Conformational selectivity of ligands depends on their pharmacological profile

In this first section, we explore the predicted conformational selectivity for a number of ligands, as a function of their pharmacological profiles. To do this, a thermodynamic cycle was designed to estimate, for a given molecular pair of e.g. agonist and antagonist, the relative affinities between the two relevant conformational states of the A_2AAR. The cycle is solved by subtracting the corresponding binding free energies (ΔG_b), calculated via an FEP transformation (agonist → antagonist) performed in the inactive state (ΔG_b,R), from the same transformation performed in the active state (ΔG_b,R*, vertical legs in Fig 1B). The difference between these estimated affinities would correspond to the difference in the conformational selectivity between these two ligands (ΔΔG_cs(L1−L2)) as:

Such model assumes that the pharmacological profile of a ligand (i.e. agonist, antagonist) is dictated by a more or less biased selectivity towards the active conformation, respectively. In each case, the pair of agonist and antagonist molecules to compare should share enough chemical similarity (e.g. bearing the same chemical scaffold) while maintaining a sufficiently distinct pharmacological profile. This scheme was consequently used to investigate the variability in the pharmacological profile within three chemical scaffolds: i) the classical ribose-containing agonists, such as NECA; ii) 2-phenylaminothiazolo[5,4-d]pyrimidines, which were recently characterized as partial agonists depending on the substituent in position 7 [28]; and iii) partial agonists derived from the 4-phenylpyridine scaffold of LUF5834 [29]. In each case, the (partial/full) agonist was compared to a chemical analog that behaves as a (neutral) antagonist (Fig 2).

The experimentally determined coordinates of classical agonists binding to the A_2AAR (Fig 3A and 3B) were used as a starting configuration for the active-state simulations (denoted with an asterisk), while the corresponding binding mode to the inactive A_2AAR was generated by receptor superposition. The binding mode of 9-cyclopentyadenine (Cyp-Ade) antagonist where the ribose group of adenosine is replaced by a cyclopentane [30], was generated using a flexible ligand alignment strategy (see Methods). The initial configuration of the two partial agonist chemotypes was generated via docking to the A_2AAR*, showing occupancy of the ribose pocket (Fig 3C and 3D), while the corresponding antagonist analogues were modeled in an analogous binding orientation in the A_2AAR inactive structure. All compounds shared similar interactions with N253^6.55 and F168^EL2, irrespective of the receptor conformation or ligand chemotype.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Binding mode of four adenosine A_2AAR agonists with different efficacies.

(A, B) Top panels show the experimental pose of full agonists adenosine (A, PDB code 2YDV [31]) and CGS21680 (B, PDB code 4UG2 [13]). (C, D) Bottom panels show the docking model obtained for partial agonists LUF5834 (C) [29] and 10n (D) [28]. The key residues common for ligand-receptor interactions (N253^6.55, E169^EL2 and F168^EL2) [18] are shown in sticks. The ribose binding site is denoted in a dotted circle, with interacting residues associated with ligand activation shown in sticks.

https://doi.org/10.1371/journal.pcbi.1009152.g003

The full agonists NECA and adenosine (ADO) show different experimental efficacies for the activation of A_2AAR. Consequently, NECA was set as a reference with a maximum efficacy of 100%, with adenosine having 45% efficacy relative to NECA [29], and the neutral antagonist Cyp-Ade having 0% efficacy. A qualitative descriptor, Δ-efficacy, can be defined as the positive difference in % efficacy values between pairs of (partial) agonist and antagonist compounds. In analogy, the thermodynamic cycle depicted in Fig 1B allows to estimate the relative preference for the active conformation for the same pair of molecules, a property that we will try to correlate with the corresponding efficacy shifts. Here, one has to note that, due to their fundamentally different formulation, a full quantitative correlation between the experimental (Δ-efficacy, percentage) and the calculated (ΔΔG, logarithmic) descriptors is not expected. However, a qualitative correlation would indicate that our end-state modeling of ligand efficacy would be useful to explain efficacy shifts between molecule pairs, and thus be of potential interest to further predict ligand pharmacological profiles. In other words, the differences in the calculated ΔΔG might be used to classify the compounds on the basis of their expected efficacy gain from a reference molecule. According to the scheme in Fig 1B, an expected positive value in the calculated ΔΔG would indicate the expected preference of the agonist for the active state. This was indeed the case for the NECA and ADO perturbations to the antagonist Cyp-Ade (Fig 4 and S1 Table). Moreover, the magnitude of the calculated ΔΔG value is double for the NECA transformation as compared to the case of adenosine, which indicates the correct ranking of this class of compounds according to their expected efficacy.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Calculated relative affinities for the active vs inactive receptor conformation.

