Efficient discovery of an agonistic anti-OX40 nanobody by epitope-directed approach to address enrichment-driven epitope bias

Abstract

Background

Conventional antibody discovery approaches that do not account for enrichment-driven biases, such as epitope immunogenicity, PCR amplification bias, or protein expression efficiency, may result in under-representation of rare yet functionally relevant clones, necessitating labor-intensive in vitro screening to identify agonistic antibodies among a large number of dominant clones. Thus, efficient screening methods for agonistic antibodies are urgently needed. OX40 is a promising target for cancer immunotherapy due to its role in enhancing T-cell activation and survival. However, effective anti-OX40 agonistic antibodies have not yet been developed.

Methods

We developed a novel screening strategy that involves the selection of nanobody clone pools enriched by biopanning against gp34-engaged and non-engaged OX40-expressing cells, next-generation sequencing, and computational clustering and subtraction analysis to identify clones recognizing the ligand-receptor interface. Representative nanobody clones underwent in vitro validation, including epitope mapping, binding affinity measurements, and functional assessments. Furthermore, we engineered the selected nanobody to enhance its in vivo efficacy. We also performed structural analysis of the nanobody-OX40 complex.

Results

Our epitope-directed approach efficiently identified nanobody clones recognizing functionally relevant epitopes distinct from dominant immunogenic regions. Notably, clone Nb479 demonstrated robust agonistic activity, closely mimicking the natural ligand gp34 with extensive OX40-binding interactions. Trimerization of Nb479 facilitated potent OX40 activation without the need for a cross-linking scaffold. Conjugation of the Nb479 trimer with an anti-serum albumin nanobody exhibited significantly improved pharmacokinetics in vivo and enhanced antitumor activity in a mouse model treated with CD19 chimeric antigen receptor T cells.

Conclusion

This study presents an innovative epitope-directed approach that greatly accelerates the discovery of functionally potent agonistic nanobodies by effectively circumventing enrichment-driven epitope bias. Our approach and engineered multivalent anti-OX40 nanobody offer a powerful platform to advance immunotherapeutic strategies for cancer treatment.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Conventional antibody discovery approaches that do not account for enrichment-driven biases may result in under-representation of rare yet functionally relevant clones, necessitating labor-intensive screening to isolate the rare agonistic antibodies. OX40 represents an attractive target in cancer immunotherapy; however, no clinically effective antibodies have yet been developed.

WHAT THIS STUDY ADDS

• This study demonstrated a novel antibody screening strategy that efficiently identifies clones recognizing the ligand-receptor interface by leveraging next-generation sequencing and nanobody display against ligand-engaged targets. As a proof of concept, we developed a novel anti-OX40 nanobody that exhibits robust agonistic activity and enhances antitumor efficacy in a mouse model treated with CD19 chimeric antigen receptor-T cells.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Our novel epitope-directed antibody screening approach can efficiently identify clones that recognize functional epitopes with structural features similar to the native ligand. Our newly developed anti-OX40 agonistic antibody has the potential to serve as an effective immunotherapy for tumors. Taken together, our novel clone screening strategy may accelerate the development of next-generation immunotherapies for cancer.

Background

Tumor necrosis factor (TNF) receptor superfamily member 4 (TNFRSF4, also known as CD134 and OX40), a T-cell costimulatory molecule, is a promising target for cancer immunotherapy due to its capacity to enhance T-cell effector function and memory cell formation and inhibit regulatory T cell-mediated suppression, thereby boosting antitumor immunity.^1–3 Although anti-OX40 agonistic antibodies have shown potential in preclinical studies,^3–7 their clinical efficacy has been limited, highlighting the urgent need for new-generation OX40 agonists.^8–14

For OX40 signal activation, the role of (1) receptor oligomerization and geometry (eg, trimerization) and (2) higher-order supramolecular assembly (eg, immune synapse formation involving CD3) have been extensively characterized.^15–17 In contrast, the contribution of intramolecular conformational changes remains poorly understood. Notably, the observation that different antibody clones targeting distinct epitopes exhibit markedly different agonistic activities led us to hypothesize that epitope-dependent, antibody-induced intramolecular conformational changes contribute to OX40 signal activation, and that antibodies targeting the same epitope as the native ligand can induce similar conformational changes and thereby promote downstream signaling.^18–20

One major challenge in developing agonistic antibodies is that conventional antibody discovery approaches may result in under-representation of rare yet functionally relevant clones due to enrichment-driven biases, such as epitope immunogenicity, PCR amplification bias, or protein expression efficiency.^21–27 Identifying clones with agonistic activity therefore requires labor-intensive screening. To address this, we developed a novel strategy using nanobody library display against ligand-engaged targets, combined with next-generation sequencing (NGS), to identify clones recognizing the ligand-receptor interface. Further leveraging the structural simplicity of nanobodies, we developed an in vivo-active OX40 agonist with enhanced antitumor efficacy.

Methods

Detailed methods are described in the online supplemental methods.

Immunization and sample collection

2 mg of human IgG1 Fc-tagged human OX40 (hOX40) extracellular domain were subcutaneously injected into an alpaca three times at 2-week intervals; paraformaldehyde (PFA)-fixed ATL-43T cells, an adult T-cell leukemia-derived cell line expressing OX40, were injected only during the first immunization. Peripheral blood mononuclear cells were isolated, and total RNA was extracted using a Direct-Zol RNA MiniPrep Kit (Zymo Research).

Library generation

Complementary DNA was synthesized from 1 mg of total RNA. The heavy-chain variable domain coding regions were amplified by PCR and purified by agarose gel electrophoresis, followed by nested PCR to amplify the nanobody-coding genes. The amplicons were subcloned into the pMES4 phagemid vector and transformed into TG1 competent cells (Agilent Technologies). Colonies from 8 mL cultured TG1 cells were harvested, pooled, and stored in 50% glycerol stocks as the mother nanobody library.

Biopanning

Growing TG1 cells (10 mL of culture) were infected with 400 µL of VCSM13 helper phage (Agilent Technologies) at 37°C for 30 min. The cells were harvested by centrifugation and cultured in 2×TY medium containing antibiotics (100 µg/mL ampicillin and 25 µg/mL kanamycin) overnight at 37°C. Phages were precipitated with 4% PEG6000 and 0.5 M NaCl. Dissolved phages in phosphate-buffered saline (PBS) containing 0.05% Tween 20, 0.5 mM NaCl, and 0.3% bovine serum albumin were absorbed into PFA-fixed control wild-type cells for 10 min at room temperature. Unbound phages were collected and incubated with PFA-fixed, hOX40-expressing cells for 30 min at room temperature. The phages bound to the cells were eluted with trypsin-EDTA solution at room temperature for 30 min with gentle agitation and used to infect TG1 cells. The infected cells were cultured in LB Miller broth containing 100 µg/mL ampicillin at 37°C overnight. The phagemids were collected using a QIAprep Miniprep Kit (Qiagen).

