Inhibition of VEGF binding to neuropilin-2 enhances chemosensitivity and inhibits metastasis in triple-negative breast cancer

Abstract

Although blocking the binding of vascular endothelial growth factor (VEGF) to neuropilin-2 (NRP2) on tumor cells is a potential strategy to treat aggressive carcinomas, a lack of effective reagents that can be used clinically has hampered this potential therapy. Here, we describe the generation of a fully humanized, high-affinity monoclonal antibody (aNRP2-10) that specifically inhibits the binding of VEGF to NRP2, conferring antitumor activity without causing toxicity. Using triple-negative breast cancer as a model, we demonstrated that aNRP2-10 could be used to isolate cancer stem cells (CSCs) from heterogeneous tumor populations and inhibit CSC function and epithelial-to-mesenchymal transition. aNRP2-10 sensitized cell lines, organoids, and xenografts to chemotherapy and inhibited metastasis by promoting the differentiation of CSCs to a state that is more responsive to chemotherapy and less prone to metastasis. These data provide justification for the initiation of clinical trials designed to improve the response of patients with aggressive tumors to chemotherapy using this monoclonal antibody.

INTRODUCTION

The challenges associated with the clinical management of highly aggressive tumor types such as triple-negative breast cancer (TNBC) are well known to patients, physicians, and scientists. TNBC represents about 15% of all breast cancers, and it occurs preferentially in younger women, with the prognosis being quite grim (1). In sobering terms, TNBC affects about 42,000 women per year, is associated with a 37% incidence of mortality within 5 years, and results in a median survival of 9 months after recurrence (1, 2). Clearly, there is an urgent need to understand the nature of aggressive tumor types such as those found in TNBC, especially their resistance to current therapies, and to develop strategies for improving patient outcomes.

Neuropilin-2 (NRP2) is a single-pass transmembrane protein that forms heterodimeric complexes with many other plasma membrane receptors, including growth factor receptors (3, 4) and integrins (5, 6). The extracellular domain (ECD) of NRP2 binds ligands, including the vascular endothelial growth factor (VEGF) and semaphorin families (6). Its short intracellular domain does not signal directly; rather, it enhances the signaling output of growth factor receptors (7), integrins, and other co-receptors (6, 8). NRP2 expression has been detected in several human tumor types (including prostate, pancreatic, gastric, kidney, colon, bladder, and breast cancers), with the highest expression observed in aggressive, treatment-resistant tumors (6, 9). For example, we have shown that NRP2 is highly expressed in TNBC compared with other breast cancer subtypes (10), and other data indicate that its expression is associated with lower survival (11). Of note, NRP2 is not expressed in normal breast epithelial tissue (11), which heightens its therapeutic potential. Publications have demonstrated that NRP2 confers aggressive properties, including therapy resistance and metastasis, by contributing to cancer cell stemness, epithelial-to-mesenchymal transition (EMT), and therapy-induced cellular plasticity in TNBC, bladder, and prostate cancers (10, 12–15). On the basis of its unique role in contributing to these properties, we believe that monoclonal antibody (mAb)–mediated inhibition of NRP2 should improve existing therapies in aggressive, metastatic tumor settings.

To date, the development of NRP2 as a therapeutic target has not progressed largely because of the absence of highly potent and selective inhibitors that can be used in the clinic and are capable of blocking specific functions of this receptor. Although NRP2 can function as a receptor for multiple ligands, the existing data indicate that its interaction with VEGFs, especially VEGF-A and VEGF-C, is responsible for its ability to contribute to the aggressive behavior of certain tumor subtypes (6, 8, 10, 12, 13, 16). Its interaction with class 3 semaphorins, in contrast, may impede some aspects of tumor behavior, such as growth, migration, and metastasis (17–21). Compelling evidence also supports the importance of semaphorins in cancer and their use as therapeutic targets (22). These observations highlight the need to generate highly specific reagents that can distinguish binding of VEGFs and semaphorins to NRP2 and enable an analysis of the differential contributions of these two disparate ligands to these aggressive phenotypes. In this context, it should also be noted that the VEGF-targeting mAb bevacizumab specifically blocks the binding of VEGF-A to receptor tyrosine kinases but not to NRP2, which may explain its lack of lasting efficacy when used as a single agent in some clinical studies (23). In this study, we describe the generation of highly specific antibodies that selectively block the binding of either VEGF or semaphorin to NRP2 and assess their preclinical potential for the therapeutic management of aggressive tumors such as those found in TNBC. The data obtained provide justification for the use of a specific mAb that inhibits the binding of VEGF to NRP2 in clinical trials.

RESULTS

Generation and characterization of inhibitory aNRP2 mAbs

The first goal of this study was to generate highly specific anti-NRP2 (aNRP2) mAbs that could be used for the therapeutic targeting of this protein. Two mAbs (aNRP2-10 and aNRP2-14) were produced that primarily target the b1 and a2 domains of NRP2, respectively (Fig. 1A, tables S1 and S2, and fig. S1A). As will be shown here, we focused on these two NRP2 mAbs because they differ in function: aNRP2-10 blocks the binding of VEGF to NRP2, and aNRP2-14 blocks the binding of semaphorin 3F (Sema3F) to NRP2. Both mAbs specifically bound to the human NRP2 (hNRP2) ECD (hNRP2-ECD) but not to the hNRP1-ECD (Fig. 1B) despite 42% identity of the amino acid sequences. Neither antibody bound rodent NRP2, indicating recognition of nonconserved residues (table S2). The humanized aNRP2-10 and aNRP2-14 mAbs bound to hNRP2-ECD with subnanomolar affinities, as measured by surface plasmon resonance (SPR). The dissociation constant (K_D) was 27.1 pM for aNRP2-10 and 131.7 pM for aNRP2-14 (table S1 and fig. S1B). They potently bound to suspension-adapted human embryonic kidney 293 cells that stably overexpress hNRP2 (Expi293F-hNRP2). The half-maximal effective concentration (EC₅₀) of the binding was 0.51 nM for aNRP2-10 and 1.51 nM for aNRP2-14 (Fig. 1C and table S1). Specificity of binding to hNRP2 was demonstrated by cell binding to wild type versus NRP2 knockout (KO) A549 cells (Fig. 1D). aNRP2-10 and aNRP2-14 bound to wild-type (WT) A549 cells with high sensitivity but showed little binding to the NRP2 KO clonal cells. In contrast, a commercial polyclonal aNRP2 antibody (R&D Systems, #AF567) showed less binding to A549 WT cells and obvious nonspecific binding to A549 NRP2-KO cells.

Fig. 1. aNRP2 mAbs specifically bind NRP2 and selectively block ligand binding.

(A) Schematic of the hNRP2 protein domains and binding sites of NRP2 ligands and aNRP2 mAbs is shown. (B) Binding of aNRP2-10 and aNRP2-14 to hNRP2_ECD and hNRP1_ECD was measured by ELISA. Consistent results were obtained in three independent experiments, and representative data with n = 2 technical replicates are shown. (C) Binding of aNRP2-10 and aNRP2-14 to Expi293F-hNRP2 cells was measured by flow cytometry. hIgG4 was used as a control in (B) and (C). Consistent results were obtained in three independent experiments, and shown are representative data with n = 2 technical replicates. The EC₅₀ of mAb binding to Expi293F-hNRP2 cells quantified from independent experiments is shown in table S1. (D) aNRP2-10 and aNRP2-14 binding to A549 cells expressing endogenous NRP2 and A549 NRP2_KO cells were analyzed. (Left) Representative flow cytometry spectra of aNRP2-10 mAb or a commercial aNRP2 polyclonal Ab (R&D Systems, #AF567) binding to A549 WT versus NRP2_KO cells are shown. Ctrl indicates isotype control. The plot shows the fold change (FC) of aNRP2-10, aNRP2-14, or AF567 MFI (median fluorescence intensity) to that of Ctrl. Dashed line is at FC = 1 and indicates background staining signal of Ctrl. Consistent results were obtained in three independent experiments, and representative data with n = 2 technical replicates are shown. (E) Ability of aNRP2-10 to block VEGF-C binding to Expi293F-hNRP2 cells was measured by flow cytometry. The IC₅₀ from three independent experiments is 0.71 ± 0.15 nM (0.412, 0.875, and 0.842 nM). Representative data with n = 2 technical replicates are shown. (F) Ability of aNRP2-14 to block Sema3F binding to Expi293F-hNRP2 cells was measured by flow cytometry. hIgG4 was used as a control in (E) and (F). The IC₅₀ from three independent experiments is 1.94 ± 0.38 nM (2.25, 2.374, and 1.186 nM). Representative data with n = 2 technical replicates are shown. (G to I) Ligand-induced receptor dimerization and antibody blocking effects were analyzed using a luciferase complementation assay. The ability of aNRP2-10 and aNRP2-14 to block NRP2/KDR dimerization induced by VEGF-A (G), NRP2/FLT4 dimerization induced by VEGF-C (H), and NRP2/PlexinA1 dimerization induced by Sema3F (I) was measured. hIgG4 was used as a control. Representative data with n = 3 biological replicates are shown. Data from three or four independent experiments are shown in fig. S1E. Antibody-mediated blocking of NRP2/co-receptor dimerization at additional doses is shown in fig. S1 (F and G). (J) Crystal structure of aNRP2-10 Fab bound to NRP2-a2b1b2 domains, with an interaction interface on the b1 domain of NRP2, in contrast to aNRP2-14 that interacts primarily with the a2 domain of NRP2, is shown (see also fig. S1I). (K) Interaction interface of NRP2 and aNRP2-10 in the crystal structure of NRP2 a2b1b2/aNRP2-10 Fab complex is shown. R100 (Kabat numbering) on the heavy chain of aNRP2-10 inserts into the binding pocket of the NRP2 b1 domain, which overlays with the C-terminal R227 of VEGF-C in a previously published NRP2/VEGF-C structure (25). Top: Structure of NRP2/aNRP2-10 aligned to the structure of VEGF-C C-terminal peptide bound to NRP2 [PDB: 4QDQ]. Bottom: VEGF-C C-terminal R227 or aNRP2-10 R100 inserting into the binding pocket of NRP2. Data are presented as means ± SEM.

