Molecular basis for variations in the sensitivity of pathogenic rhodopsin variants to 9-cis-retinal

Over 100 mutations in the rhodopsin gene have been linked to a spectrum of retinopathies that include retinitis pigmentosa and congenital stationary night blindness. Though most of these variants exhibit a loss of function, the molecular defects caused by these underlying mutations vary considerably. In this work, we utilize deep mutational scanning to quantitatively compare the plasma membrane expression of 123 known pathogenic rhodopsin variants in the presence and absence of the stabilizing cofactor 9-cis-retinal. We identify 69 retinopathy variants, including 20 previously uncharacterized variants, that exhibit diminished plasma membrane expression in HEK293T cells. Of these apparent class II variants, 67 exhibit a measurable increase in expression in the presence of 9-cis-retinal. However, the magnitude of the response to this molecule varies considerably across this spectrum of mutations. Evaluation of the observed shifts relative to thermodynamic estimates for the coupling between binding and folding suggests underlying differences in stability constrains the magnitude of their response to retinal. Nevertheless, estimates from computational modeling suggest that many of the least sensitive variants also directly compromise binding. Finally, we evaluate the functional properties of three previous uncharacterized, retinal-sensitive variants (ΔN73, S131P, and R135G) and show that two of these retain residual function in vitro. Together, our results provide a comprehensive experimental characterization of the proteostatic properties of retinopathy variants and their response to retinal.

Over 100 mutations in the rhodopsin gene have been linked to a spectrum of retinopathies that include retinitis pigmentosa and congenital stationary night blindness. Though most of these variants exhibit a loss of function, the molecular defects caused by these underlying mutations vary considerably. In this work, we utilize deep mutational scanning to quantitatively compare the plasma membrane expression of 123 known pathogenic rhodopsin variants in the presence and absence of the stabilizing cofactor 9-cis-retinal. We identify 69 retinopathy variants, including 20 previously uncharacterized variants, that exhibit diminished plasma membrane expression in HEK293T cells. Of these apparent class II variants, 67 exhibit a measurable increase in expression in the presence of 9-cisretinal. However, the magnitude of the response to this molecule varies considerably across this spectrum of mutations. Evaluation of the observed shifts relative to thermodynamic estimates for the coupling between binding and folding suggests underlying differences in stability constrains the magnitude of their response to retinal. Nevertheless, estimates from computational modeling suggest that many of the least sensitive variants also directly compromise binding. Finally, we evaluate the functional properties of three previous uncharacterized, retinal-sensitive variants (ΔN73, S131P, and R135G) and show that two of these retain residual function in vitro. Together, our results provide a comprehensive experimental characterization of the proteostatic properties of retinopathy variants and their response to retinal.
Mutations in integral membrane proteins are responsible for a variety of genetic diseases (1,2). Most such mutations generate a loss of function as a result of one or more molecular defects, including the disruption of protein folding, the attenuation of protein expression, changes in protein localization, and/or the perturbation of a protein's intrinsic activity (2,3). An understanding of the molecular defects caused by specific mutations can provide a decisive advantage in drug discovery and targeting (4). For instance, mechanistic knowledge of the effects of common cystic fibrosis transmembrane conductance regulator variants facilitated the successful development of both corrector compounds that restore the expression of misfolded variants (class II) and potentiator compounds that enhance the gating of inactive variants (class III & IV) that are responsible for cystic fibrosis (5,6). Though the emergence of clinical sequencing has accelerated the discovery of MP variants that are associated with a variety of diseases, experimental efforts to characterize the molecular effects of their mutations have not kept pace (4). Therefore, new experimental approaches that enable rapid characterization of disease-linked membrane protein variants and how they respond to therapeutic compounds are needed to guide the development of precision therapeutics (2).
There are currently over 100 known mutations in the rhodopsin G protein-coupled receptor that cause a spectrum of visual retinopathies including autosomal dominant retinitis pigmentosa (adRP) and congenital stationary night blindness (CSNB) (7). Most experimentally characterized adRP variants accumulate in the endoplasmic reticulum in a manner that compromises their maturation within the secretory pathway (8)(9)(10). In contrast, CSNB variants are typically well expressed but exhibit constitutive activation (11). Nevertheless, there are wide variations in the age of onset and severity of the retinopathies that likely reflect differences in the molecular effects of these mutations and other uncharacterized variants (7). The expression of many of retinopathy variants can be partially restored by analogs of rhodopsin's native 11-cis-retinal cofactor and/or other small molecules that bind and stabilize the opsin apoprotein (12)(13)(14)(15)(16). Despite the discovery of numerous therapeutic lead compounds and the initiation of several clinical trials, there are currently no approved treatments for these disorders. An improved understanding of the molecular effects of the spectrum of clinical rhodopsin variants may help to identify a subset of "correctable" rhodopsin variants that could be targeted in future clinical trials.
In the following, we apply deep mutational scanning (DMS) to quantitatively compare the plasma membrane expression (PME) of 123 known rhodopsin variants that are associated with visual retinopathies, including 42 that were previously uncharacterized. We show that 69 of these 123 variants exhibit deficient PME in HEK293T cells, including 20 that were previously uncharacterized. Our results reveal that the mutations that have the most severe proteostatic effects on the opsin apoprotein cluster within the protein core and/or retinal binding pocket. Of the 69 putative class II variants, 67 exhibit a measurable increase in expression in the presence of 9-cisretinal-a photostable isomer of rhodopsin's native 11-cisretinal cofactor. However, response to retinal varies greatly across this spectrum of variants. A comparison of the observed effects of retinal to theoretical estimates of the stabilization afforded by retinal binding suggests that responses are generally constrained by stability. Nevertheless, binding calculations imply that many of the least responsive variants also directly disrupt retinal binding. Finally, we show that two of the three previously uncharacterized variants that exhibit the largest change in PME in the presence of retinal are capable of regenerating rhodopsin pigments that retain residual signaling activity in vitro. Together, our findings provide a comprehensive overview of the proteostatic effects of pathogenic rhodopsin variants that may help to guide the discovery and targeting of rhodopsin corrector molecules.

