Discovery of antibodies and cognate surface targets for ovarian cancer by surface profiling

Significance

Target selection for the development of antibody-directed therapies is commonly driven by a specific hypothesis or based on expression profile analysis, which require laborious experimental validation. We developed a pooled screening platform for the unbiased, high-throughput identification of multiple cancer-specific targets in a single screen. Applying this technology resulted in the identification of multiple therapeutic target candidates in ovarian cancer. In addition to the selection of cell type-specific targets, the platform simultaneously allows the discovery of antibodies with therapeutic potential, thereby bypassing the need for lengthy antibody discovery campaigns.

Abstract

Although antibodies targeting specific tumor-expressed antigens are the standard of care for some cancers, the identification of cancer-specific targets amenable to antibody binding has remained a bottleneck in development of new therapeutics. To overcome this challenge, we developed a high-throughput platform that allows for the unbiased, simultaneous discovery of antibodies and targets based on phenotypic binding profiles. Applying this platform to ovarian cancer, we identified a wide diversity of cancer targets including receptor tyrosine kinases, adhesion and migration proteins, proteases and proteins regulating angiogenesis in a single round of screening using genomics, flow cytometry, and mass spectrometry. In particular, we identified BCAM as a promising candidate for targeted therapy in high-grade serous ovarian cancers. More generally, this approach provides a rapid and flexible framework to identify cancer targets and antibodies.

Several therapeutic antibodies are approved for the treatment of specific cancers. Some of these antibodies target proteins essential for the malignant phenotype, such as Human epidermal growth factor receptor 2 (HER2); while others target proteins aberrantly expressed by tumor cells, such as Vascular endothelial growth factor (VEGF). More recently, antibodies, such as those that target Programmed cell death protein 1 (PD1), that modulate immune activity have exhibited dramatic responses in subsets of patients (1, 2). Most efforts to identify new antigens for antibody-based therapies require prior knowledge of targets or arduous validation schemes.

Since antibodies cannot easily penetrate the cell membrane, antibody-based therapeutics rely on their reactivity to cell surface proteins. One major challenge in the development of clinically effective biologics has been off-tumor cytotoxicity, mostly driven by on-target effects mediated by expression of the antibody target on nonmalignant tissues (2). Other challenges include immune escape through target downregulation as well as target heterogeneity within the tumor (2). To overcome these obstacles, methods to facilitate the discovery of alternative cell/tumor type-specific targets with high expression in malignant tissue are essential.

The target selection for the development of antibody-directed therapies has primarily been guided by exploring expression databases for potential cancer or tissue-specific cell surface markers, and by hypothesis-driven approaches based on the study of specific oncogenes. While these strategies have led to the approval of several promising therapies, they have several important limitations. First, the measurement of mRNA as readout for expression has historically been done on bulk samples resulting in loss of information on the cellular distribution of expression within the tumor. Second, gene expression does not always correlate well with protein expression due to translational control mechanisms. Third, gene expression data generally do not provide critical information about the surface localization of many receptors, transporters, and channels, which can be tightly regulated and vary widely depending on the microenvironmental context. Accordingly, the tumor microenvironment can alter the surface abundance of proteins without detectable differences in gene expression profiles (3, 4), which would then be missed by conventional gene expression analysis. Similarly, many surface proteins have been shown to be posttranslationally modified, or to be expressed in cancer-specific protein complexes, which might affect their conformation, rendering them cancer-specific targets despite uniform expression. Given these shortcomings, targets predicted by gene expression approaches require in depth experimental follow-up validation, which can be costly and time consuming.

To address these challenges, we developed the Phenotypic Antibody and Simultaneous Target (PhAST)-discovery platform, an approach utilizing a bacteriophage display-based Single variable domain on a heavy chain (VHH) library to select for antibodies that bind with desired cell surface–binding specificity on live cells followed by mass spectrometric identification of the antibody target. This approach allows the discovery of multiple antibody–target pairs specific to the native or cancer-specific state in a single round of screening. Applying this platform, we identified a set of therapeutic target candidates in ovarian cancer, which we chose because of the high mortality rate and lack of therapies beyond standard chemotherapy and surgery.

Results

To identify targets with cell surface expression phenotypes specific for ovarian cancer, we performed a phage display antibody library screen on live cells. We chose an array of ovarian cancer cell lines across multiple histologies to try to identify not just ovarian cancer-specific, but also possibly subtype-specific targets (5). Antibodies with desired binding properties were then matched to their targets using a proteomic approach. The individual steps of the approach are summarized in Fig. 1.

Fig. 1.

Schematic overview of the PhAST-discovery platform. Enrichment of the input library for binders to ovarian cancer cell lines was performed by 1 round of live cell biopanning with lymphocytes for negative selection and a pool of ovarian cell lines for positive selection. The nonbinders from an additional round of negative selection were then subjected to biopanning against each positive and negative cell line individually. The rescued output libraries were characterized by NGS, and sequences of interest were selected based on differential analysis. Candidates are then reformatted to VHH-hIgG-Fc antibodies, produced in a mammalian expression system, and binding specificity is characterized in a live cell multiplex FACS binding assay. Targets of antibodies with desired binding specificity are identified using an antibody-directed crosslinking and biotin transfer-based protocol on live cells followed by proteomic analysis.

Our goal was to identify antibodies that exhibited specific binding to ovarian cancer cell lines but failed to bind to lymphocytes, fibroblasts, and nonovarian cancer cells. We used a commercially available VHH-bacteriophage display library that was generated from 10 naïve llamas (Abcore) and a library constructed after immunization of two lamas with whole-cell plasma-membrane preparations (Prosci). To deplete the VHH libraries of lymphocyte-specific VHHs, we performed a negative selection using primary peripheral blood mononuclear cells (PBMCs), followed by enrichment of the unbound fraction for ovarian cell-specific binders using a pool of 6 ovarian cancer cell lines (positive cell lines). After rigorous washing, phages were eluted by low pH treatment followed by bacterial amplification. The amplified output library was further depleted of unwanted VHHs by a second round of negative selection on PBMCs. Unbound phages were then subjected to biopanning against each ovarian line, and each negative cell line (PBMCs, an immortalized fibroblast cell line, and a pancreatic cell line) individually. The output libraries were characterized by massively parallel sequencing and compared against reported camelid V-gene and J-gene alleles in the IMGT/GENE-DB. We excluded sequences that were dissimilar to reported alleles or that showed alternations to the expected conserved amino acids. Extracted full-length VHH sequences were translated and clustered across cell lines based on the amino acid sequence similarity of the CDR3s. To select sequences that were specifically enriched in the cell lines of interest, we performed differential analysis between positive and negative cell lines across this set of CDR3 clusters (Fig. 2A).

