Computational design of a specific heavy chain/κ light chain interface for expressing fully IgG bispecific antibodies

Abstract

The use of bispecific antibodies (BsAbs) to treat human diseases is on the rise. Increasingly complex and powerful therapeutic mechanisms made possible by BsAbs are spurring innovation of novel BsAb formats and methods for their production. The long‐lived in vivo pharmacokinetics, optimal biophysical properties and potential effector functions of natural IgG monoclonal (and monospecific) antibodies has resulted in a push to generate fully IgG BsAb formats with the same quaternary structure as monoclonal IgGs. The production of fully IgG BsAbs is challenging because of the highly heterogeneous pairing of heavy chains (HCs) and light chains (LCs) when produced in mammalian cells with two IgG HCs and two LCs. A solution to the HC heterodimerization aspect of IgG BsAb production was first discovered two decades ago; however, addressing the LC mispairing issue has remained intractable until recently. Here, we use computational and rational engineering to develop novel designs to the HC/LC pairing issue, and particularly for κ LCs. Crystal structures of these designs highlight the interactions that provide HC/LC specificity. We produce and characterize multiple fully IgG BsAbs using these novel designs. We demonstrate the importance of specificity engineering in both the variable and constant domains to achieve robust HC/LC specificity within all the BsAbs. These solutions facilitate the production of fully IgG BsAbs for clinical use.

Short abstract

PDB Code(s): 5VR9; 5VSI; 5VSH

Abbreviations

BsAb

bispecific antibody

CDR

complementary determining region

DSC

differential scanning calorimetry

ELISA

enzyme linked immunosorbent assay

Fab

fragment antigen binding

Fc

fragment crystallizable

IgG

immunoglobulin γ

HC

heavy chain

LC

light chain

mAb

monoclonal antibody

MSD

multistate design

PK

pharmacokinetic

RP HPLC

reverse phase high pressure liquid chromatography

SPR

surface plasmon resonance.

Introduction

Bispecific antibodies (BsAbs) incorporate the antigen binding properties of two monoclonal antibodies (mAbs) into a single molecule. As therapeutics, BsAbs not only represent a combination therapy within a single molecule with the capacity to drive down cost, but may be designed to elicit novel biological mechanisms such as co‐clustering of signaling receptors,1 precise targeting of specific cell/tissue types,2, 3 or targeting immune cells to kill cancer cells.4 A growing number of bispecific therapeutic mechanisms enabled by BsAbs are being recognized and brought to the clinic to treat various diseases.5 Indeed there has been a revival in BsAb activity over the past few years because of advances in their production and the therapeutic benefits they may provide.6

Over the past couple decades, an extraordinary amount of research has been directed toward the development of novel BsAb platforms.7 The production and development of these molecules has been challenging for many reasons.8 Many BsAb platforms use V_H, V_L, or Fv antibody fragments (Fig. 1). These formats extract the variable domains from their native Fab environment, which can result in biophysical or pharmacokinetic (PK) issues that need to be addressed through protein engineering.9 Antibody fragments have on occasion been recognized as non‐native by the human immune system with pre‐existing antibodies to these fragments in patient sera.10, 11 The addition of an IgG‐Fc to a bispecific construct can impart BsAbs with immunoglobulin‐like PK through binding FcRn and recycling into sera (Fig. 1).12 The presence of a standard IgG‐Fc also results in obligate tetravalency (Fig. 1).13, 14 Tetravalency, or more specifically the ability of these BsAbs to bind each target using two functional antigen recognition units, may be beneficial for certain therapeutic applications where avidity or receptor hyper‐crosslinking is important, but may not be required or indeed wanted for other applications such as the fine tuning necessary for T‐cell redirected tumor targeting15 or the inhibition of certain receptor tyrosine kinases (RTKs) that require monovalent recognition to avoid unwanted agonism.16, 17

Figure 1.

Schematic diagrams of various BsAb modalities. The top left are schematics of two monoclonal IgGs with their unique antigen‐binding regions colored red and blue. Each Ig‐fold domain is depicted as an oval. The LCs comprising VL and CL could be κ or λ isotype. The top right depicts examples of domain or Fv based BsAbs. The center panel depicts commonly employed IgG‐like BsAbs that use IgG or IgG‐Fc to impart IgG‐like pharmacokinetics combined with antibody fragments to enable bispecific binding. The bottom panel depicts two different fully IgG BsAb modalities. The work described in this report focuses on the generation of monovalent IgG BsAbs as depicted on the left in the blue square.

Fully IgG BsAbs offer the potential to capture the biophysical and biological (PK, FcγR‐engagement) properties of native IgGs and the ability to monovalently engage antigens (Fig. 1). The ability to generate BsAbs with fully native IgG architecture has only recently become possible through significant protein engineering efforts, even though human IgG4 naturally heterodimerizes to form IgG BsAbs.18 Most fully IgG BsAb platforms use heterodimeric Fc designs of greater or lesser complexity to enable different antibody heavy chains (HCs) to heterodimerize instead of forming homodimers.7, 19, 20 Fc heterodimerization was first described two decades ago; however, the incidence of light chain (LC) mispairing within native IgG BsAbs hindered the production of IgG BsAbs.21 Some methods have been described to generate IgG BsAbs by separate expression of each of the IgG components (ultimately requiring two manufacturing cell lines if used to produce material for clinical studies), purification of the two parental mAbs, followed by biochemical heterodimerization under the proper redox conditions.22, 23 Spiess et al., have generated IgG BsAbs by co‐culturing two transformed E. coli cell lines each harboring one of the parental mAbs together under redox conditions favoring heterodimerization in the culture supernatant.24 These methods can produce large quantities of IgG BsAbs, but require additional processing steps, which would be time consuming, complex and costly to produce material for clinical trials or drug products.

Ideally, fully IgG BsAbs could be generated from two naturally occurring IgGs within a single mammalian cell line, which is the current standard for mAb manufacturing.25 This would dramatically reduce the complexity and cost of their production. Given the advances and many solutions to Fc heterodimerization,19 the main hurdle is light chain (LC) mispairing. If two mAbs with C_H3 (Fc) heterodimer designs, but lacking specificity designs in the Fab region, are co‐expressed, a significant level of LC mispairing is theoretically and practically observed.20, 26 Therefore, to reduce these bi‐products, one must either use parental mAbs with the same LC (prohibitive for the general research community) or design each LC of the parental mAbs to pair with their appropriate HC.20 We and others have developed methods to facilitate specific LC pairing when expressing 2 HCs and 2 LCs in a single CHO cell.3, 26, 27, 28 These methods provide a path for producing IgG BsAbs directly from a single transfection.

The focus of our previous work was to redesign conserved regions of the V_H/V_L and C_H1/C_L interfaces to improve specific LC pairing.26 The goal was for the designs to be useful for LCs of both the κ and λ isotype and direct such LCs to their appropriate HC partner within a heterodimeric HC pair. During the design process, we intentionally modified amino acids that were identical between C_H1/Cλ and C_H1/Cκ with the hope that the designs would be applicable to both.29 However, the amino acid context around the modified residues differs, and while we achieved remarkable HC/LC specificity within C_H1/Cλ, the designs provided weaker LC/HC specificity when placed in C_H1/Cκ. A solution we have applied to achieve HC/LC specificity within IgG BsAbs that have two κ LCs is to pair one of the V_H/Vκ Fv regions with the designed C_H1/Cλ that contains the specificity designs.26 This has worked well for many IgG BsAbs in‐house. In fact, Fabs with mixed V_H/Vκ and C_H1/Cλ domains have significantly improved biophysical properties over scFvs containing only V_H/Vκ.²⁹ However, there is thermodynamic cooperativity between V_H/Vκ and C_H1/Cκ pairings that is absent in Fabs with λ LCs (V_H/Vλ and C_H1/Cλ pairings). We found that V_H/Vκ pairings with C_H1/Cλ do not benefit from strong and cooperative thermodynamic stabilization.29

Given these caveats, we deemed it useful to generate novel C_H1/Cκ designs that could improve the HC/LC specificity for generating IgG BsAbs in a single cell line while maintaining the unique cooperative Fab thermodynamics observed within κ LCs. The new designs were generated using a computational technique that we recently developed within the modeling software Rosetta for increasing the energy gap between target and off‐target interactions.30 The accuracy and utility of the designs were assessed using association assays and by solving high resolution crystal structures. We also present modifications to our original V_H/V_L designs26 to achieve improved HC/LC specificity. Finally, we demonstrate the utility of the designs by making several IgG BsAbs using different parental mAbs.

Results

A V_H/V_L charge swap that improves correct HC/LC assembly of IgG BsAbs

Based on our previous experience, we understood that engineering specific interfaces for both the variable and constant domains of Fabs is important for proper IgG BsAb assembly.26 Our original designs provided good specificity, however, we felt based on our experience with numerous IgG BsAbs (both published and internal) that additional variable domain designs may improve HC/LC specificity.

