A Multipronged Unbiased Strategy Guides the Development of an Anti-EGFR/EPHA2–Bispecific Antibody for Combination Cancer Therapy

A critical limitation of both traditional and emerging single-agent targeting cancer therapies is their common tendency to leave behind resistant cancer cell populations. This often happens because of the innate heterogeneity of molecular mechanisms supporting tumor cells, which allows survival and eventually, selective evolution of treatment-refractory and more aggressive clones. Accumulation of these resistant cells practically unavoidably concludes in tumor relapse and patient lethality (1). Therefore, combination therapies with their individual components targeting distinct cancer-related molecular mechanisms in a precise coordination are urgently required to improve therapeutic effects. Currently, the systematic design of optimal combinatorial approaches is only beginning to be explored and is not yet commonly used (1–3). To assist in addressing this challenge, we describe here an unbiased multifaceted strategy for selecting effective target combinations for novel dual-specificity reagents for cancer treatment.

The proposed strategy is based on the application of genome-wide loss of function ex vivo screening (4) for genetic dependencies acquired by xenograft tumor cells in response to a single-target reagent of choice. This is complemented by the profiling of the target's interactome using BioID proximity proteomics (5). Bioinformatics analyses–based overlapping the screening and interactome datasets allow to predict most promising target combinations (Fig. 1A). The relevance of the predicted target combinations is further assessed by gene expression analyses of patient data from the Cancer Genome Atlas (TCGA) database to select candidate co-targets frequently co-expressed in tumors. Finally, the anticancer efficiency of suppressing the selected target combinations is examined in cell culture and xenograft models (Fig. 1A).

The EPHA2 receptor tyrosine kinase (RTK) often supports tumor aggressiveness, is considered a promising target for cancer therapy (6) and is prioritized by NIH as a cancer-related antigen (7). Multiple attempts have been made to interfere with EPHA2 activity in cancer cells and although some of these approaches hold promise (6), none of the attempts have yet produced a clinically relevant therapeutic reagent. Therefore, to test the applicability of our approach, we focused our efforts on the EPHA2 receptor.

To implement our strategy, we generated a novel synthetic anti-EPHA2 antibody, used it to treat xenograft tumors and generated ex vivo cell populations from the treated and matching control xenografts. Application of our multipronged unbiased approach in this model guided us with a high level of confidence toward choosing the EGF receptor (EGFR) for co-targeting with EPHA2. Following this lead and to assess the accuracy of the selection, we used a therapeutic anti-EGFR antibody, cetuximab, as a prototype to design a human bispecific anti-EGFR/EPHA2 antibody, which as predicted, produced strong antitumor effects in xenograft models.

Taken together, our work reported here outlined and successfully validated an efficacious unbiased strategy that can be used for identifying optimal candidate targets for designing novel, highly potent bi-specific therapeutics.

Details are provided in Supplementary Materials and Methods. All animal work protocols were approved by the Animal Research Ethics Board at the University of Saskatchewan. All in vivo experiments were carried out in accordance with established guidelines and regulations.

EPHA2-specific Fab fragments selected from antibody libraries were re-cloned into the pET22a plasmid and expressed in the periplasm of BLgold. After growing bacteria to OD₆₀₀ = 0.5, Fab expression was induced by addition of 1 mmol/L IPTG and incubation of culture at 28°C for 15 hours. Fabs were expressed with heavy chain C-terminal c-Myc and His6 tags (CH-AAA-c-Myc-GAALE-His6) for assisting purification and detection. Fab fragments were purified using Ni⁺-resin (Profinity IMAC Resin, Ni-charged, Bio-Rad, 1560135).

IgGs were expressed in CHO-S culture grown in Dynamis medium (Thermo Fisher Scientific, A2661501) following transfection with the FectoPro reagent (Polyplus Transfection, 116–010). The antibodies were harvested on days 12–14 after transfection and purified using MabSelect resin (GE Healthcare, 17–5199–01). After binding in PBS, the antibodies were eluted with 100 mmol/L citrate buffer (pH 3.0) followed by immediate titration to pH = 5.5 using 1 mol/L Tris-HCl, pH = 8.0. The eluted protein was dialyzed against and stored in the formulation buffer composed of 20 mmol/L l-Histidine, 30 mmol/L Citric acid, 32 mmol/L Na₂HPO₄, pH = 6.0, 1% trehalose, 0,05% Tween-20.

Spearman rank correlation was used to compute the correlation between EGFR and EPHA2 in TCGA patient data. Similarly, the log₂ of RSEM (RNA-Seq Expectation Maximization) normalized expression of EPHA2, EGFR, and c-MET of multiple tissue types as well as Luminal B subtype of breast cancer were extracted from the TCGA patient data and significance was calculated using the non-parametric Wilcoxon rank-sum test between the expression of c-MET and EPHA2. To identify the Luminal B samples, we used the PAM50 signature (9). The BAGEL (Bayesian Analysis of Gene EssentiaLity) algorithm (10) was used to evaluate the quality of the pooled shRNA screen. GSEA (Gene Set Enrichment Analysis) software (11) was used to identify the significantly enriched process of the hits generated from various experiments. A network image was generated using cytoscape software (12), where the top hits from shRNA screen and Bio-ID data were analyzed using GSEA and only hits from the top 15 significantly enriched GO processes with highest normalized enrichment scores were incorporated into the network. We used version 6.2 of the Molecular Signatures Database (MSigDB v 6.2) for GSEAs.

