CHO-PD-1 and HEK293F cells were procured from the National Collection of Authenticated Cell Cultures (Shanghai, China). CHO-PD-1 cells were cultured in CD CHO medium (Gibco, New York, USA) supplemented with 0.7 mg/ml geneticin (Gibco, Paisley, UK) and 10 mM glutamine (Gibco, New York, USA). HEK293F cells were cultured in OPM-293 CD05 medium (OPM BiosciencesCo., Ltd., Shanghai, China). Human peripheral blood mononuclear cells (PBMCs) were purchased from MiaoTong Biotechnology (Shanghai, China). E.coli DH10B was a product of Invitrogen(Carlsbad, CA, USA). E.coli XL-1Blue was obtained from Agilent Technologies(USA). M13KO7 helper phage was obtained from NEB(Beijing, China).

PrimeSTAR ^® HS DNA Polymerase was from Takara(Beijing, China). T4 DNA ligase and T4 DNA polymerase were purchased from NEB(Beijing, China). Triethylamine was purchased from Sigma(Shanghai, China), and the nitrocellulose filter was purchased from Whatman(Germany). Polyethylenimine (PEI) was purchased from Polysciences(Warrington, PA, USA). Goat anti-human kappa-unlabeled antibodies were procured from Southern Biotech(Birmingham, AL, USA), and PE-labeled goat-anti-human antibodies were purchased from BioLegend(San Diego, CA, USA). Enzymes EcoR I and Nhe I, streptavidin magnetic beads, 1% casein, Immunotube, TMB Substrate Kit, neutravidin-alkaline phosphatase-conjugated, alkaline phosphatase substrate (NBT-BCIP), 5000×SYPRO Orange, and high sensitivity streptavidin-HRP were purchased from Thermo Fisher Scientific. Additional reagents included anti-M13 antibody HRP, PD-1-6xHis, PD-L1-6xHis, GM-CSF, IL-4 TNF-α, IL-1β, IL-6, IL-10 ELISA Kit, IFN-γ ELISA Kit from SinoBiological(Beijing, China). PGE2 was sourced from Absin(Shanghai, China). The EasySep™ Human CD4+ T Cell Isolation Kit was purchased from Stemcell Technologies(Canada), and the CD14 MicroBeads human lyophilized kit was obtained from Miltenyi Biotec(Bergisch Gladbach, Germany).

The amino acid sequence of murine antibody XM PD-1 was discovered in a previous study (18)and incorporated into the library. Molecular Operating Environment (MOE, version 2019.0102) was utilized for modeling and humanization.

The murine amino acid sequence was loaded into MOE, with annotations made for the regions of FRs and CDRs. The 3D structure model of the antibody was constructed based on FR and CDR templates, respectively. The best mode was selected based on sequence similarity and energy minimization. MOE identified three types of canonical structure determining residues, and the roles of residues in maintaining conformation were confirmed through 3D visualization of the structure. Subsequently, a few human FRs with the highest homology in the Fab sequence database were chosen, and the murine CDRs were grafted onto the selected human FRs. Finally, key residues were designed for back-mutation based on the previously identified canonical structure determining residues and predictions of their effects on structure maintenance and binding ability.

Regarding the CDR grafting method, the VH and VL sequences were synthesized as designed by Sangon Biotech (shanghai). Conversely, for the FR shuffling method, the VH and VL containing CDRs and shuffled FRs were identified through phage sequencing. The VH and VL regions of the two methods were fused with human IgG1 heavy and light constant domains using DNA recombinant technology. The expression plasmids of heavy chains (HC) and light chains (LC) were co-transfected into HEK293 cells at the ratio of HC: LC=1:3 using polyethylenimine (PEI). The supernatant was harvested after 5-6 days of cell culture, and the antibodies were purified using a protein A column (Smart-lifesciences) and eluted with 0.1M glycine on the AKTA Pure FPLC system (GE Healthcare).

Analytical SEC was performed using the Agilent 1260 HPLC system (Agilent), which was equipped with a Thermo MAbPac SEC-1, 5 µm, 7.8 × 300 mm column (P/N 088460, Thermo Fisher Scientific). A total of 10 µl of sample was injected into the column at a concentration of 1mg/ml. Sodium phosphate (pH 6.8) served as the mobile phase with a flow rate of 1 ml/min. Protein detection was carried out using a UV detector at OD280.