Amplicon sequencing

The nanobody-coding regions within the mother library and target-panned sublibraries were PCR-amplified and purified using AMPure XP beads (Beckman Coulter). Dual-indexed libraries were prepared and sequenced using an Illumina MiSeq (Illumina). The unique amino acid sequences in each library were counted using a custom Python script combining SeqKit V.0.10.1²⁸ and USEARCH V.1143.²⁹ The p values of the χ² tests for changes of each clone in existing ratios in the target sublibraries with respect to the mother libraries were calculated. Clones enriched by biopanning against any of the OX40-expressing baits (p values<0.2) were subjected to clustering analysis. Amino acid sequences with >93% identity were clustered using a variant of the UCLUST algorithm.²⁹

Nanobody and OX40 expression and purification

ClearColi BL21 (DE3) competent cells (LGC Biosearch Technologies) transformed with pMES4 plasmids encoding monomeric nanobodies were plated on LB agar containing ampicillin. A single colony was resuspended in 10 mL of LB-Miller medium containing 100 µg/mL ampicillin and cultured with shaking at 37°C overnight. The culture was transferred to 250 mL of LB medium, cultured at 37°C until the OD600 reached 0.6, and then further cultured for 4 hours at 37°C with 1 mM isopropyl-β-D-thiogalactopyranoside. The bacterial pellet was collected and resuspended in TES buffer (200 mM Tris, 0.65 mM EDTA, and 500 mM sucrose; the pH was adjusted to 8.0 with HCl) with protease inhibitors (Roche) and mixed on a shaker for 1 hour at 4°C. The mixtures were further incubated with 2×volume of fourfold diluted TES buffer with protease inhibitors for 45 min at 4°C. The resulting solutions were centrifuged, and the supernatant was recovered as a periplasmic extract. His-tagged nanobodies were purified using a nickel column, eluted with imidazole, and dialyzed against PBS.

pcDNA3.1(+) plasmids encoding multimeric nanobodies or OX40 extracellular domain were transfected into 0.75×10⁸ Expi293F cells using a Gxpress 293 transfection kit (Gmep) and incubated in 25 mL of HE400AZ culture media on an orbital shaker rotating at 125 rpm at 37°C in a humidified 8% CO₂ incubator. 2 or 3 days after transfection, His-tagged multimeric nanobodies in the supernatant were purified, as described above. FLAG-tagged recombinant OX40 proteins were purified using anti-FLAG M2 affinity gel chromatography (Sigma Aldrich), followed by size exclusion chromatography.

Staining with nanobodies and epitope competition assay

Cells were incubated with nanobodies (4 µg/mL) for 30 min on ice. After washing, the cells were stained with PE anti-6×His tag antibody (Abcam) for 20 min on ice. To assess the competitivity, hOX40-expressing Jurkat cells (Jurkat-hOX40) were pretreated with recombinant gp34 (Abcam) or mouse-derived anti-hOX40 agonistic antibody (Clone 315; 315Ab)³⁰ for 30 min on ice, and incubated with each nanobody (4 µg/mL) for 30 min on ice. After washing, the cells were stained with PE anti-6×His tag antibody for 20 min on ice. The cells were then washed and analyzed using a flow cytometer (BD FACSLyric).

Cell proliferation assay

To assess the agonistic activities of the immobilized nanobodies, microwells of a non-treated 96-well culture plate (IWAKI) were coated with PBS containing anti-CD3 antibody and the indicated antibodies, recombinant gp34, or nanobodies on the day before the experiment. The carboxyfluorescein succinimidyl ester (CFSE; CellTrace, Thermo Fisher)-labeled T cells (0.5×10⁵ cells) were cultured in the antibody and/or nanobody-coated microwells for 96 hours, and the dividing cells were assessed by CFSE dye dilution and counted using Flow-Count Fluorospheres (Beckman Coulter) on flow cytometry (BD FACSLyric). To assess the agonistic activities of nanobodies in soluble form, 1 µg/mL of nanobodies or 2.5 µg/mL of antibodies were directly added to the culture media.

Cytokine measurement of culture supernatant

The concentrations of interferon-γ (IFN-γ) and granulocyte-macrophage colony-stimulating factor (GM-CSF) in the cell-free supernatants were measured using an ELISA (BioLegend).

Luciferase reporter assay for OX40 signaling activation by nanobodies

Jurkat-hOX40 cells (1×10⁶ cells) were transfected with 1.6 µg of NF-κ-Luciferase (Firefly) reporter vector and 0.32 µg of EF1α-Luciferase (Renilla) reporter vector using TransIT-Jurkat transfection reagent (Mirus) and incubated for 24 hours. Each microwell of the non-treated 96-well assay plate was coated with IgG2a isotype control antibody, 315Ab, recombinant gp34, or nanobodies the day before the assay. The transfected cells (1×10⁵ cells) were cultured in each microwell for 24 hours and harvested to measure luciferase activity using a Dual-Luciferase Reporter Assay System (Promega) and ARVO X3 plate reader (PerkinElmer). When nanobodies were tested in soluble form, IgG2a isotype control antibody, 315Ab, recombinant gp34, and nanobodies were directly added to the culture media.

Epitope mapping

Human OX40 wild-type or human-mouse chimeras were immobilized in microwells of a 96-well MaxiSorp plate (Invitrogen). After blocking with PBS containing 10% fetal bovine serum, goat anti-mouse OX40 (mOX40) polyclonal IgG (0.5 µg/mL; R&D Systems), human Fc-conjugated recombinant gp34 (0.4 µg/mL), 315Ab (0.5 µg/mL), or nanobodies (10 µg/mL) were allowed to bind for 2 hours at 4°C. After washing, binding was detected using horseradish peroxidase-conjugated secondary antibodies at 4°C for 30 min, followed by incubation with tetramethylbenzidine substrate (Thermo Fisher).

Nuclear magnetic resonance structure determination

All nuclear magnetic resonance (NMR) experiments were conducted using a Bruker AVANCE III HD 600 MHz spectrometer equipped with a triple-resonance cryogenic inverse probe (Bruker). To improve solubility, a cysteine-to-alanine mutant at residues 50 and 110 of Nb479, referred to as Nb479C(50,110)A, was used. NMR spectra were acquired at 298 K for 470 µM [¹³C, ¹⁵N]-Nb479C(50,110)A in 20 mM sodium phosphate buffer (pH 6.0) with 5% D₂O. The NMR data were processed using NMRPipe/NMRDraw and qMDD.^{31 32} Spectral analysis was conducted semi-automatically using MagRO-NMRView,^33–36 and automated signal assignment and structural calculations were performed using FLYA.³⁷ TALOS+was used to predict backbone dihedral angles.³⁸ Automated NOE assignments and structural calculations were performed using CYANA.³⁹ The final refinement was performed using AMBER12 with the generalized Born solvent model (GB model) and ff99SB force field to obtain the 20 lowest energy structures.⁴⁰

To build the structural model of Nb479, the NOE distance restraints from Nb479C(50,110)A, along with distance restraints for the disulfide bond between cysteine residues 50 and 110, were applied to the CYANA calculations. NOE distance restraints concerning the alanine residues 50 and 110 of Nb479C(50,110)A were omitted. The lowest energy structures (20 models) were obtained by the final refinement using AMBER12 with the GB model and ff12SB force field. The NMR assignments have been deposited in BioMagResBank (ID 36708), and the coordinates are available in the Protein Data Bank (accession code 9KFW).

