Selection of a Novel and Highly Specific Tumor Necrosis Factor α (TNFα) Antagonist

Inhibition of tumor necrosis factor α (TNFα) is a favorable way of treating several important diseases such as rheumatoid arthritis, Crohn disease, and psoriasis. Therefore, an extensive range of TNFα inhibitory proteins, most of them based upon an antibody scaffold, has been developed and used with variable success as therapeutics. We have developed a novel technology platform using C-type lectins as a vehicle for the creation of novel trimeric therapeutic proteins with increased avidity and unique properties as compared with current protein therapeutics. We chose human TNFα as a test target to validate this new technology because of the extensive experience available with protein-based TNFα antagonists. Here, we present a novel and highly specific TNFα antagonist developed using this technology. Furthermore, we have solved the three-dimensional structure of the antagonist-TNFα complex by x-ray crystallography, and this structure is presented here. The structure has given us a unique insight into how the selection procedure works at a molecular level. Surprisingly little change is observed in the C-type lectin-like domain structure outside of the randomized regions, whereas a substantial change is observed within the randomized loops. Thus, the overall integrity of the C-type lectin-like domain is maintained, whereas specificity and binding affinity are changed by the introduction of a number of specific contacts with TNFα.

Inhibition of tumor necrosis factor ␣ (TNF␣) is a favorable way of treating several important diseases such as rheumatoid arthritis, Crohn disease, and psoriasis. Therefore, an extensive range of TNF␣ inhibitory proteins, most of them based upon an antibody scaffold, has been developed and used with variable success as therapeutics. We have developed a novel technology platform using C-type lectins as a vehicle for the creation of novel trimeric therapeutic proteins with increased avidity and unique properties as compared with current protein therapeutics. We chose human TNF␣ as a test target to validate this new technology because of the extensive experience available with protein-based TNF␣ antagonists. Here, we present a novel and highly specific TNF␣ antagonist developed using this technology. Furthermore, we have solved the three-dimensional structure of the antagonist-TNF␣ complex by x-ray crystallography, and this structure is presented here. The structure has given us a unique insight into how the selection procedure works at a molecular level. Surprisingly little change is observed in the C-type lectin-like domain structure outside of the randomized regions, whereas a substantial change is observed within the randomized loops. Thus, the overall integrity of the C-type lectin-like domain is maintained, whereas specificity and binding affinity are changed by the introduction of a number of specific contacts with TNF␣.
Tetranectin belongs to the large class of C-type lectins characterized by a common fold known as the C-type lectin-like domain (CTLD) 3 (1). Tetranectin is a homotrimeric human protein found in both plasma and tissue. This protein binds the lysine-binding kringle domains from apolipoprotein A (2), plasminogen (3), and angiostatin (4). Tetranectin is a 60-kDa protein built from a structural unit composed of three identical chains, each with a CTLD domain located C-terminally to a trimerizing coil-coil region (5). The CTLD domains retain their structural integrity as separate protein domains, (6,7) and, moreover, it was shown that their binding to the known tetranectin ligand, plasminogen kringle-4, exhibits the same thermodynamic parameters, irrespective of whether it was analyzed in the form of free monomeric domains or as tethered domains in the complete homotrimeric protein (3). In addition, the thermodynamic analysis showed that the formation of the trimer led to an apparent 100-fold affinity increase, which most likely is due to the avidity effect caused by the three-fold clustering of CTLD domains in the complete protein. Comparison of ensembles of natural CTLD domains for which the known structure and ligand specificity are known shows that the ligand-binding site can accommodate a diverse range of ligands. We therefore concluded that the tetranectin CTLD might be a useful scaffold for designing novel protein therapeutics. We could change the sequence of loops within the CTLD scaffold in the monomeric version without perturbing the overall structure. Thus, CTLD serves as an efficient starting point for the in vitro selection of a high affinity antagonist with a low immunogenicity. In this procedure, we use the monomeric CTLD domain during the in vitro selection procedure but use the naturally occurring trimeric version in downstream applications.
