Structure-Based Discovery of Novel Chemotypes for Adenosine A_2A Receptor Antagonists

Abstract

The recent progress in crystallography of G-protein coupled receptors opens an unprecedented venue for structure-based GPCR drug discovery. To test efficiency of the structure-based approach, we performed molecular docking and virtual ligand screening (VLS) of more than 4 million commercially available “drug-like” and ‘‘lead-like’’ compounds against the A_2AAR 2.6 Å resolution crystal structure. Out of 56 high ranking compounds tested in A_2AAR binding assays, 23 showed affinities under 10 µM, eleven of those had sub-µM affinities, and two compounds had affinities under 60 nM. The identified hits represent at least 9 different chemical scaffolds and are characterized by very high ligand efficiency (0.3–0.5 kcal/mol per heavy atom). Significant A_2AAR antagonist activities were confirmed for 10 out of 13 ligands tested in functional assays. High success rate, novelty and diversity of the chemical scaffolds and strong ligand efficiency of the A_2AAR antagonists identified in this study suggest practical applicability of receptor-based VLS in GPCR drug discovery.

Introduction

G-protein Coupled Receptor (GPCR) signaling is a key to molecular pathways involved in the normal function and pathologies of nervous, cardiovascular, immune and other major systems in human body. Among about 400 non-olfactory receptors of the GPCR superfamily, more than 50 represent molecular targets for prescription drugs^1–2, and the range of clinically validated GPCRs and investigational drugs is rapidly expanding. One of the major spotlights is on the adenosine receptor (AR) subfamily³, all four subtypes of which (A₁, A_2A, A_2B and A₃) have been considered as potential therapies for neurodegenerative^4–5, cardiac^6–7, immune and inflammatory disorders^8–9 and cancer¹⁰. Despite some obstacles in clinical development of early drug candidates for ARs^11–12, the year 2008 has been marked by successful FDA approval of the new generation A_2AAR selective agonist regadenoson as a coronary vasodilator for use in myocardial perfusion imaging¹³. This breakthrough, along with other advances in pre-clinical and clinical studies³ boosts interest to development of a new generation of bioavailable and safe agonists and antagonists for adenosine receptors.

Historically, GPCR drug discovery relied on known natural ligands or screening assay hits as starting points for optimization of affinity, subtype selectivity and pharmacokinetic properties^14–16. The most useful scaffolds for design of AR ligands have been provided by adenosine^17–18 and xanthine^19–20 chemotypes. Thus, adenosine derivatives with various substitutions in position 2 or N⁶ of the adenine ring and 3’,4’, or 5’ position of the ribose ring¹² have been developed as selective agonists for all four AR subtypes; only a few other chemotypes²¹ have been found with agonist activity. Since the early discovery of caffeine and theophylline as non-selective AR antagonists^19–20,22 derivatization of the xanthine scaffold yielded a number of high affinity subtype selective antagonists¹². Several other chemotypes for AR antagonists have been discovered over the last decade^23–30 using a combination of experimental screening and ligand-based methods. The ligand-based approaches, however, require preexisting knowledge of ligand structure activity relationships (SAR), and are largely limited to relatively close analogues of known ligands (Chart 1).

Chart 1.

Chemical structures of representative antagonists (1 and 2) and agonists (3 and 3a) of A_2A adenosine receptor.

The breakthroughs in GPCR crystallography, including determination of high resolution structures of β-adrenergic receptors (human β₂AR^31–33 and turkey β₁AR³⁴), and most recently of human A_2A adenosine receptor³⁵ (A_2AAR) in complex with antagonist 1 (ZM241385³⁶), open an opportunity for alternative, receptor-based approaches to finding new GPCR ligand chemotypes³⁷. Indeed, the β₂AR crystal structure-based models have already proved to be efficient in virtual screening for antagonists/inverse agonists, and, with some modifications³⁸, also for full and partial agonists of the receptor^39–40. Recently, virtual ligand screening for the β₂AR allowed identification of new nanomolar and submicromolar inverse agonists for this receptor⁴¹.

In this study we assessed performance of the human A_2AAR structure³⁵ (PDB code: 3EML) in virtual screening, which resulted in identification of several novel ligand chemotypes for ARs. The initial benchmarking with known antagonists showed enrichment factors for the A_2AAR on par with previously published assessments for the β₂AR^39–40. Performance of the screening model was further improved by retaining several highly structured water molecules in the binding site and refining side chains in the binding pocket. The optimized model was used for virtual screening of more than 4 million commercially available lead-like and drug-like compounds. Out of 56 high scoring compounds, which were subsequently tested in radioligand binding assays, 23 compounds were identified as A_2AAR ligands with affinity K_i <10 µM (41% hit rate), of which 11 had sub-µM affinity, and two compounds showed K_i under 0.06 µM. Functional assays confirmed significant A_2AAR antagonist activities for 10 out of 13 newly identified ligands. The novel ligands represent at least nine novel chemotypes, and include low molecular weight compounds with high ligand binding efficiency (LE>0.3 kcal/mol per heavy atom), potentially suitable as leads for drug discovery. Overall, a very good success rate of VLS, high ligand efficiency of the hits, as well as diversity and novelty of the identified chemical scaffolds suggest applicability of the receptor-based screening to discovery of new candidate drugs to adenosine receptors and potentially other GPCRs.