A working solution of SYPRO Orange 250× was freshly prepared by diluting 1 μl of 5,000×SYPRO Orange in 19μl of water. Subsequently, 1μl of SYPRO Orange working solution was added to 24 μl of PBS-diluted antibodies at a 20 μM. Fluorescence intensities were measured at every 1°C interval within the temperature range of 25 to 95°C using the CF×96 Touch qPCR machine (Bio-Rad). The melting temperature (Tm) was calculated from the derivative Relative Fluorescence Unit (RFU) against temperature (dRFU/dT) curve.

The 96-well plate (Greiner) was coated with 2 μg/ml PD1-Fc at 4°C overnight, then blocked with 1% casein for 1 h. After washing, three-fold series dilutions from 100nM of antibodies were added to the plates. The color was developed using goat anti human kappa-HRP and TMB substrate at room temperature. After being stopped by 2M H₂SO₄, the absorbance was measured at OD₄₅₀ using SpectraMax M5e (Molecular Devices).

Binding kinetics analysis was conducted on the Octet RED96e system (ForteBio). Protein A biosensors (Sartorius) were pre-wetted in the kinetics buffer (PBS with 0.05% Tween 20) for 10 min. The antibodies were immobilized on the protein A biosensor at a signal level of 1.5 nm. Various concentrations of PD-1-his were applied in a two-fold series dilution in the kinetics buffer (from 100 to 1.56 nM). The association step was set for 120 s, and the disassociation step was set for 180 s. The binding curves were fitted using a global fit 1:1 binding model with ForteBio Data Analysis software 9.0 (ForteBio).

In the self-binding assay, a protein A biosensor captured the antibody to be tested. To ensure specificity, a non-binding antibody blocked any remaining Fc binding sites on the biosensor. The antibody to be tested then underwent a 120s association followed by a 180s dissociation. The signal measured during the association was used to assess the self-binding capabilities of the antibody to be tested, indicating its tendency to interact with itself.

The binding kinetics of CDR grafting humanized antibodies were verified by SPR using Biacore T200 (GE Healthcare). The running buffer employed was HBS-EP, consisting of 10 mM HEPES, 150 mM NaCl, 3 mM EDTA at pH 7.4 and 0.005% (v/v) Tween-20. Purified antibodies were diluted to 2 μg/ml and immobilized on Series S Sensor Chip Protein A (GE Healthcare, Cat: 29127556) at 250 response units (RU). The gradient PD-1-ECD-his flowed over the immobilized antibodies starting from 50nM in a two-fold serial dilution, at a flow rate of 30ul/min, with 180 s for association and 1200 s for disassociation. After each cycle, the sensor chip was regenerated with 10mM glycine at pH1.5 for 30 s. The binding kinetics were analyzed using Biacore T200 Evaluation software version 3.1 with a 1:1 Langmuir binding model.

Binding of the humanized antibodies to the cell surface PD-1 was measured using Fluorescence-Activated Cell Sorting (FACS). CHO cells overexpressing PD-1 (CHO-PD-1) were aliquoted into FACS tubes at 3×10⁵ cells/tube and then incubated with 2 μg/ml PD-1 antibody at 4°C for 30 min. After washing with PBS twice, the cells were incubated in 100 μl staining buffer containing PE-labeled goat-anti human antibody in the dark for 30 min. The mean fluorescence intensity (MFI) of binding was measured by CytoFLEX flow cytometer (Beckman), and the data were analyzed using FlowJo software (BD Biosciences).

A concentration of 1 μg/ml of recombinant human PD-1-6xHis was coated on 96-well plates (Greiner) at 4°C overnight. Testing antibodies in a 2-fold series dilution were then added to the plates along with 1 μg/ml biotinylated PD-L1. Following a 1 h incubation, the bound ligand was detected using streptavidin-HRP and developed with TMB substrate. The reaction was halted by 2M H₂SO₄, and the absorbance was measured using SpectraMax M5e (Molecular Devices).

CD14⁺ and CD4⁺ cells were isolated from frozen PBMCs from different donors using the CD14 MicroBeads human lyophilized kit and EasySep™ Human CD4⁺ T Cell Isolation Kit. The isolated CD14⁺ cells were cultured with 500 ng/ml GM-CSF and 500 ng/ml IL-4 for 5 days to generate immature DCs. Subsequently, they were treated with 30 ng/ml TNF-α, 300 ng/ml IL-1β, 300 ng/ml IL-6, and 3 μg/ml PGE2 for an additional two days to induce DC maturation. On day 8, CD4⁺ T cells (1×10⁵ cells) were co-cultured with allogeneic DCs (1×10⁴ cells) in the presence of T5, Pembrolizumab, Nivolumab, and isotype. After 5 days of culture, the supernatant was collected to measure the production of IL-10 and INF-γ using ELISA.

