Discovery of nanobodies against SARS-CoV-2 and an uncommon neutralizing mechanism

SARS-CoV-2 and its variants continue to threaten public health. The virus recognizes the host cell by attaching its Spike receptor-binding domain (RBD) to the host receptor ACE2. Therefore, RBD is a primary target for neutralizing antibodies and vaccines. Here we report the isolation, and biological and structural characterization of two single-chain antibodies (nanobodies, DL4 and DL28) from RBD-immunized alpaca. Both nanobodies bind Spike with affinities that exceeded the detection limit (picomolar) of the biolayer interferometry assay and neutralize the original SARS-CoV- 2 strain with IC50 of 0.086 μg mL-1 (DL4) and 0.385 μg mL-1 (DL28). DL4 and a more potent, rationally designed mutant, neutralizes the Alpha variant as potently as the original strain but only displays marginal activity against the Beta variant. By contrast, the neutralizing activity of DL28, when in the Fc-fused divalent form, was less affected by the mutations in the Beta variant (IC50 of 0.414 μg mL-1 for Alpha, 1.060 μg mL-1 for Beta). Crystal structure studies reveal that DL4 blocks ACE2-binding by direct competition, while DL28 neutralizes SARS-CoV-2 by an uncommon mechanism through which DL28 distorts the receptor-binding motif in RBD and hence prevents ACE2-binding. Our work provides two neutralizing nanobodies for potential therapeutic development and reveals an uncommon mechanism to design and screen novel neutralizing antibodies against SARS-CoV-2.


INTRODUCTION
A key step for SARS-CoV-2 infection is the molecular engagement between the receptor-binding domain (RBD) on the Spike protein (S) and the human receptor angiotensin-converting enzyme 2 (ACE2) [1][2][3][4] . S is a heavily glycosylated trimeric protein that in the pre-form contains 1273 amino acid residues. Upon cleavage by host proteases, S breaks down to two subunits S1 and S2 at a region near residue 685. RBD (residues 330-526) is contained in the S1 subunit 4 . In the pre-fusion state, S exists in multiple conformations regarding the relative position of RBD to the rest of the protein.
In its 'closed' conformation, all three subunits are very similar and the receptor-binding motif (RBM) of the RBD is buried by adjacent N-terminal domains (NTDs) of S1. The RBD in the closed S is referred to as the 'down' conformation and they are incompetent to engage with ACE2. In the 'open' state, one, two, or all three RBDs could assume the 'up' conformation, exposing the RBM to engage with ACE2 1,2,5,6 . Reflecting the importance of ACE2-RBD binding in viral infection, hundreds of existing neutralizing antibodies target this event by direct blockage, steric hindrance, or locking the RBDs in the 'down' conformation 7 .
The single-chain camelids-derived antibodies possess attractive features 8 . The variable region of the heavy-chain antibodies is referred to as nanobodies owing to their small sizes (~14 kDa). Despite having a single chain, nanobodies can target antigens with comparable selectivity and affinity to conventional antibodies. Being small, nanobodies are ultra-stable, relatively easy to produce (in microbial) with low costs and high yields, and amenable to protein engineering such as fusion in various forms. Such fusion can result in improved potency -binding affinity and neutralizing activity can increase by hundreds to thousands of fold [9][10][11] . In addition, nanobodies that recognize non-competing epitopes can be conveniently fused to make biparatopic nanobodies that are potentially more tolerant to escape mutant strains 9,10,12 . The heat stability of nanobodies opens the possibility of using them as inhaling drugs for respiratory diseases 8 (and indeed potentially for SARS-CoV-2 as demonstrated in hamsters 13 ) and offers convenience in storage and transport. In the past months, dozens of neutralizing nanobodies against SARS-CoV-2 have been reported [10][11][12][13][14][15][16][17][18][19][20][21] .
A challenge in developing neutralizing antibodies and vaccines against viruses is their ability to mutate. In particular, mutations in RBD that retain its structural integrity and function (ACE2-binding) may escape neutralizing antibodies by altering the binding surface either in composition or in conformation, or both [22][23][24][25] . In the past months, strains such as the lineage B1.1.7 and B1.351, referred to as the UK and South African (SA) variant based on the region they first emerge, or the Alpha and Beta strain by the recent recommendation from the World Health Organization, have caused outbreaks and concerns about how they could change the course of the pandemic because their high virulence and their general resistance against antibodies and vaccines that were developed using previous strains 26,27 . Indeed, the Delta strain which is the prevalent strain in the recent global outbreaks 28 contains both the mutations seen in UK and SA variants although it is yet to be established how much this 'double mutant' compromises the protective effect of the current vaccines. Of relevance, a laboratorygenerated mutant, E406W 29 , could escape a Regeneron cocktail that contains two mAbs recognizing different epitopes on RBD. Given the large number of active cases, it is reasonable to assume that more escape mutants are almost certain to emerge. Due to the lag phase between outbreaks caused by new mutants and the development of vaccines/mAbs against the mutants, it is of vital importance to have different antibodies and to test and develop strategies to identify antibodies with broad reactivity.
Here, we report the selection and structural characterization of two RBD-targeting neutralizing nanobodies (dubbed DL4 and DL28) isolated from immunized alpaca.
DL4 binds the Spike tightly at the RBM with a KD below the detection limit of our biolayer interferometry assay. DL4 neutralizes the Alpha but not the Beta variant. By contrast, DL28 recognizes RBD at a region adjacent to RBM and is less affected by the mutations in the Beta variant. Structural characterizations rationalize their variable potency against different SARS-CoV-2 strains and suggest an unreported neutralization mechanism by which DL28 distorts the RBM and diminishes ACE2-binding. Our work adds more evidence that RBM-antibodies are more prone to escape mutants and identifies nanobodies and its associated epitope for therapeutic development against SARS-CoV-2 mutants.

