An accurate score function for detecting the most native-like models among a huge number of decoy sets is essential to the protein structure prediction. In this work, we developed a novel integrated score function (SVR_CAF) to discriminate native structures from decoys, as well as to rank near-native structures and select best decoys when native structures are absent. SVR_CAF is a machine learning score, which incorporates the contact energy based score (CE_score), amino acid network based score (AAN_score), and the fast Fourier transform based score (FFT_score). The score function was evaluated with four decoy sets for its discriminative ability and it shows higher overall performance than the state-of-the-art score functions.
Keywords: amino acid network; contact energy; fast Fourier transform; protein native structure selection; support vector regression.
Copyright © 2013 Wiley Periodicals, Inc.