Recently, we developed a new approach to protein-structure prediction, which combines template-based modeling with the physics-based coarse-grained UNited RESidue (UNRES) force field. In this approach, restrained multiplexed replica exchange molecular dynamics simulations with UNRES, with the Cα-distance and virtual-bond-dihedral-angle restraints derived from knowledge-based models are carried out. In this work, we report a test of this approach in the 11th Community Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction (CASP11), in which we used the template-based models from early-stage predictions by the LEE group CASP11 server (group 038, called \"nns\"), and further improvement of the method. The quality of the models obtained in CASP11 was better than that resulting from unrestrained UNRES simulations; however, the obtained models were generally worse than the final nns models. Calculations with the final nns models, performed after CASP11, resulted in substantial improvement, especially for multi-domain proteins. Based on these results, we modified the procedure by deriving restraints from models from multiple servers, in this study the four top-performing servers in CASP11 (nns, BAKER-ROSETTASERVER, Zhang-server, and QUARK), and implementing either all restraints or only the restraints on the fragments that appear similar in the majority of models (the
consensus fragments), outlier models discarded. Tests with 29 CASP11 human-prediction targets with length less than 400 amino-acid residues demonstrated that the
consensus-fragment approach gave better results, i.e., lower α-carbon root-mean-square deviation from the experimental structures, higher template modeling score, and global distance test total score values than the best of the parent server models. Apart from global improvement (repacking and improving the orientation of domains and other substructures), improvement was also reached for template-based modeling targets, indicating that the approach has refinement capacity. Therefore, the
consensus-fragment analysis is able to remove lower-quality models and poor-quality parts of the models without knowing the experimental structure.