Combining few neural networks for effective secondary structure prediction

Abstract

The prediction of secondary structure is treated with a simple and efficient method. Combining only three neural networks, an average Q/sub 3/ accuracy prediction by residues of 75.93% is achieved. This value is better than the best results reported on the same test and training database, CB396, using the same validation method. For a second database, RS126, an average Q/sub 3/ accuracy of 74.13% is attained, which is better than each individual method, being defeated only by CONSENSUS, a rather intricate engine, which is a combination of several methods. The networks are trained with RPROP an efficient variation of the back-propagation algorithm. Five combination rules are applied independently afterwards. Each one increases the accuracy of prediction by at least 1%, due to the fact that each network used converges to a different local minimum. The Product rule derives the best results. The predictor described here can be accessed at http://biolab.cin.ufpe.br/tools/.