Structural Mechanism for Modulation of Synaptic Neuroligin-Neurexin Signaling by MDGA Proteins

Summary Neuroligin-neurexin (NL-NRX) complexes are fundamental synaptic organizers in the central nervous system. An accurate spatial and temporal control of NL-NRX signaling is crucial to balance excitatory and inhibitory neurotransmission, and perturbations are linked with neurodevelopmental and psychiatric disorders. MDGA proteins bind NLs and control their function and interaction with NRXs via unknown mechanisms. Here, we report crystal structures of MDGA1, the NL1-MDGA1 complex, and a spliced NL1 isoform. Two large, multi-domain MDGA molecules fold into rigid triangular structures, cradling a dimeric NL to prevent NRX binding. Structural analyses guided the discovery of a broad, splicing-modulated interaction network between MDGA and NL family members and helped rationalize the impact of autism-linked mutations. We demonstrate that expression levels largely determine whether MDGAs act selectively or suppress the synapse organizing function of multiple NLs. These results illustrate a potentially brain-wide regulatory mechanism for NL-NRX signaling modulation.


INTRODUCTION
Cell-surface synaptic organizing proteins play a central role in the assembly, maturation, stabilization, and plasticity of neuronal synapses (Siddiqui and Craig, 2011). Members of the presynaptic neurexin (NRX) and postsynaptic neuroligin (NL) transmembrane protein families form the axis of a signaling pathway that is crucial for the formation and function of excitatory and inhibitory synap-ses throughout the brain (S€ udhof, 2008). The NL-NRX complexes promote synaptic cell adhesion via direct extracellular interactions and recruit the molecular machinery for neurotransmitter release and reception. NLs recruit ionotropic glutamate and GABA A receptors through direct interactions or using DLG (Discs large) family or gephyrin and collybistin accessory proteins, respectively (Bemben et al., 2015). NRXs interact intracellularly with CASK and Mint PDZ domain proteins and the synaptic vesicle protein synaptotagmin; a-NRXs also functionally link to presynaptic voltage-gated Ca 2+ channels (Reissner et al., 2013).
NLs are generated from five genes in humans or four genes in mice, and further diversified by two sites of alternative splicing: spliced sequences A (SSA) and B (SSB). Mammalian NRXs show even greater diversity: over a thousand variants are generated from three genes, two promoters (a and b), and six sites of alternative splicing (SS1-6) (Schreiner et al., 2014;Ullrich et al., 1995). The extracellular region of the NLs contains a cholinesterase-like domain that forms a stable interaction with the a/b-NRX1-3 LNS6 (laminin, NRX, sex-hormone-binding globulin) domain (Araç et al., 2007;Chen et al., 2008;Fabrichny et al., 2007). NL1(+B) binds only b-NRXs (Boucard et al., 2005) and functions at glutamatergic synapses (Song et al., 1999), while NL2 binds all NRXs and functions at GABAergic synapses (Graf et al., 2004;Varoqueaux et al., 2004).
In contrast to all these positive effectors and modulators, the discovery of the Ig superfamily (IgSF) MDGA (meprin, A-5 protein, and receptor protein-tyrosine phosphatase mu [MAM] domaincontaining glycosylphosphatidylinositol anchor) proteins as negative modulators of NL is remarkable. MDGA1 was found to block the interaction of NL2 with NRX and suppress inhibitory synapse development in cultured neurons (Pettem et al., 2013), while MDGA2 blocks the interaction of NL1 and NL2 with NRX and can suppress excitatory and inhibitory synapse development (Connor et al., 2016). MDGA proteins are attached to the postsynaptic membrane via a C-terminal GPI anchor, and their large ($900 amino acids) extracellular domain consists of six immunoglobulin-like domains (Ig 1-6 ), a fibronectin type III-like (FnIII 7 ) domain, and a memprin, A5, mu (MAM 8 ) domain.
Aberrant signaling in the NL-NRX pathway is strongly linked to autism spectrum disorders (ASDs) and schizophrenia (S€ udhof, 2008). Similarly, intronic SNPs in MDGA1 are linked to schizophrenia (K€ ahler et al., 2008;Li et al., 2011), and MDGA2 lossof-function truncations were found in unrelated cases of ASD (Bucan et al., 2009). Single-allele knockout of the Mdga2 gene in mice elevated both excitatory neurotransmission and functional connectivity and produced behavioral phenotypes related to ASD (Connor et al., 2016). Mdga2 haploinsufficiency phenotypes were associated with elevated levels of NL1 and DLG family proteins and proposed to be due to diminished block of NL1-NRX signaling (Connor et al., 2016). However, based on a novel synaptic cleft tagging strategy in cell culture, another recent study proposed a role for MDGA2 selectively at inhibitory synapses and MDGA1 at excitatory synapses (Loh et al., 2016), raising controversy about the precise functions of MDGAs and revealing a need for more in-depth comprehensive analyses.
Despite the recent focus on mapping the complex molecular landscape of NL-NRX signaling modulators, a structural and mechanistic understanding of these processes is still lacking. In this study, we present the crystal structure of the near-complete MDGA1 extracellular domain and that of its prototypical complex with NL1, providing detailed insight into the structural basis of the modulation of NL-NRX signaling by MDGA proteins. We show that human MDGA1 and MDGA2 have the ability to interact with human NL1-5, thereby extending the previously proposed restricted, binary NL-MDGA interaction code (Connor et al., 2016;Lee et al., 2013;Pettem et al., 2013). Furthermore, we demonstrate that MDGA1 and MDGA2 are able to broadly block NL synaptogenic activity in a concentration-and splice insert-dependent fashion. Given the broad distribution of MDGA and NL-NRX complexes, our work provides a framework for understanding potential brain-wide modulation of NL-NRX signaling by MDGA proteins.

