Structure- and machine learning-guided engineering demonstrate that a non-canonical disulfide in an anti-PD-1 rabbit antibody does not impede antibody developability

ABSTRACT

Rabbits produce robust antibody responses and have unique features in their antibody repertoire that make them an attractive alternative to rodents for in vivo discovery. However, the frequent occurrence of a non-canonical disulfide bond between complementarity-determining region (CDR) H1 (C35a) and CDRH2 (C50) is often seen as a liability for therapeutic antibody development, despite limited reports of its effect on antibody binding, function, and stability. Here, we describe the discovery and humanization of a human-mouse cross-reactive anti-programmed cell death (PD-1) monoclonal rabbit antibody, termed h1340.CC, which possesses this non-canonical disulfide bond. Initial removal of the non-canonical disulfide resulted in a loss of PD-1 affinity and cross-reactivity, which led us to explore protein engineering approaches to recover these. First, guided by the sequence of a related clone and the crystal structure of h1340.CC in complex with PD-1, we generated variant h1340.SA.LV with a potency and cross-reactivity similar to h1340.CC, but only partially recovered affinity. Side-by-side developability assessment of both h1340.CC and h1340.SA.LV indicate that they possess similar, favorable properties. Next, and prompted by recent developments in machine learning (ML)-guided protein engineering, we used an unbiased ML- and structure-guided approach to rapidly and efficiently generate a different variant with recovered affinity. Our case study thus indicates that, while the non-canonical inter-CDR disulfide bond found in rabbit antibodies does not necessarily constitute an obstacle to therapeutic antibody development, combining structure- and ML-guided approaches can provide a fast and efficient way to improve antibody properties and remove potential liabilities.

KEYWORDS:

• Complementarity determining regions (CDRs)
• disulfide bond
• humanization
• ML/AI-guided protein design
• Programmed cell death (PD-1)
• Rabbit monoclonal antibody (mab)

Introduction

Rabbit antibodies have emerged as outstanding reagents in research, diagnostic, and therapeutic applications¹ because of their high affinity, diverse epitope recognition, specificity, and stability.² In addition, their structure is unique compared with antibodies from other species.^3–5 These characteristics make rabbits an attractive alternative to rodents for in vivo discovery. In contrast to rodent IgGs, one of the distinguishing features of rabbit IgGs is the presence of non-canonical disulfide bonds, which are frequently found between the kappa light chain inter-domain of VL:Cys80 and CL:Cys170 (C80-C170), and the heavy chain intra-domain of complementarity-determining region (CDR) H1:Cys35a and CDRH2:Cys50 (C35a-C50) by Kabat numbering.⁶ Unlike the C80-C170 disulfide bond, which has been well characterized and contributes to antibody thermal stability, the role of the C35a-C50 disulfide bond has not been studied as extensively, especially with regard to functionality and developability.⁷ As its presence is typically considered a liability, rabbit-derived antibodies with a C35a-C50 disulfide bond but otherwise favorable properties are generally not considered clinical candidates.

Programmed cell death (PD-1), an immune checkpoint receptor for the PD-1 ligand-1 (PD-L1) and PD-1 ligand-2 (PD-L2) proteins, has been validated as one of the most promising targets for cancer immunotherapy.⁸ Despite being valuable tools for efficacy, pharmacokinetic, and toxicity studies in preclinical mouse models, developable mouse/human cross-reactive PD-1 antibodies that could potentially become clinical candidates for combination or bispecific antibody therapies are scarce. Hoping to leverage the unique properties of the rabbit antibody repertoire, we performed a rabbit antibody discovery campaign against PD-1. Surprisingly, all the human/mouse PD-1 cross-reactive and PD-L1 blocking antibodies we identified contained the C35a-C50 disulfide bond, suggesting not only a functional role for this disulfide, but also a potential liability preventing further development of any human/mouse PD-1 cross-reactive antibody after humanization.

To understand the contribution of this non-canonical disulfide bond to antibody stability and antigen binding, we selected the human/mouse cross-reactive PD-1 blocking rabbit antibody with the highest affinity (termed rbt1340) and generated its humanized version (termed h1340.CC; where CC refers to the presence of the C35a-C50 disulfide bond). We determined the structure of h1340.CC in complex with PD-1, revealing the detailed epitope of a human/mouse cross-reactive PD-1 blocking antibody that has not been disclosed before. Using structure- and sequence-guided protein engineering, we generated a variant (termed h1340.SA.LV) that lacks the C35a-C50 disulfide but maintains PD-1 binding and PD-L1 blocking properties similar to h1340.CC, albeit with lower affinity. Developability assessment of the parent molecule (h1340.CC) and the non-disulfide variant (h1340.SA.LV) indicates that both antibodies have similar physicochemical properties and immunogenicity profiles and are therefore equally developable into clinical candidate molecules. In addition, to test whether we could rapidly generate an alternate h1340 variant without the C35a-C50 disulfide with further improved properties, we used an unbiased structure- and machine learning (ML)-guided engineering approach to generate a new variant with fully recovered affinity.⁹

To our knowledge, our study is the first to directly compare the molecular properties of a humanized rabbit antibody with or without its non-canonical C35a-C50 heavy chain disulfide bond, and to suggest that such disulfide does not necessarily need to be engineered out to enable therapeutic antibody development, while also proposing a rapid, structure- and ML-guided route to engineer variants with improved properties.

Results

Anti-PD-1 antibody discovery in rabbits

In our anti-PD-1 rabbit antibody campaign (see Materials and methods section), we successfully applied our previously described single B cell antibody discovery platform to generate 3 panels of monoclonal antibodies (mAbs) that were either specific to human PD-1 (hPD-1), mouse PD-1 (mPD-1) or cross-reactive to both (h/mPD-1).¹⁰ In the h/mPD-1 panel, we identified a total of2 unique antibodies with low sequence homology in CDRs (55%) and long CDRH3 length (18–22 amino acids), which all possess a non-canonical disulfide bond between CDRH1 (C35a) and CDRH2 (C50). These mAbs all bind with similar affinities to hPD-1 (0.1–7.3 nM) and cynomolgus monkey PD-1 (cynoPD-1) (0.2–8.9 nM), whereas the affinities for mPD-1 span a much broader range (0.3–58.9 nM) (Table S1). These cross-reactive mAbs also demonstrated strong cell surface binding to hPD-1 and mPD-1 expressing cells over a non-expressing control cell line, as indicated by a signal increase of more than two orders of magnitude in fluorescence-activated cell sorting (FACS) measurements (Table S1). Most mAbs showed complete blocking (maximum inhibition greater than 90%) for both hPD-1 and mPD-1, with only a few clones (rbt1183, rbt1168, rbt1339, and rbt1133) showing complete blocking for hPD-1, but partial blocking for mPD-1 (maximum mPD-L1 inhibition between 64% and 80%) (Table S1). The observed weaker mPD-1 inhibition is likely due to the weaker (double-digit nM) affinity for mPD-1 (Table S1). Epitope binning against hPD-1 indicated that all of the cross-reactive mAbs cluster in the same bin and overlap with the epitope of pembrolizumab (Table S1).

