We developed a system to build, stably express, and assay individual variants of the β₂AR in human cell lines. The β₂AR primarily signals through the heterotrimeric G_s protein, activating adenylyl cyclase upon agonist binding. In our platform, cAMP production stimulates transcription of a barcoded reporter gene, controlled by multimerized cAMP response elements (CRE, thus referred to as the CRE reporter for the rest of the manuscript), which can be quantified by RNA-seq (Figure 1A). Initially, we generated a HEK293T-derived cell line for stable integration of the GPCR-reporter construct (Figure 1B, Figure 1—figure supplement 1A,B). We also modified a previously developed Bxb1-landing pad system to allow for stable, once-only integration at the transcriptionally-silent H11 safe-harbor locus to avoid placing the CRE reporter within transcribed genes (Cheung et al., 2019; Duportet et al., 2014; Matreyek et al., 2017). To prevent endogenous signaling, we knocked out the gene encoding for β₂AR, ADRB2, and verified loss of CRE reporter gene activity in response to the β₂AR agonist, isoproterenol (Figure 1—figure supplement 1C). Our donor vector configuration ensures the receptor and resistance marker are only activated upon successful integration into the landing pad (Figure 1B). Lastly, we included several sequence elements in the donor vector to improve signal-to-noise of the assay: an insulator upstream of the CRE reporter and an N-terminal affinity tag to the receptor (Figure 1—figure supplement 1D,E). As a result, upon integration of a donor vector expressing wild-type (WT) β₂AR, isoproterenol induces CRE reporter gene expression in a dose-dependent manner (Figure 1B).

We designed and synthesized the receptor’s 7828 possible missense variants in eight segments on oligonucleotide microarrays (Figure 1C). We amplified the mutant oligos, attaching a random 15 nucleotide barcode sequence, and cloned them into one of eight background vectors encoding the upstream, wild-type portion of the gene. In this configuration, we mapped barcode-variant pairs with next-generation sequencing and subsequently utilized Type IIS restriction enzymes to insert the remaining sequence elements between the receptor and barcode. In the resulting mature donor vector, the barcode is located in the 3’ untranslated region (UTR) of the CRE reporter gene. We integrated the library into our engineered cell line, and developed protocols to ensure proper quantification of library members, most notably vastly increasing the numbers of cells we assayed and RNA processed for the RNA-seq (Figure 1—figure supplement 1F, Figure 2—figure supplement 1A).

We screened the mutant library at four concentrations of the β₂AR full-agonist isoproterenol: vehicle control, an empirically determined half-maximal activity (EC₅₀, 150 nM), full activity (EC₁₀₀, 625 nM), and beyond saturation of the WT receptor (E_max, 5 µM). We obtained reliable measurements (coefficient of variation <1) for 95–99% (7,461–7,749/7,828 depending on the agonist concentration) of possible missense variants (412 residues * 19 amino acids = 7828 possible missense variants) with two biological replicates at each condition (Figure 1C). We normalized these measurements against forskolin treatment, which induces cAMP signaling independent of the β₂AR (Insel and Ostrom, 2003). Forskolin treatment maximally induces the CRE reporter gene, therefore the relative barcode expression is proportional to the physical composition of the library. Each cell contains a single copy of the same CRE reporter sequence, therefore any differences in maximum transcriptional output between barcodes will be due to differences in the frequency of each barcode within the cell library. Finally, we define activity as the ratio of this value to the mean frameshift (Materials and methods). Each variant was represented by 10 barcodes (median), with biological replicates displaying Pearson’s correlations of 0.87 to 0.90 at the barcode level and 0.66 to 0.75 when summarized by individual variants (Figure 1—figure supplement 1G,H, Figure 2—figure supplement 1A). Of note, we aimed for 10 barcodes per variant in order to account for any effects individual barcodes will have on CRE reporter transcription and serve as statistical replicates for each variant.

