Soluble Endoglin Specifically Binds Bone Morphogenetic Proteins 9 and 10 via Its Orphan Domain, Inhibits Blood Vessel Formation, and Suppresses Tumor Growth*

Endoglin (CD105), a transmembrane protein of the transforming growth factor β superfamily, plays a crucial role in angiogenesis. Mutations in endoglin result in the vascular defect known as hereditary hemorrhagic telangiectasia (HHT1). The soluble form of endoglin was suggested to contribute to the pathogenesis of preeclampsia. To obtain further insight into its function, we cloned, expressed, purified, and characterized the extracellular domain (ECD) of mouse and human endoglin fused to an immunoglobulin Fc domain. We found that mouse and human endoglin ECD-Fc bound directly, specifically, and with high affinity to bone morphogenetic proteins 9 and 10 (BMP9 and BMP10) in surface plasmon resonance (Biacore) and cell-based assays. We performed a function mapping analysis of the different domains of endoglin by examining their contributions to the selectivity and biological activity of the protein. The BMP9/BMP10 binding site was localized to the orphan domain of human endoglin composed of the amino acid sequence 26–359. We established that endoglin and type II receptors bind to overlapping sites on BMP9. In the in vivo chick chorioallantoic membrane assay, the mouse and the truncated human endoglin ECD-Fc both significantly reduced VEGF-induced vessel formation. Finally, murine endoglin ECD-Fc acted as an anti-angiogenic factor that decreased blood vessel sprouting in VEGF/FGF-induced angiogenesis in in vivo angioreactors and reduced the tumor burden in the colon-26 mouse tumor model. Together our findings indicate an important role of soluble endoglin ECD in the regulation of angiogenesis and highlight efficacy of endoglin-Fc as a potential anti-angiogenesis therapeutic agent.

cells (18). Co-immunoprecipitation experiments with overexpressed endoglin have shown some binding of radiolabeled activin A, BMP2, and BMP7 to endoglin only in the presence of their respective type I and type II receptors (19), but the role of these complexes is not clear. Some direct binding between the membrane-bound endoglin and BMP9 has also been observed through cross-linking experiments (20). Other studies demonstrated that endoglin specifically binds to type I or II receptors of the TGF␤ superfamily (21) and that this interaction is absolutely required for the association of endoglin with the TGF␤ ligand (19,22,23). Of particular importance are findings suggesting that type I receptors ALK1 and ALK5 interact with endoglin, promoting the TGF␤/ALK1 signaling pathway, but inhibiting the TGF␤/ALK5 pathway. We (24) and others (25) have recently demonstrated that ALK1 binds specifically and with very high affinity to BMP9 and BMP10; however, the potential role of endoglin in these interactions has not been fully addressed.
Here we describe the use of surface plasmon resonance (SPR)-based assay that allowed us to characterize binding of different TGF␤ family ligands to endoglin ECD-Fc (Eng ECD -Fc) chimeras and assess the effect of such complex formation on the interactions with type I and II TGF␤ receptors. We demonstrate for the first time that endoglin-ECD binds directly, specifically, and with high affinity to BMP9 and BMP10. We mapped the binding site of these ligands to the orphan domain of the ECD through the production and evaluation of a series of truncations of the protein and demonstrated that endoglin and type II receptors share the same binding site on BMP9. We suggest that endoglin-ECD acts as a trap for specific ligands, impairing their signaling pathways. We further demonstrated the use of Eng ECD -Fc fusion protein to mediate a significant loss of blood vessel formation in vivo. In the oncology setting we showed that Eng ECD -Fc inhibits colon-26 tumor growth, suggesting its potential use as an anti-angiogenic therapeutic agent.
The filtered conditioned media material containing mEng ECD -mFc 27-581, hEng ECD -hFc 26 -586, or truncated variants of human endoglin was purified by affinity chromatography using mAb Select SuRe protein A chromatography. The protein mEng ECD -mFc 27-581 was subsequently purified by Q-Sepharose column chromatography. hEng ECD -hFc 26 -437 was further purified by flow through a ceramic hydroxyapatite (CHT-II, Bio-Rad) column in 5 mM sodium phosphate, pH 7.0, to remove high and low molecular weight impurities. hEng ECD -hFc 26 -359 was purified by MabSelect SuRe and ceramic hydroxyapatite columns as described above for hEng ECD -hFc 26 -437 followed by a Q-Sepharose chromatography as for mEng ECD -mFc 27-581. The purified proteins were dialyzed into 10 mM Tris, 137 mM NaCl, 2.7 mM KCl, pH 7.14 (modified TBS). The purity of the protein preparations was analyzed by SDS-PAGE under reducing and non-reducing conditions and visualized by SilverSNAP Stain for mass spectrometry (Pierce). Analytical size-exclusion chromatography (SEC) was performed on Agilent 1100 Series HPLC with a Tosoh column TSK gel G3000 SWxl, 7.88 (ID) ϫ 30 cm (length), 5-m particle size. The purity of the protein preparations was greater than 95%. The protein concentration was measured at 280 nm with Nano-Drop 1000 (Thermo) using the extinction coefficient and the molecular weight calculated from the theoretical protein sequence. The N terminus of each construct was verified by N-terminal sequencing using Procise 494 Sequencer (Applied  Biosystems).
