1. Applications of genetic algorithms to the prediction of blood-brain partitioning of a large set of drugs

In this study, the relationships between the brain-blood concentration ratio of 96 structurally diverse compounds with a large number of structurally derived descriptors were investigated. The linear models were based on molecular descriptors that can be calculated for any compound simply from a knowledge of its molecular structure. The linear correlation coefficients of the mod- els were optimized by genetic algorithms (GAs), and the descriptors used in the linear models were automatically selected from 27 structurally derived descriptors. The GA optimizations resulted in a group of linear models with three or four molecular descriptors with good statis- tical significance. The change of descriptor use as the evolution proceeds demonstrates that the octane/water partition coefficient and the partial negative solvent- accessible surface area multiplied by the negative charge are crucial to brain-blood barrier permeability. More- over, we found that the predictions using multiple QSPR models from GA optimization gave quite good results in spite of the diversity of structures, which was better than the predictions using the best single model. The predic- tions for the two external sets with 37 diverse com- pounds using multiple QSPR models indicate that the best linear models with four descriptors are sufficiently effective for predictive use. Considering the ease of computation of the descriptors, the linear models may be used as general utilities to screen the blood-brain barrier partitioning of drugs in a high-throughput fashion.