Next, we analyzed the capability of aNRP2-10 and aNRP2-14 to block the binding of known NRP2 ligands: VEGFs and semaphorins. By flow cytometry analysis, VEGF-C was determined to bind to Expi293F-hNRP2 cells with an EC₅₀ of 0.6 nM (fig. S1C). Preincubation of cells with aNRP2-10 inhibited VEGF-C binding to Expi293F-hNRP2 cells in a dose-dependent manner, whereas the isotype control human immunoglobulin G4 (hIgG4) did not (Fig. 1E). aNRP2-10 blocked VEGF-C binding completely, with a half-maximal inhibitory concentration (IC₅₀) of 0.71 nM. The semaphorin, Sema3F, bound to Expi293F-hNRP2 cells with an EC₅₀ of 5.7 nM (fig. S1D). aNRP2-14, but not the isotype control, blocked Sema3F binding completely with an IC₅₀ of 1.94 nM (Fig. 1F). To further evaluate the impact of these antibodies in blocking of ligand-induced receptor/co-receptor dimerization, we used a split NanoLuc complementation assay from Promega. Expi293F cells were transfected to overexpress NRP2 C-terminally tagged with the NanoLuc large fragment (LgBiT) and co-receptors C-terminally tagged with the small fragment (SmBiT) and then assayed for reconstituted luciferase activity upon ligand-induced receptor dimerization. aNRP2-10, but not aNRP2-14, effectively inhibited VEGF-A–induced dimerization of NRP2 with the kinase insert domain receptor [KDR; also known as VEGF receptor-2 (VEGFR2)] and VEGF-C–induced dimerization of NRP2 with the fms-related tyrosine kinase 4 (FLT4; also known as VEGFR3) (Fig. 1, G and H, and fig. S1, E and F). However, aNRP2-10 did not block Sema3F-induced NRP2/PlexinA1 dimerization (Fig. 1I), which suggests that aNRP2-10 can specifically block NRP2/co-receptor dimerization and downstream signaling induced by VEGFs, but not Sema3F. In contrast, aNRP2-14 inhibited Sema3F-induced NRP2/PlexinA1 dimerization in a dose-dependent manner, but it had no effect on VEGF-A– or VEGF-C–induced NRP2/VEGFR dimerization (Fig. 1, G to I, and fig. S1, E and G). Sema3F primarily binds the a1 and a2 domains of NRP2, whereas VEGF ligands bind the b1 and b2 domains (24). The ligand blocking profiles of the antibodies are consistent with how their epitopes overlap with binding regions of the respective ligands (Fig. 1A). This suggests that the antibodies bind and block the sites for ligand binding and thus inhibit the ligand-induced NRP2/co-receptor dimerization. In summary, we generated two anitbodies that bind NRP2 with high affinity but differ in their specificity: aNRP2-10 inhibits the binding of VEGF-A and VEGF-C to NRP2, and aNRP2-14 inhibits the binding of Sema3F.

To understand the structural basis of how aNRP2-10 blocks the binding of VEGFs, we crystallized the protein complex of hNRP2-a2b1b2 and aNRP2-10 Fab (antigen-binding domain). The crystal structure was solved at a resolution of 3.2 Å with well-resolved binding site contacts (Fig. 1J). The structure shows that residue R100 on the heavy chain of aNRP2-10 inserts into a binding pocket on the b1 domain of NRP2 (Fig. 1K). When overlaid with a previously published complex structure of NRP2 and the C-terminal peptide VEGF-C [Protein Data Bank (PDB): 4QDQ], the R100 residue in aNRP2-10 overlays well with the C-terminal R227 of VEGF-C, which is the key residue for the interaction with NRP2 (25). VEGF-A also binds NRP2 similarly (25, 26), and, therefore, the structure results demonstrate that aNRP2-10 occupies the binding site of VEGF-A or VEGF-C and blocks their access to the NRP2 binding pocket. aNRP2-10 binds strongly to hNRP2 but poorly to mouse NRP2 (mNRP2; table S2), although the mNRP2 b1 domain shares 96% identity with the human counterpart and adopts a similar fold. This suggests that aNRP2-10 has distinct epitopes from the known Genentech aNRP2 antibody, aNRP2B, which interacts with the b1-b2 domain of NRP2 and binds human and mNRP2 equally well at about 5 nM by SPR measurement (27). Binding experiments in which binding of one antibody does not interfere with subsequent binding of the second antibody demonstrate that aNRP2-10 and aNRP2B bind to distinct, nonoverlapping epitopes (fig. S1H). These results highlight aNRP2-10 as a VEGF-blocking antibody that may have distinct functional properties.

Separately, we solved the crystal structure of the protein complex of hNRP2-a2b1b2 and aNRP2-14 (fig. S1I), which shows that aNRP2-14 interacts primarily with the a2 domain of NRP2 that encompasses Sema3F binding sites. This finding agrees with the epitope mapping results (fig. S1A and tables S1 and S2) and that aNRP2-14 could block Sema3F for binding to NRP2 (Fig. 1F).

aNRP2-10 selects for cancer stem cells in TNBC and inhibits their self-renewal

Given that breast cancer stem cells (CSCs) have higher expression of NRP2 compared with non-CSCs (10, 12, 13), we hypothesized that aNRP2-10 could be used to enrich for CSCs in heterogeneous populations of tumor cells. To assess this hypothesis, we sorted BT-549 TNBC cells into NRP2^High and NRP2^Low populations and observed that the NRP2^High population comprised 2.47% of the cells (Fig. 2A). We compared the self-renewal potential of these populations by serial mammosphere formation and observed a significant (P = 0.0043) difference between these populations (Fig. 2B). Next, we compared the ability of aNRP2-10 and aNRP2-14 to inhibit mammosphere formation using the NRP2^High population that had been sorted from BT549. aNRP2-10 significantly (P = 0.0255) inhibited mammosphere formation, but no effect was observed with the Sema3F blocking mAb aNRP2-14 (Fig. 2C). Moreover, aNRP2-10 also inhibited serial mammosphere formation using other TNBC cell lines MDA-MB-468 and Hs578T (Fig. 2, D and E). We extended this analysis to TNBC tumors and found that aNRP2-10 could inhibit the self-renewal of a CD44^High/CD24^Low CSC population isolated from patient-derived TNBC cells (Fig. 2, F and G).

Fig. 2. aNRP2-10 selects for CSCs in TNBC and inhibits their self-renewal.

(A) BT-549 cells were sorted into NRP2High and NRP2Low populations by FACS using aNRP2-10. Flow spectra and the percentage of NRP2^High cells in the parental cells or NRP2^Low and NRP2^High populations 10 days after sorting are shown. (B) Number (no.) of mammospheres formed by NRP2^High and NRP2^Low BT-549 cells in serial passages is shown. P1 indicates passage 1, and P2 indicates passage 2 of cells after sorting. Consistent results were obtained in two independent experiments, and representative data with n = 3 biological replicates are shown. Statistical significance was determined by two-sided, unpaired t test. **P < 0.01. (C) Ability of aNRP2-10 and aNRP2-14 to inhibit mammosphere formation of sorted NRP2^High BT-549 cells is shown. Consistent results were obtained in three independent experiments, and representative data with n = 3 biological replicates are shown. Statistical significance was determined by two-sided, unpaired t test. *P < 0.05. (D) Ability of aNRP2-10 and aNRP2-14 to inhibit mammosphere formation of sorted NRP2^High MDA-MB-468 cells is shown. Consistent results were obtained in two independent experiments, and representative data with n = 3 biological replicates are shown. Statistical significance was determined by two-sided, unpaired t test. **P < 0.01. (E) Ability of aNRP2-10 to inhibit mammosphere formation of Hs578T cells as compared with hIgG4 as the isotype control IgG is shown. n = 3 biological replicates. Statistical significance was determined by two-sided, unpaired t test. *P < 0.05. (F) A CD44^High/CD24^Low population enriched in CSCs was sorted from TNBC patient–derived cells by FACS. (G) Ability of aNRP2-10 to inhibit serial mammosphere formation by the sorted CD44^High/CD24^Low population is shown. P3 indicates passage 3 of cells after sorting. n = 3 biological replicates. Statistical significance was determined by two-sided, unpaired t test. *P < 0.05 and **P < 0.01. (H) Representative flow spectra (top) and quantification (bottom) of the CSC population in 4T1 tumors after aNRP2-28 or control IgG treatment are shown. Mice were randomized to three mice per group and administered with aNRP2-28 (25 mg/kg) or IgG (25 mg/kg) by intraperitoneal injection. Treatments were administered three times a week for 3 weeks. Statistical significance was determined by two-sided, unpaired t test. **P < 0.01. FITC, fluorescein isothiocyanate; PE, phycoerythrin. (I) Ability of aNRP2-28 to reduce ZEB1 gene expression in 4T1 tumors is shown. Tumor samples were collected 24 hours after the last treatment in the study (H), and gene expression was quantified by qPCR. n = 2 or 3 mice in each group. Statistical significance was determined by two-sided, unpaired t test. *P < 0.05. Data in (B) to (I) are presented as means ± SEM. (J) Ability of aNRP2-28 to inhibit 4T1 tumor initiation by in vivo limiting dilution assay is shown. 4T1 tumors from mice treated with aNRP2-28 or IgG in study (H) were harvested, dissociated into a single-cell suspension, and injected into the mammary fat pads of BALB/c mice using a 10-fold serial dilution at n = 5 per group. Tumors were harvested after 3 weeks, and the frequency of CSCs was quantified by extreme limiting dilution assay. Data are presented as a log-log plot.

