The trillions of cells in the human body rely on receptors that sit in their cell membranes to communicate with each other. Hundreds of different receptors belong to the G protein-coupled receptor superfamily (called GPCRs for short) and play vital roles in the all organs and bodily systems. Indeed, GPCRs are the targets for almost 40% of therapeutic drugs. As such, deciphering the shape and activity of GPCRs is key to understanding the normal workings of the human biology and could help scientists discover new treatments for various diseases, from depression to high blood pressure to cancer. These receptors, however, are notoriously flimsy and unstable, making them difficult to work with in the laboratory.

Different approaches have been developed to make GPCRs more stable, usually by swapping one or a few of the amino acid building blocks in the protein for other amino acids. Currently, this requires a costly and slow trial-and-error approach in which each amino acid out of 300-400 in the protein is mutated and tested experimentally.

To speed up and reduce the cost of the process, Popov et al. asked if a computer could predict which mutations in the protein would stabilize it, meaning that fewer proteins would actually need to be tested. Four computer algorithms based on four different principles were developed and verified. The first one compares the target GPCR to other closely related receptors, trying to detect variations that cause the instability. The second tries to build in specific stabilizing interactions, or “bridges”, between different parts of the receptor. The third algorithm searches the known structures of other GPCRs for useful mutations. Finally, the fourth one uses accumulated data on the stability of hundreds of mutations in different GPCRs to train a machine learning predictor to recognize stabilizing mutations.

All four algorithms produced useful predictions in a real-life project. Indeed, when combined in one computational tool, named CompoMug, the algorithms made it possible to detect optimal mutations in a human GPCR called 5-HT_2C. This made the protein much easier to work with in the laboratory, and ultimately helped to solve its three-dimensional structure (which was reported in a separate study, published earlier in 2018)

The 5-HT_2C receptor is involved in regulating, among other things, mood and appetite. Details of its structure might therefore help researchers to design new antidepressants and obesity treatments. Moreover, CompoMug is already helping structural biologists to solve the structures of other GPCRs, which will further facilitate many aspects of GPCR drug discovery.

G-protein coupled receptors (GPCRs) represent the largest family of transmembrane proteins, which is involved in regulation of all major physiological functions and comprises more than 25% of established therapeutic targets (Lagerström and Schiöth, 2008; Rask-Andersen et al., 2014). However, high conformational flexibility and low thermostability of these receptors have always presented major challenges for their structural, biophysical, and biochemical characterization. With exception of visual rhodopsin, structural characterization of all other 50 GPCRs so far required substantial efforts in protein engineering to design GPCR constructs suitable for crystallization (Cherezov et al., 2007; Warne et al., 2009; Chun et al., 2012; Katritch et al., 2013; Stevens et al., 2013; Pándy-Szekeres et al., 2018). The design typically involves truncations of N- and C- termini, replacements of flexible loops and/or termini with soluble fusion domains (Chun et al., 2012), stabilizing co-crystallization partners (Zhang et al., 2015), and in many cases introduction of one or more point mutations (reviewed in (Heydenreich et al., 2015)).

Point mutations have shown to be especially important for thermostabilizing GPCR and making them amenable for structure-based drug design applications, which involve receptor co-crystallization with typically low-affinity hit or lead compounds. For example, point mutations were used to thermostabilize β₁-adrenergic (ADRB1) and A_2A adenosine (A_2AAR) receptors in both agonist and antagonist bound states, resulting in more than a dozen structures for each receptor including co-crystals with ligands in a micromolar affinity range (Moukhametzianov et al., 2011; Christopher et al., 2013; Warne et al., 2008; Warne et al., 2011). In the case of thermostabilized A_2AAR, structural and biophysical characterization of initial hits led to structure-based discovery and optimization of preclinical candidates for Parkinson disease (Langmead et al., 2012). Moreover, thermostabilized GPCR constructs can streamline biochemical characterization of ligand binding in sensor-based high-throughput screening (HTS) (reviewed in (Kumari et al., 2015)) and measurements of ligand-binding kinetics by surface plasmon resonance (SPR) (Christopher et al., 2013; Congreve et al., 2011; Rich et al., 2009).

However, currently employed experimental identification of stabilizing mutations by alanine scanning (Errey et al., 2015) or directed evolution approaches (Egloff et al., 2014; Schlinkmann et al., 2012) is a very resource consuming process, and only a few GPCRs have been successfully stabilized so far (reviewed in (Vaidehi et al., 2016)). Furthermore, stabilizing mutations obtained by these methods have shown very limited transferability between different GPCRs (Heydenreich et al., 2015; Serrano-Vega and Tate, 2009), requiring extensive stabilization campaigns to be performed for each individual receptor.

Computational approaches could provide a cost- and time-effective alternative for GPCR stabilization. The already existing in silico prediction tools for soluble proteins (Kumar et al., 2006; Khan and Vihinen, 2010), however, are not effective for GPCRs because they do not take into account peculiarities of the 7-transmembrane (7TM) nature of the receptors. At the same time, although some of the recently developed GPCR-specific methodologies can be successful in explaining known experimentally-derived mutations (Vaidehi et al., 2016; Bhattacharya et al., 2014), their success in prediction of new stabilizing mutations has been limited so far, and has not resulted yet in successfully solved crystal structures of new GPCRs.

