Co-translational formation of disulfides guides folding of the SARS-CoV-2 receptor binding domain

Abstract

Many secreted proteins, including viral proteins, contain multiple disulfide bonds. How disulfide formation is coupled to protein folding in the cell remains poorly understood at the molecular level. Here, we combine experiment and simulation to address this question as it pertains to the SARS-CoV-2 receptor binding domain (RBD). We show that the RBD can only refold reversibly if its native disulfides are present before folding. But in their absence, the RBD spontaneously misfolds into a nonnative, molten-globule-like state that is structurally incompatible with complete disulfide formation and that is highly prone to aggregation. Thus, the RBD native structure represents a metastable state on the protein’s energy landscape with reduced disulfides, indicating that nonequilibrium mechanisms are needed to ensure native disulfides form before folding. Our atomistic simulations suggest that this may be achieved via co-translational folding during RBD secretion into the endoplasmic reticulum. Namely, at intermediate translation lengths, native disulfide pairs are predicted to come together with high probability, and thus, under suitable kinetic conditions, this process may lock the protein into its native state and circumvent highly aggregation-prone nonnative intermediates. This detailed molecular picture of the RBD folding landscape may shed light on SARS-CoV-2 pathology and molecular constraints governing SARS-CoV-2 evolution.

Significance

Viral proteins are often highly prone to misfolding, which significantly constrains virus evolution. But the specific molecular factors that account for this misfolding and how host cell mechanisms resolve these deficiencies remain poorly understood. As a model system, we show experimentally that the RBD of the SARS-CoV-2 virus can only fold spontaneously if its native disulfide bonds are present before folding. But in their absence, significant nonnative misfolding and aggregation ensue. Atomistic simulations reveal how this issue can be resolved if disulfides form co-translationally, which puts the protein on the correct path to form its native state. This detailed picture of disulfide-coupled protein folding may inform future studies that link SARS-CoV-2 protein misfolding to COVID-19 pathology and evolution of new variants.

Introduction

In recent decades, a growing body of work has cast doubt on the generality of Anfinsen’s thermodynamic hypothesis, which states that a protein’s native state corresponds to the thermodynamic minimum of its folding energy landscape (1). Instead, myriad examples have emerged of proteins that cannot reliably refold from a denatured state on physiological timescales, instead populating highly stable nonnative intermediates that may be prone to aggregation (2,3,4,5,6,7,8,9). In extreme cases, these intermediates may even correspond to global minima on the free energy landscape, whereas the native state represents a high-energy state—an energetic feature that may give proteins dynamic flexibility in response to stimuli or environmental changes (10,11,12,13). For such proteins, which cannot fold under thermodynamic control, two major questions arise: 1) how do cellular mechanisms ensure the native state is attained under kinetic control, and 2) what factors ensure that the native state persists over biological timescales despite the existence of alternative stable states?

A relevant feature of many proteins, pertaining to the above questions, is the presence of disulfide bonds (14). These covalent linkages between cysteine residues can confer as much as 5–6 kcal/mol of stability per bond by reducing the entropy of the denatured state, thus significantly increasing the population of the native state. (15,16,17). At what stage during protein folding do these highly stabilizing disulfide bonds typically form? For certain proteins, the correct disulfides can spontaneously form after dilution from a denatured, reduced state into a suitable redox buffer (1,18,19,20). In these cases, the noncovalent free energy landscape readily guides the protein to a native-like structure in which the cysteines are suitably positioned for spontaneous oxidation and disulfide formation. In other cases, the underlying folding landscape may favor the native state, but kinetic intermediates along the folding pathway may promote temporary nonnative disulfide formation or even random disulfide scrambling (21,22). These nonnative bonds can then isomerize to the native disulfides, either unassisted or with the aid of enzymes such as protein disulfide isomerase (23,24,25). But in a certain subset of proteins, the native disulfides cannot be reformed at all starting from a denatured, reduced state, as the underlying energy landscape strongly favors alternative folds that are structurally incompatible with these disulfides (22,26,27). Such proteins must necessarily rely on cellular mechanisms to ensure the correct native disulfides are present before folding.

In the cell, many proteins can begin forming their disulfide bonds co-translationally as they are being secreted into the endoplasmic reticulum (ER). The ER is a highly oxidizing environment and contains multiple enzymes such as protein disulfide isomerase (PDI) and Ero1 that catalyze disulfide formation in substrates (28,29,30). Numerous studies have shown that co-translational folding in the ER can significantly enhance the efficiency of native folding and disulfide formation while reducing the formation of nonnative intermediates relative to posttranslational oxidative refolding (23,27,31,32,33,34). This improvement may stem from the vectorial nature of synthesis, which can allow native structural units to form sequentially and thus mitigate misfolding that arises posttranslationally (35,36,37,38,39,40,41). However, the specific molecular details underlying these processes remain poorly understood in general.

In this work, we address these questions as they apply to the receptor binding domain (RBD) of the spike protein in SARS-CoV-2. This virus, which is behind the COVID-19 pandemic, has killed millions of people worldwide to date and continues to wreak havoc on lives and economies across the globe (42). Thus there is a pressing need for novel antiviral therapies that can broadly target novel, highly contagious variants as they emerge. The SARS-CoV-2 RBD, which comprises amino acids 333 through 527 of the S1 subunit of the spike protein, contains multiple disulfide bonds that have previously been shown to be crucial for ensuring the domain’s stability and functional binding to the human ACE2 receptor (43,44)—an interaction that directly leads to viral-host fusion (45,46). However, the mechanism through which RBD disulfide formation is coupled to its folding has not been investigated.

Using fluorescence assays, we show here that the RBD can only be refolded reversibly from denaturant if its native disulfide bonds are present before refolding. In contrast, oxidative refolding from a denatured and fully reduced state leads to misfolding into a nonnative molten-globule-like state that is highly prone to aggregation. This implies that the native state is kinetically trapped by its disulfides, and that the intrinsic free energy landscape of the reduced RBD favors alternative nonnative structures. Using a combination of gel-based assays, we then show that, upon oxidation, the nonnative structures form fewer disulfide bonds than the native state. Atomistic simulations recapitulate these findings and generate predictions of specific disulfides that are most likely to be absent in this state. These simulations additionally predict that, at intermediate lengths during RBD synthesis, native cysteine pairs frequently come into contact, whereas nonnative pairs typically remain far apart, suggesting that co-translational oxidation in the ER may help commit the RBD to its metastable native state. Together, these results paint a detailed molecular picture of how disulfide formation is coupled to folding in a biomedically crucial model protein. This study may in turn further our understanding of SARS-CoV-2 pathology and molecular evolution and inspire novel antiviral therapies that interfere with RBD folding.

Materials and methods

Preparation of refolded samples

The RBD from SARS-CoV-2 S1 protein (amino acids 319–541), expressed in Hek293 cells and purified to $\ge 95\%$ purity, was purchased from Thermo Fisher (RP87704). The lyophilized protein in PBS buffer was reconstituted to a concentration of 0.5–1 mg/mL ($\sim 20\mu$) in MilliQ water, and the purity was confirmed by SDS-PAGE. The protein was then denatured in 5 M guanidine hydrochloride (pH 7.4) supplemented with either 1 mM of tris(2-carboxyethyl)phosphine hydrochloride (TCEP, Research Products International) to produce the Red $\to$ Ox samples or an equivalent volume of MillliQ water to produce the Ox $\to$ Ox samples. The protein concentration was then adjusted, as necessary, either via dilution in respective denaturing buffer or concentration with Amicon Ultra 0.5-mL centrifugal filters. To achieve maximal protein concentration (in Fig. 2 c), the lyophilized protein was directly reconstituted in 5 M GdnHCl buffer. After a minimum of 6 h incubation under denaturing conditions, the protein was refolded via 10-fold dilution in either 5 mM oxidized glutathione (Millipore 3542) or 3.3 mM oxidized glutathione + 1.7 mM reduced glutathione dissolved in PBS buffer (pH 7.4) for at least 6 h to overnight before measurement. For samples that were to be loaded on a gel (as in Fig. 3), the denaturation step was performed with 8 M urea instead of GdnHCl; we show in Fig. S1 that both denaturants produce very similar final spectra after refolding.

