A classification model for blood brain barrier penetration

Traditional experimental approaches to evaluate the Blood-Brain Barrier (BBB) permeability of a drug are expensive and time consuming. Hence, several computational models have been developed over time to estimate propensities of compounds to penetrate the BBB. In this study, we aimed to build improved BBB classification models using a large curated dataset of 605 compounds with two classification thresholds (threshold-1: Brain/Plasma ≥ 0.6 as BBB+ and Brain/Plasma<0.6 as BBB- and threshold-2: Brain/Plasma>0.6 as BBB+ and Brain/Plasma<0.3 as BBB-). This dataset was split into a training set of 479 compounds for threshold-1, 432 compounds for threshold-2 and a test set of 126 compounds for threshold-1 and 110 compounds for threshold-2. A single model could not predict similar results for each dataset in case of two thresholds. Hence, consensus model building was employed on the modelling set that gave similar results for each of the datasets for two thresholds. The consensus model performed better on overall prediction datasets (a test set with 126 compounds and a WDI dataset with 1425 compounds for threshold-1 and a test set with 110 compounds and the WDI dataset for threshold-2), with accuracies of 86% and 87% for threshold-1 and threshold-2, respectively. The prediction performance of our consensus model was better than other existing models, by the criteria of percent accuracy, Matthew's correlation coefficient, sensitivity, specificity and Correct Classification Ratio. An analysis of substructure moieties among BBB + compounds showed a list of moieties that were present more among BBB + compounds than among BBB- compounds. These findings corroborate with the results of similar analyses reported earlier. The BBB prediction models developed in this study would be quite useful for screening compounds very early on in drug discovery projects.