Humanized Affinity-matured Monoclonal Antibody 8H9 Has Potent Antitumor Activity and Binds to FG Loop of Tumor Antigen B7-H3*

Background: B7-H3 is an immune inhibitory ligand and an antigen on many solid tumors. Results: Antibody 8H9 was humanized and affinity-matured, and its epitope was mapped to the FG loop of B7-H3. Conclusion: hu8H9 antibodies had potent antitumor activity and may modulate immune inhibitory properties of B7-H3. Significance: Antibodies were developed to target solid tumors and affect immune checkpoint blockade. B7-H3 (CD276) is both an inhibitory ligand for natural killer cells and T cells and a tumor antigen that is widely expressed among human solid tumors. Anti-B7-H3 mouse monoclonal antibody 8H9 has been successfully used for radioimmunotherapy for patients with B7-H3(+) tumors. We present the humanization, affinity maturation, and epitope mapping of 8H9 based on structure determination, modeling, and yeast display methods. The crystal structure of ch8H9 Fab fragment was solved to 2.5-Å resolution and used as a template for humanization. By displaying the humanized 8H9 single chain Fv (scFv) on the surface of yeast, the affinity was matured by sequential random mutagenesis and fluorescence-activated cell sorting. Six mutations (three in the complementarity-determining region and three in the framework regions) were identified and incorporated into an affinity-matured humanized 8H9 construct (hu8H9-6m) and an affinity-matured chimeric 8H9 construct (ch8H9-6m). The hu8H9-6m scFv had a 160-fold improvement in affinity (0.9 nm KD) compared with parental hu8H9 scFv (144 nm KD). The IgG formats of ch8H9-6m and hu8H9-6m (nanomolar to subnanomolar KD) had 2–9-fold enhancements in affinity compared with their parental forms, potent in vitro antibody-dependent cell-mediated cytotoxicity (0.1–0.3 μg/ml EC50), and high tumor uptake in mouse xenografts. Based on in silico docking studies and experimental validation, the molecular epitope of 8H9 was determined to be dependent on the FG loop of B7-H3, a region critical to its function in immunologic blockade and unique among anti-B7-H3 antibodies published to date.

8H9 is a murine IgG1 mAb that targets B7-H3 (36,37). By immunostaining, it is broadly reactive with human solid tumors, including embryonal tumors and carcinomas (36). It has shown favorable tumor uptake for both sarcoma and brain tumors in xenograft models (38,39). When conjugated to cobra venom factor, it induces efficient complement-mediated tumor lysis (40). In its single chain Fv (scFv) format, it targets a potent immunotoxin to sarcoma and glioma in preclinical models (41,42). As a chimeric antigen receptor, it redirects natural killer cells to kill B7-H3(ϩ)-positive tumor cells (43). In early phase human clinical trials, 131 I-8H9 prolongs survival among high risk patients with solid tumors suffering from central nervous system (CNS) metastasis (44 -46). It is a promising target for radioimmunotherapy of leptomeningeal metastases (NCT00089245) (46), diffuse intrinsic pontine glioma (NCT01502917) (34), and peritoneal metastases (NCT01099644). The only other antibody in phase I trial is MGA271, a humanized IgG1 mAb (NCT0139114) (47). B7-H3 expression at the mRNA and protein levels is regulated by microRNA miR-29 in tumors and normal tissues (37). When neuroblastoma metastasized to the brain, its miR-29 was downregulated, which was associated with increased expression of several oncoproteins as well as B7-H3 (48).
Here we describe a strategy to humanize and affinity mature mAb 8H9 based on x-ray crystallography, in silico modeling, and yeast display. We were additionally able to predict and validate the molecular epitope on B7-H3 to be dependent on the FG loop, which may affect the immune modulatory properties of B7-H3.

Experimental Procedures
X-ray Crystallography-Fab fragments of ch8H9 were generated by papain digestion using a standard Fab preparation kit (Pierce). The purified Fab fragment was concentrated to 10 mg/ml in 20 mM HEPES, pH 6.5, and was crystallized in a hanging drop by vapor diffusion at 16°C against a reservoir containing Hampton Index reagent containing 0.2 M lithium sulfate monohydrate, 0.1 M Bis-Tris, pH 6.5, 25% (w/v) polyethylene glycol 3,350 (Hampton Research, Aliso Viejo, CA). The droplet was formed by mixing 1 l of protein solution and 1 l of reservoir solution. Data were collected at the Argonne Advanced Photon Source beamline 24IDE. The Fab structure was solved by molecular replacement using MOLREP (49). The best molecular replacement model was refined using PHENIX (50), and manual fitting was performed with O (51).
