An exploration of using a modular neural network to predict the secondary structures through a variation in the description of the output states

Fumiyoshi SASAGAWA (sasagawa@iias.flab.fujitsu.co.jp)
Koji TAJIMA (tajima@iias.flab.fujitsu.co.jp)

International Institute For Advanced Study Of Social Information Science, Fujitsu Laboratories Ltd.

We study the prediction of globular protein secondary structures by a neural network and super-computer. The application of a neural network with a modular architecture to prediction of protein secondary structures (alpha-helix, beta-sheet and coil) is presented. Each module is a three layer neural network. We compare the results from the neural network with a modular architecture and with a simple three layer structure. The prediction accuracy by a neural network with a modular architecture is higher than of the ordinary neural network. The 3, 4 and 8 state classification scheme of secondary structures are considered in the ordinary three layer neural network. The percentage of correct prediction depends on these state classification scheme.