Three types of sequence-based phylogenetic trees were built, namely: one tree that was based on the full 7-TM sequence, one tree employing 30 residues described by Surgand et al. 15 , and one tree which was based on the set of 44 residues described by Gloriam et al. 17 . Note that the three sequence-based trees presented here are different from those published in the referenced original work 1 15 17 , since in the current study orphan receptors, receptors with a low number of ligands, and singleton receptors were left out. Singleton receptors are receptors that are the only (available) member in their respective subfamily. Due to the chemogenomic nature of this study, we focus on the phylogenetic tree based on the set of Gloriam et al. since it represents the ligand perspective best; this set is referenced as the GSK set 17 . The two other trees are provided for reference purposes in Additional file 1 - Phylogenetic trees based on 7TM domain and selected residues. The tree that was built based on the multiple sequence alignment of the GSK set is shown in Figure 1. The GPCR subtypes in this tree are grouped as branches in the tree according to subfamily and target since it resembles the sequence-based phylogenetic tree on which GPCR classification is based 1 . For instance, the opioid receptor subtypes δ, κ, μ, and NOP cluster together, as well as the α- and β-adrenoceptor subtypes. The fact that clustering follows the receptor classification is expected since the classification of GPCRs was based on sequence similarity 24 25 . Four clusters are clearly defined in the tree: the aminergic receptors, the adenosine receptors, the prostanoid receptors, and the peptide-binding receptors.

The ligand-based receptor classification, which we will compare to the sequence-based classification, is provided in Figure 2. Subfamilies in this tree are more scattered; however, most subfamilies cluster together. For instance, except for the two purinergic receptors (P2Y₁ and P2Y₁₂) and the two glycoprotein hormone receptors (FSH and LH), all other receptors represented by only two subtypes, such as the melatonin or the leukotriene B₄ receptors, are clustered together. The adenosine receptors A₁ (ADORA1), A_2A (ADORA2A), A_2B (ADORA2B), and A₃ (ADORA3) group together, indicating overlap in ligand profiles. This may imply that ligands for these receptor subtypes are non-selective, such as the adenosine receptor antagonists caffeine and theophylline. Additionally, receptor selectivity may vary with relatively small changes in ligand structure: an 8-cycloalkyl substituent on theophylline confers A₁ receptor selectivity, whereas a phenylstyryl substituent on the same position in caffeine renders these compounds selective for the A_2A receptor. The purinergic receptor P2Y₁₂ is found near the adenosine receptors owing to the purine core typical for ligands of both these subfamilies. In agreement with the ligand selectivity reported for the α₁-, α₂-, and β-adrenoceptor subfamilies, these receptors form three distinct clusters 26 ; furthermore, the α_1B and α_1D receptors are the closest in the distance matrix. The muscarinic acetylcholine receptors M₁, M₃, M₄, and M₅ (CHRM1/3/4/5, in Figure 2) cluster together as one group, supporting the low subtype selectivity of muscarinic antagonists 27 . However, the acetylcholine receptor M₂ is found more distant from this cluster. This indicates the presence of distinct chemical classes in the ligand set of the M₂ receptor, which may be the result of inclusion of allosteric ligands. For instance, gallamine is an allosteric modulator of the muscarinic M₂ receptor 28 that is also present in the GLIDA database 29 , classified as an M₂ antagonist. In general, the remaining aminergic receptors (serotonergic, dopaminergic, histaminergic and cholinergic) are more scattered throughout the substructure tree. This means that targets share ligands or ligand substructures among subfamilies/subtypes, which is in line with the high level of polypharmacology observed for these aminergic GPCRs 30 . For instance, the serotonin receptor 5-HT_1A clusters together with the D₂ dopamine receptor, which fits with reports on antipsychotic compounds combining dopamine D₂ receptor antagonism and serotonin 5-HT_1A receptor agonism 31 32 . Structurally similar ligands may act on diverse targets, for instance, when ligands have a GPCR-privileged structure at their core 33 34 . The grouping of the eight prostanoid receptors (Figure 2) indicates similarity in substructure profiles of the ligands. This is based on the fact that most prostanoid receptor ligands are direct derivatives of the endogenous ligands 35 36 , the so-called eicosanoids. These ligands are highly similar, all consisting of large aliphatic, lipophilic alkyl chains. The presence of the leukotriene and cannabinoid receptors in this lipid cluster may seem strange at first. Leukotrienes are however also eicosanoids, which clarifies the position of the leukotriene B₄ and cysteinyl-leukotriene receptors in this cluster 37 38 . In addition, arachidonic acid is the common precursor for eicosanoids and two derivatives of arachidonic acid, anandamide and 2-arachidonylglycerol, both of which are endogenous ligands ('endocannabinoids') of the cannabinoid receptors.