Survey of the PME of retinopathy variants
To measure the proteostatic effects of retinopathy mutations by DMS, we first assembled a pooled genetic library of containing 119 adRP variants and four CSNB variants. This group of 123 missense and single-codon deletion variants includes 42 previously uncharacterized variants, 57 known class II variants, and 24 variants with other classifications (classes I, and III-VII, Table S1). These mutations are distributed across the primary structure of rhodopsin (Table S1). To ensure even sampling, we generated a set of individual plasmids in which each variant can be matched to a single unique molecular identifier sequence. A stoichiometric mixture of these plasmids was then used to create a pool of recombinant HEK293T cells in which each cell inducibly expresses a single retinopathy variant from a defined genomic locus, as was described previously (17,18). A flow cytometry analysis of opsin variant surface immunostaining reveals that 51% of these cells express variants with comparable PME to wildtype (WT), while the remaining cells express variants with comparable staining to the class II P23H variant (Fig. 1). The bimodal nature of this distribution reflects the fact that some retinopathy variants compromise PME (class II) while others simply perturb signaling (classes I and III-VII) (7). The relative proportion of cells expressing P23H-like variants decreases by 6% in the presence of 5 μM 9-cis-retinal, which suggests the PME of many class II variants can be partially restored by this investigational corrector.
To estimate the PME of individual variants, we utilized fluorescence activated cell sorting (FACS) to fractionate these cells according to surface immunostaining, extracted the genomic DNA from each fraction, and used deep sequencing of the recombined unique molecular identifier sequences to quantify the relative abundance of each variant within each fraction. Sequencing data were then used to estimate the surface immunostaining intensity of each variant, as was described previously (18). Consistent with expectations, known class II variants exhibit lower intensities relative to WT (Average Intensity = 17,400 ± 1800) and other variants that exhibit other types of conformational defects (Mann-Whitney p = 1.9 × 10 −11 , Fig. 2A). The distribution of surface immunostaining intensities among previously uncharacterized variants spans the range in between those of previously characterized variants ( Fig. 2A). This collection of retinopathy variants features a prominent cluster with little to no detectable plasma membrane opsin (severe class II, 44 variants) and a cluster with surface immunostaining intensities that are comparable to WT (other classes, 54 variants, Fig. 2B). Nevertheless, there are also several variants with intermediate surface immunostaining intensities (moderate class II, 25 variants, Fig. 2B). A projection of variant intensity values onto the structure of rhodopsin reveals that mutations of buried residues near the retinal binding pocket generally cause the largest reduction in PME (Fig. 2, A and C). By comparison, mutations within the disordered C-terminal tail have similar expression to WT (Table S1). Variants that are predicted to compromise the translocon-mediated membrane integration of TM domains (19) or to disrupt the stability of the native fold generally exhibit attenuated expression ( Fig. 2A). Together, these results unambiguously identify a comprehensive set of retinopathy variants with attenuated PME and suggest that their proteostatic effects generally arise from perturbations of cotranslational and/or posttranslational folding.

