Repulsive guidance molecules lock growth differentiation factor 5 in an inhibitory complex

Since previous studies revealed a high degree of structural similarity between the GDF5– and BMP2–receptor complexes (8, 26), we wondered whether the BMP2 coreceptor family of the RGMs has a similar effect on GDF5 signaling. We performed a GDF/BMP-responsive luciferase reporter assay in LLC-PK1 kidney cells, which are highly responsive to many BMP/GDF family members and were used to discover that RGMs act as BMP coreceptors (9–11, 17, 21, 27). Here, we show that all full-length membrane-anchored human RGMs enhance BMP2 signaling, consistent with previous observations (Fig. 1 A and B and SI Appendix, Fig. S1 A and B) (10, 17). When BMP2 was applied at 3 nM (which is below its half-maximal effective concentration, EC₅₀), the signaling increased about threefold in cells expressing RGMB and RGMC, and less prominently in cells expressing RGMA. The different levels of BMP2 signal enhancement might be related to different expression levels (SI Appendix, Fig. S2 A and B) and RGM–BMP2 binding affinities (17). RGM constructs lacking the N-terminal domain (RGM_ΔN) [that was previously shown to be essential for BMP2 binding (17)] failed to enhance BMP2 signaling (Fig. 1B and SI Appendix, Fig. S1 A–D), suggesting that modulation of BMP signaling requires RGM–BMP2 binding.

Unexpectedly and in contrast to BMP2 signaling, membrane-anchored RGMs inhibited GDF5 signaling in the same cellular assay. GDF5 signaling was down-regulated about threefold in cells expressing full-length RGMB and twofold for RGMA and RGMC (Fig. 1C and SI Appendix, Fig. S1 E and F). This effect was again dependent on the N-terminal domain of RGM, since constructs lacking this domain showed no effect on GDF5 signaling (Fig. 1C). Next, we tested whether soluble RGM fragments would also affect signaling. Soluble N-terminal domains of RGMA and RGMB did not attenuate GDF5 signaling as efficiently as membrane-anchored full-length RGMs in LLC-PK1 cells (SI Appendix, Fig. S2 C–F). This may potentially be due to avidity and the requirement for cell surface attachment, where two membrane-bound RGMs are favorably arranged to bind to the GDF5 dimer. The full extracellular domain of RGMB (RGMB_ECD) inhibited GDF5 signaling similarly to RGMB_ND, suggesting that the RGMB_ND, but not the NEO1-binding region of RGMB_ECD, is essential for the inhibitory function of RGMB (SI Appendix, Fig. S2F). Intriguingly, when BMP2 and GDF5 were added simultaneously, both the RGM-mediated inhibition of GDF5 signaling and the RGM-mediated activation of BMP2 signaling were diminished (SI Appendix, Fig. S2G). This suggests that GDF5 can down-regulate RGM-mediated activation of BMP2 signaling, a mechanism that could be essential for a number of biological processes where BMP2 and GDF5 are coexpressed, (e.g., limb and joint development) (28, 29; for recent reviews, see refs. 30 and 31).

Soluble BMPR1A ectodomain protein did not inhibit GDF5 signaling at the concentration used (SI Appendix, Fig. S2H). However, the fact that neutralizing anti-BMPR1A antibodies down-regulated GDF5 signaling (SI Appendix, Fig. S2 I and J) clearly shows that GDF5 signals are transmitted via BMPR1A into LLC-PK1 cells. Inhibition of the GDF5 signaling by anti-BMPR1A antibodies was not complete and could be further increased by the expression of full-length RGMB (SI Appendix, Fig. S2 I and J). These results suggest that the concentrations of the anti-BMPR1A antibodies (0.5 μM, which relates to a concentration of only about 10- to 25-fold of the equilibrium binding constant measured for the interaction of the respective Fab to soluble BMPR1A protein) were not sufficient to fully abrogate GDF5 signaling. In addition, the anti-BMPR1A antibodies used in this study are monovalent Fab antibody fragments that are much weaker competitors compared to a classic bivalent antibody that can simultaneously bind to two BMPR1A moieties at the cell surface. As would be expected for molecules that bind to an epitope highly overlapping with that for BMP type 1 receptors, soluble RGM fragments could down-regulate GDF5 (as well as BMP2) signals in other cell lines, such as the prechondrogenic (ATDC5) and the mesenchymal (C3H10T1/2) stem cell lines even though the half-maximal inhibitory concentration (IC₅₀) varied significantly (SI Appendix, Fig. S3).

