CryoEM structure of ALK2:BMP6 reveals distinct mechanism that allow ALK2 to interact with both BMP and activin ligands

Significance

Activin receptor-like kinase-2 interacts with two different classes of transforming growth factor β ligands: the Bone Morphogenetic Proteins (BMPs) and the activins. With BMP ligands, ALK2 signals to play roles in bone modeling, while with activins, ALK2 regulates signaling by forming nonsignaling complexes. We demonstrate that ALK2 binds to favored ligand, BMP6, by utilizing the ligand wrist-helix stabilized by a glycosylation, rather than a charged interaction observed with similar receptor, ALK3. We also show that ALK2 interaction with activin ligand, Activin A, is reliant on a single interaction at the opposite receptor:ligand interface. Thus, this study elucidates the structure of ALK2 in complex with a ligand and provides the molecular insight into how ALK2 binds to BMP and activin ligands.

Abstract

Ligands in the transforming growth factor β (TGF-β) family [activins, Bone Morphogenetic Proteins (BMPs), and TGF-βs] signal by bringing together two type I and two type II receptors. Activin receptor-like kinase-2 (ALK2) is the only type I receptor among the seven TGF-β type I receptors that interacts with both activin and BMP ligands. With BMPs, ALK2 acts as a signaling receptor to activate small mothers against decapentaplegic 1 (SMAD1)/5/8 signaling. Alternatively, with activins, such as Activin A (ActA), ALK2 forms nonsignaling complexes that negatively regulate ALK2 and ActA signaling. To gain insight into how ALK2 interacts with two distinct classes of ligands, we resolved the cryoelectron microscopy structure of ALK2 in complex with the type II receptor, ActRIIB, and the ligand, BMP6, in parallel with the corresponding structure with ALK3 for direct comparison. These structures demonstrate that ALK2 and ALK3 utilize different mechanisms to interact with BMP6 at the wrist interface, with ALK2 relying on BMP6 glycosylation and ALK3 relying on a salt bridge. Modeling of ALK2:ActA reveals that binding relies on ActA’s fingertip region, mirroring the interaction of ActA with its other receptor, ALK4. Our results demonstrate that ALK2 is a “hybrid” receptor that incorporates features of BMP type I receptors such as ALK3 at the wrist interface and an activin type I receptor such as ALK4 at the fingertip.

The transforming growth factor β (TGF-β) family is composed of over 30 structurally similar ligands that fall into three major classes, each defined by downstream SMAD activation profiles and receptor specificity (1). Signaling is initiated by the formation of complexes composed of one dimeric ligand, two type II receptors, and two type I receptors with each ligand class using distinct mechanisms of complex formation (2). With five type II and seven type I receptors in mammals, there is a receptor bottleneck where an elaborate web of specificity, affinity, and competition has evolved to govern ligand:receptor interactions (1, 3). Bone Morphogenetic Proteins (BMPs), which contain more than 20 members, are the largest of the classes and activate the small mothers against decapentaplegic 1 (SMAD1)/5/8 pathway through three type II receptors (ActRIIA, ActRIIB, and BMPR2) and four type I receptors (ALK1/ACVRL1, ALK2/ACVR1, ALK3/BMPR1a, and ALK6/BMPR1b). Each ligand:receptor combination possesses distinct signaling effects resulting from the cell-specific receptor repertoire, ligand availability, as well as binding potency (3). ALK2 and its ligands present an example of this observation in that the response of ALK2 to BMPs can be negatively modulated by Activin A (ActA) both in vitro and in vivo (4, 5).

Interestingly, initial studies identified ALK2 (Activin receptor type 1) as an activin type I receptor (6, 7). However, further experiments revealed that ActA does not activate ALK2, whereas BMP ligands, such as BMP6 and BMP7, do (8). Hence, ALK2 was reclassified as a BMP receptor and the observed interaction with ActA was considered artifactual (8). This view was later revised when it was shown that in the genetic disorder, fibrodysplasia ossificans progressiva (FOP), single amino acid changes in the intracellular domain of ALK2 render this receptor responsive to ActA (4, 9). Subsequently, we showed that wild-type ALK2 forms nonsignaling complexes with ActA and associated type II receptors (4, 5, 10) and that these nonsignaling complexes traffic to the lysosome wherein their components are degraded (11). These findings reestablished ALK2 as an activin receptor and indicated that ALK2 has evolved the unique capability to interact with two distinct classes of ligands: the Activins and the BMPs.

To date, there is no published structure of ALK2 in complex with a ligand. This is likely due to difficulties not only producing active recombinant receptor for structural studies but validating the receptor’s proper folding and activity as well. Previous efforts to determine binding kinetics for ALK2 have yielded only mild success for several BMP ligands (5). This is in direct contrast to the nanomolar affinity seen between the other BMP type I receptors such as ALK1 and ALK3 (12–14). Furthermore, the ligand with the highest affinity for ALK2, BMP6, is reliant on N-linked glycosylation to bind (15). Interestingly, whereas this posttranslational modification on BMP6 is necessary for interaction with ALK2, it is dispensable for ALK3 or ALK6, suggesting differences in the manner of type I engagement.

