A comparative study of the developability of full-length antibodies, fragments, and bispecific formats reveals higher stability risks for engineered constructs

ABSTRACT

Engineered antibody formats, such as antibody fragments and bispecifics, have the potential to offer improved therapeutic efficacy compared to traditional full-length monoclonal antibodies (mAbs). However, the translation of these non-natural molecules into successful therapeutics can be hampered by developability challenges. Here, we systematically analyzed 64 different antibody constructs targeting Tumor Necrosis Factor (TNF) which cover 8 distinct molecular format families, encompassing full-length antibodies, various types of single chain variable fragments, and bispecifics. We measured 15 biophysical properties related to activity, manufacturing, and stability, scoring variants with a flag-based risk approach and a recent in silico developability profiler. Our comparative assessment revealed that overall developability is higher for the natural full-length antibody format. Bispecific antibodies, antibodies with scFv fragments at the C-terminus of the light chain, and single-chain Fv antibody fragments (scFvs) have intermediate developability properties, while more complicated formats, such as scFv- scFv, bispecific mAbs with one Fab exchanged with a scFv, and diabody formats are collectively more challenging. In particular, our study highlights the propensity for fragmentation and aggregation, both in bulk and at interfaces, for many current engineered formats.

Introduction

As of May 2024, 127 unique antibodies were approved by the US Food and Drug Administration (FDA) as treatments for diseases such as immunological disorders and cancers.¹ While approximately 80% of these molecules are canonical monoclonal antibodies (mAbs), engineered antibody formats are attracting increasing interest as they offer the possibility to make smaller molecules with a broader range of applications compared to standard immunoglobulins.^2–6

For instance, antibody fragments can be developed into multispecific formats that target different molecules or epitopes with the same molecule^7,8 and they can potentially lead to lower full manufacturing costs as they can be expressed in high yield in simple prokaryote or yeast systems, whereas canonical mAbs are expressed in mammalian cell lines.⁹ Furthermore, smaller antibody fragments have the potential for improved tissue penetration due to faster diffusion rates,¹⁰ for targeting solid tumors and for topological ocular indications.¹¹ In addition, they are more effective in crossing the blood–brain barrier, which is a major challenge in drug discovery against central nervous system (CNS) diseases such as Alzheimer’s or Parkinson’s disease.⁹ One downside of fragments compared to the conventional mAb format is the common poor in vivo half-life due to their small size, which leads to clearance in the kidney, and the lack of Fc domain protraction.¹² However, there are ongoing advances within the development of protraction mechanisms for improving the half-life of fragments based on albumin protraction that are currently being tested in clinic.¹³

Despite offering many benefits over canonical mAbs, fragments account for only a small percentage of the antibodies approved by the FDA, and relatively few of these are bispecific antibodies (bsAb).¹ A major challenge for the development of engineered constructs is their lower stability compared to full-length mAbs, perhaps a consequence of the fact that these molecules have not evolved naturally. In particular, scFv constructs have proven to be problematic.^14,15 Although the developability of full-length mAbs has been extensively explored,^{16 – 22} a systematic analysis of the developability potential of alternative engineered constructs is lacking.

Here, we compare the developability of 64 different antibody constructs against tumor necrosis factor (TNF) covering 8 format families. We measured 15 biochemical and biophysical properties associated with activity, manufacturing, short-term physical attributes and long-term stability, including an interfacial stress assay to assess surface-induced liabilities. Natural full-length antibodies outperformed the engineered variants in terms of stability, with modified formats disproportionately suffering from fragmentation and aggregation, both in bulk and at interfaces. Nevertheless, our study highlights certain formats that can be as stable as full-length IgGs and provides an initial standardized ranking of format developability.

Results

Features of the antibody library

We built our full antibody format library based on two commercial antibody formats, namely adalimumab (hIgG1) and certolizumab (hIgG1 format), which are both specific to tumor necrosis factor (TNF). The library contains 11 regular IgG antibodies, 5 scFv compounds, 11 tandem-scFvs, 8 diabody compounds, 6 scFv-Fc-scFv compounds,

4scFv-Fc compounds, 7 mAb-scFv compounds, 4 compounds where a scFv-Fc is combined with a half antibody designated scFv-Fc_HC_LC, and 8 bispecific antibody compounds HC1_LC1_HC2_LC2. The bispecific antibody compounds were evaluated by generating biparatopic molecules in which both arms bind to TNF. In addition, we included a control set of an IgG4 variant library based on an anti-hapten antibody that spans a broad range of solubility potential and has been previously characterized.^17,23 A schematic representation of the antibody library is shown in Figure 1. Within the set of 11 regular antibodies, differences were based on single mutations included in variable domains or within the Fc region.^24,25 For the scFv-containing molecules, scFv variants were generated with or without stabilizing mutations (Adalimumab VH_R16G_D30S_S49G_S78T_Y101S) and with or without insertion of a stabilizing disulfide bond (VH_44C-VL_100C) as described in Reiter et al.²⁶ For compounds where heterodimerization of the Fc region was needed (biscFv, Fab-scFv-Fc, and biAb), knob-in-hole mutations were introduced to stimulate correct compound assembly.²⁵ The compound designs are described in the Supplementary Dataset S1.

Figure 1.

Schematic representation of the 73 antibody constructs used in this study. Our antibody library is composed of full-IgGs, fragments, bispecifics, and bispecific-fragments. The HzATNP variants (Var1-Var9) have been previously described and they have been used as the control antibody set in this study. The subset of bispecifics (IgG/biAb) includes three IgGs (Var21, Var22 and ar23) which serve as controls and are technically not bispecific antibodies because they target a single epitope. Except for the HzATNPs which binds the ATNP, the rest of the variants target TNF-α.