Results

High-performance A2AAR screening model based on the crystal structure

The A_2AAR crystal structure³⁵ (PDB code :3EML) and ligand-refined models of the A_2AAR were evaluated for their ability to select known A_2AAR specific antagonists from random decoy compounds in a docking and virtual ligand screening (VLS) benchmark test. As illustrated in Figure 1 examples, the predicted binding poses for majority of the known A_2AAR antagonists in the diverse benchmark set display key similarities with the binding mode of antagonist 1 in the crystal structure. This common binding motif involves stacking between aromatic moieties of the ligands and the conserved Phe168(5.29) side chain of the receptor, as well as polar interactions with conserved Asn253(6.55) side chain (Residue numbers in parentheses are based on Ballesteros-Weinstein nomenclature⁴²). Compounds shown in Figure 1 also form a hydrogen bond between the unsubstituted exocyclic amine group and the Glu169(5.30) side chain. In addition to these core interactions, most high affinity A_2AAR antagonists have an aromatic group extending deeper into the binding pocket and/or flexible extensions towards the extracellular opening of the pocket.

Figure 1.

Binding of known antagonists into human adenosine A_2A receptor models. (A) Co-crystal structure with compounds 1 as in PDB entry 3EML³⁵, ligand carbons are shown in magenta color. (B–D) Optimized model 3EML_{W3_opt} with 3 top scoring compounds from the benchmark set (B) mantri(n)²⁷, (C) mantri(f) ²⁷ (D) mantri(k)²⁷. For all ligands, the A_2A binding pocket is shown as transparent skin colored by properties (green: hydrophobic, red: acceptor, blue: donor of H-bond). Water molecules are shown by thin red lines and with labels for the three structured waters with lowest B-factors. Hydrogen bonds are shown by cyan spheres.

The results in Figure 2 assess overall performance of several screening models, which were generated from the crystal structure (3EML) by optimizing polar hydrogens and structured water configuration. The Receiver Operating Characteristic (ROC) curves, in semi-logarithmic scale to emphasize initial enrichment, suggest very good overall screening performance for all six models, with the area under curve (AUC) values within 89–91% range. However, the model with three structured water molecules in the binding pocket, 3EML_W3 achieved a significantly higher initial enrichment factor, EF(1%) = 70%, as compared to the 3EML_W0 model without water, EF(1%)=43 %, or any of the other 13 water configurations tested . Analysis of the crystal structure shows that the three waters selected for the 3EML_W3 model (wa, wa14 and wa5 in the 3EML PDB entry) have the lowest B-factor values and form an extended hydrogen bonding network with the binding pocket residues, suggesting their highly structured nature. While the presence of water in the models does not contribute significantly to the predicted binding scores of the known ligands, the selected structured water molecules occupy highly polar sub-pockets in the A_2AAR and apparently prevent adverse binding of some decoy compounds into these sub-pockets.

Figure 2.

Performance of A_2AAR screening models with different number of structured water molecules (W0 to W4) and with conformational optimization (W3_opt). Receiver Operating Characteristic (ROC) curves are shown in semi-logarithmic scale.

Further modest improvement of initial recognition performance (EF(1%)=78) was achieved for 3EML_{W3_opt} model by ligand-guided optimization of side chains in the binding site of the A_2AAR structure, as described in Methods . More than 200 candidate models were obtained by docking representative antagonists into the all-atom model of the A_2AAR with fully flexible side chains of the binding pocket. The best performing model was selected based on the same VLS benchmark test as described above, using a normalized square root AUC (NSQ_AUC) as a metric for model selection (see Methods for NSQ_AUC definition). Note that the conformational changes between the crystal structure and the 3EML_{W3_opt} model are minor, with RMSD about 0.27 Å for the 12 side chains of the binding pocket.

Identification of candidate A_2AAR antagonists by virtual compound screening

The optimized model of the A_2AAR, 3EML_{W3_opt}, was then used to screen for candidate binders from about 4.3 million commercially available drug- and lead-like compounds collected in the Molsoft ScreenPub non-redundant screening database. About 1800 of the library compounds docked into the A_2AAR model with the ICM binding score better than the standard threshold (−32 kJ/mol). These candidate hits were clustered into 340 groups for chemical similarity (Tanimoto distance cutoff >0.3); the cluster size ranged from 1 to 277 compounds. In each cluster we selected from 1 to 10 different compounds according to their predicted binding score and predicted ligand efficiency, LE_pred. The LE_pred, which is defined as binding score per heavy atom of the compound, may correlate with compound suitability as a lead for chemical optimization. Other criteria for the compound selection included their predicted logP (octanol/water partition) and logS (aqueous solubility) values, as well as immediate availability of compounds from major chemical vendors. Compounds with less than 0.3 Tanimoto distance to known AR ligands in GLIDA database⁴³ were removed to maintain novelty of the compounds in the testing library. This resulted in a final set of 56 diverse drug-like or lead-like compounds that were ordered and screened in experimental assays (see Supporting InformationTable SI1 for the full list of tested compounds).

Figure 3 illustrates predicted A_2AAR binding poses of selected candidate compounds representing 16 different chemical scaffolds. All candidate chemotypes except the L cluster were predicted to share some key features with antagonist (1) binding in the A_2AARcrystal structure, including stacking interaction with Phe168(5.29) and hydrogen bonds between exocyclic amine donor and acceptors in Asn253(6.55) and Glu169(5.30) side chains. Most compounds have an additional acceptor, which forms a hydrogen bond to Asn253(6.55) donor in our models.

Figure 3.

