Prediction of odours of aliphatic alcohols and carbonylated compounds using fuzzy partition and self organising maps (SOM)

Abstract

A set of 114 olfactory molecules divided into fruity, ethereal and camphoraceous compounds, was submitted to an analysis by Kohonen Neural Networks, also known as Self Organising Map (SOM). The compounds are represented in a hyperspace derived from their molecular descriptors and SOM gives a useful projection of this hyperspace onto a 2D map. Owing to the complexity of the olfaction mechanism, evidenced by the fact that one compound can exhibit simultaneously different properties, SOM alone is unable to take into account the olfaction diversity of the original 114 compounds, Then, a Fuzzy Partition method was applied on the Kohonen map previously developed. The obtained results allowed delineating different representative zones for the three odours, expressing more closely the olfactory richness. The ability of the Hybrid System combining SO and Fuzzy Partition to model the three odours was validated by dividing the 114 compounds into a training set and a test set, including 86 and 28 molecules, respectively. The most important olfactory characteristics were reproduced satisfactorily for the entire test set compounds.