Figure 2.

Reduced RBD folding landscape intrinsically favors a highly-aggregation-prone nonnative intermediate. (a) Size exclusion chromatography UV absorbance traces, normalized to maximum value, for native RBD and refolded RBD per Red $\to$ Ox protocol, defined as in Fig. 1. Dashed vertical lines indicate elution positions for standards with indicated molecular weights. (b) Two possible scenarios that may explain observed misfolding and aggregation. Solid curves with three minima represent energy landscapes for RBD monomers in the absence of disulfides. For details, see main text. (c) Ratio of intrinsic fluorescence intensity (solid red dots) at 370 nm to intensity at 335 nm as a function of time after manual-mix refolding (5 s dead time) averaged over two traces. The final protein concentration is $0.5\mu$M. The solid blue and gray lines depict the expected final ratio for RBD in the native and denatured + reduced state, respectively. The inset shows the refolding kinetics zoomed in on early times. (d) Ratio of intrinsic fluorescence intensity at 370 nm to intensity at 320 nm (black solid line) and ThT fluorescence (gray dashed line) as a function of time after reduction of RBD under native conditions. The protein and ThT concentrations are 6 and 7 $\mu M$, respectively, and the excitation and emission wavelengths used for ThT fluorescence were 450 and 485 nm, respectively.

Figure 3.

RBD nonnative state shows incomplete disulfide formation and is prone to forming cross-linked aggregates. (a) Schematic of cysteines and disulfides (blue curves) in the RBD construct used in our experiments, derived from the crystal structure with PDB: 6YOR. In the full spike protein, C205 is bonded with an additional cysteine (C590 in full spike numbering) not included in our construct. In our numbering scheme, residue 1 of the RBD corresponds to residue 333 of the full spike protein. (b) SDS-PAGE gel, stained with Coomassie G-250 dye, loaded with Ox $\to$ Ox and Red $\to$ Ox refolded samples with or without 10 mM TCEP added before loading, as indicated above respective lanes. The inset on the left shows labeled molecular weight standards.

To produce the native control, Red $\to$ Ox, native, and reduced samples (Fig. 1), the protein was prepared identically as above, except that during the first incubation step, the denaturant was replaced with an equivalent volume of PBS buffer (pH 7.4) supplemented with 5 mM TCEP (in Red $\to$ Ox native, and reduced samples) or an equivalent amount of water (in native control). The native control and Red $\to$ Ox samples were then diluted in 5 mM glutathione buffer spiked with GdnHCl for a final concentration of 0.5 M to match the final concentration in refolding measurements, whereas the reduced sample was diluted in PBS buffer spiked with TCEP and GdnHCl for final concentrations of 5 mM and 0.5 M, respectively. As before, both steps were performed for a minimum of 6 h.

Figure 1.

RBD can only refold reversibly in the presence of disulfides. (a) Schematic of Ox $\to$ Ox and Red $\to$ Ox protocols. RBD is denatured in 5 M GdnHCl in the presence (Red $\to$ Ox) or absence (Ox $\to$ Ox) of 1 mM TCEP to reduce disulfide bonds. Refolding is then initiated via 10-fold dilution into oxidizing buffer containing 5 mM oxidized glutathione. For details, see main text and materials and methods. (b) Intrinsic fluorescence spectra (excitation wavelength of 280 nm) of 1.2 $\mu M$ refolded RBD prepared via the Ox $\to$ Ox (labeled O$\to$O)and Red $\to$ Ox (labeled R$\to$O) protocols, as well as 1.2 $\mu \mathrm{M}$ native protein in 0.5 M GdnHCl to match final denaturant concentration after refolding. Each spectrum represents an average of two independent replicates. (c) Same spectra as in (b) normalized to maximum emission value to emphasize peak shift, alongside normalized spectra of 1.2 $\mu M$ Red $\to$ Ox native sample in which reduction was performed under native conditions (labeled R$\to$O nat.) and sample reduced under native conditions that was not re-oxidized (labeled Red. nat.), both in 0.5 M GdnHCl. All unnormalized spectra are shown in Fig. S1c. (d) bis-ANS fluorescence spectra (excitation wavelength of 385 nm) for all samples in (c). The bis-ANS concentration is $8\mu M$. The colors and line styles match those in (c). (e) bis-ANS fluorescence at 500 nm (385-nm excitation wavelength) as a function of temperature for samples prepared as in (b)–(d). The colors match those in (c).

Fluorescence measurements

For steady-state intrinsic fluorescence measurements, a minimum of 70 μL of protein sample was transferred into a 1-cm Helma fluorescence cuvette. The cuvette was then loaded into a Varian Cary Eclipse Fluorescence Spectrophotometer, and the sample was excited at 280 nm with a 5-nm excitation slit width and a 10-nm emission slit width at a scanning speed of 150 nm/min, averaging time of 0.2 s, and data collection interval of 0.5 nm. A blank measurement of respective buffer + protein replaced by an equivalent volume of PBS was also acquired. Samples were then recovered, and 4,4′-dianilino-1,1′-binaphthyl-5,5′-disulfonic acid, dipotassium salt (bis-ANS) was added to a final concentration of 8 μM. The samples were incubated for 10–15 min to allow the dye to bind the protein, and the resulting bis-ANS fluorescence was measured using an excitation wavelength of 385 nm. All other measurement parameters were the same as those used for intrinsic fluorescence, and a blank, bis-ANS-containing sample was similarly prepared and measured. All data processing, including blank correction and normalization to maximum, was carried out using custom Python scripts.

For thermal denaturation assays, protein and blank samples were prepared with 8 μM bis-ANS as above. The bis-ANS fluorescence (${\lambda }_{ex.}=385$ nm, ${\lambda }_{em.}=500$ nm) was then measured over a temperature range from 20 $°C$ to 80 $°C$ using a scanning speed of $0.5°\mathrm{C}/\min$ and 1 s averaging time. A blank sample was also measured, although it showed negligible change in fluorescence with temperature compared to the samples. For each sample, fluorescence at all temperatures is then divided by the maximum value observed to produce normalized melting curves (as in Fig. 1 d).

The kinetics measurements were initialized with an empty cuvette in the cell holder. At 1 s, the denatured protein was mixed with 90 μL of refolding buffer (5 mM oxidized glutathione in PBS, pH 7.4), and immediately transferred to the cuvette. To monitor kinetics with intrinsic fluorescence we used an excitation wavelength of 280 nm and a pair of two emission wavelengths, namely either 370 and 320 nm or 370 and 335 nm (1 ms integration time, excitation slit 5 nm, emission slit 10 nm). In experiments that monitored thioflavin T (ThT) fluorescence, $7\mu M$ ThT was added to the refolding buffer, and excitation/emission wavelengths of 450 and 485 nm, respectively, were used. The dead time of mixing, accounting for the time needed to transfer refolded protein to cuvette and initial mixing artifacts, is 5 s.

