Structural basis of myelin-associated glycoprotein adhesion and signalling

Myelin-associated glycoprotein (MAG) is a myelin-expressed cell-adhesion and bi-directional signalling molecule. MAG maintains the myelin–axon spacing by interacting with specific neuronal glycolipids (gangliosides), inhibits axon regeneration and controls myelin formation. The mechanisms underlying MAG adhesion and signalling are unresolved. We present crystal structures of the MAG full ectodomain, which reveal an extended conformation of five Ig domains and a homodimeric arrangement involving membrane-proximal domains Ig4 and Ig5. MAG-oligosaccharide complex structures and biophysical assays show how MAG engages axonal gangliosides at domain Ig1. Two post-translational modifications were identified—N-linked glycosylation at the dimerization interface and tryptophan C-mannosylation proximal to the ganglioside binding site—that appear to have regulatory functions. Structure-guided mutations and neurite outgrowth assays demonstrate MAG dimerization and carbohydrate recognition are essential for its regeneration-inhibiting properties. The combination of trans ganglioside binding and cis homodimerization explains how MAG maintains the myelin–axon spacing and provides a mechanism for MAG-mediated bi-directional signalling.

M yelination of axons enables enhanced conductance velocity in both the central and peripheral nervous system (CNS and PNS) of vertebrates. It also provides electrical insulation and a decrease of the capacitance, as well as physical protection and metabolic support of long axons 1 . Myelin-associated glycoprotein (MAG) adhesion and signalling at the myelin-axon interface regulates the formation and maintenance of myelinated axons, thus playing an important role in the development and function of the nervous system 2,3 . Aberrant MAG function, for example from mutations that likely cause misfolding, or anti-MAG autoimmunity, has been associated with demyelination and neurodegenerative disorders, such as corticospinal motor neuron disease also known as hereditary spastic paraplegias 4 , Pelizaeus-Merzbacher disease-like disorder 5 , demyelinating anti-MAG peripheral neuropathy 6,7 and multiple sclerosis 2,8 .
MAG is a type 1 single-pass transmembrane protein expressed on myelinating oligodendrocytes in the CNS and Schwann cells in the PNS 2,3 . MAG is the fifth highest expressed protein in myelin of the CNS 9 . It is highly enriched at the innermost (adaxonal) myelin membrane along the internode, where it contacts the axon. MAG is also found on other myelin structures, such as the mesaxon, Schmidt-Lanterman incisures and paranodal loops 2,3 . MAG adhesion maintains the myelin-axon spacing (periaxonal diameter) by interacting with specific neuronal gangliosides (glycolipids), such as the major brain gangliosides GT1b and GD1a (refs 10-13). More recently, the Nectin-like (Necl) proteins 1 and 4 have also been found to contribute to myelin-axon adhesion along the internode 14,15 , although they are expressed less than MAG in mature myelin 9 and knockout of Necl4 does not affect myelination 16 .
MAG, also known as Siglec4a, is evolutionarily the oldest member of the Siglec family 17 . Unlike all other Siglecs, MAG plays no role in the immune system and is exclusively expressed in the nervous system 17 . On the basis of the primary sequence its extracellular region is predicted to consist of five Ig domains; an N-terminal V-type Ig domain that is typical for Siglecs and four C2-type Ig domains. This is followed by a single membranespanning helix and an intracellular region predicted to be unstructured and of different length for two MAG isoforms, L-MAG and S-MAG. Like other Siglecs, MAG recognizes sialic acid groups and the specificity of MAG has been established to be Neu5Ac-a2,3-Gal-b1, . This trisaccharide is part of several neuronal gangliosides, most notably the major brain gangliosides GT1b and GD1a, but also GM1b, GT1b and GQ1ba. MAG bridges the periaxonal space by interacting with these axonal gangliosides in trans via the canonical Siglec site at a conserved arginine (R118 in MAG) in the N-terminal domain 19,20 .
MAG signalling is bidirectional, engaging in both axon-tomyelin as well as myelin-to-axon signalling. MAG has been extensively studied as one of three classic myelin-associated inhibitors of central nervous system regeneration, the other ligands being Nogo66 and oligodendrocyte myelin glycoprotein 2,3 . MAG inhibits neurite outgrowth and collapses axonal growth cones in a sialic acid binding-dependent manner. It does so as full-length transmembrane 20,21 , but also as a proteolytically shed and soluble form called dMAG 22 . As a receptor, MAG controls myelin formation and integrity. How MAG transduces the extracellular signal into the myelinating cell is not well understood, but it has been shown that the cytosolic domain of the L-MAG isoform binds to the cytoplasmic non-receptor tyrosine kinase Fyn 23 and that antibody-induced crosslinking of L-MAG triggers its localization to lipid rafts 24 and activates Fyn in oligodendrocytes 23 . This activation of Fyn is essential for the initiation of myelination 25 . In contrast, the shorter MAG isoform S-MAG binds to zinc and microtubules and this is postulated to have a structural function in mature myelin 26,27 .