Relative affinities expressed as ΔΔG = ΔG_b,R* -ΔG_b,R (kcal/mol, blue bars, average values ± SEM) for pairs of agonists/antagonist of analogous chemotype (depicted in the top, with the chemical groups varied along the FEP simulations in red). For each ligand pair, the relative shift in experimental efficacy is shown as Δ-efficacy (orange bars, E_max,ago-E_max,antago, in % ± SEM), with corresponding Emax values reported relative to the reference full agonists NECA [28], or CGS21680 -for the LUF chemotype [29], see text.

https://doi.org/10.1371/journal.pcbi.1009152.g004

We then moved on to examine the molecular determinants of the agonistic properties for a series of 7-(prolinol-N-yl)-2-phenylaminothiazolo[5,4-d]pyrimidines on the A_2AAR [28]. The SAR and earlier molecular modeling of these series revealed that the prolinol substituent was essential to maintain the partial agonist profile, presumably by partially mimicking the ribose interactions (see compounds, 10b, 10d, 10g, 10j, 10l, 10n, Fig 2), in contrast to the corresponding pyrrolidine substituted compounds (10a, 10c, 10f, 10i, 10k and 10m, Fig 2) which are all neutral antagonists (i.e. with efficacy not significantly different from 0%) [28]. Following our preliminary results on this chemotype [32], we systematically applied the thermodynamic cycle in Fig 1B to all corresponding partial agonist/antagonist pairs within this series, the results shown in Fig 4 and S1 Table. A deeper look into the experimental data shows that the efficacy of the partial agonists is modulated by decorations of the exocyclic amino group (R, see Fig 2). Our calculations indicate that compound 10m would show the highest preference for the active state of the receptor, followed by compound 10k (Fig 4), in line with their efficacy rank experimentally observed. Our modeling shows a tendency on prolinol-containing compounds to favor the binding to the active receptor, as compared to their pyrrolidine substituted compounds, with the exception of the pairs and 10g/f and, to some extent, 10b/a where no significant conformational selection is calculated (Fig 4). Explaining the role of this substituent in the partial agonist profile of this chemotype. The structural interpretation of this data reinforces the role of the hydroxyl group in the 7-prolinol in engaging the residues that trigger A_2AAR activation (Fig 3), explaining the importance of this substituent for a partial agonist profile.

The third chemotype here explored is derived from the 4-phenylpyridine scaffold, characteristic of a family of A_2AAR partial agonists represented by LUF5833 (Fig 2) [29]. Hydroxy decorations on the phenyl ring have been shown to affect ligand efficacy: the p-OH-phenyl in LUF5834 is equipotent to the unsubstituted LUF5833 (55% ± 15% as compared to the efficacy of the reference full agonist CGS21680), while the m-OH-phenyl in LUF3835 results in an increased value of this relative efficacy of 80% ± 10% (see Fig 2) [19]. In their original report, Lane et al. postulated that these compounds bind with the phenyl substituent deep in the binding pocket of the A_2AAR, making different interactions with activation-related residues as compared to ribosidic agonists [19]. We herein examined if such a binding mode hypothesis (Fig 3) could explain the differences in ligand efficacy following the FEP approach outlined in Fig 1B. In this case, the dehydroxylated partial agonist LUF5833 was used as the reference ligand in two pair comparisons. Since the three compounds show comparable experimental binding affinities for the A_2AAR, the experimental increase in efficacy for a given derivative can be directly correlated with its increased relative affinity for the active (R*) over inactive (R) receptor conformation, which is precisely the output of our calculations. The results (Fig 4 and S1 Table) indicate that the introduction of a m-OH-phenyl in LUF5835 would lead to a significant conformational selection for the active receptor, in line with the experimental gain of approximately 30% efficacy [19]. The structural interpretation of this predicted efficacy increase is located on the hydrogen bond between the OH group in meta position of LUF5835 with T88^3.36, observed in the simulations with R*, an interaction that is well known to be involved in agonist recognition [18]. However, the introduction of a p-OH-phenyl substitution in the reference ligand leads to a reduced predicted efficacy, since this substituent would not be making any preferred interaction in the active state as compared to the inactive, with the experimental data showing equipotency of this compound pair.

