Structure-Based Sequence Alignment of the Transmembrane Domains of All Human GPCRs: Phylogenetic, Structural and Functional Implications

Abstract

The understanding of G-protein coupled receptors (GPCRs) is undergoing a revolution due to increased information about their signaling and the experimental determination of structures for more than 25 receptors. The availability of at least one receptor structure for each of the GPCR classes, well separated in sequence space, enables an integrated superfamily-wide analysis to identify signatures involving the role of conserved residues, conserved contacts, and downstream signaling in the context of receptor structures. In this study, we align the transmembrane (TM) domains of all experimental GPCR structures to maximize the conserved inter-helical contacts. The resulting superfamily-wide GpcR Sequence-Structure (GRoSS) alignment of the TM domains for all human GPCR sequences is sufficient to generate a phylogenetic tree that correctly distinguishes all different GPCR classes, suggesting that the class-level differences in the GPCR superfamily are encoded at least partly in the TM domains. The inter-helical contacts conserved across all GPCR classes describe the evolutionarily conserved GPCR structural fold. The corresponding structural alignment of the inactive and active conformations, available for a few GPCRs, identifies activation hot-spot residues in the TM domains that get rewired upon activation. Many GPCR mutations, known to alter receptor signaling and cause disease, are located at these conserved contact and activation hot-spot residue positions. The GRoSS alignment places the chemosensory receptor subfamilies for bitter taste (TAS2R) and pheromones (Vomeronasal, VN1R) in the rhodopsin family, known to contain the chemosensory olfactory receptor subfamily. The GRoSS alignment also enables the quantification of the structural variability in the TM regions of experimental structures, useful for homology modeling and structure prediction of receptors. Furthermore, this alignment identifies structurally and functionally important residues in all human GPCRs. These residues can be used to make testable hypotheses about the structural basis of receptor function and about the molecular basis of disease-associated single nucleotide polymorphisms.

Author Summary

G-protein coupled receptors (GPCRs) are a large superfamily of integral membrane proteins that share a characteristic 7 transmembrane helix fold. They detect various molecules outside of the cell and signal their presence to the inside of the cell. At least half of the 800 human GPCRs are potential drug targets, so understanding their structure and function is critical. Experimental structures are now available for at least one receptor from each GPCR class. The structure of the 7 helix fold is highly conserved even for receptors with very low sequence similarity. We analyze the available experimental structures and compare the common inter-helical contacts. Our analysis leads to a unified sequence-structure alignment of the GPCR superfamily that can then be used as the starting point for structure prediction of all other GPCRs. A key result of our analysis is a list of conserved contact residues and activation “hot-spots” residues that are critical for GPCR folding and function. We propose that mutations and natural variants of amino acids at these locations in the GPCRs can dramatically influence their activation state and alter intracellular signaling. This provides hypotheses for the molecular mechanisms underlying disease causing mutants for any GPCR.

Introduction

Structural revolution in the GPCR superfamily

G protein-coupled receptors (GPCRs) comprise the largest superfamily of integral membrane proteins, covering ∼3% of the human proteome. They mediate transmembrane (TM) signal transduction by allosterically facilitating information transfer across the cellular membrane in response to extracellular signals [1, 2]. The GPCRs are pleiotropic proteins responsible for sensing a diverse set of extracellular signals ranging from photons and small molecules (neurotransmitters, metabolites, odorants, tastants) to large oligopeptides (chemokines, incretins), and converting them into one or more intracellular signaling cascades. This critical role of GPCRs in cellular signaling makes them therapeutic targets in a large number of diseases, either due to their direct role in the pathophysiology of a specific disease or due to their ability to modulate a set of signaling cascades implicated in a disease [3]. Currently, about 30–50% of all drugs and 20% of recently FDA approved drugs act through modulating GPCR functions [4].

Experimental structures are now available for more than 25 different GPCRs covering all four major GPCR phylogenetic classes. Several GPCRs have been crystallized in complex with ligands, and some have been crystallized in active conformations, capable of coupling to G proteins or arrestins. Furthermore, efforts to crystalize most human GPCR proteins are underway [5]. The human β₂ adrenergic receptor was crystallized in an active conformation in complex with the full heterotrimeric Gs protein, providing a snapshot of the conformational changes in both the receptor and the cognate G protein during GPCR activation [6, 7]. Rhodopsin has been recently crystallized with arrestin [8] providing the first detailed snapshot of the receptor before and during internalization. These new structures inspire a full spectrum of mechanistic studies into the GPCR biology, and will guide the functional understanding and pharmacological targeting of these receptors [9–11].

The cellular membrane partitions a GPCR protein into 3 domains: extracellular, transmembrane, and intracellular [12, 13]. The N-terminus and three extracellular loops (EC1, EC2, EC3) lie outside of the cell; seven transmembrane (TM) helices (TM1-7) span the membrane; and inside the cell, there are three intracellular loops (IC1, IC2, IC3) together with the C-terminus, which typically contains a shorter helix 8 resting parallel to the membrane. The N-terminus, intracellular loops, extracellular loops, and C-terminus can have very different lengths, not just across the major GPCR classes but also within a GPCR class [14]. The loop regions are flexible and display different conformations among the known GPCR crystal structures. However, the packing of the TM helices is remarkably well conserved (Fig 1) even for proteins with very small sequence similarity (down to 20%, S2 Fig).

Fig 1. TM domains of the available crystal structures.

Top: Two views of the 24 inactive crystal structures from classes A, B, C, and F (aligned to β₂) show the general GPCR fold of the transmembrane (TM) bundle. Class A in green, class B in blue (CRF1, GLR), class C in orange (MGLU1, MGLU5), class F in magenta (SMO). Bottom: Same views for only the 19 inactive class A structures showing the highly conserved class A TM fold. A detailed view of the conserved hydrogen bonding networks is shown in S1 Fig.

For class A GPCRs this is not a surprise, because several amino acids forming inter-helical hydrogen bonds are highly conserved. Indeed, the conserved residues directly correspond to each other in the structural alignment, as can be seen on S1 Fig. Within class A, the sequence similarity between the corresponding transmembrane regions for the receptors with experimentally determined structures is 35–100% (S2 Fig), which is high enough that a sequence alignment algorithm (such as Clustal Omega) typically aligns the transmembrane regions correctly. However, the sequence similarity between class A and other classes is so low, typically 20–30%, that sequence alignment algorithms generally fail.

At the time of this study, there were 5 non-class A GPCR structures available. Fig 1 shows that when these structures are aligned to class A using the GRoSS alignment (introduced below in the methods section), their transmembrane helices are in the same fold as that of class A GPCRs. A certain degree of similarity was expected for the intracellular face of the receptor, where they couple to the same set of G proteins and arrestins, but the similarity seems to be present in the full TM bundle. With this comparison of the structures of GPCRs from different classes, we are now able to examine the features that survived the evolutionary divergence in sequence potentially on the basis of their functional importance, enabling us to gain insights into the general features of the GPCR structural fold.

Transmembrane domains as a basis for modeling conformational diversity

Computational protein folding works best for small water soluble proteins and it has proved effective for up to about 100 residues [15]. The energy models used for soluble proteins have been extended to include approximations of the membrane environment [16, 17]. However, the accuracy of these methods is not sufficient for drug design, partly due to the size of many membrane proteins. GPCRs are large proteins with more than 300 residues, making them difficult to model with standard protein modeling methods. Recent assessments [18–20] of the GPCR structure prediction community showed that approaches based on homology models remain the most successful [21–23].