To explore, at the molecular level, how ALK2 interacts with its cognate ligands, we utilized cryoelectron microscopy (cryoEM) to resolve the ternary complex structures of ALK2:ActRIIB:BMP6 and ALK3:ActRIIB:BMP6. Comparisons of these structures, along with biochemical validation, illustrate two mechanisms for stabilizing type I:BMP6 interactions which dictate the ability to bind to both ALK2 and ALK3. Extending our analysis of ALK2 through modeling and mutagenesis, we pinpoint a key residue in ALK2 that is critical for interaction with ActA, but not BMP6.

Results

ALK2-ActRIIB and ALK3-ActRIIB Bind BMP6 with High Affinity.

To alleviate potential concerns with the low ligand affinity of ALK2, we designed fusion molecules featuring ALK2 tethered to the type II receptor, ActRIIB, as the latter generally has higher affinity for BMP ligands than ALK2 (5). This strategy of tethering the lower-affinity type I receptor to the higher-affinity type II receptor has been used previously to capture low-affinity interactions as well as for engineering ligand-blocking therapeutics, commonly referred to as ligand traps (16–19). For this study, the extracellular domains (ECDs) of human ALK2 (1 to 123) or ALK3 (1 to 152) and ActRIIB (1 to 133) were fused to IgG1 Fc domains (SI Appendix, Fig. S1). ALK2-ActRIIB-Fc and ALK3-ActRIIB-Fc were generated through coexpression and purified to homogeneity prior to biochemical validation (SI Appendix, Fig. S1).

To determine whether the receptor ECDs within our traps were properly folded, we utilized surface plasmon resonance (SPR) to kinetically measure the binding of both heterodimeric traps to mammalian expressed (glycosylated) BMP6 (SI Appendix, Fig. S2). Additionally, we measured binding of two additional traps: one with a single ActRIIB ECD and another containing two ActRIIB ECDs (ActRIIB₁-Fc and ActRIIB₂-Fc, respectively). Comparison of the affinities of the different traps revealed that the heterodimeric molecules displayed a high binding affinity for BMP6 similar to ActRIIB₂-Fc (36.6 pM; ALK2-ActRIIB and 30.0 pM; ALK3-ActRIIB versus 42.1 pM; ActRIIB₂-Fc), and higher than that observed with ActRIIB₁-Fc (108 pM) (SI Appendix, Fig. S2 and Table S1). Consistent with the SPR results, we were able to form stable complexes between both traps and BMP6 as determined by size exclusion chromatography (SEC) (SI Appendix, Fig. S1). These data indicated that in both heterodimeric traps, each of the receptor ECDs is active and binding to BMP6.

CryoEM Structures of ALK2-ActRIIB and ALK3-ActRIIB Bound to BMP6.

We next sought to resolve the structure of ALK2-ActRIIB-Fc:BMP6. Initial cryoEM data collection of the complex revealed that the Fc portion of the trap was interfering with the resolution of the structure. Removing the Fc portion of the heterodimeric traps through IDeS (Immunoglobulin G-degrading enzyme) digestion, which cleaves below the disulfide hinge of antibodies (20), had no significant impact on the ability of ALK2-ActRIIB and ALK3-ActRIIB (cleaved from the Fc-domain, but still joined together by the hinge disulfides), to form stable complexes with BMP6 as determined by SEC (SI Appendix, Fig. S1). Data collection on this smaller, ALK2-ActRIIB:BMP6 complex yielded a high-resolution structure containing a complete complex with two ActRIIB-ECD and two ALK2-ECD bound. The cryoEM structure of ALK2-ActRIIB:BMP6 was resolved to 3.2Å, and the map was well defined for residues: ALK2 (31 to 109), ActRIIB (26 to 118), and the mature BMP6 ligand (410 to 513; 82.3% of mature domain) (Fig. 1 and SI Appendix, Fig. S3 and Table S2). We used a similar experimental setup to resolve the structure of ALK3-ActRIIB bound to BMP6 to 3.3Å with residues, ALK3 (50 to 141), ActRIIB (26 to 117), and BMP6 mature domain (410 to 513), being well-resolved. In both structures, BMP6 adopts the canonical dimeric fold of TGF-β ligands, akin to what is commonly referred to as “two hands”: two monomeric molecules held together by a cystine knot, where each contains two long β-ribbons of two antiparallel β-strands each (fingers) and a large alpha-helix (wrist) (Fig. 1) (1, 21). The overall complex assembly of ligand:receptors is consistent with previous BMP complex structures, with the type II receptor binding at the convex, “knuckle” of the β-ribbons and the type I receptor binding at the interface built from the concave fold of the fingers and the wrist helix (22).

Fig. 1.

CryoEM structures of BMP6 in complex with ActRIIB-ALK2 and ActRIIB-ALK3. (A and B) Two views of the 3.23 Å and 3.3 Å resolution cryoEM maps and resulting structures of BMP6 in complex with ActRIIB-ALK2 (A) and ActRIIB-ALK3 (B), respectively. ALK2 (wheat) and ALK3 (gray) bind in the composite Type I interface built from both BMP6 monomers (purple/pink). ActRIIB (orange) binds at the ligand knuckle interface, dependent on a single monomer.

