Local Enrichment of Unsaturated Chains around the A_2A Adenosine Receptor

Abstract

Two 15 μs all-atom simulations of the A_2A adenosine receptor were obtained in a ternary mixture of cholesterol, saturated phosphatidylcholine lipids, and unsaturated phosphatidylcholine lipids. An analysis of local lipid solvation is reported on the basis of a Voronoi tessellation of the upper and lower leaflets, identifying first and second solvation shells. The local environments of both the inactive state and the partially active state of the receptor are significantly enriched with unsaturated chains but depleted of cholesterol and saturated chains, relative to the bulk membrane composition. In spite of the local depletion of cholesterol, the partially active receptor binds cholesterol at three locations during the entire simulation trajectory. These long-lived interactions represent the extreme of a very broad distribution of first-solvation shell lipid lifetimes, confounding sharp distinctions between lipid interactions. The broad distributions of lifetimes also make equilibrating the local lipid environment difficult, necessitating long simulation times.

Graphical Abstract

The fact that proteins may have evolved specific structural and chemical features for functional interactions with lipids has been discussed for decades.¹ Over the past 20 years, advances in membrane protein crystallization protocols have led to a proliferation of high-resolution structures of membrane proteins and, with them, reports of specifically bound lipids and cholesterol, obtained both by X-ray diffraction^2,3 and by electron microscopy of two-dimensional crystals⁴ and single proteins in nanodiscs.⁵ Mass spectrometry has also proven to be a valuable tool for identifying specifically bound lipids.⁶ In many cases, these bound lipids must survive the process of solubilization in detergent, purification, and crystallization, suggesting a high-affinity interaction. Moreover, in the structural methods, bound lipids must necessarily be in a well-defined configuration to be resolved.

Complementing experimental data, molecular dynamics simulations have found broad application in the identification of specific (sometimes called “non-annular”) lipid binding sites (for a recent comprehensive review, see ref 7). Interactions with charged lipids are often robust compared to other lipid—protein interactions and more easily understood using standard ideas regarding chemical complementarity developed in the context of small molecule binding. Examples in the literature include inositol binding to ion channels^8,9 and GPCRs¹⁰ and cardiolipin binding to respiratory complex proteins.¹¹ There are also examples of other headgroup specific interactions, such as an allosteric mechanism by which PE stabilizes the ionic lock of the β2 adrenergic receptor.¹² Cholesterol is a somewhat different case. On one hand, there is abundant structural evidence for cholesterol bound to GPCRs^13,14 and of functional consequences on signaling,^15,16 but the molecular mechanism responsible for the interaction is apparently more subtle. In retrospect, early simulation efforts to detect cholesterol binding at GPCRs were probably not long enough to yield statistically robust conclusions;^17,18 more recent work finds relatively long-lived cholesterol binding events (>5 μs),^19–21 challenging sampling by all-atom simulation. In recent work, we reported localized, specific interactions with cholesterol based on coincidence of predictions from coarse-grained (Martini) and all-atom simulations.^19,20 In the all-atom simulations, these interactions persisted for significant fractions (sometimes the entirety) of 5 and 6 μs all-atom trajectories. The robustness of these interactions is likely responsible for their reproducibility but raises questions regarding the local solvation of the receptor and whether it recruits a specific lipid environment beyond a few tightly bound cholesterols.

Distinct from specific and localized lipid–protein interactions is the hypothesis that integral membrane proteins recruit a particular lipid environment that differs from the bulk membrane composition. In the literature, this lipid environment is sometimes described as an “annulus” to distinguish it from the bulk lipid environment and is defined to be lipids that are in slower exchange with the bulk than expected for freely diffusing lipids.²² Early evidence of such a distinction came from electron spin resonance measurements of spin-labeled lipids around a Ca²⁺ ATPase.²³ Using a very nearly native membrane and protein, Soubias and Gawrisch clearly demonstrated by saturation transfer difference nuclear magnetic resonance that rhodopsin prefers solvation by polyunsaturated chains,²⁴ later confirmed by molecular simulation.²⁵ Cholesterol in particular has been extensively discussed in this context, with early reports of enrichment of cholesterol around the Ca²⁺ ATPase²⁶ and the nicotinic acetylcholine receptor²⁷ based on fluorescence quenching by brominated cholesterol. However, such experiments are significantly less common than structural evidence, due mainly to the inherent challenges of identifying relatively weak, short-lived interactions. Simulations are therefore a natural complement to experiments for the determination of preferential lipid solvation of membrane proteins. In the GPCR literature, cholesterol enrichment has been reported on the basis of molecular simulation in several cases, including the serotonin 1A receptor²⁸ and rhodopsin.²⁹