RESULTS
Crystal and Solution Structure of MDGA1 As a first step toward solving the structure of an NL-MDGA complex, we targeted the full-length apo MDGA1 extracellular domain for crystallization. Following an extensive screen of constructs from various species, we obtained diffraction-quality crystals and solved the structure of the complete chicken MDGA1 extracellular region (cMDGA1 ECTO ; Ig 1 -Mam 8 ; Gln19-Lys919; 79.5% sequence identity and 88.4% sequence similarity with human MDGA1 ECTO ; Figure S1) using selenomethionine single-wavelength anomalous diffraction (Se-SAD) at 3.20 Å (Figures 1A and 1B; Table S1). cMDGA1 ECTO was treated with endoglycosidase F1 (Endo F1) prior to crystallization, leaving a single N-linked N-acetylglucosamine monosaccharide on glycosylated Asn residues after enzymatic cleavage. Seven domains (Ig 1-6 to FnIII 7 ) could be unequivocally resolved in the electron density maps; however, the C-terminal MAM 8 domain was not visible and most likely highly mobile and accommodated in the solvent channels of the crystal. The cMDGA1 ECTO Ig 1-6 -FnIII 7 domains form a surprisingly compact, folded structure that is $120 Å wide, $110 Å high, and $50 Å deep, fitting comfortably within the typical height of the synaptic cleft ($20-25 nm). Its approximately triangular shape, unique among the cell-surface receptors crystallized to date, is a consequence of sharp-angled Ig 2 -Ig 3 , Ig 4 -Ig 5 , and Ig 6 -FnIII 7 inter-domain linkers that are stabilized by numerous inter-domain contacts.
The Ig 2 -Ig 3 domain contacts (341 Å 2 buried surface area [BSA]) are formed between (1) the Ig 2 b strands bA and bG and (2) the loop structure connecting Ig 2 and Ig 3 , and also Ig 3 loops BC and FG. The Ig 4 -Ig 5 domain contacts (598 Å 2 BSA) are formed between (1) the Ig 4 b strand bA and loop AB and (2) Ig 5 loops BC and FG. The Ig 6 -FnIII 7 domain contacts (396 Å 2 BSA) are formed between (1) the Ig6 b strand bA 0 and loops A 0 B and EF and (2) the loop connecting Ig 6 and FnIII 7 and FnIII 7 loops BC and FG. Finally, the Ig 1 -FnIII 7 domain contacts (395 Å 2 BSA) close the cMDGA1 ECTO triangle and are formed between (1) the Ig 1 N-terminal stretch (Gln19-Tyr22) and loop BC and (2) FnIII 7 loops AA 0 and C 0 E, and b strands bA and bB ( Figure 1C). The linear orientation of Ig 1 and Ig 2 is stabilized by a disulfide bond, distinct from the core Ig domain disulfide bonds, between Cys36 located on Ig 1 loop AA 0 and Cys222 located on Ig 2 loop FG ( Figures 1A and 1B).
In the crystal, two MDGA molecules form an unexpected intertwined dimeric arrangement with individual C-terminal ends pointing in opposite directions ( Figure S2A). Homophilic interfaces are formed between domain pairs Ig 1 -Ig 5 *, Ig 2 -Ig 2 *, and Ig 6 -FnIII 7 * (where * denotes contributions from the second MDGA monomer); their combined BSA is 2,666 Å 2 , suggesting a stable association. Interestingly, this arrangement is compatible with both a potential cis-or trans-homophilic interaction and might indicate formation of an adhesive or self-inhibitory complex ( Figure S2B). Recombinantly expressed MDGA1 targets to axons and dendrites and partially co-localizes with inhibitory and excitatory postsynaptic markers in cultured hippocampal rodent neurons (Loh et al., 2016;Pettem et al., 2013). Native MDGA1 and MDGA2 were observed in axon tracts in chicken (Fujimura et al., 2006) and zebrafish (Ingold et al., 2015), and a putative trans-homophilic interaction of MDGA2 was proposed to function in directed axonal growth (Joset et al., 2011).
To investigate the dimerization potential of the MDGA1 extracellular region in solution, we pursued multiple experimental avenues. First, we determined the cMDGA1 ECTO solution structure using small-angle X-ray scattering (SAXS) at a concentration of 30 mM. The scattering data were unambiguously incompatible with a dimeric MDGA1 molecule but were instead accurately (c 2 = 1.17) modeled as a limited ensemble of monomeric conformers with pronounced flexibility at the FnIII 7 -Mam 8 domain linkage ( Figure S2C). In accordance with our SAXS data, we determined using analytical ultracentrifugation (AUC) that human MDGA1 ECTO is monomeric at a concentration of 60 mM ( Figures  S2D and S2E). Finally, to probe whether potential MDGA1 selfassociation might instead be transient, we performed surface plasmon resonance (SPR) experiments in which wild-type (A) Schematic representation of the chicken MDGA1 (cMDGA1) domain structure. Gln19-Lys919, spanning Ig 1 -Mam 8 , was used for structure determination. Black diamonds indicate Asn residues with crystallographically confirmed N-linked glycosylation (nine positions). Open diamonds indicate Asn residues with predicted but crystallographically unconfirmed N-linked glycosylation (four positions). Orange lines connect cysteine residues engaged in disulfide bonds. (B) Crystal structure of cMDGA1 ECTO . Disulfide bridges are shown as yellow spheres. Glycan moieties visible in the electron density maps are shown in ball and stick representation. N and C termini, b strands, and selected Ig 1-2 loop structures are annotated to the structure. The MAM 8 domain was not visible in the electron density maps, probably due to a flexible FnIII 7 -MAM 8 linker.
(C) Details of the cMDGA1 ECTO Ig 1 -FnIII 7 , Ig 2 -Ig 3 , Ig 4 -Ig 5 , and Ig 6 -FnIII 7 domain contacts. Putative hydrogen bonds and hydrophilic interactions are indicated with black dashed lines. See also Figures S1 and S2. cMDGA1 ECTO was compared with a negative control mutant that contained three N-linked glycans inserted at distinct homophilic interfaces (Arg156Asn in Ig 2 , Ser502Asn in Ig 5 , and Arg680Asn in FnIII 7 ) for binding to wild-type cMDGA1 ECTO . Both cMDGA1 ECTO variants failed to interact with wild-type cMDGA1 ECTO up to a concentration of 100 mM ( Figure S2F), indicating that no homophilic cMDGA1 ECTO interactions occurred. Together, our results provide no biochemical evidence for an MDGA1 cis-or trans-homophilic dimer, and we propose that opening of the triangular cMDGA1 ECTO structure by transient disruption of the limited Ig 1 -FnIII 7 interface allowed formation of the dimeric arrangement in the crystal lattice.

Crystal Structure of an NL-MDGA Complex
We performed an extensive crystallization screening of the NL-MDGA complexes formed between MDGA1-2 ECTO and NL1-2 ECTO constructs from various species, and succeeded in generating diffraction-quality crystals and determining the structure of the Endo F1-treated complex formed between cMDGA1 ECTO and the human NL1 cholinesterase domain lacking splice inserts (hNL1 ECTO ; Gln46-Asp635; Figure S3) at 3.30 Å (Figures  2A and 2B; Table S1). The hNL1 ECTO -cMDGA1 ECTO complex has a 2:2 stoichiometry and overall dimensions of $180 Å wide, $110 Å high, and $120 Å deep. Two MDGA1 monomers flank the NL1 dimer to form a 2-fold symmetric complex. Remarkably, the overall root-mean-square deviation (RMSD) between apo and NL1-bound cMDGA1 ECTO structures is only 1.5 Å over 647 Ca atoms, underlining the stability and importance of this unusual multi-domain architecture. The NL1 and MDGA1 C termini point in the same direction and thus confirm an interaction in cis, situated on the postsynaptic membrane (Figures 2B and 2C). Each MDGA1 molecule spans the NL1 dimer using two large, separate interaction sites located on both NL1 monomers (Sites I and II) ( Figure 3A). The Ig 1-3 domains mediate all MDGA1 contacts, consistent with previous domain-deletion experiments (Pettem et al., 2013). In contrast with the NL-NRX complex (Araç et al., 2007;Chen et al., 2008;Fabrichny et al., 2007), there was no evidence for the presence of coordinated calcium atoms at either Site I or II interfaces.
The smaller Site I (859 Å 2 BSA) is formed between residues from (1) MDGA1 Ig1 b strands C, F, and G and loop CE and (2) NL1 loops Leu289-Gln307, Ile377-Asp385 (part of ''loop L1''), Gln392-Tyr398, and Phe496-Pro499 and helices a2 (4,5) and a4(6,7) ( Figure 3B). His NL1 291, Tyr NL1 292, Asp NL1 384, and Glu NL1 394 are at the core of Site I. His NL1 291 and Tyr NL1 292 make Van der Waals (VdW) contacts and form putative hydrogen (A) Schematic representation of the constructs used for co-crystallization of the hNL1(-A-B) ECTO -cMDGA1 ECTO complex. Orange lines connect cysteine residues engaged in disulfide bonds. SSA and SSB depict the position of spliced sequences A and B on NL1, respectively. The MDGA1 Mam 8 domain was included in the crystallization construct but was not observed in the electron density, similar to the free cMDGA1 ECTO structure. (B) Front, 120 rotated side, and 90 rotated bottom views of the hNL1(-A-B) ECTO -cMDGA1 ECTO complex, shown in surface (NL1) and cartoon (MDGA1) representation. Disulfide bridges are shown as yellow spheres. Glycan moieties visible in the electron density maps are shown in ball and stick representation. The C termini of MDGA1 and NL1 point in the same direction, suggesting a complex formed in cis, located on the postsynaptic membrane. (C) Schematic representation of the postsynaptic NL1-MDGA1 cis complex. See also Figure S3.
The NL1 ''Cys loop'' (part of loop Ala110-Pro132) and ''loop L1'' (part of loop Asp361-Asp385) occlude the ''gorge'' that, in AChE, leads to the enzyme active site. Interestingly, these loop structures form an integral part of the NL-MDGA interface. In this sense, MDGA resembles the snake toxin fasciculin (Fas) for binding to AChE (Bourne et al., 1995;Harel et al., 1995). There are, however, no indications that Fas might bind NL and interfere with MDGA binding.
The function of the NL Leu449-Arg450-Glu451 (LRE) adhesion motif, conserved in all NLs and located in the a3(7,8) helix ( Figures 7B and S3C), is not clear. The LRE motif was first identified in the extracellular matrix protein laminin b2, where it is involved in binding the Ca V 2.2 voltage-gated calcium channel. Furthermore, the LRE motif is present in the majority of mammalian AChEs, and besides in NL, it is also observed in the cholinesterase-like adhesion molecules neurotactin and glutactin (Johnson and Moore, 2013). Both Arg450 and Glu451 form an integral part of the NL-MDGA interface and interact with Tyr187 and Leu190, respectively, on MDGA1 loop DE Ig2 (Figures 7A and S4A), offering a first functional role for this LRE-tripeptide in NLs. (B) Overview of the NL1 secondary structure elements contacted by MDGA Ig1 to form Site I, and MDGA Ig2-3 to form Site II. (C) View of the NL1 and MDGA1 interaction interfaces, color-coded by sequence conservation in vertebrate NL1, NL2, NL3, NL4, and NL5 (1,046 total sequences), and vertebrate MDGA1 and MDGA2 (420 total sequences). (D) View of the MDGA1 interaction interface. Site I and Site II interfaces are outlined by yellow and green lines, respectively. Per residue position, equivalent residues in human MDGA1 and MDGA2 are annotated to highlight overall sequence conservation of the interaction interfaces. Star symbols (*) indicate residues for which side chain electron density was not clearly discernable. See also Figure S4.
Sequence conservation analysis indicated that both Site I and Site II interfaces are highly conserved in vertebrate MDGAs and NLs (Figures 3C, 3D, and S4B); this observation strongly points toward a common binding mode between all MDGA and NL family members.
We mapped all predicted N-glycosylation sites for human MDGA1-2 and NL1-5 (NLs lacking splice inserts) on the cMDGA1 ECTO and hNL1 ECTO structures ( Figures S4C and S4D). The MDGA1-specific N-glycan at Asn307, experimentally confirmed by identifying the corresponding N-acetylglucosamine monosaccharide in the hNL1 ECTO -cMDGA1 ECTO electron density map, is the only glycan that is proximal to the binding interface and is situated in Ig 3 at the edge of Site II. Analysis of the complex structure, however, indicated that all putative N-linked glycans can project into the solvent, thereby avoiding interference with complex formation. Proteins for subsequent biophysical and cellular experiments were expressed in HEK293T and COS-7 cells, respectively, and were not deglycosylated.
We set up an isothermal titration calorimetry (ITC) assay to investigate whether MDGA1 ECTO competes with b-NRX1 LNS6 lacking SS4 (b-NRX1 LNS6 (-4)) for binding to NL1 ECTO . Titration of b-NRX1 LNS6 (-4) into NL1 ECTO alone revealed a strong exothermic interaction and a K D of $390 nM. Application of an equimolar amount of MDGA1 ECTO to NL1 ECTO in the titration cell did not fully block the NL1 ECTO -b-NRX1 LNS6 (-4) interaction, but decreased its apparent K D (K D,app ) $12-fold to 4.76 mM. Application of a 2.5-fold molar excess of MDGA1 ECTO over NL1 ECTO was required to fully block binding of b-NRX1 LNS6 (-4) to NL1 ECTO ( Figure 4C). These results are consistent with the notion that MDGA is not an ultra-high-affinity decoy receptor, and that by varying the levels of MDGA, the level of NL-NRX complex formation can be tuned.