Examples of binding poses in the 3EML_{W3_opt} receptor model for predicted A_2AAR candidate ligands representing 16 different chemical clusters. The letters (A to P) show cluster names, the numbers in brackets indicate compound IDs. Labels for the (i) contact side chains, (ii) TM domains and (iii) structured waters of the receptor model are shown in panels A to C. Side chains of the receptor are shown by sticks with white carbon atoms, in ligands the carbon atoms are colored yellow. Structured water molecules in the model are shown by purple sticks. Hydrogen bonds are colored according to their predicted strength, from blue (strongest) to red (weakest). The binding pocket surface is shown by purple skin, which is clipped in foreground for better view.

Apart from these core interactions, many predicted compounds in Figure 3 (clusters A, D, F, G,H,I,K,P) have moieties extending deep in the binding pocket towards Trp246(6.48) and His250(6.52) and making additional, mostly hydrophobic contacts. Many compounds also have extensions towards the extracellular opening of the pocket (clusters A,B,D,E,F,G,J,K,M,N,O) and/or towards helix TM1 (clusters B,E,H,N,O) which may play a role in their subtype selectivity.

Results of A_A2AR Binding Assay in insect cell membranes

Experimental testing of the 56 selected compounds resulted in identification of 23 new A_A2AR ligand with Ki better than 10 µM (41% hit rate), and 11 ligands with Ki better than 1 µM. The most efficient A_2AAR binders are presented in Table 1; this includes compounds 9 (K_i =0.032 µM) and 15 (K_i =0.06 µM) and other compounds with sub-µM affinity or exceptional ligand efficiency. For the results of the binding test for all 56 compounds see Supporting Information, Table SI1.

Table1.

Representative hits clustered by different chemotypes, with their A_2AAR binding characteristics and predicted ICM Scores. Clusters are named as in Figure 3. The last column shows compounds in GLIDA database of ~30,000 GPCR ligands⁴³ with highest chemical similarity to the identified hits.

ID	Cluster	Chemical Structure	Ki µM	95% confidence interval	Maxa Inhibition %	LEb	ICM Scorec	Closest Analog in GLIDA (Tanimoto Distance)
9	A		0.032	(0.023–0.043)	80	0.38	−36.3
8	B		0.63	(0.12–3.3)	54	0.3	−34.3
15	B		0.06	0.0082–0.45)	91	0.45	−32.4
49	C		0.25	(0.21–0.3)	100	0.5	−36.4
35	C		0.64	(0.43–0.94)	96	0.4	−39.1
29	C		0.32	(0.25–0.4)	88	0.49	−37.5
26	C		0.53	(0.29–0.98)	90	0.45	−37.8
6	D		0.3	(0.15–0.62)	70	0.35	−35
14	D		1	(0.73–1.4)	78	0.35	−34.8
10	D		2.7	(1.4–5.3)	92	0.3	−35.6
50	E		0.8	(0.44–1.5)	78	0.39	−37.8
17	E		0.56	(0.19–1.6)	80	0.35	−38.4
21	I		2.9	(1.7–5.1)	66	0.4	−38.5

^aMaximum Inhibition was achieved at 10 µM concentration for compound 9 and at 32 µM for compound 15. All other compounds showed maximum inhibition at 1 mM.

^bExperimental ligand efficiency, kcal/mol , calculated as ligand binding energy per heavy atom.

^cPredicted ICM binding score, kJ/mol

Diversity of the identified A_2AAR ligands is supported by the fact that compounds in Table 1 belong to six different clusters (A,B,C,D,E,I in Figure 3), and an additional three clusters (F,G,H) contain other binders with K_i better than 10 µM. The novelty of the ligands was also assessed by searches in GLIDA database that comprises 795 specific binders of ARs used in clinical studies and research. We confirmed that all compounds in Table 1 have Tanimoto distances exceeding 0.3 to their closest analogues in the GLIDA database⁴³.

Figure 4 illustrates the binding curves for the compounds in Table 1. Antagonists 1 and 2 (theophylline), as well as agonist NECA⁴⁴ (3) were used as the reference compounds, and their observed K_i values (0.0006, 0.140 and 4.4 µM respectively) correlated well with the K_i values published previously¹². The K_i value and the binding curve for compound 9 are shown for concentrations below 2 µM; at this concentration the compound displaces up to 80% of the hot ligand. In contrast, at higher concentrations of this compound (10 to 1000 µM) displacement of the hot ligand drops to lower levels (see Figure SI1). This unusual behavior of compound 9, confirmed in 6 runs of the binding assay, can be attributed to cooperative aggregation, as also suggested by visible sedimentation of the compound at concentrations above 10 µM.

Figure 4.

Examples of competition binding curves for the identified A_2AAR ligands, as compared for known A_2AAR antagonists 1, 2, and agonist 3. Tritiated compound 1 was used as the radioligand.

Note that all A_2AAR binders shown in Table 1 have ligand efficiency better tahn 0.3 kcal/mol per heavy atom, which is considered to be optimal for lead-like compounds⁴⁵. The highest value LE=0.5 kcal/mol per heavy atom was detected for compound 49 , (18 heavy atoms and MW=257) and two of its analogues in cluster C. Smaller compounds, like compound 21 in cluster I with LE=0.4 kcal/mol per heavy atom, may be especially valuable as lead scaffolds for drug discovery, since they provide more room for chemical optimization⁴⁵. Each of the chemical clusters and some of the binding features of the ligands are discussed in more details in the Discussion section.