From earlier rotary-shadowed electron microscopy (EM) and sedimentation velocity analytical ultracentrifugation (AUC) studies it was hypothesized that the extracellular segment of MAG has a back-folded Ig-horseshoe type structure, but the estimated maximum dimensions of 8.8 and 18.5 nm determined by AUC and EM, respectively, deviate substantially 28,29 . In the absence of any high-resolution structural data on MAG or its interaction with ganglioside ligands, the conformation of the five Ig domains, the extracellular specificity-determining parameters and the mechanisms underlying MAG adhesion and bi-directional signalling are unresolved. Using a combination of structural, biophysical and cellular techniques, we provide the structural basis of MAG-mediated adhesion and identify a dimerization-dependent mechanism that explains how MAG regulates axon-to-myelin and myelin-to-axon signalling, and controls myelin-axon spacing.

Results
MAG has an extended conformation. We determined crystal structures of the full extracellular segment of mouse MAG (MAG [1][2][3][4][5] in two different crystal forms that diffracted to a maximum resolution of 3.8 and 4.3 Å. These crystals were obtained by enzymatic deglycosylation of MAG 1-5 or reductive lysine methylation of glycosylated MAG 1-5 (see 'Methods' section). In addition, crystals of a shorter construct, consisting of the three N-terminal domains (MAG 1-3 ), diffracted to a maximum resolution of 2.1 Å. The structures were solved by molecular replacement with individual Ig domains from homologous proteins. The exceptionally high-solvent content of the two MAG 1-5 crystal forms (91 and 85%, Supplementary  Fig. 1) aided in obtaining phases of sufficient quality for initial model building (see also Table 1 and 'Methods' section for details).
In all three crystal forms MAG has an extended collinear conformation (Fig. 1). Only consecutive Ig domains interact with each other via hydrophobic interfaces (buried-surface area ranging from 243 to 690 Å 2 ) and short inter-domain linkers of up to two residues (Fig. 1a). Domains Ig1 and Ig2 form the largest interface in which the Ig2 loops at the N-terminal 'head' side interact with the A2-B (Ig domain b-strand numbering) side of Ig1 (Fig. 1a). The three other inter-domain interfaces are exclusively formed in a head-to-tail manner involving loops at the N-terminal 'head' and C-terminal 'tail' side of the Ig domains (Fig. 1a). As predicted from the primary sequence the N-terminal Ig1 domain of MAG has a V-type Ig fold like other Siglec family members [30][31][32] and domains Ig3 and Ig4 are of the C2 type. Domains Ig2 and Ig5, however, have a C1-type Ig fold, contrary to the predicted C2-fold ( Supplementary Fig. 2). The three crystal structures of MAG are similar to each other with only small differences within the domains (r.m.s.d. ranging from 0.93 to 2.13 Å) and inter-domain angle rotation differences ranging from 3.4 to 17.4°, the largest difference is in the domain Ig2-Ig3 angle (Fig. 1b). The combination of hydrophobic inter-domain interfaces, the lack of flexible linker residues and a previously predicted inter-domain disulfide between Ig1 and Ig2 (C37-C165) 33 explains the limited inter-domain flexibility observed between the three different crystal forms.
MAG is post-translationally modified. The structures reveal MAG is post-translationally modified at several sites. MAG contains seven disulfides, five of which are canonical for Ig domains. Cysteines 37 and 165 form an inter-domain disulfide between Ig1 and Ig2, and cysteines 421 and 430 form an additional intra-domain disulfide in Ig5, as shown previously 33 ( Supplementary Fig. 3). In addition, MAG carries N-linked glycans and previously eight N-linked glycosylation sites were determined in human MAG by mass spectrometry analysis 34 . We observe glycan electron density for five of those equivalent sites in the mouse MAG structures (on asparagine 99, 223, 246, 315 and 406; Supplementary Fig. 4). The differences in these observations arise from one N-linked glycosylation site that is not conserved (N106 in human MAG is a threonine in mouse) and from poorly resolved electron density for the other two sites (on N450 and N454) that are situated in a flexible loop. One additional N-linked glycan is revealed by clear electron density on N332, in contrast to the previous study that did not find this residue in a glycopeptide analysis 34 . Electron density at the N332-linked glycan suggest it is fucosylated in our recombinantly produced MAG ( Supplementary Fig. 5). In native mass spectrometry experiments of intact recombinant MAG we also observed internal mass shifts of þ 147 Da on MAG monomers ( Supplementary Fig. 6), further indicating fucosylation as a post-translational modification. Thus the combination of our data and that of others 34 indicates that mouse MAG has eight N-linked glycosylation sites (on N99, N223, N246, N315, N323, N406, N450 and N454).