Impacts of 9-cis-retinal on the PME of retinopathy variants
The proteostatic effects of retinal varies considerably across the spectrum of destabilized rhodopsin variants (16,18,20). We therefore repeated these experiments in the presence of 5 μM 9-cis-retinal to identify retinopathy variants that are most amenable to correction. Consistent with our recent findings (20), the results show that poorly expressed variants exhibit the largest change in expression in the presence of retinal (Fig. 3A). There are clusters of highly responsive variants that include several previously uncharacterized mutations near the binding pocket (L47R and T289P) and the cytosolic interface (L57R and ΔN73) (Table S1 and Fig. 3B). Nevertheless, a structural map of mutated side chains reveals that responsive and nonresponsive mutants are found throughout the structure (Fig. 3B). This observation implies that, outside of a few select variants, increases in expression are unlikely to arise from structurally specific interactions between mutated side chains and retinal itself. Our recent findings suggest, instead, that these variations likely reflect the composite effects of mutations on various molecular features that are tied, directly or indirectly, to changes in conformational stability (20). Consistent with this possibility, we find that absolute increases in expression are most pronounced among moderate class II variants (Fig. 3C), which perturb different structural regions but generate a comparable reduction in expression. Together, these results identify a subset of "correctable" retinopathy variants and imply that proteostatic responses may be tied to differences in opsin/rhodopsin stability. A, a box and whisker plot depicts the distributions of surface immunostaining intensities among variants that are known to cause misfolding (class II) relative to those known to cause other conformational defects (other classes) and those that are previously uncharacterized (not classified). Distributions are also shown for subsets of variants that are grouped according to whether they occur at positions that are close to the retinal centroid (<15 Å), far from the retinal centroid (>15 Å), buried in the protein core (<15% relative surface area), or solvent exposed (>15% relative surface area). Distributions for mutations that are predicted by Rosetta to stabilize (ΔΔG < 0) or destabilize (ΔΔG > 0) the native conformation or that are predicted by the biological hydrophobicity scale to enhance (ΔΔG < 0) or disrupt (ΔΔG > 0) translocon-mediated membrane integration are shown for reference. p-values were calculated using a two-tailed Mann-Whitney U-test. B, a histogram depicts the range of observed surface immunostaining intensities among the 123 retinopathy variants. Blue, gray, and red regions reflect the intensity intervals corresponding to the designations for severe class II, moderate class II, and other classifications, respectively. C, intensity values for individual variants are projected onto the corresponding mutated side chains in the three-dimensional structure of rhodopsin (PDB 3CAP). Side chain Cβ atoms (or glycine H) are colored according to the average intensity from three replicate DMS experiments. Blue indicates poor expression, white indicates intermediated expression, and red indicates expression levels comparable to WT.