An important question that arose from our data is whether inhibiting VEGF/NRP2 using function-blocking mAbs diminished the frequency of CSCs. To address this issue, we performed secondary transplantation experiments using 4T1 cells, a murine cell line that is widely used as a syngeneic model of TNBC. Specifically, we harvested 4T1-derived tumors that had been treated with either mouse IgG (mIgG) or aNRP2-28 (a functional surrogate antibody for aNRP2-10 that binds the b1 domain of mNRP2; table S3) and quantified the number of CSCs using CD44 and CD24 as markers. aNRP2-28 caused a significant (P = 0.0035) reduction in the frequency of CD44^high/CD24^low CSCs compared with the IgG control (about 70%) (Fig. 2H). This decrease in CSC frequency was also associated with reduced expression of ZEB1, a marker of CSCs and a master regulator of EMT (Fig. 2I). To substantiate these data, we digested these tumors and reimplanted them in syngeneic mice in a limiting dilution assay. As shown in Fig. 2J, the aNRP2-28–treated group exhibited reduced CSC frequency as assessed by extreme limiting dilution analysis (ELDA), as well as reduced secondary tumor uptake compared with the IgG-treated group.

aNRP2-10 sensitizes TNBC to chemotherapy in vitro

Because the major goal of this study was to develop highly specific NRP2 mAbs that could be used to improve existing therapies for TNBC, we sought to assess the ability of aNRP2-10 to sensitize TNBC to chemotherapy. The feasibility of this goal was strengthened by the data described above indicating that aNRP2-10 inhibits CSCs by reducing their self-renewal potential and that the high frequency of CSCs in TNBC is associated with chemotherapy resistance (28). To assess the effect of aNRP2-10 on chemosensitivity, we initially investigated the ability of aNRP2-10 to sensitize TNBC cell lines maintained as adherent, two-dimensional (2D) cultures to cisplatin and observed no substantial effect (fig. S2A). On the basis of compelling evidence that cells respond differently in a more physiological 3D environment (29), we prepared methylcellulose cultures of MDA-MB-231 cells for a 3D viability assay and used this assay to evaluate a role for NRP2 in chemoresistance. We found that aNRP2-10 had a significant (P < 0.0001 or P = 0.0013) effect on sensitizing these cells in 3D culture to cisplatin-induced cytotoxicity (Fig. 3A). Subsequently, we expanded the 3D viability assay to 11 other cell lines representing different subtypes of breast cancer (Fig. 3B). Consistent with our data on MDA-MB-231 cells, aNRP2-10 alone did not cause overt cytotoxicity in these cell lines (fig. S2B). However, it did increase sensitivity to cisplatin in a subset of these cells (termed responders), but it had no benefit in others (nonresponders). Further analysis revealed that the responders had higher EMT scores (30) and expression of NRP2 and ZEB1 genes [based on data from the Cancer Cell Line Encyclopedia database (CCLE)] compared with the nonresponders, indicating that aNRP2-10 increases the chemosensitivity of more mesenchymal TNBC cells with enriched NRP2 expression (Fig. 3B). The RNA sequencing data from other groups (31, 32) showed concurrent up-regulation of NRP2 and ZEB1 in chemoresistant breast cancer cells compared with control cells, suggesting the role and potential connection of these two genes in the chemoresistance of breast cancer (fig. S2, C and D).

Fig. 3. aNRP2-10 enhances chemosensitivity of mesenchymal breast cancer cell lines and TNBC patient–derived organoids.

(A) Ability of aNRP2-10 alone or in combination with cisplatin (at either C1 = 6.1 or C2 = 12.1 μM within IC₁₀ to IC₃₀ based on the cisplatin monotherapy tests) to reduce the viability of MDA-MB-231 cells grown in methylcellulose is shown. hIgG4 was used as an isotype control. Consistent results were obtained in three independent experiments, and representative data with n = 3 biological replicates are shown. Statistical significance was determined by two-sided, unpaired t test. **P < 0.01 and ****P < 0.0001. (B) Breast cancer (BC) cell lines covering Luminal, Her2⁺, and TNBC subtypes were used in methylcellulose 3D viability assays. Cells responding to aNRP2-10 treatment in the combination therapy with chemo drug cisplatin (responders) showed higher EMT scores and enriched gene expression of NRP2 and ZEB1 compared with nonresponder cells. EMT color scheme is based on EMT scores by Le et al. (30) (dark orange: 0.5 to 1, indicating mesenchymal cells; light orange: 0 to 0.5, and light blue: −0.5 to 0, indicating intermediate cells; dark blue: −1 to −0.5, indicating epithelial cells). Cell line gene expression data were from CCLE [expressed as Log₂(TPM + 1) in the file: OmicsExpressionProteinCoding-Genes TPMLogp1.csv downloaded from the CCLE website]. TPM, transcript per million. (C) NRP2 gene expression in chemo-resistant organoids (R) and chemosensitive organoids (S) derived from two TNBC patient tumors (TPDO-1 and TPDO-2) is shown. The chemoresistant organoids were generated by culturing them in the presence of either 5-FU or cisplatin for 2 weeks. Organoids were pooled from 24 wells, and qPCR was performed with technical triplicates. Statistical significance was determined by two-sided, unpaired t test comparing R and S organoids from the same tumors. ****P < 0.0001. (D) TPDO-1 TNBC organoids were treated with aNRP2-10 or aNRP2-14 and either 5-FU or cisplatin, and viability was compared with either treatment alone. n = 4 biological replicates. Statistical significance was determined by two-sided, unpaired t test. ****P < 0.0001. (E) Effect of NRP2 siRNA and aNRP2-10, either alone or in combination, on the sensitivity of TNBC organoids to 5-FU was measured and plotted as relative percentage viability. n = 3 biological replicates. Statistical significance was determined by two-sided, unpaired t test. ***P < 0.001; ns, not significant. KD, knockdown. (F) CSC and EMT markers were analyzed after aNRP2-10 treatment of the TPDO-1 organoids using Fluidigm 84-gene panels. Organoids were pooled from 24 wells, and the Fluidigm experiment was performed with technical triplicates. Green indicates >2-fold down-regulation, and orange indicates >2-fold up-regulation of gene expression by aNRP2-10 versus vehicle treatment. (G) Expression of CSC pluripotency, EMT, and epithelial markers was measured in the second patient-derived organoids, TPDO-2, by qPCR analysis after aNRP2-10 treatment. hIgG4 was used as an isotype control. Organoids were pooled from 24 wells, and qPCR was performed with technical triplicates. Statistical significance was determined by two-sided, unpaired t test. *P < 0.05, **P < 0.01, and ***P < 0.001. (H) Effect of overexpression of ZEB1 on TPDO-1 viability was measured for the indicated treatment groups. n = 3 biological replicates. Statistical significance was determined by two-sided, unpaired t test. **P < 0.01. Data are presented as means ± SEM.

A role for NRP2 in chemoresistance is substantiated by our observation that NRP2 and ZEB1 expression is significantly (P < 0.0001) higher in TNBC organoids that are resistant to chemotherapy than those that are sensitive (Fig. 3C). This observation prompted us to investigate the ability of aNRP2-10 to sensitize TNBC organoids to cisplatin and 5-fluorouracil (5-FU). Neither aNRP2-10 nor chemotherapeutic drugs alone had a robust effect on the viability of cells in the organoids. However, the combination of these drugs with aNRP-10 resulted in a significant (P < 0.0001) increase in cytotoxicity (Fig. 3D). We confirmed this finding in a second patient-derived organoid. Similarly, aNRP2-10 alone did not exhibit an effect on the viability, whereas its combination with 5-FU resulted in a significant (P < 0.0001) viability reduction of the organoid (fig. S2E). These results suggest that aNRP2-10 by itself is not overtly cytotoxic to the TNBC cells but that it enhances their sensitivity to chemotherapeutic drugs. In contrast, the semaphore-in-blocking antibody aNRP2-14 had no impact on the viability of either organoid, indicating a specific efficacy of aNRP2-10 on chemosensitization mediated through selective blocking of VEGF-NRP2 signaling.

The role of autocrine VEGF in chemoresistance was assessed by knocking down either VEGF-A or VEGF-C expression using small interfering RNA (siRNAs) in the TNBC organoids. We observed that down-regulation of either VEGF increased sensitivity to 5-FU (fig. S2F). To investigate the contribution of signaling independently of VEGF/NRP2, such as transforming growth factor–β (TGF-β), hepatocyte growth factor (HGF), and any potential role of cytoplasmic NRP2 in chemoresistance, we used siRNA to down-regulate endogenous NRP2 in TNBC organoids. Our results showed similar degrees of chemosensitization by either aNRP2-10 or NRP2 siRNA monotargeting, and no additive effect was observed by their combination (Fig. 3E). This result indicates that VEGF/NRP2 signaling is the major driver of chemoresistance in TNBC organoids and that the maximum effect on chemosensitization can be achieved by aNRP2-10.

To understand how aNRP2-10 enhances chemosensitivity in TNBC organoids, we performed gene expression studies. The expression of key CSC markers was reduced more than twofold in response to aNRP2-10 alone compared with control IgG (fig. S2G). In contrast, most of these CSC genes exhibited increased expression in response to 5-FU treatment, which is consistent with the hypothesis that chemotherapy selects for CSCs (33). For this reason, we focused on the ability of aNRP2-10 to alter gene expression in a more extended analysis covering 84 marker genes in both of the CSC and EMT pathways. aNRP2-10 reduced the expression of multiple CSC and EMT gene markers defined in the QIAGEN CSC/EMT pathway (Fig. 3F and table S4). These include critical CSC pluripotency markers and EMT markers such as ZEB1, a dual CSC/EMT marker that plays an important role in the EMT process (fig. S2H) (34, 35). Quantitative polymerase chain reaction (qPCR) validation further confirmed the down-regulation of ZEB1 by aNRP2-10 but not by the semaphorin-blocker aNRP2-14 (fig. S2I). Subsequently, we validated these gene expression data by quantifying the expression of key genes in a second TNBC organoid by qPCR. Consistently, the results demonstrate the down-regulation of CSC pluripotency markers and EMT markers and up-regulation of certain epithelial markers (Fig. 3G). Next, we validated aNRP2-10 inhibition of ZEB1 at the protein level (fig. S2J) and showed that the overexpression of ZEB1 abolished the chemosensitizing effect of aNRP2-10 (Fig. 3H and fig. S2K). The expression of the genes encoding VEGFR2 (KDR) and VEGFR3 (FLT4) was not detectable in the TNBC organoids (fig. S2L). Similarly, the expression of KDR and FLT4 was low in most TNBC cells that responded to aNRP2-10 in this study (table S5), suggesting that the effect of aNRP2-10 on CSC/EMT inhibition and chemosensitization was mediated through a VEGFR-independent pathway. To investigate the mechanism by which VEGF/NRP2 signaling sustains ZEB1 expression to promote chemoresistance, we focused on focal adhesion kinase (FAK) because we had shown previously that the VEGF/NRP2/FAK pathway is critical for the function of TNBC CSCs (5, 10, 36). To assess this possible mechanism, we inhibited FAK in TNBC organoids and observed a significant (P = 0.0225) reduction in ZEB1 expression (fig. S2M). These results support that aNRP2-10 enhances the TNBC chemosensitivity by the inhibition of CSC/EMT pathway and agree with the earlier data of responsive breast cancer cells being more mesenchymal.