In this study, we present a set of complementary approaches for predicting stabilizing mutations in GPCRs combined into a CompoMug tool (COMputational Predictions Of MUtations in GPCRs). CompoMug consists of four modules: knowledge-based, sequence-based, structure-based, and machine-learning-based, taking maximum advantage of accumulated structural and biophysical data. We applied CompoMug to identify stabilizing point mutations for the 5-HT_2C receptor, which is an important pharmacological target for the treatment of obesity and neuropsychiatric disorders. Experimental assessment showed that 10 out of the 39 predicted mutations improved stability of the receptor by more than 1.5°C, and one mutation resulted in increase of the apparent melting temperature by up to ~8.8 ± 1.3°C, which is among the highest reported improvements in thermostability by a single point mutation in GPCRs. Moreover, combinations of two or three mutations led to even higher thermostability gains, some of which were compatible with both agonist and antagonist binding. Finally, the mutants predicted by CompoMug allowed for the determination of two 5-HT_2C crystal structures in both agonist-bound and antagonist-bound complexes. The CompoMug provides a computational platform for thermostabilization of other GPCRs and can be further evolved with an accumulation of experimental mutation data.

CompoMug consists of four modules: knowledge-based, sequence-based, structure-based, and machine-learning-based, - combining several approaches to compose a list of the candidate point mutations, which can improve the stability of a GPCR. The general workflow of CompoMug is presented in Figure 1. Below we describe each of the modules in details.

Figure 1

Download asset Open asset

Sequence-based module

Request a detailed protocol

The sequence-based module looks for residues of the target receptor that deviate from a standard conservation pattern in an evolutionarily related group of GPCRs, e.g. receptor orthologs, a subfamily or a branch of the GPCR tree. We hypothesized that such residues in GPCRs are more likely to be destabilizing, and restoring conserved amino acids in such positions might result in receptor stabilization. In CompoMug, the ‘deviation score’ for an amino acid residue is estimated based on multiple sequence alignment (MSA) of evolutionary related homologous sequences:

(1) ${\displaystyle Scor{e}_k^{aa}=\frac{C_k^{max}-C_k^{aa}}{N_{MSA}}-\frac{C_k^{aa}}{C_k^{max}},}$

where $N_{MSA}$ is the total number of sequences in the MSA, $C_k^{max}$ is the number of sequences with the most conserved amino acid residue at the position k, and $C_k^{aa}$ is the number of sequences that have the same residue aa as the target sequence in this position. As one can see from Equation 1, the first term is the highest when the target sequence has the most infrequent amino acid in the position k, that is, it approaches 1, when $C_k^{aa}=1$ and $C_k^{max}\approx N_{MSA}$. The second term penalizes the position k if it lacks a dominating conserved amino acid at the position, that is, the penalty is increased as $C_k^{max}$ is decreased. The total score varies from -1.0 to 1.0, where maximum score 1.0 is ascribed to a deviating amino acid in a super-conserved position (e.g. x.50 in GPCRs). In other words, the preference is given to mutations of rarely observed amino acids in the otherwise highly conserved positions. Figure 2A and B schematically show the score computation given an MSA. Apparently, any conservation-related score depends on the set of sequences used to construct the MSA. For example, orthologs share very high sequence similarity with respect to the target GPCR resulting in a few, but usually very clear deviation patterns. On the other hand, comparison with GPCR sequences from different branches has a much more complex conservation pattern that may result in many false positive candidates. To capture the sequence deviations at different levels of GPCR hierarchy, we composed several sets of sequences to construct various MSAs. Specifically, we used five MSAs: (1) ortholog sequences corresponding to the species variations of the target receptor, (2) sequences corresponding to the common sub-family (sequence identity for the TM regions >40%), (3) sequences corresponding to the common GPCR branch (sequence identity for the TM regions >30%), (4) sequences corresponding to the whole non-olfactory class A GPCR (Rios et al., 2015), and (5) sequences corresponding to the crystallized receptors. MSAs were generated with the structure-based alignment tool of the GPCRdb (Isberg et al., 2015), and in case of the whole class A alignment updated using MAFFT software (Katoh and Frith, 2012).

Figure 2

Download asset Open asset

Although the last MSA is not directly related to the evolutionary variation, it may contain information relevant for the GPCR stability and propensity for crystallization. At the same time, the MSA for whole class A GPCR would capture rare variations in the most conserved residue positions of class A, including N^1.50, D(E)R^3.50Y, FxxxCWxP^6.50 and NP^7.50xxY. Given all five MSAs, we computed positional scores for each MSA, as well as the global score as the average of the individual MSA scores. In a special case of non-conserved Gly residues, we multiplied the ‘deviation score’ by factor of 2, to account for Gly usually destabilizing effect on α-helical secondary structure in the transmembrane helices of the receptor. Figure 2 schematically shows the workflow of the sequence-based module applied to the 5-HT_2C receptor.

Structure-based module

Request a detailed protocol

The structure-based module is focused on identifying pairs of residues, which could form a salt bridge (also called ionic lock) when replaced with charged amino acids, or disulfide bonds when replaced with cysteines. Such ionic locks and covalent bonds can help to restrict the conformational flexibility of the receptor and improve stability. A successful use of the structure-based approach requires an accurate 3D structural model, which can be derived based on the close homology with a known crystallographic structure. In this study, structural models were obtained using the template-based homology modeling implemented in ICM-Pro v.3.8 molecular modeling suite (molsoft.com), followed by the backbone regularization and exhaustive Monte-Carlo side-chain refinement in internal coordinates. To predict potential ionic locks in the structural model, the search is performed for pairs of residues that satisfy the following criteria: i) residues are separated in sequence by at least five residues to exclude pairs of residues belonging to the same α-helix, ii) side chains point toward each other and do not point to the lipid membrane, iii) residue’s C_β-C_β distance lies in the range from 7.0 Å to 10.0 Å, and iv) mutations of residues to at least one of four charged pairs (E-K, E-R, D-K, D-R) improve predicted free energy of the receptor after thorough local conformational optimization of the mutants (Equation 2)

(2) ${\displaystyle E_{folded}^{mut}-E_{unfolded}^{mut}<E_{folded}^{wt}-E_{unfolded}^{wt}}$

We used energy calculation implemented in the Molsoft ICM-Pro v.3.8. software (molsoft.com). The structural model of the mutant type was obtained by mutation of a given residue followed by Monte Carlo sampling of the flexible side chains for the mutated residue and the neighboring residues. Then the free energy of the unfolded and folded states for the wild and mutant types was approximated by a sum of the empirically derived residue-specific energies.

