A structure-based engineering approach to abrogate pre-existing antibody binding to biotherapeutics

Development of biotherapeutics is hampered by the inherent risk of immunogenicity, which requires extensive clinical assessment and possible re-engineering efforts for mitigation. The focus in the pre-clinical phase is to determine the likelihood of developing treatment-emergent anti-drug antibodies (TE-ADA) and presence of pre-existing ADA in drug-naïve individuals as risk-profiling strategies. Pre-existing ADAs are routinely identified during clinical immunogenicity assessment, but their origin and impact on drug safety and efficacy have not been fully elucidated. One specific class of pre-existing ADAs has been described, which targets neoepitopes of antibody fragments, including Fabs, VH, or VHH domains in isolation from their IgG context. With the increasing number of antibody fragments and other small binding scaffolds entering the clinic, a widely applicable method to mitigate pre-existing reactivity against these molecules is desirable. Here is described a structure-based engineering approach to abrogate pre-existing ADA reactivity to the C-terminal neoepitope of VH(H)s. On the basis of 3D structures, small modifications applicable to any VH(H) are devised that would not impact developability or antigen binding. In-silico B cell epitope mapping algorithms were used to rank the modified VHH variants by antigenicity; however, the limited discriminating capacity of the computational methods prompted an experimental evaluation of the engineered molecules. The results identified numerous modifications capable of reducing pre-existing ADA binding. The most efficient consisted of the addition of two proline residues at the VHH C-terminus, which led to no detectable pre-existing ADA reactivity while maintaining favorable developability characteristics. The method described, and the modifications identified thereby, may provide a broadly applicable solution to mitigate immunogenicity risk of antibody-fragments in the clinic and increase safety and efficacy of this promising new class of biotherapeutics.


Introduction
Biotherapeutics are the most rapidly evolving drug class with monoclonal antibodies constituting the majority of biomolecules in development [1]. Recently, novel antibody-derived a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 or VH domains alone or in complex with other proteins. The largest number of VHH PDB entries were derived from llama (Lama glama) (n = 120), alpaca (Vicugna pacos) (n = 82), and camel (Camelus dromedarius) (n = 62). Fourteen structures were retrieved for autonomous human VH domains (refinement resolution range of 1.5 to 2.8Å). Due to their large number, VHH structures were divided into 3 subsets by species and only high-resolution structures with refinement resolutions better than 2Å were included for further analysis. This filtering process resulted in 47 structures for llama VHHs, 41 for alpaca VHHs, and 57 for camel VHHs. The program MOE (Molecular Operating Environment (MOE), 2019.01; Chemical Computing Group ULC) was used to remove non-VH(H) proteins, solvent atoms and ligands from the PDB files. VHH or VH CDRs were annotated according Kabat's CDR definitions [19]. CDR regions, as expected, displayed high sequence and structural variability, while superimpositions of only framework residues yielded RMSD values well below 1Å, indicating remarkable structure conservation. Both amino acid sequences and their structural arrangement in the C-terminal portion of VHs and VHHs were found to be highly conserved. This information was then used to design VHH variants with structural changes to abrogate preexisting ADA binding to the VHH C-terminus.

Homology modeling
Homology models of VHHs with unmodified C-terminus and various C-terminal designs to abrogate pre-existing ADA binding were generated using the protein modeler tool in the "Antibody Modeling" suite of the program MOE (Molecular Operating Environment (MOE), 2019.01; Chemical Computing Group ULC). Multiple VHH templates were generated from the pool of existing VHH structures from the RCSB PDB database (rcsb.org; [18]). Preferred templates for CDR loops based on amino acid identity and the number of amino acids in each of corresponding CDR loops were selected. Constructed models were protonated according to pKa predictions at pH 7.4, then the Amber10 ETH force field [20] was assigned. This permitted full minimization of the VHH homology models in implicit solvent simulated by a Generalized Born Solvation model [21]. Graphical representations were generated with Pymol (The PyMOL Molecular Graphics System, Version 2.0 Schrödinger, LLC).

