Pharmacological characterisation of novel adenosine A3 receptor antagonists

The adenosine A3 receptor (A3R) belongs to a family of four adenosine receptor (AR) subtypes which all play distinct roles throughout the body. A3R antagonists have been described as potential treatments for numerous diseases including asthma. Given the similarity between (adenosine receptors) orthosteric binding sites, obtaining highly selective antagonists is a challenging but critical task. Here we screen 39 potential A3R, antagonists using agonist-induced inhibition of cAMP. Positive hits were assessed for AR subtype selectivity through cAMP accumulation assays. The antagonist affinity was determined using Schild analysis (pA2 values) and fluorescent ligand binding. Structure–activity relationship investigations revealed that loss of the 3-(dichlorophenyl)-isoxazolyl moiety or the aromatic nitrogen heterocycle with nitrogen at α-position to the carbon of carboximidamide group significantly attenuated K18 antagonistic potency. Mutagenic studies supported by molecular dynamic simulations combined with Molecular Mechanics—Poisson Boltzmann Surface Area calculations identified the residues important for binding in the A3R orthosteric site. We demonstrate that K18, which contains a 3-(dichlorophenyl)-isoxazole group connected through carbonyloxycarboximidamide fragment with a 1,3-thiazole ring, is a specific A3R (< 1 µM) competitive antagonist. Finally, we introduce a model that enables estimates of the equilibrium binding affinity for rapidly disassociating compounds from real-time fluorescent ligand-binding studies. These results demonstrate the pharmacological characterisation of a selective competitive A3R antagonist and the description of its orthosteric binding mode. Our findings may provide new insights for drug discovery.

Co-stimulation with 10 μM of both forskolin and NECA reduced the cAMP accumulation when compared to 10 μM forskolin alone and this was reversed with the known A 3 R antagonist MRS 1220 (Table 1 and Supplementary Fig. 1). Compounds K1, K10, K11, K17, K18, K20, K23, K25 and K32 were identified as potential antagonists at the A 3 R through their ability to elevate cAMP accumulation when compared to forskolin and NECA co-stimulation. Of the nine potential A 3 R antagonists, eight (excluding K11) appeared to be antagonists at the tested concentration of 1 μM (Supplementary Fig. 2

and Supplementary Table 2).
A number of compounds previously documented (K5, K9, K21, K22 and K24; 19 or determined in this study (K26, K27 and K34) to have sub-micromolar binding affinities for A 3 R showed no activity in our cAMP-based screen ( Table 1, Supplementary Table 1). To ensure robustness of our functional screen, full inhibition curves of NECA in the presence or absence of tested compounds (1 μM or 10 μM) were constructed in A 3 R Flp-In CHO cells ( Supplementary Fig. 3, Supplementary Table 3). In this preliminary data all nine compounds (K5, K9, K11, K21, K22, K24, K26, K27 and K34) appeared to reduce the NECA potency at the highest tested concentration (10 μM) but showed no effect at 1 μM and thus appear to be low potency antagonists at the A 3 R.
AR subtype selectivity and specificity. The Table 1. Mean cAMP accumulation as measured in Flp-In CHO cells stably expressing A3R following stimulation with 10 μM forskolin only (DMSO) or 10 μM forskolin, NECA at the predetermined IC80 concentration and 1 μM test compound/MRS 1220/DMSO control. Binding affinities were obtained through radioligand binding assays against the A 1 R, A 2A R and A 3 R. a cAMP accumulation mean ± SEM expressed as %10 μM forskolin response where n = 3 independent experimental repeats, conducted in duplicate. Potential antagonists were selected for further investigation based on a high mean cAMP accumulation (> 80%). b Difference between the mean cAMP accumulation between 'NECA' and each compound expressed as %10 μM forskolin response. c Binding affinity measured in three independent experiments and where indicated, previously published in Lagarias et al. 19 . Bold denotes binding affinity < 10 μM. 1 Indicates previously published in Lagarias et al. 19 .

Figure 1. Characterisation of A 3 R antagonist at all AR subtypes. A 3 R Flp-In CHO cells or CHO-K1 cells (2000 cells/well
) stably expressing one of the remaining AR subtypes were exposed to forskolin in the case of G icoupled A 1 R and A 3 R (1 μM or 10 μM, respectively) or DMSO control in the case of G s -coupled A 2A R and A 2B R, NECA and test compound (10 μM) for 30 min and cAMP accumulation detected. All values are mean ± SEM expressed as percentage forskolin inhibition (A 1 R and A 3 R) or stimulation (A 2A R and A 2B R), relative to NECA. n ≥ 3 independent experimental repeats, conducted in duplicate. www.nature.com/scientificreports/ selectivity findings agree with our previously published radioligand binding data (Lagarias et al. 19 ) and are summarised in Table 2.
In addition, we wanted to determine if K18 could also antagonise the activity of the A 3 R when an alternative downstream signalling component was measured; ERK1/2 phosphorylation (Fig. 3). In line with previously reported findings 11,20 , agonists at the A 3 R increased ERK1/2 phosphorylation after 5 min, with IB-MECA tenfold more potent than NECA (Supplementary Fig. 9) and preliminary data suggests this was entirely G i/o -mediated (pERK1/2 levels were abolished upon addition of PTX). K18 was able to antagonise A 3 R-mediated phosphorylation of ERK1/2 with the antagonist potency (pA 2 values) not significantly different compared to the cAMPinhibition assay (Fig. 3C).