The relationship between target clustering in the substructure tree (Figure 2) and ligand promiscuity suggests that the substructure tree may be used to identify possible side effects on receptors that are close neighbors in this tree. For instance, off-target activity of ligands can be identified. If inspection reveals a ligand to bind to receptor(s) that are phylogenetically related to the target of interest, a more detailed experimental follow-up with respect to receptor selectivity would be worthwhile.

Visual comparison of the sequence tree (Figure 1) with the substructure tree (Figure 2) reveals that the overall phylogenetic organization is similar. For instance, with the exclusion of the glycoprotein, P2Y, angiotensin, and bradykinin receptors, all other receptors represented by two subtypes occur in pairs in both the ligand tree and the sequence tree. This is also true for receptors with three subtypes present in the dataset, e.g. the three members of the α₁, the α₂, and the β₁ adrenoceptors, as well as the bombesin receptors. Exceptions to this rule are the neuropeptide Y and vasopressin receptors. In addition, the prostanoid receptors largely group together in both trees, as do most of the aminergic receptors.

The clear distinction between the two dopamine receptor types, i.e. D₁ and D₅ (D₁-like) versus D₂, D₃, and D₄ (D₂-like), exists both in the sequence-based classification and ligand-based classification. This is in agreement with a previous study 39 and also known from drugs on the market such as the benzazepines that favor D₁-like over D₂-like dopamine receptors. Similarly, antipsychotics such as chlorpromazine have a higher affinity for the D₂-like subtypes than D₁-like receptors 40 .

The fact that many clusters arise in both trees indicates that the receptors in these clusters have similar sequences and similar ligands, that is, ligands with substantially overlapping substructure sets. However, there are also receptor targets for which this is clearly not the case. The (qualitative) similarities and differences among sequence and substructure trees are discussed in the following. A delta-delta plot was constructed to compare how pairs of receptors change. This plot, provided in Figure 3 (and described in detail in the Materials and Methods section), visualizes how receptor distances deviate between the sequence-based tree and the ligand-based classification of receptors. In sequence space, receptor distances indicate the (dis)similarly between protein sequences, while in ligand space, receptor distances reflect the overlap in structural features found in ligands for these receptors. For each receptor, the mean distance to all other receptors is plotted. From the delta-delta plot, it becomes apparent that the prostanoid receptors and P2Y₁ receptor are on average the most distant receptors from the rest of the classes. The distances of the purine P2Y₁ receptor, the prostanoid FP receptor, and leukotriene receptor CysLT₂ towards the other classes are all larger in substructure space than in sequence space, implicating that overall their ligands show little resemblance with ligands of the other GPCRs. In contrast, for most aminergic receptors, e.g. for the α_2B-adrenoceptors and the 5-HT_2B serotonin receptor in Figure 3, distances are smaller in substructure space compared to sequence space. This, again, corresponds with the high polypharmacology found for aminergic ligands, such as for most atypical antipsychotics 41 , with clozapine as a prominent example 42 . With the exception of a few targets (FSH, LH), the distribution of targets in the delta-delta plot is more scattered along the x-axis (substructure space) than the y-axis (sequence space). This may be a reflection of the evolutionary relationship between sequences, which results in coverage of a small region of the overall sequence space. The ligands for these targets do not have such a direct relationship and thus cover a broader range in overall substructure space.

The difference between ligand-based and target-based classifications may be due to convergent evolution 43 . Functional convergence denotes how proteins that differ in sequence may fulfill the same protein function. The protein sequence of GPCR subtypes will be similar in parts that are involved in the endogenous ligand recognition but may be different in other parts, for instance those parts that play a role in recognition of other, exogenous, ligands (e.g. synthetic drugs). These may therefore have a different selectivity profile compared to the endogenous ligand.

