Large library docking identifies positive allosteric modulators of the calcium-sensing receptor

Abstract

Positive allosteric modulator (PAM) drugs enhance the activation of the calcium-sensing receptor (CaSR) and suppress parathyroid hormone (PTH) secretion. Unfortunately, these hyperparathyroidism-treating drugs can induce hypocalcemia and arrhythmias. Seeking improved modulators, we docked libraries of 2.7 million and 1.2 billion molecules against the CaSR structure. The billion-molecule docking found PAMs with a 2.7-fold higher hit rate than the million-molecule library, with hits up to 37-fold more potent. Structure-based optimization led to nanomolar leads. In ex vivo organ assays, one of these PAMs was 100-fold more potent than the standard of care, cinacalcet, and reduced serum PTH levels in mice without the hypocalcemia typical of CaSR drugs. As determined from cryo–electron microscopy structures, the PAMs identified here promote CaSR conformations that more closely resemble the activated state than those induced by the established drugs.

Graphical Abstract

INTRODUCTION:

Even before the advent of molecular pharmacology, efforts were directed toward developing calcimimetic and calcilytic drugs to regulate parathyroid hormone (PTH) secretion and blood calcium levels. The activity of these drugs on the calcium-sensing receptor (CaSR), a G protein–coupled receptor, was confirmed after its cloning. CaSR is present in almost every organ system and is highly expressed in the parathyroid glands and kidneys, where it maintains calcium homeostasis by regulating PTH secretion and renal calcium reabsorption and excretion. Loss or gain of function in CaSR results in conditions such as familial hypocalciuric hypercalcemia, primary hyperparathyroidism, and autosomal dominant hypocalcemia. CaSR is also involved in nutrient sensing, vascular tone, and insulin secretion, with implications in osteoporosis and several cancers.

RATIONALE:

Present CaSR drugs, which act as positive allosteric modulators (PAMs) of CaSR, can cause side effects such as hypocalcemia, gastrointestinal problems, hypotension, and adynamic bone disease. Our recent high-resolution cryo–electron microscopy (cryo-EM) studies showed that in the active-state receptor, two CaSR PAMs, cinacalcet and the related evocalcet, adopt an “extended” conformation within the seven-transmembrane domain (7TM) of one CaSR monomer and a “bent” conformation in the second monomer of the dimer. These different conformations reflect asymmetric changes in the two allosteric PAM sites upon receptor activation. To exploit these structures, we used library docking to discover new CaSR PAMs that are topologically unrelated to those previously investigated. Our hope was that diverse chemotypes explored by the large library screen would yield PAMs with new pharmacology, potentially modulating PTH secretion without causing the dose-limiting hypocalcemia of the approved drugs.

RESULTS:

Seeking new modulators, we docked libraries of 2.7 million and 1.2 billion molecules against the CaSR structure. The billion-molecule docking found new PAMs with a 2.7-fold higher hit rate than the million-molecule library, with hits up to 37-fold more potent. Structure-based optimization led to nanomolar leads. In ex vivo organ assays, one of these new PAMs is 100-fold more potent than the standard of care, cinacalcet, and reduces serum PTH levels in mice without the hypocalcemia typical of CaSR drugs. Cryo-EM structures show that the new PAMs promote CaSR dimer conformations that more closely resemble the activated state than those induced by the established drugs, perhaps contributing to their high potency and new pharmacology.

CONCLUSION:

Our study demonstrates that expanding docking libraries improves the discovery of new CaSR PAMs. By leveraging high-resolution cryo-EM structures and large library docking, we identified previously unknown chemotypes that can modulate PTH secretion, apparently without inducing hypocalcemia. This approach offers a promising strategy for developing safer and more effective drugs for conditions such as secondary hyperparathyroidism and parathyroid cancer. Our findings highlight the importance of library size in molecular docking and provide a foundation for discovering therapeutic modulators for other challenging targets.

Large library docking identifies PAMs of the CaSR. Docking a library of 1.2 billion molecules led to a higher hit rate and more potent CaSR ligands than docking a library of 2.7 million molecules (top). Compound ‘54149 promotes a newly observed conformation of the CaSR dimer (bottom left). Compound ‘54149 is more potent in suppressing PTH secretion than present CaSR calcimimetic drugs but does not induce hypocalcemia in mice (bottom middle and right). sPTH, serum PTH.

Introduction

Well before the advent of molecular pharmacology, much effort had been directed toward developing calcimimetic and calcilytic drugs to promote or suppress the calcium-sensing abilities of parathyroid cells and to regulate parathyroid hormone (PTH) secretion and blood calcium levels. The activity of these drugs on the calcium-sensing receptor (CaSR), a G protein–coupled receptor (GPCR), was confirmed after its cloning (1). CaSR is present in almost every organ system but is most highly expressed in the parathyroid glands and in the kidneys, where it maintains calcium homeostasis by sensing changes in extracellular calcium levels to regulate PTH secretion, renal calcium reabsorption, and excretion (2, 3). Loss or gain of function in CaSR results in familial hypocalciuric hypercalcemia, primary hyperparathyroidism, and autosomal dominant hypocalcemia (4–8). Through its widespread expression, CaSR is also involved in other physiological mechanisms, notably gastrointestinal nutrient sensing, vascular tone, and secretion of insulin, with alterations in receptor activity implicated in the development of osteoporosis and in several cancers (3).

Efforts to target CaSR therapeutically have focused on the development of positive and negative allosteric modulators (PAMs and NAMs), which potentiate the receptor’s activation or its inactivation, respectively, while binding at a nonorthosteric site (here, a noncalcium site). PAMs enhance the physiological response to calcium but display little or no agonist activity on their own. In the past two decades, the small-molecule PAM drug cinacalcet and the peptide-based PAM drug etelcalcetide (9) were approved for human use, but only for the treatment of secondary hyperparathyroidism (SHPT) in patients with chronic kidney disease who were undergoing dialysis (usually stage 5), although cinacalcet is also approved to treat high levels of calcium in patients with parathyroid cancer. The limited indications reflect the side effects associated with the present PAMs, including hypocalcemia, gastrointestinal problems, hypotension, and adynamic bone disease (10). Hypocalcemia is life-threatening because it can cause seizures and heart failure (11–16). Chronic kidney disease affects more than 10% of people worldwide and, considering the prevalence of SHPT (10, 17–21), drugs that decrease PTH levels without causing hypocalcemia are much needed.

CaSR belongs to the family C of GPCRs, a relatively exotic group of receptors that operate as homo- or heterodimers with extracellular domains (ECDs) constituting the orthosteric ligand binding site. The ECD of a CaSR monomer is connected through a linker region to the seven-transmembrane domain (7TM), which primarily signals through G_q/11 and G_i/o G protein subtypes (22, 23). Upon calcium binding to the ECDs, the CaSR homodimer undergoes extensive conformational transitions that bring the 7TM bundles into close proximity through a TM6-TM6 interface, a configuration that is associated with receptor coupling to a G protein (24, 25). PAMs bind to the 7TM bundles and enhance the receptor’s sensitivity to calcium (Fig. 1A). Our recent high-resolution cryo–electron microscopy (cryo-EM) studies showed that in the active-state receptor, cinacalcet and the related evocalcet, which was recently approved for therapeutic use in Japan (23), both adopt an “extended” conformation within the 7TM of one CaSR monomer and a “bent” conformation in the second monomer of the dimer. The two different conformations adopted by the same ligands reflect asymmetric changes in the two allosteric PAM sites upon activation (23).

We sought to exploit these structures by adopting a library docking approach to discover new CaSR PAMs that are topologically unrelated to those previously investigated (26). In the past 5 years, large library docking has found potent ligands, with activities often in the mid- to low-nM concentration range, for a wide range of targets (26–32). Our hope was that from diverse chemotypes explored by the large library screen, PAMs with new pharmacology would emerge—those that could potentially modulate PTH secretion without leading to the dose-limiting hypocalcemia of the approved drugs.

If our first priority was discovering new CaSR chemotypes with new pharmacology, a second interest was to test how library size affects docking experimental results. A feature driving the success of recent docking campaigns is thought to be the expansion of docking libraries, which have grown more than 1000-fold, from millions to billions of molecules. Indeed, simulations have suggested that as the libraries expand, more molecules with high docking scores are found and better scores are reached (33). Although this pattern is encouraging, it has not been tested experimentally, and there are reasons why library expansion might have little effect on docking results or even make them worse. Docking scoring functions are notoriously approximate and leave out important terms (34, 35). It is easy to imagine that as our libraries expand, more and more molecules will exploit these gaps and act as artifacts. Indeed, several studies suggest that this does occur as libraries grow (26, 29, 33). Also, because only a thin slice of the very-best-ranked molecules are tested in docking (often fewer than 100 molecules from billions computationally screened), unless docking can reliably prioritize the better library molecules to the top of the ranked list, it is not certain that larger libraries will lead to better results. Finally, it is conceivable that the libraries are growing in number of molecules and even in molecular diversity, but, in terms of pharmacophores, the range of molecules might be much less. One might thus be able to achieve the same hit rates and affinities with a small library as with a much larger one. To experimentally test the proposition that larger libraries lead to better results in molecular docking, we compared the in vitro hit rates and hit affinities of molecules prioritized from docking a 2.7-million-molecule library versus docking a library of 1.2 billion molecules against CaSR.