RNA-Seq data and tumor growth parameters of 66 patient-derived xenograft (PDX) models representing non–small cell lung carcinoma (NSCLC) and colorectal cancer treated or not with cetuximab (132 samples total) were curated from the previously published study (13). Only these 132 PDX samples had tumor growth data for cetuximab-treated/untreated conditions. We classified these samples based on the top and bottom quartiles of EPHA2 expression levels (17 PDX models in each category). The AUC of the tumor growth was calculated using Xeva package (v.1.99.20; ref. 14). Following this, we grouped AUC of the NSCLC and colorectal cancer samples into four groups as (i) low expression of EPHA2 with cetuximab treatment; (ii) low expression of EPHA2 in untreated condition; (iii) high expression of EPHA2 with cetuximab treatment, and (iv) untreated with high expression of EPHA2. The AUC of the 4 groups were plotted in boxplots and the significance between the treated conditions of low EPHA2 samples and high EPHA2 samples was calculated using an unpaired non-parametric Kolmogorov–Smirnov test and the P value was found to be 0.0463 (P < 0.05).

Recombinant human EPHA2 receptor extracellular domain (sEphA2; R&D Systems, #3035-A2–100) was immobilized to the chip ProteOn GLH (Bio-Rad, #176–5013) through amino groups. The measurements were done using Bio-Rad ProteOn XPR36 in 10 mmol/L HEPES, 150 mmol/L NaCl, 0,05% Tween 20, pH 7.4 buffer at 25°C. BMX-066 and BMX-661 were applied at 500 nmol/L concentrations. After washing with the buffer, 100 nmol/L sEGFR (R&D Systems, #1095-ER-002) was applied, as indicated, and after binding saturation, the unbound protein was again washed out.

Quaternary structure of BMX-661 solution (0.16 mg/mL) in the formulation buffer was measured using Zetasizer Nano ZS spectrometer (Malvern Instruments) at 25°C. The data represent averaged values from three consecutive measurements. The buffer reflection index n_D = 1.3354 and buffer dynamic viscosity η = 0.95 mPa.s at 25°C were measured using latex nanoparticles. For proteins, we used the reflection index n_D = 1,450 and absorption А = 0.01. Calculations of protein molecular weight were performed using the Mark–Houwink equation with K = 7.67·10⁻⁵ sm²·s⁻¹ and a = 0.428.

Prominence HPLC system from Shimadzu Corporation equipped with UV/Vis detector was used with the size exclusion chromatography column Phenomenox Biosec SEC-s3000 with the exclusion range 5–700 kDa.

Cells were stained with indicated mono- or bispecific antibodies and secondary FITC or PE-conjugated anti-human IgG, Fcγ fragment specific. The stained cells were analyzed with Miltenyi Biotec MACSQuant VYB (Miltenyi Biotec Inc.) or Beckman Coulter CytoFLEX (Beckman Coulter) flow cytometers and FlowJo or CytExpert software.

To assess subcellular enrichment of YFP-BirA*-FLAG or EPHA2-BirA*-FLAG fusion proteins, HEK293 T-Rex cells were pre-fixed with 2% paraformaldehyde (BioShop) for 5 minutes, washed with PBS and fixed with 4% paraformaldehyde for 15 minutes. Cells were permeabilized with 0.2% Triton X-100 (Sigma-Aldrich) for 15 minutes, blocked in 10% goat serum (Wisent) with 0.1% NP-40 (Sigma-Aldrich) in PBS and incubated with rabbit anti-FLAG (Sigma-Aldrich, #F7425) primary antibody in blocking solution for 1 hour. Following 3 washes in PBS with 0.1% NP-40, cells were incubated for 1 hour with Alexa Fluor 488 goat anti-rabbit IgG (Jackson ImmunoResearch, #111–545–003) secondary antibody. Cells were washed 3 times with 0.1% NP-40 in PBS and nuclei were stained with DAPI (Sigma-Aldrich) for 5 minutes. Slides were washed twice with PBS and mounted in ProLong Gold Antifade (Thermo Fisher Scientific). Images were acquired on a confocal laser scanning microscope (Zeiss LSM 700) using Plan-NeoFluar 40x OIL (numerical aperture 1.30) and processed using the ImageJ software.