Ligand binding and AR subtype selectivity in mammalian cell membranes

In addition to Sf9 insect cells, binding of some A_2AAR ligands was also assessed for A₁ and A₃ AR subtypes expressed in mammalian cells (A_2B subtype selectivity was not measured here). The left columns in Table 2 show the results of this assay for the A_2AAR in HEK293 cell membranes. These results confirm the K_i values under 1 µM in mammalian based assays for all submicromolar binders shown in Table 1. The observed variations in affinities are expected between Sf9 and HEK based assay systems, as these assays involve a different pattern of posttranslational modifications, lipid composition and a different set of experimental conditions.. The most significant deviation was found for compound 10, which did not show any substantial binding to the HEK293 expressed A_2AAR protein. Results of binding to the A₁AR subtype in the right columns of Table 2 suggest only a modest decrease in binding affinities for majority of the tested compounds, as compared to A_2AAR binding. Substantial (more than 10 fold) selectivity for the A_2AAR over the A₁AR was observed for two compounds, 17 and 50, which both belong to the same chemotype (cluster E). One compound, 8, had a pronounced selectivity to the A₁AR subtype (20 fold). Full affinity measurements for A₃AR subtype were not attempted, because single point assays suggested that none of the compounds had apparent selectivity to the A₃ subtype, while a majority was characterized by the impaired A₃AR binding (see Figure SI2).

Table 2.

Competition binding of the hit compounds to mammalian expressed adenosine receptor subtypes A_2A and A₁.

	Displacement of radioligand at A2AAR-HEK membranes			Displacement of radioligand A1AR-CHO membranes			A2A / A1 selectivity
ID	Ki, µM	(95% Confidence Interval), µM	Max Displacementa, %	Ki, µM	(95% Confidence Interval), µM	Max Displacementa, %	fold
6	0.06	(0.03–0.12)	48	0.09	(0.04–0.19)	32	1.42
8	1.14	(0.67–1.94)	53	0.06	(0.05–0.08)	100	0.05
9	0.49	(0.20–1.17)	41	1.44	(0.56–3.70)	41	2.94
10	>10	ND	ND	1.12	(0.44–2.88)	28	ND
14	0.06	(0.05–0.09)	54	0.14	(0.04–0.52)	25	2.24
15	0.56	(0.31–0.99)	73	0.48	(0.12–1.96)	27	0.87
17	0.72	(0.55–0.95)	73	>10	ND	ND	>10.
21	3.01	(0.69–13.1)	57	0.39	(0.31–0.50)	86	0.13
26	0.10	(0.08–0.11)	100	0.26	(0.21–0.33)	100	2.74
29	0.11	(0.09–0.14)	100	0.28	(0.23–0.33)	100	2.51
35	0.20	(0.15–0.25)	100	0.15	(0.13–0.17)	100	0.75
49	0.94	(0.78–1.13)	100	0.19	(0.16–0.22)	100	2.01
50	0.15	(0.11–0.21)	50	1.75	(0.45–6.81)	27	11.42

^aMax displacement refers to the maximal displacement of bound radioligand observed using 10 µM of test compound.

The identified compounds inhibit A_2AAR-mediated cAMP production in functional assays

We performed cell-based functional assays to evaluate activity of the newly identified adenosine A_2A receptor binders as receptor antagonists. HEK293T cells stably expressing the A_2AAR were stimulated with either 10 nM or 10 µM of the A_2AAR agonist CGS21680 (K_i=27 nM)¹², and intracellular cAMP levels were determined using the LANCE cAMP TrFRET kit (Perkin Elmer), as described in Methods. The ability of the various compounds to block A_2A receptor mediated cAMP generation was assessed at a 10 µM concentration of the compounds. The results in Figure 5 show that 11 out of the 14 tested A_2AAR binders were able to effectively block more than 75% of cAMP generation at 10 nM concentration of agonist 3a (CGS21680), which strongly supports their A_2AAR antagonist activity. Most effective antagonists (compounds 29, 35, 49) inhibited more than 90% of cAMP production for cells stimulated with 10 nM of 3a. Due to the comparatively low affinity of the tested compounds as compared to the agonist 3a the majority of compounds were unable to block the effect of this agonist used at a saturating concentration of 10 µM. No A_2AAR specific activity was observed for compound 8. Compounds 6 and 10 blocked cAMP production at both 10 nM and 10 µM concentrations of 3a to the same extent, suggesting their ability to inhibit cAMP production was not A_2AAR -specific. Subsequent experiments revealed that compound 6 inhibited forskolin stimulated cAMP generation, clearly indicating an effect independent of the A_2AAR (data not shown).

Figure 5.

Inhibition of agonist-induced cAMP production by the A_2AAR hit compounds. Activity is measured at 10 µM concentration of the compounds, as compared to no compound in the control. Data is normalized where 0 % represents the unstimulated condition (black bars) and 100 % represents the accumulation of intracellular cAMP observed for stimulation with compound 3a at 10 nM or 10 µM concentrations (shaded and open bars respectively).

Discussion

Our results suggest that receptor-based virtual ligand screening is an efficient tool for discovery of novel chemically diverse antagonists of adenosine A_2A receptor. A number of A_2AAR ligands selected here by VLS not only display submicromolar affinity and significant functional activity in mammalian cells, but also have small molecular weight and high ligand efficiency (LE> 0.3 kcal/mol per heavy atom) suitable for lead optimization. Below we discuss each of the diverse chemical scaffolds identified and their A_2AAR interaction modes as predicted by VLS docking (see Figure 3 and Table 1 above).