Note: some images looked low resolution in the pdf file but were fine here on the on-line system. We can provide higher resolution images if required for the final pdf.

Biophysical properties assessed during initial stages of discovery

We first evaluated all 64 TNF binders in an in-house developed cell-reporter assay to confirm that potency was retained (Figure 2a). High potency was observed for all molecules, confirming the successful reformatting of all sets. We note an improvement in the potency of scFv and scFv-scFv compared to other formats.

Figure 2.

Biophysical properties of antibodies. The figure shows the critical properties assessed at the very early-stage of discovery. The lines represent the average values of the molecules within a set. The dashed line represents the average of all molecules. Green, yellow and red background indicate good, intermediate and bad behavior, respectively, based on the cutoff listed in Table 1. No major differences were observed in terms of activity. Bispecifics are characterized by low purity.

Next, we investigated the Fc–FcRn interaction, crucial for guaranteeing long in vivo half-life due to the ability of the Fc domain to be recycled when internalized by cells. For the bispecific variants, in particular, formats where scFvs are linked to the C-terminal of the heavy and light chain, it is important to confirm that the Fc domain is still able to interact with the FcRn receptor. Therefore, for all formats with an Fc domain, we analyzed the Fc–FcRn interaction using surface plasmon resonance (SPR) analysis, observing no significant differences upon reformatting (Figure 2b).

We also characterized properties related to large-scale manufacturing,²⁷ namely purity and the presence of soluble aggregates after purification, as measured by size exclusion HPLC (SE-HPLC). To compare the relative performance of the different molecules, we took as a reference the average values of our control set of HzATNP molecules and the full-length antibody IgGs (formatted from tandem scFvs). These cut-off values, as well as the thresholds of the properties discussed in the following paragraphs, are summarized in Table 1. In some cases, thresholds were taken from the literature, while for other properties were based on the average of all variants.

Table 1.

Thresholds of biophysical properties derived from analysis on samples corresponding to the 73 antibodies of the library. Red and yellow values are marked as a flag.

Property assessed	Green (low risk)	Yellow (medium risk)	Red (high risk)	Reference
Affinity to the human Fc receptor (hFcRn)	<10× difference in binding relative to the WT (Var9)	10–20×difference in binding relative to the WT (Var9)	>20× difference in binding relative to the WT (Var9) or non-binding (NB)	Mieczkowski et al., (28)
In vitro activity (IC50)	Below average of all variants	Close to average (less than + 20%)	Above average (more than + 20%)
SEC purity (Monomer %)	Above 95%	Between 90 and 95%	Below 90%	Bailly et al., (18)
Aggregates at t0	Below average of all variants	Close to average (less than + 5%)	Above average (more than + 5%)
Unfolding (Tm)	Above 60°C	Between 55 and 60°C	Below 55°C	Jain et al., (17) Jain et al., (16)
Aggregation (Tagg)	Above 60°C	Between 55 and 60°C	Below 55°C	Jain et al., (17) Jain et al., (16)
Self-association (kd)	Above −10 ml/g at high salt	Between −10 and −15 ml/g	Below −15 ml/g	Kingsbury et al., (22)
Agitation stress (Monomer loss %)	Below average values of all variants measured	Close to average (less than + 5%)	Above average (more than + 5%)
Interfacial stress (Monomer loss %)	Above average values of all variants measured	Close to average (less than + 5%)	Above average (more than + 5%)	Kopp et al., (32)
Long-term stability at 30°C/40°C: Aggregation and fragmentation	Below average of all variants measured	Close to average (less than + 5%)	Above average (more than + 5%)
Long-term stability at 30°C/40°C: Intact monomer	Above average of all variants measured	Close to average (max. −5%)	Below average (more than −5%)

As seen in Figure 2c, bispecifics are characterized by low purity. This could represent a challenge for the development of bispecifics, since the need of additional purification unit steps leads to increased costs and reduced final yields.

Biophysical stability properties

Next, we characterized several biophysical properties that are indicators of conformational and colloidal stability of the molecules, in bulk and at interfaces. These biophysical properties are overall indicators of the propensity of molecules to undergo self-association and aggregation, which are problematic physical instabilities associated with loss of monomer activity and increased risk of immunogenicity. All assays were conducted in a high-salt formulation (His 10 mM, pH 6, NaCl 140 mM), with the exception of the nanoparticle assay, which was performed in a no-salt formulation (His 10 mM, pH 6).

The thermal stability of the conventional and bispecific antibodies was characterized in terms of melting temperature (T_m) and the onset of aggregation temperature (T_agg). While fragment antibodies show a single transition, for full-length mAbs two transitions were observed in the Tm measurements (Dataset S2 Assay Results). For all antibody variants, we refer to the first melting temperature since undesirable product changes start occurring from the first thermal unfolding. Since most marketed antibodies have T_m higher than 60°C^16,17 and T_agg higher than 60°C,^16,17 we considered these values as thresholds to define developable molecules. Molecules below this threshold were marked as problematic, while molecules with Tm and T_agg between 55 and 60°C were considered of intermediate risk. As shown in Figure 3a,b, considering these cutoff ranges, there were no major differences among the antibody formats.

Figure 3.

