Predicted Structures of Agonist and Antagonist Bound Complexes of Adenosine A₃ Receptor

Abstract

We used the GEnSeMBLE Monte Carlo method to predict ensemble of the 20 best packings (helix rotations and tilts) based on the neutral total energy (E) from a vast number (10 trillion) of potential packings for each of the 4 subtypes of the adenosine G protein-coupled receptors (GPCRs), which are involved in many cytoprotective functions. We then used the DarwinDock Monte Carlo methods to predict the binding pose for the human A₃ adenosine receptor (hAA₃R) for subtype selective agonists and antagonists.

We find that all four A₃ agonists stabilize the 15^th lowest conformation of apo-hAA₃R while also binding strongly to the 1^st and 3^rd. In contrast the four A₃ antagonists stabilize the 2^nd or 3^rd lowest conformation. These results show that different ligands can stabilize different GPCR conformations, which will likely affect function, complicating the design of functionally unique ligands.

Interestingly all agonists lead to a trans χ1 angle for W6.48 that experiments on other GPCRs associate with G-protein activation while all 20 apo-AA₃R conformations have a W6.48 gauche+ χ1 angle associated experimentally with inactive GPCRs for other systems. Thus docking calculations have identified critical ligand-GPCR structures involved with activation.

We find that the predicted binding site for selective agonist Cl-IB-MECA to the predicted structure of hAA₃R shows favorable interactions to three subtype variable residues, I253^6.58, V169^EL2, and Q167^EL2, while the predicted structure for hAA_2AR shows weakened to the corresponding amino acids: T256^6.58, E169^EL2, and L167^EL2, explaining the observed subtype selectivity.

Introduction

Adenosine receptors (ARs), G protein-coupled receptors (GPCRs) with 7 transmembrane (TM) spanning helices,¹ are involved in many cytoprotective functions, making them important pharmacological targets for treatment of a variety of diseases because of their key roles in controlling numerous physiological processes. Many therapeutic agents under development for treatment of central nervous system disorders, inflammatory diseases, asthma, kidney failure, and ischemic injuries exert their effects via direct interactions with ARs. However the target is often just one of the four subtypes of human ARs, designated A₁, A_2A, A_2B, and A₃. The A_2A and A_2B adenosine subtypes are linked to Gs proteins, increasing the intracellular levels of cAMP upon receptor activation, while the A₁ and A₃ adenosine subtypes are linked to Gi proteins inhibiting cAMP formation. The A₃ and A_2B subtypes can also activate the inositol phosphate formation through coupling to Gq proteins.² Thus activation of human adenosine A₁ receptor (hAA₁R), decreases the oxygen demand of a stressed organ by slowing the heart rate. Conversely, activation of hAA_2AR can increase oxygen supply via vasodilation and inhibition of platelet aggregation. Also a development of anti-Parkinson’s agents utilizes hAA_2AR antagonists.³ On the other hand hAA_2BR activation, has an angiogenic effect, while activation of the cardiac hAA₃R preconditions cardiac myocytes against ischemic damage⁴ and protects against toxicity. This can be seen in the case of the cardiotoxic effects of the anticancer drug doxorubicin.⁵ Moreover, hAA₃R antagonists are of interest as potential anti-glaucoma agents.⁶ In particular, hAA₃R agonists, such as IB-MECA (N⁶-(3-Iodobenzyl)-5’-N-methylcarboxamidoadenosine) and Cl-IB-MECA (2-Chloro-N⁶-(3-iodobenzyl)-5’-N-methylcarboxamidoadenosine) that were developed in the Jacobson lab. at National Institute of Health, are now advancing into Phase II clinical trials for treating diseases such as cancer, arthritis, and psoriasis.²

Since these AR targets generally involve just one subtype, it is essential that new ligands be developed that are subtype specific. To design such subtype-selective drugs requires detailed 3D predicted structures for all four receptors and for the low-lying conformation that might be stabilized by the prospective new ligands. Particularly this is important for agonists, since it is likely that the process by which agonists activate the G-Protein involve the agonist interacting with more than one GPCR conformation.

Previously, we reported the predicted structures of the human adenosine subtypes (A₁, A_2A, A_2B, and A₃) based on the human β₂ adrenergic receptor template (PDB ID: 2rh1). There we found that the predicted binding energies of 8 antagonists to the predicted structure of A_2A were found to correlate quite well with experiment. In addition the predicted binding sites for ZM241385 inverse agonist bound to hAA_2AR⁷ compares well with x-ray crystal structure.

Here, we use the improved GEnSeMBLE method to predict new structures for the four subtypes (A₁, A_2A, A_2B, and A₃) of human AR starting with a template based on the x-ray structure for ZM241385 bound to hAA_2AR (PDB ID: 3eml). GEnSeMBLE selects from a vast number (10 trillion) of potential packings a final ensemble of the 20 best packings (helix rotations and tilts) based on predicted total E, that we consider might play a role in binding to various ligands and in activation.

In addition, we used the DarwinDock Monte Carlo method to predict the binding sites and energies for four agonists and four antagonists (structures shown in Chart 1) to each of the 20 most stable conformations of hAA₃R. We find that all four agonists prefer the 15^th conformation of the apo-hAA₃R, whereas the four antagonists prefer either the 2^nd or 3^rd. Moreover all four agonists have the W6.48 the horizontal orientation (trans χ₁) believed from experiments on other GPCRs to be associated with activation, while 20 lowest conformations of the apo-hAA₃R have W6.48 in the vertical orientation (gauche+ χ₁) believed from experiments on other GPCRs to be inactive.^8-10

Chart 1.

Structures of four agonists (Ag) and four antagonists (Ant) exhibiting selective binding to the adenosine A₃ receptor with respect to A_2a. Binding affinities (K_i in nM) are shown for A₃ with relative efficacies in parenthesis compared to the endogenous adenosine.

These predictions confirm the high selectivity toward hAA₃R for the agonist, Cl-IB-MECA, and the high selectively for adenosine toward hAA_2AR, the endogenous agonist, validating the accuracy of these methods for explaining the subtype selectivity.

We expect that improved understanding of how these GPCRs function at the molecular level will be useful for the development of novel classes of AA₃R-targeted drugs.

Materials and Methods

In order to predict the 3D structures for the various conformations needed to understand the function of GPCRs and help design new ligands, we use the GEnSeMBLE (GPCR Ensemble of Structures in Membrane BiLayer Environment) method, which provides a very complete sampling over possible rotations and tilts and leading to a ensemble of low lying predicted structures expected to include those conformations energetically accessible for binding of ligands. This replaces our earlier MembStruk method.¹¹

We now use the DarwinDock to predict the binding sites of ligands to the GPCRs. This replaces our earlier HierDock¹² and MSCDock¹³ methods, providing a much more complete sampling of possible poses. These earlier methods were validated by a series of applications to various GPCRs: human D₂ dopamine receptor (DR),¹⁴ human β₂ adrenergic receptor,^{15, 16} human M₁ muscarinic receptor,¹⁷ human Chemokine (C-C) motif receptor 1 (CCR1),¹⁸ mouse MrgC11 (Mas Related Gene) for the molluscan peptide FMRF-amide (FMRFa),^{19, 20} human prostanoid DP receptor,²¹ and human Serotonin 2C ²² receptor.

GEnSeMBLE consists of the following key steps:

1. Generation of the 7-transmembranic helix (TMH) bundle.

   1. PredicTM: To predict the seven TM domains our current approach (PredicTM) combines hydrophobicity analysis with information from multiple sequence alignments, involving six steps; i) retrieval of similar protein sequences from a database, ii) multiple sequence alignment of sequences using the MAFFT²³ program, iii) Hydrophobic profile in the multiple sequence alignment (using the thermodynamic and biological hydrophobic scales from White and von Heijne^{24, 25}), iv) initial TMH region predictions in Fig. S1 of Supporting Information, v) Application of capping rules and extension of consensus helices based on the APSSP2²⁶ and PORTER²⁷ servers, vi) identification of hydrophobic centers in Table S1 of Supporting Information. More details are described in Supporting Information.

   2. OptHelix: To predict the kinks often induced by proline residues, we replace the longer side chains and helix termini (save Ser and Thr near Pro) with alanines and carry out 2 ns of molecular dynamics (MD) on each helix, followed by a restoration of the correct side chains using SCREAM²⁸ and a minimization of the structure. This procedure works well and predicts helices within 1-2 Å of the crystal.

   3. Template Generation: To pack the 7 helices from OptHelix into a bundle requires that six quantities be specified for each helix: the x-, y-, and z-coordinate of the hydrophobic center, the tilt (θ) of each axis from the z-axis, the azimuthal orientation (φ) of this tilt; and the rotation (η) of the helix about the helical axis, when the x-axis is defined along the axis from the center point of TMH3 to the center point of TMH2 in the midplane (see Fig. S2 of Supporting Information). We obtain z from PredicTM while θ, φ, η are sampled as described below. The remaining two coordinates, x and y, for each helix are taken from template structures, either experimental or validated theoretical. Currently, we can consider eight such templates: bovRho (PDB ID: 1u19)²⁹, hβ₂-adrenergic receptor (2rh1),¹⁰ hAA_2AR (3eml),⁹ turkey β₁-adrenergic receptor (2vt4),⁸ and bovine opsin (3cap) ³⁰ from experiment, as well as the predicted and validated DP Prostaglandin,²¹ MrgC11,^{19, 20} and CCR1 Chemokine¹⁸ receptors. GEnSeMBLE allows for each of these templates to be used in separate predictions, providing an ensemble of bundles from which we can select on the basis of bundle E or ligand binding E.

      As a reference for the rotation angle (η) of each helix, we use a highly conserved residue (N1.50, D2.50, W4.50, P5.50, P6.50, and P7.50) in family A and D3.32 that is conserved among amine receptors, and match the rotation angle of its Cα projection on the x-y plane to that of the corresponding residue in the template structure. The helical axis for rotation is defined as the one corresponding to the least moment of inertia axis obtained using all backbone atoms.