In order to predict stabilizing disulfide bonds in the receptor, we first employed the DbD software (Craig and Dombkowski, 2013) to obtain the initial list of candidates. DbD scans all pairs of residues in a protein and selects those that satisfy geometrical parameters of the disulfide bond, when replaced with cysteines. The geometrical parameters, e.g. $\chi 3$ angle and $C_{\beta }-C_{\beta }$ distance, were obtained from analysis of protein structures in PDB. Given the DbD predictions, the final list of candidates was derived using the energy criterion implemented in ICM-Pro (see Equation 2). Figure 3 schematically represents the structure-based module.

Figure 3

Download asset Open asset

Machine learning module

With the accumulation of experimental data on the stability of GPCR mutants, it becomes feasible to derive powerful prediction models using machine learning techniques. Our prediction model is derived using (i) a training benchmark, composed from site-specific mutations performed on GPCRs with known structure, (ii) a feature vector, consisting of structure-based and energy-based descriptors, which reflect important changes in the protein upon a point mutation, and (iii) a support vector machine method as implemented in the libsvm package (Chang and Lin, 2011). Each of these steps is described below in details.

Training benchmark

Request a detailed protocol

To compose the training benchmark we used available alanine scanning mutagenesis data for three GPCR receptors: neurotensin receptor NTS1 (Shibata et al., 2009), A_2A adenosine receptor (Magnani et al., 2008), and β₁ adrenergic receptor ADRB1 (Serrano-Vega et al., 2008). Point mutations that improve thermostability of these receptors were used as positive examples, while reverse mutations were used as negative examples for training. Further, in order to expand the training benchmark, we considered the remaining alanine mutations, that is, those which were not reported as stabilizing, as negative examples. It is worth to note that such assumptions may introduce some false negative examples into the training set, because some of the alanine mutations were filtered out due to the lower expression level, rather than due to a decrease in the receptor stability. Overall, the training benchmark consists of 79 stabilizing point mutations and 923 non-stabilizing point mutations.

Feature vector

Request a detailed protocol

Given the training set, we projected each point mutation as a vector onto a feature space, where the coordinates of the feature vector encode information relevant to a change in the receptor stability upon introducing the point mutation. To compose a feature vector, we used characteristics of three different types. Namely, for wild type and mutated residues we used sequence-based characteristics, which could be extracted from the primary structure of the protein (hydrophobicity, polarity, charge, side chain volume, solvent-accessible area, polarizability), structure-based characteristics, which could be extracted from the secondary and the tertiary structures of the protein (number of polar, charged, hydrophobic, and aromatic contacts, residue exposure, contact area, void volume, relative accessible solvent area), and energy-based characteristics, which could be extracted from the tertiary structures of the protein given the force-field (potential of mean force, electrostatic, van der Waals, solvation, hydrogen bond, and total energies). To obtain a structural model of a mutant type we mutated a given residue and performed Monte Carlo minimization with flexible side chains of the mutated residue and its neighboring residues, keeping the rest of the receptor rigid, using the Molsoft ICM-Pro v.3.8. software (molsoft.com). To calculate components of the feature vector we used built-in functions of ICM-Pro.

Support Vector Machine classifier

Request a detailed protocol

Feature vectors computed for each point mutation in the training benchmark are then combined into the feature matrix. Each entry in the feature matrix is labeled with +1, if the corresponding point mutation stabilizes receptor, or −1 otherwise. Given this mapping of point mutations into the feature space, one can construct a hypersurface which separates +1 feature vectors from −1 feature vectors, using the support vector machine (SVM) approach. We used the libsvm package libraries (Chang and Lin, 2011) to accomplish this task. There are two free parameters in this classification problem. Namely, the regularization parameter C, which is a tradeoff between the misclassification and ‘smoothness’ of the separating hypersurface, and the kernel parameter γ, which corresponds to the variance of the radial basis function. Optimal values for these parameters were defined using the two-fold cross-validation procedure (see below). Figure 4 schematically represents the machine-learning-based module.

Figure 4

Download asset Open asset

Figure 5

Download asset Open asset

Cross-validation of the machine learning classifier

Request a detailed protocol

To validate the machine learning classifier the data set was randomly split into two parts: the training part, which consists of 65% of the training set, and the validation part, which consists of the remaining 35%. Each subset was adjusted to retain the ratio of the +1 feature vectors, corresponding to stabilizing point mutations, and −1 feature vectors, corresponding to non-stabilizing point mutations. To optimize the performance of the algorithm, we scanned values of the two free parameters (C and γ) on a grid [0.0, 50.0]×[0.0, 50.0] with a step size of 0.2 for both parameters, thus, yielded 250*250 = 62,500 different prediction models. Then for each prediction model, we calculated the positive predictive value (PPV), which is the ratio of true positive rate (TPR) and the sum of true positive and false positive rates (FPR) predictions:

(3) $PPV=\frac{TPR}{TPR+FPR}$

We repeated this procedure 10 times and scored parameters, based on the rank of each pair (rank one corresponds to the maximum PPV) in each cross-validation run:

(4) ${\displaystyle Score\left(C,\gamma \right)=\sum _{j=1}^{10}\frac{Rank\left(PP{V}_i^j\right)}{ma{x}_{i}Score{\left(C,\gamma \right)}_i}}$

We identified four pairs of parameters C and γ in different grid regions, which resulted in approximately the same low score, but provided different expected true positive and false negative rates (see Figure 5).

