The structural coordinates of C_H1-C_L (PDB:3BKY) and C_H3:C_H3 (1L6X) were used to determine inter-chain inter-residue contacts including putative hydrogen bonds (including water-bridged ones), disulfide bonds, pi-bonds, polar interactions, salt bridges, and Van der Waals interactions (non-hydrogen) occurring between pairs of residues as described previously (21). These data were assembled into an array of eight atomic interaction matrices. A weighted sum of the eight atomic interaction matrices were computed to produce a single matrix that accounts for the strength of atomic interaction between residue pairs, using weights derived from relative atomic interaction energies. The inter-residue interaction network calculated in this fashion generates a matrix that describes all the contacts made between pairs of residues. Each element i, j is the sum of the path scores of all paths between residues i and j. The degree of networking score for each residue was computed by summing across the rows of the matrix, which corresponds to the extent of “networking” for each residue. The degree of networking score was normalized (SIN score) with the maximum score for each protein so that the score varies from 0 (absence of any network) to 1 (most networked). In alignment with our previous study, residues whose SIN score is above 0.4 is considered highly networked.

Bispecific interfaces (PDB: 1N8Z, 3P11 for C_H1-C_L, PDB: 1L6X for C_H3-C_H3) were analyzed with the help of 2D graph-based connectivity networks, a concept we had developed earlier (22). Although it was developed to analyze epitope-paratope interactions (22), the same methodology is applicable to other protein-protein interfaces. Basically, the inter-residue interactions between an interfacial residue (e.g., C_H1) and its neighboring residues on its partnering domain (e.g., C_L) are rendered in a 2D graph format as a visual aid, where nodes represent amino acids, and the edges represent inter-residue non-covalent interactions. Interactions considered include putative hydrogen bonds (including water-bridged ones), disulfide bonds, pi-bonds, polar interactions, salt bridges, and Van der Waals interactions (non-hydrogen) as described previously (21). Mutations or positions (when mutated) that contribute to more favorable or unfavorable interactions, as evaluated by this network analysis, were selected for experimental investigation.

Site directed mutagenesis was performed using QuikChange II XL (Agilent Technologies). Sequence verified plasmid was amplified using PureLink HiPure Maxiprep Kit (Life Technologies).

In this study, we designed two BsAbs, i.e., anti-EGFR/HER2 BsAb and anti-CD20/CD20 BsAb. The anti-EGFR/HER2 BsAb was constructed by combining a light and heavy chain of pertuzumab (PTAB), an anti-HER2 antibody, and a light and heavy chain of DL11, an anti-EGFR/HER3 antibody. The anti-CD20/CD20 BsAb was constructed by combining a light and heavy chain of rituximab (Rxm), a type I anti-CD20 antibody, and obinutuzumab, a type II anti-CD20 antibody. For improving the BsAb production, we mutated amino acids in C_H1/C_L interface and C_H3 constant regions. The DNA fragments of the desired heavy chains and light chains were generated by gene synthesis and cloned into pcDNA 3.3 expression vector. For the parental monoclonal antibodies, we cloned DNA coding pertuzumab or rituximab into pcDNA 3.1 (+) expression vector. The sequence of the inserted DNA fragment was confirmed by DNA sequencing.

We used the ExpiFectamine 293 transfection kit for transient transfection. First, HEK cells were seeded at 3.0 × 10⁶ cells/mL and incubated for 24 hours. After incubation, cells were harvested and diluted to 2.5 × 10⁶ cells/mL. For the preparation of transfection reaction, 80 µg of plasmid DNA was diluted in 4 mL of Opti-MEM medium (Gibco). In a separate tube, 219 µL of ExpiFectamine 213 reagent was diluted in 4 mL of Opti-MEM medium (Gibco). After the mixture was incubated for 5 minutes, the diluted DNA was added. Then, DNA-reagent complex was incubated for 20 minutes and added to HEK cell suspension. After incubation for 16-18 hours, 400 µL of Transfection Enhancers 1 and 4 mL of Transfection Enhancer 2, which were provided in transfection kit, were added. Cells were incubated for 6 days before harvesting.