MDGA1 and MDGA2 Bind All NL Isoforms
We hypothesized that the interactions between human NLs and MDGAs are not limited to certain pairs of isoforms, given the high level of conservation of the Site I and Site II interface residues among human NL1-5 and MDGA1-2 ( Figures 3C and S4B). To test this, we determined the binding strengths of all pairwise NL-MDGA ectodomain interactions using SPR. We initially focused on the unspliced NL variants for these interaction studies. As a control, we measured the pairwise interactions between NL1-5 ECTO and b-NRX1 LNS6 with and without SS4 (b-NRX1 LNS6 (±4)). The reference interaction of NL1 ECTO with b-NRX1 LNS6 (-4) showed an approximately 2-fold higher equilibrium dissociation constant (K D ) than the one derived from ITC (K D of 718 ± 14 nM versus 388 ± 23 nM, respectively; Figures 4C and 5A).
We sought to validate the interaction of MDGA1 and MDGA2 with multiple NLs. To this end, we fused the rat MDGA1 and MDGA2 ectodomains to the Fc region of human IgG. MDGA1and MDGA2-Fc proteins were then used as bait to identify NLs in postnatal day 21 (P21) rat brain synaptosome extracts, using affinity chromatography coupled with mass spectrometry and bioinformatics analysis . For extraction, we used the detergent Triton X-100 at 1% w/v concentration. In two independent MDGA1-Fc pull-down experiments, we identified NL3, NL2, and NL1, ranked by spectral count ( Figure S5B; Table S2). No peptides for NLs were detected in control experiments using Fc alone or using MDGA lacking Ig 1-3 (MDGA1DIg1-3) (Table S2), demonstrating specificity in the assay. In two independent MDGA2-Fc pull-down experiments, we identified NL2 and, to a lesser extent, NL3 ( Figure S5B). In the pull-downs, no NL4 or NL5 was detected; NL4 is of very low abundance (e.g., only $3% of the total NL in mouse brain; Varoqueaux et al., 2006) and NL5 is restricted to humans. The pull-down results are consistent with our SPR data that indicated a stronger binding of NL3 to MDGA1 than to MDGA2 ( Figure 5A).