Cell pellets from four 150-mm dishes were lysed in RIPA buffer (50 mmol/L tris-HCl pH 7.5, 150 mmol/L NaCl, 1 mmol/L EDTA, 1 mmol/L EGTA, 1% NP-40, 0.1% SDS, 0.5% sodium deoxycholate) supplemented with protease inhibitors (1 mmol/L PMSF, 10 μg.mL⁻¹ aprotinin, 10 μg.mL⁻¹ leupeptin, and 10 μg.mL⁻¹ pepstatin) and 250 U of benzonase. Following 1 hour incubation with agitation, lysates were sonicated on ice using three 10 seconds bursts with 2 seconds rest in between and centrifuged 30 minutes at 20,000 × g. Cleared supernatants were further incubated with streptavidin agarose beads (Sigma-Aldrich) for 3 hours with agitation at 4°C. Beads were washed three times with lysis buffer, followed by two washes with 20 mmol/L Tris-HCl (pH 7.4). Proteins were eluted with 50 mmol/L phosphoric acid, three times for 10 minutes and stored at −80°C. Eluted proteins were digested with Promega Sequencing Grade Modified Trypsin (Thermo Fisher Scientific) as described in ref. (15). The resulting peptides were desalted using StageTips (16) and dried down by vacuum centrifugation before LC-MS/MS analyses.

Dried peptides were resuspended in 15 μL of loading solvent (2% acetonitrile, 0.05% TFA), and 5 μL was used for injection onto a 300-μm inner diameter x 5 mm C-18 Pepmap cartridge precolumn (Dionex/Thermo Fisher Scientific) at 20 μL/min in loading solvent. After 5 minutes of desalting, the precolumn was switched online with a 75 μmol/L inner diameter x 50 cm separation column packed with 3 μmol/L ReproSil-Pur C18-AQ resin (Dr. Maisch HPLC GmbH) equilibrated in 95% solvent A (2% acetonitrile, 0.1% formic acid) and 5% solvent B (80% acetonitrile, 0.1% formic acid). Peptides were separated and eluted over a 90 minutes gradient of 5% to 40% solvent B at 300 nL/min flow rate generated by an UltiMate 3000 RSLCnano system (Dionex/Thermo Fisher Scientific) and analyzed on an Orbitrap Fusion mass spectrometer equipped with a nanoelectrospray ion source (Thermo Fisher Scientific). The Orbitrap Fusion was operated in data-dependent acquisition mode with the XCalibur software version 3.0.63 (Thermo Fisher Scientific). Survey MS scans were acquired in the Orbitrap on the 350 to 1,800 m/z range using an automatic gain control (AGC) target of 4e5, a maximum injection time of 50 ms and a resolution of 120,000. The most intense ions per each survey scan were isolated using the quadrupole analyzer in a window of 1.6 m/z and selected for Higher energy Collision-induced Dissociation (HCD) fragmentation with 35% collision energy. The resulting fragments were detected by the linear ion trap at a rapid scan rate with an AGC target of 1e4 and a maximum injection time of 50 ms. Dynamic exclusion was used within a period of 20 seconds and a tolerance of 10 ppm to prevent selection of previously fragmented peptides.

All MS/MS peak lists were generated using Thermo Proteome Discoverer version 2.1 (Thermo Fisher Scientific). MGF sample files were then analyzed using Mascot (Matrix Science; version 2.5.1). Mascot was set up to search UniProt Homo Sapiens database (April 2018 release, 93683 entries) supplemented with “common contaminants” from the Global Proteome Machine (GPM, thegpm.org, July 2017 release) and in house generated BissonN_flags dataset, assuming the digestion enzyme trypsin. Mascot was searched with a fragment ion mass tolerance of 0.60 Da and a parent ion tolerance of 10 ppm. Carbamidomethyl of cysteine was specified as a fixed modification. Oxidation of methionine and phosphorylation of serine, threonine and tyrosine were specified in Mascot as variable modifications. Two miscleavages were allowed. Scaffold (version 4.11.0, Proteome Software Inc.) was used to validate MS-MS–based peptide and protein identifications. Peptide identifications were accepted if they could be established at greater than 91.0% probability to achieve an FDR less than 1.0% by the Scaffold Local FDR algorithm. Protein identifications were accepted if they could be established at greater than 95.0% probability to achieve an FDR less than 1.0% and contained at least 2 identified peptides. Protein probabilities were assigned by the Protein Prophet algorithm (17). Proteins that contained similar peptides and could not be differentiated on the basis of MS-MS analysis alone were grouped to satisfy the principles of parsimony.

EPHA2 BioID was performed in biological triplicate and was processed independently on different days using cells from successive passages. YFP-BirA*-FLAG control was treated concomitantly to EPHA2-BirA*-FLAG experimental sample. To minimize carry-over issues during liquid chromatography, extensive washes were performed between each sample. To distinguish background contaminants and non-specific interactions from bona fide protein associations, MS data were analyzed with SAINTexpress, a simplified version of the Significance Analysis of INTeractome method (18) via the CRAPome website. The SAINTexpress probability value of each potential protein–protein interaction compared with background contaminants was calculated using default parameters. The three control samples were used in uncompressed mode. The complete list of SAINTexpress analysis results is provided in (Supplementary Table S6). The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository (19) with the dataset identifier PXD023034 and 10.6019/PXD023034.