Fig. 2.

Antibody selection strategy and overview of results. (A) Schematic overview of the NGS analysis and antibody selection pipeline. Each output library is characterized by paired-end MiSeq. Following fragment stitching and several quality control steps, full-length VHH sequences are clustered based on CDR3 sequence homology and clusters subjected to differential analysis to identify sequences enriched in ovarian cancer cell lines over negative control lines. (B) Schematic outline of key steps. Numbers indicate the predicted candidates, antibodies followed up on, or validated in each step. (C) Heatmap showing the flow cytometry binding pattern of validated antibodies tested in a set of ovarian cell lines that was used for selection, lymphocytes and fibroblasts as negative cell lines, and a set of additional nonovarian cancer cell lines. Hierarchical clustering was performed based on antibody binding patterns. Color coding is based on % antibody binding over negative controls: gray, <65%; dark blue, 100%; and dark gray, not analyzed. Only antibodies with >65% positive binding to at least one ovarian cell line are shown. Clusters of antibodies with similar binding patterns are highlighted.

A total of 1,032 clusters showed enrichment in at least one ovarian cell line over the negative samples (Fig. 2B). For follow-up binding analysis, we selected 200 sequences that were enriched in at least one HGSOC ovarian line or that showed high selectivity to one specific cell type. These sequences were synthesized, fused to human IgG1-Fc, expressed in the Expi293 expression system before validating their predicted binding profiles in a FACS-based multiplex binding assay. Of the 200 antibody supernatants tested, only one showed binding to the negative Jurkat lymphocyte cell line (data not shown), while 36 antibodies showed specific binding to at least one ovarian cell line, with a subset weakly also binding to fibroblasts and a pancreatic cell line (Fig. 2C). The remaining antibodies exhibited only weak or absence of binding to any cell line tested, a subset of which could be explained by low antibody abundance in the supernatant as suggested by expression analysis of antibody supernatants by SDS-PAGE Coomassie staining. To further evaluate the selectivity of the specific binders to ovarian cell lines, we tested binding to an additional panel of pancreatic and fibroblast cell lines. As summarized in Fig. 2C, the majority of antibodies showed binding to only a small subset of cell lines, while a small subset bound to a much broader range of cell lines. Four clusters of antibodies shared similar binding patterns (clusters A-D highlighted in Fig. 2C), and other individual antibodies showed distinct binding profiles.

To identify the targets of these antibodies, we prioritized antibodies within the 4 clusters as well as ones with weak or no cross-reactivity to fibroblasts. We used an in vivo biotin transfer-based crosslinking approach (6). Specifically, we incubated an antibody labeled with a trifunctional aminooxy-sulfhydryl-biotin (ASB) crosslinker with oxidized live cells to induce formation of crosslinks with aldehyde-containing glycans on the antibody-bound cell surface protein. Subsequent reduction in the disulfide bond triggered biotin transfer from the antibody to the surface protein. Upon cell lysis, biotinylated proteins were enriched using streptavidin beads followed by mass spectrometry. For data analysis, we quantified relative peptide enrichment against similarly labeled IgG or unrelated antibody controls. Using this method, we identified 10 targets for 19 antibodies. A summary of the antibody–target pairs is listed in Table 1. The targets belong to a diverse set of protein classes, including two receptor tyrosine kinases, five adhesion molecules, two proteases, and one protein reported to regulate angiogenesis.

Table 1.

Summary of identified targets and some associated properties

Protein class	Target	Amplified in ovarian cancer, %	Overexpressed in ovarian cancer	Overexpressed or amplified in 1+ other cancer	Essential in ovarian cancer cells
Receptor tyrosine kinases	HER2	4	+	+++	Yes
	EPHA2	4	+	++	No
Adhesion/migration proteins	ITGA3	3	++	++	Yes
	ITGA6	6	+/−	++	No
	BCAM	5	+++	++	No
	ICAM1	12	++	++	No
	CADM1	5	++	++	No
Proteases	MME	16	+/−	+	No
	ANPEP	5	++	+++	No
Angiogenesis regulating proteins	ENG	2	+	++	No

Cluster A showed particularly high specificity as it selectively bound to only one ovarian cancer cell line (SKOV3) (Fig. 3A). This observation is particularly interesting as SKOV3 is a p53 wild-type cell line with an uncertain histology differing from most of the other lines in our panel (5). Comparison of their CDR3 regions confirmed that these antibodies were indeed distinct molecules (Fig. 3B). Using our proteomics approach, we identified the target of these antibodies as Her2 (Fig. 3C). To confirm the mass spectrometry results, we drove Her2 expression using CRISPRa in OVCAR8, a cell line that this antibody failed to bind. Overexpression of Her2 (Fig. 3 D, Right) resulted in strong antibody binding (Fig. 3 D, Left), while we failed to detect binding in the parental OVCAR8 dCas9 cell line. Conversely siRNA-mediated knockdown of Her2 in SKOV3 cells lead to loss of binding (Fig. 3E) comparable to a no antibody control, while strong binding was apparent in nontargeting siRNA control transfected cells. Next, we tested if the antibodies immunoprecipitated Her2 from SKOV3 whole-cell extracts. We found that three of the four antibodies immunoprecipitated Her2 as assessed by immunoblotting immune complexes using a commercially available anti-Her2 antibody as probe (Fig. 3F). Together, these observations validate Her2 as the target of cluster A antibodies.

Fig. 3.