Borean Pharma established an extended and coherent technology platform to use C-type lectins as a vehicle for the creation of novel trimeric therapeutic proteins with increased avidity and unique properties as compared with current protein therapeutics, such as antibodies and small protein scaffolds. Human tetranectin may be readily tailored to meet specific therapeutic needs by "reprogramming" CTLD. Each CTLD has five loop regions, each 6 -9 amino acids in length, which determine the binding specificities. Reprogramming is performed by creating phage libraries displaying CTLD, where specific loops are randomized, followed by selection. Randomization can be repeated either sequentially or iteratively. Furthermore, the use of the CTLD platform ensures selected protein candidates, which are highly homologous to a native human secreted protein and thus of low immunogenicity. Initial validation of this novel scaffold technology was achieved by selecting an antagonist of hTNF␣, as described herein. Subsequently, the platform has been effectively used on a number of diverse and therapeutically relevant targets. Current potency of the tetranectin-derived hTNF␣ antagonist was obtained through carefully man-aged in vitro evolution steps (Fig. 1). After each step, the binding kinetics were measured, as well as the ability of the trimeric version to inhibit hTNF␣-induced apoptosis in L929 cells (see Table 1).

EXPERIMENTAL PROCEDURES
Development and Construction of a Truncated TNF␣ Antagonist Gene-TN-2-B-1-C31 was developed by the use of phage display selection technology. A phage library displaying the monomeric scaffold structure of the C-type lectin domain of human tetranectin containing random amino acid residues in positions 116 -122 (loop 1) and 145-150 (loop 4) was constructed and used for the selection of hTNF-binding CTLDs. A candidate from the primary selection, TN-2, was taken through three rounds of affinity maturation using general phage selection techniques. A construct expressing the full-length protein of TN-2-B-1-C31 was created by cloning the coding region of TN-2-B-1-C31 into the BamHI and EcoRI sites of the vector pT7CIIH6-AmpR (8).
CTLD Purification-Escherichia coli (strain BL21 DE3) cells were harvested by centrifugation and lysed in 100 ml of lysis buffer by sonication followed by the addition of detergent buffer and inclusion body recovery for refolding in the present first generation procedure. Inclusion bodies were washed with 0.5% Triton X-100, 1 mM EDTA, pH 8, and resuspended in 6 M guanidine, 50 mM dithiothreitol, 50 mM Tris-HCl, pH 8. After inclusion body solubilization, the buffer was changed to 8 M urea, 1 M NaCl, 5 mM ␤-mercaptoethanol, 50 mM Tris-HCl, pH 8, and the fusion CTLD was captured on a nickel-nitrilotriacetic acid-agarose column (Qiagen). After elution with 8 M urea, 1 M NaCl, 5 mM ␤-mercaptoethanol, 20 mM EDTA, 50 mM Tris-HCl, pH 8, the fusion CTLD was refolded by dilution with 3 M urea, 250 mM NaCl, 50 mM glycine, pH 9.5, 2 mM CaCl 2 , 0.5/0.3 mM GSH/GSSG overnight at a rate of 0.15 ml/min to a final dilution of 10ϫ sample volume at ϩ7°C with mixing. After refolding, the CTLD was purified further on SP-Sepharose FF in 8 M urea, 50 mM NaAc, pH 4.5, 0 -1 M NaCl gradient. The fusion tag was cleaved by recombinant human granzyme B in 1 M urea, 20 mM Tris-HCl, pH 7.3, 150 mM NaCl, 3 mM CaCl 2 overnight before a final purification on a SOURCE 15S column in 8 M urea, 50 mM NaAc, pH 4.5, 0 -1 M NaCl gradient and finally buffer-exchanged into 50 mM NaAc, pH 5.0, buffer.
hTNF␣ Purification-E. coli strain BL-21 (Invitrogen) was transformed with a pT7-CIIH6-TNF␣ expression vector encoding human TNF␣ and grown to an A 600 of 0.5 in standard tryptone-yeast extract medium with 100 g/ml ampicillin and 10 mM MgSO 4 before induction with CE46 phage and rifampicin addition to 200 g/ml. Cells from 6 liters of culture were harvested by centrifugation, washed (0.1 M Tris-HCl, pH 8, 1 M NaCl), sonicated, and centrifuged. The supernatant was loaded on a nickel-nitrilotriacetic acid column (Qiagen), and the fusion TNF␣ was captured, eluted, and further purified on an SP-Sepharose FF column (GE Healthcare). Digestion with factor Xa at ϩ5°C overnight was applied afterward. The protease and cleaved fusion tag were captured on soybean trypsin inhibitor and nickel-nitrilotriacetic columns, respectively. A final Q-Sepharose FF step yielded a pure hTNF␣ batch.