Lead antibody humanization

To advance and enable development of a potential PD-1 antibody for non-human primate pharmacokinetics, efficacy, and safety studies, we humanized the lead molecule rbt1340, which displayed the highest affinity against all three PD-1 species we tested and the highest maximum inhibition of human PD-1/PD-L1 and mouse PD-1/PD-L1 interactions (Table S1). The closest human frameworks for this antibody were identified as IGKV1–6 × 01 for the variable light chain (VL) and IGHV3–23 × 04 for the variable heavy chain (VH) from our database (AbGrafter, Genentech). Subsequently, rbt1340 CDRs, which covered the definition of Kabat and Chothia,⁶ along with2 rabbit framework residues at the Vernier zone (position 2, 4, and 43 of VL, and position 2, 48, 49, 71, 73, 75, 76, 78, and 91 of VH) were grafted onto their respective acceptor frameworks to generate humanized VL and VH gene segments (Figure S1). This primary humanized version, referred to as h1340.CC, can be further optimized by framework Vernier permutation from original rabbit residues to the corresponding human germline residues and meets the World Health Organization standards for a humanized antibody (data not shown). Importantly, h1340.CC retained PD-1 cross-reactivity, with less than 2-fold loss in binding affinity to human and cyno PD-1 and less than 3-fold loss in binding affinity to mouse PD-1 compared to rbt1340 (Figure S2).

As part of our humanization efforts, we also generated an initial h1340.CC variant, termed h1340.SA, where we removed the non-canonical C35a-C50 disulfide, assuming it would be a liability preventing further development. In the h1340.SA variant, we mutated C35a to serine and C50 to alanine to match the residues found in positions 35 and 50 of the related IGHV3–23 × 04 VH framework, respectively. However, whereas h1340.CC only displayed 2–3-fold affinity losses across species compared to rbt1340, h1340.SA displayed a 4-fold, 6-fold, and1-fold decrease in affinity for human, cyno and mouse PD-1, respectively (Figure S2). These results indicated the need for a more detailed understanding of the h1340.CC epitope to allow structure-guided removal of the non-canonical C35a-C50 disulfide without compromising the cross-species PD-1 binding affinity, prompting us to solve the crystal structure of PD-1 in complex with the fragment antigen-binding region (Fab) of h1340.CC.

Structure of h1340.CC fab bound to PD-1

We solved the crystal structure of h1340.CC Fab bound to human PD-1 to a resolution of 2.3 Å (Figure). We were able to model PD-1 from residues Arg30 to Arg143, with the exception of disordered loop regions corresponding to residues 59–60, 71–72, and 85–92. As expected, we observe an extensive interaction interface between Fab h1340.CC and PD-1 that partially overlaps with the binding interface of PD-L1 and other PD-L1 blocking antibodies (Figure S3). All CDRs of h1340.CC are involved in binding to PD-1, and form three main interaction interfaces with PD-1 (Figurea).

Figure. Structure of h1340.CC fab bound to human PD-1. (a) overview structure, with interfacing residues in HC, LC and PD-1 colored dark green, dark purple and orange, respectively. The non-canonical disulfide is shown in red. (b-d) Close-up view of main interaction interfaces labeled in panel a, with key residues forming contacts shown as sticks. Polar contacts are shown as dashed yellow lines.

A first major interaction interface is formed by PD-1 residues Trp32, Lys131, and Gln133, which bind in a pocket formed by CDR L1, L2, L3, and H3 residues (Figureb). The PD-1 residues Trp32 and Gln133 form H-bonds with CDRL1 residues Asn31 and Tyr28, respectively. Additional key contacts involve PD-1 residue Lys131, which forms a hydrogen bond with the CDRL3 Tyr 92 backbone carbonyl and forms salt bridges with CDRL1 residue Asp32 and CDRL3 residue Asp95, while the backbone of PD-1 residue Ala132 forms a hydrogen bond to CDRL3 Asp95. A second interface is formed by sidechain – backbone mediated polar interactions between PD-1 residues Thr76, Glu136 and Arg139 and CDRH3 residues Ala98, Gly99, and Gly100a, as well as backbone-mediated hydrogen bonds between PD-1 residue Ile134 and CDRH3 residue Tyr100c (Figurec). Finally, a third major interaction interface is formed by a hydrogen bond between PD-1 residue Asn66 and CDRH1 residue Ser31, as well as a hydrophobic patch on PD-1 (residues Val64, Ile126, Leu128, Ala132, and Ala134) that interacts with aliphatic portions of aromatic and polar residues from CDRH1, 2, and 3 (Figured).

Comparison of the human, cyno, and mouse PD-1 sequences reveals that the epitope of h1340.CC is highly conserved, explaining the species cross-reactivity of this antibody (Figure 2). The2 key hPD-1 residues that make direct contacts with the h1340.CC Fab CDRs are all conserved in cyno PD-1, while in mouse PD-1, three of these residues differ. These observations are consistent with the similar affinities of h1340.CC for human and cyno PD-1 (K_D = 0.30 nM and 0.42 nM, respectively) but the weaker affinity for mouse PD-1 (K_D = 1.72 nM) (Figure S2). Indeed, while the conservative Gln133Lys substitution in mouse PD-1 would likely maintain the hydrogen bond to CDRL1 residue Asn31, the Tyr68Asn and Arg139Gly substitutions would result in loss of potentially up to two H-bonds to CDRH1 and CDRH3.

Figure 2. Epitope comparison for human, cyno and mouse PD-1. (a) Structure of hPD-1, with residues at the fab interface (within 4.5 Å) highlighted in orange, and residues contacting the CDRs shown as sticks and labeled. (b) Sequence alignment of residues–150 of human, cyno and mouse PD-1 (uniprot identifiers Q15116, B0LAJ3 and Q02242, respectively), with key interacting residues boxed in orange. Residues conserved across all 3 species are highlighted in green, while residues that differ in mouse are highlighted in blue. The signal sequence is shown in gray, and N- and C-terminal residues that are resolved in the structure are indicated by arrows.

Structure- and sequence-guided C35a-C50 disulfide removal

Guided by the structure, we undertook engineering efforts to try to mitigate the observed loss of affinity that resulted from removing the non-canonical C35a and C50 disulfide bond by mutation to Ser and Ala, respectively. Inspection of the disulfide bond indicates that cysteine residues C35a and C50 are located in adjacent beta strands that face the light chain, but are not involved in any direct interactions with hPD-1. However, CDRH3 residues L100e and L100h pack against the disulfide, which may in turn be important in stabilizing the CDRH3 conformation and its extensive interaction interface with hPD-1. In addition, in close proximity to the disulfide bond, CDRH1 residue W34 and CDRH2 residue Y52 form an extensive hydrophobic patch with hPD-1 residues Ile126, Leu128, Ala132 and Ile134 (Figure 3a).

Figure 3. Structural context of the h1340.CC disulfide bond suggests engineering opportunities beyond C35a and C50. The direct environment of the C35a – C50 disulfide bond facing toward (a) and away (b) from hPD-1 suggests residue engineering opportunities to recover hPD-1 affinity after disulfide removal. The light chain of h1340.CC has been omitted for clarity. Key residues (colored dark green) on the heavy chain (colored light green) of h1340.CC that pack against the disulfide (colored red) or mediate hydrophobic interactions with hPD-1 residues Ile126, Leu128, Ala132, and Ile134 are shown in sphere and stick representation. Residues that differ from human to mouse PD-1 are shown in bright pink (Y68N & Q133K).

