Specificity and Structure of a High Affinity Activin Receptor-like Kinase 1 (ALK1) Signaling Complex

Background: Activin receptor-like kinase 1 (ALK1) is an important regulator of normal blood vessel formation and pathological tumor angiogenesis. Results: Crystal structure of ALK1-BMP9-ACTRIIB signaling complex and kinetic and thermodynamic properties of receptor-ligand interactions are described. Conclusions: ALK1 achieves high specificity for BMP9/10 through unusual receptor positioning and unique receptor-ligand interface. Significance: Structural and mechanistic insights into ALK1 signaling provide a framework for novel anti-angiogenic therapies. Activin receptor-like kinase 1 (ALK1), an endothelial cell-specific type I receptor of the TGF-β superfamily, is an important regulator of normal blood vessel development as well as pathological tumor angiogenesis. As such, ALK1 is an important therapeutic target. Thus, several ALK1-directed agents are currently in clinical trials as anti-angiogenic cancer therapeutics. Given the biological and clinical importance of the ALK1 signaling pathway, we sought to elucidate the biophysical and structural basis underlying ALK1 signaling. The TGF-β family ligands BMP9 and BMP10 as well as the three type II TGF-β family receptors ActRIIA, ActRIIB, and BMPRII have been implicated in ALK1 signaling. Here, we provide a kinetic and thermodynamic analysis of BMP9 and BMP10 interactions with ALK1 and type II receptors. Our data show that BMP9 displays a significant discrimination in type II receptor binding, whereas BMP10 does not. We also report the crystal structure of a fully assembled ternary complex of BMP9 with the extracellular domains of ALK1 and ActRIIB. The structure reveals that the high specificity of ALK1 for BMP9/10 is determined by a novel orientation of ALK1 with respect to BMP9, which leads to a unique set of receptor-ligand interactions. In addition, the structure explains how BMP9 discriminates between low and high affinity type II receptors. Taken together, our findings provide structural and mechanistic insights into ALK1 signaling that could serve as a basis for novel anti-angiogenic therapies.

TGF-␤ superfamily proteins are potent determinants of cell fate in metazoans. In mammals this family consists of more than 30 ligands, including bone morphogenetic proteins (BMP), 5 growth and differentiation factors (GDF), activins, seven type I receptors (ALK1-7), and five type II receptors (ActRIIA, ActRIIB, BMPRII, TGF␤RII, and AMHRII), as well as several co-receptors and accessory proteins (1). TGF-␤ family proteins commonly act by forming a ternary signaling complex that comprises dimeric ligand, two type I receptors, and two type II receptors. According to currently held views, signaling complexes can form in two ways: either the ligand binds first with high affinity to the type II receptor followed by recruitment of the type I receptor, as is generally observed for TGF-␤ and activin receptors (2,3), or the ligand binds first with high affinity to the type I receptor, followed by type II receptor recruitment, as is typical of BMP receptors (4,5). In either case, ligand binding leads to trans-phosphorylation of the type I receptor kinase domain by the apposed, constitutively active type II receptor kinase domain. Type I receptor phosphorylation then activates SMAD signal transduction pathways to regulate expression of target genes (1).
A remarkable feature of the TGF-␤ superfamily is the broad specificity possessed by most of its members, which results in a great combinatorial complexity of partnering between ligands, type I and type II receptors. For example, BMP2 has been shown to bind two type I receptors (ALK3, ALK6) with comparable affinities and three type II receptors (BMPRII, ActRIIA, and ActRIIB) also with comparable affinities (6,7). Similarly, multiple ligands, including activin A and B, GDF8 and GDF11, bind to ActRIIB with comparable affinities (8). An interesting exception to this general pattern is the type I receptor ALK1, which seems to have a strikingly discriminate relationship with BMP9 and BMP10. Only ALK1, but not other type I receptors, binds these two ligands with high affinity in vitro (9). Conversely, BMP9 and BMP10, which exhibit 65% sequence identity at the protein level, appear to signal exclusively via ALK1 in endothelial cells (10 -12), although published data suggest that they could utilize other low affinity type I receptors in other cell types where ALK1 is not expressed. In addition, it is unclear which type II receptors act in combination with ALK1/BMP9 and ALK1/BMP10. Based on chemical cross-linking and limited in vitro experiments BMPRII, ActRIIA, and ActRIIB have all been implicated in BMP9 signaling (10 -13), but these interactions have not been analyzed systematically; even less is known about BMP10 type II receptor utilization.
ALK1 is predominantly expressed in vascular endothelial cells (14). Mutations in the ALK1 gene have been shown to cause hereditary hemorrhagic telangiectasia type 2 (HHT2), a vascular disorder characterized by abnormal capillaries, recurrent nosebleeds, mucocutaneous telangiectasias, and arteriovenous malformations in the brain, lungs, liver, and gastrointestinal tract (15). Multiple lines of genetic, pharmacological, and histopathological evidence support a critical role for ALK1 signaling in the regulation of blood vessel formation in tumors as well as normal tissues (16 -19), explaining the growing interest in ALK1 as a therapeutic target to complement existing anti-angiogenesis treatments. Currently, a soluble ALK1 ligand trap (Acceleron Pharma) and an anti-ALK1 antibody (Pfizer) are in clinical trials for treatment of advanced solid tumors (20).
