Engineering hydrophobicity and manufacturability for optimized biparatopic antibody–drug conjugates targeting c-MET

ABSTRACT

Optimal combinations of paratopes assembled into a biparatopic antibody have the capacity to mediate high-grade target cross-linking on cell membranes, leading to degradation of the target, as well as antibody and payload delivery in the case of an antibody-drug conjugate (ADC). In the work presented here, molecular docking suggested a suitable paratope combination targeting c-MET, but hydrophobic patches in essential binding regions of one moiety necessitated engineering. In addition to rational design of HCDR2 and HCDR3 mutations, site-specific spiking libraries were generated and screened in yeast and mammalian surface display approaches. Comparative analyses revealed similar positions amendable for hydrophobicity reduction, with a broad combinatorial diversity obtained from library outputs. Optimized variants showed high stability, strongly reduced hydrophobicity, retained affinities supporting the desired functionality and enhanced producibility. The resulting biparatopic anti-c-MET ADCs were comparably active on c-MET expressing tumor cell lines as REGN5093 exatecan DAR6 ADC. Structural molecular modeling of paratope combinations for preferential inter-target binding combined with protein engineering for manufacturability yielded deep insights into the capabilities of rational and library approaches. The methodologies of in silico hydrophobicity identification and sequence optimization could serve as a blueprint for rapid development of optimal biparatopic ADCs targeting further tumor-associated antigens in the future.

KEYWORDS:

• Antibody engineering
• Biparatopic ADC
• c-MET
• hydrophobicity
• manufacturability

Introduction

c-mesenchymal-epithelial transition (c-MET) also known as hepatocyte growth factor receptor (HGFR) plays a key role as tumor driver in tumor indications of high medical need such as non-small cell lung cancer (NSCLC) and gastrointestinal cancers¹ and as a mechanism of resistance to targeted therapies in NSCLC.^2–5

c-MET is a tyrosine kinase receptor activated following binding of its ligand, HGF.⁶ This interaction leads to activation of several oncogenic signaling pathways.⁷ c-MET signaling is deregulated in diverse tumor types, including lung cancer, via c-MET overexpression, genomic amplification, autocrine/paracrine ligand stimulation, translocations, point mutations, and alternative splicing.⁸ In addition to small molecule inhibitors such as tepotinib,¹ several monoclonal antibodies (mAbs) or more complex biotherapeutics targeting c-MET have been developed.⁹ Amongst those, early monovalent and bivalent mAb approaches relied on antagonism and antibody-dependent cellular cytotoxicity (ADCC) as the main modes of action and failed in clinical development.^10–12 Next-generation biologics were designed as bispecific antibodies, such as amivantamab¹³ or MM-131,¹⁴ or focused on target degradation.^15–17 One molecular mechanism for effective target degradation is crosslinking of more than two target molecules via dual binding of suitable combinations of paratopes and their optimal orientation in a biparatopic antibody to favor inter- over intra-molecular binding (Figure 1a).^18,19 The underlying modes of action are enhanced apparent affinity for cellular binding, increased internalization by cross-linking, degradation²⁰ and with it signal transduction inhibition, as well as higher potency cytotoxicity. Such enhanced target removal also provokes degradation of the therapeutic mAb or a corresponding antibody–drug conjugate (ADC), leading to enhanced intracellular payload delivery.^21–23

Figure 1. Biparatopic anti-c-MET antibodies favoring inter- over intra-target binding. (a) Scheme of inter- versus intra-molecular binding of a biparatopic antibody. Simultaneously addressing two epitopes on the same c-MET monomer yields high apparent affinity and optimally doubled saturation binding. Combinations of paratopes mediating binding to different c-MET monomers ideally results in enhanced cross-linking, internalization, target degradation and, if ADCs are applied, enhanced payload release. (b) In silico docking against c-MET (PDB code 2UZY) suggests differential fab arm positions for selected reference antibodies and CS06 and B10v5. Onartuzumab epitope derived from available X-ray structure with PDB code 4K3J. (c) Fab arms 1 and 2 of biparatopic reference antibody REGN5093 show orientations supporting inter-molecular c-MET engagement. Visualization obtained via docking with restraints from experimental data taken from DaSilva et al. clin cancer res 2020;26:1408–19. (d) Proposed inter-c-MET cross-linking property of a B10v5 × CS06 biparatopic antibody.

Diverse antibody binding sites to c-MET and biparatopic combinations favoring c-MET cross-linking.

In 2016, we reported identification of human antibodies to distinct epitopes on SEMA and IPT1 domains of c-MET extracellular domain (ECD), termed B10v5 and CS06, respectively.²⁴ Comparative docking studies were guided by experimental epitope binning data²⁴ and suggested paratope orientation likely supporting inter- rather than intra-molecular c-MET ECD binding. Initial biparatopic designs yielded high apparent affinities and strong internalization capacities indicating applicability for c-MET degradation or biparatopic ADC approaches. Both sequences, B10v5 and CS06, were further sequence-optimized, pursuing framework-based germline humanization and removal of chemically labile residues. While a sequence-optimized variant of B10v5 revealed favorable binding and hydrophobicity properties in hydrophobic interaction chromatography (HIC), CS06 revealed a significantly delayed retention time in HIC that was attributed to structure-based predicted hydrophobic patches in CS06 HCDR2 + 3. In the subsequent workflow, this property impeded conjugation of cytotoxic payloads during ADC generation. Due to their structural complexity, ADCs can generally be challenging to manufacture and prone to aggregation, representing an increased risk for immunogenic reactions and rapid clearance rates, particularly if hydrophobic linkers and payloads are involved.^25–27 Therefore, it is generally important to engineer antibodies for ADC generation toward lower hydrophobicity and to carefully evaluate the physical stability properties.^28–30