We speculate that the loss of the disulfide could affect packing of this CDRH1-H2 region to CDRH3 residues L100e and L100h and thereby affect CDRH1, H2, or H3 conformation, resulting in allosteric modulation of hPD-1 binding. As noted, mutations C35aS and C50A result in a ~ 4-fold decrease in affinity to mPD-1, but only ~ 2.5-fold decrease in affinity to hPD-1. From our sequence-structure analysis, there is no obvious rationale for the stronger loss of binding to mPD-1 over hPD-1 upon disulfide removal, but we speculate that the potential loss of two hydrogen bonds due to the Y68N and R139G sequence difference in mPD-1 could make mPD-1 more susceptible to weakened interactions resulting from the removal of the CDR H1-H2 disulfide.

Since C35aS and C50A may not provide the ideal environment for packing of the hydrophobic L100e and L100h and could therefore affect the conformation and binding of the heavy chain (Figure 3a), we reasoned that mutations to bulkier hydrophobic side chains (Val, Ile, and Leu) at both positions may help recover affinity. Additionally, since we had not tested these yet, we also included conservative mutations to Ser or Thr for residue C50, yielding a total of5 variants (Table S2: Set). However, none of these variants yielded improved affinities compared to h1340.SA (Figure 4 and Table S3).

Figure 4. h1340.CC variants binding affinity to hPD-1 assessments by SPR. The red circle indicates clone h1340.SA.LV in set 3 with improved hPD-1 affinity compared to the initial h1340.SA variant (K_D: 0.7 nM) and within two-fold of affinity of parental clone h1340.CC (K_D: 0.1 nM). The K_D value is determined at 25°C.

Suspecting that the observed loss in affinity upon mutagenesis of C35a and C50 could be due to an indirect effect resulting in weakened hydrophobic interactions between hPD-1 and h1340 CDRH1 residue W34 and CDRH2 residue Y52, we inspected residues immediately adjacent to these. Careful analysis of the molecular context of Y33, I35, I51, and A52a suggests that mutating residue I35 and/or A52a to bulkier residues could provide an opportunity to improve the packing of the hydrophobic core that supports the beta sheets and thereby improve or stabilize the interactions of W34 and Y52 with hPD-1 (Figure 3b). Intriguingly, sequence comparison of clone rbt1340 with the closely related clone rbt1182, which also contains the same non-canonical disulfide bond and which was the second-highest affinity clone from our discovery campaign (Table S1), indicated that while residues 33 and 51 are identical in both clones, residue 35 is a methionine instead of an isoleucine and residue 52a a valine instead of an alanine in clone182. We therefore tested mutation of I35 to Leu or Met and A52a to Val, Ile or Leu in the context of h1340.VS (C35aV.C50S), the best variant from set (Table S2 ; Set 2) and in the context of the original h1340.SA variant (Table S2 ; Set 3). To our satisfaction, the h1340.SA variant with additional mutations I35L and A52aV, termed h1340.SA.LV, yielded an improved affinity for hPD-1 within two-fold of the parental clone h1340.CC (Figure 4 and Table 1).

Table 1. Functional characterization of h1340.CC and engineering variants in binding. Parameters (kon, association rate constant; koff, dissociation rate constant; KD, equilibrium constant) are determined at 37°C as mean ± S.D from duplicates.

Antibody	Human PD-1			Cyno PD-1			Mouse PD-1
	kon (105 M-1 S-1)	koff (10–4 S-1)	KD(nM)	kon (105 M-1 S-1)	koff (10–4 S-1)	KD(nM)	kon (105 M-1 S-1)	koff (10–4 S-1)	KD(nM)
h1340.CC	4.63 ± 0.45	2.06 ± 0.14	0.45 ± 0.07	3.78 ± 0.51	2.35 ± 0.22	0.63 ± 0.14	5.73 ± 0.71	10.8 ± 0.01	1.91 ± 0.23
h1340.SA.LV	3.46 ± 0.11	2.76 ± 0.08	0.80 ± 0.05	2.79 ± 0.12	4.19 ± 0.06	1.51 ± 0.08	4.44 ± 0.24	17.0 ± 0.04	3.82 ± 0.11
h1340.SA.7	3.81 ± 0.41	2.35 ± 0.12	0.62 ± 0.03	3.26 ± 0.50	3.41 ± 0.22	1.05 ± 0.09	4.63 ± 0.45	16.9 ± 0.01	3.66 ± 0.35
h1340.AA.1	4.01 ± 0.28	2.40 ± 0.14	0.60 ± 0.08	3.36 ± 0.32	3.65 ± 0.26	1.09 ± 0.18	4.96 ± 0.52	21.4 ± 0.01	4.34 ± 0.48

To further assess our engineering approach, we solved the crystal structure of h1340.SA.LV in complex with hPD-1 to a resolution of 2.5 Å and compared it to our structure of h1340.CC in complex with hPD-1 (Figure S4A). As expected, both complex structures are highly similar, with an overall root-mean-square deviation of atomic positions (RMSD) of 0.640 Å (Figure S4B). Inspection of residues 35 and 52a in both structures indicates that mutations I35L and A52aV likely provide improved packing of the hydrophobic core (Figure S4C).

Functionality and immunogenicity comparison of h1340.CC and h1340.SA.LV

To assess the impact of our engineering efforts on antibody function, we compared the affinity and activity of h1340.CC and h1340.SA.LV for human, cyno and mouse PD-1. Both variants yielded similar binding affinities to recombinant PD-1, as assessed by surface plasmon resonance (SPR) (K_D values less than 2.5-fold apart, Table 1), and similar binding affinities to cell surface-expressed PD-1, as assessed by FACS (similar EC₅₀ values, Figure 5a). Both variants performed equivalently and comparably to the marketed drugs pembrolizumab and nivolumab in blocking PD-1/PD-L1 interaction, with similar IC₅₀ values and maximum inhibition values at the highest antibody concentration (Figure 5b).

Figure 5. Functional characterization of h1340.CC & h1340.SA.LV in binding and blocking. (a) Cell surface PD-1 binding analysis by FACS. CHO-hPD-1: CHO expressed human PD-1; CHO-mPD-1: CHO expressed mouse PD-1; CHO: No PD-1 expression. The error bar in the curve is mean ± S.D of MFI from triplicates (b) titration of antibody variants, using percent of inhibition (relative to00% as Pembrolizumab at 20 µg/mL) in the PD-1-PD-L1 blockade assay. The h1340 variants performed equivalently to Pembrolizumab and Nivolumab in terms of maximum inhibition and IC50. The negative isotype control (NIST) showed no PD-1 binding and blocking as expected.