Size exclusion chromatography

Proteins were separated on a Superdex 75 Increase 10/300 GL column (Cytiva) that was preequilibrated with PBS buffer (pH 7.4). Samples were prepared at 9 μM and centrifuged at 15,000 g for 15 min to remove any insoluble aggregates. The supernatant (200 μL) was loaded at a flow rate of 0.8 mL/min, and elution was monitored by UV absorbance at 280 nm.

Cross-linking assay

100 μL of Ox $\to$ Ox and Red $\to$ Ox samples were prepared, with the denaturation step performed in 8 M urea. Each sample was then split into two $50\mu$L subsamples, one of which was mixed with LDS loading dye + iodoacetamide (IAA) to a final concentration of 10 mM (Ox $\to$ Ox and Red $\to$ Ox lanes), whereas the other was mixed with loading dye, IAA, and TCEP at a final concentration of 5 mM (Ox $\to$ Ox reduced and Red $\to$ Ox reduced lanes). The samples were incubated in the dark for 1 h to allow IAA to alkylate free cysteines (to prevent disulfide scrambling) and TCEP to reduce disulfides in the reduced subsamples. All subsamples were then loaded into an SDS-PAGE gel ($\sim 5\mu$ protein per lane), which was run at 150 V for 1 h and 15 min. The gel was then stained with Coomassie-G250 dye, photographed, and brightness and contrast were adjusted uniformly across the image to maximize intensity of bands.

Atomistic simulations

Atomistic simulations were run using MCPU software (47) with enhanced sampling and replica exchange as described previously (48). MCPU uses a knowledge-based potential (49) to model (noncovalent) interactions between atoms, as well as torsional and bond-angle terms. In addition, in constructs with constraints, a harmonic term with respect to cysteine-cysteine distance, centered at an alpha-carbon distance of 5 Å, was added for all native cysteine pairs. The initial structure of RBD was prepared from 334 to 528 of SARS-CoV-2 S1 protein (PDB: 6YOR), which was equilibrated for 50 million Monte-Carlo (MC) steps in our potential. The resulting energy-minimized structure was then used for a production run of 500 million MC steps, the last 400 million of which were used to compute equilibrium properties: at this point, simulations were judged to be well converged.

To compute equilibrium probability distributions for a desired variable X (e.g., topological configuration, native contacts, or intercysteine distance), the value of X was extracted for all simulation snapshots. In the case of topological configurations, substructures were created from the RBD contact map, and snapshots were assigned to configurations as described previously (48). For variable X of interest, we then computed a potential of mean force or free energy as a function of X using the multi-Bennett acceptance ratio method (50). From these free energy profiles, the Boltzmann probability distribution over X can readily be computed, and the mean of this distribution can be obtained. To ensure robustness, most quantities were computed at multiple simulation temperatures below the temperature at which the RBD was observed to melt.

Co-translational folding simulations were run and analyzed in much the same way as for full-length RBD, described above except for the design of the constraints. For constraints of co-translation folding simulations, the harmonic term was centered at a beta-carbon distance of 4.6 Å, instead of alpha-carbon. The RBD was truncated at the end of the nearest secondary structure element following each cysteine of interest (Cys 3,28, 58, 99, 155, and 195) resulting in the chain lengths indicated in Fig. 5. We likewise ran simulations at each length, except the first, in which all prior cysteines except the most C-terminal one were constrained to their native disulfide pairs: these results are shown in the supporting material.

Figure 5.

Atomistic simulations predict that co-translational folding significantly improves native RBD formation rate and yield. (a) Probability, as a function of translation length, that cysteine pairs (as indicated in legend) make contact within 5 Å of each other as predicted by atomistic simulation. (b) Schematic of kinetic model, detailed in the main text and materials and methods sections. In our kinetic model, opening and closing transitions that unite or separate cysteine pairs (double arrows) are assumed to occur on much faster timescales than oxidation (single arrows pointing right) (per Eq. 3). At a given length, any permutation of new disulfides involving unpaired cysteines may form, provided these cysteines have been translated; one possible sequence of disulfide formation events involving the first two translation lengths is shown here. (c) Solutions to kinetic model showing probability, as a function of time, of forming native species containing various permutations of the four native disulfides (indicated above respective curves) as well as trapped species containing the 28–58 nonnative disulfide (red line), assuming either that folding begins co-translationally per the kinetic model in (b) (top panel) or does not begin until synthesis is complete (bottom panel). The x axis depicts time normalized by the time needed to translate the full RBD; we note that synthesis must be complete before the final disulfide involving C-terminal residue 192 can form. We note that each line corresponds to the probability of a state with “only” the indicated disulfides formed. Formation of additional disulfides corresponds to a transition out of that state, hence why the probabilities that states with incomplete disulfides (e.g., the 3–28 and 46–99 state) always drop after an initial rise. The insets show zoomed-in traces corresponding to the state with all disulfides formed between normalized time values 1 (corresponding to immediately after translation) and 2. In this panel, the unknown intrinsic oxidation/translation rate ratio, $k_{ox}/k_{trans}$ is set to 5.

Kinetic model

To build a kinetic model, we define a set of states $\left\{s\right\}$, each of which corresponds to the collection of microstates in which some subset of observed disulfides is formed. A state is defined as containing either any permutation of the four native disulfides 3–28, 46–199, 147–155, and 58–152 or the observed nonnative disulfide 28–58 irrespective of whether other disulfides are present; this nonnative state is treated as an unproductive absorbing boundary. We assume state s can convert to state $s^{\prime }$ if $s^{\prime }$ contains all the same disulfides as s plus one additional disulfide d, and we assume that formation of d, and thus this s to $s^{\prime }$ transition, is irreversible. Although this is not strictly true given that disulfides can isomerize, the timescales for isomerization are relatively slow and may compete with nonnative aggregation—another absorbing boundary not considered here. For disulfide d to form starting from state s at translation length L, we require two things to happen, namely 1) the cysteines comprising d need to be brought into contact (defined as a distance less than 5 Å) via structural fluctuations, which occurs with equilibrium probability $P_{s,d}^L$ at length L, and 2) the cysteine pair must be oxidized by PDI—a process that, for simplicity, is assumed to occur at a rate $k_{ox}$ that is independent of length and disulfide pair identity. We also make the additional assumption that, at a given length, the probability of any cysteine pair forming is independent of whether other disulfides have already formed, and thus we can write $P_{s,d}^L=P_d^L$; this is justifiable because cysteine pairs in RBD are located in distinct structural regions. We further justify this assumption by running simulations up to each length at which a new cysteine pair has been synthesized in which all previous cysteine pairs have been constrained to form disulfides, and we observe that these constraints do not significantly affect the probability that the new pair forms (Fig. S9). We also assume that structural fluctuations are rapid compared to disulfide oxidation such that the effective disulfide formation rate can be modulated by the equilibrium constant involving the probabilities that the cysteines are paired; this is akin to EX2 kinetics in hydrogen-exchange experiments.

Under these assumptions, the effective formation rate for native disulfide d at length L is given by Eq. 3 in the results section. Likewise, all states in which neither disulfides 3–28 nor 58–192 have formed are assumed to be capable of transitioning to the nonnative state, indexed by n, in which the 28–58 disulfide at a rate given by

$$k_{28-58}^L=\frac{P_{28-58}^L}{1-P_{28-58}^L}{k}_{ox}$$ (1)

We then construct a transition matrix ${\mathbf{M}}^{\mathbf{L}}$ corresponding to translation length L with entries given by the following:

$$M_{s^{\prime },s}^L=\{\begin{array}{cc}{k}_d^L{\delta }_{s^{\prime }-s,d}& \text{if}s\ne s^{\prime }\\ k_{28-58}^L& \text{if}\mathrm{C}28\text{and}\mathrm{C}58\text{unpaired}\text{in}{s}^{\prime }\text{and}s=n\\ -\sum _{s^{″}}{M}_{s^{″},s}^L& \text{if}s=s^{\prime }\\ 0& \text{otherwise}\end{array}$$ (2)

Here, ${\delta }_{s^{\prime }-s,d}$ is a Kronecker delta function that takes on value 1 only if $s^{\prime }$ contains all the same disulfides as s plus additional disulfide d. We note that this transition matrix conserves probability, and thus all columns sum to 0.

