High accuracy protein modeling from its sequence information is an important step toward revealing the sequence-structure-function relationship of proteins and nowadays it becomes increasingly more useful for practical purposes such as in drug discovery and in protein design. We have developed a protocol for protein structure prediction that can generate highly accurate protein models in terms of backbone structure, side-chain orientation, hydrogen bonding and binding sites of ligands. In order to obtain accurate protein models, we have combined a powerful global optimization method with traditional homology modeling procedures such as multiple sequence alignment, chain building and side-chain re-modeling. We have built a series of specific score functions for these steps, and optimized them by utilizing conformational space annealing, which is one of the most successful combinatorial optimization algorithms currently available.