NMR-binding assay

A 2D ¹H-¹⁵N HSQC spectrum was acquired for 70 µM [¹⁵N]-Nb479C(50,110)A in 20 mM sodium phosphate buffer (pH 6.8), 50 mM NaCl, and 5% D₂O at 298 K with or without 29 µM non-labeled hOX40. The spectra were processed using TopSpin 3.6.4 (Bruker). The analysis was performed using Sparky.⁴¹ Combined chemical shift perturbation values for the H^N and N atoms were defined and calculated using the following formula:

$${\displaystyle {\displaystyle \sqrt{{\left(\mathrm{\Delta }{\sigma }_{HN}\right)}^2+{\left(0.2\times \mathrm{\Delta }{\sigma }_N\right)}^2}}}$$

where $∆{\sigma }_{HN}$ and $∆{\sigma }_N$ represent the chemical shift perturbations for the H^N and N resonances, respectively.

Co-culture of CAR-T cells with FL218 cells

Chimeric antigen receptor (CAR)-T cells were cultured with 50 Gy gamma-ray irradiated FL218 cells (0.2×10⁵ cells) in microwells of a 96-well round-bottom plate supplied with nanobodies (1 µg/mL) in soluble form for 2 days, and then Luciferase and ZsGreen stably expressing FL218 (FL218-LucZsG) cells (0.2×10⁵ cells) were added. 2 days later, the remaining FL218-LucZsG cells and CAR-T cells were stained and counted using Flow-Count Fluorospheres on a flow cytometer (BD FACSCanto II).

Mouse experiments

FL218-LucZsG cells (1×10⁶ cells) were implanted intravenously into 7–8 weeks old female NOD.Cg-Prkdc^scidIl2rg^tm1Sug/ShiJic mice (In-Vivo Science). 5 days later, the mice received CAR-T cells (1×10⁶ cells) intravenously, with an intravenous injection of each nanobody (100 µg) every 3 days. Tumor distribution was monitored by capturing bioluminescence images (IVIS Lumina LT Series III, Perkin Elmer) 15 min after intraperitoneal injection of 3 mg VivoGlo luciferin (Promega).

For analysis of tumor-infiltrating CAR-T cells, FL218-LucZsG cells (5×10⁶ cells) were inoculated subcutaneously. 3 weeks later, CAR-T cells (1×10⁶ cells) were administered intravenously, followed by intravenous injection of each nanobody every 3 days. 12 days later, the mice were euthanized, and subcutaneous tumors were collected. The tumor tissues were cut into small pieces using scissors and filtrated through a 100 µm cell strainer. The cell suspension was then stained for flow cytometry analysis.

Results

Establishment of nanobody clone sublibraries using biopanning

We aimed to identify agonistic nanobody clones using biopanning and NGS screening, guided by the hypothesis that targeting the receptor-ligand interaction site may facilitate the identification of native ligand-like agonistic clones. The workflow is illustrated in figure 1A.

Figure 1. Estimation of binders to the ligand-binding domain of OX40. (A) Schematic of the methodology for estimating the binders of the OX40 receptor-ligand interaction site, validation of epitope and agonistic function, and engineering for in vivo application. (B) Biopanning of the phage-displayed nanobody library was performed against OX40-expressing cell lines, with wild-type cells as references. (C) Amino acid sequences were aligned using ClustalW and phylogeny was inferred using the maximum likelihood method. The resulting tree was visualized using the clone with the highest read count as the root. (D) Clones enriched by biopanning against Jurkat-hOX40 or ATL-43T cells with Jurkat-WT as the control, and clones enriched by biopanning against HEK293T-hOX40 with HEK293T-WT as the control are aligned based on cluster. Clones (upper panel) or clusters (lower panel) are colored according to the p values on χ² tests of biopanning against Jurkat-hOX40, HEK293T-hOX40, and ATL-43T cells, in order from top to bottom. Diluted clones or clusters are colored gray, and those with no reads in both control and hOX40-expressing cell-biopanning are colored black. (E) Clones (upper panel) or clusters (lower panel) are colored based on the p values of χ² tests between Jurkat-hOX40 versus Jurkat-WT, Jurkat-hOX40-315Ab versus Jurkat-hOX40, and Jurkat-hOX40-gp34 versus Jurkat-hOX40. (F) Jurkat-hOX40 stained with each nanobody clone. The vertical axis shows the mean fluorescence intensity, and each bar is colored according to the p values of biopanning against Jurkat-hOX40. (G) ATL-43T cells stained with each nanobody clone. Each bar is colored according to the p values of biopanning against ATL-43T cells. (H) OX40-expressing activated CD4 T cells stained with each nanobody clone after CD4 T-cell stimulation with anti-CD3/CD28 microbeads. The left and right halves of each bar are colored according to the p values of biopanning against Jurkat-hOX40 and ATL-43T cells, respectively. (I) Jurkat-hOX40 cells pretreated with 315Ab at 0.025, 0.25, 2.5, and 25 µg/mL were incubated with each nanobody (4 µg/mL). The percentage of MFI of each nanobody without pretreatment with 315Ab was calculated. (J) Jurkat-hOX40 cells pretreated with recombinant gp34 at 0.02, 0.2, 2, and 20 µg/mL were incubated with each nanobody (4 µg/mL). The percentage of MFI of each nanobody without pretreatment with recombinant gp34 was calculated. (I, J) The experiments were performed in triplicate, and the mean±SEM of the four representative nanobody clones (Nb8, Nb54, Nb479, and Nb1014) is shown. Curve fitting was performed using GraphPad Prism V.9. cDNA, complementary DNA; NGS, next-generation sequencing.

We generated a mother nanobody library by immunizing alpacas with hOX40 and obtained seven nanobody sublibraries by biopanning against the following cells: (1) wild-type Jurkat cells (Jurkat-WT), (2) Jurkat-hOX40, (3) Jurkat-hOX40 cells treated with 315Ab, (4) Jurkat-hOX40 cells treated with recombinant hOX40 ligand (gp34), (5) ATL-43T cells, (6) wild-type HEK293T cells (HEK293T-WT), and (7) hOX40-expressing HEK293T cells (HEK293T-hOX40) (figure 1B). Finally, we obtained 48,886, 62,810, 59,996, 56,434, 59,534, 33,111, and 43,224 reads, and identified 5,177, 1,660, 2,229, 4,380, 3,132, 3,401, and 2,404 nanobody clones in seven libraries, respectively, totaling 16,795 individual clones with unique amino acid sequences that bind to any of the seven cell types.