Cloning, Expression, and Purification of mEng ECD -mFc 27-581, hEng ECD -hFc 26 -586, and hEng ECD -hFc 26 -359 from CHO Cells-The mEng ECD -mFc 27-581, hEng ECD -hFc 26 -586, and hEng ECD -hFc 26 -359 constructs were subcloned into a UCOE vector (Millipore). 25 g of linearized UCOE construct was transfected into CHO cells using TransIT transfec-tion reagent (Mirus); cells were incubated at 37°C with 5% CO 2 in alpha media (Sigma) supplemented with 10% dialyzed FBS (Invitrogen) and 10 g/ml ADT mixture (Sigma) with one media change after 16 h. On the 2nd day post-transfection, the cells were split 1:6 into selection media containing methotrexate (5 and 10 nM) and grown for 5 days before the media were changed with the appropriate methotrexate concentration. Fourteen days post-transfection, pools were selected by anti-Fc ELISA or Western blotting. Selected pools were then adapted to serum-free conditions in CHO P3 media (Irvine Scientific). The protein was purified from the filtered conditioned media in two steps (affinity and ion-exchange chromatography) as described above for mEng ECD -mFc 27-581 from HEK293 cells.
SPR Assay-Receptor-ligand binding affinities were determined by SPR using Biacore 3000 and T100 instruments (GE Healthcare).
Kinetic Assay in a Capture Format-Goat anti-human Fcspecific IgG or rabbit anti-mouse IgG was immobilized onto research grade CM5 chip using standard amine coupling chemistry following the manufacturer's protocol. hEng ECD -hFc 26 -586 or mEng ECD -mFc 27-581 and truncated variants thereof were captured on the experimental flow cell. Another flow cell was used as a reference (control) to subtract for nonspecific binding, drift, and bulk refractive index. The endoglinantibody complex was stable over the time course of each ligand binding cycle. A concentration series of BMP9 and BMP10 (0.00977-1.25 nM, with 2-fold serial dilutions) was injected over experimental and control flow cells at a flow rate of 70 l/min at 20°C and 37°C. Duplicate injections of each ligand concentration were performed, with buffer blanks injected periodically for double referencing. The antibody surface was regenerated between binding cycles by injection of 3 M MgCl 2 (for Biacore antibody) or 10 mM glycine, pH 1.7 (for The Jackson ImmunoResearch Laboratories or Thermo Fisher Scientific antibodies). Buffer containing 0.01 M HEPES, 0.5 M NaCl, 3 mM EDTA, 0.005% v/v Surfactant P20, pH 7.4, supplemented with 0.5 mg/ml BSA was used as the running buffer. All sensorgrams were processed by double referencing (subtraction of the responses from the reference surface and from an average of blank buffer injections). To obtain kinetic rate constants, the corrected data were fit to a 1:1 interaction model that includes a term for mass transport using BIAevaluation software (GE Healthcare). The equilibrium dissociation constant K D was determined by the ratio of binding rate constants k d /k a .
Reporter Gene Assay in A204 Cells-A reporter gene assay in A204 cells was used to evaluate the effects of hEng ECD -hFc and mEng ECD -mFc fusion proteins on signaling by BMP9 and BMP10. This assay is based on human rhabdomyosarcoma cell line A204 transfected with a pGL3 BRE-luciferase reporter plasmid (25,26) as well as a Renilla reporter plasmid (pRLCMV-luciferase) to control for transfection efficiency. The BRE motifs are present in BMP-responsive genes (Id1 promoter), so this vector is of general use for factors signaling through Smad1 and/or Smad5. In the absence of hEng ECD -hFc and mEng ECD -mFc fusion proteins, BMP9 and BMP10 stimulate signaling in A204 cells in a dose-dependent manner.
A204 cells (ATCC HTB-82) were distributed in 48-well plates at 10 5 cells per well. The following day a solution containing 12 g of pGL3 BRE-luciferase, 0.1 g of pRLCMV-luciferase, 30 l of FuGENE 6 (Roche Diagnostics), and 970 l of Opti-MEM (Invitrogen) was preincubated for 30 min at room temperature before adding to 24 ml of assay buffer (McCoy's medium supplemented with 0.1% BSA). This mixture was applied to the plated cells (500 l/well) for incubation overnight at 37°C. On the third day, medium was removed and replaced with test substances (250 l/well) diluted in assay buffer, which included BMP9 and BMP10 at a final concentration of 5 ng/ml. After an overnight incubation at 37°C, the cells were rinsed and lysed with passive lysis buffer (Promega) and frozen at Ϫ70°C. Before the assay, the plates were warmed to room temperature with gentle shaking. Cell lysates were transferred in duplicate to a chemiluminescence plate (96-well) and analyzed in a luminometer with reagents from a Dual-Luciferase Reporter Assay system (Promega) to determine normalized luciferase activity.
Directed in Vivo Angiogenesis Assay-Directed in vivo angiogenesis assay (Trevigen, Gaithersburg, MD) was used as a quantitative method for assaying angiogenesis. Implant-grade silicone cylinders were filled with 20 l of basement membrane extract premixed with or without a combination of vascular endothelial growth factor (VEGF, 600 ng) and fibroblast growth factor (FGF-2, 1.8 g). The silicone angioreactors were implanted subcutaneously in the dorsal flank of athymic nude mice, 2 angioreactors on each side of the animal. Mice were dosed subcutaneously daily with either vehicle (modified TBS) or mEng ECD -mFc 27-581 (10 mg/kg). After 11 days, animals received an intravenous injection of FITC-dextran (20 mg/kg) and were euthanized 20 min later by CO 2 inhalation. Angioreactors were immediately excised from the animal, photographed, and solubilized. The amount of FITC-dextran in the angioreactor was quantified at 520-nm emission and 485-nm excitation using a fluorescence plate reader (Infinite m200, Tecan, Mannedorf, Switzerland). The intensity of the signal is proportional to the amount of blood vessels in each angioreactor.
