TETRALIN, INDAN AND NITROBENZENE COMPOUND STRUCTURE-MUSK ODOR RELATIONSHIP USING NEURAL NETWORKS

Structure-musk odor relationships were established by means of a neural network (NN) on a sample of 105 molecules comprising nitrobenzene compounds, carbonyl tetralins and carbonyl indans. Each molecule was described by eight variables (six steric hindrance descriptors and two electronegativity descriptors, coding six substituents on a common underlying skeleton (benzene ring with a tertiobutyl or pseudo-tertiobutyl group in the meta position relative to a nitro or carbonyl group)). Odor was coded by a binary variable. Seventeen molecules could be coded by two different descriptions, and so the back-propagation NN was trained on 122 descriptions. The NN gave a better classification (98.1%) than that obtained by discriminant analysis (65.7%). To test the reliability of the NN approach, training was made on 11 training subsets of 111 or 110 descriptions with a correct classification rate of 96.4 to 98.6%. For the 11 corresponding test subsets of 11 or 12 descriptions a correct prediction rate of 80% was obtained for the 105 molecules. The contribution of each descriptor was evaluated. The results show the possibility of establishing structure-musk odor relationships for molecules with different structures and confirm the previously mentioned effect of steric hindrance in the ortho position of the functional groups interacting with the receptor by hydrogen bonding.