(Received: January 10, 2001; Accepted for publication: October 10, 2001; Published on Web: December 7, 2001)
We developed a neural network simulator for structure-activity correlation of molecules: Neco. A self-organized network model for high-speed learning was included in Neco, a perceptron type with three layers. In the hidden layer the neurons are self-organized by using Mahalanobis generalized distance.
This report proposes an improved training algorithm to the network. A self-organizing module decides the number of neurons in the hidden layer, at first. Then, a neuron in the hidden layer has two informations which describe a characteristic of the neuron. In this way, the network can evaluate stochastic characteristics from input data better.
Using this simulator, the hydrophobic parameter, logP, of perillartine derivatives was predicted. We used for inputs a set of six parameters: five STERIMOL (L, Wl, Wu, Wr, and Wd) and the sweet/bitter activity. The 22 sampled data are used for training. Our neural network can accurately predict hydrophobic parameter, logP. Compared with a normal perceptron network, the learning ability of our network is somewhat higher and its convergence speed is greatly much larger.
This simulator doesn't depend on the machine environment because it codes by the Java programming language.
Keywords: Self-organized network, Neural network, Structure-Activity Correlation, Perillartine derivatives, Hydrophobic parameter
Text in Japanese