Variable-selection approaches to generate QSAR models for a set of antichagasic semicarbazones and analogues
Luciana Scotti, Luciana Scotti, Hamilton Mitsugu Ishiki, Leticia Marques de Peron, Leandro de Rezende, Antônia Tavares do Amaral · Chemometrics and Intelligent Laboratory Systems · 2016
Quantitative structure–activity relationship (QSAR) models were proposed to correlate structural features or property descriptors of compounds with their corresponding biological activities. Because of the huge number of descriptors that encode different structural features used to generate valid QSAR models, variable selection becomes a fundamental step in building predictive and interpretative models. In this study, we applied a combined approach using multiple linear regression (MLR) and partial least-squares regression (PLS) to generate robust QSAR models with only a few descriptors applied to a set of cruzain inhibitors, namely, 61 semicarbazones and analogues, taken from the literature. From the 4885 descriptors generated by the Dragon program, we selected only five descriptors, applying the "Best-First" algorithm and PLS, followed by analysis of frequency and, finally, the genetic algorithm with MLR. The most significant QSAR equation encodes important steric and electronic structural features, which helps to identify in the set, structures that increase or decrease the pIC50 values measured against cruzain.