Fully Human VH Single Domains That Rival the Stability and Cleft Recognition of Camelid Antibodies*

Background: Camelid antibody domains are naturally stable and capable of cleft binding. Results: Protein engineering can endow human antibody domains with such properties. Conclusion: Our strategy does not require undesirable antibody framework changes. Significance: Robust building blocks for human therapeutic applications. Human VH single domains represent a promising class of antibody fragments with applications as therapeutic modalities. Unfortunately, isolated human VH domains also generally display poor biophysical properties and a propensity to aggregate. This has encouraged the development of non-human antibody domains as alternative means of antigen recognition and, in particular, camelid (VHH) domains. Naturally devoid of light chain partners, these domains are characterized by favorable biophysical properties and propensity for cleft binding, a highly desirable characteristic, allowing the targeting of cryptic epitopes. In contrast, previously reported structures of human VH single domains had failed to recapitulate this property. Here we report the engineering and characterization of phage display libraries of stable human VH domains and the selection of binders against a diverse set of antigens. Unlike “camelized” human domains, the domains do not rely on potentially immunogenic framework mutations and maintain the structure of the VH/VL interface. Structure determination in complex with hen egg white lysozyme revealed an extended VH binding interface, with complementarity-determining region 3 deeply penetrating into the active site cleft, highly reminiscent of what has been observed for camelid domains. Taken together, our results demonstrate that fully human VH domains can be constructed that are not only stable and well expressed but also rival the cleft binding properties of camelid antibodies.

Human antibodies are exclusively expressed as paired species, containing both heavy and light chains. In marked contrast to humans, llamas and camels have been shown to also produce non-paired species (1,2). Antigen binding of these heavy chainonly antibodies resides exclusively within a single variable domain (V H H domain) (3). Since their discovery in the mid-1990s (1), camelid V H H domains have attracted considerable interest and investment, due to their potential as building blocks for therapeutic and biotechnology applications (4). The domains are biophysically and structurally well characterized and have been shown to possess high thermodynamic and colloidal stabilities, together with a preference for binding within pockets of protein structure (high cleft binding propensity) (5,6). These features are achieved by means of an extended CDR3 3 loop, which has been shown to be capable of protruding deeply into enzyme active sites and otherwise cryptic epitopes of viruses and G-protein-coupled receptors (7)(8)(9). Residues in the light chain interface differ between in human domains and V H H domains, resulting in a hydrophilic nature of the camelid surface (2). In addition, the extended CDR3 loop of camelid domains is capable of looping backwards, further shielding the V H /V L interface area through interactions with framework residues (5). However, this feature has also been shown to restrict camelid CDR3 diversity, which is highlighted by randomization studies revealing a strong consensus toward naturally occurring hydrophobic sequence motifs (10).
The favorable properties of camelid domains have inspired attempts to generate binders based on isolated human V H domains, although initially with limited success. The main problem holding back the field has not been thermodynamic stability but rather colloidal stability (aggregation propensity) (11,12). Whereas thermodynamic stability can be achieved by the use of stable human V H families (such as V H 3) (13), colloidal stability remains difficult to control (14). This results in overall poor biophysical properties, as indicated by low solubility, expression, and purification yields, and a lack of heat refoldability (11,15).
"Camelization" strategies, as pioneered by Riechmann and co-workers (16), have been developed to improve the biophysical properties of human domains. These strategies have focused on a set of framework residues, which are hydrophobic in humans but are largely hydrophilic in camelid domains (V H H tetrad; positions 37, 44, 45, and 47 (2); all numbering according to Kabat et al. (17)). However, initial attempts at directly trans-planting such residues into a human framework were met with mixed success, with the resulting domains suffering from limited stabilities and structural deformations in NMR experiments (18). Structural changes have also been observed upon engineering of human domains through non-native disulfide links (19). Recent camelization attempts have therefore focused on other interface positions (such as position 39) and on the introduction of novel substitutions at tetrad positions. A study of human single domains derived from the antibody therapeutic trastuzumab (Herceptin 4D5) (20) revealed several such novel stabilizing framework mutations (10). The crystal structure of a soluble 4D5 quadruple mutant was reported in the same study (H35G/Q39E/L45E/R50S; clone B1a). Using an elegant display approach on phage (retention of superantigen binding), the authors also demonstrated compatibility of these mutations with CDR3 diversity, suggesting that domains stabilized in this manner may be capable of antigen binding. Indeed, the same group recently reported the structure of a V H 4D5 triple mutant in complex with vascular epithelial growth factor (VEGF-V H V1a complex) (21).