In addition, MAG carries a tryptophan C-mannosylation on W22. In all three crystal forms, electron density proximal to the side chain of W22 suggests this residue is C-mannosylated ( Supplementary Fig. 6). Indeed, analysis of the MAG primary sequence reveals that this tryptophan is part of the canonical WxxW motif (W22 is the first tryptophan) for C-mannosylation, a rare post-translational modification present on several secreted proteins 35 . The W22-attached a-D-mannopyranosyl group has an unusual ring-flipped 1 C 4 chair conformation in the structures (regular mannose is 4 C 1 , Supplementary Fig. 6). This ring-flipped conformation is in agreement with previous nuclear magnetic resonance studies on mannosyl-tryptophan 36 and can be explained by the preference of the bulky tryptophan, covalently attached to the C1 of the mannose, to be in the equatorial position. We confirmed the C-mannosylation of W22 by in-gel trypsin digestion of MAG 1-5 followed by liquid chromatographymass spectroscopy (LC-MS/MS) analysis of the (glyco)peptide fragments ( Supplementary Fig. 7). In addition, native mass spectrometry of wild type and mutated MAG 1-5 , in which the second tryptophan of the C-mannosylation motif is substituted for glutamine (W25Q, resulting in WxxQ), showed a mass shift of À 221 Da in accordance with a loss of C-mannosylation and confirming the importance of the WxxW motif ( Supplementary  Fig. 6).
MAG 1-5 crystal structures reveal a dimeric arrangement. MAG 1-5 forms a symmetry-related dimer at a crystallographic two-fold rotation axis in both MAG 1-5 crystal forms. They share the same interface (Fig. 2a) on domains Ig4 and Ig5, which buries a surface area of 2,037 Å 2 . The CC 0 FG face of Ig4 binds to the ABDE face of Ig5 of the symmetry-related molecule and vice versa, thus forming two equivalent hemi-interfaces. The interface is mostly hydrophobic apart from the negatively charged E395 in Ig4, with hydrophilic residues lining the edges of the interface (Fig. 2b).
To validate the interface we generated two interface mutations based on the structures, that we predicted to either disrupt dimerization (I473 to E) or enhance it (N406 to Q). The hydrophobic I473 in the middle of the hydrophobic interface of Ig5 was mutated to a negatively charged glutamate, to ensure disturbance of the hydrophobic effect as well as introducing electrostatic repulsion with the opposing E395 in Ig4 (Fig. 2c). N406 carries an N-linked glycan and in the glycosylated lysine-methylated crystals of MAG 1-5 , the density for this glycan suggests that it sterically clashes with its symmetry partner in the dimer (Fig. 2d). We hypothesized that a dimer with increased affinity would form, were this glycan absent. Remarkably, the N406Q mutant that lacks this glycan was the only MAG construct that showed a distinct dimer peak in size exclusion chromatography (SEC; Supplementary Fig. 8). The other glycans are not expected to interfere with dimerization, also not when considering myelin-specific N-linked glycans 37 . We confirmed that MAG dimerizes in solution via the Ig4-Ig5 interface by analysing glycosylated and deglycosylated forms of MAG 1-5 , MAG 1-3 (that lacks the dimerization domains) and the aforementioned mutants in small angle X-ray scattering (SAXS) and AUC experiments.
Furthermore, the MAG 1-5 I473E pair distance distribution function P(r) and derived D max , the ab initio models as well as the Kratky plots confirm that MAG has an extended conformation and behaves as a semi-rigid rod in solution ( Fig. 2f,g,i). Whereas the SAXS data for glycosylated MAG 1-3 and MAG 1-5 I473E fit best to scattering curves calculated from single chains of the crystal structures of MAG 1-3 and MAG 1-5 , respectively (w ¼ 2.87 and 2.95, Supplementary Fig. 10), the glycosylated MAG 1-5 N406Q SAXS data fits best to scattering curves calculated from the dimer structure (w ¼ 3.45, Supplementary Fig. 10). Both glycosylated and deglycosylated MAG 1-5 wt SAXS data fit best to scattering curves calculated from a combination of monomeric and dimeric crystal structures (w ¼ 4.05 for glycosylated and 3.72 for deglycosylated MAG 1-5 , Fig. 2j and Supplementary Fig. 11). Furthermore, ab initio models based on the SAXS data of deglycosylated MAG 1-5 I473E and MAG 1-5 N406Q agree remarkably well with the crystal structures of the monomer and dimer of MAG 1-5 , respectively (Fig. 2g, Supplementary Fig. 12, w 2 of the model-to-data fit are 1.05 and 1.33 for MAG 1-5 I473E and MAG 1-5 N406Q, respectively). These data confirm MAG dimerizes via domains Ig4 and Ig5 and that the MAG 1-5 chains have an extended and relatively rigid conformation.