Results

Docking smaller and larger libraries against CaSR for new PAMs

We began by docking the smaller, in-stock library of 2.7 million molecules at both 7TM sites of CaSR (Fig. 1A and fig. S1A). In the site accommodating the bent conformation of cinacalcet (7TM^B site), an average of 3927 orientations of each library molecule were sampled, each in an average of 330 conformations, or 1.2 trillion configurations overall; the calculation took just under 1 hour of elapsed time on a 1000-core cluster by using DOCK3.7 (36). Molecules were scored for van der Waals (37) and Poisson-Boltzmann–based electrostatic complementarity (38) corrected for generalized-Born ligand desolvation (39). Conformationally strained molecules were deprioritized (40), high-ranking molecules were clustered for similarity to each other using an ECFP4-based Tanimoto coefficient (Tc) of 0.5, and molecules were also filtered against similarity to known CaSR ligands. Comparable numbers of ligand orientations, conformations, and docking configurations were sampled and calculated for the extended site (7TM^A site). From among the top 1% of the docking-ranked molecules, we ultimately selected 26 and 22 compounds from docking at the 7TM^A and 7TM^B sites, respectively (fig. S1, A and E). These molecules were tested in cell-based bioluminescence resonance energy transfer (BRET) assays for CaSR-induced G_i3 activation (41) using an extracellular calcium concentration of 0.5 mM. One PAM emerged from those selected for the 7TM^A site, with >10% of maximal efficacy (E_max) induced by cinacalcet, and three PAMs were found for the 7TM^B site (Fig. 1B and table S1), representing hit rates of 3.8% (1/26) and 13.6% (3/22), respectively (fig. S1A). The higher hit rate for the 7TM^B site may reflect its more enclosed pocket, which better excluded molecules unlikely to bind and led to better ligand complementarity. As a technical aside, we note that all 7TM^B hits would have ranked well in the 7TM^A site (two in the top 0.25% and a third in the top 2.5% of the docking-ranked library), and the 7TM^A hit, although it received a favorable score in the 7TM^B site, would have only ranked among the top 18% of the docking-ranked list for that site.

Fig. 1. Ligands identified from the in-stock and large library screens targeting the 7TM sites of CaSR.

(A) Larger-scale docking against the 7TM^B site of CaSR resulted in a higher hit rate (13.6% in the 2.7-million-molecule docking campaign versus 36.5% in the 1.2-billion-molecule docking campaign). Hit rates were defined by more than 10% BRET response compared with cinacalcet at 100 μM. The overall structure of the homodimeric CaSR is shown on the left, highlighting the binding sites for extracellular ligands—calcium (orthosteric ligand), tryptophan, and phosphate [PDB ID 7M3G (23)]. PAMs bind in transmembrane domains in different fashions. On the right, a zoomed-in view of the 7TM^B site shows evocalcet (an example of a PAM, shown in mauve) adopting a bent conformation. C, Cys. (B) BRET response (normalized to cinacalcet) of the initial hits at 100 μM (mean ± SEM of one to three biological replicates). (C) Hit-rate comparison of the 2.7-million- and 1.2-billion-molecule screens with different affinity definitions. The overall hit rate of the 1.2-billion-molecule screen is significantly better than the in-stock 2.7-million-molecule screen (p < 0.05 by z test). (D) Total docking energies of top-scoring molecules out of the large-scale docking screen compared with the in-stock screen (only molecules with DOCK scores <−35 kcal/mol are plotted). (E) Examples of the docking hits in comparison to the known PAM drugs cinacalcet and evocalcet (colors represent the different moieties fulfilling the same role). Docked poses of the new representative PAMs at the 7TM^B site are shown.

To measure the impact of larger libraries in our search for more potent PAMs, we screened a library of 1.2 billion make-on-demand (“tangible”) molecules (42) against the more enclosed 7TM^B site (Fig. 1A). This calculation took about 16 days of elapsed time on a 1000-core cluster. Top-scoring molecules were again filtered and clustered as described above, with 1002 cluster-heads passing all criteria. These molecules were further docked against the 7TM^A site, and 96 molecules that ranked well in both sites were prioritized for synthesis from the tangible space. Of these, 74 were successfully made, a 77% fulfillment rate (Fig. 1A).

Upon experimental testing, 27 of the 74 molecules produced >10% of the E_max induced by cinacalcet, a 36.5% hit rate that was almost threefold higher than the hit rate from the 2.7-million-molecule library (Fig. 1, B to C). The hits from the larger library were also more potent than those from the smaller library, with more than 70% having median effective concentration (EC₅₀) values better than 10 μM and 20% having an EC₅₀ better than 1 μM (Fig. 1C and table S1); the most potent had an EC₅₀ of 273 nM. Given the number of molecules tested, the hit-rate difference between the larger and smaller library screens was significant (p < 0.05) and, indeed, is only as good as it is for the smaller library when we count as hits molecules with EC₅₀ values worse than 10 μM. In the 1 to 10 μM and in the 0.1 to 1 μM ranges, no hits emerged from the smaller library. These results provide experimental support for simulations predicting better performance from larger libraries (33) (Fig. 1C). Correspondingly, the best-scoring molecules from the larger library had better scores, reaching the −50 range, than did those from the smaller library (−45 range of scores) (Fig. 1D). We do note a deviation from normality on the better-scoring side of the large library screen to even better scores—this likely reflects a thin wedge of molecules that “cheat” our scoring function, something we have observed before (29, 33).

The ultralarge library was also better at providing more new chemotypes than emerged from the small library screen. Whereas the compounds identified from the in-stock screens were topologically distinct from known ligands in the ChEMBL (chemical database of bioactive molecules with drug-like properties) and International Union of Basic and Clinical Pharmacology (IUPHAR) databases, with ECFP4 (extended-connectivity fingerprints with a diameter of four bonds) Tc’s less than 0.35 (table S1), they nevertheless exhibited physical similarities to established PAMs. These similarities include a buried aromatic ring, a bridging methylamine linker, and a distal aromatic ring (Fig. 1E). Conversely, the PAMs from the ultralarge library showcased a diverse range of heteroaromatic anchors, various linker types, or even the absence of a linker, as in the case of compound Z5208267909 (‘7909) (fig. S1F and table S1). More quantitatively, each of the PAMs from the large library screen represented a different Bemis-Murcko scaffold; although this was also true for the actives from the in-stock library screen, only three of these were active, so from a scaffold-centric view, the actives from the large library screen covered far more ground. By ECFP4-based Tc similarity, the average pairwise similarity within each set was essentially random for each screen, at 0.24 for the large library PAMs and 0.15 for the in-stock PAMs (fig. S1G). Tc similarities of pairs of molecules between the sets was typically lower than 0.16, attesting to not only the high diversity of the large library PAMs but also their topological divergence from the smaller, in-stock library actives. Notably, many of these compounds, such as Z1895745670 (‘5670) and Z1591490522 (‘0522), lacked the methyl group adjacent to the cation, a feature commonly found in CaSR PAMs (Fig. 1F). If we compare the new PAMs to known ligands not by topological similarity but by three-dimensional (3D) and functional group similarity using rapid overlay of chemical structures (ROCS) (43), similarity scores spanned a range from 0.84 to 1.15, broadly consistent with their topological dissimilarity to the previously known ligands (fig. S1D).

Our recent cryo-EM structures have shown that the cationic amine of the drugs cinacalcet and evocalcet hydrogen-bonds and ion-pairs with Q681^3.33 and E837^7.32 of CaSR (Q, Gln; E, Glu), interactions that are critical for PAM recognition (23, 44). Meanwhile, the highly conserved methyl group α to this cationic amine fits into a hydrophobic pocket formed by I841^7.36, F684^3.36, and F668^2.56 (I, Ile; F, Phe); alanine substitutions to these residues abolish or much decrease binding affinities for CaSR PAMs (45). In their bent conformations bound to 7TM^B, the naphthalene ring common to both drugs T-stacks with F684^3.36 and W818^6.50 (W, Trp), and the benzene forms an edge-to-π interaction with W818^6.50. Notably, although their linker lengths differ from those of the known drugs, most of the new PAMs also adopt bent conformations in their docked poses within the 7TM^B pocket (Fig. 1E and fig. S1F). Although most of them retain hydrophobic flanking groups that dock into the aryl sites defined by cinacalcet and evocalcet, all do so with different moieties (Figs. 1E and 2A).