Historical knowledge on CS06 sequence analogues obtained from affinity maturation efforts and molecular modeling including in silico developability assessment led to three comparative approaches for further sequence and property optimization. The first was rational design in two iterative cycles for hydrophobicity reduction: single mutations of predicted solvent-exposed hydrophobic residues in HCDR2 or HCDR3 against polar amino acids, followed by combinations of the best mutations within both HCDR2 + 3 after experimental profiling for c-MET binding and HIC. The second used a rationally designed combinatorial library exchanging certain percentages of the same predicted hydrophobic key residues in HCDR2 and HCDR3 against any polar amino acid and yeast surface display (YSD) sorting for display and retained target-specific binding versus the third approach, presenting the same library via mammalian display on the surface of Chinese hamster ovary (CHO) cells, sorted for high display and retained target binding. Comparative analyses revealed concordant essential versus variable positions in all three panels and a higher amino acid variability from library approaches reflecting the significantly higher combinatorial space screened. While the variants obtained after two rational design cycles already yielded distinctly reduced hydrophobicity, the library approaches enabled optimization in a broader combinatorial screening space that resulted in more diversified amino acid exchanges and ultimately further enhanced properties. Clones derived from both display approaches showed enhanced thermal stability and significantly lowered hydrophobicity. While YSD and Sanger sequencing yielded several variants with retained binding properties, CHO display and next-generation sequencing (NGS) evaluation resulted in lower hit rates for retained cellular binding.

Resulting optimized variants CS06 VH6.18 and VH6.21 showed strongly reduced hydrophobicity while maintaining pico-molar affinities and elevated cell binding properties as biparatopic antibodies, appropriate in silico developability profiles and, most relevant, improved capability for payload conjugation. In vitro potency of resulting biparatopic ADCs on several c-MET-expressing tumor cell lines was in the sub-nanomolar to one-digit nanomolar range.

This study represents, to our knowledge, the first report on rational combinatorial paratope choice for enhanced biparatopic antibodies and the first comparison of rational versus yeast or mammalian display for antibody engineering. The path to highly potent biparatopic ADCs could serve as a blueprint for similar efforts to identify effective next-generation therapeutics with innovative and differentiated modes of action.

Results

Docking for educated choice of optimal paratope combination

We previously reported the identification of human antibodies with pico-molar affinities to distinct epitopes on SEMA and IPT1 domains of c-MET ECD, B10v5 and CS06.²⁴ Comparative docking studies of the B10v5 and CS06 variable domains against c-MET ECD (PDB code 2UZY) suggested paratope orientations likely supporting inter- rather than intra-molecular c-MET ECD binding (Figure 1b). Originally identified by single-chain variable fragment (scFv) phage display, both binding moieties could be adapted to both antigen-binding fragment (Fab) and scFv format²⁴ to build Fab x scFv heterodimeric Fc biparatopic antibodies. Initial biparatopic strand-exchanged engineered domain (SEED)³¹ and DuoBody³² designs yielded high apparent affinities (Table 1) and strong internalization capacities indicating applicability Table 1 for c-MET degradation or biparatopic ADC approaches.

Table 1. Overview on key properties of final biparatopic lead candidates. Constructs marked bold were selected for subsequent antibody drug conjugation. HIC retention time reference for clinical stage antibody was cetuximab with 5.8 min. n.D., not determined.

Name	Short name	Final yield (mg/L)	Purity after ProteinA (%)	Final purity (%)	Tm1 (°C)	Tagg onset (°C)	HIC ret. time (min)	AC-SINS mean score	PSR-BLI mean response (nm)	KD human c-MET ECD His (nM)	kon (1/Ms)	kdis (1/s)	EC50 EBC-1 (nM)
one-armed CS06_VH1.0_Fab		180	92.8	100	65	61.3	7.9	4,334	n.d.	0.34	6.81E + 05	2.29E–04	0.40
one-armed CS06_VH6.21_Fab		392	94.5	100	68.5	67.9	5.1	0.315	n.d.	125.7	5.85E + 05	7.35E–02	31.97
one-armed CS06_VH6.18_Fab		315	94	100	67	66.6	4.8	−1,217	n.d.	216.0	5.94E + 05	1.28E–01	46.55
one-armed CS06_VH1.0_scFv		143	69.1	94.2	49.8	47.5	8.2	0.641	n.d.	0.25	6.64E + 05	1.66E–04	0.36
one-armed CS06_VH6.21_scFv		274	77.3	100	57.2	55.2	5.2	−1,201	n.d.	95.1	3.50E + 03	3.32E–04	low
one-armed CS06_VH6.18_scFv		229	77.3	87.1	55	53.9	5.0	−1,573	n.d.	262.6	1.05E + 06	2.76E–01	low
one-armed B10v5 Fab		n.d.	n.d.	n.d.	n.d.	n.d.	n.d.	n.d.	n.d.	0.62	2.99E + 05	1.84E–04	1.60
B10v5 scFv x CS06 VH1.0_Fab	1.0_Fab	28	78.1	94.5	60.1	55.8	8.2	25,018	0.205	<0.001	1.27E + 06	<1.0E–07	0.33
B10v5 scFv x CS06 VH6.21_Fab	6.21_Fab	95	70.1	98.7	59.4	57.6	5.7	9,973	0.081	<0.001	7.10E + 05	<1.0E–07	0.40
B10v5 scFv x CS06 VH6.18_Fab	6.18_Fab	50	54	96.9	58.6	56.3	5.4	1,632	0.024	<0.001	7.54E + 05	<1.0E–07	0.21
B10v5 Fab x CS06 VH1.0_scFv		153	88.9	93.9	51.2	48.5	8.2	26,425	n.d.	<0.001	6.67E + 05	<1.0E–07	0.56
B10v5 Fab x CS06 VH6.21_scFv	6.21_scFv	189	87.1	100	57.6	54.4	5.4	14,290	0.034	<0.001	5.26E + 05	<1.0E–07	0.55
B10v5 Fab x CS06 VH6.18_scFv		245	83.2	93.3	55.8	53.9	5.2	17,369	n.d.	<0.001	5.12E + 05	<1.0E–07	0.65
Trastuzumab		n.d.	n.d.	n.d.	n.d.	n.d.	n.d.	0.154	−0.040	n.d.	n.d.	n.d.	n.d.
Briakinumab		n.d.	n.d.	n.d.	n.d.	n.d.	n.d.	26,465	0.135	n.d.	n.d.	n.d.	n.d.

