Welcome to B3-Pred

B3-Pred is a machine learning-driven webapp, based on python scripts designed to predict blood-brain barrier permeability (BBBP) of drug-like molecules using an optimised XGBoost model. This model integrates global physicochemical descriptors, structural fragments from BRICS decomposition, and molecular composition data to deliver robust and interpretable permeability predictions.

About the Model

The B3-Pred model was developed using the B3DB dataset, containing 7807 molecules classified as BBB-penetrating (BBB+) or non-penetrating (BBB-). Each molecule was fragmented using BRICS rules and custom fragmentation algorithms to capture rings and side chains. Features from whole molecules and fragments were computed using RDKit and optimised using Optuna.

Scientific Approach

B3-Pred combines physicochemical properties (e.g., LogP, TPSA, MW) with composition counts/descriptors and fragment counts. The model employs SHAP and gain-based feature importance analysis to ensure interpretability, highlighting both global descriptors and predictive substructures.

Performance and Validation

The XGBoost model achieved a 40-optomised median model ROC-AUC of 0.96, specificity of 84%, sensitivity of 94%, and MCC of 0.79. Performance was validated using 10-fold cross-validation, with final evaluation on an independent test set. Comparisons with data from PubChem and ChEMBL ensured robustness, and SHAP analysis confirmed alignment with known permeability rules such as Lipinski's Rule of 5.