Molecular Simulations of ch8H9 Structure-The crystal structure was simulated using CHARMm (Chemistry at Harvard Molecular Mechanics) force fields, and the effect of each point mutation was calculated from the difference between the folding free energy of the mutated structure and the wild type protein. Generalized Born approximation was used to account for the effect of the solvent, and all electrostatic terms were calculated as a sum of coulombic interactions and polar contributions to the solvation energy. A weighted sum of the van der Waals, electrostatic, entropy, and non-polar terms was calculated for each point mutation. All calculations were performed using Discovery Studio (Dassault Systems, San Diego, CA). Docking simulations were generated using ZDOCK (52), and homology modeling of B7-H3 was done using MODELLER (53). Electrostatic surface potentials were generated using Del-Phi (54). All images were rendered using Discovery Studio.
Biotinylation of B7-H3 Antigen-The gene of human B7-H3mouse Fc (mFc) fusion was optimized for CHO cells (Genscript). Using the Bluescript vector, the gene was transfected into DG44 cells and expressed as described previously (55). Biotinylation of B7-H3-mFc was performed using EZ-Link sulfo-NHS-biotin and biotinylation kit (Thermo Scientific) according manufacturer's instructions. The biotinylated protein was subsequently concentrated using a 50,000-molecular weight cutoff Vivaspin centrifuge tube (Sartorius Stedim) and tested for its biotinylation in streptavidin binding ELISA.
Mutagenesis by Error-prone PCR-Error-prone PCR of the entire hu8H9 scFv (V3) gene was performed using Stratagene GeneMorph II Random Mutagenesis kit according to the instructions of the manufacturer. Reaction products were purified by 1% agarose gel electrophoresis and concentrated by ultrafiltration with water.
Selection of hu8H9 Mutants from the Yeast Libraries-Construction and growth of yeast libraries for affinity maturation were modified from previous publications (56,57). Before fluorescence-activated cell sorting (FACS), the yeast library (1 ϫ 10 9 cells) was preincubated with control mouse 3F8 IgG, sequentially panned against 10 g of B7-H3-mFc-conjugated magnetic beads, and separated with a magnetic stand. For the first yeast library, three FACS selections were performed with 10, 3, and 1 g/ml biotinylated B7-H3-mFc, respectively. Yeast plasmids were isolated using Zymoprep Yeast Plasmid Mini-prep II kit (Zymo Research) according to the manufacturer's instructions and used for templates of the second library construction. The construction and selection of the second library were the same as for the first yeast library. However, four FACS selections were performed in the second yeast library with 1, 0.2, 0.05, and 0.01 g/ml biotinylated B7-H3-mFc, respectively. Cells from the last selection were spread on SD-CAA (18.2% sorbitol, 2.0% glucose, 0.67% yeast nitrogen base, 0.5% casamino acids, 0.54% disodium phosphate, 0.86% monosodium phosphate, and 1.5% agar) plates. Monoclonal yeast cells were characterized, and isolated antibodies with improved affinity were sequenced.
Expression and Purification of Soluble scFv-scFv fragments were expressed and purified as described previously (56). All recombinant scFv fragments had FLAG and His tags.
Expression and Purification of IgG Constructs-Using a Bluescript vector (Eureka Therapeutics), the heavy and light chain genes of 8H9 constructs were transfected into CHO-S cells and selected with G418 (Invitrogen). 8H9 producer lines were cultured in OptiCHO serum-free medium (Invitrogen), and the mature supernatant was harvested. A protein A affinity column was pre-equilibrated with 25 mM sodium citrate buffer with 0.15 M NaCl, pH 8.2. Bound 8H9 was eluted with 0.1 M citric acid, sodium citrate buffer, pH 3.9, and dialyzed in 25 mM sodium citrate, 150 mM NaCl, pH 8.2.