Finally, we re-derived four prediction models on the whole training set using the obtained parameters C and γ. The estimated TPR and FPR for each derived prediction model are presented in Figure 5B. As one can see we chose prediction models with different TPR and FPR. The reason for that is to control the number of output predictions (note that ‘all −1’ model, that is, model that treats all point mutations as non-stabilizing, has perfect FPR, while 'all +1' model, that is, model that treats all point mutations as stabilizing has perfect TPR). For example, the first prediction model outputs only a few predictions, but with high confidence, in contrast to the fourth prediction model, which outputs more predictions, but also increases the number of false positive ones. The total score of a point mutation is, thus, weighted according to the prediction model.

Here we introduced a new comprehensive CompoMug approach for the prediction of stabilizing mutations in GPCRs and demonstrated its first application. In the 5-HT_2C case, CompoMug achieved a very high success rate of 25% for mutations with substantial, more than 1.5°C, improvement in Tm values, which is 5 to 15 times higher than corresponding hit rates found in the experimental alanine mutagenesis approach for the adenosine or endothelin type B receptors (Lebon et al., 2011; Okuta et al., 2016). The CompoMug predictions resulted in the discovery of 10 new stabilizing mutations and enabled the structure determination of the 5-HT_2C receptor in complexes with an agonist and an antagonist. Importantly, our first results suggest that each module of CompoMug is important and can additively contribute to the discovery of key stabilizing mutations. Below we discuss strengths and limitations of the individual CompoMug modules and how they can be further improved with an accumulation of structural and mutation data for the GPCR family.

Knowledge-based is the most established and straightforward approach, which directly copies some of the well-described mutations that already proved beneficial in crystallization of several GPCRs. However, while the mutations in the knowledge-based list are considered the most transferable at least within class A GPCR, they still often fail to give any substantial gain in stability as it happened in the 5-HT_2C case for the D99^2.50N and M143^3.41W mutations tested. Moreover, even in those cases when some of such known transferable mutants are beneficial, they may be not sufficient to get the optimal crystallization construct, nor to drive thermostability to the point required for structure-based drug design applications. Ongoing accumulation of mutation results for more GPCRs, e.g. discovered by other modules, may allow to add some additional mutations to the knowledge-based module. For example, X^3.41- > W mutation was initially discovered based on a sequence analysis and structure-based energy evaluations (Roth et al., 2008). At the same time, the recently discovered mutations in the functionally conserved sites (Table 1), e.g. in the P-I-F motif, DRY motif or sodium site (Katritch et al., 2014), apparently remove the key ‘gears’ from the activation mechanism, thus confining receptor to the inactive state and reducing conformational heterogeneity of the system, which can be beneficial for both thermostabilization and crystallization. In this respect, the sodium site residues provide the most opportunities, as it has several highly conserved sodium- and water-coordinating side chains that can be mutated.

The sequence-based module is fast and does not require structural knowledge. More than half of the tested point mutations (23 out of 39) came from the sequence-based module, with 6 of them yielding increased thermostability of the receptor, and 8 of them showing neutral effect. One of the advantages of this module is that mutation candidates have a lower probability to damage the receptor, because the candidate amino acid naturally occurs in other GPCRs. Moreover, in the particular case of 5-HT_2C receptor, the highest improvement in thermostability was observed for the C360^7.45N point mutation predicted by the sequence-based module. Similar idea of using deviations in the residue conservation pattern as a potential target for mutation has been used by Chen et al to stabilize a variant of the ADRB1 receptor (Chen et al., 2012), though its application was limited to the cases of unusual polar or charged residues. In this study, we used an empirical score that is applicable to all types of residues and allows to quantify the predictions (see Equation1). The score can be further optimized, for example, one can add specific weights with respect to MSAs, e.g. species variation may have a higher impact on the score compared to the common GPCR branch variation. A regression analysis required to adjust the optimal weights for the alignments, however would require further accumulation of additional stability data for GPCRs.

The structure-based module employs detailed information about residue interactions, and can potentially be highly predictive. In the 5-HT_2C test case, nine tested point mutations were selected using the structure-based module, and two of them (I374^8.49D and A98^2.49C/A140^3.38C) increased the apparent melting temperature of the receptor. Importantly, the structure-based module can be very effective in differentiating between active and inactive state stabilizing conformations, and indeed the A98^2.49C/A140^3.38C disulfide bond 5-HT_2C was shown to exclusively stabilize antagonist-bound complexes, but destabilize an agonist-bound complex (Figure 6). Previously, possibility to differentially stabilize GPCRs in agonist or antagonist-bund states was shown by experimental alanine mutagenesis study for the A_2A receptor (Magnani et al., 2008). The structure-based module can do it very effectively, however, it requires a high-quality structural model, due to the high sensitivity of disulfide bonds, ionic locks and the corresponding energy terms to the receptor conformation. An increasing structural coverage of GPCR family, including both active- and inactive-state conformations will, in turn, allow more accurate models, improving performance and applicability of the structure-based module.

The machine-learning-based (ML) module allows to identify stabilizing point mutations overlooked by the other modules, since it uses complex feature vectors trained on previously obtained experimental data on mutations, rather than pure sequence or structural information. The machine learning, however, makes the resulting point mutations to be more difficult to interpret. We used eight point mutations selected by this module, of which four point mutations also showed high scores in the sequence-based module. Three of the ML point mutations improved the thermostability of the receptor. This module critically depends not only on the structural model, but also on the training set. For the current study, we used the available alanine mutations data, thus, its prediction power for the residues other than alanine can be limited. This situation will improve with an accumulation of novel stability data, including both experimental results coming from full receptor scanning and incorporation of mutations predicted by CompoMug in more than a dozen GPCRs.