The heatmap representation of the variant-activity landscape reveals global and regional trends in response to specific subtypes of mutations (Figure 2A). For example, the transmembrane domain and intracellular helix eight are more sensitive to substitution than the termini or loops, and this effect becomes more pronounced at higher agonist concentrations (Figure 2A; all p<0.001; Mann-Whitney U). The transmembrane domain and intracellular helix eight are also sensitive to helix-disrupting proline substitutions (Figure 2B, Figure 2—figure supplement 1B; all p<<0.001 except TM vs Helix-8; Mann-Whitney U). Microarray-derived DNA often contains single-base deletions (47% of oligos in our library) that will introduce frameshift mutations into our library (LeProust et al., 2010). As expected, frameshifts consistently display lower activity than missense mutations regardless of agonist concentration (Figure 2C; p<<0.001; Mann-Whitney U). Furthermore, the effect of frameshifts are markedly decreased in the C-terminus of the protein (Figure 2D; p<<0.001; Mann-Whitney U). We also built and integrated previously characterized mutants (Elling et al., 1999; Sato et al., 1999; Shenoy et al., 2006) into our system individually and measured activity with a luciferase CRE reporter gene at the same induction conditions (Figure 2E and Figure 2—figure supplement 1C). As expected, known null mutations (D113A and I135W) have significantly diminished activity relative to WT in both systems, even at E_max (all p<<0.001; Wald Test). Known hypomorphic mutations (S203A and S204A) also have a significant decrease in activity relative to WT at EC₁₀₀ (all p<<0.001; Wald Test), but are not significantly different than WT at E_max as expected (all p>0.01; Wald Test).

Metrics for sequence conservation and covariation are often used to predict the effects a mutation will have on protein function (Adzhubei et al., 2013; Capra and Singh, 2007; Hopf et al., 2017). Mutational tolerance, the mean activity of all amino acid substitutions per residue at each agonist concentration, is highly correlated to conservation, both across species for the β₂AR (Figure 3—figure supplement 1A; Spearman's ρ = −0.74; 55 orthologs, predominantly mammals but including a few other vertebrates as well as a small number of invertebrate beta-like sequences, identified from the OMA Database, Supplementary file 1), and across all Class A GPCRs (Spearman's ρ = −0.68; Figure 3A and Figure 3—figure supplement 1B; Altenhoff et al., 2018; Capra and Singh, 2007; Hopf et al., 2017) at EC₁₀₀. From this point on, any use of the words tolerance or intolerance in this manuscript refer to mutational tolerance. Correlation between our data and both predictors increases with agonist concentration up to EC₁₀₀ (Figure 3—figure supplement 1A,B). We found a subset of residues in extracellular loop 2 (ECL2), including C184 and C190 that form an intraloop disulfide bridge, that were more intolerant to mutation than expected given their conservation across Class A GPCRs. This suggests a fairly specific functional role for this motif in the β₂AR (Figure 3A). On an individual variant level, mutational responses correlate (Spearman's ρ = 0.520) with EVmutation, a predictor of mutational effects from sequence covariation (Figure 3B and Figure 3—figure supplement 1C; Altenhoff et al., 2018; Capra and Singh, 2007; Hopf et al., 2017).

In addition to evolutionary metrics, understanding the functional distribution of ADRB2 variants found within the human population is important given the extensive variation found among GPCR drug targets (Hauser et al., 2018). The Genome Aggregation Database (gnomAD) reports variants found across 141,456 individuals (Karczewski et al., 2019), and many of the 180 ADRB2 missense variants are of unknown significance. We classified 11 of these variants as potentially loss of function, by comparing their activity to the distribution of frameshift mutations found in our assay (Figure 3C; see Materials and methods). Given that measurements of individual mutations are noisy (average coefficient of variation = 0.55), this analysis is best suited as a funnel to guide further characterization (see Discussion).

However, our analysis is more robust when we aggregate the signal of multiple mutations at a given position. Therefore, we compiled a list of the 100 most activating mutations at vehicle control and the 100 least active mutations at EC₁₀₀ and mapped their location on the β₂AR structure. As expected, the least active mutations tended to reside within the core of the transmembrane domain (Figure 3—figure supplement 1D,E). Alternatively, the most activating mutations mapped to TM1, TM5, TM6, and residues that typically face away from the internal core of the receptor (Figure 3D,E). Of note, a group of these mutations in TM5 face TM6, which undergoes a large structural rearrangement during receptor activation (Weis and Kobilka, 2018). Activating mutants are also enriched in the termini, ICL3, and Helix 8. Concentration at the termini is unsurprising, as these regions have known involvement in surface expression and our current assay does not discriminate between increased signaling potency and expression (see discussion; Dong et al., 2007). However, there are cases of constitutively active mutations in the N terminus that increase signaling potency without affecting surface expression, such as T11S of the melanocortin 4 (MC4R) (Lotta et al., 2019). Similarly, the enrichment of activating mutants in ICL3 appears to reflect its role in G-protein binding (Ozcan et al., 2013; Ozgur et al., 2016; West et al., 2011). Lastly, we observe a number of activating mutations in the terminal residue, L413. A recent study of genetic variation in human MC4R also found a gain-of-function mutation at the terminal residue of the receptor, suggesting a possible conserved role for this position in regulating basal activity of GPCRs (Lotta et al., 2019).