A 3 R constitutive activity and inverse agonism. We next determined if any of our competitive antagonist could function as inverse agonists of the A 3 R. In our hands, the A 3 R, when expressed in Flp-In-CHO cells, displays constitutive activity ( Supplementary Fig. 7). All eight characterised A 3 R antagonists showed a concentration-dependent inverse agonism of the A 3 R when compared to DMSO control (Fig. 2) and were determined to be A 3 R dependent ( Supplementary Fig. 8). This was also found to be the case for K20 and K23 at the A 1 R ( Supplementary Fig. 5). Notably, DMSO showed a concentration-dependent elevation in cAMP accumulation above that of forskolin alone.
Evaluation of the binding mode of K18 at A 3 R. We have previously investigated the binding mode of K18 at the human A 3 R using a homology model of the human A 3 R and simulations 21 . The computational model showed that the 3-(dichlorophenyl) group is positioned close to V169 5.30 , M177 5.40 , I249 6.54 and L264 7.34 of the A 3 R orthosteric binding site forming attractive vdW interactions. The amino group of the ligand is hydrogenbonded to the amido group of N250 6.55 . The isoxazole ring is engaged in an aromatic π-π stacking interaction with the phenyl group of F168 5.29 (Fig. 4). The thiazole ring is oriented deeper into the receptor favouring interactions with L246 6.51 , L90 3.32 and I268 7.39 (for full details see 19 ).
Of the identified residues predicted to mediate an interaction between K18 and the A 3 R, the ones which showed -according to the molecular dynamic (MD) simulations-the most frequent and the most important contacts were chosen for investigation and included amino acids L90 3.32 , F168 5.29 , V169 5.30 , M177 5.40 , L246 6.51 , I249 6.54 , N250 6.55 , L264 7.34 and I268 7.39 (Fig. 4). Site-directed mutagenesis was performed replacing each residue with an alanine in the A 3 R and expressed in the Flp-In-CHO™ cells lines. Each mutant was then screened for their ability to suppress forskolin-induced cAMP accumulation in response to NECA/IB MECA stimulation in the presence and absence of K18 and the detailed results are presented in Table 5 and Fig. 5. Mutation of residues F168 5.29 , L246 6.51 , N250 6.55 and I268 7.39 abolished agonist induced suppression of forskolin-induced cAMP accumulation and were discontinued in this study 22 . The affinity (pA 2 ) of K18 at the WT and mutant A 3 R were compared in order to determine the potential antagonist binding site (Fig. 5, Table 5). As described previously 21 , K18 displayed no difference in affinity at I249A 6.54 (compared to WT), whereas M177A 5.40 and V169A 5.30 were significantly smaller. Interestingly, we found an increase in the affinity for L90A 3.32 and L264A 7.34 when compared to WT. As would be expected, the K18 affinity at the A 3 R mutants was not different between agonists, confirming agonist independence ( Supplementary Fig. 11). These experimental findings are reflected in our predicted binding pose of K18 at the WT A 3 R (Fig. 4) 21 .
We further investigated the suggested binding mode through comparatively modelled K5 and K17 binding at the A 3 R given their structural similarity with K18; K17 and K18 possess one and two chlorine atoms attached to the phenyl ring of the 3-phenyl-isoxazole moiety, respectively, whereas K5 has none (Fig. 4B,C). The suggested binding mode of K18 was further verified since the MM-PBSA calculated rankings in binding free energies were in agreement with experimental differences in binding affinities using radio-labelled assays and BRET (K5 < K17 < K18 < MRS 1220- Table 6) and antagonistic potencies with 5 having no chlorine atoms in 3-phenylisoxazolyl lacking any antagonistic potency. Finally, following MD simulations of compounds K26, K27, K29-K34 and K36-K39, we observed that K26 (and K34 for that matter) displayed a similar binding pose ( Supplementary  Fig. 10) to that of K18 (Fig. 4). K26, K34, K5, K17 and K18 have all similar binding affinities, measured by radiolabelled assays (Sup. Table 1). K26 and K34 contain the o-diphenyl-carbonyl instead of the 3-phenyl-isoxazole moiety in K5 but all three functionally showed weak antagonistic potency (> 1 μM, Supplementary Fig. 3), in contrast to K18 or K17 which contain a 3-(chlorophenyl)-isoxazolyl moiety or 3-(dichlorophenyl)-isoxazolyl moiety,  www.nature.com/scientificreports/ respectively. These findings suggest that a more complex binding mode is present which will be investigated in the future, through synthetic installation of chlorine atoms in the ended phenyl ring of o-diphenyl-carbonyl moiety.