Affinity Determination by Surface Plasmon Resonance-In brief, B7-H3 antigen (either 2Ig-B7-H3, 4Ig-B7-H3, 2Ig-B7-H3-mFc, or 4Ig-B7-H3-mFc) was immobilized onto the CM5 sensor chip via amine coupling. A reference surface was set up for nonspecific binding and refractive index changes. For analysis of the kinetics of interactions, varying concentrations of scFv fragments (25-800 nM) or IgG (6.25-400 nM) were injected at a flow rate of 30 l/min using running buffer containing 10 mM HEPES, 150 mM NaCl, 3 mM EDTA, 0.05% Surfactant P-20, pH 7.4. The association and dissociation phase data were fitted simultaneously to a 1:1 model by using BIAevaluation 3.2. The goodness of fits was determined by examining the 2 relative to R max and the standard errors of the calculated k a and k d values. For IgG samples, the apparent k a and k d values may represent a bivalent interaction. All the experiments were done at 25°C. For epitope determination, WT 2Ig-B7-H3, R127A, and D128A were immobilized onto the chip, and 400 nM ch8H9, MIH42 (Thermo Scientific Pierce), 6A1 (Thermo Scientific Pierce), and clone 185504 (R&D Systems) were injected as described above (contact time, 180 s; dissociate time, 600 s).
Tumor Cell Culture and Antibody-dependent Cell-mediated Cytotoxicity (ADCC) Assay-Neuroblastoma LAN-1 tumor cells were obtained from Children's Hospital of Los Angeles. Cells were cultured in RPMI 1640 medium (Cellgro, Manassas, VA) supplemented with 10% fetal bovine serum (FBS; Life Technologies) at 37°C in a 5% CO 2 -humidified incubator. ADCC assays were performed as described previously (55). Briefly, LAN-1 neuroblastoma tumor cells were radiolabeled with 51 Cr, and peripheral blood mononuclear cells from healthy donors were used as effectors at a 25:1 effector-to-target ratio. Samples were incubated for 4 h at 37°C and 5% CO 2 , and then the cytotoxicity was measured by 51 Cr release.
Biodistribution of Antibody in Xenografted Mice-Female athymic nude mice (6 -8 weeks old) were purchased from Harlan Sprague-Dawley, Inc. All procedures were carried out in accordance with the protocols approved by our Institutional Animal Care and Use Committee and institutional guidelines for the proper and humane use of animals in research. Tumor cells (LAN-1) were harvested, and subcutaneously implanted to the flank of the mice (5 ϫ 10 6 cells/mouse). When the tumor volumes were ϳ200 mm 3 , randomized groups of mice (n ϭ 4 -5/group; one radiolabeled antibody preparation/group) were intravenously injected with 50 Ci of 131 I-radioiodinated antibody (prepared according to the IODO-GEN method (58,59) followed by gel filtration purification using commercial prepacked Sephadex G-25 columns into PBS ϩ 1% BSA; specific activity, 1.67-3.81 mCi/mg; 13-30 g of antibody/dose). The immunoreactivity of each tracer was evaluated using an in vitro cell binding assay using freshly harvested LAN-1. Mice were sacrificed 48 h postinjection, organs were removed, and radio-  . These organs included skin, liver, spleen, kidney, adrenal gland, stomach, small intestine, large intestine, femur, muscle, tumor, heart, lung, spine, and brain. Count rates were background-and decay-corrected, converted to activities using a system calibration factor specific for the isotope, normalized to the administered activity, and expressed as percent injected dose/g. Tumor-to-non-tumor ratios of percent injected dose/g were also calculated.

Results
Crystallization of ch8H9 -The crystal structure of ch8H9 Fab fragment was determined by molecular replacement using Protein Data Bank code 3D85 and refined to 2.5-Å resolution. Collection and refinement statistics are shown in Table 1. The final model was deposited in the Protein Data Bank (access code 5CMA).
The ch8H9 Fab structure has immunoglobulin fold domains common to all Fab structures, and the six complementaritydetermining region (CDR) loops (H1, H2, H3, L1, L2, and L3) that form the antigen recognition site had well defined electron densities (see Fig. 1). The electrostatic surface potential of the antigen binding site was calculated using DelPhi (54) (Fig. 1C). The center of the binding site has a large area of positively charged surface area dominated by the L3 and H3 loops. Two pockets of negatively charged surface area (at the H2 and L1 loop regions) flank the central region, indicating that the antigen recognized by 8H9 has a mixed surface charge distribution.