The MAG 1-5 dimer is weak in solution with a K d of 3.8 Â 10 2 lM. To quantify the affinity of MAG dimerization in solution, we performed sedimentation equilibrium AUC (SE-AUC) experiments for wt MAG 1-5 , deglycosylated wt MAG 1-5 , MAG 1-3 , MAG 1-5 I473E and MAG 1-5 N406Q constructs. For each sample a global analysis was performed for different concentrations, centrifugal speeds and wavelengths together (Supplementary Figs 13-16), except for MAG 1-5 N406, which appeared to suffer from aggregation during the experiment. Similar to the SAXS analysis, the MAG 1-5 I473E and MAG 1-3 SE-AUC data fit best to a single species that agrees with the M m of a monomer (Table 3). For wt MAG 1-5 and deglycosylated MAG 1-5 , we could fit the data to a monomer-dimer equilibrium, with K d s of 3.8 Â 10 2 and 1.7 Â 10 2 mM, respectively (Table 3). On the basis of a monomerdimer equilibrium with these K d s, dimer fractions of 18% for glycosylated wt MAG 1-5 and 24% for deglycosylated wt MAG [1][2][3][4][5] are expected to be present in the SAXS experiments (calculated at 52.2 and 37.5 mM for glycosylated and deglycosylated MAG, respectively). Indeed, the presence of a mix of monomers and dimers is observed in the SAXS data of both glycosylated and deglycosylated MAG 1-5 (Fig. 2j, Supplementary Fig. 11). The lack of dimers in the AUC experiments for MAG 1-5 I473E and MAG 1-3 and the higher affinity for deglycosylated MAG 1-5 compared with glycosylated MAG 1-5 further confirm that MAG forms dimers via domains Ig4 and Ig5 (Table 3).
Structural basis of ligand recognition by Ig1. MAG binds to sialic acids of gangliosides with its N-terminal V-type Ig domain and has a preference for a Neu5Ac-a2,3-Gal-b1,3-GalNAc moiety 18 . We observed unmodelled electron density in the 2F o À F c and F o À F c maps of the deglycosylated MAG 1-5 structure close to R118 in the putative ligand-binding site 20 (Fig. 3a). Native mass spectrometry of purified MAG 1-5 revealed a mixture of free and two ligand-bound MAG 1-5 forms with mass differences of 835±2 and 854±2 Da compared with free MAG 1-5 (Fig. 3b). The 835 Da ligand possibly corresponds to a tetrasaccharide comprising the aforementioned Neu5Ac-a2,3-Gal-b1,3-GalNAc trisaccharide plus another hexose. The 854 Da ligand might correspond to a similar tetrasaccharide where Neu5Ac is replaced by Neu5Gc, a mammalian sialic acid variant not produced by humans. These ligands are likely co-purified in complex with MAG from the expression medium that contains beef digest (Primatone). We do not observe clear electron density for any ligands in the two other crystal forms (lysine-methylated MAG 1-5 and MAG 1-3 ). In these crystal forms the unliganded MAG is apparently preferentially crystallized over the ligandbound forms. The resolution of the MAG 1-5 -ligand complex (to 3.8 Å) is not sufficient to determine the detailed structure of the bound oligosaccharide. However, the unmodelled electron density is compatible with either of the suggested compounds.
To obtain more detailed information on MAG-ligand interactions, the commercially available trisaccharide 3 0 -N-Acetylneuraminyl-N-acetyllactosamine (Neu5Ac-a2,3-Gal-b1, 4-GlcNAc) was soaked into the MAG 1-3 crystals, as these provided higher resolution data. For the soaked crystals, diffraction data was collected to 2.3 Å resolution. The crystals were isomorphous to the unsoaked MAG 1-3 crystals (Table 1) and a F o (soaked) À F o (unsoaked) map revealed clear electron density in the oligosaccharide binding site for one of the two MAG 1-3 chains in the asymmetric unit (Fig. 3c). Most likely the other chain remained unliganded due to occlusion of the binding site by crystal packing.
The ganglioside binding side is formed by the CC 0 loop and the F and G b-strands of the N-terminal V-type Ig1 domain. In agreement with previous data 20 , the side chain of R118 in strand F forms a key salt bridge with the carboxylic acid group of the sialic acid in the trisaccharide ligand. Y65 in the CC 0 loop forms extensive Van der Waals' contacts with the ligand as well as a hydrogen bond to the O9 of the Neu5Ac sialic acid. Other contributing interactions are made by: (1) the backbone carbonyls of N126 and T128 in strand G that form hydrogen bonds with H-N5 and the H-O9 of the sialic acid respectively, (2) the backbone amide proton of T128 that forms a hydrogen bond with the O8 of the sialic acid and (3) the side chain of Y127 that forms Van der Waals' contacts with the glycerol group (C7-C9) of the sialic acid (Fig. 3d).
We validated the ganglioside binding properties of MAG by mutating ligand-binding residues to alanine in MAG 1-5 and probed their interaction with GT1b ganglioside incorporated into liposomes. The set-up we used, MAG 1-5 coupled at the C-terminus to a streptavidin-coated surface plasmon resonance (SPR) chip and GT1b-containing liposomes in the mobile phase, enables avidityenhanced interactions that also occur in trans between cells (Fig. 3f, see 'Methods' section). Indeed we observed specific binding of GT1b liposomes to wt MAG [1][2][3][4][5] , no interactions with the ligand-binding mutants MAG 1-5 R118A, T128A and Y127A and reduced interaction with MAG 1-5 Y65A (Fig. 3f). In addition, the MAG 1-5 W25Q mutant that lacks the tryptophan mannosylation on W22, still interacted with GT1b liposomes in this assay. Remarkably, this W25Q mutant appeared to have higher affinity for the GT1b liposomes compared with wt MAG. This suggests that rather than contributing to the interaction strength, this tryptophan mannosylation on W22 may play a regulatory role in binding (membraneembedded) gangliosides. In summary, we have shown that MAG interacts with membrane-bound gangliosides via the side chain of R118, the CC 0 -loop and the F and G b-strands of the N-terminal V-type Ig1 domain and that the W22 mannosylation does not enhance ganglioside binding.