Isolation of a high-affinity neutralizing nanobody from immunized alpaca
An adult female alpaca was immunized four times using recombinantly expressed RBD. ELISA test of sera showed an antibody titer of ~1 × 10 6 after four rounds of immunization compared with the pre-immunization sample. mRNA isolated from peripheral blood lymphocytes of RBD-immunized alpaca was reverse-transcripted into cDNA for the construction of a phage display library (Fig. 1A). Three rounds of solution panning were performed with increasingly stringent conditions and an off-selection step to screen high-affinity nanobodies. Subsequence screening using ELISA and fluorescence-detection size exclusion chromatography (FSEC) 10 identified binders with ELISA signal that is at least 3 times higher than a control nanobody, as well as the ability to shift the gel filtration peak of fluorescently labeled RBD at 0.5 M (Fig. 1A).
We identified 28 unique clones as positive clones, among which DL4 was first chosen based on its ability to cause earlier elution of RBD in FSEC (Fig. 1B) and its exceptional binding kinetics (Fig. 1C). The binding affinity between DL4 and S exceeded the detection limit for the biolayer interferometry assay on an Octet system, reporting a KD of <1 pM and a slow koff of <1.0 × 10 -7 s -1 (Fig. 1C). A neutralization assay using SARS-CoV-2 pseudotyped particles (pp) bearing the S from the original Wuhan strain displayed an IC50 of 0.086 g mL -1 for DL4 (Fig. 1D). Table S1 summarizes the sequence and neutralizing activity of DL4 and all other nanobodies in this study.

Structural characterization of the DL4 epitope
To accurately characterize the epitope of DL4, we crystallized the DL4-RBD complex in the space group of P22121 and solved its structure to 1.93 Å resolution by molecular replacement using published RBD and nanobody structures as search models.
The structure was refined to Rwork / Rfree of 0.1973 / 0.2351 with no geometry violations (Table S2). Each asymmetric unit contains two DL4-RBD complexes which are highly similar with C RMSD of 0.207 Å. Chains A and B are used for structure description.
Specifically, the shared site includes 15 residues, some of which, such as Gln493' and Glu484' are key residues for both the receptor-and DL4-binding. Consistent with the structural observation, pre-incubation with DL4 completely blocked the binding between ACE2 and RBD (Fig. 3C).
The DL4 structure was also aligned to the S to assess RBD-binding in the context of the trimer structure. As shown in Fig. S1, the DL4-epitope is well exposed and no  before incubated with a DL4-containing solution with (blue) or without (red) ACE2. As a control, the ACE2-RBD binding profile (black) was recorded using the same procedure without DL4 on a biolayer interferometry (BLI) system.