aNRP2-10 sensitizes TNBC to chemotherapy and reduces spontaneous metastasis in vivo

Next, we performed in vivo studies with TNBC xenograft models to evaluate the efficacy of aNRP2-10 in monotherapy and combination therapy with chemotherapeutic drugs. First, given that NRP2 is expressed on tumor cells as well as stromal and immune cells in the tumor microenvironment (9), we sought to compare the effect of blocking NRP2 specifically on tumor cells alone (using human-specific aNRP2-10) versus blocking NRP2 on cells in the tumor microenvironment using the mouse-specific aNRP2-28. Neither treatment resulted in substantial tumor growth inhibition in the MDA-MB-231 tumor xenograft model compared with the IgG control in this model (fig. S3A). This lack of a monotherapy effect is consistent with our in vitro finding that aNRP2-10 monotherapy itself is not cytotoxic. However, given that aNRP2-10 used in combination with chemotherapy serves to sensitize tumor cells to chemotherapy-induced cytotoxicity in vitro (Fig. 3), we next explored whether there is a potential synergy of NRP2 blockade and cisplatin treatment with respect to tumor growth inhibition in the MDA-MB-231 xenograft model. The mice in this study were given three rounds of cisplatin with decreasing concentrations in combination with either IgG control or NRP2 blocking antibodies (aNRP2-10 blocking hNRP2 and aNRP2-28 blocking mNRP2). The aNRP2/cisplatin combination group showed a 67% reduction in tumor growth compared with the combination of cisplatin and IgG control (Fig. 4A). The tumor control index [comprising tumor inhibition score, tumor rejection score, and tumor stability score (37)] was enhanced eightfold compared with the IgG/cisplatin treatment (Fig. 4B). In a separate study, addition of aNRP2-28 to the combination of aNRP2-10 and cisplatin did not further enhance the antitumor effect (fig. S3B). These data, in addition to our aNRP2-28 monotherapy data, suggest that the observed synergistic effect between NRP2 blockade and cisplatin is mainly driven through a direct effect of aNRP2-10 on the cancer cells. We then performed gene expression analysis on the MDA-MB-231 xenografts collected at the end of the study and found that the CSC/EMT markers POU5F1, VIM, and ZEB1 were down-regulated by aNRP2-10 monotherapy compared with control IgG (Fig. 4C).

Fig. 4. aNRP2-10 enhances chemosensitivity and reduces metastasis in TNBC models in vivo.

(A and B) Tumor growth (A) and tumor control index (B) for MDA-MB231 tumor-bearing mice are shown. Mice were randomized and received their first dose on day 19 after tumor implantation. They were dosed with control IgG (25 mg/kg) or aNRP2-10 (25 mg/kg) and aNRP2-28 (25 mg/kg), a murine surrogate mAb of aNRP2-10, in combination with cisplatin (n = 9 mice in each group). Treatments were administered twice weekly by intraperitoneal injection until day 32 after the first dose. After termination of dosing, the animals were monitored for two more weeks until day 46 (recovery period). In (A), statistical significance was determined by multiple comparisons two-way ANOVA (**P < 0.01, ***P < 0.001, and ****P < 0.0001). (C) Gene expression analysis from an aNRP2-10 monotherapy study separate from the study for (A) and (B) is shown. Mice were injected intraperitoneally twice weekly from day 7 after inoculation until day 59. Tumors allocated for analysis were about 800 mm³. Terminal tumor samples collected from MDA-MB-231 tumor–bearing mice treated with either control IgG (10 mg/kg) or aNRP2-10 (10 mg/kg) were used for analysis. n = 3 mice in each group. Three technical replicates were analyzed for each mouse, and mean was used in the plot. Statistical significance was determined by two-sided, unpaired t test. *P < 0.05 and **P < 0.01. (D and E) Tumor growth (D) and tumor control index (E) for Hs578T tumor–bearing mice are shown, and survival is shown in fig. S3C. Mice were randomized to four groups and treated by intraperitoneal injection after tumor implantation with control IgG (10 mg/kg) or aNRP2-10 (10 mg/kg) in combination with 5-FU. hIgG4 group: n = 9; other groups: n = 10 mice in each group. Treatments were administered twice weekly. In (D), statistical significance was determined by multiple comparisons two-way ANOVA (*P < 0.05 and ****P < 0.0001). (F) Gene expression analysis is shown for tumor samples collected 24 hours after the last treatment from Hs578T tumor–bearing mice treated with control IgG or aNRP2-10 in the same study as (D) and (E). n = 3 mice in each group. Three technical replicates were analyzed for each mouse, and mean was used in the plot. Statistical significance was determined by two-sided, unpaired t test. *P < 0.05. (G) Spontaneous metastasis to the lung was quantitated in 4T1 tumor–bearing mice treated with control IgG (25 mg/kg) or aNRP2-28 (25 mg/kg) twice weekly by intraperitoneal injection starting 1 day before cell inoculation. The study was terminated, and lungs were harvested on day 23 after cell inoculation. mIgG1 group: n = 8; aNRP2-28 group: n = 10. (H) Representative histology (left) and frequency of lung metastasis (right) are shown for MDA-MB-231 tumor–bearing humanized mice (average humanization ratio = 70%) treated with control IgG (25 mg/kg) or aNRP2-10 (25 mg/kg) twice weekly by intraperitoneal injection. Mice were randomized and received their first dose on day 15 after inoculation. Study animals were treated for 3 weeks until day 36 after inoculation, when the study was terminated and lungs were harvested for histological analysis. hIgG4 group: n = 4; aNRP2-10 group: n = 5. Statistical significance was determined by two-sided, unpaired t test. **P < 0.01. Data are presented as means ± SEM.

Subsequently, we used another TNBC xenograft model (Hs578T) to assess the ability of aNRP2-10 to sensitize tumors to a different chemotherapeutic drug (5-FU). We observed that aNRP2-10 enhanced the efficacy of 5-FU in tumor growth inhibition (Fig. 4D). Whereas 5-FU or aNRP2-10 showed little effect as a monotherapy, the combination of 5-FU and aNRP2-10 significantly (P < 0.05 or P < 0.0001) reduced tumor growth (Fig. 4D) and increased the tumor control index 10-fold in comparison with 5-FU alone (Fig. 4E) and enhanced the survival rate (fig. S3C). Tumor samples from the Hs578T xenograft model showed the reduction of NANOG and ZEB1 expression after aNRP2-10 treatment compared with control (Fig. 4F). This agrees with our findings in vitro and in MDA-MB-231 xenografts and supports the mechanism of action for aNRP2-10 through inhibiting the CSC/EMT pathways.

Given that aNRP2-10 monotherapy down-regulates CSC/EMT pathways that play an important role in metastasis, we further investigated the antibody effect on lung metastasis using both xenograft and syngeneic models of TNBC (Fig. 4, G and H). We did not observe a reduction in the number of metastatic nodules per lung section in the syngeneic 4T1 model (Fig. 4G). However, we found that aNRP2-10 lead to a significant (P = 0.0076) inhibition of spontaneous lung metastasis in a MDA-MB-231 xenograft tumor model in humanized mice (Fig. 4H). In addition, to evaluate whether the function-blocking NRP2 mAbs that we used (aNRP-10 or aNRP2-28) affect the lymphatic system, we tested aNRP2-28 in a murine model of corneal injury and observed no difference in neolymphangiogenesis compared with the control IgG group (fig. S3D).

Last, we evaluated aNRP2-10 in a patient-derived xenograft (PDX) model. aNRP2-10 alone or a suboptimal dose of 5-FU had little effect on tumor burden. However, aNRP2-10 in combination with 5-FU robustly inhibited tumor growth (Fig. 5A). Gene expression studies also demonstrate down-regulation of critical CSC/EMT genes NANOG, VIM, and ZEB1 (Fig. 5B). Collectively, the in vitro and in vivo data presented here demonstrate that aNRP2-10 enhances chemosensitivity and reduces spontaneous metastasis in TNBC (fig. S3E).

Fig. 5. aNRP2-10 sensitizes PDXs to chemotherapy and inhibits CSC and EMT marker gene expression.

(A) Effect of aNRP2-10 on sensitizing PDXs, generated by implanting organoids in mammary fat pads of NSG mice, to 5-FU chemotherapy is shown. n = 4 mice in each group. Mice were injected intraperitoneally weekly. Statistical significance was determined by multiple comparisons two-way ANOVA (*P < 0.05, ***P < 0.001, and ****P < 0.0001). (B) qPCR was used to quantitate EMT markers in PDX tumors treated with aNRP2-10 after 8 weeks of treatment in three mice per treatment group. Tumor samples were collected 24 hours after the last treatment. Three technical replicates were analyzed for each mouse, and mean was used in the plot. NANOG and VIM expression in a hIgG4-treated mouse was greater than mean + 2 SD, which were defined as outliers and excluded from the plot. Statistical significance was determined by two-sided, unpaired t test. *P < 0.05 and **P < 0.01. Data are presented as means ± SEM.

aNRP2-10 has no observed adverse effects in a nonhuman primate toxicity study

In a 28-day repeat dose Good Laboratory Practice (GLP) toxicity study in cynomolgus monkeys, aNRP2-10 [test article (TA)] was well tolerated. No TA-related effects were noted in ophthalmic or electrocardiogram examinations; body weights or body weight changes; or hematology, coagulation, clinical chemistry, urinalysis, organ weights, macroscopic, or microscopic observations. There were no TA-related changes in peripheral blood white blood cell subsets or cytokine expressions. The no observed adverse effect level was 200 mg/kg per week, the highest dose tested (fig. S4 and table S6).

DISCUSSION

There is compelling evidence to support that blocking the binding of VEGF to NRP2 could be an effective therapeutic approach for improving the clinical management of breast and other cancers. For example, the use of anti-NRP2 antibodies has been shown to inhibit aggressive cancer properties in murine models of breast cancer and glioma (38) as well as pancreatic cancer xenografts (39). However, to date, well-characterized, highly specific, and clinical grade anti-NRP2 antibodies have not been available for clinical trials. Here we describe for the first time a clinical-grade, high-affinity molecule that blocks VEGF/NRP2 signaling with demonstrated preclinical efficacy and a favorable safety profile, as demonstrated in a 28-day repeat dose GLP toxicity study in cynomolgus monkeys. We generated and characterized a fully humanized mAb (aNRP2-10) that has promise as a therapeutic agent for TNBC and potentially other aggressive cancers that frequently develop resistance to chemotherapy. A distinguishing feature of aNRP2-10 is that it binds to a distinct epitope on NRP2 compared with the existing NRP2 antibodies, which likely accounts for its ability to selectively inhibit the binding of VEGF ligands to NRP2 without disrupting the interaction between NRP2 and semaphorins. The biochemical properties of aNRP2-10 that we have characterized underlie its ability to target CSC cell populations with high NRP2 expression and inhibit the ability of NRP2 to sustain EMT properties in these cells. Moreover, we have demonstrated that aNRP2-10 sensitizes chemoresistant cells and tumors to both cisplatin and 5-FU treatment by diminishing stemness and EMT properties. Overall, our data provide strong support for the use of aNRP2-10 in clinical trials aimed at improving the response of patients to chemotherapy.