MDGA1 and MDGA2 Modulate NL-Induced Recruitment of Hippocampal Synaptic Terminals
To assess whether MDGA1 and MDGA2 are able to broadly modulate NL-NRX-induced synapse formation, we set up a cellular hemi-synapse formation assay in which COS-7 cells co-expressing full-length (FL) N-terminally myc-tagged NL1-4 (myc-NL1-4 FL ) and full-length N-terminally HA-tagged MDGA1-2 (HA-MDGA1-2 FL ) variants were co-cultured with rat hippocampal neurons (Figures 5B, 5C, S6C, and S6D). These neurons express the -SS4 and +SS4 forms of all three aand b-NRXs (a/b-NRX1-2-3) (Aoto et al., 2013). Accordingly, this assay integrates signals from multiple NRX isoforms, in contrast with our SPR or ITC assays, which only used b-NRX1(±4) as reference interactions ( Figures 4C and 5A). To test our hypothesis that by varying the expression levels of MDGA1-2, the extent of NL-NRX complex formation and hence recruitment of synaptic terminals can be influenced, we tested three different plasmid ratios of MDGA1 and MDGA2. For MDGA1, low, medium, and high plasmid ratios designate a 2.2-, 3.5-, and 5.0-fold excess of plasmid DNA over NL, respectively. For MDGA2, these ratios were chosen to be 1.5-fold higher to achieve similar surface protein levels as MDGA1 ( Figure S6A). The low ratios used here were similar to the ratios used in our previous co-culture assays of rodent MDGA1-2 with NL1 and NL2 (Connor et al., 2016;Pettem et al., 2013). Similar results were found here for human MDGA1-2 with NL1-2 (see Figures 5B and 8D, low ratio results). However, these earlier studies did not assess the effects of NL alternative splicing, varying ratios of MDGA to NL, or MDGA on NL3-4.
We observed here that MDGA1 and MDGA2 appeared to reduce the ability of all NLs to recruit presynaptic terminals, but with different potency. MDGA1 and MDGA2 both blocked NL1-induced recruitment of synaptic terminals, although a (B) COS-7 cells expressing myc-NL1-4 were co-transfected with HA-CD4 control, HA-MDGA1, or HA-MDGA2 and co-cultured with hippocampal neurons. The ability of the co-transfected cells to induce synapsin clustering was measured and normalized to the area of tau-positive axon contact. The bar graphs represent the mean of three independent experiments for low, medium, and high plasmid ratios ( Figure S6A) of human HA-MDGA1-2:myc-NL1-4 (n > 24 total cells for each condition) with the CD4:myc-NL1-4 co-transfected controls normalized to 100% to show the relative change of synapsin integrated intensity at each ratio. Significance is shown for CD4 control versus MDGAs for each NL (one-way ANOVA with Bonferroni post hoc comparison). Error bars represent the SEM. *p < 0.05, **p < 0.01, ***p < 0.001; n.s., not significant. A detailed statistical quantification can be found in Table S3. (C) Representative images of co-cultures immunostained for surface myc-NL (blue), surface HA-MDGA or CD4 control (data not shown), synapsin (red), and tau axonal marker (green). The isolated synapsin signal (white) is shown next to each color image. Scale bar, 30 mm. See also Figures S5 and S6.
higher ratio was needed to obtain this effect for MDGA1 than for MDGA2 ( Figure 5B; Table S3). Both MDGA1 and MDGA2 potently blocked NL2-induced recruitment of synaptic terminals. Thus, there was a weaker effect of MDGA1 on NL1 relative to NL2 activity in this neuron culture-based assay in comparison with similar binding seen with purified proteins in our equilibrium SPR experiments ( Figure 5A). Differential effects in the co-culture were not due to any differences in surface levels of MDGAs or NLs ( Figure S6B). We observed a stronger differential effect when evaluating NL3-induced synapse formation. Whereas MDGA1 was able to block recruitment of synaptic terminals at all ratios, MDGA2 was not. This is consistent with our SPR analysis, which derived lower responses and corresponding lower interaction affinities for the NL3-MDGA2 interaction ( Figures  5A and S5A). Finally, both MDGA1 and MDGA2 were unable to significantly block NL4-induced synapse formation, although there was a trend toward suppression; this agrees with our SPR analysis that indicated that b-NRX1(-4) binds NL4 $6fold stronger than MDGA1-2. We suggest that even higher MDGA:NL plasmid ratios would be needed to fully block NRX binding. However, these conditions were not experimentally accessible in the assay format used, which imposed limits on the total amount of plasmid DNA that can be reliably transfected.
Overall, our results confirm that MDGA1 and MDGA2 can interfere with a broad range of NL-NRX interactions to modulate presynaptic differentiation. The functional outcome will ultimately be influenced by the relative abundances of all molecular players.
Assessment of Binding of NL1 with Hevin, Thrombospondin-1, and the NMDAR Given that the interactions of thrombospondin 1 (TSP1) (Xu et al., 2010), hevin (Singh et al., 2016), and the NMDAR (Budreck et al., 2013) with NL1 are all dependent on the coupling of their respective extracellular domains, we hypothesized that MDGA might have the potential to also block binding of these proteins to NL, thereby assigning a more general inhibitory function to MDGA. To test this, we first set out to reproduce the interactions of NL1 with recombinant hevin, TSP1, and NMDAR using SPR. In our setup, secreted human hevin and TSP1 and detergent-solubilized rat NMDAR (GluN1a-GluN2B heterotetramer) (Karakas and Furukawa, 2014) were immobilized on the chip surface (Figure S7A). We found that, in contrast to the reference interaction of NL1(-A-B) ECTO with mouse a-NRX1 ECTO (-4), all three proteins failed to interact with NL1(-A-B) ECTO up to a concentration of 25 mM ( Figures S7B and S7C).
Uncoupling of MDGA and NRX Binding to NL Given that the NL-MDGA crystal structure revealed a composite Site I-II interface, whereas NL-NRX uses only Site I ( Figure 4A), we hypothesized that NRX and MDGA binding can be uncoupled, i.e., NL can be rendered insensitive for modulation by MDGA by mutating the Site II interface. We introduced four core interface mutations into the NL1 Site I interface (NL1 DSite I : His291Ala, Tyr292Ala, Asp384Ala, and Glu394Ala) and five into the Site II interface (NL1 DSite II : Asp429Ala, Phe430Ala, Ser433Ala, Asn434Ala, and Arg450Ala) ( Figure 6A). We opted to combine multiple mutations of key interface residues instead of using single-position alanine mutants to maximize our chances of obtaining a clear binding differential and cellular phenotype.
Consistent with both NL-MDGA and NL-b-NRX1 complex structures, we found using SPR that the DSite II mutant blocked MDGA1 binding but maintained binding of b-NRX1, whereas the DSite I and combined DSite I+II mutants fully abolished both b-NRX1 and MDGA1 interactions ( Figures 6B and S8A).
Using the co-culture assay, we tested the impact of the DSite I and DSite II mutations on the recruitment of synaptic terminals by full-length NL1. Consistent with our SPR analysis, introduction of the NL1 DSite I and NL1 DSite I+II mutations, but not the NL1 DSite II mutations, prevented NL-NRX-induced synapse formation (Figures 6C and 6D; Table S4). Simultaneously, coexpression at high plasmid ratio of MDGA1 or MDGA2 with NL1 carrying the DSite II mutations did not lead to diminished recruitment of synaptic terminals (Figures 6C and 6D; Table  S4). We concluded that the NL DSite II mutant selectively uncoupled NL-NRX binding and recruitment of synaptic terminals from inhibition by MDGA.