Sublibrary comparison identifies potential hOX40-binders

Next, to screen for hOX40-binders, we selected nanobody clones enriched in the sublibraries of hOX40-expressing cells compared with those of non-expressing cells. Specifically, comparisons were made between Jurkat-hOX40 versus Jurkat-WT, Jurkat-hOX40 treated with 315Ab versus Jurkat-WT, Jurkat-hOX40 treated with gp34 versus Jurkat-WT, and HEK293T-hOX40 versus HEK293T-WT. The sublibrary of ATL-43T was compared with that of Jurkat-WT (figure 1B, online supplemental figure 1A–E). Clones with an increased read-count proportion (p<0.2, χ² test) in any comparison were considered positive, identifying 4,476 potential binders. Of them, 381 clones were enriched by all three hOX40-expressing cell lines, whereas 597 clones were enriched only by the ATL-43T cells (online supplemental figure 1F). Some of the latter clones may be specific for molecules other than OX40, which are expressed at higher levels in ATL-43T cells than in Jurkat-WT.

NGS identifies ligand interaction site binding nanobody clones

The NGS screening was performed as follows: first, to obtain a global view of the relationships among the selected nanobody clones, we performed clustering analysis based on the amino acid sequences of the nanobody clones (figure 1C), classifying them into 26 major clusters labeled alphabetically (A, B, C, etc) and subclusters by branch numbers (A-1-1, A-1-2,…, A-2, A-3, etc). Clones within the same cluster were hypothesized to share similar epitopes. Consistently, clones in the same cluster exhibited similar enrichment profiles (increasing or decreasing patterns) for the sublibraries of hOX40-expressing cells (figure 1D, online supplemental figure 2). However, some clones within the same cluster showed different enrichment profiles (enriched by two or more hOX40-expressing cells), indicating that minor amino acid changes could alter the epitope specificity (arrows in figure 1D). Therefore, we screened for clones by considering the data at both the cluster and clone levels.

Cluster A-1, the largest cluster, contained clones highly enriched in all three hOX40-expressing cell-panned sublibraries (Jurkat-hOX40, HEK293-hOX40, and ATL-43T), comprising 71–95% of the total read counts in these sublibraries but less than 20% in the Jurkat-WT or HEK293T-WT sublibraries (online supplemental figure 3). This suggests that cluster A-1 clones target the immunodominant epitope(s) of hOX40. The other clusters contained fewer clones that were enriched in only one or two hOX40-expressing cell lines.

For further screening, we evaluated nanobody clones based on the biopanning results from Jurkat-hOX40 treated with or without 315Ab or gp34. Specifically, comparisons were conducted between Jurkat-hOX40 treated with 315Ab versus Jurkat-hOX40 and Jurkat-hOX40 treated with gp34 versus Jurkat-hOX40 (figure 1E, online supplemental figures 4 and 5). The proportion of total read counts of clones in cluster A-1 remarkably decreased with 315Ab and gp34 treatment (95% in Jurkat-hOX40, 87% in Jurkat-hOX40-315Ab, and 54% in Jurkat-hOX40-gp34), whereas the proportion in the other clusters increased. This suggests that cluster A-1 recognizes the gp34 or 315Ab binding site vicinity on OX40, whereas the non-A-1 clusters might recognize other epitopes or non-OX40 molecules.

Notably, at the clone level, some non-A-1 clones enriched by Jurkat-hOX40 competed with 315Ab or gp34 (arrows in figure 1E). In particular, two clones from cluster C showed the unique feature of competing with gp34 but not with 315Ab, suggesting that they recognize distinct epitopes within the gp34-binding site away from the 315Ab-binding site.

Finally, based on the clustering analysis and binding profiles, we selected the following nanobody clones for in vitro testing (online supplemental figure 6): (1) 16 clones from the immunodominant cluster A-1; (2) clone Nb479 from cluster C with a unique binding profile for hOX40 and the gp34 interaction site; (3) 7 clones with the highest read counts in clusters A-2, B, D, E, M, and X; (4) 10 clones enriched by cell lines other than Jurkat-hOX40; and (5) 3 negative control clones not enriched by any hOX40-expressing cells.

Flow cytometry confirms specific nanobody clone binding to hOX40

We performed a cell-based flow cytometry assay to confirm the specific binding of the nanobody clones to hOX40. The results largely aligned with the biopanning data, confirming the binding capability of the enriched clones, although some discrepancies were noted between Jurkat-hOX40 and ATL-43T binding, likely due to the presence of clones recognizing ATL-43T-specific antigens other than OX40 (figure 1F,G).

Next, to verify binding to naturally occurring hOX40, we tested human primary CD4 T cells. Most clones enriched by both Jurkat-hOX40 and ATL-43T cells bound to OX40-expressing CD4 T cells (figure 1H). However, clones enriched only by ATL-43T but not by Jurkat-hOX40 failed to bind to CD4 T cells, except for Nb1027, suggesting that these clones may recognize ATL43-specific molecules distinct from OX40 that are not expressed on Jurkat-WT. Notably, almost all clones enriched in hOX40-expressing cells did not bind to Jurkat-WT, suggesting high specificity for hOX40 (online supplemental figure 7).

Finally, we tested hOX40-binders for binding to Jurkat-hOX40 pretreated with serial concentrations of 315Ab or gp34 (figure 1I,J, online supplemental figure 8). Consistent with the biopanning results, the binding of cluster A-1 clones, such as Nb8 and Nb54, was blocked by both 315Ab and gp34. In contrast, the binding of Nb479 from cluster C and Nb1014 from cluster E was blocked by gp34 but not by 315Ab, confirming the correlation between cluster classification and epitope specificity.

Nanobodies enhance proliferation and cytokines in primary T cells

To identify the nanobody with the best T-cell stimulatory activity, we performed cell-based functional assays. As OX40 belongs to the TNFRSF and requires trimerization for activation, we immobilized nanobodies on a culture plate to cross-link OX40. CD4 T cells were cultured for 96 hours in microwells coated with anti-CD3 antibody and each nanobody, and cell proliferation and cytokine production were assessed. Anti-CD28 induced the most robust proliferation and cytokine production, likely because CD28 is constitutively expressed and does not require prior priming, whereas anti-OX40 agonism depends solely on anti-CD3 antibody-mediated priming for OX40 upregulation. Plate-bound Nb479 was the most effective in promoting proliferation (figure 2A,B) and enhancing IFN-γ and GM-CSF production (figure 2C,D), whereas other nanobodies, such as Nb8 and Nb54, also exhibited significant but lower activity. In CD8 T cells, Nb479 significantly stimulated IFN-γ production, but did not affect proliferation or GM-CSF production (figure 2E–G). As OX40 transduces activating signals to nuclear factor kappa B (NF-κB),^{42 43} we next performed a luciferase reporter assay using Jurkat-hOX40 to assess the potency of each nanobody in inducing NF-κB activation. The luciferase reporter assay in Jurkat-hOX40 confirmed that immobilized Nb479 was the most potent in inducing NF-κB activity (figure 2H). Accordingly, we selected the Nb479 clone for further development.

