Coordinate Refinement on All Atoms of the Protein Backbone with Support Vector Regression

For the past decades, many efforts have been made in the fields of protein structure prediction. Among these, the protein backbone reconstruction problem (PBRP) has attracted much attention. The goal of PBRP is to reconstruct the 3D coordinates of all atoms along the protein backbone for given a target protein sequence and its Cα coordinates. In order to improve the prediction accuracy, we attempt to refine the 3D coordinates of all backbone atoms by incorporating the stateof-the-art prediction softwares and support vector regression (SVR). We use the predicted coordinates of two excellent methods, PD2 and BBQ, as our feature candidates. Accordingly, we define more than 100 possible features. By means of the correlation analysis, we can identify several significant features deeply related to the prediction target. Then, a 5-fold cross validation is carried out to perform the experiments, in which the involved datasets range from CASP7 to CASP11. As the experimental results show, our method yields about 8% improvement in RMSD over PD2, which is the most accurate predictor for the problem.