αEPHA2-Cond and cIgG-Tr cells were SILAC labeled in arginine- and lysine-free high glucose DMEM (Caisson Laboratories, DML51–500ML) supplemented with 10% (v/v) dialyzed FBS (Invitrogen, 26400044), and 0.4 mmol/L light arginine and 0.8 mmol/L light lysine for cIgG-Tr cells, or 0.4 mmol/L heavy arginine (¹³C₆, Cambridge Isotopes, CLM-2265-H-PK) and 0.8 mmol/L heavy lysine (D₄, Cambridge Isotopes, DLM-2640-PK) for αEPHA2-Cond cells. Cells were grown at 37°C, 5% CO₂, and until the heavy amino acids were fully incorporated, requiring a month of culturing in SILAC medium.

Single 15-cm petri dish of cells were grown to high confluency, washed with cold 1X PBS three times, scraped to harvest, and boiled in 1% sodium deoxycholate in 50 mmol/L ammonium bicarbonate. The protein concentration was estimated by BCA assay (Pierce, 23227) and 100 μg of protein from each label was mixed, reduced, alkylated, and digested with trypsin (Promega, V5113) as described previously (20). Resulting peptides were cleaned with C-18 STop and Go Extraction (STAGE) tips (21) and eluted peptides were proceeded to lactic acid–modified-TiO₂ enrichment without performing vacuum centrifugation. Equal volume of buffer C consisting of 300 mg/mL DL-lactic acid (Acros Organics, 412965000) in 60% (v/v) acetonitrile and 0.075% (v/v) TFA solution was added to the peptides and passed through two 0.5 mg of bulk TiO₂ beads (GL Sciences, 5010–21315) sitting on C8 STAGE tips (3M, 2214). The tips were then washed with buffer C, and again with buffer B consisting of 80% (v/v) acetonitrile and 0.1% (v/v) TFA, and sequentially eluted with 5% (v/v) NH₄OH and 0.5% (v/v) pyrrolidine. The eluted phosphopeptides were pooled and immediately passed through C18 STAGE tip again and dried with vacuum centrifuge (Eppendorf, Vacufuge) before LC/MS-MS analysis. Three biological replicates were analyzed and each biological replicate was analyzed twice on the LC/MS-MS.

Each set of cleaned peptides was analyzed by a quadrupole–time of flight mass spectrometer (Impact II; Bruker Daltonics) coupled to an Easy nano LC 1000 HPLC (Thermo Fisher Scientific) using a Captive nanospray ionization source (Bruker Daltonics). The column was constructed in-house and included a fritted 2-cm-long trap column made with 100-μm-inner diameter fused silica and 5-μm Aqua C-18 beads (Phenomenex, 04A-4331) packing material, and a 40–50-cm-long analytic column with an integrated spray tip (6 – 8-μm-diameter opening, pulled on a P-2000 laser puller from Sutter Instruments) made with 75-μm-inner diameter fused silica and 1.9-μm-diameter Reprosil-Pur C-18-AQ beads (Dr. Maisch, r119.aq.0003) packing material. The column was operated at 50°C with an in-house built column heater. Buffer A for the LC consisted of 0.1% (v/v) formic acid, and buffer B for the LC consisted of 0.1% (v/v) formic acid and 80% (v/v) acetonitrile. A standard 90-minute peptide separation was done, and the column was washed with 100% buffer B before re-equilibration with buffer A. The Impact II was set to acquire in a data-dependent auto-MS/MS mode with inactive focus fragmenting the 20 most abundant ions (one at the time at an 18-Hz rate) after each full-range scan from m/z 200 to m/z 2,000 at 5 Hz rate. The isolation window for MS/MS was 2–3 depending on the parent ion mass to charge ratio, and the collision energy ranged from 23 to 65 eV depending on ion mass and charge. Parent ions were then excluded from MS/MS for the next 0.4 minutes and reconsidered whether their intensity increased more than five times. Singly charged ions were excluded from fragmentation.

Analysis of mass spectrometry data were performed using MaxQuant 1.5.3.30 (22). The search was performed against protein sequences from Uniprot Homo sapiens (retrieved on June 13, 2012 with 86,747 entries) plus common contaminants (245 entries) with variable modifications of methionine oxidation, N-acetylation of the proteins, and phosphorylation of serine, tyrosine and threonine in addition to the heavy amino acids used for quantitation and fixed modification of carbamidomethylation of cysteines. The precursor ion mass tolerance was set at 0.006 Da and 40 ppm for fragment ion tolerance. Default match-between-run and requantify parameters were enabled. Data were filtered to 1% FDR at both protein and peptide level based on parallel reverse hits. Phosphoproteomic data are available via ProteomeXchange (ref. 19; http://www.ebi.ac.uk/pride) with identifier PXD023087. Differential expression between αEPHA2-Cond and cIgG-Tr cells was tested using a one-sample t test on H/L SILAC ratios. Both the protein group abundance and phosphorylated peptide abundance were tested (proteinGroups.txt and phosphoSites.txt MaxQuant output tables). Only proteins or phosphosites with at least two H/L SILAC ratios were tested.

Data availability for proteomics, Phosphoproteomics, and BioID experiments is described in detail in the related sections of Materials and Methods. Microarray datasets related to the shRNA screens can be accessed at the GEO database using the accession number GSE171920 and access code, “yxarweuenxcbnch.” All other data and materials are available upon reasonable requests, BMX-066 and BMX-661—pending their production by Biomirex.