Energetic interpretation of the observed trends in variant expression
The sensitivity of moderate class II variants to retinal potentially arises from the energetic coupling between binding and folding. To rationalize the energetic basis of these observed trends, we used a series of simplifying assumptions to approximate the stability of each variant, how much retinal binding should increase variant stabilities, and how much this stabilization should increase variant expression levels (see Materials and Methods). These gross simplifications provide a lens to understand how differences in stability shape the response to retinal. First, we note that retinal should have no effect on variants that compromise binding. These variants should therefore fall along a diagonal when rhodopsin variant intensities (+retinal) are plotted against their corresponding opsin intensities (apo) (red dashes, Fig. 4A). In contrast, variants with native binding energetics should exhibit an increase in intensity proportional to the change in the combined fraction of folded opsin and rhodopsin (f fold ), which can be calculated by combining ΔG fold values estimated from the expression of each variant with the estimated free energy of binding (ΔΔG fold 1.1 kcal/mol) (blue dashes, Fig. 4A). The observed changes in variant immunostaining intensities generally fall between these bounds (Fig. 4A). While our data lack true internal measurements for WT, estimated intensities from independent measurements (21) place the shift in WT expression close to the upper bound (Fig. 4A, purple). Consistent with the observed trends ( Fig. 3C), the general properties of this model and the resulting shape of the upper boundary suggest marginally stable variants (apparent ΔG fold 0 kcal/mol) should exhibit the largest absolute increases in intensity (blue dashes, Figs. 4A and S1). Nevertheless, relatively few variants exhibit gains in expression that approach the upper bound (Fig. 4A), which suggests the response to retinal is also likely sensitive to other effects of these mutations (20).
The insensitivity of certain variants to retinal could reflect secondary effects of these mutations on the binding affinity (20). To evaluate whether changes in binding energetics appreciably contribute to variations in the proteostatic effects of retinal, we analyzed the observed expression patterns in relation to computational estimates for the effects of mutations on binding. We first assumed that deviations from the upper bound arise solely from the effects of mutations on the binding energy. Using the framework described above, we estimated the impact of retinal on the f fold and ΔG fold for each variant from observed changes in immunostaining intensity. Based on the change in the apparent ΔG fold (ΔΔG fold ), we then calculated the corresponding change in the retinal equilibrium dissociation constant relative to WT (K d Mut /K d WT ).
To determine whether these projected variations in binding can be reconciled with the structural effects of individual mutations, we carried out two distinct structural analyses. We first utilized Rosetta to generate structural models of both the opsin (apo) and rhodopsin (+retinal) forms of the target variants. To survey perturbations of the initial binding reaction, we used RosettaLigand (22) to dock 9-cis-retinal into the binding pocket of each WT opsin and used this docked structure as a template to build structural models of each variant. We then used a convolutional neural network (K DEEP ) (23) to predict changes in the free energy of binding. To survey perturbations of the mature pigment, we used Rosetta to calculate the change in the protein-ligand interface energy in the context of rhodopsin variant models bearing the native Schiff base linkage between K296 and 9-cis-retinal. Both sets of calculations suggest retinal-sensitive mutants (K d Mut /K d WT ≤ 10) vary with respect to their predicted effects on the retinal binding energy (Fig. 4, B and C). However, both analyses also show that the variants that are least sensitive to retinal (K d Mut /K d WT ≥ 10) are generally predicted compromise binding (Fig. 4, B and C). We should note that we restricted this analysis to the 25 Values represent the average of three biological replicates and brackets reflect the standard deviation. Blue, gray, and red regions reflect the intensity intervals corresponding to the designations for severe class II, moderate class II, and other classifications, respectively. The blue dashed line reflects the upper boundary for the increase in intensity for variants that retain WT binding affinity. The red dashed line reflects the lower intensity boundary for variants that do not bind retinal. B, 9-cis-retinal was noncovalently docked into structural models of the 25 intermediate class II variants, and the corresponding binding energy was predicted using the KDEEP web server (23). Predicted binding energies are plotted against thermodynamic approximations for the increase in K d that were calculated based on the observed changes in variant immunostaining intensities. C, Rosetta was used to estimate the protein-ligand interface energy in the context of structural models of the 25 intermediate class II variants featuring a Schiff base between 9-cisretinal and K296. Rosetta interface energies are plotted against thermodynamic approximations for the increase in K d that were calculated based on the observed changes in variant immunostaining intensities.
intermediate class II variants due to dynamic range constraints; the difference between the upper and lower bounds approaches the magnitude of experimental variation for variants with the highest or lowest expression (Fig. 4A). Thus, it is unclear whether changes in binding energetics are likely to factor into the retinal-sensitivity of this entire spectrum of variants. Nevertheless, these results suggest that mutationspecific responses can be generally reconciled with energetic perturbations of binding and/or folding equilibria.

Functional characterization of retinal-sensitive variants of uncertain significance
Though our analyses identify several previously unclassified variants that exhibit enhanced expression in the presence of 9cis-retinal (Table S1), it is unclear whether any of these variants are likely to regain function. To assess the potential functional relevance of these observed proteostatic effects, we characterized the biochemical properties of two previously unclassified intermediate class II variants (ΔN73 and R135G) and one previously unclassified severe class II variant (S131P) that recover expression in the presence of retinal (Table S1). Briefly, we transiently expressed each of these variants in HEK293T cells in the presence of 9-cis-retinal, harvested cellular membranes, and purified each variant into n-dodecyl-β-D-maltopyranoside (DDM) micelles. Absorbance spectra of each purified variant reveals that both intermediate class II variants (ΔN73 and R135) are at least partially capable of binding retinal and recovering the native absorbance of WT rhodopsin (Fig. 5A). Moreover, these variants retain the ability to partially activate G t in response to photoactivation (Fig. 5B). In contrast, the severe class II variant S131P failed to regenerate the native pigment even though treatment with this molecule nearly doubles its PME (Fig. 5A and Table S1). This observation potentially suggests this proteostatic response arises from cotranslational interactions (20). Together, these findings reveal that certain intermediate class II variants are likely to retain residual activity upon recovery of expression but raise doubts about the druggability of severe class II variants.