BioPhi was a platform designed for antibody design, humanization, and humanness evaluation, which leveraged natural antibody repertoires and deep learning technologies (21). It incorporated a component called OASis, which provided a detailed, interpretable humanness score. The sequences of the light chain and heavy chain were inputted on this website https://biophi.dichlab.org, and the OASis score was outputted.

T20 humanness analyzer was another classic prediction tool (22). The sequences of the light and heavy chains were inputted on this website http://abAnalyzer.lakepharma.com, and the T20 score was outputted.

The structures of the murine antibody was obtained by homology modeling by MOE. The antibodies were annotated according to Kabat. The antibody Fv was solvated in a cubic water box that is sufficiently large to provide a minimum buffer zone of 14 Å between biological material and the cubic system boundaries. Na⁺ and Cl^- ions were randomly placed to neutralize the system electrostatically at a physiological salt concentration of 0.150 M. The CHARMM36m force field (23) and the three-site OPC water model were used subject to periodic boundary conditions (24). The initial energy minimization was performed using the steepest descent method. Later, the system was equilibrated in the NVT ensemble at 310 K for 100 ps, using a small integration time step of 2 fs. Production trajectories were recorded in the NPT ensemble at 310 K and 1 atm atmospheric pressure using a 2 fs of integration time steps for a total of 100 ns. Atomic coordinates were saved every 10 ps. The long simulation utilized similar parameter settings, with a total duration of 1 μs. Gromacs 2024.1 was used for simulation setups and trajectory collection. Gromacs and in-house Python scripts were used for all analyses.

We developed a comprehensive method to analyze protein-protein interactions within a structure, utilizing the Bio.PDB module from the BioPython library. Our method focuses on identifying four primary types of molecular interactions critical to protein function and stability:

Hydrogen Bonds: Analyzed between nitrogen (N) and oxygen (O) atoms within a 3.5 Å threshold.

Hydrophobic Contacts: Evaluated between side-chain carbons of hydrophobic residues (Leucine, Isoleucine, Valine, Phenylalanine, Tryptophan, Methionine) excluding alpha carbons, with a distance limit of 5.0 Å.

Salt Bridges: Assessed between oppositely charged residues (Arginine, Lysine, Histidine; Aspartic acid, Glutamic acid) within a 4.0 Å range.

Pi-Pi Stacking: Investigated between aromatic rings (Phenylalanine, Tyrosine, Tryptophan, Histidine) up to 6.5 Å apart.

Statistical evaluation was conducted using GraphPad Prism 8.0 (GraphPad Software Inc.). P values were calculated using a one-way ANOVA multiple comparison test (*P < 0.05, **P < 0.01, ***P < 0.001, ns: not significant).

The design of CDR-grafting humanization was conducted using software Molecular Operating Environment (MOE) (25). The annotation of antibodies was performed using the IMGT(The International ImMunoGeneTics information system). The murine antibody structure model was built based on the FR and CDR templates retrieved from the PDB database. The templates and corresponding similarity scores were shown in Supplementary Table 1 . Three types of canonical structure determining residues were identified through the predicted 3D structure ( Supplementary Tables 2 , 3 ). Type 1 residues, located at the binding interface between VL and VH, play key roles in the packing of the two structural domains. Type 2 residues, positioned close to the CDR region and embedded in the protein interior, may affect the overall antigen-antibody binding conformation. Type 3 residues have direct interactions with the CDR region, including hydrophobic interactions, hydrogen bonds, and salt bridges. The robustness of our MOE template modeling has been validated through comparative analyses with models generated by ABodyBuilder (26) and ImmuneBuilder (27). The results, as shown in Supplementary Figure 1 , indicate that the overall structural frameworks were consistent across the models, with RMSD values of 0.48 Å between MOE and ABodyBuilder, and 0.393 Å between MOE and ImmuneBuilder. This consistency underscores the reliability of the MOE model, especially within the framework regions (FR).The RMSDs calculated separately for the CDR loops and the remaining Fv regions were shown in Supplementary Table 4 ). Static modeling provides a single, often idealized snapshot of the molecular structure, without accounting for the natural fluctuations and movements that molecules undergo in their native environments. Molecular Dynamics (MD) simulation captures the time-dependent behavior of molecules and incorporates environmental effects, enabling researchers to observe how structures and interactions evolve over time. Supplementary Figure 2A displayed the Root Mean Square Deviation (RMSD) values for XM Ch PD1 in MD simulations (100 ns), showing fluctuations between 6.5 and 12.5 ns, yet not exceeding 0.23 nm. After approximately 15 ns of MD simulation, the system began to stabilize, ultimately settling at 0.09-0.17 nm. This suggested that the simulated structure closely matched the predicted structure from homology modeling. Additionally, long simulations (1 µs) were employed to gain deeper insights into the dynamic behavior of conformation-stabilizing contacts and to potentially correct minor inaccuracies in the modeled structures. The RMSD values, ranging from 0.069 to 0.258 nm ( Supplementary Figure 2B ), indicated that the antibody maintained a relatively stable structure with some conformational flexibility over the 1 μs MD simulation. This further indicated that the simulation conditions were appropriately set, and the initial model was reliable, without resulting in significant structural changes.