Identification of a set of antibodies specific to Her2. (A) FACS binding profile of clone 1A12. Flow cytometry staining was performed with 1A12-hIgG1-Fc followed by α-human-APC secondary antibody in a multiplexed format. (B) Alignment of the CDR3 sequences of a set of 5 antibodies with similar FACS binding profiles. (C) Target identification of ASB crosslinked antibodies. Representative mass spectrometry results for 1A68 are plotted against hIgG1-negative control. The total number of peptides/protein is shown. (D) FACS binding profile of CRISPRa induced Her2 expressing OVCAR8 cells (blue) and the parental cell line (gray) stained with 1A12-hIgG1-Fc and α-human-APC antibody (Left). In parallel, cells were lysed and Her2 expression analyzed by immunoblotting (Right). (E) FACS staining with 1A12-hIgG1-Fc of SKOV3 cells with (blue) or without (green) Her2 siRNA transfection (Left). Her2 knockdown efficiency was verified by immunoblotting (Right). (F) Immunoprecipitation with indicated antibodies from SKOV3 whole-cell extracts. Her2 precipitation was analyzed by western blotting with an α-Her2-specific antibody. (G) Western blot of indicated cell lines probed with α-Her2 antibody. Equal loading was verified by probing with α-αTubulin. (H) Binding of serial dilutions of the indicated antibodies to SKOV3 cells was analyzed by flow cytometry. APC mean fluorescent intensity was plotted against antibody dilutions in a nonlinear three parameter fit (Left) and EC50 values calculated (Right) using PRISM software. (I) Antigen-dependent cytotoxicity was evaluated by quantifying antibody-induced apoptosis mediated by PBMCS in SKOV3 target cells by Annexin V-488 staining and subsequent flow cytometry analysis. To distinguish effector and target cells, SKOV3 cells were labeled with CellTrace-Violet prior to the assay. Results represent % of annexin V-positive SKOV3 cells relative to a no antibody control. (J) For internalization studies, antibodies were labeled with pHrodo iFL Red and incubated with cells O/N before fluorescence analysis by flow cytometry. Internalization was quantified by mean fluorescent intensity of PE. Representative results from at least 2 repetitions are shown. Results shown in panel A, C, D, and E are representative of all five identified antibodies.

Immunoblot analysis with an established Her2 antibody showed that protein expression of Her2 was high in SKOV3 cells and weak or undetectable in the other ovarian cell lines we analyzed (Fig. 3G). In agreement, analysis of Her2 gene expression across CCLE cell lines showed high expression of Her2 in SKOV3 cells comparable to high expressing breast cancer cell lines, while expression in other ovarian cell lines was considerably lower (SI Appendix, Fig. S1A), correlating well with the selectivity of cluster A antibodies to Her2. Interestingly, in the Dependency Map portal, an initiative to systematically identify cancer cell line vulnerabilities (7), the SKOV3 cell line exhibited high Her2 dependency, further confirming our findings.

Her2 is a well-known target in the subset of breast cancers that harbor Her2 amplifications as well as metastatic uterine serous and gastric cancer that overexpress Her2 (8). One mode of action of Her2 targeting therapeutic antibodies is their ability to induce antibody-dependent cellular cytotoxicity (ADCC) (9, 10). Accordingly, we tested whether Trastuzumab, an anti-Her2 humanized monoclonal antibody in clinical use, and the cluster A chimeric single-domain antibodies induced ADCC in the context of ovarian cancer cell lines. The affinity to the target and target density on the cell surface are important factors necessary for ADCC (11). Therefore, we first compared the affinity of Trastuzumab and the cluster A antibodies for binding to SKOV3 cells. As shown in Fig. 3H, 4 out of 5 of the antibodies had a low nanomolar affinity between 1 and 10 nM, comparable to Trastuzumab (1 nM). Only 1A68 had a somewhat lower affinity of 42.2 nM. Epitope binning experiments demonstrated that all antibodies had distinct epitopes from Trastuzumab and Pertuzumab (SI Appendix, Fig. S1 C and D). In a FACS-based assay in which PBMCs were used as effector cells and SKOV3 as target cells, Trastuzumab induced potent ADCC (Fig. 3I). We found that all of the cluster A antibodies induced ADCC to comparable levels as Trastuzumab when used in saturating concentrations. To further characterize the antibodies, we tested their internalization properties by measuring the induction of fluorescent signal that is triggered by uptake of pHrodo iFL-labeled antibodies into low-pH endosomes. As shown in Fig. 3J, we detected internalization of all antibodies. Together, these observations suggest that in addition to breast, uterine serous, and gastric cancers, Her2 is an attractive target in Her2 overexpressing ovarian cancers, and that the antibodies we discovered in our screen have features comparable to antibodies in clinical use.

Several antibodies showed intriguing specificity to HGSOC cell lines. Of these, cluster B antibodies showed high binding specificity to two HGSOC cell lines (OVSAHO and Kuramochi) and one clear cell ovarian cell line (OC314) and exhibited low or no binding to control cells (Figs. 2C and 4A). Using the in vivo crosslinking-mass spectrometry method, we identified the target of these antibodies as CADM1 (Fig. 4B). To validate these observations, we knocked down CADM1 in Kuramochi cells and tested antibody binding by flow cytometry. As shown in Fig. 4C, while strong binding of 6N2_38 was detected in control transfected cells, binding was abolished in siRNA transfected cells, confirming the specificity of cluster B antibodies to CADM1.

Fig. 4.

The discovery of antibodies with binding specificity to high-grade ovarian cell lines. (A) FACS binding profile of 6N2_41 representative of cluster B antibodies. Staining was performed with 6N2_41 antibody followed by incubation with APC-conjugated α-human Fc secondary antibody. HGSOC lines are underlined. (B) Mass spectrometry identification of CADM1 representative for cluster B antibodies. Total number of peptides obtained with 6N2_41 vs hIgG control antibody is plotted. (C) FACS binding of 6N2_38 to Kuramochi cells transfected with siRNA targeting CADM1 (blue), control transfected Kuramochi cells (green), and secondary only antibody (gray). (D) FACS binding profile of clone 6N2_22. Flow cytometry staining was performed with 6N2_22-hIgG1-Fc followed by α-human-APC secondary antibody in a multiplexed format. High-grade ovarian cell lines are underlined. (E) Target ID of 6N2_22-hIgG1-F-ASB antibody. Mass spectrometry results for 6N2_22 are plotted in comparison to the hIgG1-negative control. The total number of peptides/protein is shown. (F) FACS staining with 6N2_22-hIgG1-Fc of Kuramochi cells with (blue) or without (green) BCAM transient transfection in 293T cells (Left). BCAM overexpression was verified by western blotting (Right). (G) FACS staining with 6N2_22-hIgG1-Fc of OVSAHO cells with (blue) or without (green) BCAM siRNA transfection (Left). BCAM knockdown efficiency was verified by western blotting (Right). (H) ELISA binding of 6N2_22-hIgG1-Fc dilutions to untreated (blue) or PNGase-treated (red) recombinant human BCAM.