Discussion
Corrector molecules hold great promise for the treatment of protein misfolding diseases (24). However, their efficacy varies widely across the spectrum of clinical mutations that enhance misfolding (16,(25)(26)(27)(28). Though an understanding of how different variants respond to correctors can aid in their development (5), the sheer volume of clinical variants and the constraints of traditional biochemical and biophysical assays (29) has historically precluded their comprehensive characterization. In this work, we utilize DMS to characterize the proteostatic effects of 123 retinopathy-linked rhodopsin variants in HEK293T cells and measure their response to the investigational corrector 9-cis-retinal. Though the effects of these mutations may differ in the native context of the rod outer segment, we show that the observed PME of previously characterized variants are highly consistent with previous classifications (Fig. 2A) and with the results of another recent high throughput investigation of retinopathy variant expression (9). Our measurements also identify 13 previously uncharacterized variants with severely deficient PME and seven that exhibit moderately deficient PME (Table S1). The 22 other uncharacterized variants characterized herein exhibit robust expression and presumably compromise other aspects of signaling. Our measurements in the presence of retinal reveal that while variants with the lowest expression exhibit the biggest change in PME (Fig. 3, A and C), variants with intermediate expression exhibit the largest absolute increases in PME (Fig. 3C). To rationalize these observed mutagenic trends, we outline a thermodynamic framework to interpret Figure 5. Functional properties of purified retinitis pigmentosa variants. The propensity of purified opsin variants to bind retinal and photoactivate Gt are shown. A, mutant rhodopsin pigments were regenerated in HEK293T cells through the addition of 9-cis-retinal then purified into DDM micelles. The relative absorbance of R135G (green), ΔN73 (blue), S131P (red), and WT (black) rhodopsins are plotted as a function of wavelength. B, purified Gt was mixed with regenerated rhodopsins, and the kinetics of Gt activation was monitored over time according to the change in fluorescence that arises from nucleotide exchange. The change in fluorescence emission at 345 nm following photoactivation of R135G (green), ΔN73 (blue), and WT (black) rhodopsins are plotted over time. Fitted rate constants are shown for reference. DDM, n-dodecyl-β-D-maltopyranoside. the effects of retinal on the PME and to infer which mutants compromise retinal binding. Together, our results provide a holistic overview of the proteostatic effects of known retinopathy variants and identify a discrete subset that are potentially amenable to pharmacological correction.
Our data suggest that 56% of retinopathy variants (69/123) exhibit a considerable attenuation of PME (Table S1)-a figure that is slightly lower than the relative proportion of class II variants associated with other diseases of protein misfolding (30)(31)(32). Of these variants, 64% (44/69) exhibit severely reduced PMEs. Consistent with expectations, variants that are predicted to disrupt translocon-mediated membrane integration and/or to destabilize the native conformation exhibit the largest decreases in expression ( Fig. 2A). Severe class II variants also exhibit minimal gain in expression in the presence of 5 μM 9-cis-retinal (Table S1 and Fig. 3B). Our thermodynamic projections imply that these mutations are insensitive to retinal because their energetic effects on stability far outweigh the stabilization that is imparted from binding (Figs. 4A and S1), though some severe class II variants may also compromise binding. The identification of such variants is complicated by the fact that the magnitude of the expected change in PME is within the error of the experiment (Fig. 4). Regardless of the mechanism, variants that are insensitive to correctors are often labeled "irreversibly" misfolded. Nevertheless, thermodynamic considerations suggests variants that retain binding affinity should generally exhibit larger gains in PME in the presence of compounds that bind with higher affinity (Fig. S1). This provides room for optimism, as 9-cis-retinal binds with much lower affinity than rhodopsin's native cofactor 11-cis-retinal (K d = 25 PM) (33). As is true for other classes of correctors (34), binding affinity is a key consideration for emerging retinoid and nonretinoid rhodopsin correctors (10,14,35). Provided the sensitivity of these variants is rooted in stability, the moderate class II variants that exhibit large shifts in response to 9-cis-retinal should generally represent the most favorable target variants for other correctors as well.
The theoretical and computational approaches described herein provide new tools to rationalize the spectrum of proteostatic effects that can arise from the interplay between mutations and small molecules. Due to the simplicity of their operative underlying principles, these approaches can potentially be generally useful for the interpretation of wider mutagenic trends in other membrane proteins. Nevertheless, it is important to recognize that this framework represents an oversimplification of the complex molecular defects that are caused by certain mutations. For instance, the modeling approaches introduced herein do not account for the kinetic assembly defects and/or changes in the interactome caused by certain mutations. It is also likely that these variants differ with respect to the basis of their attenuated PME-some may enhance degradation within the endoplasmic reticulum while others may exhibit accelerated internalization and/or degradation at the plasma membrane. Though these variables surely factor into deviations in the expression and retinal response of certain variants, absent additional information, these factors cannot be easily incorporated into modeling approaches. Nevertheless, our framework provides a means to form mechanistic hypotheses through deductive logic-moderate class II variants that do not respond to retinal are perhaps more likely to compromise binding or to exhibit assembly defects that remodel their interactome. Expanding upon the approaches detailed herein will require the development of new approaches to compare other properties of these variants at scale. For instance, repeating these measurements in cell lines that are deficient in the expression of various molecular chaperones may help differentiate the various classes of misfolded variants and expand our understanding of mechanistic basis for deviations in their pharmacological properties.
Despite many efforts to develop correctors for misfolded rhodopsin variants, there are currently no treatments for adRP or CSNB. Most efforts to discover correctors have been evaluated in relation to their effects on the expression and/or maturation of the P23H variant-the most common pathogenic rhodopsin variant (12)(13)(14)(15)36). However, P23H is also among the most poorly expressed variants and exhibits only modest sensitivity to retinal (Table S1). Based on this consideration, ongoing corrector screens could potentially achieve better sensitivity by targeting moderate class II variants that exhibit a greater response (i.e., ΔN73, R135G, or Y191C, Table S1). We should also note that the DMS approach described herein could potentially provide an efficient approach to compare the mutation-specific responses to emerging lead compounds. Such experiments will provide key insights as to whether different correctors can rescue distinct variants or whether there is one common set of "correctable" variants that can be targeted in clinical trials. These considerations highlight new applications of DMS for precision pharmacology.

