We have developed a new energy function, termed GEM SCORE, for the protein structure prediction, which is an emergent problem in the field of computational structural biology. The GEMSCORE combines knowledge-based and physics-based energy functions. Instead of hundreds and thousands parameters used in many physics-based energy functions, we optimized nine weights of energy terms in the GEMSCORE by using a generic evolutionary method. These nine energy terms are the electrostatic, the der Waals, the hydrogen-bonding potential, and six terms for solvation potentials. The GEMSCORE has been evaluated on six decoy sets, including 96 proteins with more 70,000 structures. The result indicates that our method is able to successfully identify 74 native proteins from these 96 proteins. Our GEMSCORE is fast and simple to discriminate between native and nonnative structures from thousands of protein structure candidates in these decoy sets. We believe that the GEMSCORE is robust and should be a useful energy function for the protein structure prediction.