To further dissect the RGM–GDF5 interactions, we tested three RGMB mutations that impair both RGMB-mediated potentiation of BMP2 signaling and RGMB–BMP2 interactions (SI Appendix, Figs. S4 and S5) (17). The two RGMB mutations Gly101Arg and Leu103Glu correspond to RGMC mutations that cause the blood iron overload disease juvenile hemochromatosis (17, 21, 27). The third mutation (RGMB His106Arg) abolished the RGMB–BMP2 interaction (17). In our assay, all three mutations impaired RGMB-mediated potentiation of BMP2 signaling as well as inhibition of GDF5 signaling (Fig. 1 D and E and SI Appendix, Fig. S1 C, D, G, and H). This suggests that both the RGMB-inhibitory effect on GDF5 and the RGMB-potentiating effect on BMP2 signaling are mediated by similar interactions between BMP2/GDF5 and membrane-anchored RGMs.

To characterize the RGM–GDF5 interactions, we determined the crystal structure of RGMB_ND in complex with GDF5 at 1.7 Å resolution (Fig. 1 F–I and Table 1). In the complex, the disulfide-linked GDF5 dimer binds two RGMB_ND molecules that are related by a noncrystallographic pseudo twofold axis (RMSD of 0.21 Å for 143 equivalent Cα atoms of the 1:1 RGMB:GDF5 complexes) (Fig. 1 F and G). RGMB_ND forms a bundle of three α-helices stabilized by three disulfide bonds (Fig. 1 F and G and SI Appendix, Fig. S5). It binds to the finger region of GDF5 and interacts with both GDF5 molecules of the dimer. The RGMB_ND–GDF5 interface comprises mixed charged and hydrophobic interactions (containing 116 nonbonded contacts and 5 hydrogen bonds) mediated by 20 residues from each binding partner and buries a total surface area of 1,929 Ų (SI Appendix, Fig. S6).

The RGMB–GDF5 interface is mainly formed by two interaction sites. First, a short linker between helices α2–α3 (Fig. 1H) harbors the “RGD-loop” in RGMA and RGMC (“RGN” in RGMB) and interacts with the GDF5 wrist region. The side-chain of RGMB Leu103 inserts into the hydrophobic pocket on GDF5 formed by Trp414, Trp417, Ile449, Tyr490, and Met493 (Fig. 1H) and is shielded by a hydrogen bond formed by RGMB Gly101 (“RGD-loop”) and GDF5 Tyr490. Second, the RGMB α3 helix stretches across the two GDF5 fingers. A major interface is formed by the side-chain of RGMB His106 that stacks onto GDF5 Trp417 (Fig. 1I). This π–π stacking is shielded from the solvent by RGMB Arg100. Our 1.7 Å crystal structure allowed us to identify a network of well-ordered water molecules contributing to the RGMB–GDF5 interactions. The side-chain of RGMB His106 forms a hydrogen bond with the amide of Gly101 from the “RGD-loop” and a water molecule that extends a hydrophilic network to Tyr105, Trp414, and Asp416 of GDF5 (Fig. 1I). Asp416 is evolutionarily conserved in 12 of 20 GDFs/BMPs and might modulate the specificity of RGM–GDF/BMP interactions (e.g., Asp416 corresponds to Ser343 in BMP9 that does not bind to RGMs) (32) (SI Appendix, Fig. S4).

We also determined the structures of the RGMA_ND–GDF5 and RGMC_ND–GDF5 complexes (Figs. 1A and 2 and Table 1). Superposition of the three RGM–GDF5 complexes reveals common interactions between all three RGMs and GDF5. However, while the GDF5 structure in the three complexes is largely invariant, the relative orientation of the three RGM_ND helix bundles towards GDF5 differs (Fig. 2 A and B). The RGM α3 helix, which is in the core of the RGM–GDF5 interface, undergoes a conformational change. The relative position of the α3 helix differs by ∼5 Å in RGMB_ND–GDF5 and RGMC_ND–GDF5 complexes (Fig. 2B), with its N terminus being the anchor point for a rotation of RGM_ND. The pivot point seems to be RGMB Leu103 (RGMA Leu97, RGMC Leu101) (Figs. 1H and 2 C and E). A relative rotation around this leucine permits the side-chain of the preceding arginine (Arg100 in RGMB) to be inserted between GDF5 fingers 1 and 2 in the RGMA– and RGMB–GDF5 complexes (Figs. 1I and 2D). In the RGMC_ND–GDF5 complex, however, this arginine side-chain is placed above finger 2 (Fig. 2F), potentially explaining the different arrangements of the RGM α3 helix. Taking these data together, we find the linker between RGM α2–α3 helices is essential for the RGM–GDF5 interactions, potentially directing the relative orientation of RGM in the complex. Intriguingly, mutations in the α2–α3 linker of RGMC (Gly99Val, Gly99Arg, Leu101Pro) cause iron overload (21, 27) and impair interactions with the closely related BMP2 (17).