Following this tradition, two all-atom simulations of the A_2A adenosine receptor were run for 15 μs each in a ternary mixture of cholesterol, dioleoylphosphatidylcholine (DOPC), and dipalmitoylphosphatidylcholine (DPPC). In one simulation, the receptor was bound to an antagonist, and in the other, the receptor was bound to an agonist. Analysis of the local solvation of the receptor shows that (i) the lipid occupancy of the first and second solvation shells requires ~10 μs to equilibrate, (ii) the receptor recruits a local environment that is significantly enriched with unsaturated chains, and (iii) the lifetimes of lipid–protein interactions span more than four decades in time, limited at the long time end by the length of the simulation. In spite of the locally unsaturated lipid environment, the agonist-bound receptor binds three cholesterols for the entire duration of the simulation. No cholesterol binds for the entire duration of the antagonist-bound receptor. Significant differences are found when the same analysis is performed on a Martini simulation of the inactive receptor, which shows no preference for any of the three lipids in the first or second solvation shells.

METHODS

All-Atom Simulation Details.

Initial coordinates for the inactive state were based on the 1.8 Å resolution structure bound to ZM241385 [Protein Data Bank (PDB) entry 4EIY].³¹ The partially active state was based on the UK432097-bound structure in which intracellular loop 3 is replaced with a BRIL fusion construct (PDB entry 3QAK); the BRIL construct was excised, and the missing residues were rebuilt as previously described.⁴⁰ The protein was embedded in a symmetric lipid bilayer containing 504 dipalmitoylphosphatidylcholine (DPPC), 132 dioleoylphosphatidylcholine, and 273 cholesterol molecules and solvated with approximately 55000 TIP3P⁴¹ waters, one sodium ion as observed in the crystal structure, and 10 chloride ions to neutralize the system. The lipid mixture (~4:1:2 DPPC:DOPC:Chol) is in the liquid–liquid coexistence region of the phase diagram³⁰ and was motivated by a recent report by Malmstadt and co-workers indicating a change in partitioning from the liquid-disordered phase to the liquid-ordered phase upon ligand binding.¹⁵ The protein and lipids were modeled with the CHARMM36 force field,^42,43 and the ligand was modeled with the CHARMM general force field⁴⁴ with atoms typed by the ParamChem server.⁴⁵

Several equilibration steps were run with NAMD version 2.9.⁴⁶ The initial configuration was relaxed by 4000 steps of steepest descent. The protein backbone was then restrained using a force constant of 2 kcal mol⁻¹ Å⁻² while the system was heated to 295 K, reassigning velocities from a Maxwell—Boltzmann distribution every four time steps, with the simulation cell volume allowed to change semi-isotropically via a Langevin piston with a damping time scale of 0.1 ps and period of 0.2 ps.⁴⁷ An additional 20000 equilibration steps were performed, rescaling velocities every 100 steps to enforce a temperature of 295 K. Finally, the Langevin equation was integrated for 1 ns with a 1.0 fs time step, followed by 10 ns with a 2.0 fs time step and all covalently bonded hydrogens constrained by SHAKE.^31,44,48 During all relaxation steps, electrostatics were computed with the particle mesh Ewald method⁴⁹ on a 1 Å grid, with a tolerance of 10⁻⁶ and fourth-order interpolation. Lennard-Jones interactions were cut off at 10 Å and shifted to zero at 12.00 Å, ensuring both continuous potential and force.