Kinetics of A 3 R antagonists determined through BRET. NanoBRET techniques have been success-
fully used to determine the real-time kinetics of ligand binding to GPCRs [23][24][25] . Here, we investigated the ability of the selective A 3 R antagonists MRS 1220, K17 or K18 to inhibit specific binding of the fluorescent A 3 R antagonist CA200645 to Nluc-A 3 R 26,27 . The kinetic parameters for CA200645 at Nluc-A 3 R were initially determined as K on (k 1 ) = 2.86 ± 0.89 × 10 7 M −1 min −1 , K off (k 2 ) = 0.4397 ± 0.014 min −1 with a K D of 17.92 ± 4.45 nM. (Supplementary Fig. 12). Our MRS 1220 kinetic data was fitted with the original 'kinetic of competitive binding' model 28 (built into GraphPad Prism 8.0) with a determined K on (k 3 ) and K off (k 4 ) rate of 3.25 ± 0.28 × 10 8 M −1 min −1 and 0.0248 ± 0.005 min −1 , respectively. This gave a residence time (RT) (RT = 1/K off ) of 40.32 min. It was noticed in the analysis for K5, K17 and K18 that the fit in some cases was ambiguous and/or the fitted value of the compound dissociation rate constant was high (k 4 > 1 min −1 , corresponding to a dissociation t 1/2 of < 42 s). In order to determine the reliability of the fitted k 4 value, data were also analysed using an equation that assumes compound dissociation is too rapid for the dissociation rate constant to be determined reliably and the fits to the two equations were compared ("Kinetics of competitive binding, rapid competitor dissociation", derived in the Appendix I). This model allowed estimate of the equilibrium binding affinity of the compound (K i ) but not the binding kinetics of K5, K17 and K18 ( Fig. 6 and Table 6). These affinity values were in agreement with those obtained via Schild analysis (except for K5) and were approximately tenfold higher than those determined through radioligand binding (Table 6). Notably, the order of affinity (K5 < K17 < K18) was consistent.

Assessing the species selectivity of A 3 R antagonists.
There remains an issue as to understanding the precise role of the A 3 R. This is, in part due, to a lack of appropriate antagonists that are able to be tested in classical animal models such as rodent since only a few molecules (including the 6-phenyl-1,4-dihydropyridine MRS 1191 and the triazoloquinazoline MRS 1220) display cross species (human and rat) reactivity 29 . Comparison of residues of the binding area between human and rat A 3 R show that they differ in residues 167 5.28 , 169 5.30 , 176 5.37 , Table 4. Binding of compounds to the rat A 3 R. Equilibrium dissociation constant of MRS 1220 and K compounds as determined through NanoBRET ligand-binding (pK i ).    (Fig. 7). The scarcity of rat A 3 R antagonists prompted us to investigate if our characterised compounds were also potential antagonists at the rat A 3 R. Using a Nluc-tagged rat A 3 R expressing HEK 293 cell line, we conducted NanoBRET ligand-binding experiments whereby we determined the ability of our compounds to inhibit specific binding of the fluorescent antagonist AV039 to Nluc-rat A 3 R. As expected, AV039 was displaced by increasing concentrations of MRS 1220 (pK i 6.788 ± 0.1) ( Fig. 7 and Table 4). We found very weak binding of K17, K18, K10 and K32, with no binding detected below the concentration of 10 μM, whereas K1, K20, K23 and K25 were determined as potential rat A 3 R antagonists (pK i 6.07 ± 0.04, 5.71 ± 0.03, 5.93 ± 0.04 and 6.37 ± 0.06, respectively) ( Fig. 7 and Table 4). K25 had a higher binding affinity for the rat A 3 R when compared to the human A 3 R (Table 1) (pKi 6.37 ± 0.1 and 5.81, respectively). MD simulations of the rat A 3 R (performed as described previously for the human A 3 R) suggested that the presence of M264 7.34 most likely hampers K18 binding due to steric hindrance of the dichloro-phenyl group (Fig. 7). In contrast, MD simulations of K25 against rat A 3 R showed the formation of stable complex (Fig. 7). Here, the 2-amido group of the thiophene ring is hydrogen-bonded to the amido group of N250 6.55 . The thiophene ring forms aromatic π-π stacking interaction with F168 5.29 and the 5-(p-chlorophenyl) is oriented deep in   (Fig. 7).
Pharmacokinetic assessments of K18. The metabolic in vitro t 1/2 (human liver microsomes, 0.1 mg/ mL) of K18 (0.1 μM) was determined (0-60 min) as 24 min and the intrinsic clearance (CL int ) calculated as 287.2 μl/min/mg ( Supplementary Fig. 13). This was comparable to the reference compound verapamil and terfenadine (0.1 μM) with t 1/2 determined as 35 and 12 min and CL int as 200.1 or 581.1 μl/min/mg, respectively. Human plasma stability assessment determined the percentage of K18 (1 μM) remaining after 120 min as 90%, with a t 1/2 of > 120 min. This is considerably higher than the reference compound propantheline (1 μM) which was determined to have a half-life of 55 min. The t 1/2 of K18 (1 μM) in PBS (pH 7.4) over 240 min was determined as > 240 min, with 87% remaining at 240 min and was comparable to the reference compound propantheline (1 μM), with a determined t 1/2 of > 240 min.

Discussion
The search for an AR subtype specific compound often leads to compounds active at more than one of the AR subtypes because of the broad and similar orthosteric binding site of ARs 30 . Given that AR subtypes play distinct roles throughout the body, obtaining highly specific receptor antagonists and agonists is crucial. The virtual    19 .