Similar to previous structural efforts, we were unable to confidently build the antibody linker between the type II and type I arms of either ALK2-ActRIIB or ALK3-ActRIIB, likely due to flexibility (19). However, when contoured to a low level, the cryoEM maps suggest that the antibody linker is between the cis-pair of receptors (i.e., the type I and type II receptors that bind on opposite sides of the same ligand monomer’s fingers) and that the two heteromeric traps bind to the ligand in a cis-manner (SI Appendix, Fig. S4). Measurements between the terminal tails for the different receptor pairs is consistent with what has been previously reported (~35 Å for the cis-pair and ~70 Å for the trans-pair) (2). It is possible, then, that tethered receptors, such as in ligand-independent receptor dimers (23), would prefer binding in a cis-manner as well. The distance between this cis-pair is also consistent with the modeled spacing of intracellular domains (24). However, it is possible that this relative positioning of the heterodimeric receptor ECDs (cis) is due to the constraints imposed by the fusion design and contributes to the slightly increased affinity when compared to ActRIIB₂-Fc in SPR.

BMP6 Binding to ActRIIB, ALK2, and ALK3.

Both structures indicate that ActRIIB binds at the knuckle of BMP6, utilizing a highly conserved set of interfacial residues observed across numerous type II complex structures containing both BMP and Activin class ligands (21). Specifically, a hydrophobic triad (Tyr⁶⁰, Trp⁷⁸, and Phe¹⁰¹) in ActRIIB engages an “IAP” motif on BMP6 (Ile²², Ala²³, and Pro²⁴) to form the core of the receptor interface. The exact receptor pair (ALK2-ActRIIB or ALK3-ActRIIB) of the trap has little effect on ActRIIB during BMP6 binding (rmsd = 0.632 Å over 92 Cα atoms), indicating there are no significant constraints induced by the tethered type I receptor ectodomain.

In both structures, the type I receptor binds within the concave, composite interface built from both ligand monomers, albeit with different surface area utilization. During interaction with ALK3, BMP6 buries a total of ~1,325.4 Ų, similar to that of ALK3:BMP2 (25), while only burying 1,113.8 Ų while engaging ALK2. Comparison between the two structures reveals that ALK2 engages a much smaller degree of surface area (203.6 Ų less) on BMP6 monomer B, and that the pattern of surface engagement is significantly different with each type I receptor engaging different regions of the ligand (Fig. 2). Notably, the β2β3 loop of ALK2 contacts a charged pocket on BMP6 forming several charged/polar interactions between several ALK2 (Ser⁴¹, Asn⁶⁰, and Asp⁶⁰) and BMP6 (Asn⁴⁵⁰, Ser⁴⁴⁶, Glu⁴⁴⁴, and Lys⁴⁷⁵) residues (Fig. 2B). The ALK3 β2β3 loop is positioned distantly and in fact, an overlay of the two structures shows a 27° shift of overall receptor positioning in the concave ligand cleft. Conversely, ALK3 caps the wrist helix with residue Gln¹¹⁷, engaging Asn⁴⁶⁹ of the posthelical loop of BMP6, whereas ALK2 does not (Fig. 2C). This interaction is facilitated by a three amino acid extension in the β4β5 loop of ALK3, not present in ALK2 (Fig. 2D). Further comparison of ALK2 and ALK3 reveals that the alpha helix that is formed in the ALK2 β4β5 loop is uniquely extended an additional half turn when compared to that of ALK3, as well as other BMP type I receptors, ALK6 or ALK1 (Fig. 2D). Despite this conformational difference, no major contacts are observed between this alpha helix extension and BMP6.

Fig. 2.

ALK2 and ALK3 utilize distinct binding interactions to bind BMP6. (A) Ligand surface representation of the Type I interface, with BMP6 residues interacting with either ALK2 or ALK3 colored in wheat and gray, respectively. Buried surface area calculations for each of the interfaces are shown following determination in PISA. (B) Comparison of the β2β3 loop interface during interaction with ALK2 (Left) and ALK3 (Right). (C) Comparison of interactions between the posthelical region of BMP6 with ALK2 (Left) and ALK3 (Right). (D) Comparison of ligand-bound BMP type I receptors: BMP6-bound ALK2 and ALK3, GDF5-bound ALK6 [blue, PDB: 3EVS (26)], and BMP9-bound ALK1 [light blue: 4FAO (14)].

Glycosylation of the BMP6 Prehelical Loop Stabilizes Interaction with ALK2.

The differences in the manner of engagement of BMP6 by ALK2 versus ALK3 are accentuated further by the fact that only glycosylated BMP6 can activate ALK2 (15, 27). In vivo, BMP6 exists as several species corresponding to glycosylated and nonglycosylated forms (28, 29). Glycosylation of the prehelix loop determines whether BMP6 can activate ALK2; enzymatically deglycosylated or bacterially produced BMP6 (which is not glycosylated) is unable to bind in SPR experiments or signal through ALK2 (15, 27). We corroborate these previous findings by showing that BMP6 treated with PNGaseF (Peptide N-glycosidase F), which cleaves between the innermost GlcNAc residue and the asparagine residue, completely removing glycosylation from BMP6 (dgBMP6), does not interact with ALK2 on SPR (SI Appendix, Fig. S2B); whereas BMP6 binds ALK2-ActRIIB-Fc with a KD of 36.6 pM, dgBMP6 has a 3.5-fold reduced KD of 127 pM, similar to the affinity observed with ActRIIB₁-Fc (110 pM) (SI Appendix, Fig. S2B and Table S1). Conversely, deglycosylation of BMP6 does not reduce binding to ALK3-ActRIIB-Fc to the same degree (KD of 30.0 pM and 35.1 pM for BMP6 and dgBMP6, respectively).