Physical stability properties. Thermal stability expressed in terms of melting temperature (T_m) and temperature of the onset of aggregation (T_agg), and colloidal stability represented by the self-interaction parameter (k_d) for the different molecules. We refer to the first melting temperature of all antibody variants since undesirable product changes start occurring from the first thermal unfolding. Interfacial stability was measured as monomer loss during a classical agitation assay and in a nanoparticle assay that accounts for different types of surface chemical properties. The lines represent the average values of the molecules within a set. The dashed line represents the average of all molecules. Green, yellow and red background indicate good, intermediate and bad behavior, respectively, based on the cutoff listed in Table 1. Overall, fragments underperform in the interfacial stress assays.

As a measure of colloidal stability, we assessed the self-interaction parameter (kd) by dynamic light scattering (DLS) for all molecules except diAbs and scFvs. The diAb molecules were not characterized in all assays due to their very low yield. The kd is related to the second virial coefficient and has been in many cases a good predictor for antibody self-assembly at high concentrations, which may cause issues such as viscosity, opalescence,²² low solubility,²⁹ or liquid–liquid phase separation.³⁰ A negative kd value corresponds to net attractive interactions, while the positive value indicates net repulsive self-interactions. The threshold to define well-behaved antibodies in this case was selected as −15 ml/g, which was previously reported to identify molecules characterized by high viscosity in a high ionic strength formulation.²² Given this threshold, all tested variants are well behaved, although the full-length mAbs have the highest values of all sets (Figure 3c).

A classical agitation stress assay was performed to assess the synergistic effect of air/water interface and flow on antibody stability, which was evaluated by measuring the monomer loss (see Materials and Methods). As a threshold value, we selected the average monomer loss measured for all variants. In analogy with the previous measurements, molecules with monomer loss higher than 5% of the threshold were flagged as poorly behaved, while variants with intermediate values were considered of intermediate risk (Figure 3d).

As a further assessment of interface-stress, we applied a recently developed nanoparticle assay,^31–33 which probes the interactions of antibodies with different model surfaces with different charges and hydrophobicity, therefore promoting different electrostatic and hydrophobic interactions. The residual monomer loss was measured in the presence of three different types of nanoparticles, hydrophobic with negative charge, hydrophobic with positive charge, and high negative charge. A summary of the properties of the nanoparticles is shown in Table S1, and details of the methods are described in Materials and Methods. We defined an overall stability score based on the average monomer loss measured for all three nanoparticles, which averages the contribution of the different types of interactions (electrostatic and hydrophobic), and enables the ranking of the molecules (Figure 3e).

As shown in Figure 3, molecular format families have different values for the tested parameters (T_m, T_agg, and k_d, agitation, and nanoparticle assay). Moreover, large variability is observed between molecules within the same set. Overall, the set of canonical mAbs has more uniform values and on average performs better than the other formats. This variation in the spreading between conventional antibodies and bispecifics highlights a greater impact of mutations on alternative constructs in contrast to conventional antibodies.

The average values for T_m, T_agg, k_d, and nanoparticle assay are rather similar across different formats, although some liabilities are observed in the T_agg for the diAb and biscFv/biscFv-Fc formats, in the agitation assay for the biscFv/biscFv-Fc, diab and scFv formats, and in the nanoparticle assay for the biscFv/biscFv-Fc, diab, and scFv-scFv formats. The difference in the average k_d values is rather small ($\mathrm{\Delta }$k_d = 6 mL/g) compared to the range of possible values reported in the literature ($\mathrm{\Delta }$k_d = 20 mL/g).²² The spreading of the average values across different formats is higher for the agitation stress assay, with the IgG/biAb and Fab-scFv-Fc formats exhibiting lower monomer losses and scFv, diAb, and scFv-scFv formats showing higher liabilities.

Long-term stability

Finally, the long-term stability of the whole library was evaluated in terms of fragmentation and aggregation after one-year incubation at 4°C, 30°C, and 40°C. Fractions of residual monomer, fragments, soluble and insoluble aggregates were evaluated by SE-HPLC (for details, see Materials and Methods section). No changes were observed at 4°C for all molecules (data not shown). At 30°C the majority of the variants were stable in terms of aggregation, with the exception of the scFv-scFv fragments, while 5–8% of fragmentation was observed for several molecules (Figure 4a-c). Significant aggregation and fragmentation were observed under accelerated conditions at 40°C, with full-length antibodies behaving in general better than fragments and bispecifics (Figure 4d-f). Based on previous in-house observations, it is likely that the pronounced fragmentation of antibody fragments is due to the GS linkers, which can be optimized by engineering (e.g., by adjusting the linker’s length or introducing disulfide bonds) or using in-silico modeling to optimize their flexibility. Note that we have not evaluated the chemical stability of the complementarity-determining regions (CDRs), e.g., isomerization, deamidation, and oxidation, which we considered to be out of scope for this study. All biophysical properties are reported in the Supplementary Dataset S2.

Figure 4.

Long-term stability data. Percentage of aggregates (HMWP), fragments (LMWP) and intact monomer (Mon) measured after one year incubation at 30°C and 40°C. The lines represent the average values of the molecules within a set. The dashed line represents the average of all molecules. Green, yellow and red background indicate good, intermediate and bad behavior, respectively, based on the cutoff listed in Table 1. The diAb variants were not characterized for all assays due to their very low yield. Fragments are particularly problematic.

In several format classes, the highest monomer loss at 40°C is observed with the Humira scFv wild type, corresponding to Var36 for the scFv format, Var63 in the scFv-scFv format, and Var30 in the mAb-scFv biAb format.