Here, we present the pharmacological characterisation of eight A 3 R antagonists identified though virtual screening. Of these eight compounds, K10, K17, K18, K20, K23, K25 and K32 were determined to be competitive. Whereas K20 and K23 are antagonists at both the A 1 R and A 3 R, K10, K17, K18, K25 and K32 were are A 3 R selective antagonists. Indeed, we found no functional activity, or indeed binding affinity (< 30 μM), at the other AR subtypes. K1, K20 and K23 showed weak antagonism of the A 2A R with no activity at the A 2B R (Fig. 1, Table 2). These selectivity findings are in agreement with our radioligand binding data (Supplementary Table 1, and Lagarias et al. 19 for K1-25, K28 and K35). However, a number of compounds previously determined to have micromolar binding affinity for A 3 R (K5, K9, K21, K22, K24, K26, K27 and K34), show no antagonistic potency in our initial functional screen. Further testing confirmed that these compounds are low potency antagonists and, although supporting the previously published radioligand binding data, confirmed the need for functional testing: not all compounds with binding affinity show high functional potency.
We show the A 3 R, when expressed in Flp-In-CHO cells, displays constitutive activity. Compounds which preferably bind to the inactive (R) state, decreasing the level of constitutive activity 33 and in the case of a G i/o -coupled GPCR leading to an elevated cAMP, are referred to as inverse agonists. In this study all the A 3 R antagonists identified and both A 1 R antagonists (K20 and K23) were found to act as inverse agonists. We also report an elevation in cAMP accumulation when cells were stimulated with DMSO, which is concentration-dependent. Given that even low concentrations of DMSO have been reported to interfere with important cellular processes 34 , the interpretation of these data should be made with caution.
We show that the presence of a chloro substituent in the phenyl ring of 3-phenyl-isoxazole favoured A 3 R affinity, as following 0Cl < 1Cl < 2Cl i.e. K5 < K17 < K18. This theory is supported by both radioligand binding, NanoBRET ligand-binding and functional data. Moreover, MD simulations show that these compounds adopt a similar binding mode at the A 3 R orthosteric binding site, but the free-energy MM-PBSA calculations show that K18, having two chlorine atoms and an increased lipophilicity, leaves the solution state more efficiently and enters the highly lipophilic binding area. Importantly, substitution of the 1,3-thiazole ring in K17 with either a 2-pyridinyl ring (K32) or a 3-pyridinyl ring (K10) but not a 4-pyridinyl ring (K11) maintains A 3 R antagonistic potency. Also, substitution of 1,3-thiazole ring in K17 with the p-iodo-phenyl in K9 loses the antagonistic potency. Although we have not directly determined the effects of similar pyridinyl ring substitutions on the higher affinity antagonist K18, we suspect there would be no significant increase in the potency of K18 given the small changes we observed for K17.
For the first time, we demonstrated the utilisation of a new model which expands on the 'Kinetic of competitive binding' model 28 (built into Prism 8.0) for fitting fast kinetics data obtained from NanoBRET experiments and assumes the unlabelled ligand rapidly equilibrates with the free receptor. Very rapid competitor dissociation can lead to failure of the fit, eliciting either an ambiguous fit (regression with Prism 8: "Ambiguous", 2019) or unrealistically large K 3 and K 4 values. Whereas we were able to successfully fit the MRS 1220 kinetic data with the Motulsky and Mahan model due to its slow dissociation, fitting of K5, K17 and K18 kinetic data with this model often resulting in an ambiguous fit. Our new model, assuming fast compound dissociation, successfully fits the data and allows the determination of binding affinity. In the cases where the data were able to fit the Motulsky and Mahan model, the dissociation constant was higher (of the order of 1 min −1 ), indicating rapid dissociation. Although we found nearly a tenfold difference in determined binding affinity for MRS 1220, K5, K17 and K18 Table 6. Binding of K5, K17, K18 and MRS 1220 to the A 3 R orthosteric binding area. Effective binding energies (ΔG eff ) and energy components (E vdW , E EL , ΔG solv ) in kcal mol −1 calculated using the MM-PBSA method. a vdW energy of binding calculated using molecular mechanics. b Electrostatic energy of binding calculated using molecular mechanics. c Difference in solvation energy between the complex, the protein and the ligand, i.e. G complex, solv -(G protein, solv + G ligand, solv ). d Effective binding free energy calculated as ΔG eff = ΔE ΜΜ + ΔG sol ; in Table 6, ΔE ΜΜ = Ε vdW + E EL (see "Materials and methods"). e Equilibrium dissociation constant of MRS 1220, K5, K17 and K18 as determined through three independent experimental approaches: Schild analysis (pK B ), NanoBRET (pK i ) or radioligand binding (pK i ). f pK B obtained through Schild analysis in A 3 R stably expressing Flp-In CHO cells. g pK i (mean ± SEM) obtained in NanoBRET binding assays using Nluc-A 3 R stably expressing HEK 293 cells and determined through fitting our "Kinetics of competitive binding, rapid competitor dissociation" model or in the case of MRS 1220 through fitting with the 'Kinetics of competitive binding' model with a determined K on (k 3 ) and K off (k 4 ) rate of 3.25 ± 0.28 × 10 8  www.nature.com/scientificreports/ between BRET ligand binding and radioligand binding assays, we demonstrated the order of affinity remains consistent. Indeed, this is seen across all three experimental approaches: Schild analysis, NanoBRET ligandbinding assay and radioligand binding. Combining MD simulations with mutagenesis data, we presented a final binding pose of K18 which appears to be within the orthosteric binding site, involving residues previously described to be involved in binding of A 3 R compounds 35 . We previously reported 22 no detectable G i/o response following co-stimulation with forskolin and NECA or IB-MECA for A 3 R mutants F168A 5.29 , L246A 6.51 , N250A 6.55 and I268A 7.39 and our findings are in line with previous mutagenesis studies investigating residues important for agonist and antagonist binding at the human A 3 R 36,37 . Through performing Schild analysis (results of which were used to inform modelling in Lagarias et al. 21 ) we experimentally determined the effect of receptor mutation on antagonist affinity for L90A 3.32 , V169A/E 5.30 , M177A 5.40 , I249A 6.54 and L264A 7.34 A 3 R. The pA 2 value for I249A 6.54 A 3 R is similar to WT, whereas M177A 5.40 and V169A 5.30 are significantly smaller suggesting these residues appear to be involved in K18 binding. Interestingly we found an increase in K18 affinity at L90A 3.32 and L264A 7.34 when compared to WT. Our detailed MD simulations, results published elsewhere 21 have investigated the selectivity profile of K18 and have demonstrated that K18 failed to bind A 1 R and A 2A R due to a more polar area close to TM5, TM6 when compared to A 3 R.
We have also performed a preliminary pharmacokinetic assessment of K18 to assess its potential as a lead compound for future use in drug design. Based upon our initial findings K18 has a metabolic half-life and intrinsic clearance equivalent to verapamil and terfenadine. K18 is highly stable in human plasma. As such these For MRS 1220, this trace demonstrates a classic tracer 'overshoot' , as has been previously described observed when the unlabelled compound has a slower off rate than the labelled CA200645 (K off of 0.0248 ± 0.005 min −1 and 0.4397 ± 0.014 min −1 respectively) (Sykes et al. 24 , Motulsky and Mahan 28 ). The data shown are representative of three independent experimental repeats (mean ± SEM) fitted with the appropriate model, as determined by statistical comparison between our new model ("Kinetics of competitive binding, rapid competitor dissociation", derived in the Appendix I) (K5, K17 and K18) or the 'kinetic of competitive binding' model (built into Prism) for MRS 1220 (see "Materials and methods" for fitting procedure and statistical comparison method). (E) The resulting concentration dependent decrease in BRET ratio at 10 min was taken to calculate pK i through fitting the Cheng-Prusoff equation 59 . Each data point represents mean ± SEM of five experiments performed in duplicate.
Antagonists that are A 3 R-selective across species or at rat A 3 R alone are useful pharmacological tools to define the role of these receptors. The human and rat A 3 R display 72% homology 38 . The lack of rat A 3 R selective antagonists prompted us to investigate if our characterised A 3 R antagonists are potential antagonists at the rat A 3 R. We reported no binding of our lead A 3 R antagonist, K18, at the rat A 3 R and MD simulations suggest this is due to steric hinderance by M264 7.34 . This finding suggests that K18 may not only be A 3 R specific within the human ARs but may also be selective across species. Of the compounds that showed rat A 3 R binding (K1, K20, K23 and K25), K25 show the highest binding affinity and represents an interesting candidate for further investigation. MD simulations show K25 forms a stable complex with rat A 3 R and we suggest a potential binding pose.
In conclusion, we present findings of a unique scaffold (K18) which is both chemically and metabolically stable and as such can be used as a starting point for detailed structure-activity relationships and represents a useful tool that warrants further assessment. Furthermore, we introduce K25 as a potential rat A 3 R antagonist which also warrants further investigation.

Materials and methods
Cell culture and transfection. Cell lines were maintained using standard subculturing routines as guided by the European Collection of Cell Culture (ECACC) and checked annually for mycoplasma infection using an EZ-PCR mycoplasma test kit from Biological Industries (Kibbutz Beit-Haemek, Israel). All procedures were performed in a sterile tissue culture hood using aseptic technique and solutions used in the propagation of each cell line were sterile and pre-warmed to 37 °C. All cells were maintained at 37 °C with 5% CO 2 , in a humidified atmosphere. This study used CHO cell lines as a model due to the lack of endogenous AR subtype expression Stable Flp-In-CHO cell lines were generated through co-transfection of the pcDNA5/FRT expression vector (Thermo Fisher Scientific) containing the gene of interest and the Flp recombinase expressing plasmid, pOG44 (Thermo Fisher Scientific). Transfection of cells seeded in a T25 flask at a confluency of ≥ 80% was performed using TransIT-CHO Transfection Kit (MIR 2174, Mirus Bio), in accordance with the manufacturer's instructions.
