As mean 6 standard error or percentage inhibition at 10 mM.Results Model Building DockingIn total, four conformational variants of the A1AR homology model were used during docking and Cucurbitacin I ligand selection (Fig. 1). Model A was the original model, refined with the two previously known MedChemExpress Bromopyruvic acid ligands 5 and 6; model B was obtained by rebuilding ECL3 and adjacent residues around ligand 8; and models C and D were generated by further adapting the binding site to the most selective ligand previously identified in this study (8; binding mode shown in Fig. 2) using logAUC and side chain orientation diversity as model selection criteria. In terms of heavy-atom RMSD, models C ??and D differed by less than 0.18 A overall 
and by less than 1.17 A in the refined residues in the binding site (Fig. 1). Docked compounds that ranked highly in at least one of the models (Figure 5 and Table S1) were selected after visual inspection and tested experimentally for receptor affinity. These diverse compounds included thiazole (7, 8, 10?3, 16, 18, 20, and 23), 1,3,5triazine (9 and 24) and other heterocyclic cores. Thiazoles and 1,2,4-triazines are known chemotypes for binding to ARs [41,42]. A xanthine derivative 19, unusual in its 1-phenyl substitution, also appeared as a hit. According to the docking predictions, this phenyl ring of 19 was oriented away from Asn2546.55 toward the pocket lined by Val622.57, Ala662.61, and Val873.32. A commonality of all compounds was that they form two hydrogen bonds with Asn2546.55 in the calculated poses. Table 1 lists all ligands that inhibited radioligand binding to at least one hAR subtype by more 
than 50 at a concentration of 10 mM and were thus classified as active. Their two-dimensional structures are shown in Figure 5. Data for molecules that did not pass this threshold are presented in Table S1. Table 2 lists the total number of molecules tested in each round. In total, we found 8 ligands for the A1AR, 15 for the A2AAR and 14 for the A3AR. The structurally most similar known AR ligand from ChEMBL for each hit, as determined by ECFP4 Tanimoto similarity, is listed in Table S2. One of the ligands (14) may be regarded as a novel AR ligand because its Tanimoto similarity to the most similar known ligand is less than 0.26, which is generally accepted as a strict cutoff [43]. By a more relaxed cutoff of 0.4 [44], five more compounds (15, 21, 22, 25, 26) are novel. Table 2 furthermore details the performance of the individual models by their ability to predict ligands. Model C was the most unproductive, having no correct ligand 1516647 predictions. It is interesting to note that there is no clear trend in the performance in terms of selectivity. One could have assumed that models productive for one AR subtype might perform badly in retrieving ligands for a different one (despite all of them being models with the A1AR sequence). This only seems to be the case for model A (retrieving more A2A and A3AR ligands than A1AR ligands), but not the other ones, which tend to find approximately equal numbers for ligands of all subtypes.Selectivity CalculationsA total of 2181 ligands from the ChEMBL database had experimentally determined non-negative Ki values against both A1 and A2A, and 1476 molecules had such measurements against A1 and A3. Only 77 of all known experimental AR ligands had ambiguous classifications as being “inactive” and “active” against at least one receptor, and were thus not investigated further. The results are presented as.As mean 6 standard error or percentage inhibition at 10 mM.Results Model Building DockingIn total, four conformational variants of the A1AR homology model were used during docking and ligand selection (Fig. 1). Model A was the original model, refined with the two previously known ligands 5 and 6; model B was obtained by rebuilding ECL3 and adjacent residues around ligand 8; and models C and D were generated by further adapting the binding site to the most selective ligand previously identified in this study (8; binding mode shown in Fig. 2) using logAUC and side chain orientation diversity as model selection criteria. In terms of heavy-atom RMSD, models C ??and D differed by less than 0.18 A overall and by less than 1.17 A in the refined residues in the binding site (Fig. 1). Docked compounds that ranked highly in at least one of the models (Figure 5 and Table S1) were selected after visual inspection and tested experimentally for receptor affinity. These diverse compounds included thiazole (7, 8, 10?3, 16, 18, 20, and 23), 1,3,5triazine (9 and 24) and other heterocyclic cores. Thiazoles and 1,2,4-triazines are known chemotypes for binding to ARs [41,42]. A xanthine derivative 19, unusual in its 1-phenyl substitution, also appeared as a hit. According to the docking predictions, this phenyl ring of 19 was oriented away from Asn2546.55 toward the pocket lined by Val622.57, Ala662.61, and Val873.32. A commonality of all compounds was that they form two hydrogen bonds with Asn2546.55 in the calculated poses. Table 1 lists all ligands that inhibited radioligand binding to at least one hAR subtype by more than 50 at a concentration of 10 mM and were thus classified as active. Their two-dimensional structures are shown in Figure 5. Data for molecules that did not pass this threshold are presented in Table S1. Table 2 lists the total number of molecules tested in each round. In total, we found 8 ligands for the A1AR, 15 for the A2AAR and 14 for the A3AR. The structurally most similar known AR ligand from ChEMBL for each hit, as determined by ECFP4 Tanimoto similarity, is listed in Table S2. One of the ligands (14) may be regarded as a novel AR ligand because its Tanimoto similarity to the most similar known ligand is less than 0.26, which is generally accepted as a strict cutoff [43]. By a more relaxed cutoff of 0.4 [44], five more compounds (15, 21, 22, 25, 26) are novel. Table 2 furthermore details the performance of the individual models by their ability to predict ligands. Model C was the most unproductive, having no correct ligand 1516647 predictions. It is interesting to note that there is no clear trend in the performance in terms of selectivity. One could have assumed that models productive for one AR subtype might perform badly in retrieving ligands for a different one (despite all of them being models with the A1AR sequence). This only seems to be the case for model A (retrieving more A2A and A3AR ligands than A1AR ligands), but not the other ones, which tend to find approximately equal numbers for ligands of all subtypes.Selectivity CalculationsA total of 2181 ligands from the ChEMBL database had experimentally determined non-negative Ki values against both A1 and A2A, and 1476 molecules had such measurements against A1 and A3. Only 77 of all known experimental AR ligands had ambiguous classifications as being “inactive” and “active” against at least one receptor, and were thus not investigated further. The results are presented as.