Neurite outgrowth inhibition depends on MAG dimerization.
We tested different MAG variants in neurite outgrowth assays to determine the role of MAG dimerization for neuronal plasticity inhibition (Fig. 4). In agreement with previous data 20, 38 , MAG 1-5 wt on coverslips inhibited neurite outgrowth of hippocampal neurons compared with poly-L-lysine (PLL)-covered slips (Fig. 4a,b). Other dimeric variants (MAG 1-5 N406Q and MAG 1-5 -Fc) inhibited neurite outgrowth to a similar level (Fig. 4d,f). The monomeric MAG 1-5 I473E and MAG 1-3 wt on the other hand showed no significant neurite outgrowth inhibition (Fig. 4c,e). Interestingly, MAG 1-5 -Fc R118A, which is dimeric but lacks sialic acid binding properties, showed neurite outgrowth stimulation instead of inhibition, compared with PLL (Fig. 4g). These data indicate that dimerization through domains Ig4-Ig5 and the ability to bind sialic acid moieties on the neuronal surface are required for neurite outgrowth inhibition signalling by MAG for hippocampal neurons.

Discussion
MAG controls adhesion and signalling between myelinating cells and axons. In contrast to earlier studies 28,29 , we find that MAG does not fold back onto itself like an Ig-horseshoe as in the L1CAM and axonin neuronal adhesion molecules. Instead, our data show that the extracellular region of MAG has an extended shape with limited inter-domain flexibility, similar to several other cell adhesion molecules such as SYG, Cadherin and Nectin family members [39][40][41] . The structure of MAG is the first of a full extracellular portion of a Siglec family member. Besides the common N-terminal V-type Ig domain for recognizing sialic acid moieties, Siglecs vary in the number of additional Ig domains; from 1 up to 16. Comparison of the structures of MAG and Siglec5 reveals a different inter-domain orientation between domains Ig1 and Ig2, likely due to differences in amino acids at the interface (Supplementary Fig. 17 and Supplementary notes).
By binding to axonal gangliosides, MAG maintains a defined spacing between the innermost myelin surface and the axon surface [10][11][12]42 . This myelin-axon spacing has been reported to be 12-14 nm based on electron micrographs of chemically fixed myelin tissue [10][11][12] . However, analysis of more recent electron micrographs of high-pressure frozen myelin that does not suffer from fixation induced artefacts 43 reveals an axon-myelin spacing of 9-12 nm. This periaxonal diameter matches well with a straightforward model that follows from our structural data of the MAG dimer and the MAG-ganglioside interaction; two opposing membrane surfaces are spaced 10 nm apart when the membraneproximal C-termini of the MAG dimer are positioned on one membrane (the structures lack only two residues to the transmembrane helix) and the MAG dimer-bound gangliosides are positioned on the other membrane (Fig. 5). Although the two crystal forms of the full extracellular segment of MAG have different inter-domain angles (3.4-17.4°), the overall arrangement and resulting structure-based axon-myelin spacing is similar (see Fig. 2a). The agreement of intermembrane distance determined from high-pressure frozen EM on myelin tissue 43 and here by structural and biophysical data on the extracellular segment of MAG indicates that in the periaxonal space, MAG is dimerized in cis via domains Ig4-Ig5 when bound to axonal gangliosides in trans.
Intriguingly, in this model, the unusual tryptophan C-mannosylation on W22 is positioned at the interface of MAG and the extracellular leaflet of the axonal membrane (Fig. 5, Supplementary Fig. 18). The WxxW motif is conserved among MAG orthologues in vertebrates from fish to human, but is absent in all other Siglec paralogs ( Supplementary Fig. 6). This suggests that tryptophan mannosylation is specific for the function of MAG. We showed that the mannosyl group does not enhance the binding of MAG to GT1b ganglioside liposomes but may weaken it. Possibly, tryptophan mannosylation of MAG provides specificity to sialylated ligand interactions. Alternatively, the close proximity of the mannosyl group to the axonal membrane during MAG-ganglioside interaction may indicate a regulatory role in axonal membrane engagement.
The buried-surface area of the MAG dimerization interface formed by Ig4 and Ig5 is large (2,037 Å 2 ) and hydrophobic. We find, however, that the affinity of MAG dimerization via this interface in solution is low (K d of 3.8 Â 10 2 mM). The weak interaction is probably important in the native context, where MAG is expressed on the cell surface, as the cis dimer can be trans stabilized by interaction with gangliosides on the opposing axonal membrane. For the N-Cadherin family of cell adhesion molecules it has been shown that affinities as weak as 10 mM in solution are functionally important in the context of a trans-stabilized cis-dimer 39 .