B cell epitope predictions
In-silico identification of linear B cell epitopes was carried out with Ellipro and Bepipred, and conformational epitopes were predicted with Ellipro, both available through the IEDB Analysis Resources. BepiPred predicts the location of linear B-cell epitopes using a combination of a hidden Markov model and a propensity scale method, resulting in a score [22]. The residues with scores above the threshold are predicted to be part of an epitope. The threshold was varied in the 0.90-1.15 range to assess changes in antigenicity associated with the different cutpoints. ElliPro predicts linear and discontinuous antibody epitopes based on a protein antigen's 3D structure [23]. ElliPro accepts as an input protein structure in PDB format. The respective PDB files from the models generated as explained in the previous paragraph were used as the input for ElliPro analysis.

Cloning, test expression, and SDS-PAGE analysis
A VHH sequence was synthesized (IDT, Coralville, IA) with a 6X His-tag fused to the N-terminus with a three amino acid linker (GGS). Nineteen variations of the His-tagged VHH sequence, including various amino acid additions, deletions, or substitutions, were designed at the C-terminus or framework 1 and were also synthesized (IDT, Coralville, IA) ( Table 1). The twenty sequences were cloned into a GS-containing expression plasmid backbone (pEE12.4based plasmid) with the viral CMV promoter. The VHH sequences were fused in frame with the coding sequence of a signal peptide sequence, METDTLLLWVLLLWVPGSTG, to enhance secretion of the VHH variants into the tissue culture medium.
The VHH variants were recombinantly expressed from 2mL cultures of transiently transfected CHOK1 suspension cells using a PEI-based method, as previously described [24,25]. At the end of the incubation period, cells were removed by low-speed centrifugation and supernatants containing the VHH proteins were harvested for subsequent SDS-PAGE analysis. The supernatants, 10μL, for all 20 VHH variants were analyzed on a 12% polyacrylamide SDS-PAGE gel (BioRad, #345-0118), stained by Simply Blue Safestain (Invitrogen, #LC6060) and destained overnight in distilled water. Reduced samples were mixed 1:20 (v/v) with reducing agent (BioRad, #161-0792), 1:4 (v/v) with 4x loading dye (BioRad, #161-0791) and heated at 95 o C for 5 min before loading onto the gel. Non-reduced samples were mixed with 4x loading dye (BioRad, #161-0791) before loading onto the gel. Images of the destained gels were acquired using a UVP EpiChemi3 Darkroom imager with white light and UVP Visionworks LS software. Image editing was limited to cropping and labeling only.

Large-scale expression and purification
The VHH variants were recombinantly expressed from transient transfections of CHOK1 suspension cells using a PEI-based method, as previously described [24,25]. At the end of the incubation period, cells were removed by low-speed centrifugation and supernatants containing the VHH proteins were harvested for subsequent 2-step purification.

Plasma stability assay
VHH variants MC6.40, MC6.43, MC6.51 and MC6.56 were tested for their stability in mouse plasma. Previously frozen mouse plasma (-80˚C) was thawed and filtered using 0.22 μm Ultrafree MC centrifugal filters (Sigma Aldrich) for 5 min at 10,000 g. For each VHH variant, 125 μg were transferred to a 1.8 mL Eppendorf tube. The total volume was brought to 1 mL with the filtered mouse plasma, achieving a greater than 10-fold dilution. Samples were vortexed and then set on a thermomixer at 37˚C and 350 rpm. At each time point (1, 6, 24, 48 h), a 200 μL (approx. 25 μg) sample was removed and the VHH variants were extracted using HIS-PurTM Ni-NTA spin columns (ThermoScientific). Spin columns were equilibrated to room temperature and the bottom tabs were removed. Each column was placed into an Eppendorf tube and centrifugated for 2 min at 700 g to remove storage buffer. Equilibration Buffer, 400 μL, was added and the column was placed in a new Eppendorf tube. After a 2 min wait, the column was centrifugated for 2 min at 700 g to remove Equilibration Buffer. The bottom plug was placed back on the column and 200 uL of mouse plasma stability sample were added to 200 μL of Equilibration Buffer in each spin column. The spin column was taped into a new Eppendorf tube and mixed on an orbital mixer at room temperature for 30 min. The bottom plug was removed, and the column centrifuged for 2 min at 700 g. The column was placed into a new Eppendorf tube, washed with 400 μL of Wash Buffer and centrifuged for 2 min at 700 g. The flow through was kept and the process repeated 2 more times. The column was placed in a new Eppendorf tube and 200 μL of Elution Buffer were added. The column was centrifuged for 2 min at 700 g and the process repeated 2 more times. All 3 elutions were combined and the volume reduced by Speed Vac to approximately 400 μL. The column was washed with 400 μL of Elution Buffer and centrifuged for 2 min at 700 g. The combined eluates were buffer-exchanged with Zeba Spin columns (7K MWCO) using 50 mM Tris-HCl pH 7.5 and dried down to 50-100 μL with a speed vac. Samples from the various timepoints were then analyzed by mass spectrometry to determine integrity of the VHH variants.