Here, a total of 6 μg of DNA (receptor to pOG44 ratio of 1:9) was transfected per flask at a DNA:Mirus reagent ratio of 1:3 (w/v). 48 h post-transfection, selection using 600 μg/mL hygromycin B (Thermo Fisher Scientific) (concentration determined through preforming a kill curve) was performed for two days prior to transferring the cells into a fresh T25 flask. Stable Flp-In-CHO cell lines expressing the receptor of interest were selected using 600 μg/mL hygromycin B whereby the media was changed every two days. Successful mutant cell line generation for non-signalling mutants were confirmed by Zeocin sensitivity (100 μg/mL).
The Nluc-tagged human A 3 R expressing HEK 293 cell line along with the Nluc-tagged rat A 3 R pcDNA3.1 + construct for the generation of stable Nluc-tagged rat A 3 R expressing HEK 293 cells were kindly gifted to us by Stephen Hill and Stephen Briddon (University of Nottingham). HEK 293 cells in a single well of 6-well plate (confluency ≥ 80%) were transfected with 2 μg of DNA using polyethyleneimine (PEI, 1 mg/ml, MW = 25,000 g/mol) (Polysciences Inc) at a DNA:PEI ratio of 1:6 (w/v). Briefly, DNA and PEI were added to separate sterile tubes containing 150 mM sodium chloride (NaCl) (total volume 50 μl), allowed to incubate at room temperature for 5 min, mixing together and incubating for a further 10 min prior to adding the combined mix dropwise to the cells. 48 h post-transfection, stable Nluc-rat A 3 R expressing HEK 293 cell were selected using 600 μg/mL Geneticin (Thermo Fisher Scientific) whereby the media was changed every two days. HEK 293 cell lines were routinely cultured in DMEM/F-12 GlutaMAX (Thermo Fisher Scientific) supplemented with 10% FBS (F9665, Sigma-Aldrich).
Constructs. The human A 3 R originally in pcDNA3.1 + (ADRA3000000, cdna.org) was cloned into the pcDNA5/FRT expression vector and co-transfected with pOG44 to generate a stable Flp-In-CHO cell line. Mutations within the A 3 R were made using the QuikChange Lightening Site-Directed Mutagenesis Kit (Agilent Technologies) in accordance with the manufacturer's instructions. The Nluc-tagged rat A 3 R pcDNA3.1 + construct, used in the generation of the stable Nluc-tagged rat A 3 R expressing HEK 293 cell line was kindly gifted to us by Stephen Hill and Stephen Briddon (University of Nottingham). All oligonucleotides used for mutagenesis were designed using the online Agilent Genomics 'QuikChange Primer Design' tool (detailed in Stamatis et al. 22 (Table S4)) and purchased from Merck. All constructs were confirmed by in-house Sanger sequencing.
Compounds. Adenosine, NECA, IB-MECA, HEMADO, DPCPX (8-cyclopentyl-1,3-dipropyl-7H-purine-2,6-dione) and MRS 1220 were purchased from Sigma-Aldrich and dissolved in dimethyl-sulphoxide (DMSO). CA200645, a high affinity AR xanthine amine congener (XAC) derivative containing a polyamide linker connected to the BY630 fluorophore, was purchased from HelloBio (Bristol, UK) and dissolved in DMSO. AV039, a highly potent and selective fluorescent antagonist of the human A 3 R based on the 1,2,4-Triazolo[4,3-a]quinoxalin-1-one linked to BY630 39 , was kindly gifted to us by Stephen Hill and Stephen Briddon (University of Nottingham). PMA was purchased from Sigma-Aldrich. Compounds under investigation were purchased from e-molecules and dissolved in DMSO. The concentration of DMSO was maintained to < 1.5% across all experiments (1.26% for all cAMP assays, 1% for pERK1/2 assays and 1.02% or 1.1% for NanoBRET ligand-binding experiments using CA200645 or AV039, respectively). In order to allow the A 1 R/A 3 R mediated G i/o response to be determined, co-stimulation with forskolin, an activator of AC 40 , at the indicated concentration (depending on cell line) was performed. Testing of potential antagonists was performed in a competition experiment where cells received a co-stimulation with forskolin, agonist and compound/DMSO control, without test compound pre-incubation. cAMP levels were then determined using a LANCE cAMP kit as described previously 41 . In order to reduce evaporation of small volumes, the plate was sealed with a ThermalSeal film (Excel Scientific) at all stages.

Scientific Reports
Phospho-ERK assay. ERK1/2 phosphorylation was measured using the homogeneous time resolved fluo- Radioligand binding. All pharmacological methods followed the procedures as described in the literature 42 .
In brief, membranes for radioligand binding were prepared from CHO cells stably transfected with hAR subtypes in a two-step procedure. In the first step, cell fragments and nuclei were removed at 1000×g and then the crude membrane fraction was sedimented from the supernatant at 100,000×g. The membrane pellet was resuspended in the buffer used for the respective binding experiments and it was frozen in liquid nitrogen and stored at -80 °C. For radioligand binding at the A NanoBRET ligand-binding. Through the use of NanoBRET, real-time quantitative pharmacology of ligand-receptor interactions can be investigated in living cells. CA200645, acts as a fluorescent antagonist at both A 1 R and A 3 R with a slow off-rate 43 . Using an N-terminally NanoLuc (Nluc)-tagged A 3 R expressing cell line, competition binding assays were conducted. The kinetic data was fitted with the 'kinetic of competitive binding' model 28 (Motulsky and Mahan, 1984; built into Prism) to determine affinity (pK i ) values and the association rate constant (K on ) and dissociation rates (K off ) for unlabelled A 3 R antagonists. In several cases this model resulted in an ambiguous fit (Regression with Prism 8: "Ambiguous", 2019). We developed a new model which expands on the 'kinetic of competitive binding' model to accommodate very rapid competitor dissociation, assuming the unlabelled ligand rapidly equilibrates with the free receptor. This method allows determination of compound affinity (pK i ) from the kinetic data.