To correlate binding to signaling capability, we tested the activity of BMP6 and dgBMP6 to activate Smad1/5/8 in HEK293 cells. Additionally, to probe the degree of glycosylation required for signaling, we treated BMP6 with EndoH (N-endoglycosidase H), which cleaves the β1-4 glycosidic bond between the first and second GlcNAc residue, leaving the first carbohydrate attached to the protein (SI Appendix, Fig. S5A). Following treatment of the Humanembryonic kidney (HEK) cells, we observed robust activation of Smad1/5/8 for each of the three molecules (SI Appendix, Fig. S5B). Next, we supplemented either an ALK2 blocking antibody or an ALK3 blocking Fab to parse out signaling through each specific receptor (30). While blocking ALK2, all BMP6 glycostates activate signaling, indicating that activation of ALK3 signaling is not significantly affected by the carbohydrate moiety. However, when interaction with ALK3 is blocked, only glycosylated BMP6 and EndoH-treated BMP6 are able to activate signaling, whereas dgBMP6 does induce signaling, indicating that the first carbohydrate moiety is required for activation of ALK2 by BMP6. These results were confirmed in mouse embryonic stem cells (mESCs) that are either wild type or are lacking either ALK2 or ALK3 (SI Appendix, Fig. S5C). Consistent with the results obtained in HEK cells, both BMP6 and dgBMP6 activate signaling in ALK2^−/− mESCs (where signaling is dependent on ALK3), whereas only fully glycosylated BMP6 can activate signaling in ALK3^−/− mESCs (where signaling is dependent on ALK2).

Several structures of BMP ligands, particularly BMP2, show that the prehelix loop is flexible but is stabilized in a conserved conformation upon type I receptor binding (14, 25, 26, 31–34). Given the requirement for glycosylation along with the flexibility of the loop that contains it, we utilized the structure of ALK2-ActRIIB:BMP6 to explore the carbohydrate’s role in ALK2 binding. The cryoEM structure of ALK2-ActRIIB:BMP6 displays a strong signal for the first three carbohydrate moieties of the prehelical Asn⁴⁵⁴-linked glycosylation (2× N-acetylglucosamine and 1× β-D-mannose). No molecular contacts are seen between carbohydrate chain of BMP6 and ALK2, suggesting that there is no direct contribution of the carbohydrate chain to ALK2 binding (Fig. 3A). Notably, previously generated structural models of ALK2 bound to BMP6 suggest that residues Lys³¹ and Tyr⁷⁴ directly interact with the glycosylation (15), whereas our structure reveals that both residues are positioned on the opposite side of the receptor and away from the wrist interface, hence precluding direct binding (SI Appendix, Fig. S6) (15). We do observe, however, several interactions between the carbohydrate chain and specific residues in the same monomer chain of BMP6. Specifically, the primary carbohydrate moiety forms a stabilizing hydrophobic interaction with His⁴⁵², as well as forming a hydrogen bond with Lys⁴⁸² (Fig. 3A). Interestingly, from our signaling data with EndoH-treated BMP6 (SI Appendix, Fig. S5) and our SPR binding data (SI Appendix, Fig. S2), these two interactions are sufficient to stabilize BMP6’s prehelix loop into a conformation suitable for ALK2 binding and signaling, mirroring previous structures of type I–bound BMP ligands. Consistent with this idea, no significant structural differences are observed in the backbone chain of BMP6 when comparing the ALK2 and ALK3-bound structures (rmsd = 0.624 Å over 208 Cα atoms), supporting the overall rigidity of BMP6 and the necessity for a specific prehelix loop conformation.

Fig. 3.

BMP6 utilizes two compensatory mechanisms to stabilize type I binding. (A) Schematic view of ALK2-ActRIIB:BMP6 with comparison of the β1β2 interface, highlighting the position of the N-linked glycosylation (primary NAG in green; Second NAG and BMA in blue) on the BMP6 prehelix and the formation of a salt bridge at the same interface between ALK3:BMP6. Engineered β1β2 loop swap between ALK2 and ALK3 colored in light blue in both structures. (B) SPR sensorgrams of BMP6, deglycosylated BMP6, and BMP2 binding to anti-human antibody-captured ALK3-Fc, ALK3.D69G-Fc, and ALK3.ALK2-β1β2Loop-Fc. Each experiment was performed in duplicate.

Given the lack of direct binding between the carbohydrate moiety of BMP6 and ALK2, it remains curious why ALK2 is reliant on the presence of the carbohydrates for BMP6 binding, whereas ALK3 is not. Comparison between the CryoEM structure of ALK2:BMP6 and ALK3:BMP6 reveals the presence of a salt bridge between ALK3 residue Asp⁶⁹ and BMP6 residue Arg⁴¹³, which sterically occludes the previously mentioned His⁴⁵² driven hydrophobic contacts in BMP6 (Fig. 3A). The β1β2 loop in ALK2 is three residues shorter and a glycine residue is in the position of ALK3 Asp⁶⁹ preventing a similar contact. Additionally, Arg⁴¹³ is positioned away from ALK2, further suggesting no interaction. We hypothesized that ALK3 utilizes this salt bridge to stabilize the prehelix loop without relying on the carbohydrate moiety.