Given the biological and clinical importance of the ALK1 pathway, we sought to elucidate the molecular and biophysical basis of ALK1 complex formation. Here, we provide a kinetic and thermodynamic characterization of BMP9 and BMP10 binding to ALK1 and different type II receptors. Our data show that BMP9 displays a significant discrimination in type II receptor binding, whereas BMP10 does not. Importantly, we present the crystal structure of a fully assembled high-affinity complex between BMP9 and the extracellular domains of the type I receptor ALK1 and the type II receptor ActRIIB refined at 3.35 Å. The structure reveals the molecular determinants of ALK1 selectivity for BMP9 and BMP10, and the comparatively discriminate binding of BMP9 to its high affinity type II receptor. Last, we map extracellular HHT2 mutations to the ALK1 structure to explain the structural mechanism of HHT2 pathogenesis.

EXPERIMENTAL PROCEDURES
Construction of Expression Vectors-The expression constructs pAID4T ALK1.EK.Fc and pAID4T ActRIIB.EK.Fc contained the extracellular domain (ECD) of human ALK1 (residues  and ActRIIB (residues 19 -134), respectively, followed by an enterokinase (EK) cleavage motif fused to the human IgG1 Fc domain. hBMP9 cDNA harboring the tissue plasminogen activator signal peptide was cloned into the pAID4 vector and transfected into a CHO DUKX B11 cell line engineered to express soluble PACE/Furin to facilitate propeptide cleavage.
Protein Expression and Purification-ALK1.EK.Fc and ActRIIB. EK.Fc were expressed by Gala Biotech using GnTI-deficient HEK293S cells (21). The conditioned media was purified over a MabSelect SuRe column (GE Healthcare). Purified ALK1 and ActRIIB ECDs were digested with enterokinase (prepared in house), treated with Endo Hf (New England Biolabs), and further purified on a Superdex S200 (GE Healthcare). A proprietary receptor-Fc fusion protein was used to prepare an affinity resin for BMP9 purification. BMP9-conditioned media was concentrated 10-fold before affinity purification. BMP9 was eluted from the affinity resin with 0.1 M glycine (pH 3.0) and further purified by reverse phase HPLC.
Structure Determination-A 3.35-Å dataset was collected inhouse (Rigaku FR-Eϩ SuperBright TM /Cu anode) at 100 K on a single crystal equilibrated in mother liquor containing 15% ethylene glycol (for processing and refinement statistics, see supplemental Table S1). A MR solution, containing six ternary complexes in the asymmetric unit, was found with PHASER (22) using crystal structures of BMP9 (PDB 1ZKZ) (13), ActRIIB (PDB 2H62) (23), and an in-house structure of ALK1 6 as search models. The model was refined using PHENIX (22), applying NCS restraints to all equivalent residues. Manual building was performed in COOT (24). Electron density was continuous for all protomers, with the only disordered regions at the N and C terminus. No waters were picked in the final model, because of the resolution. The final structure has an R free of 26.1% and R cryst of 21.9%. Ramachandran analysis by Mol-Probity (25) indicates 88% of residues reside in the most favorable regions. Contact maps and buried surface area values were calculated using the Protein Interfaces, Surfaces, and Assemblies (PISA) server (26). Structural figures were prepared using PyMOL (27). The atomic coordinates and structure factor amplitudes for the ALK1 ECD /BMP9/ActRIIB ECD structure have been deposited in the Protein Data Bank (PDB ID code 4FAO).
Surface Plasmon Resonance Analysis-Experiments were performed using a Biacore T100 biosensor (Biacore/GE Healthcare) essentially as described (28). Briefly, goat anti-human Fcspecific antibody was immobilized onto a Series S research grade CM5 sensor chip (Biacore/GE Healthcare) using standard amino-coupling chemistry. Receptor-Fc fusions were captured on the flow cell at a density of ϳ100 Resonance units. For kinetic analysis at 25 and 37°C, a concentrated series of BMP9 or BMP10 (RnD) were injected in duplicates or triplicates over the flow cells at a flow rate of 50 l/ml. To obtain kinetic rate constants the corrected data were fit to a 1:1 interaction model using BiaEvaluation software (GE Healthcare). The equilibrium binding constant K D was determined by the ratio of binding rate constants k d /k a .

BMP9 Exhibits Strikingly Different Binding Kinetics and
Affinities for Type II Receptors-Previous studies have suggested that BMP9 signaling can occur via several type II receptors, namely BMPRII, ActRIIA, and ActRIIB (10 -13); however, BMP9-receptor interactions have not been measured directly and thus the importance of a particular receptor interaction is difficult to ascertain. We therefore determined the relative affinities of BMP9 for its type II receptors using surface plasmon resonance (SPR). We captured ActRIIB ECD -Fc, ActRII-A ECD -Fc, and BMPRII ECD -Fc in separate trials on a CM5 chip with immobilized anti-hFc IgG to mimic the dimeric nature of the receptors on the cell surface and exposed each type II receptor to a concentration series of BMP9 at 25°C. Our results reveal that BMP9 binds ActRIIB with ϳ30-fold higher affinity (K D of 21.6 pM) than BMPRII (K D 627.6 pM) (Fig. 1). Interestingly, the BMP9 interaction with ActRIIA is ϳ300-fold weaker (K D 6.43 nM) than with ActRIIB and is characterized by very fast association and dissociation of the complex. Other known type II receptors, TGF␤RII and anti-Müllerian hormone type II receptor (AMHRII), did not show any binding to BMP9 at a concentration of 100 nM (data not shown). We repeated the experiment at 37°C and confirmed the overall trend in receptor specificity at physiological temperature ( Fig. 1). Notably, BMP9/ActRIIB binding was least affected (K D ϭ 33.0 pM) by temperature, confirming the high specificity of this interaction. At 37°C, the BMPRII/BMP9 interaction was ϳ64-fold weaker (K D ϭ 2.11 nM) compared with ActRIIB, mainly due to a significant decrease in receptor-ligand complex stability.