In addition, we discovered 6N2_22, an antibody which showed remarkable specificity to the two HGSOC cell lines Kuramochi and OVSAHO, but did not bind to any other cell line we tested (Fig. 4D). Using our proteomics approach, we identified Basal Cell Adhesion Molecule (BCAM), a transmembrane glycoprotein that acts as a receptor for Laminin a5 (LAMA5) as the target of 6N2_22 (Fig. 4E). When we ectopically expressed human BCAM in 293T cells, we found robust binding of 6N2_22 to BCAM overexpressing cells, while no binding was detected in control transfected cells (Fig. 4F). Conversely, silencing of BCAM in the Kuramochi cell line led to loss of 6N2_22 binding (Fig. 4G), while control siRNA transfection did not affect antibody binding.

To further confirm the specificity of 6N2_22 to BCAM, we measured its binding affinity to recombinant BCAM by ELISA. 6N2_22 showed an affinity of 3.5 nM to recombinant BCAM, compared with an affinity of ~7 nM on Kuramochi cells (Fig. 4H and SI Appendix, Fig. S2A). Since BCAM is known to be heavily glycosylated (12), we tested whether 6N2_22 binding depends on BCAM glycosylation status. Deglycosylation with PNGase resulted in a ~20-kDa shift in BCAM migration on a Coomassie gel indicating successful deglycosylation (SI Appendix, Fig. S2B). ELISA showed that the binding affinity of 6N2_22 was unaffected by BCAM glycosylation status, indicating that the antibody recognizes BCAM irrespective of glycosylation (Fig. 4H).

To further characterize the 6N2_22 antibody, we mapped the epitope for 6N2_22. BCAM belongs to the immunoglobulin superfamily, and the extracellular region is composed of five immunoglobulin like domains (V1-2, C1-3) (13, 14). To narrow down the region necessary for antibody binding, we constructed chimeras between BCAM and MCAM, a closely related protein with similar Ig-like domain architecture that 6N2_22 does not bind to (15, 16) (Fig. 5A). Chimeras were tested for antibody binding by flow cytometry of 293T cells transiently transfected with the respective constructs. As expected, the antibody did not bind to MCAM, and swapping domains V1 and V2 to those of BCAM had no effect on the ability of the antibody to bind (Fig. 5A). The additional exchange of domain C1 however resulted in antibody binding comparable to that of full-length BCAM, indicating that the epitope is located within domains V1, V2 and C1. Conversely, swapping BCAMs V1 domain with V1 of MCAM resulted in binding comparable to full-length BCAM, suggesting that this region is dispensable for binding. BCAM binding was lost when V2 was replaced with the respective MCAM domain. Further replacement failed to restore the ability of these antibodies to bind BCAM. Together these studies demonstrate that the epitope is located on V2 and C1 of BCAM. To further map the amino acids involved in binding, we performed mutagenesis within this region, replacing structure-predicted surface-exposed charged residues with alanine (13, 14). Expression of all constructs were comparable (Fig. 5 B, Lower). Although most mutations had no effect on binding, mutation of aspartic acids 310 and 312 to alanine both abolished 6N2_22 binding to BCAM, indicating that these residues are an essential part of the binding interface (Fig. 5 B, Upper). Several polymorphisms in the extracellular domain of BCAM have been reported (17). We therefore tested 6N2_22 binding to the most common polymorphism and found that 6N2_22 binding to all mutants was comparable to wild-type BCAM, except for Lu12, which was poorly expressed (SI Appendix, Fig. S4).

Fig. 5.

Characterization of α-BCAM 6N2_22. (A) Epitope mapping of 6N2_22 was performed in 293T cells transiently transfected with indicated BCAM/MCAM chimeras followed by flow cytometry. Binding was quantified as % APC-positive cells compared with secondary only antibody staining. Results from one representative experiment are shown. (B) 6N2_22 antibody binding to indicated Myc-tagged BCAM point mutants was tested by flow cytometry of transfected 293T cells (Top). Expression was verified by immunoblotting using α-Myc antibody. Results from one representative experiment are shown. (C) Antigen-dependent cytotoxicity was evaluated by quantifying antibody-induced apoptosis mediated by PBMCs in CellTrace Violet-labeled parental or BCAM ko Kuramochi target cells by Annexin V-488 staining and subsequent flow cytometry analysis. Results represent % of Annexin V/Violet-positive cells relative to a no antibody control. Avarage +/− SD is shown.

We then assessed whether 6N2_22 mediates killing of BCAM overexpressing cells. Specifically, we performed an ADCC assay, using PBMCs as effectors and CSFE labeled Kuramochi cells as target cells. Upon incubation with 6N2_22 or control antibody, cells were stained with Annexin V-488 and analyzed by flow cytometry. As shown in Fig. 5C, 6N2_22 potently induced ADCC of Kuramochi cells in a dose-dependent manner. The activity was dependent on BCAM expression as the antibody lost its ability to induce ADCC of BCAM KO cells.

To assess BCAM as a potential target for ovarian cancer, we first performed immunoblot analysis with a commercially available BCAM antibody in ovarian cancer cell lines and patient-derived organoids. As shown in Fig. 6A, high protein expression was detected in Kuramochi and OVSAHO cell lines, while expression was considerably weaker in the other ovarian cell lines analyzed, correlating well with the 6N2_22 cell line binding pattern. In agreement, gene expression data derived from CCLE showed Kuramochi and OVSAHO cell lines among the highest BCAM-expressing ovarian cell lines (Fig. 6B). Given the selectivity of the antibody to two high-grade cell lines, we tested whether high BCAM expression is enriched in HGSOC over other ovarian subtypes. Indeed, comparing BCAM expression across CCLE ovarian cell lines showed highly significant enrichment for high BCAM expression in HGSOC cell lines. We next evaluated patient-derived ovarian cancer-derived organoids. As shown in Fig. 6C, 8 of the 9 organoids analyzed showed strong 6N2_22 staining in at least 65% of cells. Together, these observations demonstrate that BCAM is preferentially expressed in HGSOC cell lines and on a large fraction of organoids derived from ovarian cancer patients.

Fig. 6.

