Abstract
We have developed a neural network model to extract the functional class, prefixed by the EC on the bases of the type of reaction catalyzed, directly from all currently existing sequences and to identify the specific location in the sequences that characterize each function. The model is similar to the ordinary multi-layered network except that it changes the topology reproductively and reductively according to the currently evaluated network performance. We examined several test enzyme data sets that are different in homology from the training data. We observed that the model resulted in preferable predictions for Oxidoreductase, Transferase and Hydrolase, even when the homologies are low. However, scores for Lyase, Isomerase and Ligase were poor because they are small in number compared with the first three. More detailed examinations and further development are needed, however, the method will have considerable utility in assisting biotechnological experiments and in developing some predictive systems.