The equilibrated binary restart files were converted to dms format for production simulation on Anton2. Force field information was added using Viparr version 4.5.34. Integration was performed under constant pressure (1 atm), temperature (295 K), and particle number with the multigrator⁵⁰ method with a 2.5 fs time step, with the temperature (295 K) controlled by a Nose-Hoover⁵¹ chain coupled every 24 time steps and pressure controlled by the Martyna—Tobias—Klein barostat (pressure of 1 atm, semi-isotropic) coupled every 480 time steps.⁵² Electrostatics was computed using the k-space Gaussian split Ewald method,⁵³ with long-range interactions computed every third time step, and hydrogens constrained by M-SHAKE.⁵⁴ Lennard-Jones interactions were cut off at ≥11 Å. Production simulations were run for 15 μs, and configurations were stored every 240 ps.

Martini Simulation Details.

The all-atom structure of the active state (described above) was converted to a coarse-grained (CG) representation using the martinize.py script, and the system was built with the insane.py script³² (both scripts freely available from http://cgmartini.nl/website). The secondary structure of the protein was restrained by an elastic network among the backbone sites³³ (cutoff distance of 0.9 nm with a force constant of 500 kJ mol⁻¹ nm⁻²), ensuring that the inactive conformation was simulated. The simulated system consists of one protein embedded in the same lipid mixture as described above. The lipids and protein were modeled with the Martini CG force field for biomolecules.^34,35 Nonbonded interactions were cut off at a distance r_cut of 1.2 nm. The Lennard-Jones potential is shifted from an r_shift of 0.9 nm to r_cut. The electrostatic potential is shifted from an r_shift of 0.0 nm to r_cut.

All simulations were performed using GROMACS simulation package version 4.5.4.³⁶ The system was energy minimized via steepest descent followed by a 20 ns integration at constant temperature, pressure, and particle number (NPT) followed by a 20 μs production simulation with a 20 fs time step, with configurations stored every 500 ps. The protein/bilayer (330 DPPC, 90 DOPC, and 180 cholesterol molecules) and the aqueous phase (10000 Martini waters) were coupled independently to external temperature baths at 295 K using a Bussi–Donadio–Parinello thermostat³⁷ with a relaxation time of 1.0 ps. The pressure was controlled by a Parinello–Rahman barostat³⁸ at 1.0 bar using a relaxation time of 12.0 ps, a semiisotropic pressure scheme, and a compressibility set to 3 × 10⁻⁴ bar⁻¹.

Average Lipid Area Density.

The last 6 μs of the trajectory was sampled every 1.92 ns. The box dimension for each frame was scaled to the average box dimension of the 6 μs trajectory. The center of mass of each lipid was assigned to a 1 Å × 1 Å pixel in each leaflet according to its lateral position, and the number of each type of lipid was averaged for each pixel, normalized by the number of lipids of each type averaged over all pixels (excluding the area of the protein).

Voronoi Analysis.

The Voronoi analysis follows the procedure developed by Beaven et al.³⁹ For the entire 15 μs trajectory, the configurations were sampled every 4.8 ns. A total of 3125 frames were then divided into 10 segments within which the statistics are totaled independently. For each leaflet, the center of mass of each lipid was used to form a 2d point set, where the lipid type of each point was tracked. After the Delaunay triangulation and protein boundary assessment had been performed, each lipid was assigned a shell number according to whether it shares a boundary with the protein (first shell), with a first shell lipid (second shell), or neither (not considered further). The number of each type of lipid was tabulated for the first and second shells within each 1.5 μs segment, and the average and standard deviation were computed.

First-Shell Lifetime Calculation.

After lipids had been assigned to first and second shells, the lifetimes of lipids in the first shell were computed by counting contiguous time spent in the first shell, with a lower cutoff of 4.8 ns, and then compiled into a histogram (i.e., a lipid must spend at least two consecutively sampled Voronoi frames in the first shell to be counted in the histogram).

RESULTS

The Local Membrane Environment Is Enriched with Unsaturated Chains.

Over the course of 15 μs, the membrane area and protein surface are well sampled by all three lipids, yet there are clear preferences for particular lipids in the immediate protein environment. Figure 1 shows the time-averaged area density for DOPC, DPPC, and cholesterol in the lower leaflet for both the active and inactive receptor states. Figure 2 shows the same analysis, but for the upper leaflet. The receptor is shown in cross section, with the cross section taken at the average position of the cholesterol rings in each leaflet. Comparison of the top and bottom panels in Figures 1 and 2 shows that the change in the shape of the receptor upon activation is confined mostly to the inner leaflet, consistent with the prevailing view of the structural rearrangements that occur upon activation.⁴⁰

Figure 1.