Structure-based design improved DL4's potency
Next, we set to engineer DL4 for higher neutralizing activity. Avidity effects are commonly exploited for nanobody engineering 10,33 and we also constructed the Fc version of DL4. Unlike those in previous reports, however, the Fc fusion did not increase neutralizing activity, displaying an IC50 of 0.132 g mL -1 , which, by molarity (3.4 nM), was similar to that of the monomeric DL4 (5.3 nM).
Previously, we have designed gain-of-function nanobody mutations based on structural information to increase binding affinity and neutralizing activity 10 . This approach was used again for DL4. Analyzing the DL4-RBD structure reveals that His56 from CDR2 is located in a hydrophobic microenvironment (Fig. 4A) and does not contribute to hydrogen bonding (Fig. 2C). To match the hydrophobic patch, His56 was mutated to Phe, Tyr, and Trp. Similarly, Gln101 in CDR3 was also mutated to the three aromatic residues to match the hydrophobic patch on the RBD made by Tyr421', Leu455', Phe456', Try473', Tyr489', and the hydrocarbon portion of Lys471' (Fig. 4B).
In addition, the G100E mutant was designed to introduce a possible salt bridge with Lys417' or the nearby Arg403'. Subsequent neutralizing assays identified H56Y, Q101F, and G100E as gain-of-function mutants with IC50 values of 0.133, 0.098, and 0.084 g mL -1 , respectively (the Fc-version was used, Fig. S2A). Consistently, the triple mutant showed a 3-fold increase of neutralizing activity, with an IC50 of 0.038 g mL -1 (0.49 nM) (Fig. 4C).

DL4 neutralizes the Alpha potently but neutralizes the Beta variant poorly
A challenge in antibody and vaccine research against SARS-CoV-2 is its ability to evolve escape mutants. During the study, two major more infectious variants, the lineages B1.1.7 (Alpha) and B1.353 (Beta), were reported. The Alpha strain contains the N501Y mutation in the RBD and the Beta strain contains two additional mutations, K417N and E484K.
Although Ans501' is in the vicinity of the CDR1, it does not form hydrogen bonds with DL4 (Fig. 2B). Therefore, mutation of Ans501' is not expected to affect DL4-RBD binding, at least directly. In addition, a tyrosine replacement appeared to be compatible with the local hydrophobic patch consisting of Phe28/29/31; and Tyr501' may even form a hydrogen bond with Glu30 ( Fig. 5A). Therefore, it was expected that DL4 should remain competent against Alpha. This was indeed the case for both DL4 and the triple mutant H56Y/G100E/Q101F; they displayed equal or slightly higher neutralizing activity against Alpha compared to the original Wuhan strain (Fig. 5B).
The Beta variant contains a lysine replacement of Glu484', a residue that forms a key salt bridge with Arg50 in CDR2. The E484K mutation would not only eliminate the salt bridge, but also introduce charge-charge repulsion with Arg50. Thus, the Beta variant was expected to escape DL4. This was also confirmed by the neutralizing assay using both Fc-DL4 and the triple mutant (Fig. 5B). Interestingly, two DL4 mutants (R50E and R50D), designed to restore the salt bridges with Lys484' in the Beta variant RBD, could not neutralize the Beta variant.

Identification of a nanobody capable of neutralizing the Beta variant
One of the focuses in the research of SARS-CoV-2 neutralizing antibodies is to identify antibodies with broad reactivity. To this end, we re-screened clones and obtained a nanobody (dubbed DL28) that showed weak neutralizing activity against the Beta variant. Increasing avidity by Fc-fusion improved the neutralizing activity, with an IC50 of 1.06 g mL -1 which is ~2 fold of that for the original Wuhan strain (Fig. 5C).
Similar to DL4, DL28 could bind to the S protein with ultra-high affinity -its KD was also below the detection limit of the BLI assay (Fig. 5D).

Structural characterization of the DL28 epitope
To characterize the epitope, we also crystallized the DL28-RBD complex. The crystals belong to the space group of P6522 and diffracted to 3.0 Å at the synchrotron.
The structure was refined to Rwork / Rfree of 0.2264 / 0.2467 with no geometry violations ( Table S2). The asymmetric unit contains two copies of complexes that are very similar (C RMSD of 0.509 Å). The chains A/C are used for structure description.
DL28 binds RBD at one side of the high-chair shape with a buried surface area of 986.3 Å 2 (Fig. 6A). In addition to three CDRs (CDR1, 41.9 Å 2 ; CDR2, 195.4 Å 2 ; CDR3, 369.2 Å 2 ), the framework region also contributed significantly to the binding with a buried surface area of 379.8 Å 2 (~40% of the total). Characteristically, most interactions are contained in CDR3 and only one residue in CDR1 is involved in the binding ( Fig.   6B-6E). Overall, the interaction involves 12 hydrogen bonds and a π-π interaction between Phe47 and Phe450'.
Similar to DL4, aligning the DL28 structure to the S structure reveals no clashes for DL28 in binding with the 'up'-RBD, and only minor clashes with the NTD from the clock-wise subunit when binding with the 'down'-RBD (Fig. S3).

