Structure-based discovery of β₂-adrenergic receptor ligands

The Binding Site of β₂AR.

The epinephrine binding site of β₂AR presents a pocket almost ideally suited to docking. It is a narrow and deep cleft that is largely concealed from solvent (Fig. 1Upper). Because the cleft is narrow, ligands can interact with both walls, offering the possibility for van der Waals contacts that span at least one principal dimension of the site. Most of the residues lining the site are hydrophobic, which contributes to potential affinity, whereas a few polar residues allow for strong directional constraints through electrostatic interactions. Most notably among these latter residues, Asp-113 [3.32 in Ballesteros–Weinstein notation (19), given as superscript hereafter] forms a salt bridge with most known ligands. Located at the back of the pocket is Ser-203^5.42, which hydrogen bonds with carazolol, the ligand present in the X-ray structure. Also Gln-312^7.39, adjacent to Asp-113^3.32, is known to accept hydrogen bonds from the β-hydroxy-amine motif, which is prevalent among β-adrenergic modulators (Fig. 1 Lower).

Docking the ZINC Database.

The 972,608 molecules of the lead-like subset of the ZINC database (20) were docked into the carazolol-bound structure of β₂AR. DOCK found poses without major clashes for 919,549 of these molecules. These were sampled in, on average, 14,930 orientations and, for each of these, 30,834 conformations (4.6·10⁸ poses, on average, for each of the 919,549 molecules fit into the site). The time of calculation was 3.9 h on 468 CPUs. The molecules were scored for receptor complementarity based on the sum of their van der Waals [using the AMBER potential (12)] and electrostatic interaction energies [using ligand probe charges in an electrostatic potential calculated by DelPhi (13, 21, 22)], corrected for ligand desolvation (adapted from ref. 14).

Based on their docking scores, the compounds were ranked from best to worst fitting, and all of the compounds prioritized for experimental testing were picked from among the top-ranking 500 molecules, or 0.05% of the docked library. In addition, these molecules were inspected visually for features not captured in the docking calculation, such as chemical diversity, actual commercial availability, and an overall balance between polar and nonpolar complementarity to the binding site. This is our common practice in prosecuting docking screens, and these criteria are consistent with previous studies. Twenty-five molecules were ultimately selected for experimental testing. To evaluate how these 25 represented the chemotypes present among the top 500 ranking molecules, we clustered these latter molecules using FCFP4 fingerprints with an average of 10 molecules per cluster and a maximum intracluster distance of 0.7 (23). This yielded 50 clusters; the 25 compounds chosen for testing fell into 13 of these, whereas what turned out to be the 6 true ligands originated from four of them (Fig. 2).

On closer inspection of these 25 selected molecules, we noticed that mutual topological resemblance, similar binding modes or physicochemical properties, would allow us to group them into 4 classes that captured common themes in their interaction with the receptor. We would like to emphasize that this classification was done a posteriori and that class membership was not a necessary condition for a molecule to be selected; it was merely a useful way to think about the compounds. “Classic” compounds featured most of the pharmacophoric characteristics of known antagonists, i.e., an aromatic ring system and an aliphatic amino group (typically charged at neutral pH). These compounds' docked poses usually overlapped well with the crystallographic pose of carazolol. Second, “bridge” compounds contained an aromatic ring system that overlapped with the location of the carbazole moiety of carazolol, but did not interact with Asp-113^3.32. Instead, they connected the 2 walls of the binding site by hydrogen bonding with Thr-195^5.34 and Tyr-308^7.35, respectively. Third, “sulfonamide” compounds shared a common aromatic ring–sulfonamide–aliphatic nitrogen motif, with binding modes that resembled that of carazolol. Finally, the fourth class contained molecules with cationic nitrogens in ring scaffolds that have not previously been described among adrenergic ligands, or indeed among ligands for aminergic GPCRs. For the 25 compounds tested, the number in each class was 12, 9, 2, and 2, respectively, which in part reflects their relative abundance in the top 500, but also their commercial availability (Fig. 3).