Although such prior camelization approaches have undoubtedly resulted in improvements of biophysical properties, these engineered human domains nevertheless fall short of what would be considered ideal for many applications, human therapy in particular. For instance, changes of multiple conserved framework residues may well result in the generation of novel B-and T-cell epitopes and an increase of immunogenicity in humans (22). In addition, the introduction of mutations into the V H /V L interface prevents pairing with light chain, which restricts the developability of the domains and limits their use in antibody bispecifics (23). Previously reported structures of human V H single domain in complex with antigen have also largely failed to recapitulate the cleft binding properties of camelid domains (21,24). To overcome these limitations, we decided to bypass the mutation of human framework residues altogether, instead exclusively focusing on the engineering of CDRs. Here we investigate strategies for library design and report biophysical and structural properties of fully human antibody V H domains.

Construction of Synthetic Human V H Antibody Repertoires-
Synthetic libraries were constructed essentially as described previously (25). For the generation of the Garvan I repertoire, single-stranded DNA encoding the model V H domain HEL4 in the phage vector FdMyc was isolated using a QIAprep spin M13 kit (Qiagen). Randomization of CDRs was carried out by combinatorial mutagenesis using partially degenerate oligonucleotides (Table 1). For the generation of the Garvan II repertoire, single-stranded DNA encoding DP47 germ line (V3-23/DP-47 V-segment and JH4b J-segment derived from the immunoglobulin heavy chain locus of chromosome 14) in the phagemid vector pHEN1 was isolated using a QIAprep spin M13 kit (Qiagen). Randomization of the CDR1 and CDR2 regions was carried out by Kunkel mutagenesis using reverse-complemented degenerated oligonucleotides. For CDR3 regions of the Garvan II library, TRIM codon mutagenesis was utilized, and diversity was introduced by splice-overlap extension PCR. After the mutagenesis steps, repertoires were transformed into Escherichia coli TG1 bacteria and validated by DNA Sanger sequencing of random clones (Ͼ100 clones analyzed). Phages were purified by two precipitations with PEG/NaCl either directly from the culture supernatant in the case of FdMyc or after a rescue step using KM13 helper phage in the case of pHEN1.
Heat Refolding on Phage-Aggregation resistance was analyzed by measuring retention of signal after heating in a phage ELISA format utilizing vector FdMyc as described (the pHEN1based Garvan II library was recloned for this purpose) (27). A Maxisorp Immunoplate was coated with Staphylococcus aureus protein A (Sigma) and blocked with 5% milk powder (w/v) in PBS buffer. Single colonies were picked from agar plates and grown overnight at 30°C in 2ϫ TY medium supplemented with 15 g/ml tetracycline. Supernatant was cleared by centrifugation, and phages were biotinylated by the addition of NHS-PEG 4 -biotin (Pierce) to a final concentration of 50 M for 2 h. The reaction was then quenched by the addition of Tris, pH 7.5, to a final concentration of 100 mM and incubated for 1 h. For heat-induced refolding, biotinylated phage supernatant was incubated at 80°C for 10 min, followed by incubation at 4°C for 10 min. Supernatant was added to wells of the blocked ELISA plate and incubated for 1 h. After washes with PBST, bound phage were detected using HRP-conjugated extravidin (Sigma) and 3,3Ј,5,5Ј-tetramethylbenzidine substrate. Finally, the level of retained protein A binding after heat treatment was calculated as a percentage of the untreated phage sample.
