Delower/dpp4-pic50-extra-trees

extra-trees-morgan1024-physchem

Features: morgan1024, physchem

Model: the original LazyPredict ExtraTrees model, a sklearn Pipeline (ColumnTransformer + ExtraTreesRegressor, 100 trees). Unpickling model.pkl requires lazypredict (listed in requirements.txt). Inference uses infer.py, which builds a pandas DataFrame with columns feature_0..feature_1029 (Morgan 1024 first, then the 6 physchem descriptors in the order MolWt, MolLogP, TPSA, NumHDonors, NumHAcceptors, NumRotatableBonds).

Metrics (original LazyPredict run)

Test set (80/10/10 split, seed 42): RMSE 0.4449, MAE 0.1437, R2 0.8668. FDA-12: RMSE 0.9637, MAE 0.6794 (11 of 12 compounds with a wetlab pIC50).

See metrics.csv and fda.csv in this repo for the full tables.

Inference

from infer import load_model, predict_smiles
model = load_model("model.pkl")
print(predict_smiles(model, ["CCO"]))

See infer.py, config.json and requirements.txt in this repo.