Experimental procedures Plasmid preparation and mutagenesis
A previously described pcDNA5 vector containing the cDNA sequence for the human rhodopsin gene (RHO), an Nterminal hemagglutinin (HA) tag, an internal ribosome entry site-dasher GFP cassette, and a Bxb1 recombination site in place of the promoter (18) was used to generate a molecular library of barcoded retinopathy variants. We first installed a randomized 10 nucleotide "barcode" region upstream of the Bxb1 recombination site using nicking mutagenesis (37). A plasmid preparation containing a mixed population of barcoded vectors was used as a template for 123 individual sitedirected mutagenesis reactions to generate a library of retinopathy variants that were found in either the Uniprot Database, the Human Gene Mutation Database, and/or the Leiden Open Variation Database. Individual clones from each reaction were generated using the GeneJET Plasmid Miniprep Kit (ThermoFisher Scientific) and sequenced to confirm the sequence of each mutated open reading frame and to determine its corresponding 10 base barcode sequence. Plasmids encoding individual variants were pooled and electroporated into electrocompetent NEB10β cells (New England Biolabs), which were then grown in liquid culture overnight and purified using the ZymoPure endotoxin-free midiprep kit (Zymo Research). The Bxb1 recombinase expression vector (pCAG-NLS-HA Bxb1) was provided by Douglas Fowler.

Production and fractionation of recombinant cell lines
A pool of recombinant stable cells expressing individual retinopathy variants was generated using a previously described stable HEK293T cell line containing a genomic Tet-Bxb1-BFP landing pad (17). Recombinant cells were generated and isolated as was previously described (18). Briefly, cells grown in 10 cm dishes in complete Dulbecco's modified Eagle medium (Gibco) supplemented with 10% fetal bovine serum (Corning) and penicillin (100 U/ml)/streptomycin (100 μg/ml) (complete media) were cotransfected with our library of retinopathy variants and the Bxb1 recombinase expression vector using Fugene 6 (Promega). Doxycycline (2 μg/ml) was added 1 day after transfection, and the cells were grown at 33 ⁰C for the following 4 days. On the fourth day, cells were sorted using the BD FACS Aria II (BD Biosciences) to isolate GFP-positive/ BFP-negative cells that had undergone recombination. These cells were grown in 10 cm dishes with complete media supplemented with doxycycline (2 μg/ml) for up to 7 days. Where indicated, cells were incubated with 5 μM 9-cis-retinal for 16 h prior to sorting. Rhodopsin expressed at the plasma membrane of recombinant cells was labeled with a DyLight 550-conjugated anti-HA antibody (ThermoFisher). Labeled cells were then fractionated into quartiles according to surface immunostaining intensity using a FACS Aria IIu fluorescence activated cell sorter (BD Biosciences). At least two million cells from each fraction were isolated to ensure exhaustive sampling. Fractionated subpopulations were expanded in 10 cm culture dishes prior to harvesting and freezing 10 to 20 million cells per quartile fraction for the downstream genetic analysis.

Extraction of genomic DNA and preparation of nextgeneration sequencing libraries
To track the expression of individual retinopathy variants, we first extracted the genomic DNA (gDNA) from each cellular fraction using the GenElute Mammalian Genomic DNA Miniprep kit (Sigma-Aldrich). A previously described seminested polymerase chain reaction (PCR) technique (17) was then used to selectively amplify the barcoded region of the recombined plasmids within the gDNA. Briefly, an initial PCR reaction was used to first amplify the region of interest from the gDNA. The product of this reaction was then used as a template for a second PCR reaction that amplified the barcoded region while installing indexed Illumina adapter sequences. Amplicons were gel-purified using the Zymoclean Gel DNA Recovery Kit (Zymo Research). The purity of each sequencing library was confirmed using an Agilent 2200 TapeStation (Agilent Technologies). Libraries were sequenced using a NextSeq 500 Mid Output 150-cycle kit (Illumina) at an average depth of 2 million reads per quartile.