Figure 2. Nanobodies enhance proliferation and cytokine production of primary T cells via OX40. (A) CFSE-labeled CD4 T cells (0.5×10⁵ cells) were cultured in microwells coated with anti-CD3 antibody and nanobodies, antibodies, or recombinant gp34 for 96 hours, and the proliferating cells determined by CFSE dye dilution using flow cytometry. (B) The absolute number of proliferating cells with diluted CFSE dye was counted using Flow-Count Fluorospheres. (C, D) After 96 hours culture, the supernatant of each microwell was harvested and the concentrations of IFN-γ and GM-CSF were measured. (E) CFSE-labeled CD8 T cells (0.5×10⁵ cells) were cultured in microwells coated with anti-CD3 antibody and nanobodies, antibodies, or recombinant gp34 for 96 hours, and the absolute number of proliferating cells was counted. (F, G) After 96 hours culture, the supernatant of each microwell was harvested and the concentrations of IFN-γ and GM-CSF were measured. (H) Jurkat-hOX40 cells transfected with NFkB-Firefly luciferase (Fluc) reporter plasmid and EF1α-Renilla luciferase (Rluc) plasmid were incubated in microwells coated with the indicated antibodies, nanobodies, or recombinant gp34 for 24 hours. Firefly luciferase activity was normalized to Renilla luciferase activity (Fluc/Rluc). (I) Schematic diagram of the dimeric or trimeric Nb479 construct. Two or three Nb479 nanobodies were tandemly linked through (G4S)₃ linker sequences, respectively. (J) Jurkat-hOX40 cells transfected with NFkB-Fluc reporter plasmid and EF1α-Rluc plasmid were incubated with the indicated nanobodies in soluble form for 24 hours. (H, J) Averages of six replicates are shown as mean±SEM. (K) CFSE-labeled CD4 T cells (0.5×10⁵ cells) were cultured in anti-CD3 antibody-coated microwells in the presence of the indicated nanobodies or antibodies in a soluble form for 96 hours, and the absolute number of proliferating cells was counted. (L, M) After 96 hours culture, the supernatant of each microwell was harvested and the concentrations of IFN-γ and GM-CSF were measured. (K, L, M) All experiments were performed with five or six replicates. The mean±SEM are shown. CFSE, carboxyfluorescein succinimidyl ester; GM-CSF, granulocyte-macrophage colony-stimulating factor; IFN, interferon.

Non-immobilized trimerized Nb479 elicits agonistic activity

Given that the anti-OX40 nanobody would ultimately be administered to patients and function in vivo, it must be able to activate OX40 without being immobilized on a plastic plate, as in in vitro experiments. To enable Nb479 to function and activate OX40 in vivo, we generated dimerized and trimerized forms of Nb479 using (G4S)₃ linkers⁴⁴ (figure 2I). There was no appreciable difference in binding when the orientation of the Nb479–NbHER2 heterodimer fusion construct was reversed, suggesting that the (G4S)₃ linker does not impose unfavorable steric constraints on receptor engagement (online supplemental figure 9). The NF-κB reporter assay revealed that the soluble form of the Nb479 monomer (Nb479m) had no activity, confirming the necessity of cross-linking, while both the Nb479 dimer (Nb479d) and trimer (Nb479t) significantly induced NF-κB activation, with Nb479t being more potent (figure 2J). Notably, human Fc conjugated gp34 induced NF-κB activation, whereas 315Ab did not. Cross-linking with anti-IgG further enhanced gp34-induced NF-κB activation, while it had no effect on 315Ab, suggesting that, in the soluble format, the binding orientation of 315Ab may not promote productive receptor clustering (online supplemental figure 10). The soluble form of Nb479t also stimulated primary CD4 T cells, unlike Nb479m (figure 2K–M). Additionally, Nb479t increased the phosphorylation of AKT and the NF-κB p65 subunit in T cells compared with NbHER2t in the presence of plate-bound anti-CD3 stimulation, confirming its efficacy in activating the OX40-PI3K/AKT pathway² (online supplemental figure 11). Heterotrimeric nanobodies (Nb479–Nb8–Nb479 and Nb479–Nb54–Nb479) did not outperform Nb479t, suggesting the importance of homotrimeric assembly of Nb479 homotrimer (online supplemental figure 12).

Wide paratope of Nb479 binds OX40 across multiple domains

To elucidate the structural mechanism underlying the agonistic activity of Nb479, we aimed to identify its binding domains on OX40. As anti-hOX40 nanobodies did not bind to mOX40, we created human-mouse OX40 chimeras with swapped cysteine-rich domains (CRDs) to determine the nanobody binding domains (figure 3A,B, online supplemental figure 13). The polyclonal anti-mOX40 antibody bound strongly to mOX40, weakly to hOX40, and moderately to the four chimeras, indicating successful domain replacement and retention of the antigenic structure. Nb8 (cluster A-1-3), Nb34 (cluster A-1-2), Nb54 (cluster A-1-1), and 315Ab did not bind to the CRD3-substituted OX40, suggesting that cluster A nanobodies and 315Ab target CRD3. In contrast, Nb479 and Nb1014 exhibited decreased binding across all chimeras, similar to gp34, indicating that Nb479 and Nb1014 bind to a broad interface. Furthermore, Nb479t completely blocked gp34 binding, and vice versa, suggesting that their epitopes are in close proximity (figure 1J, online supplemental figure 14).

Figure 3. Epitope mapping of nanobodies and structural model of Nb479. (A) Schematic representation of the four human-mouse OX40 chimeras. (B) The targeting CRD of each ligand or nanobody was determined by an ELISA using human-mouse OX40 chimeras. The means±SEM of four replicates are shown. (C) Overlay of 20 NMR structures of Nb479C(50,110)A. (D) Overall NMR structure of Nb479C(50,110)A (top). β-strands and helical regions are shown in green and orange, respectively. The side chains of the CDR3 are highlighted with nitrogen, oxygen, and sulfur atoms colored blue, red, and yellow, respectively. Electrostatic surface potential of Nb479C(50,110)A (bottom). Red and blue indicate negative and positive charges on the surface, respectively. (E) NMR binding assay of [15N]-Nb479C(50,110)A to hOX40. Residues in [15N]-Nb479C(50,110)A exhibiting CSP on binding to hOX40 are labeled. Magenta: CSP >AV plus 2 SD; cyan: CSP >AV plus 1 SD. (F) Mapping of residues exhibiting CSP onto the Nb479C(50,110)A structure. Magenta and cyan are the same as in figure 3F. Figure 3C,D (left), and F (left) are viewed in the same orientation. (G) Overall NMR structure of Nb479C(50,110)A (left) and the structural model of Nb479m (right), viewed from the top of CDR3. These structures are tilted 50$°$ around the horizontal axis relative to those in figure 3C,D (left), and F (left). (G, top) β-strand and helical regions are shown in green and orange, respectively. The side chains of the CDR3 residues are highlighted with nitrogen, oxygen, and sulfur atoms colored blue, red, and yellow, respectively. (G, bottom) Electrostatic surface potentials of Nb479C(50,110)A (left) and Nb479m (right). Red and blue represent negative and positive charges, respectively. AV, average value; CSP, chemical shift perturbation; CRD, chemical shift perturbation; NMR, nuclear magnetic resonance.

