BBB Prediction Benchmarks

Arion's 13-signal ML ensemble vs. published rule-based methods on MoleculeNet BBBP held-out test set (408 compounds).

99.6%
Arion AUC
77.8%
SwissADME AUC
8113
Training Compounds
Method Comparison
Method Type AUC-ROC Accuracy Balanced Accuracy F1 Score
Arion Ensemble (RF+XGB) ML ML Ensemble
0.996 		0.985 0.980 0.990
Arion RF (Morgan FP) ML ML Ensemble
0.992 		0.968 0.947 0.979
Arion XGBoost (Descriptors) ML ML Ensemble
0.997 		0.980 0.976 0.987
SwissADME BOILED-Egg Rule Rule-based
0.778 		0.760 0.778 0.826
BBBscore (Gupta 2019) Rule Rule-based
0.827 		0.789 0.678 0.866
Clark Rules (2003) Rule Rule-based
0.742 		0.733 0.742 0.806
Lipinski CNS Rule Rule-based
0.728 		0.706 0.728 0.781
ROC Curves
Calibration (Ensemble)
Test Set: MoleculeNet BBBP (held-out) | Size: 408 compounds | BBB+: 312 (76%) | Training: 8113 compounds (B3DB + MoleculeNet BBBP)
Try BBB Predictor