Fig. 2. Initial hits to high-affinity analogs.

(A) Contact analysis of the initial docking hits versus cinacalcet. Hydrogen bonds (distance ≤3.2 Å, donor angle ≥120°, acceptor angle ≥90°), stacking (aromatic groups face to face or face to edge), salt bridges (electrostatic interaction, oppositely charged residues within 3.5 Å), and van der Waals interactions (attractions between atoms within a 3.5 Å cutoff) are shown. G, Gly; R, Arg; T, Thr. (B) 1.2-billion-molecule screen docking hit ‘5250 (three biological replicates) and its optimized analog ‘2021 (a diastereomer of ‘6783; three biological replicates). (C) 1.2-billion-molecule screen docking hit ‘5670 (three biological replicates) and its optimized analog ‘2460 (an enantiomer of ‘6218; four biological replicates). (D) 1.2-billion-screen docking hit ‘0522 (mean ± SEM of three biological replicates) and its optimized analog ‘5526 (three technical replicates). (E) 2.7-million-molecule screen docking hit ‘21374 (three technical replicates) and its optimized analog ‘54149 (two biological replicates). EC₅₀ was determined by monitoring G_i activation by CaSR upon compound addition at [Ca²⁺] = 0.5 mM. The efficacy of the compounds is normalized to the maximum BRET response induced by cinacalcet. The EC₅₀ of cinacalcet is 71 nM (48 to 106 nM; 12 biological replicates). In (B) to (D), error bars represent the SEM.

Structure-based optimization of new PAMs

A core goal of this study was to find chemo-types that confer distinctive pharmacological outcomes. We therefore prioritized high-ranking docked molecules on the basis of both potency and topological dissimilarity to known CaSR PAMs for further optimization. To increase the affinity of the initial hits, we sought to optimize interactions with residues that had proven important in other series (44, 45), including F668^2.56, Q681^3.33, E837^7.32, I841^7.36, F684^3.36, and W818^6.50, through medicinal chemistry followed by docking (Fig. 2A). The greater polarity of the docking hits, whose calculated octanol:water partition coefficient (cLogP) ranged from 2.3 to 4.0 versus a cLogP of 5.1 for the more hydrophobic cinacalcet, gave us freedom to operate in the hydrophobic CaSR site.

To fill a gap in the interface with L773^5.40 and Y825^6.57 (L, Leu; Y, Tyr) and to stiffen the linker in the docking-derived PAM Z1420545250 (‘5250) (EC₅₀ 444 nM), a second methyl was added proximal to the cationic nitrogen. This change improved potency fivefold, to an EC₅₀ of 90 nM, and synthetic resolution of the diastereomers improved it another 130-fold. The resulting compound Z8554052021 (‘2021; stereo-chemically purified but as yet not defined), with an EC₅₀ of 3.3 nM, is among the most potent CaSRs and indeed GPCR PAMs of which we are aware (Fig. 2B).

The tetrahydrobenzapine of compound ‘5670 (Fig. 2C) separates it from the naphthalene equivalent of cinacalcet and evocalcet and gives it a relatively polar and 3D character versus the equivalent groups of other CaSR PAMs. Substitution of the terminal phenyl-furan with a more compact and more polar benzothiazole, which can be well accommodated in the hydrophobic site defined by residues I777^5.44, W818^6.50, and Y825^6.57, improved potency sevenfold [compound Z2592185946 (‘5946)], whereas its N-methylation led to Z6660636218 (‘6218) (EC₅₀ 0.25 μM) (fig. S2B). Enantiomeric purification of ‘6218 led to Z8456582460 (‘2460) (we denoted this as R-‘6218 on the basis of the cryo-EM structure in the next section), a 177 nM CaSR PAM (Fig. 3C). Despite its 80-fold potency improvement, the molecular weight and cLogP values of Z8456582460 (‘2460) were reduced versus the parental docking hit, improving lipophilic ligand efficiency from 1.0 to 3.4. Furthermore, substituting the nitrogen atom in tetrahydrobenzapine with oxygen, sulfur, or carbon resulted in the inactivation of the compounds, potentially by introducing electrostatic repulsion, thereby making them useful probe pairs for physiological studies (fig. S2B).

Fig. 3. Structural comparison between docked and experimentally determined poses for ‘54149 and ‘6218.

(A) Close-up view of ‘6218 in the 7TM^A site, with its EM density shown (level = 0.668). Surrounding residues are in green. (B) Superposition of docked and experimentally determined poses of ‘6218 in the 7TM^A site. (C) Close-up view of ‘6218 in the 7TM^B site, with its EM density (level = 0.58). Surrounding residues are in blue. (D) Superposition of docked and experimentally determined poses of ‘6218 in the 7TM^B site. (E) Close-up view of ‘54149 in the 7TM^A site, with its EM density (level = 0.215). The surrounding residues are in green. (F) Superposition of docked and experimentally determined poses of ‘54149 in the 7TM^A site. (G) Close-up view of ‘54149 in the 7TM^B site, with its EM density (level = 0.191). The surrounding residues are in blue. (H) Super-position of docked and experimentally determined poses of ‘54149 in the 7TM^B site. In (B), (D), (F), and (H), the residues under-going conformational changes in the experimental structures are shown. Docked poses and protein residues in the docked structures are in cyan.

Similar changes in the equivalent aryl groups, engaging the hydrophobic site defined by residues F668^2.56 and I841^7.36, led to improvements in docking hits ‘7909 and ‘0522 (Fig. 2D and fig. S2D). For the former, the EC₅₀ improved from more than 100 μM to 1.7 μM [Z6562953161 (‘3161); fig. S2D], and efficacy was much improved even though molecular weight was, again, decreased. Meanwhile, the analog of ‘0522, Z6923555526 (‘5526), saw the introduction of the same benzothiazole as in ‘2460, along with a simplification of the linker, giving better complementarity with the hydrophobic site defined by I777^5.44, W818^6.50, and Y825^6.57 (fig. S2C) and improving the EC₅₀ 95-fold to 0.48 μM.

We also sought to optimize the early PAMs revealed by the in-stock library which, for reasons of project staging, we had several months to evaluate and improve the initial hits before we had hits from the ultralarge library docking. Although these molecules began with weak EC₅₀ values, we were able to optimize three molecules to between 30 and 160 nM by analog searching through a 46-billion-molecule make-on-demand library using SmallWorld (https://sw.docking.org/) followed by docking (Fig. 2E and fig. S3). As an aside, we note that although these molecules originated from the smaller in-stock library, their optimization depended on the far larger space represented by the make-on-demand virtual library. Most compelling was the improvement of ZINC000057421374 (‘21374). Here, simplification of the linker and installation of a benzothiazole, as in ‘6218 and ‘5526 above, led to Z5449585339 (‘85339), with an EC₅₀ of 174 nM. Stereochemical purification to R-‘85339 [Z6269954149 (‘54149)] revealed a 41 nM PAM. Compared with our most potent lead, ‘2021, the synthesis of ‘54149, with known stereochemistry, is more straightforward. We thus selected ‘54149—with relatively high potency (1.8-fold improved on that of cinacalcet), favorable cLogP (2.9), and distinctive chemo-type receptor contacts—to ultimately take forward into in vivo studies. Although optimization of chemotypes can alter them to resemble known ligands, here the optimized PAMs retained their high dissimilarity to the established CaSR ligands (ECFP4-based Tc <0.35) (table S1).

Cryo-EM structures of the ‘6218- and ‘54149-CaSR complexes

To understand the molecular basis of recognition and to template subsequent optimization, we determined structures of CaSR in complex with two PAMs, ‘6218 and ‘54159 (R-‘85339), derived from the 1.2-billion- and 2.7-million-molecule screens, respectively. For the CaSR-‘6218 complex, the map was determined at a global nominal resolution of 2.8 Å, with locally refined maps at resolutions of 2.7 and 3.4 Å for ECD-linker and linker-7TM regions, respectively (figs. S4 and S5). For the CaSR-‘54149 complex, the map was determined at a global nominal resolution of 2.7 Å, with locally refined maps at resolutions of 2.6 and 3.6 Å for ECD-linker and linker-7TM regions, respectively (figs. S4 and S5). Similar to the structures of cinacalcet- and evocalcet-bound CaSR complexes, both the CaSR-‘6218 and -‘54159 complexes exhibit closed-closed configuration of the Venus flytrap domains (VFTs) bound to Ca²⁺, tryptophan, and phosphate, with the cysteine-rich domains (CRDs) positioned in close proximity. The 7TMs between two protomers adopt an asymmetric arrangement characterized by a raised position adopted by the TM6 of 7TM^A relative to the opposing TM6 of 7TM^B (fig. S6).

