with each compound's predictions ranked against approved drugs and combined into a composite score.
- Results are purely in silico with no experimental validation, and the main outcome is a flagged liability profile (notably cardiac and liver risk) plus an accessible, FAIR-style template rather than a benchmark against other tools.
💡 Points Based on the abstract
Treat this as a template, not evidence. Flagged hERG and liver risk across a kinase series is exactly what an over-sensitive model produces, and without experimental data or a comparison to other predictors the flag rates are uninterpretable. Useful next step: run it on your own series with known assay results and check calibration before trusting the ranking.