Competition Assay and Similarity to Known Antagonists.

All 25 compounds were tested in a radioligand displacement assay at 20 μM, and 8 molecules measurably displaced ³H-dihydroalprenolol (DHA). Of these, the 6 compounds with percentage of displacement >10% were further characterized to obtain full DHA competition binding curves (Table 1 and Fig. 4). The measured percentage of displacement values obtained were: 100 (compound 1), 56 (compound 2), 85 (compound 3), 79 (compound 4), 67 (compound 5), and 39 (compound 6). All six had K_i values of <4 μM, with the most potent compound, 1, binding with a K_i of only 9 nM (95% confidence interval 7–10 nM). For this compound, only a racemic mixture was available for experiments. Because adrenergic receptors typically prefer one chiral form for compounds similar to 1, it is likely that the affinity of one of the enantiomers is even higher. Although docking cannot reliably rank order ligands by affinity, we note that 1 has the highest rank in DOCK score of all tested compounds, ranking 9th of the ≈1 million compounds.

The 6 confirmed ligands originated from 2 classes. Four of them, including compound 1, belong to the classic category: their docking geometry corresponds to that of carazolol and more broadly to what is expected for a β₂AR ligand. Fig. 3 shows, as examples of canonical hits, the predicted poses for 1 in A and 2 in B. Compound 1, in particular, overlaps almost perfectly with carazolol and maintains all of its polar interactions. There are 2 main differences, one being the smaller aromatic ring system of compound 1 and the other one the ethoxycarbonyl substituent on this ring system. The latter is not involved in any obvious polar interactions in the docked pose. It can be speculated, however, that the side chain of Ser-204^5.43 will rotate to form a hydrogen bond with the carbonyl oxygen of compound 1. As discussed below, compound 1 appears to be the most efficacious inverse agonist yet discovered for β₂AR, but how this relates to these interactions remains unclear, because inverse agonism is convoluted with conformational changes distal from the orthosteric site.

The remaining 2 ligands (5 and 6) contain cyclic cations that are predicted to interact with Asp-113^3.32 in manners previously unknown for adrenergic ligands. Compound 5 bears a constitutive positive charge and does not feature a hydroxy group. Because of this positive charge, which is dislocated between the 2 nitrogen atoms, the alkyl substituents on them are positively polarized. In the docking pose, these alkyl substituents form charge–charge interactions with Asp-113^3.32 at 3.2 Å and 3.4 Å, respectively (Fig. 3C). Similarly, compound 6 interacts with Asp-113^3.32 through the N-pyrrolidinomethylene substituent and the hydroxy group that is in the ortho position to it on the benzofuran moiety (Fig. 3D). Whereas this interaction resembles the classic aliphatic amine-hydroxyl motif in its spatial arrangement, it has not been described as a ligand chemotype.

The similarities and dissimilarities of these new ligands may be quantified by comparing their topological fingerprints with those of the 8,053 adrenergic ligands of the WOrld of Molecular BioAcTivity (WOMBAT) database (24) (see Materials and Methods and Table 1, column 5). Consistent with their categorization, the 4 classic molecules broadly resemble known adrenergic ligands. This can be seen by simply comparing them visually with their nearest neighbors among the WOMBAT ligands (Table 1) or, quantitatively, by calculating their Tanimoto coefficients to these nearest neighbors (their Tanimoto coefficients, based on ECFP4 fingerprints, range from 0.33 to 0.41). On the other hand, compounds 5 and 6 have low Tanimoto similarity values to any of the compounds in the database, and by visual inspection do not resemble their nearest neighbor in that database. These 2 molecules, although certainly cationic, seem to be novel chemotypes not previously explored. At the same time, it must be admitted that the most novel chemotypes to emerge from the docking, the bridge and sulfonamide compounds, were all nonbinders (Fig. 3 E and F).

Testing the Binding Mode of the Aryl Cations.