Our approach of directly targeting NRP2 using aNRP2-10 as a therapeutic strategy is based on the realization that tumor cells, especially CSCs, express high concentrations of NRP2 and are dependent on NRP2/VEGF signaling to sustain their more mesenchymal properties (10, 40), independent of the role of VEGF receptors in angiogenesis (40–43). Although the use of angiogenesis inhibitors such as bevacizumab as a single therapy or in combination with chemotherapeutic drugs has shown efficacy in progression-free survival, bevacizumab has not been effective for the overall survival in breast cancer (44, 45). In this context, it is worth noting that bevacizumab inhibits the binding of VEGF-A to VEGF receptor tyrosine kinases, notably VEGFR2, but it does not block the binding of VEGF-C to NRPs or VEGFRs (46). For this reason, it has minimal direct impact on CSCs; several studies have concluded that VEGFR2 signaling does not contribute to the function of these cells and that VEGF signaling is driven largely by the NRPs, especially NRP2 (40–42). Our data here substantiate the observation that VEGFR2 expression is low in TNBC organoids and cell lines. Therefore, the clinical success of bevacizumab is likely attributed to its impact on angiogenesis (45), whereas its ultimate failure to improve overall survival may be due, in part, to its inability to effectively target the cancer cell population that drives the cancer to an aggressive state. Despite this information, targeting NRPs on tumor cells has not been pursued as a potential therapeutic strategy in patients for several reasons. These include the early failure of Genentech’s anti-NRP1 antibody in combination with bevacizumab (47) and the prevailing notion that anti-VEGF therapies for cancer-targeting angiogenesis lack long-term efficacy in TNBC (48). Further, targeting NRPs has been hampered by the absence of an effective reagent that specifically and effectively inhibits the binding of VEGFs to NRP2 on tumor cells and that is capable of effectively sensitizing those tumor cells to chemotherapy. The data presented in this study demonstrate that aNRP2-10 meets these needs and represents a therapeutic strategy that merits clinical evaluation in patients with TNBC.

A question that arises from our data is whether the function-blocking NRP2 mAbs that we used (aNRP-10 or aNRP2-28) affect the lymphatic system based on the reported involvement of NRP2 in lymphangiogenesis (49, 50). In the 4-week repeat dose GLP toxicity study in nonhuman primates, aNRP2-10 was tested at doses of up to 200 mg/kg, and no signs of lymphedema were observed. NRP2 KO mice have impaired lymphatic vessel development during early development, but the adult mutant mice exhibit normal lymphatic drainage and show no signs of edema (49). This result indicates that a complete loss of NRP2 does not affect the overall functionality of the lymphatic system. Therefore, we believe that aNRP2-10 therapy may primarily affect the process of neovascularization upon insult or tumor development and not adversely affect established lymphatic vessels, but this possibility would be closely monitored in human clinical trials.

The mechanism of action of aNRP2-10 merits discussion because this mechanism relates to the plasticity of tumor cells, a key issue for understanding the biology of tumors and for developing effective therapies. More specifically, TNBC is characterized by a high degree of intratumoral heterogeneity that is manifested by the coexistence of subpopulations of tumor cells that range in their degrees of differentiation and other properties on the basis of the microenvironment surrounding them (51). Considerable evidence supports the hypothesis that the more poorly differentiated slower-growing subpopulations of cells have stem cell properties that contribute to the aggressive behavior of the entire TNBC population by providing chemoresistant cells with the ability to repopulate and promote tumor recurrence (52, 53). There is also evidence that plasticity exists in these subpopulations and that poorly differentiated cells can acquire more differentiated characteristics in response to changes in the tumor microenvironment or therapy (53). The strategy of promoting differentiation as an approach to improving therapeutic response is gaining acceptance (54, 55). The data we present here reveal that aNRP2-10 functions in this manner by stimulating the differentiation of subpopulations of TNBC and diminishing their stem cell properties. Consequently, these cells remain viable but are more sensitive to chemotherapy. A particular strength of our work is that we have generated a highly specific mAb that functions in vivo to promote differentiation and enhance chemosensitivity.

Our data revealed that aNRP2-10 suppresses the expression of stem cell and pluripotency genes as well as genes associated with a mesenchymal phenotype, and it increases the expression of key epithelial genes, including CD24 and ESRP1. Among the genes that are repressed by aNRP2-10, we focused on ZEB1 for several reasons. First and foremost, ZEB1 is a “master regulator” of the CSC and EMT states. A seminal study in this regard demonstrated that the induction of ZEB1 transcription converts non-CSCs into CSCs and promotes an EMT in basal breast cancer, which comprises most TNBCs (56). This study also highlighted the plasticity of the non-CSC and CSC states in basal breast cancer. ZEB1 represents a prime therapeutic target for these reasons. Our finding that aNRP2-10 suppresses ZEB1 expression provides a mechanism for how this mAb sensitizes TNBC to chemotherapy, even in immunosuppressed models, and justifies its use as a therapy for this disease. It is worth noting that there is also evidence that ZEB1 can regulate NRP2 expression in lung cancer cells (57), which suggests that a positive feedback loop exists between NRP2 and ZEB1.

We observed that monotherapy using aNRP2-10 was sufficient to cause a decrease in metastasis in the humanized MDA-MB-231 xenograft tumor model. The antimetastatic effect of aNRP2-10 is consistent with the report that the dissemination of cancer cells is an early event in breast tumorigenesis and that these disseminated cells have CSC- and EMT-like features (58, 59). Although the role of EMT in metastatic cancer cell dormancy is not well understood, there is evidence to suggest that ZEB1 can reactivate dormant disseminated cancer cells to promote metastatic colony formation (60). Our finding that aNRP2-10 suppresses ZEB1 expression may suggest that the observed inhibition of metastasis is due to the inability of ZEB1 to promote the reactivation of dormant disseminated cancer cells.

A limitation of our study is that, although the ability of the function-blocking NRP2 mAbs to sensitize tumors to chemotherapy was tested in multiple xenograft models, further work using other models of TNBC, including genetically engineered mouse models, could strengthen the rationale for use of aNRP2-10 in TNBC patients clinically. Second, our in vivo studies were performed using 5-FU or platinum-based chemotherapy, both of which are used in the treatment of TNBC. However, subsequent studies could expand to evaluate other drugs that are standard of care for TNBC such as anthracyclines and taxanes to provide further guidance for clinical protocol design. Last, although the expression of NRP2 and the frequency of CSCs is lower in other subtypes of breast cancer than in TNBC, it will be worth investigating the therapeutic potential of our mAbs in these subtypes.

The mechanism proposed for aNRP2-10 in TNBC could also apply to other cancer types where NRP2 expression is associated with cancer progression, metastasis, therapy resistance, and poor prognosis (6). These include a variety of cancer types such as non-small cell lung cancer and clear cell renal cell carcinoma. Expression of VEGF-C and related genes has been linked to induction of EMT and chemoresistance in breast, bladder, and skin cancers (61–63). Furthermore, VEGF-C/NRP2 signaling was reported to mediate tumor growth and metastasis through promoting EMT-epithelial breast cancer cell cross-talk (16). More work is needed to characterize the activity of aNRP2-10 (named ATYR2810 for the clinical compound made under guanosine 3′,5′-monophosphate conditions) in other cancer types, but it may have the potential to sensitize a variety of solid tumor types to chemotherapy and may be particularly useful in a chemotherapy-sparing setting. We also note the ability of our VEGF/NRP2 function-blocking mAbs to activate antitumor immunity in prostate cancer. In a separate study, we reported that VEGF/NRP2 signaling sustains programmed death-ligand 1 expression and that its inhibition using these mAbs increases immune-mediated tumor cell killing (64).

In summary, we have developed a highly specific and effective mAb that blocks the binding of VEGFs to NRP2. This mAb, aNRP2-10, inhibits CSC properties and EMT features and promotes a more differentiated phenotype in TNBC by a mechanism that involves its ability to suppress ZEB1 expression by blocking VEGF/NRP2/FAK signaling. Other VEGF/VEGFR-targeting anticancer therapies, such as anti-VEGFA mAbs and tyrosine kinase inhibitors, do not inhibit this signaling axis (fig. S3E). Moreover, the specificity of aNRP2-10 enables us to conclude that these functions of NRP2 are dependent on its binding to VEGFs and not to semaphorins. As a consequence of its mechanism of action, aNRP2-10 sensitizes TNBC to chemotherapy. Overall, these studies validate the approach of targeting NRP2 on tumor cells as a therapeutic strategy and support the evaluation of aNRP2-10 in clinical trials aimed at improving the therapeutic response of TNBC to chemotherapy. More specifically, patients with TNBC who had previously failed chemotherapy or immunotherapy and whose tumors highly express NRP2 may benefit from treatment with aNRP2-10 in combination with drugs such as capecitabine, a pro-drug of 5-FU and a treatment option for TNBC patients with advanced disease (65). On the basis of the proposed mechanism of action, addition of aNRP2-10 to capecitabine (or a comparable chemotherapeutic drug) may lead to a deeper and prolonged clinical response in TNBC patients.

MATERIALS AND METHODS

Study design

This study was designed to generate mAbs that inhibit the binding of VEGF to NRP2 specifically and that can be used in clinical trials. The mAbs were generated and characterized for inhibition of either VEGF or Sema3F binding to NRP2. We determined the VEGF/NRP2 inhibitory mAb (aNRP2-10) but not the semaphorin/NRP2 mAb (aNRP-14) selected for CSCs in TNBC and inhibited their self-renewal. We also evaluated the ability of aNRP2-10 to sensitize TNBC to chemotherapy and inhibit spontaneous metastasis using in vitro and in vivo models. All raw, individual-level data are presented in data file S1. Reagents used in the study are presented in data file S2.