After transient transfection, antibodies in the cell culture supernatant were collected and filtered through the 0.2 µm Thermo Scientific Nalgene Rapid-Flow disposable filter unit with PES membrane (ThermoScientific). Antibodies were purified using HiTrap MabSelect Sure affinity column (GE Healthcare) using an AKTA Pure chromatography system (GE Healthcare). First, cell culture supernatant was loaded onto HiTrap MabSelect Sure affinity column pre-equilibrated with binding buffer (0.02 M phosphate, 0.15 M NaCl, pH 7.4). Protein was eluted using elution buffer (0.1 M sodium citrate, pH 3.0). Eluted fractions were collected in tubes containing 10% neutralization buffer (1 M Tris HCl, pH 8.8). For buffer exchange, protein sample was added into the filter column (30 kDa Ultra-4 Centrifugal Filter Units, Amicon) pre-washed with phosphate buffer saline, pH 7.4 (Gibco). Then, the filter column was centrifuged at 2800 g for 15 minutes at 4°C. This step was repeated until all of sample was loaded into filter column. Then, PBS was added into the filter column; and the filter column was centrifuged at 2800 g for 15 minutes at 4°C. This step was repeated until the elution buffer is less than 1%. After buffer exchange, the purified IgG was filtered using 0.2 µm filter in type II biological safety cabinet to obtain sterile antibodies for cell-based experiment.

The integrity of the antibody was also examined using reducing and non-reducing SDS-PAGE. For reducing condition, 600 ng of proteins in glycine/SDS solution (50 mM Tris-HCl pH 6.8, 2% SDS, 10% glycerol, 1% β-mercaptoethanol, 0.02% bromophenol blue) with 12.5 mM DTT were loaded on NuPAGE 4 to 12%, Bis-Tris, 1.0 mm, Mini Protein Gel, 12-well (ThermoFisher SCIENTIFIC). For non-reducing condition, 400 ng of proteins in glycine/SDS solution without DTT were loaded on NuPAGE 4 to 12%, Bis-Tris, 1.0 mm, Mini Protein Gel, 12-well (ThermoFisher SCIENTIFIC). Each sample was subjected to electrophoresis at 100 mA for approximately 120 min. Protein bands were visualized using InstantBlue^® Coomassie Protein stain (abcam) followed by destaining until the background was clear. The protein bands were visualized by ChemiDoc Touch Imaging System (BIO-RAD).

In this study, BsAbs and parental mAbs were subjected to size exclusion chromatography (SEC). Antibodies were diluted to the concentration of 5 mg/mL and filtered with 0.2 µM Nanosep MF Centrifugal Devices with Bio-Inert Membrane (PALL). SEC was performed using Agilent High Performance Liquid Chromatography (Agilent Technologies) consisting of a binary pump with a degasser, an autosampler was maintained at 4°C and a photo diode array detector. A TSKgel G3000SW_XL HPLC column (0.78 × 30 cm) (TOSOH CORPORATION) was operated at 25°C. The pre-filtered mobile phase was 0.2 M potassium phosphate buffer with 0.25 M potassium chloride at pH 6.2 with constantly at 0.7 mL/min. Antibodies were eluted isocratically over 11-14 minutes. The percentage of monomer and aggregates were calculated by area under the curve using OpenLab CDS 3.4 (Agilent). Mix standard proteins (ovalbumin; 43 kDa, p-aminobenzoic acid; 137.36 kDa ribonuclease A; 13.7 kDa, thyroglobulin; 660 kDa) were used for MW calculation of the antibodies.

Bispecific antibodies (BsAbs) were deglycosylated using Rapid PNGase F enzyme (NEW ENGLAND BioLabs) by incubating at 37°C for 18 hours prior intact mass analysis. Intact mass spectral data of deglycosylated BsAbs were acquired using an Orbitrap Q-Exactive HF Mass Spectrometer coupled to an Ultimate 3000 UHPLC system (Thermo Scientific). Buffer A is 0.1% formic acid in water and buffer B is 0.1% formic acid in acetonitrile. Approximately 3 µg of BsAbs were loaded to MAbPac RP Column (4 µm, 2.1 mm x 50 mm, Thermo Scientific) preconditioned with 20% buffer B. The first 2 minutes of gradient were maintained at 20% buffer B. The separation was attained using solvent gradient ramping from 20% to 65% buffer B over 4.5 minutes. Washing was done using 90% buffer B for 5 minutes. Column was then re-equilibrated with 20% buffer B for 5.5 minutes. The flow rate was set to 0.3 mL/min throughout all gradient steps. The column temperature was maintained at 80°C throughout all gradient steps. The HESI-II ion source parameters were set as following; spray voltage 4.20 kV, capillary temperature 325°C, S-lens RF 70%, sheath gas flow 50 units, and auxiliary gas flow rate 15 units. The data acquisition parameters were set as following; scan range m/z 800-4500, resolution 15,000, positive mode, FT Full scan AGC target 3×10⁶, fixed AGC mode, and maximum injection time 150 ms. BioPharma Finder version 3.0 (Thermo Scientific) software was used to deconvolute intact mass spectrum. An average molecular weight of BsAbs and their corresponding mispairing were calculated using Agilent MassHunter Sequence Manager B.09.00 (Agilent Technologies). Lysine clipping and glycine loss at C-terminus were considered as they are common variants found for antibody.

