Refinement and Selection of Near-native Protein Structures

In recent years \textit{in silico} protein structure prediction reached a level where a variety of servers can generate large pools of near-native structures. However, the identification and further refinement of the best structures from the pool of decoys continue to be problematic. To address these issues, we have developed a selective refinement protocol (based on the Rosetta software package), and a molecular dynamics (MD) simulation based ranking method (MDR). The refinement of the selected structures is done by employing Rosetta's relax mode, subject to certain constraints. The selection of the final best models is done with MDR by testing their relative stability against gradual heating during all atom MD simulations. We have implemented the selective refinement protocol and the MDR method in our fully automated server Mufold-MD. Assessments of the performance of the Mufold-MD server in the CASP10 competition and other tests will be presented.