Development of NEural network simulator for structure-activity COrrelation of moldecules: Neco

Yoshimi ISU*1, Umpei NAGASHIMA*1, Haruo HOSOYA*1 and Tomoo AOYAMA*2

*1 Faculty of Science, Ochanomizu University, Bunkyo-ku, Tokyo, 112 JAPAN
*2 Faculty of Engineering, Miyazaki University, Gakuenkihanadai, Miyazaki, 889-21 JAPAN

(Received: October 11, 1994; Accepted for publication: November 18, 1994)

A perceptron type neural network simulator for structure-activity correlation of molecules, Neco, has been developed with two different pre-education methods: back propagation method and reconstruction of weight matrix method.
Since the program was written in C, it is executable on popular Unix workstations. The number of intermediate layered units in intermediate layer and intermediate layer itself can be set arbitrarily by the user.
As an example of application, conformations of norbornene isomers were predicted using ^{13}C-NMR data. The predicted conformations are in excellent agreement with experiments.
By reconstructing the weight matrix, it was suggested that ^{13}C-NMR data of only two specified carbons int the norbornene skeleton have strong correlation with the conformation (exo and endo) of the main branch. This was verified by the parameter scan method.

Keywords: Perceptron type neural network, Back propagation, Reconstruction

Abstract in Japanese

Text in Japanese

Author Index
Keyword Index