We let $P^L\left(t\right)$ denote the vector of probabilities of occupying different states at time t at translation length L. This vector satisfies the master equation (Eq. 4).

To predict the time evolution of the disulfides while accounting for changing translation length, we first solve Eq. 4 numerically at a given translation length L, using the cysteine pairing probabilities drawn from the simulation at that length, for an amount of time given by ${\tau }_L={\tau }_{AA}\ast \left(L^{\prime }-L\right)$, where ${\tau }_{AA}$ is the average time to translate 1 amino acid, and $L^{\prime }$ is the next translation length simulated. The resulting probabilities, $P^L\left({\tau }_L\right)$, are then used as the initial probabilities at the next length $L^{\prime }$. At the first length $L=37$, the initial probability is assumed to be 1 at the state with no disulfides and 0 at all other states. Once the final length is reached, the time evolution is solved for an additional amount of time shown in the figures.

Results

RBD can only refold reversibly if native disulfides are present before folding

To begin, we investigated whether the RBD can refold reversibly from a denatured state, depending on whether or not the domain’s native disulfide bonds are disrupted during denaturation (Fig. 1 a). To this end, we chemically denature the RBD in 5 M guanidine hydrochloride (GdnHCl) in the presence or absence of 1 mM TCEP as a disulfide reducing agent; then we refold the protein in oxidizing buffer containing 5 mM oxidized glutathione. The refolded protein’s intrinsic fluorescence spectrum was then compared to that of the native protein to assess if refolding was successful.

We observe that, if denaturation is performed in the absence of reducing agent—a protocol we refer to as Ox $\to$ Ox to reflect that both denaturation and refolding occur under oxidizing conditions—then the refolded protein’s intrinsic spectrum perfectly overlaps with that of the native state (Fig. 1 b). Thus, refolding is reversible provided the native disulfides are maintained intact during denaturation. But if the denatured protein is reduced with TCEP and then re-oxidized during refolding—a protocol referred to as Red $\to$ Ox—then the refolded protein’s fluorescence spectrum shows a pronounced red shift and decrease in intensity relative to the native state (Fig. 1 b and c). This finding, which is robust to the composition of the glutathione buffer used during refolding (Fig. S1), suggests that the refolded, re-oxidized protein adopts a nonnative conformation in which the RBD tryptophan residues (two per subunit) are more exposed to solvent than in the native state. As a second structural probe for refolding, we incubated all constructs with bis-ANS—a dye that fluoresces upon binding hydrophobic cavities—and measured the resulting fluorescence. Strikingly, we observe that the Red $\to$ Ox construct shows significantly higher bis-ANS fluorescence at room temperature than the native and Ox $\to$ Ox constructs, suggesting that the former may exhibit molten-globule-like character, whereas the latter constructs are well folded (Fig. 1 d). To further confirm this, we measured the dependence of bis-ANS fluorescence on temperature for each construct (Fig. 1 e). The native and Ox $\to$ Ox melting curves show nonmonotonic fluorescence with temperature that peaks at $50°C$—an indicator of a well folded protein in which hydrophobic residues are compact at room temperature but become looser and more exposed at intermediate temperatures before complete denaturation. These two curves overlap, further indicating that the Ox $\to$ Ox leads to native refolding. In stark contrast, the Red $\to$ Ox construct shows monotonically decreasing fluorescence, with half-maximal intensity occurring around $37°\mathrm{C}$. This further points toward nonnative molten-globule-like character in Red $\to$ Ox and hints that this construct may be prone to high-temperature aggregation resulting in loss of fluorescence; similar behavior has previously been observed in destabilizing mutants of the enzyme dihydrofolate reductase (51).

These results indicate that RBD can only attain its correct structure if its native disulfides are present before folding, whereas disulfide reduction during denaturation followed by oxidative refolding leads to misfolding. One mechanism that may explain this misfolding is if, early during refolding, the chain becomes locked into kinetically accessible but high-energy states via semirandom disulfide formation—a behavior previously observed in proteins such as hirudin (21,28) and RNAse-A (52). To assess this possibility, we determined whether similar nonnative misfolding occurs if the protein is reduced in the absence of denaturing agent; under these conditions, the protein should only populate states that are low in free energy under native conditions. Indeed, this preparation produces a nearly identical intrinsic fluorescence red shift to that observed in the Red $\to$ Ox construct (Fig. 1 c) as well as increased bis-ANS fluorescence (Fig. 1 d) and a monotonically decreasing melting curve (Fig. 1 e). These results, which are observed whether or not this reduction under native conditions is followed by re-oxidation (samples referred to Red $\to$ Ox native and reduced native, respectively), rule out the possibility that these nonnative states correspond to trapped high-energy intermediates and indicate that misfolding is irreversible upon disruption of disulfides. Interestingly, the “magnitudes” of the bis-ANS and intrinsic fluorescence in these Red $\to$ Ox native and reduced constructs differ from those in Red $\to$ Ox (Figs. 1 d and S1 c), suggesting that these protocols may lead to slightly different nonnative structures; in the case of reduced native, such a difference is expected given that the protein was not re-oxidized.

RBD oxidative refolding produces a stable nonnative intermediate that is highly aggregation prone

Our observation that the refolded, re-oxidized RBD is more hydrophobic than the native state and shows molten-globule-like behavior hints that this species may be prone to aggregation. To test this, we ran our Red $\to$ Ox species, as well as native RBD, on a size exclusion chromatography column (Fig. 2 a). We find that, whereas the native RBD elutes primarily as a monomer as expected (MW $\sim 35$ kDa accounting for N-linked glycans), the Red $\to$ Ox sample largely elutes near the column void volume suggesting it is forming large oligomers whose molecular weight exceeds $\sim 100$ kDa (the exclusion limit for our Superdex 75 column.) Notably, centrifuging this Red $\to$ Ox sample before loading does not produce any visible precipitate, suggesting these oligomers are soluble. We observe a similar void peak if protein is prepared via the Red $\to$ Ox native preparation (Fig. S2), and this sample shows fluorescence upon incubation with ThT—an indicator of amyloid-like aggregation—to a significantly greater extent than the native protein (Fig. S3 a and b). Interestingly, we note that our standard Red $\to$ Ox preparation in which RBD is denatured during reduction does not lead to appreciable ThT fluorescence, perhaps because residual denaturant during refolding affects the aggregates’ structure such that they are no longer amyloid-like (Fig. S3 c). Together, these findings indicate that the oxidative refolding causes RBD to populate a molten globule intermediate that is highly prone to forming soluble aggregates.

We next sought to relate this observed aggregation to the folding landscape of RBD monomers. Our findings thus far are consistent with two possible scenarios (Fig. 2 b).

• (1)The native state represents the free energy minimum in the reduced RBD’s folding landscape, and thus most molecules spontaneously adopt a native-like state upon refolding. However, transient structural excursions into a high-energy intermediate, which occur before the native state can become locked in place via disulfide formation, lead to aggregation. Disulfide formation then further locks the monomers in this nonnative state.
• (2)The aggregation-prone monomeric intermediate (likely an ensemble of rapidly interconverting molten globule structures) represents the minimum free energy state in the “reduced RBD’s” folding landscape, and thus most of the population spontaneously adopts it upon refolding and then proceeds to aggregate, followed by disulfide formation per scenario 1.