Dimerization of MAG may serve two purposes. It provides a mechanism to restrain the intermembrane distance, since a cis-trans stabilized MAG dimer would restrict angular freedom with respect to the membrane more than a trans only stabilized MAG monomer. In this sense, the MAG dimer could function as a 'molecular leaf spring' that maintains the well-defined spacing between the axonal membrane and the adaxonal myelin membrane along the internode. In addition, dimerization of MAG could enable compaction of the periaxonal space. The weak cis-interaction of MAG, if not stabilized in trans, may ensure enough monomer is available to bridge a wider periaxonal spacing (of up to 16 nm, based on the length of a MAG monomer) that may exist during myelin formation. Even greater distances could be bridged if MAG binds to sialylated N-or O-linked glycans of axonal surface glycoproteins before reaching  the gangliosides. When trans interactions with gangliosides have been established and possibly local concentrations are elevated due to the abundance of ganglioside ligands, formation of MAG cis-dimers is triggered concomitant with compaction of the periaxonal diameter to its final spacing (Fig. 5). Dimerization of MAG can be regulated by modulating glycosylation on N406. Large and charged glycans on N406 can obstruct dimerization because of steric clashes and coulombic repulsion. Trimming all MAG N-linked glycans down to single N-Acetylglucosamines by deglycosylation with Endo-H f or preventing glycosylation on N406 by mutating it to glutamine enhances dimerization (Fig. 2, Supplementary Figs 8-12 and 16, Tables 2 and 3). This glycosylation site might play a regulatory role during myelin development and myelination-related pathologies. MAG glycosylation changes during development 44 and abnormal glycosylation of MAG correlates with myelination deficiencies [45][46][47] . Possibly, modulation of N406 glycosylation, either at the biosynthesis level or by extracellular trimming, affects MAG dimerization and thereby impacts on the myelin-axon interaction (see Supplementary notes for details).
Myelin-to-neuron signalling with MAG as a ligand can inhibit neurite outgrowth 2,3,20,21,48 . Studies based on MAG truncations, chimeras and mutant versions suggested that the inhibitory properties reside in domains Ig4 and Ig5, and in the sialic acid binding site 20,49,50 . For example, a chimeric protein consisting of Siglec1 Ig domains 1-3 plus MAG Ig4-5, but not Siglec1 domains 1-3 alone, has neurite outgrowth inhibition properties similar to MAG 49 . We confirmed that the sialic acid-binding properties of MAG are required for neurite outgrowth inhibition in hippocampal neurons as the MAG 1-5 -Fc R118A mutant that lacks ganglioside binding properties does not inhibit, but surprisingly, stimulates neurite outgrowth (Fig. 4g). This has not been reported before and may be an interesting new avenue for therapeutic intervention to enhance central nervous system regeneration. In addition we show that domains Ig4 and Ig5 are essential for MAG dimerization and that neurite outgrowth inhibition is abrogated by the I473E point mutation that monomerizes MAG (Fig. 4e,c). Thus, it is the dimerization of MAG that is required for neurite outgrowth inhibition, rather than direct interactions of domains Ig4 and Ig5 with neuronal receptors, as previously suggested 49,50 . Although other protein receptors have been identified that mediate the neurite outgrowth-inhibiting signalling by MAG 3 , it has been shown that direct clustering of gangliosides by antibodies also leads to inhibition of neurite outgrowth of hippocampal neurons 20 . The combination of our data and that of others indicates that MAG dimerization at domains Ig4 and Ig5 and sialic acid binding at domain Ig1 induces neurite outgrowth inhibition for hippocampal neurons, by clustering of gangliosides.
Axon-to-myelin signalling with MAG as a receptor controls myelin formation. Antibody-mediated extracellular clustering of the L-MAG isoform activates Fyn kinase 23 and Fyn activation is essential for the initiation of myelination 25 . Our structures show that the C-termini of MAG 1-5 are separated by 5.4 nm in the dimer, bringing the cytosolic regions into close proximity (Fig. 5). Probably, L-MAG dimerization as a result of trans interaction   Fig. 1) and the tryptophan mannosylation on W22 (yellow spheres). Whereas monomeric MAG could span intermembrane distances of B16 nm (left), the cis dimerization of MAG restricts the periaxonal diameter to 10 nm (right). The arrow at the ganglioside binding site in the left panel indicates that the third sugar, which has the highest B-factors in the deglycosylated MAG 1-5 structure and is not visible in the ligand-bound MAG 1-3 structure, needs to be in a different conformation to accommodate such binding. The dimerization of MAG brings the cytosolic regions into close proximity. This possibly triggers signalling into the myelin-forming oligodendrocyte cell. See also Supplementary Fig. 18. NATURE COMMUNICATIONS | DOI: 10.1038/ncomms13584 ARTICLE with gangliosides on the axon brings the cytosolic regions of MAG into close proximity to trigger activation of Fyn, similar to Fyn activation by signalling lymphocytic activation molecule clustering in immune cells 51 . Whether MAG forms higher-order clusters that are triggered by dimerization needs to be established, but preference of both MAG and Fyn for lipid rafts 24,52 suggests that both proteins can be locally enriched in the membrane to assist clustering.