In the CaSR-‘6218 complex, the PAM binding sites show the density of ‘6218 in extended and bent conformations, recapitulating those of cinacalcet and evocalcet (Fig. 3, A and C, and fig. S7, A and B) (23). Compound ‘6218 interacts with the same overall residues in both monomers, making conserved as well as site-specific interactions. Despite ‘6218 being a racemic mixture, BRET assays revealed that one enantiomer (‘2460; EC₅₀ 177 nM) exhibits higher activity than the other [Z8456572888 (‘2888); EC₅₀ 7.2 μM] (fig. S2B). Our modeling indicates that the cationic amine of R-‘6218 hydrogen-bonds with Q681^3.33 and ion-pairs with E837^7.32 in both the 7TM^A and 7TM^B sites, whereas the interactions are potentially weakened by a longer distance when binding to S-‘6218 (fig. S8, A and B). This is consistent with previous observations of the superior activity of the R configuration of arylalkylamine PAMs (46). Consequently, we modeled R-‘6218 in our structure and denoted it as ‘2460. In the 7TM^B site, the methyl-benzazepine ring forms π-π interactions with F684^3.36 and W818^6.50, recapitulating the interactions formed by the naphthalene in cinacalcet and evocalcet. The benzoisothiazole ring makes π-π interactions with F821^6.53 and Y825^6.57 (Fig. 3C). In the 7TM^A site, whereas W818^6.50 swings out by 120° and Y825^6.57 moves down, the π-π interactions are still maintained. Conversely, the interaction with F821^6.53 is lost as it swings out and is no longer part of the allosteric pocket (Fig. 3A).

The docking-predicted pose for ‘6218 super-poses well with its experimental structure in both monomers (Fig. 3, B and D). Both the docked and experimental poses of ‘6218 have an extended conformation in the 7TM^A site (Fig. 3B), but they adopt a bent conformation in the 7TM^B site (Fig. 3D). The same bent and extended conformations were observed for the initial docking hits; in this sense, this level of geometric fidelity emerged directly from the docking screen (Fig. 1E and fig. S1F). The docked and experimental structures super-imposed with a 1.88-Å root mean square deviation (RMSD) in the bent conformation monomer and with a 2.23-Å RMSD in the extended conformation monomer. Although the experimental results broadly support the docking prediction, there were important differences in the receptor structures. Compared with the cinacalcet complex against which we docked (7TM^B), in the 7TM^B of the CaSR-‘6218 complex, F821^6.53 swings 120° into the site to become part of the binding pocket, making a π-π interaction with the benzoisothiazole ring of ‘6218 (Fig. 3D). This conformation is not adopted in the cinacalcet or the evocalcet complex, likely because the mobile groups of cinacalcet [1-propyl-3-(trifluoromethyl) benzene] and evocalcet (2-phenylacetic acid) are bulkier and would clash with this phenylalanine (fig. S9). Meanwhile, in the extended monomer’s binding site (7TM^A), W818^6.50 moves 120° to swing outside of the binding pocket in the ‘6218 complex.

Similar to ‘6218, the cryo-EM structure of CaSR with ‘54149 shows that the new PAM also adopts an extended conformation in the 7TM^A site and a bent conformation in the 7TM^B site, inducing similar rearrangements of W818^6.50 and F821^6.53 in the 7TM^A and 7TM^B sites, respectively (Fig. 3, E to H, and fig. S7, C and D). Compounds ‘54149 and ‘6218 share a benzoisothiazole group that is flexible in the two sites, suggesting that the conformational changes of W818^6.50 and F821^6.53 are benzoisothiazole specific. At the 7TM^B site, the benzodioxole group interacts with F684^3.36 and W818^6.50 through π-π stacking, and the benzoisothiazole forms π-π interactions with F821^6.53 and Y825^6.57. The cationic amine hydrogen-bonds with Q681^3.33 and ion-pairs with E837^7.32, and the adjacent methyl packs with I841^7.36 (Fig. 3G). The interactions with F821^6.53 and Y825^6.57 are lost in the 7TM^A site as F821^6.53 swings out of the pocket and Y825 swings down (Fig. 3E). The docked and experimental structures superposed to 0.91-Å RMSD in the 7TM^A site and to 2.68-Å RMSD in the 7TM^B site (Fig. 3, F and H). Docking predicted ‘54149 to adopt both extended conformations in the binding pocket, but we observed signs of conformational heterogeneity in the 7TM^A site. The EM density suggests that ‘54149 adopts an alternative “folded-over” conformation at this site, which has not been previously observed (figs. S7C and S10). In this folded-over configuration, ‘54149 establishes favorable interactions with CaSR; the benzoisothiazole ring makes additional contacts by edge-to-π stacking with F814^6.46 and is surrounded by a hydrophobic pocket created by A840^7.35, I841^7.36, A844^7.39, and V817^6.49 (A, Ala; V, Val). Among these, A840^7.35 and I841^7.36 are particularly important for the recognition of CaSR PAMs (44, 45). Unlike methyl-benzazepine (in ‘6218) and naphthalene (in cinacalcet and evocalcet), ‘54149 uses a smaller benzodioxole as the stationary binding component, possibly allowing more configurations in the pocket. Together, the conformational disparity in the structure of these complexes highlights the ongoing importance of cycles of docking and structure determination in drug discovery efforts.

‘54149 promotes a distinct active-state CaSR dimer conformation

Compared with CaSR-cinacalcet alone, the structure against which we docked, our recent structure of the receptor in complex with cinacalcet and G_iβγ [Protein Data Bank (PDB) ID 8SZH] (47) revealed that G protein binding promotes an additional conformational change that brings the two 7TMs into closer contact, in a configuration that is in line with the activation of other family C receptors (24, 48). From the inactive state to the G protein–bound active state, the interface contact area increases from 178.9 Ų (inactive; NPS-2143–bound; PDB ID 7M3E) to 206.2 Ų (cinacalcet-bound; PDB ID 7M3F) to 682.7 Ų (cinacalcet, G_iβγ-bound) (calculated by PDBePISA). By aligning the 7TM^B sites, we observe that in the complexes with ‘54149 and ‘6218, 7TM^A moves down toward the cytoplasm. This increases the interface contact area to 351.3 and 271.5 Å, respectively, compared with that for cinacalcet-bound CaSR (Fig. 4A). The downward shift brings the two 7TM bundles into a conformation that is closer to the G protein–bound structure, especially for that induced by ‘54149, suggesting that ‘54149 promotes a dimer that may favor G protein activation versus those stabilized by the other compounds (fig. S11), which may contribute to its efficacy and potentially confer a different pharmacology.

Fig. 4. ‘54149 increases the TM6-TM6 interface and is more effective in suppressing PTH secretion in ex vivo parathyroid glands.

(A) From the cinacalcet-bound to the ‘54149-bound to the Gi-bound CaSR, the 7TM^A protomer undergoes a downward and rotational movement, bringing TM6 closer to the 7TM^B. Cinacalcet-bound CaSR is in gray, ‘54149-bound CaSR is in orange, ‘6218-bound CaSR is in pink, and G_i-bound CaSR is in blue. (B) Parathyroid glands of 4-week-old WT C57/B6 mice were sequentially incubated with increasing concentrations of ‘54149, cinacalcet, and evocalcet from 0.01 nM to 50 μM in the presence of 0.75 mM extracellular calcium ([Ca²⁺]_e). The median inhibitory concentrations (IC₅₀) of ‘54149, evocalcet, and cinacalcet in suppressing PTH secretion are 583 nM (122 to 4727 nM), 998 nM (412 to 4018 nM), and 53 μM, respectively. (C and D) Parathyroid glands were sequentially incubated with increasing [Ca²⁺]_e from 0.5 to 3.0 mM in the presence of vehicle [0.1% dimethyl sulfoxide (DMSO)] or 50 nM (C) or 500 nM (D) ‘54149, cinacalcet, or evocalcet. Shown at the top are changes in the rate of PTH secretion on a per-gland and per-hour basis with increasing [Ca²⁺]_e to compare the secreted PTH at a mean maximal rate (PTH-max). Shown at the bottom are the normalized PTH secretion rates (the highest rates are normalized to the basal rate at 0.5 mM [Ca²⁺]_e of the vehicle, and the lowest rates are normalized to the rate at 3.0 mM [Ca²⁺]_e) to better assess changes in the Ca²⁺ set point ([Ca²⁺]_e needed to suppress 50% of [Ca²⁺]_e-suppressible PTH secretion). Dotted vertical lines indicate Ca²⁺ set points for the corresponding treatments. In (B) to (D), error bars represent the SEM of n = 8 groups of parathyroid glands for each treatment. Dashed lines are used for visual guidance only and do not represent curve fits.