Chick Chorioallantoic Membrane (CAM) Assay-Nine-dayold fertilized chick embryos were maintained in a 48-place tabletop egg incubator at 37°C and specific humidity (60%). Prominent blood vessels were visualized through the eggshell with the aid of an egg lamp. The area of the outer eggshell, where the prominent blood vessels are located, was swabbed with 70% alcohol, a small hole was made, and the air sack was displaced causing a "blister" to form between the shell membrane and the CAM; finally, a small window was cut through the eggshell with a hobby grinding wheel (Dremel Emerson Electric Co., Racine, WI). Small filter disks infused with the angiogenic substance VEGF (50 ng) either in the presence or absence of mEng ECD -hFc 27-581 (3 doses of 14 g) in Biacore running buffer, pH 7.4, was then placed at the opening. For hEng ECD -hFc 26 -359, 3 doses of 20 g were used in the assay. Each group (8 eggs) was treated daily for 3 days, and CAM results were analyzed on the fourth day. Quantification of the number of resulting vessels per disk post-treatment was made visually with the help of an egg lamp. Statistical analysis of the results was performed with Excel and SigmaStat software.
Colon-26 Cancer Model-Five-to six-week-old BALB/c mice were obtained from Harlan Laboratories. They were supplied autoclaved food and water ad libitum. Colon 26 (ATCC# CRL-CL25) cells were grown in T-175 flasks at 37°C ϩ 5% CO 2 in RPMI 1640 medium (ATCC) ϩ 10% FBS (Invitrogen). Cells were trypsinized with TrypLE Select (Invitrogen) and counted using a CEDEX cell counter. Cells were washed once with PBS and resuspended in sterile water at a concentration of 10 ϫ 10 6 cells/ml. 100 l of the cell suspension was injected subcutaneously in the dorsal right flank. After 3 days, subcutaneous tumors became palpable, and animals were subcutaneously administered daily either mEng ECD -mFc 27-581 (10 mg/kg) or modified TBS. Tumor volumes were measured manually using digital calipers (VWR International), and tumor volume was estimated using the formula (0.5 ϫ length ϫ width 2 ). At the conclusion of the study, animals were euthanized by CO 2 inhalation, and tumors were excised and frozen in liquid nitrogencooled isopentane.

Cloning, Expression, and Purification of Mouse and Human
Endoglin ECD-The full-length ECDs of the mouse and human endoglin were cloned and fused to the Fc portion from an IgG of corresponding species producing mEng ECD -mFc 27-581 and hEng ECD -hFc 26 -586 constructs (Fig. 1, A and D). Expression of each construct was achieved by transient transfection of HEK293 cells or stably transfected CHO cells. Proteins were purified from conditioned media by column chromatography as described under "Experimental Procedures." A silver-stained SDS-PAGE gel revealed protein bands of ϳ110 kDa under reducing and ϳ220 kDa under non-reducing conditions (Fig. 1, B and E). Analytical SEC showed that both mEng ECD -mFc 27-581 and hEng ECD -hFc 26 -586 were ϳ95% pure with the absence of aggregates (Fig. 1, C and F) irrespective of the cell expression system used.
Endoglin ECD Does Not Bind to Either Type I or Type II Receptors in Vitro-Previous studies performed on membrane-bound receptors have shown that endoglin can directly associate with type II receptor TGF␤RII (19) as well as type I receptors ALK5 and ALK1 (21,27) without TGF␤ family ligands present. It has been suggested that these interactions occur through the ECDs of these receptors (28). To understand what role endoglin ECD plays in modulating interactions of TGF␤ family ligands with their receptors, we tested binding of hEng ECD 26 -586 to the ECD of 5 human type II and 7 type I TGF␤ family receptors fused to hFc using SPR technology (Biacore). Anti-hFc IgG was immobilized on a CM5 chip followed by capture of receptor-hFc fusion proteins. The assay design allows the receptor molecule to adopt homogeneous orientation at the surface of the chip. The construct hEng ECD 26 -586 was then injected over the captured receptors. We did not observe any binding of endoglin ECD to captured receptors (supplemental Table I). Similarly, mEng ECD -mFc 27-581 captured on an anti-mFc IgG chip failed to show any binding to the ECD of mouse type II and type I receptors (data not shown). Thus, we conclude that the ECD form of endoglin is not able to associate directly with the ECD of TGF␤ family receptors in vitro. Screening for Endoglin Binding Ligands-Because previous studies suggested that endoglin might bind to TGF␤ and several other ligands directly, thus affecting a variety of different signaling pathways (18,20,23,25), we used SPR to screen multiple TGF␤ family ligands for binding to human and mouse endoglin ECDs. In-house-generated human endoglin (hEng ECD -hFc 26 -586) and the commercially available mouse endoglin (mEng ECD -hFc 27-581) were captured on an anti-hFc IgG CM5 chip. Subsequently, different TGF␤ superfamily ligands (listed under "Experimental Procedures") were injected over captured endoglin molecules at concentrations well above the physiological values usually found in serum (29,30). Of all the ligands tested, in the absence of other receptors, only BMP9 and BMP10 showed any significant binding to both endoglin constructs, as defined by a K D value lower than 10 Ϫ6 M (supplemental Table II). A similar experiment was performed with in-house-generated mouse endoglin (mEng ECD -mFc 27-581), which also showed binding only to BMP9 and BMP10 (data not shown). To confirm these findings and ensure that the Fc portion of the fusion protein did not interfere with ligand binding, we performed additional ligand screening with hEng ECD 26 -586 that lacks any Fc. hEng ECD 26 -586 was immobilized directly onto a CM5 chip and screened with the same ligands. Again, we observed only binding of BMP9 and BMP10 but not other TGF␤ ligands (data not shown), suggesting that endoglin ECD is capable of binding to only two TGF␤ family ligands: BMP9 and BMP10.