To investigate this further, we produced and purified the homodimeric Fc-fusion of ALK3-ECD, and a corresponding version where Asp⁶⁹ is replaced with the homologous ALK2 residue, a glycine (ALK3.D69G). We also generated an ALK3 ECD variant where the β1β2 loop of ALK3, C67-C77, was replaced with the homologous residues C42-C48 of ALK2 (ALK3.ALK2-β1β2Loop) (Fig. 3A). The effects on ligand specificity and binding were determined through SPR binding studies with the receptor fusions captured on the biosensor via the Fc domain (Fig. 3B). Binding was observed between ALK3-Fc and both BMP6 and dgBMP6. However, significantly reduced binding of dgBMP6 was observed with ALK3.D69G. Similar effects were seen during binding to ALK3.ALK2-β1β2Loop. In contrast, BMP2 binding was similar between ALK3 and ALK3.D69G, where only introduction of the β1β2 from ALK2 reduces binding significantly (Fig. 3B). These data support that the salt bridge between ALK3 Asp⁶⁹ and BMP6 Arg⁴¹³ stabilizes binding and constitutes an alternative strategy to stabilize the prehelix loop in the absence of glycosylation. In contrast, ALK2, with a glycine in this position, is unable to form the salt bridge with BMP6 and is thus reliant on the presence of the glycosylation to stabilize the prehelical loop’s conformation for binding.

The Binding between ActA and ALK2 Is Reliant on Unique Fingertip:β4β5 Interactions.

Activin ligands, such as ActA or GDF11, uniquely rely heavily on their fingertips for type I receptor affinity as well as specificity, where extensive hydrogen bond networks are formed between residues within the ligand fingertip and the β4β5 of the type I receptor (2, 19). While a structure of ActA in complex with ALK2 proved difficult to resolve, we can draw insight from the structure of ALK2 bound to BMP6 for possible mechanisms of ALK2:ActA binding.

Analysis of the fingertip:β4β5 interface in our structures reveals that no direct interactions occur between the fingertip of BMP6 and ALK2 (Fig. 4A). This is consistent with previously reported type I:BMP structures, including our cryoEM structure of ALK3:BMP6. However, the additional helical turn in the β4β5 loop of ALK2 positions the C alpha carbon of Tyr⁷⁴ 3.8 Å closer to the ligand fingertip than the corresponding Tyr¹⁰³ residue in ALK3 (Fig. 4A). Next, we aligned both the structure of ALK4:ActA (PDB: 7OLY) to ALK2:BMP6 (SI Appendix, Fig. S7) as well as utilized AlphaFold to model ALK2:ActA (Fig. 4B) in order to visualize whether this shift in ALK2’s relative β4β5 position could facilitate unique contacts with ActA (35). Both approaches indicated that Asp⁴⁰⁶ of the ActA fingertip is well positioned to form a hydrogen bond with Tyr⁷⁴ of ALK2. Thus, we hypothesized that ALK2 is able to bind to ActA due to the unique positioning and conformation of the β4β5 loop and specifically residue Tyr⁷⁴.

Fig. 4.

ALK2 uniquely engages the ActA fingertip for interaction and signaling. (A) Schematic view of ALK2-ActRIIB:BMP6 with comparison of the fingertip interface between BMP6-bound ALK2, ALK3. (B) AlphaFold model of the ALK2:ActA complex highlighting the fingertip interface. (C) Luciferase reporter assays in HEK cells, stably expressing either ALK2.R206H or respective variants. Cells were treated with a titration of BMP6 (black lines) and ActA (red lines).

To probe this specific contact and its possible role on ActA:ALK2 interaction, we generated a set of stable cell lines expressing various ALK2 mutants starting with a parental HEK293 cell line expressing a Smad1/5/8 responsive luciferase reporter: three variations of Tyr⁷⁴ as well as a variation of knob-in-hole residue, Met⁷⁹. Given that wild-type ALK2 is not activated by ActA, we utilized the FOP variant, ALK2.R206H in our construct, hence enabling Smad1/5/8 signaling as the readout. As expected, both ActA and BMP6 activated ALK2.R206H cells (Fig. 4C). When Tyr⁷⁴ was changed to neutral residues, alanine or valine, ActA signaling was effectively ablated, while having minimal effect on BMP6-induced signaling (Fig. 4C). Substituting Tyr⁷⁴ of ALK2 to the corresponding residue in ALK3, glutamate, also rendered ALK2 unresponsive to ActA, but not BMP6. This corroborates with our previous finding that loss of the fingertip residue, Asp⁴⁰⁶, in ActA significantly reduced the ability of ActA to assemble ALK2 and type II receptor complexes, due to a reduction in ALK2 binding (5). Unexpectedly, altering the knob-in-hole residue, Met⁷⁹, to alanine did not ablate the signaling of either ligand, indicating that ALK2 relies more heavily on interactions outside of the conserved anchor interface to stabilize ligand binding. These results support our structural modeling and indicate that ActA engages ALK2 specifically with the ligand fingertip, relying on the unique structure of the ALK2 β4β5 loop and residue, Tyr⁷⁴.

Discussion

ALK2 is a type I receptor that binds two different ligand classes of the BMP/TGF-β superfamily: the BMPs and the Activins. To investigate how ALK2 accomplishes this task, we solved the cryoEM structures of both ALK2-ActRIIB and ALK3-ActRIIB bound to BMP6 and then modeled the interaction of ALK2 with ActA.