BMP10 Exhibits Comparable Binding Kinetics and Affinities for Type II Receptors-Because little is known about which type II receptors are involved in BMP10 signaling, we investigated the selectivity profile of BMP10 for type II receptors, in part to compare BMP10 selectivity to that of BMP9. We first screened known type II receptors (ActRIIA, ActRIIB, BMPRII, TGF␤RII, and AMHRII) for binding to BMP10 using SPR. Whereas TGF␤RII and AMHRII did not display any binding to BMP10 as expected (data not shown), all three receptors previously shown to bind BMP9 also bound BMP10. We therefore performed a detailed kinetic analysis of BMP10 interactions with ActRIIB, ActRIIA, and BMPRII, at 25 and 37°C (Fig. 2). In contrast to BMP9, we found that BMP10 did not have a distinct preference for one of these three type II receptors at 25°C, as these receptors bound BMP10 with very similar kinetics and high affinities (K D ϭ 20.7, 88.6, and 59.2 pM, respectively). However, analysis of these interactions at 37°C suggested that, at physiological FIGURE 1. Kinetic and thermodynamic analysis of BMP9 binding to ALK1, ActRIIB, ActRIIA, and BMPRII. A, SPR sensograms of ALK1 ECD -Fc, ActRIIB ECD -Fc, ActRIIA ECD -Fc, and BMPRII ECD -Fc binding to BMP9 at 25 and 37°C. Raw data (colored lines) were overlaid with a global fit to a 1:1 model with mass transport limitations (black lines). ActRIIA ECD -Fc data were analyzed using affinity fit in SigmaPlot. B, kinetic parameters for receptor-ligand interactions. C, comparison of thermodynamic parameters of ALK1 ECD -Fc, ActRIIB ECD -Fc, and ActRIIA ECD -Fc interactions with BMP9 at 25°C. temperature, BMP10 exhibits a slight preference for ActRIIB and ActRIIA (K D ϭ 17.3 and 33.3 pM, respectively) over BMPRII. This is due mainly to the greater stability of the BMP10-ActRIIB complex and faster rate of BMP10-ActRIIA complex formation.
Kinetic and Thermodynamic Analysis of Receptor Discrimination-To permit a direct comparison of affinity measurements between the type I and II receptors, we also performed the SPR assay with captured ALK1 ECD -Fc at 25 and 37°C. Our results (Figs. 1 and 2) show that ALK1 binds BMP10 with ϳ2-4-fold better affinity than BMP9 (K D ϭ 10.3 pM for BMP10 and 45.2 pM for BMP9), which correlates well with previously published data (9). The temperature dependence of ligand binding to ALK1 was negligible, as shown by similar equilibrium binding constants at 37°C (K D ϭ 31.4 pM for BMP9 and 14.8 pM for BMP10) (Figs. 1B and 2B). In summary, our kinetic data reveal that BMP9 and BMP10 bind with picomolar affinities to both type I receptor ALK1 and type II receptor ActRIIB.
To investigate whether receptor discrimination is thermodynamically driven, we performed a thermodynamic characterization of BMP9 and BMP10 binding to ALK1 and ActRIIB by SPR on a Biacore T100 (see supplemental "Experimental Procedures" and Fig. S1) using the same receptor-Fc capture for-mat described above. Thermodynamic parameters indicate that the free energies of binding (⌬G 0 ) for BMP9 and BMP10 to their cognate type I and II receptors are almost identical within experimental error: Ϫ14.11 Ϯ 0.02 kcal mol Ϫ1 for BMP9/ALK1 and Ϫ14.30 Ϯ 0.30 kcal mol Ϫ1 for BMP9/ActRIIB; Ϫ14.88 Ϯ 0.05 kcal mol Ϫ1 for BMP10/ALK1 and Ϫ14.52 Ϯ 0.05 kcal mol Ϫ1 for BMP10/ActRIIB (Figs. 1C and 2C, and Table 1). Thus, there is no thermodynamic preference for the order in which either ligand (BMP9 or BMP10) binds type I receptor (ALK1) versus type II receptor (ActRIIB) in vitro.