Generating specificity designs in the variable domains requires a number of considerations. First, roughly 30% of the V_H/V_L interface is composed of the H3 and L3 complementarity determining regions (CDRs). Any changes to these loops or interface residues in their vicinity may have a significant impact on antigen binding and should be avoided to maintain antigen binding affinity. This leaves approximately 50% of the interface for design. Additionally, the designs should be at conserved positions across variable domain germlines if they are to be generally useful to a broad range of parental mAbs.

Through structure gazing, we evaluated numerous V_H/V_L residue pairs at the interface and found the V_H_Q105/V_L_K/Q42 (Kabat numbering31) residue pair as a good candidate for imparting specificity (Supporting Information Fig. S1). In particular, we evaluated the introduction of charged residues at these positions as charge swaps have been shown in other applications to provide specificity.19, 32 A small charge library was constructed within the pertuzumab IgG33 at these positions, expressed and screened for thermal stability using a thermal challenge assay. The most stable charge pair, V_H_Q105R/V_L_K42D, was significantly more stable than the wild‐type (WT) pairing [Fig. 2(A)]. We were interested whether this charge swap (denoted VRD3) could improve the pairing observed with two previously described variable domain designs (VRD1 and VRD2).26 We constructed the pertuzumab HC to carry either VRD1 or VRD2 with or without VRD3 (the VRD2 and VRD3 combination was denoted VRD2_3). We performed a competition study to see if HCs carrying VRD1 or VRD2/VRD2_3 would specifically bind LCs carrying VRD1 or VRD2/VRD2_3 if co‐expressed in the presence of both VRD1 and VRD2/VRD2_3 κ LCs. The relative ratio of VRD1 or VRD2/VRD2_3 bound to each HC was determined by reduced mass spectrometry, using the assumption that both LCs ionize equally. A bias was observed for both VRD1‐containing LCs to bind both VRD1 and VRD2‐containing HCs [Fig. 2(B)]. For the pertuzumab HC containing VRD2, 62% bound the pertuzumab LC with VRD1. When VRD3 was added to VRD2, the mispairing of HC VRD2_3 with LC VRD1 was decreased modestly, suggesting an improvement in specificity.

Figure 2.

Screening and structure of a novel charge swap in antibody variable domains to enhance specific HC/LC pairing. A: Sandwich ELISA binding of thermally challenged pertuzumab variants that contain a charged residue library at V_H_105/V_L_42. B: Reduced intact mass spectra of two fully κ LCs expressed simultaneously with a single pertuzumab HC containing the VRD2 specificity design. Note, the constant domain CRD2 design provides poor specificity when used in Cκ and is optimal for Cλ.²⁶ The lower molecular weight (MW) LC indicates the level of proper pairing with a LC containing the VRD2 specificity design, while the higher molecular weight peak is a mispairing with a LC containing VRD1_CRD2 designs. The bottom curve is in the absence and the top curve is in the presence of the VRD3 (V_H_Q105R/V_L_K42D) design. C: Schematic diagram of the variable domains from an X‐ray crystal structure of the pertuzumab Fab containing the VRD2 and VRD3 design (PDB 5VSH). There were four Fabs in the unit cell. The Fabs were overlaid on one‐another. Residues V_H_105R and V_L_42D (design VRD3) are shown in stick format. Significant mobility was observed at position 105R that placed the sidechain charged groups from 6 to 10 Å apart (not counting hydrogens), which suggests why the specificity gain was modest. D: Overlay of the wild‐type and VRD3‐containing Fabs. Movement of residue V_H_105 closer to V_L_42 is observed.

Next, we generated IgG BsAbs in the presence or absence of VRD3. We found previously that the HC/LC specificity imparted by our original C_H1/C_L specificity designs (CRD2) was very good in the context of C_H1/Cλ, but weaker in the context of C_H1/Cκ.²⁶ Indeed, IgG BsAbs made using these designs with fully κ LCs did not assemble with ≥90% correct HC/LC pairings (Supporting Information Table S1). These constructs provided a window to observe the impact of VRD3 on specific HC/LC pairing. Thus, we generated four IgG BsAbs with varied parental mAb components. The IgG BsAbs contained VRD1_CRD2 in one of the parental mAbs and VRD2 with or without VRD3 in the other parental mAb. These IgG BsAbs were expressed at small scale (in triplicate or more), purified by protein G, and assessed for proper assembly using non‐reduced intact mass spectrometry. A trend was observed indicating that VRD3 does impart modest gains in specificity (∼10% if averaged across all four IgG BsAbs, Supporting Information Table S1).

A high resolution crystal structure of the pertuzumab Fab with the VRD2_3 designs was obtained to help our understanding of how the V_H_Q105R/V_L_K42D charge introduction may improve HC/LC specificity. There were four copies of the pertuzumab Fab in each unit cell. Interestingly, there were clear dynamics observed for residue V_H_105R [Fig. 2(C)]. The B‐values were high for the sidechain in each of the Fabs and the apparent position of the sidechain varied within each Fab of the unit cell. The distance (without hydrogens) to the guanidinyl sidechain was between 6 and 10 Å from V_L_42D. One orientation has the V_H_105R sidechain pointing directly at the V_L_42D sidechain. These distances correlate well with the modest increase in HC/LC specificity observed with the introduction of these charges as clearly the two sidechains do not form a neatly ordered salt bridge. In the WT structure, V_H_105Q makes a hydrogen bond with the V_H backbone, but upon mutation, the V_H_105R sidechain reorients toward solvent and closer to V_L_Q42D, indicating that some modestly stabilizing interaction may occur via the mutation [Fig. 2(D)], although interactions with other residues within in the protein may also play a role.

Novel designs to improve specific C_H1/Cκ heterodimerization

We next turned to generating novel designs to improve C_H1/Cκ specificity given that our original designs were less effective in C_H1/Cκ.²⁶ We employed a computational multistate design (MSD) strategy to design an orthogonal pair of C_H1s and Cκs; an Aa pair and a Bb pair (capital letters = C_H1; lower‐case = Cκ) designed to prevent the formation of Ab and Ba. To design sequences capable of disfavoring one interaction while still favoring another requires explicit representation of both interactions; it would be ineffective to try and find a low‐energy Aa pair without considering whether this pair was capable of avoiding interactions with the Bb pair. The MSD algorithm in Rosetta uses an outer loop to search sequence space, and for each sequence it examines, it threads the sequence onto each of many different “states” in an inner loop. An objective function evaluates each sequence based on the energies of the states with the sequence threaded onto them; thus a design can be evaluated for multiple criteria simultaneously.

For speed, Rosetta's MSD algorithm uses fixed‐backbone side‐chain optimization in its inner loop. This can lead to artifacts in the sequences it produces. If a single backbone conformation is used for the off‐target interactions (the negative states), then MSD tends to find sequences that produce collisions across the interface [Fig. 3(A), red overlap] but that relax out of collision when the backbones are allowed to move. To circumvent this problem, we build up a negative state repertoire30 of alternate low‐energy docked conformations and feed these additional states back into MSD so that if MSD finds a sequence that induces a collision in some but not all conformations, then MSD is able to correctly decide that the sequence is imperfect and keep searching for a better one.

Figure 3.

Flow charts for iterative multi‐state design. A: Initial fixed‐backbone multistate design (MSD) starting from the C_H1/Cκ crystal structure produces high‐energy collisions (red) in the off‐target interactions (Ab & Ba), but these collisions can often by accommodated by small rigid‐body motions of the two chains. After MSD, we run rigid‐body docking on the resulting sequences. We collect the low‐energy conformations found by docking the off‐target interactions, and add these conformations to a growing negative state repertoire. These alternate conformations then become part of the inputs to the next MSD iteration, allowing MSD to eventually find collisions that the off‐target interactions cannot readily accommodate. B: Our first round designs were constructed from an initial set of 5K MSD trajectories. We took from these simulations those designs that destabilized both the off‐target interactions (Ab and Ba) and analyzed them for sets of co‐occurring mutations. The frequent co‐occurring mutations were grouped as families (usually mutating 3–5 positions) and within families all combinations of mutations were analyzed. Mutation combinations that appeared to destabilize one or both of the off‐target interactions were selected for experimental characterization.

After running MSD, we performed rigid‐body docking on the resulting sequences, four per design (Aa, Bb, Ab, and Ba), and filtered designs that successfully destabilized the off‐target interactions [Fig. 3(B)]. We analyzed the remaining designs looking for mutations that co‐occurred frequently, organizing them into twelve families. We enumerated all combinations of mutations in the families and interrogated each mutation set using a flexible‐backbone relaxation protocol. This flexible‐backbone protocol was very noisy with replicates showing standard deviations in the range of the binding energy differences. The protocol was still helpful in identifying sequences that appeared to destabilize the off‐target interactions in a fixed‐backbone context, but whose destabilization disappeared when the backbone was allowed to move. The C_H1_S188A/Cκ_S176I design (described below) was incorrectly filtered at this stage, though we serendipitously tested it in combination with other designs later. Mutation sets advanced through this stage if they destabilized one of the two off‐target interactions. We selected 32 designs from eight families using this procedure (Supporting Information Table S2, design families 1–11).