Cluster A is represented by only one compound 9 based on 1,3,5-triazine-2,4-diamine scaffold, which has the best A_A2AR affinity in the set (K_i = 0.03 µM in Sf9 based assay). The functional activity of compound 9 in HEK293 cells is also one of the highest among the tested ligands (see Figure 4). As seen in Figure 3A, compound 9 is predicted to form a hydrogen bond network with Asn253(6.55) and Glu169(5.30). The compound’s 3-methyl-benzofuran moiety is extended downward and reaches deeper into the binding pocket than the furan moiety of the antagonist 1 in the crystal structure³⁵, suggesting that this part of the pocket can be exploited in ligand design. The 5-chloro-2-methoxyphenyl moiety stretches towards TM2 helix and forms predominantly hydrophobic contacts with Ile66(2.64), Ala63(2.61) and Ile274(7.39), although an alternative extended orientation of this ring has also been observed with only slightly inferior binding scores.

Cluster B includes compounds 15 and 8 (K_i = 0.06 µM and 0.63 µM respectively in Sf9 assays), which have a common 2-amino-3-cyano-4-phenylpyridine motif, though in a different context of bi- and tricyclic aromatic ring systems. For both compounds the amine and cyano moieties were predicted to form a strong hydrogen bonding network with Asn253(6.55) and Glu169(5.30). Interestingly, while compound 1 and most other known antagonists have nearly flat configurations deep in the binding pocket, the phenyl ring of both compounds 15 and 8 is intrinsically out of plane with the core pyridine ring. The non-planar configuration is apparently accommodated by the pocket as it is in the crystal structure, but this feature may result in different conformational preferences for the A_2AAR receptor, potentially leading to unusual binding and /or functional properties of these ligands. Indeed, we found that compound 8 does not have any functional activity in A_2AAR cells (see Figure 4); at the same time it is the only compound in our set with significant binding selectivity to A₁AR subtype.

Cluster C is the most populous one and represents as many as nine 1-amino-anthraquinone compounds in our test list. Of those, four (26, 29, 35, 49 in Table 1) have sub-µM affinity for the A_2AAR and another three (38, 20, 28 in Table SI1) have low-µM affinity. Quite remarkably, all four sub-µM ligands in this cluster have significant functional activity as antagonists in mammalian cells. Some of the smaller compounds in this group are characterized by an exceptionally high ligand efficiency (LE) that exceeds 0.45 kcal/mol per heavy atom (compounds 49,29,26). These compounds have different small groups in anthraquinone position 4, which seems to be rather permissive to substitutions. This site may be useful for optimization of the scaffold affinity and chemical properties, though a few bulky substituents tested in this initial screening (38, 20, 18, 23) were apparently suboptimal. Compound 35 (K_i =0.64 µM in Sf9) has a modified scaffold with an additional conjugated ring, which extends towards TM2 and makes contacts with the binding pocket side chains. Interestingly, anthraquinone-based analogues of Reactive Blue 2 dye have been found that inhibit distantly related purinergic P2Y receptors⁴⁶, but no AR activity for this chemotype had previously been reported to the best of our knowledge.

Cluster D comprises a relatively diverse set of compounds with a common thieno[2,3-b]pyridin-3-amine core motif and a variety of amide or ketone substituents in position 2; some of the highest affinity compounds in this cluster (6 (K_i=0.3 µM in Sf9 ) ,14,51,10) are presented in Table 1. In addition to the common set of core interactions, cluster D compounds have aromatic groups predicted to extend deep into the binding pocket, though too bulky or highly polar extensions apparently compromise binding (e.g. compound 13 in Table SI1). Functional assays for compound 6, however, suggest that while this compound blocks cAMP signaling, the effect does not depend on agonist concentration. Additional experiments (results not shown) confirm that compound 6 might inhibit cAMP production through an A_2AAR–independent mechanism.

Cluster E includes 2,3-diaminonaphthalene-1,4-dione compounds 50 and 17 with K_i < 1 µM. Both compounds have an aryl substitution at one of the amines, which is predicted to extend towards extracellular loops. An interesting feature suggested by docking for both compounds is a polar interaction of both unsubstituted and substituted amine groups with the Glu169(5.30) side chain, as shown in Figure 3E. Both compounds show significant activity in functional assays. Interestingly, compounds 50 and 17 are the only compounds in the set with a pronounced A_2A subtype selectivity (>10 fold). Given their small size and very good ligand efficiency LE~0.4 kcal/mol per heavy atom, they could provide reasonable leads for optimization of A2A selectivity.

Cluster F is represented in our set by only one compound 7 with K_i =1.9 µM (Table SI1). It has a pattern of interaction somewhat similar to cluster D, though its tricyclic system has a different orientation and extends more towards TM2 helix.

Cluster G, which has a 1,3,5-triazine-2,4-diamine moiety similar to cluster A, lacks the methyl-benzofuran extension and is likely to have a different binding mode with the phenylamine ring oriented deep into the binding pocket. Seven out of the 9 tested compounds of this cluster showed some A_2AAR affinity, though only two had K_i under 10 µM (compounds 36 and 31 in Table SI1).

Cluster H includes two compounds (16 and 46 in Table SI1) based on different chemical scaffolds sharing a similar core polar interaction motif, i.e. amine donor and carboxyl acceptor connected to a 5-membered ring. Though the compounds have only modest affinity, their small size leaves room for optimization (ligand efficiency LE=0.38 and 0.32 kcal/mol per heavy atom respectively).

Cluster I has only one representative compound 21, based on 4-amino-2-isoindole-1,3-dione scaffold, shown in Table 1 . Though affinity of this small compound is modest, it has a very high ligand efficiency (LE=0.38 kcal/mol per heavy atom) and may represent an attractive lead for optimization.