To further investigate the structural properties of Nb479, particularly its wide interface, we analyzed its solution structure using NMR spectroscopy. Nb479m contained four cysteine residues (C21, C50, C96, and C110), with C21 and C96 forming the canonical disulfide bond connecting framework regions 1 and 3 (FR1 and FR3) in the two β-strands of the immunoglobulin domain.⁴⁵ C50 and C110 formed an additional disulfide bond connecting FR2 and complementarity-determining region 3 (CDR3) within the β-sheet and the long loop, respectively.⁴⁵ The structural ensemble of the Nb479C(50,110)A models, where C50 and C110 were substituted by alanine, showed good convergence with an average root mean square deviation of 0.53 Å for the N, Cα, and C atoms (figure 3C, tables 1 and 2). Nb479C(50,110)A exhibited a typical immunoglobulin domain structure, consisting of a bilayer sandwich of nine antiparallel β-strands arranged in two β-sheets (figure 3D, top). The NMR structure of Nb479C(50,110)A revealed that the CDR3 loop was located on top of a β-sheet and was negatively charged (figure 3D, bottom). An NMR binding assay identified 14 residues (V2, Q5, S53, Y60, A75, E89, A98, D101, V106, A110, V111, E113, D118, and S129) in Nb479C(50,110)A that showed chemical shift perturbations on binding with hOX40, predominantly within or near CDR3 (residues 99–119), indicating its role in binding (figure 3E,F, online supplemental figure 15).

Table 1. NMR and refinement statistics for Nb479C(50,110)A structures.

	Nb479C(50,110)A
NMR distance and dihedral restraints
Number of NOE distance restraints	2,457
Intra-residue	532
Inter-residue
Sequential (|i−j|=1)	620
Medium-range (1<|i−j|<5)	231
Long-range (|i−j|≥5)	1,074
Hydrogen bonds	32
Total dihedral angle restraints
χ1 and χ2 angles	9
Structure statistics
Violations (mean and SD)
Distance restraints (Å)	0.0044±0.0005
Dihedral angle restraints (°)	0.580±0.120
Max. distance violations (Å)	0.33
Max. dihedral angle violation (°)	2.58

NMR, nuclear magnetic resonance.

Table 2. AMBER refinement statistics for Nb479C(50,110)A structures.

	Nb479C(50,110)A
Structure statistics
Energy (mean and SD)
E-AMBER (kcal/mol)	−4,256.85±20.70
NMR NOE	10.26±1.37
NMR dihedral	0.88±0.26
Violations
Number of distance restraints >0.1 Å, >0.5 Å	2.0±1.4, 0.0±0.0
Number of dihedral angle restraints >2.5°, >5.0°	0.0±0.0, 0.0±0.0
Max. distance violations (Å)	0.24
Max. dihedral angle violation (°)	n/a
Ramachandran plot statistics (%)*
Residues in favored regions	88.8
Residues in allowed regions	11.2
Residues in outlier regions	0.1
Average pairwise r.m.s. deviation (Å)*
Heavy	0.896
Backbone (N, Cα, C)	0.530

^*Fit_Robot was used to obtain the residues that are well ordered; 1–21, 23–41, 43–53, 57–95, 97–103, and 116–130 were identified.

NMR, nuclear magnetic resonance; r.m.s., root mean square.

We built a structural model of Nb479m, incorporating the additional disulfide bond between C50 and C110, based on the NMR data obtained for Nb479C(50,110)A (figure 3G). The bilayer β-sheet structure of Nb479m was similar to that of Nb479C(50,110)A; however, the CDR3 backbone trace in the Nb479m model was wider than that in Nb479C(50,110)A (figure 3G, top). The molecular surface of the CDR3 of Nb479m was wider and more planar with some concave regions (figure 3G, bottom). Consequently, Nb479m is suggested to have a large, negatively charged surface. These findings imply that the Nb479m-hOX40 complex has a wide binding interface across multiple CRDs on hOX40, similar to the gp34-hOX40 complex.⁴⁶

Trimeric Nb479 enhances CD19 CAR-T cell antitumor function

To apply Nb479t to in vivo tumor treatment, we first investigated whether OX40 signaling via Nb479 could enhance T-cell function using the CD19 CAR-T cell system. CD19 CAR-T cells incorporating 4–1BB and CD3ζ intracellular domains were tested against FL218 cells, a CD19+B cell line from a patient with transformed follicular lymphoma⁴⁷ (online supplemental figure 16A). As OX40 expression peaked at 24 hours and lasted at least 48 hours (online supplemental figure 16B), we preloaded irradiated FL218 cells to induce OX40 upregulation before testing cytotoxicity (figure 4A). Nb479t treatment demonstrated significantly superior cytotoxicity along with a trend of higher CAR-T cell proliferation, whereas Nb479m enhanced cytotoxicity in the immobilized form but not in the soluble form (figure 4B, online supplemental figure 16C,D). The levels of phosphorylated AKT and p65 were increased in CAR-T cells co-cultured with FL218 cells under Nb479t treatment (online supplemental figure 16E).

Figure 4. Trimeric Nb479 enhances the antitumor function of CD19 CAR-T cells both in vitro and in vivo. (A) CD19-specific CAR-T cells (0.2×10⁵ cells/well) were preloaded with irradiated CD19+FL218 cells (0.2×10⁵ cells/well) and then co-incubated with FL218-LucZsG target cells (0.2×10⁵ cells/well). (B) The indicated nanobodies were supplied in soluble form. The vertical axis indicates (100 – (percentage of remaining FL218-LucZsG cells to absolute cell counts of FL218-LucZsG cells cultured without CAR-T cells)). The averages of six replicates are shown as mean±SEM. (C) Schematic of the Nb479t-MSA construct. Anti-MSA nanobody was tandemly linked using a (G4S)₃ linker. (D) 5 days after intravenous inoculation with FL218-LucZsG cells, CAR-T cells were injected and nanobodies were administered every 3 days. (E) Bioluminescence images of tumor growth. (F) Bioluminescence intensity in each mouse. (G) The mean±SEM of each group is shown. (H) Survival curves. (I) Percentage of tumor-infiltrating CAR-T cells among living cells. CAR, chimeric antigen receptor; MSA, mouse serum albumin.