BCAM is a potential therapeutic target in HGSOC. (A) Immunoblot analysis of expression of BCAM in indicated ovarian cell lines. Equal loading was verified with probing for α-Tubulin. (B) Box plot comparing BCAM gene expression between HGSOC (blue) and other ovarian CCLE cell lines (gray). Kuramochi and OVSAHO (6N2_22 binders) are highlighted in red. The statistical significance was tested using a Kruskal–Wallis test P = 6.7 x 10⁻⁵. (C) Representative facs histogram of ovarian cancer-derived organoids stained for BCAM (blue) or IgG control (gray) (Left). BCAM expression is shown as % positive cells compared with IgG control (Right). (D) Representative images of IHC staining of ovarian tumor tissue microarrays. Microarrays were stained with α-BCAM and fluorescently labeled α-rabbit secondary antibody (green) and counterstained with DAPI (blue). Top 2 panels represent cores from HGSOC, bottom left mucinous adenocarcinoma, bottom right endometrioid adenocarcinoma. (E) Quantification of BCAM expression from tissue microarray of 36 HGSOV and 33 other ovarian subtypes. Statistical significance was tested using unpaired t test P < 0.0001.

To examine BCAM levels on primary tumors, we analyzed ovarian tissue microarrays for BCAM. The microarray included 36 HGSOC cores and 33 cores from other ovarian cancer subtypes. Strong staining was detected on the surface of HGSOC tumor cells, while no or weak BCAM expression was observed on adjacent stromal cells or tumor cores from other tumor subtypes (Fig. 6D). The cell staining pattern suggested that BCAM is expressed on the cell surface. When we compared the percentage of tumor cells expressing BCAM (positivity score) between HGSOC and other cancer subtypes, we found a highly significant enrichment for BCAM expression on HGSOC (Fig. 6E). Interestingly when we costained these samples with laminin5a (LAMA5), BCAM primary ligand (18, 19), we found no correlation between BCAM and Laminin 5 expression neither on tumor cells nor stroma (SI Appendix, Fig. S5C). Of note, BCAM expression was weak or absent in healthy ovarian, kidney, and thyroid tissues (SI Appendix, Fig. S5D). BCAM has been reported to be expressed on red blood cells (RBCs), in particular RBCs in patients with Sickle cell anemia (20, 21). To assess the level of BCAM expression on healthy RBCs, we performed flow cytometry binding studies with 6N2_22 on whole blood. As illustrated in SI Appendix, Fig. S6, we did not detect appreciable binding of 6N2_22 to RBCs, while expression of the RBC marker CD235 and of CD47, a surface protein well known to be expressed at high levels on RBC were readily detectable. We obtained similar results with a commercial BCAM antibody and confirmed the low levels of expression of BCAM on RBCs by immunoblotting (SI Appendix, Fig. S6, Right). Together these data suggest BCAM as an attractive therapeutic target in a subset of HGSOC.

Discussion

Although antibodies are a well-established therapeutic modality, target nomination remains a bottleneck in the development of targeted cancer therapeutics, in part due to the requirement for intensive research to identify potential candidates that then need to undergo rigorous validation to confirm their suitability as cancer targets. Using a phenotypic library screening approach, we report a platform that overcomes some of these challenges and identified numerous physiologically relevant ovarian cancer-specific surface targets in an unbiased manner. Selecting for antibodies with desired cell line binding profiles allowed us to biochemically identify several highly cell type-specific targets bypassing the need for extensive target validation.

In this study, we used single-domain antibody libraries, but the platform is highly versatile and with some adjustments can readily be adopted to other formats, such as scFv-phage display libraries. We used two antibody library sources, a library derived from 2 llamas immunized with ovarian cancer membrane preparations, and a naïve VHH library. Although both libraries resulted in the identification of a number of surface targets, the identity of the targets differed. The low diversity (~10⁴) immunized library likely shows a bias toward highly expressed and highly immunogenic targets (such as Her2), while the more diverse naïve library (10⁷) has the potential to identify targets with varying expression patterns, and might thus be better suited for this approach. Although we performed follow-up studies on the limited number of 200 candidates, we anticipate that deeper candidate mining and the use of more diverse libraries would lead to the identification of many additional targets.

While the target identification method we applied proved to be successful for about 75% of antibodies, we were unfortunately unable to identify the target of a few antibodies with interesting binding patterns, including one HGSOC-specific antibody cluster (cluster C). Alternative genome-based approaches such as CRISPR knockout or CRISPR activation screens could be considered for these challenging to identify antibody targets.

Anti-Her2 antibodies have been successfully used in treatment of Her2-overexpressing breast, uterine serous, and metastatic gastric cancers (8, 22). The discovery of a set of Her2 antibodies in the context of ovarian cell lines suggests Her2 should be evaluated as a target in a subset of ovarian cancers. Indeed, analysis of CCLE data shows high expression of Her2 in some ovarian cancer cell lines of which a small subset is highly Her2 dependent; for example, the ovarian cell line SKOV3 shows levels of Her2 overexpression and dependency similar to the SKBR3 breast cancer cell line (SI Appendix, Fig. S1A). Furthermore, HER2 overexpression/amplification has been reported in ovarian cancer, especially in clear cell and mucinous tumors (23). Her2 targeting has been explored in a limited number of clinical trials in ovarian cancer using Trastuzumab or Pertuzumab. These trials showed limited clinical benefit in ovarian cancer patients, particularly when compared with combination chemotherapy (24–26). However, in breast cancer, single agent Her2 therapy showed a similar limited response. Combining anti-Her2 therapy with chemotherapy or as Her2-conjugated antibodies led to vastly improved responses, suggesting that a similar approach might be worth considering for future ovarian cancer trials.

HGSOC is the most common and lethal subtype of ovarian cancers, with the vast majority of women diagnosed at an advanced stage of disease. The current standard treatment is surgical debulking combined with chemotherapy. While standard therapy induces an initial response, tumors ultimately recur, and 70% of patients die within 5 y of diagnosis (27). To achieve better outcomes, new therapeutic targets are needed. Our screen led to the identification of CADM1 and BCAM, two adhesion proteins that are highly expressed on the surface of HGSOC cell lines. Previous studies have demonstrated that CADM1 can act as tumor suppressor and is frequently inactivated by promoter hypermethylation in many solid tumors, including pancreatic, lung, melanoma, esophageal, and cervical cancer (28–33). However, CADM1 has also been reported to be overexpressed and protumorigenic in T cell leukemia, lymphoma, and small cell lung cancer (34, 35). Ectopic expression of CADM1 has been suggested to inhibit cell proliferation and migration in an endometrial ovarian cancer cell line model (36). The high endogenous expression of CADM1 in the two HGSOC cell lines in this study does not appear to support these findings and suggests a more complex, possibly context-dependent function. The physiological roles of CADM1 in ovarian cancer will need further investigation to determine its suitability for targeted therapy of HGSOC.