By design, the four CompoMug modules are based on different principles and use different types of input information, so they are expected to complement each other, rather than overlap. Indeed, our results suggest that most of the successful mutations for 5-HT_2C were predicted by only one of the modules, and overlap in module predictions did not correlate with improved chances for the successful mutation, at least in this test case. Moreover, some of the trends in different modules can be opposite. For example, some of the specific knowledge-based mutations of highly conserved residues (e.g. D^2.50N), can render the receptor dysfunctional, but at the same time help its stability (White et al., 2018). On the other hand, the sequence-based module is based on a premise that residues deviating from the local conservation pattern are likely to be destabilizing. There is no contradiction here, as the evolutionary selection acted differently on different GPCR sites, in some positions selecting for general stability, but in other positions preserving conformationally unstable, but functionally critical residues.

The CompoMug tool is being applied to a number of GPCR targets, showing consistently high hit rates and helping structure determination of several GPCRs, including non-class A receptors (unpublished results). In principle, structure-based and sequence-based modules of CompoMug in their current form can be also applied to other membrane proteins families. The feature set of the machine-learning-based module may also have a more general utility, however, the model would require retraining on mutation stability data for the target family, where available.

All data generated in this study are included in the manuscript and supporting files. Source data for Figure 6 are provided in Supplementary Information table. CompoMug modules are available on GitLab (https://gitlab.com/pp_lab/CompoMug.git)

1. 1. Alexandrov AI
   2. Mileni M
   3. Chien EY
   4. Hanson MA
   5. Stevens RC

   (2008) Microscale fluorescent thermal stability assay for membrane proteins

   Structure 16:351–359.

   https://doi.org/10.1016/j.str.2008.02.004

   • PubMed
   • Google Scholar
2. 1. Bhattacharya S
   2. Lee S
   3. Grisshammer R
   4. Tate CG
   5. Vaidehi N

   (2014) Rapid computational prediction of thermostabilizing mutations for G Protein-Coupled receptors

   Journal of Chemical Theory and Computation 10:5149–5160.

   https://doi.org/10.1021/ct500616v

   • PubMed
   • Google Scholar
3. 1. Chang C-C
   2. Lin C-J

   (2011) LIBSVM: a library for support vector machines

   ACM Transactions on Intelligent Systems and Technology 2:1–27.

   https://doi.org/10.1145/1961189.1961199

   • Google Scholar
4. 1. Chen KY
   2. Zhou F
   3. Fryszczyn BG
   4. Barth P

   (2012) Naturally evolved G protein-coupled receptors adopt metastable conformations

   PNAS 109:13284–13289.

   https://doi.org/10.1073/pnas.1205512109

   • PubMed
   • Google Scholar
5. 1. Cherezov V
   2. Rosenbaum DM
   3. Hanson MA
   4. Rasmussen SG
   5. Thian FS
   6. Kobilka TS
   7. Choi HJ
   8. Kuhn P
   9. Weis WI
   10. Kobilka BK
   11. Stevens RC

   (2007) High-resolution crystal structure of an engineered human beta2-adrenergic G protein-coupled receptor

   Science 318:1258–1265.

   https://doi.org/10.1126/science.1150577

   • PubMed
   • Google Scholar
6. 1. Christopher JA
   2. Brown J
   3. Doré AS
   4. Errey JC
   5. Koglin M
   6. Marshall FH
   7. Myszka DG
   8. Rich RL
   9. Tate CG
   10. Tehan B
   11. Warne T
   12. Congreve M

   (2013) Biophysical fragment screening of the β1-adrenergic receptor: identification of high affinity arylpiperazine leads using structure-based drug design

   Journal of Medicinal Chemistry 56:3446–3455.

   https://doi.org/10.1021/jm400140q

   • PubMed
   • Google Scholar
7. 1. Chun E
   2. Thompson AA
   3. Liu W
   4. Roth CB
   5. Griffith MT
   6. Katritch V
   7. Kunken J
   8. Xu F
   9. Cherezov V
   10. Hanson MA
   11. Stevens RC

   (2012) Fusion partner toolchest for the stabilization and crystallization of G protein-coupled receptors

   Structure 20:967–976.

   https://doi.org/10.1016/j.str.2012.04.010

   • PubMed
   • Google Scholar
8. 1. Congreve M
   2. Rich RL
   3. Myszka DG
   4. Figaroa F
   5. Siegal G
   6. Marshall FH

   (2011) Fragment screening of stabilized G-protein-coupled receptors using biophysical methods

   Methods in Enzymology 493:115–136.

   https://doi.org/10.1016/B978-0-12-381274-2.00005-4

   • PubMed
   • Google Scholar
9. 1. Craig DB
   2. Dombkowski AA

   (2013) Disulfide by Design 2.0: a web-based tool for disulfide engineering in proteins

   BMC Bioinformatics 14:346.

   https://doi.org/10.1186/1471-2105-14-346

   • PubMed
   • Google Scholar
10. 1. Egloff P
    2. Hillenbrand M
    3. Klenk C
    4. Batyuk A
    5. Heine P
    6. Balada S
    7. Schlinkmann KM
    8. Scott DJ
    9. Schütz M
    10. Plückthun A

    (2014) Structure of signaling-competent neurotensin receptor 1 obtained by directed evolution in Escherichia coli

    PNAS 111:E655–E662.

    https://doi.org/10.1073/pnas.1317903111

    • PubMed
    • Google Scholar
11. 1. Errey JC
    2. Doré AS
    3. Zhukov A
    4. Marshall FH
    5. Cooke RM

    (2015) Purification of stabilized GPCRs for structural and biophysical analyses

    Methods in Molecular Biology 1335:1–15.

    https://doi.org/10.1007/978-1-4939-2914-6_1

    • PubMed
    • Google Scholar
12. 1. Fenalti G
    2. Giguere PM
    3. Katritch V
    4. Huang XP
    5. Thompson AA
    6. Cherezov V
    7. Roth BL
    8. Stevens RC

    (2014) Molecular control of δ-opioid receptor signalling

    Nature 506:191–196.

    https://doi.org/10.1038/nature12944

    • PubMed
    • Google Scholar
13. 1. Heydenreich FM
    2. Vuckovic Z
    3. Matkovic M
    4. Veprintsev DB

