Quantitative structure–retention relationship study of analgesic drugs by application of combined data splitting-feature selection strategy and genetic algorithm-partial least square

Abstract

The combined data splitting feature selection (CDFS) is a new strategy in quantitative structure–property relation (QSPR) analysis, in which the sampling of training set is performed repeatedly to find a subset of molecular descriptors producing a stable QSPR model insensitive to the presence or absence of one or more compounds in the training set. Here, we used genetic algorithm-partial least square (GA-PLS) as modeling method in CDFS methodology and applied it to QSPR study of the GC retention of the analgesic drugs. A set of 58 analgesic drugs with known Kovats retention index were selected and a large number of theoretical descriptors was calculated for each molecule. The random sampling of the training set (80% of data) was performed 20 times and the remaining molecules were used as validation set. Each time, the most appropriate QSPR model was produced by GA-PLS. The selected descriptors of each run were then analyzed for similarity and frequency distribution. The final QSPR model, which obtained from the common descriptors between 50% of runs, possessed squared correlation coefficient of 0.924, 0.865 and 0.903 for training, validation and cross-validation, respectively. In addition, it was able to reproduce 85% of variances in the retention factor of external test set compounds.