Identifying Functional Hotspot Residues for Biased Ligand Design in G-Protein-Coupled Receptors S

Supplemental material to this article can be found at: http://molpharm.aspetjournals.org/content/suppl/2018/01/24/mol.117.110395.DC1

1521-0111/93/4/288–296$35.00 https://doi.org/10.1124/mol.117.110395 MOLECULAR PHARMACOLOGY Mol Pharmacol 93:288–296, April 2018 Copyright ª 2018 by The American Society for Pharmacology and Experimental Therapeutics

Identifying Functional Hotspot Residues for Biased Ligand Design in G-Protein-Coupled Receptors s

Anita K. Nivedha, Christofer S. Tautermann, Supriyo Bhattacharya, Sangbae Lee, Paola Casarosa, Ines Kollak, Tobias Kiechle, and Nagarajan Vaidehi Department of Molecular Immunology, Beckman Research Institute of the City of Hope, Duarte, California (A.K.N., S.B., S.L., N.V.); Departments of Medicinal Chemistry (C.S.T.) and Immunology and Respiratory Diseases Research (I.K., T.K.), Boehringer Ingelheim Pharma GmbH & Co. KG, Biberach, Germany; and Corporate Department of Business Development and Licensing, C.H. Boehringer Sohn, Ingelheim, Germany (P.C.)

Received August 24, 2017; accepted January 16, 2018 Downloaded from

ABSTRACT G-protein-coupled receptors (GPCRs) mediate multiple signal- coupling region in GPCRs, we have developed a computational ing pathways in the cell, depending on the agonist that activates method to calculate ligand bias ahead of experiments. We have

the receptor and multiple cellular factors. Agonists that show validated the method for several b-arrestin–biased agonists in molpharm.aspetjournals.org higher potency to specific signaling pathways over others are b2-adrenergic receptor (b2AR), serotonin receptors 5-HT1B and known as “biased agonists” and have been shown to have better 5-HT2B and for G-protein–biased agonists in the k-opioid therapeutic index. Although biased agonists are desirable, their receptor. Using this computational method, we also performed design poses several challenges to date. The number of assays a blind prediction followed by experimental testing and showed to identify biased agonists seems expensive and tedious. that the agonist carmoterol is b-arrestin–biased in b2AR. Therefore, computational methods that can reliably calculate Additionally, we have identified amino acid residues in the the possible bias of various ligands ahead of experiments and biased agonist binding site in both b2AR and k-opioid receptors provide guidance, will be both cost and time effective. In this that are involved in potentiating the ligand bias. We call these work, using the mechanism of allosteric communication from residues functional hotspots, and they can be used to derive

the extracellular region to the intracellular transducer protein pharmacophores to design biased agonists in GPCRs. at ASPET Journals on October 1, 2021

Introduction lowering the unfavorable side effects of a drug. This property of biased agonists can markedly improve their therapeutic G-protein-coupled receptors (GPCRs) are a superfamily of index when compared with unbiased or balanced agonists signaling proteins of critical importance to cellular function, (Mailman, 2007; Neve, 2009; Violin et al., 2014). Design of and comprise the largest group of drug targets. The binding of biased agonists remains a daunting challenge due to the endogenous agonists to their target GPCRs results in coupling complexity of the cellular and structural factors that contrib- to a variety of intracellular transducer proteins, including ute to the bias. Therefore, an understanding of the structural various heterotrimeric G-proteins and b-arrestins, that medi- underpinnings of biased signaling of both G-protein bias as ate different downstream signaling pathways. Certain GPCR well as b-arrestin bias of GPCRs is needed to design biased agonists elicit differential responses to different signaling agonists for GPCRs. pathways and are known as biased agonists, and the phenom- Linking GPCR Conformation to Cellular Function. enon is called functional selectivity or biased signaling (Urban Several factors contribute to biased signaling by an agonist– et al., 2007; Rajagopal et al., 2010; Rasmussen et al., 2011; GPCR pair in cells: 1) structural features of the agonist–GPCR– Wisler et al., 2014). G-protein or agonist–GPCR–b-arrestin complex (Shukla et al., The ability of biased agonists to cause differential signaling 2014) and 2) cell-specific and tissue-specific factors (Zhou and resulting in more focused cellular changes could translate to Bohn, 2014). Designing biased agonists is a challenge due to the lack of structural information on how an agonist–GPCR This work was funded by Boehringer-Ingelheim and by National Institutes of pair selectively couples to specific G-proteins or b-arrestins. Health National Institute of General Medical Sciences [Grant R01-GM097296]. Moreover, because GPCRs are dynamic, with various possi- https://doi.org/10.1124/mol.117.110395. s This article has supplemental material available at molpharm. ble active-state conformations, the structural ensemble of aspetjournals.org. the agonist–GPCR–transducer complex is one of the key

ABBREVIATIONS: 3D, three-dimensional; b2AR, b 2-adrenergic receptor; BAI, b-arrestin interface; BI-167107, 5-hydroxy-8-[1-hydroxy-2-[[2- methyl-1-(2-methylphenyl)propan-2-yl]amino]ethyl]-4H-1,4-benzoxazin-3-one; EC, extracellular; ECL2, extracellular loop 2; GPCR, G-protein- coupled receptor; GPI, G-protein interface; 5-HT1B, 5-hydroxytryptamine (serotonin) receptor 1B; 5-HT2B, 5-hydroxytryptamine (serotonin) receptor 2B; MD, molecular dynamics; kOR, k-opioid receptor; PDB, Protein Data Bank; PR, allosteric pipeline ratio; SAR, structure–activity relationship; TM, transmembrane domain; U69593, N-methyl-2-phenyl-N-[(5R,7S,8S)-7-pyrrolidin-1-yl-1-oxaspiro[4.5]decan-8-yl]acetamide.

288 Identifying Functional Hotspots for Biased Ligand Design 289 determinants driving the biased signaling (Shukla et al., 2014). the m-opioid receptor because the active state structure of kOR The selectivity of GPCRs to trigger distinct intracellular is not available and we wanted to test our computational signaling profiles is very ligand dependent. Spectroscopic method using a homology model of the receptor that makes the studies have now established that ligands differentially stabi- testing more realistic. lize a range of GPCR conformations (Yao et al., 2006; Kahsai The serotonin receptors 5-hydroxytryptamine 1B (5-HT1B) et al., 2011; Liu et al., 2012; Nygaard et al., 2013). Therefore, and 5-HT2B have been crystallized with the same agonist knowledge about how the agonist affects the agonist–GPCR–G- ergotamine, yet ergotamine shows b-arrestin bias in 5-HT2B protein structural ensemble will significantly aid the design of and not in 5-HT1B (Wacker et al., 2013). We chose to study the biased agonists. system to find out which residues in 5-HT2B cause the bias in Challenges in Biased Ligand Identification and ergotamine. This example would provide a validation for the Design. Given the huge challenges in crystallography and same ligand showing different biases in two related receptors. NMR in determining structural ensembles of agonist-GPCR- Another similar example is the bias behavior of epinephrine in effector complexes, computational methods can greatly aid in the triple mutant T68F, Y132G, Y219A b2AR, denoted here as delineating the structural factors that determine agonist b2ARTYY. This mutant was designed using evolutionary trace bias and its outcome. Despite a wealth of biochemical and analysis to identify residues that could cause bias in b2ARTYY biophysical studies on inactive conformations, there is (Shenoy et al., 2006). Thus, choosing the variety of different