2. Homology of the 7-TMH bundle.

   1. HomologyTM: In addition to PredicTM helices, we also generated the homology TM helices where we took advantage of the high sequence identity of hAA₁R, hAA_2BR and hAA₃R with the x-ray structure of hAA_2AR, both at the whole sequence (A₁: 45.39%, A_2B: 38.59%, A₃:31.31%) and TMH region (A₁: 68.84%, A_2B: 59.94%, A₃: 51.01%) level. Using the sequence alignment of each subtype to hAA_2AR, the side chain of each amino acid in the isolated helices from the hAA_2AR bundle was mutated to the corresponding amino acid of the given subtype by our rotamer placement method SCREAM²⁸.

      For hAA_2AR, we constructed a TMH bundle from the x-ray structure, removing the N/C terminals and loops. The following residue number ranges describe the final TMH regions with α-helix structure: 5-32 for TMH1, 41-67 for TMH2, 75-106 for TMH3, 119-140 for TMH4, 175-204 for TMH5, 222-258 for TMH6, and 269-291 for TMH7.

   2. OptimizeHelix: The geometry of each isolated helix was optimized by 100 steps with 0.5 terminal gradient using DREIDING 3 force field (FF) (referred to as DREIDING-III or D3FF)³¹ to remove seriously bad contacts among helices.

3. BiHelix Optimization of the rotation angles

   1. BiHelix: The next step is to optimize the rotation angles (η) for each of the 7 TMHs. The general procedure is to simultaneously rotate each TMH through 360° in 30° increments, leading to (12)⁷ = 35,000,000 configures (Fig. S2 of Supporting Information). To rapidly identify the best 1000 of these, we first consider each helix interacting with just one other helix, leading to 12*12 =144 combinations for each of the 12 pairs of interacting helices (1-2, 2-4, 4-5, 5-6, 6-7, 1-7, 1-3, 2-3, 3-4, 3-5, 3-6, 3-7). For each of these 144*12 pairs we optimize the sidechains using SCREAM.²⁸ Then these 1728 “mean field” numbers are used to estimate the energies of all 35,000,000 conformations, from which we pick the best 1000 by total E for subsequent CombiHelix analysis.

      SCREAM uses a library of residue conformations ranging from a root mean square of a Cα atom (CRMS) diversity of 0.4 Å to 1.6 Å in conjunction with a Monte Carlo sampling using full valence, hydrogen bond and electrostatic interactions, but with special van der Waals (vdW) potentials that somewhat reduce the penalty for contacts that are slightly too short while retaining the normal attractive interactions at full strength. With SCREAM, we find that we can now base the selections on the total E, Escream, without separate considerations for valence, electrostatic, hydrogen bond, and vdW terms.

      However for the ARs, we restricted rotations ± 45° rotation with 15° increments resulting in (7)⁷ = 823,543 combinations. Then we used the BiHelix to combine the Escream energies of the 49 combinations for each helix pair (using a total of 49 × 12 helix pairs = 551 energies) to estimate the E for all 823,543 conformational combinations for the ± 45° rotational range.

   2. CombiHelix: The top 1,000 predicted structures from BiHelix analysis are built explicitly using the rotations specified for each helix in the combination. The multiple rotational combinations are minimized by their side chains using SCREAM in order to evaluate the total E of the optimized multi-helix bundle. To avoid overestimation of long-range charge interaction, we neutralize the system by adding protons into the negatively charged residues, Asp and Glu, or by deleting protons from the positively charged amino acids, Lys and Arg. After neutralizing the top 100 predicted structures based on the charged total E, the final 100 low E predicted structures are ordered by the neutral total E of the optimized multi-helix bundle.

      We validated this procedure for all five experimental GPCR structures. Using the BiHelix method, we estimated the energies for all (12)⁷ ~ 35 million combinations from 0 to 360° angles by 30° increments. The experimental structure corresponds to the angle rotation combination 0_0_0_0_0_0_0 (the first zero corresponds to TMH 1, the second zero corresponds to TMH 2, and so on). The results indicate that the crystal structures rank as the top 1 lowest E predicted structure at the end of the procedure for all five cases.

4. SuperBiHelix optimization of the rotation and tilt angles

   1. SuperBiHelix: For the lowest E predicted structures from BiHelix for each of the 4 subtypes, we determine the optimum rotations (θ, φ, η) for the packing of the 7 helices into a 7-TMH bundle. Here we considered a ±10° sampling of θ tilt angle and a ±30° sampling of both φ and η angles, leading to a total of (3×5×5)⁷ ~ 10 quadrillion combinations. To make this huge sampling computationally feasible, use SuperBiHelix sampling as follows. For each of the 12 pairs of helices, we sample all combinations of θ, φ, η for each of the (3×5×5)² = 5,625 rotational-tilt combinations of each of the 12 pairs. In each case we optimize the side chains using SCREAM²⁸ with all-atom DREIDING FF.³¹

      The total energies for each of these (12)×(5,625) helix pair combinations are used to estimate the E for all 10 quadrillion 7-TMH bundle conformational combinations using the following procedure. In order to determine the conformations for each helix that lead to low E bundles. Then the 7-TMH bundle is partitioned into three quadhelix bundles. Next, the total energies of the three quadhelix bundles are calculated using the BiHelical energies. This is feasible because only 3× (3×5×5)⁴ ≈ 10⁸ bundle energies must be calculated. The 2,000 predicted structures with the lowest E for each quadhelix are listed by increasing E and used to select the top 36 orientations preferred by each helix. Finally, these conformations are used to calculate the E of (36)⁷ ≈ 8 × 10¹⁰ full bundles and output the 1000 best E predicted structures from this procedure.

   2. SuperComBiHelix: The top 1,000 predicted structures from the SuperBiHelix analysis were built explicitly using the rotations and tilts specified for each helix in the combination (they will take different conformations than were found in the BiHelical mode). The 7 TMH predicted structure is then minimized for 10 steps. This ensemble of low-lying predicted structures from SuperComBiHelix is used to in the docking studies for various agonists and antagonists.

      The 100 lowest E predicted structures based on the charge total E, are then neutralized and reordered by the neutral total E of their optimized multi-helix bundle. From this we select the ensemble of ~20 low-E predicted structures, each of which is docked to agonists, antagonists, and inverse agonists to provide the structures likely to be involved in GPCR activation.

      This procedure has been validated by applying it to all GPCRs.⁷

      Our goal in developing these in silico methods is to predict the 3D structures of GPCRs that are remote from the systems currently available from experiment. Thus we have emphasized first principles approaches rather than homology methods. We believe that the studies reported in this paper indicate the potential power of this approach. Thus we find that the ensemble of configurations of the apo-A₃ provides the flexibility to bind both to agonists and antagonists. As experimental methods are developed to obtained x-ray structures of agonist-bound systems, we will obtain quantitative measures for the accuracy of our predictions.

5. Ligand docking

   The lowest E predicted structures of the hAA_2AR and hAA₃R from BiHelix were used for docking several ligands, using the DarwinDock. The structure and charges of the ligands in Chart 1 were calculated using quantum mechanics (B3LYP with the 6-311G** basis set). The Histidines forming the H-bonding with Glu were treated as the protonated His with +1 charge. Those His are H272^7.43 in the hAA₃R and H264^EC3/ H278^7.43in the hAA_2AR.

   1. Scanning the receptor for potential binding regions: Starting with the predicted structure, we predicted putative ligands binding regions as follows. We first alanized the entire protein (replacing the 6 hydrophobic residues, I, L, V, F, Y, and W with A) and scanned for potential binding regions with no assumption about the binding site. The entire molecular surface of the predicted structure was mapped with spheres representing the empty volume of the protein (currently using the Sphgen procedure in DOCK4.0 suite of programs). The entire set of protein spheres is partitioned into ~30 to 50 overlapping cubes of 10 to 14 Å sides. We then generated 1,000 poses for each of these 30 - 50 regions. These results are compared to select the most promising two or three putative binding regions.

   2. DarwinDock: Our earlier studies used the HierDock and GenMSCDock, which have now been replaced by DarwinDock. For each ligand conformation, DarwinDock iteratively generates ~50,000 poses into the putative binding regions of the bulky-residue-alanized protein. This is followed by the E scoring of family heads to select the top 10% ordered by total E. The top 100 conformations are chosen for further optimization. For each of these we dealanize the protein side-chains (using SCREAM) to find the optimum side chains for each of the best 100 poses. Then we neutralize the protein and ligand by transferring protons appropriately within salt bridges and protonating or deprotonating exterior ligands, followed by further full geometry minimization.

      This same procedure was followed for each of 7 ligand conformations generated as follows. Starting from the x-ray structure of adenosine crystal, we rotated the torsion angles C_ar-C_ar-NH-C_al and C_ar-NH-C_al-C_ar between the adenine ring and the benzyl substituent by 60 ° increments to generate 36 conformations. These were generated with the Maestro software and minimized with the D3FF. We selected the lowest 7 ligand conformations (within 3 kcal/mol of the best E) for docking in Table S2 of Supporting Information as described above. The final docked structure with the best binding E from all ligand conformations was selected.

      DarwinDock has been validated for a number of x-ray co-crystals including 3 crystal structures of ligand/ GPCR complexes: human β₂-adrenergic receptor (0.4 Å RMSD),¹⁰ human AA_2AR (0.8 Å RMSD),⁹ and turkey β₁-adrenergic receptor (0.1 Å RMSD).⁸ This shows that we can accurately identify ligand binding sites in proteins, which can then be used to optimize the ligands with desirable properties.

   3. Neutralization for scoring E: Quantum mechanics (QM) calculations show that for an effective dielectric constant is < 8Å, the extra proton on a Lys or Arg transfers back to the negative carboxylate of an Asp or Glu.²² Thus we expect that buried salt bridges will have neutral residues. We consider that use of these neutral residue charges improves the accuracy for comparing different docked structures. Of course the final bond energy relative to ligand in the solvent and binding site exposed to solvent must be corrected by the effective pK of the bare Lys, Arg, Glu, Asp or of the ligand, a correction that is simple to make. For example, if the pK of a carboxylate is 4.5 and the solvent is taken to have a pH of 7.4 we must correct by 2.9*1.38 kcal/mol.

