CalibratedClassifierCV adjusts classifier probability estimates so confidence better reflects observed outcomes. In CYP3A4 inhibition prediction, it supports Platt scaling with sigmoid regression or flexible isotonic regression. section 5.1.3
Use calibration carefully:
- Fit the calibrator on validation data and assess it on untouched test data; reusing data produces optimistically biased results. section 5.1.3
- Prefer holistic assessment with reliability diagrams and calibration metrics; isotonic regression may lower RMSCE while leaving probability bins empty. section 5.1.3
Calibrating a logistic regressor
Wrap a fitted logistic model in FrozenEstimator, fit CalibratedClassifierCV on validation fingerprints, and generate test probabilities with predict_proba using method='sigmoid' or method='isotonic'.