To evaluate the potential immunogenicity of both variants, we used peripheral blood mononuclear cells (PBMCs) derived from healthy donors and monitored the proliferation of co-incubated CD4 T cells by bromodeoxyuridine (BrdU) staining in FACS after 7 days of stimulation. The stimulation index (SI), a percent ratio of BrdU⁺/CD3⁺CD4⁺ T cells in the presence or absence of test antibodies, was calculated to determine the response of each donor (see Materials and methods section). Overall, both antibody variants induced T cell proliferations with similar SI values in positive donors (13–17%), indicating they both possess low immunogenicity potential (Figure S5).

Developability assessment of h1340.CC and h1340.SA.LV

We assessed the impact of the non-canonical disulfide bond on the potential developability of h1340.CC and h1340.SA.LV by measuring several physicochemical properties (Table 2). We thermally stressed variants h1340.CC and h1340.SA.LV at high concentration (150 mg/mL) to exacerbate their physical instability while also assessing their developability into a high concentration formulation. After storage at 30°C for month, we evaluated the changes in aggregation and fragmentation by size exclusion chromatography (SEC) and the changes in charge variants by ion exchange chromatography (IEC) (Table 2 Figure S6). Although the non-canonical disulfide could potentially lead to unpaired cysteines or disulfide scrambling, which could cause aggregation, no increase in aggregation was observed for h1340.CC compared to h1340.SA.LV. Furthermore, while a bigger decrease in the main peak was observed for h1340.SA.LV by IEC, the change was within the assay variability, indicating that both variants behave similarly when thermally stressed at high concentration.

Table 2. Physicochemical property analysis of h1340.CC and h1340.SA.LV. Property changes after storage at 30°C for month are measured by size exclusion chromatography (SEC) and ion exchange chromatography (IEC). HMWF: high molecular weight forms. LMWF : low molecular weight forms.

Antibody	SEC			IEC			Viscosity [cP]	Turbidity [OD]
	Change in HMWF (%)	Change in Monomer (%)	Change in LMWF (%)	Change in Acidics (%)	Change in Main Peak (%)	Change in Basics (%)
h1340.CC	0.9	−1.3	0.4	5.0	−7.0	2.0	13.0	0.22
h1340.SA.LV	0.8	−1.2	0.4	6.7	−7.7	1.0	11.0	0.22

When we chemically stressed both variants using 2,2’-azobis(2-amidinopropane) dihydrochloride (AAPH), we observed higher oxidation at W34 in h1340.SA.LV compared to h1340.CC (36% vs.1%, respectively), suggesting increased solvent accessibility of this region. In line with our structural analysis, the close proximity of this residue to the non-canonical disulfide bond suggests that removal of the disulfide bond destabilizes this region and increases the flexibility of the CDRH1 loop. No additional differences in oxidation were observed by peptide mapping for any asparagine, aspartic acid, methionine, or tryptophan residue in the CDRs of either variant after thermal or oxidative stress (data not shown).

Besides thermal and chemical stability, the viscosity of the antibody is a key characteristic that can affect the manufacturing process.¹¹ We therefore measured the viscosity at high concentration in high-ionic strength arginine succinate buffer for both variants. While the viscosity of h1340.CC was slightly higher than h1340.SA.LV (Table 2), the observed difference can likely be explained by the difference in protein concentration between samples (189 mg/mL for h1340.CC vs82 mg/mL for h1340.SA.LV). In addition, both are well below the threshold of what is typically considered a non-manufacturable viscosity and therefore pose no risk to the manufacturing process.

Finally, to understand the impact on drug administration, we evaluated the solubility of both variants by dialyzing high concentration samples into phosphate-buffered saline (PBS), which serves as a surrogate for physiological conditions. The turbidity of each sample was measured post dialysis and similar to the other results, no differences between both variants were observed (Table 2).

Structure- and ML-guided C35a-C50 disulfide removal

While our conventional structure- and sequence-guided protein engineering approach yielded variant h1340.SA.LV with a potency, cross-reactivity and developability similar to h1340.CC, it only partially recovered affinity compared to the parental clone. Prompted by recent developments in ML-guided protein engineering, we wondered whether an unbiased ML- and structure-guided approach could rapidly generate an antibody variant without the C35a-C50 disulfide of h1340.CC, but with properties superior to h1340.SA.LV.

We therefore chose to leverage ProteinMPNN, a deep learning – based protein sequence design method that generates amino acid sequences based on computed or experimental structures using protein backbone features as input, to generate new designs.⁹ Using our structure of Fab h1340.CC in complex with hPD-1 as input, and keeping the sequence of hPD-1 and the h1340.CC light chain fixed, we used ProteinMPNN to generate00 sequences for the h1340 heavy chain (Table S5). Since our goal was to guide C35a-C50 disulfide removal, we first analyzed the residue co-frequencies at the 35a-50 pair (Table 3). About half of the generated sequences contained the C35a-C50 disulfide, giving us confidence that ProteinMPNN was correctly interpreting protein backbone features, thus validating our design approach. Of interest to us, 22% of the sequences contained an alanine-alanine pair at residues 35a-50, while only1% contained a serine-alanine pair, the substitution we chose in our sequence-guided approach.

Table 3. Residue co-frequencies for top 35a-50 pairs in ProteinMPNN designs of the heavy chain of h1340.H1340. The residue co-frequency for the 35a-50 residue pair was analyzed for one hundred h1340 heavy chain sequence designs generated by ProteinMPNN.

Variant name	35a − 50 residue pair	35a − 50 residue co-frequency (%)
h1340.CC	Cys-Cys	47
h1340.AA	Ala-Ala	22
h1340.SA	Ser-Ala	11
h1340.VA	Val-Ala	9
h1340.GA	Gly-Ala	6
h1340.LA	Leu-Ala	5

Focusing on the top three non-cysteine pairs (AA, SA and VA), which accounted for 42% of all sequence designs, we next analyzed all other residues in this subset of sequences and only kept mutations that occurred with a frequency equal to or higher than 50%. This approach yielded a shortlist of 58 candidate positions to consider for mutagenesis. To further narrow down this list, we then inspected each position in the context of our h1340.CC – hPD-1 structure to assess the potential effect of the suggested mutations on hPD-1 affinity and h1340 stability (Table S6). Excluding residues that were solvent-facing, distal from the antigen binding site, deemed to potentially reduce affinity or with no expected improvement, this yielded a list of 6 positions to test experimentally (Table 4).

Table 4. Top h1340 heavy chain mutations from structure- and ML-guided analysis. The frequencies for selected mutations occurring within sequences containing an SA, AA or VA substitution at the non-canonical disulfide pair are shown and color-coded from light to dark green, according to the frequency at which they occur. A structural rationale as to why these mutations may improve hPD-1 affinity and were chosen as top candidates is provided. The numbering of the position follows the kabat numbering system.