We further confirmed these results by preincubating the mEng ECD -mFc 27-581 construct with either BMP9, BMP10, activin A, or TGF␤1 and analyzing the potential complexes on a native PAGE gel (supplemental Fig. S1). Although we have not seen any complex formation between endoglin and TGF␤1 (supplemental Fig. S1, lane 5) or activin A (supplemental Fig.  S1, lane 4), we observed decreased mobility as evidenced by a visible upward shift of the endoglin band in the presence of BMP9 and BMP10 (supplemental Fig. S1, lanes 2 and 3), which indicates formation of strong complexes between endoglin ECD and these ligands. We wanted to confirm the lack of interaction between TGF␤1 and endoglin ECD in a cell-based assay format. We performed a reporter gene assay with TGF␤1 (mixed or not with mEng ECD -mFc 27-581) as a ligand in HepG2 cells transfected with the CAGA 12 -luciferase plasmid (TGF␤1/Smad 2/3-responsive). The mEng ECD -mFc 27-581 did not inhibit the TGF␤ signaling, confirming our in vitro results (data not shown).
Several previous reports suggested that endoglin binding to TGF␤1 requires the ligand to be complexed with type II receptor TGF␤RII (23). We attempted to mimic this system in an SPR experiment by injecting TGF␤1 over captured hTGF␤RII-hFc (containing ECD of human TGF␤RII fused to hFc) followed by injection of hEng ECD 26 -586 (supplemental Fig. S2). Although TGF␤1 formed a stable high affinity complex with TGF␤RII, no binding of hEng ECD 26 -586 to the TGF␤1-TGF␤RII receptor-ligand complex was observed (supplemental Fig. S2A). Additionally, the assay was reversed by capturing mEng ECD -hFc 27-581 on the Biacore chip, then injecting preformed TGF␤1-hTGF␤RII complex over captured endoglin. No binding of the preformed TGF␤1-hTGF␤RII complex was observed by SPR in this configuration either (data not shown). To test if endoglin might alter TGF␤ affinity for its receptor, we premixed TGF␤1 with hEng ECD 26 -586 or buffer and injected both mixtures over captured TGF␤RII-hFc receptor (supplemental Fig. S2B). No difference between the two injections was observed, suggesting that endoglin ECD does not affect TGF␤ binding to the receptor. Thus, we conclude that in vitro mouse and human endoglin ECDs bound to BMP9 and BMP10 but did not associate with TGF␤1 either free or in complex with the type II receptor.
High Resolution Kinetic Analysis of mEng ECD -mFc 27-581 and hEng ECD -hFc 26 -586 Binding to BMP9 and BMP10 and Potency Testing in the A204 Cell-based Assay-To characterize more fully BMP9 and BMP10 binding to endoglin, we performed a high resolution kinetic study using captured mEng ECD -mFc 27-581 or hEng ECD -hFc 26 -586 proteins expressed either in HEK293 or CHO cells. A representative kinetic analysis is shown in Fig. 2 (CHO) and supplemental Table SIII (HEK293). The interaction between ligands and endoglin-Fc is partially mass transport-limited. To minimize the mass transport, we used a high flow rate (70 l/min) and a low level of the captured endoglin; however, we were not able to remove the limitations completely. Nevertheless, by including a mass transport term into 1:1 binding model, we were able to achieve an excellent fit of the experimental data as shown by overlay of the simulated binding responses (Fig. 2, A-D, red  lines). BMP9 showed high affinity binding to mouse and human endoglin constructs with the equilibrium binding constants K D of 39 pM (HEK293-derived material) and 22 pM (CHO-derived material) for mEng ECD -mFc 27-581 and 26 pM (HEK293) and 33 pM (CHO) for hEng ECD -hFc 26 -586 ( Fig. 2E and supplemental Table III). The kinetic rate constants for BMP9 binding to mouse and human endoglin constructs were very similar. Interestingly, BMP10 seems to have different modes of binding to human and murine endoglin, which can be deduced from markedly different K D values obtained for mEng ECD -mFc 27-581 (76 pM (HEK293) and 71 pM (CHO)) and hEng ECD -hFc 26 -586 (241 pM (HEK293)) and 491 pM (CHO)) proteins, the difference being due mainly to a faster dissociation of the BMP10-hEng ECD -hFc complex. To understand how endoglin interaction with ligand changes at physiological temperature, we further performed SPR analysis at 37°C. We found that hEng ECD -hFc 26 -586 (CHO) maintained its affinity for BMP9 (27.8 pM) and BMP10 (408 pM) (data not shown). Thus, the ECD of human endoglin binds BMP9 and BMP10 specifically and with high affinity at physiological temperature.