The conventional CDR-grafting method employs a single human germline sequence for grafting parent CDRs, often failing to satisfy all key requirements: low immunogenicity, high stability, and high expression. To address this, each non-human FR was individually humanized by selecting the most homologous human germline FR, allowing the preservation of the original antibody’s properties to the greatest extent. The chosen germlines of CDR grafting were shown in Table 1 . After grafting the CDRs onto human FRs, the designed molecule was named ‘human-germline-grafted’. By comparing the designed one and the murine antibody XM Ch PD-1, the residues different in the FR region were identified. Three types of canonical structure determining residues listed by the MOE were only for reference, we determined the key residues by comprehensive considerations. To determine which key residues need back-mutation to retain the binding ability, their roles in conformation maintenance were checked using the constructed parental structure model. The sites involved in the mutations include K19, T40, E42, R44, P61, A63, V70, E76, E77, I78, K87, and S88 in the heavy chain and the sites involved mutations include M37, H38, and P47 in the light chain( Supplementary Figure 3 ). Finally, eight humanized heavy chains and four humanized light chains were designed. We generated 32 humanized antibodies. Through single point ELISA screening, four candidates with the highest binding, named BS#3, BS#5, BS#11, BS#31 were selected for subsequent studies.

FR-shuffling humanization was executed by selecting different human heavy and light (κ) germline genes into each of the four FRs while keeping the CDRs fixed (16). The full-length variable structural domains were generated by fusing CDRs with the FR pool. The gene diversity of the library exceeded 4×10⁶, which was ample for selecting the optimal matches between CDRs and FRs. After multiple rounds of solution panning and immunotube panning, the capture lift assay was developed for the initial screening of the large library (19, 20). Only deep colored spots were isolated as positive clones. Single point ELISA was employed for the secondary screening. The schematic diagrams of the FR shuffling and CDR grafting methods are shown in Figure 1 . We selected the top nine candidate molecules with the highest OD values from a single-point ELISA binding assay, representing the molecules with the best binding affinity, named T1-T9.

The humanized heavy and light chains were completely shuffled as they derived from various human germline families. For example, T1 Ab contained VH5/VH3 and VK6/VK3, respectively ( Table 2 ). With a more nuanced insight, three FRs of individual heavy and light chains may derive from different human germline genes with in a given family, such as T5 Ab compromised VH3-15 and VH3-53 in the heavy chain. Overall, the humanized PD-1 antibodies’ heavy chain showed a preference for the VH3 germline family. This inclination could be attributed to the greater homology of this class with the parental FR, thereby increasing the likelihood of preserving essential parental residues (16). Besides, various germline families were exhibited in the humanized light chain.

The top mammalian expression titers of selected CDR grafting antibodies (BS# series) were approximately 100 mg/L, and most of the FR shuffling-humanized antibodies (T series) yielded titers higher than 100 mg/L, except T6 and T8 ( Table 3 ). Size exclusion chromatography (SEC) were used to assess the purities of antibodies after a single-step purification with Protein A columns. Peaks that elute faster in the chromatogram represented larger aggregates. All the CDR grafting-humanized antibodies aggregated to an extent of about 20% ( Figure 2A ). In comparison, all nine FR shuffling-humanized antibodies exhibited minimal aggregation with >90% purity as revealed by analytical SEC ( Figure 2B ). Thus, T1-5, T7 and T9 were selected for further studies. For subsequent head-to-head comparison, we used SEC for secondary purification to remove aggregates and improved the antibody purities to nearly 100%. Once the aggregates were removed, the CDR grafting-humanized remained stable in the PBS buffer.