Biacore Assay-Surface plasmon resonance binding analysis was performed on a Biacore 3000 instrument. The hTNF␣ capture antibody (AHC3712) to be immobilized was dissolved in 10 mM NaAc, pH 5.0, and then immobilized on a CM5 Biacore sensor chip using the amine coupling kit (Biacore). Binding analysis was performed at a flow rate of 5 l/min. Before loading of the protein sample, the chip was equilibrated in 10 mM Tris, pH 8.0, 150 mM NaCl, 2 mM CaCl 2 , 50 M EDTA, and 0.005% surfactant P20. Recombinant human TNF␣ was dissolved in 10 mM Tris, pH 8.0, 150 mM NaCl, 2 mM CaCl 2 at 10 g/ml, and 10 l was injected. Aliquots (20 l) of CTLD C31 were injected using the KINJECT option. 5 minutes of dissociation were allowed before the chip was regenerated with sequential injection of 0.05% SDS and 10 mM glycine, pH 2.5. Binding of CTLD C31 was analyzed at six different protein concentrations, ranging from 1.5 to 50 nM. Binding data were evaluated using the BIAevaluation program version 3.2 (Biacore).
Inhibition of hTNF␣-induced Apoptosis-The effect of inhibition of hTNF␣ by trimeric TN-2-B1-C31 was analyzed in a standard cell assay for the antagonistic effect of hTNF␣ binders on the cytotoxic effect of hTNF␣ on the murine L929 cell line as described in Ref. 9.
Crystallization, Data Collection, and Structure Determination-The complex was formed by mixing hTNF␣ with CTLD at a 1:1 molar ratio. Initial crystallization screening was performed with the Crystal Screen  The structure of the CTLD-hTNF␣ complex was solved by molecular replacement using Phaser (11). We used a search model composed of one wild type hTNF␣ monomer (Protein Data Bank (PDB) code 1TNF) (12) and the wild type CTLD (6) (PDB code 1TN3), excluding the loops, which were changed during the in vitro evolution procedure. The resulting model was used as a starting point for wARP/ARP (13), which did build 98% of the complex (including the maturation loops). The model was finalized with Coot (14) and refined in REFMAC5 (15). The final model consists of residues 9 -157 for hTNF␣, residues 46 -180 for CTLD excluding the loop region 52-56, for which no electron density was seen, 244 water molecules, and one magnesium ion. 99.1% of the residues are in the allowed regions of the Ramachandran plot, with 0.4% (1 residue) in the generously allowed region and Arg-120 from CTLD in the disallowed region. Figures were made using the program PyMOL (19).

RESULTS AND DISCUSSION
A primary phage display library was generated by randomizing loops 1 and 4 in the CTLD domain of tetranectin, and selection for binding to hTNF␣ was performed by panning. This resulted in a candidate molecule of high specificity but suboptimal affinity (Clone TN-2). This clone was then subjected to three consecutive steps of affinity maturation. First, loops 1 and 4 were randomized individually to provide full sequence coverage of each loop, and phage display selection was performed to select for higher affinity. This did not result in any changes in loop 1 (the loop 1 sequence was reselected), but 3 out of 4 amino acids were changed in loop 4 (only the 1st proline was conserved and would remain conserved throughout the maturation). An additional unexpected change occurred outside the randomized area, where D CTLD 157 was changed to Gly. The significance of this change, although not fully clear, may have helped stabilize the CTLD structure and resulted in substantially improved affinity (Table 1). However, the TN-2-B clone was incapable of inhibiting hTNF␣-induced apoptosis. In the second maturation step, loop 1 was randomized in increments of 3 amino acids, and the selection procedure was changed to favor clones with low off-rates. This was done by including prolonged washing steps and competition by free hTNF␣. The resulting TN-B-1 clone showed a 300-fold reduction in off-rate and improved bioactivity. During the third and final stage of maturation, loop 3 was randomized together with two positions in loop 4. This only led to a marginal improvement in the binding kinetics (35-fold in affinity and 6-fold in off-rate for the monomeric CTLD). However, the bioactivity improved by almost 200-fold for the trimeric tetranectin derivative.
The improved bioactivity of clones obtained by applying selective pressure for "better" dissociation kinetics demonstrated the potential of this strategy to improve the efficacy of tetranectinbased therapeutics. It should be noted that the selection as well as the binding data originate from the monomeric version of CTLD, whereas the trimeric version was used for measuring the biological activity. As compared with the moderate improvement in binding of the monomeric CTLD after the third round of maturation, the 200-fold improvement in biological activity of the trimeric tetranectin derivative might well be explained by a large increase in avidity of the trimeric tetranectin derivative.