BCAM, first shown to be highly expressed on sickle RBCs (37, 38), is overexpressed in a number of tumors (35, 36), notably highest in HGSOC, while its expression appears relatively low in normal tissues, with moderate expression in the kidney and the thyroid (SI Appendix, Fig. S4 A and B). We found that BCAM shows high expression in about 35 to 40% of primary HGSOC tumors and low to undetectable levels in the kidney and thyroid (SI Appendix, Fig. S4C). BCAM has previously been suggested as an ovarian-specific target (21). We found that BCAM expression is very low or undetectable on the surface of healthy RBCs, alleviating potential off-tumor toxicity concerns. A recurrent BCAM–AKT2 fusion has also been described in some HGSOC (39). Although we did not find evidence for this fusion in our cell line and organoid models, targeting the BCAM extracellular domain would allow the elimination of BCAM-expressing cells irrespective of their fusion status.

BCAM is a transmembrane glycoprotein with 5 immunoglobulin-like domains that acts as a receptor for LAMA5 (18, 19). Their interaction was demonstrated to promote adhesion and migration of carcinoma cells (18, 19). Accordingly, inhibition of BCAM-LAMA interaction has been demonstrated to have an inhibitory effect on migration (40, 41). Surprisingly, we observed no correlation between Laminin 5 and BCAM expression in HGSOC in tissue microarrays (SI Appendix, Fig. S4D), raising the question as to whether laminin 5 is the primary ligand for BCAM in ovarian cancer. In agreement with this, although our mutagenesis data suggest that the 6N2_22 epitope at least partially overlaps with the LAMA5-binding region, it does not have an apparent effect on cell adhesion (SI Appendix, Fig. S3B). A previous study described an α-BCAM antibody-drug conjugate that induced cancer cell killing (42). We did not observe evidence that 6N2_22 triggers BCAM internalization in 2 cell lines tested (SI Appendix, Fig. S3A), but it showed potent ADCC activity in vitro, underscoring its potential use as therapeutic antibody in the VHH-hIgG1-Fc format. Further in vivo studies, which require antibody engineering and suitable cancer models, are necessary to assess the consequences of targeting BCAM with suitable antibodies.

Together, using an unbiased approach focused on identifying antibodies with specific binding patterns, we identified a number of surprising surface proteins as candidates for targeted therapy against subsets of ovarian cancer. Applying similar screens to other cancers, with a focus on individual cancer subtypes, or in the context of different microenvironmental conditions will likely lead to the discovery of many new highly specific targets that will accelerate the design of innovative new single or combination cancer therapies.

Materials and Methods

Cell Culture and Cell Line Generation.

293T, A549, IMR, JIMTI, KP4, MIAPACA, PANC1, PATU8902, and PATU 8988T cell lines were cultured in DMEM media (Life Technologies). The JHOC5 cell line was cultured in DMEMF12 (Life Technologies). RMUGS cells were cultured in HAM’s F12 (Fischer Scientific). SKBR3, SKOV3, and HT29 lines were cultured in McCoy’s 5A media (Life Technologies). ASPC1, BXPC3, ES2, HCC1395, HCC202, Jurkat, Kuramochi, OC314, OVCAR8, OVSAHO, and PANC1005 lines were all cultured in RPMI (Life Technologies). All cells were grown at 37 °C and 5% CO2 and supplemented with 10% fetal bovine serum (FBS) and 1% penicillin streptomycin (Life Technologies). Organoid cultures were grown as previously described (43).

Expression Vectors and Cloning.

For expression of VHH-hFc chimeric antibodies pcDNA3 was modified to carry the IgGκSP (METDTLLLWVLLLWVPGSTG) signal peptide for antibody secretion and human IgG1-Fc by Gibson cloning. VHH sequences were synthesized and cloned in-frame into the modified vector using AgeI/EcoRI restriction sites (Genscript). Vectors for CRISPR ko and CRISPRa have been obtained from the Genomics Perturbations Platform at the Broad Institute of Harvard and MIT (Cambridge, MA). ORF expression vectors for human and mouse BCAM, MCAM, and Her2 were obtained from Origene and sequence verified. Chimeras and point mutants were generated by overlapping PCR and Gibson cloning into EcoRI/XhoI cut pcDNA3.

Transfections and Lentiviral Transduction.

For overexpression experiments, 293T cells were transfected using Lipofectamine P3000 according to manufacturer instructions. Cells were analyzed by western blotting or Flow cytometry 2 to 3 d post transfection. For siRNA transfection, lipofectamine RNAiMAX (Life Technologies) was used according to manufacturer instructions. Typically 10 μM siRNA was transfected and knockdown validated by immunoblot 2 to 3 d post transfection.

For lentiviral production, virus was produced by cotransfecting 293T cells with the lentiviral vector, D8.9 packaging construct, and VSV-G using Lipofectamine P3000 reagent (Life Technologies) according to manufacturer protocol. Media were changed the following day, and virus harvested 2 d post transfection. After filtration through a 0.45-μM syringe filter (Fisher Scientific), cell lines were infected in the presence of polybrene (Santa Cruz). Media were changed 24 h post infection and selection with puromycin (Fisher Scientific) or blasticidin (Fisher Scientific) was started 2 days post infection.

Immunoblotting.

Confluent plates of cells were harvested, washed with PBS, and lysed with cold RIPA buffer (Sigma Aldrich) with protease inhibitor and phosphatase inhibitor tablets (Sigma Aldrich). Lysates were cleared by centrifugation and protein quantified using the Thermo Fisher BCA Protein Assay protocol. Equal amounts of proteins were prepared in SDS loading buffer supplemented with β-mercaptoethanol, boiled at 95 °C to denature proteins, loaded onto precast 4 to 12% Bis–Tris gels (Life Technologies), and subjected to electrophoresis at 100 V. They were then transferred to PVDF membranes (Life Technologies) with the iBlot2 Transfer System for 7 min or 1h wet transfer at 100V. Membranes were blocked in Intercept Blocking Buffer (LICOR Biosciences) followed by incubation with indicated primary and IRDye-labeled secondary antibodies (LICOR Biosciences). Bands were visualized with the Odyssey® Imaging Systems. Primary antibodies used were as follows: α-Her-2 (Cell Signaling), α-BCAM (R&D), α-α-Tubulin (Abcam), α-Myc-tag (Upstate).