Given that our data spans thousands of mutations across several treatment conditions, we used unsupervised learning methods to reveal hidden regularities within groups of residues’ response to mutation. In particular, we applied Uniform Manifold Approximation and Projection (UMAP) (McInnes and Healy, 2018) to learn multiple different lower dimensional representations of our data and clustered the output with density-based hierarchical clustering (HDBSCAN; Figure 4—figure supplement 1; Campello et al., 2013). We found residues consistently separated into six clusters that exhibit distinct responses to mutation (Figure 4A,B). Clusters 1 and 2 are globally intolerant to all substitutions, whereas Cluster 3 is vulnerable to proline and charged substitutions. Cluster 4 is particularly inhibited by negatively charged substitutions and Cluster five by proline substitutions, while Cluster 6 is unaffected by any mutation. Mapping these clusters onto a 2D snake plot representation shows Clusters 1–5 primarily comprise the transmembrane domain, while Cluster 6 resides in the loops and termini (Figure 4C). These flexible regions are often truncated before crystal structure determination to minimize conformational variability (Rosenbaum et al., 2007). Surprisingly, a number of residues from Cluster five also map there, suggesting potential structured regions. However, Cluster 5 assignment is largely based on the response of a single proline mutation, and thus is more susceptible to noise than the other clusters (see Discussion).

Next, we projected the clusters onto the hydroxybenzyl isoproterenol-bound structure (Figure 5—figure supplement 1A; PDB: 4LDL). The globally intolerant Clusters 1 and 2 segregate to the core of the protein, while the charge-sensitive Cluster 3 is enriched in the lipid-facing portion (Figure 5—figure supplement 1B). This suggests that differential patterns of response to hydrophobic and charged substitutions could correlate with side chain orientation within the transmembrane domain. Indeed, residues that are uniquely charge sensitive are significantly more lipid-facing than those that are sensitive to both hydrophobic and charged mutations at EC₁₀₀ (Figure 5A, Figure 5—figure supplement 1C–D, p=0.000036; Mann-Whitney U) (Mitternacht, 2016).

Decades of research have revealed how ligand binding is coupled to G-protein activation through a series of conserved motifs (Weis and Kobilka, 2018). This comprehensive, unbiased screen enables us to systematically evaluate and rank the functional importance of every implicated residue. The globally intolerant UMAP clusters (1 and 2) highlight many residues from these motifs and suggest novel residues for investigation (Figure 5B). We can further resolve the significance of individual residues within these motifs by ranking the mutational tolerance of positions in these clusters at EC₁₀₀ (Figure 5C). In fact, 11 of the 15 most mutationally intolerant positions belong to the PIF, CWxP, and NPxxY motif, orthosteric site, a water-mediated bond network, an extracellular disulfide bond, and a cholesterol-binding site. Interestingly, the second most intolerant residue is the uncharacterized G315^7x41 (GPCRdb numbering in superscript Isberg et al., 2016). In the active state, G315’s alpha carbon points directly at W286^6x48 of the CWxP motif, the fourth most intolerant residue, and any substitution at G315^7.x41 will likely clash with W286^6x48 (Figure 5D). We confirmed G315’s intolerance with a luciferase CRE reporter gene assay, where mutants G315T and G315L resulted in complete loss of function (Figure 5—figure supplement 2A).

Recent simulations suggest water-mediated hydrogen bond networks play a critical role in GPCR function (Venkatakrishnan et al., 2018; Venkatakrishnan et al., 2019). The third most intolerant residue in our assay, Y326^7x53 of the NPxxY motif, is especially important as it switches between two of these networks during receptor activation. In the inactive state, Y326^7x53 contacts N51^1x50 and D79^2x50, two of the top 15 most intolerant positions (Figure 5E). N51L and N51Y also result in complete loss of function when assayed individually (Figure 5—figure supplement 2A). The movement of Y326^7x53 is also part of a broader rearrangement of residue contacts that are conserved across Class A GPCRs, with the majority of these residues being intolerant to mutation (Figure 5—figure supplement 2B; Venkatakrishnan et al., 2016). Aside from G315^7x41, the other uncharacterized residues in the top 15 include W99^23x50, S319^7x46, and G83^2x54. Given the correlation between mutational tolerance and functional relevance, further investigation of these residues will likely reveal insights into GPCR biology.

