At the A1AR determined using [3H]DPCPX in the absence and presence of 10 mM CGS15943 (N-[9-chloro-2-(2-furanyl) [1,2,4]triazolo[1,5-c]quinazolin-5-amine), respectively. b ECFP4 Tanimoto similarity for the most structurally similar known AR ligand (Table S3). *percent inhibition at 10 mM compound concentration. **n = 
1. doi:10.1371/journal.pone.0049910.tglobular proteins, its usefulness for assessing models of membrane proteins such as GPCRs was unclear. 18325633 Thus, globular regions were extracted from the modeled A1AR structures by selecting residues ?in a 6 A sphere around C7, C11, and C12 of 1. This extraction resulted in 100 approximately globular protein fragments. Thesefragments were scored with DOPE and DOPE_HR (DOPE high resolution) and the top five scoring models were inspected visually. Criteria in this visual inspection were the absence of obvious steric clashes with 1, the absence of kinks in the helices, an orientation of the sidechain of Asn2546.55 away from the main chain, and preservation of the disulfide bonds between Cys803.25-Cys169 and Cys2606.61-Cys2637.28 (superscripts denote Ballesteros-Weinstein numbers [19]). The model that was chosen among the top five according to these criteria was denoted as model O. Table 2. Performance of the four homology models against the three AR subtypes.A1 MODEL A B C D A/TaA2A 7 42 0 33 21 A/T 5/15 7/12 0/6 3/6 15/39 33 58 0 50 38A3 A/T 7/15 4/12 0/6 3/6 14/39 47 33 0 50 36 Round 1 2 31/15 5/12 0/6 2/6 8/Figure 2. Calculated binding mode of compound 8, the ligand hit with the highest selectivity towards A1AR. The protein is model A. Orange dotted lines denote hydrogen bonds formed with Asn2546.55. Helices are labeled with roman numerals. doi:10.1371/journal.pone.0049910.gSbabnumber of actives (A) over number of molecules tested (T). sum: overall hit rate for all tested ligands. doi:10.1371/journal.pone.0049910.tIn Silico Screening for A1AR AntagonistsFigure 3. Comparing the selectivity of ligands from this work with ChEMBL data. Selectivity statistics for experimentally measured affinities of molecules from the ChEMBL database (outer shell) and our screen (inner donut). Selectivity ratios have been binned into log-sized bins, ranging from more than 1000-fold selectivity in either direction to 1. doi:10.1371/journal.pone.0049910.gModel RefinementAs shown previously, adapting the orthosteric sites of GPCR homology models to known ligands improves pose fidelity and hit rates [20]. Thus, for optimization of model O, binding site residues ?within a 6 A radius around the equivalent position of 1 (the ligand in 3EML) were iteratively refined with CHARMM [21] and MODELLER. The residues selected for optimization were also compared to mutagenesis studies of the A1AR in recognition of agonists and antagonists [22,23]. Residues that caused major changes in binding affinity (up to 100-fold BIBS39 decrease) after alanine substitution were checked against the selection of residues within 6 ?A of the ligand. In all cases, the residues that contributed to a loss of binding affinity after alanine substitution were included in the selection. For the part of the refinement using CHARMM, the CHARMm22 force field (Accelrys, Inc.) was used, and harmonic ?restraints with a force constant of 50 kcal/mol?A2 and a minimum ?at 2.4 A were assigned to the hydrogen bonds formed between the respective ligand and Asn2546.55, the key recognition residue in the A1AR binding CASIN pocket. A known ligand of the A1AR (.At the A1AR determined using [3H]DPCPX in the absence and presence of 10 mM CGS15943 (N-[9-chloro-2-(2-furanyl) [1,2,4]triazolo[1,5-c]quinazolin-5-amine), respectively. b ECFP4 Tanimoto similarity for the most structurally similar known AR ligand (Table S3). *percent inhibition at 
10 mM compound concentration. **n = 1. doi:10.1371/journal.pone.0049910.tglobular proteins, its usefulness for assessing models of membrane proteins such as GPCRs was unclear. 18325633 Thus, globular regions were extracted from the modeled A1AR structures by selecting residues ?in a 6 A sphere around C7, C11, and C12 of 1. This extraction resulted in 100 approximately globular protein fragments. Thesefragments were scored with DOPE and DOPE_HR (DOPE high resolution) and the top five scoring models were inspected visually. Criteria in this visual inspection were the absence of obvious steric clashes with 1, the absence of kinks in the helices, an orientation of the sidechain of Asn2546.55 away from the main chain, and preservation of the disulfide bonds between Cys803.25-Cys169 and Cys2606.61-Cys2637.28 (superscripts denote Ballesteros-Weinstein numbers [19]). The model that was chosen among the top five according to these criteria was denoted as model O. Table 2. Performance of the four homology models against the three AR subtypes.A1 MODEL A B C D A/TaA2A 7 42 0 33 21 A/T 5/15 7/12 0/6 3/6 15/39 33 58 0 50 38A3 A/T 7/15 4/12 0/6 3/6 14/39 47 33 0 50 36 Round 1 2 31/15 5/12 0/6 2/6 8/Figure 2. Calculated binding mode of compound 8, the ligand hit with the highest selectivity towards A1AR. The protein is model A. Orange dotted lines denote hydrogen bonds formed with Asn2546.55. Helices are labeled with roman numerals. doi:10.1371/journal.pone.0049910.gSbabnumber of actives (A) over number of molecules tested (T). sum: overall hit rate for all tested ligands. doi:10.1371/journal.pone.0049910.tIn Silico Screening for A1AR AntagonistsFigure 3. Comparing the selectivity of ligands from this work with ChEMBL data. Selectivity statistics for experimentally measured affinities of molecules from the ChEMBL database (outer shell) and our screen (inner donut). Selectivity ratios have been binned into log-sized bins, ranging from more than 1000-fold selectivity in either direction to 1. doi:10.1371/journal.pone.0049910.gModel RefinementAs shown previously, adapting the orthosteric sites of GPCR homology models to known ligands improves pose fidelity and hit rates [20]. Thus, for optimization of model O, binding site residues ?within a 6 A radius around the equivalent position of 1 (the ligand in 3EML) were iteratively refined with CHARMM [21] and MODELLER. The residues selected for optimization were also compared to mutagenesis studies of the A1AR in recognition of agonists and antagonists [22,23]. Residues that caused major changes in binding affinity (up to 100-fold decrease) after alanine substitution were checked against the selection of residues within 6 ?A of the ligand. In all cases, the residues that contributed to a loss of binding affinity after alanine substitution were included in the selection. For the part of the refinement using CHARMM, the CHARMm22 force field (Accelrys, Inc.) was used, and harmonic ?restraints with a force constant of 50 kcal/mol?A2 and a minimum ?at 2.4 A were assigned to the hydrogen bonds formed between the respective ligand and Asn2546.55, the key recognition residue in the A1AR binding pocket. A known ligand of the A1AR (.