The ASD-Linked NL3 Mutation Arg451Cys Prevents Suppression of Synapse Formation by MDGA1
The well-characterized NL3 mutation Arg451Cys (R451C) leads to a number of ASD-linked phenotypes in mice (Tabuchi et al., 2007). In this knockin mouse model, R451C acts as a gain-offunction mutation by actually increasing inhibitory synaptic transmission, a result that is seemingly at odds with the severe reduction of NL3 in these mutant mice (Tabuchi et al., 2007). Indeed, complete knockout of NL3 has no such effect (Tabuchi et al., 2007). (C) COS-7 cells expressing myc-NLs were co-transfected with HA-CD4 control, HA-MDGA1, or HA-MDGA2 and co-cultured with hippocampal neurons. The ability of the co-transfected cells to induce synapsin clustering was measured and normalized to the area of tau-positive axon contact. The bar graphs represent the mean of three independent experiments for high plasmid ratios ( Figure S6A) of human HA-MDGA1-2:myc-NL1 (n > 22 total cells for each ratio). Significance is shown for CD4 control versus MDGA1-2 for each NL1 variant (one-way ANOVA with Bonferroni post hoc comparison). Error bars represent the SEM. ***p < 0.001; n.s., not significant. Mutation of Site II renders NL1(-A-B) insensitive to suppression of synapse formation by MDGA1 and MDGA2. A detailed statistical quantification can be found in Table S4. (D) Representative images of co-cultures immunostained for surface myc-NL (blue), surface HA-MDGA or CD4 control (data not shown), synapsin (red), and tau axonal marker (green). The isolated synapsin signal (white) is shown next to each color image. Scale bar, 30 mm. See also Figure S8.
Our hNL1 ECTO -cMDGA1 ECTO complex crystal structure shows that NL1 Arg450, which is equivalent to NL3 Arg451 and part of the NL1 Leu449-Arg450-Glu451 (LRE) motif (Figures 7B and S3C), is an integral part of the Site II interface ( Figure 7A). We introduced the Arg450Cys (R450C) and Arg451Cys (R451C) mutations into NL1(-A-B) ECTO and NL3(-A) ECTO , respectively. We observed diminished secretion for the mutants as compared to wild-type proteins ( Figure S8B), consistent with reported trafficking defects and protein destabilization (Chih et al., 2004;Comoletti et al., 2004;Tabuchi et al., 2007). Using SPR, we then measured the interaction of b-NRX1(±4) and MDGA1-2 with these mutant proteins and compared them to the wild-type interactions. Our measurements revealed that for both NL1 and NL3, introduction of the R450/451C mutation nearly completely abolished binding of both MDGA1 and MDGA2, while leaving the binding of b-NRX1(±4) unaffected ( Figures 7C and S8B). This is consistent with the fact that the R450/451C mutation is situated in the MDGA-specific Site II interface. In this sense, the mutation thus phenocopies our NL1 DSite II mutant ( Figure 6B).
Using the co-culture assay, we tested the impact of the R451C mutation on the recruitment of synaptic terminals by full-length NL3. Importantly, although impaired relative to wild-type NL3, the R451C mutant can traffic to the surface of transfected COS-7 cells ( Figure S8C) and rat hippocampal neurons (Figure S8D;consistent with Chih et al., 2004). Thus, for the co-culture analysis, we again selected COS-7 cells that displayed equal amounts of surface NL to ensure meaningful readout of synapse formation ( Figure S8C). Consistent with our SPR analysis, introduction of the R451C mutation had no impact on NL-NRX-induced synapse formation when compared to wild-type NL3 ( Figure 7D). Then, co-expression at low plasmid ratio of MDGA1, but not MDGA2, with NL3 wild-type led to diminished recruitment of synaptic terminals. This result closely reproduces our earlier observation ( Figure 5B). Introduction of R451C, however, prevented the diminished recruitment of synaptic terminals mediated by MDGA1 (Figures 7D and 7E; Table S5). We concluded that R451C selectively uncoupled NL3-NRX binding and recruitment of synaptic terminals from inhibition by MDGA1.
Tuning of the NL-MDGA Interaction by NL SSA and SSB Alternative splicing leads to insertion of SSA and SSB onto the NL cholinesterase scaffold. SSB is restricted to NL1, whereas distinct SSA sequences are present in NL1, NL2, and NL3. In NL1 and NL3, the two possible SSA sequences (A1 and A2) can also occur in tandem (denoted as A1A2) ( Figure S3B). Whereas NL1 mRNAs containing and lacking splice insert A are detected at similar levels at hippocampal, cortical, and cerebellar excitatory synapses, mRNA coding for NL1(+B) is more abundant than for NL1(-B) (Chih et al., 2006). Simultaneously, the insertion point for SSB in NL1 is in close proximity to the Site I interface (Koehnke et al., 2010), suggesting that presence of SSB might affect MDGA binding. These observations prompted us to investigate the effect of insertion of SSA and SSB on the NL-MDGA complex formation. First, we mapped SSA, derived from a published NL1(+A1) crystal structure (PDB: 3VKF; Tanaka et al., 2012), onto the NL1-MDGA1 (0.399 Å RMSD over 477 NL1 Ca positions) and NL1-NRX1 (0.375 Å RMSD over 453 NL1 Ca positions) structures ( Figure 8B).
Interestingly, although SSA is spatially distant from both Site I and Site II binding interfaces, it is in close proximity to the MDGA Ig 5 and Ig 6 domains ( Figure 8B). As such, SSA might have the potential to either clash with Ig 5 -Ig 6 or, conversely, provide an additional binding site for MDGA. We tested using SPR whether insertion of the distinct SSA sequences into NL1, NL2, or NL3 had an effect on the NL-MDGA or NL-NRX interactions. We were unable to detect a robust or meaningful impact of the SSA sequences on the binding strength of any NL1-3 ECTO -MDGA1-2 ECTO or NL1-3 ECTO -b-NRX1 LNS6 (±4) pair ( Figures  S9A, S9B, S10A, and S10B), suggesting that SSA possesses sufficient conformational freedom to not perturb the core NL-MDGA interaction. Accordingly, we suggest that SSA is not involved in modulating the NL-MDGA interaction.
Next, we determined the crystal structure of human NL1 containing SSB (hNL1(+B)) at 2.55 Å ( Figure 8A; Table S1). The nineresidue SSB (NRWSNSTKG), inserted between Gly295 and Leu305, was clearly visible in the electron density; the N-linked glycan at Asn300, a modulator of the NL-NRX interaction (Chih et al., 2006;Comoletti et al., 2003), was, however, not fully resolved due to conformational flexibility ( Figure 8A). Superposition of NL1(+B) and NL1-MDGA1 (0.308 Å RMSD over 434 NL1 Ca positions) or NL1-NRX1 (0.306 Å RMSD over 428 NL1 Ca positions) structures revealed that SSB is spatially immediately adjacent to both Site I interfaces ( Figure 8B). We found, using SPR, that insertion of SSB weakened the NL1-MDGA1-2 interaction $7-fold, while reducing the NL1-b-NRX1(±4) interaction less than 2-fold ( Figure 8C). We propose that this differential effect is due to the much larger molecular footprint of MDGA and the resulting close proximity of the MDGA Ig 5 and Ig 6 domains to the N-linked glycan at Asn300, suggesting that SSB reduces the NL-MDGA interaction due to steric hindrance ( Figure 8B). Indeed, removal of the N-linked glycan (Asn300Gln mutant) partially recovered the NL1-MDGA1-2 interaction affinity, whereas it had almost no effect on the NL1-b-NRX1(±4) interaction ( Figure 8C).
Using the co-culture assay, we tested the effect of the presence of SSB on the ability of MDGA1-2 to block recruitment of synaptic terminals by full-length NL1(-B) and NL1(+B). Coexpression of MDGA1-2 at low, medium, and high plasmid ratios (as previously defined; Figure S6A) led to a decreased recruitment of terminals by both NL1(-B) and NL1(+B) (Figures 8D and 8E ; Table S3); however, we found a concentration-dependent decrease for MDGA1. At the medium ratio, MDGA1 significantly blocked recruitment of terminals by NL1(-B), but not NL1(+B), consistent with the difference in binding observed in the SPR assay ( Figure 8C). Taken together, our results suggest that, despite the proximity of SSB to the Site I interface, its presence does not eliminate the ability of MDGA to block NL-NRX signaling. Rather, SSB provides a way to fine-tune the NL1-MDGA1-2 interaction at excitatory synapses.