During the preparation of this manuscript, a new crystal structure of compound LUF5833 in complex with the inactive conformation of the A_2AAR was published [33]. Comparison of our model with this structure revealed an almost identical inactive conformation for the receptor (RMSD_A2AAR = 0.78 Å for the Cα trace). The ligand position was also in good agreement with our MD simulations with an overall RMSD_LUF5833 = 2.20 ± 0.45 Å, calculated over the last 10% of the FEP trajectories. Specifically, most variability was located on the flexible 1H-imidazol- 2-ylmethylsulfanyl substituent oriented towards the extracellular cavity, with the core moiety being more stable along the MD simulation (RMSD_{LUF5833_core} = 1.38 ± 0.51 Å). The structure also demonstrates that the partial agonist LUF5833 can actually bind to the inactive conformation of the receptor, which supports the line of reasoning of this study.

The effect of point mutations on basal activity and conformational selectivity

According to the experimental mutagenesis data, the activation trigger induced by the partial agonist LUF5834 would involve different interactions with residues in the ribose binding site, as compared to ribose-containing full agonists [19]. In that study, CGS21680, a C2-substituted variation of NECA (Figs 2 and 3), was used as a reference full agonist in the pharmacological characterization of the LUF series of compounds. Particularly intriguing was the effect of two mutations, T88^3.36A and S277^7.42A, on the modulation of the efficacy of these two molecules. While the binding affinity and potency of CGS21680 was severely reduced by both mutations, the potency of LUF5834 was unaltered or even slightly increased [19].

Consequently, we wondered about the molecular mechanism of the different mutational-induced shifts on the internal efficacy of the two ligands. This is indeed a complex question, as one can imagine two mechanisms via which a mutation can modulate the efficacy of an agonist: On the one hand, the mutation might affect the basal activity of the receptor, by selectively stabilizing one conformation. The T88^3.36A mutant is a CIM that reduces the basal activity [19], by selective thermal stabilization of the antagonist-bound conformation [20]; however no significant effect was observed on the basal activity for the S277^7.42A mutant [19]. On the other hand, the same mutation can directly modulate the binding affinity of the (partial) agonist for the effective, active conformation. While there is no experimental data for the conformational affinity, our approach allows to model each of these effects independently and combine them a posteriori, taking advantage of the possibility of combining two thermodynamic cycles if they share a common leg.

As shown in Fig 1C, the effect of a mutation on the predicted basal activity can be modeled by designing a thermodynamic cycle that represents the effect of the corresponding Ala mutation on the conformational selectivity. The cycle can be solved through FEP simulations of the vertical legs, i.e. annihilation of the sidechain of interest (wt → Ala perturbation) in each receptor state (ΔG_m,R and ΔG_m,R* for the inactive and active state, respectively) as:

Such model assumes that the mutational effects on the GPCR basal activity are indeed due to differences on the conformational selection of the receptor (ΔG_cs, for wt or mut versions of the receptor). This model can be combined with a second thermodynamic cycle, accounting for the mutational shifts in ligand binding affinity for the active conformation, which we extensively used to explain mutagenesis effects on A_2AAR agonist binding [34]. The addition of these two cycles (Fig 5A) should yield as a net result the effect the mutation on the ligand internal efficacy, estimated as the difference between the two vertical legs delimiting the combined thermodynamic cycle in Fig 5A, i.e. (Fig 5B, solid blue columns). It should be noted that, in this case, the experimental parameter that we are trying to match with these relative free energy calculations is not the maximum efficacy (%E_max), but the internal efficacy of each compound (EC₅₀), which can vary upon receptor mutations. Thus, a numerical (quantitative) correlation between calculated and experimental values can be attempted in this case, by expressing the experimental shift in EC₅₀ values induced by a mutation as (Fig 5B, solid orange columns). Finally, the additional calculation of the shared vertical leg, ΔG_R*, which is canceled in the combined cycle, allows to separate the effect on basal activity (ΔΔG_R*, Fig 5A, left side, Fig 5B, dense dashed columns) and on ligand affinity for R* (ΔΔG_L-R*, Fig 5A, right side; Fig 5B, light dashed columns). As we will see, this can provide valuable information for the interpretation of the calculated data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Effect of protein mutations on ligand efficacy.