paucity of structural information on active conformations of systems described here provides a test for the robustness of Downloaded from GPCRs. More importantly the dynamics of the agonist- our computational method. GPCR-effector complex is one of the major factors leading To summarize, we have studied b2AR and its biased mutant to a conformational ensemble, which governs the functional b2ARTYY with several agonists, 5HT1B and 5HT2B with the same selectivity of biased signaling. agonist, and kOR with three agonists including two G-protein– A better understanding of the structural basis of G protein/ biased agonists. These systems are also listed in Supplemental b-arrestin selection will provide for a more precise targeting of Table 1 of the Supplemental Information for convenience. molpharm.aspetjournals.org GPCRs with pharmaceuticals and also inform ongoing structure-based drug discovery efforts. Hence, our goal in this work is to identify structural elements in the GPCR that Materials and Methods confer functional selectivity either to the G-protein signaling A New Computational Method for Calculating Ligand Bias pathway or to the b-arrestin signaling pathway. Our unan- The experimentally measured ligand bias factor depends on the swered question has been, how do the various biased agonists 21 ligand affinity (KA ) and its efficacy (t) toward a specific signaling influence the role of these structural determinants in biasing pathway. The bias factor is calculated with respect to a reference b the signaling to effect -arrestin signaling or G-protein ligand using the Black and Leff operational model (Kenakin et al., signaling? Here we have developed a computational method 2012). Because macroscopic properties such as ligand potency and at ASPET Journals on October 1, 2021 to calculate the ligand bias ahead of experiments, which would coupling efficacy are still intractable computationally, we developed a aid the design of biased agonists. The two-part goal in this method to calculate the ligand bias from atomic-level properties of the work is: 1) to develop a method to compute the ligand bias structural ensemble of the agonist-GPCR complex that potentiates the ahead of experiments and 2) to identify the residues in the agonist bias. agonist binding site that potentiate the ligand mediated bias. Agonist and G-protein binding to GPCRs leads to G-protein We define these residues as functional hotspots. To this end activation. Such activation at a distance is potentiated by the allosteric communication between the ligand binding site and the we chose systems that cover a wider base of bias systems. b b G-protein coupling site in the fully active state of GPCRs. We recently The human 2-adrenergic receptor ( 2AR) and angiotensin developed a computational method to calculate the strength of the receptor AT1R have been studied extensively for discovery of allosteric communication pipelines from the extracellular region of the b-arrestin–biased ligands (Violin et al., 2014; Wisler et al., GPCR to the intracellular regions where the G-protein or b-arrestin 2014). Therefore, robust experimental data on the quantita- couples to the receptor. In this work we hypothesize that the strength tive ligand bias factor calculated using the Black-Leff model of the allosteric communication pipelines from the extracellular (EC) (Black and Leff, 1983) are available for these receptors. We region passing through the residues in the ligand-binding site (ligand chose b2AR for our study because the three-dimensional (3D) shown in blue spheres in Fig. 1) to the residues in the G-protein and b-arrestin coupling interface in the GPCRs is related to the ligand crystal structure of the fully active Gs-protein bound state is available and this constitutes a robust test case for the potency and its coupling efficacy. Figure 1 shows the scheme used here for calculating the ligand bias computational method. The experimental values of the ligand in GPCRs: the allosteric communication pipelines that connect to the bias factor have been calculated using cAMP secondary G-protein interface are shown as pink sticks, and those to the messenger assays for assessing the G-protein coupling po- b-arrestin interface are shown as green sticks. We defined the ligand tency for the b2AR and using Tango or other b-arrestin bias as the ratio of the strength of the allosteric communication recruitment assays for calculating the ligand potency toward pipelines to the G-protein pathway to that of the b-arrestin signaling b-arrestin–mediated signaling pathways (Rajagopal et al., pathway for a test ligand, to that of the same quantity for the reference 2011; Weiss et al., 2013; Zhou et al., 2013). ligand. We used a reference ligand in these calculations so that these There has been a surge in efforts to develop G-protein– quantities can be directly compared with the experimental bias factor. biased opioid receptor agonists that serve as better analgesics with reduced side effects coming from receptor desensitization Computational Methods (Soergel et al., 2014; Luttrell et al., 2015; Stahl et al., 2015). The computational methods for calculating the allosteric commu- We tested three agonists, two of which showed different nication pipelines in GPCRs involve the following steps: 1) preparing extents of G-protein bias for the k-opioid receptor (kOR) the GPCR structure with agonists for the molecular dynamics (MD) (Zhou et al., 2013). We chose to study the kOR rather than simulations and 2) performing MD simulations in explicit membrane 290 Nivedha et al.

residues in the agonist-binding site to the residues on the b-arrestin or G-protein interacting interfaces. The list of residues in each of these regions is given in Supplemental Table 2. For each MD run, the top 10% of allosteric pathways ranked by the total mutual information were used for further calculation of the ligand bias. The strength of an allosteric communication pipeline is the number of overlapping allosteric communication pathways contained in the pipeline (Bhattacharya and Vaidehi, 2014a; Vaidehi and Bhattacharya, 2016).

Residue Groups Definitions Here we list the residues that define the various regions in the GPCRs. These residues were included in the calculation of the allosteric communication pipelines. Ligand Binding Site Residues. The residues located within 5 Å of the agonist-binding site in the GPCR and contacting the agonist in more than 40% of the snapshots from the MD simulation trajectories were chosen as the binding site residues for that agonist. The super set of

binding site residues for all agonists to a particular receptor were listed by Downloaded from combining the residues list for all the agonists to the receptor studied. This superset of residues was used to calculate the allosteric communication pipeline from the EC loops going through the ligand binding site to the G-protein or b-arrestin coupling interface. The list of residues in the ligand binding site used in this study is shown in Supplemental Table 2. b-Arrestin Interface Residues and G-Protein Interface

Fig. 1. The model used for calculating the ligand bias in GPCRs. The Residues. The list of residues in the GPCR interacting with the molpharm.aspetjournals.org allosteric communication pipelines starting from the residues in the extracellular loop region and passing through the residues in the agonist G-protein was obtained from Supplementary Table 1 of the crystal b binding site (agonist shown as blue spheres) and terminating in the structure paper of 2AR-Gs protein complex (Rasmussen et al., 2011). residues that couple to the G-protein (GPI shown as pink surfaces) are The Protein Data Bank (PDB) ID of this crystal structure is 3SN6. The shown as pink sticks. The allosteric communication pipelines that connect list of residues interacting with the b-arrestin interface was obtained to the b-arrestin interface (BAI shown as green surfaces) are shown in by aligning the b2AR structure from PDB ID: 3SN6 to the crystal green sticks. The ratio of the strength of these two pipelines for a test structure of rhodopsin bound to visual arrestin (PDB ID: 5DGY) (Zhou ligand with respect to a reference ligand is defined as the ligand bias. et al., 2016) and identifying the residues in the receptor that were within 5 Å of visual arrestin in the aligned structure. The kOR bilayer and water. The list of receptors used in this work is shown in structure was aligned to that of b2AR, and the b-arrestin interface

Supplemental Table 1. The two-dimensional representation of the (BAI) and G-protein interface (GPI) residues were translated for the at ASPET Journals on October 1, 2021 agonists used in this study is shown in Supplemental Fig. 1. The kOR based on the alignment of residues and corresponding details of the receptor structure preparation, docking of the agonists, Ballesteros-Weinstein numbers. The list of residues in the GPI and and MD simulations are provided in the Supplemental Information BAI used in this study are shown in Supplemental Table 2. file. We performed brute force MD simulations with no constraints for EC Region Residues. Residues comprising the EC loops of the wild-type b2AR with several ligands listed in Supplemental Table 1, receptor in addition to those residues forming two turns in the EC TYY b2AR mutant with epinephrine, serotonin receptors 5-HT1B and region of each transmembrane helix formed the set of EC residues. 5-HT2B with ergotamine, and homology model of kOR with three The EC loop residues included in the calculation of allosteric different agonists bound. communication pipelines for the GPCRs studied here are provided in Supplemental Table 2. Allosteric Pipeline Analysis The MD simulation trajectories were used to calculate the alloste- Definition of the Computational Ligand Bias ric communication pipelines for each agonist-GPCR pair using the Using the MD trajectories, we calculated the allosteric communication Allosteer computational method (Bhattacharya and Vaidehi, 2014a,b; pipelines of every agonist-GPCR complex to the G-protein and b-arrestin Bhattacharya et al., 2016; Vaidehi and Bhattacharya, 2016), which we interface. We then calculated the ratio of the strength of the allosteric have developed previously elsewhere. The details of the Allosteer communication pipeline (PR) for the b-arrestin coupling interface to that computational method for calculating the allosteric communication pipe- of the G-protein coupling interface residues. We did the same calculation lines in GPCRs are described by Bhattacharya and Vaidehi (2014a,b). for all agonists, including the reference agonist, which is usually a Here we provide a brief description of the method as used in this work. balanced agonist, used in the experimental assays for the bias factor Using the MD trajectories, we first calculated the correlated calculations. The computational ligand bias values is defined as: movement between residues on the EC surface of the receptor and the residues in G-protein or b-arrestin coupling interface. Here we PRlig have used the trajectories from the five 200-nanosecond MD simula- Computational ligand bias; b 5 2 1 (1) PRref tions collected as a total of 1-microsecond of MD simulations to calculate the mutual information in torsion angle space between all where PRlig is the ratio of the strength of the allosteric communication pairs of residues in each receptor-ligand complex. Using the mutual pipeline for a given agonist-receptor pair for the b-arrestin to the G-protein information, we then calculated the shortest pathway from the EC coupling interfaces, and PRref isthesameforareferenceagonist. region to the G-protein or b-arrestin coupling surfaces that maxi- mizes the correlated movement. This method is called Allosteer Testing the Robustness of the Method to Calculate the (Bhattacharya and Vaidehi, 2014a). Using the Allosteer method we then calculated the allosteric Ligand Bias communication pipelines starting from the residues in the EC loops The computational ligand bias shown in eq. 1 is a ratio that is (defined in the section Residue Group Definitions) passing through the obtained by calculating the strength of the allosteric communication Identifying Functional Hotspots for Biased Ligand Design 291 pipelines to the b-arrestin and the G-protein coupling interfaces. The Black and Leff Model Used for Calculating the Bias Factor error in the calculation of the strength of the allosteric pipelines is The concentration–response data for each agonist at the G protein dependent on: 1) the number of snapshots used from the MD and b-arrestin signaling assays were fitted to the Black-Leff model of trajectories and 2) the simulation time length of each MD simulation. agonism (Black and Leff, 1983): The dependence on the number of snapshots stored in the MD simulations is because the state of the system has to be captured ½AntnðE 2 BasalÞ Response 2 Basal 5 m (2) often enough to enable the recording of all significant, meaningful n n n ½A t 1 ð½A 1 KAÞ correlations. However, recording snapshots at extremely frequent intervals would increase the probability of registering spurious where [A] is the agonist concentration, t is equal to RT/KE (RT: receptor correlations and dependencies, which in turn will introduce noise into density; KE: intrinsic agonist efficacy), n is the transducer slope, and Em the bias calculations. is the maximal response of the system. The transduction coefficient for