The DSC measurements were obtained using Differential Scanning Calorimeter 3+ (DSC3+) Auto Sample Robot (Mettler Toledo). Antibodies were concentrated at 15 mg/mL in phosphate buffer pH 7.4. Antibodies were scanned from 30 to 95°C at a heating rate of 1°C/minute. The corresponding buffer was used as a baseline and subtracted from sample data. The thermal transition midpoint (T_m) was obtained from the maximum temperature of the DSC thermogram. Data analysis was done using STARe Default DB V.16.

The Far-UV CD spectroscopy was performed to analyze secondary structure of antibodies using a Jasco J-815 spectropolarimeter (Jasco). The UV CD spectra were obtained in the intervals of 190-250 nm, in a quartz cuvette with a path length of 0.1 cm. The spectra were collected by continuous scanning at 0.1 nm intervals, a scanning speed of 50 nm/min, a response time of 1 s, a bandwidth of 1 nm, and 1 accumulation. The protein concentration was 0.1 mg/mL. The baseline was corrected in all experiments using a milliQ water as control.

Dual-binding ELISA for EGFR and HER2 was developed to determine the binding activity of anti-EGFR/HER2 BsAb. A pre-absorbed Maxisorp 96-well plates (ThermoFisher SCIENTIFIC) was coated with 1 μg/mL of human EGFR protein (Sino Biological) in phosphate-buffered saline (PBS) and incubated at 4°C overnight. Plate was washed three times with PBS. A blocking solution of 1% BSA in PBS plus 0.05% tween (PBST) was added. Then, the plate was incubated at room temperature for 1 hour and washed three times by PBST. Serial dilution (2.5-fold) of the anti-EGFR/HER2 BsAb was done with the starting concentration of 8 µg/mL for 11 dilutions. BsAb was diluted in PBS and in duplicate with 3 different batches of production. The plate was incubated at room temperature for 2 hours and then washed with PBST. Incubated plate was then wash three times with PBST and added with 100 µL of 1 µg/mL human HER2 protein (Sino Biological) to each well. Plate was incubated at room temperature for 2 hours and washed three times with PBST. After that, 0.1 µg/mL of secondary antibody (rabbit monoclonal anti-human HER2 antibody (HRP) (Sino Biological) was added. Plate was incubated at room temperature for 1 hour and then washed with PBST. For detection, TMB substrate (SeraCare) was added and incubated for 5 minutes. The reaction between TMB substrate and HRP on secondary antibody was stopped by adding 1N sulfuric acid. Plate was read using a microplate reader (SpectraMax iD3) at 450 nm.

HER2 in phosphate-buffered saline at concentration of 1 μg/mL was immobilized on Maxisorp 96-well plate (ThermoFisher SCIENTIFIC) at 4°C overnight. The antigen coated plate was then washed with PBS three times to remove excess antigens. Plate then was blocked with 1% BSA in PBST for 1 hour following by incubation with serially diluted antibodies with the starting concentration of 8 µg/mL (2.5-fold) for 2 hours. A total of 100 µL/well of goat anti-human IgG H&L (abcam; 1/12500 dilution) was added to plate and incubated for 1 hour. Washing 3 times with PBST was performed after each incubation step. After final wash, the plates were developed with TMB (3,3’,5,5’-tetramethylbenzidine) substrate (SeraCare). After the reaction between TMB and HRP was stopped by 1N sulfuric acid, absorbance at 450 nm was recorded using a microplate reader (SpectraMax iD3).