      For external residues not involved in binding, we find it expedient to neutralize the external residues exposed to solvent or membrane. Here the issue is that the force fields commonly used in MD calculations involve fixed charges, usually based on QM. In reality any net partial charges are shielded by the dielectric polarization of the surrounding protein and solvent so that there is negligible effect by 10Å. However with fixed charges the electrostatic interaction energy between two point charges separated by 10Å is 33 kcal/mol. The result is that small changes in geometries of charged ligands far from the binding site can lead to large differential binding energies, even 10 to 30 kcal/mol. We find that neutralizing these exposed residues removes the sensitivity to details of the distances of charged residues (and counter ions) remote from the active site. This neutralization leads to differential binding energies that are dominated by the local cavity interactions and leads to much smaller solvation energies.

6. MD studies

   In order to validate the stability of the Monte Carlo structures predicted herein, each ligand-protein predicted structure was inserted into a fully equilibrated hydrated POPC lipid bilayer having cell size in the XY plane of 75 Å x75 Å (172 POPC molecules) and solvated with 6,689 water molecules using Membrane builder in the VMD program. This was done for predicted structure 1 of the LJ1251 antagonist bound to predicted structure 2 and the Cl-IB-MECA agonist bound to predicted structure 15 with the disulfide bridge between 88 and 166. The procedure here was identical to that in Kim. et al.³²

   The lipid tails were almost fully extended, allowing for easy insertion of the protein into the membrane. The distance between the layers was set at c=72.3 Å to fit the actual membrane thickness plus the solvent thickness. The lateral supercell was set to fit the actual surface density of lipid molecules, a=69.9 and b=68.4 Å. We introduced disorder into the lipid bilayer patches to better resemble a membrane at 300K by allowing random orientation of each lipid about its axis with a short (1ps) equilibration at 300K in vacuum. This eliminated steric contacts between the lipid atoms but left the lipid tails mostly extended. These steps in our procedure reduce the time required for equilibration of the lipid/protein complex. After inserting the protein into the lipid-water cell, lipids overlapping within 1 Å of protein and waters overlapping within 5 Å were removed.

   For the particle mesh Ewald (PME) in the electrostatics calculation,³³ the charge of system was balanced through replacing waters into 13 Cl^- ions. These systems contain 37,661 atoms (A₃ selective agonist Cl-IB-MECA-hAA₃R), and 367,645 atoms (A₃ selective antagonist LJ1251-hAA₃R). This predicted structure was then equilibrated at 300K for 1 ns using the NAMD 2.5 (NAnoscale Molecular Dynamics) program.³⁴ We used the CHARMM22 force field parameters for the protein, the TIP3 model for water,³⁵ and the CHARMM27 force field parameters for the lipids.³⁶ Minimization of both ligands in vacuum for these ligands led to bond lengths, angles, and dihedrals similar to the quantum mechanical results and the previous predicted structures determined with DREIDING.³¹ Quantum charges from DFT/6311G** method were used for these ligands.

Results and Discussions

1. GEnSeMBLE predictions of the ensemble of low lying conformations of all four ARs

As summarized in Methods, the GEnSeMBLE method for predicting 3D structures of GPCRs aims at sampling a complete set of helix conformation with the emphasis on the rotations of the helices about their axis and on the tilts of these helices and then selecting the top ~20 based on energetic. We start with a template to define the six degrees of freedom for each helix (x, y, z, θ, φ, η) for each of the 7 helices and first consider all possible rotation angles (η) about the helix axes. Generally GEnSeMBLE allows each η angle to go over 360°, in increments of 30°, leading to (7)¹² = 35,000,000 different combinations. However, for the ARs, we considered each η angle from -45° to +45° in 15° increments leading to (7)⁷ = 823,000 combinations. This is because an x-ray structure was available for hAA_2AR and the similarity of the four subtypes from the multiple sequence alignment in Fig. 1 suggests that the rotational angles would change by less than 45°. Supporting this, the analysis of relative differences of the six degrees of freedom for each helix (x, y, z, θ, φ, η) among the x-ray structures of GPCRs, β₁ and β₂ adrenergic receptor between two subtypes differs less than ±10 ° in the rotational (η) angle of each helix, as shown in Table S3. Indeed despite the major differences between hAA_2AR and human β₂ adrenergic receptor, the variation of the rotational (η) angle of each helix is within ±20 °. In addition, the comparison of bovine rhodopsin to opsin (which is believed to be an equilibrium active state structure) suggests a range of ±30 ° in rotational angle variations (Table S4). Thus, our sampling angle of ±45 ° can generate the ground-state structure as well as the active state structure.

Figure 1.

Alignments of the four adenosine receptor (AR) subtypes, A₁, A_2A, A_2B, and A₃ from the PredicTM method. The predicted TMH regions from PredicTM are displayed in colored boxes (TMH1 in purple, TMH2 in blue, TMH3 in cyan, TMH4 in green, TMH5 in yellow, TMH6 in orange, TMH7 in red), while the TMH region from the x-ray structure of the AA_2AR (PDB: 3eml) is underlined. Highly conserved residues in Family A G protein-coupled receptors (GPCRs) are shown in red in TMH 1-6 and white in TMH7. Variable amino acids among the four subtypes in the upper TMH regions are marked with red asterisks and subtype selective residues predicted from the cavity analysis are boxed. The disulfide bridges in yellow were assigned (A₁: 80-169, 260-263, A_2A: 71-159, 74-146, 77-166, 259-262, A_2B: 72-167, 78-171, A₃: 83-166) based on the sequence alignment. We use Ballesteros-Weinstein numbering consisting of the TMH helix number followed by residue number relative to the highly conserved residue in the helix, numbered as 50. H-bonding is indicated by arrows, and subtype selective residues are shown in red.

As described in Methods, we used the BiHelix mean field strategy to rapidly estimate the energies of the 823,000 packings, selecting the lowest E 1000 to build into full 7-TMH bundles with optimum side chains (CombiHelix). The top 10 predicted structures of hAA_2AR and hAA₃R ranked by neutral total E are shown in Table 1. The other top 10 predicted structures of hAA₁R and hAA_2BR are displayed in Table S5. We find that the best packing predicted structure for hAA_2AR has exactly the same rotation angles as the experimental hAA_2AR (denoted as the all-0° predicted structure).

Table 1.

Top 10 predicted structures of human adenosine A_2A and A₃ receptors from the CombiHelix analysis of the (7)⁷ = 823,000 BiHelix packing geometries within ± 45 ° angle range by 15 ° increments. All 1,000 predicted structures from CombiHelix were selected for neutralization by their charge total energy score (ChargeTot: kcal/mol). The final 100 predicted structures are ordered by neutral total energy (NeutTot: kcal/mol). The case with η = 0 ° for all 7 helices (same as in the experimental antagonist bound A_2AR structure) is shown in grey shading. Thus there is a change only for AA₃R in TMH4 by +15°.

A2A
H1	H2	H3	H4	H5	H6	H7	ChargeIH	ChargeTot	NeutIH	NeuTot	W6.48 χ1
0	0	0	0	0	0	0	-379.3	158.8	-397.8	13.7	-79.5
0	0	0	15	0	0	0	-356.2	169.7	-380.3	14.2	-79.5
0	0	0	-15	0	0	0	-380.2	189.7	-399.6	47.0	-79.5
-15	0	0	0	0	0	0	-364.6	216.9	-382.1	60.2	-79.5
-15	0	0	15	0	0	0	-347.3	220.3	-369.5	60.2	-79.5
0	0	0	0	15	0	0	-328.1	245.6	-350.2	84.2	-79.2
0	0	0	0	15	15	0	-325.0	250.0	-344.8	90.7	-145.0
0	0	15	0	0	0	0	-354.4	254.8	-383.4	93.1	-79.5
15	0	0	15	0	0	0	-323.1	262.1	-344.9	93.2	-79.5
0	0	0	15	15	15	0	-311.9	248.2	-330.6	93.8	-81.0
A3
H1	H2	H3	H4	H5	H6	H7	ChargeIH	ChargeTot	NeutIH	NeuTot	W6.48 χ1
0	0	0	15	0	0	0	-354.9	430.3	-327.6	292.5	174.9
0	0	0	0	0	0	0	-359.2	418.7	-341.5	299.6	-79.1
0	0	15	15	0	15	0	-347.5	421.1	-331.4	309.4	-174.5
0	0	0	0	0	30	-15	-347.7	443.1	-337.0	311.5	163.9
0	0	0	15	0	15	0	-381.0	445.4	-356.0	315.3	-179.6
0	-15	15	15	0	15	0	-339.0	414.1	-308.2	319.5	-179.7
0	0	0	0	15	15	0	-325.0	464.4	-320.7	321.5	-179.5
0	0	0	-15	0	0	0	-354.5	442.7	-317.5	322.7	-79.1
0	0	15	-15	0	15	0	-368.9	421.6	-322.7	324.0	-174.6
0	0	0	15	15	15	0	-318.4	479.0	-313.2	324.8	153.4

- ChargeIH (Charge interhelical energy), ChargeTot (Charge total energy), NeutIH (Neutral interhelical energy), NeutTot (Neutral total energy)

The same procedure was used for all four subtypes, but except for hAA₃R, all ARs prefer the same η=0° angles as hAA_2AR.

For hAA_2AR, we find that the top 10 predicted structures have variations of ± 15° for TMHs 1 and 4 and +15° for TMHs 3, 5, and 6.

For hAA₁R we find that the top 10 predicted structures have variations of +15° for TMHs 3 and 7, ± 15° for TMH5, +15/ +45° for TMH6, and ± 15/ +30° for TMH4.

For hAA_2BR we find helix rotation preferences similar to hAA_2AR (reasonable based on the sequence identity of ~60% in the TMH region between these two ARs), with rotational variations of +15° for TMHs 4 and 6, ± 15° for TMH5, and -15 ° for TMH1.

For hAA₃R, the lowest E predicted structure has TMH4-rotated by +15°. The top 10 predicted structures show rotational variations of -15° for TMHs 2 and 7, +15° for TMH3, ± 15° for TMH4, and +15, +30 ° for TMH6.