In order to identify if the characterised compounds also bound the rat A 3 R, we conducted competition binding assays using Nluc-tagged rat A 3 R expressing HEK 293 cells and the fluorescent compound AV039 39 rather than xanthine based CA200645, which have previously been reported as inactive at rat A 3 R 44 . For both human and rat A 3 R experiments, filtered light emission at 450 nm and > 610 nm (640-685 nm band pass filter) was measured using a Mithras LB 940 and the raw BRET ratio calculated by dividing the 610 nm emission with the 450 nm emission. Here, Nluc on the N-terminus of A 3 R acted as the BRET donor (luciferase oxidizing its substrate) and CA200645/AV039 acted as the fluorescent acceptor. CA200645 was used at 25 nM, as previously reported 26 and AV039 was used at 100 nM (pre-determined K D , 102 ± 7.59 nM). BRET was measured following the addition Figure 7. Pharmacological characterisation of K series of compounds at the rat A 3 R. (A) Comparison of the residues of the orthosteric binding area in human and rat A 3 Rs. (B) Saturation binding experiment with AV039 with a K D of 102 ± 7.59 nM. (C) Inhibition of BRET between Nluc and AV039 at the rat A 3 R by MRS 1220 and K compounds. HEK293 cells stably expressing Nluc-rat A 3 R were treated with 100 nM AV039 and increasing concentrations of unlabelled compound. The resulting concentration dependent decrease in BRET ratio at 5 minutes was taken to calculate pK i through fitting the Cheng-Prusoff equation 59 . Each data point represents mean ± SEM of n (n = 5 for MRS 1220, K1, K20, K23 and K25, n = 3 for K10, K17, K18 and K32) experiments, performed in duplicate. (D) Top and side (E) views of Rat A 3 R in complex with K18. Starting pose (carbons of the ligand in green), after 100 ns MD simulation (carbons of the ligand in orange). Light blue sticks show residues conserved with human A 3 R. M264 7.34 most likely hampers K18 binding due to steric hindrance of the dichloro-phenyl group. (F) Top and side views (G) of the average structure of rat A 3 R in complex with K25 from 100 ns MD simulations (carbons of the ligand are shown in orange sticks and light blue sticks show residues in contact with K25). K25 was docked into the orthosteric site of the rat A 3 R using the GoldScore scoring function and the highest scoring pose was inserted in a hydrated POPE bilayer. The complexes were subjected to MD simulations with Amber14ff. and K25 adopts a potential binding pose within the orthosteric binding area.

Scientific Reports
| (2020) 10:20781 | https://doi.org/10.1038/s41598-020-74521-y www.nature.com/scientificreports/ of the Nluc substrate, furimazine (0.1 μM). Nonspecific binding was determined using a high concentration of unlabelled antagonist, MRS 1220 at 10 nM or 10 μM, for human and rat A 3 R, respectively. Data were also analysed using an equation that assumes compound dissociation is too rapid for the dissociation rate constant to be determined reliably and the fits to the two equations compared ("Kinetics of competitive binding, rapid competitor dissociation", derived in the Appendix I, Supplementary material). This equation assumes rapid equilibration between compound and receptor and consequently provides an estimate of the equilibrium binding affinity of the compound (K i ) but not the binding kinetics of the compound. The equation is, where ρ I is fractional occupancy of receptors not bound by L:

Receptor binding kinetics data analysis.
and k obs,+ I is the observed association rate of tracer in the presence of competitor, defined as, The fits to the two equations were compared statistically using a partial F-test in Prism 8.
Pharmacokinetic assessments of K18. Preliminary pharmacokinetic assessments of K18 was outsourced to Eurofins Panlabs (Missouri, U.S.A) and including tests for intrinsic clearance (human liver microsomes), plasma (human) stability and half-life in PBS. These tests were conducted in duplicate using a single concentration of K18 (0.1 μM or 1 μM) using the substrate depletion method. Here, the percentage of K18 remaining at various incubation times was detected using high-performance liquid chromatography mass spectrometry (HPLC-MS). Reference compounds (verapamil, terfenadine and propantheline) were supplied and tested alongside K18. The half-life (t 1/2 ) was estimated from the slope (k) of percentage compound remaining (In(%K18 remaining)) versus time (t 1/2 =-In(2)/k), assuming first order kinetics. The intrinsic clearance (CL int, in μl/min/mg) was calculated according to the following formula: Data and statistical analysis. All in vitro assay data was analysed using Prism 8.0 (GraphPad software, San Diego, CA), with all dose-inhibition or response curves being fitted using a 3-parameter logistic equation to calculate response range or E max and IC/EC 50 . Experimental design ensured random distribution of treatments across 96/384-well plates to avoid systematic bias. Agonist stimulation alone was used as an intrinsic control across all experiments. Although initial screening of the 50 compounds was blinded, due to limitations in resources, it was not possible to perform all of our experiments in a blinded manner. Normalisation was used to control for unwanted sources of variation between experimental repeats.