To further understand the three-dimensional interactions responsible for the high affinity binding and to assure that the overall structure of tetranectin was maintained, we determined the crystal structure of the CTLD-hTNF␣ complex by molecular replacement at 2.1 Å resolution ( Fig. 2A). The final model consists of residues 9 -157 for hTNF␣ and residues 46 -180 for CTLD, excluding residues 52-56, for which no electron density was observed ( Table 2). The secondary structure of hTNF␣ bound to CTLD closely resembles that of free hTNF␣, with a root mean square deviation of 1.2 Å for C␣ atoms (12,16). Likewise, the non-randomized part of the CTLD scaffold exhibited little change in structure (root mean square deviation for C␣ atoms with free native CTLD is less than 1 Å), which validates the use of human tetranectin as a scaffold or platform for engineering new biologics. As for the CTLD loops, the structure has changed substantially in the randomized loop 3 and loop 4 (supplemental Fig. 1). In contrast, the randomized loop 1 backbone resembles that of native CTLD despite a complete change of sequence. The side chains of the randomized maturation loops display a systematic directionality toward hTNF␣, creating a visibly depressed surface in the previously rather flat hTNF␣ area. The buried, accessible surface area at the interface between hTNF␣ and the CTLD is 865 Å 2 , and the interaction surface features a mixture of polar, van der Waals, and hydrophobic interactions. Hydrophobic interactions are often the driving force in tight protein-protein interactions. However, we observed a more diverse set of interactions, which in this case is most likely found because the CTLD must adapt to the hTNF␣ surface, including neutralizing the charge present.
Residues K CTLD 116 and SRYF CTLD -(119 -122) in loop 1 are conserved throughout the maturation procedure. We were unable to select new loop 1 sequences in the first maturation round by complete randomization of loop 1. In the second round of maturation, where we performed partial randomization of loop 1, K CTLD 116 and SRYF CTLD -(119 -122) were reselected, and only VR CTLD -(117-118) was changed to RW. This suggests that these residues play an essential role in the binding of CTLD to hTNF␣. The structure shows that they interact with a well organized loop structure in hTNF␣ consisting of residues 138 -141 (Fig. 2B). This loop in hTNF␣ has multiple internal contacts; R TNF 138 hydrogen-bonds with D TNF 140 and is further prevented from lateral flexibility by close van der Waals a Primary, first, and second clones were analyzed with ϳ500 RU ligand immobilized, and the third clone was analyzed with ϳ50 RU. T-values Ͼ 100. RU designates Biacore response unit; T-value is defined as parameter value divided by S.E. b ND, not detectable. packing with Y TNF 141, thus providing a stable binding epitope for the CTLD. R CTLD 120 and Y CTLD 121 form hydrogen bonds with D TNF 140 and R TNF 138 of hTNF␣, respectively. Considerable strain is put on R CTLD 120, making it fit into the tight space at the complex interface, which results in / angles outside the allowed regions of the Ramachandran plot. The hydroxyl group of S CTLD 119 forms multiple hydrogen bonds with the backbone of K CTLD 116 and F CTLD 122 and thus locks the conformation of the randomized loop 1. Furthermore, S CTLD 119 is situated at the center of the CTLD-TNF␣ interaction surface pointing into the core of the CTLD, and thus there is substantial steric restriction on this residue. K CTLD 116 is also conserved in the selection procedure but does not contact hTNF␣. However, K CTLD 116 forms several contacts with both the backbone and the side chain of Q CTLD 148 and the side chain of Q CTLD 151. Thus, the length and charge of lysine are favorable at position 116, stabilizing the structure of the randomized loops. The selection scheme not only produces strong interactions with hTNF␣ but also ensures the stability of the interacting framework. Both T CTLD 121 and F CTLD 122 participate in close van der Waals packing with hTNF␣. F CTLD 122 also stacks closely with R CTLD 130 and redirects the side chain of this residue toward hTNF␣. R CTLD 130 is found well outside the randomized part, but the side chain has adopted a completely different conformation from that found in free CTLD. The guanidine group of R CTLD 130 forms hydrogen bonds with the side chain of E TNF 104 and the backbone of T TNF 105 (Fig. 2C). Furthermore, the backbone of R CTLD 130 forms hydrogen bonds with the backbone of G TNF 108 and A TNF 109 of hTNF␣. Salt bridges between flexible residues, such as glutamic acid and arginine, are usually rather unstable, but here F CTLD 122 restricts the flexibility of R CTLD 130, and R CTLD 130 is caught in a pocket between F CTLD 122 and hTNF␣. Thus, not only do the randomized residues contact hTNF␣ themselves, but they also reprogram part of the non-randomized part of the structure to interact with hTNF␣. R CTLD 117 and W CTLD 118 appear during the second round of maturation, where there is a selection for low off-rate of the antagonist. W CTLD 118 increases the buried hydrophobic surface substantially, and it also stacks with the guanidine group of R CTLD 117 and stabilizes R CTLD 117 hydrogen binding with hTNF␣.