DL28 impairs ACE2-binding mainly by RBM distortion
Cross-competition binding assays showed that DL28 also blocked receptor binding to near completion (Fig. 7A). Aligning the DL28-RBD structure to the ACE2-RBD 31 revealed that DL28 and ACE2 approach RBD at the opposite sides of the 'seat' region. Unlike DL4, only minor clashes were observed between aligned DL28 and ACE2. Specifically, Gln44 of DL28 would clash with the mainchain of the ACE2 helix 20-52 (subscript numbers refer to the start-end residues) (Fig. 7B), which contains most of the key receptor-RBD interactions 31 . Mutating Gln44 to glycine resulted in a slight increase in neutralizing activity (Fig. S4, Table S2). Further mutation of the adjacent Lys43 to glycine resulted in a somewhat weaker activity for ACE2-blocking ( Fig. 7C) and neutralization (Fig. S4, Table S2). Possibly, the tri-glycine motif (together with Gly42) introduces structural instability to the nanobody framework and affects the orientation of the CDRs for tight binding. The results suggest that the steric hindrance is not the main factor for neutralization. (iii) for RBD-binding. A sensor coated with RBD was first saturated with DL28 before incubated with a DL28-containing solution with (blue) or without (red) other antibodies.
As a control, the binding between RBD and other antibodies (black) was recorded using the same procedure without DL4 on a biolayer interferometry (BLI) system.
Mapping the DL28 epitope and the RBM to the RBD reveals that they overlap by four residues, namely Gly446', Tyr449', Glu484', and Gln493' (Fig. 7C). However, none of the four RBD residues interacts with DL28 by sidechain (Gly446' has no sidechain). Gly446' and Gln493' are only in the proximity without specific hydrogen bonds with DL28; and although Tyr449' and Glu484' form hydrogen bonds with Arg45 and Tyr58, the interactions only involved mainchain of Tyr449'/Gln493'. In addition, both the mainchain and sidechain of the four RBD residues showed negligible differences in conformation between the ACE2-and DL28-bound forms (Fig. 7D).
Therefore, the RBD conformation at this overlapping region appears to be compatible for simultaneous binding with ACE2 and DL28. The analysis also supports the abovementioned idea of additional mechanisms for DL28's receptor-blocking activity ( Fig. 7A).
Aligning the RBD structures in the receptor-and DL28-binding mode reveals that the backbone of the RBM is largely similar but displays noticeable distortion at the 'backrest' region ( Fig. 4A) between residues Ile472' and Leu492' (Fig. 7E, 7F).
Specifically, DL28 nudges this loop toward the direction of ACE2 by ~ 2 Å. Notably, the pushing by DL28 was not mediated by sidechains or loop regions which may tolerate clashes by assuming alternative conformations. Rather, it was mediated by a 4residue -sheet (56-59) which is part of the stable nanobody framework made of four stacking -sheets (Fig. 7E). The tight binding (Fig. 5D) and the rigidness of the nanobody core should therefore force and lock the loop in the left-ward position. As a consequence, the distorted loop clashes with the 20-52 of ACE2. Specifically, Phe486', Ans487', and Tyr489' from RBD would come into close contact with residues Thr27 and Phe28 in the ACE2 20-52 and Tyr83 in an adjacent helix. Therefore, the conformational change in the 'backrest' loop appears to be incompatible with ACE2binding, unless ACE2 can adapt to the conformational change, which, as reasoned below, would be unlikely.
The ACE2 20-52 lies on top of RBD like a lever. The C-terminal half of the helix binds the 'seat' region of the RBD, and the N-terminal end binds with the 'backrest' region. A 2-Å distortion at the 'backrest' area acts like forces pushing the lever at one end (Fig. 7G). The -helix would have to deform/break to adapt to such a dramatic distortion. However, -helices are generally rigid and 20-52 contains no helixdestabilizing residues such as proline and glycine. Therefore, we propose that DL28 neutralizes SARS-CoV-2 by an 'RBM distortion' mechanism.