The second factor, the length of the MD simulations, is important each agonist for a given pathway was calculated as log(t/KA) (Kenakin because longer MD simulations capture the collective correlated and Christopoulos, 2013). For fitting the data, the Black-Leff equation motions of torsion angles in proteins. However, a tradeoff has to be was recast to a different form according to Tschammer et al. (2011), and made between the cost of computational resources and accuracy of log(t/KA) was directly obtained from the fit. Em was estimated as the ligand bias calculated here. Therefore, we examined the optimum maximum value of the signaling response for a given assay. number of snapshots to be stored and length of MD simulations to Setting the value of n 5 1 gave a good fit for all the dose–response reduce the significant errors in the ligand bias calculations. Thus, we data. In practice, n was allowed to vary within a very narrow range (0.9– calculated the variation in the S.E. in the pipeline strength as a 1.2) to account for statistical variability. The fitting was performed Downloaded from function of these two variables. The S.E. in these calculations was using the Genetic Algorithm (GA) module in MATLAB (MathWorks, analyzed with the aim of choosing optimal values for both (Supple- Natick, MA) (the operational model fitting of GraphPad Prism did not mental Figs. 2 and 3). Upon analysis of the S.E. values, storing always find a solution for all data sets). Instead of starting from a single snapshots from MD simulations at every 20-picosecond time step in a initial guess of the solution, 10,000 initial guesses were randomly

200-nanosecond MD simulation (10,000 snapshots) was chosen as the generated within a prescribed range. The provided range for KA was 215 optimum frame rate, and 200 nanoseconds was chosen as the 10 to 1, while for log(t/KA) it was 0–15 (the range for n was stated simulation time length for the calculations. For more details on these earlier). Using the different initial guesses, the algorithm converged to a molpharm.aspetjournals.org tests we refer the reader to the Supplemental Information. solution within the provided tolerance limit of 1028. Experimental Methods. For each agonist and a given pathway, Dlog(t/KA) was calculated as log(t/ b-Arrestin Recruitment Assay and cAMP Detection Assay. To test KA) 2 log(t/KA)ref,wherelog(t/KA)ref is the value for the reference agonist the predictivity of the computational methods, three compounds (iso- isoproterenol. DDlog(t/KA) was then calculated as Dlog(t/KA)arrestin 2 DDlog(t/KA) proterenol, formoterol, and carmoterol) were tested in a b-arrestin assay Dlog(t/KA)Gprotein.Thus,thebiasfactorisgivenby10 . as well as a cAMP assay on human b2-adrenoceptors. Isoproterenol and To estimate the variability in the calculated log(t/KA) values, we 2 formoterol were purchased from Sigma-Aldrich Chemie, (Hoxter, calculated Sij for agonist i and pathway j as: Germany), and carmoterol was synthesized using the protocol described 1 nj in the US Patent US2010/0113790 A1 (Kankan et al., 2010). For 2 5 + 2 2 Sij yijk ymean (3) 2 at ASPET Journals on October 1, 2021 b-arrestin assay we chose the PathHunter b-Arrestin Assay system nij 1 k 5 1 (DiscoveRx, Fremont, CA); for cAMP, the AlphaScreen (PerkinElmer where k denotes the number of experiments and y corresponds to the Life and Analytical Sciences, Waltham, MA) technology was employed. log(t/K ) for the agonist/pathway pair. The total variability of the b-Arrestin Recruitment Assay. The PathHunter b-Arrestin Assay A estimates S is given by: system (DiscoveRx) was used to determine the b-arrestin recruitment pooled sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi to human b -adrenoceptors. Chinese hamster ovary K1 (CHO-K1) 2 +2 +3 2 5 5 Sij cells stably expressing b2-adrenoceptors fused to ProLink and the i 1 j 1 Spooled 5 (4) b-arrestin-enzyme acceptor were seeded onto opaque 384-well plates dferror at 5000 cells per well in Dulbecco’s modified Eagle’s medium/Ham’s F-12 (1:1) without phenol red supplemented with 1% heat-inactivated where dferror is the degree of freedom given by: fetal bovine serum. The cells were maintained at 37°C in a 5% CO , 2 df 5 +2 +3 n 2 1 (5) humidified atmosphere overnight. The next day the cells were error i 5 1 j 5 1 ij stimulated in growth medium with different concentrations of ago- The 95% confidence levels for the calculated log(t/K ) values are nists at 37°C for 60, 30, 15, and 5 minutes. Arrestin recruitment A given by: resulting in b-galactosidase enzyme complementation was detected by use of the PathHunter Flash detection kit (DiscoveRx). Luminescence Confidence level 5 logðt=KAÞ 6 Tðdf error; 0:975ÞðStandard errorÞ was read on an Envision plate reader (PerkinElmer Life and (6) Analytical Sciences). We performed 13 experiments for isoproterenol, 9 for formoterol, and 4 for carmoterol. where S.E. is given by: cAMP Detection Assay. Changes in intracellular cAMP levels were sffiffiffiffiffiffi determined by AlphaScreen technology (PerkinElmer Life and Ana- 1 Standard error 5 S (7) lytical Sciences) in CHO-K1 cells stably expressing b2-adrenoceptors pooled nij following the manufacturer’s instructions. Cells in suspension (15,000 cells per well) were stimulated in opaque 384-well plates with and T corresponds to a two-tailed t test with 95% confidence. For respective agonists at different concentrations in Hank’s buffered calculating the confidence levels for DDlog(t/KA), we used the same saline solution supplemented with 5 mM HEPES, 0.1% bovine serum method (eq. 6), except the S.E. is given by: albumin, and 500 mM 3-isobutyl-1-methylxanthine for 30 minutes at sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi room temperature. The cells were lysed by using Alphascreen 1 1 1 1 Standard error 5 Spooled 1 1 1 reagents. After 2 hours, the plates were read on an Envision plate nG protein nb arrestin nG protein; ref nb arrestin; ref reader (PerkinElmer Life and Analytical Sciences). The concentration (8) of cAMP in the samples was calculated from a standard curve. We performed 12 experiments each for isoproterenol, formoterol, and where n refers to the number of experiments for the given pathway, carmoterol. and nref is the number of experiments for the reference ligand. 292 Nivedha et al.

Results in ligand potency and efficacy that manifests as ligand bias. Thus, the computational method provides a rapid method to Calculation of Ligand Bias Factor Using Black-Leff identify biased ligands ahead of experiments, and is useful in Operational Model for Carmoterol and Formoterol. We designing other biased ligands. measured the ligand bias factor of carmoterol and formoterol The agonists epinephrine, isoproterenol, and salbutamol using isoproterenol as the reference ligand. Using the are considered balanced agonists to b2AR (Rajagopal concentration-dependent curves measured using the cAMP et al., 2011) whereas formoterol, 5-hydroxy-8-[1-hydroxy- assay for G-protein activation and PathHunter assay for 2-[[2-methyl-1-(2-methylphenyl)propan-2-yl]amino]ethyl]- b-arrestin recruitment, as detailed in the Materials and 4H-1,4-benzoxazin-3-one (BI-167107), and carvedilol are Methods section, we calculated the ligand bias factor for b-arrestin–biased, with carvedilol acting as an inverse agonist carmoterol and formoterol with reference to isoproterenol in to G-protein (Wisler et al., 2007). In the study by Rajagopal b2AR. We used the Black-Leff model to obtain bias factors as et al. (2011), the bias factors of epinephrine, isoproterenol, described in the Materials and Methods section. The calcu- formoterol and salbutamol was measured, and the investiga- lated bias factors are plotted in Fig. 2A and are also shown in tors concluded that formoterol was a b-arrestin–biased ago- Supplemental Table 3. Our results show that carmoterol has a nist, while the remaining ligands failed to exhibit any 2-fold higher bias factor toward the b-arrestin signaling considerable bias toward either arrestin or G-protein. pathway compared with formoterol in b2AR.

As shown in Fig. 2A, in agreement with experiments, Downloaded from Comparison of Computational Ligand Bias to Experi- formoterol is more biased toward b-arrestin compared with mental Bias Factor. Figure 2 shows the comparison of isoproterenol or salbutamol in b2AR. Similarly, the super the computed ligand bias to the experimental ligand bias agonist BI-167107 is more biased toward b-arrestin compared factor for various b-arrestin and G-protein–biased agonists with isoproterenol. Carvedilol is a weak b-arrestin–biased for the b2AR and kOR, respectively. In one set of experiments agonist and G-protein antagonist, so experimental bias epinephrine was used as the reference ligand to calculate the factor values could not be obtained. However, the computa- molpharm.aspetjournals.org bias factor from experimental titration curves (Rajagopal tional ligand bias values indicate that carvedilol is indeed et al., 2011), and isoproterenol was used as a reference ligand b-arrestin–biased. As a blind test to our method using MD in another set of experimental measurements (Weiss et al., simulations and calculation of ligand bias, we further predicted 2013). Supplemental Fig. 4, A and B, shows the same that carmoterol is a b-arrestin–biased ligand compared with comparisons for 5-HT receptors and b2ARTYY mutant. isoproterenol. We then performed the cAMP and the Path- As seen in Fig. 2 the computed ligand bias recapitulates the Hunter assays as described in the Results section and calcu- rank ordering of the ligand bias factors for all the agonists in lated the bias factors shown in Supplemental Table 3. As all the four receptors for both b-arrestin and G-protein–biased predicted, carmoterol is indeed b-arrestin–biased with refer- agonists. It is important to note that the ligand bias ordering ence to isoproterenol. has been captured for the G-protein–biased agonists where we In kOR (Fig. 2B), we correctly predicted that the agonists 1.1 at ASPET Journals on October 1, 2021 used a homology model for kOR. The magnitude of computed and 2.1 are G-protein–biased compared with N-methyl-2- ligand bias shown in Fig. 2A is an order of magnitude lower phenyl-N-[(5R,7S,8S)-7-pyrrolidin-1-yl-1-oxaspiro[4.5]decan-8-yl] than the experimental ligand bias factor. This is because we acetamide (U69593). The experimental results (Zhou et al., 2013) have not computed the macroscopic property of the ligand bias showed that when using the ligand U69593 as the reference factor as measured by experiments. Instead we have com- agonist, probe 1.1 and probe 2.1 are G-protein–biased, with probe puted the relative strength of the allosteric communication 1.1beingastrongerG-protein– biased agonist compared with 2.1. pipelines for the test ligand to that of the reference ligand and Theoretical ligand bias calculations were obtained for these three this is a molecular factor that potentiates and leads to change ligand-receptor systems, and the values were found to be in