Structure-based design of VHH variants
Pre-existing antibodies, like antibodies in general, recognize both conformational as well as linear epitopes. Thus, to abrogate pre-existing ADA binding to the C-terminal neoepitope of isolated VH and VHH domains, a structure-based approach was utilized. First, a total of 159 crystal structures of VHHs and autonomous human VH domains from the RCSB PDB database (rcsb.org; [18]) were compared. It was observed that their overall structures as well as the amino acid sequences and structural arrangements in their C-terminal portion are highly conserved (S1-S5 Figs in S1 File). A network of molecular interactions occurs at the C-terminus: Residues Thr110, Val111, and Ser112 are part of a β-sheet composed of the G and A' β-strands of the IgV domain, which also includes residues Leu11 and Val12 (Fig 1A and 1B). The C-terminal Ser113 forms two main chain H-bonds to the side chain of Gln13. In addition, the structural arrangement is stabilized by hydrophobic packing interactions between side chains of Leu11 and Val12 with Thr110 and Val111, respectively ( Fig 1B).
For potential modifications that would abrogate pre-existing ADA binding to the VH(H) C-terminus, the following constraints were considered: 1. Structural changes should remain localized to the VHH C-terminus to avoid negative impact on overall structure. This enables the same pre-existing ADA designs across VHs and VHHs without impacting antigen binding or destabilizing the molecule.
2. Changes should have a low likelihood of inducing de novo immunogenicity. Therefore, preferably proline and glycine residues were utilized in the pre-existing ADA designs as those have been reported to confer low complex stability in HLA-II-binding peptides [26].
Based on those constraints, a total of 18 new VHH variants were generated, which featured the addition of extra amino acids, disruption of H-bonds to reorient amino acid side chains, changes of main-chain conformation, and the removal of side chains (Table 1). Some designs were also included to interrogate the potential contribution of Leu11 to the C-terminal preexisting ADA epitope. Leu11 is known for its close interactions with residues in the antibody constant region [27]. In isolated VH and VHHs, the Leu11 side chain becomes highly solventexposed, and it was hypothesized that this new hydrophobic surface might contribute to the pre-existing ADA epitope.