Intrigued by the high occurrence of the AA pair at positions 35a-50, we first tested the hPD-1 affinity for the h1340.AA variant (Table 5). Strikingly, the h1340.AA variant had a significantly better affinity than the h1340.SA variant, suggesting that the presence of a serine at position 35a is destabilizing the interface with hPD-1. To quantify this effect, we used the PyRosetta4 software to calculate the delta delta G (ΔΔG), a metric for predicting how a single-point mutation will affect protein stability,^12,13 for the S35aA mutation in the context of our h1340.SA.LV – hPD-1 complex structure. The computed ΔΔG for the S35aA mutation was −14 Rosetta Energy Units (REU, data not shown), with REU approximately equivalent to kcal/mol,¹⁴ a significant improvement considering that a threshold of −1 kcal/mol is often used to consider the relevance of mutations on structural stability.¹⁵

Table 5. Relative affinities for h1340.CC variants selected by structure- and ML-guided engineering. KD (variant)/KD (h1340.CC) indicates the fold change in binding affinity relative to h1340.CC. Sequence differences in variants compared to h1340.SA are highlighted in orange. The non-canonical disulfide bond is labeled in red. Single mutants with improved affinity compared to h1340.SA are highlighted in green, and the best combinatorial mutant in the SA or AA backbone is highlighted in blue and yellow, respectively. The numbering of each residue position follows the kabat numbering system. The relative KD values were determined at 25°C.

Next, and to allow for a direct comparison with the mutants we previously generated, we measured the effect of the six non-disulfide mutations we selected within the h1340.SA background on the hPD-1 affinity relative to h1340.CC (Table 5). Of the six mutants tested (h1340.SA.1–6), three mutants (h1340.SA.3, h1340.SA.5, and h1340.SA.6) showed improved affinities over h1340.SA, which can be rationalized by analyzing their structural context. Mutation S30D likely creates an additional polar interaction with PD-1 residue K78, conserved in human, mouse, and cyno (Figure 6a). Similar to the A52aV mutation we previously identified, mutation A52aL provides improved hydrophobic packing, with a computed ΔΔG of −22 kcal/mol in the h1340.CC:hPD-1 complex structure (data not shown) (Figure 6b). Mutation A95S likely creates a hydrogen bond with the backbone carbonyl of T100d that locks the CDRH3 loop conformation (Figure 6c). Combining these three mutations together in the h1340.SA background (h1340.SA.7) or in the h1340.AA background (h1340.AA.1) yielded variants with affinity fold-changes relative to the initial h1340.CC variant close to (Table 5). Detailed measurements of association and dissociation rate constants and the resulting affinities for human, cyno and mouse PD-1 indicate that both variants showed affinity differences relative to h1340.CC of less than1.5-fold for human and cyno PD-1 and less than 2.5-fold for mouse PD-1 (Table 1). Thus, our structure and ML-guided approach allowed for rapid and efficient generation of an antibody variant with improved properties.

Figure 6. Structural analysis of structure- and ML- guided top mutations. Close-up view of key structural context rationalizing the improved affinity observed for the top 3 ProteinMPNN-generated mutations. Point mutations were generated using the h1340.CC – hPD-1 structure as a template and energy-minimized using the rosetta software suite. (a) Mutation S30D creates a strong hydrogen bond with the conserved PD-1 residue K78. (b) Mutation A52aL provides improved packing of the hydrophobic core supporting key CDRH1 residue W34 and CDRH2 Y52 residues that interact with PD-1. (c) Mutation A95S creates a hydrogen bond with the carbonyl backbone of residue T100d, thereby locking the CDRH3 loop conformation.

Discussion

Non-canonical disulfide bonds are prevalent in rabbit, chicken, llama, shark, and cow antibodies.⁵ Their roles are essential in the generation of diverse antibody repertoires to expand paratope space for mediating distinct structural and functional interaction with antigens.^4,16 It is generally assumed that, in order to transform valuable mAbs derived from these species into therapeutics, the non-canonical disulfide bonds inside and outside the CDRs must be removed as part of the drug development process. In rabbit antibodies, the non-canonical C80-C170 disulfide bond (100% prevalent in our internal discovery campaigns) between the light chain domains can be readily removed by alanine mutagenesis without detrimental effects in protein expression and strain-specific animal usage (Basilea mutant and b9 allotype rabbits) for in vitro phage-displayed immunized library generation.^7,17 In contrast, the non-canonical C35a-C50 disulfide bond between CDRH1 and CDRH2 (30–100% appearance in our internal discovery campaigns) cannot be readily engineered out. As a result, antibody clones belonging to this group are often deprioritized despite otherwise favorable characteristics and properties.

Here, we report the identification of a cross-reactive anti-PD-1 rabbit antibody harboring a non-canonical C35a-C50 disulfide bond and demonstrate that combining structure- and sequence- guided protein engineering approaches allows its removal without compromising cross-species affinity and activity. Unexpectedly, physicochemical characterization and immunogenicity assessment of both variants suggest that the presence of this non-canonical disulfide bond has no detrimental effect on the developability of the molecule. Additionally, we demonstrate that when perceived liabilities are present, a simple structure- and ML-guided protein engineering approach can also be leveraged to rapidly remove these while maintaining key antibody properties such as cross-species affinity.

ML-based methods are increasingly being deployed to assist with challenging and complex tasks such as de novo antibody design or affinity maturation, which require dedicated computational efforts, large datasets, and multiple iterations. We show here that existing structure-based ML methods, in particular ProteinMPNN, can also be used to rapidly suggest a tractable number of sequence designs for a well-defined antibody engineering problem such as disulfide removal. For comparison, in our conventional structure- and sequence-based engineering approach we tested about 40 variants over 4 iterations, informed by the sequence of a closely related clone and the experimental structure of the parent antibody – antigen complex. In our unbiased structure- and ML- based approach we experimentally tested2 variants in 2 iterations, solely based on using the parent antibody – antigen complex structure as input to generate 58 candidate mutants and to narrow the selection to the top 6 candidates, yielding a superior candidate molecule using approximately half as much time and resources. While our approach was applied to a single use case and does not constitute a general solution for each antibody engineering problem, it demonstrates how ML-based methods can be a simple and effective additional tool to guide engineering efforts and suggest design ideas.

Ultimately, more studies and clinical validation will be required to understand how generalizable our findings are and whether other developability aspects are affected, but our case study does suggest that rabbit-derived antibodies with non-canonical C35a-C50 disulfide bonds should not be immediately disregarded or disfavored. This opens opportunities to explore antibody clones from rabbit discovery campaigns more broadly and in a less biased way, with the potential to find and, if needed, further engineer, variants with more desirable properties that could result in better clinical candidates.

Materials and methods

PD-1 antibody discovery and humanization

PD-1 antibodies discovered from rabbits are generated by single B cell sorting, culturing, and cloning technology, as previously described.¹⁰ Briefly, New Zealand White (NZW) rabbits purchased from Charles River Laboratories were co-immunized with recombinant human PD-1 (hPD-1) and mouse PD-1 (mPD-1) His tagged proteins (hPD-1: 8986-PD; mPD-1: 9047-PD, R&D Systems). All animal works were approved by Roche Institutional Animal Care and User Committee (Roche IACUC) and performed following the Genentech Laboratory Animal Resources (LAR) guidelines. Rabbit PBMCs were isolated and enriched for IgG⁺ B cells that were then sorted with fluorescently labeled hPD-1 and mPD-1 by flow cytometry. After 7 days of culture at 37°C, B cell supernatants containing rabbit IgGs were screened with hPD-1 and mPD-1 in a high-throughput enzyme-linked immunosorbent assay (ELISA) to identify positive clones for the subsequent flow cytometry-based screening using the same supernatants and Chinese hamster ovary (CHO) cells expressing hPD-1 or mPD-1. The total RNAs from individual FACS⁺ (hPD-1 or mPD-1 specific, or h/mPD-1 cross-reactive) B cells were isolated for molecular cloning, sequencing and recombinant IgGs expression for further characterization.