    (2015) Stabilization of G protein-coupled receptors by point mutations

    Frontiers in Pharmacology 6:82.

    https://doi.org/10.3389/fphar.2015.00082

    • PubMed
    • Google Scholar
14. 1. Isberg V
    2. de Graaf C
    3. Bortolato A
    4. Cherezov V
    5. Katritch V
    6. Marshall FH
    7. Mordalski S
    8. Pin JP
    9. Stevens RC
    10. Vriend G
    11. Gloriam DE

    (2015) Generic GPCR residue numbers - aligning topology maps while minding the gaps

    Trends in Pharmacological Sciences 36:22–31.

    https://doi.org/10.1016/j.tips.2014.11.001

    • PubMed
    • Google Scholar
15. 1. Katoh K
    2. Frith MC

    (2012) Adding unaligned sequences into an existing alignment using MAFFT and LAST

    Bioinformatics 28:3144–3146.

    https://doi.org/10.1093/bioinformatics/bts578

    • PubMed
    • Google Scholar
16. 1. Katritch V
    2. Cherezov V
    3. Stevens RC

    (2013) Structure-function of the G protein-coupled receptor superfamily

    Annual Review of Pharmacology and Toxicology 53:531–556.

    https://doi.org/10.1146/annurev-pharmtox-032112-135923

    • PubMed
    • Google Scholar
17. 1. Katritch V
    2. Fenalti G
    3. Abola EE
    4. Roth BL
    5. Cherezov V
    6. Stevens RC

    (2014) Allosteric sodium in class A GPCR signaling

    Trends in Biochemical Sciences 39:233–244.

    https://doi.org/10.1016/j.tibs.2014.03.002

    • PubMed
    • Google Scholar
18. 1. Khan S
    2. Vihinen M

    (2010) Performance of protein stability predictors

    Human Mutation 31:675–684.

    https://doi.org/10.1002/humu.21242

    • PubMed
    • Google Scholar
19. 1. Kruse AC
    2. Hu J
    3. Pan AC
    4. Arlow DH
    5. Rosenbaum DM
    6. Rosemond E
    7. Green HF
    8. Liu T
    9. Chae PS
    10. Dror RO
    11. Shaw DE
    12. Weis WI
    13. Wess J
    14. Kobilka BK

    (2012) Structure and dynamics of the M3 muscarinic acetylcholine receptor

    Nature 482:552–556.

    https://doi.org/10.1038/nature10867

    • PubMed
    • Google Scholar
20. 1. Kumar MD
    2. Bava KA
    3. Gromiha MM
    4. Prabakaran P
    5. Kitajima K
    6. Uedaira H
    7. Sarai A

    (2006) ProTherm and ProNIT: thermodynamic databases for proteins and protein-nucleic acid interactions

    Nucleic Acids Research 34:D204–D206.

    https://doi.org/10.1093/nar/gkj103

    • PubMed
    • Google Scholar
21. 1. Kumari P
    2. Ghosh E
    3. Shukla AK

    (2015) Emerging approaches to GPCR ligand screening for drug discovery

    Trends in Molecular Medicine 21:687–701.

    https://doi.org/10.1016/j.molmed.2015.09.002

    • PubMed
    • Google Scholar
22. 1. Lagerström MC
    2. Schiöth HB

    (2008) Structural diversity of G protein-coupled receptors and significance for drug discovery

    Nature Reviews Drug Discovery 7:339–357.

    https://doi.org/10.1038/nrd2518

    • PubMed
    • Google Scholar
23. 1. Langmead CJ
    2. Andrews SP
    3. Congreve M
    4. Errey JC
    5. Hurrell E
    6. Marshall FH
    7. Mason JS
    8. Richardson CM
    9. Robertson N
    10. Zhukov A
    11. Weir M

    (2012) Identification of novel adenosine A(2A) receptor antagonists by virtual screening

    Journal of Medicinal Chemistry 55:1904–1909.

    https://doi.org/10.1021/jm201455y

    • PubMed
    • Google Scholar
24. 1. Lebon G
    2. Bennett K
    3. Jazayeri A
    4. Tate CG

    (2011) Thermostabilisation of an agonist-bound conformation of the human adenosine A(2A) receptor

    Journal of Molecular Biology 409:298–310.

    https://doi.org/10.1016/j.jmb.2011.03.075

    • PubMed
    • Google Scholar
25. 1. Liu W
    2. Chun E
    3. Thompson AA
    4. Chubukov P
    5. Xu F
    6. Katritch V
    7. Han GW
    8. Roth CB
    9. Heitman LH
    10. IJzerman AP
    11. Cherezov V
    12. Stevens RC

    (2012) Structural basis for allosteric regulation of GPCRs by sodium ions

    Science 337:232–236.

    https://doi.org/10.1126/science.1219218

    • PubMed
    • Google Scholar
26. 1. Magnani F
    2. Shibata Y
    3. Serrano-Vega MJ
    4. Tate CG

    (2008) Co-evolving stability and conformational homogeneity of the human adenosine A2a receptor

    PNAS 105:10744–10749.

    https://doi.org/10.1073/pnas.0804396105

    • PubMed
    • Google Scholar
27. 1. Massink A
    2. Gutiérrez-de-Terán H
    3. Lenselink EB
    4. Ortiz Zacarías NV
    5. Xia L
    6. Heitman LH
    7. Katritch V
    8. Stevens RC
    9. IJzerman AP

    (2015) Sodium ion binding pocket mutations and adenosine A2A receptor function

    Molecular Pharmacology 87:305–313.

    https://doi.org/10.1124/mol.114.095737

    • PubMed
    • Google Scholar
28. 1. Moukhametzianov R
    2. Warne T
    3. Edwards PC
    4. Serrano-Vega MJ
    5. Leslie AG
    6. Tate CG
    7. Schertler GF