To test the predicted binding mode of 5, nine conservative derivatives were docked with an eye to identifying those that could not be fit in the bound geometry. From these calculations, the N-butyl derivative 7 (Table 1 and supporting information (SI) Fig. S1) appeared to clash with β₂AR in a geometry that would otherwise maintain the key hydrophobic interactions between the quinoline moiety and residues Val-114^3.33, Phe-289^6.51 and Phe-290^6.52 (Fig. 1 Lower). Consistent with this view, 7 does not detectably displace bound radioligand on experimental testing.

Ligand Efficacy.

In the functional assay, 5 of the 6 ligands behaved as inverse agonists (Fig. 5); the activity of the sixth ligand, compound 5, could not be classified because of nonspecific effects on G_αs, even though its affinity was clear by ligand displacement assay. It is interesting to note that compound 1 exhibits inverse agonist activity as good or better than ICI 118551, heretofore the strongest and most potent inverse agonist for β₂AR. The specificity of the inverse agonism of compound 1 was confirmed by blocking this activity with alprenolol. This behavior of the compounds is consistent with the binding site adopting a conformation adapted to the inverse agonist with which it was crystallized, carazolol, a point to which we will return.

Library Bias.

We wished to investigate how biased the docking library might be toward GPCR ligands. The lead-like subset of the ZINC database (20) contains 972,608 molecules and we compared each of these with the annotated ligands of the WOMBAT database (24) using the Similarity Ensemble Approach (25). At an E-value (confidence level) of 1·10⁻¹⁰, 103,563 molecules or 10.6% have an assignment to a GPCR. This compares with 41,182 (4.2%), 22,198 (2.3%), and 9,041 (0.9%) molecules that were assigned to kinases, proteases, and ligand-gated ion channels (LGICs), respectively. Comparing β₂AR ligands specifically with those of an enzyme we have studied in the lab, β-lactamase TEM-1, there were 864 molecules in the ZINC subset with E-values of <1·10⁻¹⁰ for β₂AR, but only 12 for TEM-1.

Protein Preparation for Docking.

The structure of the β₂AR in complex with carazolol [PDB ID code 2RH1 (9, 10)] was used in the docking calculations. All water, organic solvent, ligand, and lipid molecules were removed. The T4-lysozyme insertion in place of loop 3 was also removed, because it carried a net formal charge of +9 that is not present in the native receptor. Protons were assigned automatically and in such a way that the protonation states of side chains and termini corresponded to pH 7. Orientations for hydroxy groups in selected binding site residues were modified to conform to the proton positions determined by the HBUILD module in CHARMM (30).

Docking Calculations.

All calculations were performed with DOCK3.5.54 against the β₂AR/carazolol complex determined by crystallography, and screening the ≈1 million compounds of the lead-like subset of ZINC (see SI Methods for details).

Similarity Calculations.

The similarity calculations were performed with Pipeline Pilot (23). As a reference database, we used a subset of the WOMBAT 2006.2 database (24) containing the molecules that had been annotated as interacting with adrenergic GPCRs. Both the database and the query molecules were generated from SMILES and compared by using ECFP4 (23) fingerprints and the Tanimoto coefficient.

Competition and Functional Assays.

All K_i values were determined based on radioligand displacement assays by using ≈60 fmol of ³H-dihydroalprenolol preincubated with the receptor. Purified unliganded β₂AR and tethered-G_αs protein was used in functional assays to determine ligand efficacy (see SI Methods for details).

Library Bias.

Similarities of the library molecules to sets of known ligands of the targets in WOMBAT (24), version 2006.2, were assessed with the Similarity Ensemble Approach (SEA) (25). This method yields a predicted target for every query molecule based on the similarity of the query to the set of known ligands of the respective target. Library bias was assessed by counting the numbers of molecules with a certain target class assignment, i.e., significant similarity of a molecule with the target ligand set. The criterion for a significant assignment was an E value of <1·10⁻¹⁰. We then calculated cumulated predictions for targets of the GPCR, protease, kinase, and LGIC classes.