Constructing a linear QSAR for some metabolizable drugs by human or pig flavin-containing monooxygenases using some molecular features selected by a genetic algorithm trained SVM

Abstract

A genetic algorithm optimized and feature selectable support vector machine (GFSVM) was designed for classifying some 71 and 151 substrates of human and pig flavin-containing monooxygenases (FMOs; EC 1.14.13.8) collected from the literatures. While a novel fitness function was designed, a feature mask for selecting (represented by bit 1) or masking (represented by bit 0) a feature was also implemented in the chromosomes generated during the evolution process. The feature selection was performed according to the ranked accumulated |w| values computed from several preliminary runs. Some numbers of top ranked features were then selected and gradually increased in a multiple linear regression process employed for building a linear quantitative structure-activity relationships (QSARs) for both human and pig FMOs. Each of these preliminary QSAR models generated was judged by both a conventional and 10 fold cross-validation statistics computed for choosing the best set of top ranked features for building the best linear QSAR model. The best linear QSAR thus constructed for human or pig FMOs was from 89 or 145 top ranked features selected, respectively. Moreover, these two linear QSARs were also found to be specific to their own top ranked features computed and selected. These two linear QSARs constructed may be useful for predicting whether or not a drug is metabolizable by human or pig FMOs if the same feature computation and ranking scheme has been applied on it beforehand.