Constructing homology models is facilitated by the fact that the transmembrane region of Class A GPCRs is relatively well conserved. The accuracy of homology models is limited, however, by the uncertainty of modeling the extra- and intracellular loops, which greatly vary in length and amino acid composition, even between otherwise closely related GPCRs. In this study, we tested the utility of homology models for docking and selecting compounds with reasonable affinity for the investigated receptor subtype. We intentionally restricted the amount of existing ligand data used to refine the binding site during model building to mimic a situation where few ligands are known. In fact, except for the very first steps of model building and optimization, only the affinity data obtained in this study was used to improve the homology models. Three sequential cycles of model refinement, docking, and ligand testing were applied, using the data acquired in previous rounds to guide the receptor model optimization in the following rounds. In parallel, we also probed the tendency of the screen to identify novel ligands of other subtypes within the same receptor family, i.e. the selectivity of a homology model-based screen against a single GPCR subtype. These findings were compared with the distribution of selectivity ratios of known ligands for the same subtypes. The adenosine receptors, which consist of the four subtypes A1, A2A, A2B, and A3, have been chosen as a suitable test case for the application of virtual screening to a closely related subtype of a known GPCR structure. There are both antagonistbound and agonist-bound X-ray structures known for the A2AAR subtype, with various ligands co-crystallized for each case. Thus, the region for orthosteric AR ligand binding has been well characterized. The first antagonist-bound structure to be determined was co-crystallized with the high affinity ligand 4- triazin-5-yl-amino]ethylphenol. An unexpected orientation of the ligand perpendicular to the plane of the membrane bilayer was observed. Extracellular loops, as well as helical TM domains, are involved in coordinating the ligand. In silico virtual screening for A2AAR antagonists has already been demonstrated to be successful based on the inactive conformation of the A2AAR, as determined by crystallography. Among the different subtypes, the A1AR is also an attractive pharmaceutical target. Its antagonists have been explored as kidney-protective agents, compounds for treating cardiac failure, cognitive enhancers, and antiasthmatic agents. Structurally diverse antagonists, such as the pyrazolopyridine derivative 2 and the 7-deazaadenine derivative 3, were previously identified, and some of these compounds were under consideration for clinical use. The prototypical AR antagonists, i.e. the 1,3- dialkylxanthines, have provided numerous high affinity antagonists with selectivity for the A1AR. One such antagonist, rolofylline 4, an alkylxanthine derivative of high content screening abmole bioscience nanomolar affinity, was previously in clinical trials for cardiac failure. The human A1AR subtype was investigated in this study because it shares a high level of sequence identity with the A2AAR. It should thus be possible to model the A1AR by homology with high confidence. While this homology model was the only three-dimensional structure of a protein employed in the screening, all compounds were also tested in receptor binding assays against two other AR subtypes in order to investigate the intrinsic selectivity of the model.. We did not exclude the molecules tested in earlier rounds of screening during the subsequent ones, yet the vast majority of ligands identified in one model did not appear in the top ranks of a screen against another one.