SPR studies were performed on a Biacore T200 instrument (GE Healthcare). Each Fc receptor was captured on a pre-immobilized with His capture kit (Cytiva) Biacore chip for kinetic studies of bispecific antibodies. Receptors used in this study were FcɣRIIIa V and F forms, FcRn, and C1q. For FcɣRIIIa binding, either 0.2 µg/mL human FcɣRIIIa V form (Human CD16a/FcɣRIIIa Protein (176 Val, His Tag), Sino Biological) or F form (CD16a Protein, Human, Recombinant (176 Phe, His Tag), Sino Biological) in Biacore buffer HBS-EP+ (10 mM HEPES, pH 7.4, 150 mM NaCl, 3 mM EDTA, 0.05% Surfactant P20, Cytiva) was run through a pre-His capture immobilized series S sensor chip CM5 (Cytiva). Later, five concentrations of each BsAb in Biacore buffer HBS-EP+ was flowed through the FcɣRIIIa-immobilized CM5 sensor chip. The binding was measured in association and dissociation phases at 30 µL/min.

For FcRn binding, Human FcRn (Recombinant Human FcγR/FcRn protein, Sino Biological) was diluted to 0.25 µg/mL in Biacore buffer HBS-EP+ (Cytiva) and was run through a pre-captured series S sensor chip CM5 (Cytiva). Serial dilution (2-fold) of BsAbs were done with the starting concentration of 2000 nM for 5 dilutions in Biacore buffer HBS-EP+ (10 mM HEPES, pH 6, 150 mM NaCl, 3 mM EDTA, 0.005% Surfactant P20, Cytiva) and run kinetic binding through previously FcRn-immobilized sensor chip. The binding was measured in association and dissociation phases at 30 µL/min. For C1q binding of anti-CD20/CD20 BsAb, human C1q (Native Human C1q protein, abcam) was diluted to 120 µg/mL in immobilization buffer (Cytiva) by using amine coupling kit (Cytiva) and was immobilized on a series S sensor chip CM5 (Cytiva). Serial dilution (2-fold) of anti-CD20/CD20 BsAb was done with the starting concentration of 8000 nM for 5 dilutions in Biacore buffer HBS-EP+ (10 mM HEPES, pH 7.4, 150 mM NaCl, 3 mM EDTA, 0.005% Surfactant P20, Cytiva) and run kinetic binding through previously C1q-immobilized sensor chip. The binding was measured in association and dissociation phases at 30 µL/min.

The C1q binding of anti-EGFR/HER2 BsAb was done by a sandwich Biacore assay. The anti-EGFR/HER2 BsAb was diluted in Biacore buffer HBS-EP+ pH 7.4 (Cytiva) to a concentration of 1 µg/mL and was captured on a sensor chip Protein L (Cytiva). Serial dilution (2-fold) of human C1q protein was done with the starting concentration of 5 nM for 5 dilutions in Biacore buffer HBS-EP pH 7.4 (Cytiva) and was flowed over the previously captured chip at 30 µL/min. All sensogram data were processed and analyzed by Biacore Evaluation software (GE Healthcare).

We employed the structures of IgG1κ anti-HER2 (PDB: 1N8Z) and anti-HER1/3 (PDB: 3P11) antibodies to design and screen for Fab (i.e., C_H1-C_L) and Fc (C_H3-C_H3) mutations. These two antibodies were chosen since their heavy and light chains possessed an intrinsic ability for heterogeneous pairing, thus presenting an excellent test case to screen for mutations that enable selective chain pairing preferences. Since the predicted Fab and Fc mutations were studied separately, the screening process for Fab mutations made use of WT Fc, and similarly, the screening process for Fc mutations employed WT Fab.

The C_H1-C_L and V_H-V_L domain pairs are each related by a pseudo-2-fold symmetry. Therefore, we attempted to engineer Fab interface mutations that would introduce asymmetry in the mispaired heavy-light chain interfaces, while retaining WT-like symmetry in the cognate heavy-light chain interfaces. With this design approach, it is expected that the heavy (or light chain) of the first antibody would pair with its cognate chain with higher affinity and specificity than to the non-cognate chain of the second antibody. To aid in the design process, we chose to mutate C_H1-C_L interface residues ( Supplementary Table 3 ) that were 1) highly interconnected by a network of non-covalent interactions or have high SIN scores (Methods), 2) proximal to the pseudo-2-fold axis ( Supplementary Table 4 ), and 3) evolutionarily conserved across Ig genes. Criteria 1 & 2 ensure that the mutations have a profound impact on the interface thus minimizing the need to make mutations in the variable domains. Criteria 3 ensures that the mutations are not specific to the system under consideration but are transferrable to other antibody combinations and isotypes. Using computational modeling, we explored the possibility of engineering two sets of oppositely charged residues, one on each Fab, such that it would create a repulsive charge-charge interaction (asymmetry) in the context of heavy-light chain mispairing. For example, the interacting residue pair Leu150 (C_H1) – Val132 (C_L) in the WT interface would be substituted Asp150 (C_H1′) – Lys132 (C_L′) in the first Fab and Arg150 (C_H1′′) – Asp132 (C_L′′) in the second Fab. In this scenario, the interface between mispaired heavy-light chains will be destabilized as a result of bringing two like charges into close proximity. Likewise, we also explored the possibility of introducing two sets of “knobs” and “holes” in place of a C_H1-C_L residue pair, one on each Fab, which would lead to a sterically unfavorable interaction (asymmetry) in the context of heavy-light mispairing. Then, we combined two or more C_H1-C_L mutation sets, where structurally feasible, to generate additional sets of C_H1-C_L mutations. A total of 44 C_H1-C_L mutations (referred henceforth as “clusters”) were generated with the help of the structure-guided process (Methods) ( Supplementary Table 5 ).

