Journal of Chemical Software
Online ISSN : 1883-8359
Print ISSN : 0918-0761
ISSN-L : 0918-0761
Development of a Neural Network Simulator for Structure-Activity Correlation of Molecules: Neco (4)
– Sweet / Bitter Classification in Perillartine Derivatives –

2000 Volume 6 Issue 3 Pages 115-126


The relationships between molecular structure and taste quality: sweet or bitter, or several perillartine derivatives were examined using a perceptron type neural network simulator for structure-activity correlation of molecules: Neco with reconstruction of weight matrix method. The reconstruction of weight matrix method was used to optimize the number of neurons in hidden layer. In the case of using six parameters: hydroforbic(log P) and the STERIMOL(L, Wl, Wu, Wr, and Wd) parameters as inputs, the number of neurons in hidden layer is minimized to one by the reconstruction learning method. Even in this case, there is no misclassified compound. The prediction rate by leave-one-out procedure was also 100%. The most important three parameters were the same as predicted by Fisher ratio. The number of input parameters was minimized by holding the number of neurons in hidden layer to one. Two parameters, namely Log P and Wr were found to describe the sweet/bitter activity of perillartine derivatives. Instead of STERIMOL parameters, atomic charges of common molecular skeleton, HOMO and LUMO energies, and HOMO-LUMO energy difference were selected as input data. MOPAC93/AM1 was used to evaluate these parameters. The optimum number of neurons in the hidden layer was also one. Atomic charges and LUMO energy are also important for sweet/bitter classification. This suggests that electronic structure around common molecular skeleton and electron elimination reaction are essential to sweet/ bitter activity of perillartine derivatives.

© 2000 by the Chemical Software Society of Japan
Previous article