We performed surface plasmon resonance (SPR)-based RGM–GDF5 equilibrium binding experiments to dissect the contribution of different RGM domains and residues to GDF5 recognition. GDF5 was immobilized on the SPR chip surface and RGMs were used as analytes. To avoid nonspecific interactions between RGMs and GDF5, which impair the analysis of the dose-dependency of equilibrium binding [as reported for the RGM–BMP2 interaction (17)], a high ionic strength buffer with a 10-fold higher detergent concentration than usually employed was used: 0.5 M NaCl, 20 mM Hepes pH 7.4, 0.05% Tween 20. This elevated detergent concentration generally weakened RGM–BMP2/GDF5 interactions compared to previous studies without altering binding specificity and thus allowed comparative analysis of our RGM mutants (17, 32). Under these conditions the RGMB ectodomain (RGMB_ECD) bound GDF5 with an apparent affinity (K_d) of 8.8 μM, and the N-terminal domain alone (RGMB_ND) exhibited a K_d of 2.7 μM, suggesting that RGMB_ND comprises the major domain for GDF5 binding (Fig. 3 A and B). Under the same conditions, RGMB_ND bound to BMP2 with a notably weaker (compared to GDF5) K_d of 20.1 μM (SI Appendix, Fig. S7A). GDF5 interacted similarly with RGMC_ND (K_d 3.9 μM) and weaker with RGMA_ND (K_d 16.7 μM) (Fig. 3 C and D). The juvenile hemochromatosis-causing mutation, Gly99Arg, in RGMC_ND and a corresponding mutation, Gly101Arg, in RGMB_ND attenuated the interactions with GDF5, K_d > 20 μM and > 150 μM, respectively (Fig. 3 E and F). Mutations of RGMB at either of two key RGMB–GDF5 interactions (Leu103Glu and His106Arg) (Fig. 1 H and I) abolished binding to GDF5 (Fig. 3 G and H). To compare RGMB–GDF5 interactions to previous BMPR1A–GDF5/BMP2 binding studies (33), we also measured binding under the conditions used in those studies (0.5 M NaCl, 10 mM Hepes pH 7.4, 3.4 mM EDTA, 0.005% Tween 20). Here, binding affinities of RGMB for GDF5 (K_d 2.7 ± 0.9 nM) and BMP2 (K_d 4.2 ± 1.2 nM) were similar (SI Appendix, Figs. S8 and S9 and Table S1). Finally, we tested the pH dependency of the RGMB–GDF5 interactions using size-exclusion chromatography (SEC) combined with multiangle light scattering (MALS). In solution the 2:2 RGMB–GDF5 complex is stable at pH 7.5, but dissociates at pH 5.5 (SI Appendix, Fig. S10), similarly to the RGMB–BMP2 complex.

Superposition of the BMPR1B–GDF5 and BMPR1A–BMP2–ActR2b complexes with binary RGM–GDF5 and RGM–BMP2 complexes, respectively, reveals that RGMs and type 1 receptors (BMPR1A and BMPR1B) bind to highly overlapping epitopes on GDF5 and BMP2 (Fig. 4 A and B). However, in contrast to previously determined RGM–BMP2 complexes (17), the relative orientations of our three RGM_NDs bound to GDF5 vary (Fig. 4 C–F). Strikingly, the superposition of GDF5-bound RGMB_ND and RGMC_ND requires translation by 5.6 Å and rotation by 15.4° (Fig. 4C). For comparison, the superposition of BMP2-bound RGMB_ND (two crystal forms) and RGMC_ND requires translation by 3.0 to 3.6 Å and rotation by 6.9 to 8.2° (Fig. 4E). While RGMs bind to at least five different BMPs (32), differences in their interactions might lead to distinct biological effects.