Time-averaged density of cholesterol (left), DOPC (middle), and DPPC (right) in the lower leaflet for the active (top) and inactive (bottom) receptor. The cross section of the protein at the average location of the cholesterol rings is colored red with the TM helices numbered 1–7 colored yellow. The scale bar shows multiplicative enrichment over the bulk average density of each species.

Figure 2.

Time-averaged density of cholesterol (left), DOPC (middle), and DPPC (right) in the upper leaflet for the active (top) and inactive (bottom) receptor. The cross section of the protein at the average location of the cholesterol rings is colored red with the TM helices numbered 1–7 colored yellow. The scale bar shows multiplicative enrichment over the bulk average density of each species.

Unsaturated chains are enriched in the immediate environment of the protein at the expense of cholesterol and saturated chains, evidenced by the darker green and black adjacent to the protein in the center column of panels in Figures 1 and 2. This is true for both leaflets and both receptor states. In contrast, the local environment is relatively depleted of cholesterol and saturated chains, which appears as a halo of white and lighter green in the first and last columns. Again, these patterns hold for both leaflets and both receptors.

In addition to the nonspecific enrichment of unsaturated chains, there also appear to be very localized regions of much higher density in the cholesterol and DOPC plots, which appear in Figures 1 and 2 as black areas, which correspond to area densities 20-fold higher than that of the bulk. For DOPC, these regions appear around both the inactive receptor and the active receptor. In contrast, only the active receptor has such localized areas of high cholesterol density, which in the lower leaflet coincide with clefts on either side of H4. No such localized regions are observed for the saturated chains. Residue-level interactions and whether these are long-lived interactions with individual lipids are discussed below, after considering the convergence of the local lipid environment.

Solvation Shell Occupancies Require at Least 10 μs To Converge.

The occupancies of the first and second solvation shells were determined by a Voronoi tessellation of the two leaflets as reported by Beaven et al. for simulations of gramicidin A³⁹ and as described in Methods. Any lipid Voronoi polyhedron that shares an edge with a protein polyhedron is identified as the first shell, and any lipid polyhedron sharing an edge with a first shell lipid is identified as the second shell. Identifying lipid solvation shells using Voronoi polyhedra rather than a simple distance criterion better represents the irregular shape of membrane proteins and the difference in size of different lipid species, particularly cholesterol.

Plotting the occupancy of the first (Figure 3) and second (Figure 4) solvation shells averaged within 10 non-overlapping equal increments reveals the time scale over which the local lipid environment equilibrates. The population of a particular lipid in a single leaflet can change by as much as 5 or 6 over the duration of the simulation, which represents a change of 30% in the first shell (of approximately 22 lipids in total in each leaflet). These changes can take more than 5 μs to occur, indicating that averages should be taken only after longer times. (For this reason, the area densities plotted in Figures 1 and 2 were averaged over the last 6 μs of the trajectories.)

Figure 3.

Lipid occupancies of the first solvation shell for the upper leaflet (top) and lower leaflet (bottom) for the active (left) and inactive (right) receptor. The left axes refer to the fractional occupancies of each lipid type, and the right axes refer to the total number of each lipid type. The points are averages within 1.5 μs trajectory windows, and the bars are the standard deviation within each window.

Figure 4.

Lipid occupancies of the second solvation shell for the upper leaflet (top) and lower leaflet (bottom) for the active (left) and inactive (right) receptor. The left axes refer to the fractional occupancies of each lipid type, and the right axes refer to the total number of each lipid type. The points are averages within 1.5 μs trajectory windows, and the bars are the standard deviation within each window.

Overall, the trend is toward enrichment of unsaturated chains in the first solvation shell, consistent with Figures 1 and 2. The mole fraction of each lipid averaged over the entire system is 0.55/0.15/0.30 DPPC/DOPC/Chol. The first solvation shell in both leaflets around both receptor states is between 45 and 55 mol % DOPC, a significant enrichment over the bulk value of 15%, and well outside the one standard deviation range plotted in Figure 3 (background shading). The saturated chains are depleted relative to the bulk value of 0.55, particularly around the inactive receptor (left panels of Figure 3), with an average in the range of 0.35–0.45 over the last 6 μs of the simulation. The saturated chain density around the active receptor (right panels of Figure 3) is closer to the bulk value, but the average over the last 6 μs is still below the bulk value.