(A) Binding mode of the agonist CGS21680 to the WT A_2AAR (left), T88A (center) and S277A (right) mutant forms of this receptor (mutations indicated in red), simulated by FEP simulations (H-bonds in yellow dashed lines with residues involved in sticks). (B) Thermodynamic cycles representing the effect of a point mutation (mut) on the distribution of inactive (R) and active (R*) states of the receptor (left side, dense dashed), and on the affinity of a ligand (L) for the active state (right side, light dashed). The combination of the two thermodynamic cycles would yield the net effect of the mutation ligand efficacy. (C) Calculated effects of point mutations on the A_2AAR constitutive activity (ΔΔG_{R* →R}, dense dashed columns), and on the ligand relative affinity for R* (ΔΔG_L-R*, light dashed columns), following the corresponding thermodynamic cycles depicted in (A). The overall effect of the mutation on the shift in ligand efficacy, (, solid blue) is calculated by combination of these two values, showing correlation with the experimental values (, solid orange, see text).

https://doi.org/10.1371/journal.pcbi.1009152.g005

We first looked into the effects on the basal activity of the receptor of the T88^3.36A and S277^7.42A mutations. According to our model, the CIM T88^3.36A should increase the relative stability of the inactive receptor, which is precisely the outcome of the corresponding FEP simulations (Fig 5B). The active state of the T88^3.36A mutant is significantly less stable than the wt, with ΔΔG_{R* →R} = 3.98 kcal/mol (S1 Table). In contrast, a similar analysis of the S277^7.42A mutation shows a negligible value for the calculated values of ΔΔG_{R* →R}, (Fig 5B and S1 Table), meaning that the conformational equilibrium should not be affected by the mutation, in line with the experimentally observed neutral effect of this mutation on the basal activity of the receptor [19].

The overall mutational effects on the modulation of ligand internal efficacies were then calculated and compared to the experimental efficacy shifts in each case (Fig 5B, solid columns). One can observe that the modeled T88^3.36A mutation does not significantly affect the predicted internal efficacy of LUF5834 (, Fig 5A and S2 Table) in line with the experimental data () [19]. A deeper look at the data allows us to envisage a mechanism for this neutral effect, since the relative increase in affinity for the active state is compensated by a decrease in the population of this active state induced by the mutation (light and dense dashed bars, respectively, in Fig 5B). Conversely, the model indicates that the S277^7.42A mutation does not affect the conformational equilibrium of the receptor, and in this case the predicted increase in affinity of LUF5834 for the active state is translated into a net increase of the efficacy of this ligand upon mutation (), in qualitative agreement with the experiment data (,) [19]. For the full agonist CGS21680, both mutations resulted in a similar decrease in estimated ligand affinities, in agreement with our previous computational modeling of the mutagenesis data for ribose-containing ligands [34]. However, this effect is amplified by the diminished availability of the active state upon T88^3.36A mutation, resulting in a much more drastic estimated reduction of CGS21680 efficacy () as compared to the S277^7.42A mutant (). These results allow a correct ranking of this pair of mutations in terms of how they affect the internal efficacy of this ligand, in line with the experimental observation where the reduction in the EC50 of CGS21680 was 10 times higher for the T88^3.36A mutant [19].

Structure preparation, membrane insertion and equilibration

The high-resolution crystal structure of the A₂AR with antagonist ZM241385 (PDB code 4EIY [12]) was used as receptor starting structure for the inactive state, whereas the same receptor in complex with agonist NECA (PDB code 2YDV [31]) was used for the active (like) state. In the first case, the engineered BRIL fusion protein was removed and missing loops (C-terminal fragment of EL2 and most of EL3) were modelled, and protonation states of residues assigned, as described elsewhere [55]. Notably, this included the allosteric sodium ion and the charged form of the coordinating residue Asp50^2.50. The active-like structure, on the other side, cannot accommodate this sodium ion though the protonation state of Asp50^2.50 remains charged [55]. The structure used here (2YDV) contained four thermostabilizing mutations L48A^2.46, A54L^2.52, T65A^2.63 and Q89A^3.37, which we reverted prior to a similar assignment of protonation states as described previously [56], and finally it was aligned to the inactive structure of the receptor. The selection of this simpler structure to represent R* in our models over the fully active G-protein bound A_2AAR (available with PDB code 5G53), was based on the fact that the RMSD between both binding sites (i.e. considering the spheres to simulate for FEP calculations, see next section) was as low as 0.69 Å for the Cα trace. In addition, it proved previously a very useful framework to reproduce, with high correlation, the effect of point mutations to agonist binding on this same structure [34, 56]. The 3D coordinates for the ligands were generated with LigPrep and subsequently docked to the prepared A_2AAR structure using Glide [57], and the antagonists structures by flexible ligand alignment to their reference agonist compound. Apo structures were generated by removing the ligand, but keeping crystallographic waters, and subsequently embedded in a solvated membrane environment using PyMemDyn [58]. This protocol embeds a structure in a pre-equilibrated POPC (1- palmitoyl-2-oleoyl phosphatidylcholine) membrane model such that the TM bundle is parallel to the vertical axis of the membrane. The system is then soaked with bulk water and inserted into a hexagonal prism-shaped box that is energy-minimized and carefully equilibrated during 5 ns, following the PyMemDyn protocol described elsewhere [58]. The standard OPLS all-atom (OPLS-AA) force field is used for all residues [59], and parameters for membrane lipids were taken from the Berger united-atom model [60]. The corresponding equilibrated holo structures were generated by restoring the docked ligands and removing overlapping water molecules.