After determining the range of optimum η angles for the top 10 conformations of the four subtypes, we then optimized the θ (tilt) and φ (sweep) angles. Here we started with the η’s of the best conformation and varied η and φ over the range of -30, -15, 0, 15, 30 ° while sampling θ for -10, 0, 10°. This leads to (75)⁷ = ~13 trillion (13,348,388,671,875) combinations. To make it practical to estimate the energies of all these packings, we used the SuperBiHelix mean field strategy described in the methods sections. Here we select the best 2000 conformations to build into full 7-TMH bundles with optimum side chains, selecting the best based on E for subsequent docking. These results for hAA₃R are shown in Table 2 where the all 0° predicted structure corresponds to the best predicted structure from CombiBiHelix, which had TMH4 rotated 15°. Among the top 20 predicted structures for hAA₃R we find

• η variations of -15° for TMHs 1 and 4

• θ differences of -10° for TMH4 and +10° for TMH5.

• φ changes of ± 15° for TMHs 4 and 7, 15° for TMH5, and ± 15, -30° for TMH6.

Table 2.

Top 20 predicted structures of human A₃ adenosine receptor from the SuperBiHelix/ CombiHelix analysis using the range (-10, 0, 10°) for θ (tilting) and the range (-30, -15, 0, 15, 30°) for φ (sweep) and η (rotation) angles. All 1,000 predicted structures from CombiHelix by charge total energy (ChargeTot: kcal/mol), were selected for neutralization and the final 100 predicted structures were ordered by neutral total energy (NeutTot: kcal/mol). The case with η = 0 ° for all 7 helices is shown in grey shading. This corresponds to the 0, 0, 0, 15, 0, 0, 0 ° lowest NeutTot energy predicted structure from Bihelix in Table 1. Table S1 includes detailed analysis of the salt-bridge and the H-bonding networks among top 20 predicted structures. All cases have the gauche+ χ1 angle for W243^6.48, the rotamer trigger angle associated with the inactive state.

#	Θ (°)							Φ (°)							η (°)							NeutTot
	H1	H2	H3	H4	H5	H6	H7	H1	H2	H3	H4	H5	H6	H7	H1	H2	H3	H4	H5	H6	H7
1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	147.4
2	0	0	0	0	0	0	0	0	0	0	0	0	15	0	0	0	0	-15	0	0	0	155.1
3	0	0	0	0	10	0	0	0	0	0	15	0	-30	0	0	0	0	-15	0	0	0	158.8
4	0	0	0	0	0	0	0	0	0	0	0	0	0	-15	0	0	0	-15	0	0	0	164.4
5	0	0	0	0	10	0	0	0	0	0	0	0	-30	0	0	0	0	-15	0	0	0	166.0
6	0	0	0	0	10	0	0	0	0	0	15	0	-15	0	0	0	0	-15	0	0	0	170.6
7	0	0	0	0	10	0	0	0	0	0	0	0	-30	0	0	0	0	0	0	0	0	171.4
8	0	0	0	0	0	0	0	0	0	0	0	0	15	15	0	0	0	0	0	0	0	172.6
9	0	0	0	0	0	0	0	0	0	0	0	0	15	-15	0	0	0	-15	0	0	0	174.2
10	0	0	0	0	10	0	0	0	0	0	15	0	-30	0	0	0	0	0	0	0	0	174.4
11	0	0	0	0	10	0	0	0	0	0	0	0	-15	0	0	0	0	-15	0	0	0	176.8
12	0	0	0	0	0	0	0	0	0	0	0	0	15	15	0	0	0	-15	0	0	0	177.8
13	0	0	0	0	10	0	0	0	0	0	15	0	-30	-15	0	0	0	-15	0	0	0	182.4
14	0	0	0	0	10	0	0	0	0	0	0	15	-30	0	0	0	0	-15	0	0	0	183.2
15	0	0	0	0	0	0	0	0	0	0	-15	0	15	0	0	0	0	-15	0	0	0	183.3
16	0	0	0	0	10	0	0	0	0	0	0	0	-15	0	-15	0	0	-15	0	0	0	183.8
17	0	0	0	0	10	0	0	0	0	0	15	0	-30	-15	0	0	0	0	0	0	0	184.7
18	0	0	0	0	10	0	0	0	0	0	15	0	-30	-30	0	0	0	0	0	0	0	185.0
19	0	0	0	-10	10	0	0	0	0	0	-15	0	-30	0	0	0	0	-15	0	0	0	186.6
20	0	0	0	0	10	0	0	0	0	0	0	0	-15	-30	0	0	0	-15	0	0	0	187.3

Thus there are favorable variations at TMHs 4, 5, 6, 7 within +15°. The various salt-bridges and the H-bonding networks of the predicted structure in the D(E)RY and NPxxY region are summarized in Table S6. Among these top 20 predicted structures, we find an average of 10 inter-helical hydrogen bonds (HB). A₃ has a specific H-bonding network among Y109^3.51, K113^3.55 and D199^5.60. These interactions are not observed in the x-ray structure of hAA_2AR, because the amino acid corresponding to D190^5.60 in hAA₃R is Arg in hAA_2AR.

The 15^th predicted structure of apo-A₃ that is best for binding of selective agonists shows a weakened interaction of TMHs 1-2-7 H-bonding networks among N30^1.50, D58^2.50, and N278^7.49, as shown in Table S6. The H-bonding distances of N30^1.50: N278^7.49 and D58^2.50: N278^7.49 are 4.6 and 4.9 Å, respectively, while the corresponding H-bond distances in the 1^st and 2^nd model are 2.9 and 2.7 Å, respectively. However the salt-bridge and the H-bonding in the D(E)RY region (D107^3.49: R111^3.53, D107^3.49: R126^4.41, R108^3.50: E225^6.30, E225^6.30: R224^6.29 K113^3.55: D199^5.60, Y109^3.51: K113^3.55) are similar among them.

We show below that several of these best 20 predicted structures might play a role in differential binding of agonists, antagonists, and inverse agonists, as well as in activation pathways.

2. Ensemble docking of hAA₃R selective agonists

We then docked the hAA₃R selective agonists and selective antagonists in Chart 1 to each of the best 20 predicted structures in Table 2 from SuperBiHelix.

We found variations in the orientation and packing of several of the TMH (especially TMHs 4, 5, 6, 7) in the top 20 predicted structures as follows: tilting ±10° for TMHs 4 and 5 , sweeping angles of ± 15° for TMHs 4/ 5 and ± 15, -30° for TMHs 6/ 7 , and rotation of -15° for TMHs 1/ 4.

As shown in Table 3 and Fig. 2, we find that all four agonists,

• Cl-IB-MECA,

• IB-MECA,

• MRS1898 ((1’S, 2’R, 3’S, 4’R, 5’S)-4-{2-Iodo-6-[(3-iodophenylmethyl)amino]-purine-9-yl}-1-(methylaminocarbonyl)bicycle-[3.1.0]-hexame-2,3-diol), and

• MRS3558 ((1’S, 2’R, 3’S, 4’R, 5’S)-4-{2-Chloro-6-[(3-iodophenylmethyl)amino]-purine-9-yl}-1-(methylaminocarbonyl)bicycle-[3.1.0]-hexame-2,3-diol),

prefer to bind to the 15^th lowest conformation from SuperBiHelix (TMH4 φ: -15°, TMH6 φ: 15°, TMH4 η: -15°).

Table 3.

SuperCombiHelix Top 20 predicted structures for the human adenosine A₃ receptor, including agonists (left) and antagonists (right), ordered by neutral total energy including solvation (TotSol Energy) (kcal/mol). The lowest energy predicted structure is shown in gray shading. Although we consider TotSol Energy to be the most accurate for ranking the predicted structures, we have included three other energies also useful in ranking ligand-bound proteins: Tot, UCav, and ComSol. The lowest state predicted by these measures are indicated with an underline. The χ1 angle of W243^6.48 is shown at the right for each case. From studies on other GPCRs, this angle of W6.48 is considered to serve as a rotamer trigger switch for activation of the G-protein, with gauche+ corresponding to the inactive state and trans corresponding to the active state. Here we put the cases with trans (active) in bold face, with the gauche+ cases (inactive) in italics. The predicted ground state for all four agonists have the trans (active form), while all 20 lowest strucrures for the apo-hAA₃R have the gauche+ (inactive form). For the antagonist systems, two 5‘-truncated antagonists (LJ1251 and MRS5127) have the gauche (inactive form) for the all 20 lowest conformations, while the other two have a trans (active state) but many low lying gauche (inactive) excited configuraions.