Heat Refolding in Solution-Refolding after heat denaturation in solution was determined using size exclusion chroma-tography. For this purpose, purified protein at 10 M in PBS was heated at 80°C for 10 min, followed by cooling at 4°C for 10 min. Samples were analyzed on a Superdex-S75 gel filtration column (GE Healthcare) using an AKTA Purifier (GE Healthcare) chromatography system and PBS running buffer. Recovery of each variant was determined by measuring the area under the curve after heating, expressed as a percentage of the unheated protein sample.
Protein Expression and Purification-Expression and purification of human V H domains were performed essentially as described (28). In brief, genes encoding V H domains were recloned from phage or phagemid vectors into the expression vector pET12a (Novagen) and transformed into E. coli BL21-Gold (Stratagene). Transformants were used to inoculate 2ϫ TY medium supplemented with 4% glucose, 100 g/ml ampicillin, 15 g/ml tetracycline at an A 600 nm of 0.05 in baffled flasks and grown in a shaking incubator at 37°C. At an A 600 nm of 0.5, bacteria were pelleted by centrifugation at 3200 ϫ g, resuspended in 2ϫ TY medium supplemented with 1 mM isopropyl 1-thio-␤-D-galactopyranoside, 100 g/ml ampicillin, 15 g/ml tetracycline and incubated for 36 -48 h at 30°C (with a second induction with 1 mM isopropyl 1-thio-␤-D-galactopyranoside after 24 h). Supernatant was cleared by centrifugation at 10,000 ϫ g, passed through a 0.45-m filtration unit, and pH-adjusted to 7.0 -8.0 by the addition of 1 M HCl. Protein A resin (GE Healthcare) was applied to a gravity flow column (Bio-Rad) and washed successively with 5 column volumes of mQ-H 2 O and PBS. Supernatant was applied to the column by gravity flow and washed with 1 column volume of PBS. Protein was eluted with 100 mM glycine, pH 2.7, and the eluted fractions were neutralized with 100 mM Tris, pH 8.8. Finally, the fractions were dialyzed extensively against PBS, and the purified domains were concentrated using Amicon-Ultra 10-kDa concentration devices (Millipore).
Surface Plasmon Resonance-Binding kinetics were determined by surface plasmon resonance using a Biacore 2000 instrument (GE Healthcare). Biotinylated antigen was immobilized on a streptavidin chip (GE Healthcare), and serial dilutions of purified V H domains in PBS were injected. Association and dissociation curves were fitted using a 1:1 Langmuir binding model and analyzed using BIAevaluation software (GE Healthcare).
Circular Dichroism-Thermal unfolding was measured by circular dichroism (CD) using a J-815 spectrometer (Jasco) in a quartz cuvette (2-mm path length). Protein samples at a final concentration of 20 M in PBS and melting curves were obtained by recording CD signal at 235 nm with a 1-nm bandwidth and 1-s integration time while heating the solutions from 20 to 80°C at 1°C/min. To measure the reversibility of unfolding, signal was measured during cooling from 80 to 20°C at 1°C/min.
Crystal Growth, Structure Solution, Refinement, and Analysis-The V H H04-hen egg white lysozyme (HEL) complex was purified by size exclusion chromatography using 50 mM Tris-HCl, pH 7.5, supplemented with 150 mM NaCl as running buffer (as described above). Peak fractions were collected and concentrated using Amicon-Ultra 10-kDa concentration devices (Millipore). Initial crystallization hits were obtained with the JCSG-plus crystal screen (Molecular Dimensions) using a Mosquito crystallization robot (TTP Labtech) and 96-well MRC2 sitting drop crystallization plates (Swissci). Crystals were grown at room temperature in a hanging-drop format, where 2 l of V H H04-HEL complex (4.37 mg/ml in 50 mM Tris-HCl, pH 7.5, supplemented with 150 mM NaCl) was combined with an equivalent volume of well solution (100 mM sodium citrate, pH 5.4, supplemented with 28% (w/v) PEG 1500 for Protein Data Bank (PDB) entry 4PGJ; 300 mM sodium citrate, pH 5.5, supplemented with 16% (w/v) PEG 3350 for PDB entry 4U3X). Crystals were snap-frozen in liquid nitrogen (100 K), and diffraction data were recorded on beamline MX2 at the Australian Synchrotron. Reflections were indexed and integrated with iMOS-FLM (29), scrutinized for symmetry with POINTLESS (30), scaled with SCALA (30), and imported into the CCP4i software package (31). Reflection data statistics are shown in Table 2. Structures were solved by molecular replacement using PHASER (32). The search models for the V H and HEL components were PDB entries 1OHQ (V H ) and 1ZVY (HEL), both stripped of side chain moieties. Two essentially identical V H -HEL complexes were found in the asymmetric unit, where H04 domains are chained A and C, and corresponding HEL molecules are chained B and D. During refinement, the two complexes were not averaged using non-crystallographic symmetry but were treated independently of each other. Initial rounds of rigid body refinement were followed by B-factor restrained refinement using REFMAC5 (33) and employed translationlibration-screw parameterization (34). During later stages of refinement, water molecules were modeled into appropriate features of difference density. Models were compared with maps and manipulated in real space using COOT (35). Models were scrutinized using the MOLPROBITY validation server (36). Buried surfaces were calculated with PDBePISA (37). Coordinates for the V H H04-HEL structure have been deposited in the PDB as entries 4PGJ and 4U3X, respectively.