Biparatopic antibody generation and optimization

Sequence optimization toward framework-based mapping to the closest human germline was successfully achieved for B10v5 in one design cycle. For CS06, adaption to the closest human germline sequence could be accomplished too, but the parental as well as the designed variants showed high HIC retention times (Figure 2d) that were subsequently attributed to pronounced solvent-exposed hydrophobic patches in HCDR2 and HCDR3 (Figure 2a) based on a structural model. In addition, these properties resulted in poor to no conjugatability (Figure 4). Sequence optimized variants are herein referred to as B10v5 and CS06, and CS06 VH variants discussed below were optimized for suitable hydrophobicity profiles. Previous knowledge obtained from affinity maturation²⁴ and structural analysis of CS06 antibody models guided the path to three comparative approaches to reduce hydrophobicity while retaining sufficient affinity (Figure 2): 1) iterative rational design of CS06 HCDR2 and HCDR3 sequence variants; 2) a library approach varying hydrophobic complementarity-determining region (CDR) residues screened by YSD; and 3) sorting the same library for display and target binding via mammalian display.

Figure 2. Rational optimized CS06 sequence and library design. (a) Predicted aggregation hot spots in CS06 fv and respective amino acids in positions marked red. (b) Positional variability in CS06 HCDR3 as observed during affinity maturation (sellmann 2016), indicating essential versus alterable amino acids. (c) Semi-rational library design varying hydrophobic residues defined from information shown in (a) and (b). (d) Sequential rational sequence engineering toward lower hydrophobicity. Single HCDR2 or HCDR3 point mutations with retained affinity but lowered hydrophobicity in 3.X series were combined into 2- to 4-point mutant sequences in 5.X variants. (e) In comparison to parental sequences (top panel) and focused rational engineering, library approaches yielded a comparatively broad diversity within the resulting cloned library (second logo plot) and distinct sequence motifs upon one CHO display sorting round (third plot) or four consecutive rounds of affinity sorting on the surface of yeast (three bottom logo plots).

CS06 HCDR2+3 sequences of rational design and library outputs.

Rational hydrophobicity engineering

To assess and optimize antibody sequences not only based on their binding affinities, but also regarding their overall developability profiles, we recently described an approach for an early in silico Sequence Assessment Using Multiple Optimization Parameters (SUMO).³³ Using SUMO, rational hydrophobicity reduction was performed in two iterative design cycles after structural antibody modeling and structure-based in silico prediction of aggregation hot spots. In the first design cycle, solvent-exposed hydrophobic residues in HCDR2 or HCDR3 were replaced by specific polar residues: in each position, an alternative residue with similar steric properties but higher polarity was introduced (I59T, F62H, Y111:113 H). Further single-point mutations were inspired from the positional variability observed in a historical affinity maturation library (Figure 2b). These distributions suggested a certain tolerability of several polar residues in positions 111,1 and 111,2 and the tolerability of Asn in position 112,1, as well as Ser in positions 112 and 113. In total, 18 single-point mutations (termed VH3.1 – VH3.18 in Figure 2d) of the parental CS06 sequence were designed. Those variants that revealed reduced hydrophobicity and favorable binding affinities after synthesis and experimental characterization were combined into a total of seven 2-, 3- or 4-point mutations in the second design cycle (VH5.1 – VH5.7 in Figure 2d). Encouragingly, the optimized candidates showed strongly reduced hydrophobicity and binding affinity comparable to the parental CS06 sequence (VH1.0). In agreement with our design rationales, the 4-point mutational variant VH5.1 showed the lowest retention time in HIC and, encouragingly, only an affinity loss of < factor 2 compared to VH1.0.

Hydrophilicity screening library design and YSD sorting

Enabling the search within a larger combinatorial screening space, knowledge obtained from affinity-tolerating hydrophobicity-reducing single-point mutations was applied to rationally design a focused combinatorial library exchanging hydrophobic HCDR2 and HCDR3 key residues to pre-defined portions against any polar or soluble amino acid (Figure 2c). The resulting amino acid distribution was assessed by Sanger sequencing of 96 single clones obtained after multiple rounds of fluorescence-activated cell sorting (FACS) and in high accordance with desired proportions (Figures 2 and 3d). YSD sorting was conducted in four rounds for display and retained target-specific binding. The resulting clones were sequenced (Figure 2e) and reproduced as IgG1 for evaluation of target-specific binding and physicochemical properties, including hydrophobicity (Table S1). All engineered variants showed higher purities, higher thermal stabilities and, most relevant, strongly reduced HIC retentions times similar or below a reference antibody (cetuximab). Affinities were reduced to varying extent, but due to avidity, reduction in cellular binding (as measured by EC50) was limited (Table 1). In silico developability assessment suggested favorable overall profiles for most variants (Table S1).

Figure 3. Yeast surface and mammalian display for hydrophobicity engineering. (a) Yeast surface display sort for display and retained antigen binding plus resulting hit panel for in vitro analyses. (b) Differential display of rationally designed parental (VH1.0) and optimized (VH5.1) CS06 indicating applicability of screening for manufacturability properties such as hydrophobicity by mammalian display. (c) CHO display sort for manufacturability (as indicated by high display) and retained target binding plus resulting most enriched output sequences. (d) Diversity of H-CDR2-CDR3 sequences illustrated using UMAP dimensionality reduction.

Yeast surface and mammalian display for CS06 hydrophobicity engineering.

Mammalian display sorting for manufacturability

As interim results had indicated a strongly reduced hydrophobicity profile for the preliminary engineered variant CS06 VH5.1, the display level on the surface of CHO cells was assessed in comparison to parental VH1.0 sequence. Indeed, display levels differed by 3 to 4-fold (Figure 3b), indicating the feasibility to sort for both enhanced physicochemical properties as well as target affinity via a mammalian display approach. Library cloning was employing the same diversity as before, and sorting was conducted in a similar fashion to YSD, with a gating strategy aiming at high displaying clones with retained target binding. Evaluation was conducted by NGS and analysis of most prevalent clones showing suitable in silico developability properties.

