SINGLE-SEQUENCE PROTEIN SECONDARY STRUCTURE PREDICTION BY NEAREST-NEIGHBOR CLASSIFICATION OF PROTEIN WORDS

Abstract

Predicting protein secondary structure is the process by which, given a sequence of amino acids as input, the secondary structure class of each position in the sequence is predicted. Our approach is built on the extraction of protein words of a fixed length from protein sequences, followed by nearest-neighbor classification in order to predict the secondary structure class of the middle position in each word. We present a new formulation for learning a distance function on protein words based on position-dependent substitution scores on amino acids. These substitution scores are learned by solving a large linear programming problem on examples of words with known secondary structures. We evaluated this approach by using a database of 5519 proteins with a total amino acid length of approximately 3000000. Presently, a test scheme using words of length 23 achieved a uniform average over word position of 65.2%. The average accuracy for alpha-classified words in the test was 63.1%, for beta-classified words was 56.6%, and for coil classified words was 71.6%.