Structure of the Ternary Complex of ALK1 ECD -BMP9-ActRIIB ECD -We determined the crystal structure of the highaffinity BMP9-ALK1 ECD -ActRIIB ECD complex at 3.35-Å resolution (see "Experimental Procedures" and supplemental Table  S1). The choice of ActRIIB as the type II receptor in the complex was based on its high affinity for BMP9 in our SPR assays and not on the prevalence of specific receptor combinations in vivo. The structure is composed of one BMP9 dimer, two ALK1 ECD molecules, and two ActRIIB ECD molecules. The overall assembly of the complex resembles that of other reported BMP-receptor ternary complexes (6,29), with the BMP9 homodimer located in the center of the complex, and two alternating molecules of ALK1 and ActRIIB bound on the outside (Fig. 3A). Specifically, one ActRIIB molecule binds on the out- side convex face of each BMP9 protomer. Two antiparallel ␤-sheets on BMP9 (fingers 1 and 2, also known as the "knuckle" region) mediate binding to a concave hydrophobic patch on ActRIIB (6) (Fig. 3A). The ALK1 interface is formed through recognition of both BMP9 protomers on the "wrist" (6), with the majority of binding occurring through packing of the ALK1 ␣1 helix against the pre-helix loop and ␣1 helix of BMP9 (Fig.  3A). Consistent with other BMP ternary complexes, the receptors make no direct contact with one another.
Because there are multiple copies of the ternary complex in this crystal structure, we also observe a higher ordered assembly composed of a dimer of ternary complexes (supplemental Fig. S2). Dimerization occurs between ActRIIB molecules on adjacent ternary complexes, facilitating the formation of a large ring-shaped structure. Interaction at the dimer interface is mediated predominantly through Asn 24 , which is glycosylated and makes glycan-mediated contacts with the opposing ActRIIB protomer. Although not reported, the same dimerization is evident in the ternary structure of BMP2-ALK3 ECD -ActRIIA ECD upon analysis of crystal packing (PDB code 2G00) (29). Other published ternary structures within the family are not glycosylated, and thus do not exhibit this arrangement.
ALK1 Receptor Orientation Shapes a Novel Interaction Surface-Despite similarities in overall architecture, there are clear differences between the structure presented here and the structures of homologous BMP complexes determined previously. Superposition with the BMP2/ALK3 ECD /ActRIIB ECD ternary structure (PDB code 2H64) (23) via one ligand protomer (Fig. 3B, red protomer) shows that ALK1 is significantly displaced relative to ALK3 with C␣ root mean squares differences of 3.74 and 4.31 Å for protomers C (61 atoms) and D (63 atoms), respectively (Fig. 3B). The receptors align closest along helix ␣1, although the helix is shifted by as much as 3.5 Å in ALK1 compared with equivalent residues on ALK3 (Fig. 3, C  and D). This shift is offset partly by a parallel rigid body movement in BMP9, where the tips of fingers 1 and 2 shift by about 2.5 Å relative to BMP2 (Fig. 3D). Thus, ALK1 pivots around helix ␣1 by about 20°relative to ALK3. This rotation repositions several key structural elements (Fig. 3, C and D), notably the F2 loop (residues 57-63), which shifts by as much as 6 Å to lie within contacting distance of BMP9. This is of particular note, as the F2 loop does not participate in ligand recognition in other type I receptor complexes. The F1 loop (residues 36 -44) also rearranges significantly, by as much as 9 Å, relative to ALK3 (Fig. 3, C and D), but in this case the effect is opposite; the loop moves away from the ligand interface and only loosely contacts BMP9. Alignment of this complex with the BMP2/ALK3 ECD / ActRIIA ECD ternary structure (PDB code 2G00) (29) and GDF5/ALK6 ECD binary structure (PDB code 3EVS) (30) shows a similar trend in ALK1 orientation, validating expectations that the ALK1-ligand interface diverges from other type I receptor-BMP interfaces. The changes in ALK1 orientation therefore contribute to a unique set of receptor-ligand interactions.
ActRIIB also shifts relative to the BMP2/ALK3 ECD /ActRII-B ECD structure, but here a rigid body movement in fingers 1 and 2 of BMP9 mirrors the shift (Fig. 3B). Thus, ActRIIB is able to maintain canonical receptor-ligand contacts. Last, there are no significant conformational changes in BMP9 upon complex formation. Similar to BMP2, the BMP9 homodimer binds as a rigid unit, with a C␣ root mean square difference of 1.01 Å compared with the apo structure (13). This is in contrast to activin, which is a highly flexible molecule that has been crystallized in multiple conformations (31,32), and GDF5, which undergoes an induced conformational change upon receptor binding (30).
The BMP9-ALK1 Binding Interface-In the complex described here, one ALK1 molecule contacts two BMP9 protomers. The binding interface is broadly separated into three sites (sites I, II, and III) (supplemental Table S2). Sites I and III overlap with those of other type I receptor-ligand complexes, whereas site II is unique to ALK1.
The interface at site I comprises BMP9 residues that contribute to the pre-helix loop (Phe 42 , Phe 43 , and Pro 44 ) and together with His 7 , His 62 , and Leu 63 form a hydrophobic interface with ALK1 residues His 40 , Val 54 , Val 56 , Arg 57 , His 66 , and His 87 (Fig.  4). Phe 43 and Pro 44 are conserved in BMP2 (Phe 49 , Pro 50 ) and GDF5 (Phe 54 , Pro 55 ) and form similar hydrophobic contacts with ALK3 and ALK6. Distinctively, residues in the F2 loop, including Glu 58 and Glu 59 , directly contact the ligand (Fig. 4B), whereas the F2 loop does not participate in recognition in the BMP2-ALK3 or GDF5-ALK6 complexes, as Glu 58 and Glu 59 are replaced by the shorter amino acid aspartate in both ALK3 (Asp 66 and Asp 67 ) and ALK6 (Asp 46 and Asp 47 ) (supplemental Fig. S3). Thus, in ALK3 and ALK6, the F2 loop is too far away to mediate any ligand interactions, suggesting that the F2 loop may play a role in conferring specificity to the ALK1-BMP9 complex.