RESULTS
Generation of the Garvan I Human V H Phage Repertoire-As a starting point for the design of a first generation repertoire of stable V H domains, we utilized the HEL4 model protein. This human V H domain had been identified by selection on phage, and its structure had been determined by crystallography (27,38). It is characterized by low aggregation propensity (high colloidal stability), heat refoldability, and excellent expression and purification yields (27,39). HEL4 differs from the DP47 germ line domain it had been derived from at over 20 CDR and framework positions (Fig. 1A), and the specific role of individual changes had largely remained unclear (27,38). We initially focused our attention onto CDR3 (Fig. 1B, blue), which provides the majority of binding energy in antibody/antigen interactions. Using site-directed mutagenesis, we introduced diversity at residues 95-100a/c, predominantly utilizing DVK codons to introduce hydrophilic amino acids into CDR3 (Table  1). Such codons have previously been successfully utilized in the generation of paired antibody fragment libraries (40 -42). Transformation into E. coli yielded a library of 3 ϫ 10 9 individual clones.
Characterization of the Garvan I Phage Repertoire-For initial characterization, resistance to heat denaturation was determined using a phage method originally developed by Jespers et al. (27). The method is based on the multivalent display of human antibody domains on phage, coupled with heating to 80°C and cooling of the phage particle. Resistant domains can be captured by the conformation-dependent superantigen protein A, which binds to folded but not to unfolded or aggregated V H domains (27,43). The phage method is an excellent predictor of solution properties of human antibody V H domains, including heat refoldability, expression, and purification yields (44). Human antibody variable domains, such as the DP47 germ line, aggregate under the conditions of the phage method; however, this is not observed for the HEL4 domain, which resists heat-induced aggregation (Fig. 1C).
Initial CDR grafting studies carried out by our group (Fig. 1C) (45) and more recently by others (46) had indicated that the aggregation determinants of HEL4 largely reside within the CDR1 region. Because this region had not been targeted in our randomization strategy (which focused on CDR3 instead), we speculated that the Garvan I library may retain the aggregationresistant properties of the HEL4 domain. Further analyses of randomly selected clones from the naive Garvan-1 repertoire revealed that this was indeed the case and that the majority of clones displayed considerable resistance against heat-induced aggregation on phage (Fig. 1D), with a median level of resistance of 64% compared with 88% for HEL4 and 2% for the DP47 germ line.
Phage Display Selection Using the Garvan I Phage Repertoire-We next decided to assess the potential of the library for the selection of antigen-specific V H domains. For this purpose, a set of representative antigens was chosen, covering a wide range of molecular weights and structural classes (human tumor necrosis factor ␣, murine interleukin 21, human prolactin receptor,  ␤-galactosidase) (Fig. 1E). The antigens were immobilized on solid supports, and single domains were selected by phage display. After three rounds of selection, phage ELISA confirmed that binders had been obtained for all antigens, with 22-57% of all clones displaying antigen binding. After recloning into the pET12a expression vector, monoclonal ELISA was performed on soluble protein, and non-redundant binders were identified by DNA sequencing. This revealed that multiple unique V H domains had been selected for each target antigen.