A third scenario, in which the aggregation results from the monomers adopting an energetically unfavorable yet highly kinetically accessible state upon refolding, can be ruled out by our previous finding that reduction under native conditions leads to similar molten-globule-like behavior (Fig. 1 c–e) and substantial aggregation (Fig. S2). Notably, in both of the above models, the fully oxidized, native RBD’s free energy is lower than that of any state on the reduced energy landscape, and thus, oxidation into the native state is treated as an irreversible transition; however, aggregation of the nonnative RBD represents a second irreversible transition that prevents the native state from being reached in our experiments.

To distinguish between scenarios 1 and 2, we directly monitored the kinetics of Red $\to$ Ox refolding by measuring the ratio of fluorescence at two wavelengths—370 nm and 335 nm—as a function of time after manual mixing of denatured protein with refolding buffer. This fluorescence ratio is expected to acquire a value of 0.4 in the native state, whereas in the nonnative state, it is closer to 0.6, and in the denatured, reduced state, it is around 1. Strikingly, the vast majority of the refolding amplitude occurs within the dead time of manual mixing (5 s) with only minimal kinetics afterward (Fig. 2 c). By 10 s, the signal has stagnated at the expected nonnative value of 0.6. These kinetics, which are qualitatively similar if an alternative wavelength pair is monitored (320 and 370 nm, see Fig. S4 a), are much too rapid to correspond to aggregation at the protein concentration used (0.5 μM) This is because the rate of aggregation cannot be any faster than the diffusion-limited association rate, which for most proteins is in the range of ${10}^5-{10}^6{\mathrm{M}}^{-1}{\mathrm{s}}^{-1}$. Protein pairs with evolved complementary charged interfaces, such as acetylcholine and its nicotinic receptor (53) and Barnase and Barstar (54), can associate faster than this owing to electrostatic steering forces, but in the case of RBD, the monomers have a net +7 positive charge at physiological pH. Thus we expect that the ${10}^5-{10}^6{\mathrm{M}}^{-1}{s}^{-1}$ range represents an upper bound on the association rate for RBD, which at submicromolar protein concentrations, translates to a 20- to 200-s timescale. The observation that the kinetics are nearly complete by 5 s and stagnant after 10 s strongly suggests that this fluorescence ratio is not reporting on aggregation at all but rather on nonnative tertiary structure within the monomers. Thus, this nonnative monomer structure is substantially populated even before aggregation, strongly pointing toward scenario 2.

In additional support of scenario 2, we observe similarly rapid kinetics and a similar nonnative final ratio if both denaturation and refolding are performed under “reducing” conditions (Fig. S4 c). This, along with our finding in Fig. 1 c–e that reduction is sufficient to produce a molten-globule-like intermediate, indicates that the folding landscape of the reduced RBD favors nonnative monomeric states. Furthermore upon reducing native RBD with TCEP, we observe a substantial separation between the timescales for intrinsic fluorescence change, which is rate limited by disulfide reduction and is largely complete within 3 h, and aggregation as monitored by ThT fluorescence, whose rate is highly concentration dependent and generally exhibits a multihour lag phase (Figs. 2 d, S4 d, and e). We note that an increase in ThT fluorescence may not be a reliable indicator of the onset of aggregation, as ThT may not bind aggregates until they have reached a certain size and/or undergone ripening. Nevertheless, the observed separation of timescales between the intrinsic fluorescence and ThT kinetics is certainly consistent with our expectation from scenario 2, and it further supports our conclusion that the former probe is most sensitive to conformational properties of RBD “monomers” rather than oligomers.

Nonnative RBD structure shows incomplete disulfide bond formation

Given that RBD can correctly refold in the presence of its native disulfides (Ox $\to$ Ox), whereas disruption and re-formation of these disulfides lead to misfolding (Red $\to$ Ox), we expect that the misfolded, re-oxidized state should exhibit a nonnative pattern of disulfide bonding. To characterize the differences between these constructs, we proteolytically digested refolded RBD under both conditions and performed tandem mass spectrometry on the resulting peptides (Fig. S5 and supporting material). For the Ox $\to$ Ox sample, we obtained substantial peptide coverage in all RBD regions except the heavily glycosylated N-terminus (Fig. S5 a), and we observed disulfide-linked peptides that largely correspond to the expected disulfides from the native crystal structure (shown in Fig. 3 a), with the exception of the 3–28 disulfide, which could not be detected due to the glycosylation of peptides containing cysteine 3 (Fig. S5 c; Table S1). Interestingly, we observed an appreciable degree of disulfide scrambling involving cysteines 58, 192, and 205. In the crystal structure, the former two cysteines are paired, whereas the latter is free, thus indicating that these protein segments may undergo structural dynamics beyond what can be observed in the isolated RBD under crystallographic conditions. We also observe cross-linked peptides, both of which contain cysteine 205, indicating that RBD monomers can form cross-linked dimers via this cysteine. But unlike in the Ox $\to$ Ox construct, we obtained minimal peptide coverage for the Red $\to$ Ox construct (Fig. S5 b), and no cross-linked peptides were observed, likely because this construct’s high aggregation tendency limited digestion efficiency. Thus we turned to alternative methods to characterize differences in the disulfide bonding between these constructs.

One possibility is that the refolded monomers contain fewer disulfide bonds than the native state, whose disulfide bonding pattern is shown in Fig. 3 a. If so, then the additional free cysteines in the misfolded state may mediate disulfide cross-links between aggregated monomer subunits. To test for such cross-links, we ran both our Ox $\to$ Ox and Red $\to$ Ox on a denaturing SDS-PAGE gel either with or without reducing agent (5 mM TCEP) added before loading the respective sample (Fig. 3 b), as well as iodoacetamide to alkylate any free cysteines and prevent disulfide scrambling. Interestingly, we observe that, whereas the Ox $\to$ Ox sample runs as a monomer independent of TCEP, the Red $\to$ Ox sample runs as a ladder of mixed oligomeric species that disappears when TCEP is added; very similar results are observed using Red $\to$ Ox native samples (Fig. S6 a). This TCEP dependence indicates that the observed higher-order oligomers are indeed disulfide linked. We thus conclude that the misfolded monomers must contain at “minimum” one fewer disulfide than the native state, which only has one free cysteine and thus cannot form cross-linked species larger than dimers. We note that, whereas this SDS-PAGE assay reveals a ladder of oligomeric species within the Red $\to$ Ox population including species as small as dimers, the size exclusion chromatography trace in Fig. 2 a reveals exclusively large oligomers whose size exceeds the column void volume. This discrepancy implies that the oligomers are held together by a mix of noncovalent interactions (which are disrupted in SDS-PAGE) and disulfide cross-links (which are maintained intact on a nonreducing gel) and that, moreover, there is a significant dispersion in the degree of cross-linking within the aggregates.

To further characterize the difference in disulfide bonding between the samples, we incubated both in the presence of PEG-maleimide—an alkylating reagent that conjugates free cysteines to a bulky PEG group—while leaving disulfide-paired cysteines intact. By running both on a gel and observing mobility shift, we find that the Red $\to$ Ox sample shows $\sim 3-4$ more PEGylation bands than Ox $\to$ Ox, suggesting that the former contain up to two fewer disulfides than the latter (Fig. S6 b–d). However, the relatively low PEGylation efficiency under our conditions, which we attribute to aggregation and disulfide cross-linking reducing the number of free cysteines, as well as a potentially inhomogeneous starting population render it difficult to obtain a clearly quantifiable difference between the constructs with this assay.