Experimental and computational developability ranking with a flag risk approach

The results described in the previous paragraphs (summarized in Suppl. Figure S2) showed large variability in the behavior of the different formats and among different mutants within the same format. To perform a quantitative comparison, we assigned to each molecule a certain number of flags, defined as the number of biophysical properties that exceed the quantitative thresholds described in the previous paragraphs and summarized in Table 1. An overview of the distribution of flags in the different sets of molecules is shown in Figure 5. Due to low material availability, a few variants could not be analyzed with all assays. Therefore, the percentage of flagged properties is presented instead of the absolute number of flags. Each set of variants exhibits a distribution of flags, which is, however, very different for different formats. Full-length IgGs have a very limited number of flags. Bispecifics biAb and scFv have medium developability properties, while the mAb-scFvbiAb, bispecifics biscFv/biscFv-Fc, the fragments scFv-scFv and the bispecific fragments Fab-scFv-Fc and diAb formats appear to be more problematic. In Figure 6 we summarize the overall flag assigned per antibody set for each of the analyzed properties. This summary provides an orthogonal overview of the flagging shown in Figure 5.

Figure 5.

Distribution of flagged variants within the different antibody formats. Green, yellow and red frames indicate the overall developability risk for the respective variants set.

Figure 6.

Risk-assessment of the developability based on biophysical properties. The assays that take more leverage on the overall developability risk are the long-term stability studies as well as the colloidal and interfacial stability. Bispecific fragments and fragments show the least favorable developability profiles.

Moreover, we investigated possible correlations between the physical properties measured in short time (Figure 3) with the long-term stability of the molecules (Figure 4). As observed in other studies,^17,34,35 poor correlations were observed, and no individual property could predict the long-term stability of the variants (Suppl. Figure S3). Interestingly, when focusing the investigation of possible correlations in the interfacial assays, we observed a good correlation between the results of the assay with hydrophobic nanoparticles with negative charge and the standard agitation assay for all sets of molecules (Suppl. Figure S4), although this correlation has to be confirmed in the future with a broader set of molecules.

Finally, we attempted to rationalize these experimental findings with the Therapeutic Antibody Profiler (TAP) tool,^36,37 which evaluates developability potential by comparing calculated 3D biophysical properties of antibodies to distributions observed across clinical-stage therapeutic molecules (CSTs). The tool provides high (red), medium (amber) and low (green) risk flags for each property (see Methods for further details). The structure-based metrics are Patches of Positive Charge (PPC), Patches of Negative Charge (PNC), Patches of Surface Hydrophobicity (PSH), and the Structural Fv Charge Symmetry Parameter (SFvCSP). The performance of TAP in distinguishing the behavior of variants that are close in sequence space has not previously been tested, and the variants investigated in this study show a low sequence diversity relative to all CSTs. Despite these limitations, a level of agreement was observed between TAP and the biochemical and biophysical assays. Specifically, across all variants, TAP assigned a green flag for three of the four structure-based metrics (PPC, PNC, and PSH) (Suppl. Figure S5). However, for the SFvCSP, 78% of the variants were predicted as medium risk, 3% as high risk, and the remaining 19% as low risk. This indicates that the variants largely occupy a risky developability space, in agreement with the experimental assays, where all variants were highly flagged. Moreover, SFvCSP flagging agreed closest with the nanoparticle assays, in particular with the negative and positive nanoparticle assays (Suppl. Figure S6).

We also investigated whether TAP could predict the difference in developability between antibody formats that was observed experimentally. No clear trend could be seen. This is in accordance with TAP only evaluating properties in the Fv region, as well as the flagging thresholds for TAP having been determined based on all available post-Phase-I CSTs, which are dominated by whole mAb formats,³⁸ reducing sensitivity to format-specific patterns.

Discussion

In addition to new conventional antibodies brought into clinical testing, smaller antibody formats are becoming increasingly popular. Relevant examples include VHH’s and other small de novo affinity ligands,^39,40 and many other examples involving Fc domains.⁴¹ Our work shows that, on average, alternative constructs are characterized by higher instability than conventional antibodies, particularly in terms of aggregation and fragmentation. Moreover, the types of liabilities can largely differ among different format families. These insights highlight some important challenges that the current field of antibody developability assessment must tackle to enable future developability predictions of alternative scaffolds.

In contrast to conventional antibodies, new molecular formats inherently bring the challenge of lack of cohort experimental data. Our work represents an important step toward filling this gap, and promotes future systematic studies toward a quantitative comparison among emerging new formats. We also show that the developability potential is strongly format dependent and that a molecule that is unstable as a fragment (e.g., scFv Var36) can be optimized by mutations (e.g., scFv Var38) or developed in another format (e.g., adalimumab IgG Var10-Var20).

Moreover, it is unclear whether in silico developability tools developed for full-length antibodies will translate to other molecular formats. In recent years, significant advances within computational predictions of the developability of conventional antibody IgG formats have been made.⁴² This progress increases the potential for replacing (or at least reducing) resource-demanding experimental readouts of developability with less- demanding and accurate computational predictions. Furthermore, the increased focus on standardizing developability methods will hopefully lead to larger comparative studies with disclosed antibody sequences, which will ultimately support both development and evaluation of in silico tools.¹⁷ In our work, we have shown how the in silico predictor TAP, which has been well validated for full-length antibody format, has limitations in predicting the developability issues of the antibody-derived formats tested here, although a level of agreement was observed between TAP and the biochemical and biophysical assays. Moreover, alternative formats are often fused with other protein domains to enable, for instance, multi-specific functionality and in vivo half-life extension, which further complicate model building.