The fraction of the first-shell lipids that is cholesterol is relatively constant over the course of the simulation staying within a range of 0.10–0.20, which is also significantly less than the bulk value of 0.30. This observation increases the significance of localized cholesterol interactions; cholesterol is locally depleted relative to the bulk yet still binds the receptor at specific locations. This observation has implications for the “annular cholesterol” concept, which will be revisited in the Discussion.

In the second solvation shell (shown in Figure 4), the unsaturated chains are also enriched, though not as dramatically as in the first shell. Around the inactive receptor, they account for approximately 0.30–0.40 of the second shell (approximately double the bulk value), while around the active receptor, they average ~0.25 of the second-shell lipids. The saturated chains are within one standard deviation of their bulk value (0.55) in the second shell around both receptor states. The cholesterol fraction also stays near its bulk value, between 0.20 and 0.30, with the exception of the lower leaflet of the active state, which averages to just below 0.20.

The slow convergence of the solvation shell numbers originates from a broad distribution of lipid–protein interaction times. The histogram of individual lipid lifetimes (Figure 5) in the first shell spans the entire range of simulated time, from 5 ns to 15 μs, with an average lifetime for all three lipids in both systems of ~100 ns. The histogram is bounded at the short time end by the time scale for a lipid to jump between local cages formed by its neighboring lipids in a homogeneous, protein free fluid membrane, approximately 5 ns. It is bounded at the long time end by the total duration of the simulation. It is clear from the histogram that many lipids are in much slower exchange with the bulk than the average lifetime. In the inactive receptor simulation, 108 lipids have first-shell lifetimes of >1 μs; in the partially active receptor, this number is 119 lipids. Several lipids do not exchange during the entire simulation, as described next.

Figure 5.

Histogram of first-shell lipid lifetimes for partially active (black circles) and inactive (red diamonds) states of the receptor. Note the logarithmic scale on the vertical axis. The counts are unnormalized so that the numbers of lipids (one or two) is clear in the very long time bins.

The Receptor Binds Several Cholesterols in the Active State but Not the Inactive State.

Despite depletion of cholesterol from the first shell, there are three sites that bind cholesterol on the active receptor during the entire 15 μs trajectory of the active state. These appear as nearly solid horizontal red lines in the right panel of Figure 6, which shows the distance from each side chain to the nearest cholesterol atom. In contrast, the longest cholesterol binding event during the inactive receptor simulation (Figure 6, left panel) was observed at helices 5 and 6, during the first half of the simulation. Many transient interactions are observed in both cases, which appear as intermittent red and white regions.

Figure 6.

Distances between cholesterol and every side chain during simulations of the inactive (left) and active (right) receptor. The sequence of the receptor is indicated on the left, and helices 1–7 are indicated on the right. The scale bar indicates the distance between the nearest heavy atoms of any cholesterol and the side chain in nanometers.

By visual inspection of the active state trajectory, it was clear that three cholesterols were responsible for the long-lived interactions with the active receptor (right panel of Figure 6). The residues involved in these interactions were found first by filtering the data in the right panel of Figure 6 to identify the most robust, long-lived interactions. A list of residues was compiled for which a heavy atom side chain–cholesterol distance of <0.45 nm was observed continuously for at least 1 μs. Some of the cholesterols identified by this filtering interacted with the protein only superficially. For example, the hydroxyl of the cholesterol might be observed to form a hydrogen bond for >1 μs but might form no other contacts with the protein. These residues were not considered further. The remaining residues participated in interactions with a single cholesterol that formed multiple, long-lived contacts with the protein. They fall into three groups, corresponding to three cholesterols that remained bound throughout the simulation. They are listed in Table 1.

Table 1.