At the same time, we tested another thirteen designs generated using a similar procedure (families 36–229). These designs contained more mutations than the first batch (often more than eight per interface) and computationally did a better job of destabilizing the off‐target interactions. However, only three of these designs expressed, and they seemed to favor the off‐target interactions, and won't be described further. We speculate that Rosetta's force field and protocol for calculating protein free energies has intrinsic errors, and therefore, with each additional mutation that is made to the protein there is an increased chance that the protocol will include a mutation that it perceives to be good, but is actually strongly destabilizing. Also, when multiple mutations are made to the protein there is a greater chance for conformational rearrangement that impacts the energy of the system, but is not correctly modeled. Finally, β‐sandwich proteins such as antibody domains are high‐contact order structures (i.e., many tertiary contacts between residues distant in primary sequence) that typically fold slowly via complicated folding pathways. Indeed, specialized chaperones naturally promote the folding of antibodies.34 During the protein design process, Rosetta does not consider the folding mechanism or potential interactions with chaperones, and therefore, may favor mutations that are favorable for the free energy of folding but promote unproductive folding pathways or disrupt interactions with chaperones.

The initial computational designs and rational approaches validated by computation (Supporting Information Table S2) were generated within an IgG‐like construct containing only the IgG1 constant domains (C_H1‐Fc) and a Cκ constant domain (i.e., lacking antibody variable domains). The Cκ/C_H1‐Fc constructs allow the ability to screen for C_H1/Cκ specificity without variable domain interference. Additionally, two Cκ constructs were designed (with or without an N‐terminal 8 histidine tag) to enable charge‐based separation of different Cκ proteins on a reverse phase HPLC column, which is necessary using the screening methodology described below. The mutations added to the Cκ/C_H1‐Fc constructs are designed to improve specific assembly of an ‘Aa’ C_H1/Cκ pair and a ‘Bb’ C_H1/Cκ pair (Supporting Information Table S2).

To test the designs for their ability to enable specific assembly of selected C_H1/Cκ complexes, ‘A’ or ‘B’ variant C_H1‐Fc constructs were co‐transfected with both ‘a’ and ‘b’ Cκ constructs in mammalian cells (HEK293F or CHO). After culturing, proteins were purified from the mammalian cell culture supernatants by protein G, reduced, and characterized by gradient reverse phase (RP) HPLC in a 2D chromatography process. The ‘a’ and ‘b’ Cκ proteins generally elute at different concentrations of organic buffer because of the charge difference induced by the 8 histidine tag on ‘b’. The peak areas of ‘a’ and ‘b’ were integrated and the percent of ‘a’ and ‘b’ bound to each C_H1‐Fc protein was calculated. The Wild‐Type (WT) ‘b’ Cκ with the histidine tag expresses slightly poorer than the WT ‘a’ Cκ without the tag resulting in a roughly 75/25 ratio in the absence of any designs [Fig. 4(A,B)]. A few designs expressed well and improved the percent of 'Aa’ and ‘Bb’. These can be found in Table 1 and include designs 1.3, 1.18, 2.3, and 12.1. Many designs improved the percent ‘Aa’ and/or ‘Bb’, but resulted in decreases in protein expression and were not considered further.

Figure 4.

Screening and characterization of computationally derived C_H1/Cκ specificity designs. A: Denaturing reverse phase chromatograms of protein G purified Cκ/C_H1‐Fc heterodimers without specificity designs reduced with TCEP prior to injection. ‘Cκb’ and ‘Cκa’ are Cκ proteins with and without a C‐terminal histag, respectively, to allow chromatographic separation. B–D: Stacked chromatograms of Cκ proteins captured by C_H1‐Fc. B: used wild‐type Cκ/CH1‐Fc (WT); (C) 14.1.2; and (D) 14.3.1. The C_H1‐Fc and Cκ proteins co‐expressed with one another are labeled on each curve. The identity of each peak was determined by expressing the ‘A’ CH1‐Fc with the ‘a’ Cκ domain and the ‘B’ C_H1‐Fc with the ‘b’ Cκ domain. Competition of ‘a’ and ‘b’ for each C_H1‐Fc protein was quantitated by the area under the curve. E: DSC traces of matuzumab IgG1 with and without C_H1/Cκ designs as well as curve fits to the three observed transitions.

Table 1.

Specificity Screening Using Reverse Phase HPLC^a

Design	Chain A CH1	Chain A Cκ	Chain B CH1	Chain B Cκ	%Aab	%Abb	%Bab	%Bbb
WT	WT	WT	WT	WT	76.0 ± 6.6	23.9 ± 6.7	76.6 ± 2.7	23.3 ± 2.7
Initial designs evaluated using reverse phase HPLC
1.3	S188G	S176I	WT	WT	100/95	0/2.4	91.6/73	8.4/25
1.4	S188G	S176I	S188T	WT	100	0	82.2	17.8
1.5	S188A	S176M	WT	WT	96.3	3.7	92.1	7.9
1.6	S188A	S176M	S188T	WT	94.4/87.1	5.6/7.4	78.3/61.7	21.7/27.2
1.7	S188G	S176M	WT	WT	100/100	0/0	91.9/94	8.1
1.8	S188G	S176M	S188T	WT	100/96.7	0/0	77.7/94.2	23.3
1.12	S188G	S176I	S188I	S176G	94 ± 1.5	3.8 ± 3.6	67.8 ± 6.3	30.7 ± 4.1
1.13	S188A	S176M	S188M	S176A	89.4	10.6	67.8	32.2
1.15	S188I	S176A	WT	WT	72.2/68.4	27.8	0/100	0/0
1.18	S188G	S176I	S188I	S176A	98.7 ± 2.3	0 ± 0	68.2 ± 3.1	27.9 ± 5.2
2.3	K145A	S131R	WT	WT	59 ± 5.9	41 ± 5.9	1.3 ± 2.2	96.3 ± 3.7
2.4	K145A	S131K	WT	WT	61.6	38.4	0	100
2.7	K145S	S131R	WT	WT	51.8	48.2	0	100
2.8	K145S	S131K	WT	WT	34.8	65.2	7.5	92.5
12.1	K221E	E123K	WT	WT	100	0	23.8c	76.2c
12.2	K221E	E123Q	WT	WT	69.1	30.9	68	32
Combination designs evaluated using reverse phase HPLC
1.3.1	K145A S188A	S131R S176I	WT	WT	77.7	22.3	12.5	87.5
1.12.1	K145A S188A	S131R S176I	S188I	S176G	72.3	27.7	36.3	63.7
1.18.1	K145A S188A	S131R S176I	S188I	S176A	70.1	29.9	40	60
14.1.2	K145A K221E	S131R E123K	WT	WT	100c	0c	4.5c	95.5c
14.3.1	S188A K145A	S176I S131R	K221E	E123K	94.7	5.3	0	100
14.3.2	S188G K145A	S176I S131R	K221E	E123K	100	0	0	100
15.1	K145A K221E S188A	S131R E123K S176I	WT	WT	97.3	2.7	5.8	94.2
15.2	K145A K221E S188G	S131R E123K S176I	WT	WT	95.7	4.3	6.1	93.9
Combination designs evaluated by LCMS d
14.1.2	K145A K221E	S131R E123K	WT	WT	99.8 ± 0.1	0.2 ± 0.04	0/0	100/100
14.3.1	S188A K145A	S176I S131R	K221E	E123K	98.2 ± 0.6	1.8 ± 0.6	0.2 ± 0.2	99.8 ± 0.2
14.3.2	S188G K145A	S176I S131R	K221E	E123K	99.8	0.2	0.1	99.0

^aThe table represents a selected subset of the designs that were constructed and tested.

^bHPLC analysis of the specific association of Cκ proteins with CH1‐Fc proteins. The CH1‐Fc proteins with ‘A’ or ‘B’ designs were co‐transfected with both Cκ domains containing the ‘A’ and ‘B’ designs, denoted ‘a’ and ‘b’. The specificity of the ‘A’ for ‘a’ and ‘B’ for ‘b’ designs is determined based on the relative amount of ‘a’ that associates with ‘A’ relative to ‘b’. The same is true for the specificity of ‘B’ with ‘b’. The values in each column represent the %area for each Cκ species, ‘a’ or ‘b’, as observed using denaturing, reverse phase chromatography with protein G purified samples that are reduced prior to injection onto the reverse phase column. Values with error (±) were run 3 or more times. The mean value is listed followed by the standard deviation. Cells with 2 values were run in duplicate and the values listed are from each replicate.