Our structures also reveal GDF5 regions engaged specifically in either BMPR1B or RGM recognition (Fig. 4 G and H). GDF5 His440 interacts with a cyclic loop of BMPR1B (Fig. 4G) (but not with any RGMs), whereas a tip of finger 2 (Phe478–Ser481) of GDF5 interacts with C-terminal residues of the RGM α2 helix (Fig. 4H). GDF5–type 2 receptor interactions are mediated by the knuckle epitope in finger 2, which is not part of the RGM–GDF5 interface. Nevertheless, RGMs were found to modulate interactions between BMP2/4 and their type 2 receptors (34) although it is unclear whether these receptors can form ternary complexes with RGMs–BMPs/GDFs. Using SEC-MALS, we show that RGMB_ND, GDF5 and ActR2b form a ternary complex in solution (Fig. 4I). This suggests that RGMs might modulate BMP/GDF–BMP type 2 receptor interactions (34) by altering the conformation of finger 2 and the knuckle epitope (Fig. 4H).

Next, to understand the opposing role of RGM on GDF5 and BMP2 signaling, we tested the effect of RGMB on signaling by two GDF5 mutants (GDF5 Arg438Leu and GDF5 Arg438Ala) that bind to BMPR1A with higher affinity, similar to that of BMP2, thus mimicking a BMP2-like behavior for BMPR1A binding (26) (SI Appendix, Table S1). Both GDF5 mutants activated signaling as found for wild-type GDF5 in LLC-PK1 cells but with a dose-dependency similar to BMP2 as would be expected for a BMP2 mimic (SI Appendix, Fig. S2 K and L). However, signaling by both GDF5 mutants was still inhibited by RGMB similar to that found for wild-type GDF5. Using an SPR-based binding assay we could furthermore show that these GDF5 mutants bind to all RGM ectodomains (RGMA_ECD, RGMB_ECD, and RGMC_ECD) with affinities identical to wild-type GDF5, indicating that only BMPR1A binding is changed by these mutations (SI Appendix, Fig. S9). Our additional analyses using these GDF5 mutants clearly indicate that the opposing effect of RGM on BMP2/GDF5 signaling cannot be explained with a simple RGM/type 1 receptor competition mechanism as this would have been addressed with these GDF5 mutants. Hence, an additional far more complex mechanism must play a role that allows recognition and discrimination of BMP2 and GDF5, despite structure analyses showing seemingly identical interaction schemes for both BMP ligands.

How the NEO1 and GDF5/BMP signaling pathways, which both control chondrogenesis and iron homeostasis, are interfacing, is only poorly understood (4, 11, 18, 22–24, 26). We hypothesized that GDF5 might form a ternary complex with RGMB and NEO1 similar to the ternary RGMB–BMP2–NEO1 assembly (17). To test this hypothesis, we determined the structure of the ternary complex between the two membrane proximal NEO1 fibronectin-type III domains (NEO1_FN56) (16), RGMB_ECD and GDF5 to 5.5 Å resolution. In the complex, the GDF5 dimer binds to two RGMB ectodomains in a similar arrangement to that in the binary RGMB_ND–GDF5 complex (Fig. 5 A and B and SI Appendix, Fig. S11). RGMB_ND is connected to the RGMB C-terminal domain (RGMB_CD) via a flexible linker (∼12 residues), as suggested by its fragmentary electron density. RGMB_CD is sandwiched between GDF5 and NEO1. A hydrophobic face of RGMB_CD (Phe162–Phe166, Arg172, Tyr268–Glu270, Phe281, Ala292) interacts with a hydrophobic region of GDF5 (Ile419–Leu422, Val485–Tyr487, Leu477–Ile479). However, this interaction is likely to play a minor role in the RGMB–GDF5 interactions, as RGMB_ND is sufficient to bind to GDF5 with high affinity (Fig. 3 A and B). RGMB_CD uses an opposite face of the β-sandwich to bind NEO1, resembling the main interface previously observed in the binary NEO1–RGMB complex (16).

