Quantitative structure—Odor relationship: Using of multidimensional data analysis and neural network approaches

Abstract Structure—odor relationships (SOR) are key issues for the synthesis of new odorant molecules. But, this relation is hard to model, due to limited understanding of olfaction phenomena and the subjectivity of odor quantity and quality as stated in Rossitier's review (1996) . Many molecular descriptors are used to correlate molecule's odor, but no universal rules emerge in this field. In this paper, we focus on the use of molecular descriptors as an alternative approach in the prediction of odors, by the mean of regression techniques. Principal Component Analysis (PCA) and Stepwise Collinearity Diagnosis (SCD) techniques are used to reduce the dimensionality of data, by the identification of significant molecular descriptors. Then, the chosen molecular descriptors are used with a neural networks algorithm to correlate the structure to molecular odor quality. The results are validated on balsamic flavor.