To measure the ability of mEng ECD -mFc 27-581 and hEng ECD -hFc 26 -586 to antagonize Smad signaling induced by BMP9 and/or BMP10, we performed a reporter gene assay in the human rhabdomyosarcoma cell line A204 transfected with a pGL3BRE-luciferase reporter plasmid (a Smad1/5-responsive construct (25,26) and plasmid pRLCMV-luciferase (internal control). The cells were treated with the corresponding ligands in the presence or absence of mouse and human endoglin ECDs and assayed for luciferase activity. As predicted by the binding affinities determined by SPR, mouse and human endoglin inhibit BMP9 and BMP10 signaling in a dose-dependent manner with inhibitory concentrations (IC 50 ) in the sub-nanomo-lar ϳ low nanomolar range (Fig. 2E). These data demonstrate that endoglin-Fc fusion protein inhibited signaling via the Smad pathway by BMP9 and BMP10. Taken together the results from two distinct in vitro systems suggest that the extracellular domain of mouse and human endoglin bind to and efficiently inhibit BMP9-and BMP10-induced signaling.
Truncations of hEng ECD -hFc Mapped the Ligand Binding Site to the Orphan Domain 26 -359-Although no crystal structure for endoglin is available, electron microscopy modeling and sequence alignment have revealed that the endoglin ECD is composed of three domains; the N-terminal domain does not show any similarity to known proteins and was thus designated "orphan domain," whereas the subsequent two domains show similarity to a zona pellucida (ZP) domain (31). To map the BMP9/BMP10 binding site on the human endoglin protein, we performed truncations of the ECD based on the electron microscopy model and the location of the protease cleavage site proposed for the generation of the soluble endoglin form found in preeclampsia (13,31). As shown in Fig. 3, we constructed hEng ECD -hFc proteins containing either the orphan domain (26 -359), the ZP (360 -586) domain, the soluble, proteasecleaved ECD suggested for the protein found in preeclampsia (26 -437) (13,31), or the remainder of the protease-cleaved endoglin ECD (438 -586). Each construct was expressed by transient transfection in HEK293 cells, purified by affinity chromatography, and tested by SPR for their ability to bind BMP9 and BMP10. As shown in Fig. 3B, constructs possessing only the ZP domain lost the ability to bind BMP9 and BMP10, whereas hEng ECD -hFc 26 -586 and the constructs containing the entire orphan domain were still able to bind to both BMP9 and BMP10.
To further characterize the BMP9/BMP10 binding site, we prepared human endoglin constructs with further truncations at the C terminus of the orphan domain (Fig. 3). The secondary structure prediction for the endoglin ECD, generated by using the Psipred and Jalview software (32,33), showed that the Ile 359 residue at the terminus of the orphan domain is right at the end of an ␣-helix (amino acids 352-358), which is located in a long disordered segment of amino acids (amino acids 328 -367) that could impair expression and purity (Fig. 3A). Hence, we decided to prepare the constructs hEng ECD -hFc 26 -378 (inclusion of an extra ␤-strand and ␣-helix segment compared with hEng ECD -hFc 26 -359), 26 -332 (removal of the ␣-helix at amino acids 352-358), and 26 -329 (terminating the orphan domain at the end of a ␤-strand at amino acids 322-327 and removal of an extra Cys (amino acid 330) potentially important for correct protein folding). The constructs 26 -329 and 26 -332 showed no binding activity to either ligand by SPR, whereas the construct 26 -378 was still able to bind to both BMP9 and BMP10 (Fig. 3B). These results highlight the potential importance of the ␣-helix at amino acids 352-358 and/or the need for this potentially disordered segment of amino acids for the flexibility of the orphan domain in relation to the Fc portion of the construct to actively bind BMP9 and BMP10.
Having established that the minimal construct capable of binding to ligands was 26 -359, we further tested binding specificity and potency of hEng ECD -hFc 26 -359 as well as other truncation constructs 26 -378 and 26 -437 in comparison with the construct containing full-length ECD by Biacore and A204 cell-based assays as shown in Table 1. The equilibrium dissociation constant (K D ) obtained in these assays was very similar to the value obtained for the full-length endoglin ECD hEng ECD -hFc 26 -586 with both BMP9 and BMP10. However, although BMP9 dissociates quite rapidly from the full-length hEng ECD -hFc 26 -586 (k d of 1.98 ϫ 10 Ϫ3 s Ϫ1 ), its dissociation is ϳ6-fold slower for the truncated construct hEng ECD -hFc 26 -359 (k d of 2.99 ϫ 10 Ϫ4 s Ϫ1 ), which represents a marked improvement in the stability of the ligand-endoglin complex. The improved binding of BMP9 was somewhat reflected by an IC 50 value of hEng ECD -hFc 26 -359 (0.16 nM) compared with hEng ECD -hFc 26 -586 (0.26 nM) and was not due to the cell type used to express the material. Based on the limitations of the assay, this change in IC 50 values suggests a trend instead of a significant difference. The truncation of the C-terminal portion of the endoglin ECD appears to allow for a more specific construct FIGURE 3. Effect of truncations in the ECD on hEng ECD -hFc binding to BMP9 and BMP10 determined by Biacore. A, hEng ECD -hFc 26 -359 secondary structure prediction, starting at amino acid 240, generated by Psipred and Jalview software (32,33), and location of truncations performed at the C terminus of the orphan domain are shown. The ␣-helix is represented by a cylinder, and the ␤-sheet is represented by an arrow. B, shown is a schematic representation of hEng ECD -hFc truncations with their constituent domains. Each fusion protein was purified from HEK293-transfected conditioned media by affinity chromatography and evaluated for binding to BMP9 and BMP10 by SPR. The endoglin fusion proteins were captured at high density (ϳ500 -600 resonance units) on an anti-hFc IgG chip. BMP9 and BMP10 were injected at a concentration of 10 nM over each construct for detection of any significant binding. For those constructs showing binding activity with BMP9 and BMP10, a high resolution kinetic analysis was performed (Table 1) with tighter binding to BMP9 compared with the full-length ECD in hEng ECD -hFc 26 -586. To further assess the specificity of hEng ECD -hFc 26 -359, we tested binding of this construct to other members of the TGF␤ family including TGF␤1, -2, and -3, activin A, BMP2, and BMP7; however, no direct binding was observed (data not shown).