Comparison of the ALK2-ActRIIB and ALK3-ActRIIB structures reveals that although the BMP6 prehelix adopts the same conformation in both structures, the mechanism employed for stabilizing this shape of the prehelix is distinct between ALK2 and ALK3. With ALK2, BMP6 utilizes a glycosylation on Asn⁴⁵⁴ to form interchain contacts that stabilize the prehelix in the active conformation, whereas with ALK3, BMP6 utilizes a salt bridge to stabilize the same interface.

The requirement for BMP6 to be glycosylated to activate ALK2 had been previously recognized (15). Our results illuminate the molecular and signaling details of this requirement but also hint to a potential regulatory role of BMP6 glycosylation. Glycosylated and deglycosylated forms of BMP6 have been reported from previous studies assaying BMP6 in serum and human brain hippocampus (26, 30). Since BMP6 can bind the type II receptor irrespective of its glycosylation state (15, 27) and since ALK2 has recently been shown to exist on the cell surface in a heterodimeric state with its partner type II receptors (36), we surmise that BMP6 drives the formation of ALK2-ActRIIB heterotetramers solely through the binding of the two type II receptors, and hence in a manner that is independent of BMP6 glycostate. Such heterotetramers form with unglycosylated BMP6 would fail to signal, as only glycosylated BMP6 can activate ALK2. Supporting this, artificial formation of the heterotetrameric complex through dimerization of type II receptors failed to produce a signal, emphasizing ligand dependence for canonical assembly (23). These complexes may act to reduce the overall level of signaling mediated by ALK2 without affecting signaling via ALK3. Furthermore, we speculate that this mechanism may be shared with other ALK2 ligands such as BMP5, BMP6, BMP7, and BMP8 which feature a conserved prehelical sequence and are glycosylated similarly to BMP6 (SI Appendix, Fig. S8).

Next, we explored the other class of ligands that interact with ALK2, the Activins, represented by ActA. Given practical difficulties in structurally resolving an ALK2–ActA complex, we modeled ALK2 interaction with ActA based on the ALK2–BMP6 complex, using AlphaFold. The model revealed a key interaction between the ALK2 β4β5 loop and the fingertip of ActA, which we verified experimentally. Furthermore, we demonstrate that upon ligand binding, ALK2 undergoes a conformational change in the β4β5 loop that is greater than that seen with other BMP type I receptors, positioning ALK2 closer to the ligand fingers. This difference is likely due to the β4β5 loop having additional amino acids in the N-terminal half of the β4β5 loop when compared to ALK1, ALK3, or ALK6 (Fig. 2D). A similar extension is seen in the activin receptors, ALK4 and ALK5 but not ALK7. Our data show that while this altered position of the β4β5 loop is not critical for BMP6, it plays roles in ActA binding. Here, modulation of this prospective interface by mutagenesis of residues from either the receptor (ALK2.Y74E) or the ligand (ActA.deltaD406) completely ablates the interaction. Interestingly, this fingertip interaction is similar to what is observed during ALK4:ActA interaction, where the ligand fingertip residue D406 directly binds to ALK4.

Sequence alignment of ALK2 across a variety of species reveals a widespread conservation of the critical residue for ActA interaction, Tyr⁷⁴. Interestingly, in zebrafish and Drosophila, there is a proline in this position and in salmon, a serine (SI Appendix, Fig. S8). The ALK2 homolog in Drosophila, Saxophone (Sax), has been demonstrated to form nonsignaling complexes that sequester BMP homologs (37). Furthermore, these complexes are similarly converted to signaling complexes when a FOP-homologous mutation, K262H, is introduced to Sax (38). However, overexpression of Sax has no effect on either ActA or ActB homologs (Dawdle and dActB), confirming a lack of interaction. This is consistent with experiments in zebrafish that display increased BMP signaling with FOP mutants, but not due to ActA (39). Thus, it appears that while the ability to engage BMP ligands and form nonsignaling complexes is conserved between Sax and ALK2, interactions with the activin class are more evolutionarily recent, perhaps as a part of the evolution of ossified bone (4, 9, 40).

ALK2 remains one of the more perplexing TGF-β receptors, with interactions between two major branches of ligands, the BMPs and the Activins. The structural resolution of ALK2-ActRIIB:BMP6 and ALK3-ActRIIB:BMP6, and AlphaFold modeling of ALK2-ActRIIB:ActA reveals ALK2 as a “hybrid” receptor, able to act similarly to a BMP type I receptor at the wrist interface or an activin type I receptor (such as ALK4) at the fingertip. Additionally, these structures allowed us to directly compare how BMP6 binds to two distinct type I receptors highlighting the importance of BMP6’s carbohydrate moiety for binding to ALK2. Future work may explore the role of the ALK2 nonsignaling complex as well as comprehensively characterize the various glycostates of BMP6 and their potential roles in vivo. Overall, our structures serve to not only to detail the interactions between ALK2 and its ligands but also to advance our molecular understanding of TGF-β type I receptors and the mechanisms of specificity used to dictate each unique receptor:ligand combination.

Methods

Hetero- and Homodimeric Receptor Trap Expression and Purification.

