A significant number of models have been published for the prediction of blood-brain barrier (BBB) permeability, but most of these publications are based on training sets with a limited number of data and collected from different bibliographic sources, implying a poor quality of the resulting predictions.

Other alternative is the generation of models able to predict the ability of compounds to cross the BBB and to act at some level in the CNS. Ajay et al. used the CMC and MDDR databases, comprising over 15,000 compounds with some kind of CNS activity and 50,000 ones with no CNS activity mentioned. A neural network was able to correctly predict about 80 % of the compounds when all the 1D and 2D descriptors were used.

In order to improve the reliability of our model, we based it on a limited set of CNS active and CNS inactive drugs (about 1,000) for which the information about CNS activity and secondary effects is very well known. The most performing and reliable QSAR model was based on a combination of 19 constitutional and topological descriptors using an axis-parallel decision tree, with about 90 % of the compounds correctly classified in the learning and in the validation set.

The model was systematically used on our collection of libraries to select the most potent CNS active compounds. A diverse selection can be downloaded on our website. Feel free to download the pdf brochure and to request a login password to have access to the structures.