Estimation of surface immunostaining levels from DMS data
Surface immunostaining levels were estimated from sequencing data using a computational approach described previously (18). Briefly, low quality reads that were likely to contain more than one error were removed from the analysis. The remaining reads containing one of the 123 barcodes corresponding to a variant of interest were then rarefied to generate subsampled datasets with a uniform number of reads for each sample. We then calculated weighted-average immunostaining intensity values for each barcode/variant using the following equation: where CID variant is the weighted-average fluorescence intensity of a given variant, CFD i is the mean fluorescence intensity associated with cells from the i th FACS quartile, and N i is the number of barcode/variant reads in the i th FACS quartile. Variant intensities from each replicate were normalized relative to one another using the mean surface immunostaining intensity of the recombinant cell population on each day to account for small variations in laser power and/or detector voltage. Intensity values reported herein represent the average normalized intensity values from three replicate experiments.

Derivation of thermodynamic bounds for variant expression
Estimates for the upper and lower bounds for the change in variant expression in the presence of 9-cis-retinal were derived based on a series of simplifying assumptions concerning the relationship between the folding energetics, binding energetics, and the expression of the mature protein at the plasma membrane. The PME of integral membrane proteins in eukaryotic cells should generally scale with their thermodynamic preference for their native fold. (21,31,38) Therefore, we first assume that surface immunostaining is proportional to the combined equilibrium fraction of folded opsin and rhodopsin (f fold ). We also assume that this collection of variants includes some that are stable (ΔG fold ≤ −3 kcal/mol) and others that are unstable (ΔG fold ≥ 3 kcal/mol). Based on this criterion, the highest and lowest variant immunostaining intensities were taken as the signal generated by cells expressing fully folded (f fold 1) and fully unfolded (f fold 0) variants, respectively. This scaling can be used to approximate the fraction of folded protein (f fold ) using the following generalizable equation: Equation 3 can be used to derive a lower boundary for the change in immunostaining intensity for arbitrary variants in the presence of retinal based on the differences in the observed experimental fluorescence intensities (red dashes, Fig. 4A).
Approximations of f fold must be cast in terms of the free energy of folding (ΔG fold ) to project the effects of binding energetics on the immunostaining intensities of variants that bind retinal. Approximated f fold values can then be used to calculate the corresponding value of ΔG fold using the following generalizable equation: where R is the gas constant, and T is the temperature. The free energy of folding in the absence of retinal (ΔG Apo ) for each variant can be determined by plugging f fold,Apo into Equation 4. The ΔG Apo value for each variant can then be used to estimate the apparent free energy of folding in the presence of retinal (ΔG App ) using the following previously derived equation (39):  Most variants fall between these upper and lower bounds (Fig. 4A). If it is assumed that the difference between this projected I Mut,Ret and the observed I Mut,Ret for a given variant arises solely from the effects of the mutation on binding energetics, then an approximation for K d,Mut can be projected from observed immunostaining intensities as follows. First, Equation 2 can be used to calculate f fold,Ret from I Max,Ret , I Min,Ret , and the observed Immunostaining intensity of the variant in the presence of retinal. The corresponding ΔG app can then be determined by plugging f fold,Ret into Equation 3. An approximated K d,Mut can then be calculated by plugging the resulting ΔG app along with the ΔG fold for the apo form of the variant and [L] into Equation 5, then solving for K d .

Computational predictions of the effects of mutations on folding energetics
The impacts of mutations on the thermodynamic stability of rhodopsin were estimated by constructing molecular models of each rhodopsin variant and comparing their stability using a membrane protein-specific Rosetta energy function as described previously. (18) Briefly, a high-resolution crystal structure of bovine rhodopsin [Protein Data Bank (PDB) 3C9L] was used to generate homology models of each human rhodopsin variant as was previously described (21). A Rosetta ΔΔG protocol was then using to estimate the effects of each mutant on the conformational stability of the native fold (40). The effect of RP mutations on cotranslational TM domain integration was estimated using a previously described ΔG predictor algorithm (https://dgpred.cbr.su.se/) (41).

Computational predictions of the effects of mutations on binding energetics
We estimated the effects of mutations on the binding energy using a series of structural models for each mutant in both its apo state (opsin) and its covalently bound state (rhodopsin). Opsin and rhodopsin structures were modeled based on a crystal structure of bovine rhodopsin (PDB 2PED), which features a linkage to the 9-cis-retinal isomer used in these studies. To facilitate Rosetta modeling, the covalently linked retinal was treated as a noncanonical amino acid (NCAA) by converting its SDF to a molfile with Open Babel. The Biochemical Library code repository (http://www.meilerlab. org/index.php/bcl commons/show/b_apps_id/1) was then used to generate conformer libraries and Rosetta params files. The MutateResidue mover was then used to mutate K296 to this NCAA representation to facilitate its correct recognition by Rosetta. We used Rosetta scripts to introduce each mutation into the WT model, then used FastRelax in dualspace with the ref2015_cart energy function to refine each model. A custom packer palette was included to expand the default type sets used during relaxation of the retinal NCAA. To estimate the effects of mutations on the initial (noncovalent) binding reaction, we first converted the retinal-conjugated K296 in the PDB 2PED model back to lysine. We then used RosettaLigand to dock a noncovalent 9-cis-retinal back into the pocket and used the resulting structure as a template to generate variant models using the cartesian_ddg interface in Rosetta. The resulting models were then relaxed without constraints, and the free energy of binding for the lowest energy structure was then calculated using the K DEEP web server (23). To estimate the effects of mutations on the free energy of the covalently bound structure, the interface energy between retinal and the portions of the protein was estimated by taking the average of the InteractionEnergyMetric of the five lowest energy models for each mutant rhodopsin structure.