RBM distortion does not affect the binding of several RBM-targeting antibodies
It might be expected that the RBM-targeting monoclonal antibodies (mAbs) are incompatible with DL28 because of DL28's ability to distort RBM. However, this was not the case for three such mAbs (that are available to us): REGN10933 34 , CV30 35 , and CB6 36 . Thus, although their epitopes would also be shifted by a similar or more extent compared to the RBM of ACE2 (Fig. 7H), the mAbs bound to RBD in the presence of DL28 (Fig. 7I).
Unlike ACE2, mAbs bind to RBD with CDRs which are usually made of, or connected to the rigid framework by, flexible loops. This may have allowed the mAbs to adapt and to remain bound with RBD. Thus, the results are seemingly contradictory to the ACE2 competition but can be rationalized by the structural analysis.
Monovalent nanobodies with KD values in the low picomolar ranges include two RBMtype binders: Nb20 (10.4 pM) 21 from immunized llama and Nanosota-1C (157 pM) by in vitro maturation of a binder from a naïve llama/alpaca library 37 . Remarkably, both DL4 and DL28 binds S with sub-picomolar affinities. This reinforces the notion that, despite their small sizes, nanobodies can bind antigens with comparable affinity with Fab which is four times in size. One of the reasons, as revealed in this study and previous structural reports, is that the framework region of the nanobodies can also participate in the antigen-binding, thus essentially expanding the binding surface and increasing the number of interactions. In addition, as revealed by DL28, nanobodies may achieve their high affinities by shape complementarity with antigen. Thus, the concave arc formed by CDR3 and the loop at the other end clamps onto the antigen.
This type of interaction has also been observed in the case of nanobodies against the KDEL receptor 38 , the κ-opioid receptor 39 , the folate transporter 40 , and the histo-blood group antigen BabB 41 .
In the literature, increasing avidity generally improves potency, although the effect can vary from dozens to thousands of times 10,11 . Interestingly, the avidity effect for both DL4 was not apparent (Table S1), and that for DL28 was only obvious for the Beta variant (Fig. 5C, Table S1). Mechanistically, fusing with Fc may introduce additional steric hindrance to prevent RBD-ACE2 binding. It may also tether two S trimers to restrict their conformational changes should the two nanobody entities bind to different S. More commonly, avidity is known to increase potency by boosting apparent binding affinity by increasing local concentration and hence a faster kon and a slower koff. In the case of DL4/DL28, the affinity may not be the limiting factor owing to their exceptional binding kinetics. This provides a possible reason for the lack of avidity effect. Despite this, the Fc fusion can increase the potency in vivo by extending the serum half-life of nanobodies from several minutes to several days 10 and thus should be still be useful for therapeutic reasons.
The fact that the DL4(3m) is more potent than DL4 is worth discussing. Thus, despite ultra-high affinity after multiple rounds of immunization, there is still space for rational design. Such practice may be applied to the existing antibodies although the effect of mutations on pharmacological behavior will have to be tested in the cases of therapeutic antibodies.
It was rationalized that the DL4 R50E/D mutants would gain at least some neutralization activity towards the Beta variant by restoring a salt bridge that was probably lost due to the E484K mutation. However, the results showed that mutant was as ineffective as the wildtype DL4 (Table S2). This may suggest that, apart from a sidechain replacement, subtle conformational changes also occur in this region and the changes are compatible with ACE2-binding but not for DL4. It is also possible that other mutations contained in the Beta variant, although not directly involved in the binding with DL4, helped escape the nanobody by allosteric effects. This highlights the challenges in the development of broadly effective neutralizing antibodies against SARS-CoV-2. Antibodies need high affinity to work best, but high affinity requires the epitope arranges in a precise three-dimensional shape. Mutation even remote from the epitope can distort the fine shape and render the antibodies ineffective. Targeting structurally rigid domains is key to develop broadly active antibodies.