Humanization of 8H9 -The m8H9/ch8H9 antibody was humanized by CDR grafting methods where the CDR loops of the heavy and light chains of m8H9 were grafted onto human IgG1 frameworks based on their homology to human germline sequences IGHV1-46 for the heavy chain and IGKV-7 for the light chain. The framework sequence was then optimized by in silico mutational analysis of the ch8H9 structure based on CHARMm (Chemistry at Harvard Molecular Mechanics) force fields where the effect of each humanizing mutation was calculated from the difference between the folding free energy of the mutated structure and the wild type protein, a method previously described for our humanization of the anti-GD2 mAb 3F8 (60). The resulting hu8H9 (see Table 2) was then used as a template for affinity maturation.
Affinity Maturation by Yeast Display-A biotinylated B7-H3 construct was used for the selection strategy. For this construct, 4Ig-B7-H3 fused to mouse Fc was expressed in DG44 cells. For biotinylation, NHS-activated biotin was used to react efficiently with primary amino groups (-NH 2 ) in the side chain of lysine residues and the N terminus of each polypeptide. Excess nonreacted biotin-NHS was removed by size exclusion using ultrafiltration columns. The biotinylation of B7-H3-mFc antigen was validated by HRP-conjugated streptavidin using ELISA, and the detection of mouse/human chimeric 8H9 IgG confirmed that biotinylated B7-H3-mFc retained specific binding to 8H9 antibody (data not shown).
To increase the affinity of hu8H9, it was converted to a scFv (VH-VL format) and randomly mutagenized by error-prone PCR. Previously, we successfully used yeast display for further maturation of antibodies because it allows fine discrimination between mutants by flow cytometry (57,61). A mutant yeast library of relatively large (up to 10 8 ) size was generated and subjected first to a preselection by using biotinylated B7-H3-mFc-conjugated magnetic beads, allowing elimination of yeast cells that did not express antibodies or bound weakly to antigen. The mutant library was then sorted several times by FACS for binding to B7-H3 with strin-   (Fig. 2). The sorted scFv fragments were further mutated by error-prone PCR of the entire gene to yield a new sublibrary. The process of sorting and mutagenesis was then cyclically repeated. The highest binding clones to B7-H3 were obtained from the final round of maturation. The Sort 7 yeast showed significantly stronger binding than yeast-displayed parental hu8H9 when they were stained by B7-H3-mFc antigen. The results indicated that the enriched yeast mutants markedly increased the binding affinity. The highest affinity clones from the final round of maturation were identified. Sequence data of strong binders were grouped by cluster analysis, and five repeatedly selected scFv fragments (S7.2, S7.17, S7.22, S7.28, and S7.29) were identified by sequence analysis (Fig. 3).
Binding Properties of Selected scFv Mutants-With 4Ig-B7-H3 antigen coated onto CM5 chips, K D values of antibody binding were compared by surface plasmon resonance using Biacore T-100. A summary for K D values of all selected mutants, including hu8H9 (parental) and S3.3, can be found in Table 3. All scFv fragments from Sort 7 respectively bound to 4IgB7-H3 with dissociation constants (K D ) of 2, 6, 1, 277, and 3 nM compared with a K D of 144 nM for parental hu8H9 and a K D of 10 nM for S3.3. Among them, scFv fragments S7.17, S7.22, and S7.29 exhibited the highest binding affinity by achieving more than a 50-fold affinity improvement. Clone S7.22 in particular resulted in over a 100-fold enhancement compared with the parental hu8H9 scFv sequence.  Determination of Key Affinity-enhancing Mutations- Table  4 lists the six mutations identified from affinity maturation sorted sequences (LC, S20T; LC, H34Y; HC, A24T; HC, E42G; HC, G56D; and HC, A102G). Four of the six mutations (LC, S20T; LC, H34Y; HC, A24T; and HC, E42G) were derived from the third round of sorting (clone S3.3) and subsequently present in most of the clones from Sort 7. The two additional mutations (HC, G56D and HC, A102G) were derived from the highest affinity clone S7.22. Three of the six mutations were in CDR regions (LC, H34Y; HC, G56D; and HC, A102G). Two of the six mutations (LC, S20T and HC, E42G) were found in the human germ line templates used for humanization, thereby increasing humanness.