‘54149 suppresses PTH secretion better than the approved PAM drugs

Upon activation, CaSR suppresses PTH secretion from parathyroid glands (49), which is the primary goal of the calcimimetic drugs. Because all PAM-bound structures were obtained under saturating calcium concentrations (10 mM), the different conformations observed are specific to each PAM and may translate into functional differences. We thus investigated the functional effects of the different PAM drugs and leads by monitoring PTH secretion in extracted parathyroid glands from wild-type (WT) C57/BL6 (B6) mice at a constant external calcium concentration of 0.75 mM. Compound ‘54149, cinacalcet, and evocalcet all inhibit PTH secretion dose-dependently, with potencies as follows: ‘54149 (580 nM) ~ evocalcet (1 μM) >> cinacalcet (53 μM) (Fig. 4B). Because PAMs positively regulate CaSR by lowering the required calcium for activation, we wanted to assess how the different compounds shift the calcium set point for PTH secretion by the glands (Fig. 4, C and D). For this assay we used two PAM concentrations, 500 and 50 nM (dashed lines in Fig. 4B). At 500 nM, ‘54149 shifted the calcium set point to 0.62 mM (from 1.5 mM), whereas at the same concentration, cinacalcet shifted the set point to 0.94 mM and evocalcet shifted it to 0.76 mM (Fig. 4D). The same trend held when the PAMs were administered at 50 nM, leading to shifts in the calcium set point from 1.47 (vehicle) to 1.23 (cinacalcet) to 1 (evocalcet) to 0.85 (‘54149) mM (Fig. 4C). It is worth noting that ‘54149 also suppresses the tonic secretion of PTH at 0.5 mM calcium, an effect not observed with the two approved drugs.

‘54149 reduces serum PTH at lower doses with less hypocalcemia than cinacalcet

Encouraged by its improved affinity and ex vivo organ efficacy, we investigated the in vivo activity of ‘54149, beginning with pharmacokinetic studies in CD-1 mice. We separately administered ‘54149 and the two drugs, cinacalcet and evocalcet, at 3 mg per kg body weight (mg/kg) subcutaneously (Fig. 5A). At this dose, ‘54149 was found in an appreciable amount in plasma—area under the curve from time 0 to infinity (AUC_0→∞) of 18,500 ng·min/ml. The peak plasma concentration (C_max) reaches 112 ng/ml (340 nM) at 15 min and stays high until 60 min (100 ng/ml). Based on the EC₅₀ of ‘54149, at this dose, ‘54149 is close to saturation over this period (Fig. 2E). By comparison, evocalcet has a much higher systemic exposure at the same dose, with a C_max of 3250 ng/ml (8.7 μM) at 60 min. Conversely, cinacalcet, which is far more widely used, has lower exposure than ‘54149, with a C_max of 58.9 ng/ml (149.5 nM) 15 min after subcutaneous administration (Fig. 5A). We note that no effort has been made to optimize ‘54149 for pharmacokinetic exposure or clearance; to the extent that it has favorable pharmaco-kinetics, this simply reflects the physical property constraints imposed in docking and ligand optimization.

Fig. 5. ‘54149 suppresses serum PTH at a lower dose and causes less of a hypocalcemia effect than cinacalcet and evocalcet.

(A) Pharmacokinetics of ‘54149 compared with that of cinacalcet and evocalcet after 3 mg/kg subcutaneous injection (n = 3 mice). (B) Serum PTH concentration change over 8 hours after 1 mg/kg subcutaneous injection of ‘54149, cinacalcet, or evocalcet (n = 3 mice). (C) Serum PTH concentration change over 8 hours after 10 mg/kg subcutaneous injection of ‘54149 or cinacalcet (n = 5 mice). (D) Comparison of ‘54149 and cinacalcet in regulating serum PTH at different doses (subcutaneous injection) after 30 min of injection. Each dose was administered to 10 mice, except for the injection at 10 mg/kg (n = 5 mice). P values were assessed by unpaired Student’s t test. (E) Plasma calcium concentration in mice after 3 mg/kg subcutaneous injection of ‘54149, cinacalcet, or evocalcet (n = 3 mice). (F) Serum calcium concentration after 1 mg/kg subcutaneous injection of ‘54149, cinacalcet, or evocalcet (n = 5 mice). For experiments in (B) to (D) and (F), the concentrations of evocalcet and cinacalcet are corrected for their molecular weight difference with ‘54149. Error bars represent the SEM.

On the basis of the pharmacokinetic study, we selected two doses to investigate the time course of PTH suppression by the PAMs in WT B6 mice. At 1 mg/kg (3.1 μmol/kg), ‘54149 and equimolar evocalcet fully suppressed PTH secretion, whereas cinacalcet was less effective at this dose (Fig. 5B). Only at 10 mg/kg (31 μmol/kg) did cinacalcet fully suppress PTH secretion (Fig. 5, C and D). Overall, ‘54149 fully suppresses serum PTH at a 10 times lower dose than cinacalcet (Fig. 5D), consistent with its ability to suppress releases of both tonic and Ca²⁺-suppressible pools of PTH (Fig. 4, C and D).

A key adverse effect of cinacalcet and etelcalcetide is decreased blood calcium (50). In SHPT, high PTH is accompanied by low or normal blood calcium concentration. The overproduction of PTH and the proliferation of parathyroid cells in patients with SHPT are largely driven by low blood calcium and high blood phosphate levels (51–53) as well as reduced CaSR expression in parathyroid cells (54). We were thus keen to compare the serum calcium concentration after injection of ‘54149 versus cinacalcet. At 3 mg/kg, ‘54149 did not substantially alter serum calcium concentration for 4 hours but slightly increased it from 2.2 to 2.4 mM after the drug dissipated in circulation 6 hours after injection (Fig. 5E). By contrast, the same dose of cinacalcet and evocalcet substantially lowered serum calcium for more than 8 hours, from 2.2 mM to 1.7 and 1.6 mM at trough, respectively. The hypocalcemic action of evocalcet is particularly robust even at a lower dose of 3.1 μmol/kg (~1 mg/kg) (Fig. 5F), whereas the same dose of ‘54149 retained the ability to maximally suppress serum PTH without hypocalcemia (Fig. 5D). Although the mechanisms underlying the different calcemic actions of these three compounds remain to be determined, their common ability to suppress PTH secretion suggests that differential calcemic actions likely take place in other calciotropic organs outside of the parathyroid glands. For example, the hypocalcemia induced by cinacalcet has been partly linked to the activation of CaSR in the thyroid, leading to increased CaSR-mediated calcitonin secretion (55). Notably, some PAMs that also display agonist activity (ago-PAMs), such as AC265347 and R,R-calcimimetic B, have a low effect on calcitonin levels but maintain the ability to suppress serum PTH (56, 57). This has been attributed to their capacity to bias signaling toward the ERK1/2 pathway, which does not influence the stimulation of calcitonin release, unlike cinacalcet (58). However, in our assays, ‘54149 does not exhibit intrinsic agonist activity toward CaSR, indicating that the favorable pharmacology is not limited to ago-PAMs (fig. S12). A systematic study to evaluate the effects of ‘54149 on CaSRs across different signaling pathways and tissues may illuminate how it avoids inducing a hypocalcemia side effect.

Discussion

Four key observations emerged from this study. First, from a structure-based screen of a 1.2-billion-molecule tangible library emerged a spectrum of diverse chemotypes that potently enhanced CaSR activation. These molecules are among the first PAMs discovered through large library docking and are among the first structure-based ligands discovered for family C GPCRs (59). The potency of the initial docking hits was relatively high, with EC₅₀ values down to 273 nM, and all were structurally dissimilar to known CaSR PAMs. Structure-based optimization improved affinity between 40- and 600-fold, leading to molecules that were up to 50-fold more potent than cinacalcet in vitro and 10- to 100-fold more potent at suppressing PTH secretion from organs ex vivo and in vivo. Second, the docking predictions were largely confirmed by the subsequent cryo-EM structures, with an important exception (see below), including selecting for and correctly predicting extended and bent conformations in the 7TM^A and 7TM^B sites of the CaSR dimer. Third, docking a library of 1.2 billion molecules led to 2.7-fold higher experimental hit rates and PAMs that were up to 37 times more potent than docking a smaller (2.7 million) molecule library (Fig. 1C). This offers among the first experimental support for simulations suggesting that docking results improve with library size increases (33). Fourth, the chemotypes reported here make interactions with the receptor that have not been seen in prior PAMs, promoting new active-state dimer interfaces that are closer to the G protein–coupled state and which were not observed with the established drugs. The more extensive dimer interfaces adopted in complex with the new PAMs, closer to that of the G protein–coupled state, may be behind the improved efficacy and reduced hypocalcemia of ‘54149 versus the established drugs (60, 61).