The 2:2:2 NEO1–RGMB–GDF5 architecture is compatible with a potential arrangement of RGMs, NEO1, and GDF5 on the cell surface (Fig. 5 A and B). C termini of RGMB and NEO1 point in the same direction and could be accommodated on the surface of one cell to control downstream signaling. This overall arrangement resembles the previously determined NEO1–RGMB–BMP2 structure (17) (Fig. 5C), since the RGMB_ND–BMP2/GDF5 and RGMB_CD–NEO1 interfaces are conserved, and the C termini of RGMB and NEO1 point in the same direction so that both can be accommodated on the surface of one cell. Both NEO1–RGMB–GDF5/BMP2 complexes demonstrate the modular architecture of RGMs where RGM_ND and RGM_CD can control BMP/GDF and NEO1 signaling, respectively. This raises the question of whether NEO1 affects GDF5 signaling and the inhibitory effect of RGMs on GDF5. To address this, we performed GDF5-responsive cellular signaling assays in LLC-PK1 cells expressing either full-length NEO1 (NEO1_FL) or a NEO1 construct replacing its cytoplasmic domain with a green fluorescent protein (NEO1_ΔC) (17). In both cases, RGMB could inhibit GDF5 signaling at similar levels compared to RGMB alone (Fig. 5 D and E). Interestingly, both NEO1_ΔC and NEO1_FL increased GDF5 signaling in the absence of RGMB (Fig. 5 D and E), a similar effect to that previously shown for BMP2 signaling in chondrocytes (23) and myoblasts (35). However, the molecular basis of NEO1-mediated enhancement of BMP2/GDF5 signaling remains unknown. In summary, our structural analysis of the ternary NEO1–RGMB–GDF5 complex combined with cellular assays reveal the modular architecture of RGMB where its N-terminal domain acts as a robust inhibitor of GDF5 signaling, whereas its C-terminal domain interacts with NEO1.

Human RGMA_ND (residues Ser46–Ser139, UniProtKB Q96B86-1), RGMB_ND (Gln53–His136, Q6NW40-1), RGMB_ECD (Gln53–Ser411, Q6NW40-1), RGMC_ND (Gln36–Pro145, Q6ZVN8-1), BMPR1A_ECD (Thr55–Val141, P36894-1) (17), fibronectin type-3 domains 5 and 6 of mouse NEO1 (Thr883–Asn1123, P97798-4) (16) were cloned into the pHLsec vector (48), resulting in an expression construct with an N-terminal secretion signal, followed by Glu-Thr-Gly, the target protein and a C-terminal Gly-Thr-Lys-His₆ tag. Proteins for crystallization were expressed by transient transfection in HEK-293T cells (ATCC CRL-11268) in the presence of the class I α-mannosidase inhibitor kifunensine for ∼3 to 5 d at 37 °C (48). Media with secreted proteins were centrifuged (7,000 × g, 1 h, 18 °C), filtered (0.22-μm polyethersulfone membrane; Millipore), and dialyzed against 200 mM NaCl, 20 mM Tris⋅HCl pH 8.0 using a QuixStand benchtop diafiltration system (GE Healthcare) (∼21 °C, ∼3 to 6 h). Proteins were purified using immobilized metal (cobalt) affinity chromatography (IMAC, TALON resin; Clontech). For crystallization, RGMs and NEO1 were deglycosylated with endoglycosidase F1 (∼10 μg per mg of target protein, ∼2 to 12 h, 21 °C) to cut the Asn-linked glycans down to one N-acetylglucosamine moiety, concentrated to ∼1 to 10 mg mL⁻¹ and further purified by SEC (typically in 150 mM NaCl, 10 mM Hepes pH 7.5) before complex formation. RGM_ND constructs and BMPR1A_ECD for cellular and binding assays were produced without kifunensine and kept fully glycosylated. RGM_ND constructs and BMPR1A_ECD used for cellular assays were dialyzed against 37 mM NaCl, 2.7 mM KCl, and 10 mM phosphate buffer (pH 7.4 at 25 °C) before use. Wild-type GDF5 and GDF5 variants Arg438Ala and Arg438Leu were produced in Escherichia coli, refolded, and purified as described previously (7). Soluble RGM extracellular domains for SPR-based binding studies between RGMs and GDF5 mutants were produced in FreeStyle 293 cells by transient expression. Expression levels of RGMs in HEK293 cells were analyzed using primary anti–His-tag antibody followed by a secondary antibody coupled to peroxidase and chemiluminescence (see SI Appendix for details).