    (2011) Two distinct conformations of helix 6 observed in antagonist-bound structures of a beta1-adrenergic receptor

    PNAS 108:8228–8232.

    https://doi.org/10.1073/pnas.1100185108

    • PubMed
    • Google Scholar
29. 1. Munk C
    2. Harpsøe K
    3. Hauser AS
    4. Isberg V
    5. Gloriam DE

    (2016) Integrating structural and mutagenesis data to elucidate GPCR ligand binding

    Current Opinion in Pharmacology 30:51–58.

    https://doi.org/10.1016/j.coph.2016.07.003

    • PubMed
    • Google Scholar
30. 1. Okuta A
    2. Tani K
    3. Nishimura S
    4. Fujiyoshi Y
    5. Doi T

    (2016) Thermostabilization of the human endothelin type B receptor

    Journal of Molecular Biology 428:2265–2274.

    https://doi.org/10.1016/j.jmb.2016.03.024

    • PubMed
    • Google Scholar
31. 1. Pándy-Szekeres G
    2. Munk C
    3. Tsonkov TM
    4. Mordalski S
    5. Harpsøe K
    6. Hauser AS
    7. Bojarski AJ
    8. Gloriam DE

    (2018) GPCRdb in 2018: adding GPCR structure models and ligands

    Nucleic Acids Research 46:D440–D446.

    https://doi.org/10.1093/nar/gkx1109

    • PubMed
    • Google Scholar
32. 1. Peng Y
    2. McCorvy JD
    3. Harpsøe K
    4. Lansu K
    5. Yuan S
    6. Popov P
    7. Qu L
    8. Pu M
    9. Che T
    10. Nikolajsen LF
    11. Huang XP
    12. Wu Y
    13. Shen L
    14. Bjørn-Yoshimoto WE
    15. Ding K
    16. Wacker D
    17. Han GW
    18. Cheng J
    19. Katritch V
    20. Jensen AA
    21. Hanson MA
    22. Zhao S
    23. Gloriam DE
    24. Roth BL
    25. Stevens RC
    26. Liu ZJ

    (2018) 5-HT_2Creceptor structures reveal the structural basis of GPCR polypharmacology

    Cell 172:719–730.

    https://doi.org/10.1016/j.cell.2018.01.001

    • PubMed
    • Google Scholar
33. 1. Rask-Andersen M
    2. Masuram S
    3. Schiöth HB

    (2014) The druggable genome: Evaluation of drug targets in clinical trials suggests major shifts in molecular class and indication

    Annual Review of Pharmacology and Toxicology 54:9–26.

    https://doi.org/10.1146/annurev-pharmtox-011613-135943

    • PubMed
    • Google Scholar
34. 1. Rich RL
    2. Miles AR
    3. Gale BK
    4. Myszka DG

    (2009) Detergent screening of a G-protein-coupled receptor using serial and array biosensor technologies

    Analytical Biochemistry 386:98–104.

    https://doi.org/10.1016/j.ab.2008.12.011

    • PubMed
    • Google Scholar
35. 1. Rios S
    2. Fernandez MF
    3. Caltabiano G
    4. Campillo M
    5. Pardo L
    6. Gonzalez A

    (2015) GPCRtm: An amino acid substitution matrix for the transmembrane region of class A G Protein-Coupled Receptors

    BMC Bioinformatics 16:206.

    https://doi.org/10.1186/s12859-015-0639-4

    • PubMed
    • Google Scholar
36. 1. Roth CB
    2. Hanson MA
    3. Stevens RC

    (2008) Stabilization of the human beta2-adrenergic receptor TM4-TM3-TM5 helix interface by mutagenesis of Glu122(3.41), a critical residue in GPCR structure

    Journal of Molecular Biology 376:1305–1319.

    https://doi.org/10.1016/j.jmb.2007.12.028

    • PubMed
    • Google Scholar
37. 1. Schlinkmann KM
    2. Honegger A
    3. Türeci E
    4. Robison KE
    5. Lipovšek D
    6. Plückthun A

    (2012) Critical features for biosynthesis, stability, and functionality of a G protein-coupled receptor uncovered by all-versus-all mutations

    PNAS 109:9810–9815.

    https://doi.org/10.1073/pnas.1202107109

    • PubMed
    • Google Scholar
38. 1. Serrano-Vega MJ
    2. Magnani F
    3. Shibata Y
    4. Tate CG

    (2008) Conformational thermostabilization of the beta1-adrenergic receptor in a detergent-resistant form

    PNAS 105:877–882.

    https://doi.org/10.1073/pnas.0711253105

    • PubMed
    • Google Scholar
39. 1. Serrano-Vega MJ
    2. Tate CG

    (2009) Transferability of thermostabilizing mutations between beta-adrenergic receptors

    Molecular Membrane Biology 26:385–396.

    https://doi.org/10.3109/09687680903208239

    • PubMed
    • Google Scholar
40. 1. Shibata Y
    2. White JF
    3. Serrano-Vega MJ
    4. Magnani F
    5. Aloia AL
    6. Grisshammer R
    7. Tate CG

    (2009) Thermostabilization of the neurotensin receptor NTS1

    Journal of Molecular Biology 390:262–277.

    https://doi.org/10.1016/j.jmb.2009.04.068

    • PubMed
    • Google Scholar
41. 1. Stevens RC
    2. Cherezov V
    3. Katritch V
    4. Abagyan R
    5. Kuhn P
    6. Rosen H
    7. Wüthrich K

    (2013) The GPCR Network: a large-scale collaboration to determine human GPCR structure and function

    Nature Reviews Drug Discovery 12:25–34.

    https://doi.org/10.1038/nrd3859

    • PubMed
    • Google Scholar
42. 1. Vaidehi N
    2. Grisshammer R
    3. Tate CG

    (2016) How can mutations thermostabilize G-Protein-Coupled receptors?

