Prediction of the Odor Thresholds of Oxygen and Nitrogen Containing Heterocycles Using the Quantitative Structure-Property Relationship Approach

Abstract

Quantitative structure-property relationship (QSPR) models are developed to correlate the odor thresholds of 50 oxygen and nitrogen containing heterocycles from their molecular structures. Three statistic methods including multiple linear regression (MLR), non-linear radial basis function neural network (RBFNN) and support vector machine (SVM) are performed to build the models. A six-descriptor equation with the squared correlation coefficient (R²) of 0.8012 and root mean square error (RMS) of 1.0011 were obtained for the training set and R² = 0.648, RMS = 1.7165 for the external test set. The radial basis function neural network model gave better results: R² = 0.8767, RMS = 0.7165 for the training set and R² = 0.7746, RMS = 1.3570 for the external test set. The SVM model gave similar results to multiple linear regression, that is, R² = 0.8023, RMS = 0.9271 for the training set and R² = 0.7033 and RMS = 1.5888 for the test set. The aim of the paper is to provide an easy, direct and relatively accurate way to estimate the odor thresholds.