Large-scale expression and purification. Constructs were transiently expressed in N-acetylglucoaminyltransferase I-deficient (GnTI À ) Epstein-Barr virus nuclear antigen I (EBNA1)-expressing HEK293 cells in suspension (U-protein express). Medium was collected 6 days after transfection and cells were spun down by 10 min of centrifugation at 1,000 g. Supernatant was concentrated fivefold and diafiltered against 500 mM NaCl, 25 mM HEPES pH 7.8 (IMAC A) using a Quixstand benchtop system (GE Healthcare) with a 10 kDa molecular weight cut-off (MWCO) membrane. Cellular debris was spun down for 10 min at 9,500 g and the concentrate was filtered with a glass fibre prefilter (Minisart, Sartorius). Protein was purified by Nickel-nitrilotriacetic acid (Ni-NTA) affinity chromatography followed by SEC on a Superdex200 Hiload 16/60 column (GE Healthcare) equilibrated in SEC buffer (150 mM NaCl, 20 mM HEPES pH 7.5). Protein was concentrated to 7-14 mg ml À 1 using a 30 kDa MWCO concentrator before plunge freezing in liquid nitrogen and storage at À 80°C.
Structure solution and refinement. Data were integrated by IMOSFLM 54 (MAG 1-5 deglycosylated, MAG 1-5 lysine-methylated and MAG 1-3 ligand bound) or XDS 55 (MAG 1-3 unliganded) and scaled and merged by the AIMLESS pipeline 56 . All structures were solved by molecular replacement using PHASER 57 [1][2][3] was solved by searching for two copies of 1URL, followed by two copies of 1CS6 and finally two copies of 4FRW. Search models were trimmed to polyalanine chains by CHAINSAW (ref. 62) to obtain better starting density. Next, deglycosylated MAG 1-5 was solved by searching for the first two (refined) domains of MAG 1-3 , followed by SCULPTOR (ref. 63)-trimmed versions of 3P3Y and 2YD6 respectively. Lastly, the Ig3 domain was searched using the refined Ig3 of MAG from MAG 1-3 as this domain had a much higher B-factor in the MAG 1-5 crystals, likely because of the lack of any crystal packing contacts for this domain. It was important to realize that the unit cell only contained a single copy of MAG, resulting in a solvent content of 91%, which was used in PHASER to estimate the total scattering. Lysine-methylated MAG 1-5 was solved by searching for the Ig1, Ig2, Ig4, Ig5 and Ig3 from the refined MAG 1-3 and deglycosylated MAG [1][2][3][4][5] structures, in that order. Again, Ig3 was searched last because of the high B-factor. After molecular replacement, models were improved by iterative density modification by DM (CCP4) 64 , manual model building in COOT 65 and refinement with REFMAC 66 . Final refinement was performed with PHENIX 67 and validation with Molprobity 68 . In both MAG 1-5 structures, Ig3 from the higher resolution MAG 1-3 structure was used as a reference structure to restrain refinement, because of the high B-factor and the resulting poor density for that domain. Ramachandran statistics were (% Ramachandran favored/% allowed/% outliers): 93/7/0 (MAG 1-5 deglycosylated), 91/9/0 (MAG 1-5 lysine-methylated), 97/3/0 (MAG 1-3 unliganded) and 96/4/0 (MAG 1-3 ligand bound). The buried surface are for MAG dimerization was calculated by PISA 69 .
Small angle X-ray scattering. SAXS was performed at the ESRF BM29 BioSAXS beamline equipped with a 2D Pilatus 1 M detector (DECTRIS, Switzerland), operated at an energy of 12.5 keV. MAG constructs were diluted with and dialyzed against SEC buffer using a 10 kDa MWCO membrane. The concentrations were determined by ultraviolet-visible spectroscopy on a nanodrop ND-1,000 spectrophotometer. Similar concentrations were selected for all samples to allow comparison (see Table 2). SAXS data were collected at 20°C unless indicated otherwise. The data were radially averaged, normalized to the intensity of the transmitted beam and exposure time and the scattering of the solvent-blank (SEC buffer) was subtracted, following standard procedures. The curve was scaled using a BSA reference so that the I 0 represents the M m of the sample. Radiation damage was monitored by comparing curves collected from the same sample, only curves without radiation damage were merged. A single concentration was used for all measurements, no extrapolation to zero concentration was performed. Data were analysed by PRIMUS 70  wt glycosylated) were used. Sedimentation equilibrium runs were performed at 20°C and at 7,500, 14,000 and 20,000 r.p.m. Absorbance was determined at 250, 280 and 300 nm with SEC buffer as reference. Buffer density and viscosity were determined by SEDNTERP as 0.99823 g ml À 1 and 0.001002 Pa s, respectively.