Locations and Side Chain Interactions of Three Cholesterols That Remain Bound during the Entire 15 μs Simulation of the Partially Active Receptor

location	residues	motif
H1 and H7 (outer leaflet)	Gly5, Val8, Tyr9, Val12, Trp268, Tyr271	unknown
H3 (inner leaflet)	Phe93, Leu96, Ala97, Ile100	unknown
H2 and H4 (inner leaflet)	Ser47, Ala50, Ala51, Lys122, Ile125, Ile129	CCM motif

As reported previously by us,¹⁹ cholesterol interacts with the cholesterol consensus motif¹⁴ during the entire simulation. There are in addition two other cholesterols that remain bound during the entire simulation, located on H1 and H7 in the outer leaflet (left panel of Figure 7) and on H3 in the inner leaflet (right panel of Figure 7). Neither of these two sets of residues is consistent with known cholesterol interaction motifs, such as the cholesterol recognition amino acid consensus (CRAC) motif.

Figure 7.

Cholesterol bound during the entirety of the simulation at (A) helices 1 and 7 and (B) helix 3.

Detailed interactions between specific cholesterol heavy atoms and the side chains listed in Table 1 are shown in Figures S1–S3. In each case, most of the long-lived interactions are with the ring system of cholesterol. The site on H3 is unusual in that the most robust amino acid contacts do not include a charged residue (both the CRAC motif and the CCM contain an Arg or Lys) to interact with the hydroxyl of cholesterol.

One might expect the Phe at position 93 to provide some specificity via a ring stack, but instead, it interacts with the tail of the cholesterol. While the interaction does not include a Lys or Arg, there is an aspartic acid at position 101 (see the right panel of Figure 7) that forms an intermittent hydrogen bond with the hydroxyl of cholesterol, formed ~75% of the time. The other site includes residues on both H1 and H7 and includes rings (Trp and Tyr), bulky hydrophobic side chains, and a glycine (also no charged side chain, although Tyr9 interacts with the hydroxyl as shown in Figure S2).

The Protein Recruits a Different Lipid Environment When Modeled with Martini.

All-atom simulations, even the long trajectories presented here, are very limited in terms of sampling. For this reason, the Martini CG model is very often used to study lipid–protein interactions, because it is possible to eliminate undersampling as a source of error. Several authors have reported on specific lipid–protein interactions obtained with Martini, such as charged head-group–protein interactions.^9–11,41 We also reported a consensus between residues that interact strongly with cholesterol in both all-atom and Martini simulations.²⁰

Although there is some consensus between the all-atom and Martini results for specific, localized cholesterol interactions, the two models yield somewhat different results for the averaged local lipid environment, particularly regarding cholesterol. Figure 8 reports the results of the Voronoi analysis performed on a Martini simulation of the inactive state for the same lipid composition as in the all-atom simulation. Comparison of the first-shell lipids as observed in the Martini simulation (Figure 8, left column) to the all-atom data (Figure 3, right column) reveals both similarities and differences. In the Martini case, the DPPC is depleted and the DOPC enriched relative to their bulk values, but not as dramatically as in the all-atom simulation. Cholesterol, by contrast, remains close to (perhaps even slightly enriched over) its bulk value, in contrast to the all-atom case, where it is depleted. In the second shell, the lipid fractions have already converged to their bulk values, similar to the all-atom data for the inactive receptor (compare the right column of Figure 4 to the right column of Figure 8).

Figure 8.

Lipid composition of the first (left) and second (right) shells for the inactive receptor as modeled with Martini 2.0. The top row is the upper leaflet, and the bottom row is the lower leaflet. To compare to the all-atom simulation data for the inactive receptor, compare the left column to the right column of Figure 3 and the right column to the right column of Figure 4.

DISCUSSION

Historically, interactions between lipids and integral membrane proteins have been described as either annular or non-annular. The former category refers to the immediate solvation layer around the protein, which exhibits differences in dynamics and conformation compared to the bulk, while the latter category refers to more tightly bound lipids, sometimes located deep inside a multimeric complex. For the data presented above, the distinction becomes one of degrees rather than quality; the distribution of first-shell lifetimes is very broad, with many lipids remaining in the first shell much longer (10–100 times) than the average first-shell lifetime. This observation is consistent with reports of similar quantities obtained using Martini for simulations of an ion channel⁴² and a GPCR.²⁰ Even simulations of the single-pass helix gramicidin A identified at least two time scales for lipid exchange at the surface of the protein: a fast time scale of ~5 ns and a slower time scale of ~75 ns.³⁹

