DescriptionHigh throughput and automatic procedures for NMR structure determination are under intensive study in the current era of structural genomics. The major steps include data collection, data processing, resonance assignment with validation, derivation of structural restraints, generation of 20 conformers satisfying the structural restraints (the ensemble), and refinement of the conformers. The final step structure refinement typically refers to further energy minimization based on certain force fields. Refinement could improve the structure quality to a large extent due to the sparseness of NMR experimental measurements. A scientific and robust refinement methodology is desired as a vital part of the standard protocols of automatic NMR structure determination. In this study, we compare the performances of two refinement methods, CNS refinement and AMBER refinement. The core algorithm of CNS refinement is simulated annealing with gradient descent while AMBER uses molecular dynamics simulated annealing. Eight protein targets are chosen randomly from the NESG depository and the two refinement methods tested on these targets. All the targets have chemical shifts and NOESY peak lists available, and 4 of them also have RDC data. Using the available NMR experiment data, initial coarse structures are generated by ASDP-CYANA. These coarse structures further go through CNS refinement and AMBER refinement. Then the CNS-refined and AMBER-refined structures are evaluated in terms of RMSD (reference to X-ray PDB structure) and DP score. We find that AMBER refinement achieves better results than CNS on 7 out of 8 targets—AMBER refined structures have smaller average RMSDs and higher ensemble-average DP scores. The differentiated performance of the two refinement methods could stem from the different algorithms and force fields implemented.