Although the EPHA2 receptor contributes to multiple human malignancies, none of the earlier attempts to interfere with its activity has translated into an actual therapy (6). Therefore, to test our strategy for choosing optimal candidates for therapeutic co-targeting (Fig. 1A), we selected the EPHA2 receptor as the initial potential target. To identify potential therapeutic partners of EPHA2, we initially screened the Biomirex human phage display Fab library with a repertoire around 4×10¹⁰ and developed EPHA2-targeting Fab fragments. His6-tagged human EPHA2 extracellular domain (ECD) and EPHA2-Fc construct (murine EPHA2 ECD fused to human IgG Fc fragment) were used for the screening to assure the cross-species reactivity. The selection of cross-reactive anti-EPHA2 Fabs without compromising the selected repertoire was possible due to high degree (83%) of amino acid sequence conservancy between human and mouse EPHA2 ECDs (Supplementary Fig. S1). Eventually, 10 Fabs were converted into human IgG1 mAbs. One of these antibodies, BMX-066, displayed a high affinity against EPHA2 in ELISA assays (Fig. 1B). BMX-066 effectively bound EPHA2 recombinantly expressed on the HEK-293 cell membrane and interacted with cancer cells intrinsically expressing this receptor (Fig. 1C–F). Administration of BMX-066 partially (by ∼2 fold; P < 0.05) suppressed tumor growth in a xenograft model generated with human triple-negative breast cancer (TNBC) cells, MDA-MB-231, in NOD.Cg-Prkdc^scid Il2rg^tm1Wjl/SzJ (NSG) mice (Fig. 1G) and we focused on selecting potential therapeutic partners.

We administered BMX-066 or control non-specific human IgG (cIgG) to NSG mice harboring MDA-MB-231 tumors. Animals were treated for 27 days, tumors were extracted and anti–EPHA2-conditioned (αEPHA2-Cond), and cIgG-treated (cIgG-Tr) ex vivo cell populations were established. The biological difference between the αEPHA2-Cond and cIgG-Tr cells was confirmed by comparing their phospho-proteome profiles. This revealed that 160 phosphorylated proteins were differentially expressed and 469 protein phosphorylation events differed significantly between the αEPHA2-Cond and cIgG-Tr cells (P < 0.05; Fig. 2A and B; Supplementary Tables S1 and S2). Specifically, phosphorylated proteins associated with cell/anchoring junction, cell substrate junction and regulation of cell death expressed differently in the αEPHA2-Cond and cIgG-Tr cells, indicating that the αEPHA2-Cond cells indeed represent a biologically distinct ex vivo cell population (Supplementary Fig. S2; Supplementary Table S3).

We next used the αEPHA2-Cond and cIgG-Tr cells in a genome-wide shRNA screen with a library containing approximately 90,000 unique shRNAs targeting approximately 18,000 human genes, as we recently described (8), to identify molecules that can be inhibited to selectively eliminate αEPHA2-Cond cells (Fig. 2C). A correlation clustergram and density plots of the two screens (αEPHA2-Cond and cIgG-Tr) showed high reproducibility among replicates (Supplementary Fig. S3A and S3B). A high recall of the set of essential genes (indicated by the F measure > 0.7; ref. 23) was achieved in both screens, confirming the high confidence in our analyses (Supplementary Fig. S3C). Overall, we identified 748 statistically significant hits (P < 0.005), whose shRNAs dropped out in the αEPHA2-Cond, but not in cIgG-Tr cell populations, indicating that inhibition of these genes should selectively kill anti–EPHA2-treated tumor cells (Fig. 2D; Supplementary Table S4). A stringent P value cutoff was applied in this analysis to select only the most significant candidates for further investigation. Thus, the shRNA screen identified several cancer-related molecules, including EGFR, SRC, KRAS, NCK1, with an enrichment in cellular processes such as cell death/survival, proliferation, migration, and signal transduction to name a few (Fig. 2F; Supplementary Fig. S4, Supplementary Table S5). The selective essentiality of EGFR, SRC, and NCK1 in αEPHA2-Cond cells did make sense, because all these molecules have been shown to be involved in EPHA2-initiated responses (6, 24, 25). Interestingly, EPHA2 suppression was previously reported to elevate KRAS activity (26), suggesting that KRAS overactivation might provide survival advantages to cancer cells in αEPHA2-treated tumors. Therefore, finding it among screening hits was also consistent with earlier observations. These matches between our screening data and previous reports indicated that inhibition of some of these molecules might potentially improve the efficiency of EPHA2 targeting in cancer cells. More importantly for our investigation, capturing these hits provided a strong support for the relevance of our ex vivo screening approach.