Agonist						Antagonist
ClIBMECA	Tot	UCav	ComSol	TotSol	W243χ1	MRS1292	Tot	UCav	ComSol	TotSol	W243χ1
15-A3	-79.8	-19.3	-267.1	-346.9	154	2-A3	-67.9	-22.4	-268.1	-336.0	177
3-A3	-81.5	-23.4	-257.8	-339.4	-71	15-A3	-64.1	-23.0	-268.2	-332.3	177
2-A3	-63.3	-20.7	-266.1	-329.4	-165	1-A3	-55.2	-15.5	-273.1	-328.3	-87
1-A3	-56.9	-20.5	-271.7	-328.6	-175	12-A3	-68.2	-21.2	-256.3	-324.5	-87
5-A3	-65.8	-18.6	-261.1	-326.8	-69	9-A3	-57.1	-15.9	-261.6	-318.8	-85
12-A3	-71.3	-24.3	-255.5	-326.8	-164	3-A3	-56.1	-19.0	-256.8	-312.8	-119
9-A3	-59.7	-18.8	-260.2	-319.9	155	4-A3	-32.8	-17.5	-274.6	-307.3	-87
4-A3	-46.9	-18.7	-272.2	-319.1	-175	8-A3	-49.2	-18.0	-254.8	-304.0	-87
8-A3	-58.9	-23.0	-255.0	-313.8	-164	13-A3	-30.0	-14.4	-269.1	-299.1	178
6-A3	-49.6	-25.4	-263.6	-313.2	-171	5-A3	-37.2	-18.7	-260.6	-297.8	-119
13-A3	-41.3	-13.1	-266.0	-307.3	-69	19-A3	-27.7	-19.5	-255.0	-282.6	177
11-A3	-34.8	-17.0	-268.6	-303.4	-171	11-A3	-10.5	-16.1	-267.0	-277.6	-119
19-A3	-45.5	-13.9	-252.8	-298.4	-69	16-A3	-24.9	-17.4	-251.3	-276.2	-119
7-A3	-35.3	-20.8	-262.2	-297.5	-71	6-A3	-13.4	-17.7	-262.5	-275.9	-118
16-A3	-41.1	-22.1	-252.4	-293.5	-171	7-A3	-10.3	-19.1	-261.3	-271.6	-119
10-A3	-30.8	-22.7	-261.7	-292.6	-71	14-A3	-3.2	-16.6	-267.1	-270.3	-119
20-A3	-21.2	-19.4	-268.8	-290.0	-171	18-A3	1.1	-10.2	-268.2	-267.1	177
14-A3	-18.8	-17.5	-267.9	-286.7	-166	17-A3	0.4	-12.9	-264.6	-264.2	177
18-A3	-18.7	-11.4	-267.3	-286.0	-69	10-A3	-3.9	-18.3	-260.2	-264.1	-118
17-A3	-13.9	-17.7	-261.9	-275.8	-69	20-A3	3.8	-18.9	-267.2	-263.4	-171
IBMECA	Tot	UCav	ComSol	TotSol	W243χ1	MRS3771	Tot	UCav	ComSol	TotSol	W243χ1
15-A3	-75.6	-19.7	-267.1	-342.6	155	2-A3	-80.9	-26.5	-267.9	-348.8	-165
1-A3	-62.8	-20.5	-271.5	-334.3	-175	15-A3	-69.0	-26.1	-268.6	-337.6	-166
3-A3	-75.3	-21.9	-257.4	-332.8	-71	9-A3	-66.8	-26.9	-260.7	-327.5	155
2-A3	-59.1	-20.0	-266.2	-325.3	-165	12-A3	-63.2	-26.8	-255.8	-318.9	-164
12-A3	-69.1	-23.1	-255.8	-325.0	-164	1-A3	-42.8	-24.2	-271.7	-314.5	-175
5-A3	-60.1	-18.1	-261.1	-321.2	-69	4-A3	-39.1	-23.2	-274.9	-314.1	-175
9-A3	-54.9	-18.4	-260.1	-315.0	155	3-A3	-54.5	-28.9	-255.9	-310.4	-79
4-A3	-43.1	-17.9	-271.8	-314.9	-175	8-A3	-48.7	-26.1	-254.3	-303.0	-165
8-A3	-56.8	-22.1	-254.5	-311.3	-164	16-A3	-49.0	-24.5	-251.4	-300.4	-171
13-A3	-36.2	-12.1	-265.6	-301.8	-70	13-A3	-26.2	-21.8	-266.5	-292.7	-166
6-A3	-36.7	-23.6	-263.4	-300.1	-171	11-A3	-22.3	-22.1	-267.7	-290.0	-171
11-A3	-30.7	-16.1	-269.0	-299.7	-171	5-A3	-29.7	-25.8	-259.4	-289.1	-82
19-A3	-39.5	-13.3	-253.5	-292.9	-69	6-A3	-13.2	-26.8	-261.9	-275.1	-164
14-A3	-23.4	-19.1	-267.4	-290.8	-166	7-A3	-10.5	-27.3	-260.0	-270.5	-79
7-A3	-28.3	-19.6	-261.8	-290.1	-71	19-A3	-17.9	-24.4	-251.7	-269.6	-82
16-A3	-36.9	-21.2	-251.6	-288.5	-171	14-A3	-1.0	-22.7	-265.7	-266.7	-167
20-A3	-19.6	-19.8	-268.3	-287.9	-171	10-A3	-3.1	-28.3	-259.2	-262.3	-79
10-A3	-24.4	-21.1	-260.9	-285.3	-71	17-A3	4.7	-20.9	-261.2	-256.5	-166
18-A3	-14.7	-12.6	-266.0	-280.7	-69	20-A3	8.6	-24.3	-262.8	-254.1	-166
17-A3	-7.5	-16.6	-260.9	-268.4	-69	18-A3	14.6	-25.3	-265.3	-250.7	-81
MRS1898	Tot	UCav	ComSol	TotSol	W243χ1	MRS5127	Tot	UCav	ComSol	TotSol	W243χ1
15-A3	69.7	-17.8	-267.5	-197.8	154	3-A3	49.2	-17.0	-260.2	-211.0	-79
3-A3	66.4	-17.0	-260.7	-194.2	-71	2-A3	56.4	-16.9	-265.7	-209.3	-80
1-A3	80.9	-12.8	-274.9	-193.9	-70	1-A3	67.8	-13.6	-273.9	-206.1	-80
2-A3	79.2	-20.1	-265.7	-186.5	-165	15-A3	69.6	-15.6	-266.5	-197.0	-80
5-A3	77.5	-17.3	-263.2	-185.8	-70	12-A3	63.7	-16.6	-255.5	-191.8	-79
9-A3	82.1	-18.7	-259.9	-177.8	154	9-A3	71.5	-14.6	-260.7	-189.2	-86
4-A3	96.9	-12.2	-273.5	-176.6	-70	4-A3	89.9	-11.3	-274.5	-184.6	-81
12-A3	80.4	-20.7	-255.5	-175.1	-164	5-A3	79.3	-15.6	-262.1	-182.8	-81
13-A3	103.5	-12.7	-268.6	-165.1	-70	6-A3	92.0	-19.7	-265.5	-173.5	-80
8-A3	94.2	-19.0	-256.1	-161.9	-164	19-A3	83.2	-15.7	-254.7	-171.5	-81
6-A3	106.3	-18.9	-264.3	-158.0	-71	13-A3	97.7	-11.7	-268.6	-170.9	-81
11-A3	112.3	-17.8	-267.5	-155.2	-70	8-A3	85.3	-14.9	-255.3	-170.1	-79
7-A3	114.1	-15.6	-264.8	-150.7	-72	11-A3	99.4	-16.5	-268.5	-169.0	-81
19-A3	103.8	-17.4	-253.4	-149.6	-70	14-A3	108.0	-16.0	-267.8	-159.8	-80
16-A3	105.4	-17.8	-254.0	-148.6	-72	16-A3	95.3	-20.1	-253.8	-158.5	-80
10-A3	118.4	-16.2	-264.9	-146.6	-71	10-A3	107.7	-16.4	-264.3	-156.6	-79
18-A3	125.1	-7.3	-269.5	-144.4	-70	7-A3	112.7	-16.1	-264.0	-151.3	-79
14-A3	127.9	-20.2	-267.8	-140.0	-166	20-A3	125.1	-6.8	-265.9	-140.8	-96
20-A3	131.6	-5.7	-265.4	-133.8	-70	17-A3	126.9	-10.9	-264.2	-137.2	-81
17-A3	135.0	-11.4	-264.0	-129.0	-70	18-A3	135.3	-4.4	-269.5	-134.3	-96
MRS3558	Tot	UCav	ComSol	TotSol	W243χ1	LJ1251	Tot	UCav	ComSol	TotSol	W243χ1
15-A3	69.2	-18.5	-265.6	-196.4	154	2-A3	-88.6	-17.5	-265.6	-354.3	-80
1-A3	79.6	-13.5	-273.1	-193.5	-70	1-A3	-78.0	-20.3	-273.1	-351.1	-80
3-A3	68.6	-17.3	-259.2	-190.6	-71	12-A3	-92.4	-17.3	-256.2	-348.6	-79
2-A3	77.7	-19.9	-265.6	-187.9	-165	15-A3	-80.7	-18.0	-266.6	-347.3	-80
5-A3	79.7	-18.1	-262.1	-182.4	-70	3-A3	-90.4	-19.3	-256.8	-347.2	-79
9-A3	81.4	-20.0	-259.2	-177.8	155	4-A3	-54.4	-11.6	-273.8	-328.2	-80
4-A3	94.4	-12.4	-272.2	-177.8	-70	9-A3	-68.4	-10.2	-259.8	-328.1	-85
12-A3	79.3	-21.5	-254.7	-175.5	-164	5-A3	-59.2	-18.0	-259.6	-318.7	-80
13-A3	105.9	-10.6	-267.1	-161.2	-70	8-A3	-63.4	-20.6	-255.0	-318.5	-79
8-A3	93.4	-20.4	-253.9	-160.5	-164	16-A3	-57.5	-18.9	-251.8	-309.2	-79
16-A3	98.8	-16.3	-253.3	-154.5	-72	6-A3	-43.6	-19.5	-263.0	-306.6	-80
6-A3	108.2	-15.0	-262.4	-154.2	-71	14-A3	-39.3	-13.6	-266.9	-306.2	-80
10-A3	112.4	-17.0	-262.6	-150.2	-71	19-A3	-53.1	-19.1	-252.3	-305.4	-80
7-A3	116.2	-14.2	-263.6	-147.4	-71	11-A3	-37.8	-17.5	-266.1	-303.9	-81
19-A3	106.1	-16.3	-252.5	-146.4	-70	13-A3	-34.5	-18.4	-265.9	-300.3	-81
11-A3	121.2	-16.5	-266.4	-145.2	-70	10-A3	-32.0	-18.7	-260.6	-292.6	-79
18-A3	132.3	-8.4	-267.1	-134.8	-70	7-A3	-28.4	-19.5	-260.8	-289.2	-79
20-A3	129.4	-9.1	-263.6	-134.2	-70	18-A3	-14.6	-16.8	-266.9	-281.5	-96
14-A3	131.7	-18.7	-265.1	-133.4	-166	20-A3	-14.8	-15.1	-264.5	-279.3	-96
17-A3	137.2	-9.3	-261.7	-124.5	-69	17-A3	-15.1	-16.7	-262.1	-277.2	-81

Figure 2.

Graphical summary of ensemble docking in Table 3. The relative energies for apo and antagonist or agonist-bound human adenosine A₃ receptors (hAA₃R) are displayed. We find that all four A₃ selective agonists stabilize the 15^th lowest conformation of hAA₃R while also binding strongly to the 1^st and 3^rd. In contrast three antagonists stabilize the 2^nd lowest conformation of hAA₃R while antagonist MRS5127 prefers the 3^rd but binding to the 2^nd is next best. Experimentally the function of MRS5127 is ambiguous. It appeared to be an antagonist but acted as a partial agonist in cAMP (cyclic Adenosine monophosphate) production in transfected cells with 45% efficacy compared to the full agonist NECA (Adenosine 5’-N-ethyluronamide). All 20 lowest predicted structures for the apo hAA₃R have the “toggle switch” W6.48 in the gauche+ χ1 (NH-C_α-C_β-C_γ) rotamer (tryptophan plane perpendicular to the membrane plane) associated with the inactive G protein-coupled receptor (GPCR) while the lowest predicted structures for the agonist bound state all have the trans χ1 (NH-C_α-C_β-C_γ) rotamer (tryptophan plane parallel to the membrane plane) associated with the active GPCR. The situation is more ambiguous for the antagonist cases. Here MRS5127 and LJ1251 retain gauche. The cases with this trans rotamer are denoted by filled boxes while the cases with the gauche+ rotamer have open boxes. For cases in which the boxes overlap, the higher energy one was shifted higher to avoid overlapping. These cases are marked with asterisks.