Homology modeling requires accurate sequence alignment to existing structures, which fails completely for alignment between classes. Within each class, sequence alignments typically conserve key motifs, but they often lead to gaps in the transmembrane helices, which indicate serious problems. Recently, Isberg et al. [24] studied in detail the available crystal structures and obtained pairwise sequence alignments from the highly conserved 7 helix fold by aligning individual TM domains of two structures at a time. Their alignment leads to many single amino acid gaps, which correct for non-matching bends and turns of the TM helices that are often caused by mismatched Pro or Gly. Nevertheless, it is not possible to predict these gap positions with sequence information alone, without the knowledge of the correct structure. Therefore it is not possible to extend their alignment to other GPCRs. Our goal was to construct a sequence alignment of the TM domains of all human GPCRs, that would be suitable for 3D structure prediction. Thus we developed a method that avoids the gaps in the TM regions.

We can consider the common GPCR structural fold to be formed by alpha-helices, allowing us to think about the structural relationships in terms of rigid helices. Past studies [25, 26] used the rigid transmembrane helices to compare the different GPCR receptors. The comparative analysis of available GPCR structures informs structure prediction methods (such as in [27]), where 3D rotations and tilts of the rigid helices are used to sample conformations of the protein in the membrane. Moving the whole helix at once can overcome large potential barriers and sample possible inactive and active states. Better definitions of the GPCR fold and its activation mechanism can help to reduce the number of coordinates needed for modeling.

Integrating sequence, structure, function

Several studies have analyzed inter-helical contacts [28], ionizable networks [29], or developed a phylogenetic analysis of class A GPCRs [30], but no one has yet reported a GPCR superfamily-wide sequence alignment that avoids gaps within the TM domains. The available GPCR structures in the PDB cover a wide sequence space, with some structures showing different ligands bound to the same receptor or the same ligand bound to different receptors, and some structures showing different conformations for the same receptor. The fact that the seven helix fold is conserved across divergent sequences can teach us about the nature of the GPCR fold and its features.

In this work, we first generate a structure-based alignment of the TM regions of the available 25 structures by maximizing the number of the corresponding inter-helical contacts. We use chemically nonspecific contacts instead of conserved residues, since, even though the amino acids are not conserved across this diverse superfamily, many inter-helical contacts are conserved across all GPCR classes as will be shown below. Our structural alignment is extended to include all human GPCRs by sequence alignments of small subgroups of similar receptors. The final GPCR sequence-structure alignment (GRoSS) includes the TM regions of all 817 human GPCRs. The GRoSS alignment is then used to generate a phylogenetic tree for all human GPCRs, which correctly distinguishes the different receptor classes, suggesting that the functional information about different GPCR classes, that are usually characterized by variable length N/C-termini and loop regions, is in fact encoded at least in part within the GPCR TM domains.

This observation that the TM domains and their relationships capture the similarities between the rather disparate set of GPCRs, has broad implications in terms of simplifying and enhancing the connection between receptor sequence and function. We identified 23 Conserved inter-HelIcal COntacts (CHICOs) that define the general GPCR structural fold. Furthermore, our structural comparisons of inactive and active conformations lead to the identification of 15 Native ACtivation “Hot-spOt” residues (NACHOs). We expect that mutations of the structurally important CHICO and NACHO residues will dramatically affect receptor function, and may be responsible for many diseases. Cross-checking these residues with available GPCR mutation databases (Uniprot [31], NAVA [32], TinyGRAP [33]), which contain disease/function association, has allowed us to identify several deleterious receptor mutations that are found at or adjacent to the “hot-spot” residues as shown later in the results section. Using the GRoSS alignment, we have mapped these important positions to all GPCRs leading us to suggest new testable hypotheses about molecular mechanisms behind natural or man-made mutations. We expect that this will create new rational drug discovery opportunities for efficacious therapeutics that minimize side-effects.

Methods

In this section we describe our procedure for sequence alignment of the TM domains of all 817 human GPCR proteins. The resulting alignment is available in S2 Table, and in fasta format in S1 File.

Human GPCR classification

After the Human Genome Project was completed, Fredriksson et al. [34] performed detailed phylogenetic analysis of GPCRs developing the GRAFS classification system [34, 35], which identified 5 main groups of receptors: Glutamate, Rhodopsin, Adhesion, Frizzled/Taste2, and Secretin. The Rhodopsin family is the largest and was partitioned into four subgroups: α, β, γ, and δ. Olfactory receptors were attached to the δ branch. The bitter taste receptors (Taste receptors type 2, Taste2, or TAS2Rs) were grouped with the Frizzled family.

Currently, the list of receptors is maintained by the International Union of Basic and Clinical Pharmacology (IUPHAR) [36], which also keeps track of known endogenous ligands and signaling mechanisms. IUPHAR identifies 6 main classes A-F: Four of them, class A (rhodopsin-like receptors), class B (secretin and adhesion family), class C (glutamate receptors), and class F (frizzled receptors) contain human receptors. Classes D, E are part of the classification that includes other species, but contain no human orthologs. In contrast to GRAFS, secretin and adhesion groups have been merged since their sequences align well. The bitter taste receptors (Taste2, or TAS2Rs) and Vomeronasal receptors (VN1Rs) are beginning to be considered similar to class A, in contrast to the GRAFS classification, which is what we also find in the phylogenetic analysis presented below.

GPCR database

We extended the list of human GPCRs of Fredriksson et al. [34], with the proteins considered by the GPCR Network [5] and by IUPHAR [36]. Furthermore we added all sequences annotated as human GPCR proteins in the Uniprot database [31]. Altogether we collected 836 candidate sequences.

Most of these proteins have been assigned to a class by previous studies [34, 36]. For sequences with unknown class, we first searched for related proteins by running BLAST, and then we aligned them against the candidate class with the program Clustal Omega [37]. If the sequence aligned without large gaps in the TM regions for any of the classes, we assigned the sequence to this class. The following 11 proteins might be GPCR proteins, but they do not align well to any of the classes, thus we ignore them in the further analysis (listed as Uniprot accession numbers, ACs):

P51810, Q5T9L3, Q5VW38, O60478, Q86V85, Q86W33, Q8N3F9, Q8NBN3, Q96K49, Q96N19, Q9NPR9.

The following 8 sequences are most likely pseudogenes, because they are missing one or more TM domains:

A6NFC9, Q32VQ0, Q8NGA4, Q8NGU1, Q8NGY7, Q8TDU5, Q96P88, Q99463.

We kept Q9P1P4 (TAAR3) which is a pseudogene in humans but functional in rodents, and we kept Q49SQ1 (GPR33), which may be functional in some people. A curious case is the protein GPR157 (Uniprot AC Q5UAW9), which is most similar to class B, but its TM1 has a gap in the alignment to class B. However, the TM1 aligns well to the class A TM1, so this protein appears to be a hybrid between these two classes.

Table 1 summarizes the count of all the sequences that we kept and S2 Table lists their Uniprot ACs. In total there are 817 candidate human GPCRs, 399 of which are non-olfactory. If the sequences in class A were present in the analysis of Fredriksson et al. [34], then we kept the subgroup labels α, β, γ, δ, otherwise we labeled them as A-other.

Table 1. Number of GPCR sequences by class.

The total number of candidate human GPCR sequences that were considered are listed. The full list of Uniprot ACs is in S2 Table.

88	Aα	16	B	5	Vomeronasal
33	Aβ	22	C	25	Taste2
57	Aγ	11	F	11	Other
58	Aδ	33	Adhesion	8	Pseudogene
51	A-other	418	Olfactory

Available crystal structures

At the time of this study, crystal structures of 25 GPCRs were available (19 of these for human sequences): 4 in both active and inactive conformation [6, 38–44], 1 in active only [45], and 20 in inactive conformation only [46–66]. When multiple crystal structures of the same protein were available, we used the one with the best resolution or the one with the best-defined transmembrane alpha helices. The PDB IDs of the structures used in analysis are listed in S1 Table.