Owing to their minute sizes, nanobodies may bind surfaces that are inaccessible for conventional antibodies. In the case of DL4 and DL28, they may be able to bind to the 'down'-RBD given their minor clashes with the 'closed' conformation of S, in addition to binding with the 'up'-RBD (Fig. S1, S3). On the other hand, the small size could mean that the destruction of S trimer by binding, as observed for several conventional antibodies, can be rare 42,43 .
Despite similar binding kinetics, DL28 showed less neutralizing activity (~5 fold) compared to DL4. In addition, the cross-competition for ACE2-binding was complete by DL4 ( Fig. 3C) but not by DL28 (Fig. 7A). We do not yet understand the structural reason for this. Possibly, ACE2 can interact weakly with RBD via the non-distorted part of the RBM. As also reported in the literature, RBM-targeting antibodies are generally more competent for neutralization, i.e., direct completion is generally more efficient for ACE2-blocking 44,45 . However, by targeting the more conserved RBD core region 31 , antibodies that do not aim at the RBM may be less susceptible to escape mutants.
Whether this is the case for DL28 remains to be investigated using replicationcompetent viruses.
Because DL28's epitope only marginally overlaps with the RBM, DL28 may be able to bind RBD in the presence of other RBM-targeting nanobodies and human monoclonal antibodies. Such pairs will allow the development of biparatopic nanobodies to increase tolerance to escape mutants, and DL28's ultra-high affinity could offer great advantages in such applications.
Although we did not test the neutralizing activity of DL4 and DL28 against the Delta strain, DL4 is expected to be a weak neutralizer because the critical Glu484' for DL4-binding is, as in the case for the Beta variant, mutated, although the replacement is glutamine instead of lysine. For DL28, the impact of the Delta mutations is not very clear from the structural analysis. As shown in Fig. S5, Leu452' is a part of a hydrophobic network comprised of Phe490', Tyr351', and Ile468' from RBD, and Phe47, Tyr37, Ile50, and Trp104 from DL28. The Delta mutation L452R would weaken the hydrophobic interactions. Besides, the RBD Arg452' may be repulsed by DL28 Arg45 in the vicinity. On the other hand, however, Arg452 may, depending on the sidechain conformations, form a cation-π interaction with Trp104, and/or form a salt bridge with Asp102. Thus, the exact effect will need to be tested in the future. In the case of weakened neutralizing activity, mutations to accommodate the Delta variant such as W104D can be designed and screened to restore neutralizing activity.
As revealed by the structure of the ACE2-RBD complex, ACE2 engages with RBD mainly through two structurally rigid -helices (20-52 and 55-82). By contrast, the counterpart in RBD is composed of loops lay on top of the core RBD region (Fig. 7B,   7E). This interaction model is perhaps suited for the RBD function. Thus, the structural flexibility of loops allows RBD to assume different, but functionally competent conformations by adjusting its backbone position while allowing escape mutants to evolve. However, the flexible feature also makes it prone to distortion, and the RBM distortion can have detrimental consequences for ACE2-binding, as demonstrated here by DL28. Our work suggests a previously unreported mechanism for SARS-CoV-2 neutralization which could be exploited for developing therapeutic nanobodies.

