OpenADMET CYP3A4 Inhibition Models (CPU)

Horus port of the OpenADMET from the OMSF workflow directory.

by Temple Compute

Run OpenADMET CYP3A4 Inhibition Models (CPU) on Horus

What it does

A Horus port of the OpenADMET Showcase from the OMSF workflow directory. It builds models that predict how strongly small molecules inhibit CYP3A4, a liver enzyme behind many drug-drug interactions, and uses the best one to score new compounds.

1. Fetch data: downloads the ChEMBL35 CYP3A4 IC50 aggregate (about 4,900 compounds, from OpenADMET data-catalogs) and the compounds to predict.
2. Clean data and write recipes: adds the OpenADMET standard columns and writes one openadmet anvil recipe per model.
3. Train (parallel): trains LightGBM, random forest, XGBoost and CatBoost at the same time on RDKit 2D descriptors + ECFP4 fingerprints, with repeated k-fold cross-validation.
4. Compare models: openadmet compare runs ANOVA, Levene and Tukey HSD tests across the models' CV scores and writes the plots.
5. Predict: picks the model with the highest mean CV R2 and predicts pIC50 for every input compound.
6. Report: one table of CV metrics per model and a list of the top 50 predicted inhibitors.

Everything runs on CPU in one pinned conda environment (openadmet-models v0.2.0, CPU builds of PyTorch and CatBoost). No GPU needed.

How to use it

Input: the URL of a CSV with the compounds to score. It needs an OPENADMET_CANONICAL_SMILES column; an OPENADMET_INCHIKEY column is copied to the top list if present. The URL must be reachable from the machine that runs the workflow (for example a raw GitHub link or a presigned S3 URL). Leave the default to score the showcase's 10,000 ZINC compounds.

Settings (keep the defaults unless you need otherwise):

| Setting | Default | Meaning | |---|---|---| | Models | lgbm, rf, xgboost, catboost | Models to train and compare. Chemprop is also accepted, but takes hours on CPU. | | CV folds | 5 | k of the k-fold cross-validation | | CV repeats | 5 (upstream) | Repeats of the cross-validation. 1 is about 5 times faster and fine for a quick look. |

Outputs

• results/metrics.csv: mean and 95% CI of R2, MAE, MSE, Spearman and Kendall tau per model, and which model was picked
• results/top_predictions.csv: the 50 compounds with the highest predicted pIC50
• results/predictions.csv: the prediction for every compound
• results/comparisons/: statistical comparison report (PDF plots and JSON)
• trained/<slot>/model/: each trained model with its CV metrics and plots

Runtime: with CV repeats 1 the whole workflow takes about 9 minutes on 16 CPUs, plus about 3 minutes the first time to build the environment. Training uses 8 threads per model, so a 32-vCPU machine trains all four at once.

Credits: science, models and data by the OpenADMET contributors (Apache-2.0); training data derived from ChEMBL 35 (CC BY-SA 3.0).