Site II interactions are ostensibly unique to this complex and absent in the other known structures described to date, BMP2-ALK3 and GDF5-ALK6. Site II comprises a band of charged residues on ALK1 helix ␣1 and loop F3 that form a network of

Structure of ALK1-BMP9-ActRIIB Complex
AUGUST 10, 2012 • VOLUME 287 • NUMBER 33 polar interactions with residues on fingers 1 and 2 of one BMP9 protomer and on helix ␣1 of the second BMP9 protomer, BMP9Ј (Fig. 4). ALK1 Arg 78 is particularly prominent in that it forms a network of salt bridges with BMP9 Asp 89 , along with hydrogen bonds to the backbone carbonyl of Lys 87 and Asp 88 , and a main chain hydrogen bond to Ser 24 (Fig. 4B). ALK1 Arg 80 and Glu 83 form a similarly prominent network of contacts with BMP9. Although Arg 78 is conserved in ALK6 (Arg 69 ) and conservatively replaced by a lysine in ALK3 (Lys 88 ), there are no functional substitutes for ALK1 Arg 80 (Ser 90 in ALK3; Thr 71 in FIGURE 3. Structure of the ALK1 ECD -BMP9-ActRIIB ECD ternary complex. A, the ternary complex comprised of one BMP9 homodimer (red/blue), two ALK1 ECD receptors (gold), and two ActRIIB ECD receptors (green). The complex is shown as viewed from above (left) and from the side with the membrane below (right). B, superposition with the BMP2-ALK3 ECD -ActRIIB ECD ternary complex (shown in ribbon) via the ligands reveals differences in receptor positioning (BMP2 homodimer colored pink/light blue; ALK3 ECD and ActRIIB ECD receptors colored light gray and dark gray, respectively; ALK1 ECD -BMP9-ActRIIB ECD complex colored as in panel A). C, superposition of ALK1 ECD and ALK3 ECD at the BMP9 interface (shown in surface representation) reveals a 20°rotation of the ALK1 protomer pivoted around helix ␣1. D, close-up view of helix ␣1, F1, F2, and F3 loops from ALK ECD and ALK3 ECD (shown as schematic). Repositioning of these elements in ALK1 results in a novel set of receptor-ligand interactions.
ALK6) or Glu 83 (Ala 93 in ALK3; Pro 74 in ALK6) (Fig. 4C). Thus, not surprisingly, the F3 loop is positioned too far away in the structures of ALK3 and ALK6 to make contact with the co-crystallized ligand. Also in contrast to those previous observations, ALK1 Glu 83 and Phe 84 contact the second BMP9 protomer, BMP9Ј (Fig. 4B), where, conversely, several unique BMP9 residues interact with ALK1. Of these, Lys 64 is invariant in BMP9 and BMP10, suggesting that the ALK1/Glu 83 -BMP9/Lys 64 interaction is important for ALK1 ligand recognition (Fig. 4D). Ser 24 is also invariant in BMP9 and BMP10 (Fig. 4D), whereas all other TGF-␤ ligands have an Asp in the corresponding position, potentially introducing an unfavorable charge interaction with the main chain carbonyl of Arg 78 .
Site III presumably holds a key ALK1 specificity determinant. It is present in all type I receptor complex structures and consists of residues along the ALK1 pre-helix loop and helix ␣1, as well as residues from the ␣1 helix of one BMP9 protomer and the concave face of the second protomer, BMP9Ј (Fig. 5A). This site is largely hydrophobic in the BMP2/ALK3 and GDF5/ ALK6 structures, characterized by a conserved phenylalanine on helix ␣1 (Phe 85 in ALK3; Phe 66 in ALK6). This phenylalanine makes a "knob into hole" packing interaction into a hydrophobic pocket on the ligand. ALK1 is the only type I receptor that does not have a phenylalanine or equivalent hydrophobic residue in this position, but instead contains a glutamate (Glu 75 ), thus raising the question of how the charged residue in this B FIGURE 4. Site I and II interactions at the ALK1/BMP9 interface. A, close-up of the interaction surfaces, with the color scheme retained from Fig. 3. ALK1 is shown in schematic, with key residues drawn as sticks; BMP9 is shown in surface representation. For clarity the overall complex is shown to the side. B, detailed close-up of the interfaces, showing specific polar contacts (dashed lines) between residues along the F2 and F3 loops of ALK1. Key residues are highlighted with asterisks. C, sequence alignment of type I receptors (ALK1-7), with contact residues highlighted in gold. D, sequence alignment of BMP9 and BMP10 with select TGF-␤ ligands, showing key contact residues from both BMP9 protomers (red and blue). Asterisks coincide with panel B. position can fulfill a similar role? The structure described here shows that Glu 75 does not point into the conserved hydrophobic pocket as presumed, but, due to the unique receptor positioning, aligns structurally with Asp 84 (ALK3) and Asp 65 (ALK6) at the n-1 position on helix ␣1 (Fig. 5B). Thus, ALK1 Glu 75 fulfills a similar role as the aspartate residues in ALK3 and ALK6, forming a hydrogen bond with Tyr 99 on the periphery of the hydrophobic pocket. Importantly, Glu 75 forms additional salt bridges with Lys 53 on helix ␣1 of the second BMP9 protomer, BMP9Ј. This lysine is "locked" into the binding site by His 73 on the pre-helix loop of ALK1. Significantly, sequence and structural alignments indicate that the interactions of His 73 and Glu 75 with Lys 53 are found exclusively at the BMP9 and BMP10/ALK1 interface, therefore suggesting that these three residues are key specificity determinants (Fig. 5, C and D). Consistent with this view, substitution of Glu 75 with phenylalanine as in ALK3-ALK6 or valine as in ALK7 almost completely abolishes BMP9 binding in an SPR assay (Fig. 5F) and reduces cell signaling activity by ϳ5-and ϳ170-fold, respectively, compared with wild-type ALK1 (Fig. 5E). Additionally, replacement of ALK1 His 73 by glycine (ALK3, -4, and -6) significantly reduces BMP9 binding as determined by SPR (data not shown) and impairs cell signaling activity ϳ93-fold, whereas substitution by aspartate (ALK5) reduces BMP9 signaling ϳ25-fold.