RGM_ND–GDF5 and NEO1–RGMB_ECD–GDF5 complexes were crystallized using sitting-drop vapor diffusion with 100-nL protein solution plus 100-nL reservoir solution per droplet in 96-well Greiner plates at 21 °C (49). Purified RGMA_ND and GDF5 were mixed (1:1 mol:mol; ∼1.5 mg mL⁻¹) in 100 mM NaCl, 27 mM Hepes pH ∼7.5, 170 mM NDSB-256, incubated for ∼12 h at 4 °C and concentrated (Amicon Ultra-4 centrifugal filters, 3-kDa molecular mass cutoff) to 7.8 mg mL⁻¹ and crystallized in 0.2 M ammonium acetate, 0.1 M sodium citrate tribasic dihydrate pH 5.5, 24% (vol/vol) polyethylene glycol (PEG) 400. Crystals were cryoprotected in reservoir solution supplemented with 10% (vol/vol) PEG 400 before transferring into liquid nitrogen. For RGMB_ND–GDF5 (crystal form 1), RGMB_ND, ActR2b, and GDF5 were mixed (3.2:2.4:1 mol:mol:mol, ∼1.3 mg mL⁻¹) in 0.9 M NaCl, 20 mM Hepes pH 7.4, incubated for ∼12 h at 4 °C, followed by SEC in 0.5 M NaCl, 20 mM Hepes pH 7.4 (HiLoad 16/60 Superdex column; GE Healthcare, 21 °C). SEC fractions containing a ternary complex were concentrated to 5.3 mg mL⁻¹ and crystallized in 0.2 M Li₂SO₄, 0.1 M Hepes pH 7.5, 25% (vol/vol) PEG 3350. Crystals were cryoprotected in reservoir solution supplemented with 30% (vol/vol) glycerol. For RGMB_ND–GDF5 (crystal form 2), RGMB_ND was deglycosylated after IMAC for ∼2 h at 21 °C in 150 mM NaCl, 10 mM Hepes pH 7.5, then methylated following the described standard procedure in Walter et al. (50) and purified by SEC in 150 mM NaCl, 10 mM Tris⋅HCl pH 8.0. Purified RGMB_ND (methylated) and GDF5 were mixed (1:1 mol:mol; 1.6 mg mL⁻¹) in 130 mM NaCl, 15 mM Tris⋅HCl pH 8.0, 130 mM NDSB-256, incubated for ∼12 h at 4 °C, concentrated to 8.2 mg mL⁻¹ and crystallized in 0.2 M (NH₄)₂SO₄, 0.1 M sodium acetate pH 4.6, 35% (wt/vol) pentaerythritol ethoxylate (15/4 EO/OH; average molecular mass 797 Da). Crystals were cryoprotected in reservoir solution supplemented with 30% (vol/vol) glycerol. Purified RGMC_ND and GDF5 were mixed (1:1 mol:mol; ∼1 mg mL⁻¹) in 60 mM NaCl, 30 mM Hepes pH ∼7.5, 0.2 M NDSB-256, incubated for ∼12 h at 4 °C, concentrated to 6.9 mg mL⁻¹ and crystallized in 1 M LiCl, 0.1 M citric acid pH 4.0, 8% γ-butyrolactone. Crystals were cryoprotected in reservoir solution supplemented with 3.5 M LiCl₂. For NEO1–RGMB_ECD–GDF5 crystallization, the NEO1–RGMB_ECD complex was formed first. NEO1 and RGMB_ECD were mixed (1.2:1 mol:mol; ∼1.3 mg mL⁻¹ in 150 mM NaCl, 10 mM Hepes pH 7.5), incubated for ∼12 h at 4 °C and purified using SEC (150 mM NaCl, 10 mM Hepes pH 7.5; HiLoad 16/60 Superdex 200 column, GE Healthcare, 21 °C). Purified NEO1–RGMB_ECD complex was mixed with GDF5 (1:1.5, mol:mol) in 90 mM NaCl, 21 mM Hepes pH 7.5, 39 mM NDSB-256, incubated for ∼3 h at 4 °C, and purified using SEC (150 mM NaCl, 10 mM Hepes pH 7.5; HiLoad 16/60 Superdex 200 column, GE Healthcare, 21 °C). SEC fractions containing NEO1–RGMB_ECD–GDF5 were concentrated to 8.0 mg mL⁻¹ and crystallized in 0.1 M NaCl, 20 mM MES pH 6.7, 6.6% (wt/vol) PEG 4000. For data collection, crystals were cryoprotected at 100 K in reservoir solution supplemented with 25% (vol/vol) glycerol.