Next, we hypothesized residues in the orthosteric site that directly contact isoproterenol would respond uniquely to mutation; however, no crystal structure of β₂AR bound to isoproterenol exists. Using the crystal structure of the β₂AR bound to the analog, hydroxybenzyl isoproterenol (PDB: 4LDL), we find that residues responsible for binding the derivatized hydroxybenzyl tail have significantly higher mutational tolerance than residues that contact the catecholamine head common to both isoproterenol and hydroxybenzyl isoproterenol at EC₁₀₀ (p=0.0162; Figure 5F, Figure 5—figure supplement 2C). Given this discrimination, we believe DMS can be a powerful tool for mapping functional ligand-receptor contacts in GPCRs.

GPCR signaling is dependent on a series of intermolecular interactions, and the numerous β₂AR crystal structures enable us to comprehensively evaluate residues mediating such interactions. For example, cholesterol is an important modulator of GPCR function (Thal et al., 2018), and the timolol-bound inactive-state β₂AR structure elucidated the location of a cholesterol-binding site (PDB: 3D4S) (Hanson et al., 2008). Of residues in this pocket, W158^4x50 is predicted to be most important for cholesterol binding, and in agreement, W158^4x50 is the most mutationally intolerant (Figure 5—figure supplement 2D). Similarly, a number of studies have mutagenized residues at the G_ɑs-β₂AR interface (Jensen et al., 2001; Moro et al., 1993; O'Dowd et al., 1988; Rasmussen et al., 2011; Sheikh et al., 1999; Swaminath et al., 2003; Valentin-Hansen et al., 2012; Valiquette et al., 1995), but a complete understanding of the relative contribution of each residue to maintaining the interface is unknown. Most residues are more mutationally tolerant than residues in the intolerant Clusters 1 and 2, but the four most intolerant positions are I135^3x54, V222^5x61, A271^6x33, and Q229^5x68 respectively (Figure 5—figure supplement 2E). Q229^5x68 appears to coordinate polar interactions between D381 and R385 of the α5 helix of G_ɑs, whereas V222^5x61 and I135^3x54 form a hydrophobic pocket on the receptor surface (Figure 5—figure supplement 2F).

Analysis of the mutational tolerance data has highlighted the functional importance of previously uncharacterized residues. In particular, W99^23x50 of extracellular loop 1 (ECL1) is the 13^th most intolerant residue, which is unusual as mutationally intolerant residues are rare in the flexible loops. Furthermore, W99^23x50 is proximal to the disulfide bond C106^3x25-C191^45x50, an important motif for stabilization of the receptor’s active state (Noda et al., 1994; Dohlman et al., 1990; Hulme, 2013; Dohlman et al., 1990; Noda et al., 1994). While aromatic residues are known to facilitate disulfide bond formation, only tryptophan is tolerated at this position (Bhattacharyya et al., 2004). We hypothesize W99’s indole group hydrogen bonds with the backbone carbonyl of the neighboring uncharacterized and mutationally intolerant G102^3x21, positioning W99^23x50 toward the disulfide bond. Other aromatic residues are unable to form this hydrogen bond and are less likely to be positioned properly. G102^3x21 also hydrogen bonds with the backbone amide of C106^3x25, further stabilizing this region. To verify this claim, we individually confirmed the mutational intolerance of both W99^23x50 and G102^3x21 (Figure 6—figure supplement 1A). Additionally, we evaluated surface expression for a subset of W99^23×50 and G102^3×21 mutants (Figure 6—figure supplement 1B). Relative to three previously characterized mutants with severely impaired surface expression (Parmar et al., 2017) and wild-type β2AR, the mutants exhibited mildly impaired to normal surface expression—supporting a role in signaling for these residues.

Interestingly, W99^23x50, G102^3x21, and C106^3x25 are almost universally conserved across Class A GPCRs (Vass et al., 2018; Figure 6A, Figure 6—figure supplement 1C). Comparison of over 25 high-resolution structures of class A GPCRs from five functionally different sub-families and six different species revealed that these residues consistently contact each other (Figure 6B,C). Based on the evolutionary and structural conservation across Class A GPCRs, we find W99^23x50, G102^3x21, and the C106^3x25-C191^45x50 disulfide bond represent a conserved WxxGxxxC motif, forming an extracellular ‘structural latch’ that is maintained consistently throughout GPCRs spanning diverse molecular functions and phylogenetic origins. While a minority of Class A GPCRs lack the Trp/Gly combination of residues in the ECL1 region, these receptors have varying structures in ECL1: an alpha helix (sphingosine S1P receptor), beta strand (adenosine receptor), or even intrinsically disordered (viral chemokine receptor US28) (Figure 6—figure supplement 1D).