Fig. 2. Comparison of computational ligand bias (y axis) to the experimental ligand bias factors (x axis) for (A) the b-arrestin–biased agonists in b2AR. epi, epinephrine; iso, isoproterenol; sal, salbutamol; form, formoterol; BI, BI-167107; carm, carmoterol. (B) the G-protein–biased agonists in kOR. u69, u69593. (A) The bias values obtained using epinephrine as the reference agonist are shown as green triangles, and those obtained using isoproterenol as the reference agonist are shown as green circles. The respective reference agonists from both studies are marked at zero. The error bars for the bias factors of all b2AR agonists represent S.E. with 95% confidence limits except for BI-167107 which is 99% confidence limit. The number of experiments for these cases were n $ 3. (B) The bias values obtained for the G-protein–biased system kOR is shown using U69593 as the reference agonist. For the opioid receptor agonists, the experimental bias factor was calculated from at least three independent experiments. All computational ligand bias values are reported with S.E. and were obtained from five independent MD simulations. Identifying Functional Hotspots for Biased Ligand Design 293 excellent agreement with the experimental bias factor values, not b2AR complex (as shown in Fig. 3B) than in the epinephrine- only indicating the G-protein bias of both probes with respect to b2AR complex. However, it is interesting to note that the U69593, but also reflecting the greater G-protein bias for probe 1.1 allosteric communication pipelines to the G-protein interface compared with probe 2.1. are equally strong in both epinephrine and formoterol Comparison of the computational ligand bias for ergotamine (Fig. 3A). Both epinephrine and formoterol show strong in 5-HT1B and 5-HT2B as well as epinephrine in b2AR and G-protein communication pipelines in the fully active state the b2ARTYY mutant showed good agreement with the of the receptor, and the allosteric pipeline strength to the experimental bias factors reported (Supplemental Fig. 4). b-arrestin coupling interface is comparatively weaker in The serotonin receptors are an example of a system with epinephrine than in formoterol. This is also true in the different receptor serotypes in complex with the same ligand, experimental assays, where Rajagopal et al. (2011) showed and one receptor system being neutral while the other is that the Emax for formoterol and epinephrine was similar for b-arrestin–biased. The computational ligand bias values were the cAMP response while that for b-arrestin recruitment was in agreement with experimental bias factor values. There is no lower for epinephrine than for formoterol. To date there is no quantitative experimental bias factor available for the known biased agonist that shows a strong preference for b2ARTYY mutant. b-arrestin and not for G-protein in b2AR. It should be noted These results for several systems and different types of that we have used the fully active G-protein–bound confor- b agonists show the robustness and the reliability of our method mation of 2AR for all these calculations, and the allosteric Downloaded from in predicting the ligand bias ahead of experiments. As de- pipeline calculations are sensitive to the conformational state scribed in detail in the Materials and Methods section, the of the receptor (Bhattacharya and Vaidehi, 2014a). confidence level for the prediction of ligand bias is 90%. Thus, One of the challenges in designing biased agonists is the the theoretical ligand bias calculations were validated against lack of structural details of biased agonist binding sites that experimental bias factor values, and they were found to be could aid structure–activity relationship (SAR) studies. The clearly indicative of whether a preference for either G-protein binding site of biased agonists looks similar to that of balanced molpharm.aspetjournals.org and/or b-arrestin exists. agonists from crystal structures (Warne et al., 2012). There- Identifying Functional Hotspot Residues in the Ligand fore, it would be beneficial to SAR studies if we could delineate Binding Site That Contribute to the Bias. Figure 3 shows the residues in the biased agonist binding site that contribute the comparison of the shape of the ligand binding sites in to the ligand bias. Toward this goal, we identified the residues epinephrine (neutral agonist) and formoterol (b-arrestin–biased in the biased agonist binding site that showed a significant agonist), respectively, bound to the fully active state of b2AR as contribution to the allosteric communication pipelines of modeled in this work. For fair comparison, this figure shows the either the b-arrestin or G-protein coupling interface by same view in the same receptor orientation for both agonists. calculating the differences in the strengths of the allosteric Unlike epinephrine, the binding site of formoterol extends to the communication pipelines to the BAI and GPI for the biased extracellular loop 2 (ECL2) region of the receptor. The residues and unbiased agonists. at ASPET Journals on October 1, 2021 D1133.32,F193ECL2,S2035.42,andS2045.43—although present in Figure 4 show the residues within 5 Å of the formoterol- and both the ligand binding sites—showed different side chain epinephrine-binding sites, respectively. The larger circles conformations (Supplemental Fig. 5) in formoterol-bound b2AR show sustained contact with the ligand. There are many compared with the epinephrine-bound b2AR. Thus, changes in residues that are common to both the epinephrine- and the rotameric states of these residues reshape the common formoterol-binding sites. As seen in Fig. 4A, the four residues regions of the cavities forming the ligand binding site for both T1103.29, F193ECL2, S2045.43, and Y3087.35 show a significant epinephrine and formoterol. contribution to the allosteric communication pipeline to the The allosteric communication pipelines from the EC loops b-arrestin interface compared with the G-protein pipelines passing through the residues in the formoterol-binding site to with reference to epinephrine (shown in Fig. 4B). These the b-arrestin coupling interface are stronger in formoterol- residues are the functional hotspot residues that confer

Fig. 3. The shape of the agonist binding cavity in b2AR of (A) a neutral agonist epinephrine and (B) formoterol, a biased agonist. The allosteric communication pipelines to the G-protein and the b-arrestin coupling interfaces are also depicted. This figure shows b2AR surfaces sliced to reveal the shape of the ligand binding cavity. 294 Nivedha et al. b-arrestin bias to the formoterol-b2AR pairing. Carmoterol- residues K1735.39, I2406.55, and Y2587.35 areuniqueto1.1. bound b2AR also shows T1103.29, F193ECL2, and S2045.43 as The residues D843.32 and Y652.64 also show a difference in the functional hotspots in its binding site. Thus, three functional side-chain rotamer conformations when bound to U69593 or hotspot residues (T1103.29, F193ECL2, and S2045.43) are com- probe 1.1 (Supplemental Fig. 7). A Y258A7.35 mutation in the mon to both formoterol and carmoterol. This shows that kOR has a strong deleterious effect on the potencies of GPCRs might have specific residues that show functional agonists U69593, dynorphin A, the octahydroisoquinolinone selectivity to signaling pathways in the receptor. carboxamide 1xx, and salvinorin A (Vardy et al., 2013). The 3D view showing the relative orientation of these Mutation of residues D843.32 to Ala/Asn and Y652.64 to residues in the formoterol-binding site is shown in the Ala/Phe negatively affected the affinity and potency of the Supplemental Fig. 6. To identify which chemical moieties in receptor in the presence of U69593, 1xx, and salvinorin A formoterol contribute to bias, we identified the moieties in (Vardy et al., 2013). In summary, we found that the formoterol that contact the functional hotspot residues in its G-protein–biased agonists in kOR displayed a rather rigid binding site. The functional hotspots T1103.29 and Y3087.35 side-chain behavior, whereas, the b-arrestin–biased agonists showed strong contact with the 4-methoxyphenyl-propanyl such as formoterol in b2AR displayed a more flexible side- group attached to the protonated amine (Supplemental chain rotamer distribution. Supplemental Fig. 8 shows the Fig. 6). This moiety that contacts the functional hotspots 3D orientation of the probe 1.1 bound to kOR.