Well-defined TM helices are a prerequisite for our analysis, however, different criteria have been used for annotating helices in different PDB files. Many of the transmembrane helices contain bends, and sometimes the helix termination is not well defined. We define the extent of each transmembrane helix as the residues positioned in the membrane (as placed by the Orientations of Proteins in Membranes (OPM) database [67]) extended until the end of the alpha helix by the DSSP secondary structure determination [68]. The helices were manually inspected and only a few manual corrections were needed. The final TM lengths used are displayed in S2 Table.

Alignment by minimizing RMSD is not unique

As suggested by Fig 1, the transmembrane helical bundles of all the GPCR structures display the same structural fold (same relative position of all seven helices). The main basis for some published structural comparisons of classes A and B, C, F was an iterative structural alignment implemented by the program ICM-Pro [69] used for GLR, MGLU1, and SMO. Also, MGLU5 was aligned to class A with an iterative algorithm (SSM algorithm [70]) and CRF1 was aligned to class A manually. The iterative structural alignment algorithm, removes from the alignment all atoms that are too far in the previous rounds. This works relatively well, but the exact sequence pairing is not uniquely defined and depends on cutoff parameters. In many cases it leaves an ambiguity of ±4 residues (1 helical turn).

In order to remove this ambiguity and to determine which alignment would minimize the root mean squared deviation (RMSD) of the full TM bundle, we start from an approximate structural alignment, and try all nearby sequence alignments (±1 helical turn on each helix). For each sequence alignment, we first select the maximal overlapping lengths of the 7 TMs. Then we compute the C_α RMSD by evaluating the least-squares superposition of the corresponding C_α atoms. But minimizing only RMSD does not necessarily lead to an optimal alignment. For several cases in class A, we find TM alignments that have a lower RMSD than the alignment that conserves the correspondence of the same Ballesteros-Weinstein (BW) residue positions [12] (some BW residue positions in each TM domain are expected to be conserved at least within a GPCR class and hence align well). The reason for this is that the extracellular ends of the helices sometimes have significantly different tilts, making the most tilted helix dominate the RMSD measure. To avoid these issues with the RMSD measure, we instead look for an alignment, which maximizes the number of conserved inter-helical contacts. A contact is defined to be conserved if the residue pair in the contact has the same BW numbering in the two structures being compared. For example, if structure X has an inter-helical contact between residues 2.45–3.42 and structure Y also has an inter-helical contact between residues 2.45–3.42, then this contact is defined as conserved in the two structures. The BW numbering of residues in structure X and structure Y implicitly assumes that the sequence for structure X and the sequence for structure Y are aligned to superimpose residues with the same BW numbering. This measure accommodates the scenario commonly seen in GPCRs and described below, where residues may not be conserved but the corresponding structural contacts have been conserved during evolution.

Class A conserved contacts

In class A, the analysis of inter-helical interactions typically focuses on hydrogen bonds, since many of the hydrogen bonding residues are highly conserved. In order to compare receptors from different classes, which have poor sequence conservation, we compare inter-helical contacts ignoring their chemical nature. We use the definition of an inter-helical contact as in [28]: any two heavy atoms from different TMs that are closer than the sum of their van der Waals radii plus 0.6 Å. The inter-helical contacts, which are present in almost all class A structures, are shown in Fig 2. This list is very similar to the contacts found by [28], but there are minor differences caused by using a different set of crystal structures.

Fig 2. Conserved inter-helical contacts.

Top left: Diagram of 40 conserved inter-helical contacts (CHICOs) present in at least 23 out of 24 studied class A structures. The contacts common to all classes are shown in purple, and contacts present only in class A in orange. Top right: List of these contacts in Ballesteros-Weinstein numbering scheme. Bottom: Extracellular view of the same contacts in the β₂ crystal structure. The contacts in the inner and outer half of the membrane are shows on the left and right respectively.

We briefly compare this list of chemically unspecific contacts to the conserved hydrogen bonds. Within class A, the focus is often on two conserved networks of hydrogen bonds (shown in S1 Fig):

4-2-3: W4.50 ↔ S/N/T2.45 ↔ S/N/T3.42

1-2-7: N1.50 ↔ D2.50 ↔ N7.49.

Fig 2 shows that the network 4-2-3 is well conserved across classes. In particular, the contact 2.45–3.42 is present in all classes, and there are many conserved contacts in its immediate vicinity, such as 2.42–3.45, 2.42–3.46, and 2.46–3.42. The contact between TM 3 and 4 is also well conserved, as the highly conserved bulky residue W4.50 leans on A3.38.

The network 1-2-7 also has many conserved contacts, including N1.50-D2.50, but interactions with N7.49 are not conserved. Even in class A, N7.49 interacts with the other residues of this hydrogen bonding network indirectly through a water molecule. Still, the side chain packing in the regions where the helices are close together is important for structural stability. Conserved water-mediated interactions could be included in future analysis, after enough crystal structures resolve them. For class A it is clear that using BW numbering greatly facilitates sequence and structure comparisons. Next we will see how these ideas can be extended to other classes.

Sequence alignment across classes based on common contacts

To compare other GPCR classes with class A, we first find the inter-helical contacts of all the current crystal structures. Each structure has about 200 contacts. For any pair of structures and a given sequence alignment, we compute the number of conserved inter-helical contacts. We consider all possible TM sequence alignments that have no gaps (±10 residues from the starting sequence alignment). For class A, the number of common contacts between any two structures is maximized by the alignment that preserves the BW numbering.

To translate the BW numbering into classes B, C, and F, we start with a sequence alignment corresponding to an approximate initial structural alignment to class A. For each non-class-A structure, we then try all possible adjustments to BW n.50 residues (±10 residues, again with no gaps) on each helix and count the number of common contacts with each of the 24 class A structures. Table 2 shows the alignments with the highest cumulative number of the common inter-helical contacts.

Table 2. Selection of the alignment between class A and classes B, C, and F.

This table shows the selection process for assigning BW.50 residues to non class A proteins. Shifting BW.50 residue on each helix renumbers the relative BW numbers, effectively changing the labels of contacts observed in these proteins. Subsequently, the number of common contacts each structure shares with the class A structures changes for different BW residue assignments. The second rightmost column shows the cumulative number of contact occurrences among the 24 class A structures (including active conformations). The BW assignment with the highest number of contacts is selected (except for MGLU5, see text). The selected alignment is in bold.

Class	Protein	BW#							Common contacts	RMSD Å
		1.50	2.50	3.50	4.50	5.50	6.50	7.50
B	CRF1	L134	F162	L213	W236	V279	L329	S360	2212	3.11
		L134	F162	L213	W236	V279	I325	A363	2084	3.61
		L134	F162	L213	W236	V279	I325	S360	2081	3.35
	GLR	L156	F184	L249	W272	A314	V364	A397	1972	2.93
		L156	F184	L249	G273	A314	V364	A397	1913	2.94
		L156	F184	L249	M276	A314	V364	A397	1876	3.11
C	MGLU1	T607	I638	I682	I714	L763	A800	L827	2017	3.02
		T607	I638	I682	S711	L763	A800	L827	2012	3.24
		T607	I638	I682	S711	C767	A800	L827	1873	3.41
	MGLU5	T594	I625	A669	F698	L750	A787	L814	1974	3.36
		T594	I625	A669	I701	L750	A787	L814	1954	3.17
		T594	I625	A669	F698	L750	I784	L814	1820	3.64
F	SMO	T245	F274	W339	W365	V411	I465	S533	2358	3.00
		T245	F274	W339	W365	V411	C469	G529	2311	2.98
		T245	F274	W339	W365	V411	C469	S533	2248	3.02
		…
		T245	F274	W339	W365	V411	S468	I530	1827	3.04

In class A, the BW n.50 (n = 1 to 7 denotes the TM) residues correspond to the most conserved residues in each TM. After the projection of the BW numbering to the other classes, the n.50 residues are not necessarily the most conserved within each class. Moreover, they correspond neither to class B specific Wooten numbering [71], nor the class C specific Pin numbering [72], nor the class F specific numbering [65], which only define labels n.50 for residues that are the most conserved within the respective classes. In the context of Isberg et al. [24], our n.50 can be referred to as n.50a and uses class A residues as a reference to number TM residues for GPCRs from all classes. This unified numbering scheme enables a view of structurally conserved or structurally similar positioned residues across all GPCR classes because conserved residues in individual GPCR classes have no structural similarity or positioning in the GPCR structural fold in cross-class comparisons.