In Vitro Tumor Killing Properties of 8H9 Antibodies-We then tested in vitro tumor killing ADCC properties of the five IgG antibody constructs (m8H9, ch8H9, ch8H9-6m, hu8H9, and hu8H9-6m) using neuroblastoma LAN-1 tumor cells as targets and human peripheral blood mononuclear cells as effector cells. Cytotoxicity was measured by chromium-51 release, and the results are shown in Fig. 4 and Table 6. m8H9 had no ADCC function with human peripheral blood mononuclear cells, typical for the murine IgG1 isotype. The ch8H9 and hu8H9 constructs exhibited potent ADCC (1.4 and 0.7 g/ml EC 50 , respectively). The ch8H9-6m (0.1 g/ml EC 50 ) had a 14-fold improvement in cytotoxicity over ch8H9, and hu8H9-6m (0.3 g/ml EC 50 ) had a 2-fold enhancement in killing over hu8H9 and a 5-fold enhancement in killing over ch8H9. We postulate that the observed effects were mediated by CD16(ϩ) NK cell engagement and not by immune checkpoint blockade, which would not have been easily observed during the 4-h time course of the assay.
In Vivo Targeting of Human Neuroblastoma Xenografts-We then determined the in vivo biodistribution of our lead preclinical candidates hu8H9-6m and ch8H9-6m compared with m8H9 and ch8H9 using mouse xenografts. All four antibodies were radiolabeled with 131 I, and their in vitro immunoreactivity against LAN-1 cells was determined (m8H9, 69%; ch8H9, 75%; ch8H9-6m, 73%; and hu8H9-6m, 54%). The in vivo biodistributions of each antibody at 48 h postinjection were analyzed using mice bearing subcutaneous neuroblastoma LAN-1 xenografts (Fig. 5 and Table 7). Tumor uptake as measured by percent injected dose/g was 8.9% for m8H9, 6.1% for ch8H9, 6.8% for   ch8H9-6m, and 10.6% for hu8H9-6m). Tumor-to-normal tissue ratios are shown in Table 8. All four antibodies had comparable tumor-to-non-tumor ratios for most tissues. hu8H9 had the highest tumor uptake, which was statistically higher than ch8H9 (p Ͻ 0.05), and a higher tumor-to-blood ratio than ch8H9 (1.26 compared with 0.72, p Ͻ 0.05). ch8H9-6m had a slightly lower tumor uptake but comparable tumor-to-blood ratio. Epitope Determination-To fully understand how mAb 8H9 recognized the B7-H3 target, in silico methods were used to predict the exact molecular epitope, which was then verified experimentally. The crystal structure of murine 2Ig-B7-H3 (Protein Data Bank code 4IOK) was used as a template to create a homology model of human 2Ig-B7-H3 using MODELLER (53). The human 2Ig-B7-H3 was then used for docking the ch8H9 Fab crystal structure using ZDOCK (52). The top docked model (Fig.  6A) predicted that mAb 8H9 binds the IRDF sequence (residues 126 -129) of the V domain of the FG loop in human 2Ig-B7-H3. In 4Ig-B7-H3, this sequence is present twice (residues 126 -129 of the V1 domain and residues 344 -347 of the V2 domain). The predicted interaction energies for each of the 2Ig-B7-H3-interacting residues of the docked complex are shown in Table 9. Arg-127 and Asp-128 had the highest interaction energies (Ϫ55.2 and Ϫ46.3 kcal/mol, respectively). Fig. 6B shows how Arg-127 was predicted to interact with the negatively charged surface on the 8H9 L1 CDR loop and how Asp-128 was predicted to interact with the positively charged surface area at the H3 CDR loop, indicating a high degree of charge complementarity.