Purification and spectroscopic characterization of rhodopsin variants
Rhodopsin pigments were purified using a modified version of a previously described protocol (42). Briefly, polyethyleneimine was used to transfect twenty 10 cm dishes of HEK293T cells with pcDNA5 expression vectors containing the cDNA for human WT, S131P, R135G, and ΔN73 rhodopsin. Cells were treated overnight with 7.5 μM 9-cisretinal beginning 24 h posttransfection and were grown in the dark to facilitate the regeneration of each pigment. Cells were harvested 48 h posttransfection under dim red light, then pelleted by centrifugation at 800g. Cell pellets were either stored at −80 C or were directly lysed by rotating the slurry in the dark for 1 h at 4 C in 20 mM Bis-tris propane (BTP, pH 7.5) containing 120 mM NaCl, 20 mM DDM, and a protease inhibitor cocktail. Lysates were then clarified by centrifugation at 100,000g for 1 h at 4 C. Pigments were purified from the supernatant using a 1D4 anti-rhodopsin immuno-affinity chromatography. 200 μl of 1D4 resin (6 mg 1D4 monoclonal anti-rhodopsin antibody/ml agarose beads) were added to the supernatant and rotated for 1 h at 4 C. The beads were then transferred to a column and washed with 12 ml of 20 mM BTP (pH 7.5) containing 120 mM NaCl and 2 mM DDM, followed by a wash with 20 mM BTP (pH 7.5) containing 500 mM NaCl and 2 mM DDM. Rhodopsin pigments were eluted with 20 mM BTP (pH 7.5) containing 120 mM NaCl, 2 mM DDM, and 0.6 mg/ml of an elution peptide (TETSQVAPA) (21,43,44). The UV-visible spectra of purified rhodopsin pigments were recorded in the dark using a Cary 60 UV-visible spectrophotometer (Varian). The concentrations of purified pigments regenerated with 9-cis-retinal rhodopsins were determined assuming a molar extinction coefficient of Ɛ 485nm = 43,600 M −1 cm −1 (45).

G t activation measurements
The ability of mutant pigments to activate G t in vitro was measured as previously described (21). Briefly, G t was extracted and purified from frozen rod outer segment membranes isolated from 100 dark-adapted bovine retinas (15,44). Purified G t was mixed with purified rhodopsin variants to final concentrations of 250 nM and 25 nM, respectively, in 20 mM BTP (pH 7.0) containing 120 mM NaCl, 1 mM MgCl 2 , and 1 mM DDM. This mixture was then illuminated for 1 min with a Fiber-Light illuminator (Dolan Jenner Industries Inc) through a 480 to 520 nm band-pass wavelength filter (Chroma Technology Corporation) in order to photoactivate the rhodopsin pigments. 10 μM GTPγS was then added following illumination, and the change in tryptophan fluorescence associated with the exchange of guanyl nucleotides within the α subunit of G t was measured for 1200 s with a FL 6500 Fluorescence Spectrometer (PerkinElmer). Excitation and emission wavelengths were set at 300 nm and 345 nm, respectively (15,43). G t activation rates were determined by fitting the change in fluorescence intensity over the initial 600 s with a single exponential function.

Data availability
Deep mutational scanning measurements and mutant classifications are reported in Table S1. All other data will be provided free of charge upon request by Jonathan Schlebach (jschleba@iu.edu). Funding and additional information-This research was supported by grants from the National Institutes of Health (NIH) (R01GM129261 to J. P. S., R01EY025214 to B. J., R01HL122010 to J. M., R01GM080403 to J. M., and R01DA046138 to J. M.). F. J. R. acknowledges receipt of a predoctoral fellowship from the Graduate Training Program in Quantitative and Chemical Biology at Indiana University (T32 GM109825). The authors also thank the Visual Science Research Center Core at Case Western Reserve University (supported by NIH grant P30EY011373). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Conflict of interest-The authors declare that they have no conflicts of interest with the contents of this article.