To validate how well our method performed as a chemogenomics method, i.e. how well it connects sequence space with small molecule space and how applicable the relationship is in practice, we conducted a 'virtual de-orphanization exercise'. For each receptor in the dataset, we pretended not to know any of its ligands by excluding them from the datasets (we 'orphanized' the receptor in this particular run of the protocol). We next predicted its ligands by considering a model derived from the closest neighbors of the receptor in sequence space (we attempted to 'de-orphanize' the receptor whose ligands we omitted from the study in the previous step). For this calculation, the distance matrix for the GSK residue set was used. The cumulative number of correctly identified ligands of every receptor is plotted against the number of closest neighbors (sequences) included to find these ligands. The (relative) area under the curve (AUC) and shape of the curve are measures of the performance of our method. In 93% of the studied receptors, de-orphanization of the pretended orphan receptor using the ligands of related receptors performed better than random (AUC > 0.5) and for 35% of receptors de-orphanization performance was good (AUC > 0.7). All AUC plots could be divided into four categories according to curve shape and AUC (the complete set of plotted scores is available as additional material in Additional file 2 - Plotted scores for the leave-one-out validation). Typical examples of the four categories are given in Figure 4. The first category is most abundant and consists of curves with a convex shape and an AUC above 0.5, marking good performance. An example of this category is the muscarinic acetylcholine receptor M₁ (CHRM1 in Figure 4) with an AUC of 0.7990. Curves of the second category display a gradual rise that is approximately equal to the diagonal of the plot. These plots have an AUC near 0.5, indicating performance that is equal to random prediction. An example is the plot of the angiotensin receptor AT₁ (AGTR1 in Figure 4) with an AUC of 0.5120. Curves of the third category perform worse than random and are characterized by a concave shape and an AUC below 0.5. Clearly the worst example is the P2Y₁ purinoceptor with an AUC value of 0.0857 (P2RY1 in Figure 4). In contrast to the first three categories, curves of the fourth category do not have a clear AUC range. This category consists of curves that are divided into several discrete parts of alternating rises and plateaus, as shown in the plot of bombesin receptor BB₃ (BRS3 in Figure 4), with an AUC of 0.8145. Performance varies from good (BRS3) to worse than random, depending on the value of the AUC. An example of such a plot with an AUC value below 0.5 is the FSH receptor (not shown, see: Additional file 2 - Plotted scores for the leave-one-out validation) with an AUC of 0.4428. The steep rises are caused by a few receptors identifying the majority of ligands. Some of these curves are steeply rising at the start, which suggests that part of its ligand set could be readily identified even though this is not reflected in the AUC. The poor performance concerning the P2Y₁ receptor is probably due to the nature of its ligands: this set consists of a small number of highly similar ligands that all possess a phosphate group, a feature not found in other ligands in the database. The number of features (substructures) shared with ligands of this receptor and other receptors is therefore small. Interestingly, the adenosine A₁ and A₃ receptors, which are also purinergic, identify most (28 out of 42) of the P2Y₁ ligands. However, in sequence space these receptors are at great distance (at positions 91 and 92, respectively).

Overall, our method proves useful for receptor de-orphanization, since for 93% of receptors studied de-ophanization performed better than random selection (AUC > 0.5) and for 35% of receptors de-orphanization performed well (AUC > 0.7).

In the present study, some targets were excluded due to insufficient availability of ligand data in the source databases. The absence of a receptor may influence the order of other receptors in the trees. Scarcity of ligand data is reflected in the substructure profiles, thereby influencing the correlations among receptors. The issue of data (in) completeness and its effect on interaction networks was recently discussed by Mestres et al. 44 . Using three datasets of increasing complexity (more connections) that linked ligands to targets based on full chemical identity, the authors showed that an increase in the number of connections rapidly leads to shifts in connection patterns. However, our study linked targets based on overlap in substructures; as a consequence sharing of substructures rather than of ligands is sufficient for targets to be identified as related. Bender et al. and Keiser et al. already showed that overlapping ligands are not necessary to predict whether targets are close in ligand space 19 20 . In addition, our method employs an exhaustive approach to analyze the structural features of ligands. Frequent substructure mining considers all possible substructures that occur in the ligands and is therefore unbiased, i.e. all possible substructures were evaluated, not only those intuitive to chemists, such as functional groups, ring systems (e.g. a phenyl ring), and linkers 45 . However, in the present study less 'obvious' substructures such as ethyl or isobutyl are also considered 21 . For a complete discussion on substructure generation and evaluation, see ref. 46 . Our method is not limited to GPCRs alone; it is easily extended to other protein families for analysis of the differences between subfamily phylogenies, given that sufficient ligand information is available. For instance, it can be applied to the realm of enzymes to complement other chemogenomics analyses 47 .

For the comparison of trees, several methods and visualizations are available; however, there is not a single definitive measure for tree difference. To visualize how the receptor positions change between two trees we employed a delta-delta plot.