Data were collected at the Diamond Light Source (United Kingdom), indexed, and integrated using XDS (51), scaled using AIMLESS (52) in xia2 (53). RGM_ND–GDF5 structures were solved by molecular replacement using Phaser (54) and RGM_NDs from RGM_ND–BMP2 complexes [PDB ID codes 4UHY, 4UI0, and 4UI1 for RGMA, RGMB, and RGMC, respectively (17)] and GDF5 [1WAQ (7)] structures as search models. Diffraction data from RGMB_ND–GDF5 (crystal form 2) and RGMC_ND–GDF5 crystals were anisotropic. Thus, scaled and merged data were anisotropy-corrected with the STARANISO webserver (Global Phasing) prior to refinement. Estimated diffraction limits of ellipsoid fitted to diffraction cutoff surface for RGMB_ND–GDF5 (crystal form 2): 3.29 Å, 3.29 Å, 3.0 Å in 0.894a*-0.447b*, b*, and c* directions, respectively; for RGMC_ND–GDF5: 2.73 Å, 2.73 Å, 2.35 Å in 0.894a*-0.447b*, b*, and c* directions, respectively. All structures were refined using Coot (55) and Phenix (56). The ternary NEO1–RGMB_ECD–GDF5 complex was solved by molecular replacement using complexes of the RGMB_ND–GDF5 (crystal form 1) and NEO1–RGMB_CD [PDB ID code 4BQ6 (16)] as search models. RGMB_ND–GDF5 and NEO1–RGMB_CD complexes were refined as rigid bodies (9 in total) with one translation, libration, screw (TLS) group per protein domain (30 in total) in Phenix. Nine rigid bodies comprised three RGMB_ND–GDF5 complexes (rigid body 1: Chains A, B, C, D; 2: Chains G, H, I, J; 3: Chains M, N, O, P) and six NEO1–RGMB_CD complexes (rigid body 4: Chains c and E; 5: Chains d and F; 6: chains i and K; 7: Chains j and L; 8: Chains o and Q; 9: Chains p and R) (SI Appendix, Fig. S11). Refinement of the NEO1–RGMB_ECD–GDF5 complex using single chains as rigid bodies (21 rigid bodies in total; disulfide linked chains A/B, G/H and M/N of GDF5 were treated as 3 distinct rigid bodies) gave a model with minor clashes (e.g., between chains j and L that appear less ordered in the crystal) (SI Appendix, Fig. S11) as quantified by increase of MolProbity all-atom clash score from 3.1 to 7.3 (57). Thus, the ternary NEO1–RGMB_ECD–GDF5 complex was refined using nine rigid bodies described above. The quality of structures was assessed using Coot (55) and MolProbity (57). Figures of structures were prepared using PyMOL (The PyMOL Molecular Graphics System, v2.0, Schrödinger).

Initial SPR experiments were performed using a ProteOn XPR36 instrument (Bio-Rad). All measurements were performed at 25 °C and a flowrate of 100 to 200 µL min⁻¹ using 10 mM Hepes pH 7.4, 500 mM NaCl, 3.4 mM EDTA, 0.005% (vol/vol) Tween20 (HBST₅₀₀). BMP/TGF-β ligands were immobilized onto GLC sensor chips (Bio-Rad) by primary amino coupling after EDAC/sulfo-NHS activation following the manufacturer’s recommendation. For SPR data acquisition, recombinant full-length mature RGM proteins lacking the C-terminal residue, which is modified to carry the membrane-anchoring glycolipid moiety, were perfused as analyte over the biosensor usually employing six different concentrations. Association was observed for 120 to 240 s, dissociation was initiated by perfusing HBST₅₀₀ buffer and monitored for 40 to 200 s. Biosensor regeneration was performed by a 60-s injection of 100 mM glycine pH 1.5 and then 4 M MgCl₂ (flow rate during regeneration 30 µL min⁻¹). The association and dissociation phase of the sensorgrams were fitted with a Langmuir-type 1:1 interaction model using the software ProteOn Manager v3.1 (Bio-Rad). Data were only interpreted if χ² values were ≤10% of the maximal signal amplitude (RU_max). Equilibrium binding constants K_d were calculated from the equation K_d = k_off/k_on. All affinities are termed “apparent” K_ds, indicating that the absolute values might differ from values obtained by other interaction analysis methods. Differences in affinities and rate constants determined by this set-up greater than twofold were considered significant.