Atomistic simulation model for nonnative state aligns with experimental observations

Next, we investigated whether atomistic simulations can generate a detailed structural model for the RBD nonnative states that is consistent with experimental observations. To this end, we made use of the DBFOLD method—an all-atom enhanced-sampling MC simulation platform and analysis pipeline that is uniquely capable of sampling both native and nonnative folding intermediates not accessible to conventional simulations on reasonable computational timescales (47,48). We start off by running simulations of RBD with or without distance constraints between native cysteine pairs, which mimic the effect of disulfide bonds (see materials and methods). Although our simulation force field is not currently equipped to model RBD’s N-linked glycans, we note that these glycans have been primarily linked to spike protein expression and cellular processing (55), and their absence does not appear to alter RBD structure, ACE2-binding, or immunogenicity (56); likewise we do not expect their omission from simulations to drastically affect our findings. Furthermore, for simplicity, only the crystallographic disulfides shown in Fig. 3 a are included in these simulations with disulfides, and we do not allow for disulfide scrambling involving cysteines 58, 192, and 205 (the latter is not even modeled in our simulations, which are based on the 6YOR crystal structure that excludes this cysteine). This simplification is unlikely to produce systematic deviations from experimental conditions as the majority of the observed mass spectrometry intensity associated with these cross-links involves either cysteine 58 and 192 (the crystallographic disulfide) or 58 and 205, which are topologically similar to each other and thus likely to have similar stabilizing effects on the native state.

After the completion of simulation runs, we asked whether the equilibrium folding landscapes with and without disulfide constraints favor different structures. To describe the folding landscape, we generated a contact map of the native RBD structure and subdivided the map into islands of proximal native contacts that are expected to form cooperatively, known as substructures (Fig. 4 a). By computing the free energy of states associated with different permutations of formed/broken substructures, known as “topological configurations” (35,48,58), we can determine the energetics associated with forming and breaking different structural units at equilibrium. Notably, we observe that whereas the folding landscape in the presence of disulfide constraints favors the fully folded state, the unconstrained landscape instead favors a nonnative configuration in which the interfaces between four beta strands are disrupted, namely the strands comprising residues 3–5, 27–30, 58–60, and 190–192 (Fig. 4 b). Thus, these simulations agree with the experimental finding that, in the absence of disulfides, the noncovalent energy landscape favors alternative nonnative states over the native state. To assess whether the disruption of these beta sheets is due to absence of disulfide constraints, we can directly monitor the probability that different cysteine pairs make contact with each other in these unconstrained simulations (Fig. 4 e). Consistent with our gel-based cross-linking and PEGylation assays (Figs. 3 b and S6), we find that the nonnative state typically contains 1–2 fewer cysteine pairs than the native state, and indeed, the two most frequently missing pairs in the absence of constraints (Cys 3–28 and 58–192) occur at the interfaces of the disrupted beta sheets (Table 1). Thus, per simulation, the RBD folding landscape favors conformations that bring together some but not all native cysteine pairs. Interestingly, our simulations predict that the unconstrained folding landscape brings together a “nonnative” cysteine pair, 28–58, with substantially higher probability ($30\%$ ) than either of the native pairs involving these two cysteines, namely 3–28 and 58–192 (which occur with probabilities $2\%$ and $15\%$, respectively). This suggests that the nonnative state may be prone to nonnative disulfide scrambling—a prediction that can be directly tested in future work.

Figure 4.

Atomistic simulations explain experimentally observed structural properties of misfolded RBD. (a) Native contact map for RBD, alongside native structure. Contacts corresponding to the substructures a to j are indicated with different colors and labeled on structure as dashed lines. (b) Free energies of distinct “topological configurations,” defined as the collection of snapshots containing some subset of folded native substructures (defined as in a). Each dot represents one topological configuration, as labeled, with its x position corresponding to its respective number of folded substructures and y position corresponding to its respective free energy relative to the lowest energy state. (c) Example snapshot of fully folded structure (configuration abcdefghij) from the simulation with constraints. (d) Example snapshot of configuration bcdefgi, the lowest energy state from simulations without disulfide constraints. Cysteine pairs CYS46 (cyan) and CYS99 (light green), as well as CYS147-CYS145 (orange) remain in contact with each other, whereas other cysteine pairs corresponding to native disulfides do not form contacts. (e) Probability distribution of number of cysteine-cysteine contacts among snapshots from the simulation without disulfide constraints. (f) Root mean-square fluctuation relative to native state (top plot) of each residue in simulations with disulfide constraints (orange) and without constraints (blue line, cysteine positions indicated with red markers) and difference in solvent-accessible surface area (SASA) between the simulations as a function of residue (bottom plot). Positive $\Delta$SASA indicates the residue is more exposed in the simulation without constraints. The color of each bar represents the hydrophobicity of the residue, quantified by the Miyazawa-Jernigan scale (57). The secondary structure of each residue is indicated below the plots (arrows = beta strands, rectangles = helices, colored as in c and d).

Table 1.

Probabilities that indicated cysteine pairs come together within a beta-carbon distance of $5\mathrm{\AA }$ in full-length RBD simulations with and without disulfide constraints

CYS Pair	With constraints	No constraints
3–28	100%	3%
46–99	100%	89%
147–155	100%	94%
58–192	100%	15%
28–58	$<0.01\%$	29%
46–192	$<0.01\%$	10%

Our atomistic simulations further generate specific structural predictions that may explain why the nonnative state shows molten-globule-like behavior and high aggregation propensity. Namely, we observe that the majority of residues are markedly more dynamic (i.e., show higher root-mean-squared fluctuation with respect to the native state) in the absence of constraints; this discrepancy is especially pronounced at the N- and C-terminal beta strands, but some degree of difference is observed throughout, even in segments without cysteines (Fig. 4 f, top). Furthermore, certain hydrophobic patches are more exposed in the unconstrained simulation, most notably the disrupted C-terminal beta strand but also to some extent the disrupted N-terminal strand as well as core segments around residues 25–30, 55–61, 108–110, and 116–121. This loosened, more exposed protein core may explain the enhanced bis-ANS fluorescence in the nonnative state, and the highly dynamic N- and-C-terminal beta strands may help mediate intermolecular aggregation (Fig. 2 a), ThT fluorescence (Fig. S3, often an indicator of amyloidogenic aggregation between beta strands), and intermolecular cross-linking via the free cysteines (Fig. 3). We note however that, although it shows partial molten-globule-like character, the nonnative state does not behave as a classical molten globule as it contains numerous ordered, well-packed regions in common with the native state. In line with this, we observe that not all hydrophobic residues are more exposed in the nonnative state; in particular, two long stretches containing multiple hydrophobic residues between AAs 62–115 and AAs 119–153 are generally “less exposed” in this alternative state. Interestingly, this latter stretch largely coincides with the disordered ACE2 binding interface, hinting that these functionally important segments may introduce energetic frustration into the protein. In addition to exposed hydrophobics, this binding interface also contains multiple positively charged residues that are closely spaced in the native state, potentially introducing further energetic strain to the native state.

Simulation model predicts that co-translational folding improves RBD disulfide formation and folding efficiencies

Together, our experimental and simulation results indicate that the fully synthesized RBD cannot be refolded from a denatured reduced state as the underlying folding landscape favors nonnative, molten-globule-like states with incomplete/incorrect disulfides and a significant aggregation propensity. Given that many proteins can begin folding co-translationally, we wondered whether disulfide formation during RBD secretion into the ER may help ensure correct disulfides form, ultimately locking the protein into the native state and improving folding efficiency. To investigate this, we used our all-atom DBFOLD algorithm to simulate RBD constructs truncated at various chain lengths shortly after the synthesis of successive cysteine residues. For each length, we assessed the probability that native and nonnative cysteine pairs come into contact (within a beta-carbon distance cutoff of 5 Å) and thus can be oxidized into a disulfide.