Comparative sequence and antibody evaluation

We compared the output of three approaches in several dimensions: First, amino acid exchanges occurred in similar positions and quantities in all three resulting panels and correlated with previous data (Figure 3 versus Figure 4). Secondly and as expected, amino acids variability was higher from library output then rational design, reflecting the significantly higher combinatorial space (Figure 3a+c), and higher from NGS analyses than Sanger sequencing due to the different nature of these sequencing technologies. Third, there was overlap of resulting sequence variants between all three approaches (d). Comparative analyses of resulting properties between antibodies derived from the different optimization approaches revealed enhanced thermal stability and lowered hydrophobicity in variants derived from all approaches, particularly from CHO display and YSD. 14 of 30 variants identified by YSD and Sanger sequencing showed affinities within 10-fold difference to the parental clone and most similar cellular binding potency due to avidity (Table S1). Hits from CHO display sorts were evaluated by NGS and chosen for high prevalence and suitable in silico developability properties. Of 12 clones, 3 still bound c-MET with affinity loss higher than 10-fold (Table S1).

Figure 4. Improved conjugatability and high potency of engineered biparatopic ADCs. (a) Scheme of final biparatopic ADC. (b) Key data on biparatopic ADC conjugation. (c) in vitro potency of biparatopic ADCs on c-MET expressing human tumor cell lines NCI-H441, HCC-827 and EBC-1 in comparison to a reproduced exatecan DAR6 REGN5093 reference. Dose-response curves of REGN5093, 6.21_scFv and 6.18_Fab depicted are from one representative experiment and had been merged with the dose-response curve of 6.21_Fab, obtained from an independent experiment. Complete data sets are listed in table S3 that indicate comparable potencies between independent experiments, serving as basis for a merged representation shown as representative graphs with error bars indicating mean ± SD of technical triplicates.

Improved conjugatability and high potency of engineered biparatopic anti-c-MET ADCs.

Based on enhanced hydrophobicity profiles, suitable physicochemical properties, retained cellular binding as IgG1 and absence of in silico developability flags, CS06 variants VH6.18 and VH6.21 were chosen for assembly of final lead candidates in combination with sequence optimized B10v5. Although suitable regarding binding and physicochemical qualifications, the VH5.1 sequence variant was excluded from further comparative studies due to a minor in silico immunogenicity risk flag (data not shown). Biparatopic antibodies were produced as N-terminal fusions of Fab and scFv to a knob-into-hole heterodimeric Fc portion.³⁴ Reference construct one-armed B10v5 was historically available as SEED construct.³¹ While yields and purities were acceptable, thermal stabilities were enhanced and HIC retention times were strongly reduced in all CS06 VH6.18 and VH6.21 variants when compared to the parental VH1.0 reference (Table 1). Affinities of one-armed engineered variants were reduced to the three-digit nanomolar range, with CS06 VH6.18 and VH6.21 Fab constructs retaining cellular binding EC50s in the range of 30–50 nM.

When rendered biparatopic, however, apparent affinities to recombinant target protein were too high to measure due to very slow off-rates and cellular binding EC50s were in the picomolar range. CS06 Fab x B10v5 scFv constructs yielded higher binding potency and maximal binding in flow cytometry studies than monovalent references (Figure S2). For further early developability assessment, the engineered constructs were evaluated in affinity-capture self-interaction nanoparticle spectroscopy (AC-SINS) as an early experimental predictor for colloidal stability³⁵ using clinical antibodies trastuzumab and briakinumab as known assay references with low and high propensities for self-interaction.³⁶ In agreement with the engineering strategy and the experimental HIC data, comparison of the AC-SINS data for the different sequence pairs (Table 1) reveals reduced self-interaction of the CS06 VH6.18 and VH6.21 variants compared to the parental VH1.0 derived variants. Finally, four biparatopic variants (B10v5 scFv x parental VH1.0_Fab and the engineered variants B10v5 scFv x VH6.18_Fab; B10v5 scFv x VH6.21_Fab and B10v5 Fab x VH6.21_scFv) were evaluated in the polyspecificity reagent (PSR) assay,³⁷ which provides insights into the general off-target interactions/specificity, again using trastuzumab as assay reference for reduced unspecific interactions and briakinumab reference indicating more pronounced polyspecificity. The engineered CS06 variants VH6.18 and VH6.21 showed reduced nonspecific binding compared to VH1.0 (Table 1).

In summary, suitable biophysical properties for the three lead combinations, B10v5 scFv x CS06 VH6.18 Fab, B10v5 scFv x CS06 VH6.21 Fab, and B10v5 Fab x CS06 VH6.21 scFv (further referred to as 6.18_Fab; 6.21_Fab; 6.21_scFv), suggested assessment of improved capability for payload conjugation in comparison to parental bispecific B10v5 scFv x CS06 VH1.0 Fab (termed 1.0 Fab).

Antibody drug conjugation and potency evaluation

The drug-linker in this study comprises a maleimide, a cleavable β-glucuronic acid-based linker, and the topoisomerase 1 inhibitor exatecan.³⁸ Topoisomerase 1 inhibitors inhibit the enzyme’s normal function, leading to the accumulation of DNA strand breaks. This disrupts replication and transcription processes, ultimately causing cell death. The drug-linker is attached to the interchain cysteines of the targeting antibody via the maleimide. Upon internalization into the target cell, the β-glucuronic acid-based linker is cleaved to release exatecan.