Other clusters (J to X) of the set do not have compounds with affinities under 10 µM, though some of them, e.g. compound 57 in cluster P with LE = 0.32 kcal/mol per heavy atom, may still present interesting leads. No active compounds were identified in clusters Q, R, T, U, V, W and X.

Although highly diverse, all high scoring VLS compounds used in experimental testing contained an exocyclic amine group. The preference for compounds with exocyclic amines is also noticeable in the evaluation benchmark with the crystal structure (3EML_W0) or optimized models (3EML_{W3_opt}) where other A_2AAR ligand types all scored below the default binding score threshold of −32 kJ/mol, with only a few xanthine analogues approaching this threshold. The docking preference for exocyclic amine compounds apparently reflects their exclusively high affinity to the A_2AAR (the best compounds have Ki < 0.1 nM)¹². At the same time, it may be at least partially attributed to some bias in the crystal structure of the A_2AAR complex with ligand 1. For example, the flexible side chain of Glu169(5.30) in the EL2 is perfectly positioned to form a strong polar interaction with the unsubstituted exocyclic amine of compound 1. As we have shown previously, binding of other ligand classes, e.g. xanthine analogues lacking an exocyclic amine may be somewhat suboptimal in this conformation on the A_2AAR⁴⁷ and thus may require a modified model or multiple models to capture the associated pocket plasticity ⁴⁸.

Most of the novel A_2AAR antagonists tested for subtype selectivity are characterized by reduced binding to A₃AR subtype (see Figure SI2). Such A_2A/A₃ selectivity is expected for chemotypes with unsubstituted exocyclic amine forming a hydrogen bond interaction network with Asn253(6.55) and Glu169(5.30) side chains in the A_2AAR, because A₃ has a hydrophobic valine in the 5.30 position. No substantial A_2A/A₁ selectivity was observed for most compounds in the set, with only two compounds 17 and 50 having 10-fold selectivity to the A_2AAR, and only one compound 8 with about 20-fold selectivity to A₁ subtype. . At the same time, most of the newly discovered scaffolds boast compounds with very high ligand efficiency LE > 0.30 kcal/mol per heavy atom. Predicted models of ligand receptor interactions can suggest specific functional groups exposed to the extracellular opening of the binding pocket, and provide an accurate structural template for rational optimization of selectivity and chemical properties for these ligands.

Conclusions

Our results present one of the first examples of virtual ligand screening, which employs a high resolution crystal structure of a GPCR target. We show that for adenosine A_2A receptor structure, which has a deep and well-defined binding pocket, virtual ligand screening can predict small molecule binders with success rates of as high as 41%. The 23 ligands selected by VLS from a 4 million compound library and confirmed active by A_2AAR binding assays represent at least nine novel chemotypes, supporting high diversity of the hits generated through structure-based virtual screening. Two of the identified compounds had K_i values of 0.06 µM and below, and 11 had submicromolar binding affinity. As expected in screening with antagonist-bound model of the receptor, most ligands (10 out of 13 tested) were proved A_2AAR antagonists in functional assays, while none showed agonistic activity. The virtual screening in this study did not explicitly target subtype selectivity, yielding hits that bind equally well to both A2A and A1 adenosine receptor subtypes. Most hit compounds, however are relatively small (<400 Da) and have high ligand binding efficiency (LE>0.3 kcal/mol per heavy atom), suitable for optimization of their subtype selectivity, affinity and chemical properties. Overall, the results demonstrate that high resolution structures of GPCR, combined with accurate docking and virtual ligand screening methods provide a highly efficient tool for identification of new GPCR antagonist chemotypes as lead candidates for drug discovery.

Experimental Section

Computational Methods

Generation of all-atom A_2AAR models

The initial all-atom model of the adenosine A_2A receptor was prepared from the PDB coordinates (PDB ID: 3EML)³⁵ using molecular conversion procedure⁴⁹, implemented in ICM molecular modeling software (Molsoft, LLC). The procedure includes the addition of hydrogen atoms to the PDB receptor structure, selection of the energetically favorable conformations of His, Asn and Gln side chains, and local minimization of polar hydrogens in the internal coordinates space.

For each of the 13 water configuration models tested, hydrogen atoms of the structured water molecules were co-optimized together with A_2AAR hydrogens, while coordinates of water oxygen atoms were retained as in the crystal structure. The 3EML_W1 model included only one water molecule, wa, with the lowest B-factor = 39 (as in the PDB entry), the 3EML_W2 model comprised wa and wa14 (B-factor = 46), the 3EML_W3 model had wa, wa14 and wa5 (B-factor=62). Other water configurations were obtained by adding water molecules in the binding pocket in the order of their increasing B-factor.

Virtual ligand libraries for optimization and validation of A_2AAR models

A diverse set of 23 known high affinity A_2AAR antagonists was compiled from 8 A_2AAR clinical candidates listed in Table 1 of ref.12, and additional 15 compounds in Figure 2 of ref.28. The all-atom molecular models of antagonists were built from their published 2-D structures. A library of 2000 decoy compounds was randomly selected from Chemdiv discovery collection of drug-like compounds (www.chemdiv.com).