ActRIIB_2-Fc along with ALK2-Fc, ALK3-Fc, and the corresponding mutant ALK2/3-Fc fusions were expressed and purified from Chinese hamster ovary (CHO) cells as previously described (17, 19). The heterodimeric traps were also produced and purified from CHO cells. Here, constructs containing each receptor-Fc fusion were designed to favor heterodimer formation through either the Fc-Star strategy (ALK2-ActRIIB-Fc) or a knob-in-hole strategy (ALK3-ActRIIB-Fc) (41, 42). To express each molecule, plasmids were cotransfected at a ratio of 60% ActRIIB to 40% type I receptor fusion into CHO cells. Following 10 d of expression, conditioned media was harvested, centrifuged, and filtered prior to loading through a MabSelect SuRe protein A column (Cytiva). Here, ALK2-ActRIIB-Fc was eluted through a pH gradient from pH 7.5 (20 mM Tris, 300 mM NaCl) to pH 3 (20 mM Sodium Citrate, 300 mM NaCl) to separate contaminating homodimeric ALK2₂-Fc. ALK3-ActRIIB-Fc was eluted in a batch manner. Semipurified protein was then further purified by SEC using a Superdex 200 increase column (Cytiva).

For structural studies, ALK2-ActRIIB-Fc and ALK3-ActRIIB-Fc were then digested with IdeS-His protease at 37 °C overnight, cleaving the Fc portion of each construct below the antibody hinge (20). The cleaved Fc and IdeS-His were then removed from ALK2-ActRIIB and ALK3-ActRIIB by both CaptureSelect (Thermo) and TALON (Cytiva) resins, respectively. To further purify the digested receptor fusions, the sample was loaded on a Superdex 200 increase column (Cytiva). 20 mM Tris 7.5, 300 mM NaCl was used as the running buffer for both resins and the final polishing SEC.

Human recombinant ActA was produced to homogeneity as previously described (5). Recombinant human BMP6 was purchased from RnD (Cat. No. 507-BP), where it was produced from CHO cells. For applicable data, BMP6 was deglycosylated with either rapid PNGase F (New England Biolabs) or EndoH (New England Biolabs) for either 15 min at 50° C or 4 h at 37° C, respectively, where digestion was confirmed through Sodium Dodecyl Sulfate Polyacrylamide gel analysis (SI Appendix, Fig. S5A).

CryoEM Sample Preparation and Data Collection.

Digested and purified ALK2-ActRIIB and ALK3-ActRIIB were mixed with BMP6 at a 2.1:1 receptor trap:ligand molar ratio and incubated at room temperature for 30 min. Complexes were purified through SEC on a Superdex 200 increase column (Cytiva). For each complex, distinct peaks consistent with expected complex molecular weights were observed and pooled for subsequent concentration for structural studies.

CryoEM grids of both purified complexes of ALK2-ActRIIB:BMP6 and ALK3-ActRIIB:BMP6 were prepared at a protein concentration of ~3 mg/mL. Samples were supplemented with PMAL-C8 amphipol (Anatrace) to a final concentration of 0.15% immediately prior to grid preparation to aid in vitrification. UltrAuFoil 1.2/1.3 grids were used for both protein samples following plasma cleaning in a Solarus II (Gatan) using a H₂O₂ gas mixture. A Vitrobot Mark IV (Thermo Fisher) operated at 4° C and 70% humidity was used for blotting the grids and plunge freezing them into liquid ethane cooled by liquid nitrogen.

Grids were loaded into a Titan Krios G3i electron microscope equipped with a Bioquantum K3 (Gatan). Images were collected in counted mode at a nominal magnification of 105,000×, yielding a pixel size of 0.839 Å. For ALK2-ActRIIB:BMP6 and ALK3-ActRIIB:BMP6, defocus ranges of −0.8 to −2.4 µm and −1.0 to −3.4 µm, respectively, were set for data collection in EPU (ThermoFisher). The energy filter was inserted with slit width 20 eV. Each movie was dose fractionated into 46 frames over a 2 s exposure and had a total dose of ~40 electrons per Ų. Further details of the data collections leading to the structures are shown in SI Appendix, Table. S2.

CryoEM Data Processing.

CryoEM data were processed using cryoSPARC v4, where the same general workflow was used for each sample (43). Exact workflows and details for each structure are available in SI Appendix, Fig. S3 and the statistics, in SI Appendix, Table S2. In summary, movies were motion and contrast transfer function corrected using cryoSPARC’s patch implementations. Micrographs with poor resolution estimates were removed from further processing. Particles were then blob-picked followed by multiple rounds of 2D classification to generate templates for particle picking and subsequent 2D and 3D classification. Once a subset of particles was obtained that resulted in a homogenous population and 3D volume, a topaz model was trained against particle coordinates from a subset of the initial exposures (44). The model was then used to pick particles across the entire dataset, which were subject to multiple rounds of 2D and 3D classification. Finally, maps were refined sequentially against alternative classes (Heterogeneous refinement) to further increase particle homogeneity and ultimately, C2 symmetry was applied to refinement, in the case of ALK2-ActRIIB:BMP6 and ALK3-ActRIIB:BMP6 (45). Sharpened and unsharpened maps were carried forward for visualization and model building.

Model Building and Refinement.