In terms of absolute sequence alignment of TM domains, our Class B (CRF1, GLR) alignment agrees with the alignment suggested in [62], which was obtained by an iterative structural alignment. Similarly, our alignment for class C (MGLU1, MGLU5), agrees with the suggested alignment in [63]. For MGLU5, we chose the second highest scoring alignment to make the alignment consistent with the MGLU1. This choice was checked visually and the corresponding residues are in a more similar position in our selected alignment.

Our alignment for the SMO receptors is the only one that differs from the published alignment to class A presented with the crystal structure [65]. This published alignment corresponds to the last row of Table 2, and it differs from our best alignment (fourth to last row of Table 2) in TM6 and TM7 by 3 residues each. The RMSD difference between this published alignment and our alignment is small, however, our alignment results in ∼500 more common structural contacts than the former alignment. Helix 7 of the SMO receptor does not have a proline residue, and so it is missing the kink that is typical for class A GPCRs. There are many inter-helical contacts in the extracellular part of the TM7, so that our chosen alignment gives a good spatial correspondence for the larger part of helix 7. We have presented a well-defined protocol for structural-sequence alignment. As more experimental structures become available, this protocol can refine the alignments where needed, which might occur especially for classes with a small number of known crystal structures.

This sequence alignment was used to generate the structural alignment shown in Fig 1. The sequence similarities between the TM domains of the crystal structures are shown in S2 Fig, and the TM domain RMSDs are shown in S3 Fig. Sequence similarities correlate reasonably well with the RMSDs (structural differences). Note that we define percent similarity as the fraction of similar residues, where two residues are similar if their BLOSUM62 [73] (or GPCRtm [74]) matrix entry is positive.

GRoSS: Extension of alignment to all known GPCR sequences

We extend the structure-based alignment derived above to all human GPCRs by anchoring each subfamily to the correct crystal structure. As a guide for the quality of the sequence alignment, we check for the presence of any gaps in the transmembrane regions. The approximate positions of the TM regions are already annotated in the Uniprot database as predicted by the TMHMM program [75]. These predictions are quite noisy, and even for similar proteins that align well, they can differ by 5–8 residues and sometimes even misclassify a TM. However, for multiple sequences the overall trend clearly identifies the approximate TM location and allows us to judge the quality of the alignment of multiple sequences. If there are gaps in the TM regions, the alignment cannot be used to successfully create homology models.

First, we try to align directly all 817 sequences of the GPCR superfamily using a multiple sequence alignment program Clustal Omega [37]. However, the overall sequence conservation is very low, and the resulting alignment has many large gaps even within TM domains. Some highly conserved residues end up aligned incorrectly. In order to avoid this problem, we aligned class A sequences separately (705, including olfactory). Again the resulting alignment has large gaps even in the TM region. It seems that the large variability of the loop region is what confuses the alignment algorithm.

Fortunately, we find that sequences in individual subgroups can be aligned using Clustal Omega without large gaps in the TM regions. We take these individual subgroup alignments and fix them into a profile—a multiple sequence file for which aligned columns are kept fixed, and from which the hidden Markov model (HMM) is computed. We then align any two profile HMMs to see how similar are the two groups. A profile alignment of Aα to each of the other class A groups (Aβ, Aγ, Aδ) has no gaps in the TM regions, and also gives the correct alignment of the BW.50 residues.

The multiple sequence alignment of the group A-other showed gaps in the TM regions for several proteins (Uniprot ACs: Q96P67, Q8TDU6, Q16570, Q86SM8, Q9NS66, Q9NS67, P60893, Q86SM5), so a separate profile was created for these sequences. After the split, both profiles of the A-other proteins aligned separately to the Aα profile without any gaps in the TM region, which anchors them to the class A alignment. Similarly, the profile of olfactory receptors (both tetrapod and fish-like) aligned to the Aα profile without any gaps in the TM region, which anchors the olfactory profile to the class A alignment. The Vomeronasal and Taste2 groups were more problematic, and are discussed in the following section.

The profile of the adhesion class aligns well to secretin class (original class B). Class B is aligned to class A using the structural analysis described above. Aligning profiles of Aα and B does not yield meaningful alignment, because the TM regions are offset and there are many gaps in the TM regions. Similarly, aligning classes A and C or A and F does not yield meaningful alignment, and again structural alignment was used for these cases.

Once the alignment is fixed, the TM lengths for new proteins can be predicted to be the average TM lengths from the available structures in the same class. For example, for sweet taste receptors (TAS1) the predicted TM length is the average TM length of GMR1 and GMR5; and for bitter taste receptors the average TM lengths of the 20 class A structures. These are meant to be the best initial guesses.

The complete listing of all the 817 human GPCR proteins is shown in S2 Table. The alignment of each helix is determined by the provided BW.50 residue. Expected TM range is also provided and it is estimated as the average TM region of the known crystal structures from the same class. For easy viewing, we provide the same alignment also in the fasta format together with the annotations of TM range and BW residues in Jalview [76] format in S1 File. The GRoSS alignment was also compared to alignments obtained by two other methods: HMM-HMM [77] and GPCRDB [24, 78]. This comparison is described in S1 Text.

Bitter taste and vomeronasal receptors

Bitter taste (Taste2, TAS2R) and vomeronasal receptors are small groups of receptors that do not easily align to the profiles for classes A-F, and so their classification has not been unique. While IUPHAR assigns the bitter taste receptors into the class A, Singh et al. [79] points out the lack of conserved amino acids between the two.

The profile of the vomeronasal group aligns better with class Aα compared to classes B and C, but there is still a gap of length 2 near the center of TM5. We remove the gap in such a way that the residue, which aligns with 5.50 stays fixed. To check that this is indeed the best alignment we explore small changes in the alignment by shifting individual TM by up to ±5 residues. In Fig 3 we see that for TMs 1 to 4, our current alignment gives the highest sequence similarity with Aα, so the alignment of these TMs is correct. However, for TM5, the alignment shifted by -1 or +2 residues gives higher similarity with Aα. Nevertheless, the similarity with groups Aβ, Aγ, Aδ, and B is the highest for our current alignment. We therefore keep the current choice.

Fig 3. Testing the robustness of the alignment of the Vomeronasal receptors with the other groups.

The table shows similarity between TMs averaged over all pairs of sequences formed from the two groups (red denotes high similarity, blue low similarity). For most TMs the optimal choices agree with the optimal alignment to Aα (full table in S5 Fig); all combinations are shown only for TM5. The same table but using the GPCRtm substitution matrix [74] instead of BLOSUM62 is shown in S7 Fig. GPCRtm was developed in particular for GPCR proteins, but in this case both matrices result in the same alignment.

We performed a similar analysis for the Taste2 receptors, for which adjustments were necessary. The profile alignment of Taste2 with Aα has some gaps, but it is still the best alignment (i.e., it has the fewest gaps) compared to aligning to classes other than class A. TM3 has two gaps in the alignment: a gap of length 4 in the middle of TM3, and a gap of length 5 at the DRY motive. As the first iteration we kept the alignment fixed on residue 3.50, then we computed the similarity to other groups for ±5 residue shifts. The shift by +3 residues gives better similarity and so it was kept. See Fig 4 for the computed similarities after the shift has been made. All class A subclasses favor this new choice, as the highest similarity has offset 0. Class B would favor shift by 2 residues, but the similarity is less than 30%.

