Prediction of protein folding using the shift-learning method with a large scale neural network

The authors previously demonstrated the utility of large neural network simulations for encoding the association between protein sequence and 3-D structure for a small heterologous training set of small proteins. They report the application of this approach to a selected homologous training set of eight proteins using a Cray 2 supercomputer. The large memory of this machine made it possible to configure a network with more than 0.3 million connections and 30,000 neural units; a network of this size was necessary to accommodate a new training/testing set with eight proteins of up to 140 amino acid residues. This training set was constructed to investigate the performance of the neural network approach in prediction of structure within the protease class of proteins. The network learned the sequence-structure association for four of the proteins within 100 iterations selected in a random order and shifted by a random offset to the left or to the right. When presented with novel sequences from related proteins, the network was able to predict 3-D structures of the four proteins in the testing set. The results suggest that a neural network trained to recognize the entire sequence of a protein using the shift-learn method can retain some of the rules of protein folding in a form which allows prediction of 3-D structures.