^cPeak overlap in the HPLC method made these values difficult to quantify.

^dValues listed in the LCMS evaluation are the deconvoluted peak areas for each of the species.

We combined the best mutations from the initial screen into several alternative combinations. These were constructed, expressed and screened in a similar fashion as the original designs. Combining some designs resulted in modest increases in specificity over the initial designs. Many combinations led to significant decreases in protein expression and were not considered further. A few, however, resulted in near complete specificity with ‘A’ C_H1‐Fc binding with near complete specificity to ‘a’ Cκ and ‘B’ C_H1‐Fc binding with near complete specificity to ‘b’ Cκ. These included the 14.1.2, 14.3.1, 14.3.2, 15.1, and 15.2 designs [Table 1, Fig. 4(C,D)]. Significant overlap in the two Cκ domain elution peaks in the reverse phase elutions of ‘a’ and ‘b’ made it difficult to accurately quantitate the level of each Cκ domain for some designs [e.g., design 14.1.2, Fig. 4(D)]. Therefore, we evaluated the C_H1/Cκ pairing of 14.1.2, 14.3.1, and 14.3.2 by intact, non‐reduced LCMS as a secondary method. The mass spectrometry results correlated with the RP HPLC results; all three designs induced near complete specificity of ‘Aa’ and ‘Bb’ formation with minimal ‘Ab’ or ‘Ba’ mispairing (Table 1).

The best combination designs are all permutations of one another with or without C_H1_S188A/Cκ_S176I or C_H1_S188G/Cκ_S176I (Supporting Information Fig. S1). Each contains the interface pair mutation C_H1_K145A/Cκ_S131R (design 2.3) and the interface pair mutation C_H1_K221E/Cκ_E123K (design 12.1). Designs 14.1.2, 15.1, and 15.2 combine designs 2.3 and 12.1 within the same C_H1/Cκ domain, while design 14.3.1 and 14.3.2 harbor 2.3 and 12.1 in separate C_H1/Cκ domains. While both at the C_H1/Cκ interface, designs 2.3 and 12.1 are spatially distant from one another; therefore, it is not surprising they are capable of providing their individual contributions to specificity whether they are both in ‘Aa’ or split into ‘Aa’ and ‘Bb’.

We next tested the best C_H1/Cκ designs for their relative stability compared with WT C_H1/Cκ using an enzyme‐linked immunoadsorbent assay (ELISA) with heat challenged supernatants and/or differential scanning calorimetry (DSC) with purified protein. Using the ELISA method, we found that the 14.1.2, 14.3.1, 14.3.2, 15.1, and 15.2 designs do not destabilize the C_H1/Cκ heterodimer subunit. In fact, the 14.3.1 and 15.1 designs stabilized C_H1/Cκ. To confirm these results, we scaled up the expression of each Cκ/C_H1‐Fc pair to 100 mL, purified each protein, buffer exchanged them into PBS, and analyzed each variant by DSC. Again, while none of the C_H1/Cκ designs were destabilizing, designs containing C_H1_S188A/Cκ_S176I or C_H1_S188G/Cκ_S176I were significantly more stable than WT C_H1/Cκ based on their midpoints of thermal denaturation (T_m) [Fig. 4(E)]. Thus, the stabilized designs were 14.3.1Aa, 15.1Aa and 15.2Aa [Fig. 4(E)].

Structural impact of the novel C_H1/Cκ designs

To evaluate how each design imparts its specificity, we obtained high resolution crystal structures of the matuzumab Fab35 with and without design 15.1. Design 15.1 contains all the most successful mutations within a single C_H1/Cκ subunit (Table 1). Design 15.1 includes design 12.1, which was a rationally designed charge‐swap (C_H1_K221E/Cκ_E123K) validated computationally using the flexible‐backbone relaxation technique described above. Interestingly, the crystal structure indicates that the mutant swap was successful not for being a simple mirror image of the original charge pair. Instead, the terminal amine group of Cκ_E123K appears to hydrogen bond with the backbone of C_H1_V121 while staying in close proximity, but not necessarily salt‐bridging with, C_H1_K221E (Fig. 5). Design 2.3, also in Design 15.1, was another design partially exposed at the C_H1/Cκ interface. The design introduces a charge clash between ‘B’ and ‘a’ by the mutation of Cκ_S131R that is alleviated in ‘A’ by mutation of C_H1_K145A (Fig. 5). When aligning the 15.1 structure with the WT matuzumab Fab structure, it is clear how the ‘Ba’ pairing introduces a strong charge clash, while the ‘Ab’ variant leaves a void at the C_H1/Cκ interface. Lastly, mutation of C_H1_S188A/Cκ_S176I introduces a steric clash within the ‘Ba’ pairing while creating a hole at the center of the C_H1/Cκ interface for the ‘Ab’ pairing (Fig. 5). Additionally, an unburied serine without a hydrogen bond partner is expected in both ‘Ba’ and ‘Ab’. The C_H1_S188A/Cκ_S176I design was not tested in isolation; however, its close homologues, designs 1.3 and 1.5, both appeared to have minimal impacts on specific C_H1/Cκ pairing (Table 1). Instead, this mutant pair increased the expression of the subunit that harbors it through stabilization of the C_H1/Cκ subunit presumably due to the burial of additional hydrophobic surface area (Fig. 5). Thus, it can be used to improve the expression of HC1 or HC2 within an IgG BsAb, which may have implications for correct IgG BsAb assembly. One commonly observed issue during IgG BsAb expression is the formation of ‘half‐antibody’. Half‐antibody is an impurity derived from the overexpression of HC1 or HC2 resulting in an unwanted non‐disulfide linked head‐to‐tail homodimer.30, 36 The ability to change the HC1 or HC2 expression ratio while stabilizing one of the two mAbs within an IgG BsAbs is a significant commodity for reducing IgG BsAb impurities. The choice of what design to use (14.1.2, 14.3.1.1, 14.3.1.2, 15.1 and 15.2) provides maneuverability for the antibody engineer to generate IgG BsAbs with the highest levels of correctly assembled protein.

Figure 5.

Schematic diagrams of the C_H1/Cκ domains from the WT (PDB ID 5VR9) and design 15.1 Aa (PDB ID 5VSI) matuzumab Fab structures. The 15.1 structure includes all three mutation pairs: L_E123K/H_K221E, L_S131R/H_K145A, and L_S176I/H_S188A. Overall, the backbone moved very little between the two sequences with a Cα RMSD of 0.588 Å over 206 residues. L_E123K forms polar contacts with both H_K221E and the backbone carbonyl on H_V121 and occupies the space that would otherwise be occupied by H_K221 in the “B” heavy chain. Similarly, L_S131R occupies the space that would be occupied by H_K145 in the “B” heavy chain. Finally, L_S176I would clash with H_S188 in the “B” heavy chain; and H_S188A would leave L_S176 buried and unsatisfied in the “b” light chain.

Generation of IgG BsAbs using the novel C_H1/Cκ specificity designs

The purpose behind the novel C_H1/Cκ designs described here was to provide an improved alternative for fully IgG BsAb assembly. In particular, the C_H1/Cκ designs should help direct correct HC/LC pairing when two HCs and two LCs are co‐expressed. To this end, the successful C_H1/Cκ designs were introduced into the HCs and LCs of 4 different proof‐of‐concept parental mAbs with the goal of producing well assembled IgG BsAbs. The parental mAbs chosen include pertuzumab,33 MetMAb,17 matuzumab,35 and BHA10.37 Pertuzumab and MetMAb were chosen to harbor the ‘Aa’ designs and BHA10 and matuzumab harbored the ‘Bb’ designs. We generated the IgG BsAbs to either not include or include variable domain interface designs that have been described previously26 and the new VRD3 design described in this report for enhanced HC/LC specificity. In particular, we wished to evaluate the impact the V_H and V_L domains have on correct HC/LC pairing. The HC sequences also included a C_H3 domain heterodimerization design, denoted 7.8.60, to induce specific HC1/HC2 heterodimerization over HC1/HC1 and HC2/HC2 homodimerization.30

Initially, we characterized the assembly properties of IgG BsAbs containing three of the successful C_H1/Cκ designs without adding any variable domain specificity designs. The HCs and LCs of each IgG BsAb were constructed in four separate expression plasmids and co‐transfected at small scale (2–4 mL) in mammalian cells. The IgG BsAbs were purified using the same HPLC protein G method described above for characterizing the Cκ/C_H1‐Fc proteins; however, the proteins were not reduced and submitted for RP chromatography. Instead, the purified proteins were collected and submitted for intact, non‐reduced LCMS analysis. LCMS provides information regarding the level of correct and incorrect assembly of each IgG BsAb given the assumption that chains with similar molecular weights have similar ionization efficiencies. Without C_H1/Cκ or V_H/V_L designs, the four IgG BsAbs were roughly 10% above the theoretical 50% mispairing level expected based on random pairing, except the MetMAb/matuzumab IgG BsAb, which achieved roughly 80% correct HC/LC assembly likely due to intrinsic V_H/V_L pairing preferences (Table 2). The roughly 60% correctly paired IgG BsAbs containing both LC1 and LC2 could be correctly paired IgG BsAb or be IgG BsAb with double LC mispairing (i.e., HC1/LC2 and HC2/LC1). However, we believe as the apparent correct pairing improves to >85%, the contribution of the doubly mispaired LC likely becomes negligible.