hEng ECD -hFc 26 -359, cloned into a UCOE vector and stably expressed in CHO cells (as described under "Experimental Procedures"), was used for subsequent studies. Fig. 4 shows that the hEng ECD -hFc 26 -359 expressed in CHO cells and purified with protein A and a Q-Sepharose purification steps yielded a protein of 91% purity as determined by SDS-PAGE gel and analytical SEC (Fig. 4, A and B). The specificity of this construct was the same as the full-length endoglin chimera, as shown by Biacore and cell-based assay data (Fig. 4, C-E) and ligand screen (supplemental Table II). This protein was used for further antiangiogenesis studies.

Endoglin-Fc Acts by Blocking BMP9 Binding to Type II
Receptors-Our previous findings (24) and experiments presented here demonstrated that both endoglin ECD and type I receptor ALK1 bind to BMP9 with high affinity and specificity; thus, they could potentially either act synergistically or compete for binding to these ligands. To understand if endoglin ECD alters BMP9 binding to ALK1, we performed a solution inhibition experiment using SPR. BMP9 and mALK1-mFc were premixed at different ratios, and the mixtures were injected over captured hEng ECD -hFc 26 -359 (Fig. 5A). We observed that the complex bound on the chip was bigger than BMP9 alone, consistent with the binding of the BMP9-mALK1-mFc complex to the endoglin-Fc. In a binding exclusion experiment, BMP9 was first injected over captured hEng ECD -hFc 26 -359 at saturating concentrations followed by injection of soluble mALK1-mFc or hALK1-His (Fig. 5B). We observed a ternary complex between endoglin, BMP9, and ALK1 (Fig. 5B, insert).

TABLE 1 Effect of truncations in the ECD on endoglin binding to BMP9 and BMP10 determined by Biacore and in A204 cell-based assays
Kinetic analysis of BMP9 and BMP10 binding to truncated endoglin variants (produced in HEK293 cells except where indicated) was performed in duplicate on Biacore 3000 at 20°C as described under "Experimental Procedures." Data were globally fit to a 1:1 binding model with the mass transfer term using BIAevaluation software. The ability of truncated endoglin variants to antagonize Smad signaling induced by BMP9 and BMP10 was tested in a reporter gene assay using A204 cells transfected with Smad-responsive construct pGL3(BRE)-luc. a CBD, could not be determined. 0% inhibition was observed at the maximum concentration tested (70 nM). b CHO-expressed material, and not HEK293-produced material, was used in the cell-based assays.

Endoglin
The same binding exclusion experiment was run in reversed format by injecting BMP9 over hALK1-hFc followed by the injection of mEng ECD -mFc 27-581 with similar results (supplemental Fig. S3). Thus, we conclude that endoglin and ALK1 bind to different sites on BMP9. BMP9 can bind to three different type II receptors, ActRIIA, ActRIIB, and BMPRII (20,34,35). To understand if endoglin ECD alters BMP9 binding to type II receptor, we performed a similar series of SPR experiments. By altering the ActRIIB: BMP9 ratio we were able to show inhibition of BMP9 binding to captured hEng ECD -hFc 26 -359 (Fig. 5C), which suggests that mActRIIB-mFc competes with hEng ECD -hFc 26 -359 for BMP9 binding. Moreover, injection of BMP9 followed by mActRIIB-mFc did not produce additional binding of the receptor to the FIGURE 5. hEng ECD -hFc 26 -359 and type II receptors bind to overlapping sites on BMP9. The hEng ECD -hFc 26 -359 (CHO-produced) was captured onto an anti-hFc IgG chip, then a mixture of BMP9 with buffer, mAlk1-mFc (2.5 or 50 nM) (A) or mActRIIB-mFc (1, 2.5 nM) (C) was injected in a solution inhibition assay. In a Biacore epitope exclusion assay, the hEng ECD -hFc 26 -359 was captured, then subsequent injections of BMP9 and buffer, mAlk1-mFc, or hAlk1-His (100 nM) (B) or mActRIIB-mFc (100 nM) (D) were performed. The inset is a magnification of the type I and II receptor injections. RU, resonance units. E, shown is a proposed model for the mode of action of endoglin ECD. The hEng ECD -hFc 26 -359 binds to the type II receptor binding site on BMP9, blocking access of a type II receptor to the ligand and impairing the formation of the ternary complex necessary for signaling. captured hEng ECD -hFc 26 -359 (Fig. 5D), suggesting the binding of BMP9 to endoglin and ActRIIB is mutually exclusive. The reversed assay with mEng ECD -mFc 27-581 with different type II receptors gave similar results (supplemental Fig. S3). Based on these observations we suggest that endoglin and type II receptor might bind to overlapping sites on BMP9. These experiments could explain the signaling inhibition observed previously in A204 cell-based assays by endoglin-Fc. Endoglin can form a complex with BMP9 by binding to its type II binding epitope, thereby blocking the type II receptor from interaction with the ligand. No ternary complex type I-ligand-type II can be formed to trigger the signaling pathway, and thus inhibition was observed in cell-based assays (Fig. 5E).