To screen for hPD-1 blocking Abs, a panel of PD-1 mAbs along with pembrolizumab (50 μg/ml) were incubated with Jurkat T cells expressing hPD-1 for 30 min at 4°C followed by incubation with Biotin-labeled hPD-L1 (BT156–050, R&D Systems) for h. After washing, the bound hPD-L1 was detected by incubating with streptavidin-phycoerythrin for 30 min followed by washing and analyzed by flow cytometry. The blocking activity (% inhibition) was calculated by subtracting the mean fluorescence intensity (MFI) of cells in the absence of Abs (maximum binding MFI) by MFI of cells in the presence of antibodies and divided by the maximum binding MFI. mPD-1 blocking Abs were screened by an ELISA-based binding assay. 384-well MaxiSorp ELISA plates (Nunc) were coated with 0.25 μg/mL mPD-1 (9047-PD, R&D Systems) in PBS overnight at 4°C. After washing three times with wash buffer (10 mM phosphate buffer, pH 7.4,50 mM NaCl, 0.05% Polysorbate 20), the plates were blocked with blocking buffer (1X PBS, pH 7.4, 0.5% bovine serum albumin (BSA),5 ppm Proclin) for h followed by washing for three times. A serially diluted PD-1 mAbs (10 μg/ml to 0.61 ng/ml) in assay diluent (1X PBS, pH7.4, 0.5% BSA, 0.05% Tween-20,0 ppm Proclin) were added to the plates and incubated for 2 h at room temperature. After washing six times, Biotin-labeled mPD-L1 (0.125 μg/mL in assay diluent) (SRP0529, Sigma Aldrich) was added and incubated for h at room temperature. The plates were washed six times followed by incubation with streptavidin-horseradish peroxidase (HRP) conjugate (RPN1231, GE Healthcare) and TMB substrate (MOSS Bio) for color development. The reaction was stopped by adding M phosphoric acid and absorbance was read at 450 nm wavelength. After subtracting the background absorbance (no streptavidin-HRP) for each well, the blocking activity (% inhibition) was calculated by subtracting the absorbance in the absence of Abs (maximum binding) by absorbance in the presence of antibodies and divided by the maximum binding absorbance.

To humanize lead antibody rabbit340 (rbt1340), the amino acid sequences in the variable region were aligned to its closest human germline gene using an informatics tool (AbGrafter, Genentech) that covered all available human immunoglobulin IGKV and IGHV repertoires in the international immunogenetics information system.¹⁸ The closest human IGKV and IGHV germline frameworks identified through the process were then served as the acceptor for grafting rbt1340 light chain and heavy chain hypervariable regions with the corresponding rabbit framework residues at the Vernier zone, respectively.¹⁹ The gene segments of the humanized antibody variable regions were synthesized, cloned into human IgG1 expression vectors, and expressed in large scale CHO cells for purification (Genewiz, Wuxi).

PD-1 antibody characterization in affinity, cell binding, blocking, and epitope binning

The binding affinities for all rabbit and h1340 variants were determined by SPR technology (Biacore™-8K for rabbit IgGs and T200 for h1340 variants, Cytiva). Briefly, Series S sensor chip Protein A was applied to capture antibody to achieve approximately00 response units (RU), followed by injection of each PD-1 recombinant protein (R&D systems catalog numbers 8986-PD, 8509-PD and 9047-PD for human, cynomolgus, and mouse PD-1, respectively) in five-fold serial dilutions from00 nM to 0.16 nM in HBS-EP buffer (100 mM 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid (HEPES) pH 7.4,50 mM NaCl, 3 mM EDTA, 0.05% (v/v) Surfactant P20) under a flow rate of 50 μl/min at 25°C or00 μl/min at 37°C. The sensorgrams were recorded, processed by reference and blank subtraction, and evaluated by a simple one-to-one Langmuir binding model (Biacore Insight and T200 evaluation software) to determine the kinetics (k_on and k_off) and equilibrium dissociation constant (K_D = k_off/k_on).

In the epitope-binning experiment, the microarray-based 96 × 96 microfluidic system (IBIS-MX96 SPRi, Carterra) was used under a classical sandwich assay format. First, each rabbit IgG (10 μg/mL in0 mM sodium acetate buffer pH 4.5) was directly spotted onto a sensor prism CMD 200 M sensor chip (CMD 200 M, XanTec Bioanalytics) using amine-coupling chemistry in the instrument of continuous flow microspotting (CFM, Carterra). Next,00 nM of human PD-1 was injected over the sensor chip for 4 min binding, followed by another 4 min injection of other antibodies (10 μg/mL in HBS-EP buffer) from cycle to cycle at 25°C. The chip surface was regenerated between cycles using0 mM Glycine pH.5 and the binding response was recorded and analyzed in Carterra microfluidics’ binning software for network plotting.

To determine the PD-1 mAbs binding to the cell surface PD-1, the rabbit IgGs (5 μg/ml) or h1340.CC variants (five-fold serial dilutions from33 nM to.7 pM) along with an isotype control antibody (NIST, Genentech) were incubated with stable transfectants of CHO cells expressing hPD-1 or mPD-1 in FACS buffer (PBS, 0.5% BSA, 2 mM EDTA) at 4°C for 30 min. After washing, the bound antibodies were detected by APC-goat anti-rabbit IgG (F0111, R&D systems;:30 dilution) for the rabbit IgGs, and APC-F(ab’)2 fragment goat anti-human IgG Fc specific (109-136-098, Jackson ImmunoResearch;:200 dilution) for the h1340.CC variants, and analyzed by the flow cytometry.

To compare the blocking activities of humanized PD-1 mAb340 (h1340.CC) and its CC-removal variant h1340.SA.LV, functional blocking assays were performed using the PD-1/PD-L1 Blockade Bioassay (J1250, Promega) following manufacturer’s instructions. Briefly, 4 × 10⁴ CHO-K1 cells expressing hPD-L1 in00 μl culture medium (RPMI-1640,0% Fetal Bovine Serum, 2 mM Glutamine) were added to each well of a white, clear flat bottom 96-well plate (Corning) and incubated overnight at 37°C in a CO₂-controlled incubator. The following day, the culture medium was removed and 5-fold serially diluted PD-1 Abs starting from 40 μg/mL (2X of the final concentration) in 60 μl culture medium were added to the cells. Pembrolizumab and nivolumab served as positive controls. After incubation for h at 37°C, 5 × 10⁴ Jurkat NFAT luciferase reporter cells expressing hPD-1 in 60 μl culture medium per well were added and incubated for additional 6 h at 37°C. After incubation, 80 μL of luciferase substrate Bio-Glo (Promega) was added to each well and incubated in the dark for 5 min at room temperature followed by luminescence measurement on a Tecan M-1000 plate reader. After subtracting the background RLU (Jurkat cell only) from each well, the blocking activity (% inhibition) was calculated by dividing the RLU of each well by the RLU of pembrolizumab at 20 μg/mL (maximum inhibition). Titration curves were generated and IC50’s were calculated using the Prism software (GraphPad).