Table 2.

Mass Spectrometry Determination of the Assembly of IgG BsAbs Using Novel CH1/Cκ Designs

VH/Vκa_CH1/Cκ Design	%BsAb	%2xLC1	%2xLC2	%AA Homodimer	%BB Homodimer
MetMAb (7.8.60a) x BHA10 (7.8.60b)
WT_WT	61.0 ± 1.0	14.6 ± 1.2	24.4 ± 0.4	0.0 ± 0.0	0.0 ± 0.0
WT_14.1.2	71.9 ± 2.1	2.7 ± 0.5	25.4 ± 1.8	0.0 ± 0.0	0.0 ± 0.0
WT_14.3.1	65.5 ± 1.2	2.7 ± 0.1	31.8 ± 1.2	0.0 ± 0.0	0.0 ± 0.0
WT_14.3.2	68.5 ± 2.3	2.8 ± 1.0	28.8 ± 1.5	0.0 ± 0.0	0.0 ± 0.0
VRD_WT	91.9 ± 3.8	0.8 ± 2.0	4.7 ± 4.7	1.2 ± 1.3	1.5 ± 2.8
VRD_14.1.2	95.4 ± 0.4	0.0 ± 0.0	3.9 ± 0.3	0.7 ± 0.7	0.0 ± 0.0
VRD_14.3.1	93.2 ± 1.6	0.0 ± 0.0	4.6 ± 0.9	2.5 ± 0.9	0.0 ± 0.0
VRD_14.3.2	85.6 ± 4.6	0.0 ± 0.0	7.2 ± 2.3	8.4 ± 3.1	0.0 ± 0.0
VRD_15.1	87.4 ± 2.5	2.3 ± 4.0	8.3 ± 2.1	0 ± 0.0	1.9 ± 1.7
VRD_15.2	91.4 ± 0.5	0.0 ± 0.0	7.2 ± 0.5	1.6 ± 0.1	0.0 ± 0.0
Pertuzumab (7.8.60a) x BHA10 (7.8.60b)
WT_WT	59.0 ± 5.3	33.9 ± 7.8	7.1 ± 2.6	0.0 ± 0.0	0.0 ± 0.0
WT_14.1.2	57.5 ± 1.7	36.0 ± 1.8	6.5 ± 0.4	0.0 ± 0.0	0.0 ± 0.0
WT_14.3.1	68.2 ± 0.3	28.3 ± 0.2	3.6 ± 0.2	0.0 ± 0.0	0.0 ± 0.0
WT_14.3.2	71.3 ± 2.6	26.9 ± 1.3	1.8 ± 1.5	0.0 ± 0.0	0.0 ± 0.0
VRD_WT	58.9 ± 2.8	39.9 ± 2.7	0.4 ± 0.4	0.0 ± 0.0	0.8 ± 0.9
VRD_14.1.2	95.9 ± 2.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	4.1 ± 2.0
VRD_14.3.1	88.3 ± 7.5	1.0 ± 0.1	0.0 ± 0.0	2.0 ± 1.8	1.7 ± 0.3
VRD_14.3.2	84.0 ± 1.7	0.9 ± 0.3	0.0 ± 0.0	14.1 ± 0.8	1.3 ± 1.2
VRD_15.1	96.8 ± 5.5	0.0 ± 0.0	0.8 ± 1.4	0.8 ± 1.4	1.6 ± 2.8
VRD_15.2	94.6 ± 3.7	0.0 ± 0.0	2.1 ± 0.9	1.7 ± 2.9	1.8 ± 3.1
MetMAb (7.8.60a) x Matuzumab (7.8.60b)
WT_WT	78.6 ± 0.5	12.3 ± 0.7	9.0 ± 0.9	0.0 ± 0.0	0.3 ± 0.3
WT_14.1.2	92.3 ± 0.4	1.5 ± 0.4	5.6 ± 0.2	0.0 ± 0.0	0.5 ± 0.2
WT_14.3.1	80.7 ± 2.2	0.0 ± 0.0	19.1 ± 2.3	0.0 ± 0.0	0.0 ± 0.0
WT_14.3.2	n.d.b	n.d.b	n.d.b	n.d.b	n.d.b
VRD_WT	91.0 ± 2.6	3.2 ± 2.3	1.2 ± 1.3	1.1 ± 1.2	3.7 ± 2.5
VRD_14.1.2	94.8 ± 0.3	0.0 ± 0.0	1.4 ± 0.3	2.3 ± 0.3	1.7 ± 0.1
VRD_14.3.1	93.7 ± 0.3	0.8 ± 0.1	1.4 ± 0.1	3.4 ± 0.1	0.6 ± 0.3
VRD_14.3.2	92.6 ± 0.5	0.9 ± 0.2	1.3 ± 0.2	5.4 ± 0.3	0.0 ± 0.0
VRD_15.1	92.3 ± 2.2	0.0 ± 0.0	4.3 ± 1.3	3.1 ± 1.9	0.6 ± 0.7
VRD_15.2	96.3 ± 2.6	0.0 ± 0.0	1.2 ± 0.6	0.3 ± 0.5	2.2 ± 2.5
Pertuzumab (7.8.60a) x Matuzumab (7.8.60b)
WT_WT	64.1 ± 5.1	29.1 ± 7.2	5.1 ± 1.8	2.9 ± 0.5	0.0 ± 0.0
WT_14.1.2	81.3 ± 0.8	18.1 ± 1.7	0.0 ± 0.0	0.7 ± 1.2	0.0 ± 0.0
WT_14.3.1	80.0 ± 2.0	11.5 ± 0.5	8.5 ± 2.3	0.0 ± 0.0	0.0 ± 0.0
WT_14.3.2	78.6 ± 1.0	12.4 ± 0.2	9.0 ± 1.2	0.0 ± 0.0	0.0 ± 0.0
VRD_WT	70.0 ± 11.0	21.6 ± 8.0	0.0 ± 0.0	0.0 ± 0.0	8.5 ± 4.0
VRD_14.1.2	97.3 ± 2.0	0.0 ± 0.0	0.0 ± 0.0	0.1 ± 0.2	2.6 ± 1.9
VRD_14.3.1	96.3 ± 1.5	0.0 ± 0.0	0.0 ± 0.0	2.6 ± 0.6	1.1 ± 1.0
VRD_14.3.2	96.2 ± 0.9	0.0 ± 0.0	0.2 ± 0.3	3.7 ± 1.2	0.0 ± 0.0
VRD_15.1	99.5 ± 0.9	0.0 ± 0.0	0.0 ± 0.0	0.5 ± 0.9	0.0 ± 0.0
VRD_15.2	99.1 ± 0.1	0.0 ± 0.0	0.0 ± 0.0	1.0 ± 0.1	0.0 ± 0.0

^a‘VRD’ indicates VRD1 in the first mAb and VRD2_3 in the second mAb.

^bn.d. = not done.

The BsAbs denoted ‘WT’ have no specificity designs in the variable domains while the BsAbs denoted ‘VRD’ contain VRD1 in the first HC/LC pair 26 and VRD2 and VRD3 in the second HC/LC pair. Values in the table represent the mean and standard deviation of at least three separate experiments. For each pair, the ‘Aa’ CH1/Cκ designs in the first HC/LC pair and ‘Bb’ designs in the second HC/LC pair

We next methodically added C_H1/Cκ designs, V_H/Vκ designs, or the combination of C_H1/Cκ and V_H/V_L designs to the HCs and LCs and characterized their impact on IgG BsAb assembly. Adding only C_H1/Cκ designs 14.1.2, 14.3.1, or 14.3.2 improved correct HC/LC assembly by roughly 10% over constructs with no designs; although two still displayed significant LC mispairing (Table 2). Adding only the V_H/V_L designs had a slightly stronger impact. In particular, the BsAbs containing MetMAb achieved ∼90% correct assembly using only V_H/V_L designs without any specificity designs in the C_H1/Cκ domains (Table 2). When combining the V_H/V_L and C_H1/Cκ designs, we obtained ∼90 correct assembly across all mAb pairings [Table 2, Fig. 6(A,B)].

Figure 6.

