Pattern recognition in the prediction of protein structural class

Algorithms that predict the secondary structure of individual amino acids based solely on their local sequence environment obtain at best 60% to 65% accuracy. The approach presented uses more global information as input to two methodologies, a modified Euclidean statistical clustering algorithm and a three-layer backpropagation network. Input to both these methods consists of the normalized frequency of the 20 amino acids as well as the frequency of six hydrophobic amino acid patterns. The average predictive accuracy for all test set proteins using the Euclidean statistic was 74.0%. The backpropagation network correctly predicted 76.2% of the test proteins. These results show that there exist patterns in the protein primary sequence that are useful for the prediction of protein structural class by certain statistical clustering algorithms and neural networks.