Next, we assessed the potential of a predicted C_H1-C_L cluster to promote preferential pairing between cognate chains by expressing the two correctly paired and two mispaired monospecific antibody species ( Supplementary Figure 2 ) and then quantifying their expression using IgG ELISA ( Supplementary Table 5 ). Supplementary Table 5 captures the expression levels of matched and mismatched antibodies relative to their unmodified versions. From this, the cluster that led to maximum expression difference between the correctly paired and mispaired species was identified ( Supplementary Table 5 ). In many cases, mutations that lowered the expression levels of the mispaired IgG species also caused moderate-to-significant impact on the expression of the correctly paired IgG species (e.g., clusters 1, 17, 18, 43 in Supplementary Table 5 ). In other cases, mutations that maintained the expression levels of the correctly paired IgG species were not detrimental to the mispaired species (e.g., clusters 23, 30, 35 in Supplementary Table 5 ). Cluster 40 yielded by far the most favorable results (in Supplementary Table 5 and Table 1 ). In this context, the correctly paired-IgG species expressed at 87% (C_L′-C_H1′) & 63% (C_L′′-C_H1′′) of their unmodified versions, whereas the mispaired-IgGs expressed at much lower levels of 0% (C_L′′-C_H1′) and 48% (C_L′-C_H1′′) ( Figures 1A, B ). Critically, one of the mispaired species represented by C_L′′ and C_H1′ chains expressed remarkably poorly, indicating the potential of this cluster to reduce or eliminate the formation of n=3 mispaired species from the mixture of ten ( Supplementary Figure 1 ). This preferential pairing is likely due to the (1) repulsive interactions created by two like-charged residues– Lys152Asp (C_H1′′)-Gln123Asp (C_L′), His173Asp (C_H1′′)-Asn136Asp (C_L′) ( Figure 1C ). Likewise, structural modeling indicates that the interface between C_H1′- C_L′′ will be destabilized by two unfavorable interactions involving like-charged residues (1) Lys152 (C_H1′) and Gln123Lys (C_L′′) and (2) His173 (C_H1′) and Asn136Lys (C_L′′) ( Figure 1D ). In the context of the cognate heavy-light interface (C_L′′- C_H1′′)κ, Lys152Asp and His173Asp of C_H1′′ will form favorable interactions with the oppositely charged Gln123Lys and Asn136Lys of C_L′′, respectively ( Figure 1E ). Likewise, the interface formed by C_L′ and C_H1′ will be stabilized by the favorable electrostatic interactions between (1) Lys152 (C_H1′) and Gln123Asp (C_L′) and (2) His173 (C_H1′) and Asn136Asp (C_L′). The Leu133Val and Leu150Ala mutations on C_H1′ create a void to accommodate the necessary Gln123Asp mutation on C_L′ ( Figure 1F ).

Next, we evaluated the specificity afforded by the C_H1-C_L mutations using a 3-chain competition assay where two heavy chains (cognate and non-cognate) were forced to compete with a light chain for assembly (Methods). The pairing preference between the chains was evaluated using LC-MS. In the absence of the C_H1-C_L mutations, 3 species are expected to form, a WT antibody, a mispaired monospecific and a common light chain BsAb. Along similar lines, competition experiments were conducted using two light chains and a heavy chain, where three species are expected to form in the absence of C_H1-C_L mutations (WT antibody, mispaired monospecific and a common heavy chain BsAb). Mutations L133V, L150A (C_H1′) and Q123K, N136K, T177A (C_L′′) completely prevented heavy-light mispairing, suggesting its potential to reduce the number of mispaired species in the context of the actual hetero-tetrameric BsAb assembly ( Supplementary Table 6 ).