Membrane Preparation of Immunization.

Membrane preparations were prepared as described (44). Briefly, 1 × 10⁷ of each ovarian cancer cell line (SKOV3, OVCAR8, Kuramochi, OVSAHO, ES2, OC314, RMUGS) were resuspend in 5 mL cold 25 mM Tris-HCl, pH 7.4, 320 mM sucrose, and protease inhibitor, lysed by sonication followed by spinning at 1,000 g for 12 min 4 °C to remove unbroken cell nuclei and cell debris. The supernatant was centrifuged at 40.000 rpm for 30 min; the pellet was resuspended in 50 mM Tris-HCl, pH 7.4, and again centrifuged at 40.000 rpm for 20 min. The pellet was resuspended in 1 mL 0.02 M bicarbonate buffer (pH 9.6) and passed through a 27-G needle to homogenize membrane fraction. Equal amounts of membrane fractions from each cell line were pooled and used for immunization. Llama immunizations and library preparation were performed by Prosci.

Biopanning and NGS.

For negative selection, PBMCs were isolated from blood collars by Ficol gradient centrifugation. Blood collars were purchased from the Brigham and Women’s Hospital Crimson core, who consents donors and ensures that they are deidentified. The phage display library was incubated with PBMCs for 1 h on ice. After centrifugation, the supernatant was transferred to the harvested positive cell lines and incubated for 2 to 4 h gentle mixing. Cells were washed with PBS/5%BSA/0.5% Tween followed by elution of bound phages with 0.1M Glycin-HCl pH 2.6 and neutralization with Tris-base. Output library was rescued in TG1 cells and amplified. For the second round of negative selection, the new sublibrary was incubated with PBMCs followed by incubation with fibroblasts. Supernatant was added to individual positive cell lines and incubated for 2 to 4 h, followed by washing and elution in Glycin–HCl. Eluted phages were rescued in TG1 cells by culturing O/N at 30 °C in presence of ampicillin and glucose. For NGS phagemids from each output library were isolated using a plasmid midiprep kit (Qiagen) followed by restriction digest with AgeI/SfiI to isolate VHH fragments. Illumina paired-end 2 × 250-bp sequencing was performed on targeted VHH sequences.

NGS Analysis.

Trimmomatic (version 0.38) was first used to remove fragments with low base calling quality (average Phred score < 30) and clip Illumina adapter sequences from all reads (45). Reads were additionally cropped at 225 bp to remove low-quality positions. Quality passing paired reads were merged using FLASh (version 1.2.11) with expected fragment length and SD set to 375 bp and 35 bp, respectively (46). Merged reads were filtered to only those which appeared to be valid VHH sequences based on expected heavy chain structure. Reference sequences for the camelid heavy chain framework regions (FR) were obtained from IGHV and IGHJ alleles of the closely related Vicugna pacos in the IMGT/V-QUEST reference directory set (release 201908-4) (47). Position weight matrices were constructed for each of the four FRs based on the reference alleles. Full-length VHH sequences were trimmed and translated to amino acid sequences. Amino acid (AA) sequences for the complementary determining region 3 (CDR3) were extracted from reads based on the previously matched FR3 and FR4 positions. CDR3 sequences shorter than 2 AAs were dropped. Unique CDR3 sequences were clustered across all samples using CD-HIT (version 4.8.1) (48, 49). CDR3 sequences were sorted by total fragment counts prior to clustering with CD-HIT. For each CDR3 cluster, we counted the number of fragments matching a CDR3 sequence in the cluster for each sample. The matrix of sample fragment counts across CDR3 clusters was next used for differential analysis (50). Selection of sequences for follow-up was based on overlap between replicates, the number of cell lines we identified with priority given to sequences found in at least 2 ovarian cell lines while absent in PBMCs, and the highest single cell line hits.

Target Identification.

Target identification was essentially done as described (6). Briefly, to prepare the ASB crosslinked antibody 100 μg purified antibody were incubated with PEG₄-SPDP (Thermo Fischer Scientific) at room temperature for 1 h followed by quenching with glycine (Santa Cruz Biotechnology). The antibody was then incubated overnight with 60 μg reduced ASB in 1 mM EDTA (Life Technologies). Antibodies were buffer exchanged with PBS in Amicon filters (Thermo Fisher Scientific). To confirm crosslinking, 1 μg of sample was run on a gel and Coomassie stained in parallel with unlabeled purified antibody. Successful crosslinking indicated an upshifted band in the labeled sample. Approximately 10⁸ cells were harvested and suspended in 2 mM sodium meta-periodate (Sigma-Aldrich) in PBS pH 6.5, followed by 4 °C incubation. Cells were incubated with 100 μg ASB-labeled antibody followed by addition of 10 mM p-phenylenediamine (Sigma-Aldrich) to catalyze crosslinking. After washing, cells were flash frozen in a dry ice and ethanol bath and stored at −80 °C. The cell pellet was lysed in 2% sodium dodecyl sulfate (Sigma–Aldrich) with protease inhibitor (Sigma–Aldrich) and benzonase (Santa Cruz Biotechnology), and cell clumps were dissociated by passing through a syringe needle (Sigma–Aldrich 22 gauge, L 1 in). Then, 50 mM DTT (Sigma-Aldrich) was added to cleared lysates and boiled to cleave biotin crosslinks. Cooled samples were treated with 375 mM IAA (Sigma-Aldrich) in 50 mM ammonium bicarbonate (Westnet Inc) in the dark, and subsequently quenched with 200 mM DTT. Biotinylated proteins were isolated from sample by incubating with Streptavidin magnetic beads (Life Technologies) followed by multiple washes with 0.5% SDS, 2M urea (Life Technologies), and 50 mM AMBIC. The samples were finally resuspended in 50 mM AMBIC and stored at 4 °C until mass spec analysis. Mass spec analysis was performed after on bead Trypsin digestion on a LTQ Orbitrap Velos Pro ion-trap mass spectrometer (Thermo Fisher) at the Harvard Medical Schools Taplin Core facility as described previously (51, 52).