3. Ensemble docking of hAA₃R selective antagonist

On the other hand three antagonists,

• MRS3771 [(2S,3S,4R,5R)-5-[2-Chloro-6-(3-iodo-benzylamino)-purin-9-yl]-3,4-dihydroxy-tetrahydrofuran-2-carboxylic acid dimethylamide],

• LJ1251 [(2R, 3R, 4S)-2-(2-Chloro-6-(3-iodobenzylamino)-9H-purin-9-yl)tetrahydrothiophene-3,4-diol] and

• MRS1292 [(2R, 3R, 4S, 5S)-2-[N⁶-3-Iodobenzyl]adenos-9’-yl]-7-aza-1-oxa-6-oxospiro[4.4]-nonan-4,5-diol],

prefer to bind to the predicted structure 2 from Table 3 (TMH6 φ: 15°, TMH4 η: -15°). whereas antagonist

• MRS5127 [(1’R, 2’R, 3’S, 4’R, 5’S)-4’-[2-chloro-6-(3-iodobenylamino)-purine]-2’, 3’-O-dihydroxybicyclo-[3.1.0]-hexane],

prefers predicted structure 3 with TMH4 φ: 15°, TMH6 φ: -30°, TMH4 η: -15°. The second best predicted structure for MRS5127 is also predicted structure 2 of Table 1.

This exception of MRS5127 might arise from a different signaling system. MRS5127 was highly AA₃R-selective (Binding affinity at three human AR subtypes: hA₁ = 3040 ± 610 nM, hA_2A = 1080 ± 310 nM, hA₃ = 1.44 ± 0.6 nM). By Schild analysis of [³⁵S]GTPγS binding to membranes from CHO (Chinese Hamster Overlay) cells expressing the hAA₃R, MRS5127 appeared to be an antagonist.³⁷ However, further analysis determined that it is a partial agonist stimulating cAMP (cyclic adenosine monophosphate) production in transfected cells with 45% efficacy compared to the full agonist NECA (Adenosine 5’-N-ethyluronamide).³⁸

LJ1251 acts more uniformly as an antagonist both in GTPγS binding and in cyclase experiments, but MRS5127 leads to differences in the functional assay.

4. Comparison of conformations for binding agonists and antagonists

The 20 most stable predicted structures for the apo protein all have a gauche+ χ₁ rotamer preference for W6.48 (leading to the vertical orientation with respect to the membrane plane) that has been implicated from experiments on other GPCRs to be associated with the inactive form. ^8-10

We predicted the bound structure of the agonist (Cl-IB-MECA) to each of these predicted 20 lowest apo-hAA₃R structures and found that the predicted best structure (to apo-15) and 60% of the 20 show a preference of W6.48 for the trans χ1 rotamer (leading to the horizontal orientation with respect to the membrane plane). Previous experiments on other GPCRs have indicated that the trans χ₁ rotamer W6.48 is involved in activation.^39,40

These results indicate that the mere binding the agonist to the apo-hAA₃R structure already can reposition the W6.48 conformation from the inactive form to the active form. This may prepare the agonist bound protein for a step that leads to activation, and hence it could be involved in activation as suggested by experiments on other GPCRs.^39.40

Previous experiments on other GPCRs have also suggested that Y7.53 and R3.50 might also serve as rotamer switch upon activation.⁴¹ However, considering each of the top 20 predicted apo-hAA₃R structures bound to either agonists or antagonists, Table S7 in Supporting Information, shows that the χ₁ angles of Y282^7.53 and R108^3.50 in the ligand-bound predicted structures have values nearly identical to the apo-protein. Thus the χ₁ angle of Y282^7.53 is -65° in apo- and antagonist-hAA₃R and -68° for agonist-hAA₃R. Similarly the χ₁ angle of R108^3.50 is -74° in apo-hAA₃R, -73° in antagonist-hAA₃R and -73° for agonist-hAA₃R. This can be compared to the χ₁ angles of are -74.31 for Y282^7.53 and -81.43 ° for R108^3.50 in opsin (PDB ID: 3cap). It may be that the Y7.53 and R3.50 residues play a later step in activation not sampled by our ensemble of structures and dynamics or it may be that hAA₃R does not involve these toggling events.

Fig. S3 compares the best predicted structure for agonist Cl-IB-MECA/hAA₃R to the opsin crystal structure. Here we see similar values for the orientation of the three toggle switch residues, W243^6.48, Y282^7.53, and R108^3.50. This would suggest that in our apo protein structures, Y282^7.53 and R108^3.50 are already in their active form.

Table S7 shows that for apo-hAA₃R the top 20 predicted structures from SuperBihelix all have the gauche+ χ1 rotamers of W6.48.

Table 3 shows that the best conformation (#15 of apo-hAA₃R) for all four agonists has the trans χ1 conformation of W6.48. Indeed, this conformation is found in 12 of the top 20 in the Cl-IB-MECA and IB-MECA-hAA₃Rs and in 6 of the top 20 for the MRS1898 and MRS350-hAA₃Rs.

Similarly Table 3 shows that the 5’-truncated antagonists, MRS5127 and LJ1251, docked to hAA₃R all show gauche+ χ1 rotamers of W6.48 (conformation 2). In contrast the 5’-modifed MRS3771 and MRS1292 prefer the trans χ1 conformation.

Fig. 3 shows the predicted structure 15 of agonist Cl-IB-MECA and the predicted structure 2 of the antagonist LJ1251 which features strongest interactions with the ligand molecule from DarwinDock (Table 3). The orientation of the W6.48 side chain in the agonist-bound hAA₃R is toward TMH5, while the W6.48 in the antagonist bound hAA₃R, is the same as in the apo-protein, which interacts with the NPxxY motif. That is, the agonists all induce a rotamer switch of W6.48 during upon binding. Indeed we find that the agonist-bound predicted structures are highly unfavorable in the gauche+ χ1 conformation of W6.48. Thus Fig. 4 shows how the total cavity E depends on a trans vs. gauche+ χ1 rotamer of W6.48. All agonist-bound complexes lead to an increase in the E of > ~12 kcal/mol in the gauche+ χ1 form of W6.48 due to steric bumps with 5’-substituents.

Figure 3.

Predicted best structures of selective nucleoside antagonist LJ1251 bound to adenosine A₃ receptor (AA₃R) (A), selective agonist Cl-IB-MECA bound to adenosine A₃ receptor (AA₃R) (B) and superimposition of antagonist (blue)/ agonist (red) bound predicted structures (C) without ligands. We predicted the binding to the top 20 predicted structures from the SuperBihelix analysis of the apo-AA₃R as shown in Table 3. LJ1251 binds to predicted structure 2 of apo-AA₃R while Cl-IB-MECA binds to predicted structure 15 of apo-AA₃R The orientation of the W6.48 side chain, shown in yellow, in the agonist-bound hAA₃R is perpendicular to the AA₃R axis and toward TMH5, while its orientation in the antagonist binding parallel to the AA₃R axis allowing it to interacts with the NPxxY motif, just as in the best predicted structure of the apo-AA₃R.

Figure 4.

The relative difference of total cavity energy depending on the trans vs. gauche+ χ1 (NH-C_α-C_β-C_γ) rotamer states of W6.48. All top agonist-bound complexes have over ~12 kcal/mol higher E in the gauche+ χ1 form of W6.48 because of steric bumps with 5’-substituents. However, all antagonist-bound predicted structures have similar energies for both conformations. In the case of nucleoside antagonists with 5’-substituents, the 5’-cyclized uronamide in MRS1292 and 5’-N,N-dimethyl uronamide in MRS3771 are < ~3 kcal/mol unfavorable in the gauche+ χ1 form of W6.48. However, MRS5127 and LJ1251, which were truncated at the 5’-position, are more stable in the gauche+ χ1 conformation of W6.48.

In contrast, we find the antagonist-bound predicted structures lead to similar cavity energies for both conformations. For nucleoside antagonists containing 5’-substituents (MRS1292 with 5’-cyclized uronamide and MRS3771 with 5’-N,N-dimethyl uronamide) are ~3 kcal/mol unfavorable in the gauche+ χ1 form of W6.48. However, MRS5127 and LJ1251, which are truncated at the 5’-position, are more stable in the gauche+ χ1 conformation of W6.48.

The predicted rotamer switch of W243^6.48 induced by binding of the agonist found in this study is consistent with some experimental observations.⁴² Thus the W243A mutation in hAA₃R dramatically impaired functional coupling but did not diminish agonist affinity. Also W243F affected function but did not diminish agonist affinity. Thus we conclude that binding the agonist induces reorientation of the W243^6.48 side chain which affects the receptor’s transition from an inactive to an active state.