Fig 4. Testing the robustness of the alignment of the Taste2 receptors with the other groups.

The table shows similarity between TMs averaged over all pairs of sequences formed from the two groups (red denotes high similarity, blue low similarity). For most TMs the optimal choices agree with the optimal alignment to Aα (full table in S6 Fig) only TM6 shows a second possible alignment at offset +4. The same table but using the GPCRtm substitution matrix instead of BLOSUM62 is in S8 Fig. Again, both matrices result in the same alignment.

TM4 has low sequence similarity, and in particular the highly conserved Trp is not present in Taste2. Again as a starting point we kept the alignment at 4.50, but later had to adjust it by 4 residues. Fig 4 shows the similarity after this shift has been made. For TM4 the similarity is only slightly higher at the new best offset than at nearby offsets.

Taste2 TM6 showed the partially conserved motif IYFLS, with S being aligned to P6.50, which we kept as an initial try. This choice is kept in Fig 4. However, we see that an offset of +4 residues, which corresponds to a one turn shift (the motif IYFLS aligns Ile with P6.50), also gives high similarity. Based solely on sequence similarity we cannot distinguish which alignment is better, and therefore both cases should be considered when building homology models and energy of the resulting structures should be used as a guide to select the best choice. These alignments will be revisited when the first experimental structure of one of the Taste2 receptors is determined.

Computation of the phylogenetic tree

We compute the similarity for each pair of sequences using the weights from the BLOSUM62 matrix (two residues are considered similar if their BLOSUM62 matrix entry is positive), and use the similarities as a distance metric to cluster the proteins. We used the unweighted pair-group clustering algorithm (implemented in Jalview [76]), which iteratively extends clusters by finding a non-member sequence with the lowest average dissimilarity over the cluster members. The phylogenetic tree constructed by this clustering algorithm was visualized using the Iterative Tree of Live toolkit [80].

List of natural variants

S3 Table lists all 2449 GPCR natural variants annotated by Uniprot. According to the GRoSS alignment 1289 of these lie in the TM regions and are listed here with the corresponding BW number. For each mutation we computed its distance to the closest NACHO (or CHICO respectively) residue on the same TM. Zero means this residue is the NACHO (CHICO) residue, in which case we also provide the multiplicity column counting to how many NACHO (CHICO) contacts this residue belongs to. We found 13 (23 including olfactory) mutations of residues on both lists, 48 (99 including olfactory) on the NACHO only list, and 161 (299 including olfactory) on the CHICO only list.

Molecular graphics

3D molecular views have been rendered using PyMOL [81].

Results and Discussion

Gaps in the alignment of TM regions

We constructed the GRoSS alignment in order to avoid gaps in the TM regions and to simplify preparation of homology models. With the BW residues correctly aligned, and without any gaps in the TM regions, we can use this alignment for direct generation of homology models of the TM helix bundle using essentially any structural template.

A general approach to preparing homology models is to create a new alignment for each target and available template, say using the HMM-HMM method [77]. However, HMM-HMM often produces false gaps in the TM regions. Recently, a sequence numbering for GPCR crystal structures was presented by Isberg et al. [24] (available at GPCRDB [78]) that used a structural alignment to identify gaps or bulges in TM regions, when comparing the same TM between any two crystal structures. The properly placed gaps, often improve structural alignment of helix kinks or loose turns. However, the best structural alignment also resulted in gaps in TM regions that can never be predicted by sequence alignment, such as HMM-HMM alone. Table A of S1 Text shows that the mismatches between the GPCRDB and HMM-HMM are common. The GPCRDB alignment is good for retrospective analysis of known structures, but cannot be used for predictions of unobserved gaps.

To quantify the differences between homology models based on the three difference alignments, we compare the RMSD, TM-score and number of common contacts in Fig A of S1 Text. Overall, these comparisons show that GRoSS performs similarly to GPCRDB for alignments within one class, and better for inter-class alignments. GRoSS performs better than HMM-HMM within a class, and significantly better between different classes.

Loop alignment

We omit loops from the GRoSS alignment, because in the GPCR protein superfamily loops are very diverse, especially the loops EC2 and IC3. EC2 is up to 171 residues long for some class A receptors, but it is shorter than 35 residues in class C, and shorter than 20 residues for all other receptors. IC3 is up to 223 residues long for some class A receptors, but it is shorter than 20 residues for all other receptors. There are likely important similarities among the loops across the GPCR classes. For example, on the intracellular side the receptors have to be sufficiently similar to accommodate G-proteins and arrestins. Furthermore, on the extracellular side, there is a highly conserved disulfide bond between TM3 and loop EC2 that is important for the assembly of the receptor in the membrane [82]. Thus it is possible that with more experimental GPCR structures, a more systematic understanding of the loop regions will emerge as well.

Sequence alignment from structural alignment

Fig 2 compares class A to the other classes for the alignment constructed by maximizing the number of common inter-helical contacts (Table 2). The purple color in Fig 2 denotes the structural contacts common to all classes, and orange denotes contacts specific to class A. Only one contact, 6.51–7.39, is present in all of class A structures (active and inactive), but it is not in the structures of the other classes. Furthermore, the interactions of TMs 1–5 are more conserved across all classes, but the TM 6 and 7 contacts are more class A specific. It is possible that during the GPCR assembly the helices 1–5 form some intermediate partially folded state before helices 6 and 7 are fully present in the membrane. This might be the reason why the contacts between helices 1–5 are more similar across the classes.

Fig 5 shows the alignment of the TM3 regions for all the known crystal structures (other TMs are shown in Fig 6). We see that the DRY motif at positions 3.49–3.51 is highly conserved within the 20 class A sequences, and even when there are mutations only similar amino acids occur: ERY, DRF (however, there exist class A GPCRs without this motif, e.g. PTGDR has ECW [83]). In classes B, C, and F the DRY motif is not conserved at all.

Fig 5. TM 3 sequence alignment for the 25 crystal structures.

Other TMs are shown in Fig 6. The sequences are taken from the selected PDB files. The TM helix residues are colored in the Zappos scheme, which captures the chemical nature of each residue (e.g. helix breakers, proline and glycine, are shown in purple). The loop residues are shown in grey. The BW n.50 residue (numbering displayed below the sequences) is the most conserved within the class A. The consensus sequence is most similar to class A, because most sequences are from this class. The largest differences are for the last 5 sequences, which belong to the classes B, C, and F. The figure was prepared using Jalview.

Fig 6. Sequence alignments for TMs 1,2,4–7 for the 25 crystal structures.

Same caption as Fig 5, where TM3 is shown.

Proline residues often cause a helix kink and are commonly found in the TM domains of membrane proteins. They are structurally important for deciding which structures should be used as templates for modeling a new protein. In Fig 5, prolines are highlighted in purple. For example, only MGLU5 has a proline in a central region of TM3, but in this case, the shape of TM3 is very similar to MGLU1, which does not have the corresponding proline.

The consensus sequence for TM3 mostly agrees with class A residues, because most of the crystal structures are from the class A. Interestingly the most conserved residue across all classes is Cys3.25, which forms a disulfide bond to the extracellular loop EC2. This bond is important for the stability of the protein, and shown to be critical for the GPCR assembly [82].

Fig 6 shows the consensus sequence and alignment of the remaining TM regions (1,2,4,5,6,7) for all experimental structures considered. Known conserved residues in these TMs for class A receptors are easily spotted. For TM1, residue 1.50 is a conserved polar residue for all GPCR classes except class B. For TM4, W4.50 residue is conserved across all classes except class C. For TM5, residue 5.60 is a positively charged residue for classes B, C, F, and most Aα receptors. For TM6, residue W6.48 is conserved for all classes except Aδ, B, and F. TM7 residue 7.45 is exceptional in being a conserved polar residue across all classes and also appears on our conserved contact residue list (see below).