Ligand docking and small scale VLS benchmark

To use ICM fast docking and VLS procedures, the receptor all-atom models were converted into energy potential maps calculated on a fine 3D grid (0.5 Å cell)⁴⁹. The grid potential maps account for van der Waals, hydrogen-bonding, hydrophobic and electrostatic interactions between ligand and receptor^50–51. The ligand was represented by an all-atom model and considered fully flexible in the potential field of the receptor. The ligand and decoy compounds were automatically docked into the A_A2AR models using the Biased Probability Monte Carlo (BPMC) global energy minimization procedure⁵² and sorted according to their ICM binding scores. To ensure convergence of the Monte Carlo optimization, three independent runs of the docking procedure were performed, and the best scoring pose per compound was kept. No distance restraints or any other experimentally derived information was used in the ligand docking procedure. The docking procedure takes about 30 seconds of Intel Xeon 2.8 Ghz CPU time per compound, and was performed using a 100 processor Linux cluster.

Metrics for VLS benchmark performance

Based on the values of ICM binding scores for the docked compounds, we used several complementary metrics to assess VLS performance⁵³. The Initial Enrichment Factor at 1% cutoff was calculated as:

$$\mathit{\text{EF}}(1\%)=(\mathit{\text{Lf}}/(\mathit{\text{Lf}}+\mathit{\text{Df}}))/(\mathit{\text{Lt}}/(\mathit{\text{Lt}}+\mathit{\text{Dt}})),$$

where Lt and Dt are the total number of known ligands and decoy compounds in the dataset respectively, while Lf and Df are the number of known ligands and decoys respectively in the top 1% scoring fraction the database. Receiver Operator Characteristic (ROC) curves were plotted with True Positive rate (TP = Lf/Lt) on the Y axis vs. False Positive rate (FP = Df/Dt) on the X axis for different fractions f of the dataset.

Area under ROC curve (AUC) was calculated along with Normalized Square root AUC (NSQ_AUC) ⁴⁷. For NSQ_AUC the area AUC* is calculated for the ROC curve plotted with X coordinate $X=\sqrt{\mathit{\text{FP}}}$. The NSQ_AUC value is then calculated as:

$$\mathit{\text{NSQ}}\_\mathit{\text{AUC}}=100(({\mathit{\text{AUC}}}^{*}-{\mathit{\text{AUC}}}_{\mathit{\text{random}}}^{*})/({\mathit{\text{AUC}}}_{\mathit{\text{perfect}}}^{*}-{\mathit{\text{AUC}}}_{\mathit{\text{random}}}^{*}).$$

The value of NSQ_AUC is more sensitive to initial enrichment than the commonly used linearAUC. The NSQ_AUC measure returns the value of 100 for any perfect separation of signal from noise and values close to 0 for a random subset of noise.

Additional ligand guided receptor optimization

A set of more than 200 conformational models with 3EML_W3 water configuration was generated by co-optimization of representative A_A2AR antagonists in the binding pocket of the receptor model. The optimization was performed using BPMC method^50,52, which allows extensive sampling of the flexible ligand and flexible receptor side chain conformations. The protein backbone and water oxygen atom coordinates were fixed to that of the crystal structure (PBD:3EML). For each of the 200 resulting conformational models, a small-scale VLS benchmarking was performed as described above and the model with the best NSQ_AUC value, 3EML_{W3_opt} was selected as an optimal model for VLS screening.

VLS of available compound libraries and selection of candidate A_2AAR ligands

Virtual screening of a large library of available compounds was performed using the same ICM VLS procedure as in the small-scale benchmark above. The screening library was prepared from the Molsoft ScreenPub database of 4.3 million unique drug-like compounds available from more than 40 vendors by automatically removing compounds with reactive groups and molecular weight exceeding 500 Da. Based on docking results, compounds with an ICM binding score better than −32 kJ/mol were selected and clustered by chemical similarity with 0.3 Tanimoto distance cutoff. Predicted values of Ligand Efficiency were calculated as a ratio of ICM binding score to the number of heavy atoms (LE_pred = −Score/N_HA). Values of logP (octanol/water partition) and logS (solubility) were also predicted using algorithms implemented in the ICM software. In each chemical cluster, from 1 to 10 compounds were selected for experimental assays according to their binding score, LE_pred values, chemical properties and on-shelf availability from vendors, resulting in 56 compounds purchased for the binding assays.

In Vitro and In Vivo Assays

The compounds selected by virtual screening were purchased from available screening collections of four vendors, Chembridge (www.Hit2Lead.com), Chemdiv (www.chemdiv.com), Enamine (www.enamine.com) and Sigma-Aldrich (www.Sigmaaldrich.com). Purity of compounds was equal to or greater than 96% as verified by liquid chromatography (HPLC) experiments performed by the vendors.