Characterization of a Stable Human Anti-TNF␣ Single Domain-For further analyses, clone G07 was expressed in large scale, and its biophysical properties and affinity were characterized. Thermal denaturation as monitored by circular dichroism revealed that the domain was fully heat-refoldable ( Fig. 2A,  T m ϭ 63°C), similar to what has been observed for HEL4 and camelid domains (but not for other human V H domains). Moreover, the domain was also well expressed in bacteria (at around 5 mg/liter in shaking flasks, compared with 3.3 mg/liter for HEL4 and 0.9 mg/liter for DP47). Affinity measurements by surface plasmon resonance Biacore revealed moderate affinity interactions with the human tumor necrosis factor ␣ (TNF␣) antigen against which the domain had been selected (k a ϭ 4.8 ϫ 10 3 s Ϫ1 M Ϫ1 , k d ϭ 8.9 ϫ 10 Ϫ3 s Ϫ1 , K D ϭ 1.9 M) (Fig. 2B).
Generation of the Garvan II Human V H Phagemid Repertoire-We next investigated strategies to further increase the affinity of the human domains. The first generation library had been constructed in a multivalent phage vector, thereby streamlining analysis of aggregation resistance using the heat/ cool method described by Jespers et al. (27). However, it is also evident that multivalent display can promote the enrichment of low affinity clones (through increased avidity, resulting in multiple interactions with antigen immobilized on a solid support) (47). We therefore decided to utilize a phagemid system for the construction of a second generation library, allowing for monovalent display on the phage surface. In addition to changes to the display system, we also considerably increased the number of CDR positions diversified in the library (Fig. 3A) while avoiding positions important for biophysical properties.
Recent work in our laboratory had demonstrated that the introduction of negatively charged amino acids at CDR1 positions significantly improved the colloidal stability of human V H domains (aspartate or glutamate at 28, 30, 31, 32, 33, and 35) (48). This is in agreement with the observations for the HEL4 model domain, which carries a negatively charged triad within this region ( 31 DED 33 ) (Fig. 3A). Importantly, introduction of multiple negatively charged amino acids at CDR1 positions is required (two or more) (48). This can be readily observed upon stepwise introduction of mutations into DP47, resulting in considerable increases of heat refoldability (Fig. 3B). In addition to reversible unfolding, the introduction of mutations at the above CDR1 positions also significantly increases expression and purification yields (44,48), mimicking important and highly desirable characteristics of camelid domains. We also retained two other CDR mutations (28R and 35G) within the libraries that are present in the HEL4 model domain (Fig. 3A). These have been reported to result in modest improvement of V H properties (10,27,38,49). In contrast to HEL4 and previously reported camelized domains, all of the mutations utilized here are located at CDR positions, fully retaining the human DP47 framework sequence (Fig. 3A).
More specifically, negatively charged amino acids were introduced at CDR positions 32 and 33 while extensively randomizing CDR1, CDR2, and CDR3 positions to obtain a maximum degree of repertoire diversity (Fig. 3 (A and C) and Table 1). Based on our experience with the Garvan I library, we predominantly utilized codons encoding hydrophilic amino acids at CDR1 and CDR2 positions (Table 1). For the diversification of CDR3, we utilized triphosphoramidite codon mutagenesis (TRIM) (50). This enabled us to encode a high proportion of amino acids commonly observed in the human repertoire (in particular, Tyr, Gly, Ala, and Ser) with other amino acids at lower defined proportions (Table 1) (51). A library of 3 ϫ 10 9 individual clones was obtained after transformation into E. coli bacteria (Garvan II library). From the naive library, clones were randomly selected and analyzed using the phage heat/cool method (27). This revealed that a large proportion of the library displayed resistance against heat-induced aggregation (with a median resistance of 40%), despite the randomization of up to 16 positions within three CDRs (Fig. 3D).