We expected that integration of genetic and physical interactions of our target should reveal fundamental connections within key functional modules of cancer cells that may become necessary for their survival (27). Therefore, to prioritize candidates among the 748 hits in a completely unbiased manner, we used BioID proteomics to delineate the EPHA2 proximity network. We produced inducible HEK293 T-Rex stable cell lines expressing YFP-BirA*-FLAG (control) or EPHA2-BirA*-FLAG using the Flp-In T-Rex system (Fig. 2E). Addition of biotin to the cell culture medium led to a strong increase in biotinylation of endogenous proteins compared with controls (Fig. 2E). We affinity-purified biotinylated proteins following the addition of biotin to YFP-BirA*-FLAG or EPHA2-BirA*-FLAG HEK293, identified them in biological triplicate with MS and eliminated non-specific interactions via SAINTexpress (18), using YFP-BirA*-FLAG-cells as controls. Only high-confidence interactions with a SAINT score ≥0.8, corresponding to a Bayesian FDR (BFDR) ≤2% were considered for analyses. This revealed 420 potential EPHA2 partners (Supplementary Table S6), with several of them predominantly associated with cell–cell and cell–substrate attachment and cell migration (Fig. 2F). The Gene Ontology analysis indicated enrichment of 122 of them in at least one of 15 gene sets selected on the basis of top enrichment scores (Supplementary Table S7, Supplementary Fig. S5). Among these proteins, only EGFR was found to be enriched in 14 gene sets (Supplementary Fig. S5), suggesting that it has a strong potential for being involved in the functional cross-talk with EPHA2, especially taking into consideration that EGFR was among the only nine hits found to be enriched in both shRNA screen and BioID assay datasets (Fig. 2F).

To further examine the relation between EphA2 and EGFR and its potential clinical relevance, we analyzed the TCGA patient dataset. We found expression levels of EPHA2 and EGFR to positively correlate in multiple if not the majority of malignancies, suggesting that these molecules probably co-operate in cancer cells and should be available for co-targeting in multiple tumor types (Fig. 3). This was consistent with previously published data, indicating that EGFR signaling enhances expression of the EPHA2 receptor in cultured cell lines (28, 29). We also assessed the efficiency of the clinically used EGFR inhibitor, erlotinib, in tumorsphere cultures produced by CRISPR-Cas9 EPHA2 knockout and mock-transduced control MDA-MB-231 cells (Fig. 4A). We selected tumorsphere models for these experiments, as tumorspheres better represent tumor biology than cells cultured in monolayers. Interestingly, EPHA2 knockout significantly increased the sensitivity of tumorsphere cells to erlotinib treatment (Fig. 4B–D), while not affecting erlotinib action in cancer cell monolayers (Fig. 4E), once again highlighting biological differences between tumorspheres and cells in 2-D cultures. Similar effects of EPHA2 knockout were also observed in pancreatic cancer MIA PaCa-2 cells (Fig. 4F–I). Moreover, our analysis of EPHA2 expression in 66 previously described independent PDX models of colorectal cancer and NSCLC (13) revealed that treatment with the therapeutic anti-EGFR antibody, cetuximab, is significantly (P < 0.05, Kolmogorov–Smirnov test) more effective in tumors with low EPHA2 levels (Fig. 4J). This finding was consistent with our observations in tumorspheres and agreed well with a recently published report, showing that the EPHA2–EGFR crosstalk contributes to the aggressiveness of both colon and lung cancers (30). Similarly, earlier data indicate that EPHA2 silencing can suppress lung cancer cells that developed resistance to EGFR inhibitor, erlotinib (31), and that lower EPHA2 levels correlate with better responsiveness of colorectal cancer patients to cetuximab (32, 33).

Overall, a prediction based on the integration of genetic and proteomics approaches with gene expression analyses of the TCGA database was successfully validated by subsequent EPHA2 knockout experiments, and analyses of independently reported PDX models. This prediction clearly pointed toward EGFR as a potential therapeutic co-targetable partner of EPHA2.

To co-target EPHA2 and EGFR and assure a synchronized delivery of anti-EPHA2 and anti-EGFR reagents to tumor cells, we generated a bispecific anti-EGFR/EPHA2 antibody. The antibody, BMX-661, was heterodimerized through the “Knob-Hole” approach using BMX-066 and cetuximab, where the amino acids of heavy chains CH3 domains were remodeled with “knob-into-holes” mutations (34). One arm of BMX-661 was composed of BMX-066–derived Fab, whereas another part was made by the fusion of VH and VL domains of cetuximab using the (G4S)3 linker to form scFv. To assist bioanalytics and purification, the non-immunogenic His5 tag (35) was incorporated between cetuximab scFv and Fc (Fig. 5A). BMX-661 was expressed in CHO-S cells and purified using Protein A followed by Ni⁺-resins. Its affinities to EPHA2 and EGFR were measured by ELISA, and proved to be similar for both antibody arms (Fig. 5B). Surface plasmon resonance measurements also confirmed that there is no steric hindrance between the anti-EPHA2 and anti-EGFR arms of BMX-661 in binding to their corresponding antigens (Fig. 5C).

BMX-661 was additionally characterized in several conventional bioanalytical assays, including dynamic light scattering (Fig. 5D), HPLC (Fig. 5E), and SDS-PAGE (Fig. 5F), which confirmed good aggregation, thermal and proteolytic characteristics. It remained stable at 5 mg/mL for up to 200 days at 4°C and three days at 37°C (Fig. 5F), although some batches were observed to lose activity after longer storage at 4°C or room temperature. The antibody was also stable in human blood serum, as determined by its incubation for seven days at 37°C followed by the assessment of its binding characteristics to EPHA2 and EGFR (Fig. 5G). BMX-661 recognized EGFR and EPHA2 on the cell membrane (Fig. 5H–J), and interacted with EPHA2- and EGFR–co-expressing cancer cells (Fig. 6A and B). Overall, the bioanalytical characterization of BMX-661 highlighted its potential for preclinical and clinical testing.