Loop 4 has changed conformation considerably as compared with the native tetranectin CTLD structure. However, the loop makes relatively few contacts with hTNF␣ that are predominantly made by W CTLD 149. P CTLD 146 appears in the first round of selection and is maintained throughout the maturation. We believe that P CTLD 146 dictates a change in the conformation of loop 4, which is required to avoid sterical conflict with hTNF␣. Thus, in the first round of selection, loop 1 provides the major part of the binding affinity, and loop 4 changes position to allow this interaction. In the later maturation stages, W CTLD 149 appears and provides additional hydrophobic interaction with hTNF␣ and probably contributes to the lower off-rate.
Loop 3 was randomized during the third and final round of selection. Loop 3 does not make any direct contact with hTNF␣. However, the position of the loop has shifted substantially relative to the native CTLD structure. It is likely that the FIGURE 2. A, overview of hTNF␣-CTLD complex structure. The hTNF␣ secondary structure is shown as green ribbon, combined with a transparent surface representation. The part of the hTNF␣ surface involved in interactions with CTLD is shown in space-fill coloring according to electrostatic potential. The antagonist is depicted as cyan ribbon, except for the randomized part, which is shown in magenta. B and C, close up of the hTNF␣-CTLD interactions. Amino acid labels in red refer to CTLD, and amino acid labels in black refer to hTNF␣. D, reconstruction of trimeric version of the TN2-B1-C31 using the structure of natural trimeric tetranectin as a model. This was done by substituting the structure of the CTLD domain found in the structure of natural tetranectin (PDB ID 1HTN) with the structure of the TN2-B1-C31 clone in complex with hTNF␣. Next we added the symmetry mates of hTNF␣ to reconstruct the naturally occurring trimeric hTNF␣ (panel E). changes in loop 3 help to stabilize the conformation of both loop 3 and loop 4 and provide a more rigid interaction platform. Loop 3 is normally involved in binding of Ca 2ϩ , but Mg 2ϩ is found bound in the former Ca 2ϩ -binding site, and crystallization is performed in the presence of Mg 2ϩ . However, none of the ions are needed for binding to hTNF␣, and selections have been performed in the absence of both ions (data not shown).
Several anti-TNF␣ drugs have been marketed, based either upon antibody technology or upon modified soluble TNF␣ receptor (17,18). We compared the CTLD clone selected here with the available drugs. In our in vitro bioassays, the TN2-B1-C31 clone compared favorably to these drugs, both in terms of binding affinity and when comparing the ability to inhibit hTNF␣-induced apoptosis (data not shown).
In summary, the data show that the CTLD scaffold can be "reprogrammed" for binding to novel targets through a short series of directed evolutionary steps while maintaining its core scaffold structure. Furthermore, the crystal structure of the CTLD-TNF␣ complex provided a unique insight into the workings of directed evolution. Initial binding affinity is provided by loop 1, which is accompanied by a change in the structure of loop 4 removing steric hindrance. During later steps of selection, in which a low off-rate was favored, additional buried hydrophobic contacts appeared, represented by W CTLD 118 and W CTLD 149. The selection also seems to favor additional rigidity of the antagonist, as exemplified by the changes of structure in loops 3 and 4. The successful selection for low off-rates is of particular importance for development of therapeutically use-ful antagonists. In subsequent selections for candidates against a number of targets, low nM affinity candidates were selected in the first round and then further affinity-matured for low offrates as described here, thereby rendering solid therapeutic drug candidates with favorable pharmacodynamic profiles.