In conclusion, we characterized 73 mAbs expressed in 9 distinct formats by measuring 15 different biophysical properties that describe activity, production, conformational and colloidal stability, as well as long-term stability. Relative comparisons were quantitatively performed by defining threshold values for each biophysical property. Different formats were characterized by different distributions of flags and developability profiles (Figure 5). The overall developability potential was higher for traditional full-length antibodies and decreased for bispecific and tandem fragments. Specifically, bispecifics biAb and mAb-scFvbiAb as well as fragments scFv exhibited medium developability properties, while the bispecifics biscFv/biscFv-Fc, the fragments scFv-scFv and the bispecific fragments Fab-scFv-Fc and diAb formats revealed a large number of flags. Poor correlations were observed among different properties, and no individual property could predict the long-term stability of the variants. Although full-length antibodies outperformed the engineered variants, this work highlights certain formats that can be almost as stable as full-length antibodies and provides an initial standardized ranking of format developability.

Materials and methods

Antibody variants expression and purification

Expression plasmids for production of all antibody formats were purchased from Geneart/Life technologies. The scFv variants, tandem scFv, and diabody formats were designed with a C-terminal HPC4 tag for purification purposes. scFvs constructs were designed with 21 residues long glycine-serine linker (SGGGGSGGGGSGGGGSGGGGS) separating the VH and VL domains. For tandem scFv constructs, the two scFv moieties were connected by 25 glycine-serine linker (GGGGSGGGGSGGGGSGGGGSGGGGS) and for diabody constructs the compounds were designed using the following setup: VH1-GGGGS-VL2-GGGGSGGGGSGGGGS- VH2-5GS-VL1. All proteins were expressed by transient transfection of Expi293 cells following the instructions of the manufacturer (Expi system, Life Technologies). Proteins were purified from cell culture supernatants using first an affinity chromatography step using either anti-HPC4 affinity resins (for scFv, tandem-scFv or diabodies) or protein-A-based resins for Fc-containing constructs. Subsequently, the molecules were purified with a Superdex200 size-exclusion purification step operating HiTrap MabSelectSure and HiLoad 26/600 Superdex 200 columns according to manufacturer’ description (Cytiva, cat. no. 11003495 and 28,989,336). HPC4 column and purification protocols were established internally at Novo Nordisk based on anti-HPC4 antibody produced recombinantly using Expi203, as described above, and the purified antibody coupled to CNBr-Activated Sepharose 4 Fast Flow resin according to the manufacturer's description (Cytiva, cat. no. 17098101). Proteins were prepared in the size exclusion running buffer 20 mM Hepes, 150 mM NaCl, pH 7.4. All molecules were subjected to quality control analysis, including SDS-PAGE, SE-HPLC, and intact mass LC-MS. All molecules were then buffer exchanged using centrifugal filters to 10 mM His, pH 6, 140 mM NaCl, and concentrations were determined based on UV280 measurements.

Activity in vitro (cell reporter assay)

A HeLa cell-based reporter assay (Novo Nordisk, in-house) was used to determine the antibody functionality by measuring TNF inhibition. NF-κB signaling is activated when TNF is bound to the TNF receptor, which in this cell line induces the transcription and intercellular production of firefly luciferase. Luciferase can be used to quantify the amount of TNF inhibition by adding its substrate (beetle luciferin) and monitoring the resulting light emission.

On day 1, the HeLa cells were washed in 50 ml phosphate-buffered saline (PBS; Gibco 14,190–094), spun at 1200 rpm for 5 min and resuspended in 1 ml assay medium containing DMEM+GlutaMAX (Gibco 10,569–010), 10% FBS (Gibco 01,190,009 M), 10 mM Hepes (Gibco 15,630–056) and 5 ml Glutamax (Gibco 35,050,061). The cells were counted on a ViaCell, diluted in an assay medium and seeded in 96-well plates (Perkin Elmer 6,005,680). 100 μl medium containing 2 ∙ 10⁴ cells was seeded in each well. The plates were incubated at 37°C with 5% CO2 overnight.

On day 2, the medium was removed from the plates and 50 μl of the antibody samples were added to the cells. Beforehand fourfold dilutions of each antibody were made in technical triplicates starting from 100 nM using the assay medium. Cells and antibodies were incubated for 30 min at 37°C before 50 μl 0.5 ng/ml TNF (R&D Systems, 210-TA) was added to each well. 1–2 columns on each plate were reserved for controls where only TNF was added to threefold dilutions. The plates were incubated for 4 hours at 37°C, to provide the cells enough time for luciferase expression. 100 μl Steady-Glo (Promega, E254B) was added to each well, and the plates were incubated in the dark for 15 min. During this time, the cells were lysed and the substrate beetle luciferin was catalyzed by the reporter enzyme firefly luciferase. The resulting luminescence was measured on an EnVision plate reader.

The luminescence signal from the luciferase cell assay was analyzed using GraphPad Prism. The intensity of the signal was plotted against the antibody concentration and visualized using a logarithmic x-axis. To estimate the concentration of half of maximum inhibition, the concentrations were transformed to log(concentration) and the reads were fitted to a dose–response curve with four parameters and a variable slope. In the case of not all variants reaching the bottom plateau, fitting was done under the restriction of a shared top and bottom. Prism automatically calculates the IC50, as the concentration half-way between the top and bottom plateau using the equation:

$$\mathit{Y}=\mathit{Bottom}+\frac{\mathit{Top}-\mathit{Bottom}}{1+{10}^{\left(\mathit{LogIC}50-\mathit{X}\right)\cdot H\mathrm{illslope}}}$$

where Y is the luminescence signal intensity, X is the transformed log(concentration), HillSlope the steepness of the slope, Top and Bottom the top and bottom plateau, respectively.