To better understand the dynamics of the structural latch, we compared the active and inactive state crystal structures of four representative GPCRs. While the overall RMSD between the inactive and active states for the β2AR, M2 muscarinic receptor, and μ opioid receptorreceptor are 1 Å,1.5 Å, and 1.7 Årespectively, the conformation of the latch in the active and inactive states is nearly identical in each receptor (Figure 6D). This suggests that the extracellular structural latch is part of a larger rigid plug present at the interface of the transmembrane and extracellular regions, which could be important for the structural integrity of the receptor and possibly guide ligand entry.

In Class A receptors lacking components of the WxxGxxxC motif, introducing the Trp-Gly interaction could increase the stability of the receptor for structural studies. In fact, in the BLT1 receptor structure, a Gly mutation at 3×21 was found to be thermostabilizing (Hori et al., 2018). Other candidate receptors lacking a Gly at 3×21 include the alpha2B receptor and the neuropeptide FF2 receptor, where the R81G and D112G mutations have potential to increase receptor stability, respectively. More broadly, these ECL1/TM3 positions conserved across Class A GPCRs could serve as candidate sites for introducing thermostabilizing mutations.

We used HEK293T cells purchased from ATCC that were genetically modified in house. The identity of the lines have been verified with STR profiling and have tested negative for mycoplasma contamination.

We incubated our cellular library with forskolin to activate the cAMP reporter in each cell, providing an agonist-independent measurement of maximal reporter activity. This measurement can be used to approximate cellular copy number. To ensure that barcodes with low cellular representation are excluded from our analyses, we require all barcodes to be present in both forskolin repeats, and filter out any barcodes with a mean reads per million less than 0.2 (~8–10 reads at our sequencing depth). We also excluded barcodes with high forskolin counts (>=10 RPM) as they are systematically less induced in the drug conditions relative to other barcodes. Next, we require that all of the barcodes in the forskolin condition are also present in our drug conditions, and set any missing barcodes to 0. We then add a pseudocount that is scaled relative to the condition with the fewest number of reads (N/min(N)), and normalize each condition to its read depth (including added pseudocounts) (Bloom, 2015). Finally, we divide this value by its associated forskolin value to control for variation in cellular abundance.

Since each variant in our library was associated with a median of 10 barcodes, we took the average of all barcodes. We then defined activity as the ratio of these values to the value of the mean frameshift. Finally, we averaged the relative activities of our two repeats together and used propagation of uncertainty to combine their standard deviations.

To calculate sequence conservation at a species level, we aligned 55 ADRB2 orthologs from the OMA database (entry: HUMAN24043) using MAFFT with the default settings (mafft --reorder --auto). For Class-A GPCRs, we retrieved the multiple sequence alignment from GPCRdb. We then used the Jensen-Shannon Divergence (Altenhoff et al., 2018; Capra and Singh, 2007; Hopf et al., 2017) to score both these alignments. We only considered conservation scores at positions in the MSA that contained residues from the β2AR. For both EVMutation and gnomAD, we simply downloaded the results for ADRB2. We considered residues in gnomAD to be potentially loss of function if their mean activity plus the standard error of the mean was less than two standard deviations from the mean of the frameshift distribution.

We performed a number of preprocessing steps before running UMAP on our data. First, we grouped amino acids into eight different classes based on their physicochemical properties ((+) - R, H, K; (-) - D, E; Aromatic - F, W, Y; Amide - N, Q; Nucleophilic - C, S, T; Hydrophobic - I, L, V, M; Small - G, A; Proline - P) and averaged their relative activities. Next, we standardized the log2 relative activity values of each group and used mean imputation to model missing data for any missing AA groups at a given position. Finally, we combined the data from every drug condition into a 412 × 32 design matrix in which the columns are an AA group at a specific condition and the rows are the positions in the protein.