A comparison of the interfacial residues from BMP9 with the sequence of BMP10 confirms that the two ligands share functional similarity, as 16 of the 24 interacting residues are fully conserved (supplemental Fig. S4). Five of the remaining 8 residues are conservatively substituted, and these substitutions may account for the small difference in binding of the two ligands to ALK1. Last, Glu 18 and Lys 87 in BMP9 (Lys 17 and Leu 86 in BMP10) interact with ALK1 through main chain atoms. Asp 89 of BMP9, which is involved in Site II interactions, is absent in BMP10.
The BMP9/ActRIIB Binding Interface-Unlike the BMP9/ ALK1 interface, the BMP9/ActRIIB interface closely resembles that of other type II receptor-ligand complexes. The majority of the ActRIIB residues that interact with BMP9 are also involved in interactions with Activin A (32). A total surface of ϳ860 Å 2 per BMP9 protomer and ϳ880 Å 2 per ActRIIB ECD are buried at the interface, with 23 residues from BMP9 and 22 residues from ActRIIB participating in the interaction (Fig. 6 and supplemen- tal Table S3). The majority of the interactions are hydrophobic with relatively few polar contacts. The conserved ActRIIB residues Tyr 42 , Trp 60 , and Phe 83 make up the core of the hydrophobic ligand-binding site and BMP9 residues Ala 28 , Leu 85 , and Leu 95 bind this site as expected (supplemental Fig. S5). These are the same residues that have been previously identified as important for ActRIIB binding to BMP2 and Activin A (29). However, uniquely observed here is an insertion of BMP9 Tyr 97 (or BMP10 Phe 95 ) that creates a small bulge, which changes the structure of an otherwise conserved ␤ strand (Fig. 6C). Tyr 97 buries into a hydrophobic pocket formed by the conserved ActRIIB Cys 31 -Cys 59 disulfide on one side and Leu 61 on the other side, whereas forming a hydrogen bond with the backbone carbonyl of ActRIIB Glu 32 (supplemental Fig. S5). Also of note is a partial helix formed by ActRIIB Asp 63 , Phe 64 , and Asn 65 . ActRIIB Phe 64 is located so that it can form a cationinteraction with BMP9 Arg 14 (supplemental Fig. S5). Similar interactions, although not as well defined, are also found in co-structures of BMP2 with ActRIIB and activin A with ActRIIB (23,32).