The conserved inter-helical contacts of class A were the basis for the alignment between the GPCR classes. These contacts show interactions that should be considered first in analysis of the structure or function of these proteins. Fig 2 shows the contacts conserved only within class A, and a similar analysis is shown for classes B, C, and F in S9 Fig. However, since only one or two structures are available in classes B, C, and F, the resulting list is not averaged as it was for class A, and it will be refined as more crystal structures from these classes become available. Thus we cannot yet determine which residues are causing most of the systematic differences between the classes and which residues are critical within each particular class.

Common contacts between different classes define the GPCR structural fold

There are 40 inter-helical contacts common to class A GPCRs as shown in orange in Fig 2. Out of these, 23 contacts (shown in purple) are present in the crystal structures from classes B, C, and F as well. These 23 conserved inter-helical contacts (CHICOs) formalize our initial insight that the TM bundles of all the different classes are similar and define the GPCR “structural fold”. As more structures become available, this structural fold will be refined.

Examining the inter-helical contacts that make up this GPCR structural fold, we find that TM6 is fully decoupled from both TM3 and TM5, whereas this was not the case for the fold that corresponds to only class A GPCRs (orange contacts in Fig 2). GPCRs across different classes couple to the same set of G proteins and arrestins, and now it is known from overwhelming structural and biophysical evidence that G proteins [6] and arrestins [8] couple to the GPCRs between TM3/TM6 or TM5/TM6 regions. The GPCRs have evolved to conserve these functionally important couplings with their intracellular signal transduction partners, but have not had the need to conserve contacts of TM6 with TM3 or TM5. This is consistent with the structural fold of GPCRs shown by purple contacts in Fig 2.

Fig 2 seems to suggest that there are two separate conserved units: TMs 1–5 and TMs 6–7. It is possible that for receptors that have very long loop IC3 (these are only in class A), the TMs 1–5 need to stabilize in the membrane prior to assembly of the last two TMs.

Another important observation is that the specific conserved contact residues defining the GPCR structural fold are not conserved across the different GPCR classes. This tactic of nature to maintain a structural fold without conserving the residues is not uncommon, e.g., the MATα2 homeodomain-operator complex in yeast and drosophila has maintained the homeodomain-fold structure to interact with DNA, even though the species are separated by millions of years and have poor sequence homology in this domain [84].

Phylogenetic tree

The sequence similarities between the TM regions of the crystal structures are shown in S2 Fig (two residues are considered similar if their BLOSUM62 matrix entry is positive). The similarities are higher than 40% for proteins within the class A branches, from 34% to 54% across the class A branches, and 18–36% across the classes. Based on the GRoSS alignment, we computed the similarity for all the human GPCR proteins. The phylogenetic tree in Fig 7 graphically captures sequence similarity between all the proteins, which also indirectly corresponds to their structural differences (we compare these below). Even though evolutionary considerations were ignored when constructing this tree (for phylogenetic analysis see e.g. the Evolutionary Trace method [85]), this phylogenetic tree clearly contains evolutionary information, but it may miss the information encoded in the loops.

Fig 7. The phylogenetic tree based only on TM similarity using the GRoSS alignment (loops were ignored).

Color coding denotes the GPCR class. Proteins with known crystal structure are emphasized with a dot. The full resolution version of this figure is in S4 Fig.

The branches near the root of our tree are very sensible: First class C separates, then class B and adhesion proteins branch off, then class F, and finally class A comprising the rest of the tree. Evolutionarily, this is consistent with the recent most detailed analysis of 83 species [86], which showed that glutamate receptors (class C) and bacterial cAMP receptors are the oldest (>1400 MYA, million years ago), followed by class B and F (∼1275 MYA), and lastly rhodopsin-like receptors (class A)(∼1100 MYA). Except for several outliers, the first major branches to separate in class A are the sensory receptors: Vomeronasal, Taste2, and Olfactory. In the olfactory branch, the first split separates the fish-like receptors (families 51–56) from the tetrapod-like receptors (families 1–13). The subdivision of the rest of the class A does not follow the α−δ subclasses, but it is close. Near the leaves (i.e., for closely related proteins), the displayed tree might not provide the best classification, since our computation of similarity ignored loops. For related proteins, it may be advantageous to include similarity of the loops as well, since loops often interact with ligands, and therefore can determine receptor specificity. Indeed, it is somewhat surprising that using only the TM domain alignment of all human GPCRs, a phylogenetic tree can be constructed that correctly gets most evolutionary signatures of GPCRs. This suggests that the TM domains of GPCRs contain a good part of the signatures of the divergent evolution of GPCRs, including evolutionary separation of different classes, whose visible differences are usually seen in their soluble domains (N/C-termini and intracellular/extracellular loops).

Conserved inter-helical contacts involved in activation provide functionally important residues

We found that 40 inter-helical contacts are present in at least 23 out of 24 class A crystal structures (CHICOs in Fig 2). We infer that these residues are important for the interactions between the helices and that any changes to these residues may cause structural stability issues for the protein. Thus, naturally occurring mutations of the residues involved in the conserved contacts could be direct causes of physiological differences and/or diseases.

Comparing the common contacts among different proteins is not straightforward because many of the sequence differences appear random. Focusing on the difference between active and inactive conformations of the same protein makes the significance of the individual residues much clearer. There are 3 active-inactive crystal structure pairs available: RHO, β₂AR, and M2 with accepted “fully active” conformations. The active structure of A2A is only partially active, and for NTS1act the inactive structure is not available.

The main signature of activation for rhodopsin is breaking the R3.50↔E6.30 salt-bridge and forming of the K5.66↔E6.30 salt-bridge. Instead of keeping track only of hydrogen bonds, our analysis of contacts allows us to determine more general changes during the activation. The changes in structural contacts upon activation are shown in Fig 8 along with the list of contact residues that change, referred to as native activation “hot-spot” residues (NACHOs).

Fig 8. Native activation “hot-spot” residues (NACHOs), which are contacts that change upon receptor activation.

The width of the green lines is proportional to the number of contacts common to all six structures (RHO, β₂AR, M2, and their active structures). Blue shows the contacts present only in inactive structures, and not in inactive structures; while red shows the opposite. The upper diagrams show contacts in the extracellular half of the membrane. We see that there is no systematic change common to the class A receptors in the conformation of the extracellular half of the TMs. This is not obvious, because there are conformational changes accompanying ligand binding. All the systematic changes, which enable G protein binding, occur in the intracellular half of the TMs. The list only contains 15 different residues in 15 different contacts. Thus many of the residues switch partners upon activation.

An important observation is that the structural contacts in the extracellular half of the receptors do not change upon activation. All structural changes occur in the intracellular region, where the contacts get rewired upon the binding of the G protein. Most of the changes occur for TM 6, since the intracellular end of helix 6 undergoes the largest movement upon activation. However, TM 7 also shows a large number of systematic changes, as it breaks a contact with TM1 and creates new contacts with TM 2 and 3. The residues 3.43 and 3.46 occur in the list of conserved contacts in both active and inactive structures, therefore the conformational changes around these residues seem to be very important for the conformational changes during activation. The class A switching mechanism seems to rely critically on a small number of NACHO residues (15 residues). If any of these residues is mutated, the energy landscape of the active and inactive states might be modified, making the receptor likely to become either constitutively active or inactive, thereby altering or breaking its natural function.

Examples of mutations and natural variants modifying the function

It has been shown experimentally that single amino acid mutations can have a dramatic effect on GPCR activity. For example, the man-made mutation T3.46A makes the receptor CB1 fully inactive, while the mutations T3.46I and L3.43A make it constitutively active [87, 88]. Both positions, 3.46 and 3.43, are on the NACHO list of residues critically involved in activation. These particular mutations were introduced by experimentalists, but the NACHOs (Fig 8) are useful for judging the effect of natural variants as well. The positions of many single nucleotide polymorphisms (SNPs) are known from genetic studies, and by using the global GPCR alignment, we can determine the BW position of each SNP residue. Then each position can be directly compared against the list of activation hot-spots (and to the list of conserved contacts) to estimate the variant’s importance: whether the mutation causes some structural defects or whether it is likely to be benign. We scanned the Uniprot database [31] for naturally occurring mutations for all human GPCRs and converted the residue numbering to the BW scheme using our alignment. Table 3 provides several examples [87–98].