We find specific interactions of the ribose moiety in its hydrophilic pocket at T3.36, S7.42, and H7.43. This complex is stabilized by inward movement of F5.43 and the characteristic outward rotation of W6.48 as found in previous docking studies.⁴³

Several previous studies on various GPCRs have suggested that W6.48 functions as a rotamer toggle switch from inactive gauche+ to active trans. The available experimental 3D structures of GPCRs bound to an inverse agonist or antagonist^{8-10, 29} all have W6.48 constrained in the inactive gauche+ conformation either by a network of interactions through water (bRho and the β₁, β₂ adrenergic receptors) or by an extension of N7.45 (hAA_2AR) to D2.50. The x-ray structure of metarhodopsin I suggests that W6.48 undergoes a conformational transition during the process of receptor activation, where it shifts from pointing toward TMH7 in the inactive gauche+ conformation to pointing toward TMH5 in the active trans conformation.³⁹ Based on the Conformational Memories biased Monte Carlo technique it was concluded that rotamer changes of C/S/T6.47, W6.48, and F6.52 in β₂ adrenergic receptor are highly correlated, representing a rotamer “toggle switch” that may modulate the TMH6 Pro-kink.⁴⁴

On the other hand site-directed mutagenesis and molecular modeling approaches suggest that the 5-hydroxytryptamine-4 receptor involves different side-chain conformational toggle switches, i.e. T3.36 from inactive gauche- to active gauche+ and W6.48 from inactive gauche+ to active trans.⁴⁰

5. MD on ligand-protein in membrane and solvent

The above predictions represented the presence of the membrane and water using implicit solvation methods. To test the effect of explicit interactions with membrane and solvent, we inserted each ligand-protein predicted structure into a fully equilibrated hydrated palmitoyl-oleyl-phosphatidylcholine (POPC) lipid bilayer and water, which was subsequently equilibrated for 10 nanoseconds. All charged residues (Asp, Glu, Lys, and Arg) were treated as charged amino acids with CHARMM22 force field parameters.³⁵

The potential E (Fig. S4 in Supporting Information) shows reasonable convergence in 10 ns, suggesting that the predicted structures are reasonably close to a local minimum.

In particular for the agonists the MD finds that the 15^th lowest predicted structure (average RMSD of protein and ligand: 1.5 and 1.3 Å) remains more stable than the first predicted structure (average RMSD of protein and ligand: 2.3 and 2.6 Å), while for the antagonist the second predicted structure (average RMSD of protein and ligand: 2.1 and 1.9 Å) remains more stable than the first predicted structure (average RMSD of protein and ligand: 2.6 and 2.1 Å).

Our predicted structures lead to no significant changes in the salt-bridges or polar contacts for these cases. However, for the agonist Cl-IB-MECA/ hAA₃R complex, we found increased stability of the salt-bridge between D107^3.49 and R126^4.41 compared with LJ1251/ hAA₃R complex. The predicted salt bridge from R108^3.50 of the D(E)RY motif of TMH3 with E225^6.30 of TMH6 remains intact during these dynamics (Fig. S5 in Supporting Information).

As shown in Fig. 5, the W6.48 rotamer for both agonist and antagonist retains the same conformation (horizontal for agonist vs. vertical for antagonist) during the dynamics. We found no significant changes in the salt-bridges or polar contacts for these cases in the predicted structures.

Figure 5.

Molecular dynamics simulation of (A) antagonist LJ1251 in yellow and (B) agonist Cl-IB-MECA in orange bound human adenosine A₃ receptor (hAA₃R) in explicit lipid and water. Bottom) χ1 torsion angle (NH-C_α-C_β-C_γ) variation of W243^6.48 in green through 10 ns simulations. Ligand at the beginning is represented by the stick model and ligand at 10ns is shown in vdW (van der Waals) drawing mode of the VMD (Visual Molecular Dynamics) program.

These limited MD studies validate the local stability of the predicted structures. However, we anticipate that it will take much longer dynamics (perhaps 1 to 1000 microseconds) to observe the changes involved with activation.

6. Subtype selectivity through docking studies using the DarwinDock method

As summarized in Methods the DarwinDock method for predicting ligand binding sites, starts by sampling the full protein to locate putative binding regions, and then aims at sampling a complete set of ligand conformation for each of which we sample a complete set of poses, from which we use the total binding E to select the best poses.

6.1 hAA_2AR selective adenosine

Since adenosine has ~ 30 fold higher selectivity to hAA_2AR than hAA₃R, we docked adenosine, the endogenous agonist, to the lowest E predicted structure of hAA_2AR of Table 1. Table S8 in Supporting Information lists the final 25 poses out of the 5,000 conformations generated by DarwinDock

All poses show similar ribose packing, and indeed nearly all adenosine agonists maintain similar ribose moieties, except for the amide substitution at the 5’-position. Indeed an intact furanose moiety is present in most potent adenosine agonists previously developed.⁴⁵ Our cavity analysis (see Table 4) of the adenosine bound to hAA_2AR shows that the three major contributing amino acids are N253^6.55 (-7.22 kcal/mol), F168^EL2 (-6.33 kcal/mol), and E169^EL2 (-5.13 kcal/mol), based on non-bonding energies [defined in the methods section as the sum of vdW, electrostatic Coulomb with 2.5 dielectric constant and H-bond energies].

Table 4.

Cavity energy of the endogenous agonist adenosine bound to the human adenosine A₃ (AA₃R) and A_2A receptors (AA_2AR). Residues are ordered by total NonBond energy (A_2A), which is the sum of van der Waals (vdW), Coulomb, and H-bond energy (kcal/mol) in the unified cavity. Subtype residues that are predicted to vary between A_2A and A₃ are displayed in bold face. The main interactions leading to subtype selectivity (underlined) in the predicted structure are E169 in A_2A vs. V169 in A₃, and V84 in A_2A vs. L90 in A₃. The color coding for contributions of each residue to binding of the adenosine ligand is: Dark grey: > 3 kcal/mol, Grey: 1 - 3 kcal/mol, Light grey: 0.5 - 1.0 kcal/mol.

AA2AR Ki: estimated 30 nM						AA3R Ki: estimated 1,000 nM
Res	#	VdW	Coulomb	H-Bond	NonBond	Res	#	VdW	Coulomb	H-Bond	NonBond
ASN	253	3.86	-3.03	-8.05	-7.22	ASN	250	2.41	-2.64	-5.59	-5.82
PHE	168	-6.49	0.16	0.00	-6.33	PHE	168	-4.57	0.05	0.00	-4.52
GLU	169	-0.42	-1.16	-3.55	-5.13	VAL	169	0.00	0.00	0.00	0.00
HSE	278	-1.02	-0.08	-1.72	-2.82	HSE	272	0.20	-0.68	-3.55	-4.03
ILE	274	-2.82	0.10	0.00	-2.72	ILE	268	-2.25	0.23	0.00	-2.02
VAL	84	-2.14	-0.24	0.00	-2.39	LEU	90	1.37	-0.30	0.00	1.07
LEU	249	-2.15	-0.04	0.00	-2.19	LEU	246	-3.31	-0.07	0.00	-3.39
SER	277	-0.45	0.08	-1.41	-1.78	SER	271	-1.28	-0.49	0.00	-1.77
LEU	85	-1.39	-0.10	0.00	-1.49	LEU	91	-0.88	-0.08	0.00	-0.96
MET	177	-1.35	-0.04	0.00	-1.39	MET	177	-1.29	-0.05	0.00	-1.34
THR	88	-0.56	-0.63	-0.11	-1.29	THR	94	-0.38	-0.42	0.00	-0.80
MET	270	-1.05	0.00	0.00	-1.05	LEU	264	-1.28	0.01	0.00	-1.27
ALA	63	-0.44	-0.05	0.00	-0.49	ALA	69	-0.33	-0.04	0.00	-0.37
HIS	250	-0.41	0.11	0.00	-0.30	SER	247	0.00	0.00	0.00	0.00
ILE	66	-0.26	0.01	0.00	-0.25	VAL	72	-0.19	0.01	0.00	-0.18
ALA	81	-0.30	0.06	0.00	-0.24	THR	87	-0.73	0.06	0.00	-0.67
ILE	252	-0.21	0.04	0.00	-0.17	ILE	249	-0.19	0.07	0.00	-0.12
MET	174	-0.20	0.03	0.00	-0.17	MET	174	0.00	0.00	0.00	0.00
GLU	13	-0.09	-0.06	0.00	-0.15	GLU	19	0.00	0.00	0.00	0.00
TRP	246	0.00	0.00	0.00	0.00	TRP	243	-1.92	-0.01	0.00	-1.94
LEU	167	0.00	0.00	0.00	0.00	GLN	167	-0.31	-0.13	0.00	-0.44
ALA	59	0.00	0.00	0.00	0.00	VAL	65	-0.33	-0.01	0.00	-0.34
THR	256	0.00	0.00	0.00	0.00	ILE	253	-0.12	0.08	0.00	-0.04
SUM					-37.56	SUM					-28.92

In the predicted structure of adenosine-hAA_2AR, we find that the exo-cyclic amino group of the adenine ring interacts with N253^6.55 and E169^EL2 in the same way that it interacts with the amino group in the tri-cyclic ring of ZM241385 seen in the x-ray structure of hAA_2AR, as shown in Fig. 6A and 7A. An additional predicted interaction (-2.82 kcal/mol) occurs between the 2’ and 3’-OH group of the nucleoside ring and the side chains of the double protonated basic H278^7.43, which also interacts with E13^1.39. Supporting this, the Ala mutation of N253^6.55 shows loss of agonist and antagonist binding, while the H278E^7.43 mutation displays substantial decrease of agonist and antagonist binding except in the case of N⁶-substituted agonists.⁴⁶

Figure 6.

Schematic of the predicted binding sites for adenosine receptor agonists based on the cavity binding energies. Left: adenosine at hAA_2AR; Right: Cl-IB-MECA at hAA₃R

Figure 7.

Predicted binding sites (A) adenosine to the adenosine A_2A receptor (AA_2AR), (B) adenosine to the adenosine A₃ receptor (AA₃R), (C) A₃ selective agonist Cl-IB-MECA at AA_2AR, (D) A₃ selective agonist Cl-IB-MECA at AA₃Rs. The high selectivity of Cl-IB-MECA for AA₃R relative to AA_2AR is due to the salt-bridge between E169 and H264 (red circle in (C)) in AA₃R. In AA_2AR we have V169 but there is no residue corresponding to H264 because of the gap in AA_2AR. H-bonding is indicated by red dots, and subtype selective residues are shown in red.