    Trends in Pharmacological Sciences 37:37–46.

    https://doi.org/10.1016/j.tips.2015.09.005

    • PubMed
    • Google Scholar
43. 1. Wacker D
    2. Wang C
    3. Katritch V
    4. Han GW
    5. Huang XP
    6. Vardy E
    7. McCorvy JD
    8. Jiang Y
    9. Chu M
    10. Siu FY
    11. Liu W
    12. Xu HE
    13. Cherezov V
    14. Roth BL
    15. Stevens RC

    (2013) Structural features for functional selectivity at serotonin receptors

    Science 340:615–619.

    https://doi.org/10.1126/science.1232808

    • PubMed
    • Google Scholar
44. 1. Warne T
    2. Serrano-Vega MJ
    3. Baker JG
    4. Moukhametzianov R
    5. Edwards PC
    6. Henderson R
    7. Leslie AG
    8. Tate CG
    9. Schertler GF

    (2008) Structure of a beta1-adrenergic G-protein-coupled receptor

    Nature 454:486–491.

    https://doi.org/10.1038/nature07101

    • PubMed
    • Google Scholar
45. 1. Warne T
    2. Serrano-Vega MJ
    3. Tate CG
    4. Schertler GF

    (2009) Development and crystallization of a minimal thermostabilised G protein-coupled receptor

    Protein Expression and Purification 65:204–213.

    https://doi.org/10.1016/j.pep.2009.01.014

    • PubMed
    • Google Scholar
46. 1. Warne T
    2. Moukhametzianov R
    3. Baker JG
    4. Nehmé R
    5. Edwards PC
    6. Leslie AG
    7. Schertler GF
    8. Tate CG

    (2011) The structural basis for agonist and partial agonist action on a β(1)-adrenergic receptor

    Nature 469:241–244.

    https://doi.org/10.1038/nature09746

    • PubMed
    • Google Scholar
47. 1. White KL
    2. Eddy MT
    3. Gao ZG
    4. Han GW
    5. Lian T
    6. Deary A
    7. Patel N
    8. Jacobson KA
    9. Katritch V
    10. Stevens RC

    (2018) Structural Connection between Activation Microswitch and Allosteric Sodium Site in GPCR Signaling

    Structure 26:259–269.

    https://doi.org/10.1016/j.str.2017.12.013

    • PubMed
    • Google Scholar
48. 1. Zhang X
    2. Stevens RC
    3. Xu F

    (2015) The importance of ligands for G protein-coupled receptor stability

    Trends in Biochemical Sciences 40:79–87.

    https://doi.org/10.1016/j.tibs.2014.12.005

    • PubMed
    • Google Scholar

Author details

1. Petr Popov

   1. Department of Biological Sciences, University of Southern California, Los Angeles, Los Angeles, United States
   2. Moscow Institute of Physics and Technology, Dolgoprudny, Russia

   Contribution

   Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft, Writing—review and editing

   Competing interests

   Filed a U.S. provisional patent application (serial no. 62/644,008) for CompoMug algorithms.

2. Yao Peng

   iHuman Institute, ShanghaiTech University, Shanghai, China

   Contribution

   Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—review and editing

   Competing interests

   No competing interests declared

3. Ling Shen

   1. iHuman Institute, ShanghaiTech University, Shanghai, China
   2. School of Life Science and Technology, ShanghaiTech University, Shanghai, China

   Contribution

   Validation, Investigation, Visualization, Writing—review and editing

   Competing interests

   No competing interests declared

4. Raymond C Stevens

   1. Department of Biological Sciences, University of Southern California, Los Angeles, Los Angeles, United States
   2. iHuman Institute, ShanghaiTech University, Shanghai, China
   3. Department of Chemistry, University of Southern California, Los Angeles, Los Angeles, United States
   4. Bridge Institute, University of Southern California, Los Angeles, Los Angeles, United States

   Contribution

   Conceptualization, Resources, Funding acquisition, Writing—review and editing

   Competing interests

   No competing interests declared

5. Vadim Cherezov

   1. Department of Biological Sciences, University of Southern California, Los Angeles, Los Angeles, United States
   2. Moscow Institute of Physics and Technology, Dolgoprudny, Russia
   3. Department of Chemistry, University of Southern California, Los Angeles, Los Angeles, United States
   4. Bridge Institute, University of Southern California, Los Angeles, Los Angeles, United States

   Contribution

   Conceptualization, Resources, Funding acquisition, Validation, Project administration, Writing—review and editing

   Competing interests

   No competing interests declared

6. Zhi-Jie Liu

   1. iHuman Institute, ShanghaiTech University, Shanghai, China
   2. School of Life Science and Technology, ShanghaiTech University, Shanghai, China
   3. Insititute of Molecular and Clinical Medicine, Kunming Medical University, Kunming, China

   Contribution

   Resources, Funding acquisition, Validation, Project administration, Writing—review and editing

   Competing interests

   No competing interests declared

7. Vsevolod Katritch

   1. Department of Biological Sciences, University of Southern California, Los Angeles, Los Angeles, United States
   2. Moscow Institute of Physics and Technology, Dolgoprudny, Russia
   3. Department of Chemistry, University of Southern California, Los Angeles, Los Angeles, United States
   4. Bridge Institute, University of Southern California, Los Angeles, Los Angeles, United States

   Contribution

   Conceptualization, Resources, Software, Supervision, Funding acquisition, Validation, Investigation, Methodology, Writing—original draft, Project administration, Writing—review and editing

   For correspondence

   katritch@usc.edu

   Competing interests

   Vsevolod Katritch: Filed a U.S. provisional patent application (serial no. 62/644,008) for CompoMug algorithms.

   "This ORCID iD identifies the author of this article:" 0000-0003-3883-4505

• 5,153

  views

• 735

  downloads

• 63

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.