Several caveats merit mentioning. We used DOCK3.7 for the screening of 2.7 million molecules and DOCK3.8 for the screening of 1.2 billion molecules; although these two methods are highly similar, they are not exactly the same. We note that if we dock random subsets of 2.7 million molecules, multiple times, drawn from the 1.2 billion molecules, their score distributions closely resemble that of the 2.7 million molecules screened using DOCK3.7 (fig. S13). In addition, we do not suggest that the molecules described here are drugs. Whereas the pharmacokinetics of ‘54149 are sufficient to support in vivo studies, and indeed in some ways to demonstrate superiority to cinacalcet, optimization of exposure and half-life of the molecule will be necessary. Although the relative lack of a hypocalcemic effect is encouraging, understanding the mechanism underlying this effect requires systematic exploration of CaSR activation in other calciotropic organs, including bones and kidneys. Whereas in three of the four cases the docking-predicted structures of the PAMs in the 7TM^A and 7TM^B monomers were confirmed by cryo-EM, in one site the docking pose was different from the experimental result. Although these molecules are among the most potent PAMs discovered for CaSR and for GPCRs more generally, this does reflect their binding to a site that represents the orthosteric ligand binding region in most receptors, such as the family A GPCRs. Finally, the improvement in docking hit rates and docking potencies from billion-molecule versus million-molecule libraries seems compelling, but the numbers experimentally tested remain relatively low given docking uncertainties.

From a lead discovery standpoint, the molecules that emerged from this structure-based approach have advantages over the present CaSR drugs: They are 10- to 100-fold more potent and do not induce the hypocalcemia that is the major dose-limiting side effect of what is now the standard of care. From a library docking standpoint, the hits that emerged from the billion-molecule library were 37-fold more potent and hit rates were 2.7-fold higher than those that emerged from docking the million-molecule-scale library against the same site. Although such a comparison merits further study, certainly with more molecules being tested to power the statistics, it is consistent with theoretical studies (33) and supports the continued expansion of readily testable libraries for drug discovery (27, 30).

Materials and methods

In-stock and ultralarge virtual ligand screening

To investigate the effect of small versus large library docking and test the docking prediction of the PAM binding sites in complex with extended or bent PAMs, we optimized two docking setups based on the cryo-EM structures of cinacalcet- or evocalcet-bound CaSR. CaSR-cinacalcet (PDB ID 7M3F) is used for the 7TM^A site, and CaSR-evocalcet (PDB ID 7M3G) is used for the 7TM^B site (23). In both sites, the positions of Q681 and E837 were manually adjusted to form stronger hydrogen bonds or a salt bridge with the secondary amine in cinacalcet or evocalcet, and in the 7TM^B site, lipid tails were added in the docking setup based on the existing electron density. 7TMs were protonated using Reduce (62) (7TM^B site) or by Protein Preparation Wizard in Maestro (7TM^A site) (2020 release) (63). Energy grids for the different energy terms of the scoring function were pregenerated—van der Waals term based on the AMBER force fields using CHEMGRID (37), Poisson-Boltzmann–based electrostatic potentials using QNIFFT (38), and context-dependent ligand desolvation was calculated using SOLVMAP (36). The volume of the low dielectric and the desolvation volume was extended out 0.8 and 0.3 Å in the 7TM^A site and 0.6 and 0.3 Å in the 7TM^B site. Different dielectric and desolvation volumes were sampled to maximize enrichment (ability to prioritize ligands over nonligands) and geometric fidelity of the known molecules and to bring the van der Waals and polar components into rough balance in retrospective control calculations (64). The experimentally determined poses of cinacalcet and evocalcet were used to generate matching spheres, which were later used by the docking software to fit pregenerated ligand conformations into the small-molecule binding sites (36).

The resulting docking setups were evaluated for their ability to enrich known CaSR ligands over property-matched decoys. Decoys are theoretical nonbinders to the receptor that are topologically dissimilar to known ligands but retain similar physical properties. We extracted 10 known PAMs from CHEMBL (https://www.ebi.ac.uk/chembl/), including cinacalcet and evocalcet. Four hundred eighty-five decoys were generated by using the DUDE-Z pipeline (65). High logAUCs of 38.89 and 31.67 were achieved for the 7TM^A and 7TM^B sites, respectively. Moreover, these docking setups offered fidelity in reproducing extended and bent poses of the known PAMs. For example, by using the 7TM^A site setup, 7 out of 10 PAMs adopted extended conformations while making sensible interactions with the surrounding key residues. By using the 7TM^B site setup, 7 out of 10 PAMs adopted bent conformations. We also used an “extrema” set of 92,552 molecules using the DUDE-Z web server (http://tldr.docking.org) to ensure that the setups did not enrich extreme physical properties. Both setups enriched more than 90% neutrals or monocations among the top-ranking molecules, which are two charge species that are precedented among CaSR PAMs.

Then, 2.7 million “lead-like” molecules (molecular weight 300 to 350 Da and logP ≤ 3.5) from the ZINC15 database (https://zinc15.docking.org/) were docked against both sites using DOCK3.7 (36). In the docking screen against the 7TM^A site, each library molecule was sampled in about 3927 orientations and, on average, 330 conformations. For the 7TM^B site, each library molecule was sampled in about 3612 orientations and, on average, 330 conformations. The best-scoring configuration for each docked molecule was relaxed by rigid-body minimization. The two screens took 956 and 917 core hours, respectively, spread over 100 cores, or slightly more than 3 days. For the 1.2-billion-molecule ultralarge library docking, each library molecule from the ZINC22 database (42) was sampled in about 1707 orientations and 425 conformations in the 7TM^B site by using DOCK3.8 (36). Overall, more than 681 trillion complexes were sampled and scored, spending 380,016 core hours spreading over 2000 cores.

Processing of the docking results

For the in-stock screen against the 7TM^B site, 5208 molecules with dock energy ≤−35 kcal/mol were filtered for uniqueness compared to known ligands using the ECFP4-based Tc against 662 CaSR ligands in CHEMBL (https://www.ebi.ac.uk/chembl/). Molecules with a Tc >0.35 were eliminated. These molecules were filtered for internal strain with criteria of total strain energy <8 and maximum dihedral torsion energy <3 (40). Moreover, the molecules were further filtered for key interactions: a hydrogen bond with Q681 and either a hydrogen bond or salt bridge with E837, using interfilter.py based on OpenEye Python Toolkits (https://docs.eyesopen.com/toolkits/python/quickstart-python/linuxosx.html). After these three filters, 103 molecules were left for further examination. Upon clustering by an ECFP4-based Tc of 0.5, 79 molecules were visually inspected for π-π interactions with W818 and F684. Twenty-eight molecules were picked, but only 22 molecules could be sourced from vendors and arrived for in vitro testing.

For the in-stock screen against the 7TM^A site, 33,321 molecules with dock energy ≤−43 kcal/mol were filtered against the same three filters, resulting in 2540 molecules for further examination. The 2540 molecules were filtered against a vendor filter to assess their purchasability, resulting in 647 molecules for further examination. The 647 molecules were clustered based on ECP4-based Tc of 0.5 and resulted in 413 clusterheads. The clusterheads were visually inspected in a similar manner, resulting in 28 candidates for ordering for purchasing, of which 26 molecules arrived for testing. For the large-scale screen, 1.2 billion molecules were screened and 1 billion molecules were scored in the 7TM^B site. The strain filter was incorporated as part of the new DOCK3.8 pipeline. Then, 2,321,171 molecules with ≤−35 kcal/mol were filtered for key interactions with Q681, E837, W818, and F684 as well as distinction to known ligands. The interaction filtering script for p-p interactions with W818 and F684 was implemented based on LUNA (https://github.com/keiserlab/LUNA) (66). After visual inspection, 1002 molecules were left. To reduce the number of candidate molecules for purchasing, these 1002 molecules were redocked against the 7TM^A site, and 907 molecules were scored in the 7TM^A site. The molecules were visually inspected again for their poses against both sites, and the remaining 212 new and nonstrained molecules were clustered by the LUNA 1024-length binary fingerprint of a Tc = 0.3, resulting in 112 clusterheads. Ultimately, 96 molecules were prioritized for purchasing based on a final round of visual inspection. The 96 molecules may be organized into three broad physical categories: (i) molecules that have two aromatic ends, which usually adopt a bent conformation in the 7TM^B site and an extended conformation in the 7TM^A site; (ii) molecules that have an aromatic moiety in the pocket and a nonaromatic structure at the distal end but still score well; and (iii) neutral molecules.