In-silico predictions of B cell epitopes at the C-termini of the VHH variants
Homology models of VHHs of unmodified C-terminus (Fig 2A) and various C-terminal designs (Fig 2B) were generated using the protein modeler tool in the "Antibody Modeling" suite of the program MOE. VHH templates were generated from the pool of existing VHH structures from the RCSB PDB data base. The C-terminal designs were implemented and each VHH variant was subjected to energy minimization. Superimposition of the homology models showed that structural perturbations indeed remained limited to the VHH C-terminus ( Fig  2B). In addition, the various designs represented a wide variety of distinct structural conformations as desired.
To predict whether the modified C-termini would constitute less likely targets for pre-existing antibodies as compared to the unmodified sequence, two in-silico B cell epitope mapping tools were used. Because antibodies can recognize both linear and conformational epitopes, both Ellipro and BepiPred were utilized to assess both scenarios. The results of these analyses are reported in Tables 2 and 3, respectively. ElliPro predicts linear and discontinuous antibody epitopes based on a 3D structure of an antigen, which can be loaded as a PDB file [23]. There was no significant difference between unmodified VHH sequence with all the VHH variants in the likelihood that a cluster of residues containing the C-terminus would constitute a conformational epitope. Instead, this   G9, G10, L11, V12, Q13, P14, G15, G16, S17, A61, D62, S63, K65, G66, R67, AN84, S85, L86,  R87, P88, E89, T91, T122, S124, S125 prediction indicated that 11 C-terminal designs would introduce a linear epitope that is not otherwise present in the parental sequence or the other seven variants. BepiPred utilizes a Hidden Markov Model trained using a data set derived from the database Antijen and the Pellequer data set and combined with Parker's hydrophilicity scale to predict the location of linear B-cell epitopes [22]. The residues with scores above the threshold are predicted to be part of an epitope. A threshold of 0.90 yields a sensitivity of 0.25 and a specificity of 0.91. Using this cutpoint, all the sequences would contain a linear epitope at the C-terminus. On the contrary, adoption of a higher threshold, such as 1.15, would decrease sensitivity (~0.18) and increase specificity (~0.93). With the higher threshold, none of the predicted linear B cell epitopes would comprise the C-terminal residues, including the parental VHH. Collectively, these results indicate that the ability to discriminate variants on the basis of the in-silico predicted reactivity to pre-existing antibodies is weak, and an experimental assessment strategy is required to rank these molecules more accurately.

Generation and characterization of VHH variant proteins
Twenty variants of a human serum albumin-binding VHH were cloned for recombinant expression in mammalian cells. This included the humanized VHH (MC6.1) with an unmodified C-terminus; the same VHH with a C-terminal alanine addition that had been previously found to reduce pre-existing ADA binding [17] (MC6.40); and 18 new structure-based designs (MC6.41-MC6.58). It should be noted that a N-terminal His-tag was included in all constructs to facilitate purification and evaluation in the pre-existing ADA assay. Small-scale test expressions demonstrated comparable expression levels for all 20 VHH variants (Fig 3A), indicating that the designs did not negatively impact intracellular folding or stability in the cell culture. All 20 VHH variants were then expressed at large scale and purified to homogeneity using a 2-step purification process. Final purification yields of all 20 VHH variants were comparable (S6 Table in S1 File).
To ensure the pre-existing ADA designs did not induce any major structure disruptions, the thermal stability of the VHH variants was evaluated in a thermofluor assay (Fig 3B). All, but two VHH variants exhibited Tm values of 75-76˚C, comparable to the unmodified VHH MC6.1 (mean Tm of 75.3˚C ± 0.88˚C for all VHHs). Interestingly, MC6.41 and MC6.46 had 1.5˚C and 3.5˚C reduced thermal stability, respectively. The statistically significant thermal stability reduction of MC6.46 (Ser112Pro, desSer113) likely results from disruption of the beta-sheet secondary structure due to changes in the main-chain conformation and hydrogen bonding from the Ser112Pro substitution. This hypothesis is also corroborated by the fact that MC6.50, which is identical to MC6.46 with exception of Ser112Gly versus Ser112Pro, does not have a reduced Tm.
To verify that the structure-based designs would not be subjected to proteolytic cleavage in biological matrices, selected VHH variants (MC6.40, MC6.43, MC6.51 and MC6.56) were evaluated in a plasma stability assay. Up to the final 48h timepoint, all 4 VHH variants proved to be stable at 37˚C without any detectable modifications in the C-terminal regions as monitored by mass spectrometry.