To test Nb479t in a mouse model, we first assessed the binding kinetics of Nb479m with recombinant hOX40 using bio-layer interferometry. Nb479m exhibited good association (Kon=9.87E+04) but rapid dissociation (Koff=2.77E−03) (online supplemental figure 17A). Flow cytometry analysis of Jurkat-hOX40 showed 50% receptor occupancy at 0.49 µg/mL for Nb479m, while Nb479t achieved the same at a much lower concentration (0.024 µg/mL), probably because of the reduced Koff (online supplemental figure 17B). The binding of Nb479t was almost saturated at 0.2 µg/mL. The full agonist function of Nb479t on CD4 T-cell proliferation was elicited at >0.008–0.2 µg/mL (online supplemental figure 17C).

To extend in vivo half-life, we fused an anti-mouse serum albumin (MSA) nanobody to the C-terminus of the trimeric nanobody. In a pharmacokinetic study in mice, Nb479t conjugated to MSA (Nb479t-MSA) exhibited a markedly prolonged serum half-life of up to 48 hours versus 1.1 hour for non-conjugated Nb479t, maintaining serum concentrations above 1 µg/mL for 7 days after intravenous injection of 100 µg (figure 4C, online supplemental figure 18), without affecting its ability to bind to hOX40 (online supplemental figure 17B). Consequently, 100 µg of Nb479t-MSA was administered every 3 days to maintain effective serum levels (>0.2 µg/mL).

In a luciferase and ZsGreen co-expressing FL218 xenograft model, CAR-T cells were infused 5 days post-tumor cell intravenous inoculation with 100 µg of Nb479t-MSA, administered intravenously every 3 days (figure 4D). Nb479t-MSA-treated mice exhibited delayed tumor progression compared with those treated with control MSA-conjugated NbHER2t (NbHER2t-MSA), as shown by bioluminescence imaging (figure 4E–G), and showed a higher survival trend (HR, 2.70: 95% CI 0.78 to 9.37) (figure 4H). Nb479t-MSA also promoted CAR-T cell infiltration into tumor tissues (figure 4I, online supplemental figure 19). In contrast, 315Ab treatment did not exhibit enhanced antitumor effects compared with the isotype IgG2a control (online supplemental figure 20A–D).

Discussion

In this study, we explored potential anti-OX40 agonistic clones derived from an extensive library of over 16,000 clones generated from an immunized alpaca. By employing biopanning and clustering analysis, we confined the candidate clones to those targeting the receptor-ligand interaction site and successfully identified a distinct clone that mimicked the native ligand gp34 and demonstrated a wide paratope and favorable agonistic activity against OX40. Moreover, we enhanced the functionality of the nanobody in vivo by generating multimeric forms and conjugating them with MSA, which significantly enhanced the induction of the molecular assembly of OX40 for activation and improved the pharmacokinetics.

Predicting the agonistic activity of a nanobody solely from its DNA sequence or binding profile is challenging. Thus, in vitro assays are essential for verifying agonistic potential. To streamline screening, it is beneficial to reduce the number of clones that require in vitro testing. Previous studies employed phage display in combination with heavy chain variable region NGS to identify top-ranked non-synonymous clones based on enrichment levels.^{48 49} However, as demonstrated in our study, selection strategies relying solely on read count-based enrichment can overlook high-affinity clones whose apparent abundance is underestimated due to enrichment-driven biases, such as antigenicity bias, PCR amplification bias, and differences in protein expression efficiency. Consequently, potentially functional clones may be missed when enrichment metrics alone are used for clone prioritization. In contrast, our approach combined biopanning against ligand-engaged targets and clustering analysis, enabling the incorporation of epitope-level information into the selection process. This strategy helped us avoid redundant assays on numerous similar clones and identify rare and unique clones within small clusters. Our study demonstrated that a clone belonging to a small cluster with a unique binding profile can be particularly valuable.

The phylogenetic tree of the clustering analysis provided a clear overview of the clonal relationships of the nanobodies. Clones within the same cluster, having similar amino acid sequences, likely share close epitope specificities, as evidenced by their binding profiles in biopanning. A large cluster may represent immunodominant epitopes that drive cluster formation. Discrepancies in the binding profiles among clones in the same cluster could be attributable to the limited sensitivity of biopanning or actual differences in epitope specificities.

Trimerization of TNFRSF molecules via ligand-binding plays a crucial role in inducing intracellular signaling.¹⁵ The native OX40 ligand gp34 has two binding sites for OX40 at a 180° angle, which induces OX40 trimerization with each gp34 molecule associated with two OX40 molecules. The binding site of gp34 to OX40 has a broad interface spanning multiple CRDs, but its significance in agonistic function remains unclear.⁴⁶ Of note, previous studies have reported that antibody clones with comparable binding affinities and similar kinetic parameters (Kon/Koff) can exhibit markedly different agonistic activities depending on the clone (epitope), suggesting that antibody-induced changes in receptor conformational dynamics may also contribute to signal activation, beyond simple receptor cross-linking. Although the relationship between antibody binding domains and downstream signaling activity has been investigated for type II TNFRSF members composed of four CRDs, such as OX40 and CD40, the results have been inconsistent.^{15 18–20} For OX40, studies have reported that antibody binding to CRD2 and CRD4 is associated with stronger signaling activation, whereas studies on CD40 have suggested that engagement of the membrane-distal domains (CRD1–2) is more important for agonistic activity. However, in both cases, only a limited number of antibody clones—at most a few—were tested per CRD. Given that each CRD represents a three-dimensionally extended structure with multiple possible epitopes, the limited sampling is likely insufficient to draw robust conclusions regarding domain-level agonistic potential. Moreover, unlike the natural ligand gp34, which engages the receptor across multiple CRDs, no antibody to date has been engineered to recapitulate such multidomain receptor engagement.⁴⁶ As a result, the structural basis by which antibody binding translates into signaling activation in type II TNFRSF members remains incompletely defined. Notably, our study showed that Nb479, which had the most robust agonistic activity, had a broad paratope. Some rotational freedom has been demonstrated between CRD3 and CRD4, whereas CRD1 and CRD2 form a rigid unit in OX40.⁴⁶ The broad binding interface of gp34 and Nb479 across multiple CRDs may contribute to stabilization of the intramolecular structure of OX40, which may be necessary to form a stable OX40 trimer. This hypothesis needs to be confirmed in future studies using X-ray crystallography or cryo-electron microscopy at higher resolution. Notably, AlphaFold3 modeled the Nb479-OX40 interface spanning across CRD1 and CRD2, exhibiting similarity to the OX40-gp34 interface, while suggesting an association of the interface of Nb8, Nb34, and Nb54 with CRD2–3⁵⁰ (online supplemental figure 21). Although confidence estimates for the nanobody-OX40 interface in AlphaFold3 were generally low (plDDT<70) and predicting variable loops remains challenging, advancements in artificial intelligence models might assist in predicting epitopes in the future. Despite its broad binding site, Nb1014 did not exhibit agonism, probably because of its low affinity for OX40.

