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

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