Adapting the Structure of a Neural Network to Extract Chemical Information. Application to Structure-Odour Relationships

Abstract

Two types of neural networks were used to establish relationships between chemical structure and musk odour of 79 nitrobenzenic compounds. Substituents on the five free sites of the benzene ring (one position was always occupied by a t−butyl group) were described using three volume descriptors and three electronegativity descriptors. Musk odour was coded by a binary variable. First a classical network with two hidden layers containing six and three neurons was used. This network gave a better classification (94%) than that obtained by linear discriminant analysis (81%). The odour was men predicted using a leave-ten-out procedure, with 77% of correct prediction for the whole sample. Then a dual two-way network was built to mimic the symmetry of the problem (two sides on a molecule, two muskophore patterns). This network recognized both patterns already known to chemists and gave 99% of correct classifications by taking into account substitution in all positions. As a side benefit of the modified network structure it was possible to evaluate the influence of each of 19 substituents in each of the five possible positions.