Hybrid Strategies in Variable Selection for PLS Regression in a QSAR Study on Indulines Derivatives as 5-HT2c Receptor Antagonists

Hybridization methods are able to combine some beneficial features of a number of chemometrics methods. On the other hand, variable selection is the most important part in regression modeling. In this work, a hybrid approach that combines genetic algorithm (GA) and variable importance in projection (VIP) is proposed to achieve a variable selection method in PLS analysis. This method was applied for QSAR studies of a set of 1–(3–Pyridylcarbamoyl) induline derivatives. Thanks to hybridization method, not only lowering of the variable selected numbers was possible, but also a low dimensional PLS model with interpretable variables was obtained. The squared regression coefficient of prediction for training and test sets obtained by PLS model were 0.911 and 0.843 respectively. The effect of each class of descriptors in the final model, on the binding affinity of indulines derivatives, thoroughly explained and discussed in a descriptive manner.