Several solutions have been proposed to prevent the homodimerization of Fcs. The most prominent is the knob-into-holes mutations (19), which generates antibody heavy chains for heterodimerization, using steric complementarity.

Our design strategy centers around the four ionic interactions in the C_H3-C_H3 interface (Asp356–Lys439, Glu357–Lys370, Lys392-Asp399 and Asp399–Lys409) (23). Each ionic interaction pair occurs twice in the C_H3-C_H3 interface as a result of the two-fold symmetry. The designs aim to disrupt the ionic interaction pair within the homodimers through a hydrophobic substitution. For instance, Asp356–Lys439 was made Asp356–Phe439 in the first antibody and Val356–Lys439 in the second antibody. The switch in the side chain character (ionic to hydrophobic) is expected to reduce the propensity for homodimerization, while favoring heterodimerization. The newly formed Fc heterodimer will contain one ionic and one hydrophobic pair instead of two ionic interactions as seen in a homodimer. Similarly, we made attempts to create asymmetry by introducing charged residues instead of hydrophobic. Where possible, designs from the two strategies were combined together in order to generate profound effects on Fc pairing.

Our criteria for screening the Fc mutations was based on the observation that transient expression of a heavy and light chain carrying the traditional Fc knob mutation (19) generates half-antibody (H-L) fragments along with the full IgG ( Figure 2A ). Using the same principle, we determined potential knob-like mutation(s) from the list of predicted mutations based on the ability to induce the formation of half-antibody fragments. Coomassie gel showed the Fc mutations Glu357Lys, Lys409Arg formed high amounts of half-antibody fragments, suggesting that these mutations possess knob-like features ( Figure 2A ). Keeping the knob mutation fixed, we then evaluated potential compensatory mutations (hole-like) based on their ability to reduce the formation of half-antibody fragments (via heterodimerization) in a competition 3-chain system (2 heavy, 1 light). Our analysis showed that maximum reduction in half-antibody fragments occurred in the presence of compensatory mutation, Lys370Glu ( Figure 2B ). Structural modeling of these mutations on the C_H3-C_H3 interface provides a rationale for the preferential heterodimer formation. The formation of homodimers (C_H3′-C_H3′; C_H3′′-C_H3′′) is prevented by the repulsive charges that are brought close together ( Figures 1G, H ) whereas the heterodimer formation (C_H3′-C_H3′′) is aided by favorable ionic interactions between closely-packed residues ( Figure 1I ). Hence, these two sets of mutations were employed for preventing homodimerization ( Figures 1G–I ).

Our structure-guided design process introduces mutations in the conserved regions of an IgG antibody. In order to assess the potential impact of the engineered mutations on increasing immunogenicity risk, we employed NetMHCIIpan version 4.0 (24) to predict MHC II binding sites on the WT and modified constant region sequences for common HLA class II alleles: DRB1_0101, DRB1_0701, DRB1_0901, DRB1_1101, DRB1_1201, DRB1_1501, DRB1_0401, DRB1_0405, DRB1_0301, DRB1_1302, DRB3_0101, DRB3_0202, DRB4_0101, and DRB5_0101 (25). The software was run using default settings. Peptides of length n=15 were searched. The software compares the predicted affinity of a given peptide with the scores of 200,000 random natural peptides of the same length and computes a percentile rank. Then, it classifies the peptide as a weak binder or a strong binder if the percentile rank is below 5% or 1%, respectively. Supplementary Table 7 lists the number of weak and strong binders in the WT and modified constant regions. Remarkably, the designs led to a net decrease in the number of weak and strong binders ( Supplementary Table 7 ). Notably, mutations in 3/4 chains led to a decrease in MHC II binders. On the other hand, Q123K, N136K, and T177A in the light chain led to a modest increase in the number of MHC II binders; additional studies with therapeutic candidate molecules will be needed to confirm whether these mutations will have a detrimental impact (e.g., anti-drug antibody response).