From these simulations (Fig. 5 a), we find that most native cysteine pairs show an appreciable probability of coming into contact at intermediate lengths. Perhaps the most salient feature of these profiles is that the 3–28 pairing probability varies nonmonotonically with length, peaking around length 100 ($\sim 20\%$ probability) before declining to less than $5\%$ at the full length (195 residues). This indicates that, whereas the full-length RBD is highly prone to adopting misfolded states that keep this cysteine pair apart, the energy landscape at intermediate lengths is more likely to bring this native pair together. In contrast, “nonnative” cysteine pairs rarely come together at intermediate lengths. The most likely such pair, namely 28–58, rarely occurs before length $\sim 150$, but this probability increases at later chain lengths, and by 195 residues (the full length), this nonnative pair probability exceeds that of the native 3–28 pair (other nonnative pairs are almost never observed at all; see Table 1).

A key parameter that will affect the efficiency of co-translational disulfide formation is the ratio between the characteristic timescale for translation of the domain ($\sim 20-100$ s, assuming a translation rate of 1–5 AAs/s (59)) and the characteristic timescale of disulfide formation by PDI, which has been shown to be capable of catalyzing disulfide formation co-translationally (28,29,30), provided a cysteine pair is in contact. To quantify the effect of this parameter, and the advantage to co-translational folding more generally, we constructed a kinetic model of co-translational disulfide formation (Fig. 5 b, materials and methods). The key assumptions underlying this model are the following:

• (1)At any given length L, the rate at which a given disulfide d forms is independent of whether other disulfides are present. This is justified by control simulations that show that, at a given length, the presence of all disulfides that could have formed before that length does not significantly affect probability of additional cysteines coming into contact (Fig. S9).
• (2)A disulfide bond can only form if the respective cysteines first come into proximity of each other; here, we use a distance cutoff of 5 Å. After such a cysteine-cysteine contact is formed, the pair is oxidized at a rate $k_{ox}$ that is assumed to be insensitive to the identity of the cysteines. In the cell, PDI is expected to catalyze this oxidation step.
• (3)Disulfide formation is assumed to be an irreversible process. Although in reality disulfide shuffling can occur, this process is typically quite slow compared to translation timescales and is thus neglected here.
• (4)The rates of structural fluctuation that bring cysteines into contact are assumed to be much faster than the oxidation rate $k_{ox}$. This assumption is in line with previously measured co-translational folding rates on the millisecond timescale (60) compared to minute timescales for PDI catalytic activity (24). Thus, combining this assumption with the first one, we can express the effective rate of formation for disulfide d as follows:

$$k_d^L=\frac{P_d^L}{1-P_d^L}{k}_{ox}\text{,}$$ (3)

where $P_d^L$ is the probability that the cysteine pair corresponding to d makes contact at length L, per Fig. 5 a, and the above ratio thus represents the equilibrium constant for these cysteines coming together. We can then write down a master equation at length L that governs the time evolution for the probability of occupying states with different permutations of formed disulfides:

$$\frac{d}{dt}{P}^L\left(t\right)={\mathbf{M}}^{\mathbf{L}}{P}^L\left(t\right)\text{,}$$ (4)

where $M_L$, the transition matrix at length L, incorporates the (irreversible) disulfide formation transitions with rates obtained as described above. We note that in addition to proximity, the rate of disulfide formation is also affected by the cysteine pair’s surface accessibility (30); however, we observe that, among snapshots where a given cysteine pair is formed, the pairs’ total solvent-accessible surface area (SASA) shows minimal dependence on chain length, and the SASA values of the two competing cysteine pairs that are most likely to determine folding outcome (3–28 and the nonnative 28–58) fall within a relatively narrow ($\sim$ twofold) range (Fig. S10). Thus, for simplicity, we neglect the effect of accessibility on catalytic rates.

For a given ratio of the characteristic oxidation rate to the translation rate for the full domain ($k_{ox}/k_{trans}$, a free parameter), we can solve for the probability, as a function of time, that species with different numbers of cysteine pairs have formed, either assuming that co-translational disulfide formation occurs or that disulfides cannot form until synthesis is complete. To begin, we set $k_{ox}/k_{trans}=5$. Assuming a cysteine pairing probability of $\sim 20\%$, this would yield an effective disulfide formation rate equal to the RBD translation rate, which is reasonable given that both processes should occur on minute timescales. Under this condition, we observe that co-translational folding significantly speeds up the initial formation rate of product with all four correct disulfides formed (Fig. 5 c, line labeled “All native disulfides”); for example, the amount of native product produced in an amount of time equivalent to twice the RBD translation time is nearly 10-fold larger when co-translational folding is permitted (inset in Fig. 5 c). In addition to decreasing the time needed for native product formation, this speedup may come with the added benefit of reducing the population lost to competing aggregation and degradation reactions, which are not accounted for in our kinetic model. Furthermore, the final native yield is roughly twofold higher when co-translational folding is permitted, and the amount of nonnative product (containing the 28–58 disulfide) is reduced. This is because, with co-translational folding, the nascent chain can take advantage of intermediate chain lengths at which the native 3–28 pair is significantly more likely to form than the competing 28–58 pair—a trend that is reversed at the full length where the 3–28 pairing probability is extremely low and nonnative disulfides are more likely (Fig. 5 a). In Fig. S11 a, we vary the oxidation/translation ratio over a reasonable range and show that this benefit to co-translational folding is robust to the exact value, although the benefit is naturally increased at higher oxidation rates. We note that this trend is somewhat more sensitive to another unknown parameter—the simulation temperature (which, in the context of our coarse-grained potential, cannot be precisely translated to real temperature)—but some degree of benefit is predicted at a wide range of temperatures below the melting transition (Fig. S11 b).

Discussion

By combining in vitro experiments with atomistic simulation, our work generates a unique, detailed molecular picture of the SARS-CoV-2 co- and posttranslational oxidative folding landscapes. These findings advance our understanding of how protein folding is coupled to disulfide formation more broadly. In some proteins, the formation of native secondary and tertiary structure precedes the disulfide bonds, and the underlying folding landscape strongly favors native over nonnative cysteine pairings; such proteins fold via a so-called “structured precursor mechanism” (1,18,19,20,23,28,30). But this mechanism is at odds with our finding that RBD can only be refolded if its native disulfides are maintained intact during denaturation, whereas disruption of these disulfides leads to misfolding and nonnative disulfide formation. Thus, the fully synthesized RBD more closely resembles proteins that fold via alternative mechanism, known as the “quasi-stochastic model,” in which the folding landscape intrinsically favors nonnative states with significant disulfide heterogeneity (22,24,25,26,27). Our simulation model agrees with our experimental finding that the RBD nonnative state contains fewer disulfides than the native state, and additionally it predicts the specific native cysteine pairs—namely 3–28 and 58-192—that are most likely to be disrupted, as well as the possibility of the nonnative 28–58 disulfide. We attempted to experimentally validate these predicted disulfide patterns using mass spectrometry but were unable to detect any cross-linked peptides for the highly-aggregation-prone Red $\to$ Ox sample; future work is needed to improve digestion conditions for this highly-aggregation-prone construct while taking care to minimize disulfide scrambling that may occur under optimal digestion conditions. Nonetheless, our mass spectrometry experiments on the natively folded Ox $\to$ Ox construct revealed some interesting structural features involving the isolated RBD that may not have been predicted from the crystal structure, including dynamic disulfide exchange between three cysteines (58, 192, and 205) and intermolecular cross-linking between natively folded monomers via cysteine 205. Although these findings pertaining to isolated RBD may not hold in the full spike protein, where cysteine 205 is disulfide bonded with a downstream partner, these results suggest that once digestion conditions are improved, the protocol could be used to successfully shed light on nonnative disulfides in the Red $\to$ Ox construct, including the possibility of dynamic disulfide shuffling predicted by simulation. Our experimental and simulation results further indicate that the nonnative Red $\to$ Ox states are molten-globule-like and highly prone to aggregation; similar aggregation-prone molten globule states have been observed in certain destabilizing mutants of the enzyme DHFR (51,61), as well as other systems. This aggregation tendency further suggests that the presence of other spike protein domains may further exacerbate the folding deficiencies observed here for isolated RBD, given that interdomain misfolding and nonnative interdomain disulfide formation are oft-observed risks (23,41,62,63). These effects are not included in our kinetic model and may even further heighten the benefit of co-translational folding.