was absent in epinephrine. Therefore, extending the agonist Thus, the functional hotspot residues in the ligand binding Downloaded from beyond the protonated amine group and involving con- site identified by our method can be used 1) for generating a tacts with T1103.29 and Y3087.35 could yield biased agonists pharmacophore for identifying newer b-arrestin–biased li- in b2AR. gands for b2AR and 2) for SAR studies for enhancing or Identifying Functional Hotspot Residues in the improving the bias factor. G-Protein–Biased Agonist Binding Site in kOR. Figure 5 shows the residues within 5 Å of the G-protein–biased kOR Discussion molpharm.aspetjournals.org agonist probe 1.1 and that of the unbiased agonist U69593, respectively. U69593 is predominantly nested in the region Using MD simulation trajectories and the Allosteer method between transmembrane helices TM2, TM3, and TM7 (Fig. to calculate the allosteric communication pipelines, we have 5B). On the other hand, the G-protein–biased agonist 1.1 is more developed a computational method to calculate the ligand bias flexible in the binding site and made contacts with residues in in GPCRs. The ligand bias is defined as the relative strength of TM2, TM3, TM5, and TM6, as shown in Fig. 5A. The residues allosteric communication pipelines starting from the residues Y652.64,D843.32, K1735.39,I2406.55, and Y2587.35 are functional in the EC loops passing through the residues in the agonist hotspot residues in the 1.1 biased agonist binding site that binding site to the residues in the b-arrestin or G-protein contribute significantly to the strength of the allosteric commu- interacting interface in the GPCR. We found that the calcu- nication pipeline to the G-protein interacting interface compared lated ligand bias shows the same ordering as the experimental at ASPET Journals on October 1, 2021 with the b-arrestin interacting interface. The residues Y652.64, bias factors. Using the computations, we correctly predicted D843.32,I2406.55,andY2587.35 are common functional hotspot that carmoterol is b-arrestin–biased in b2AR with respect to residues to both the 1.1 and 2.1. The residues D843.32 and Y652.64 isoproterenol. We have identified the functional hotspot are common to both the 1.1 and U69593 binding sites, but the residues in both b2AR and kOR, as shown in Figs. 4A and

Fig. 4. (A) The residues within 5Å of formoterol in b2AR. (B) The binding site residues within 5 Å of epinephrine in b2AR. The size of the circles is proportional to the percentage of snapshots from the MD simulations that show the contacts. Large circles: agonist-residue contacts present in more than 75% of the snapshots from MD simulations. Medium-size circles: between 60% and 75% of the snapshots. Small circles: between 40% and 60% of the snapshots. Residues shown in red text with yellow highlighting are the functional hotspot residues. Identifying Functional Hotspots for Biased Ligand Design 295 Downloaded from Fig. 5. (A) The residues within 5Å of the G-protein–biased agonist 1.1 in kOR. (B) The binding site residues within 5 Å of the balanced agonist U69593 in kOR. The size of the circle is proportionate to the percentage of snapshots from the MD simulations that show the contacts. Large circles: agonist-residue contacts present in more than 75% of the snapshots from MD simulations. Medium-size circles: between 60% and 75% of the snapshots. Small circles: between 40% and 60% of the snapshots. Residues shown in red text with yellow highlight are the functional hotspot residues.

5A, respectively, that can be used in designing b-arrestin– or identify biased agonists from virtual screening and 2) to use G-protein–biased ligands. the functional hotpots to modify any hit ligand as a starting molpharm.aspetjournals.org Even to date, the identification and optimization of biased point for bias optimization. This allows us to pick starting ligands is serendipity-driven. Usually the starting points for points for lead optimization based on crucial parameters such ligand optimization are identified in finding campaigns, such as pharmacokinetic properties and off-target profiles rather as high-throughput screening, fragment-based screening, or than the bias of the initial hit. The ligand bias can be built in virtual screening. The hit compounds are tested for their by modifications to contact the functional hotspots. ligand bias experimentally, and the promising starting points Another important, sought-after tool in the pharmaceutical for further optimization have to show the desired biased industry is the ability to assess whether a particular target profile to begin. The knowledge of functional hotspots can GPCR is amenable to bias or “biasable,” similar to assessing revolutionize this canonical hit-finding process. Using our the “druggability” of a target protein. Therefore, we asked the at ASPET Journals on October 1, 2021 computational method can be beneficial in two ways: 1) to question, which of the two receptors, b2AR or kOR, is more

Fig. 6. Heat map depicting the extent of bias in b2AR and in kOR. The thickness of the cartoon is directly proportional to the contribution of that residue to arrestin or G-protein bias. (A) b2AR in complex with formoterol (b-arrestin–biased). Regions of the receptor that are strongly b-arrestin–biased are represented by shades of blue to red. (B) kOR in complex with probe 1.1 (G-protein–biased). Regions of the receptor that are strongly G-protein–biased are represented by shades of blue to red. 296 Nivedha et al.

“biasable”? Figure 6 shows a heat map of the residue Nygaard R, Zou Y, Dror RO, Mildorf TJ, Arlow DH, Manglik A, Pan AC, Liu CW, b b Fung JJ, Bokoch MP, et al. (2013) The dynamic process of 2-adrenergic receptor contribution to the differential -arrestin and G-protein activation. Cell 152:532–542. allosteric communication pipelines in the formoterol-bound Rajagopal S, Ahn S, Rominger DH, Gowen-MacDonald W, Lam CM, Dewire SM, b k Violin JD, and Lefkowitz RJ (2011) Quantifying ligand bias at seven- active state of 2AR and the 1.1-bound active state of OR, transmembrane receptors. Mol Pharmacol 80:367–377. respectively. The heat map is an indication of the difference in Rajagopal S, Rajagopal K, and Lefkowitz RJ (2010) Teaching old receptors new tricks: biasing seven-transmembrane receptors. Nat Rev Drug Discov 9:373–386. the strength of allosteric communication by each residue Rasmussen SGF, DeVree BT, Zou Y, Kruse AC, Chung KY, Kobilka TS, Thian FS, toward b-arrestin in b2AR and toward G-protein in kOR. Chae PS, Pardon E, Calinski D, et al. (2011) Crystal structure of the b2 adrenergic receptor-Gs protein complex. Nature 477:549–555. The more red the residue and thicker the size of the backbone Shenoy SK, Drake MT, Nelson CD, Houtz DA, Xiao K, Madabushi S, Reiter E, Pre- cartoon in Fig. 6, the greater is its contribution to the ligand mont RT, Lichtarge O, and Lefkowitz RJ (2006) Beta-arrestin-dependent, bias. We observed that there are large clusters of residues on G protein-independent ERK1/2 activation by the b2 adrenergic receptor. J Biol Chem 281:1261–1273. TM5 and TM6 with high strength of allosteric communication Shukla AK, Singh G, and Ghosh E (2014) Emerging structural insights into biased toward the G-protein coupling interface in kOR. Such strong GPCR signaling. Trends Biochem Sci 39:594–602. Soergel DG, Subach RA, Burnham N, Lark MW, James IE, Sadler BM, Skobieranda functional hotspot residues are absent in the b2AR. This could F, Violin JD, and Webster LR (2014) Biased agonism of the m-opioid receptor by mean that kOR is probably more readily biasable compared TRV130 increases analgesia and reduces on-target adverse effects versus mor- phine: a randomized, double-blind, placebo-controlled, crossover study in healthy with b2AR. volunteers. Pain 155:1829–1835. In summary, our computational method to calculate the Stahl EL, Zhou L, Ehlert FJ, and Bohn LM (2015) A novel method for analyzing extremely biased agonism at G protein-coupled receptors. Mol Pharmacol 87: ligand bias and identify functional hotspot residues in GPCRs 866–877. Downloaded from is extensible to other class A GPCRs, which will greatly aid in Tschammer N, Bollinger S, Kenakin T, and Gmeiner P (2011) Histidine 6.55 is a major determinant of ligand-biased signaling in dopamine D2L receptor. Mol biased ligand design for GPCRs. Pharmacol 79:575–585. Urban JD, Clarke WP, von Zastrow M, Nichols DE, Kobilka B, Weinstein H, Javitch Authorship Contributions JA, Roth BL, Christopoulos A, Sexton PM, et al. (2007) Functional selectivity and classical concepts of quantitative pharmacology. J Pharmacol Exp Ther 320:1–13. Participated in research design: Nivedha, Tautermann, Vaidehi. Vaidehi N and Bhattacharya S (2016) Allosteric communication pipelines in Conducted experiments: Nivedha, Bhattacharya, Lee, Kollak, G-protein-coupled receptors. Curr Opin Pharmacol 30:76–83.