Endoglin Prevents Formation of Blood Vessels in the CAM Assay-Using SPR and the A204 cell-based assay, we have demonstrated that mEng ECD -mFc 27-581 and hEng ECD -hFc 26 -359 bind to BMP9 and BMP10 with high affinity and can block their intracellular signaling. There are several publications suggesting that endoglin plays a role in angiogenesis (2); therefore, we further explored the ability of these two endoglin constructs to affect vasculature in the in vivo chick CAM assay. Individual filter discs were treated with VEGF in the presence or absence of endoglin-Fc and inserted between the shell membrane and the CAM of 9-week-old fertilized eggs. Quantification of the number of the resulting vessels per disk post-treatment showed that mouse endoglin acted as an anti-angiogenic factor, which decreased the amount of blood vessels by 65% compared with the VEGF-treated group (Fig. 6A). The same experiments were repeated with the human truncated endoglin hEng ECD -hFc 26 -359 with the same results (Fig. 6B). To assess any direct binding between VEGF and mouse or human endoglin ECD constructs, a Biacore experiment was performed. No binding of VEGF to endoglin was observed (data not shown). Thus, we conclude that both mouse endoglin-Fc and the truncated human version containing the orphan domain are able to inhibit angiogenesis.
Endoglin-Fc Prevents Formation of Blood Vessels in the Directed in Vivo Angiogenesis Assay-We investigated the antiangiogenic effect of mEng ECD -mFc 27-581 in an angiogenesis assay containing VEGF and FGF. Silicone angioreactors were filled with basement membrane extract premixed with or without a combination of VEGF and FGF. Mouse endoglin acted as an anti-angiogenic factor by reducing the number of sprouting blood vessels compared with the control groups with and without growth factors (Fig. 6C). Quantification of the blood vessel sprouts using FITC-dextran confirmed these results, showing a statistically significant decrease in the relative fluorescence of the mEng ECD -mFc 27-581 treated angioreactors compared with the vehicle controls (Fig. 6D).
Endoglin-Fc Reduces Tumor Burden in the Colon-26 Cancer Model-Anti-endoglin antibodies have shown efficacy in decreasing the tumor burden in a colon-26 adenocarcinoma cell cancer model (36). We investigated whether our endoglin-Fc construct would behave similarly in that cancer model by treating mice with mEng ECD -mFc 27-581 or modified TBS (vehicle) following the subcutaneous implantation of colon-26 adenocarcinoma cells. Tumor volume, measured by digital calipers, was found to be significantly delayed in the mEng ECD -mFc-treated animals as early as day 12 after cell implantation, as shown by the p value lower than 0.01 (Fig. 6E).

DISCUSSION
Multiple lines of evidence suggested that endoglin plays a crucial role in angiogenesis (1, 2), cancer (12,17), and preeclampsia (13). In this paper we have systematically studied the structurefunction relationships and interaction of the endoglin ECD with its ligands and receptors. Using endoglin ECD (hEng ECD 26 -586) as well as proteins containing the ECD fused to IgG Fc, we found that both forms bind specifically and with high affinity to only BMP9 and BMP10 ligands (Fig. 2). We were surprised that endoglin ECD failed to bind TGF␤1 or TGF␤3 in HepG2 cell-based assays and in SPR experiments even when complexed with TGF␤RII, particularly as several publications had suggested the importance of direct interactions between endoglin and these ligands in the TGF␤ signaling pathway in vivo.
Several published investigations have reported that type I receptors (ALK1 and ALK5) and type II receptor TGF␤RII are needed for endoglin interactions with TGF␤ ligands (19,(21)(22)(23). Thus, we initially investigated whether the receptors are able to bind directly to endoglin ECD and, secondly, whether TGF␤1 or TGF␤3 ligands bound to the receptor TGF␤RII facilitate the interaction with endoglin. We did not observe any interactions between the ECD of type I or type II receptors and endoglin, in contrast to reported experiments with overexpression of these molecules on the cell surface. We cannot rule out the possibility of intracellular domains promoting the interaction (37), but our data strongly suggest that the binding cannot be attributed to interactions between the extracellular domains. Furthermore, by varying the sequence of assembly of the TGF␤ complex on the surface of the Biacore chip, we explored multiple potential ways for endoglin ECD to be involved in the TGF␤ ligand-receptor complex formation. However, we did not detect any effect of endoglin ECD in this system (supplemental Fig. S2). Our observation that endoglin ECD is unable to bind TGF␤1 complexed with type II receptor TGF␤RII indicates that ligand-receptor association does not alter the specificity of endoglin ECD and that the balance between soluble versus membrane-bound endoglin, which could possess different specificities, might add an additional level of regulation to the system.