In the rare cases where significant outliers were identified through the ROUT method (performed in Prism with Q set to 2% (defines the chance of falsely identifying one or more outliers)) (Statistics with Prism 7: "How to: www.nature.com/scientificreports/ Identify outliers", 2019), these were excluded from data analysis and presentation. Dose-inhibition/dose-response curves were normalised to forskolin, expressed as percentage forskolin inhibition for G i -coupled A 1 R and A 3 R (1 μM or 10 μM, respectively) or stimulation for A 2A R and A 2B R (100 μM, representing the maximum cAMP accumulation of the system), relative to NECA/IB-MECA (agonist allowing comparison across AR subtypes and a selective A 3 R agonist, respectively). For cAMP experiments on A 3 R mutants, data was normalised to 100 μM forskolin, representing the maximum cAMP accumulation possible for each cell line. In the case of pERK1/2 response, normalisation was performed to PMA, a direct PKC activator providing the maximum pERK1/2 level of the system. Schild analysis was performed to obtain pA 2 values (the negative logarithm to base 10 of the molar concentration of an antagonist that makes it necessary to double the concentration of the agonist to elicit the original submaximal response obtained by agonist alone 45 ) for the potential antagonists. In cases where the Schild slope did not differ significantly from unity, the slope was constrained to unity giving an estimate of antagonist affinity (pK B ). pA 2 and pK B coincide when the slope is exactly unity. Through performing Schild analysis, whereby the pA 2 is independent of agonist, we were able to experimentally determine the effect of receptor mutation on antagonist binding. The pA 2 values obtained through conducting Schild analysis of K18 at WT and mutant A 3 R were compared in order to indicate important residues involved in K18 binding. Whereas an increase in the pA 2 for a particular mutant when compared to WT suggested the antagonist was more potent, a decrease indicated a reduced potency.
All experiments were conducted in duplicate (technical replicates) to ensure the reliability of single values. The data and statistical analysis comply with the recommendations on experimental design and analysis in pharmacology 46 . Statistical analysis, performed using Prism 8.0, was undertaken for experiments where the group size was at least n = 5 and these independent values used to calculate statistical significance (*p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001) using a one-way ANOVA with a Dunnett's post-test for multiple comparisons or Students' t-test, as appropriate. Any experiments conducted n < 5 should be considered preliminary. Compounds taken forward for further investigation after our preliminary screening (n = 3) were selected based on a high mean cAMP accumulation (> 80%).

Computational biochemistry. MD simulations. Preparation of the complexes between human A 3 R
with K5, K17, K18 or MRS 1220 and rat A 3 R with K18 or K25 was based on a homology model of A 2A R (see Appendix II in Supplementary material) and are detailed in Lagarias et al. 21 . Each ligand−protein complex was embedded in hydrated POPE bilayers. A simulation box of the protein-ligand complexes in POPE lipids, water and ions was built using the System Builder utility of Desmond (Desmond Molecular Dynamics System, version 3.0; D.E. Shaw Res. New York, 2011; Maest. Interoperability Tools, 3.1; Schrodinger Res. New York, 2012). A buffered orthorhombic system in 10 Å distance from the solute atoms with periodic boundary conditions was constructed for all the complexes. The MD simulations were performed with Amber14 and each complexbilayer system was processed by the LEaP module in AmberTools14 under the AMBER14 software package 47 . Amber ff14SB force field parameters 48 were applied to the protein, lipid14 to the lipids, GAFF to the ligands 49 and TIP3P 50 to the water molecules for the calculation of bonded, vdW parameters and electrostatic interactions. Atomic charges were computed according to the RESP procedure 51 using Gaussian03 52 and antechamber of AmberTools14 47 . The temperature of 310 K was used in MD simulations in order to ensure that the membrane state is above the main phase transition temperature of 298 K for POPE bilayers 53 . In the production phase, the relaxed systems were simulated in the NPT ensemble conditions for 100 ns. The visualization of produced trajectories and structures was performed using the programs Chimera 54 and VMD 55 . All the MD simulations were run on GTX 1060 GPUs in lab workstations or on the ARIS Supercomputer.
MM-PBSA calculations. Relative binding free energies of the complexes between K5, K17, K18, MRS 1220 and A 3 R was estimated by the 1-trajectory MM-PBSA approach 56 . Effective binding energies (ΔG eff ) were computed considering the gas phase energy and solvation free energy contributions to binding-see appendix II 21,22 .
Nomenclature of targets and ligands. Key protein targets and ligands in this article are hyperlinked to corresponding entries in https ://www.guide topha rmaco logy.org, the common portal for data from the IUPHAR/ BPS Guide to Pharmacology 57 , and are permanently archived in the Concise Guide to Pharmacology 2017/18 58 .

Data availability
The data are available from the corresponding authors on reasonable request.