QSPR study on melting point of carbocyclic nitroaromatic compounds by multiple linear regression and artificial neural network

Abstract

A quantitative structure–property relationship study was performed to correlate descriptors representing molecular structures to the melting point of carbocyclic nitroaromatic compounds. The complete set of 60 compounds was divided into a training set of 45 compounds and a test set of 15 compounds by using the DUPLEX algorithm. Multiple linear regression analysis was used to select the best subset of descriptors and to build linear models; nonlinear models were developed by means of an artificial neural network. The robustness of the obtained models was assessed by leave-one-out and leave-many-out cross-validation, Y-randomization test, and external validation through test set. The obtained models with six descriptors show good predictive power: the linear model has the average absolute relative deviation (AARD) of 5.31% and 3.98% for the training and test sets, respectively; while the nonlinear model performs better than the linear model, with the AARD of 4.42% and 3.82% for the training and test sets. In addition, the applicability domain of the models was analyzed based on the Williams plot.