The therapeutic mAbs were generated by aTyr Pharma Inc. and provided to the University of Massachusetts (UMass) Chan Medical School. The mAbs for in vitro mammosphere formation assays were labeled with codes for blind tests, and the decoding was done only after the submission of the results to aTyr. For animal studies performed at UMass Chan Medical School, the sample size was determined by the Bioinformatics Core in the Department of Molecular, Cell and Cancer Biology to ensure that all studies were sufficiently powered, and data were properly analyzed and interpreted. In animal studies, all treatment and control groups had 4 to 10 mice per group (indicated in the figure legends) and were randomized according to tumor volume at the start of treatment (about 100 mm³). Body weight and tumor size were measured thrice weekly, and mice were euthanized when tumor volume reached 1000 mm³. If mice were euthanized before 20 days from the start of the treatment due to large tumor size, then the data were derived from the extrapolated tumor size at that day. The primary end point for animal studies was tumor volume. All animal use was in accordance with the guidelines of the Animal Care and Use Committee of the UMass Chan Medical School and conformed to the recommendations in the Guide for the Care and Use of Laboratory Animals (Institute of Laboratory Animal Resources, National Research Council, National Academy of Sciences, 1996). All animal study protocols (UMMS protocol ID: PROTO202000145) were approved by the UMass Medical School Institutional Animal Care and Use Committee (IACUC). Biological replicates for experiments are indicated in the figure legends.

All study protocols for in vivo studies performed at aTyr Pharma were governed and approved by the Explora Biolabs’ IACUC (#EB17-010). In vivo studies performed at aTyr Pharma had 10 mice per treatment group, except for the MDA-MB-231 xenograft study in humanized mice, which had five mice per treatment group. In tumor studies, mice were randomized into treatment groups when tumor volumes reached 40 to 100 mm³. The 4T1 syngeneic tumor model and the MDA-MB-231 xenograft studies in non-humanized mice were conducted in a blinded manner. Body weights were taken once a week, and tumors were measured two or three times per week depending on the study. Animals were euthanized before the scheduled study termination if tumor volumes exceeded 1500 mm³ or an animal was found moribund. Mice with a greater than 20% body weight loss from baseline were carefully assessed for potential euthanasia. The primary end point for the tumor studies was tumor volume, and the secondary end point was metastatic spread to the lung.

The neovascularization study was performed at Comparative Biosciences Inc. Each treatment group had 10 animals. The primary end point was quantification of Lyve-1⁺ vessels in the cornea by immunohistochemistry.

Generation of aNRP2 monoclonal antibodies

We first generated mouse anti-hNRP2 mAbs by immunizations and hybridoma fusion. SJL/J mice were immunized with Expi293F cells [suspension-adapted human embryonic kidney (HEK) 293 cells; Thermo Fisher Scientific] stably expressing cytomegalovirus-driven NRP2 (v2, NM_003872.3) and boosted with a recombinant fusion protein, NRP2 (v2, amino acids 23 to 855)-Fc. Hybridoma supernatants were then screened by enzyme-linked immunosorbent assay (ELISA) to confirm their binding to recombinant hNRP2 ECDs (containing a1, a2, b1, b2, and c domains). Epitope mapping was also performed to identify the antibody-targeted domain(s) on hNRP2 by ELISA or SPR measurement. After the initial screening, we prioritized a panel of aNRP2 mAbs with strong hNRP2 binding and distinct epitopes, including aNRP2-10 and aNRP2-14. These antibodies were humanized by complementarity-determining region grafting onto human frameworks, and affinity was matured by NNK (where N = A/C/G/T and K = G/T) mutagenesis and somatic mutation transfer as described in Burman et al. (66).

Recombinant proteins

hNRP1 (NP_001019799) and hNRP2 (NP_003863) were cloned into mammalian expression vectors and recombinantly expressed in Expi293F cells. The hNRP1-ECD (amino acids 22 to 602) and hNRP2-ECD (amino acids 23 to 855) proteins used in ELISA binding and SPR affinity measurements were fused to a His tag and were purified on a Ni Sepharose Excel column (Cytiva). The hNRP2-a1a2b1b2c (amino acids 23 to 855), hNRP2-a1a2b1b2 (amino acids 23 to 595), hNRP2-a2b1b2 (amino acids 145 to 595), hNRP2-b1b2 (amino acids 276 to 595), and hNRP2-b1 (amino acids 276 to 430) proteins used in epitope mapping studies were fused to a human Fc-tag. These proteins were purified on a HiTrap MabSelect column (Cytiva). Human Sema3F from the mature N terminus to the C-terminal furin cleavage site (amino acids 19 to 779) was fused to a signal peptide from the SPARC gene (MRAWIFFLLCLAGRALA), a 6xHis-Myc tag (HHHHHHEQKLISEEDLG), and a GS linker (GGGGS) at the N terminus of Sema3F. The protein was mutated at R583A and R586A to remove an internal furin cleavage site and subsequently mutated to remove the 31–amino acid fragment modifying it from the isoform 1 (NP_004177.3) to isoform 2 (NP_001305729.1), which showed a higher potency than the isoform 1 in the NRP2/co-receptor dimerization assay. Untagged human VEGF-A or VEGF-C recombinant proteins were purchased from R&D Systems (#293-VE and #9199-VC-025).

Binding assay and epitope mapping by ELISA

A 96-well immuno-4HXB plate (Thermo Fisher Scientific) was coated with NRP1-ECD or NRP2-ECD diluted to 2 μg/ml in 1× phosphate-buffered saline (PBS) pH 7.4 (Gibco), with 50 μl per well and overnight incubation at 4°C. After blocking with casein (100 μl per well; Thermo Fisher Scientific), the ELISA plate was washed three times with a wash buffer (0.05% Tween 20 in 1× PBS), and antibody titrations (50 μl per well) for aNRP2-10, aNRP2-14, and hIgG4 were added to the assay plate. After incubation for 1 hour at room temperature with shaking (400 rpm), the plate was washed, and goat anti-human horseradish peroxidase (HRP; 50 μl per well; Jackson ImmunoResearch) was added at a 1:5000 dilution into 1% bovine serum albumin (BSA) in PBS. After incubation for 1 hour, the plate was washed, and bound antibody signals were developed using 1-Step Ultra 3,3′,5,5′-tetramethylbenzidine substrate (50 μl per well; Thermo Fisher Scientific) and stopped by the addition of stop solution (50 μl per well; BioL-egend). Optical density signals were measured at 450 nm using a PowerWave HT plate reader (BioTek).

For epitope mapping, 96-well immuno-4HXB plates (Thermo Fisher Scientific) were coated with hNRP2-a1a2b1b2c, hNRP2-a1a2b1b2, hNRP2-a2b1b2, hNRP2-b1b2, and hNRP2-b1 recombinant proteins as well as mNRP2 (SinoBio) and rat NRP2 (R&D Systems). The ELISA binding assay was performed as described above with aNRP2-10 and aNRP2-14 antibodies diluted 1:100, 1:1000, and 1:5000 in casein (a buffer blank well was also included). After the blocking step was complete, plates were washed three times with wash buffer and antibody dilutions (50 μl per well) were added to the assay plate.

Epitope mapping and binding kinetics by SPR

SPR experiments were performed on a Bio-Rad ProteOn XPR36 instrument. The running buffer was 20 mM Hepes, 300 mM NaCl, 5 mM CaCl₂, and 0.005% Tween 20 (pH 7.4), and the regeneration buffer was 10 mM glycine (pH 2.0). For epitope mapping, goat anti-mouse IgG Fcγ (Jackson ImmunoResearch) was covalently immobilized on a ProteOn GLC chip (Bio-Rad) through 1-ethyl-3-(3-dimethylaminopropyl)carbodiimid/N-hydroxysuccinimide coupling. The aNRP2 antibodies were captured on this anti-mouse surface. Each recombinant NRP2 protein (50 nM) was flowed over the antibodies. For affinity measurements, CaptureSelect Human Fab-kappa Kinetics biotin-conjugated antibody (Thermo Fisher Scientific) was captured and immobilized on a ProteOn NLC chip (Bio-Rad). The humanized anti-NRP2 antibodies were subsequently captured on this anti-human surface. A titration of hNRP2-His (150, 50, 16.67, 5.56, and 1.85 nM) was flowed as analyte over the antibodies. Data were double-referenced against a surface with no aNRP2 antibody captured (immobilized anti-human antibody only) and a buffer-only blank. Affinity constants were derived by globally fitting to a 1:1 binding model in the ProteOn Manager software.

aNRP2 binding to NRP2 on cells and blocking of ligand binding

The Expi293F-hNRP2 clonal cells were generated by the transfection of a linearized plasmid encoding the untagged hNRP2 (v2, NM_003872.3) into Expi293F cells. Transfected cells were selected with G418 (ProteOn Manager) for 3 weeks, followed by limiting dilution into 96-well plates to grow single clones. Clonal cells that overexpress hNRP2 were validated by aNRP2 cell surface staining followed by flow cytometry analysis. To test the binding of aNRP2 to Expi293F-hNRP2 clonal cells, cells were added to a 96-well V-bottom plate at 100,000 cells per well in 1× Dulbecco’s PBS (DPBS) and stained with Zombie Violet viability stain as per manufacturer’s instructions (BioLegend). aNRP2 antibodies were tested at final concentrations of 0.006 to 30 nM diluted in the wash buffer [FWB, DPBS plus 2% fetal bovine serum (FBS), and 0.1% sodium azide to inhibit receptor internalization] and detected by Cy3-conjugated goat anti-human IgG Fc (Jackson ImmunoResearch). All samples were acquired on the CytoFLEX flow cytometer (Beckman). Data were analyzed by FlowJo software (BD Biosciences) with gated live singlet cells and plotted using GraphPad Prism. A four-parameter variable-slope curve was fitted to the data ([agonist] versus response) using nonlinear regression, and the EC₅₀ for each curve was determined.

A549 KO clonal cells were generated by the CRISPR gene KO approach using a mix of three plasmids encoding Cas9 guide RNAs targeting various loci in the Exon2 of human NRP2 gene (Applied Biological Materials). Clonal cells were generated by limited dilution. Complete KOs were validated by genomic sequencing of NRP2 that confirmed frame-shifting deletions in Exon2, Western blot analysis of NRP2 protein expression in total lysates using a polyclonal aNRP2 antibody (R&D Systems, #AF567), and aNRP2 cell surface staining followed by flow cytometry analysis. Binding of aNRP2 at a concentration of 10 μg/ml to A549 WT or KO clonal cells was performed as described above for cell surface aNRP2 staining.