To determine whether these mutations will have any impact on solubility, we employed CamSol (26) to predict the solubility of WT C_H1-C_L and C_H3-C_H3 regions and their mutants. CamSol has been successfully employed for rational design of protein variants having improved solubility (26). The goal of this exercise was to determine whether the introduced changes will lead to a decrease in solubility character. CamSol was run using the homology models of different constant region interfaces (C_H1-C_L (WT), C_H1′-C_L′, C_H1′′-C_L′′, C_H3-C_H3 (WT), C_H3′-C_H3′′) under default parameter settings: pH 7 and patch radius 10. Interestingly, the structurally-corrected intrinsic solubility scores of the modified interfaces were higher than that of the native interface, suggesting that the introduced mutations will not lead to reduced solubility (or increased hydrophobicity). Specifically, the solubility scores for C_H1-C_L (WT), C_H1′-C_L′, C_H1′′-C_L′′, C_H3-C_H3 (WT), and C_H3′-C_H3′′ were found to be -0.055342, 0.103082, 0.114590, 0.572991, and 0.609848, respectively.

We employed anti-EGFR/HER2 and anti-CD20/CD20 as two systems to study and understand the utility of the mutations ( Supplementary Figure 3 ). For the anti-EGFR/HER2 BsAb, we employed the variable domain sequences of pertuzumab (anti-HER2) (27) and duligotuzumab/DL11 (anti-EGFR/HER3) (28). Previous studies have shown that simultaneous targeting of EGFR, HER2, and HER3 with multispecific antibodies can prevent compensatory tumor promoting mechanisms within the HER family, providing a scientific rationale for the development of a pan-HER antibody (29). For the anti-CD20/CD20 BsAb, we used the variable domain amino acid sequences from the type I anti-CD20 monoclonal antibody, rituximab (30), and a type II anti-CD20 monoclonal antibody, obinutuzumab (31).

Briefly, the EGFR/HER2 and CD20/CD20 BsAbs were expressed in ExpiFectamine 293 transfection system and purified using protein A column chromatography as described in Methods. Supplementary Table 8 contains the amino acid sequences of the two bispecific antibodies that were expressed. After purification, protein integrity was determined by SDS-PAGE together with Coomassie blue staining. Under non-reduced conditions, anti-EGFR/HER2 and anti-CD20/CD20 BsAbs showed a major band at around 150 kDa, which was similar to the parental mAbs, ( Figure 3A ). Under reducing conditions, both BsAbs bands around 50 kDa and 25 kDa corresponding to the heavy and light chains, respectively. As expected, the parental mAbs also showed two bands corresponding to the heavy and light chains ( Figure 3B ). Analytical size exclusion chromatography (SEC) was performed to assess the homogeneity of the bispecific antibodies as well as parental mAbs. Both BsAbs were predominantly (>95%) monomeric in content. All antibodies eluted from the column with retention times that are consistent with their expected molecular weight of around 144-145 kDa ( Figure 4 ).

To further characterize the BsAbs, intact mass analyses were performed using mass spectrometry. In case of the anti-EGFR/HER2 BsAb, the observed mass of this main peak (144597.05 Da) ( Figure 5A ) was close to the theoretical calculated mass for the anti-EGFR/HER2 BsAb with C-terminal lysine clipping (144599.10 Da) with 14.18 ppm mass difference. We also observed a peak corresponding to a variant (glycine loss) of the BsAb. Taken together, the purity of the correct forms of anti-EGFR/HER2 BsAb was 84.8% ( Figure 5A ). In addition to these two peaks, we observed four distinct peaks corresponding to other species: two monospecific antibodies (0.6%), single (1.4%) and dual (13.3%) pertuzumab light chain mispairs ( Figure 5A ). In the anti-CD20/CD20 BsAb example, the mass of the major peak was 145146.85 Da with 15.64 ppm difference from the theoretical calculated mass of the BsAb with C-terminal lysine clipping (145149.12 Da) ( Figure 5B ). Like in the case of the anti-EGFR/Her2 BsAb, we observed a neighboring peak corresponding to a glycine loss variant of anti-CD20/CD20 BsAb. Together, the purity of the correct forms of anti-CD20/CD20 BsAb was 78.3% ( Figure 5B ). Besides these, we also observed peaks corresponding to the two parental monospecific antibodies which accounted for a total of 21.7% of the sample ( Figure 5B ). While we observed pertuzumab light chain mispairs (but not DL11 light chain mispairs) in the case of anti-EGFR/Her2 BsAb, the profile for the anti-CD20/CD20 BsAb shows complete absence of light chain mispairs. This observation also indicates that the chances of formation of dual light chain mispaired species, which has the same mass as the BsAb, are unlikely in both cases. The byproducts can be controlled or minimized through additional Fab/Fc interface engineering, alteration of chain transfection ratios (20), introduction of protein A ablation mutation (32), pI engineering (33), and/or other secondary purification techniques.

