OpenADMET CYP3A4 Inhibition Models

Train and cross-validate OpenADMET models on CYP3A4 IC50 data in parallel, compare them statistically, and predict new compounds with the best one.

by Temple Compute

Run OpenADMET CYP3A4 Inhibition Models on Horus

W-38 ยท OpenADMET CYP3A4 Inhibition Models

Train and cross-validate OpenADMET models on CYP3A4 IC50 data in parallel, compare them statistically, and predict new compounds with the best one.

Domain: OMSF

Overview

Predicts CYP3A4 inhibition (pIC50) with OpenADMET v0.2.0 models, ported from the OpenADMET Showcase in the OMSF workflow directory. It writes one openadmet anvil cross-validation recipe per model, trains all models in parallel (one map clone each), compares them statistically with openadmet compare and predicts the compound set with the model with the best mean CV R2. The workflow has 5 tasks.

Pipeline

make_recipes  cleaned training data       --> data/cleaned.csv, recipes/<model>/{recipe.yaml,data.csv}
train         openadmet anvil, 1 per model --> trained/<slot>/model/
compare       openadmet compare           --> results/comparisons/
predict       best mean CV R2 model       --> results/predictions.csv
report        CV metrics + top compounds  --> results/metrics.csv, results/top_predictions.csv

compare and predict both run after train; predict picks the best model itself.

Inputs / Outputs

Parameters

| id | Default | Meaning | |---|---|---| | dataset | ChEMBL35 CYP3A4 parquet | Training set | | compounds | examples/compounds_for_inference.csv | Compounds to predict | | target_name | CYP3A4 | Biological target label, written to the cleaned data's Target column and the recipes' biotargets | | models | lgbm, rf, xgboost, catboost | Models to train; chemprop can be added | | cv_folds | 5 | k of the repeated k-fold CV (2-10) | | cv_repeats | 5 | Repeats of the k-fold CV (1-10) |

References