DISCUSSION
In this work, we present the structure of the near-complete MDGA1 extracellular domain and its complex with NL1, establishing the general recognition paradigm between these synaptic organizing molecules. Simultaneously, our structural analyses (A) NL1 Arg450, equivalent to NL3 Arg451, is engaged in p-stacking interactions with Tyr MDGA1 187 and its side chain is oriented by charged interactions with Asp NL1 447 and Glu NL1 451. (B) Sequence alignment of human, mouse, and rat NL1-5. Helices a2(7,8) and a3(7,8) of Hs_NL1 are annotated above the alignment. NL residues unique to the ''core'' and ''rim'' of the NL-MDGA interface are highlighted in black and gray, respectively. The Leu-Arg-Glu (LRE) motif, conserved in all NLs and located in the a3(7,8) helix, is boxed in yellow. The equivalent NL1 Arg450 and NL3 Arg451 residues are part of the Site II interface and central to the LRE motif. Hs; Homo sapiens, Mm; Mus musculus, Rn; Rattus norvegicus. (C) Schematic representation of the SPR setup, summary of K D values, and binding isotherms for the interaction of NL1, NL1 Arg450Cys, NL3, and NL3 Arg451Cys with MDGA1-2 ECTO and b-NRX1 LNS6 (±4). (D) COS-7 cells expressing myc-NL3 wild-type or myc-NL3 Arg451Cys were co-transfected with HA-CD4 control, HA-MDGA1, or HA-MDGA2 and co-cultured with hippocampal neurons. The ability of the co-transfected cells to induce synapsin clustering was measured and normalized to the area of tau-positive axon contact. The bar graphs represent the mean of three independent experiments for the low plasmid ratio ( Figure S6A) of human HA-MDGA1-2:myc-NL3 (n > 21 total cells for each condition). Significance is shown for CD4 control versus MDGAs (one-way ANOVA with Bonferroni post hoc comparison). Error bars represent the SEM. ***p < 0.001; n.s., not significant. A detailed statistical quantification can be found in Table S5. (E) Representative images of co-cultures immunostained for surface myc-NL3 (blue), surface HA-MDGA or CD4 control (data not shown), synapsin (red), and tau axonal marker (green). The isolated synapsin signal (white) is shown next to each color image. Scale bar, 30 mm. See also Figure S8. guided the discovery of a broad splicing-modulated interaction network between all MDGA and NL isoforms that is able to block NL-NRX complex formation and modulate NL-induced recruitment of synaptic terminals.
Two large, triangular MDGA1 molecules cradle dimeric NL to shield it from interacting with NRX. We tested whether this arrangement also has the potential to negatively influence the interaction of NL with the astrocyte-secreted proteins TSP1 (Xu et al., 2010) and hevin (Singh et al., 2016) and with the NMDAR (Budreck et al., 2013). However, we failed to reproduce these interactions using SPR. Our results suggest that, at least using isolated recombinantly produced proteins and in an SPR setup with defined components and buffer conditions, these interactions are very weak, require the membrane environment, or are mediated through as-yet-unidentified auxiliary proteins or small-molecule ligands. Future studies will have to identify the exact molecular components required for these interactions.
The structure of the NL1-MDGA1 complex uncovers Site II, a hitherto unrecognized interaction site on NL that is distinct from the canonical NL-NRX Site I interface, highlighting the ability of the NL cholinesterase fold to accommodate a diverse array of ligand interaction modes. Furthermore, the NL DSite II mutant is a useful molecular tool to selectively uncouple NL-NRX complex formation from inhibition by MDGA, or from other proteins that would utilize Site II.
MDGA Ig domains 1-3 mediate all contacts with NL (Figure 3A). We speculate that MDGA might have more binding partners besides NL. Indeed, the MDGA1 MAM domain binds a receptor on axons (Fujimura et al., 2006) and enhances cell motility and adhesion to non-MDGA1-expressing cells (Díaz-López et al., 2010). The Ig domains 4-6 are reported to play a role in determining synaptic localization of MDGA1 and MDGA2 (Loh et al., 2016). Adhesive interactions of MDGA with as-yet-unidentified partners may be responsible for the MDGA-dependent aggregation of basal progenitor cells in the subventricular zone (Perez-Garcia and O'Leary, 2016), radial migration of cortical neurons (Takeuchi and O'Leary, 2006), and directed axon outgrowth (Ingold et al., 2015;Joset et al., 2011). The widespread expression of NLs (Varoqueaux et al., 2006) and NRXs (Brown et al., 2011;Gó recki et al., 1999) from early postnatal ages also raises the interesting possibility that MDGAs may func-tion to shield NLs at the stage of process outgrowth to prevent premature axon-dendrite adhesion and synaptogenesis.
Given the similar interaction affinities of MDGA1-2 and NRX with NL, the balance between NL-NRX and NL-MDGA complex formation will be determined by their relative abundances and binding availability at each synapse in vivo. The net effect of MDGA on synaptic NL-NRX signaling may be influenced by the presence of other protein partners of MDGA, NRX (LRRTMs, calsyntenin 3, dystroglycan, latrophilin 1, cerebellins, and C1q-like proteins), and NL (hevin, thrombospondin, and NMDARs). The complexity of NL-NRX signaling is compounded even further by the existence of postsynaptic cis NL-NRX silencing complexes (Taniguchi et al., 2007) and by the recent report of MDGA-like functions for g-protocadherins (Molumby et al., 2017).
The capacity of NLs to form heterodimers (Poulopoulos et al., 2012) will differentially affect MDGA and NRX binding since the MDGA interface spans both NL monomers, whereas the NRX interface does not. For example, NL1/3, the most prevalent NL heterodimer located at excitatory synapses (Budreck and Scheiffele, 2007;Poulopoulos et al., 2012), would harbor an asymmetric set of Site I-II interfaces: Site II on one side of the dimer will come from NL3, while Site I will be donated by NL1. At the other side of the dimer, this will be inverted. Since NL3 interacts $10-fold more weakly with MDGA than NL1(-B) (Figure 5A), a composite interface will likely lead to an intermediate strength binding event. Insertion of SSB into NL1 near Site I, however, brings the affinity of NL1 for MDGA in the range of that of NL3 ( Figure 8C).
The direct interaction affinities with NL1 and NL2 do not seem to account for selectivity of MDGAs to suppress excitatory or inhibitory synapses. Consistent with the role of MDGA2 to suppress excitatory synapses in vivo (Connor et al., 2016), MDGA2, but not MDGA1, suppressed the synaptogenic activity of the major NL at excitatory synapses, NL1(+B), in co-culture experiments at low-medium ratios ( Figure 8D). Yet MDGA2 showed $12-fold and MDGA1 $6-fold greater affinity for the major NL at inhibitory synapses, NL2, than for NL1(+B) (Figures 5A  and 8C). Factors other than direct MDGA-NL1-2 binding affinities that may contribute include differential glycosylation, although we could find no indication for such ( Figures S4C and  S4D); additional interacting proteins; or differential cell-type (D) COS-7 cells expressing myc-NLs were co-transfected with HA-CD4 control, HA-MDGA1, or HA-MDGA2 and co-cultured with hippocampal neurons. The ability of the co-transfected cells to induce synapsin clustering was measured and normalized to the area of tau-positive axon contact. The bar graphs represent the mean of three independent experiments for low, medium, and high plasmid ratios ( Figure S6A) of human HA-MDGA1-2:myc-NL1(-A ± B) (n > 24 total cells for each ratio) with the CD4:myc-NL1(-A ± B) co-transfected controls normalized to 100% to show the relative change of synapsin integrated intensity at each ratio. Significance is shown for CD4 control versus MDGAs for each NL1 (one-way ANOVA with Bonferroni post hoc comparison). Error bars represent the SEM. *p < 0.05, **p < 0.01, ***p < 0.001; n.s., not significant. A detailed statistical quantification can be found in Table S3. (E) Representative images of co-cultures immunostained for surface myc-NL (blue), surface HA-MDGA or CD4 control (data not shown), synapsin (red), and tau axonal marker (green). The isolated synapsin signal (white) is shown next to each color image. Scale bar, 30 mm. See also Figures S9-S11. expression and subcellular targeting in the brain. As summarized in the introduction, there are conflicting reports on the roles of MDGAs at excitatory versus inhibitory synapses, perhaps related to the use of different model systems, reinforcing the need to consider the native abundance of each molecular player. The newly discovered interaction of MDGAs with NL3 and NL4, particularly the strong association of MDGA1 with NL3 in the pull-down assay and functional modulation of NL3 by MDGA1 in co-culture, may help in better understanding the roles of MDGAs in specific circuits in vivo.
In the rat and mouse brain, MDGA1 and MDGA2 are widely expressed by neuronal populations in both the central and peripheral nervous systems. These include neurons of the basilar pons, inferior olivary nucleus, cerebellum, cerebral cortex, olfactory bulb, spinal cord, dorsal root and trigeminal ganglia, and hippocampus (Connor et al., 2016;Lee et al., 2013;Litwack et al., 2004;Takeuchi et al., 2007). There are regional differences: for example, MDGA1 is more abundant in superficial cortical layers and MDGA2 in deep layers. NL and NRX are also very widely expressed in the mouse brain, such that most neurons likely express NL1-4 and NRX1-3 at varying levels (Hoon et al., 2011;Ullrich et al., 1995;Varoqueaux et al., 2006). We propose that the structural mechanism we described here will be representative for the full range of CNS synapses at which NL, NRX, and MDGA family members are present. Through NL2 and NL4, the range of synapses modulated by MDGA is likely to include glycinergic synapses, not just GABAergic and glutamatergic synapses as shown previously. The differential affinities of specific MDGA and NL isoforms as well as isoform selective interactions of NL with NRX, interactions with other partners regulating bioavailability, and cell-type expression patterns of all molecular players will serve to fine-tune MDGA modulation of synapse development and function.
An important finding of this study is the discovery that both MDGAs interact with and regulate NL3 and NL4. This is of particular interest since rare mutations in NLs, particularly NL3 and NL4, have been associated with ASD and schizophrenia in human genetic studies (S€ udhof, 2008; Simons Foundation Autism Research Initiative database, https://gene.sfari.org). Interestingly, two mutations in the MDGA interaction-selective Site II of NL3 have been reported in patients with ASD: Arg451Cys (R451C; corresponding to NL1 residue Arg450 and part of the Leu449-Arg450-Glu451 LRE motif) and Gly426Ser (G426S; corresponding to NL1 residue Ala425) (Jamain et al., 2003;Xu et al., 2014) (Figures 7A and 7B). This raises the possibility that selective modulation of MDGA binding to NLs in patients carrying mutations in Site II could contribute to the development of ASD. R451C was characterized as an NL3 gain-of-function mutation in mice, leading to both increased inhibitory synaptic transmission in the somatosensory cortex (Tabuchi et al., 2007) and increased excitatory synaptic transmission in the hippocampus (Etherton et al., 2011), despite resulting in trafficking defects and protein destabilization (Chih et al., 2004;Comoletti et al., 2004). Nonetheless, we observed surface expression of the mutant in both transfected COS-7 cells ( Figure S8C) and rat hippocampal neurons ( Figure S8D). The latter observation agrees with a report showing cell-surface expression of NL3 R451C in a subset of transfected hippocampal neurons with high expres-sion level (Chih et al., 2004). This also led to an increase in the number of contacting presynaptic terminals, suggesting that the NL3 R451C that trafficked to the surface is functional (Chih et al., 2004). Importantly, we found that, similarly to the NL DSite II mutant, the NL3 R451C mutation selectively uncoupled NL3-NRX binding and recruitment of synaptic terminals from inhibition by MDGA1 ( Figure 7D), suggesting that the R451C gainof-function phenotype is achieved by preventing the inhibition of NL3 by MDGA1, thereby leading to disruption of the overall balance of excitatory/inhibitory (E/I) synaptic transmission.
An E/I imbalance surpassing the capacity of neuronal populations and circuits to regulate synaptic homeostasis is a proposed hallmark of ASD (Nelson and Valakh, 2015;Rubenstein and Merzenich, 2003). Disruptions in the regulatory NL-MDGA network we report here contribute to ASD based on human genetics (Bucan et al., 2009;K€ ahler et al., 2008;Li et al., 2011;S€ udhof, 2008) and can generate such an E/I imbalance in animal models (e.g., Connor et al., 2016;Tabuchi et al., 2007). Our findings considerably broadened this interaction network beyond that previously envisioned. Moreover, our structural studies constitute an essential guide toward the generation of directed therapies targeting these gene products to restore E/I balance.