To assess therapeutic properties of BMX-661, we tested it in xenograft models of human malignancies. To examine its effect on tumor growth and metastasis, we exploited as in Fig. 1G, MDA-MB-231 cells that metastasize into lungs (36). MDA-MB-231 cells stably expressing EGFP were used to develop tumors in mammary fat pads of NSG mice. Tumors were allowed to grow to a measurable size and the mice were treated with BMX-661. This very effectively, by up to approximately 4-fold (t test, P < 0.05), inhibited tumor growth and strongly suppressed metastasis into lungs, demonstrating both antitumor efficiency and anti-metastatic action of BMX-661 (Fig. 6C). Targeted treatment approaches are usually most effective, when applied in combinations with traditional chemotherapies and when applied at a lower dose, BMX-661 proved to enhance the therapeutic action of a DNA-damaging drug cisplatin in MDA-MB-231 xenografts (Fig. 6D). The antitumor activity of BMX-661, was also tested in the HCI-010 PDX that represents TNBC and closely mimics properties of the original tumor (37). Excitingly, treatment with BMX-661 strongly suppressed tumors in this model (Fig. 6E). The therapeutic effect was not restricted to TNBC, because BMX-661 administration also significantly inhibited tumor growth in a xenograft model representing pancreatic cancer (Fig. 6F).

Taken together, these data convincingly confirmed that EGFR is the molecule of choice for co-targeting with EPHA2 in tumor cells and highlighted the potential of our unbiased strategy for selecting efficient target combinations for bispecific therapeutic reagents.

Accumulating advances in the understanding of the molecular machinery responsible for tumor development have triggered the generation of multiple therapeutic approaches that target individual components of cancer-promoting mechanisms. These treatments depend on the development of selective small-molecule inhibitors and therapeutic antibodies (38, 39). Unfortunately, tumors expressing matching targets often produce low responses or acquire resistance to these compounds, leading to relapse. This is at least in part driven by the innate tumor heterogeneity, which allows survival of treatment-resistant cancer cell populations (1, 40). One of the strategies of choice for minimizing cancer resistance is the development of selective dual-specificity regents, which are currently mostly represented by bispecific antibodies aimed at two different oncogenic cell surface molecules (41–43). The bispecific nature of these therapeutics allows perfectly synchronized action on both targets, optimizing tumor-suppressing efficiency, tumor targeting, and minimizing the evolutionary potential of cancer cells. The relevance of this notion is illustrated by our data, showing that high expression of two pro-malignant RTKs, EGFR and EPHA2, coincides in multiple malignancies (Fig. 3), which suggests that their co-targeting with bispecific reagents is likely to produce beneficial therapeutic effects. Indeed, more than 50 bispecific antibodies have already reached clinical trials (43). Most recently, the anti-EGFR/c-MET antibody, amivantamab, has been fast-tracked into clinics by Janssen (44) https://www.fda.gov/drugs/resources-information-approved-drugs/fda-grants-accelerated-approval-amivantamab-vmjw-metastatic-non-small-cell-lung-cancer and two similar bispecific anti-EGFR/c-MET antibody programs have been successfully developed by Eli Lilly (45) and Merck KGaA (46).

Until now, target combinations for bispecific reagents have been mostly chosen on the basis of the pre-existing knowledge of functions of cancer-related molecules, which may often result in suboptimal tumor-suppressing properties (1–3). To address this, we suggest a systematic and completely unbiased strategy for designing bispecific therapeutics. Thus, to identify optimal candidates for co-targeting, we suggest to pre-condition cancer cells in xenograft tumors to the action of a mono-specific reagent aimed at the selected candidate target. These pre-conditioned cells can be subsequently used ex vivo, in genome-wide loss of function screening to identify acquired dependencies. Cross-referencing the genetic dependencies with the delineated physical target's interactome should reveal key functional connections essential for cancer cell survival. The final selection of the biologically optimal co-target is based on the unrestricted bioinformatics analysis of the TCGA database, to assure that the chosen combination of molecules occurs at a significant frequency in human malignancies and therefore, should have a high level of therapeutic relevance. In our investigation, this strategy pointed toward two RTKs, namely EPHA2 and EGFR as an optimal combination for co-targeting. Remarkably, this prediction was consistent with previous findings that reported a functional crosstalk between EPHA2 and EGFR in cancer cells and implicated EPHA2 in the resistance to EGFR inhibitors in some malignancies, including colorectal cancer and non-NSCLC (24). The close match between our prediction and previous independent reports highlights the high level of accuracy of our unbiased strategy. It also suggests that additional potential target combinations can be identified with a high level of confidence using our coordinated genome-wide/proteome-wide data (Supplementary Tables and Fig. 2F).

