Predicting the 3-dimentional structure of protein has been a major interest in the modern computational biology. While lots of successful methods can generate models with 3~5Å root-mean-square deviation (RMSD) from the solution, the progress of refining these models is quite slow. It is therefore urgently needed to develop effective methods to bring low-quality models to higher-accuracy ranges (e.g., less than 2 Å RMSD).

In this thesis, I present several novel computational methods to address the high-accuracy refinement problem. First, an enhanced sampling method, named parallel continuous simulated tempering (PCST), is developed to accelerate the molecular dynamics (MD) simulation. Second, two energy biasing methods, Structure-Based Model (SBM) and Ensemble-Based Model (EBM), are introduced to perform targeted sampling around important conformations. Third, a three-step method is developed to blindly select high-quality models along the MD simulation. These methods work together to make significant refinement of low-quality models without any knowledge of the solution.

The effectiveness of these methods is examined in different applications. Using the PCST-SBM method, models with higher global distance test scores (GDT_TS) are generated and selected in the MD simulation of 18 targets from the refinement category of the 10th Critical Assessment of Structure Prediction (CASP10). In addition, in the refinement test of two CASP10 targets using the PCST-EBM method, it is indicated that EBM may bring the initial model to even higher-quality levels. Furthermore, a multi-round refinement protocol of PCST-SBM improves the model quality of a protein to the level that is sufficient high for the molecular replacement in X-ray crystallography. Our results justify the crucial position of enhanced sampling in the protein structure prediction and demonstrate that a considerable improvement of low-accuracy structures is still achievable with current force fields.