In silico prediction of blood-brain partitioning using a chemometric method called Genetic Algorithm Based Variable Selection

A chemometric method called Genetic Algorithm Based Variable Selection (GAVS) was developed for Quantitative Structure-Property Relationship (QSPR) study here. Different from other GA based techniques, improvements have been made in GAVS, including fitness function and the strategy of elite warehouse. In the case of modeling the Blood-Brain Barrier (BBB) penetration, GAVS performed very well in variable selection. The training set consists of 151 structurally diverse compounds 154 molecular descriptors of which were calculated by E-Dragon. The best model generated by GAVS contained six descriptors with r=0.848, q=0.817, F=61.40. The model was further validated by two external test sets, with the predicted correlation coefficient of 0.837 for Test set I which contained 28 compounds and correct rate of 82.42% for Test set 11 which included 91 drugs or drug-like compounds with known BBB indication. To the CNS library containing 1283 CNS activities, the model performed with a success rate of 97.3% in predicting CNS active compounds. In comparison with other methods, GAVS performed best among Genetic Function Approximate (GFA) method and other published models. The established predictive model could be applied in design of combinatorial libraries or virtual screening against Central Nervous System (CNS) targets.