After being heated at 60°C for one hour, the purity of the humanized antibodies was analyzed by SEC to determine their thermal stabilities. Protein peaks that elute after the main peak were typically degraded smaller fragments, which traveled through the gel matrix slower than the intact antibody molecules. The four candidates obtained by CDR grafting were not thermally stable, and degraded more when compared to the chimeric antibody XM Ch PD-1 ( Figure 3A ). In contrast, the nine candidates obtained by FR shuffling were thermally stable after being heated ( Figure 3B ).

Additionally, Differential Scanning Fluorimetry (DSF) was used to measure the thermal stability of the antibodies. Using extrinsic fluorescent probes and real-time PCR apparatus, DSF can monitor unfolding transition of specific proteins with increasing temperature (28). As shown in Figure 3C , all the Tm values of the antibodies’ Fc domains were at 68.2°C, consistent with the chimeric antibody. The melting temperatures were shown in Supplementary Table 5 . The Fab domains of T1, T3, T5 and T6-9 exhibited enhanced thermal stability compared with its murine counterpart XM Ch PD-1, and the Tm values of the antibodies were all above 73.9°C. Thus, the selected FRs and CDRs are compatible to each other. In contrast, the Tm values of the Fabs of the four CDR grafting-generated variants were under 68.2°C ( Figure 3C ). Thus, they are inferior to FR shuffling-generated variants generated in thermal stability.

The binding affinities of humanized antibodies were compared by ELISA. The two CDR grafting-humanized variant BS#5 and BS#11, and the nine FR shuffling-humanized variants all showed high affinities similar to that of the chimeric antibody ( Figures 4A, B ).

The binding kinetics was evaluated by biolayer interferometry (BLI), which monitors the rates of association and disassociation in real-time with precision and accuracy (29). The nine FR shuffling-humanized variants showed high binding affinities in BLI ( Figure 4C ), with T5 as the highest affinity antibody. In contrast, CDR grafting-humanized variant BS#3, BS#5, BS#11 and BS#31 exhibited much lower binding affinities than XM Ch PD-1 and the positive control pembrolizumab. Surface plasmon resonance (SPR) is another reliable label-free detection method for biomolecular interactions (30), which showed similar results to BLI ( Figures 4D, E ).

Next, we compared the binding activities of the variants after being heated, which demonstrated that FR shuffling-humanized variants were superior to CDR grafting-humanized ones ( Figure 5A ).

Based on the results above, we focused on developing molecules of FR shuffling. The nine FR-shuffling variants were assayed for binding abilities to PD-1 overexpressed on CHO cells. They showed comparable binding activities to XM Ch PD-1 and pembrolizumab, with T5 as the highest-binding variant ( Figure 5B ).

The effects of the FR shuffling-humanized variants on the interaction between PD-1 and PD-L1, were measured (31). All nine variants could effectively block the binding between PD-1 and PD-L1. The IC₅₀ values of the variants ranged from 0.45 to 0.84 nM, slightly better than pembrolizumab. Among them, T5 was the best variants with IC₅₀ values of 0.45 nM ( Figure 6A ). In addition, the functional activity of FR-shuffling variant T5 was analyzed by MLR assay (32). As shown in Figure 6B , T5 significantly stimulated IL-10 and INF-γ production at both 1 and 10 μg/ml, which was comparable to positive control pembrolizumab and nivolumab.

We used two different methods to predict the humanness of the humanized antibodies, BioPhi and T20 humanness analyzer, as shown in Table 4 . The results showed that the variants from both humanization strategies were similar in degrees of humanness and all had improved humanness compared to the murine counterpart. FR-shuffling variant T5 was one of the variants with the highest humanness.

Analyzing the potential T cell epitopes has been suggested to provide valuable forecasts for immunogenicity estimation. Thus, netMHCIIpan (4.0) was implemented to predict the potential peptides binding to HLA II. Due to the extensive polymorphism of HLA II in the general population, which presents a formidable obstacle in T cell epitope identification, we chose the most frequent alleles in loci DRB1 to represent the coverage of HLA II in the population (33, 34). As shown in Table 5 , T5 demonstrated a reduction in affinity for HLA II alleles when compared to its murine counterpart, suggesting a decrease in potential immunogenicity.