Phage Display Selections Using the Garvan II Phagemid Repertoire-Antigen-specific binders were next selected from the second generation repertoire. As described for the Garvan I library, antigens were immobilized on solid supports, and after three to four rounds of selection, binders were identified by phagemid ELISA and verified by soluble ELISA. After selection, a large proportion of clones (14 -58%, depending on the target antigen) displayed antigen binding, with DNA sequencing revealing non-redundant sets of 1-6 (unique) clones per target (Fig. 3E). Two of the antigens (VEGF and CD25) represent validated therapeutic targets and had been included in the selections for this reason. Biosensor measurements of representative V H domains selected against these targets revealed high affinity interactions with equilibrium binding constants in the midnanomolar range (VEGF, k a ϭ 5.7 ϫ 10 4 s Ϫ1 M Ϫ1 , k d ϭ 2.3 ϫ 10 Ϫ2 s Ϫ1 , K D ϭ 407 nM; CD25, k a ϭ 6.6 ϫ 10 4 s Ϫ1 M Ϫ1 , k d ϭ 1.4 ϫ 10 Ϫ2 s Ϫ1 , K D ϭ 208 nM).

Characterization of Garvan II V H Single Domains and Epitope
Binning-In addition to therapeutic targets, we had also included the common model antigen HEL in the selections to streamline comparisons with previously reported structures of camelid V H H complexes (52). The selected anti-HEL V H single domains were cloned into the pET12a expression vector, expressed in large scale, and further characterized (clones H04, G08, D05, C01, and F05). Surface plasmon resonance was used to determine binding kinetics, revealing equilibrium binding down to the low nanomolar range (Fig. 4 (A and B); 26 -257 nM). In addition to high affinity antigen binding, all of the selected clones also displayed high levels of resistance against heat-induced aggregation (Fig. 4B).
To gain further insight into the epitope distribution of the selected clones, we performed competition assays against three previously reported anti-HEL camelid domains (clones D3L11, D2L19, and D2L24) (52). D3L11 and D2L19 bind to the active site cleft of the HEL antigen (Fig. 4C, asterisk). This represents the dominant mode of binding of camelid domains to this antigen, which is well characterized through previously reported co-crystal structures (5,52). In contrast, D2L24 binds to a planar epitope on the back of the molecule, distant from the active site cleft (Fig. 4C) (52). Epitope binning experiments were carried out using surface plasmon resonance and one of the human V H single domains selected here (clone H04 characterized by a high affinity and stable surface plasmon resonance baselines). These experiments revealed that H04 competed for binding with D3L11 and D2L19 but not with D2L24 (Fig. 4D). Moreover, H04 also completed for binding to antigen with G08, D05, C01, and F05 (Fig. 4E). This strongly indicated that all of the human domains selected in this study bind to epitopes within or proximal to the lysozyme active site cleft.
Crystal Structure of a Human V H Domain in Complex with Hen Egg White Lysozyme-To gain further insights into the interaction of V H H04 with HEL antigen, molecular details were investigated by x-ray crystallography. The V H -antigen complex was isolated by size exclusion chromatography, and crystals were obtained by hanging drop vapor diffusion under two distinct conditions (see "Experimental Procedures" for details). Diffraction data were collected at the Australian Synchrotron, and structures solved by molecular replacement and refined to 2.60 and 2.26 Å resolution ( Table  2). The structures obtained from both data sets displayed two complexes within the asymmetric unit and overall high levels of structural similarity (with root mean square devia-  (Garvan II). A, amino acid sequence of HEL4 model domain, DP47 germ line, and Garvan II library with numbering according to Kabat, with CDR1, CDR2, and CDR3 regions highlighted in green, pink, and blue, respectively. B, thermal unfolding (red) and refolding (blue) of HEL4, DP47, DP47, and CDR1 double and triple mutants as monitored by circular dichroism. C, schematic representation of human V H single domain with CDR1, CDR2, and CDR3 regions highlighted. D, colloidal stability of Garvan II library, as measured by retained binding to superantigen after heating to 80°C. E, selection of binders.  