Abstract
Relationships between structure and green or bell-pepper or nutty odours were established for a set of 88 pyrazines and 32 thiazoles by means of backpropagation neural networks (BNN). Each molecule was described by six steric hindrance or electronegativity descriptors. Excellent classification rates were obtained for the green note (97.7%, much better than the 67% obtained by discriminant analysis) and for the nutty note (96.6%). Correct prediction rates of 97.9% and 96.6% were obtained for the green and nutty notes respectively. A further test on 32 thiazoles gave a correct prediction rate of 87.5% for the bell-pepper note, and 81 3% for the green note.
In another part of this study, based on similarity calculations, bell-pepper, green and nutty notes were again considered on the same compounds.
First, Euclidean distances were calculated for all pairs of pyrazines, using the same descriptors as with BNN. A pyrazine possessing a given odour was considered as correctly classified if two or more of its four nearest neighbours had the same odour. This rule gave a correct classification rate of 92.8% for the 14 bell-pepper pyrazines and of 91.9% for the 37 green pyrazines, whereas much less satisfactory results (69.6% of correct-classification rate) have been obtained for 23 nutty pyrazines.
Second, a “best representative pyrazine” was chosen as reference compound, one for each olfactory class, and Euclidean distances to each of the three reference compounds were considered. Similarities were then calculated and graphically represented. The odour attributed to a given pyrazine was that of the nearest reference molecule. This method gave a correct classification rate of 92.9% for the bell-pepper pyrazines and 82.6% for the 23 nutty pyrazines. However, the 37 green pyrazines gave poorer results (73% of correct classification rate).
Comparison among these three methods shows that good results were obtained using BNN for all notes. Pairwise comparisons gave correct results for green and bell pepper notes and poorer results for the nutty note. Similarity comparisons gave correct results for the nutty and bell-pepper notes and a less satisfactory classification for the green note.