G protein-coupled receptors (GPCRs) are therapeutic targets for many diseases, but progress in developing active and selective therapeutics has been severely hampered by the difficulty in obtaining accurate structures. We have been developing methods for predicting the structures for GPCR ligand complexes, but validation has been hampered by a lack of experimental structures with which to compare our predictions. We report here the predicted structures of the human adenosine GPCR subtypes (A 1 , A 2A , A 2B , and A 3 ) and the binding sites for adenosine agonist and eight antagonists to this predicted structure, making no use of structural data, and compare with recent experimental crystal structure for ZM241385 bound human A 2A receptor. The predicted structure correctly identifies 9 of the 12 crystal binding site residues. Moreover, the predicted binding energies of eight antagonists to the predicted structure of A 2A correlate quite well with experiment. These excellent predictions resulted when we used Monte Carlo techniques to optimize the loop structures, particularly the cysteine linkages. Ignoring these linkages led to a much worse predicted binding site (identifying only 3 of the 12 important residues).These results indicate that computational methods can predict the three-dimensional structure of GPCR membrane proteins sufficiently accurately for use in designing subtype selective ligands for important GPCR therapeutics targets.