To evaluate the impact of manufacturability optimization on ADC production efficiency and yields, conjugation trials with the parental 1.0_Fab and the engineered variants 6.18_Fab; 6.21_Fab and 6.21_scFv were performed. First, analytical test reactions were set up to assess conjugation efficiency and monomeric purity of the conjugates. To attach the thiol-reactive drug-linker, interchain disulfide bridges were reduced with Tris(2-carboxyethyl)phosphine (TCEP) and afterward incubated with the drug-linker for 2 h. Subsequently, monomeric purity was analyzed via analytical size exclusion chromatography (SEC) and drug-to-antibody ratio (DAR) via polymeric reversed-phase HPLC (PLRP). While conjugation efficiency appeared to be most efficient for 1.0_Fab and 6.18_Fab (both DAR = 5.8, Figure 4b), it was slightly less efficient for 6.21_scFv (DAR = 5.3) and 6.21_Fab (DAR = 5.1). In contrast, the monomeric purity of the 1.0_Fab conjugate was rather low (76.7%) and a substantial amount of low molecular weight species (LMW, 23.3%) was present, indicating a low stability of this conjugate. The engineered constructs, however, showed a much higher purity, with 6.18_Fab: 90.0%; 6.21_scFv: 91.3%; and 6.21_Fab: 92.6%. In a next step, larger reactions were carried out including preparative SEC purification and concentration to assess the behavior of the respective conjugates throughout different downstream processing (DSP) steps. Analytical SEC and PLRP analyses yielded similar observations as before of a high degree of LMWs in 1.0_Fab preparative SEC. Although the LMWs could be separated, subsequent ultrafiltration was not successful as all material was adsorbed to the filter membrane, omitting recovery of significant sample amounts. Such effects can be critical during manufacturing since many DSP steps for ADCs typically include ultrafiltration steps. Although the conditions for conjugation were not optimized, improved yields were obtained for engineered variants with best final yield for 6.18_Fab with 36%, followed by 6.21_scFv with 23%, and 6.21_Fab with 16%. In addition, an internally reproduced REGN5093 biparatopic antibody was used to generate a comparable DAR6 exatecan ADC to enable comparison of antibody backbones’ suitability for potent cytotoxicity.

Overall, while several approaches failed to produce an ADC from parental CS06 VH1.0 containing biparatopic references, all three engineered biparatopic ADCs could be conjugated and produced with reasonable yields, demonstrating that CS06 hydrophobicity engineering had a positive impact on ADC producibility. To demonstrate that the hydrophobicity engineering did not alter ADC functionality, in vitro potency on several tumor cell lines was evaluated and shown to be in the sub-nanomolar to one-digit nanomolar range for high (Figure 4) and two-digit nanomolar range for low c-MET-expressing cell lines (Table S3). Throughout all levels of expression, optimized biparatopic ADCs showed potencies similar to an internal REGN5093 biparatopic antibody-based exatecan DAR6 reproduced ADC and higher than IgG1-based DAR8 reference ADCs (Figure 4c and Table S3).

Discussion

Next-generation designs can overcome past obstacles of therapeutic anti-c-MET biologics by supporting differentiated modes of action. Biparatopic antibodies are a promising approach that can lead to enhanced cellular binding, increased internalization by target cross-linking, and degradation, resulting in signal transduction inhibition and higher potency cytotoxicity when applied as ADCs. Several approaches have been described for degradation-enhancing mono-specific biologics,^15,16 bispecific antibodies aiming at crosslinking of c-MET with another tumor-associated antigen,^14,39 or further approaches for antibody-induced target degradation,^40,41 which have been reviewed recently by Anh et al.⁴² and Lin and coworkers.⁴³ There is broad evidence that an optimal combination of paratopes can foster inter-molecular target binding and its degradation,^18,19,44,45 but, in addition to target removal, the forced degradation of antibodies or, if applied, ADCs can yield enhanced payload delivery and cytotoxic potency.^22,46 While optimal paratope combinations may be identified in large unbiased screens as well,^47,48 modeling-supported choice of paratope combination for preferential inter-target binding could save resources and shorten discovery times.

Whereas our initial biparatopic designs based on CS06 and B10v5 provided high apparent affinities and strong internalization capacities for c-MET degradation and ADC approaches, subsequent experimental assessment revealed a substantial sequence liability that was attributed to in silico predicted pronounced hydrophobic patches in H-CDR2 and H-CDR3 of CS06. The approach described here represents, to our knowledge, the first comparative protein engineering for manufacturability and applicability to ADC conjugation and yields deep insights into the capabilities of rationale and library approaches, as well as enhanced conjugation properties and a potent biparatopic anti-c-MET ADC.

A comparison of CHO versus YSD was based on recent reports on manufacturability screening on the surface of mammalian^49–51 or yeast cells.⁵² Differential display of optimized to parental CS06 variants, to our knowledge, proves for the first time the ability to screen for developability properties beyond aggregation propensities such as a suitable hydrophobicity profile. Different numbers of sorting rounds and post-sort sequence identification methodologies do not allow a full comparison of yeast versus CHO display outputs. Both approaches yielded optimized variants with high thermal stabilities and hydrophilicities, indicating successful enrichment for “manufacturability”. Both sort outputs indicate conversion over time to parental tyrosine residues in essential positions (111; 112,2; 112,1; 112) and distinct suitable amino acids, e.g., H/R in position 111,1 or H/D in position 111,2.

Whereas sequence information was available for rounds 2–4 of YSD-enabling variant selection from yet diverse round 2/3 output while neglecting fully converged round 4 samples, only one round of CHO display sorting was conducted and available for hit identification. The resulting success rates for retained target binding were higher for yeast-derived clones. Although the choice of most prevalent variants in sort output NGS datasets can yield suitable hits,⁵³ CHO hit rates could be further enhanced by application of deep learning focusing on pre- versus post-sort enriched sequences⁵⁴ and iterative sorting in 2 to 3 rounds. Full in vitro and in silico qualification of resulting CS06 variants indicated strongly enhanced physicochemical properties, but in turn mostly reduced target affinities (Table S1). The combination of a B10v5 and CS06 binding moiety into a heterodimeric biparatopic antibody restored and exaggerated apparent affinities comparable to a similar recent report (DaSilva 2020) and suitable for subsequent ADC generation. Although not a focus of this work, ADCs with lower c-MET affinities could yield advantages of lower normal tissue on-target engagement, resulting in an enhanced therapeutic index.⁵⁵

While several approaches failed to produce an ADC from parental CS06 VH1.0 containing biparatopic references, all three engineered biparatopic ADCs could be conjugated and produced with reasonable yields, indicating CS06 hydrophobicity engineering had a positive impact on ADC producibility. Engineering and exatecan conjugation resulted in highly active biparatopic ADCs showing similar in vitro potency on high to low level c-MET expressing cell lines as a reproduced REGN5093 exatecan ADC containing the same drug-linker and comparable DAR (Figure 4 and Table S3). The molecule format chosen here restricted the maximal achievable DAR to 6, yet enabled sufficient internalization and payload delivery to obtain sub- to low nanomolar potencies, supporting the therapeutic applicability of such a biparatopic ADC.⁵⁶ As the main focus of the work presented herein was on optimization of antibody sequences, please note that conjugation efficiencies can depend not only on antibody backbones, but also on drug-linker, the comparability is currently limited to in vitro and that further optimization of conjugation processes is warranted to yield suitable cost of goods supporting commercial development of such therapeutic candidates.