Based on the molecular modeling predictions, R127A and D128A point mutations were experimentally made in two separate 2Ig-B7-H3-Fc constructs, recombinantly expressed, and tested for binding by surface plasmon resonance. Fig. 7 shows   the binding kinetics of ch8H9 IgG to the R127A and D128A mutants compared with WT 2Ig-B7-H3 proteins along with the binding of three commercially available anti-B7-H3 mAbs (MIH42, 6A1, and clone 185504). The maximum response units of binding to WT 2Ig-B7-H3-Fc as well as the relative binding of mutants R127A and D128A are shown in Table 10. ch8H9 lost 99.9% of binding to the R127A mutant and 40% binding to the D128A mutant. Antibodies MIH42 and clone 185504 retained ϳ70 -75% of their respective binding to both R127A and D128A mutants. Antibody 6A1 had weak overall binding to the WT 2Ig-B7-H3 and either the same binding to R127A or elevated binding to the D128A mutants. These binding studies demonstrated that ch8H9 was uniquely specific to the IRDF sequence of B7H3. We also verified that the humanized 8H9 constructs lost binding to the R127A and D128A B7-H3 mutants by surface plasmon resonance (data not shown).

Discussion
We have developed high affinity chimeric and humanized anti-B7-H3 antibodies based on in silico modeling and yeast display affinity maturation. Both ch8H9-6m and hu8H9-6m achieved nanomolar to subnanomolar K D affinity to B7-H3, showing high potency in vitro ADCC and high in vivo tumor uptake in mouse xenografts. We have additionally used com-pletely in silico molecular docking methods to predict the binding epitope on B7-H3 to be dependent on the IRDF sequence in the FG loop.
The FG loop plays a critical role in receptor recognition in several B7 family member proteins, including PD-L1 (62), PD-L2 (63), B7-1 (64), and B7-2 (65). Although the receptor of B7-H3 is still not definitively known, a structural study of mouse B7-H3 (33) has shown that the FG loop is responsible for the ability of B7-H3 to inhibit T cell proliferation. Mouse B7-H3 only exists as a 2Ig-B7-H3, and the critical sequence responsible for this inhibitory function is IQDF (residues 126 -129). We have demonstrated that mAb 8H9 binds to the homologous IRDF (residues 126 -129) in human B7-H3 with highest specificity to Arg-127. Additionally, mAb 8H9 and its humanized forms do not bind to the murine B7-H3, which contains Gln-127. Unlike other anti-B7-H3 antibodies tested, only 8H9 can be used to bind the FG loop of B7-H3, which may become relevant for blocking the immune inhibitory functions of human B7-H3. Although we have shown that the binding of mAb 8H9 is FG loop-dependent, we cannot exclude the possibility that other residues may be involved in the binding of this conformational epitope, which is the subject of further investigation. We are also currently testing the ability of mAb 8H9 to inhibit immune checkpoints for cancer immunotherapy.
Unlike any other tumor-associated antigens, B7-H3 is widely expressed among solid tumors and is strongly correlated with recurrence, metastasis, and death. These include prostate can-   cer, non-small cell lung cancer, pancreatic cancer, colorectal cancer, breast cancer, liver cancer, renal cell carcinoma, urothelial cell carcinoma, ovarian cancer, brain tumors, head and neck cancers, melanoma, soft tissue sarcomas, osteosarcoma, neuroblastoma, diffuse intrinsic pontine glioma, mesothelioma, and desmoplastic small round cell tumor. Such prevalent B7-H3 expression is usually associated with highly aggressive clinical behavior. The discovery of B7-H3 on tumor vasculature has further expanded its utility as a cancer immunotherapy target. Clinically proven targets for antibody-based therapy share some general properties. Their expression on the cell surface is usually high but low to absent among normal tissues, thereby achieving optimal therapeutic indices. They should be present on tumor stem cells to suppress tumor regrowth. As radioim-munoconjugates they acquire the ability to destroy antigennegative tumor clones. Alternatively, if the target is on the tumor vasculature, vital supplies to the tumor can be choked off irrespective of whether the antigen is present on all tumor cells. Finally, if the target antigen is critical in immune checkpoints, blocking antibodies have the potential to reverse the negative immune microenvironment. B7-H3 can be exploited for all of these antitumor mechanisms.
We have developed potent chimeric and humanized anti-B7-H3 mAbs that may have clinical potential in Fc␥ receptordependent immunotherapy, immunoconjugates, and checkpoint blockade. This combination presents a unique and powerful opportunity for maximal antitumor effects directed at a single target for most solid tumors.  The relative binding (measured as response units) of the anti-B7-H3 antibodies to the mutants R127A and D128A compared with WT B7-H3 are presented in Table 10.