Blood-Brain Barrier Exclusion (BBBX) Score: Designing Drugs Not to Enter the Brain

The blood-brain barrier (BBB) is an important physiological barrier that protects the brain by preventing entry of harmful substances into the central nervous system (CNS). While this barrier is effective at blocking most drugs from disrupting the tightly controlled environment of the CNS, certain classes of non-neurologic drugs can cross the BBB, potentially causing neurological side effects. Developing safer alternatives to these drugs requires a reliable method for predicting the BBB non-penetrance, whilst maintaining a high level of drug-like properties to the molecules. Existing regression methods for BBB penetration are inaccurate for this task, tending to overestimate logBB values due to imbalances in available datasets. To address this, we have developed a LightGBM-based model with a custom multi-task loss function that learns from both regression and classification data. Improving performance across classification and regression metrics, our model outperforms existing models in identifying non-penetrant molecules and achieves regression accuracy approximately equal to experimental uncertainty. Based on this, we introduce the BBB Exclusion (BBBX) Score, a scoring method which stratifies compounds by predicted BBB penetrance and confidence. To interpret model outputs, SHAP analysis was used to identify features contributing most to model outputs. We use these features to define threshold values associated with increased likelihood of BBB Exclusion whilst maintaining high oral bioavailability and drug-like properties required for developing a safe, non-CNS drug.