Characterization of IgG BsAbs generated using novel C_H1/Cκ designs. A, B: Stacked LCMS traces of IgG bispecifics (A. pertuzumab/matuzumab, HER2 × EGFR; B. pertuzumab × BHA10; HER2xLTβR) expressed in the presence of variable domain designs and in the absence (WT) or presence of C_H1/Cκ designs (14.1.2 or 14.3.1) after 1‐step protein G purification. Figures on the bottom depict the peaks corresponding to the correctly assembled IgG bispecific as well as peaks corresponding to specific mispaired proteins. C, D: Non‐reduced and reduced SDS‐PAGE analysis (C) and analytical SEC (D) of protein G/CaptoMMC two‐step purified IgG bispecifics. E: Surface plasmon resonance traces of 25 nM EGFR then HER2 injections (bottom) or HER2 then EGFR injections (top) over sensorchip surfaces with captured IgG bispecifics (Cκ/C_H1 from bottom to top: WT, red; 14.1.2, blue; 14.3.1, green; 15.1, black).

One aspect not captured in Table 2 was the level of half‐antibody observed for each of the IgG BsAbs. For both BsAbs containing MetMAb (MetMAb × matuzumab and MetMAb × BHA10), very little half‐antibody was observed in the absence of designs. This indicates that the matuzumab and BHA10 mAbs express at a similar level as MetMAb. Adding the C_H1/Cκ design 14.1.2 to either IgG BsAb (L_E123K, L_S131R and H_K145A, H_K221E) did not significantly impact half‐antibody levels. However, adding 14.3.1 to each of the BsAbs resulted in increased MetMAb half‐antibody, either due to an increase in MetMAb expression as part of the ‘Aa’ design (L_S131R, L_S176I and H_K145A, H_S188A) or due to a decrease in BHA10 and matuzumab expression resulting from the charge swap L_E123K/H_K221E in ‘Bb’. When stacking all the designs into the MetMAb Fab as with 15.1, we also observed a significant increase in MetMAb half‐antibody, suggesting that the L_S176I/H_S188A double mutation that stabilizes C_H1/Cκ domains results in enhanced expression of the antibody that harbors it. The trend was similar, but less consistent with the pertuzumab × BHA10 and pertuzumab × matuzumab IgG BsAbs.

The pertuzumab × matuzumab (HER2 × EGFR) IgG BsAbs with and without the C_H1/Cκ designs were scaled up for purification and characterization. These BsAbs were purified using two steps to reduce half‐antibody. Transient transfections were performed with equal amounts of the four plasmids harboring each HC and LC26, 38 (i.e., no optimization to reduce half‐antibody was performed). Half‐antibody ranged from 20–40% after the initial protein A affinity chromatography step. The second Capto MMC chromatography step helped eliminate half‐antibody and also separated IgG BsAbs with mispaired LC, which was particularly prevalent for the WT (control) IgG BsAb lacking C_H1/Cκ designs. The quality of the IgG BsAbs after purification was assessed by SDS‐PAGE and analytical SEC with in‐line static light scattering. The IgG BsAbs had the expected molecular weight by all three analytical methods [Fig. 6(C,D)]. Very little residual half‐antibody was observed by LCMS for these samples. However, roughly 5% half‐antibody is apparent on the SDS‐PAGE gel eventhough N‐ethyl maleimide (NEM) was used during SDS‐PAGE sample preparation to reduce artefactual half‐antibody production39 [Fig. 6(C)]. The reason for this discrepancy is not understood.

Potential impact of the designs on antigen binding affinity of the HER2 × EGFR IgG BsAbs was assessed using surface plasmon resonance. No difference in antigen binding was observed between any of the purified IgG BsAbs with or without C_H1/Cκ designs. The matuzumab (anti‐EGFR) equilibrium binding constant (K_D) was measured to be 11.5 ± 1.3 nM (identical to what was published previously40) and the pertuzumab (anti‐HER2) K_D was measured as 0.90 ± 0.07 nM across all the BsAbs [Fig. 6(E)]. The ability to bind two antigens simultaneously was assessed by sequential addition of the soluble receptor extracellular domains [Fig. 6(E)]. Modifying the order of addition of the soluble receptors did not impact IgG BsAb binding indicating that neither IgG BsAb binding site occludes the other from binding its respective target [Fig. 6(E)].

Discussion

We present here further designs to reduce the LC mispairing issue that has been a challenging hurdle for the development of native‐like IgG BsAbs. Our solution is particularly useful with C_H1/Cκ domains and adds to the V_H/V_L designs that we published previously.26 Prior to the engineering of the novel C_H1/Cκ designs described here, we had generated an alternative using C_H1/Cλ that was less robust in C_H1/Cκ.26, 30 Thus, to create robustly assembled IgG BsAbs using two parental mAbs both with κ LCs, our lab would generate a chimeric VκCλ LC to obtain highly specific HC/LC pairing when both parental antibodies of the IgG BsAb harbor κ LCs. This strategy worked well with medium to highly stable parental mAbs, but suffered when we used Fabs of lower than average stability. The novel C_H1/Cκ designs described here improve the ability to generate IgG BsAbs using two fully κ LCs.26, 30 The ability to add stabilizing interactions within C_H1/Cκ using the 14.3 and 15 design series, which harbor C_H1_S188A/Cκ_S176I, should enable the use of parental IgGs of modest folded stability without having to engineer them for enhanced stability. Clearly, solutions to the HC/LC pairing problem have become a highly valued commodity as many groups have published their own solutions to the problem very recently through electrostatic steering,28, 32 non‐native disulfide engineering,3 or domain swapping.27

We also demonstrate the strong influence the variable domains can have on correct/incorrect HC/LC pairing. We felt this was surprising given that (i) roughly 30% of the Fv interface consists of residues from the hypervariable HC and LC CDR3 loops; (ii) the framework region of each Fv interface is unique to each variable domain pairing and only a fraction of the V_H/Vκ interface is conserved across all germlines. Biophysically, it has been shown that the V_H/Vκ subunit often does not unfold cooperatively in isolation (as is the case for pertuzumab); however, the C_H1/Cκ subunit does unfold cooperatively.29, 41 In fact, the C_H1 domain of IgG is intrinsically unfolded and only folds in the presence of Cκ,³⁴ and thus the interaction between Cκ and C_H1 must be quite strong to induce folding of the C_H1 domain. Yet, we achieved an average 12.3% increase in specific HC/LC pairing within the four IgG BsAbs described here using only variable domain designs. Alternatively, while the C_H1/Cκ designs imparted near complete HC/LC specificity when variable domains were excised from the IgG proteins, the C_H1/Cκ designs only afforded an average 8%–10% increase in specificity in the four IgG BsAbs when variable domain specificity designs were absent. This demonstrates the significant impact the variable domain interactions can have on specific HC/LC pairing.

But what leads to the unexpectedly strong impact of the variable domains on HC/LC recognition? There are a number of factors that could come to play. First, the C_H1 domain is known to be intrinsically unfolded.34, 41 Induced folding of C_H1 by Cκ may attenuate the kinetic binding of Cκ to C_H1and allow V_H/Vκ to influence (or at times dominate) HC/LC specificity. Second, during cellular synthesis and secretion, HCs bind the chaperone BiP through the C_H1 domain to avoid aggregating prior to LC association.34 Binding of unfolded C_H1 to BiP may additionally hinder C_H1/Cκ association kinetics and provide a kinetic advantage for V_H/Vκ association to influence specific HC/LC recognition. Still, both subunits clearly had an impact on specificity. We only achieved >90% correct HC/LC assembly for BsAbs when both V_H/Vκ and C_H1/Cκ designs were present. These results suggest that designs focusing on C_H1/Cκ interactions may suffer when V_H/V_L interactions are particularly strong and benefit by the addition of variable domain specificity designs.3, 27

In conclusion, there are many applications for BsAbs. While only two BsAbs have been approved for clinical use thus far,5 many more are entering clinical trials. There are still significant lessons to be had in terms of the complexity of the biology; however, construction of BsAbs in as native an IgG architecture as possible will likely have advantages in terms of pharmacokinetics, effector function, and immunogenicity. Additionally, the monovalent nature allows the ability to affinity tune towards particular targets, which may open avenues to specific target cell engagement.2, 3, 42 smaller immune complexes, and fine tuning of the pharmacodynamic response. As BsAbs represent a much more complicated therapeutic paradigm than their monoclonal antibody predecessors, likely such subtleties will be important to intervene appropriately in various disease indications.