Affinity to the human Fc receptor

Since the biscFv/biscFv-Fc (Var45 – Var54) proteins lack light chains that can be used as the affinity tag in capture setups, we immobilized them and the anti- trinitrophenol (ATNP) hIgG4(S214P) on the sensor chip via amine coupling. To do this, these antibodies were first diluted at 3–8 ug/ml in 10 mM sodium acetate (pH 4.5) (Cytiva Life Sciences, catalog # BR100350) and then were immobilized on a CMD50L sensor chip (XanTec Bioanalytics GmbH) in flow cells 2, 3 or 4 at ~100 RU with 1× HBS-EP⁺ (pH 7.4) (Cytiva Life Sciences, catalog # BR-1006-69) used as the running buffer. Although the ATNP hIgG4(S214P) is in the regular hIgG4 format, it was immobilized as the control sample on the antibody side. Flow cell 1 was used as the reference flow cell and thus no antibody was immobilized in this flow cell. Next, the hFcRn ectodomain diluted at 6400 nM with twofold serial dilutions was injected through each flow cell for 90 sec to allow for binding of the hFcRn ectodomain to these immobilized antibodies, followed by dissociation of the hFcRn ectodomain for 90 sec. 1× CBS-EP⁺-CMD (pH 5.6) and 1× PBS-EP⁺-CMD (pH 7.4) were used as the dilution buffers for the hFcRn ectodomain and the running buffers in each binding cycle at the flow rate of 20 ul/min. After each binding cycle, the sensor chip was regenerated with two successive 60 sec injections of 1× PBS-EP⁺ (pH 7.4) at the flow rate of 30 ul/min. This regeneration procedure left the flow cell surfaces with the antibodies intact, while the bound hFcRn ectodomain were removed, allowing for the binding of new hFcRn ectodomain samples in the next cycle.

The other antibodies have light chains that can be used as the affinity tag in capture setups. As the capturing reagent, a F(ab’)2 anti-human IgG (Fab specific) (Sigma-Aldrich, catalog # I3266-1 ml) was immobilized on the sensor chip via amine coupling. To do this, this F(ab’)2 was first diluted by 100-fold in 10 mM sodium acetate (pH 4.5) (Cytiva Life Sciences, catalog # BR100350) and then were immobilized on a CMD50L sensor chip in flow cells 1, 2, 3, and 4 at ~1500 RU with 1× HBS-EP⁺ (pH 7.4) used as the running buffer. Next, these antibodies were diluted at 10–25 µg/ml and then were injected for 30 sec at a flow rate of 10 µl/min in flow cell 2, 3 or 4 to allow for capture of these antibodies at 70–400 RU. The ATNP hIgG4(S214P) was captured as a control sample on the antibody side. Flow cell 1 was used as the reference flow cell and thus no antibody was captured in this flow cell. Following the capture steps, the hFcRn ectodomain diluted at 6400 nM with twofold serial dilutions was injected through each flow cells for 90 sec to allow for binding of the hFcRn ectodomain to these captured antibodies, followed by dissociation of the hFcRn ectodomain for 90 sec. 1× CBS-EP⁺-CMD (pH 5.6) and 1× PBS-EP⁺-CMD (pH 7.4) were used as the dilution buffers for the antibodies and the hFcRn ectodomain and the running buffers in each binding cycle at a flow rate of 20 ul/min. After each binding cycle, the sensor chip was regenerated with two successive 20 sec injections of 10 mM glycine-HCl (pH 1.7) (regeneration solution in the kit, Cytiva Life Sciences, Catalog # BR100838) at the flow rate of 30 µl/min. This regeneration procedure left the flow cell surfaces with the F(ab’)2 anti-human IgG (Fab specific) intact, while the captured antibodies were removed, allowing for the capture of new antibody samples and binding of new hFcRn ectodomain samples in the next cycle.

In the immobilization steps, the surface of the sensor chip in flow cells 1, 2, 3, and 4 were activated by 100 mM N-hydroxysuccinimide (NHS) and 400 mM N-ethyl-N’- dimethylaminopropyl carbodiimide hydrochloride (EDC) mixed at equal volume before the immobilization of any protein and blocked by 1 M ethanolamine hydrochloride-NaOH (pH 8.5) after the immobilization of the proteins. NHS, EDC, and ethanolamine were obtained from the Amine Coupling Kit, type 2 (Cytiva Life Sciences, catalog # BR100633).

The 1× PBS-EP⁺-CMD (pH 7.4) buffer was prepared by diluting 0.5 M EDTA (pH 8.0) (Invitrogen, catalog # 15575038) at 3 mM and dissolving carboxymethyldextran (CMD) (Sigma-Aldrich, catalog # 86524-100 G-F) at 1 mg/ml in 1× PBS-P⁺ (Cytiva Life Sciences, catalog # 28995084). The 1× CBS-EP⁺-CMD (pH 5.6) buffer was prepared by diluting 0.5 M EDTA (pH 8.0) at 3 mM, diluting 10% (v/v) surfactant P20 (Cytiva Life Sciences, catalog # BR100054) at 0.05% (v/v), dissolving CMD at 1 mg/ml, dissolving sodium citrate at 10 mM and dissolving sodium chloride at 150 mM, followed by pH adjustment to 5.6 using ~10 M HCl.

