An Artificial Neural Network Classifier for the Prediction of Protein Structural Classes

As there are quite a few difficulties for us to predict a protein structural class directly from its primary sequence, the protein structural prediction based on the predicted secondary structure will undoubtedly be the first choice we would like to take. Protein structural classes are generally defined as four classes: α, β, α/β, α +β. The protein secondary structure describes the local structural conformation of the polypeptide backbone, and it can be obtained fairly accurately from the primary sequence, all of these very features make the protein secondary prediction a critical way to predict the structural class. We constructed a more balanced PSIPRED (a neural network predictor with psi-blast, original method first proposed by Rost & Sander in 1994) algorithm to predict the protein secondary structure. Finally the features about Chaos Game Representation of the predicted secondary structure sequence were selected as the input of neural network classifier. As a result, the predictor has got an overall accuracy score of 71.2% on 40% identity dataset of astral on Structural Classification of Proteins database. Such situation proved that the predictor via secondary structure prediction is an effective approach to classify the structural classes.