The fact that some lipids at the protein surface exchange quite slowly with the bulk leads to slow (on the all-atom MD simulation time scale) equilibration of the first-shell composition; the composition of the first shell appears to still be changing up to (and perhaps beyond) 10 μs, making the problem just barely tractable with standard computational resources. Lipid–protein interactions are easily converged using the Martini coarse-grained model, but key differences in the local lipid environment were found when compared to the all-atom results. This may be due the difficulty of capturing the difference between a single unsaturated bond and a fully saturated chain at the Martini level (four CG sites per chain in both cases), so it will be interesting to see if the results change with a newer version of Martini or with a version specifically optimized for the DPPC/DOPC/Chol phase diagram.⁴³ In the absence of experimental data, it is impossible to say which model is more correct.

Despite the limitation imposed by the time scale, the all-atom simulation data clearly show that the local lipid environment around both the inactive receptor and the active receptor is significantly enriched with unsaturated chains. Rhodopsin also prefers solvation by unsaturated chains,^24,29 but this might have been expected given that the rod cell outer membrane is highly enriched with polyunsaturated chains. The fact that A_2A also prefers an unsaturated solvation shell suggests that this may be a more generic feature, perhaps because unsaturated chains are more conformationally flexible⁴⁴ and therefore more compatible with the irregular surface of a GPCR. Note that because this is a very localized effect, it is not inconsistent with the recent report by Malmstadt and co-workers that liganded A_2A partitions into the liquid-ordered phase in phase-separated vesicles.¹⁵

The lipid occupancies of sites that are in slower exchange (i.e., lifetimes of more than a few microseconds) are unlikely to equilibrate even on the 15 μs time scale, as evidenced by the three sites that are occupied by cholesterol during the entire trajectory of the active receptor. The interaction of cholesterol with the active receptor and not the inactive state is consistent with numerous reports that cholesterol supports GPCR signaling in general^16,45 and A_2A in particular^15,19,46 but contradicts a recent report that cholesterol inhibits A_2A ligand binding.⁴⁷ The simulation results should be interpreted cautiously given the concerns regarding convergence of the occupancy of these locations by cholesterol and should in the future be cross-validated against other computational predictions such as coarse-grained simulation²⁰ or docking.⁴⁸

The affinity of a particular site for a particular lipid or cholesterol depends on the lipids that constitute the rest of the membrane environment. The case presented here provides an illustrative example: Local enrichment of unsaturation around the protein likely increases the affinity of the observed cholesterol interactions, because partitioning of cholesterol into unsaturated chains is less favorable than into a saturated chain environment. Would similar interactions and enrichment be observed in a more realistic lipid environment? The answer may even be leaflet-dependent. The inner leaflet is enriched in polyunsaturated chains, while the outer leaflet is enriched in sphingolipids. Headgroup interactions complicate the picture on the inner leaflet, where there is an abundance of negative charge. These questions will be addressed in future work, which will revisit GPCR–lipid interactions in the context of a more physiologically realistic, asymmetric lipid mixture.

The local lipid environment recruited to the protein is worth considering in the context of continuum models^49,50 for the role of the lipid matrix in membrane protein function. Among the GPCRs, it is well established that the function of rhodopsin is modulated by the elastic curvature stress of the membrane.^51,52 Can such models be cleanly distinguished from localized lipid interactions in a complex mixture of lipids? Perhaps not. In the extreme case of a lipid bound so tightly that it can be considered as a part of the structure of the protein, it is clear that the lipid alters the shape of the protein presented to the membrane and therefore changes the coupling between the protein and curvature elastic properties of the membrane. It seems reasonable that lipid redistribution in response to curvature stress exists along a continuum, depending on both curvature stress and the composition of the surrounding lipid matrix, in a manner similar to cholesterol redistribution across the two leaflets of the plasma membrane.⁵³ Indeed, the interplay between lipid recruitment and curvature stress was shown recently for gramicidin by Sodt, Beaven, and co-workers.^39,54 Similar effects are likely at play in larger membrane proteins like GPCRs and ion channels, but resolving the details is challenged by the slow time scale for lipid redistribution.