Expression, purification, and crystallization of human PD-1 with fab h1340.CC or fab h1340.SA.LV

Human PD-1 (amino acids 25–146) was subcloned into the pAcgp67 vector (BD Biosciences) with an N-terminal His tag. A cysteine to serine mutation was introduced at position 93 of PD-1 to facilitate the expression and folding.²⁰ To generate glycan-minimized PD-1, kifunensine-treated Trichoplusiani (Tni) insect cells were co-infected with PD-1 virus together with the virus carrying the endoglycosidase-H (endoH) cDNA. At 48 h post infection, protein was harvested by centrifugation to remove cells. After filtering with 0.45 micro filter to remove precipitated materials, the clarified supernatant was loaded over Ni-NTA agarose (Invitrogen) equilibrated with buffer A (25 mM Tris, pH 7.5, 300 mM NaCl, 5 mM imidazole). The column was washed and eluted with buffer A supplemented with 25 mM imidazole and 300 mM imidazole, respectively. In order to remove N-terminal his-tag, the eluted protein was incubated with TEV protease and dialyzed against buffer B (25 mM Tris, pH 7.5,50 mM NaCl) overnight. The digested PD-1 protein was reloaded to Ni-NTA resin to remove undigested PD-1 protein and further separated by Superdex 756/600 chromatography column (Cytiva) in buffer B. The purified PD-1 protein was further de-glycosylated by incubation with Endo F2 (NEB) and polished by Superdex 200 Increase0/300 size-exclusion chromatography column (Cytiva) in buffer B.

Fabs of h1340.CC and h1340.SA.LV were obtained by LysC and Papain cleavage from the humanized antibodies. The Fab/PD-1 complex was initially formed by adding a.25 molar excess of PD-1 to each Fab and stoichiometric complex was then isolated using a Superdex 200 Increase0/300 size-exclusion chromatography column (Cytiva) equilibrated in buffer B. For crystallization, the stoichiometric complex was concentrated to0–13 mg/ml. The crystal of h1340.CC Fab/PD-1 was obtained by vapor diffusion at9°C in a solution containing% w/v Tryptone, 0.001 M Sodium azide, 0.05 M HEPES sodium pH 7.0, 20% w/v PEG 3,350. The crystal of h1340.SA.LV Fab/PD-1 grew in precipitant solution containing 20%w/v PEG 3350, 0.2 M Na2 Malon, 0.1 M BIS-TRIS prop 8.5 pH. Crystals were cryoprotected with 25% glycerol.

Data collection and structure determination

Datasets of h1340.CC Fab/PD-1 and h1340.SA.LV Fab/PD-1 were collected at the Advanced Light Source beamline 5.0.2 and Stanford Synchrotron Radiation Lightsource2–1, respectively. The collected datasets were processed with autoPROC.²¹ Crystal structures were solved by molecular replacement using Phaser.²² As for the h1340.CC Fab/PD-1 crystal structure, the PD-1 triple mutant crystal structure²³ (PDB code: 6UMU) served as a search model for PD-1. The heavy chain of constant portion of hu4D5 Fab crystal structure²⁴ (PDB code:N8Z), the heavy chain of variable portion of N5-i5 Fab structure (PDB code: 3TNN), and the light chain of VD20_5A4 Fab crystal structure²⁵ (PDB code: 6U3Z) were used as search templates for h1340.CC. The model was built and adjusted with Coot.²⁶ The final model was refined with Phenix.²⁷ The h1340.SA.LV Fab/PD-1 crystal structure was solved by molecular replacement using the h1340.CC Fab/PD-1 structure as search template. The model was built and refined the same as h1340.CC Fab/PD-1 structure. PyMOL (Version 2.4.1, Schrodinger, 2010) was used to render the crystal structures. Data collection and refinement statistics are summarized in Table S4.

Immunogenicity assay by in vitro CD4 T cell proliferation

To examine the immunogenicity potential of both humanized antibodies, a previously reported PBMC-based CD4 T cell proliferation assay was applied.²⁸ Briefly, PBMCs from 24 healthy donors were isolated and cryopreserved to cover a diverse set of HLA-II-encoding alleles. After thawing, cells were prepared using 50 ml Aim-V media (Gibco, Cat#2055091) containing 3% human serum (Sigma-Aldrich, Cat# H3667), treated with Dnase I (Stemcell Technologies, Cat#00–0762) to reduce cell clumping, and seeded on a 24-well plate (4 ×0⁶ cells/ml/well) in Lonza X-vivo15 media (Lonza, Cat# 04-418Q). The test antibodies (50 ug/ml) were then added to the designated wells containing cells at 37°C. On day 6, BrdU (BD Bioscience, Cat# 557891) was added to the cells. After an additional 24-h incubation, cells were harvested and stained with Aqua Live/Dead (Thermo Fisher, Cat# L34957), anti-BV421-CD3 (BD Bioscience, Cat# 562426), anti-PE-Cy-7-CD4 (BD Bioscience, Cat# 557852), anti-APC-CD14 (BD Bioscience, Cat# 555399), anti-APC-CD19 (BD Bioscience, Cat# 555415) and anti-BrdU (BD Bioscience, Cat# 557891) antibodies for FACS analysis in Attune (Invitrogen™) with FlowJo software (Tree Star, Inc.).The stimulation index (SI), a percent ratio of BrdU⁺/CD3⁺CD4⁺ T cells in the presence or absence of test antibodies was then calculated to determine the response of each donor. The positive criterion used a 95% tolerance interval of bevacizumab low positive control (Avastin®; a humanized anti-VEGF mAb with low immunogenicity risk; United States Prescribing Information (USPI, 2020). Donors with an SI greater or equal to 3 in this study were considered positive. Based on this threshold, 80% of the donors responded well to keyhole limpet hemocyanin (KLH) high positive control (a highly immunogenic antigen) and 50% to bococizumab medium positive control (anti-proprotein convertase subtilisin-kexin type 9 mAb; induction of immunogenicity in 48% of patients reported previously).²⁹

Oxidative stress assay

To examine the impact of the disulfide bond on chemical accessibility and stability, oxidative stress was performed. Oxidatively stressed samples were prepared by incubating mM AAPH (2,2′-Azobis (2-amidinopropane) dihydrochloride) (Cat# 0554100–6, Cayman Company) with mg of antibody in low-ionic histidine-acetate, pH 5.5, for6 h at 40°C. After6 h, the reaction was quenched with excess methionine (Met) at a ratio 20:1 (Met to AAPH). Control samples were spiked with water instead of AAPH. Control and stress samples were buffer-exchanged prior to analysis by LC-MS/MS peptide mapping.