This study highlights the relevance of taking early developability aspects into consideration for sequence selection and optimization. Pronounced hydrophobicity might not only result in poor physical stability, poly-specificity and increased in vivo clearance, but also hinder the manufacturing and, as demonstrated here, the payload conjugation process for ADCs. Encouragingly, such labilities can be detected by early in silico and experimental assessment and can be overcome by rational design and library approaches, even in cases where these liabilities are located within critical CDR regions and when an experimental crystal structure is not available.

In conclusion, this study highlights the substantial advancements made in the field of biparatopic antibody engineering for targeting c-MET. By strategically optimizing hydrophobicity and using rational design and combinatorial library approaches, we successfully developed highly potent and manufacturable biparatopic ADCs. These methodologies could serve as a blueprint for rapid development of optimized biparatopic ADCs targeting further tumor-associated antigens in the future.

Materials and methods

Molecular modeling of mono-specific and multi-specific antibody structures

Structural models of variable domains and full-length antibodies were created using the antibody modeler tool in the molecular modeling software package MOE.⁵⁷ scFv constructs and biparatopic constructs were built by adding linkers via MOE’s protein builder, followed by a conformational search of the linker via MOE’s linker modeler. Finally, the modeled constructs were subjected to energy minimization. Visualization of 3D structures was done with PyMOL.⁵⁸

Antibody-antigen docking

The crystal structure of v-MET was taken from PDB code 2UZY. Docking hypotheses of the variable antibody regions shown in Figure 1b were generated using PIPER⁵⁹ via Schrödinger’s BioLuminate.⁶⁰

In silico developability assessment

The in silico developability profiles (shown in Supplemental Table S1) were computed using an internal pipeline termed SUMO.³³ This approach automatically generates antibody models based on the provided sequences of the variable regions, identifies the human-likeness by sequence comparison to the most similar human germline sequence, determines structure-based surface-exposed chemical liability motifs (unpaired cysteines, methionines, asparagine deamidation motifs and aspartate deamidation sites), as well as sites susceptible to post-translational modification (N-linked glycosylation). Moreover, a small set of orthogonal computed physico-chemical descriptors including the isoelectric point (pI) of the variable domain, Schrödinger’s AggScore⁶¹ as a predictor for hydrophobicity and aggregation tendency calculated for the complete variable domain as well as the complementarity-determining regions (CDRs) only and the calculated positive patch energy of the CDRs were determined. These scores were complemented with a green to yellow to red color coding, indicating scores within one standard deviation from the mean over a benchmarking dataset of multiple biotherapeutics approved for human application⁶² as green, scores above one standard deviation as yellow and those above two standard deviations as red. For the AggScore values, these cutoffs were slightly adjusted based on correlation analyses to internal experimental HIC data.

Combinatorial variant library design and generation

DNA libraries were generated via synthesis with S. cerevisiae codon usage and mutations incorporated into the parental CS06 VH sequence as shown in Figure 1b (Twist Biosciences):.two positions in H-CDR2 and seven positions in H-CDR3 were exchanged against hydrophilic amino acids. The library was constrained to contain 5% single-point and 10% double-point mutations in H-CDR1 and 15% double-point as well as 70% triple-point mutations in H-CDR3. In addition, specific amino acid motifs, relating to glycosylation motifs, asparagine deamidation, aspartate deamidation, lysine glycation, integrin binding, CD11c/CD18 binding, fragmentation and hydrophobicity were explicitly excluded.

For YSD Fab library generation, randomized CS06 VH gene strings were amplified using CS06 FR1 and FR4-specific primers carrying yeast gap repair overhangs (CS06_Twist_fwd: TGTTTTTCAATATTTTCTGTTATTGCTAGCGTTTTAGCAGG-Gcaagtccaattagttcaa and CS06_Twist_rev: AGAAGATGGAGCCAATGGAAAAA-CAGATGGACCTTTTGTAGAAGCagaagagacagtgac). CS06 VL gene was synthesized as a gene string at GeneArt and was also amplified using specific primers to introduce gap repair overhangs (CS06_VL_fwd: GCCAGCATTGCTGCTAAAGAAGAAGGGG-TACAACTCGATAAAAGAcaattggtcttgactcaatc and CS06_VL_rev: GGATGGCG-GGAACAGAGTGACCGAAGGGGCGGCCTTCGGCTGACCagaacggtcaattttgtac). Next, VH library was cloned into YSD vector using S. cerevisiae strain EBY100 MATa (URA3–52 trp1 leu2D1 his3D200 pep4:HIS3 prb1D1.6 R can1 GAL (pIU211:URA3)) and the VL chain was cloned using BJ5464 cells (MATalpha URA3–52 trp1 leu2D1his3D200 pep4:HIS3 prb1D1.6 R can1 GAL) via gap repair cloning according to the optimized protocol of Benatuil et al.⁶³ EBY100 cells comprising the randomized heavy chain diversity were mated with BJ5464 cells comprising the CS06 VL chain to produce diploid yeast cells capable of Fab display.⁶⁴

CS06 display CHO library was cloned in a similar fashion as described above to realize direct fusion of the randomized CS06 heavy chain to GGGGS-linked transmembrane domain of PDGFR. The library was subsequently generated as reported in Gaa et al. ⁶⁵

Library sorting

YSD sorting was performed as described in references.^66,67 Briefly, CS06 Fab display was induced by incubation of diploid yeast library cells in SG-Trp-Leu medium + 10% (w/v) polyethylene glycol 8000 for 48 h at 20°C, 120 rpm agitation. For sorting rounds 1 and 2, yeast cells were incubated with 1 µM (rh)c-MET-ECD-His6 (inhouse) in phosphate-buffered saline (PBS), while for sorting rounds 3 and 4 the antigen concentration was reduced to 8 nM and 0.5 nM c-MET-ECD-His6, respectively. Surface display of correctly assembled Fab molecules was detected using 25 µg/ml goat-F(ab’)₂ anti-human lambda-PE antibody (Southern Biotech) and specific c-MET-ECD-His6 antigen binding was detected using 5 µg/ml SureLight APC anti-6×His tag antibody (Abcam) in PBS. Cell sorting was performed using BD FACSAria Fusion flow cytometer. Following FACS, sorted cells were transferred to 50 ml SD-Trp-Leu cultivation medium and were expanded at 30°C, 120 rpm agitation for 48–72 h.

