EnsembleBBB: Enhanced accuracy in predicting drug blood-brain barrier permeability with a Machine Learning Ensemble model

Abstract

Abstract The blood-brain barrier (BBB) serves as a selective and semi-permeable barrier, crucial for maintaining homeostasis within the central nervous system. When developing drugs that act on the brain, understanding the permeability of compounds across the BBB is of utmost importance. However, succinctly formulating this consideration poses a challenge. Clinical experiments are the most accurate method for assessing BBB permeability, but they are time and cost consuming. Hence, computational methods have been explored as an alternative approach to predict BBB permeability. Nevertheless, the issue of accuracy has persistently plagued BBB permeability prediction models. To enhance the precision of BBB permeability prediction, we employed ensemble methods based on popular machine learning algorithms. Our models were trained using a dataset of 7,807 diverse compounds, each encoded with different molecular binary fingerprints. The predictive performance of the developed models was assessed and compared with the literature. We found out that Random Forest algorithm and MACCS fingerprints perform best. Notably, the ensemble model with MACCS fingerprints yielded an AUC of 0.95 in the testing set and a mean AUC of 0.94 in 5-fold cross validation. The applicability domain was evaluated using the William plot, which indicated that the MACCS dataset had the fewest outliers, while the PubChem dataset had the highest number of outliers. The most important features were calculated using the built-in features importance in Random Forest. Furthermore, our model achieved an impressive accuracy of 95% and an AUC of 0.92 in predicting BBB permeability of an external dataset used for benchmarking. This model holds significant promise for facilitating the screening of compounds based on their BBB permeability during the preliminary stages of drug development.