STAR+METHODS
Detailed methods are provided in the online version of this paper and include the following:

Expression and purification of recombinant NMDA receptor
The rat GluN1a-GluN2B heterotetrameric NMDA receptor (NMDAR) was expressed and purified as previously described (Karakas and Furukawa, 2014), with the exception that the OneStrep tag was not cleaved. The final purification buffer was 200 mM NaCl, 20 mM HEPES pH 7.4, 10 mM Glycine, 10 mM Glutamate, 0.0025% LMNG.
Gene splicing and site-directed mutagenesis A multiple-step overlap-extension PCR (Pyrobest Polymerase, Takara Bio) was used for site-directed mutagenesis, construction of chimeric protein constructs and introduction or deletion of splice inserts (Heckman and Pease, 2007); the resulting PCR products were cloned into the pHLsec-His6, pHLsec-Avitag3, or derived vectors .

Protein crystallization
Crystallization trials, using 100 nL protein solution plus 100 nL reservoir solution in sitting drop vapor diffusion format, were set up in 96-well Greiner plates using a Cartesian Technologies robot (Walter et al., 2005). Purified chicken MDGA1 ECTO (cMDGA1 ECTO ; Gln19-Lys919), containing the Arg120Lys mutation, concentrated to 5.0 g/L and treated with endoglycosidase F1 (Endo F1; 1:100 w/w) for 30 min at 294K immediately prior to dispensing the crystallization drops, crystallized in 0.1M HEPES pH7.5, 4% w/v polyethylene glycol 8000. The Arg120Lys mutation was introduced into cMDGA1 ECTO to bring the sequence in line with rat, mouse and human isoforms ( Figure S1B).
Crystals of cMDGA1 ECTO grown in this condition were fragmented, and the obtained seed stock (Walter et al., 2008) was used as an additive during crystallization trials of selenomethionine-(SeMet) labeled cMDGA1 ECTO . Matrix screens were performed using precipitant concentration and seed stock dilution as variables. SeMet-labeled cMDGA1 ECTO , concentrated to 5.0 g/L, crystallized in 0.1M HEPES pH7.5, 3% w/v polyethylene glycol 8000, using a 32-fold diluted native cMDGA1 ECTO seed stock dispensed in 20 nL drops. Crystals were cryoprotected using reservoir solution containing 20% (v/v) PEG200.
Crystallographic data collection and structure determination Diffraction data for cMDGA1 ECTO were collected at Diamond Light Source (DLS) beamline I03 to a nominal resolution of 3.20 Å in space group (SG) P2 1 2 1 2 1 . X-ray fluorescence wavelength scans were performed to experimentally determine the Selenium absorption K-edge peak. The cMDGA1 ECTO structure was determined using Single Anomalous Diffraction (SAD); the heavy-atom Selenium substructure was solved using SHELXD (Schneider and Sheldrick, 2002) at 3.70 Å , and phase determination, phase extension and density modification was performed using PHENIX Autosol (Terwilliger et al., 2009). Automated model building programs failed to reliably place stretches of b strand, necessitating manual model building of the complete structure.
Diffraction data for hNL1(-A+B) ECTO were collected at Diamond Light Source (DLS) beamline I24 to a nominal resolution of 2.55 Å in SG P22 1 2 1 . The structure was determined by molecular replacement using the program Phaser (McCoy et al., 2007), and using the mouse NL1 (PDB: 3BIX) crystal structure as search model.
Diffraction data for the hNL1(-A-B) ECTO -cMDGA1 ECTO complex were collected at Diamond Light Source (DLS) beamline I04-1 to a nominal resolution of 3.30 Å in SG P2 1 2 1 2. The structure was determined by molecular replacement using the program Phaser (McCoy et al., 2007), employing the refined hNL1(-A+B) ECTO (in which the spliced sequence B was excised from the molecular model) and cMDGA1 ECTO crystal structures we determined here as search models.
All data were indexed, integrated, and scaled using the automated XIA2 expert system (Winter et al., 2013), using the Labelit (Sauter et al., 2004), POINTLESS and AIMLESS (Evans, 2006), and XDS (Kabsch, 2010 programs. Crystallographic data collection and refinement statistics are presented in Table S1.

Crystallographic refinement and model analysis
Maximum-likelihood refinement of cMDGA1 ECTO , hNL1(-A+B) ECTO and the hNL1(-A-B) ECTO -cMDGA1 ECTO complex was initially performed with Refmac using ''jelly body '' restraints (Murshudov et al., 2011), and finally with the PHENIX suite (Adams et al., 2010), with automated X-ray and atomic displacement parameter (ADP) weight optimization applied throughout, and torsion angle non-crystallographic symmetry (NCS) and high-resolution reference structure restraints applied where suitable. All manual model building was performed using Coot (Emsley et al., 2010). Structure validation was performed with the PHENIX program suite using MolProbity routines (Adams et al., 2010;Chen et al., 2010).