tions of 0.38 -0.52 Å for 231-246 C␣ atoms). Further investigations therefore focused on the 2.26 Å resolution data set. Nature of the Single Domain-Antigen Interaction and Comparative Analyses-Initial structural analyses revealed a 1:1 stoichiometry and binding of a single V H H04 domain to a single HEL antigen (Fig. 5A). The interaction is centered on the CDR3 region of V H H04, which protrudes deeply into the lysozyme active site cleft (Fig. 5B), burying extensive surface area (Table 3; 535 Å 2 ). Notably, this interaction surface is achieved within the context of a CDR3 loop of nine residues, considerably shorter than what had been observed for camelid V H H and shark V NAR domains raised against HEL antigen (12-18 residues; Table 3) (52)(53)(54). More specifically, an extensive network of hydrogen bonds is observed between CDR3 residues (Tyr-95, Ser-97, Pro-99, Gln-100, Asn-100a, and His-100b) and the HEL cleft surface (Fig. 5B). Intriguingly, two prolines (at positions 98 and 99) are observed in the center of the loop. The presence of sequential prolines in CDR3 is unusual and results in a distinct bend of backbone orientation, projecting key contact residues toward the HEL active site. Complementing CDR3 interactions, minor contacts from CDR1 and CDR2 are observed in the H04-HEL complex (buried surface areas of 20 and 60 Å 2 , respectively). Notably, an aspartate residue at CDR1 position 33, which is a determinant of aggregation resistance (48), participates in a salt bridge with Arg-61 of the antigen (Fig. 5B). Outside the CDRs of V H H04, an additional 168 Å 2 of single domain surface is buried by the HEL surface. These interactions involve a number of mainly hydrophobic residues (Val-37, Leu-45, and Trp-47) within the light chain interface of the domain, packing against the surface of the HEL antigen (Fig. 5C).
In total, a surface area of 783 Å 2 is buried by V H H04 in the lysozyme complex, which is comparable with that of a shark V NAR single domain (755 Å 2 ) but higher than what is observed for a set of camelid domains raised against this antigen (561-645 Å 2 ) ( Table 3). All of the analyzed structures display a large degree of shape complementarity between single domains and the HEL antigen (H04, 0.81; V NAR , 0.77; camelids, 0.76 -0.80). In the majority (5 of 6) of complexes, this high level of complementarity is achieved by means of an extended CDR3 interacting with the active site cleft of the lysozyme antigen. In contrast to H04, shark, and camelid domains, previously reported structures of human single domains (V H 9 (24) and V1a (21)) display lower levels of shape complementarity with antigen (0.69 and 0.70). This is reflected by the essentially planar epitope surfaces of their respective antigens (MDM4 and VEGF) (Fig. 6). Although both domains bind to antigen predominantly through CDR3 contacts, the loop is not inserted into structural clefts but rather packs against a hydrophobic helix surface (in the case of V H 9-MDM4) (24) or interfaces with an outer ␤ strand (in the case of V1a-VEGF) (21). Structural Features of CDR3 and the V H /V L Interface-In addition to contacts with antigen, the structure of the human H04 single domain also revealed important structural roles of CDR3 residues. As outlined above, a hallmark of camelid domains involves the presence of a hydrophobic pocket formed by CDR3 packing against hydrophobic chains of the V H H framework (5,52,(55)(56)(57) (as observed in the structure of the D2L24 anti-lysozyme V H H domain; Fig. 7A). Although these framework residues are not strictly conserved between human and camelid V H H domains, an analogous core is nonetheless observed in the V H H04 structure with a tyrosine residue at CDR position 95 packing against a hydrophobic framework core formed by residues at positions 35, 37, 93, and 100d (Fig.  7B). Tyrosines are common at Garvan II CDR3 positions and within the human repertoire, rendering it likely that similar structural features could be observed in other Garvan II domains (Table 1) (51). This finding is in contrast to the previ-ously reported structure of the camelized B1a domain, in which the interaction is mediated through a rare CDR tryptophan residue (Fig. 7C) (10). Taken together, the formation of a hydrophobic core at the base of H04 CDR3 and the restriction of backbone dihedrals through proline residues result in the stabilization of CDR conformations, allowing antigen binding through extensive cleft interactions.