CONCLUSIONS
We obtained two alpaca nanobodies that target RBD with ultra-high affinities and neutralize SARS-CoV-2 with high potencies. DL4 neutralizes SARS-CoV-2 by direct competition with ACE2 for RBD-binding, whereas DL28 distorts the ACE2-binding site and forces RBD to a conformation incompatible with receptor engagement. DL28 can neutralize the Alpha and Beta variants which are more infectious than the original SARS-CoV-2 strain. Our study provides two tight nanobodies for research and potential therapeutic applications and suggests an uncommon mechanism for SARS-CoV-2 neutralization.

Protein expression and purification -Spike (S)
The polypeptide containing, from N-to C-terminus, residues Met1 -Gln1208 For crystallization, RBD eluted from the Ni-NTA column was desalted using a desalting column, and digested with home-purified 3C protease to remove the Cterminal tags. The resulted tag-free RBD was mixed with nanobodies (see below) at a molar ratio of 1:1.3 and the mix was loaded onto a Superdex Increase 200 10/300 GL column for gel filtration. Fractions containing the complex were pooled, concentrated to 10 mg mL -1 for crystallization.

Protein expression and purification -monovalent nanobodies in Escherichia coli
Monovalent nanobodies were expressed with a C-terminally Myc tag and a hexahistidine tag in E. coli MC1061 cells. Briefly, cells carrying nanobody-encoding pSb-init plasmids 46  buffer-exchanged into PBS using a desalting column (Bio-Rad). mAbs were quantified using their theoretical molar extinction coefficients that are calculated based on the contents of aromatic residues and with absorbance at 280 nM measured using a Nanodrop machine.

Alpaca immunization and antibody titer determination
Purified RBD (1 mL at 2 mg mL -1 ) was mixed with an equal volume of the Gerbu adjuvant (Cat. 3111) by vortexing. The resulted emulsion was injected by the subcutaneous route at ten sites near the bow lymph node in the neck base of an adult female alpaca (3-years old). The immunization process was repeated 3 times (a total of 4 rounds) with 4 days between each injection.
To determine the antibody titer, 3 mL of blood samples before and after each injection were collected. After 2 h at room temperature (RT, 20-25 °C), the clotted sample was centrifuged at 3,000 g for 5 min at RT to collect sera in the supernatant. Wells of 96-well plates (Maxisorp, Nunc Thermo Fisher Scientific) were coated overnight at 4 °C with 100 μL of 2 μg mL -1 biotinylated RBD in TBS (150mM NaCl, 20mM Tris, pH8.0) and blocked with 0.5% bovine serum albumin (BSA) in TBS. After washing five times with TBS, serially diluted alpaca sera were added and incubated for 1 h. After washing, the bound nanobody was detected by HRP-conjugated Goat anti-Alpaca IgG (Cat. S001P, NBbiolab) using Tetramethylbenzidine (TMB) (Merck, Cat.T2885) as a substrate for HRP. ELISA test of sera showed an antibody titer of ~1 × 10 6 after four rounds of immunization compared with the pre-immunization sample.

Phage display library construction and panning
Eighty milliliters of blood were collected from the immunized alpaca in EDTA-

Biolayer interferometry for S-nanobody binding and competitive binding
The binding kinetics was measured by a bio-layer interferometry (BLI) assay using an Octet RED96 system (ForteBio). A streptavidin-coated SA sensor (Cat. 18-5019, Sartorius) was coated with 5 g mL -1 biotinylated nanobodies for approximately 1 min.
The sensor was equilibrated in a nanobody-free buffer for ~30 s, before bathing in solutions containing various concentrations (association) of Spike (analytes) for 120 s (DL4) or 360 s (DL28).
For competition between ACE2 and nanobodies, biotinylated RBD (2 g mL -1 ) was immobilized on an SA sensor by incubating with the sensor in the BLI Buffer in liquid nitrogen before X-ray diffraction data collection.
X-ray data collection and structure determination X-ray diffraction data were collected at beamline BL18U1 at Shanghai Synchrotron Radiation Facility with a 50 × 50 μm beam on a Pilatus detector at a distance of 300 -500 mm, with oscillation of 0.5° and a wavelength of 0.97915 Å. Data were integrated using XDS 47,48 , and scaled and merged using Aimless 49 . The structure was solved by molecular replacement using Phaser 50 with the RBD structure (PDB 6M0J) 31 and a nanobody structure (PDB 5M13) 46 as the search model. The model was built with 2Fo-Fc maps in Coot 51 , and refined using Phenix 52 . Structures were visualized using PyMol.

Data availability
The structure factors and coordinates are available through the protein data bank (PDB) under accession codes 7F5G (DL4-RBD) and 7F5H (DL28-RBD).

Supplementary Information
Table S1 and S2 Fig. S1-S5    The sequences for the mature monovalent nanobodies include 'GSSS' at the N-terminal, and 'AGRAGEQKLISEEDLNSAVDHHHHHH' at the C-terminal which contains a myc-tag (italic) for ELISA and a hexahistidine tag for purification.
b Not determined.
c IC50 value could not be determined owing to weak or the lack of neutralizing activities.      Leu452 was virtually mutated to arginine. The surrounding residues on both RBD (white) and DL28 (light blue with CDR1 in Cyan, CDR2 in magenta, CDR3 in orange, and RBDinteracting framework residues in yellow) are shown as sticks. RBD residues are labeled with a prime. Leu452' is part of a hydrophobic network formed by the shown residues. In the Delta variant, Arg452' may become incompatible with the hydrophobic microenvironment. On the other hand, however, Arg452' may, depending on the sidechain conformations, form a salt bridge with DL28 Asp102.