This repo provides starter notebooks and example workflows for the OpenADMET CYP Inhibition Blind Challenge: A community benchmark for predicting Cytochrome P450 (CYP) inhibition across four major human isoforms: CYP3A4, CYP2C9, CYP2D6, and CYP1A2.
CYP enzymes drive the metabolism of most marketed small-molecule drugs. Unexpected CYP inhibition can unexpected trigger drug–drug interactions (DDIs) by slowing the clearance of co-administered drugs. Furthermore, time-dependent (mechanism-based) inhibition (TDI) can cause DDIs that persist long after the parent compound has been cleared. This challenge asks participants to build predictive models that predicts both CYP direct inhibition, as well as TDI.
For full challenge details and background, see the challenge announcement.
You'll need conda or mamba installed to set up the Python environment.
Create and activate the environment from environment.yaml:
conda env create -f environment.yaml # or mamba env create -f environment.yaml
conda activate oadmet_cyp_tutorial # or mamba activate oadmet_cyp_tutorialThis installs Python along with RDKit, scikit-learn, LightGBM, JupyterLab, and the other dependencies needed to run the tutorial notebooks.
Training and test data will be available on Hugging Face:
openadmet/cyp-challenge-train-test
The dataset covers a high-throughput dose-response curve (DRC) campaign across the four isoforms, including a dose-response training set (~1,500 DRCs per isoform), primary single-concentration data for those compounds, and a 750-compound analog-expansion test set assayed against all four CYPs.
-
Direct Inhibition Track (Regression): Predict direct-inhibition pIC50 for all 4 isoforms (CYP3A4, CYP2C9, CYP2D6, CYP1A2) across the 750 test-set compounds (4 continuous regression targets per compound).
- Primary metric: Macro-Averaged Soft-Threshold RAE (MA-ST-RAE), where error is measured as the distance between the predicted value and the credible interval bound of the fitted dose-response curve; predictions falling anywhere inside the credible interval incur zero error.
- Secondary metrics (MAE, R², Spearman ρ, Kendall's τ) are also reported with bootstrap confidence intervals.
-
Time-Dependent Inhibition (TDI) Track (Classification): Classify whether a compound shows significant time-dependent inhibition, evaluated on CYP3A4 and CYP2D6 only.
- Primary metric: Matthews Correlation Coefficient (MCC).
- Target Isoforms: Evaluated on CYP3A4 and CYP2D6 only (where TDI phenomena are clinically prominent).
- Label Definition: Positive means the TDI-arm IC50 shift exceeds 2-fold relative to direct inhibition, including inferred positives among low-activity compounds where a shift can't be measured directly.
Join the Discord server for Q&A and discussion in the #cyp-challenge channel.
All submission deadlines are 23:59 UTC.
| Date | Action |
|---|---|
| July 29, 2026 | Challenge announced |
| August 17, 2026 | Training/Test sets released; submissions open 🚀 |
| September 24, 2026 | Deadline for intermediate leaderboard submissions |
| September 25, 2026 | Intermediate leaderboard released (one-time full test set reveal) |
| November 3, 2026 | Submissions close 🏁 |
| From November 4, 2026 | Final leaderboard released, webinars, blog posts, and wrap-up |
In addition to leaderboard rankings, the OpenADMET team will present an award for the Most Innovative Machine Learning Approach(es), recognizing architectural novelty, creative data usage, and novel uncertainty quantification, regardless of final leaderboard rank.