In-gel digestion and LC-MS/MS. MAG 1-5 was separated by SDS-PAGE, and cut from gel for digestion with trypsin (Promega). The gel band was cut to small pieces, washed in Milli-Q water and treated with acetonitrile to shrink the gel pieces. The sample was then incubated in 1 g l À 1 1,4-dithiothreitol for 60 min at 60°C, treated with acetonitrile, alkylated with 10 g l À 1 iodoacetamide for 30 min at room temperature in the dark and subsequently washed with ammonium bicarbonate and treated with acetonitrile, twice. The gel pieces were then incubated on ice for 90 min with 30 mg l À 1 trypsin. Excess trypsin was removed, the gel pieces were covered in ammonium bicarbonate, and the samples were subsequently incubated overnight at 37°C. The digested samples were collected and remaining sample was extracted from the gel pieces by treatment with acetonitrile. The solution with the peptides was subsequently dried in a speedvac and the peptides resuspended in 10% formic acid, 5% dimethylsulfoxide in water.
Peptides were separated by reversed phase LC coupled on-line to an Orbitrap Elite for MS/MS analysis. The nano-LC consists of an Agilent 1200 series LC system equipped with a 20 mm ReproSil-Pur C18-AQ (Dr Maisch GmbH) trapping column (packed in-house, i.d., 100 mm; resin, 5 mm) and a 400 mm ReproSil-Pur C18-AQ (Dr Maisch GmbH) analytical column (packed in-house, i.d., 50 mm; resin, 3 mm) arranged in a vented-column configuration. The flow was passively split to 100 nl min À 1 . We used a standard 45 min gradient from 7-30% acetonitrile in aqueous 0.1% formic acid. All precursors were fragmented by both ETcaD and HCD. Data were searched against a custom database of recombinant protein sequences, including the MAG constructs used here, with trypsin as protease, allowing up to two missed cleavages. We used a 50 p.p.m. precursor mass window and 0.02 Da fragment mass window. The C-mannosylated peptide, with 3 þ precursor charge, eluted after 33-34 min and was identified by both HCD and ETcaD MS/MS, with matched fragment ions supporting site localization for the C-mannosylation.
Native mass spectrometry. Purified protein samples were deglycosylated with Endo-H f as for crystallization, followed by buffer exchange to 150 mM ammonium acetate pH 7.5, using Vivaspin500 10 kDa MWCO centrifugal filter units. Samples were loaded onto gold-coated borosilicate capillaries prepared in-house for nanoelectrospray ionization. Samples were analysed on a modified Orbitrap extended mass range (Thermo Fisher) for high mass ions 75 .
Surface plasmon resonance. MAG 1-5 wt and mutants cloned in-frame with an N-terminal cystatin secretion signal and a C-terminal biotin acceptor peptide and His 6 -tag (sequence AAAGSGLNDIFEAQKIEWHEGRTKHHHHHH), were biotinylated in HEK293 cells by co-transfection with E. coli BirA biotin ligase with a sub-optimal secretion signal (in a pUPE5.02 vector), using a DNA ratio of 9:1 (MAG DNA: BirA DNA, m per m)). Additional sterile biotin (100 ml of 10 mg ml À 1 Tris-buffered biotin per 4 ml HEK293 culture) was supplemented to the medium a few hours after transfection. MAG mutants were purified from the medium by Ni-NTA affinity purification. Purity was evaluated by SDS-PAGE and coomassie staining and biotinylation by a streptavidin gel-shift assay followed by a-His 6 Western blot (Supplementary Fig. 19). C-terminally biotinylated MAG proteins were spotted on a G-STREP SensEye chip (Ssens) with a Continuous Flow Microspotter (Wasatch Microfluidics) using a 8 Â 6 format. The C-terminal coupling of MAG to the surface mimics the native, membrane attached topology. SEC buffer with 0.005% tween was used as a spotting buffer and the coupling was quenched using 1 mM biotin in SEC buffer solution.
GT1b ganglioside liposomes were prepared as described previously 76 . In brief, the lipids dipalmitoyl phosphatidylcholine, dipalmitoyl phosphatidylglycerol, cholesterol and GT1b gangliosides were mixed in a molar ratio of 40.3:4.2:40.9:1.3 in a chloroform/methanol mixture (6:1, v per v). The lipid mixture was dried under vacuum on a rotary evaporator to create a thin film of lipids. Liposomes were formed by addition of 1 ml of SEC buffer per 21.7 mmol of lipid mixture and 11 freeze-thaw cycles. As a negative control, liposomes with the same lipid composition but lacking GT1b were prepared using the same protocol. Liposomes were extruded through a 100 nm membrane with a mini-extruder (Avanti Polar Lipids) at 70°C. SPR experiments, with the liposomes in the mobile phase and the MAG constructs attached to the surface, were performed on an MX96 SPRi instrument (IBIS Technologies) equipped with a CX flowcell, using the CX vesicle run protocol and an association time of 60 min at a temperature of 25°C. As a running buffer, SEC buffer without any detergent was used. Preliminary removal of co-purified ligands before the runs and regeneration after runs was performed by washes with 0.5% SDS in phosphate-buffered saline (PBS) followed by 5M NaCl. Data was zeroed and referenced using SprintX 1.11 (IBIS Technologies).
Animals. All animal use and care was in accordance with institutional guidelines and approved by the animal experimentation committee (DEC). Littermate C57BL/6 (Charles River) mice were killed by decapitation at postnatal day 1 (P1) before the brain was removed to prepare hippocampal neuronal cultures.