Table 3. Examples of natural variants and mutations that are associated with functional change or disease and which coincide with the NACHO residues.

Class	Protein	Uniprot	G-protein	Mutation	BW#	Activity Change	Disease Association	Reference
A-alpha	CB1	P21554	Gi/Go, Gs	T210A	3.46	Inactive	None	[87]
				T210I	3.46	Highly constitutively active	None	[87]
				L207A	3.43	Highly constitutively active	None	[88]
A-beta	V2R	P30518	Gs	R137Ca	3.50	Constitutively active	NSIAD	[89]
				R137La	3.50	Constitutively active	NSIAD	[89]
A-gamma	CCR5	P51681	Gi/Go	R126N	3.50	Disables G-protein coupling	None	[90]
A-delta	FSHR	P23945	Gs	R573Ca	6.36	Reduces AC stimulation	Ovarian dysgenesis 1	[91]
B	PTH1R	Q03431	Gs, Gq/G11	T410Pa	6.37	Constitutively active	JMC	[92]
				T410Ra	6.37	Active (less than T410P)	JMC	[93]
				H223Ra	2.43	Constitutively active	JMC	[92]
C	CASR	P41180	Gi/Go, Gq/G11, G12/G13	F788Ca	5.55	More active than wild type	Hypocalcemia	[94]
				F806Sa	6.36	No significant activating effect	Hypocalcemia	[95, 96]
				F788La	5.55	More active than wild type	Hypocalcemia	[97]
F	FZD4	Q9ULV1	G12/G13	K436Ta	6.36	Not known	Colorectal cancer	[98]
Predictions
A-alpha	DRD5	P21918	Gs	T297Pa	6.36	Predicted change of function	Not known
Adhesion	GPR56	Q9Y653	Gq/G11, G12/G13	M493Ta	3.43	Predicted change of function	Not known

^a Natural variant.

For example, the natural variants R3.50C and R3.50L cause the vasopressin V2 receptor to be constitutively active. This causes “nephrogenic syndrome of inappropriate antidiuresis”, which presents itself as an inability to excrete a free water load, resulting in low sodium levels [89]. The mutations of R3.50 clearly interfere with arginine’s ability to form hydrogen bonds, and so they disrupt the activation mechanism.

Similarly the natural variant H2.43R in Parathyroid hormone receptor causes its constitutive activity. This mutation of class B receptor causes “Jansen metaphyseal chondrodysplasia”, which is characterized by short-limbed dwarfism [99]. Since the same G proteins couple to different GPCR classes, we can expect the same or similar structural signatures of activation in class B as in class A.

For both of the above examples, the mutations are known to cause constitutive activity. However, there are many observed natural variants, for which the effect is unknown. For example, we predict that the natural variant M3.43T of GPR56 will influence its activation, because the residue 3.43 has to switch contact residues during activation. This adhesion GPCR is involved in cell adhesion as well as in cell to cell interactions, and regulates the migration of neural precursor cells; thus the mutation likely has serious consequences. No databases of single nucleotide polymorphisms contain any functional information about this mutation (we checked Uniprot, and the GPCR specific TinyGRAP [33] and NAVA [32] databases), therefore this is a new prediction based on the analysis of the GPCR fold presented here. Another prediction can be made for the natural variant T6.36P of the D₅ dopamine receptor. This is a class A receptor and it influences the activity of adenylyl cyclase. Again, we predict that the natural variant T6.36P dramatically changes activation response of this receptor, either to be more constitutively active or less active.

We have illustrated the importance of the NACHO residues by finding disease associations that are caused by single mutations at these positions. The list of NACHO residues only contains 15 residues, which is about 5% of the transmembrane domain. Similarly, we hypothesize that mutations at the CHICO positions (that define the structural fold) can dramatically change the receptor function. There are many known natural variants whose effect has not been experimentally studied yet, and these criteria can be used to focus (experimental) attention on variants, which cause dramatic changes.

From Uniprot, we collected all 2449 GPCR natural variants, of which about half (1289) lie in the TM regions. These are listed in S3 Table together with their functional or disease associations, if available on Uniprot. Table 4 summarizes the limited disease-association data available for mutations in GPCRs. It shows that about half (∼53%) of the GPCR TM residue mutations have been found to be associated with diseases. This number jumps to about two-thirds for CHICO or NACHO residues (∼67% and ∼66% respectively) and almost all (12 out of 13 or ∼92%) for residues that appear on both CHICO and NACHO lists. This strongly suggests that NACHO and CHICO residues can help prioritize mutation sites to guide experimental validation of the structural and functional hypotheses presented by these specific residues.

Table 4. Summary of SNPs annotated on Uniprot. The complete list is in S3 Table.

	Number of SNPs	With disease annotation	% with disease
All GPCRs	2489	694	27.9
All TM regions	1289	363	28.2
Excluding olfactory and unassigned	1463	635	43.4
TM	652	346	53.1
Non TM (Nterm+loops+Cterm)	811	289	35.6
CHICO only	161	105	65.2
CHICO	174	117	67.2
NACHO only	48	28	58.3
NACHO	61	40	65.6
Both CHICO and NACHO	13	12	92.3

There are still many SNPs in S3 Table that have unknown functional implications. We sort them with respect to a score capturing their relative position to the CHICO and NACHO residues: “distance to the closest NACHO + distance to CHICO - multiplicity of the closest NACHO - multiplicity of CHICO + Blosum62 of the mutation”. We hypothesize that the entries with the lowest score are very likely to cause dramatic changes in the receptor structure and function. The full list thus provides a large number of testable hypotheses about the molecular basis of disease-associated SNPs.

The CHICO and NACHO residues are results of a structural comparison, but the functional relevance of mutations is often obtained from phylogenetic considerations instead of structural ones. In S4 Table we compare these two approaches. We consider variations in sequences among a curated list of 77 P2Y12 orthologs [100], and among orthologs in an uncurated database for multiple proteins [101]. The CHICO and NACHO positions are more conserved than other TM residues in all GPCR classes among orthologs; and residues present on both lists are even more conserved. Thus both approaches are consistent, and should be combined to form more detailed insights.

Size of helix movement in available crystal structures and implications for homology modeling

By analysis of the inter-helical contacts we constructed the GRoSS alignment between all the GPCR proteins, from which new homology models can be derived. For structure prediction we would like to know how far the homology models are from the target structure. The variability of the TM bundle can be measured using the available crystal structures.

Fig 9 shows the observed move sizes, when the individual TM helices are treated as rigid bodies. Each pair of known structures was first aligned together, then each helix of the first protein was individually aligned to the corresponding helix of the second protein and the size of the move was measured. The center of mass translation was broken down into the direction along the helical axis and a direction perpendicular to it. The “tilt of axis” measures how much axis 1 had to be rotated to axis 2. And finally the “rotation around axis” measures the necessary rotation around the axis to map the corresponding atoms to each other.

Fig 9. Magnitude of the rigid body moves of the helices necessary to map one structure to another.

All TMs 1–7 from all available structure pairs were compared and each symbol denotes which TM is the data point from. The coordinate system is defined in the text. The maximal observed deviation is approximately proportional to the sequence dissimilarity of the two compared TMs, and it follows the same trend within class A (blue symbols) and across the GPCR superfamily (green symbols). The red symbols, which correspond to the active-inactive structure pairs, show rigid body moves caused by receptor activation. S10 Fig has an analogous plot of residual RMSD vs. similarity for each helix after the best rigid body transformation. RMSD shows a similar trend as the plots in this figure.