Molecular Basis of BMP9 Selectivity for ActRIIB over ActRIIA-Our SPR results show that BMP9 binds ActRIIB with ϳ300fold greater affinity than ActRIIA. In contrast, the closely related ligand BMP10 has a more relaxed receptor binding profile and is able to bind ActRIIB and ActRIIA with similar, high affinities. Alignment of the BMP9 and BMP10 sequences reveals that nearly half of the residues (11 of 23) at the BMP9/ ActRIIB interface differ between the two ligands (supplemental Fig. S4). Also, only 11 of 22 interfacial residues are conserved between ActRIIA and ActRIIB (Fig. 6B). Thus, it is not surprising that BMP9 and BMP10 show differences in type II receptor binding. Notably, BMP9 residues Arg 14 , Lys 30 , and Ser 80 are likely to hinder binding to ActRIIA (Fig. 6D). Lys 30 is within van der Waals bonding distance of ActRIIB Glu 76 , but in ActRIIA this glutamate is replaced by a lysine (Lys 76 ), which introduces a potential steric clash (Fig. 6D) and the prospect of charge repulsion. BMP10 has a proline (Pro 29 ) in place of Lys 30 , suggesting a mechanism for recognition tolerance at this position. BMP9 Ser 80 is also likely involved in receptor discrimination as it forms a hydrogen bond with ActRIIB Glu 34 ; replacement of Glu 34 with aspartate in ActRIIA (Asp 34 ) likely abolishes this stabilizing interaction. Last, ActRIIB Phe 64 , which forms a cation-interaction with BMP9 Arg 14 , is replaced by isoleucine in ActRIIA (Ile 64 ). Loss of this stabilizing interaction may be partially compensated for in BMP10 with the aromatic tyrosine (Tyr 13 ) in place of BMP9 Arg 14 . Mapping HHT2 Mutations to the ALK1 Structure-Mutations in the ALK1 gene are associated with HHT2, an autosomal dominant vascular dysplasia (15). We mapped 32 HHT2 mutations (www.hhtmutation.org) located in the ALK1 ECD to the ALK1 structure as determined here and were able to confirm which mutations play a role in destabilizing core structural elements and which ones disrupt interactions with BMP9 (Fig.  7A). Consistent with previous predictions, mutations that map to the highly conserved cysteine residues disrupt the core intramolecular disulfide bonds (33,34). We found that only three mutations (H66P, G79R, and H87D) directly affect residues at the BMP9 interface. His 66 and His 87 form part of the conserved hydrophobic pocket in site I (Fig. 4B); mutation of His 66 (H66P) is expected to deform strand ␤4, whereas mutation of His 87 (H87D) introduces an unfavorable negative charge directly into the binding site. ALK1 Gly 79 is on the F3 recognition loop (Fig.  4B); mutation to a bulky charged residue (G79R) would be predicted to cause clashing at the BMP9 interface and prevent binding. The remaining mutations are distal to the BMP9 binding site and are predicted to have an effect on protein folding and stability. Notably, Gly 48 , Ala 49 , and Trp 50 connect ␤2-␤3 strands and, along with Arg 47 on ␤2 and Thr 52 on ␤3, stabilize the conformation of the pre-helix loop and helix ␣1, which harbor BMP9 recognition residues Asn 71 , Leu 72 , His 73 , and Glu 75 (Fig. 7B). Mutations of Arg 47 , Gly 48 , and Ala 49 to bulkier (G48R, G48E, A49L) or conformation-restrained (R47P, A49P) residues would deform the backbone and ablate these stabilizing interactions. Indeed, introduction of the corresponding mutations in ALK1 ECD by site-directed mutagenesis resulted in protein misfolding and aggregation, which led to abrogated binding to BMP9. 7 Mutation of Thr 52 (T52A) would also lose stabilizing contacts with Asn 71 . Similarly, Trp 50 mutations (W50G, W50C) eliminate stacking interactions with residues on helix ␣1 (notably Arg 74 ) and introduce an unpaired cysteine. Other residues of note include Arg 67 and Asn 96 , which form a large network of charged interactions between residues on strand ␤4 and residues at the C terminus (Fig. 7C). ALK1 Arg 67 forms hydrogen bonds with residues Asn 96 , His 97 , and Val 99 , as well as a salt bridge to a neighboring Glu 65 . Similarly, ALK1 Asn 96 , which is invariant in the type I receptors, forms a salt bridge and backbone contact to Thr 35 , as well as a hydrogen bond to Thr 52 and Gly 68 . Adding negative charge (N96D and R67Q) or removing charge (R67W) would be expected to disrupt this network of interactions.

DISCUSSION
ALK1 has long been described as an "orphan" receptor as its signaling partners remained obscure. In an early attempt to identify its signaling partners, chemical cross-linking experiments suggested that TGF-␤ can interact with ALK1 in endothelial cells (16); however, direct binding of TGF-␤ to ALK1 could not be confirmed with SPR (9). More recently, BMP9 and BMP10 were identified as ALK1 ligands (9 -11, 13, 35). In a thorough kinetic and thermodynamic analysis, we conclusively demonstrate here that ALK1 binds BMP9 and BMP10 with high affinity and specificity. Interestingly, whereas the affinities, kinetics, and Gibbs free energies of binding (⌬G 0 ) for 7 R. Castonguay and A. V. Grinberg, unpublished data. ALK1/BMP9 and ALK1/BMP10 are very similar, the thermodynamic signatures of interactions are different. The BMP9-ALK1 complex is driven by strongly favored entropy (ϪT⌬S 0 ϭ Ϫ23.6 Ϯ 0.5 kcal mol Ϫ1 ), which is significantly offset by positive enthalpy (⌬H 0 ϭ 9.5 Ϯ 0.5 kcal mol Ϫ1 ), whereas the BMP10-ALK1 complex is characterized by a favorable input of entropy (ϪT⌬S 0 ϭ Ϫ17.2 Ϯ 1.1 kcal mol Ϫ1 ) and only mild offset of positive enthalpy (⌬H 0 ϭ 2.4 Ϯ 1.1 kcal mol Ϫ1 ).