The affinities (equilibrium dissociation constants, KD) were determined via steady- state fitting of the binding curves using the 1:1 binding model in Biacore T200 Evaluation Software 2.0 (Cytiva Life Sciences). Each interaction was tested in two independent replicates on two different sensor chips. The avarage KD values of each interaction were calculated. The error bars for the average KD values were calculated as the half of the difference between the KD values (diff/2) of the two replicates if diff/2 is larger than both the standard errors (SEs) for the fitted KD values of the two replicates or the larger SE for the fitted KD value in one replicate if diff/2 is smaller than it.

Tm and Tagg measurements

The melting temperature (T_m) and the temperature of the onset of aggregation (T_agg) for each variant were measured via nano differential scanning fluorimetry (nanoDSF) on a Prometheus NT.48 system (NanoTemper, PR001). Samples were prepared by filling standard capillaries (NanoTemper, PR-C006) with 10 µL of 1 mg/mL protein solution. The intrinsic tryptophan fluorescence was measured at 330 nm and 350 nm, while heating the sample from 20°C to 90°C using a 1°C/min temperature ramp. The data were analyzed using NanoTemper PR Control software. Tm was derived from the maximum of the first derivative of the fluorescence ratio at 350 nm and 330 nm.⁴³

T_agg was determined in parallel to T_m on the same platform by measuring light scattering at different temperatures. Data analysis for determining the Tagg was performed using a python script. Data were fitted by a sigmoidal curve and the onset of aggregation was defined as the temperature at which the scattering signal relative to the pre-transition baseline has reached 10% of the amplitude of the curve.⁴⁴

k_d measurements

The self-interaction parameter (k_d) was measured using DLS as described in Dingfelder et al.⁴⁵ Briefly, the diffusion coefficient was measured at different protein concentrations from 1 to 20 mg/ml in high-salt buffer (10 mM His, pH 6, 140 mM NaCl). 20 µl of each protein dilution was pipetted in triplicates into a 384-well plate (Corning 3540) and sealed with transparent foil (ClearSeal). The plate was centrifuged (1500 rpm, 5 min) before measuring in a Plate Reader (Wyatt DynaPro) at 25°C. k_d was calculated from the linear relationship of the diffusion coefficient D versus protein concentration c:

$$\mathit{D}=D_0\cdot \left(1+k_d\cdot \mathit{c}\right)$$

where D0 is the diffusion coefficient at infinite dilution.

Agitation stress assay

Stock solutions of the variants (0.5 mg/ml) were prepared in high-salt buffer (10 mM His, 140 mM NaCl, pH 6). 100 µl of each sample were transferred in quadruplicates into 1.5 ml screw neck vials (clear) (BGB, 32 × 11.6 mm) and sealed with parafilm. Samples were shaken in duplicates at 1400 rpm and 25°C for 4 h and transferred into 1.5 ml Eppendorf tubes immediately after shaking. As a control, two replicates of samples incubated under stagnant conditions were also analyzed. The four Eppendorf tubes were centrifuged for 1.5 h at 17,200 rpm. The absorbance at 280 nm of the supernatant was then measured twice for all samples (Nanodrop). The monomer loss due to the agitation stress was calculated using the following expression:

$$\mathit{Monomer}\mathit{loss}\left(\mathrm{\%}\right)=\frac{\mathit{Ab}{\mathit{s}}_{\mathit{ref}}-\mathit{Ab}{s}_{\mathit{sample}}}{\mathit{Ab}{s}_{\mathit{ref}}}\ast 100\mathrm{\%}$$

where Abs_ref and Abs_sample are the absorbance values measured in the absence and presence of agitation, respectively.

Nanoparticle synthesis

The nanoparticles were synthesized by emulsion polymerization as described previously.^32,33 In brief, positively and negatively charged nanoparticles were synthesized from methyl methacrylate (MMA, Sigma Aldrich, 99%) and [2-(methacryloyloxy)ethyl]-trimethylammonium chloride (MA-Ch⁺, Sigma-Aldrich, 80% solution in water) or 3-sulfopropyl methacrylate potassium salt (MA-SO3⁻, Sigma-Aldrich, 98%) in distilled water. 2,2′-azobis(2-ethylpropionamidine) dihydrochloride (V50, Sigma-Aldrich, 97%) or potassium persulfate (KPS, Merck) were used as radical initiators for positive and negative nanoparticles, respectively.

Hydrophobic nanoparticles with a residual negative charge (Hb⁻) were synthesized from butyl acrylate (BA, Sigma-Aldrich, 99%) as a hydrophobic monomer in the presence of sodium dodecylsulfate (SDS, Sigma-Aldrich, 99,85%) and water. Hydrophobic nanoparticles with a residual positive charge (Hb⁺) were also prepared using BA as well as MA-Ch⁺ as a stabilizer in combination with tert-Butyl hydroperoxide (tBuOOH, ABCR, 70%) and ascorbic acid (AsAc, Fisher Chemicals) as radical initiators.

Nanoparticle assay

The nanoparticle assay was performed as previously described.³² Briefly, the antibody variant was mixed in 1:1 ratio with the nanoparticles in 1.5 ml reaction tubes (Eppendorf), to reach a final antibody concentration of 0.5 mg/ml. Nanoparticles were introduced at a surface ratio of 25:1 protein to nanoparticle. All nanoparticle assays were performed in a no-salt buffer (10 mM His pH 6) due to the colloidal instability of the nanoparticles at higher salt concentrations.