Materials and Methods

Computational methods

The C_H1 and Cκ domains from the 4NZU crystal structure were used as a starting point. The structure was prepared for multistate design runs using the dock‐pert protocol to optimize the side chains and the rigid‐body orientation between the two domains; the backbone degrees of freedom were held fixed.43

Multistate design (MSD) trajectories were run using the mpi_msd application in Rosetta.44 Design was allowed at 15 positions (HC residues 122, 124, 139, 143, 145, 174, 188, & 190; LC residues 116, 118, 124, 131, 133, 135, & 176) and to all amino acids, except cysteine. The objective function penalized making more than a certain threshold number of mutations, with this threshold being varied between 4 and 6 in different MSD trajectories. A negative‐state repertoire30 of 17 low‐energy docked conformations was manually selected from an initial set of MSD trajectories using only the optimized WT sequence for the off‐target interactions. Production runs included the initial backbone conformation for each of the two on‐target interactions, the initial backbone conformation and the 17 docked conformations for the two off‐target interactions, and one undocked‐conformation of the backbone for each of the four possible complexes, for a total of 42 states. The objective function computed binding energies for each of the four interactions by taking the difference between the lowest‐energy docked conformation and the energy of the undocked conformation, clipped at 0 (binding energies cannot be positive). It rewarded sequences that (1) had low total energies for the on‐target docked conformations, and (2) introduced binding energy differences between the on‐target and off‐target interactions, favoring the destabilization of both off‐target interactions more highly than complete destabilization of one interaction and partial destabilization of the other. Input files for one of the runs that produced the mutation pair in design 1.3 are given in the supplemental materials. For these runs, a variant of the Talaris2014 energy function45 was used with an atom‐pair interaction distance cutoff of 9.0Å (instead of the typical 6.0Å); reference energies were fit using the optE protocol.46 In a previous study, the incidence of charge‐based designs within the MSD simulations using the atom‐pair interaction cutoff of 6.0 Å was low30; however, it is has been shown that charge‐based destabilization of the off‐target states can be a powerful tool to drive specificity for the on‐target interactions.26, 47 Use of a 9.0Å cutoff, as benchmarked using monomer ΔΔG predictions within the original ‘score12’ energy function,48 led to more of these charge‐charge repulsions within the negative states.

Following MSD, the output structures for all four interactions were run through the dock‐pert protocol. Binding energies were computed using the interface‐analyzer application.49 Sets of mutations that appeared frequently in designs that successfully destabilized both off‐target interactions were identified and isolated. These mutation sets were threaded onto docked and undocked conformations of all four species and all eight conformations were optimized with 25 trajectories of the flexible‐backbone FastRelax protocol.50 Binding energies were computed as the difference between the average of the three lowest‐energy docked and undocked conformations. From this pool, 32 designs were selected for initial characterization along with another 14 designs produced by a slightly different computational protocol, none of which proved useful.

After initial characterization, pairs of promising designs were combined computationally, and optimized using the flexible‐backbone protocol described above. From these pairs, 9 more designs were tested.

Cloning of wild‐type Cκ/C_H1‐Fc constucts and incorporation of designs for specificity and stability screening

To interrogate the ability of C_H1/Cκ designs to provide a new and specific interface that discriminates from Wild‐Type or alternately designed C_H1/Cκ interfaces, it is useful to remove the variable domains.26 To this end, human IgG1 HC and κ LC constructs lacking variable genes were constructed in pEHG1 and pEHK vectors. The pEHG1 and pEHK were originally from Lonza and were modified in‐house for general antibody HC and LC expression.26 The Wild‐Type (no designs) pEHG1_C_H1‐Fc plasmid was generated by direct recombinase cloning of two separate and overlapping GeneBlocks (gBlocks, synthesized at Integrated DNA Technologies or IDT) and coding for the C_H1/hinge region and separately for the C_H2‐C_H3 region into pEHG1 (Lonza) using restriction sites HindIII and EcoRI (SeqID1). The Wild‐Type (no designs) pEHK_Cκ (SeqID2) and pEHK_8XHIS_Cκ (SeqID3) and construct designs were created by recombinase cloning of single gBlocks (IDT) into pEHK (Lonza) using restriction AgeI and EcoRI.

The computational and rational designs were introduced into the pEHG1_CH1‐Fc, pEHK_Cκ, pEHK_8XHIS_Cκ, and/or plasmids in one of two ways. The first procedure was site directed primer based mutagenesis. Briefly, the site‐directed mutagenesis protocol employs the supercoiled double‐stranded DNA vector and two synthetic oligonucleotide primers (generated at IDT) containing the desired mutation(s). The oligonucleotide primers, each complementary to opposite strands of the vector, were extended during thermal cycling by the DNA polymerase (HotStar HiFidelity Kit, Qiagen Cat.#202602) to generate an entirely new mutated plasmid. Following temperature cycling, the product was treated with Dpn I enzyme (New England BioLabs, Cat #R0176). The Dpn I enzyme cleaves only methylated parental DNA derived from the parental plasmid that was prepared in E. coli. Each newly generated mutant plasmid pool was then transformed into E. coli strain TOP 10 competent cells (Life Technologies). The second method employs synthesized dual or single gBlocks (IDT) containing 15 base pair 5′ and 3′ overlaps to allow recombinase‐based cloning into pEHG1_C_H1‐Fc using restriction sites AgeI and BamHI or cloning into pEHK_Cκ and/or pEHK_8XHIS_Cκ digested with AgeI and EcoRI. The recombinase‐based cloning was performed using the In‐Fusion protocol (Clontech Laboratories, Inc.) to generate the clone. The in‐Fusion reaction was transformed into E. coli strain TOP 10 competent cells (Life Technologies). Colonies were picked and clonal DNA produced by miniprepping (According to Qiagen MiniPrep protocol procedures) and sequenced at our in‐house DNA sequencing core. Medium and large scale plasmid purifications were performed according to the instructions within the Plasmid Plus Midiprep Kit (Qiagen Cat.# 12945) and Maxiprep Plus Kit (Qiagen Cat. # 12965), respectively.

Cloning of IgG BsAbs harboring novel V_H/Vκ and C_H1/Cκ designs

IgG1 BsAbs containing the novel V_H/Vκ and/or C_H1/Cκ designs, an N297Q mutation in the HC to eliminate N‐linked glycosylation and the 7.8.60 Fc heterodimerization30 were constructed in pEHG1 and pEHK vectors as described above using either of the two methods (gBlock recombinase cloning or site‐directed mutagenesis). HC and LC vectors were built using variable domains from Pertuzumab,33 MetMAb,17 Matuzumab,35 or BHA1037 as proof‐of‐concept. Plasmid isolation, sequencing and scale‐up were as described above for the constructs lacking variable domains.

Protein expression in human embryonic kidney cells (HEK293) or Chinese hamster ovary (CHO) cells

All constructs were expressed transiently either in HEK293 or CHO cells according to protocols described previously in the literature.26, 51, 52, 53 For screening the different designs described in the computational section above as well as further combinatorial designs of positive hits, (i) pEHG1_C_H1‐Fc, (ii) pEHK Cκ 8XHIS V_L Minus, and (iii) pEHK Cκ V_L Minus plasmid were all transfected transiently into HEK293F or CHO cells using a 1:1.5:1.5 plasmid ratio, respectively. As a reminder, the ‘A’ side designs (Table 1) were typically cloned into the pEHK Cκ V_L Minus plasmid while the ‘B’ side designs (Table 1) were typically cloned into pEHK Cκ 8XHIS V_L Minus plasmid. Thus, the C_H1‐Fc protein is exposed to both Cκ proteins and the 2D UPLC method described below measures whether the designs induce a preference for binding one of the Cκ proteins over the other. To identify which peak in the reverse phase elution profile belongs to Cκ versus 8XHis_Cκ, each pEHG1_C_H1‐Fc plasmid was co‐transfected only with its designed counterpart pEHK Cκ V_L Minus or pEHK Cκ 8XHIS V_L Minus plasmid using a 1:3 HC/LC plasmid ratio for every set of designs.

For IgG BsAb production, four plasmids individually harboring each of the 2 HC and 2 LC DNA inserts were transfected using 1:3 HC:LC plasmid ratios in HEK293F cells and using 1:1 HC:LC plasmid ratios for CHO as described previously.25, 26 For both HEK293 and CHO, secreted protein material was harvested by centrifugation at 5 K rpm for 5 min at the end of the culture period. Supernatants were passed through 0.22 μm filters (small scale) for either large or small scale purification.

Two dimensional high pressure liquid chromatography (2D‐UPLC) method for specificity screening

The 2D‐UPLC purification/characterization method (tandem protein G + reverse phase‐high pressure liquid chromatography (HPLC) with in‐vial reduction) was performed using Dionex Ultimate 3000 Dual Rapid Separation Liquid Chromatography.

The first dimension protein G step purifies the protein. It uses a protein G column (POROS® G 20 µm Column, 2.1 x 30 mm, 0.1 mL part # 2–1002‐00) equilibrated with 1x PBS prior to sample load. All flow rates are 1 mL/min except the final post elution column wash at 2 mL/min. 450 μL of sample (filtered cell culture media) is injected onto the protein G column. After washing with 1x PBS, the protein is eluted from the column with 100 mM sodium phosphate, pH 2.2 (2 minutes). Titers are obtained using the UV₂₈₀ peak area obtained from the protein G eluant using an in‐house IgG1 standard curve. Protein G eluted peaks are collected into vials pre‐filled with 20 μL 1M TCEP (tris(2‐carboxyethyl)phosphine) in an auto sampler held at ambient temperature (optimal condition for reduction).