Because ALK1-specific ligands have not been recognized until recently, it has not been possible to determine which type II receptors partner with ALK1 to form a functional signaling complex. Three different type II receptors, BMPRII, ActRIIA, and ActRIIB, had been suggested for BMP9 signaling, and thus indirectly implicated in ALK1 signaling (11)(12)(13); less is known about type II receptors involved in BMP10 signaling. Here we identify the type II receptors that interact with BMP9 and BMP10, and consequently form part of the ALK1 signaling pathway. Among receptor candidates suggested previously, we find that BMP9 is more selective for the type II receptor ActRIIB. Binding to the putative cognate type II receptor BMPRII is significantly diminished, whereas binding to ActRIIA appears to be transient in vitro, due to the extremely fast dissociation of the BMP9-ActRIIA complex. This is especially true at physiological temperatures, where BMP9 barely binds ActRIIA, whereas its affinity for ActRIIB remains unchanged. BMP10, however, exhibits a less discriminating type II receptor binding profile. We show that BMP10 can bind ActRIIB, ActRIIA, and BMPRII with similar affinities; although its binding profile is slightly more stringent at physiological temperatures. Nonetheless, BMP10 likely has the ability to form signaling complexes with ALK1 and all three type II receptors ActRIIA, ActRIIB, and BMPRII in vivo. Thus despite apparent sequence homology and possibly similar functions BMP9 and BMP10 could act through different type II receptors.
BMP9 and BMP10 are also noteworthy as potentially the only TGF-␤ ligands that can bind to both type I and II receptors with high affinity. As shown here, both ligands bind their cognate type I and II receptors with almost indistinguishable kinetic and thermodynamic profiles. Moreover, we have not observed any significant cooperativity between ECDs of ALK1 and ActRIIB receptors, as the affinity of ActRIIB for BMP9 in complex with ALK1 is similar to the affinity of ActRIIB for the uncomplexed ligand (supplemental Fig. S6). These data are corroborated by the absence of any conformational changes in the structure of ALK1, ActRIIB, or BMP9 upon complex formation, as well as the absence of any direct contacts between the receptor ECDs. This situation contrasts with other known TGF-␤ signaling complexes, which follow a two-state binding model where the ligand first binds with high affinity to either the type I or II receptor (as the case may be), followed by recruitment of the lower affinity receptor. Thus, our results suggest a new, nondiscriminative mechanism of ALK1-BMP9-ActRIIB complex assembly that is different from other TGF-␤ family signaling complexes. The significance of this finding in vivo remains to be determined. Critical analysis of literature data suggests that the type I/type II receptor combination in the BMP9 signaling might be cell type specific. It is possible that high affinity of BMP9 for ActRIIB is essential for signaling in cells where ALK1 is not expressed; whereas other lower affinity type II receptors might be recruited to a preformed ALK1-BMP9 complex in endothelial cells.
To better understand the structural basis for the strict ligand-receptor specificity in ALK1 signaling, we elucidated the structure of the fully assembled ALK1-BMP9-ActRIIB signaling complex. The structure shows that ALK1 achieves its particular specificity through a unique receptor orientation and a distinct set of interfacial contacts, as could not be predicted from previously available structures. ALK1 does not make the canonical knobs into hole packing interaction conserved among all other type I receptors, but instead a distinctive conformation of its helix ␣1 positions ALK1 Glu 75 as would be ALK3 Asp 84 and ALK6 Asp 65 and not, as expected, in position of ALK3 Phe 85 and ALK6 Phe 66 . Indeed, replacement of ALK1 Glu 75 with a hydrophobic residue to mimic the canonical phenylalanine in this position abrogates the ability of ALK1 to bind BMP9. A consequence of this ALK1-specific conformation is that helix ␣1 and loop F3 bulge outward to within contacting distance of the second BMP9 protomer. ALK1 loop F2 also contacts the ligand, whereas it does not participate in recognition in the BMP2-ALK3 or GDF5-ALK6 complexes. Thus, given the unique set of receptor-ligand interactions as well as the offset in receptor positioning, it is not surprising that ALK1 has such a discriminating ligand binding profile.
The structure presented here also clarifies the structural consequences of specific pathogenic mutations associated with HHT2. HHT2 is a rare disease characterized by vascular malformations in the lung, liver, and brain that is often referred to as a conformational disease, as the mutant protein is rendered nonfunctional through misfolding or aggregation (15,36). In the absence of a structure, however, the molecular mechanism of pathogenesis cannot be confirmed. Here we show that the majority of HHT2 mutations affect residues that are integral to the core structure, whereas only three residues are predicted to have a direct effect in BMP9 binding. These results may be useful for building a more comprehensive understanding of the contribution of particular mutations to the clinical pathology and severity of the disease in HHT2 patients.
Over the last decade, our understanding of the molecular mechanisms of angiogenesis has improved dramatically. This has led to the approval of several anti-angiogenic therapies targeting the VEGF pathway for treatment of various cancers. Although highly effective, these therapies do give rise to tumors that resist anti-VEGF therapies. Notably, it has been suggested that ALK1 is responsible for compensatory signaling and vascularization in tumors that are refractory to VEGF inhibition. Thus it is not surprising that application of ALK1 ECD -Fc resulted in blockade of several VEGF but also non-VEGF angiogenic pathways, suppression of tumor growth in MCF7 mammary adenocarcinoma (9), and antitumor activity in patients with advanced solid tumors in Phase I clinical trials, 8 emphasizing that development of ALK1 directed therapies could pro-vide a mechanism to block the escape pathways of tumor progression, especially of VEGF refractory tumors. Taken together our results provide the first structural and mechanistic insight into ALK1 angiogenic signaling response, and as such establish a firm structural basis for a novel class of angiogenesis inhibitors.