Predicting blood-brain barrier permeability of chemicals by machine learning modeling.

In vitro studies are widely used to measure chemical neurotoxicity or neuroactivity. To show similar activity in vivo, chemicals need to cross the blood-brain barrier (BBB), a membrane that controls chemical movement from the bloodstream to the brain. Machine learning (ML) models present a rapid, cost-effective alternative to experimental methods for predicting BBB permeability. We developed and validated a set of ML models to predict BBB permeability by employing various molecular descriptors and ML algorithms. Models were rigorously evaluated using 5-fold cross-validation and scaffold splitting approaches. Balancing techniques were utilized to address dataset imbalance. Predictions from individual models were combined into a consensus model, which demonstrated the most robust and stable performance. After applying data-balancing techniques, the final consensus model achieved a balanced accuracy of 0.81 for the scaffold splitting validation. SHAP (SHapley Additive exPlanations) analysis identified the topological polar surface area (TopoPSA) as the most critical molecular feature affecting BBB permeability. Additionally, we identified 88 molecular fragments frequently found in BBB-permeable compounds, with nitrogen-containing rings being the most common. Regression model development was more challenging; the Graph Neural Network model performed best with a coefficient of determination (R2) of 0.48. The practical utility of the consensus model was demonstrated by screening 6031 compounds identified by in vitro assay as dopamine D2 receptor (DRD2) antagonists. The consensus model predicted that 27% of these compounds would not cross the BBB in vivo, highlighting how this model can be used to deprioritize in vitro assay testing of chemicals that will likely produce false positive results.