Abstract
The inefficiency of some medications to cross the blood-brain barrier (BBB) is often attributed to their poor physicochemical or pharmacokinetic properties. Recent studies have demonstrated promising outcomes using machine learning algorithms to predict drug permeability across the BBB. In light of these findings, our study was conducted to explore the potential of machine learning in predicting the permeability of drugs across the BBB. We utilized the B3DB dataset, a comprehensive BBB permeability molecular database, to build machine learning models. The dataset comprises 7,807 molecules, including information on their permeability, stereochemistry, and physicochemical properties. After preprocessing and cleaning, various machine learning algorithms were implemented using the Python library Pycaret to predict permeability. The extra trees classifier model outperformed others when using Morgan fingerprints and Mordred chemical descriptors (MCDs), achieving an area under the curve (AUC) of 0.93 and 0.95 on the test dataset. Additionally, we conducted an experiment to train a voting classifier combining the top three performing models. The best-blended model, trained on MCDs, achieved an AUC of 0.96. Furthermore, Shapley additive exPlanations (SHAP) analysis was applied to our best-performing single model, the extra trees classifier trained on MCDs, identifying the Lipinski rule of five as the most significant feature in predicting BBB permeability. In conclusion, our combined model trained on MCDs achieved an AUC of 0.96, an F1 Score of 0.91, and an MCC of 0.74. These results are consistent with prior studies on CNS drug permeability, highlighting the potential of machine learning in this domain.