The maximal move sizes that need to be considered get smaller as the similarity of the TM sequence increases. If we are predicting a structure starting from a homology model with higher than 50% similarity, then we need only consider translating the helices up to 1.5 Å in any direction, tilting them up to 10°, and rotating around their axis by 40°. This is a very useful bound for refining homology models.

The same comparison can be applied to a single protein in multiple conformations. The red points in Fig 9 show the magnitude of rigid body moves undergone during activation for the 3 available pairs of active-inactive structures. Activation involves mainly the movement of TMs 5, 6, and 7. The computation of the move sizes ignores the bending of TM6 during activation, so it should be understood as an approximate description only.

Conclusion

A conceptual understanding of the molecular mechanisms behind biased signaling and functional selectivity is emerging around the conformational flexibility and dynamics of GPCRs. This is supported by an ever-increasing number of experimental and computational studies [102–107] that point to different ensembles of receptor conformations behind the pleiotropic signaling of GPCRs. These mechanistic studies have gotten a significant boost from the recent dramatic developments in GPCR structure determination methods.

We constructed the GRoSS sequence alignment of the transmembrane regions for all known human GPCRs. Although the inter-helical contact residue correspondence in the GRoSS alignment is in many cases approximate (intra-class) or non-existent (inter-class), several inter-helical contacts are highly conserved across the classes, which suggests their importance in the evolutionarily conserved GPCR fold. Our conserved contact analysis of the experimentally observed inactive and active conformations of Rhodopsin, Muscarinic M2, and adrenergic β₂AR identified 15 residue positions in the TM regions that change molecular contacts upon activation. Targeted or natural mutations of these residues are known to cause dramatic changes in the receptor signaling. We recommend that these be the starting point for examining mechanisms for activation and for deleterious mutations. The GRoSS alignment of the TM domains also leads to a functional phylogenetic tree that captures many evolutionary signatures of GPCR evolution. This shows that the class-level differences in the GPCR superfamily are encoded in the TM domains even though the divergence in the loop domains is usually used to distinguish the classes from each other.

The GRoSS alignment is also a promising starting point for structure prediction, as there are no gaps present in the TM domains. For a protein in question one can build homology models based on any of the available templates by mutating the corresponding amino acids. While the increasing coverage of proteins by the crystal structures makes it easier to find a close template, the GRoSS alignment allows remote homology modeling, which can be particularly useful for modeling active states for all GPCR classes. GPCRs are too large (>300 residues) for exploring larger conformational changes only using molecular dynamics. Comparing the available structures with respect to the GRoSS alignment gives approximate bounds on the size of rigid body moves needed for the TM helices to reach the target structure.

The GRoSS alignment is unique in aligning all human GPCR sequences by maximizing the number of conserved inter-helical contacts. These conserved contacts provide a basis for defining a GPCR superfamily-wide structural fold, functionally conserved residue positions (even if residue type may not be conserved), and activation hot-spot residues (NACHOs).

Supporting Information

S1 Table. List of studied GPCR crystal structures.

When multiple structures are available, then the one with the highest resolution or the one with least deformed TM helices is used.

(PDF)

S2 Table. GRoSS sequence alignment for all 817 human GPCRs.

S1 File has this alignment in fasta format. Since there are no gaps in the TM domains, the alignment of each protein is uniquely determined by the BW.50 residues for each TM 1 through 7. We list also the expected range of the helical TM regions, which is estimated as the average TM region in the known crystal structures from the same class. In the discussion of the bitter taste receptors (TAS2Rs), we identified two possible alignments of TM6, but only the first one is presented in the following table. The second choice is to decrease the start, end, and BW50 residue of TM6 by 4.

(CSV)

S3 Table. GPCR natural variants annotated by Uniprot mapped to BW numbering and indicating their proximity to the NACHO and CHICO residues.

The mutations are ordered according to the following score: “distance to the closest NACHO + distance to CHICO - multiplicity of the closest NACHO - multiplicity of CHICO + Blosum62 of the mutation”.

(CSV)

S4 Table. Conservation of CHICO and NACHO residues among orthologs.

For orthologs of several proteins we computed average amino acids conservation over TM, and over CHICO/NACHO residues. The data shows that CHICO and NACHO positions are more conserved than other TM residues in all GPCR classes. Residues present on both lists are even more conserved. Two measures of conservation provided by Jalview are used: Consensus is the percentage of orthologs sharing the human amino acid; and Conservation is a qualitative measure counting the number of conserved chemical properties. For P2Y12, we used a curated list of 77 orthologs from [100]. For other proteins, we collected predicted orthologs from the MetaPhOrs database (release 201405 [101]), aligned them with Clustal Omega, and then removed sequences with gaps in the TM regions.

(PDF)

S1 Fig. Detailed view of conserved motifs in class A GPCRs.

The conserved residues in 24 different structures (including active) have very similar positions, which shows that the class A GPCR fold is highly conserved. The full TM bundle is shown in Fig 1.

(PDF)

S2 Fig. Sequence similarity (%) of the TM bundles between crystal structures for the final sequence alignment.

Two residues are similar if their BLOSUM62 entry is positive.

(PDF)

S3 Fig. Backbone (atoms N, C_α, C, O) RMSD of the TM bundles for the final sequence alignment.

For a given pair of structures, there may exist a different sequence alignment, which results in a lower RMSD than the listed one.

(PDF)

S4 Fig. High-resolution phylogenetic tree (Fig 7) based on TM similarity only.

The pdf file is searchable for the UNIPROT accession numbers. Loops were ignored. Color coding denotes the GPCR class. Proteins with known crystal structure are emphasized with a dot.

(PDF)

S5 Fig. Testing the robustness of the alignment of the Vomeronasal receptors with the other groups.

This is an extended version of Fig 3, same caption.

(PDF)

S6 Fig. Testing the robustness of the alignment of the Taste2 receptors with the other groups.

This is an extended version of Fig 4, same caption.

(PDF)

S7 Fig. Testing the robustness of the alignment of the Vomeronasal receptors with the GPCRtm substitution matrix.

Same caption as in Fig 3.

(PDF)

S8 Fig. Testing the robustness of the alignment of the Taste2 receptors with the GPCRtm substitution matrix.

Same caption as in Fig 4.

(PDF)

S9 Fig. Diagram of interhelical contacts present in classes B, C, and F.

The width of the line connecting two TMs is proportional to the number of contacts present in all structures from the given class. The list in red font shows the contacts not present in any available structure from other classes.

(PDF)

S10 Fig. RMSD of helices after best rigid body move.

Same caption as Fig 9.

(PDF)

S1 Text. Comparison of the GRoSS alignment to the HMM-HMM alignment [77] and to the GPCRDB alignment [24, 78].

(DOCX)

S1 File. The GRoSS alignment in fasta format and annotation of the TM regions and BW residues in Jalview format.

The first 29 sequences are the actual sequences from the PDB files of used crystal structures; the rest of the sequences are from Uniprot. N-terminal, loops and C-terminal are not aligned. For interactive work it is useful to also highlight the TM regions and BW residues using the Jalview annotation gross-alignment.gff file.

(ZIP)

Data Availability

All relevant data are within the paper and its Supporting Information files. They are also available in a convenient searchable form at http://www.gpcrs.org.

Funding Statement

This work was supported by: Gifts to the Materials and Process Simulation Center at Caltech, NIH (R01NS073115 and R01AI040567), Defense University Research Instrumentation Program (DURIP) grant (N00014-12-1-0818), and NSF (EFRI-1332411) to WAG; College of Science & Engineering and the University of Minnesota Supercomputing Institute to VC; and Cedars-Sinai Medical Center (Startup Funds) to RA. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.