After incubating the solution for 30 min at room temperature, 50 mM MgCl2 was introduced to destabilize the suspensions of proteins and nanoparticles. The mixture was transferred to filter plates (Corning 96-well filter plates) and centrifuged for 30 min at 2500 rpm at 20°C. The absorbance at 280 nm of the filtrates was then measured in duplicates on a Nanodrop Spectrophotomer (3917 Lite, ThermoScientific). The experiment was conducted in duplicates per each formulation, and in addition, two reference samples were measured in the absence of nanoparticles. Control samples of nanoparticles without protein were also measured in duplicates, to account for background. The % monomer loss was calculated using the following equation:

$$\mathit{Monomer}\mathit{loss}\left(\mathrm{\%}\right)=\frac{\mathit{Ab}{\mathit{s}}_{\mathit{ref}}-\mathit{Ab}{s}_{\mathit{sample}}}{\mathit{Ab}{s}_{\mathit{ref}}}\ast 100\mathrm{\%}$$

where Abs_refand Abs_sample are the absorbance values measured in the absence and presence of nanoparticle, respectively.

While the nanoparticle assay was performed in a no-salt formulation (10 mM His, pH 6), the other biophysical parameters were determined in a standard high ionic strength formulation (10 mM His, 140 mM NaCl, pH 6). This difference could affect the correlations between the different assays.

Long-term stability studies

Six replicates of 100 µl of each variant at 1 mg/ml were incubated into HPLC glass vials (BGB, 32 × 11.6 mm) at 4°C, 30°C, and 40°C. Three reference samples were prepared, two of them were measured immediately with SE-HPLC and the third one was frozen for storage at −80°C.

After one year of incubation, the samples were analyzed by injecting into a size-exclusion column (TSK G3000 SWXL, Tosoh Bioscience) assembled on an Agilent HPLC system. Samples were analyzed at a column temperature of 20°C and a flow rate of 0.8 ml/min for 25 min, using as running buffer a solution of 122 mM Na2HPO4, 78 mM NaH2PO4, 300 mM NaCl, 4% 2-propanol at pH 6.8.

Monomers, aggregates, and fragments peak areas were determined from the SEC chromatograms measured before and after incubation. Due to solvent evaporation, the total area of the SEC chromatograms of some of the samples increased after 1 year of incubation. To take this into account, the individual areas of monomers, fragments, and soluble aggregates were normalized to the total area after 1-year incubation. This approach neglects the formation of insoluble aggregates, which are unable to enter the column, and the estimated monomer conversions are therefore lower boundaries. The following equations were used to calculate the percentage of aggregates, fragments, and intact monomer after 1 year of incubation at 30°C and 40°C:

$$\mathit{Aggregates}\left(\mathrm{\%}\right)=\frac{{\mathit{A}}_{\mathit{HMWP},\mathit{t}}}{A_{\mathit{total},\mathit{t}}}\ast 100\mathrm{\%}$$

$$\mathit{Fragments}\left(\mathrm{\%}\right)=\frac{{\mathit{A}}_{\mathit{LMWP},\mathit{t}}}{A_{\mathit{total},\mathit{t}}}\ast 100\mathrm{\%}$$

$$\mathit{Monomer}\left(\mathrm{\%}\right)=\frac{{\mathit{A}}_{\mathit{Monomer},\mathit{t}}}{A_{\mathit{total},\mathit{t}}}\ast 100\mathrm{\%}$$

Computational developability profiling (TAP)

The Fv region of all antibody variants was structurally modeled using ABodyBuilder2⁴⁶ and all hydrogen atoms were removed. The modeled structures were then run through TAP³⁶ to calculate the five developability metrics – Total IMGT CDR Length, PSH with the Kyte and Doolittle scale, PPC, PNC, and the SFvCSP. The PSH, PPC, and PNC metrics are based on residues in the CDR vicinity which is defined as any residue in IMGT-defined CDR positions $\pm$ 2 residues on either side, as well as any other surface-exposed residues within 4.5 Å of these residues. We used the four structure-based metrics from TAP (PSH, PPC, PNC, and SFvCSP) to assign high (red), medium (amber), or low (green) risk flags to each antibody. The amber and red flagging thresholds are defined according to the distribution observed in clinical-stage therapeutics (CSTs) as reported in the latest TAP update paper.³⁷ For bispecific formats, each unique Fv region was structurally modeled and assigned developability flags separately, and we assigned the highest risk flag for each metric to the variant.

Funding Statement

The work was supported by the Novo Nordisk STAR Fellowship.

Disclosure statement

N.L., T.E., J.R.B., S.G., Z.C., A.B., are present employees of Novo Nordisk A/S and are shareholders of Novo Nordisk A/S.

Author contributions

P.A., and N.L.Z designed research. I.C.M., F.D., and B.P. performed measurements and analyzed data. I.W. analyzed data. J.R.B. and T.E. engineered, produced, and purified antibody library, Z.C. measured the binding affinities, S.N.J. and A.B. performed in vitro activity measurements. O.M.T., A.M.H., M.I.J.R., and C.M.D. performed in silico analysis. I.C.M., P.A., and N.L.Z. wrote the initial version of the manuscript. All authors provided inputs and reviewed the manuscript.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2024.2403156