Synthesis of molecules

The in-stock prioritized molecules were sourced from Enamine; Vitas-M laboratory, Ltd.; UkrOrg-Synthesis, Ltd.; ChemBridge Corporation; and Sigma. Ninety-six molecules prioritized for purchasing were synthesized by Enamine for a total fulfilment rate of 74%. Compounds were sourced from the Enamine REAL database (https://enamine.net/compound-collections/real-compounds). The purities of active molecules were at least 90% and typically greater than 95%. The detailed chemical synthesis can be found in the Chemical synthesis and analytical investigations section (fig. S14).

Hit optimization

Potential analogs of the hits were identified through a combination of similarity and substructure searches of the SmallWorld database (https://sw.docking.org/) from a 46-billion-molecule make-on-demand library. Potential analogs were docked to the CaSR 7TM^B binding site using DOCK3.8. As was true in the primary screen, the resulting docked poses were manually evaluated for specific interactions and compatibility with the site, and prioritized analogs were acquired and tested experimentally.

Molecular cloning

Full-length and the truncated CaSR (residues 20 to 894) were cloned into a pFastBac1 vector (for expression in insect cells) or a pcDNA3.1(+) vector [for expression in human embryonic kidney (HEK) 293S cells], with an N-terminal haemagglutinin (HA) signal sequence followed by a FLAG tag. To improve the protein yield of CaSR, the DNA sequence of the C-terminal tail from GABA_B1 or GABA_B2 and an endoplasmic reticulum retention motif were inserted at the C terminus of pFastBac1-FLAG-CaSR (residues 20 to 894) to generate CaSR-C1 and CaSR-C2 constructs, which have been shown to have comparable G protein–signaling profiles to that of the WT CaSR homodimer (23). The FLAG tag of the CaSR-C1 construct was then replaced by a Twin-Strep-tag (WSHPQFEKGGGSGGG-SGGSAWSHPQFEK). All plasmids used were sequence-verified.

BRET TRUPATH assay

BRET assays were performed and analyzed similar to previously described methods (41). HEK293S cells grown in FreeStyle 293 suspension media (ThermoFisher) were cotransfected with 150 ng of pCDNA3.1-CaSR FL, Gai3-Rluc8, Gβ, and Gγ-GFP2 per 1 ml of cells at a density of 1 × 10⁶ cells/ml using a DNA:polyethylineimine ratio of 1:5, and incubated at 130 rpm. and 37°C. Cells were harvested 48 hours after transfection and washed in assay buffer (Hank’s balanced salt solution with 25 mM HEPES pH 7.5) supplemented with 0.5 mM EGTA, followed by another wash in assay buffer. The cells were then resuspended in an assay buffer with 5 μg/ml coelenterazine 400a (GoldBio) and placed in white 96-well plates (136101, Thermo Scientific) in a volume of 60 μl per well. Thirty microliters of ligands prepared at three times the final concentrations in assay buffer with 1.5 mM CaCl_2, 0.1% bovine serum albumin (BSA), and 3% dimethyl sulfoxide (DMSO) were added to plated cells (final concentrations of 0.5 mM CaCl_2, 0.033% BSA, and 1% DMSO). After 5 min of incubation, the emissions at 410 and 515 nm were read using a SpectraMax iD5 plate reader with a 1-s integration time per well. The BRET ratios (GFP2/RLuc8 emission) were calculated and normalized to ligand-free control before further analysis. The efficacy and potency of the molecules were calculated by fitting the concentrations of molecules and the BRET ratios to a four-parameter logistic equation in Prism (GraphPad software). To evaluate the intrinsic agonist activity of PAMs, a similar approach was used except that the concentration-response curves were generated over a range of extracellular Ca²⁺ concentrations in the presence of the indicated concentrations of PAMs. The responses were normalized to the maximum response detected in the absence of Ca²⁺ and fit to an operational model of allosterism (67).

Protein expression and purification

CaSR-C1 and CaSR-C2 were overexpressed in Spodoptera frugiperda Sf9 cells using a Bac-to-Bac baculovirus expression system. Sf9 cells grown to a density of 3 × 10⁶ cells/ml were co-infected with CaSR-C1 and CaSR-C2 baculoviruses for 48 hours at 27°C. Cells were then harvested and stored at −80 °C. Purifications of CaSR in complex with compounds ‘6218 and ‘54159 followed a similar protocol. Cell pellets were thawed, resuspended, and lysed by nitrogen cavitation in the lysis buffer containing 20 mM HEPES pH 7.5, 150 mM NaCl, 10 mM CaCl₂, 10% glycerol, 10 mM l-tryptophan, protease inhibitors, benzonase, and 50 μM of a specific compound. The lysates were centrifuged at 1000g for 10 min to remove nuclei and unlysed cells. The membranes were harvested by centrifugation at 100,000g for 30 min and solubilized in the lysis buffer supplemented with 1% (w/v) lauryl maltose neopentyl glycol (LMNG; Anatrace) and 0.2% (w/v) cholesteryl hemisuccinate (CHS; Anatrace) for 3 hours, followed by centrifugation at 100,000g for 30 min. The supernatant was incubated with Strep-TactinXT 4Flow resin (IBA) overnight at 4°C. The resin was then loaded into a gravity column and washed with 10 column volumes of the washing buffer, which contained 20 mM HEPES 7.5, 150 mM NaCl, 10 mM CaCl₂, 5% glycerol, 40 μM l-tryptophan, and 50 μM compound, supplemented with 0.01% (w/v) LMNG and 0.002% (w/v) CHS, followed by a second wash with 10 column volumes of washing buffer with 0.001% (w/v) LMNG and 0.0002% (w/v) CHS. Proteins were eluted by Strep-TactinXT elution buffer (IBA) supplemented with 10 mM CaCl₂, 40 μM l-tryptophan, 50 μM compound, 0.00075% (w/v) LMNG, 0.00025% (w/v) glycodiosgenin (GDN; Anatrace) and 0.00015% (w/v) CHS, and further purified by a Superose 6 column (Cytiva) using a buffer containing 20 mM HEPES 7.5, 150 mM NaCl, 10 mM CaCl₂, 40 μM l-tryptophan, 50 μM compound, 0.00075% (w/v) LMNG, 0.00025% (w/v) GDN, and 0.00015% (w/v) CHS. The peak fractions were pooled and concentrated for cryo-EM studies.

Cryo-EM data acquisition and data processing

Three microliters of CaSR-‘6218 or CaSR-‘54159 at 6 mg/ml was applied to glow-discharged 300-mesh R1.2/R1.3 UltrAuFoil holey gold grids (Quantifoil) under 100% humidity at 4°C. Excess sample was blotted away, and the grids were plunge-frozen into liquid ethane using a Vitrobot Mark IV (Thermo Fisher Scientific). For cryo-EM imaging of the CaSR-‘6218 complex, movies were collected using a Titan Krios G2 (Thermo Fisher Scientific) transmission electron microscope equipped with a Gatan K3 direct detector and a postcolumn energy filter with a 20-eV slit width. The microscope was operated at 300 kV, with a nominal magnification of 130,000×, resulting in a pixel size of 0.8677 Å. Movies were automatically recorded in counting mode using SerialEM (68), with a total exposure of 55 electrons/Ų over 60 frames, and the defocus range was set from −0.5 to −1.5 μm. For cryo-EM imaging of the CaSR-‘54159 complex, movies were collected using a Titan Krios G2 transmission electron microscope equipped with a Falcon 4i Direct Electron Detector and a postcolumn energy filter with a 20-eV slit width. The microscope was operated at 300 kV, with a nominal magnification of 165,000×, resulting in a pixel size of 0.75 Å. Movies were recorded in counting mode using EPU 3.6 (Thermo Fisher Scientific) with a total exposure of 50 electrons/Ų over 50 frames, and the defocus range was set from −0.5 to −1.5 μm.

For a detailed workflow of data processing, please refer to fig. S4. All data underwent processing using similar strategies using cryoSPARC 3.0 (69) and Relion 3 (70). Movies were imported into cryoSPARC and subjected to patch motion correction, followed by the contrast transfer function (CTF) estimation using patch CTF estimation. Micrographs with CTF estimations worse than 4 Å were excluded, resulting in a total of 11,926 micrographs for the CaSR-‘6218 complex and 17,625 micrographs for the CaSR-‘54149 complex, which were selected for further processing. Particles were autopicked, extracted from the micrographs, and subjected to three to five rounds of 2D classification. Particles classified into “good” classes were selected and subjected to iterative rounds of 3D ab initio reconstruction using multiple classes, followed by 3D heterogeneous refinement to remove particles from bad classes. For the early rounds of 3D classification, particles from “bad” classes were further classified by 2D classification, and good particles were retained for subsequent heterogeneous refinement. The resulting high-quality particle projections were then imported into Relion, where they were subjected to C2 symmetry expansion, followed by two or three rounds of focused 3D classification (without applying symmetry) without alignment with a mask covering the two 7TMs of CaSR. Finally, the particles from one of the two best 3D classes with C1 symmetry were selected and imported to cryoSPARC for CTF refinement and local nonuniform refinement with a soft mask covering CRD-7TM and ECD-CRD to obtain high-resolution maps. The focused maps were used to generate composite maps for refinement.

