Learning the pattern-based CRF for prediction of a protein local structure

Abstract

We describe a pattern-based conditional random field model for the prediction of dihedral angles of an all-alpha protein from its primary structure. Such conditional random fields appear naturally in sequence labeling problems of bioinformatics and can be considered relative to the Hidden Markov Models. The learning of parameters of the model is done by the structural SVM technique. The accuracy that we achieved in predicting dihedral angles, φ and ψ, equals 22.8 and 48.3 degrees, respectively. The MDA score, defined as the percentage of residues that are found in correctly predicted eight-residue segments, attained 56.5%.