Antibody Expression.

Expi293F cells (Life Technologies) were grown with Expi293 Expression medium (Fisher Scientific) at 120 rpm at 37 °C with ≥80% relative humidity and 8% CO2. Cells at a density between 3 to 5 × 10⁶ viable cells/mL were transfected with 2.5 μg of desired antibody expression plasmid with the ExpiFectamine 293 transfection kit reagent (Life Technologies) according to manufacturer instructions. Cell supernatants containing the antibodies were harvested 3 d later by centrifugation. Expression of antibodies was verified by SDS–PAGE/Coomassie staining.

Antibody Purification.

For antibody purification, crude antibody supernatant was incubated with Protein A Plus agarose beads (Pierce) for 2 h followed by washing with PBS. Beads were collected by centrifugation at 1,000 rpm and antibodies eluted in 4× bead volumes of Elution Buffer at pH2.0 (Pierce). The eluate was neutralized with 3 M Tris base. Antibodies were run on SDS-PAGE to confirm size and purity.

Flow Cytometry.

For antibody binding studies, cells were washed with PBS and harvested using CellStripper (Thermo Fisher). Antibody staining was performed in PBS/5%BSA with indicated amounts of primary antibody. Additional staining with corresponding fluorescently labeled secondary antibody was performed when primary antibody was not directly conjugated. All wash steps were performed in PBS/5%BSA. To assess cell viability, cells were stained with Fixable Viability Dye eFluor 780 (Biolegend). For the multiplexed screening assay, each cell line was labeled individually with different concentrations of CellTrace CSFE and/or CellTraceViolet. Briefly, 1.2 x 10⁶ cells per cell line were labeled with Violet and CFSE dye combinations (CellTrace Violet 40 μM, Violet 5 μM, CFSE 10 μM, and CFSE 1 μM) for 30 min. The reaction was quenched with FBS. For antibody screening, labeled cell lines were mixed and aliquoted to 3 x 10⁵ total cells/well. All antibody staining reactions were performed on ice, protected from light, and wash steps were performed with PBS/5%BSA. Cells were incubated with 20 μg/mL antibody in 100 μL PBS/5%BSA for 30 min, followed by incubation with APC-conjugated α-human IgG Fc (BioLegend) at 5 μg/mL in 100 μL PBS/5%BSA for 30 min. Fixable Viability Dye eFluor 780 (eBioscience) at 1000x dilution was added for 15 min. For antibody competition assays, antibodies were labeled with Alexa Fluor 488 according to manufacturer instructions (Thermo Fisher). For staining 3 x 10⁵ cells were aliquoted into each well. For blocking, unlabeled antibodies were added to the samples at saturating amounts (20 μg/mL) for 30 min on ice. Control cells were incubated with no antibody or an unrelated antibody. After washing, 488-labeled antibodies were added for 30 min to blocked and control samples. Upon washing, cells were resuspended in PBS/5%BSA for flow analysis. Facs analysis was performed on the BD Fortessa. All data were analyzed using FlowJo software. Commercial primary antibodies used were as follows: α-hERBB2-APC (FAB1129A, R&D), α-human-APC (HP6017, Biolegend), α-human-488 (A10631, Invitrogen); Trastuzumab and Pertuzumab were purchased from the Dana-Farber Cancer Institute Pharmacy.

ADCC Assay.

Effector PBMCs were isolated from buffy coats using Percoll (Sigma) density gradient centrifugation and stimulated with 100 ng/mL IL-2 overnight. Target cells were stained with CellTrace Violet (Life Technologies) according to manufacturer instructions. A total of 10⁴ violet-stained cells were seeded into round bottom 96-well plates in RPMI/5% human AB serum (Sigma). Indicated amounts of antibodies were incubated at 37 °C/5% CO₂ for 30 min. 2.5 ×10⁵ PBMCs were added and incubated at 37 °C/5% CO2 for 4 h. The media were replaced with 1:20 Annexin V-488 (Life Technologies) diluted in Annexin V buffer (Life Technologies) and incubated at room temperature for 30 min. The samples were adjusted to 200 μL before being assessed on BD Fortessa II Cytometer and analyzed on FlowJo™.

BCAM ELISA.

First, 20 μg/mL recombinant BCAM (Sino) was coated onto 96-well high-attachment plates and incubated at 4 °C O/N. Then, the plate was washed two times with PBS (Thermo Fisher) and blocked with PBS/10% bovine serum albumin (BSA) for 1 h at room temperature. Primary antibodies in PBS/10% BSA were added in 1:4 dilutions from 10 μg/mL to 0.01 μg/mL and incubated for 1 h at room temperature. The plate was washed three times with PBS/0.1%Tween before adding the secondary anti-human-HRP (Cell Technologies) in a 1:1,000 ratio diluted in PBS/10%BSA and incubated for 30 min at room temperature. The samples were washed three times with PBS/0.1%Tween, and TMB substrate (Pierce) was added to the wells for 15 min at room temperature. Then, 2 M sulfuric acid was added directly to the TMB substrate to stop the reaction. Absorbance was recorded at 450 nm on SpectraMax M5E (Molecular Devices).

Immunohistochemistry.

Tissue microarrays (BC11012b, OV812) were purchased from US Biomax. Fluorescence immunohistochemistry staining was performed on 5-µm sections of the TMAs for detection of BCAM on the automated Leica Bond RX system, with α-BCAM antibody PA5-50376 (Thermo Fisher), at dilution of 1:50 for 15 min, a tyramide-conjugated FITC fluorophore and a DAPI counterstain. For spectral imaging & analysis, the stained slides were scanned on a Vectra 3 imaging system (Akoya Bio) and analyzed using Halo (Indica Labs). We run an algorithm learning tool utilizing the Halo training for the epithelial cells and stroma regions, and subsequently completed cell segmentation. The reported expression intensity for a cell is the average intensity of the total pixel values in the antigen presented subcomponent of the cell. The threshold for BCAM was set based on the staining intensity normalized by the background slide. Cells with the intensity above the setting threshold were defined as positive.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Raw NGS sequencing data have been deposited in National Library of Medicine BioProject. Accession: PRJNA910291. All study data are included in the article and/or SI Appendix.