Vardy E, Mosier PD, Frankowski KJ, Wu H, Katritch V, Westkaemper RB, Aubé J, molpharm.aspetjournals.org Kiechle. Stevens RC, and Roth BL (2013) Chemotype-selective modes of action of k-opioid Performed data analysis: Nivedha, Tautermann, Bhattacharya, receptor agonists. J Biol Chem 288:34470–34483. Vaidehi. Violin JD, Crombie AL, Soergel DG, and Lark MW (2014) Biased ligands at G-protein-coupled receptors: promise and progress. Trends Pharmacol Sci 35: Wrote or contributed to the writing of the manuscript: Nivedha, 308–316. Tautermann. Wacker D, Wang C, Katritch V, Han GW, Huang X-P, Vardy E, McCorvy JD, Jiang Y, Chu M, Siu FY, et al. (2013) Structural features for functional selectivity at sero- – References tonin receptors. Science 340:615 619. Warne T, Edwards PC, Leslie AGW, and Tate CG (2012) Crystal structures of a Bhattacharya S, Salomon-Ferrer R, Lee S, and Vaidehi N (2016) Conserved mecha- stabilized b1-adrenoceptor bound to the biased agonists bucindolol and carvedilol. nism of conformational stability and dynamics in G-protein-coupled receptors. Structure 20:841–849. J Chem Theory Comput 12:5575–5584. Weiss DR, Ahn S, Sassano MF, Kleist A, Zhu X, Strachan R, Roth BL, Lefkowitz RJ, Bhattacharya S and Vaidehi N (2014a) Differences in allosteric communication and Shoichet BK (2013) Conformation guides molecular efficacy in docking screens pipelines in the inactive and active states of a GPCR. Biophys J 107:422–434. of activated b-2 adrenergic G protein coupled receptor. ACS Chem Biol 8: at ASPET Journals on October 1, 2021 Bhattacharya S and Vaidehi N (2014b) Mapping allosteric communication pipelines 1018–1026. in GPCRs from microsecond timescale molecular dynamics simulations. Biophys Wisler JW, DeWire SM, Whalen EJ, Violin JD, Drake MT, Ahn S, Shenoy SK, J 106:635a. and Lefkowitz RJ (2007) A unique mechanism of beta-blocker action: carvedilol Black JW and Leff P (1983) Operational models of pharmacological agonism. Proc R stimulates beta-arrestin signaling. Proc Natl Acad Sci USA 104:16657–16662. Soc Lond B Biol Sci 220:141–162. Wisler JW, Xiao K, Thomsen ARB, and Lefkowitz RJ (2014) Recent developments in Kahsai AW, Xiao K, Rajagopal S, Ahn S, Shukla AK, Sun J, Oas TG, and Lefkowitz biased agonism. Curr Opin Cell Biol 27:18–24. RJ (2011) Multiple ligand-specific conformations of the b2-adrenergic receptor. Nat Yao X, Parnot C, Deupi X, Ratnala VR, Swaminath G, Farrens D, and Kobilka B Chem Biol 7:692–700. (2006) Coupling ligand structure to specific conformational switches in the Kankan RN, Rao DR, Birari D, and Sawant AA (2010) inventors, Cipla Limited, assignee. b2-adrenoceptor. Nat Chem Biol 2:417–422. Process for preparing isomers of carmoterol. U.S. patent 8,236,959. 2012 Aug 7. Zhou L and Bohn LM (2014) Functional selectivity of GPCR signaling in animals. Kenakin T and Christopoulos A (2013) Signalling bias in new drug discovery: de- Curr Opin Cell Biol 27:102–108. tection, quantification and therapeutic impact. Nat Rev Drug Discov 12:205–216. Zhou L, Lovell KM, Frankowski KJ, Slauson SR, Phillips AM, Streicher JM, Stahl E, Kenakin T, Watson C, Muniz-Medina V, Christopoulos A, and Novick S (2012) A Schmid CL, Hodder P, Madoux F, et al. (2013) Development of functionally se- simple method for quantifying functional selectivity and agonist bias. ACS Chem lective, small molecule agonists at kappa opioid receptors. J Biol Chem 288: Neurosci 3:193–203. 36703–36716. Liu JJ, Horst R, Katritch V, Stevens RC, and Wüthrich K (2012) Biased signaling Zhou XE, Gao X, Barty A, Kang Y, He Y, Liu W, Ishchenko A, White TA, Yefanov O, pathways in b2-adrenergic receptor characterized by 19F-NMR. Science 335: Han GW, et al. (2016) X-ray laser diffraction for structure determination of the 1106–1110. rhodopsin-arrestin complex. Sci Data 3:160021. Luttrell LM, Maudsley S, and Bohn LM (2015) Fulfilling the promise of “biased” G protein-coupled receptor agonism. Mol Pharmacol 88:579–588. Mailman RB (2007) GPCR functional selectivity has therapeutic impact. Trends Address correspondence to: Dr. Nagarajan Vaidehi, Department of Pharmacol Sci 28:390–396. Molecular Immunology, Beckman Research Institute of the City of Hope, Neve K (2009) Functional Selectivity of G Protein-Coupled Receptor Ligands: New 1500 E. Duarte Road, Duarte, CA 91010. E-mail: [email protected] Opportunities for Drug Discovery, Humana Press, New York. Molecular Pharmacology

Identifying Functional Hotspot Residues for Biased Ligand Design in G-protein- coupled Receptors Anita K. Nivedha, Christofer S. Tautermann, Supriyo Bhattacharya, Sangbae Lee, Paola Casarosa, Ines Kollak, Tobias Kiechle, and Nagarajan Vaidehi

Department of Molecular Immunology, Beckman Research Institute of the City of Hope, 1500, E. Duarte Road, Duarte, CA – 91010. (A.N., S.B., S.L., N.V.) Dept. Medicinal Chemistry, Boehringer Ingelheim Pharma GmbH & Co. KG, Birkendorfer Straße 65, 88397 Biberach, Germany (C.T.) Dept. Imm. and Resp. Diseases Research, Boehringer Ingelheim Pharma GmbH & Co. KG, Birkendorfer Straße 65, 88397 Biberach, Germany (I.K., T.K.) Corp. Dept. Bus. Development & Licensing, C.H. Boehringer Sohn, Binger Str. 173, 55216 Ingelheim, Germany (P.C.)

Supplementary Information

Molecular Pharmacology

Materials and Methods

Computational Methods

Receptor Structure Preparation β2 Adrenergic Receptor (β2AR): The β2AR structure for our study was taken from the nanobody-stabilized active state crystal structure of the receptor in complex with an agonist (PDB ID: 3P0G) (Rasmussen et al., 2011). The starting structure for epinephrine, carvedilol, isoproterenol and carmeterol were obtained from PDB’s 4LDO (Ring et al., 2013), 4AMJ (Warne et al., 2012), 2Y03 and 2Y02 (Warne et al., 2011). The initial positions of formeterol and BI-167107 were obtained by making modifications to the structure of carmeterol from PDB 2Y02, and aligning the two receptors (from 2Y02 and 3P0G), while the starting structure for salbutamol was obtained by making modifications to the structure of isoproterenol from PDB 2Y03 and aligning the two receptors (from 2Y03 and 3P0G).

Serotonin Receptors: The three-dimensional structure for the two serotonin receptor serotypes, 5HT1B and 5HT2B, in complex with ergotamine, were obtained from PDBs 4IAR and 4IB4 (Wang et al., 2013), respectively. For both 5HT1B and 5HT2B, missing loops were added using the MMTSB Toolbox (Feig et al., 2004), followed by annealing the added loops keeping the rest of the protein structures fixed.

Κ-Opioid Receptor (κOR): The active state structure of κOR is not available but the structure of the inactive state of κOR with antagonist bound (PDB ID: 4DJH) (Wu et al., 2012) and the nanobody bound active state of μOR (PDB ID: 5C1M) (Huang et al., 2015) are available. In order to generate a homology model of the active-state κOR we first generated a chimeric template PDB by using the structures of the inactive κOR and the active μOR. TMs1-4, ECL2 and helix 8 were obtained from the inactive κOR and TMs5-7 were obtained from the active μOR by aligning the two receptors using PyMOL (“The PyMOL Molecular Graphics System, Version 1.6 Schrödinger, LLC.,”), and connecting the interface residues using the Draw mode in Maestro. A local minimization of the regions which were joined to create the chimeric template model (choosing ± 2 residues 151-156, 173-177, 289-293 at the three points of chimerization) was performed, followed by unrestrained minimization using MacroModel. Missing residues were added using Prime. Homology modeling was performed using MODELLER 9.0 (Šali and Blundell, 1993) and the top 20 models were analyzed. The best homology model was chosen based on the lowest RMSD (0.91 Å) to the chimeric template model.

The homology modeled active state structure of the κOR was passed through the Protein Preparation Wizard in Maestro ver. 9.1.207, followed by Receptor Grid Generation. A van der Waals scaling factor of 0.5 was used, while the partial charge cut-off was set to 1.0. The grid box was centered at the centroid of residues determined as important for binding based on mutation studies of the receptor (Zhou et al., 2013). The inner box dimension was set as 15x15x15Å.

Molecular Pharmacology

Ligand Preparation & Docking: The ligands for β2AR and κOR were built using Maestro ver. 9.1.207. They were minimized in MacroModel using the OPLS 2005 forcefield in MacroModel, with a distance-dependent dielectric. A single-point quantum calculation was performed using the Hartree-Fock theory with 6-31G** basis set and a PBF water solvation model using Jaguar (Bochevarov et al., 2013). Prior to docking, a conformational ensemble for each ligand was generated using the MacroModel Conformational Search. This ensemble of structures was then docked to the receptor using Standard Precision docking in Glide. A van der Waals scaling factor of 0.5 and a partial charge cut-off of 1.0 was used, and the output format requested was to write out at most 10 poses per ligand. The docked poses were evaluated based on the formation of critical interactions within the binding pocket, determined using mutation studies and based on similarity to conformations of comparable structures obtained from the Cambridge Structural Database. The interactions deemed as important for the binding of U69593 were with residues D1383.32, Y1192.64 and Y3207.43. For probe 1.1, it was D1383.32, Y1192.64, Y3137.36, I2627.39 and Y3207.43, and finally for probe 2.1 the residues selected were D1383.32, Y1192.64, Y3137.36 and Y3207.43 (Mansour et al., 1997; Kane et al., 2006, 2008, Frankowski et al., 2012; Martinez-Mayorga et al., 2013; Vardy et al., 2013).