Evaluation of pre-existing ADA binding to VHH variants
All 20 VHH variants were evaluated against a pool of sera of 10 healthy human donors using an ELISA assay (Fig 4A). In the assay, the VHH with an unmodified C-terminus, MC6.1, showed a high level of pre-existing ADA binding (Fig 4B). The addition of a single alanine residue Ala114 in MC6.40, as previously described [17], led to a significant reduction of pre-existing ADA reactivity and validated the assay setup. MC6.47 (Ser113Ala) had comparable reactivity as MC6.1, indicating that the conservative serine to alanine modification of the Cterminal residue had no impact on recognition by pre-existing ADA. In contrast, a change to Removal of side chains of the C-terminal residues did show some promise, depending on the applied change. Interestingly, while combining a Ser113Ala change with the addition of an extra proline at residue 114 (MC6.48) had only a modest impact on pre-existing ADA reactivity, the same design using a Ser113Gly change (MC6.52) lowered pre-existing ADA reactivity to buffer background levels. Consistently, the promising designs were changes to or additions of proline residues, i.e., adding a single proline at position 114 (MC6.42) reduced signals from nearly 8-fold to 2-fold over buffer control while addition of two or three prolines (MC6.43 and MC6.44) lowered pre-existing ADA reactivity to the buffer control.
To compare pre-existing ADA reactivity across the various VHH variants, for each molecule the values of absorbance measured in response to the serum of each of the 70 donors was plotted, and the mean and standard deviation was calculated ( Fig 6A). As expected, the unmodified MC6.1 had the highest level of pre-existing ADA reactivity. MC6.51, MC6.52 and MC6.56 had strongly reduced pre-existing ADA reactivities comparable to Cordy's design [17] (MC6.40). MC6.43 was the most effective design with pre-existing ADA reactivity comparable to the buffer control, indicating that addition of prolines 114 and 115 efficiently abrogated pre-existing ADA binding. This can be explained by the significant structural differences the pre-existing ADA design imparted on the C-terminal VHH region. Structure models of the unmodified MC6.1 VHH (Fig 6B) and the MC6.43 variant (Fig 6C) highlight how the addition of the extra two proline residues reshapes the molecular surface at the C-terminus. Proline 114 hydrophobically stacks against Leu11, the main chain of Val12 and the side chain of Gln13 ( Fig 6C). Numerous hydrogen bonds to the side chain of Gln13 stabilize the structural  conformation and may contribute to both thermal and serum stability. Pre-existing ADA, like antibodies in general, recognize both conformational as well as linear epitopes. By reshaping the molecular surface of the C-terminal VHH epitope of MC6.43, it is possible to efficiently abrogate pre-existing ADA reactivity.

Conclusion
Protein therapeutics, particularly engineered molecules, bear an inherent immunogenicity risk with potential impact on PK, safety and efficacy [5]. Pre-existing ADAs, in addition to treatment-induced anti-drug antibodies, are recognized as a development challenge for biotherapeutics [7][8][9]. Although some progress has been accomplished towards understanding origins and clinical impact of pre-existing ADA, much remains to be learned; e.g., understanding the specificity of pre-existing ADA is a pre-requisite to enable engineering strategies to abrogate pre-existing ADA binding and reduce potential impact on PK, efficacy, and safety [7,8]. In this research, a structure-based approach was utilized to specifically engineer a known pre-existing ADA epitope at the C-terminus of isolated VH and VHH domains by leveraging 3D-structures to design modifications applicable to any VH(H), which should not impact developability or antigen binding. In-silico B cell epitope prediction algorithms were adopted for ranking of the modified VHH variants based on antigenicity; however, this analysis was insufficient in differentiating the parental from the engineered sequences. The limited applicability of in-silico immunogenicity predictions for pre-existing ADA reactivity required an experimental evaluation of numerous VHH variants, in which several designs were identified to have greatly reduced pre-existing ADA binding. Of particular note, the addition of two proline residues at the VHH C-terminus possessed no detectable pre-existing ADA binding resulting from its significant reshaped molecular surface and favorable developability characteristics. Thus, this particular modification may provide a broadly applicable solution to mitigate immunogenicity risk of VH(H)s and other antibody fragments in the clinic to increase safety and efficacy of this promising new class of biotherapeutics.