Case study
ML Prediction of Compound Bioactivity
Machine-learning bioactivity classifier over 18,432 ChEMBL compounds for virtual-screening triage — 208 RDKit descriptors plus Morgan fingerprints, four benchmarked models (XGBoost best, ROC-AUC 0.938), probability calibration, and SHAP feature attribution.
Problem
Virtual screening produces far more candidates than can be tested. The aim was a classifier that triages compounds by predicted bioactivity.
Method
A classifier was trained on 18,432 ChEMBL compounds described by 208 RDKit descriptors plus Morgan fingerprints. Four models were benchmarked, with probability calibration and SHAP feature attribution.
Result
XGBoost performed best, with a ROC-AUC of 0.938.