The observed hydrophobic core is partially formed by a residue at CDR1 position 35, which is occupied by glycine in H04 and conserved in the Garvan II library. This amino acid is also observed in D2L24 (Fig. 7A), B1a (Fig. 7C), and HEL4, which results in moderate improvements of gel filtration profiles (but not heat refoldabilty) (27,38). This is in agreement with the findings of Barthelmy et al. (10), who observed an enrichment of glycine at position 35 when selecting for protein A superantigen binding as a (gentle) proxy of biophysical properties (no heat selection was utilized). In the HEL4 structure, the presence  of Gly-35 results in a marked conformational shift of Trp-47 within the V H /V L interface (Fig. 5D, red), resulting in an increase of surface hydrophilicity (38). However, such changes are not observed in the V H H04 structure reported here with the conformation of Trp-47 closely aligned with that observed in representative human Fab antibody structures (Fig. 5D). This suggests that the main role of Gly-35 may not be the rearrange-ment of Trp-47 conformations (38) but rather the formation of a hydrophobic pocket centered on positions 35, 37, 93, and 100x, which can act as an acceptor for hydrophobic CDR3 residues (as seen in the H04, B1a, and D2L24 structures; Fig. 7, A-C).
In addition to Trp-47, the conformation of the remainder of the H04 V H /V L interface (Fig. 8A), formed by residues at posi-  tions 37, 44, and 45, is also highly similar to that of native (light chain paired) human antibodies ( Fig. 8B; PDB code 3QOS). In contrast, the H04 interface is noticeably different from that of camelized human domains (Fig. 8C) and camelid V H H domains (Fig. 8D).

DISCUSSION
Our study of human V H single domains demonstrates that, unlike what has been reported for camelized human domains, improvements of biophysical properties do not necessarily rely on mutational changes of framework residues. Rather, we demonstrate that CDR composition alone can endow human single domains with favorable properties. Importantly, we demonstrate that these properties are largely independent of CDR3 diversity, allowing the randomization of this important antigen-binding region. This enabled us to generate a stable first generation library (capable of refolding after heating to 80°C) and to select antigen binders against a diverse set of antigens.
In a second step, we extensively randomized additional CDR1 and CDR2 positions while maintaining key determinants of biophysical properties (two or more negatively charged amino acids at Kabat positions 28,30,31,32,33,and 35). This second generation V H library retained a high proportion of stable V H domains. Whereas we had observed the selection of predominantly low-medium affinity clones for the Garvan I library, affinities in the nanomolar range were observed for the second generation Garvan II repertoire, comparable with what has been observed for camelid and shark domains (Table 3). Moreover, the selected human single domains also displayed structural modes (cleft binding) and shape complementarities that were highly reminiscent of shark and camelid domains. This is in marked contrast to previously reported structures of human and camelized domains, which are characterized by the absence of cleft interactions and lower shape complementarities.
In addition to the camelization of human V H , strategies for the humanization of camelid single domains have also become available in recent years. For instance, in a detailed study, Vinke et al. (58) have reported the generation of a partially humanized domain (h-NbBCII10 FGLA ) suitable as an acceptor framework for CDR grafting approaches. However, the study also highlighted limitations of this strategy, namely the loss of heat refoldability upon humanization and the requirement to maintain multiple camelid residues within the framework 2 region (58).
Taken together, our results demonstrate that libraries of human V H single domains can be constructed that rival the stability and cleft binding properties of camelid domains. The availability of fully human domains will minimize potential immunogenicity risks by removing the necessity to change large conserved antibody framework residues. Although the effects of framework changes on the immunogenicity of human antibody therapeutics are unknown, changes to CDRs are well studied, with the majority of antibodies in clinical practice having been generated through CDR grafting or CDR affinity maturation approaches (59,60). It is becoming increasingly evident that such changes have little or no detectable effect on immunogenicity in patients (22), further strengthening the rationale for the CDR-only approach outlined here. Moreover, by maintaining the structure of the human V H /V L interface, the domains developed in this study retain the potential of light chain pairing; this may allow the modular assembly of bispecifics through native chain interactions. We conclude that the availability of fully human V H domains has the potential to open up new and exciting opportunities for the development of human therapeutics.