For proteins with a significant nonnative disulfide propensity, co-translational folding during secretion into the ER may help improve the speed and efficiency of native disulfide formation and folding. Our atomistic simulations suggest that the co-translational oxidative folding of RBD proceeds via a hybrid between the structured precursor and quasi-stochastic models discussed in the previous paragraph. On the one hand, our simulations predict that at intermediate translation lengths, the folding landscape favors structures that bring together native disulfide pairs over nonnative pairs, in line with the structured precursor model. Similar behavior has been observed in certain proteins such as influenza hemagglutinin (31) and the hemagglutinin-neuraminidase glycoprotein of Newcastle disease virus (26), although in other cases, co-translational folding intermediates may not be able to form disulfides at all (27,34) or may even be prone to nonnative disulfide formation (23,34) These differences between systems may relate to disulfide topology—where most native disulfides in RBD involve cysteines that are relatively close in sequence—the relative stabilities of native-like versus nonnative co-translational folding intermediates, and other factors that have yet to be fully understood (30,34). But despite the RBD’s bias toward native co-translational disulfide formation, our simulations predict that the co-translational intermediates are not fully ordered, and native secondary/tertiary structure does not fully form until after synthesis, provided the disulfides have begun forming co-translationally (Fig. S8). In this sense, the RBD’s co-translational mechanism more closely resembles the quasi-stochastic model, in which completion of folding typically follows disulfide formation. These structural predictions can be tested in vivo using pulse-chase assays (31,32,33) or using in vitro translation systems alongside pulse-labeling techniques such as hydrogen-deuterium exchange (64). Furthermore, our simulations predict that disrupting co-translational oxidation of the early-forming 3–28 disulfide bond in particular (e.g., via circular permutation) will drastically decrease the native folding efficiency and increase the population of misfolded proteins, some fraction of which will contain the scrambled 28–58 disulfide.

Our results here may also shed further light on previous findings that reduction of RBD’s disulfides leads to altered secondary and tertiary structure, reduced stability, and impaired ACE2 binding (43,44). Consistent with this, it is known that expressing RBD in E. coli leads to a disulfide-lacking species that is highly prone to aggregation and must be refolded from inclusion bodies (56,65). One possible mechanism to explain these observations is that, in the absence of correct disulfides, RBD is more prone to structural fluctuations that expose hydrophobic segments, leading to aggregation (Scenario 1 in Fig. 2), as observed in disulfides-containing proteins such as hen lysozyme and SOD1 (66,67). But our work points to a more unique mechanism in the case of RBD—namely that the native structure represents a metastable state that is disfavored thermodynamically in reduced, disulfide-free energy landscape. (Scenario 2 in Fig. 2). Thus, RBD cannot efficiently fold under standard conditions, and it is expected to benefit significantly from nonequilibrium mechanisms that promote folding under kinetic control. Although this work emphasizes the role of co-translational folding, we note that a previous study showed that RBD can successfully be expressed in E. coli—an expression system with limited disulfide formation capacity—and subsequently oxidized in suitable buffer, provided the RBD is fused to a solubility-enhancing peptide tag (56). However, this finding does not contradict our results here, which reveals that the nonnative state corresponds to the free energy minimum in the “absence” of RBD’s native disulfides (Fig 2 b, Scenario 2). But in the presence of these disulfides, the native state indeed corresponds to the energetic minimum. Thus, given enough time, as well as conditions that sufficiently suppress aggregation (e.g., low protein concentration or the presence of a solubility-enhancing peptide tag), the oxidized RBD will eventually attain its native state—a process that will likely involve some scrambling of nonnative disulfides. Aggregation may also be sufficiently suppressed if resolubilized RBD inclusion bodies are refolded dropwise in oxidizing buffer, as has been previously shown (65). But in the crowded cellular environment, we expect it is impossible to suppress aggregation and interdomain misfolding to a sufficient degree that the native state can be reached autonomously. Thus, co-translational folding is a more likely mechanism for achieving efficient native folding in vivo.

A key outstanding question left by this work is what factors contribute to the stability gap between the metastable RBD native state and the low-energy nonnative state we observe, and could these factors paradoxically be linked to RBD’s function? One pertinent observation from our atomistic simulations is that disordered segments in RBD tend to exhibit higher flexibility in the nonnative state than in the native state, indicating that conformational entropy may be important in stabilizing the former (Fig. 4 f). Yet these disordered regions also form a key region of the ACE2 binding interface (45,68,69,70,71,72) and are conserved across evolution (72), strongly pointing toward a functional role; perhaps this disorder increases the domain’s avidity toward ACE2 by allowing for multivalent interactions (73). These disordered regions are further enriched with positively charged residues that allow the RBD to form strong ionic interactions with the negatively charged binding interface (45). But this cluster of like charges may introduce energetic frustration to the native state, and indeed, we observe that nonnative structures tend to be more expanded with these charges further apart. Additional studies are needed to confirm these physical tradeoffs between folding and function and to determine how they may shape the evolution of SARS-CoV-2 variants with different ACE2 binding affinities. We further note that, in addition to our finding that the RBD structure represents a metastable state, it is known that the prefusion conformation of the spike protein as a whole is likewise kinetically trapped (74). This metastability allows the spike to undergo a functionally crucial conformational change into the lower-energy postfusion state. However, in the case of the RBD, we do not anticipate our observed metastability to be functionally relevant as the lower-energy state is highly aggregation prone.

Finally, our work may lay the groundwork for future therapeutic directions against SARS-CoV-2. Recent studies have suggested that abnormal blood clotting, caused by fibrin aggregation, may be linked to various COVID-19 long-term sequelae (75,76). Notably, such fibrin aggregates have been shown to contain spike protein molecules, hinting that misfolded or aggregated spike protein can nucleate fibrin aggregation (77). It would be interesting to examine whether variability in redox metabolism across patients could modulate the folding and disulfide formation efficiency in spike protein and thus affect subsequent aggregation and clinical outcome. This knowledge can also be used to improve efficiency and reduce side effects of vaccines that use the RBD as an antigenic agent. In the alternative scenario that the misfolded RBD state is clinically benign, our study can inspire the design of folding inhibitor drugs that selectively stabilize nonnative structures and promote RBD misfolding, thus inhibiting virus-host membrane fusion and replication.

Author contributions

A.B., K.P., and E.I.S. designed research. A.B. and K.P. performed research and analyzed data. A.B., K.P., E.S., and E.I.S. wrote the manuscript.

Editor: Robert Best.

Supporting citations

Reference (78) appears in the supporting material.