With our data processed, we used the R implementation of UMAP to run hyperparameter search (https://github.com/jlmelville/uwot; copy archived at swh:1:rev:b449908402ba0ab5348c22cd2620efe23de01012; Melville, 2019) of all combinations of UMAP embeddings with the parameters n_neighbors = (4, 8, 16, 32) and n_components = (2, 3, 4, 5, 6, 7, 8, 9, 10), holding min_dist = 0 and n_epochs = 2000 constant. This provided a variety of different representations of our data that we used HDBSCAN (Campello et al., 2013) to search for clusters in these embeddings (R package dbscan; minPts = 10). To ease interpretation of the clustering, we plotted the HDBSCAN results onto a 2D UMAP embedding with the following parameters: n_neighbors = 4, min_dist = 0, n_components = 2, n_epochs = 2000, and random_state = 3308004 using the Python implementation (McInnes and Healy, 2018) (https://github.com/lmcinnes/umap; copy archived at swh:1:rev:2b9a2521b4c6d5f084278b2e967040e3020dac9d). We found the cluster assignments to be largely robust across the different embeddings, and used them to guide our manual cluster assignment.

Molecular graphics and analyses were performed with UCSF Chimera (Pettersen et al., 2004) and PyMol. To determine if a given position in the β₂AR points into the core of the protein or into the lipid membrane, we used FreeSASA (Mitternacht, 2016) (version 2.0.3) to calculate the Solvent Accessible Surface Area (SASA) of the G_s-bound β₂AR (PDB: 3SN6). The G_s occludes the intracellular surface of the β₂AR thereby reducing the SASA of residues on the intracellular surface. Similarly, the extracellular surface is mostly blocked by the extracellular loops. Finally, we used the Orientations of Proteins in Membranes (OPM) database (Lomize et al., 2012) to filter out any residues outside of the lipid membrane from our analyses. To quantify charge sensitivity, we calculated the average activity for H, K, R, D, and E substitutions at each agonist concentration for residues in the lipid membrane. We then multiplied the values by −1 and standardized the results within each agonist concentration group such that the values were mean-centered and scaled by their standard deviation. We calculated hydrophobic sensitivity (I, L, V, M) in an analogous manner. Next, we classified residues that had above average charge sensitivity and below average hydrophobic sensitivity as being exclusively charge sensitive. Conversely, we classified residues that had above average charge sensitivity and above average hydrophobic sensitivity as being intolerant.

For the structural analysis, the crystal structures of class A GPCRs were obtained from PDB (Berman et al., 2000). In order to compare structures across the different sub-families of class A GPCRs, structures of representative examples from each subfamily were chosen. In order to compare structures across the conformational states in different GPCRs, the structures of beta-2 adrenergic receptor, M2 muscarinic receptor, kappa-opioid receptor, and mu-opioid receptor were chosen. A2A receptor and mammalian rhodopsin were excluded as they lacked a Trp residue in the first extracellular loop. These receptors were chosen due to the availability of pairs of inactive state and active state structures. The inactive state structures were identified based on the presence of co-crystallized antagonist/inverse-agonist and the active state structures were identified based on the presence of a co-crystallized agonist and co-complexed interacting partner at the G-protein-coupling site. Structural alignment and measurement of inter-atomic distances were performed using PyMOL (https://pymol.org). The structure alignment was performed over the sequence stretch between the conserved Trp/Phe residue in the first extracellular loop and the canonical disulfide bridge forming Cys3 × 25 (GPCRdb number) present on TM3.

For the sequence analysis, the sequence alignment of Class A GPCRs was obtained from GPCRdb (Pándy-Szekeres et al., 2018). The alignment was filtered for receptors that contained the canonical disulfide bridge forming on TM3 residue in ECL1, which gave a total of 202 GPCRs sequences. Using this alignment, the sequence logo was made using the Weblogo program (https://weblogo.berkeley.edu/logo.cgi).

All statistical tests unless otherwise noted are the two-sided Mann-Whitney U test and were performed in R (version 3.5.x) using the wilcox.test function.

All codes are available at (https://www.github.com/KosuriLab/b2-dms; Jones, 2020 copy archived at swh:1:dir:09b4931491e1c9f9ee2c90c5687f44efa6464373). Sequencing data can be accessed from the sequencing read archive (SRA) with the accession number SRP247450 or from the Gene Expression Omnibus (GEO) with the accession number GSE144819. To avoid potential visual distortions in the heatmap, we used perceptually uniform color maps (Crameri, 2018). For parallelization, we employed GNU Parallel (Tange O, 2011).