Molecular Dynamics (MD) simulation protocol: All MD simulations were performed using the GROMACS ver. 5.1 package and GROMOS forcefield (Oostenbrink et al., 2004) at 300K. The receptor-agonist complex was placed in a POPC lipid bilayer box with 128 POPC and surrounded by water molecules and neutralizing Cl- ions. The topology file (.itp) and GROMACS file (.gro) for the ligand was obtained using the webserver PRODRG (Schüttelkopf and van Aalten, 2004). The solvent and solute were coupled to a temperature bath with a relaxation time of 0.1 ps. The pressure was calculated using a molecular virial and was held constant by weak coupling to a pressure bath with a relaxation time of 0.5 ps. For all the equilibration simulations, positional restraints were applied, and the simulations were performed at constant volume (NVT), during which, no pressure coupling was applied. The bond lengths and geometry of water molecules were constrained using the SHAKE algorithm. The equations of motion were integrated at a timestep of 2 fs and using the leapfrog algorithm. Center of mass motion was removed every 20 fs. The short-range van der Waals and electrostatic interactions were evaluated at every time step by using a charge-group pair list with a cutoff radius of 8 Å between the centers of geometry of the charge groups. Longer-range van der Waals and electrostatic interactions, between pairs at a distance greater than 8 Å and shorter than a long-range cutoff (14 Å), were evaluated every fifth time step, during which the pair list was also updated, and kept unchanged between updates. The receptor-agonist complexes were equilibrated under NVT conditions, for 10ns in the case of β2AR, β2ARTYY and serotonin receptors and for 25ns for the homology model of κOR. During the equilibration of the receptor−agonist complexes, the atoms of the complex were restrained in their positions using a harmonic restraining force with a force constant of 1000 kJ/(mol·nm) during the 200 ps of equilibration. In this step, the water molecules and lipid bilayer were allowed to move to optimize their packing around the receptor. The system was further equilibrated using NPT ensemble, while the force constant of the restraining force was set to 5 kcal/mol and reduced to zero stepwise at each 5 ns, and the Molecular Pharmacology

pressure coupling was switched on. This results in a total equilibration time of 40 ns. We performed an additional 10 ns of simulations without restraints before the production runs. Five independent simulations each to 200 ns were performed with different starting velocities amounting to a total of 1μs of total simulation time per system. The average representative structures from MD simulations were calculated by first performing conformation clustering on the combined MD simulation trajectories from the five simulations for any given system. The most representative structure of the most populated cluster was calculated as the frame that has the smallest RSMD to the center of this cluster of conformations. All MD simulation runs were analyzed for convergence by plotting the RMSDs of the Cα atoms of the residues in the transmembrane regions of the receptor and that of the ligand.

Analysis of the standard error in the ligand bias calculations The error in the calculation of the strength of the allosteric pipelines is dependent on two variables: (i) The number of snapshots used from the MD trajectories and (ii) the simulation time length of each MD simulation. We calculated the variation in the standard error in the pipeline strength as a function of the number of snapshots stored from MD simulations. We stored snapshots ranging from 5000 to 40,000 within the 200ns of MD simulation length. The standard error in the calculation of the strength of the allosteric communication pipeline using different number of snapshots is shown in Figure S2. Upon analysis of the standard error values, we concluded that storing snapshots from MD simulations every 20ps should be adequate for a 200ns length simulation time. We collect statistics from MD simulations by running multiple simulations rather than one long simulation. This provides a wider conformational sampling given the limitations of reaching ergodicity. To analyze how long of a simulation time should each MD simulation run be, we analyzed the standard error in the ratio of the strength of the allosteric pipelines to the beta-arrestin to that of the G-protein PRlig in equation (1) in the main text, as a function of simulation time as shown in Figure S3. The standard error is calculated as Std. error = Std. Dev./sqrt(N) where N = 5. Running individual simulations for 200ns shows lower standard errors in PRlig values calculated for several ligand-GPCR systems. To assess the sensitivity of the ligand bias calculations to the docked structure of the agonist we performed the following test. We performed MD simulations following the same protocol starting from a docked structure of carmoterol that is 0.9Å away from the crystal structure of carmoterol used in this study. Additionally, we intentionally chose a carmoterol docked structure that did not have contacts between the catechol group and all the three serines, S203, S204 and S207, of the receptor as seen in the crystal structure of carmoterol bound β-1-adrenergic receptor. The calculated ligand bias is 0.1395 with mean error of 0.016. The ligand bias value calculated previously using MDs starting from the crystal structure was 0.14. Thus, the docked structure leads to ligand bias that is similar to that of the crystal structure.

Calculation of the residue-based allosteric hub score We calculated the extent of contribution to the allosteric pipelines to both the - arrestin and G-protein coupling interfaces by each residue in the GPCR. We define this as the allosteric hub score of each residue. The allosteric hub score is the simply the number of allosteric communication pathways passing through each residue in the GPCR either to Molecular Pharmacology

the -arrestin or G-protein coupling interface. For more details see reference Bhattacharya 2014, and Bhattacharya 2016.

Molecular Pharmacology

Table S1: List of all systems used in the current study, including the Computational Ligand Bias values and details about the studies from which the Experimental Bias Factors were obtained.

Reference Biased Computational System Description Experimental Ref. Receptor Agonist Receptor Agonist Ligand Bias Bias β2AR WT isoproterenol 0.12 β2AR WT epinephrine β2AR WT formoterol β-arrestin 0.21 [ref] • Same WT receptor β2AR WT salbutamol -0.02 • Different β-arrestin biased ligands β2AR WT isoproterenol β2AR WT BI-167107 β-arrestin 0.04 [ref] β2AR WT formoterol 0.08 β2AR WT isoproterenol β-arrestin SI Table S3 β2AR WT carmeterol 0.14 • Same WT receptor κOR WT probe 1.1 -0.13 • Different G-protein κOR WT u69593 G-protein biased ligands κOR WT probe 2.1 -0.04 • Different receptor serotypes 5HT1B ergotamine 5HT2B ergotamine β-arrestin 0.32 [ref] • Same ligand • WT and Mutant receptors β2AR WT epinephrine β2AR TYY epinephrine β-arrestin 0.40 Unavailable • Same ligand

Molecular Pharmacology

Table S2: The list of residues in the GPCR β-arrestin interface and G-protein interface that we have used in this study for calculating the allosteric pipelines for the different GPCRs. G-protein Extracellul β-arrestin interface interface Binding Site -ar Region System residues (BAI) residues residues (BS) residues (GPI) (EC) 63, 64, 65, 66, 131, 134, 135, 136, 131, 134, 135, 136, 93, 109, 110, 113, 114, 117, 30-41, 88-110, β-2AR 137, 138, 139, 140, 142, 143, 146, 138, 139, 141, 142, 165, 169, 191, 192, 193, 195, 162-204, 291- and 222, 225, 271, 274, 278, 329, 331 143, 222, 225, 226, 200, 203, 204, 207, 208, 286, 313 TYY 229, 230, 232, 233, 289, 290, 293, 294, 297, 305, β-2AR 239, 271, 274, 275 308, 309, 312, 316 35, 38, 40, 41, 70, 73, 102, 105, 106, 102, 105, 106, 107, 8, 9, 12, 54, 55, 57, 58, 61, 62, 1-9, 60-79, 135- 107, 108, 109, 110, 112, 113, 114, 109, 110, 112, 113, 64, 65, 66, 68, 70, 80, 81, 82, 173, 238-259 195, 198, 199, 202, 203, 205, 209, 114, 193, 196, 197, 83, 84, 85, 87, 88, 156, 158, κ-OR 210, 211, 213, 214, 221, 222, 280, 200, 201, 203, 204, 172, 173, 176, 177, 232, 233, 281, 282 210, 217, 220, 221 235, 236, 237, 239, 240, 243, 255, 258, 259, 262, 265, 266, 268, 269 79, 80, 82, 84, 85, 147, 150, 151, 147, 150, 151, 152, 38, 39, 109, 110, 125, 126, 129, 38-49, 96-118, 152, 153, 154, 155, 156, 158, 159, 157, 158, 159, 231, 130, 133, 134, 199, 200, 201, 170-212, 340- 5HT1B 161, 162, 231, 234, 237, 238, 239, 234, 235, 238, 239, 212, 213, 217, 327, 330, 331, 362 306, 307, 310, 311, 312, 313, 314, 313, 314, 317 334, 351, 352, 355, 359 315, 316, 370, 375 84, 85, 87, 42, 43, 153, 156, 157, 153, 156, 157, 158, 114, 115, 121, 131, 132, 135, 48-60, 108-132, 158, 159, 160, 161, 162, 163, 164, 160, 161, 162, 163, 136, 139, 140, 207, 208, 209, 186-221, 342- 165, 166, 170, 240, 243, 244, 246, 240, 243, 244, 247, 217, 218, 221, 222, 225, 337, 362 5HT2B 247, 248, 314, 315, 316, 317, 318, 323, 326, 327 340, 341, 344, 347, 359, 362, 319, 320, 321, 322, 323, 324, 325, 363, 366, 370 326, 380, 383, 384, 386

Molecular Pharmacology

Table S3: Experimental bias factors for agonists formeterol and carmeterol using reference agonist isoproterenol. We used the Black-Leff model to calculate the bias factors as described in the Methods section of the main text. The error bars correspond to 95% confidence limits corresponding to 2 tailed t-test. The number of experiments is n=3.