The secondary structure of anti-EGFR/HER2 and anti-CD20/CD20 bispecific antibodies was assessed by far-UV CD analysis (200 to 240 nm). The spectrum of both BsAbs showed a similar negative maximum at approximately 218 nm similar to their parental mAbs ( Figure 6 ). These results demonstrate that anti-EGFR/HER2 and anti-CD20/CD20 bispecific antibodies predominantly contain beta-sheet consistent with the quaternary structure of an IgG molecule (34–36).

Next, the thermal stability profile of the two purified BsAbs was assessed by using differential scanning calorimetry (DSC) and compared to their parental mAbs. A mid-point temperature (T_m) of BsAbs, which represents the temperature at which 50% of the protein is in its native conformation and 50% in denatured conformation, was used as an indicator for thermal stability. As shown in Figure 7A , the T_m values of the anti-EGFR/HER2 BsAb were observed to be 77.63°C for the C_H2/Fab domains and 84.83°C for the C_H3 domain, which were similar to the parental monospecific antibody pertuzumab, which has the wild-type sequence (C_H2/Fab: 76.32°C; C_H3: 83.18°C). Another (third) peak with T_m at 67.52°C was found in the anti-EGFR/HER2 ( Figure 7A ). This is likely to correspond to the Fab/C_H2 region of the second parental monospecific antibody DL11 and/or the contaminants carrying the heavy chain of DL11 and light chain of pertuzumab ( Figure 5A ). This is supported by the fact that DL11 is reported to have lower T_m than pertuzumab (37). For the anti-CD20/CD20 BsAb, the T_m values were observed at 73.32°C for C_H2/Fab domains and 81.27°C for C_H3 domain, which were similar to the parental monospecific antibody rituximab, which has the wild-type sequence (C_H2/Fab: 74.42°C; C_H3: 82.02°C) ( Figure 7B ). Taken together, the T_m values of the two BsAbs are similar to their parental mAbs, as well as, in the range observed for naturally occurring antibodies, indicating that the mutations do not lead to significant destabilization (38, 39).

We evaluated the dual antigen binding capabilities of the BsAb using a sandwich ELISA, where the detection signal from the secondary antibody is obtained only when the BsAb is bound to both target antigens (Figures 8A, B) (Methods). As shown in Figure 8B , the anti-EGFR/HER2 BsAb bound to their target proteins in a dose-response manner. These results indicate that the modifications made to the constant regions did not impact the Fab-based binding activity. The anti-CD20/CD20 (type ½) was not amenable to a sandwich ELISA. We also performed the HER2-binding ELISA to confirm that anti-EGFR/HER2 BsAb can bind to HER2 antigen Figures 8C, D. The EC₅₀ values of BsAb and pertuzumab were 0.41 and 0.26 nM, respectively ( Figure 8D ). The HER2 binding EC₅₀ values of the bispecific antibody was 1.6 times weaker than that of pertuzumab, indicating no major loss in binding as a result of the new format.

Alterations to the Fc peptide backbone or glycan structure are known to critically affect the effector properties (CDC and ADCC) and FcRn-dependent pharmacokinetics of IgG1 therapeutic antibodies (40). To determine whether the introduced mutations had any impact on binding to the effector molecules, surface plasmon resonance (SPR) was used to measure the binding activity of the two BsAbs to Fc receptors including FcɣRIIIa V and F forms, FcRn, and C1q. The binding of anti-EGFR/HER2 BsAb is shown in Figure 9A . The equilibrium dissociation constants (K _D) for binding to immobilized human FcɣRIIIa V form (V158) and F form (F158) were 185 ± 5.3 nM and 735.8 ± 27.4 nM, respectively. The K _D of FcRn binding to anti-EGFR/HER2 BsAb was 11.5 ± 2.5 nM. To measure complement binding, soluble human C1q was run though the BsAb-captured protein-L chip. The K _D was measured at 6.4 ± 0.3 nM. Binding of soluble Fc receptors and C1q to human IgG1 monospecific antibodies has been extensively studied (41–45). The equilibrium dissociation constants (K _D) we determined match closely with literature (41–45), implying that the format change did not affect binding to Fc receptors and C1q. The binding affinity of Fc receptors and C1q to the anti-CD20/CD20 BsAb was found to be in a similar range, as expected ( Figure 9B ).