The second dimension characterizes the relative ratio of each Cκ protein (‘A’ side designs in Cκ OR ‘B’ side designs in 8XHIS_Cκ, see Table 1) that binds to each C_H1‐Fc protein. The method utilizes two buffers: Buffer A being 100% H₂O and 0.05% trifluoroacetic acid (TFA); Buffer B being 100% acetonitrile (CAN), 0.05% TFA. All flow rates are 1 mL/min. The method injects each purified and reduced sample onto a Waters Symmetry C18 column, 4.6 × 75mm², 3.5µm (WAT066224) equilibrated in 95% Buffer A/5% Buffer B. Once each sample is captured onto the column, a gradient is applied starting at 10% ACN and linearly going to 40% ACN in 13 minutes. The column is then flushed with up to 70% ACN for 2 minutes and re‐equilibrated with 5% ACN for 2 minutes prior to the next injection. Comparisons by reversed‐phase chromatography are made by overlaying the control samples with the test samples. The areas under each peak are obtained to calculate percent correct assembly.

Thermal challenge assay

Enzyme‐linked immunosorbent assays (ELISAs) that detect intact protein from thermally challenged cell culture media were performed to compare the stability of the designed samples against the Wild‐Type control proteins. Briefly, 96‐well U‐bottom high protein binding 96‐well plates (Greiner bio‐one, cat#650061) were coated overnight at 4°C with 100 μL/well with 1μg/ml of sheep anti human IgG (Fd) (Meridian Life Science Cat. # W90075C‐1) in a 0.05 M NaHCO₃ buffer, pH 8.3. The plates were then washed four times with PBS with 0.1% Tween (PBST) and blocked for 1 hour with casein (Thermo Scientific, cat#37528) at 37°C. The plates were washed again and 100 μL/well of HEK293F culture supernatants containing the “variable minus” C_H1‐Fc proteins expressed with complementary designs within “variable minus” Cκ normalized to 0.1 μg/mL concentration were incubated for 1 hr at 37°C. The supernatant were pre‐exposed to various temperatures for 1 hr using a Thermal cycler with a 25°C thermal gradient window (55°C to 80°C). The plates were then washed and goat‐anti‐human Kappa‐HRP (Southern Biotech Cat.# 2060–05) at 1:8000 in casein was added and incubated for 1 hr at room temperature. The plates were then washed and 1 step Ultra TMB ELISA substrate (Thermo Scientific Cat. # 34208) was added at 50 μL/well. The reaction was allowed to proceed for 1–15minutes then quenched by the addition of (50 μL 0.25 M H₂SO₄. The absorbance at 450 nm was read using a SpectraMax 190 UV plate reader (Molecular Devices).

Protein purification and characterization

All protein purification was performed on an AKTA Explorer (GE Healthcare). Cκ/C_H1‐Fc, BsAb, and mAb proteins from 100 mL or greater in transient HEK293F or CHO were passed over a protein A column (GE Healthcare) to capture the protein, washed with phosphate buffered‐saline, eluted with 0.1 M glycine pH 3.4, neutralized with 1 M Tris‐HCl pH 8.5 and dialyzed against PBS for characterization. For quantitative binding analyses, some IgG BsAbs were subjected to a second, Capto MMC multimodal (GE Healthcare) column. The protein was bound and washed using PBS and eluted using a 0–100% gradient of PBS (buffer A) and 25 mM Na₂HPO₄, 1 M NaCl pH 6.3 (buffer B). Proteins were subsequently dialyzed into PBS prior to detailed analyses.

Analytical size exclusion chromatography (SEC) with in‐line static light scattering were performed as described previously26 to insure purity. Sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS‐PAGE) and differential scanning calorimetry (DSC) analysis were also performed as described previously.29

Crystallization and structure determination of the matuzumab Fab with and without C_H1/Cκ design 15.1

Generation of pertuzumab and matuzumab Fabs for crystallography was performed by papain cleavage from full‐length IgG1s as described previously.54 Wild type matuzumab Fab was crystallized by the vapor diffusion method with a well solution of 0.1 M tri‐sodium citrate, 20% PEG 4000, and 20% isopropanol, mixing 1.6 µL of protein with 1.6 µL of well solution. Drops were streak seeded immediately after setup with crystals previously grown at identical conditions. Crystals were harvested and frozen in liquid nitrogen after a one‐minute soak in 75% well solution supplemented with 25% (final) glycerol. Crystals of matuzumab Fab design 15.1 were obtained by the vapor diffusion method with a well solution of 0.1 M tri‐sodium citrate pH 8.0, 1.9 M ammonium sulfate, and 0.05 M sodium potassium tartrate, mixing 1.6 µL of protein with 1.6 µL of well solution. Drops were streak seeded immediately after setup and on the following day with previously grown crystals (well solution of 0.1 M tri‐sodium citrate, 2.5 M ammonium sulfate, and 0.2 M sodium potassium tartrate). Crystals were harvested and frozen in liquid nitrogen after a one‐minute soak in 76% well solution supplemented with 12% (final) glycerol and 12% (final) ethylene glycol. pertuzumab Fab containing variable domain VRD2_VRD3 was crystallized by the vapor diffusion method with a well solution of 6% glycerol, 28% PEG 4000, and 0.2 M magnesium sulfate, mixing 1.6 µL of protein with 1.6 µL of well solution. Drops were streak seeded immediately after setup and on the two following days with previously grown crystals (well solution of 10% Glycerol, 20% PEG 4000, and 0.2 M magnesium sulfate). Crystals were harvested and frozen in liquid nitrogen after a one‐minute soak in 76% well solution supplemented with 12% (final) Glycerol and 12% (final) Ethylene glycol.

Diffraction data were collected at Lilly Research Laboratories CAT, sector 31ID of the Advanced Photon Source at Argonne National Laboratory, Chicago, Illinois. The wavelength used was 0.9793 Å collecting the images on a Rayonix 225‐HE CCD detector. The images were indexed with MOSFLM or XDS and further processed with Scala and Truncate from CCP4. Phases for the initial structure determination were determined by Molecular Replacement using PHASER and as template the public domain structure of matuzumab Fab with the access code ‘3c08’ for the wild type matuzumab Fab data. The final structure of the wild type matuzumab and the public domain structure of pertuzumab with the access code ‘1l7i’ were used as a template for the Molecular Replacement solution of matuzumab Fab design 15.1 and pertuzumab containing VRD2_VRD3, respectively. The initial models were refined using Refmac5 for the matuzumab data and Buster for the pertuzumab data. Model building was performed with Coot and final structure validation with MolProbity and CCP4 validation tools.

Liquid chromatography/mass spectrometry to determine percent correct assembly of IgG BsAbs

IgG BsAb samples were purified from supernatant using the UPLC method described above for specificity screening. Instead of reducing the samples in TCEP and running the 2^nd dimension reverse phase column, the IgG samples were submitted for LCMS characterization as described previously.26

Surface plasmon resonance

Surface plasmon resonance experiments were performed on a Biacore3000 (GE Healthcare). A CM5 sensorchip was prepared by immobilizing goat anti‐human IgG‐Fc (Jackson ImmunoResearch Cat.#109–005‐098) in a 10 mM Acetate pH 5 solution to the second flowcell using standard amine coupling methods. The first flowcell was blocked after activation using ethanolamine and used as reference. IgG BsAbs at 100 nM in HBS‐EP buffer (GE Healthcare) were captured onto the chip by 40 μL injection at 20 μL/min. Soluble, monomeric human EGFR and HER‐2 (or vice versa) at 25 nM concentrations were injected (80 μL at 20 μL/min) back‐to‐back 9 minutes apart to demonstrate the bispecific binding activity of the IgG BsAbs. Injections of IgG BsAb followed by blank injections (HBS‐EP) were used for double referencing. The chip surface was regenerated by setting the flow rate to 50 μL/min followed by two 5 μL injections of 0.1 M Glycine pH 2.0 spaced 1 minute apart. Data was processed using the BiaEvaluation Software.

Protein Data Bank Accession Codes

The atomic coordinates and structure factors of the three Fab structures described in this report were submitted to the RCSB Protein Data Bank, with accession codes 5VR9 (matuzumab Fab_WT), 5VSI (matuzumab Fab_15.1), and 5VSH (Pertuzumab Fab_VRD2_3).

Conflict of Interest Statement

A. Leaver‐Fay and B. Kuhlman are equity holders in Dualogics LLC, a company aimed at commercializing bispecific antibodies. K.J. Froning, X. Wu, S. Phan, F. Huang, A. Pustilnik, M. Bacica, Q. Chai, J.R. Fitchett, J. Hendle, and S.J. Demarest are employees of Eli Lilly, a biopharmaceutical company.

Supporting information

Supporting Information