Blood-Brain barrier Penetration
The blood-brain barrier (BBB) is composed of tightly packed endothelial cells that restrict the ability of substances carried in the bloodstream from passing through.
Why is Blood-Brain Barrier Penetration Important?
While protecting the bodies central nervous system (CNS), this presents significant difficulties for the development of drugs specifically designed to target problems within the CNS. Conversely, it should be remembered that researchers developing drugs targeting non-CNS targets will want to ensure that these drugs do not cross the BBB otherwise there will be an elevated risk of side-effects.
Asteris has two blood-brain barrier penetration models, log([brain]/[blood]) (logBB) and a classification model (BBB Category), both of which have been included to allow determination of a consensus score. This approach provides a higher level of confidence because the models were developed independently. Hence, compounds predicted to cross the blood-brain barrier by both models will have a higher consensus score than those only predicted to cross the blood-brain barrier by one model.
The data set used to build the logBB model consists of 509 structures with a reported logarithm of the concentration ratio between brain tissue and plasma (log(BB)) which were derived from various literature sources. The structures were assigned randomly to training (n = 359), internal evaluation (n = 75) and independent test (n = 75) sets. The latter was excluded from the model development process. The model was built automatically by the StarDrop Auto-Modeller using the standard settings. The model is a Radial Basis Function technique combined with a genetic algorithm to assist in descriptor selection (GA-RBF).
The model predicts the log(BB) value for each compound, along with an estimate of the RMSE in prediction. The distance of each predicted compound from the descriptor-space of the training set, referred to as the chemical space of the model, is calculated in order to gauge the validity of the results. The RMSE in prediction for compounds within the chemical space is 0.36 log units and the RMSE in prediction for compounds outside, but in close proximity to, the chemical space is 0.54 log units.
It is a feature of the RBF technique that it will usually provide a perfect fit for the training set. However, on the test set the model achieves an R2 of 0.72 with an RMSE of prediction of 0.36 log units.
The data used to build the BBB classification model consists of 201 structures classified as BBB+ and BBB- that are reported in literature models. This data was divided into a training set containing 101 compounds with an even distribution between BBB+ and BBB- compounds and an internal evaluation set of 48 compounds, with a 3.5:1 ratio between BBB+ and BBB- compounds, which was used to monitor the training of the model. The remaining 52 structures were utilized as an independent test set with a 1:2 ratio of BBB+ and BBB- compounds. The model generates a prediction for each compound as BBB crossing (+) or non-crossing (-). This is based on a nominal classification boundary of log(BB)=-0.5 between BBB- and BBB+ compounds. For the independent test set, 91% of BBB- predictions were correct in relation to the known category, whereas BBB+ predictions were correct in 83% of cases. Overall training and test set classifications were 96% and 93% correct respectively. The model also correctly predicted the BBB+ category for 18 of the top 20 best-selling drugs in 2003. The only incorrect predictions were for compounds identified as substrates for active uptake or efflux transporter proteins.