Manual model building was conducted in Coot 0.9.6, and real space refinement of models was conducted using Phenix 1.21 (37, 38). The initial model for building ALK2-ActRIIB’:BMP6 was constructed by combining ALK2 (PDB code: 7YRU), ActRIIB (PDB: 7MRZ), and BMP6 (2QCW). Each individual component was aligned to ALK3:ActRIIA:BMP2. This structure was then used as a starting model for building ALK3-ActRIIB’BMP6 after ALK3 (PDB code: 2GOO) was aligned in place of ALK2 (25). PyMOL and USCF Chimera X were used to visualize models and maps (46). Buried surface area and subsequent related calculations were carried out in Phenix 1.21-Pisa (38). The model of ALK2 bound to ActA was generated utilizing AlphaFold 2 (35). Five AMBER relaxed complex models were generated from input sequences featuring the ALK2-ECD and ActA-mature domain and the highest ranked model was used for analysis.

SPR.

SPR affinity analysis was performed on a Cytiva T-200 instrument using T200 Control Software version 3.2.1 (Cytiva) for data acquisition. A CM4 sensor chip (Cytiva) was prepared by EDC/NHS [1-ethyl-3-(3-dimethylaminopropyl) carbodiimide/N-hydroxysuccinimide] immobilization of a goat antihuman Fc-specific IgG (in-house) of approximately 3,000 RU. Experiments were carried out in N-2-Hydroxyethylpiperazaine-N’-2-ethanesulfonic acid (HEPES) Buffered Saline-EP+ buffer (10 mM HEPES pH 7.4, 350 mM NaCl, 3.4 mM ethylenediaminetetraacetic acid, 0.005% P-20 surfactant) at 25° C. A five-step, twofold serial dilution was performed in the aforementioned buffer for each ligand, with an initial concentration of 5 nM for the studies with ActRIIB₂-Fc, ALK2-ActRIIB-Fc, ALK3-ActRIIB-Fc, and ActRIIB₁-Fc and 20 nM for the studies with ALK2-Fc, ALK3-Fc, and the corresponding mutants. Each cycle had a ligand association and dissociation time of 300 s. The flow rate for kinetics was maintained at 30 μL/min. SPR chips were regenerated with 20 mM H₃PO₄ pH 2.5. Kinetic analysis was conducted using the Biacore T200 evaluation software using a 1:1 fit model with mass transport limitation (red lines).

Smad Activation Assays and Immunoblotting.

HEK cells were seeded in a 12-well format (0.22 × 10⁶ cells/well) and propagated for 18 h. ALK2^−/− and ALK3^−/− (ALK3^−/−; ALK2^{[R258H]FlEx/+}) mESCs were generated and propagated previously using CRISPR guides to biallelically ablate the target gene (36). Following propagation, growth medium was then replaced with serum-free medium supplemented with 0.1% bovine serum albumin (BSA) for 2 h. Medium was again replaced with serum-free medium supplemented with the desired ligand and, if applicable, an in-house ALK2-blocking antibody (REGN5168) or ALK3-blocking Fab (Bio-Rad; Clone AbD01564) (23, 30). Whole-cell lysates were collected with radio immunoprecipitation assay buffer lysis buffer containing 2× protease and phosphatase inhibitor and 0.1 U/mL nuclease (Thermo Fisher Scientific, 89901, 78440, and 88701, respectively). Protein concentration in the whole cell lysate was determined using the Pierce BCA protein assay kit (Thermo Fisher Scientific, 23227). For HEK cells, Phosphorylated Smad1/5/8 (1:500; Cell Signaling; Rabbit; 13820S) and Smad1 (1:1,000; Cell Signaling; Rabbit; 9743S) were detected utilizing the SimpleWestern platform with 2.4 μg total protein loaded into each lane (47). For mESCs, lysates were resolved under reducing conditions and transferred to PVDF membranes and probed for pSmad1/5 (1:1,000, Cell Signaling, 41D10) and Cyclophilin B (1:10,000, Cell Signaling, D1V5J).

Luciferase Reporter Assays.

Assays using the BRE (BMP responsive element) luciferase reporter cells were performed in a manner similar to those previously described (4). Briefly, cells were plated in a 96-well format (3 × 10⁴ cells/well) and propagated for 24 h. For EC50 experiments (Fig. 4C), growth medium was replaced with serum-free medium supplemented with 0.1% BSA and the desired ligand, where a threefold serial dilution was performed with a starting concentration of 54 nM (ActA and BMP6). Cells were incubated for 18 h before lysing and assaying for luminescence using a Victor X Light Multilabel Reader (Perkin Elmer, Waltham, MA). For the assays featuring stable expression of ALK2 and corresponding mutants, cell lines were generated through lentiviral transduction similar to previous work using reagents obtained from Origene (48). Here, constructs containing the full-length ALK2 receptor were designed in a pLVX vector backbone and cotransfected with lentiviral packaging plasmids into HEK293T cells. Viral particles were harvested 48 h later and used to transduce BRE luciferase reporter cells. Following selection of stable pools with hygromycin, cells were sorted for expression utilizing antibodies against ALK2 (Mouse; Mab637; RnD) and Mouse IgG (Donkey Alexa Fluor 647 conjugated; A-31571; Life Technologies). The activity data were imported into GraphPad Prism and fit using a nonlinear regression to calculate the EC₅₀ or IC₅₀.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The structures of ALK2-ActRIIB:BMP6 and ALK3-ActRIIB:BMP6 described in this paper have been deposited to both the Protein Data Bank (9N4K (49) and 9MIR (50), respectively) and Electron Microscopy Data Bank (EMD-48883 (51) and EMD-48301 (52), respectively).