Model building and refinement

The initial models of CaSR were built on the structure of the active-state CaSR (PDB ID 7M3F) and manually docked into the cryo-EM maps in Chimera (71). The models were then subjected to iterative rounds of manual refinement in Coot (72) and automatic real-space refinement in Phenix (73). The models for the CRD-7TM and ECD-CRD regions were refined using the focused maps that cover these regions first and then combined for further refinement using the composite maps. The final models were analyzed and validated using MolProbity (74). The refinement statistics are shown in table S2. Structure figures were generated using ChimeraX (75).

Screening for potential colloidal aggregation

Dynamic light scattering

Samples were prepared in filtered 50 mM KPi buffer, pH 7.0 with final DMSO concentration at 1% (v/v). Colloidal particle formation was detected using DynaPro Plate Reader III (Wyatt Technologies). All compounds were screened in triplicate at 100 or 10 μM. Colloidal particles were defined as having a scattering intensity of at least one order of magnitude greater than background scattering and a hydrodynamic radius greater than 100 nm in size as previously described (76). If colloids were detected at the initial screening concentration, eight-point half-log dilutions of compounds were performed on dynamic light scattering in triplicate. To determine the critical aggregation concentration, data for each compound were spilt into two datasets based on aggregating (i.e., >10⁷ scattering intensity) and nonaggregating (i.e., <10⁷ scattering intensity) and were fitted with separate nonlinear regression curves, and the point of intersection was determined using GraphPad Prism software version 9.1.1 (San Diego, CA).

Enzyme inhibition assays

Enzyme inhibition assays were performed at room temperature using CLARIOstar Plate Reader (BMG Labtech). Samples were prepared in 50 mM KPi buffer, pH 7.0 with final DMSO concentration at 1% (v/v). Compounds were incubated with 2 nM AmpC β-lactamase (AmpC) [purified as previously described (77)] or malate dehydrogenase (MDH) (Sigma Aldrich, 442610) for 5 min. AmpC reactions were initiated by the addition of 50 μM CENTA chromogenic substrate (Sigma Aldrich, 219475). The change in absorbance was monitored at 405 nm for 80 s. MDH reactions were initiated by the addition of 200 μM nicotinamide adenine dinucleotide (NADH) (Sigma Aldrich, 54839) and 200 μM oxaloacetic acid (Sigma Aldrich, 324427). The change in absorbance was monitored at 340 nm, also for 80 s. Each compound was screened at 100 or 10 μM in triplicate. Experimental rates were divided by the DMSO control rate to determine percent enzyme activity (%). All compounds were initially screened against MDH, and if compounds did not display ≥40% inhibition against enzyme but formed colloidal-like particles by dynamic light scattering, compounds were additionally screened against enzyme, AmpC. Data were analyzed using GraphPad Prism software version 9.1.1 (San Diego, CA).

Pharmacokinetics

Pharmacokinetic experiments of ‘54149, cinacalcet, and evocalcet were performed by Bienta Enamine Biology Sciences (Kiev, Ukarine) in accordance with the study protocols P092622a, P050723b, and P050723a. Plasma pharmacokinetics of ‘54149, cinacalcet, and evocalcet were measured after a single 3 mg/kg dose, administered subcutaneously at time points of 5, 15, 30, 60, 120, 240, 360, 480, and 1440 min. All animals were fasted for 4 hours before dosing. ‘54149 was formulated in 2-HPbCD:saline (30:70, v/v). Cinacalcet and evocalcet were formulated in DMSO:20% captisol in saline w/v (10:90, v/v). Testing was done in healthy male CD-1 mice (9 weeks old) weighing 32.7 ± 2.1, 32.8 ± 1.9, or 32.9 ± 2.4 g in the three studies. For all three studies, each of the time-point treatment groups included three animals with a control group of one animal dosed with vehicle. In total, 28 animals were used in each study. Mice were injected intraperitoneally with 2,2,2-tribromoethanol at the dose of 150 mg/kg before drawing the blood. Blood collection was performed from the orbital sinus in microtainers containing K3EDTA and tubes with clot activator. Animals were sacrificed by cervical dislocation after the blood sample collection. Blood samples were centrifuged for 10 min to obtain plasma (15 min to obtain serum) at 3000 rpm. All samples were immediately processed, flash-frozen, and stored at −70°C until subsequent analysis. The concentrations of the test compound below the lower limits of quantitation (2 ng/ml) were designated as zero. The pharmacokinetic data analysis was performed using noncompartmental, bolus injection, or extravascular input analysis models in WinNonlin 5.2 (PharSight). Data below the lower limits of quantitation were presented as missing to improve validity of T1/2 calculations. For each treatment condition, the final concentration values obtained at each time point were analyzed for outliers using Grubbs’ test with the level of significance set at p < 0.05.

Sample processing

Plasma samples (40 μl) were mixed with 200 μl of internal standard solution. After mixing by pipetting and centrifuging for 4 min at 6000 rpm, 2 μl of each supernatant was injected into a liquid chromatography–tandem mass spectrometry (LC-MS/MS) system. Solution of compound verapamil (200 ng/ml in water-methanol mixture 1:9, v/v) was used as internal standard for quantification of ‘54149 in plasma samples. A solution of Prometryn (100 ng/ml in water-methanol mixture 1:9, v/v) was used as internal standard for quantification of cinacalcet in plasma samples. A solution of imipramine (50 ng/ml in water-methanol mixture 1:9, v/v) was used as an internal standard for quantification of evocalcet in plasma samples.

Data analysis

C_max and time for the peak plasma concentration (T_max) were the observed values. The areas under the concentration time curve (AUC_last and AUC_inf) were calculated by the linear trapezoidal rule. The terminal elimination rate constant, ke was determined by regression analysis of the linear terminal portion of the log plasma concentration-time curve. Mean, SD, and percent coefficient of variation (CV) were calculated for each analyte.

Serum calcium measurement

Serum calcium level was determined using commercial kits according to the manufacturer’s instructions. The principle of the method is the ability of calcium to form a blue-colored complex with Arsenazo III dye at neutral pH, the intensity of which is proportional to the concentration of calcium. Interference with magnesium is eliminated by the addition of 8-hydroxyquinoline-5-sulfonic acid. The reproducibility of the experiment was demonstrated with a CV of 2.91%.

Animal studies

All animal studies were performed on 12- to 16-week-old male C57/B6 mice (Jackson Laboratory; Bar Harbor, ME, USA), which were approved by the Institutional Animal Care and Use Committee of the San Francisco Department of Veteran Affairs Medical Center (protocol nos. 2021–005 and 2021–016). For the latter studies, test compounds with specified doses were injected subcutaneously at six different time points (15, 30, 60, 120, 240, and 480 min), followed by isoflurane overdose before blood collections by cardiac puncture. Sera were prepared by centrifugation (2000g) in a microtainer (Becton Dickinson, SST 365967) and assayed for PTH levels by enzyme-linked immunosorbent assay (ELISA) (Quidel, 60–2305) and total calcium using Alfa Wassermann ACE Axcel Vet Chemistry Analyzer.

Ex vivo parathyroid gland culture

Mouse parathyroid glands were isolated from 4-week-old male C57/B6 mice, dissected free of thyroid and surrounding fibrous tissues, and cultured to assess PTH secretion rate (ng per gland per hour) and Ca²⁺ set point (extracellular calcium, [Ca²⁺]_e, needed to suppress 50% of PTH_max) (8, 78). Briefly, parathyroid glands were incubated sequentially with a series of Dulbecco’s modified Eagle’s medium (DMEM) containing increasing concentrations of PAM at 0.75 mM calcium or containing increasing [Ca²⁺]_e with PAM (50 or 500 nM) or without PAM in preparation for testing. Intact PTH levels in culture media were assessed by ELISA and used to calculate the EC or Ca²⁺ 50 set points for each PAM.

Data and materials availability:

Code for DOCK3.7 and DOCK3.8 are available without charge for academic use from https://dock.compbio.ucsf.edu/. All underlying data from this study are included among the primary figures and tables and in the supplementary materials. Relevant files relating to CaSR virtual screening can be found in https://lsd.docking.org/. All molecules tested are available from Enamine and may be accessed via their ZINC numbers (table S1). Plasmids and reagents to conduct BRET signaling assays are available from corresponding author G.S. The cryo-EM maps and corresponding coordinates are available in the Electron Microscopy Data Bank (EMDB) and Protein Data Bank (PDB) under the accession codes EMD-45127 and 9C1P (CaSR-‘6218) and EMD-45156 and 9C2F (CaSR-‘54149), respectively.