To further address the discrepancy between in vitro SPR data for endoglin ECD and previously published in vivo literature data obtained for membrane-bound endoglin, we performed cell-based assays (Fig. 2) that confirmed the specificity of endoglin ECD observed in the SPR experiments. Thus, we conclude that the effect of endoglin ECD is attributed to its inhibition of signaling triggered by BMP9 and BMP10 and not TGF␤1 or TGF␤3. Endoglin previously has been shown to bind BMP9 by means of a cross-linking assay with a radiolabeled ligand (20). Here we demonstrate for the first time a detailed kinetic characterization of BMP9 and BMP10 binding to endoglin ECD using SPR technology. Recent studies have established BMP9 and BMP10 to be major ligands in the angiogenic pathway of ALK1 (24,25). Mutations in both endoglin and ALK1 result in similar disease symptoms in hereditary hemorrhagic telangiectasia HHT1 and HHT2, respectively, which supports our find-ing that both ALK1 and endoglin act through the same ligands and are implicated in the same signaling pathway. Also, several mutations found in patients with HHT1 and HHT2 prevented endoglin and ALK1 binding to BMP9 or BMP10 ligands. 4 Given the role of endoglin in vascular function and disease, the identification of ligands for soluble endoglin is of utmost importance. In fact, soluble endoglin from the placenta has been shown to be present at much lower levels in healthy pregnant women than in preeclamptic women and barely present in non-pregnant women and thus is implicated in severe vascular problems (13). Our identification of the binding site for the ligands BMP9 and BMP10 in the orphan domain of human endoglin, comprised of the amino acid sequence 26 to 359 (Figs. 3 and 4), is consistent with reports of physiological and ligand binding activity for the soluble form of endoglin in preeclamp-FIGURE 6. Endoglin ECD is able to inhibit angiogenesis in vivo in the CAM assay and in angioreactors as well as decrease tumor burden in a colon-26 cancer model. A, shown is blood vessel count with VEGF (50 ng) as the angiogenic factor and 3 doses of 14 g of mEng ECD -hFc 27-581 (NS0-produced; R&D Systems). B, shown is blood vessel count with VEGF (50 ng) as the angiogenic factor and 3 doses (20 g each) of CHO-produced hEng ECD -hFc 26 -359. C, silicone angioreactors were removed from the mice after 11 days of treatment with mEng ECD -mFc 27-581 (from CHO cells). Blood vessel sprouting into the angioreactors appears dark red. D, quantification of vessel formation in the angioreactors was done by injecting FITC-dextran into the animals before euthanasia. The amount of fluorescence in the angioreactor correlates to the amount of blood vessels present. In the presence of angiogenic growth factors, angioreactors from animals treated with mEng ECD -mFc 27-581 had less relative fluorescence than the vehicle-treated controls. E, tumor volume, measured by digital calipers, in a mouse model of colon-26 cancer treated with modified TBS or mEng ECD -mFc 27-581 is shown. *, p Ͻ 0.05; **, p Ͻ 0.01; ***, p Ͻ 0.001. sia containing residues 26 -437. Our observation that small portions of the orphan domain of endoglin (e.g. residues 26 -378 and 26 -359) have higher affinities for BMP9 and BMP10 than the full-length ECD (residues 26 -586) further restricts the ligand binding site and indicates an opportunity for the design of anti-angiogenic, endoglin-based therapy with improved ligand binding parameters (Table 1). The two domains contained in the remainder of the extracellular domain not required for ligand binding are members of a large class of structures, the ZP domains, which perform highly diverse functions in other proteins, most notably filament formation (38). In endoglin these domains may act as spacers between the ligand binding site of endoglin and the cell surface and are suggested, based on published studies, to be involved in endoglin oligomerization and in ligand-independent heteromeric interactions with receptors and activin receptor-like kinase (21).
Although the exact mechanism is a subject of ongoing investigation and beyond the scope of this paper, we hypothesize that the main role of soluble endoglin may be scavenging BMP9 or BMP10 ligands, which could affect the delicate balance required for angiogenesis in normal versus tumor blood vessels (Fig. 5). Another interesting observation, that endoglin and type II receptors share the same interface on BMP9, suggests an additional level of signaling control in the system and will require additional investigation under physiological conditions.
The efficacy of Eng ECD -Fc to block angiogenesis was demonstrated in the CAM and angioreactor assays in which endoglin dramatically inhibited vessel formation induced by VEGF (Fig.  6). These data agree with previously published observations that Eng-Fc reduced spontaneous and VEGF-induced endothelial sprouting in HUVEC cells (15). Anti-angiogenic activity of endoglin ECD was further manifested as anti-tumor activity in the colon-26 model, where the mEng ECD -mFc construct was effective in reducing the tumor burden. These findings can help explain the disruption of normal vasculature development in instances when soluble endoglin is present, for example in preeclampsia (13), and indicate potential roles for soluble forms of endoglin as anti-angiogenic factors.
In summary, the orphan domain of endoglin binds BMP9 and BMP10 with a very high affinity and specificity. It inhibits the interaction of BMP9 with type II receptors, thereby blocking multiple signaling pathways at once, which could be advantageous over other therapeutic agents targeting a single component of the TGF␤ pathway. More experiments in animal models of cancer and angiogenesis are needed to establish the use of endoglin ECD as a modulator of BMP9/BMP10 activity.