To further validate our prediction, we have generated and characterized a human anti-EGFR/EPHA2 antibody, BMX-661. As anticipated, BMX-661 very efficiently suppressed aggressive TNBC and pancreatic cancer tumors in xenograft models, when compared with its prototypic anti-EGFR antibody, cetuximab. BMX-661 proved to be effective in xenograft cancer models at a relatively low dose of 2 mg/kg/wk compared with doses exceeding 15 mg/kg/wk often used for testing RTK-targeting antibodies for therapeutic efficiency (47–49). When applied at an even lower dose of 1.5 mg/kg/wk, BMX-661 worked well in combination with cisplatin and significantly improved the antitumor action of this conventional chemotherapeutic drug. Although the exact mechanism of BMX-661 action is still being dissected, its clinical potential is highlighted by our data, showing that targets for BMX-661 are co-expressed in multiple human malignancies and that the antibody produces strong antitumor responses. Currently, monospecific EGFR-inhibiting reagents are efficient in only a few malignancies and inflict incomplete therapeutic effects (50–52). We expect BMX-661 to improve the therapeutic effectiveness of EGFR inhibition in a variety of cancers expressing both antigens. Thus, co-suppressing EGFR and c-MET has a high therapeutic potential (44), and our work indicates that co-targeting EGFR and EPHA2 would provide a very effective complementary treatment approach. This could be especially relevant in EGFR-expressing tumors with EPHA2 levels similar or higher compared with c-MET (Supplementary Fig. S6), suggesting a need for further preclinical evaluation in matching models.

Importantly, our work with BMX-661 demonstrates the feasibility of our multipronged unbiased strategy for selecting optimal co-targets for combination therapies and for supporting generation of efficient bispecific therapeutic reagents. We certainly realize that this strategy relies on the accurate choice of the first target for a combination treatment, and also on several advanced approaches that require sufficient resources and expertise. That said, we also believe that current rapid advances in research technologies and in the understanding of cancer biology should allow to effectively use our approach in a broad variety of drug development projects.

I. Alexandrov reports a patent for antibody pending. B. Wang reports grants from NIH during the conduct of the study. B.M. Toosi reports nonfinancial support from Biomirex Inc. during the conduct of the study. N. Bisson reports grants from FRQNT during the conduct of the study. T.A. Mirzabekov reports a patent for bispecific anti-EGFR/anti-EphA2 antibody pending; in addition, T.A. Mirzabekov reports employment as chief executive officer and president of Biomirex, Inc., which owns the anti-EGFR/anti-EphA2 bispecific antibody. F.J. Vizeacoumar reports nonfinancial support from Biomirex Inc. during the conduct of the study. A. Freywald reports nonfinancial support from Biomirex Inc. during the conduct of the study; in addition, A. Freywald reports an advisory role on the scientific board of Biomirex Inc., which is a volunteer/public service activity to support the development of new cancer therapies for which A. Freywald has received no compensation. No disclosures were reported by the other authors.

A. El Zawily: Investigation, methodology. F.S. Vizeacoumar: Data curation, investigation, methodology. R. Dahiya: Investigation, methodology. S.L. Banerjee: Data curation, investigation, methodology. K.K. Bhanumathy: Investigation, methodology. H. Elhasasna: Investigation, methodology. G. Hanover: Investigation, methodology. J.C. Sharpe: Investigation, methodology. M.G. Sanchez: Investigation, methodology. P. Greidanus: Investigation, methodology. R.G. Stacey: Data curation, investigation, methodology. K.-M. Moon: Investigation, methodology. I. Alexandrov: Investigation, methodology. J.P. Himanen: Investigation, writing–review and editing. D.B. Nikolov: Investigation, writing–review and editing. H. Fonge: Investigation, writing–review and editing. A.P. White: Investigation, methodology. L.J. Foster: Investigation, methodology, writing–review and editing. B. Wang: Conceptualization, writing–review and editing. B.M. Toosi: Conceptualization, validation, methodology, writing–original draft, writing–review and editing. N. Bisson: Conceptualization, investigation, methodology, writing–original draft, writing–review and editing. T.A. Mirzabekov: Conceptualization, investigation, methodology, writing–original draft, writing–review and editing. F.J. Vizeacoumar: Conceptualization, funding acquisition, writing–original draft, writing–review and editing. A. Freywald: Conceptualization, funding acquisition, writing–original draft, writing–review and editing.

We thank Dr. A.L. Welm, Co-Director, Cell Response and Regulation Program, Huntsman Cancer Institute, University of Utah, for providing the HCI-010 PDX and for the related technical advice. This work was supported by the Canadian Institutes of Health Research grants (PJT-156401; PJT-156309); College of Medicine, University of Saskatchewan, and Be Like Bruce funding (to A. Freywald and F.J. Vizeacoumar); and Genome Canada (264PRO; to L.J. Foster). R. Dahiya was supported by an SHRF fellowship; F.S. Vizeacoumar is supported by the College of Medicine, University of Saskatchewan; and G. Hanover, M.G. Sanchez, and H. Elhasasna were supported by CoMGRAD awards, College of Medicine, University of Saskatchewan.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.