Sensorgrams of RGMB–BMP2/GDF5 interactions derived with the initial set-up were difficult to analyze due to biphasic binding hinting towards nonspecific binding and usually required all fitting parameters to be treated local instead employing global fitting routines. Hence, additional SPR experiments were performed using a different set-up and employing a Biacore T200 instrument (GE Healthcare) at 25 °C. GDF5 and BMP2 were covalently linked to the surface of a CM5 chip (Biacore) via primary amine coupling. A relatively high ionic strength running buffer with detergent (0.5 M NaCl, 20 mM Hepes pH 7.4, 0.05% Tween 20; flow rate 15 µL min⁻¹) was used to minimize any nonspecific protein–protein and protein–chip surface interactions. Site-directed mutagenesis of RGMs was performed by a two-step overlap extension PCR and constructs were verified by DNA sequencing. All RGM constructs were purified by SEC in SPR running buffer before use, and a 1:2 dilution series was prepared. After each measurement, the chip surface was regenerated with 4 M MgCl₂ (100 μL min⁻¹, 2 min). The signal from experimental flow cells was corrected by subtraction of a buffer and reference signal from a flow cell without coupled protein. Equilibrium dissociation constants (K_ds) were calculated using BIAevaluation 3.0 program (Biacore) and a 1:1 Langmuir binding isotherm model: y = (B_max × x)/(K_d + x); where y is binding response, x is the analyte concentration, and B_max is the maximum analyte binding. Analyte concentrations were determined from the absorbance at 280 nm using a NanoDrop ND-1000 spectrophotometer (Thermo Fisher Scientific) and the calculated molar extinction coefficients from ProtParam webserver.

LLC-PK1 pig kidney cells (ATCC CL-101) were grown in DMEM high-glucose media (D6546, Sigma) supplemented with 2 mM l-glutamine (Gibco), minimum essential medium nonessential amino acids (Gibco), and 10% FBS (10270, Gibco) at 37 °C, 5% CO₂. For assays, cells were plated in complete DMEM, 10% FBS at a density of 50,000 cells per well (100 μL per well) in a 96-well plate (Nunc-Immuno MicroWell polystyrene plates with Nunclon delta surface; P8616, Sigma). After 24 h, cells were transfected with Lipofectamine 2000 transfection reagent (Thermo Fisher Scientific) according to the manufacturer’s protocol (0.5 μL per well), with 40 ng pGL3 BRE firefly luciferase plasmid (58), 30 ng pRL-TK Renilla luciferase control plasmid (Promega), and 20 ng of empty pHLsec vector control or RGM constructs (or 10 ng of NEO1 plus 10 ng of RGMB constructs) (Fig. 5) in complete DMEM, 0.1% FBS. Twenty-four hours after transfection, media was changed to complete DMEM, 0.1% FBS supplemented with purified BMP2 (3-nM final concentration) or GDF5 (30 nM). After 48-h incubation with BMP2 or GDF5, cells were lysed and the activity of two different luciferases was measured using the Dual-Glo luciferase assay system (Promega). Luminescence was measured using a Luminoskan Ascent luminometer (Labsystems). Fab fragments of anti-BMPR1A antibodies AbD1556 and AbD1564 were produced as described previously (59). Additional BMP2 and GDF5 cellular assays were performed in C3H10T1/2 cells stably transfected with a BMP-responsive luciferase reporter, and ATDC5 cells overexpressing alkaline phosphatase upon stimulation with BMP2/GDF5 (see SI Appendix for details).

A SEC-MALS experiment was performed using a Wyatt Dawn HELEOS-II 8-angle light-scattering detector (with 663.8-nm laser) and a Wyatt Optilab rEX refractive index monitor linked to a Shimadzu HPLC system comprising LC-20AD pump, SIL-20A autosampler and SPD20A UV/Vis detector. SEC-MALS of the RGMB_ND–GDF5–ActR2b_ECD complex (1 mg mL⁻¹, 100 μL per injection) was performed using Superdex 200 HR 10/30 column equilibrated in 0.5 M NaCl, 20 mM Hepes pH 7.4, 0.5 mL min⁻¹ flow rate, 21 °C. GDF5 variant Tyr487Lys/Gln489Asp with increased affinity for the type 2 receptor was used to reconstitute the RGMB_ND–GDF5–ActR2b_ECD complex. SEC-MALS of the RGMB_ND–GDF5 complex (1 mg mL⁻¹, 100 μL per injection) and RGMB_ND (1 mg mL⁻¹, 100 μL per injection) were performed using a Superdex 200 HR 10/30 column equilibrated either in 0.5 M NaCl, 20 mM Hepes pH 7.5, or 0.5 M NaCl, 20 mM MES pH 5.5 at 0.5 mL min⁻¹ flow rate, 21 °C. Scattering data were analyzed and molecular weight was calculated using ASTRA 6 software (Wyatt Technology).