CHO library sorting for high display and target binding was conducted applying a Sony SH800S flow cytometer after staining with 100 nM c-MET-ECD-His6 in PBS and detected using 15 µg/ml anti-Penta·His Alexa Fluor 647 conjugate (Qiagen) in PBS following manufacturer’s recommendations. Sorting was conducted applying a 100 µm nozzle in sort mode “purity” with gates adjusted to include approximately 2% events, as shown in Figure 3c and Figure S1f-j.

Sequence evaluation and comparative analyses

Sanger sequencing was used for sequence analyses of YSD output diversities, and the resulting forward + reverse consensus sequences were analyzed with Geneious Prime software version 2022.1.1.

For CS06 mammalian display sort output evaluation, bulk cell output cDNA was generated as previously described.⁶⁵ The CS06 VH sequences were amplified using fusion primers targeting the flanking vector sequence. The VH amplicons were purified with AMPure (Beckman Coulter) and amplified with index primers for Illumina sequencing. The final sequencing library was purified using a Pippin Prep (Sage Science). The VH domains were sequenced on a MiSeq using the v3 600 cycle kit according to the manufacturer’s protocol. FASTQ files were uploaded to Geneious Biologics (https://www.geneious.com/biopharma) for analysis. Reads were overlapped, filtered for length, and the VH sequences were annotated using the IMGT/V-QUEST reference library (https://www.imgt.org/vquest/refseqh.html). Normalized counts for HCDR2 and HCDR3 were used to identify amino acid variants of high prevalence.

The in silico developability profile was computed using SUMO.³³ In brief, automatically generated VH models allow evaluation of “humanlikeness” by sequence comparison to the most similar human germline and determine structure-based surface-exposed chemical liability motifs and N-linked glycosylation. In addition, orthogonal physicochemical descriptors such as the pI of the variable domain, Schrödinger’s AggScore as predictor for hydrophobicity and aggregation tendency, and the calculated positive patch energy of the CDRs are determined. These scores were complemented by color coding, indicating scores within one standard deviation from the mean over a benchmarking dataset (Ahmed Gupta Martin 2021) as green, scores above one standard deviation as yellow, and above two standard deviations as red. For comparative visualization of sequence panels derived from three different approaches, the respective sequences pools were projected into a two-dimensional space using UMAP.⁶⁸

Engineered antibody qualification

Recombinant transient antibody expression and purification and thermal stability determination were conducted as described by Yanakieva et al. ⁶⁹ HIC analysis was conducted via two methods with varying gradients. For all processes, a butyl-non-porous resin (NPR, 2.5 µm 4.6 mm × 100 mm) column (TOSOH Bioscience 42,168) was applied at 25°C and protein samples formulated in PBS pH 7.4 were mixed with ammonium sulfate at 1 M end concentration prior to analyses. A long gradient method was run at a flow rate of 0.75 ml/min using a linear gradient of 50 mM sodium phosphate +1.5 M ammonium sulfate pH 7.0 to 50 mM sodium phosphate + 5% isopropanol pH 7.0 in 33 min. A short gradient method was applied with a flow rate of 0.5 ml/min using a linear gradient of 1.2 M ammonium sulfate, 1× PBS, pH 6.47, 170.1 mS/cm to 50% methanol, 0.1× PBS, pH 8.39, 0.998 mS/cm in 15 min. Typically, 20 μg of protein sample were loaded onto the column. Absorbance was monitored at 214 nm using a multi-wavelength detector (Agilent). ChemStation software (Agilent) was used to integrate the peak areas.

Affinity determination was conducted by biolayer interferometry (BLI) as previously described (Sellmann 2016) EC50 values were determined via flow cytometry (iQue 3 screener, Sartorius) in 1:2 titration series from 500 nM to 0.01 nM on c-MET expressing human lung carcinoma EBC-1 cells (Riken Bioresource Center Cel Bank JCRB0920 031496, cultured under recommended conditions) and AF-488 AffiniPure Fab Fragment Goat-anti-Human IgG, Fcγ fragment specific detection antibody (#109-547-008, Jackson ImmunoResearch) Analyses were conducted with GraphPad Prism 9.1.2 software (GraphPad Software LLC).

ADC generation

ADCs were generated via interchain reduction and subsequent reaction with maleimide-containing drug-linkers (DL). MAbs were thawed at 20°C and mAb concentration was adjusted to 5 mg/ml using conjugation buffer (50 mM histidine pH 6.5, 100 mM NaCl). Subsequently, interchain disulfides were reduced by incubating the mAb solution with 10 molar equivalents (relative to the mAb) of TCEP at 20°C for 2 h. Afterwards, the solution was incubated with 12 molar equivalents (relative to the mAb) of DL for 1 h at 20°C to conjugate the DL to the antibody. The reaction was stopped by addition of 12 molar equivalents (relative to the mAb) of N-acetyl cysteine (NAC) for 20 min at 20°C. ADCs were separated from unconjugated DL via SEC using a HiLoad Superdex 200 Increase column in combination with an Äkta LC system (Cytiva) and conjugation buffer as running buffer. Fractionated samples containing the ADC material were pooled and concentrated using Amicon Ultra 15 50K centrifugal (Millipore) followed by a final buffer exchange into 10 mM Histidine, 40 mM NaCl, 6% Trehalose, 0.05% TWEEN, pH 5.5 using HiTrap Desalting columns in combination with an Äkta LC system (Cytiva). Final ADC material was filtered through a 0.22 µm sterile filter unit (Millipore) and shock frozen in liquid nitrogen until further use.