We then matched the predicted ligand binding structure from hAA_2AR to hAA₃R, leading to common interactions at N250^6.55 (-5.82 kcal/mol), F168^EL2 (-4.52 kcal/mol), and H272^7.43 (-4.03 kcal/mol). However, V169^EL2 in hAA₃R corresponds with E169^EL2 in hAA_2AR losing the interaction with the terminal amino group, resulting in unfavorable interactions in hAA₃R because of the different loop geometry of EL2 (Table 4 and Fig. 7B). The result is a dramatic decrease in binding affinity (cavity sum = -37.56 for hAA_2AR vs. -28.92 for hAA₃R) in agreement with the dramatically decreased experimental binding affinity of endogenous adenosine at both rat ARs (rAA_2AR estimated K_i: 30 nM vs. rAA₃R estimated K_i, 1,000 nM).⁴⁵

6.2 hAA₃R selective Cl-IB-MECA agonist

We docked Cl-IB-MECA to the lowest E predicted structure of hAA₃R of Table 1. From DarwinDock we selected the final 125 poses (five different ligand conformations with 25 poses each) out of 25,000 conformations. These predicted structures were ordered by the average rank of neutral total and Unified cavity E (UniCav E) in Table S9 in Supporting Information. From the cavity analysis of A₃ subtype selective Cl-IB-MECA bound to hAA₃R, Table 5 shows that the major contributing amino acids are F168^EL2 (-9.79 kcal/mol), N250^6.55 (-4.44 kcal/mol), L246^6.51 (-4.28 kcal/mol), and H272^7.43 (-3.35 kcal/mol). The terminal benzyl group of Cl-IB-MECA is surrounded by hydrophobic residues, V169^EL2 and I253^6.58, as shown in Fig. 6B and 7D. The N⁶-bulky group extends an additional interaction between EL2 and upper TMH6, making stronger interactions at hAA₃R.

Table 5.

Cavity energy of the A₃ selective Cl-IB-MECA bound to the human adenosine A₃ (AA₃R) and A_2A receptors (AA_2AR). The residues are ordered by the total NonBond energy (A₃), which is the sum of van der Waals (vdW), Coulomb, and H-bond energy (kcal/mol) in the unified cavity. Subtype residues that vary between A_2A and A₃ are displayed in bold face. The main interactions leading to subtype selectivity (underlined) are I253 in A₃ vs T256 in A_2A, V169 in A₃ vs E169 in A_2A, and Q167 in A₃ vs L167 in A_2A, leading to 3.85 of the predicted 7.31 differential binding. The other major differences, involving conserved residues are for F168, S273, M174 favoring A₃ and N250, I268 favoring A_2A, leading to a differential of 2.16 of the predicted 7.31 differential binding. The color coding for contributions of each residue to binding of the adenosine ligand is: Dark grey: > 3 kcal/mol, Grey: 1 - 3 kcal/mol, Light grey: 0.5 - 1.0 kcal/mol

AA3R Ki: 1.4 nM						AA2AR Ki: 5,360 nM
Res	#	VdW	Coulomb	H-Bond	NonBond	Res	#	VdW	Coulomb	H-Bond	NonBond
PHE	168	-9.22	-0.56	0.00	-9.79	PHE	168	-7.54	-0.46	0.00	-7.99
ASN	250	2.46	-1.15	-5.74	-4.44	ASN	253	6.32	-1.64	-10.54	-5.86
LEU	246	-4.49	0.22	0.00	-4.28	LEU	249	-4.42	0.23	0.00	-4.19
HSE	272	-1.15	-0.57	-1.64	-3.35	HSE	278	-1.44	-0.24	-0.88	-2.55
LEU	264	-2.77	-0.10	0.00	-2.87	MET	270	-3.50	0.01	0.00	-3.49
ILE	253	-2.71	-0.16	0.00	-2.87	THR	256	-0.92	-0.37	0.00	-1.29
VAL	169	-2.45	-0.37	0.00	-2.82	GLU	169	-1.04	-0.89	-0.01	-1.93
TRP	243	-2.02	0.00	0.00	-2.02	TRP	246	-0.98	-0.22	0.00	-1.20
ILE	268	-1.87	-0.04	0.00	-1.91	ILE	274	-2.99	-0.11	0.00	-3.10
MET	177	-1.63	-0.09	0.00	-1.72	MET	177	-1.62	-0.11	0.00	-1.73
GLN	167	-0.64	-0.57	-0.17	-1.38	LEU	167	0.00	0.00	0.00	0.00
THR	87	-0.86	-0.51	0.00	-1.37	ALA	81	-0.35	-0.39	0.00	-0.75
LEU	90	-1.25	-0.03	0.00	-1.28	VAL	84	-1.67	-0.07	0.00	-1.74
SER	271	-0.84	-0.21	0.00	-1.05	SER	277	0.20	-0.03	0.00	0.17
MET	174	-1.09	0.04	0.00	-1.05	MET	174	0.34	0.38	0.00	0.72
TYR	265	-0.59	-0.08	0.00	-0.67	TYR	271	-0.48	-0.14	0.00	-0.63
MET	172	-0.63	-0.02	0.00	-0.65	VAL	172	-0.61	-0.06	0.00	-0.66
THR	94	-0.39	-0.22	0.00	-0.61	THR	88	-0.06	0.06	0.00	0.01
ILE	249	-0.57	0.03	0.00	-0.53	ILE	252	-0.61	0.04	0.00	-0.57
LEU	91	-0.50	0.02	0.00	-0.48	LEU	85	-0.93	0.01	0.00	-0.92
VAL	72	-0.33	-0.07	0.00	-0.40	ILE	66	-0.47	0.05	0.00	-0.42
ALA	69	-0.37	0.00	0.00	-0.38	ALA	63	-0.57	-0.05	0.00	-0.62
VAL	65	-0.35	0.11	0.00	-0.25	ALA	59	-0.16	0.17	0.00	0.01
SER	242	-0.13	0.07	0.00	-0.05	CYS	245	-0.10	-0.05	0.00	-0.15
ARN	173	0.00	0.00	0.00	0.00	PRO	173	-0.26	0.02	0.00	-0.24
GLY	257	0.00	0.00	0.00	0.00	PRO	260	-0.36	-0.03	0.00	-0.40
						HIS	264	0.95	-0.31	0.00	0.64
SUM					-46.17	SUM					-38.86

To understand the origin of subtype selectivity, we matched the predicted best binding pose of the highly A₃-selective ligand Cl-IB-MECA to the hAA_2AR predicted structure, and then we used SCREAM²⁸ to predict the optimum side chain position of residues in the binding pocket, after which we minimized the final ligand/protein complex post neutralization. In hAA_2AR, the residue-ligand interactions were hampered by the constrained loop between E169^EL2 and H264^EL3 via a salt-bridge, as shown in Fig. 7C. The corresponding amino acids of I253^6.58, V169^EL2, and Q167^EL2 in hAA₃R are T256^6.58, E169^EL2, and L167^EL2 in hAA_2AR. All three subtype variable residues in hAA_2AR show weakened interactions in the cavity of hAA_2AR in Table 5, resulting in a 7.31 kcal/mol favorable interaction in hAA₃R. This predicted structure is consistent with the experimental binding affinity of A₃ subtype selective Cl-IB-MECA (AA₃R Ki: 1.4 nM vs. AA_2AR Ki, 5,360 nM).⁴⁷ The docking result also agrees with experimental observations, where the subtype selectivity of adenosine derivatives as AR agonists has been probed extensively, principally through modification of the N⁶-amine moiety (where large hydrophobic groups tend to produce AA₁R and AA₃R selectivity) and the C2 position (where large hydrophobic groups tend to produce AA_2AR selectivity).⁴⁸

Based on the docking studies of the subtype selective antagonists, we suggest that I253^6.58 and V169^EL2 in hAA₃R are involved in vdw interactions with the N⁶-bulky group in the A₃ selective ligands, while E169^EL2 in hAA_2AR (which is the corresponding amino acid of V169^EL2 in hAA₃R) interacts with the free amino group of the A_2A selective ligands through additional H-bonding. Thus this predicted structure explains the increase of A₃ selectivity for the additional N⁶-bulky substituent at hAA₃R. However, Q167^EL2 in hAA₃R is located in the proximity of the C2 substituent which is oriented toward the EC2 in the docking model. The corresponding amino acid L167^EL2 in hAA_2AR might explain the increase of A_2A selectivity for large hydrophobic C2 substituents which are found in most A_2A selective antagonists.⁴⁸

Conclusions

We used the GEnSeMBLE Monte Carlo Technique to predict an ensemble of the 20 top apo-protein predicted structures for all 4 subtypes of the human adenosine receptor. For hAA₃R we used the DarwinDock Monte Carlo Technique to predict the best binding pose for four subtype selective agonists and antagonists.

We find that all four selective agonists prefer what was the 15^th predicted structure of the apo protein. In addition, all four change the W6.48 conformational state from gauche+ (vertical) to trans (horizontal). This is consistent with observations in other GPCRs that switching of W6.48 from gauche+ (vertical) to trans (horizontal) is associated with activation of the G-Protein.

In contrast the four selective antagonists all prefer ether the 2^nd or 3^rd conformation of the apo protein. However the situation with respect to the toggle switch is mixed. MRS5127 and LJ1251, for which the ribose is dramatically modified, retain the gauche+ conformational of W6.48 found in the apo protein for all 20 low lying conformations. However MRS1292 and MRS3771, which retain the ribose unit of the adenosine agonist but with substitutions for the CH₂OH, prefer the trans (active) conformation in the lowest E conformation, but has the gauche+ conformation for various low lying cases.

These results provide strong support for the role of W6.48 as a rotamer toggle switch affecting activation, suggesting that our methods have successfully found the ensemble of low lying conformations relevant for ligand binding and for activation of the GPCR.

We also predicted the low lying conformations of the adenosine A_2B, A_1, and A₃ receptors subtypes finding:

1. Excellent agreement of subtype specificities for binding of the A₃ selective Cl-IB-MECA and A_2A selective adenosine agonists between these two ARs.

2. Excellent agreement between the binding site of the adenosine agonist bound to hAA_2AR with available mutation data.

These results suggest that these computational methods can predict the ensemble of low lying predicted structures accurately enough to extract information relevant for understanding activation and accurately enough to predict the changes in conformation for different subtypes. Moreover the detailed binding sites are accurate enough to predict the differential binding characteristics between different subtypes and between agonists and antagonists binding to the same subtype. This suggests that these methods may be useful for designing subtype selective ligands for GPCR targets, even when no experimental structural information is available.

The success of these studies relies in part on having an experimental structure of hAA_2AR, indicating the value of having a crystal structure of a related system. The theory can then build on this structure to predict the structures of other subtypes. Moreover the theory can then predict the structures for agonists not yet available from experiment.

Abbreviation

AA₃R

Adenosine A₃ receptor

GPCR

G Protein-Coupled Receptor

vdW

van der Waals

TMH

transmembranic helix