FEP simulations

The receptor-ligand structures in the equilibrated membrane were subsequently transferred to the MD package Q (version used available at https://github.com/esguerra/q6), in order to perform FEP calculations under spherical boundary conditions [61]. A 50 Å diameter sphere was centered on the center of geometry of ZM241385 (or equivalent point in the remaining structures), where solvent atoms are subject to polarization and radial restrains using the surface constrained all-atom solvent (SCAAS) model to mimic the properties of bulk water at the sphere surface [62]. Atoms lying outside the simulation sphere are tightly constrained (200 kcal/mol/Ų force constant) and excluded from the calculation of non-bonded interactions. Within the simulation sphere, long range electrostatics interactions beyond a 10 Å cut off were treated with the local reaction field method [63], except for the atoms undergoing the FEP transformation, where no cutoff was applied. Solvent bond and angles were constrained using the SHAKE algorithm [64]. All ionizable residues outside the sphere and those within the boundary were considered in their neutral form as described elsewhere [25]. Residue parameters were translated from the latest version of the OPLSAA/M force field [59], whereas ligand OPLS2005 parameters were retrieved from Schrodinger’s ffld_server [65], and translated to Q following the QligFEP protocol [25]. The simulation sphere was heated from 0.1 to 298 K during a first equilibration period of 0.61 nanoseconds, where the initial restraint of 25 kcal/mol·Å² imposed on all heavy atoms was slowly released. Thereafter the system was subject to unrestrained MD simulations, starting with a 0.25 nanosecond unbiased equilibration period which is followed by the FEP sampling, applying different protocols for sidechain and ligand perturbations as detailed below. In both cases, atom transformations occur between initial and ending states, evenly divided into a number of λ windows that depend on the FEP protocol adopted (see below). This sampling is replicated in 10 independent MD simulations with different initial velocities, each of them consisting of 10 ps sampling per λ window using a 1 fs time step in all cases.

The generalized FEP protocol for amino acid mutations, QresFEP [26,66] was used to estimate the effects of single point mutations on ligand binding (available at https://github.com/qusers/qligfep). Briefly, QresFEP is a single-topology FEP protocol that divides the sidechain perturbation to alanine into separate stages, where atom annihilations occur gradually for each charge group (as defined on the OPLS force field), starting from the most topologically distant from the C_β atom, in four consecutive stages [66]: 1) the partial charges are initially removed, 2) van der Waals potentials are transformed into smoother soft-core potentials, 3) annihilation of the corresponding group of atoms., 4) restoring the partial charges of the final species. The number of perturbation stages needed for the full annihilation depends on the nature of the sidechain involved, in this case ranging from four (Ser) to five (Thr), where each of the subsequent stages is evenly divided into 20 λ windows. To fulfill a thermodynamic cycle, the same sidechain annihilation is simulated in the apo structure of the protein, so that the energetic difference between these two processes equals the binding affinity shift due to the mutation. It follows that the sampling time for the sidechain perturbation here considered was 8–10 ns. These sampling times have been repeatedly shown to be sufficient for conformational sampling at equilibrium [26,66], assuming that only minor local conformational changes occur as is the case of the models here presented.

All ligand perturbations were performed with our dual-topology QligFEP protocol [25], where the full transformation of one ligand into another is performed along a linear λ sampling consisting of 50 windows (code available at https://github.com/qusers/qligfep). In order to fulfill the thermodynamic cycle, the simulation is performed in the two receptor states to compare, i.e. active (R*) and inactive (R), and in this case the difference between the ligand transformation in the two states equals the difference in relative binding affinity between the two ligands.

In both residue and ligand transformations, the sampling time was considered sufficient as judged from the low s.e.m. estimated over replica simulations (see Results). Relative binding free energies were estimated by solving the thermodynamic cycle utilizing the Bennet acceptance ratio method (BAR) [67] as (1) where the constants C_i are optimized iteratively so that the two ensemble averages become equal, yielding ΔG_i = C_i. Average values ± SEM are reported from the 10 independent MD simulations in each relevant state.

Starting structures, input files and submission scripts needed to reproduce the simulations, together with raw output data, are available at https://zenodo.org/record/5602896#.YXlGHZ5ByF4.