In vitro potency assessment

In vitro potency assessment of the ADCs was determined on c-MET expressing human lung cancer cell lines HCC-827 (lung adenocarcinoma, ATCC^Ⓡ CRL-2868^TM, originally obtained from the American Type Culture Collection (ATCC), Manassas, VA, USA), NCI-H441 (lung adenocarcinoma, papillary, ATCC^Ⓡ HTB-174^TM), EBC1 (squamous cell carcinoma, Health Science Research Resources Bank, now National Institutes of Biomedical Innovation, Health and Nutrition, Japanese Cancer Research Bank, JCRB0920 031496) and A549 (lung carcinoma, ATCC^Ⓡ CCL-185 ^TM) using the CellTiter Glo^Ⓡ Luminescent Cell Viability assay (#G7575, Promega Corporation, Madison, WI, USA).

The cell lines HCC-827 and NCI-H441 were cultured in RPMI1640 medium with GlutaMAX ^TM supplement (#61870–010, Gibco ^TM, purchased from Thermo Fisher Scientific, Waltham, MA, USA), 1 mM sodium pyruvate (#113670–070, Gibco ^TM, Thermo Fisher Scientific), 2.5 g/L Glucose (#G8769, Sigma Aldrich, St. Louis, MO, USA or #A24949–01, Gibco ^TM, Thermo Fisher Scientific) and 10% fetal bovine serum FBS) (#S0615, Sigma Aldrich). The cell line EBC-1 was cultured in MEM Eagle Medium (#M2279, Sigma Aldrich) including 2 mM glutamine (#35050–061, Gibco ^TM GlutaMAX ^TM, Thermo Fisher Scientific) and 10% FBS (#S0615, Sigma Aldrich). The cell culture medium of A549 cell line DMEM including 2 mM glutamine (#41964–039, Gibco ^TM) and 10% FBS.

The day before treatment, 2500 cells/well (NCI-H441 or EBC1) or 1250 cells/well (HCC827) or 625 cells/well (A549) were plated in sterile 96-well flat-bottom microplates (90 µl volume each, #165303, Thermo Fisher Scientific) and incubated at 5% CO2, 37°C. Background wells were supplied with respective culture medium. After overnight incubation, a 10× starting concentration of the ADCs and a 1:4 serial dilution was prepared with RPMI1640 medium with GlutaMAX ^TM supplement, 1 mM sodium pyruvate and 10% FBS. A total of 10 µl was added to the respective wells in technical triplicates. Control wells were treated with a respective amount of RPMI1640 medium. Following 6 days of incubation, 100 µl CellTiter Glo _® reagent was added to each well and plates were incubated for 2 min with shaking at 300 rpm and for additional 20 min at room temperature, protected from light. Afterwards, luminescence was measured on a Varioskan Flash plate reader (Thermo Fisher Scientific). Relative light units (RLU) were processed by subtracting the background and by normalization the data to untreated control cells. The processed data was used to describe the dose-response by %effect vs. concentration [M] with the equation log(inhibitor) vs. response – variable slope (four parameters) (GraphPad Prism version 8.2.0 for Windows, GraphPad software, La Jolla, California, USA). Graphs were displayed with error bars indicating the standard deviation (SD) of technical triplicates. For determining IC₅₀ values, data were processed by using Genedata Screener (Genedata). Experiments were performed several times and Geometric mean values of determined IC₅₀ values were indicated as geomean IC₅₀ [M].

Ac sins

Molecules were captured onto particles via immobilized capture antibodies and self-association was judged in PBS buffer at pH 7.4 by shifts in the plasmon wavelengths.⁷⁰ Trastuzumab was used as control indicating favorable biophysical properties with mean Δλmax values of ~0.2 nm after subtraction of buffer blanks. Final AC-SINS scores for molecules were calculated via subtraction of blank and trastuzumab scores.

PSR-BLI

To assess nonspecific antibody interactions to PSR, a published cytometric assay³⁷ was adapted for the application of fast and sensitive BLI. PSR was derived from soluble membrane proteins (SMP) of CHO and HEK293-6E cells as described by Xu et al.³⁷ Assays were performed at 25°C with orbital sensor agitation at 1,000 rpm in 200 µL volume with DPBS. Pre-hydrated AHC biosensors were loaded with antibody (10 µg/mL) for 300 s. Afterwards biosensors were blocked with 1% bovine serum albumin for 200 s and a baseline was established by rinsing in DPBS for 60 s. Association with 20 µg/mL PSR (1:1 mixture of CHO and HEK293-6E SMP) was performed for 100 s. As reference, association was performed in DPBS. To calculate the PSR-BLI score, the binding response from the association step was normalized to the reference measurement by subtraction, followed by subsequent subtraction with non-loading control (DPBS).

Abbreviations

AI	=	artificial intelligence
ADC	=	antibody-drug conjugate
AF-488	=	Alexa Fluor 488 nm
BLI	=	biolayer interferometry
c-MET	=	mesenchymal-epithelial transition receptor
CHO	=	Chinese hamster ovary
DAR	=	drug-to-antibody ratio
DL	=	drug-linker
DSP	=	down-stream process
FBS	=	fetal bovine serum
HCDR2/3	=	heavy chain complementary determining region 2 or 3
HIC	=	hydrophobic interaction chromatography
IC50	=	half-maximal inhibitory concentration
LMW	=	low molecular weight
NGS	=	next-generation sequencing
VH	=	variable domain of an antibody heavy chain
VL	=	variable domain of an antibody light chain
PLRP	=	polymeric reversed-phase HPLC
SEC	=	size exclusion chromatography
SUMO	=	sequence assessment using multiple optimization parameters
TCEP	=	Tris(2-carboxyethyl)phosphine
YSD	=	yeast surface display

Acknowledgments

We thank Thomas Clarke for support with NGS analyses; Carolin Sellmann for supplying initial CS06 HCDR3 information; Dirk Mueller-Pompalla, Mamatha Chandar and team for antibody qualification data; Iris Schmidt and Irina Onofrei for in vitro potency data.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2024.2302386.

Additional information

Funding

The author(s) reported there is no funding associated with the work featured in this article.