Binding assays in Sf9 membranes

Membranes highly expressing human A_2AAR receptor were produced from the Sf9-baculoviral expression system as previously described. Frozen aliquots of cells were thawed and then resuspended in homogenization buffer (50 mM Hepes, pH 7.4). The cells were subjected to homogenization using a Nitrogen Cavitation Pump (30 min, 800 psi) following 30 strokes with a Dounce homogenizer. The cell debris and nucleolus were removed by centrifugation at 900 × g for 10 min, followed by centrifugation at 100,000 × g for 45 min to isolate the raw membrane fraction. The resulting membrane pellet was resuspended in the buffer containing 50 mM Hepes, pH 7.4, 800 mM NaCl, and the protein concentration was assayed using the BCA protein assay kit from Pierce (Rockford, IL, USA) using Bovine serum albumin (BSA) as a standard for the protein assay. The centrifugation-resuspension cycle was repeated until supernatant fraction did not contain any trance of soluble protein based on the BCA protein assay (typically six cycles). Prior to the ligand binding assays, the membrane pellets were resuspended in ligand binding buffer (TME: 50 mM Tris–HCl, 10 mM MgCl₂, 0.5 mM EDTA, pH 7.4). The samples were tested for binding with tritiated compound 1 ([³H]ZM241385) (27.4 Ci/mmol) obtained from ARC Inc., St. Louis, USA. Crude plasma membranes (0.2–1 µg of total protein per reaction) were incubated for 30 min at room temperature with serial dilutions of the radioligand (0.05 – 10 nM). Incubations were rapidly terminated by filtration using a Tomtec Mach III cell harvester (Tomtec) through a 96-well GF/B filter plate (MultiScreen Harvest plate, Millipore Corp.), and rinsed five times with 500 µl of ice-cold buffer (50 mM Tris–HCl, pH 7.4). The harvest plates were dried and 30 µl of OptiPhase - HiSafe III scintillation liquid (Perkin-Elmer Life Sciences) was added. The bound radioactivity was measured using a Perkin Elmer Wallac Jet 1450 Microbeta scintillation counter. Nonspecific binding was determined in parallel reactions in the presence of an excess of compound 2 (100 µM, Sigma–Aldrich, USA), and specific binding was defined as the difference between total and nonspecific binding. All incubations were performed in triplicate, and independent experiments were repeated at least twice. Equilibrium dissociation constants (K_i) and maximal receptor levels (Bmax) were calculated from the results of saturation experiments using GraphPad Prism version 5 Software. For competition binding studies, the crude membranes were resuspended in ice-cold binding buffer (TME: 50 mM Tris-HCl, 10 mM MgCl₂, 1 mM EDTA, pH 7.4), containing protease inhibitors (Complete protease inhibitor cocktail tablet, Roche Applied Science) and homogenized for 30 strokes with a Dounce homogenizer. Test ligand stock solutions were prepared in DMSO at a final concentration of 10 mM. Some of the test ligands have strong dye property that might cause problem in the assays at the high concentration of the compound, and therefore we tested ligands using full completion range. Identical aliquots of crude plasma membranes (0.2–5 µg of total protein per reaction) containing 10–14 different concentrations of the competing unlabelled ligands with 20 nM of tritiated compound 1 in 0.1 mL total volume. The reaction components were added in order of crude membrane, test ligand and radioligand. The final DMSO concentration was 10%. Reference compounds 1, 2 and 3 were tested with and without 10% DMSO in the reactions. After 60 min at room temperature with shaking, reactions were rapidly terminated by filtration and counted as mentioned above. Initial testing was preformed with in triplicate, and was verified with six replicates. Individual assay results are represented.

Experimental ligand efficiency⁴⁵ (LE) was defined as LE = ΔG/N_HA , where ΔG was the free energy of ligand binding (ΔG = −RT ln K_i), and N_HA is the number of heavy atoms in the compound.

Binding and selectivity assays in mammalian membranes

[³H]DPCPX and [¹²⁵I]AB-MECA were purchased from Amersham Biosciences (NL), tritiated compound 1 was obtained from Tocris Cookson, Ltd. (UK). CHO cells expressing the human adenosine A₁ receptor were provided by Dr. Andrea Townsend-Nicholson, University College London. HEK293 cells stably expressing the human adenosine A_2A receptor and CHO cells expressing the human adenosine A₃ receptor were kind gifts from Dr. J. Wang (Biogen/IDEC, Cambridge, MA) Dr. K. N. Klotz (University of Wurzburg, Germany), respectively.

All compounds were tested in radioligand binding assays to determine their affinities at human adenosine A₁ ([³H]DPCPX), A_2A (tritiated compound 1), and A₃ ([¹²⁵I]AB-MECA) receptors as described previously in literature⁵⁴ with the exception that nonspecific binding to the A_2A receptor was determined in the presence of 10 µM concentration of 3a instead of 100 µM CPA and non-specific binding to the A₃ receptor was determined by the addition of 100 µM compound 3. The incubation was terminated by filtration over Whatman GF/C filters under reduced pressure with a Brandel harvester (Gaithersburg, MD). Filters were washed three times with ice cold 50 mM Tris/HCl pH 7.4, placed in vials and counted.

Functional cAMP Accumulation Assays

HEK293 cells expressing the human adenosine A_2A receptor were grown as a monolayer on 10 cm culture plates. For cAMP production and determination, 7500 cells/well were used on 384-well plates. The cells were incubated for 45 min at room temperature with or without test compounds at a 10 nM concentration of the agonist 3a (approximately an EC₂₀ concentration), and with or without test compounds with a saturating concentration of 3a (10 µM). The assay medium contained cilostamide (50 µM), rolipram (50 µM), and adenosine deaminase (0.8 IU/mL). Incubation was stopped with detection mix and antibody solution was added, these two steps according to the instructions of the supplier. The assay was performed with the Lance cAMP 384 kit from Perkin-Elmer based on the competition of the sample’s cAMP with a europium-labeled cAMP tracer complex for binding sites on cAMP-specific antibodies labeled with Alexa Fluor dye.

Abbreviations

GPCR

G protein-coupled receptor

AR

adenosine receptor

A_2AAR

subtype A_2A

β₂AR

β₂ adrenergic receptor

TM

transmembrane

EL2

extracellular loop 2

VLS

virtual ligands screening

PDB

protein databank

RMSD

root mean square deviation

LE

ligand efficiency

ROC

receiver operating characteristic

AUC

area under ROC curve

NSQ_AUC

normalized square root AUC

BPMC

biased probability Monte Carlo method