Binding of recombinant human untagged VEGF-C or Myc-tagged Sema3F to Expi293F-NRP2 cells was determined by immunostaining and flow cytometry. Studies were conducted with cells incubated with each of these proteins at a series of concentrations on ice for 30 to 40 min. VEGF-C bound on the cell surface was probed by a rabbit anti–VEGF-C antibody (Abcam #AB9546), followed by the detection by an AF647-conjugated goat anti-rabbit IgG (Jackson ImmunoResearch). Sema3F was detected by an AF555-conjugated anti-Myc antibody (Thermo Fisher Scientific). Binding curves were fit, and EC₅₀ values were determined as described above. To analyze aNRP2 blocking of ligand binding, Expi293F-hNRP2 cells were first incubated with aNRP2 for 30 min on ice. Without washing, VEGF-C at 10 nM or Sema3F at 15 nM was added to the cells. After 40-min incubation on ice, cells were washed, and the bound ligands were detected as described above. Data analysis was performed using GraphPad Prism. A four-parameter variable-slope curve was fitted to the data ([Inhibitor] versus response) using nonlinear regression, and the IC₅₀ for each curve was determined.

Ligand-induced NRP2/co-receptor dimerization

Receptor dimerization was monitored through complementation of a split luciferase. cDNA was obtained for NRP2 (RC220706, OriGene), KDR (GenScript, OHu27183), FLT4 (R&D Systems, RDC1286), and PLXNA1 (R&D Systems, RDC0967). The ECDs and transmembrane helices of each receptor were cloned into pBiT vectors containing both C-terminal spacers and either the large (LgBiT, 18 kDa) or small (SmBiT, 1.3 kDa) fragment of the NanoLuc Luciferase (Promega). Plasmids were transfected into Expi293F cells with ExpiFectamine (Thermo Fisher Scientific) at a reduced density of 1 million cells per ml and then harvested about 20 hours after transfection. Cell were counted, washed once with assay buffer (Live Cell Imaging Solution, Thermo Fisher Scientific), and plated in luminometer plates at 100,000 cells per well. Cell-permeable luciferase substrate was added, baseline was established by reading on a GloMax96 (Promega) until luminescence peaked, then ligand was added and luminescence was recorded over time. To calculate a normalized response, each individual well was normalized to its baseline value, and then triplicates were normalized to a no-ligand control to remove baseline drift due to reagent decay. Blocking of ligand-induced receptor dimerization was performed similarly to dimerization experiments. Cells were transfected and harvested as described, but once baseline was established, antibodies were added, and a new baseline was established. Ligand was then added at a subsaturating concentration (VEGF-A at 2 nM, VEGF-C at 2 or 20 nM, and Sema3F at 200 nM), and luminescence was again monitored over time.

Protein purification and crystallization

His-tagged hNRP2 (amino acids 146 to 595) was expressed in Expi293 cells (Thermo Fisher Scientific), and aNRP2-10 Fab with a His-tagged heavy chain and untagged light chain was expressed in ExpiCHO cells (Thermo Fisher Scientific). The hNRP2 (amino acids 146 to 595) and aNRP2-10 Fab were separately purified over Ni Sepharose Excel (Cytiva) columns. The purified proteins were mixed together at a ratio of 20 mg of hNRP2 to 23 mg of Fab to form a complex. The hNRP2-Fab complex was purified over a POROS XS (Thermo Fisher Scientific) cation exchange column followed by size exchange chromatography on a Superdex 200 (Cytiva) column and concentrated to 19.4 mg/ml in 50 mM tris and 150 mM NaCl (pH 7.4). The proteins were initially screened using the sitting drop method (0.5 μl of protein plus 0.5 μl of precipitant) and incubated at 16°C. After 2 to 4 days of incubation, the complex crystal grew in the condition containing 21% (w/v) polyethylene glycol, molecular weight 1000; 0.1 M sodium citrate tribasic dihydrate (pH 5.5); and 0.05 M lithium sulfate monohydrate. However, the crystals were too small to be used for diffraction. These crystals were collected as seed stocks and used to do matrix seeding. To set up matrix seeding, the seed stock was diluted by mother liquid with several gradients. The proteins were mixed with the diluted seeds and precipitant at a volume ratio of 1 to 0.33 to 0.66. The condition of mother liquid remained the same as before. After 1 week, larger crystals grew in these conditions. The crystals were cryoprotected in the same condition with the addition of 25% (v/v) glycerol before being flash-cooled and stored in liquid nitrogen.

The purified protein complexes of hNRP2 (v2, amino acids 146 to 595) and aNRP2-14 Fab were concentrated to 18.1 mg/ml in 1× PBS (pH 7.4). The proteins were screened for crystallization hits by the sitting drop method (0.5 μl of protein plus 0.5 μl of precipitant). After 2 to 4 days of incubation at 16°C, the complex crystals grew in the condition containing 30% (v/v) polyethylene glycol, molecular weight 300, and 0.1 M sodium acetate trihydrate (pH 4.5). Crystals were cryoprotected in the same conditions with the addition of 15% (v/v) glycerol before being flash-cooled and stored in liquid nitrogen.

Crystal structure determination

All of the diffraction data of the complex crystal were collected on a PILATUS3 6M detector at the Shanghai Synchrotron Radiation Facility beamline BL19U, at 100 K, and indexed and processed with the HKL3000 software (67). The structure of the complex was determined by molecular replacements with phaser (68) in CCP4 suite (69) using the previous published models as the search model [PDB codes: 2QQK (NRP2) and 2D03 (Fab)]. The final models were generated through multiple steps of building in Coot (70) and refinement in Refmac (71) in CCP4 suite (69). All the structure pictures and alignments were generated with PyMOL (v.1.8). Buried surface areas were calculated by the European Bioinformatics Institute Proteins, Interfaces, Structures and Assemblies (EBI PISA) server (www.ebi.ac.uk/msd-srv/prot_int/pistart.html).

aNRP2-10 and aNRP2B epitope binning

The Genentech aNRP2B antibody was cloned on the basis of the patent sequence (Seq ID: AHL58550.1) (27). Antibody binding experiments were carried out on an Octet RED96e (ForteBio) biolayer interferometry instrument. Biotinylated NRP2 was captured on streptavidin biosensors and then dipped sequentially into 200 nM of the first antibody followed by 100 nM of the second antibody.

Anti-mNRP2 antibody aNRP2-28

aNRP2-28 is an antibody that recognizes mNRP2, which was used as a mouse-reactive surrogate of aNRP2-10 for in vivo mouse studies. It was generated by immunization of rabbits with recombinant mNRP2 (amino acids 23 to 595) and identified by single-cell enrichment of plasma cells, sequence amplification, recombinant expression, and supernatant screening (ExonBio). The antibody variable domains were sequenced and cloned as a rabbit/mouse chimera into mammalian expression vectors to generate recombinant antibody. The aNRP2-28 antibody was shown to have similar affinities as aNRP2-10 for their respective antigens (mNRP2 for aNRP2-28 and hNRP2 for aNRP2-10). Both bind to the b1 domain of NRP2, and both block VEGF binding to NRP2 with comparable IC₅₀ (0.71 nM for aNRP2-10 and 1.02 nM for aNRP2-28) while not disrupting Sema3F binding (table S3).

Mammosphere assays

The BT549 cell line was provided by D. Kim (UMass Medical School). Cells were cultured in RPMI-1640 supplemented with insulin and 10% FBS. Cells were stained with aNRP2 antibody (aTyr) and anti-mouse fluorescein isothiocyanate (FITC; the Jackson Laboratory). The NRP2^High and NRP2^Low populations were separated using a Sony SH800 sorter. For mammosphere assays, cells were plated in ultralow attachment plates (5000 cells per well of six-well plate; Corning) in Dulbecco’s modified Eagle’s medium/F12 medium supplemented with B27 (Invitrogen), epidermal growth factor (20 ng/ml), and basic fibroblast growth factor (20 ng/ml) as described before (10). The number of mammospheres in each well was counted. For serial passage, mammospheres were collected by centrifugation and dissociated into single cells by incubating at 37°C for 10 min with 0.05% trypsin. Dissociated cells were replated as described above.

To harvest CSCs from human TNBC, we used PDX maintained in immunocompromised NOD scid gamma (NSG) mice (NOD.CgPrkdcscid IL2rgtm1Wjl). Freshly harvested tumor was minced and digested using a Tissue Dissociation kit (Miltenyi Biotec) and a gentle MACS Dissociator. Digested cells were washed and passed through a cell strainer (100 μm). Cells were washed twice and plated in the presence of serum for 2 hours to remove mammary fibroblasts. The suspended cells were stained using anti-human CD44-FITC (clone IM7; BioLegend, #103006) and CD24-APC (clone ML5; BioLegend, #311118) at 1:100 dilution. CSC (CD44^High/CD24^Low) were sorted by fluorescence-activated cell sorting (FACS) and plated for mammosphere assays.

4T1 tumor CSC staining and in vivo limiting dilution assay

For tumor growth, 14 BALB/c female mice (6 weeks of age) were inoculated with 1 × 10⁵ 4T1 cells into the mammary fat pad in 40-μl volume. The mice were obtained from Charles River Laboratories and acclimatized to the facility for at least 1 week before mammary fat pad injection. After injection, the mice were randomly assigned to the control [IgG (25 mg/kg)] or treatment group [aNRP2-28 (25 mg/kg)] with three mice per group. Treatments were intraperitoneally injected. Single-cell suspensions of 4T1-derived tumors were stained using anti-mouse CD44-phycoerythrin (clone IM7; BD Biosciences, #553134) and CD24-FITC (clone M1/69; BioLegend, #101805) at 1:100 dilution for flow cytometry.

For the limiting dilution experiment, 4T1 tumors were harvested and minced using a razor blade and then dissociated into a single-cell suspension using a Tumor Dissociation kit (Miltenyi Biotec). Primary tumor–derived cells were inoculated into the mammary fat pads of 6-week-old BALB/c mice at varying cell densities ranging from 500 to 50,000 cells in 40-μl volume (n = 5 for each group). The mice were euthanized 3 weeks after injection. The frequencyof CSCs was calculated using the ELDA (72) webtool (http://bioinf.wehi.edu.au/software/elda).