Sequence alignments and conservation analysis
Mining of protein sequence databases was performed using the Delta-Blast program (Altschul et al., 1990). Sequence lists were manually curated and sequences were aligned using the program MUSCLE (Edgar, 2004). Sequence conservation scores for individual residue positions of NL1, À2, À3, À4, and À5 (1046 total unique sequences) and MDGA1 and À2 (420 total unique sequences) homologs were assigned to NL1 and MDGA1 structural templates extracted from the hNL1(-A-B) ECTO -cMDGA1 ECTO complex, respectively, using the ConSurf web server (Ashkenazy et al., 2010). Sequence alignments were visualized using the program ALINE (Bond and Sch€ uttelkopf, 2009).
Small-angle X-ray scattering (SAXS) Purified cMDGA1 ECTO (Gln19-Lys919; with Arg120Lys mutation) was treated with endoglycosidase F1 (Endo F1; 1:100 w/w) for 12 hr at 294K and re-purified using SEC in 10 mM HEPES pH 7.50, 150 mM NaCl. SAXS data were collected at beamline BM29 of the European Synchrotron Radiation Facility (ESRF, Grenoble, France) (Pernot et al., 2013) at 293 K within a momentum transfer (q) range of 0.01 Å À1 < q < 0.45 Å À1 , where q = 4psin(q)/l, and 2q is the scattering angle. The X-ray wavelength was 0.9950 Å , and data were collected on a Pilatus 1M detector. cMDGA1 ECTO was measured at concentrations of 1.50 and 3.36 g/L in 10 mM HEPES pH 7.50, 150mM NaCl. Data reduction and calculation of invariants was carried out using standard procedures implemented in the ATSAS (Petoukhov et al., 2012) and ScÅ tter (Rambo and Tainer, 2013) suites. A merged dataset was obtained by merging the low-angle part of the low-concentration dataset with the high-angle part of the high-concentration dataset.
A molecular model for the C-terminal Mam 8 domain was generated by homology modeling starting from the crystal structure of the N-terminal RPTPmu MAM domain (PDB: 2C9A, UniProt: P28827) (Aricescu et al., 2006a) using the SWISS-MODEL server (Biasini et al., 2014). This model was concatenated with the cMDGA1 ECTO crystal structure, and manually placed near the C terminus of the FnIII 7 domain. Missing side chains, loops, and C-terminal His6-tag were added to the resulting assembled model using the MODELER (Webb and Sali, 2014) ''Model/Refine Loops'' routine as implemented in Chimera (Pettersen et al., 2004).
Coarse-grained molecular dynamics (MD) simulations were performed using the program Allosmod (Weinkam et al., 2012). Five independent runs were performed, each consisting of 30 independent trajectories generating 100 models. From this total pool of 15,000 models, automated selection of the minimal set of models that best described the scattering data was performed with the program MES (Hammel, 2012), and calculation and fitting of scattering patterns were performed with the program FoXS (Schneidman-Duhovny et al., 2013). This whole procedure was automated with the AllosMod-FoXS web server (Guttman et al., 2013). The MDGA1 solution structure was accurately (c 2 = 1.17) modeled as a five-membered ensemble of monomeric conformers with pronounced flexibility at the FnIII 7 -Mam 8 domain linkage.
Surface plasmon resonance (SPR) with soluble proteins cDNA for the immobilized proteins was cloned into the pHLsec-Avitag3 vector , resulting in proteins carrying a C-terminal biotin ligase (BirA) recognition sequence (Avitag). Constructs were co-transfected with pDisplay-BirA-ER (Addgene plasmid 20856; coding for an ER-resident biotin ligase) (Howarth et al., 2008) for in vivo biotinylation in HEK293T cells in small-scale 6-or 12-well plates in a 3:1 pHLsec:pDisplay stoichiometric ratio. A final concentration of 100 mM D-biotin was maintained in the expression medium to ensure near-complete biotinylation of the recognition sequence. After 48 hr of expression, conditioned medium was collected and dialysed against 10 mM Tris pH 7.4, 150 mM sodium chloride, 3 mM calcium chloride and 0.005% (v/v) Tween-20 (TBS-CT). SPR experiments were performed on a Biacore T200 machine (GE Healthcare) operated at a data collection frequency of 10 Hz; i.e. a temporal resolution of 0.1 s. Streptavidin (Sigma-Aldrich) was chemically coupled via amine coupling chemistry onto CM5 chips to a response unit (RU) level of 5000 RU. Then, biotinylated proteins were captured to the desired RU level. In each instance, for every two analyte binding cycles, a buffer injection was performed, allowing for double referencing of the binding responses (Myszka, 1999).
Due to (i) sample consumption associated with equilibrium affinity experiments of high-nanomolar to low-micromolar interactions and (ii) the limited production yield of MDGA1 and À2 proteins, we prioritized testing the full matrix of NL-MDGA isoform interactions over performing replicate experiments of only a selected number of interactions. Interaction of chicken MDGA1 ECTO with chicken MDGA1 ECTO and MDGA1 ECTO GLYCAN WEDGE cMDGA1 ECTO and cMDGA1 ECTO GLYCAN WEDGE (triple glycan wedge (GW) mutant; Arg680Asn-Ser502Asn-Arg156Asn) variants were immobilized at a level of 2000 RU to maximize the likelihood of detecting a potentially weak binding event. SPR running buffer composition was TBS-CT supplemented with 1.0 g/L bovine serum albumin (BSA; yielding TBS-CTB buffer) as passivating agent to prevent binding to the carboxymethyldextran-based SPR chips. MDGA1 ECTO was prepared by SEC in TBS-CT. BSA was added to the concentrated stock solutions to a final concentration of 1.0 g/L. Injection of 18 concentrations of cMDGA1 ECTO prepared in a two-fold dilution series from a 100 mM stock concentration was performed in order of increasing concentration. Each sample was injected for 150 s at a flow rate of 25 mL/min, followed by a 180 s dissociation phase. No self-association binding event could be detected. each protein to have two peptide matches and each peptide to have at least 1 tryptic terminus and an overall protein false discovery rate (FDR) < 1.2% for each dataset. Proteins shown in Figure S5B and Table S2 are the complete set of proteins found in both MDGA1-or MDGA2-Fc purifications after removing background proteins identified in Fc negative control purifications. Only proteins identified with two or more spectral counts were included in the analysis.

DATA AND SOFTWARE AVAILABILITY
The accession number for the crystal structure of human NL1(-A+B) ECTO reported in this paper is PDB: 5OJK. The accession number for the crystal structure of chicken MDGA1 ECTO reported in this paper is PDB: 5OJ2. The accession number for the crystal structure of the complex between human NL1(-A-B) ECTO and chicken MDGA1 ECTO reported in this paper is PDB: 5OJ6. (B) Schematic representation of the potential MDGA1 homophilic trans-dimer or cis-dimer, GPIanchored to the pre-and/or postsynaptic membranes.
(C) Solution structure of cMDGA1ECTO. Experimental scattering curves (black) and calculated scattering patterns (colored) are shown to a maximal momentum transfer of q = 0.40 Å -1 .
Individual data:fit pairs are displaced along an arbitrary y axis to allow for better visualization.
Bottom curve: cMDGA1 crystallographic dimer (red). Second curve from bottom: extracted cMDGA1ECTO crystallographic monomer (purple). Second curve from top: best single monomeric model (orange). Top curve: best five-membered minimal ensemble (green). Atomic models, corresponding to the individual curves, are shown in ribbon and surface representation.
The inset shows the experimental (black line and grey shade) and calculated (colored) pairwise distance distribution (P(r)) functions and derived maximum intra-particle distance (DMAX) values.   residues for which side chain electron density was not clearly discernable. The PheNL1430-PheMDGA1154 π-π sandwich stacking interaction, as well as ArgNL1450 and GluNL1451 of the NL1 LRE motif, are highlighted with shaded ovals.
(B) View of the NL1 interaction interface. Site I and Site II interfaces are outlined by yellow and green lines, respectively. Per residue position, equivalent residues in human NL1, -2, -3, -4(X) and -5(=4(Y)) are annotated to highlight overall sequence conservation of the interaction interfaces. Star symbols (*) indicate residues for which side chain electron density was not clearly discernable.
(C) N-linked glycosylation sites common to human MDGA1 and -2 (pink), or unique to human MDGA1 (red) or human MDGA2 (blue) are annotated onto the cMDGA1ECTO structure. Site I and Site II interfaces are shown in surface representation. The hMDGA1-specific Asn307 (N307) is proximal to the edge of the Site II interface.
(D) N-linked glycosylation sites common to human NL1-2-3-4-5 (pink), or unique to human NL1 (red) or human NL2 (blue) are annotated onto the hNL1(-A-B)ECTO structure. Site I and Site II interfaces are shown in surface representation as in Figure 3A.  Figure 5A. (B) The graphs show the summed spectra count (spec count) for all surface proteins identified in two independent MDGA1-Fc or MDGA2-Fc pulldown experiments from rat brain synaptosome extracts, each compared to two negative control Fc experiments. NL1-3 are the main surface proteins specifically identified by MDGA1-Fc. MDGA2-Fc bait protein identifies NL2 and NL3 as interactors. in each co-culture experiment. For each condition (low ratio MDGA1, medium ratio MDGA1, high ratio MDGA1, low ratio MDGA2, medium ratio MDGA2, and high ratio MDGA2) the mean myc signal (for surface NLs) and mean HA signal (for surface MDGAs) was measured and cells that did not have similar levels of each protein were excluded from analysis. One-way ANOVA with Bonferonni post hoc comparison was used to determine statistical significance.

(C)
Examples are shown for all the channels imaged for myc-NL1(-A+B) co-transfected with HA-CD4 or HA-MDGA1, and for the HA-MDGA1 only (no NL) negative control. Scale bar is 30 µm.

(D)
The bar graph shows the hemi-synapse formation assay data for MDGA1 and MDGA2 at medium ratio without normalization. Each NL co-transfected with CD4 shows a different baseline level of synapsin clustering, with NL2 being the most potent and NL4 being the least potent at hemi-synapse induction. Error bars represent the SEM. *; p < 0.05, ***; p < 0.001.  (C) Comparable levels of surface NL3 and MDGAs in the hemi-synapse formation assay. COS-7 cells chosen for analysis displayed similar levels of myc-NL3 and HA-MDGA1-2 in each coculture experiment. The mean myc signal (for surface NL3s) and mean HA signal (for surface MDGAs) was measured and cells that did not have similar levels of each protein were excluded from analysis. One-way ANOVA with Bonferonni post hoc comparison was used to determine statistical significance. Error bars represent the SEM.     Figure S5.
Mass spectrometry summary data file for MDGA1 (two independent experiments), MDGA2 (two independent experiments), and MDGA1ΔIg1-3 (one experiment). Column headers containing spectra counts are highlighted in bold. Protein descriptions are listed in the far-right column. Data are sorted by MDGA1-Fc spectra count in descending order.