Thermal stress, solubility, and viscosity measurements

Protein samples were buffer-exchanged into high-ionic strength arginine succinate buffer, pH 5.5 using Slide-A-Lyzer Mini dialysis devices (10 kDa molecular weight cutoff (MWCO), Thermo Fisher Scientific) and then concentrated to50 mg/mL using0 kDa MWCO Amicon-Ultra centrifugal tubes (Millipore). Samples were then thermally stressed for 4 weeks at 30°C while the control samples were stored at − 70°C until analysis by SEC, IEC, and LC-MS/MS peptide mapping. Solubility assessment was performed by dialyzing the unstressed high concentration samples into PBS, pH 7.4 over 24 hours at 37°C and measuring the turbidity at 350 nm. Viscosity measurements were performed on a TA Instruments Discovery HR 30 rheometer (TA Instruments) equipped with a 20 mm geometry with a° degree angle (Part 511,204.945). For each antibody, a 40 uL sample at80 mg/mL in high-ionic strength arginine succinate buffer, pH 5.5 was pipetted onto the sample plate. Measurements were collected at 25°C over 90 seconds (10 second intervals) using a constant shear rate of,000 s⁻¹ and the average value is reported.

LC-MS/MS peptide mapping

A 250 µg sample of humanized antibody was reduced with 20 mM dithiothreitol in 6 M guanidine hydrochloride, 360 mM Tris, and 2 mM EDTA at pH 7.0 for hour. The reduced samples were cooled to room temperature and alkylated using iodoacetic acid (final concentration of 50 mM) for5 min in the dark. Samples were then buffer-exchanged into the digestion buffer (50 mM Tris, 2 mM CaCl2, pH 7.5) for trypsin digestion h at 37°C (1:20 (w/w) enzyme to substrate ratio). The reaction was stopped at a final concentration of 2.2 mM L-Methionine and 3% formic acid. Samples were then analyzed using a Vanquish UHPLC (Thermo Fisher Scientific) coupled to a Exploris240 (Thermo Fisher Scientific) mass spectrophotometer. Separation of0 µg of digested sample was performed on an Acquity UPLC peptide CSH C18 column with.7 µm,30 Å particles (Waters) running a flow rate of 0.2 mL/min at 77°C. Mobile phase A was 0.1% formic acid in water and mobile phase B was 0.1% formic acid in acetonitrile. The separation gradient was 2 min of% mobile phase B, 5 min of–13% mobile phase B, 35 min of3–35% mobile phase B, 2 min of 35–95% mobile phase B, 2 min of 95% mobile phase B. Mass spectrometry data was collected in positive ion mode using a Top 8 data dependent scan with a resolution of 60,000 for MS scans and 30,000 for MS2 scans. An external calibration of the instrument was performed prior to sample analysis. Mass spectrometry data was processed using PMI-Byos software (Protein Metrics) using the PTM workflow. Relative quantitation was calculated using the top two charge states for both the native tryptic peptide and the modified counterpart. The percent change of asparagine (N) deamidation, aspartic acid (D) isomerization following thermal stress, and methionine (M), tryptophan (W) oxidation following oxidative stress were calculated for all residues within CDRs.³⁰

Size exclusion chromatography and Ion Exchange chromatography

In SEC, thermal stressed samples were analyzed using a Acquity H-Class UHPLC (Waters) equipped with a PDA detector (Waters). A 30 µg sample injection was separated using a Tosoh TSKgel UP-SW3000 (2 µm, 4.6 × 300 mm) column running a flow rate of 0.3 mL/min at 25°C. The mobile phase was composed of 0.2 M potassium phosphate, 0.25 M potassium chloride, pH 6.2. The UV data was collected at 280 nm and integrated using Chromeleon software (Thermo).

In IEC, thermal stressed samples were analyzed using a Acquity H-Class UHPLC (Waters) equipped with a TUV detector (Waters). A 30 µg sample injection was separated using a Thermo Fisher MabPac^R SCX-10 (10 µm, 4 × 150 mm) column running a flow rate of 0.8 mL/min at 37°C. Mobile phase A was0 mM HEPES free acid, pH 5.5 and mobile phase B was0 mM HEPES potassium salt, 35 mM potassium sulfate. The separation gradient was.0 min of 4% mobile phase B, 51 minutes of 4–80% mobile phase B.9 minutes of00% mobile phase B. The UV data were collected at 280 nm and integrated using Chromeleon software (Thermo).

Generation and analysis of h1340 heavy chain variable domain sequences by ProteinMPNN

Protein design of mutant variants was performed using ProteinMPNN.⁹ The structure of h1340.CC in complex with hPD-1 was used as input to generate sequences of the Fab heavy chain while fixing the Fab light chain and the antigen to their original sequence. One hundred sequences were designed, with a sampling temperature of 0.1 and random seeds. Frequency analysis of mutations at specific positions, as well as global frequencies with a higher than 90% and 50% mutation ratios were performed using custom-made Python scripts.

Computational mutagenesis and ddG calculations using Rosetta

Mutant variants structure modeling and energy calculations were performed using PyRosetta4 (2021.27 release).¹³ After initializing the PyRosetta Python module with standard options and random seed generation, native and mutated poses were scored using the REF2015 scoring function.³¹ Single residue mutations were performed via the mutate_residue pyrosetta toolbox function, using a0 Å side chain repacking cutoff. Mutant energy scores were calculated after side chain repacking. Calculated ∆∆G values are reported in Rosetta energy values.

Abbreviations

AI	=	artificial intelligence
AAPH	=	2,2’-azobis(2-amidinopropane) dihydrochloride
BrdU	=	bromodeoxyuridine
CC	=	C35a-C50 disulfide bond
CDR	=	complementarity-determining region
ELISA	=	enzyme-linked immunosorbent assay
Fab	=	fragment antigen-binding region
FACS	=	fluorescence-activated cell sorting
IEC	=	ion exchange chromatography
mAb	=	monoclonal antibody
MFI	=	mean fluorescence intensity
ML	=	machine learning
MPNN	=	message-passing neural network
PBMCs	=	peripheral blood mononuclear cells
PBS	=	phosphate buffered saline
PD-1	=	programmed cell death
PD-L1	=	programmed cell death ligand-1
REU	=	Rosetta Energy Units
SEC	=	size exclusion chromatography
SPR	=	surface plasmon resonance

Author contributions

WCL, HX, DS, JZ, ZY, MM, JP, MM, and YW designed the research, interpreted the data and wrote the paper. WCL, HX, DS, NS, RC, YL performed the experiments. LDA performed all computational structural biology analyses.

Data availability materials

All study data are included in the article and/or SI Appendix. The coordinates of the h1340.CC-hPD-1 and h1340.SA.LV-hPD-1 complexes have been deposited in the PDB with accession codes 8U31 and 8U32, respectively.

Acknowledgments

We thank the Antibody Production and Automation Technologies (APAT) group for antibody purification support, Genentech FACS core and Laboratory Animal Resources (LAR) group for technical assistance. We thank Jeremy Murray for help with x-ray data processing. Datasets of h1340.CC Fab/hPD-1 and h1340.SA.LV Fab/hPD-1 were collected at the Advanced Light Source beamline 5.0.2 and at the Stanford Synchrotron Radiation Lightsource beamline2-1, respectively. We thank Dan Qin and Sien Tam for immunogenicity data generation.

Disclosure statement

All authors are current or previous employees of Genentech Inc. / Roche.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2024.2309685

Additional information

Funding

All authors are current or previous employees of Genentech Inc/Roche. This study was supported by internal Genentech funds.