In vitro 3D viability assay

The breast cancer cells were obtained from the American Type Culture Collection and the Leibniz Institute DSMZ and maintained at early passages (≤P5) until 70 to 85% confluence. The cells were dissociated by cell dissociation buffer (Thermo Fisher Scientific) or Accutase (Sigma-Aldrich), whichever better maintained the cell viability for each cell type. The dissociated cells were well suspended and filtered through a cell strainer (Sysmex) to obtain single-cell suspensions and seeded at 1 × 10³ to 10 × 10³ cells per well in a 96-well plate based on a preliminary test of optimal seeding density that allows the formation of well-dispersed single colonies for each cell type. Next, a cisplatin monotherapy test was performed to determine the monotherapy inhibition curve on these cells. Cells in medium were mixed with the 3D matrix methylcellulose (at a final concentration of 0.65%; Sigma-Aldrich) and seeded at 70 μl in low attachment round-bottom 96-well plates (Corning), which were prelayered with 20 μl of matrix alone to prevent cells sedimenting to the bottom of the plate. After overnight incubation to stabilize the cells, 10 μl of cisplatin at 10 times the final concentrations were added on top of the solidified cell/matrix for a 1× final concentration in a total of 100 μl of methylcelluose cell medium mixture per well). In the follow-up combination treatment, aNRP2-10 or its isotype control hIgG4 (CrownBio) was incubated with the cell/methylcellulose mixture since seeding. Cisplatin was added the next day, typically at IC₃₀ (high dose) and one-half or one-third IC₃₀ (low dose), on the basis of its monotherapy inhibition curve. Reagents were renewed by spiking-in after 3 days, and cells were treated for another 3 days. Seven days after seeding, CellTiter-Glo 3D (Promega) was added to each well, which produced a fluorescence signal proportional to viable cells in the colonies. A reduction in viability indicates the decrease of colony number, colony size, or both. All treatments were performed in triplicates, and at least two separate experiments for each cell line were performed to obtain consistent results for determining whether a cell was responsive or nonresponsive to aNRP2-10 in mono- and combo therapies.

Patient-derived organoid studies

Tumor tissues from freshly resected biopsies from patients with TNBC were obtained from the UMass Cancer Center Tumor Bank. These tumors were digested using a Tissue Dissociation kit (Miltenyi Biotec) and a gentleMACS Dissociator. The digested tumors were washed three times using PBS, and partially digested tumor pieces were embedded into reduced growth factor basement membrane extract (RGF-BME) (Cultrex). The embedded organoids were cultured in organoid medium as described before (73). The drug-resistant organoids were generated by culturing them in the presence of either 5-FU or cisplatin for 2 weeks. mRNAwas harvested from the sensitive and resistant organoids, and expression of NRP2 was quantitated using qPCR. Primer sequences were obtained from the Harvard Primer Bank. To measure the effect of various treatment on viability of organoids, the organoids were harvested and incubated with aNRP2-10 for 30 min in ultralow attachment plates. For the 3D viability assay, the organoids were embedded in RGF-BME and plated in 24-well plates. The solidified dome containing organoids was fed with various drug combinations. After 96-hour incubation, cell viability was measured using CellTiter-Glo 3D.

For knockdown experiments using VEGF or NRP2 siRNA, the harvested organoids were incubated with the siRNA pool (Santa Cruz Biotechnology) or nontargeting control for 8 hours in ultralow attachment plates. The organoids were further incubated with aNRP2-10 for 30 min before embedding into RGF-BME.

To investigate the effect of FAK inhibition on ZEB1 expression, we incubated the organoids with the FAK inhibitor PF 573228 at 1 μM (Tocris, CAS#869288-64-2) or vehicle for 48 hours. RNA was isolated for ZEB1 gene expression analysis by qPCR.

TNBC mouse models

To establish the MDA-MB-231 tumor xenograft model, 1.25 × 10⁶ cells were implanted into the mammary fat pads of either female NSG mice (the Jackson Laboratory, #005557) or female NOD CRISPR Prkdc Il2r gamma (NCG) mice previously engrafted with huCD34⁺ cells (HuCD34 NCG, Charles River Laboratories). Tumors were monitored biweekly, and mice were randomized into study groups when mean tumor volume reached about 50 mm³. Cisplatin (1 to 5 mg/kg), aNRP2-10 (25 mg/kg), and IgG controls (25 mg/kg) were injected intraperitoneally twice weekly. Cisplatin was given for three cycles with descending doses: 5, 2.5, and 1 mg/kg. Tumors were measured using calipers, and tumor volumes were calculated.

Hs578T cells (1 × 10⁶ per mouse) were implanted into the mammary fat pads of 7-week-old female NSG mice. Tumors were monitored every day, and mice were separated into four groups once tumor volume reached about 100 mm³. The drugs were injected intraperitoneally twice weekly: aNRP2-10 (10 mg/kg), 5-FU (10 mg/kg), or vehicle (saline). Tumor size was measured using calipers, and tumor volume was calculated. All experiments were approved by IACUC.

For evaluation of metastasis, at study, termination lungs were inflated with PBS through the trachea (in situ), excised, and placed in tissue cassettes in 10% neutral buffer formalin for at least 24 hours. After 24 hours, formalin was replaced with 70% ethanol, and tissues were embedded in paraffin within 1 week. Paraffin blocks were then processed, stained with hematoxylin and eosin, and scanned for software-assisted quantification using the HALO image analysis platform. Area quantification module algorithms were adjusted to quantify the area positive for hematoxylin, because cancer cells are more heavily stained. In addition, the HALO classifier module was used to classify different areas within the lungs (normal lung tissue, air space, bronchi, metastasis, and blood cells), which allowed the calculation of percentage of metastasis to lung tissue. The degree of metastasis was either expressed as the number of metastatic nodules per section or percentage of area per section.

TNBC PDXs model

PDX tumors established from patients with TNBC were implanted into the mammary fat pads of NSG mice. Mice were monitored daily, and the first palpable tumor was detected after 6 weeks of implantation. Once the tumors reached an average tumor volume of 20 to 100 mm³, the animals were randomized into four groups: hIgG, saline; aNRP2-10, saline; and hIgG, 5-FU and 5-FU plus aNRP2-10. The first dosing day was reported as day 0. All reagents were administered weekly by intraperitoneal injection. Tumor measurements were taken weekly, and tumor volumes were calculated as presented in the figure.

Gene expression analysis

For gene expression analysis, the organoids were harvested by pooling samples of the same treatment from multiple wells. Total RNA was extracted using the RNeasy kit (QIAGEN). The cDNA was synthesized using the iScript Kit (Bio-Rad). The qPCR experiments were performed with SYBR green qPCR kits (Thermo Fisher Scientific) on the Applied Biosystems ViiA7 real-time PCR system. The microfluidics-based qPCR was performed using the Fluidigm Biomark HD system. The CSC and EMT gene panels are each composed of 84 pathway-focused genes defined in QIAGEN RT2 Profiler PCR Arrays (GeneGlobe IDs: PAHS-176Z and PAHS-090Z). DELTAgene primers targeting these genes and four housekeeping genes (ACTB, GAPDH, HPRT1, and RPLP0) were ordered from Fluidigm. Sample preparation, priming, amplification, and data collection were performed following the manufacturer’s instructions for the 96.96 integrated fluidic circuit delta gene assays on pre-amplified samples. The experiment was performed with technical triplicates. For data analysis, NormFinder (74) was used to identify the most stable genes across samples, and three genes (ACTB, GAPDH, and HPRT1) were lastly used for the normalization of expressions of target genes. For the gene expression study with in vivo samples, tumors were harvested at the end of the study, and total RNA was extracted and transcribed into cDNA followed by qPCR analysis as described above.

Immunoblotting

The NRP2^High population was sorted using FACS from the BT549 cell line. The sorted cells were incubated with the aNRP2-10 antibody for 96 hours. Cells were lysed using radioimmunoprecipitation assay buffer (Sigma-Aldrich), and proteins were separated using 8% reducing SDS–polyacrylamide gel electrophoresis. The proteins were transferred on a nitrocellulose membrane and blocked for 1 hour using 5% nonfat milk prepared in 1× tris buffered saline with 0.1% Tween 20 (TBST) buffer. The immunoblotting was performed using a-ZEB1 antibody (clone D80D3; Cell Signaling Technology, #3396S) at 1:1000 dilution by incubating overnight at 4°C. The membrane was washed three times (10 min per wash) with 1× TBST and incubated with HRP-conjugated secondary antibodies for 1 hour. Tubulin expression was used as a loading control. Imaging was done by chemiluminescence on a Bio-Rad ChemiDoc. Protein bands were quantified using ImageJ.

ZEB1 studies

Plasmid expressing full-length human ZEB1 cloned in lentiviral vector (pLV) was purchased from VectorBuilder. Lentiviral particles expressing ZEB1 were prepared by transfecting this plasmid along with packaging plasmids (psPAX2 and pMD2.G) in HEK293 cells using Lipofectamine 3000 (Thermo Fisher Scientific). Lentiviral particles were collected in organoid base medium. Organoids were extracted from Matrigel with trypsin and washed twice. Organoids were resuspended in 1 ml of base medium containing lentivirus particles along with polybrene (8 μg/ml) in an ultralow attachment plate (Corning) for 6 hours. Organoids were collected by centrifugation and incubated with the aNRP2-10 antibody for 30 min. The organoid 3D viability assay was performed as described above.

Statistical analysis

The statistical analysis for comparing two groups was performed using unpaired t tests (GraphPad Prism 9.0). Means were taken to be significantly different if P < 0.05. In figures, * indicates P < 0.05, ** indicates P < 0.01, *** indicates P < 0.001, **** indicates P < 0.0001, and not significant (ns) indicates P ≥ 0.05 for the indicated pairwise comparison. The tumor volumes in the in vivo studies were analyzed by two-way analysis of variance (ANOVA) followed by Šídák multiple comparison tests using Prism. Error bars in all figures indicate SEM unless stated otherwise.

Data and materials availability:

All data associated with this study are present in the paper or the Supplementary Materials. The structural data are available through the PDB (PDB IDs: 8IVW and 8IVX). The function-blocking NRP2 mAbs will be available to all qualified academic researchers and provided under material transfer agreement for research use only.