Reference ligand Biased Experimental Error Experimental Bias Factor Receptor Agonist Receptor Agonist Negative Positive

formeterol 1.06 isoproterenol β2AR 0.43 0.55 β2AR WT WT carmeterol 2.14 0.77 1.01

Molecular Pharmacology

Figure S1: Two-dimensional representation of the agonists studied here

Molecular Pharmacology

0.25

0.20

Frame Rate No. of

i l

R (200ns) Snapshots P

/ 0.15

r 5ps 40000

r r

e 10ps 20000

d 0.10 t

S 20ps 10000 40ps 5000 0.05

0.00 5000 10000 20000 40000 epinephrine carvedilol formoterol isoproterenol carmoterol salbutamol bi-167107 tyy

Figure S2: Variation of mean error in allosteric pipeline strength as a function of the number of snapshots from MD simulations. Molecular Pharmacology

0.14"

0.12"

r 0.1"

r E

0.08"

d t

S 0.06"

0.04"

0.02"

0" 50ns" 100ns" 150ns" 200ns" epinephrine" carvedilol" formoterol" isoproterenol" carmoterol" salbutamol" bi<167107" tyy"

Figure S3: Plot of the standard error in the ratio of the strength of the allosteric pipelines to the beta-arrestin to that of the G-protein PRlig in equation (1) in the main text, as a function of simulation time. Std. error = Std. Dev./sqrt(N); N = 5. Running individual simulations for 200ns shows lower standard errors in PRlig values calculated for several ligand-GPCR systems.

Molecular Pharmacology

A 0.50 B 250 0.40 225 200 0.40 0.30 175 0.30 150 125 0.20

100 Bias 0.20 Factors 75 0.10 50

0.10 Theoretical Bias Theoretical Experimental Bias Experimental 25

0 0.00 Ligand Computational 0.00 5HT1B 5HT2B

Figure S4: A. Calculated Ligand Bias for the β2ARTYY mutant in complex with epinephrine compared to that of the WT β2AR in complex with epinephrine. There is no experimental quantitative bias factor available for epinephrine in the β2ARTYY mutant. B. Comparison of computational ligand bias values to experimental bias factors for ergotamine in the two serotonin receptors. The bias factors were taken from reference (Wacker et al., 2013).

Molecular Pharmacology

Epi 3.32 A Asp-113 Form 50 Ser-2035.42 50 40 40

30 30

20 20

10 10

0 0 Populatio Distribution Populatio Population Distribution Population 0 60 120 180 240 300 360 0 60 120 180 240 300 360 -10 -10 Χ1 (deg) Χ1 (deg)

50 50 Ser-2045.43 Phe-193(ECL2) 40 40

30 30

20 20

10 10

0 0 Population Distribution Population 0 60 120 180 240 300 360 Distribution Population 0 60 120 180 240 300 360 -10 -10 Χ (deg) Χ (deg) 1 2

Figure S5: A. Side-chain rotameric angle distribution of residues altering the binding site shape for formoterol vs. epinephrine. The residues F193 and S204 are functional hotspot residues. These residues are common to epinephrine and formeterol that shape the binding site in β2AR. Molecular Pharmacology

Formoterol B D192ECL2 Y3087.35 N2936.55

S2035.42 F193ECL2

5.43 7.39 S204 I3097.36 N312

S2075.46 W1093.28

T1103.29 D1133.32 T1183.37 V1173.36

Figure S6: Three-dimensional orientation of the formeterol binding site in 2AR. The residues shown are present in the formoterol binding cavity in WT β2AR with contacts for over 75% of the MD simulation snapshots. The residues that show differences in rotamer populations between the unbiased epinephrine and biased formeterol are also shown (F193, S203, S204, D113). The agonist, formoterol, is shown in ball and stick representation. The connection to the functional hotspot residues are shown by dashed lines.

Molecular Pharmacology

Figure S7: Side-chain rotameric angle distribution of residues altering the binding site shape for U69593 vs. Probe 1.1. These residues are additionally functional hotspot residues in 1.1.

Molecular Pharmacology

Figure S8: The residues in the probe 1.1 binding cavity in KOR that contact the agonist for over 75% of the MD simulation are shown. The residues that show differences in rotamer populations between the unbiased U69593 and biased probe 1.1 are also shown (Y65, D84). The agonist probe 1.1, is shown in ball and stick representation. Y65, I240, Y258, K173 and D84 are binding site hotspot residues in probe 1.1, and dashed lines show the connections to these residues.

Molecular Pharmacology

References: Bochevarov A, Harder E, Hughes TF, Greenwood JR, Braden DA, Philipp DM, Rinaldo D, Halls, MD, Zhang J, and Friesner RA (2013) Jaguar: A high-performance quantum chemistry software program with strengths in life and materials sciences. J Quantum Chem 113:2110-2142. Feig M, Karanicolas J, and Brooks C (2004) MMTSB Tool Set: enhanced sampling and multiscale modeling methods for applications in structural biology. J Mol Graph Model 22:377-395. Frankowski KJ, Hedrick MP, Gosalia P, Li K, Shi S, Whipple D, Ghosh P, Prisinzano TE, Schoenen FJ, Su Y, Vasile S, Sergienko E, Gray W, Hariharan S, Milan L, Heynen- Genel S, Mangravita-Novo A, Vicchiarelli M, Smith LH, Streicher JM, Caron MG, Barak LS, Bohn LM, Chung TDY, and Aubé J (2012) Discovery of small molecule kappa opioid receptor agonist and antagonist chemotypes through a HTS and Hit refinement strategy. ACS Chem Neurosci 3:221–236. Huang W, Manglik A, and Venkatakrishnan A (2015) Structural insights into μ-opioid receptor activation. Nature 524:315-321. Kane BE, McCurdy CR, and Ferguson DM (2008) Toward a structure-based model of salvinorin A recognition of the kappa-opioid receptor. J Med Chem 51:1824–1830. Kane BE, Nieto MJ, McCurdy CR, and Ferguson DM (2006) A unique binding epitope for salvinorin A, a non-nitrogenous kappa opioid receptor agonist. FEBS J 273:1966– 1974. Mansour a, Taylor LP, Fine JL, Thompson RC, Hoversten MT, Mosberg HI, Watson SJ, and Akil H (1997) Key residues defining the mu-opioid receptor binding pocket: a site-directed mutagenesis study. J Neurochem 68:344–353. Martinez-Mayorga K, Byler KG, Yongye AB, Giulianotti MA, Dooley CT, and Houghten RA (2013) Ligand/kappa-opioid receptor interactions: Insights from the X-ray crystal structure. Eur J Med Chem 66:114–121. Oostenbrink C, Villa A, Mark AE, and Van Gunsteren WF (2004) A biomolecular force field based on the free enthalpy of hydration and solvation: The GROMOS force-field parameter sets 53A5 and 53A6. J Comput Chem 25:1656–1676. Rasmussen SGF, Choi H-J, Fung JJ, Pardon E, Casarosa P, Chae PS, Devree BT, Rosenbaum DM, Thian FS, Kobilka TS, Schnapp A, Konetzki I, Sunahara RK, Gellman SH, Pautsch A, Steyaert J, Weis WI, and Kobilka BK (2011) Structure of a nanobody-stabilized active state of the β(2) adrenoceptor. Nature 469:175–80. Ring A, Manglik A, Kruse A, Enos M, and Weis W (2013) Adrenaline-activated structure of the β2-adrenoceptor stabilized by an engineered nanobody. Nature 502:575-579. Šali A, and Blundell T (1993) Comparative protein modelling by satisfaction of spatial restraints. J Mol Biol 234:779-815. Schüttelkopf AW, and van Aalten DMF (2004) PRODRG : a tool for high-throughput crystallography of protein–ligand complexes. Acta Crystallogr Sect D Biol Crystallogr 60:1355–1363. The PyMOL Molecular Graphics System, Version 1.6 Schrödinger, LLC. Vardy E, Mosier PD, Frankowski KJ, Wu H, Katritch V, Westkaemper RB, Aubé J, Stevens RC, and Roth BL (2013) Chemotype-selective modes of action of kappa- opioid receptor agonists. J Biol Chem 288:34470–34483. Wacker D, Wang C, Katritch V, Han GW, Huang X-P, Vardy E, McCorvy JD, Jiang Y, Molecular Pharmacology

Chu M, Siu FY, Liu W, Xu HE, Cherezov V, Roth BL, and Stevens RC (2013) Structural Features for Functional Selectivity at Serotonin Receptors. Science 340:615-619. Wang C, Jiang Y, Ma J, Wu H, and Wacker D (2013) Structural basis for molecular recognition at serotonin receptors. Science 340:610-614. Warne T, Edwards P, Leslie A, and Tate C (2012) Crystal structures of a stabilized β1- adrenoceptor bound to the biased agonists bucindolol and carvedilol. Structure 20:81- 849. Warne T, Moukhametzianov R, Baker J, and Nehmé R (2011) The structural basis for agonist and partial agonist action on a β1-adrenergic receptor. Nature 469:241-244. Wu H, Wacker D, Katritch V, Mileni M, Han G, and Vardy E (2012) Structure of the human kappa opioid receptor in complex with JDTic. Nature 485:327-332. Zhou L, Lovell KM, Frankowski KJ, Slauson SR, Phillips AM, Streicher JM, Stahl E, Schmid CL, Hodde P, Madoux F, Cameron MD, Prisinzano TE, Aubé J, and Bohn LM (2013) Development of functionally selective, small molecule agonists at kappa opioid receptors. J Biol Chem 288:36703-36716.