Protein torsion angle class prediction by a hybrid architecture of bayesian and neural networks

Protein torsion angle class prediction by a hybrid architecture of bayesian and neural networks

Zafer Aydin, James Thompson, Jeffrey Bilmes, David Baker and William Stafford Noble

Proceedings of the International Conference on Bioinformatics and Computational Biology (BIOCOMP'12). Editors: Hamid R. Arabnia and Quoc-Nam Tran. pp. 2012-2018.

Abstract

Protein torsion angles provide essential information about the three-dimensional structure of a protein. Accurate prediction of backbone angles can enchance the quality of tertiary (3D) structure prediction, sequence alignment and fold recognition. In this paper, we introduce a machine learning classiﬁer that is able to predict the torsion angle category of an amino acid with high accuracy. Our method combines dynamic Bayesian networks with a neural network and is capable of incorporating information from multiple input representations such as position speciﬁc scoring matrices (PSSM) derived using sequence alignment methods. We show that 3D structure prediction accuracy of the widely used Rosetta program improves in the ab initio setting when the predicted torsion class information is used during the fragment selection step.

Proceedings of BIOCOMP12
Home