STRUCTURE ODOR RELATIONSHIPS USING NEURAL NETWORKS

Models of relationships between structure and musk odor of 79 nitrobenzenic compounds were constructed by means of a multilayer neural network using the backpropagation algorithm. Substituents on the 5 free sites on the benzenic ring (one position is always substituted by a terbutyl group) were described using 3 steric hindrance descriptors and 3 electronegativity descriptors. Odor was coded by a binary variable. Our neural network gives a better classification (94%) than that obtained by discriminant analysis (81 %). The activity of the 79 molecules was then predicted (77% correct prediction) for 8 subsets of 10 (or 9) compounds using 69 (or 70) other molecules as a neural network training set.