Previous studies have demonstrated the agonistic potential of modified gp34.^51–53 However, gp34 exhibits substantially weaker binding affinity compared with Nb479t. Based on competition assays, the concentration required to achieve 50% receptor occupancy for Nb479t is 0.014 µg/mL (95% CI 0.010 to 0.018), whereas Fc conjugated gp34 requires 1.018 µg/mL (95% CI 0.621 to 1.607) (figure 1J and online supplemental figure 14). Moreover, while the orientation of tandem conjugation does not affect the binding of Nb479 (online supplemental figure 9), gp34 has been reported to exhibit orientation-dependent effects on both binding and agonistic activity when configured in tandem formats.⁵¹ However, it should be noted that our study did not directly compare the signaling potency of Nb479-based constructs with engineered forms of gp34, such as trimerized or half-life–extended variants, despite their apparent structural and mechanistic similarity to the trimeric organization of gp34. As a naturally occurring ligand, gp34 may offer advantages, including potentially lower immunogenicity, whereas Nb479 and its derivatives could carry a risk of immunogenicity. Therefore, although Nb479 provides advantages in terms of affinity and engineering flexibility, it remains unclear whether it is superior to optimized gp34-based approaches. Further studies will be required to systematically compare these strategies and determine their relative therapeutic potential.

For effective in vivo application, it is crucial to achieve cross-linking of OX40 with the soluble form of nanobody and to maintain a stable serum concentration. Huet et al demonstrated that a multimerized nanobody in soluble form successfully elicited antitumor activity against DR5, a member of the TNFRSF.⁴⁴ In addition, conjugation with anti-MSA nanobody significantly extends nanobody half-life in vivo.⁵⁴ Accordingly, our study found that the MSA-conjugated Nb479t successfully enhanced the antitumor effects of CAR-T cells in a xenograft mouse model, highlighting the scalability of nanobody technology for in vivo application according to the characteristics of the target receptor.

Despite the promising potential of Nb479t-MSA, several aspects could be further optimized to facilitate clinical translation. It remains possible that alternative linker designs beyond the (G4S)₃ linker may further enhance agonistic potency or improve physicochemical properties such as solubility and stability. In addition, given that increased valency is associated with enhanced agonistic activity, the optimal valency of the construct remains to be determined.^{44 55} Finally, optimization of pharmacokinetic properties, including half-life and dosing strategy, will be critical for achieving robust efficacy in humans.

Although we used alpaca immunization in this study, the conceptual framework of our approach is broadly applicable to other antibody generation strategies, including (1) nanobody versus conventional antibody formats, (2) immunization-based versus naïve libraries, and (3) natural versus synthetic libraries. Regarding antibody format, nanobodies offer advantages for single-domain architectures and facilitate downstream NGS-based analysis and engineering. However, conventional antibodies may still be advantageous for targeting linear epitopes or for achieving high binding affinities. In addition, phylogenomic analyses indicate that the genomic divergence between Laurasiatheria (including camelids) and Euarchontoglires (including primates and rodents) is greater than that between humans and mice.⁵⁶ Therefore, compared with mice, the use of alpacas may benefit from their greater genetic distance from humans, which reduces immune tolerance to conserved mammalian proteins and increases the likelihood of generating antibodies against a broader epitope repertoire. With respect to library generation, immunization enables affinity maturation in vivo, whereas naïve libraries may be preferable when antigens are toxic. Furthermore, synthetic libraries offer complementary advantages by enabling access to epitopes that are highly conserved across species or essential for biological function, which are often poorly immunogenic in vivo due to self-tolerance mechanisms.⁵⁷ Taken together, the optimal antibody generation strategy should be selected based on the target biology and the desired functional properties of the antibody. Importantly, our proposed approach, which integrates ligand-engaged targets with high-throughput NGS-based screening, is expected to enable systematic extraction of epitope-level information and provide a synergistic advantage irrespective of the specific platform used.

This study has limitations. We did not evaluate which factor, OX40 trimerization or the epitope, is more critical for agonistic activity. Nonetheless, the fact that non-immobilized gp34 and Nb479d induced NF-κB activity in the luciferase reporter assay, whereas 315Ab did not, suggests the contribution of the epitope to the activity. In addition, we did not perform direct comparisons between Nb479 and other clones (Nb8, Nb34, Nb54), nor with previously reported antibodies or antibodies currently in clinical development, in more clinically relevant models. Surprisingly, in the bioluminescence imaging, the antitumor activity of CAR-T cells was significantly attenuated by 315Ab treatment (online supplemental figure 20A–C). We did not observe a difference in peripheral CAR-T expansion (online supplemental figure 20E). The exact reason remains unknown; however, because conventional antibodies rely on Fc receptor-mediated cross-linking for agonism and antibody-dependent cell-mediated cytotoxicity for regulatory T-cell depletion, our xenograft model using immunodeficient mice was not appropriate for evaluating such antibody treatments. Furthermore, the mouse model showed only modest improvement with treatment. We presume that CAR-T cells already incorporate 4–1BB, which may functionally overlap with OX40 signaling and potentially have obscured the results. Therefore, we believe an OX40-humanized, immune-competent mouse model without overlapping T-cell co-stimulatory molecules will address these limitations. Lastly, we observed a bimodal peak in size-exclusion chromatography of Nb479t, with the estimated molecular weight of the larger peak being approximately 70 kDa, whereas the theoretical molecular weight of Nb479t is 44.8 kDa, suggesting weak intermolecular association (online supplemental figure 22A–D). Given that OX40 is activated through higher-order clustering, this effect could partially account for the observed activity. A detailed dissection of the relative contributions of intrinsic agonism versus association-driven clustering will require further investigation.

In summary, our study demonstrates a novel epitope-directed strategy to develop agonistic antibodies that specifically target the ligand-receptor interface. Using this approach, we generated a multivalent anti-OX40 nanobody with robust agonistic activity capable of directly stimulating OX40-expressing T cells without requiring secondary cross-linking, offering a potentially effective tumor immunotherapy. The epitope-focused strategy is broadly applicable, including the development of virus-neutralizing and tumor-specific antibodies. For example, our group has recently applied similar approaches to generate nanobodies targeting the receptor-binding domain of SARS-CoV-2 and a nanobody panel targeting triple-negative breast cancer cell lines.^58–60 Overall, the epitope-directed antibody discovery strategy provides a powerful platform for advancing next-generation antibody-based therapeutics.

The funder did not influence the results/outcomes of the study despite author affiliations with the funder.

Data availability statement

Data are available upon reasonable request. All data relevant to the study are included in the article or uploaded as supplementary information.